Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
An Introduction
Klaus Schmitt
Department of Mathematics
University of Utah
Russell C. Thompson
Department of Mathematics and Statistics
Utah State University
November 11, 2004
ii
Copyright c _1998 by K. Schmitt and R. Thompson
iii
Preface
The subject of Diﬀerential Equations is a well established part of mathe
matics and its systematic development goes back to the early days of the de
velopment of Calculus. Many recent advances in mathematics, paralleled by
a renewed and ﬂourishing interaction between mathematics, the sciences, and
engineering, have again shown that many phenomena in the applied sciences,
modelled by diﬀerential equations will yield some mathematical explanation of
these phenomena (at least in some approximate sense).
The intent of this set of notes is to present several of the important existence
theorems for solutions of various types of problems associated with diﬀerential
equations and provide qualitative and quantitative descriptions of solutions. At
the same time, we develop methods of analysis which may be applied to carry
out the above and which have applications in many other areas of mathematics,
as well.
As methods and theories are developed, we shall also pay particular attention
to illustrate how these ﬁndings may be used and shall throughout consider
examples from areas where the theory may be applied.
As diﬀerential equations are equations which involve functions and their
derivatives as unknowns, we shall adopt throughout the view that diﬀeren
tial equations are equations in spaces of functions. We therefore shall, as we
progress, develop existence theories for equations deﬁned in various types of
function spaces, which usually will be function spaces which are in some sense
natural for the given problem.
iv
Table of Contents
I Nonlinear Analysis 1
Chapter I. Analysis In Banach Spaces 3
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Banach Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3 Diﬀerentiability, Taylor’s Theorem . . . . . . . . . . . . . . . . . 8
4 Some Special Mappings . . . . . . . . . . . . . . . . . . . . . . . 11
5 Inverse Function Theorems . . . . . . . . . . . . . . . . . . . . . 20
6 The Dugundji Extension Theorem . . . . . . . . . . . . . . . . . 22
7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Chapter II. The Method of LyapunovSchmidt 27
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2 Splitting Equations . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3 Bifurcation at a Simple Eigenvalue . . . . . . . . . . . . . . . . . 30
Chapter III. Degree Theory 33
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2 Deﬁnition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3 Properties of the Brouwer Degree . . . . . . . . . . . . . . . . . . 38
4 Completely Continuous Perturbations . . . . . . . . . . . . . . . 42
5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Chapter IV. Global Solution Theorems 49
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2 Continuation Principle . . . . . . . . . . . . . . . . . . . . . . . . 49
3 A Globalization of the Implicit Function Theorem . . . . . . . . 52
4 The Theorem of KreinRutman . . . . . . . . . . . . . . . . . . . 54
5 Global Bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
II Ordinary Diﬀerential Equations 63
Chapter V. Existence and Uniqueness Theorems 65
v
vi TABLE OF CONTENTS
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2 The PicardLindel¨of Theorem . . . . . . . . . . . . . . . . . . . . 66
3 The CauchyPeano Theorem . . . . . . . . . . . . . . . . . . . . . 67
4 Extension Theorems . . . . . . . . . . . . . . . . . . . . . . . . . 69
5 Dependence upon Initial Conditions . . . . . . . . . . . . . . . . 72
6 Diﬀerential Inequalities . . . . . . . . . . . . . . . . . . . . . . . 74
7 Uniqueness Theorems . . . . . . . . . . . . . . . . . . . . . . . . 78
8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Chapter VI. Linear Ordinary Diﬀerential Equations 81
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3 Constant Coeﬃcient Systems . . . . . . . . . . . . . . . . . . . . 83
4 Floquet Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Chapter VII. Periodic Solutions 91
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3 Perturbations of Nonresonant Equations . . . . . . . . . . . . . . 92
4 Resonant Equations . . . . . . . . . . . . . . . . . . . . . . . . . 94
5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
Chapter VIII. Stability Theory 103
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
2 Stability Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3 Stability of Linear Equations . . . . . . . . . . . . . . . . . . . . 105
4 Stability of Nonlinear Equations . . . . . . . . . . . . . . . . . . 108
5 Lyapunov Stability . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
Chapter IX. Invariant Sets 123
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
2 Orbits and Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3 Invariant Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4 Limit Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5 Two Dimensional Systems . . . . . . . . . . . . . . . . . . . . . . 129
6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Chapter X. Hopf Bifurcation 133
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
2 A Hopf Bifurcation Theorem . . . . . . . . . . . . . . . . . . . . 133
TABLE OF CONTENTS vii
Chapter XI. SturmLiouville Boundary Value Problems 139
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
2 Linear Boundary Value Problems . . . . . . . . . . . . . . . . . . 139
3 Completeness of Eigenfunctions . . . . . . . . . . . . . . . . . . . 142
4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
Bibliography 147
Index 149
viii TABLE OF CONTENTS
Part I
Nonlinear Analysis
1
Chapter I
Analysis In Banach Spaces
1 Introduction
This chapter is devoted to developing some tools from Banach space val
ued function theory which will be needed in the following chapters. We ﬁrst
deﬁne the concept of a Banach space and introduce a number of examples of
such which will be used later. We then discuss the notion of diﬀerentiability of
Banach–space valued functions and state an inﬁnite dimensional version of Tay
lor’s theorem. As we shall see, a crucial result is the implicit function theorem
in Banach spaces, a version of this important result, suitable for our purposes
is stated and proved. As a consequence we derive the Inverse Function theorem
in Banach spaces and close this chapter with an extension theorem for func
tions deﬁned on proper subsets of the domain space (the Dugundji extension
theorem).
In this chapter we shall mainly be concerned with results for not necessarily
linear functions; results about linear operators which are needed in these notes
will be quoted as needed.
2 Banach Spaces
Let E be a real (or complex) vector space which is equipped with a norm
 , i.e. a function   : E →R
+
having the properties:
i) u ≥ 0, for every u ∈ E,
ii) u = 0 is equivalent to u = 0 ∈ E,
iii) λu = [λ[u, for every scalar λ and every u ∈ E,
iv) u +v ≤ u +v, for all u, v, ∈ E (triangle inequality).
A norm   deﬁnes a metric d : E E → R
+
by d(u, v) = u − v and
(E,  ) or simply E (if it is understood which norm is being used) is called a
3
4
Banach space if the metric space (E, d), d deﬁned as above, is complete (i.e. all
Cauchy sequences have limits in E).
If E is a real (or complex) vector space which is equipped with an inner product,
i.e. a mapping
¸, ) : E E →R (or C (the complex numbers))
satisfying
i) ¸u, v) = ¸v, u), u, v ∈ E
ii) ¸u +v, w) = ¸u, w) +¸v, w), u, v, w ∈ E
iii) ¸λu, v) = λ¸u, v), λ ∈ C, u, v, ∈ E
iv) ¸u, u) ≥ 0, u ∈ E, and ¸u, u) = 0 if and only if u = 0,
then E is a normed space with the norm deﬁned by
u =
_
¸u, u), u ∈ E.
If E is complete with respect to this norm, then E is called a Hilbert space.
An inner product is a special case of what is known as a conjugate linear
form, i.e. a mapping b : E E → C having the properties (i)–(iv) above (with
¸, ) replaced by b(, )); in case E is a real vector space, then b is called a bilinear
form.
The following collection of spaces are examples of Banach spaces. They will
frequently be employed in the applications presented later. The veriﬁcation that
the spaces deﬁned are Banach spaces may be found in the standard literature
on analysis.
2.1 Spaces of continuous functions
Let Ω be an open subset of R
n
, deﬁne
C
0
(Ω, R
m
) = ¦f : Ω →R
m
such that f is continuous on Ω¦.
Let
f
0
= sup
x∈Ω
[f(x)[, (1)
where [ [ is a norm in R
m
.
Since the uniform limit of a sequence of continuous functions is again con
tinuous, it follows that the space
E = ¦f ∈ C
0
(Ω, R
m
) : f
0
< ∞¦
is a Banach space.
If Ω is as above and Ω
′
is an open set with
¯
Ω ⊂ Ω
′
, we let C
0
(
¯
Ω, R
m
) = ¦the
restriction to
¯
Ω of f ∈ C
0
(Ω
′
, R
m
)¦. If Ω is bounded and f ∈ C
0
(
¯
Ω, R
m
),
then f
0
< +∞. Hence C
0
(
¯
Ω, R
m
) is a Banach space.
2. BANACH SPACES 5
2.2 Spaces of diﬀerentiable functions
Let Ω be an open subset of R
n
. Let β = (i
1
, , i
n
) be a multiindex, i.e.
i
k
∈ Z (the nonnegative integers), 1 ≤ k ≤ n. We let [β[ =
n
k=1
i
k
. Let
f : Ω →R
m
, then the partial derivative of f of order β, D
β
f(x), is given by
D
β
f(x) =
∂
β
f(x)
∂
i1
x
1
∂
in
x
n
,
where x = (x
1
, , x
n
). Deﬁne C
j
(Ω, R
m
) = ¦f : Ω → R
m
such that D
β
f is
continuous for all β, [β[ ≤ j¦.
Let
f
j
=
j
k=0
max
β≤k
D
β
f
0
. (2)
Then, using further convergence results for families of diﬀerentiable functions it
follows that the space
E = ¦f ∈ C
j
(Ω, R
m
) : f
j
< +∞¦
is a Banach space.
The space C
j
(
¯
Ω, R
m
) is deﬁned in a manner similar to the space C
0
(
¯
Ω, R
m
)
and if Ω is bounded C
j
(
¯
Ω, R
m
) is a Banach space.
2.3 H¨older spaces
Let Ω be an open set in R
n
. A function f : Ω → R
m
is called H¨ older
continuous with exponent α, 0 < α ≤ 1, at a point x ∈ Ω, if
sup
y=x
[f(x) −f(y)[
[x −y[
α
< ∞,
and H¨ older continuous with exponent α, 0 < α ≤ 1, on Ω if it is H¨ older contin
uous with the same exponent α at every x ∈ Ω. For such f we deﬁne
H
α
Ω
(f) = sup
x=y
x,y∈Ω
[f(x) −f(y)[
[x −y[
α
. (3)
If f ∈ C
j
(Ω, R
m
) and D
β
f, [β[ = j, is H¨ older continuous with exponent α on
Ω, we say f ∈ C
j,α
(Ω, R
m
). Let
f
j,α
= f
j
+ max
β=j
H
α
Ω
(D
β
f),
then the space
E = ¦f ∈ C
j,α
(Ω, R
m
) : f
j,α
< ∞¦
is a Banach space.
6
As above, one may deﬁne the space C
j,α
(Ω, R
m
). And again, if Ω is bounded,
C
j,α
(Ω, R
m
) is a Banach space.
We shall also employ the following convention
C
j,0
(Ω, R
m
) = C
j
(Ω, R
m
)
and
C
j,0
(
¯
Ω, R
m
) = C
j
(
¯
Ω, R
m
).
2.4 Functions with compact support
Let Ω be an open subset of R
n
. A function f : Ω → R
m
is said to have
compact support in Ω if the set
supp f = closure¦x ∈ Ω : f(x) ,= 0¦ = ¦x ∈ Ω : f(x) ,= 0¦
is compact.
We let
C
j,α
0
(Ω, R
m
) = ¦f ∈ C
j,α
(Ω, R
m
) : supp f is a compact subset of Ω¦
and deﬁne C
j,α
0
(
¯
Ω, R
m
) similarly.
Then, again, if Ω is bounded, the space C
j,α
0
(
¯
Ω, R
m
) is a Banach space and
C
j,α
0
(
¯
Ω, R
m
) = ¦f ∈ C
j,α
(
¯
Ω, R
m
) : f(x) = 0, x ∈ ∂Ω¦.
2.5 L
p
spaces
Let Ω be a Lebesgue measurable subset of R
n
and let f : Ω → R
m
be a
measurable function. Let, for 1 ≤ p < ∞,
f
L
p =
__
Ω
[f(x)[
p
dx
_
1/p
,
and for p = ∞, let
f
L
∞ = essup
x∈Ω
[f(x)[,
where essup denotes the essential supremum.
For 1 ≤ p ≤ ∞, let
L
p
(Ω, R
m
) = ¦f : f
L
p < +∞¦.
Then L
p
(Ω, R
m
) is a Banach space for 1 ≤ p ≤ ∞. The space L
2
(Ω, R
m
) is a
Hilbert space with inner product deﬁned by
¸f, g) =
_
Ω
f(x) g(x)dx,
where f(x) g(x) is the inner product of f(x) and g(x) in the Hilbert space
(Euclidean space) R
n
.
2. BANACH SPACES 7
2.6 Weak derivatives
Let Ω be an open subset of R
n
. A function f : Ω →R
m
is said to belong to
class L
p
loc
(Ω, R
m
), if for every compact subset Ω
′
⊂ Ω, f ∈ L
p
(Ω
′
, R
m
). Let β =
(β
1
, . . . , β
n
) be a multiindex. Then a locally integrable function v ∈ L
1
loc
(Ω, R
m
)
is called the β
th
weak derivative of f if it satisﬁes
_
Ω
vφdx = (−1)
β
_
Ω
fD
β
φdx, for all φ ∈ C
∞
0
(Ω). (4)
We write v = D
β
f and note that, up to a set of measure zero, v is uniquely
determined. The concept of weak derivative extends the classical concept of
derivative and has many similar properties (see e.g. [18]).
2.7 Sobolev spaces
We say that f ∈ W
k
(Ω, R
m
), if f has weak derivatives up to order k, and
set
W
k,p
(Ω, R
m
) = ¦f ∈ W
k
(Ω, R
m
) : D
β
f ∈ L
p
(Ω, R
m
), [β[ ≤ k¦.
Then the vector space W
k,p
(Ω, R
m
) equipped with the norm
f
W
k,p =
_
_
_
Ω
β≤k
[D
β
f[
p
dx
_
_
1/p
(5)
is a Banach space. The space C
k
0
(Ω, R
m
) is a subspace of W
k,p
(Ω, R
m
), it’s
closure in W
k,p
(Ω, R
m
), denoted by W
k,p
0
(Ω, R
m
), is a Banach subspace which,
in general, is a proper subspace.
For p = 2, the spaces W
k,2
(Ω, R
m
) and W
k,2
0
(Ω, R
m
) are Hilbert spaces with
inner product ¸f, g) given by
¸f, g) =
_
Ω
α≤k
D
α
f D
α
gdx. (6)
These spaces play a special role in the linear theory of partial diﬀerential equa
tions, and in case Ω satisﬁes suﬃcient regularity conditions (see [1], [41]), they
may be identiﬁed with the following spaces.
Consider the space C
k
(
¯
Ω, R
m
) as a normed space using the  
W
k,p norm.
It’s completion is denoted by H
k,p
(Ω, R
m
). If p = 2 it is a Hilbert space with
inner product given by (6). H
k,p
0
(Ω, R
m
) is the completion of C
∞
0
(Ω, R
m
) in
H
k,p
(Ω, R
m
).
2.8 Spaces of linear operators
Let E and X be normed linear spaces with norms 
E
and 
X
, respectively.
Let
L(E; X) = ¦f : E →X such that f is linear and continuous¦.
8
For f ∈ L(E; X), let
f
L
= sup
xE≤1
f(x)
X
. (7)
Then  
L
is a norm for L(E; X). This space is a Banach space, whenever X
is.
Let E
1
, . . . , E
n
and X be n+1 normed linear spaces, let L(E
1
, . . . , E
n
; X) =
¦f : E
1
E
n
→ X such that f is multilinear (i.e. f is linear in each
variable separately) and continuous¦. Let
f = sup¦f(x
1
, . . . , x
n
)
X
: x
1

E1
≤ 1, . . . , x
n

En
≤ 1¦, (8)
then L(E
1
, . . . , E
n
; X) with the norm deﬁned by (8) is a normed linear space.
It again is a Banach space, whenever X is.
If E and X are normed spaces, one may deﬁne the spaces
L
1
(E; X) = L(E; X)
L
2
(E; X) = L(E; L
1
(E; X))
.
.
.
.
.
.
.
.
.
L
n
(E; X) = L(E; L
n−1
(E; X)), n ≥ 2.
We leave it as an exercise to show that the spaces L(E, . . . , E; X) (E repeated
n times) and L
n
(E; X) may be identiﬁed (i.e. there exists an isomorphism
between these spaces which is norm preserving (see e.g. [43]).
3 Diﬀerentiability, Taylor’s Theorem
3.1 Gˆateaux and Fr´echet diﬀerentiability
Let E and X be Banach spaces and let U be an open subset of E. Let
f : U →X
be a function. Let x
0
∈ U, then f is said to be Gˆateaux diﬀerentiable (G
diﬀerentiable) at x
0
in direction h, if
lim
t→0
1
t
¦f(x
0
+th) −f(x
0
)¦ (9)
exists. It said to be Fr´echet diﬀerentiable (Fdiﬀerentiable) at x
0
, if there exists
T ∈ L(E; X) such that
f(x
0
+h) −f(x
0
) = T(h) +o(h) (10)
for h small, here o(h) means that
lim
h→0
o(h)
h
= 0.
3. DIFFERENTIABILITY, TAYLOR’S THEOREM 9
(We shall use the symbol   to denote both the norm of E and the norm of
X, since it will be clear from the context in which space we are working.) We
note that Fr´echet diﬀerentiability is a more restrictive concept.
It follows from this deﬁnition that the Fr´echet–derivative of f at x
0
, if it
exists, is unique. We shall use the following symbols interchangeably for the
Fr´echet–derivative of f at x
0
; Df(x
0
), f
′
(x
0
), df(x
0
), where the latter is usually
used in case X = R. We say that f is of class C
1
in a neighborhood of x
0
if f
is Fr´echet diﬀerentiable there and if the mapping
Df : x →Df(x)
is a continuous mapping into the Banach space L(E; X).
If the mapping
Df : U →L(E; X)
is Fr´echet–diﬀerentiable at x
0
∈ U, we say that f is twice Fr´echet–diﬀerentiable
and we denote the second F–derivative by D
2
f(x
0
), or f
′′
(x
0
), or d
2
f(x
0
). Thus
D
2
f(x
0
) ∈ L
2
(E; X). In an analogous way one deﬁnes higher order diﬀerentia
bility.
If h ∈ E, we shall write
D
n
f(x
0
)(h, , h) as D
n
f(x
0
)h
n
,
(see Subsection 2.8).
If f : R
n
→ R
n
is Fr´echet diﬀerentiable at x
0
, then Df(x
0
) is given by the
Jacobian matrix
Df(x
0
) =
_
∂f
i
∂x
j
¸
¸
¸
¸
x=x0
_
,
and if f : R
n
→ R is Fr´echet diﬀerentiable, then Df(x
0
) is represented by the
gradient vector ∇f(x
0
), i.e.
Df(x
0
)(h) = ∇f(x
0
) h,
where “” is the dot product in R
n
, and the second derivative D
2
f(x
0
) is given
by the Hessian matrix
_
∂
2
f
∂xi∂xj
_
¸
¸
¸
¸
x=x0
, i.e.
D
2
f(x
0
)h
2
= h
T
_
∂
2
f
∂x
i
∂x
j
_
h,
where h
T
is the transpose of the vector h.
10
3.2 Taylor’s formula
1 Theorem Let f : E → X and all of its Fr´echet–derivatives Df, , D
m
f be
continuous on an open set U. Let x and x + h be such that the line segment
connecting these points lies in U. Then
f(x +h) −f(x) =
m−1
k=1
1
k!
D
k
f(x)h
k
+
1
m!
D
m
f(z)h
m
, (11)
where z is a point on the line segment connecting x to x + h. The remainder
1
m!
D
m
f(z)h
m
is also given by
1
(m−1)!
_
1
0
(1 −s)
m−1
D
m
f(x
0
+sh)h
m
ds. (12)
We shall not give a proof of this result here, since the proof is similar to the one
for functions f : R
n
→R
m
(see e.g. [14]).
3.3 EulerLagrange equations
In this example we shall discuss a fundamental problem of variational cal
culus to illustrate the concepts of diﬀerentiation just introduced; speciﬁcally we
shall derive the so called EulerLagrange diﬀerential equations. The equations
derived give necessary conditions for the existence of minima (or maxima) of
certain functionals.
Let g : [a, b] R
2
→R be twice continuously diﬀerentiable. Let E = C
2
0
[a, b]
and let T : E →R be given by
T(u) =
_
b
a
g(t, u(t), u
′
(t))dt.
It then follows from elementary properties of the integral, that T is of class C
1
.
Let u
0
∈ E be such that there exists an open neighborhood U of u
0
such that
T(u
0
) ≤ T(u) (13)
for all u ∈ U (u
0
is called an extremal of T). Since T is of class C
1
we obtain
that for u ∈ U
T(u) = T(u
0
) +DT(u
0
)(u −u
0
) +o(u −u
0
).
Hence for ﬁxed v ∈ E and ǫ ∈ R small,
T(u
0
+ǫv) = T(u
0
) +DT(u
0
)(ǫv) + o([ǫ[v).
It follows from (13) that
0 ≤ DT(u
0
)(ǫv) +o([ǫ[v)
and hence, dividing by ǫv,
0 ≤ DT(u
0
)
_
±
v
v
_
+
o(ǫv)
ǫv
,
4. SOME SPECIAL MAPPINGS 11
where   is the norm in E. It therefore follows, letting ǫ → 0, that for
every v ∈ E, DT(u
0
)(v) = 0. To derive the Euler–Lagrange equation, we must
compute DT(u
0
). For arbitrary h ∈ E we have
T(u
0
+h) =
_
b
a
g(t, u
0
(t) +h(t), u
′
0
(t) +h
′
(t))dt
=
_
b
a
g(t, u
0
(t), u
′
0
(t)dt
+
_
b
a
∂g
∂p
(t, u
0
(t), u
′
0
(t))h(t)dt
+
_
b
a
∂g
∂q
(t, u
0
(t), u
′
0
(t))h
′
(t)dt +o(h),
where p and q denote generic second, respectively, third variables in g. Thus
DT(u
0
)(h) =
_
b
a
∂g
∂p
(t, u
0
(t), u
′
0
(t))h(t)dt +
_
b
a
∂g
∂q
(t, u
0
(t), u
′
0
(t))h
′
(t)dt.
For notation’s sake we shall now drop the arguments in g and its partial deriva
tives. We compute
_
b
a
∂g
∂q
h
′
dt =
_
∂g
∂q
h
_
b
a
−
_
b
a
h
d
dt
_
∂g
∂q
_
dt,
and since h ∈ E, it follows that
DT(u
0
)(h) =
_
b
a
_
∂g
∂p
−
d
dt
∂g
∂q
_
hdt. (14)
Since DT(u
0
)(h) = 0 for all h ∈ E, it follows that
∂g
∂p
(t, u
0
(t), u
′
0
(t)) −
d
dt
∂g
∂q
(t, u
0
(t), u
′
0
(t)) = 0, (15)
for a ≤ t ≤ b (this fact is often referred to as the fundamental lemma of the
calculus of variations). Equation (15) is called the EulerLagrange equation. If
g is twice continuously diﬀerentiable (15) becomes
∂g
∂p
−
∂
2
g
∂t∂q
−
∂
2
g
∂p∂q
u
′
0
−
∂
2
g
∂q
2
u
′′
0
= 0, (16)
where it again is understood that all partial derivatives are to be evaluated at
(t, u
0
(t), u
′
0
(t)). We hence conclude that an extremal u
0
∈ E must solve the
nonlinear diﬀerential equation (16).
4 Some Special Mappings
Throughout our text we shall have occasion to study equations deﬁned by
mappings which enjoy special kinds of properties. We shall brieﬂy review some
such properties and refer the reader for more detailed discussions to standard
texts on analysis and functional analysis (e.g. [14]).
12
4.1 Completely continuous mappings
Let E and X be Banach spaces and let Ω be an open subset of E, let
f : Ω →X
be a mapping. Then f is called compact, whenever f(Ω
′
) is precompact in
X for every bounded subset Ω
′
of Ω (i.e. f(Ω
′
) is compact in X). We call f
completely continuous whenever f is compact and continuous. We note that if
f is linear and compact, then f is completely continuous.
2 Lemma Let Ω be an open set in E and let f : Ω →X be completely continuous,
let f be Fdiﬀerentiable at a point x
0
∈ Ω. Then the linear mapping T = Df(x
0
)
is compact, hence completely continuous.
Proof. Since T is linear it suﬃces to show that T(¦x : x ≤ 1¦) is precompact
in X. (We again shall use the symbol   to denote both the norm in E and
in X.) If this were not the case, there exists ǫ > 0 and a sequence ¦x
n
¦
∞
n=1
⊂
E, x
n
 ≤ 1, n = 1, 2, 3, such that
Tx
n
−Tx
m
 ≥ ǫ, n ,= m.
Choose 0 < δ < 1 such that
f(x
0
+h) −f(x
0
) −Th <
ǫ
3
h,
for h ∈ E, h ≤ δ. Then for n ,= m
f(x
0
+δx
n
) −f(x
0
+δx
m
) ≥ δTx
n
−Tx
m

− f(x
0
+δx
n
) −f(x
0
) −δTx
n
 −f(x
0
+δx
m
) −f(x
0
) −δTx
m

≥ δǫ −
δǫ
3
−
δǫ
3
=
δǫ
3
.
Hence the sequence ¦f(x
0
+ δx
n
)¦
∞
n=1
has no convergent subsequence. On the
other hand, for δ > 0, small, the set ¦x
0
+ δx
n
¦
∞
n=1
⊂ Ω, and is bounded,
implying by the complete continuity of f that ¦f(x
0
+δx
n
)¦
∞
n=1
is precompact.
We have hence arrived at a contradiction.
4.2 Proper mappings
Let M ⊂ E, Y ⊂ X and let f : M → Y be continuous, then f is called
a proper mapping if for every compact subset K of Y , f
−1
(K) is compact in
M. (Here we consider M and Y as metric spaces with metrics induced by the
norms of E and X, respectively.)
4. SOME SPECIAL MAPPINGS 13
3 Lemma Let h : E →X be completely continuous and let g : E →X be proper,
then f = g −h is a proper mapping, provided that f is coercive, i.e.
f(x) →∞ as x →∞. (17)
Proof. Let K be a compact subset of X and let N = f
−1
(K). Let ¦x
n
¦
∞
n=1
be a sequence in N. Then there exists ¦y
n
¦
∞
n=1
⊂ K such that
y
n
= g(x
n
) −h(x
n
). (18)
Since K is compact, the sequence ¦y
n
¦
∞
n=1
has a convergent subsequence, and
since f is coercive the sequence ¦x
n
¦
∞
n=1
must be bounded, further, because
h is completely continuous, the sequence ¦h(x
n
)¦
∞
n=1
must have a convergent
subsequence. It follows that the sequence ¦g(x
n
)¦
∞
n=1
has a convergent sub
sequence. Relabeling, if necessary, we may assume that all three sequences
¦y
n
¦
∞
n=1
, ¦g(x
n
)¦
∞
n=1
and ¦h(x
n
)¦
∞
n=1
are convergent. Since
g(x
n
) = y
n
+h(x
n
)
and g is proper, it follows that ¦x
n
¦
∞
n=1
converges also, say x
n
→ x; hence N
is precompact. That N is also closed follows from the fact that g and h are
continuous.
4 Corollary Let h : E → E be a completely continuous mapping, and let f =
id −h be coercive, then f is proper (here id is the identity mapping).
Proof. We note that id : E →E is a proper mapping.
In ﬁnite dimensional spaces the concepts of coercivity and properness are
equivalent, i.e. we have:
5 Lemma Let f : R
n
→ R
m
be continuous, then f is proper if and only if f is
coercive.
4.3 Contraction mappings, the Banach ﬁxed point theo
rem
Let M be a subset of a Banach space E. A function f : M → E is called a
contraction mapping if there exists a constant k, 0 ≤ k < 1 such that
f(x) −f(y) ≤ kx −y, for all x, y ∈ M. (19)
6 Theorem Let M be a closed subset of E and f : M → M be a contraction
mapping, then f has a unique ﬁxed point in M; i.e. there exists a unique x ∈ M
such that
f(x) = x . (20)
14
Proof. If x, y ∈ M both satisfy (20), then
x −y = f(x) −f(y) ≤ kx −y,
hence, since k < 1, x must equal y, establishing uniqueness of a ﬁxed point.
To proof existence, we deﬁne a sequence ¦x
n
¦
∞
n=0
⊂ M inductively as follows:
Choose x
0
∈ M and let
x
n
= f(x
n−1
), n ≥ 1. (21)
(21) implies that for any j ≥ 1
f(x
j
) −f(x
j−1
) ≤ k
j
x
1
−x
0
,
and hence if m > n
x
m
−x
n
= x
m
− x
m−1
+x
m−1
−. . . +x
n+1
−x
n
= f(x
m−1
) −f(x
m−2
) +. . . +f(x
n
) −f(x
n−1
)
and therefore
x
m
−x
n
 ≤ x
1
−x
0
(k
n
+. . . +k
m−1
) =
k
n
−k
m
1 −k
x
1
−x
0
. (22)
It follows from (22) that ¦x
n
¦
∞
n=0
is a Cauchy sequence in E, hence
lim
n→∞
x
n
= x
exists and since M is closed, x ∈ M. Using (21) we obtain that (20) holds.
7 Remark We note that the above theorem, Theorem 6, also holds if E is a
complete metric space with metric d. This is easily seen by replacing x−y by
d(x, y) in the proof.
In the following example we provide an elementary approach to the existence
and uniqueness of a solution of a nonlinear boundary value problem (see [13]).
The approach is based on the L
p
theory of certain linear diﬀerential operators
subject to boundary constraints.
Let T > 0 be given and let
f : [0, T] R R →R
be a mapping satisfying Carath´eodory conditions; i.e. f(t, u, u
′
) is continuous
in (u, u
′
) for almost all t and measurable in t for ﬁxed (u, u
′
).
We consider the Dirichlet problem, i.e. the problem of ﬁnding a function u
satisfying the following diﬀerential equation subject to boundary conditions
_
u
′′
= f(t, u, u
′
), 0 < t < T,
u = 0, t ∈ ¦0, T¦.
(23)
In what is to follow, we shall employ the notation that [ [ stands for absolute
value in R and  
2
the norm in L
2
(0, T).
We have the following results:
4. SOME SPECIAL MAPPINGS 15
8 Theorem Let f satisfy
[f(t, u, v)−f(t, ˜ u, ˜ v)[ ≤ a[u−˜ u[+b[v−˜ v[, ∀u, ˜ u, v, ˜ v ∈ R, 0 < t < T,(24)
where a, b are nonegative constants such that
a
λ
1
+
b
√
λ
1
< 1, (25)
and λ
1
is the principal eigenvalue of −u
′′
subject to the Dirichlet boundary
conditions u(0) = 0 = u(T) (i.e. the smallest number λ such that the problem
_
−u
′′
= λu, 0 < t < T,
u = 0, t ∈ ¦0, T¦.
(26)
has a nontrivial solution). Then problem (23) has a unique solution u ∈
C
1
0
([0, T]), with u
′
absolutely continuous and the equation (23) being satisﬁed
almost everywhere.
Proof. Results from elementary diﬀerential equations tell us that λ
1
is the
ﬁrst positive number λ such that the problem (26) has a nontrivial solution, i.e.
λ
1
=
π
2
T
2
.
To prove the theorem, let us, for v ∈ L
1
(0, T), put
Av = f(, w, w
′
), (27)
where
w(t) = −
t
T
_
T
0
_
τ
0
v(s)dsdτ +
_
t
0
_
τ
0
v(s)dsdτ,
which, in turn may be rewritten as
w(t) =
_
T
0
G(t, s)v(s)ds, (28)
where
G(t, s) = −
1
T
_
(T −t)s, if 0 ≤ s ≤ t
t(T −s), if t ≤ s ≤ T.
(29)
It follows from (24) that the operator A is a mapping of L
1
(0, T) to any
L
q
(0, T), q ≥ 1. On the other hand we have that the imbedding
L
q
(0, T) ֒→L
1
(0, T), q ≥ 1,
u ∈ L
q
(0, T) →u ∈ L
1
(0, T),
is a continuous mapping, since
u
L
1 ≤ T
q
q−1
u
L
q.
16
We hence may consider
A : L
q
(0, T) →L
q
(0, T),
for any q ≥ 1. In carrying out the computations in the case q = 2, the follow
ing inequalities will be used; their proofs may be obtained using Fourier series
methods, and will be left as an exercise. We have for w(t) =
_
T
0
G(t, s)v(s)ds
that
w
L
2 ≤
1
λ
1
v
L
2,
from which easily follows, via an integration by parts, that
w
′

L
2 ≤
1
√
λ
1
v
L
2
Using these facts in the computations one obtains the result that A is a con
traction mapping.
On the other hand, if v ∈ L
2
(0, T) is a ﬁxed point of A, then
u(t) =
_
T
0
G(t, s)v(s)ds
is in C
1
0
(0, T) and u
′′
∈ L
2
(0, T) and u solves (23).
9 Remark It is clear from the proof that in the above the real line R may be
replaced by R
m
thus obtaining a result for systems of boundary value problems.
10 Remark In case T = π, λ
1
= 1 and condition (25) becomes
a +b < 1,
whereas a classical result of Picard requires
a
π
2
8
+b
π
2
< 1,
(see [21] where also other results are cited).
11 Remark Theorem 8 may be somewhat extended using a result of Opial [31]
which says that for u ∈ C
0
[0, T], with u
′
absolutely continuous, we have that
_
T
0
[u(x)[[u
′
(x)[dx ≤
T
4
_
T
0
[u
′
(x)[
2
dx. (30)
The derivation of such a statement is left as an exercise.
4. SOME SPECIAL MAPPINGS 17
4.4 The implicit function theorem
Let us now assume we have Banach spaces E, X, Λ and let
f : U V →X,
(where U is open in E, V is open in Λ) be a continuous mapping satisfying the
following condition:
• For each λ ∈ V the map f(, λ) : U → X is Fr´echetdiﬀerentiable on U
with Fr´echet derivative
D
u
f(u, λ) (31)
and the mapping (u, λ) →D
u
f(u, λ) is a continuous mapping from U V
to L(E, X).
12 Theorem (Implicit Function Theorem) Let f satisfy (31) and let there ex
ist (u
0
, λ
0
) ∈ U V such that D
u
f(u
0
, λ
0
) is a linear homeomorphism of E onto
X (i.e. D
u
f(u
0
, λ
0
) ∈ L(E, X) and [D
u
f(u
0
, λ
0
)]
−1
∈ L(X, E)). Then there
exist δ > 0 and r > 0 and unique mapping u : B
δ
(λ
0
) = ¦λ : λ−λ
0
 ≤ δ¦ →E
such that
f(u(λ), λ) = f(u
0
, λ
0
), (32)
and u(λ) − u
0
 ≤ r, u(λ
0
) = u
0
.
Proof. Let us consider the equation
f(u, λ) = f(u
0
, λ
0
)
which is equivalent to
[D
u
f(u
0
, λ
0
)]
−1
(f(u, λ) −f(u
0
, λ
0
)) = 0, (33)
or
u = u −[D
u
f(u
0
, λ
0
)]
−1
(f(u, λ) −f(u
0
, λ
0
))
def
= G(u, λ). (34)
The mapping G has the following properties:
i) G(u
0
, λ
0
) = u
0
,
ii) G and D
u
G are continuous in (u, λ),
iii) D
u
G(u
0
, λ
0
) = 0.
18
Hence
G(u
1
, λ) −G(u
2
, λ)
≤
_
sup
0≤t≤1
D
u
G(u
1
+t(u
2
−u
1
), λ)
_
u
1
−u
2

≤
1
2
u
1
−u
2
,
(35)
provided u
1
−u
0
 ≤ r, u
2
−u
0
 ≤ r, where r is small enough. Now
G(u, λ) −u
0
 = G(u, λ) −G(u
0
, λ
0
) ≤ G(u, λ) −G(u
0
, λ)
+G(u
0
, λ) −G(u
0
, λ
0
) ≤
1
2
u −u
0
 +G(u
0
, λ) −G(u
0
, λ
0
)
≤
1
2
r +
1
2
r,
provided λ −λ
0
 ≤ δ is small enough so that G(u
0
, λ) −G(u
0
, λ
0
) ≤
1
2
r.
Let B
δ
(λ
0
) = ¦λ : λ−λ
0
 ≤ δ¦ and deﬁne M = ¦u : B
δ
(λ
0
) →E such that
u is continuous, u(λ
0
) = u
0
, u(λ)−u
0

0
≤ r, and u
0
= sup
λ∈B
δ
(λ0)
u(λ) <
+∞¦. Then M is a closed subset of a Banach space and (35) deﬁnes an equation
u(λ) = G(u(λ), λ) (36)
in M.
Deﬁne g by (here we think of u as an element of M)
g(u)(λ) = G(u(λ), λ),
then g : M →M and it follows by (36) that
g(u) −g(v)
0
≤
1
2
u −v
0
,
hence g has a unique ﬁxed point by the contraction mapping principle (Theorem
6).
13 Remark If in the implicit function theorem f is k times continuously diﬀeren
tiable, then the mapping λ →u(λ) inherits this property.
14 Example As an example let us consider the nonlinear boundary value problem
u
′′
+λe
u
= 0, 0 < t < π, u(0) = 0 = u(π). (37)
This is a one spacedimensional mathematical model from the theory of com
bustion (cf [3]) and u represents a dimensionless temperature. We shall show,
by an application of Theorem 12, that for λ ∈ R, in a neighborhood of 0, (37)
has a unique solution of small norm in C
2
([0, π], R).
4. SOME SPECIAL MAPPINGS 19
To this end we deﬁne
E = C
2
0
([0, π], R)
X = C
0
[0, π]
Λ = R,
these spaces being equipped with their usual norms (see earlier examples). Let
f : E Λ →X
be given by
f(u, λ) = u
′′
+λe
u
.
Then f is continuous and f(0, 0) = 0. (When λ = 0 (no heat generation) the
unique solution is u ≡ 0.) Furthermore, for u
0
∈ E, D
u
f(u
0
, λ) is given by (the
reader should carry out the veriﬁcation)
D
u
f(u
0
, λ)v = v
′′
+λe
u0(x)
v,
and hence the mapping
(u, λ) →D
u
f(u, λ)
is continuous. Let us consider the linear mapping
T = D
u
f(0, 0) : E →X.
We must show that this mapping is a linear homeomorphism. To see this we
note that for every h ∈ X, the unique solution of
v
′′
= h(t), 0 < t < π, v(0) = 0 = v(π),
is given by (see also (28))
v(t) =
_
π
0
G(t, s)h(s)ds, (38)
where
G(x, s) =
_
−
1
π
(π − t)s, 0 ≤ s ≤ t
−
1
π
t(π −s), t ≤ s ≤ π.
From the representation (38) we may conclude that there exists a constant c
such that
v
2
= T
−1
h
2
≤ ch
0
,
20
i.e. T
−1
is one to one and continuous. Hence all conditions of the implicit func
tion theorem are satisﬁed and we may conclude that for each λ, λ suﬃciently
small, (37) has a unique small solution u ∈ C
2
([0, π], R), furthermore the map
λ →u(λ) is continuous from a neighborhood of 0 ∈ R to C
2
([0, π], R). We later
shall show that this ‘solution branch’ (λ, u(λ)) may be globally continued. To
this end we note here that the set ¦λ > 0 : (37) has a solution ¦ is bounded
above. We observe that if λ > 0 is such that (37) has a solution, then the
corresponding solution u must be positive, u(x) > 0, 0 < x < π. Hence
0 = u
′′
+λe
u
> u
′′
+λu. (39)
Let v(t) = sin t, then v satisﬁes
v
′′
+v = 0, 0 < t < π, v(0) = 0 = v(π). (40)
From (39) and (40) we obtain
0 >
_
π
0
(u
′′
v −v
′′
u)dt + (λ −1)
_
π
0
uvdt,
and hence, integrating by parts,
0 > (λ −1)
_
π
0
uvdx,
implying that λ < 1.
5 Inverse Function Theorems
We next proceed to the study of the inverse of a given mapping and provide
two inverse function theorems. Since the ﬁrst result is proved in exactly the
same way as its ﬁnite dimensional analogue (it is an immediate consequence of
the implicit function theorem) we shall not prove it here (see again [14]).
15 Theorem Let E and X be Banach spaces and let U be an open neighborhood
of a ∈ E. Let f : U →X be a C
1
mapping with Df(a) a linear homeomorphism
of E onto X. Then there exist open sets U
′
and V , a ∈ U
′
, f(a) ∈ V and a
uniquely determined function g such that:
i) V = f(U
′
),
ii) f is one to one on U
′
,
iii) g : V →U
′
, g(V ) = U
′
, g(f(u)) = u, for every u ∈ U
′
,
iv) g is a C
1
function on V and Dg(f(a)) = [Df(a)]
−1
.
5. INVERSE FUNCTION THEOREMS 21
16 Example Consider the forced nonlinear oscillator (periodic boundary value
problem)
u
′′
+λu +u
2
= g, u(0) = u(2π), u
′
(0) = u
′
(2π) (41)
where g is a continuous 2π −periodic function and λ ∈ R, is a parameter. Let
E = C
2
([0, 2π], R) ∩¦u : u(0) = u(2π), u
′
(0) = u
′
(2π)¦, and X = C
0
([0, 2π], R),
where both spaces are equipped with the norms discussed earlier. Then for
certain values of λ, (41) has a unique solution for all forcing terms g of small
norm.
Let
f : E →X
be given by
f(u) = u
′′
+λu +u
2
.
Then Df(u) is deﬁned by
(Df(u))(v) = v
′′
+λv + 2uv,
and hence the mapping
u →Df(u)
is a continuous mapping of E to L(E; X), i.e. f is a C
1
mapping. It follows
from elementary diﬀerential equations theory (see eg. [4]) that the problem
v
′′
+λv = h,
has a unique 2π–periodic solution for every 2π–periodic h as long as λ ,= n
2
,
n = 1, 2, . . ., and that v
2
≤ Ch
0
for some constant C (only depending upon
λ). Hence Df(0) is a linear homeomorphism of E onto X. We thus conclude
that for given λ ,= n
2
, (32) has a unique solution u ∈ E of small norm for every
g ∈ X of small norm.
We note that the above example is prototypical for forced nonlinear oscilla
tors. Virtually the same arguments can be applied (the reader might carry out
the necessary calculations) to conclude that the forced pendulum equation
u
′′
+λsin u = g (42)
has a 2π periodic response of small norm for every 2π  periodic forcing term
g of small norm, as long as λ ,= n
2
, n = 1, 2, . . . .
In many physical situations (see the example below) it is of interest to know
the number of solutions of the equation describing this situation. The following
result describes a class of problems where the precise number of solutions (for
every given forcing term) may be obtained by simply knowing the number of
solutions for some ﬁxed forcing term.
Let M and Y metric spaces (e.g. subsets of Banach spaces with metric
induced by the norms).
22
17 Theorem Let f : M → Y be continuous, proper and locally invertible (e.g.
Theorem 15 is applicable at each point). For y ∈ Y let
N(y) = cardinal number of ¦f
−1
(y)¦ = #¦f
−1
(y)¦.
Then the mapping
y →N(y)
is ﬁnite and locally constant.
Proof. We ﬁrst show that for each y ∈ Y , N(y) is ﬁnite. Since ¦y¦ is compact
¦f
−1
(y)¦ is compact also, because f is a proper mapping. Since f is locally
invertible, there exists, for each u ∈ ¦f
−1
(y)¦ a neighborhood O
u
such that
O
u
∩ (¦f
−1
(y)¦¸¦u¦) = ∅,
and thus ¦f
−1
(y)¦ is a discrete and compact set, hence ﬁnite.
We next show that N is a continuous mapping to the nonnegative integers,
which will imply that N is constant–valued. Let y ∈ Y and let ¦f
−1
(y)¦ =
¦u
1
, . . . , u
n
¦. We choose disjoint open neighborhoods O
i
of u
i
, 1 ≤ i ≤ n and
let I =
n
i=1
f(O
i
). Then there exist open sets V
i
, u
i
∈ V
i
such that f is a
homeomorphism from V
i
to I. We next claim that there exists a neighborhood
W ⊂ I of y such that N is constant on W. For if not, there will exist a sequence
¦y
m
¦, with y
m
→ y, such that as m → ∞, N(y
m
) > N(y) (note that for
any v ∈ I, v has a preimage in each V
i
which implies N(v) ≥ N(y), v ∈ I!).
Hence there exists a sequence ¦ξ
m
¦, ξ
m
,∈
n
i=1
V
i
, such that f(ξ
m
) = y
m
.
Since f
−1
(¦y
n
¦ ∪ ¦y¦) is compact, the sequence ¦ξ
n
¦ will have a convergent
subsequence, say ξ
nj
→ ξ. And since f is continuous, f(ξ) = y. Hence ξ = u
i
,
for some i, a contradiction to ξ ,∈
n
i=1
V
i
.
18 Corollary Assume Y is connected, then N(Y ) is constant.
Examples illustrating this result will be given later in the text.
6 The Dugundji Extension Theorem
In the course of developing the Brouwer and Leray–Schauder degree and in
proving some of the classical ﬁxed point theorems we need to extend mappings
deﬁned on proper subsets of a Banach space to the whole space in a suitable
manner. The result which guarantees the existence of extensions having the
desired properties is the Dugundji extension theorem ([16]) which will be estab
lished in this section.
In proving the theorem we need a result from general topology which we
state here for convenience (see e.g [16]). We ﬁrst give some terminology.
Let M be a metric space and let ¦O
λ
¦
λ∈Λ
, where Λ is an index set, be an
open cover of M. Then ¦O
λ
¦
λ∈Λ
is called locally ﬁnite if every point u ∈ M
has a neighborhood U such that U intersects at most ﬁnitely many elements of
¦O
λ
¦
λ∈Λ
.
6. THE DUGUNDJI EXTENSION THEOREM 23
19 Lemma Let M be a metric space. Then every open cover of M has a locally
ﬁnite reﬁnement.
20 Theorem Let E and X be Banach spaces and let f : C →K be a continuous
mapping, where C is closed in E and K is convex in X. Then there exists a
continuous mapping
˜
f : E →K
such that
˜
f(u) = f(u), u ∈ C.
Proof. For each u ∈ E¸C let
r
u
=
1
3
dist(u, C)
and
B
u
= ¦v ∈ E : v −u < r
u
¦.
Then
diamB
u
≤ dist(B
u
, C).
The collection ¦B
u
¦
u∈E\C
is an open cover of the metric space E¸C and hence
has a locally ﬁnite reﬁnement ¦O
λ
¦
λ∈Λ
, i.e.
i)
λ∈Λ
O
λ
⊃ E¸C,
ii) for each λ ∈ Λ there exists B
u
such that O
λ
⊂ B
u
iii) ¦O
λ
¦
λ∈Λ
is locally ﬁnite.
Deﬁne
q : E¸C →(0, ∞)
by
q(u) =
λ∈Λ
dist(u, E¸O
λ
). (43)
The sum in the right hand side of (43) contains only ﬁnitely many terms, since
¦O
λ
¦
λ∈Λ
is locally ﬁnite. This also implies that q is a continuous function.
Deﬁne
ρ
λ
(u) =
dist(u, E¸O
λ
)
q(u)
, λ ∈ Λ, u ∈ E¸C.
It follows for λ ∈ Λ and u ∈ E¸C that
0 ≤ ρ
λ
(u) ≤ 1,
λ∈Λ
ρ
λ
(u) = 1.
24
For each λ ∈ Λ choose u
λ
∈ C such that
dist(u
λ
, O
λ
) ≤ 2dist(C, O
λ
)
and deﬁne
˜
f(u) =
_
f(u), u ∈ C
λ∈Λ
ρ
λ
(u)f(u
λ
), u ,∈ C.
Then
˜
f has the following properties:
i)
˜
f is deﬁned on E and is an extension of f.
ii)
˜
f is continuous on the interior of C.
iii)
˜
f is continuous on E¸C.
These properties follow immediately from the deﬁnition of
˜
f. To show that
˜
f is
continuous on E it suﬃces therefore to show that
˜
f is continuous on ∂C. Let
u ∈ ∂C, then since f is continuous we may, for given ǫ > 0, ﬁnd 0 < δ = δ(u, ǫ)
such that
f(u) −f(v) ≤ ǫ, if u −v ≤ δ, v ∈ C.
Now for v ∈ E¸C

˜
f(u) −
˜
f(v) = f(u) −
λ∈Λ
ρ
λ
(v)f(u
λ
) ≤
λ∈Λ
ρ
λ
(v)f(u) −f(u
λ
).
If ρ
λ
(v) ,= 0, λ ∈ Λ, then dist(v, E¸O
λ
) > 0, i.e. v ∈ O
λ
. Hence v − u
λ
 ≤
v −w +w−u
λ
 for any w ∈ O
λ
. Since v −w ≤ diamO
λ
we may take the
inﬁmum for w ∈ O
λ
and obtain
v −u
λ
 ≤ diamO
λ
+ dist(u
λ
, O
λ
).
Now O
λ
⊂ B
u1
for some u
1
∈ E¸C. Hence, since
diamO
λ
≤ diamB
u1
≤ dist(B
u1
, C) ≤ dist(C, O
λ
),
we get
v −u
λ
 ≤ 3dist(C, O
λ
) ≤ 3v −u.
Thus for λ such that ρ
λ
(v) ,= 0 we get u−u
λ
 ≤ v−u+v−u
λ
 ≤ 4u−v.
Therefore if u − v ≤ δ/4, then u − u
λ
 ≤ δ, and f(u) − f(u
λ
) ≤ ǫ, and
therefore

˜
f(u) −
˜
f(v) ≤ ǫ
λ∈Λ
ρ
λ
(v) = ǫ.
7. EXERCISES 25
21 Corollary Let E, X be Banach spaces and let f : C →X be continuous, where
C is closed in E. Then f has a continuous extension
˜
f to E such that
˜
f(E) ⊂ cof(C),
where cof(C) is the convex hull of f(C).
22 Corollary Let K be a closed convex subset of a Banach space E. Then there
exists a continuous mapping f : E → K such that f(E) = K and f(u) = u,
u ∈ K, i.e. K is a continuous retract of E.
Proof. Let id : K → K be the identity mapping. This map is continuous.
Since K is closed and convex we may apply Corollary 21 to obtain the desired
conclusion.
7 Exercises
1. Supply all the details for the proof of Theorem 8.
2. Compare the requirements discussed in Remark 10.
3. Derive an improved result as suggested by Remark 11.
4. Establish the assertion of Remark 13.
5. Supply the details of the proof of Example 14.
6. Prove Theorem 15.
7. Carry out the program laid out by Example 16 to discuss the nonlinear
oscillator given by (42).
26
Chapter II
The Method of
LyapunovSchmidt
1 Introduction
In this chapter we shall develop an approach to bifurcation theory which,
is one of the original approaches to the theory. The results obtained, since the
implicit function theorem plays an important role, will be of a local nature. We
ﬁrst develop the method of Liapunov–Schmidt and then use it to obtain a local
bifurcation result. We then use this result in several examples. Later in the
text it will be used to derive a Hopf bifurcation theorem.
2 Splitting Equations
Let X and Y be real Banach spaces and let F be a mapping
F : X R →Y (1)
and let F satisfy the following conditions:
F(0, λ) = 0, ∀λ ∈ R, (2)
and
F is C
2
in a neighborhood of ¦0¦ R. (3)
We shall be interested in obtaining existence of nontrivial solutions (i.e. u ,= 0)
of the equation
F(u, λ) = 0 (4)
We call λ
0
a bifurcation value or (0, λ
0
) a bifurcation point for (4) provided
every neighborhood of (0, λ
0
) in X R contains solutions of (4) with u ,= 0.
It then follows from the implicit function theorem that the following holds.
27
28
1 Theorem If the point (0, λ
0
) is a bifurcation point for the equation
F(u, λ) = 0, (5)
then the Fr´echet derivative F
u
(0, λ
0
) cannot be a linear homeomorphism of X
to Y.
The types of linear operators F
u
(0, λ
0
) we shall consider are socalled Fred
holm operators.
2 Deﬁnition A linear operator L : X → Y is called a Fredholm operator pro
vided:
• The kernel of L, kerL, is ﬁnite dimensional.
• The range of L, imL, is closed in Y.
• The cokernel of L, cokerL, is ﬁnite dimensional.
The following lemma which is a basic result in functional analysis will be
important for the development to follow, its proof may be found in any standard
text, see e.g. [38].
3 Lemma Let F
u
(0, λ
0
) be a Fredholm operator from X to Y with kernel V and
cokernel Z. Then there exists a closed subspace W of X and a closed subspace
T of Y such that
X = V ⊕W
Y = Z ⊕T.
The operator F
u
(0, λ
0
) restricted to W, F
u
(0, λ
0
)[
W
: W → T, is bijective and
since T is closed it has a continuous inverse. Hence F
u
(0, λ
0
)[
W
is a linear
homeomorphism of W onto T.
We recall that W and Z are not uniquely given.
Using Lemma 3 we may now decompose every u ∈ X and F uniquely as
follows:
u = u
1
+u
2
, u
1
∈ V, u
2
∈ W,
F = F
1
+F
2
, F
1
: X →Z, F
2
: X →T.
(6)
Hence equation (5) is equivalent to the system of equations
F
1
(u
1
, u
2
, λ) = 0,
F
2
(u
1
, u
2
, λ) = 0.
(7)
We next let L = F
u
(0, λ
0
) and using a Taylor expansion we may write
F(u, λ) = F(0, λ
0
) +F
u
(0, λ
0
)u +N(u, λ). (8)
or
Lu +N(u, λ) = 0, (9)
2. SPLITTING EQUATIONS 29
where
N : X R →Y. (10)
Using the decomposition of X we may write equation (9) as
Lu
2
+N(u
1
+u
2
, λ) = 0. (11)
Let Q : Y →Z and I −Q : Y →T be projections determined by the decompo
sition, then equation (10) implies that
QN(u, λ) = 0. (12)
Since by Lemma 3, L[
W
: W → T has an inverse L
−1
: T → W we obtain
from equation (11) the equivalent system
u
2
+L
−1
(I −Q)N(u
1
+u
2
, λ) = 0. (13)
We note, that since Z is ﬁnite dimensional, equation (12) is an equation in a
ﬁnite dimensional space, hence if u
2
can be determined as a function of u
1
and λ,
this equation will be a ﬁnite set of equations in ﬁnitely many variables (u
1
∈ V,
which is also assumed ﬁnite dimensional!)
Concerning equation (12) we have the following result.
4 Lemma Assume that F
u
(0, λ
0
) is a Fredholm operator with W nontrivial.
Then there exist ǫ > 0, δ > 0 and a unique solution u
2
(u
1
, λ) of equation
(13) deﬁned for [λ − λ
0
[ + u
1
 < ǫ with u
2
(u
1
, λ) < δ. This function solves
the equation F
2
(u
1
, u
2
(u
1
, λ), λ) = 0.
Proof. We employ the implicit function theorem to analyze equation (13).
That this may be done follows from the fact that at u
1
= 0 and λ = λ
0
equation
(13) has the unique solution u
2
= 0 and the Fr´echet derivative at this point
with respect to u
2
is simply the identity mapping on W.
Hence, using Lemma 4, we will have nontrivial solutions of equation (5) once
we can solve
F
1
(u
1
, u
2
(u
1
, λ), λ) = QF(u
1
+u
2
(u
1
, λ), λ) = 0 (14)
for u
1
, whenever [λ−λ
0
[+u
1
 < ǫ. This latter set of equations, usually referred
to as the set of bifurcation equations, is, even though a ﬁnite set of equations in
ﬁnitely many unknowns, the more diﬃcult part in the solution of equation (5).
The next sections present situations where these equations may be solved.
5 Remark We note that in the above considerations at no point was it required
that λ be a one dimensional parameter.
30
3 Bifurcation at a Simple Eigenvalue
In this section we shall consider the analysis of the bifurcation equation (14)
in the particular case that the kernel V and the cokernel Z of F
u
(0, λ
0
) both
have dimension 1.
We have the following theorem.
6 Theorem In the notation of the previous section assume that the kernel V and
the cokernel Z of F
u
(0, λ
0
) both have dimension 1. Let V = span¦φ¦ and let
Q be a projection of Y onto Z. Furthermore assume that the second Fr´echet
derivative F
uλ
satisﬁes
QF
uλ
(0, λ
0
)(φ, 1) ,= 0. (15)
Then (0, λ
0
) is a bifurcation point and there exists a unique curve
u = u(α), λ = λ(α),
deﬁned for α ∈ R in a neighborhood of 0 so that
u(0) = 0, u(α) ,= 0, α ,= 0, λ(0) = λ
0
and
F(u(α), λ(α)) = 0.
Proof. Since V is one dimensional u
1
= αφ. Hence for [α[ small and λ near λ
0
there exists a unique u
2
(α, λ) such that
F
2
(αφ, u
2
(α, λ), λ) = 0.
We hence need to solve for λ = λ(α) in the equation
QF(αφ +u
2
(α, λ), λ) = 0.
We let µ = λ −λ
0
and deﬁne
g(α, µ) = QF(αφ +u
2
(α, λ), λ).
Then g maps a neighborhood of the origin of R
2
into R.
Using Taylor’s theorem we may write
F(u, λ) = F
u
u+F
λ
µ+
1
2
¦F
uu
(u, u) + 2F
uλ
(u, λ) +F
λλ
(µ, µ)¦+R,(16)
where R contains higher order remainder terms and all Fr´echet derivatives above
are evaluated at (0, λ
0
).
Because of (2) we have that F
λ
and F
λλ
in the above are the zero operators,
hence, by applying Q to (16) we obtain
QF(u, λ) =
1
2
¦QF
uu
(u, u) + 2QF
uλ
(u, µ)¦ +QR, (17)
3. BIFURCATION AT A SIMPLE EIGENVALUE 31
and for α ,= 0
g(α,µ)
α
=
1
2
_
QF
uu
(φ +
u2(α,λ)
α
, αφ +u
2
(α, λ))
+ 2QF
uλ
(φ +
u2(α,λ)
α
, µ)
_
+
1
α
QR.
(18)
It follows from Lemma 4 that the term
u2(α,λ)
α
is bounded for α in a neigh
borhood of 0. The remainder formula of Taylor’s theorem implies a similar
statement for the term
1
α
QR. Hence
h(α, µ) =
g(α, µ)
α
= O(α), as α →0.
We note that in fact h(0, 0) = 0, and
∂h(0, 0)
∂µ
= QF
uλ
(0, λ
0
)(φ, 1) ,= 0.
We hence conclude by the implicit function theorem that there exists a unique
function µ = µ(α) deﬁned in a neighborhood of 0 such that
h(α, µ(α)) = 0.
We next set
u(α) = αφ +u
2
(α, λ
0
+µ(α)), λ = λ
0
+µ(α).
This proves the theorem.
The following example will serve to illustrate the theorem just established.
7 Example The point (0, 0) is a bifurcation point for the ordinary diﬀerential
equation
u
′′
+λ(u +u
3
) = 0 (19)
subject to the periodic boundary conditions
u(0) = u(2π), u
′
(0) = u
′
(2π). (20)
To see this, we choose
X = C
2
[0, 2π] ∩ ¦u : u(0) = u(2π), u
′
(0) = u
′
(2π), u
′′
(0) = u
′′
(2π)¦,
Y = C[0, 2π] ∩ ¦u : u(0) = u(2π)¦,
both equipped with the usual norms, and
F : X R →Y
(u, λ) →u
′′
+λ(u +u
3
).
Then F belongs to class C
2
with Fr´echet derivative
F
u
(0, λ
0
)u = u
′′
+λ
0
u. (21)
32
This linear operator has a nontrivial kernel whenever λ
0
= n
2
, n = 0, 1, .
The kernel being one dimensional if and only if λ
0
= 0.
We see that h belongs to the range of F
u
(0, 0) if and only if
_
2π
0
h(s)ds = 0,
and hence the cokernel will have dimension 1 also. A projection Q : Y →
Z then is given by Qh =
1
2π
_
2π
0
h(s)ds. Computing further, we ﬁnd that
F
uλ
(0, 0)(u, λ) = λu, and hence, since we may choose φ = 1, F
uλ
(0, 0)(1, 1) = 1.
Applying Q we get Q1 = 1. We may therefore conclude by Theorem 6 that
equation (19) has a solution u satisfying the boundary conditions (20) which is
of the form
u(α) = α +u
2
(α, λ(α)).
Chapter III
Degree Theory
1 Introduction
In this chapter we shall introduce an important tool for the study of non
linear equations, the degree of a mapping. We shall mainly follow the analytic
development commenced by Heinz in [22] and Nagumo in [29]. For a brief
historical account we refer to [42].
2 Deﬁnition of the Degree of a Mapping
Let Ω be a bounded open set in R
n
and let f :
¯
Ω →R
n
be a mapping which
satisﬁes
• f ∈ C
1
(Ω, R
n
) ∩ C(
¯
Ω, R
n
), (1)
• y ∈ R
n
is such that
y / ∈ f(∂Ω), (2)
• if x ∈ Ω is such that f(x) = y then
f
′
(x) = Df(x) (3)
is nonsingular.
1 Proposition If f satisﬁes (1), (2), (3), then the equation
f(x) = y (4)
has at most a ﬁnite number of solutions in Ω.
2 Deﬁnition Let f satisfy (1), (2), (3). Deﬁne
d(f, Ω, y) =
k
i=1
sgn det f
′
(x
i
) (5)
33
34
where x
1
, , x
k
are the solutions of (4) in Ω and
sgn det f
′
(x
i
) =
_
_
_
+1, if det f
′
(x
i
) > 0
−1, if det f
′
(x
i
) < 0, i = 1, , k.
If (4) has no solutions in Ω we let d(f, Ω, y) = 0.
The Brouwer degree d(f, Ω, y) to be deﬁned for mappings f ∈ C(
¯
Ω, R
n
)
which satisfy (2) will coincide with the number just deﬁned in case f satisﬁes
(1), (2), (3). In order to give this deﬁnition in the more general case we need a
sequence of auxiliary results.
The proof of the ﬁrst result, which follows readily by making suitable changes
of variables, will be left as an exercise.
3 Lemma Let φ : [0, ∞) →R be continuous and satisfy
φ(0) = 0, φ(t) ≡ 0, t ≥ r > 0,
_
R
n
φ([x[)dx = 1. (6)
Let f satisfy the conditions (1), (2), (3). Then
d(f, Ω, y) =
_
Ω
φ([f(x) −y[)det f
′
(x)dx, (7)
provided r is suﬃciently small.
4 Lemma Let f satisfy (1) and (2) and let r > 0 be such that [f(x) −y[ > r, x ∈
∂Ω. Let φ : [0, ∞) →R be continuous and satisfy:
φ(s) = 0, s = 0, r ≤ s, and
_
∞
0
s
n−1
φ(s)ds = 0. (8)
Then
_
Ω
φ([f(x) −y[) det f
′
(x)dx = 0. (9)
Proof. We note ﬁrst that it suﬃces to proof the lemma for functions f which
are are of class C
∞
and for functions φ that vanish in a neighborhood of 0. We
also note that the function φ([f(x) − y[) det f
′
(x) vanishes in a neighborhood
of ∂Ω, hence we may extend that function to be identically zero outside Ω and
_
Ω
φ([f(x) −y[) det f
′
(x)dx =
_
Ω
′
φ([f(x) −y[) det f
′
(x)dx,
where Ω
′
is any domain with smooth boundary containing Ω.
We let
ψ(s) =
_
s
−n
_
s
0
ρ
n−1
φ(ρ)dρ, 0 < s < ∞
0, s = 0.
(10)
2. DEFINITION 35
Then ψ, so deﬁned is a C
1
function, it vanishes in a neighborhood of 0 and in
the interval [r, ∞). Further ψ satisﬁes the diﬀerential equation
sψ
′
(s) +nψ(s) = φ(s). (11)
It follows that the functions
g
j
(x) = ψ([x[)x
j
, j = 1, , n
belong to class C
1
and
g
j
(x) = 0, [x[ ≥ r,
and furthermore that for j = 1, , n the functions g
j
(f(x)−y) are C
1
functions
which vanish in a neighborhood of ∂Ω. If we denote by a
ji
(x) the cofactor of
the element
∂fi
∂xj
in the Jacobian matrix f
′
(x), it follows that
div (a
j1
(x), a
j2
(x), , a
jn
(x)) = 0, j = 1, , n.
We next deﬁne for i = 1, , n
v
i
(x) =
n
j=1
a
ji
(x)g
j
(f(x) −y)
and show that the function v = (v
1
, v
2
, , v
n
) has the property that
divv = φ([f(x) −y[) det f
′
(x),
and hence the result follows from the divergence theorem.
5 Lemma Let f satisfy (1) and (2) and let φ : [0, ∞) →R be continuous, φ(0) =
0, φ(s) ≡ 0 for s ≥ r, where 0 < r ≤ min
x∈∂Ω
[f(x) − y[,
_
R
n
φ([x[)dx = 1.
Then for all such φ, the integrals
_
Ω
φ([f(x) −y[) det f
′
(x)dx (12)
have a common value.
Proof. Let Φ = ¦φ ∈ C([0, ∞), R) : φ(0) = 0, φ(s) ≡ 0, s ≥ r¦. Put
Lφ =
_
∞
0
s
n−1
φ(s)ds
Mφ =
_
R
n
φ([x[)dx
Nφ =
_
Ω
φ([f(x) −y[) det f
′
(x)dx.
36
Then L, M, N are linear functionals. It follows from Lemma 4 that Mφ = 0
and Nφ = 0, whenever Lφ = 0. Let φ
1
, φ
2
∈ Φ with Mφ
1
= Mφ
2
= 1, then
L((Lφ
2
)φ
1
−(Lφ
1
)φ
2
) = 0.
It follows that
(Lφ
2
)(Mφ
1
) −(Lφ
1
)(Mφ
2
) = 0,
Lφ
2
−Lφ
1
= L(φ
2
−φ
1
) = 0,
and
N(φ
2
−φ
1
) = 0,
i.e.
Nφ
2
= Nφ
1
.
6 Lemma Let f
1
and f
2
satisfy (1), (2), (3) and let ǫ > 0 be such that
[f
i
(x) −y[ > 7ǫ, x ∈ ∂Ω, i = 1, 2, (13)
[f
1
(x) −f
2
(x)[ < ǫ, x ∈
¯
Ω, (14)
then
d(f
1
, Ω, y) = d(f
2
, Ω, y).
Proof. We may, without loss, assume that y = 0, since by Deﬁnition 2
d(f, Ω, y) = d(f −y, Ω, 0).
let g ∈ C
1
[0, ∞) be such that
g(s) = 1, 0 ≤ s ≤ 2ǫ
0 ≤ g(r) ≤ 1, 2ǫ ≤ r < 3ǫ
g(r) = 0, 3ǫ ≤ r < ∞. (15)
Consider
f
3
(x) = [1 −g([f
1
(x)[)]f
1
(x) +g([f
1
(x)[)f
2
(x),
then
f
3
∈ C
1
(Ω, R
n
) ∩ C(
¯
Ω, R
n
)
and
[f
i
(x) −f
k
(x)[ < ǫ, i, k = 1, 2, 3, x ∈
¯
Ω
2. DEFINITION 37
[f
i
(x)[ > 6ǫ, x ∈ ∂Ω, i = 1, 2, 3.
Let φ
i
∈ C[0, ∞), i = 1, 2 be continuous and be such that
φ
1
(t) = 0, 0 ≤ t ≤ 4ǫ, 5ǫ ≤ t ≤ ∞
φ
2
(t) = 0, ǫ ≤ t < ∞, φ
2
(0) = 0
_
R
n
φ
i
([x[)dx = 1, i = 1, 2.
We note that
f
3
≡ f
1
, if [f
1
[ > 3ǫ
f
3
≡ f
2
, if [f
1
[ < 2ǫ.
Therefore
φ
1
([f
3
(x)[)det f
′
3
(x) = φ
1
([f
1
(x)[)det f
′
1
(x)
φ
2
([f
3
(x)[)det f
′
3
(x) = φ
2
([f
2
(x)[)det f
′
2
(x).
(16)
Integrating both sides of (16) over Ω and using Lemmas 4 and 5 we obtain the
desired conclusion.
7 Corollary Let f satisfy conditions (1), (2), (3), then for ǫ > 0 suﬃciently
small any function g which also satisﬁes these conditions and which is such that
[f(x) −g(x)[ < ǫ, x ∈
¯
Ω, has the property that d(f, Ω, y) = d(g, Ω, y).
Up to now we have shown that if f and g satisfy conditions (1), (2), (3)
and if they are suﬃciently “close” then they have the same degree. In order
to extend this deﬁnition to a broader class of functions, namely those which do
not satisfy (3) we need a version of Sard’s Theorem (Lemma 8) (an important
lemma of Diﬀerential Topology) whose proof may be found in [40], see also [44].
8 Lemma If Ω is a bounded open set in R
n
, f satisﬁes (1), (2), and
E = ¦x ∈ Ω : det f
′
(x) = 0¦. (17)
Then f(E) does not contain a sphere of the form ¦z : [z −y[ < r¦.
This lemma has as a consequence the obvious corollary:
9 Corollary Let
F = ¦h ∈ R
n
: y +h ∈ f(E)¦, (18)
where E is given by (17), then F is dense in a neighborhood of 0 ∈ R
n
and
f(x) = y +h, x ∈ Ω, h ∈ F
implies that f
′
(x) is nonsingular.
38
We thus conclude that for all ǫ > 0, suﬃciently small, there exists h ∈ F,
0 < [h[ < ǫ, such that d(f, Ω, y +h) = d(f −h, Ω, y) is deﬁned by Deﬁnition 2.
It also follows from Lemma 6 that for such h, d(f, Ω, y + h) is constant. This
justiﬁes the following deﬁnition.
10 Deﬁnition Let f satisfy (1) and (2). We deﬁne
d(f, Ω, y) = lim
h →0
h ∈ F
d(f −h, Ω, y). (19)
Where F is given by (18) and d(f −h, Ω, y) is deﬁned by Deﬁnition 2.
We next assume that f ∈ C(
¯
Ω, R
n
) and satisﬁes (2). Then for ǫ > 0 suﬃ
ciently small there exists g ∈ C
1
(Ω, R
n
)
C(
¯
Ω, R
n
) such that y ,∈ g(∂Ω) and
f −g = max
x∈
¯
Ω
[f(x) −g(x)[ < ǫ/4
and there exists, by Lemma 8, ˜ g satisfying (1), (2), (3) such that g − ˜ g <
ǫ/4 and if
˜
h satisﬁes (1), (2), (3) and g −
˜
h < ǫ/4, ˜ g −
˜
h < ǫ/2, then
d(˜ g, Ω, y) = d(
˜
h, Ω, y) provided ǫ is small enough. Thus f may be approximated
by functions ˜ g satisfying (1), (2), (3) and d(˜ g, Ω, y) = constant provided f −˜ g
is small enough. We therefore may deﬁne d(f, Ω, y) as follows.
11 Deﬁnition Let f ∈ C(
¯
Ω, R
n
) be such that y ,∈ f(∂Ω). Let
d(f, Ω, y) = lim
g→f
d(g, Ω, y) (20)
where g satisﬁes (1), (2), (3).
The number deﬁned by (20) is called the Brouwer degree of f at y relative
to Ω.
It follows from our considerations above that d(f, Ω, y) is also given by for
mula (7), for any φ which satisﬁes:
φ ∈ C([0, ∞), R), φ(0) = 0, φ(s) ≡ 0, s ≥ r > 0,
_
R
n
φ([x[)dx = 1,
where r < min
x∈∂Ω
[f(x) −y[.
3 Properties of the Brouwer Degree
We next proceed to establish some properties of the Brouwer degree of a
mapping which will be of use in computing the degree and also in extending the
deﬁnition to mappings deﬁned in inﬁnite dimensional spaces and in establishing
global solution results for parameter dependent equations.
3. PROPERTIES OF THE BROUWER DEGREE 39
12 Proposition (Solution property) Let f ∈ C(
¯
Ω, R
n
) be such that y ,∈ f(∂Ω)
and assume that d(f, Ω, y) ,= 0. Then the equation
f(x) = y (21)
has a solution in Ω.
The proof is a straightforward consequence of Deﬁnition 11 and is left as an
exercise.
13 Proposition (Continuity property) Let f ∈ C(
¯
Ω, R
n
) and y ∈ R
n
be such
that d(f, Ω, y) is deﬁned. Then there exists ǫ > 0 such that for all g ∈ C(
¯
Ω, R
n
)
and ˆ y ∈ R with f −g +[y − ˆ y[ < ǫ
d(f, Ω, y) = d(g, Ω, ˆ y).
The proof again is left as an exercise.
The proposition has the following important interpretation.
14 Remark If we let C = ¦f ∈ C(
¯
Ω, R
n
) : y / ∈ f(∂Ω)¦ then C is a metric space
with metric ρ deﬁned by ρ(f, g) = f −g. If we deﬁne the mapping d : C →N
(integers) by d(f) = d(f, Ω, y), then the theorem asserts that d is a continuous
function from C to N (equipped with the discrete topology). Thus d will be
constant on connected components of C.
Using this remark one may establish the following result.
15 Proposition (Homotopy invariance property) Let f, g ∈ C(
¯
Ω, R
n
) with
f(x) and g(x) ,= y for x ∈ ∂Ω and let h : [a, b]
¯
Ω → R
n
be continuous such
that h(t, x) ,= y, (t, x) ∈ [a, b] ∂Ω. Further let h(a, x) = f(x), h(b, x) = g(x),
x ∈
¯
Ω. Then
d(f, Ω, y) = d(g, Ω, y);
more generally, d(h(t, ), Ω, y) = constant for a ≤ t ≤ b.
The next corollary may be viewed as an extension of Rouch´e’s theorem concern
ing the equal number of zeros of certain analytic functions. This extension will
be the content of one of the exercises at the end of this chapter.
16 Corollary Let f ∈ C(
¯
Ω, R
n
) be such that d(f, Ω, y) is deﬁned. Let g ∈
C(
¯
Ω, R
n
) be such that [f(x) − g(x)[ < [f(x) − y[, x ∈ ∂Ω. Then d(f, Ω, y) =
d(g, Ω, y).
Proof. For 0 ≤ t ≤ 1 and x ∈ ∂Ω we have that
[y −tg(x) −(1 −t)f(x)[ = [(y −f(x)) −t(g(x) −f(x))[
≥ [y −f(x)[ −t[g(x) −f(x)[
40
> 0 since 0 ≤ t ≤ 1,
hence h : [0, 1]
¯
Ω → R
n
given by h(t, x) = tg(x) + (1 − t)f(x) satisﬁes the
conditions of Proposition 15 and the conclusion follows from that proposition.
As an immediate corollary we have the following:
17 Corollary Assume that f and g are mappings such that f(x) = g(x), x ∈ ∂Ω,
then d(f, Ω, y) = d(g, Ω, y) if the degree is deﬁned, i.e. the degree only depends
on the boundary data.
18 Proposition (Additivity property) Let Ω be a bounded open set which is
the union of m disjoint open sets Ω
1
, , Ω
m
, and let f ∈ C(
¯
Ω, R
n
) and y ∈ R
n
be such that y ,∈ f(∂Ω
i
), i = 1, , m. Then
d(f, Ω, y) =
m
i=1
d(f, Ω
i
, y).
19 Proposition (Excision property) Let f ∈ C(
¯
Ω, R
n
) and let K be a closed
subset of
¯
Ω such that y ,∈ f(∂Ω ∪ K). Then
d(f, Ω, y) = d(f, Ω ¸ K, y).
20 Proposition (Cartesian product formula) Assume that Ω = Ω
1
Ω
2
is a
bounded open set in R
n
with Ω
1
open in R
p
and Ω
2
open in R
q
, p + q = n.
For x ∈ R
n
write x = (x
1
, x
2
), x
1
∈ R
p
, x
2
∈ R
q
. Suppose that f(x) =
(f
1
(x
1
), f
2
(x
2
)) where f
1
:
¯
Ω
1
→ R
p
, f
2
:
¯
Ω
2
→ R
q
are continuous. Suppose
y = (y
1
, y
2
) ∈ R
n
is such that y
i
/ ∈ f
i
(∂Ω
i
), i = 1, 2. Then
d(f, Ω, y) = d(f
1
, Ω
1
, y
1
)d(f
2
, Ω
2
, y
2
). (22)
Proof. Using an approximation argument, we may assume that f, f
1
and f
2
satisfy also (1) and (3) (interpreted appropriately). For such functions we have
d(f, Ω, y) =
x∈f
−1
(y)
sgn det f
′
(x)
=
x∈f
−1
(y)
sgn det
_
_
f
′
1
(x
1
) 0
0 f
′
2
(x
2
)
_
_
=
x
i
∈ f
−1
(y
i
)
i = 1, 2
sgn det f
′
1
(x
1
) sgn det f
′
2
(x
2
)
=
2
i=1
xi∈f
−1
i
(yi)
sgn det f
′
i
(x
i
) = d(f
1
, Ω
1
, y
1
)d(f
2
, Ω
2
, y
2
).
To give an example to show how the above properties may be used we prove
Borsuk’s theorem and the Brouwer ﬁxed point theorem.
3. PROPERTIES OF THE BROUWER DEGREE 41
3.1 The theorems of Borsuk and Brouwer
21 Theorem (Borsuk) Let Ω be a symmetric bounded open neighborhood of
0 ∈ R
n
(i.e. if x ∈ Ω, then −x ∈ Ω) and let f ∈ C(
¯
Ω, R
n
) be an odd mapping
(i.e. f(x) = −f(−x)). Let 0 / ∈ f(∂Ω), then d(f, Ω, 0) is an odd integer.
Proof. Choose ǫ > 0 such that B
ǫ
(0) = ¦x ∈ R
n
: [x[ < ǫ¦ ⊂ Ω. Let
α : R
n
→R be a continuous function such that
α(x) ≡ 1, [x[ ≤ ǫ, α(x) = 0, x ∈ ∂Ω, 0 ≤ α(x) ≤ 1, x ∈ R
n
and put
g(x) = α(x)x + (1 −α(x))f(x)
h(x) =
1
2
[g(x) −g(−x)] ,
then h ∈ C(
¯
Ω, R
n
) is odd and h(x) = f(x), x ∈ ∂Ω, h(x) = x, [x[ < ǫ. Thus by
Corollary 16 and the remark following it
d(f, Ω, 0) = d(h, Ω, 0).
On the other hand, the excision and additivity property imply that
d(h, Ω, 0) = d(h, B
ǫ
(0), 0) + d(h, Ω ¸ B
ǫ
(0), 0),
where ∂B
ǫ
(0) has been excised. It follows from Deﬁnition 2 that d(h, B
ǫ
(0), 0) =
1, and it therefore suﬃces to show that (letting Θ = Ω¸ B
ǫ
(0)) d(h, Θ, 0) is an
even integer. Since
¯
Θ is symmetric, 0 / ∈ Θ, 0 / ∈ h(∂Θ), and h is odd one may
show (see [40]) that there exists
˜
h ∈ C(
¯
Ω, R
n
) which is odd,
˜
h(x) = h(x), x ∈ ∂Θ
and is such that
˜
h(x) ,= 0, for those x ∈ Θ with x
n
= 0. Hence
d(h, Θ, 0) = d(
˜
h, Θ, 0) = d(
˜
h, Θ¸ ¦x : x
n
= 0¦, 0) (23)
where we have used the excision property. We let
Θ
1
= ¦x ∈ Θ : x
n
> 0¦, Θ
2
= ¦x ∈ Θ : x
n
< 0¦,
then by the additivity property
d(
˜
h, Θ¸ ¦x : x
n
= 0¦, 0) = d(
˜
h, Θ
1
, 0) + d(
˜
h, Θ
2
, 0).
Since Θ
2
= ¦x : −x ∈ Θ
1
¦ and
˜
h is odd one may now employ approximation
arguments to conclude that d(
˜
h, Θ
1
, 0) = d(
˜
h, Θ
2
, 0), and hence conclude that
the integer given by (23) is even.
22 Theorem (Brouwer ﬁxed point theorem) Let f ∈ C(
¯
Ω, R
n
), Ω = ¦x ∈
R
n
: [x[ < 1¦, be such that f :
¯
Ω →
¯
Ω. Then f has a ﬁxed point in Ω, i.e. there
exists x ∈
¯
Ω such that f(x) = x.
42
Proof. Assume f has no ﬁxed points in ∂Ω. Let h(t, x) = x−tf(x), 0 ≤ t ≤ 1.
Then h(t, x) ,= 0, 0 ≤ t ≤ 1, x ∈ ∂Ω and thus d(h(t, 0), Ω, 0) = d(h(0, 0), Ω, 0)
by the homotopy property. Since d(id, Ω, 0) = 1 it follows from the solution
property that the equation x −f(x) = 0 has a solution in Ω.
Theorem 22 remains valid if the unit ball of R
n
is replaced by any set home
omorphic to the unit ball (replace f by g
−1
fg where g is the homeomorphism).
That the Theorem also remains valid if the unit ball is replaced in arbitrary
compact convex set (or a set homeomorphic to it) may be proved using the
extension theorem of Dugundji (Theorem I.20).
4 Completely Continuous Perturbations of the
Identity in a Banach Space
4.1 Deﬁnition of the degree
Let E be a real Banach space with norm   and let Ω ⊂ E be a bounded
open set. Let F :
¯
Ω → E be continuous and let F(
¯
Ω) be contained in a ﬁnite
dimensional subspace of E. The mapping
f(x) = x +F(x) = (id +F)(x) (24)
is called a ﬁnite dimensional perturbation of the identity in E.
Let y be a point in E and let
˜
E be a ﬁnite dimensional subspace of E
containing y and F(
¯
Ω) and assume
y / ∈ f(∂Ω). (25)
Select a basis e
1
, , e
n
of
˜
E and deﬁne the linear homeomorphism T :
˜
E →R
n
by
T
_
n
i=1
c
i
e
i
_
= (c
1
, , c
n
) ∈ R
n
.
Consider the mapping
TFT
−1
: T(
¯
Ω ∩
˜
E) →R
n
,
then, since y / ∈ f(∂Ω), it follows that
T(y) / ∈ TfT
−1
(T(∂Ω ∩
˜
E)).
Let Ω
0
denote the bounded open set T(Ω∩
˜
E) in R
n
and let
˜
f = TfT
−1
, y
0
=
T(y). Then d(
˜
f, Ω
0
, y
0
) is deﬁned.
It is an easy exercise in linear algebra to show that the following lemma
holds.
4. COMPLETELY CONTINUOUS PERTURBATIONS 43
23 Lemma The integer d(
˜
f, Ω
0
, y
0
) calculated above is independent of the choice
the ﬁnite dimensional space
˜
E containing y and F(
¯
Ω) and the choice of basis of
˜
E.
We hence may deﬁne
d(f, Ω, y) = d(
˜
f, Ω
0
, y
0
),
where
˜
f, Ω
0
, y
0
are as above.
Recall that a mapping F :
¯
Ω → E is called completely continuous if F is
continuous and F(
¯
Ω) is precompact in E (i.e. F(Ω) is compact). More generally
if D is any subset of E and F ∈ C(D, E), then F is called completely continuous
if F(V ) is precompact for any bounded subset V of D.
We shall now demonstrate that if f = id+F, with F completely continuous
(f is called a completely continuous perturbation of the identity) and y ,∈ f(∂Ω),
then an integer valued function d(f, Ω, y) (the Leray Schauder degree of f at y
relative to Ω) may be deﬁned having much the same properties as the Brouwer
degree. In order to accomplish this we need the following lemma.
24 Lemma Let f :
¯
Ω →E be a completely continuous perturbation of the identity
and let y / ∈ f(∂Ω). Then there exists an integer d with the following property:
If h :
¯
Ω →E is a ﬁnite dimensional continuous perturbation of the identity such
that
sup
x∈Ω
f(x) −h(x) < inf
x∈∂Ω
f(x) −y, (26)
then y / ∈ h(∂Ω) and d(h, Ω, y) = d.
Proof. That y / ∈ h(∂Ω) follows from (26). Let h
1
and h
2
be any two such
mappings. Let k(t, x) = th
1
(x) + (1 − t)h
2
(x), 0 ≤ t ≤ 1, x ∈
¯
Ω, then if
t ∈ (0, 1) and x ∈
¯
Ω are such that k(t, x) = y it follows that
f(x) −y = f(x) −th
1
(x) −(1 −t)h
2
(x)
= t(f −(x) −h
1
(x)) + (1 −t)(f(x) −h
2
(x))
≤ tf(x) −h
1
(x) + (1 −t)f(x) −h
2
(x)
< inf
x∈∂Ω
f(x) −y (see (26)),
from which follows that x / ∈ ∂Ω. Let
˜
E be a ﬁnite dimensional subspace contain
ing y and (h
i
− id)(
¯
Ω), then d(h
i
, Ω, y) = d(Th
i
T
−1
, T(Ω ∩
˜
E), T(y)), where T
is given as above. Then (k(t, ) −id)(
¯
Ω) is contained in
˜
E and Tk(t, )T
−1
(x) ,=
T(y), x ∈ T(∂Ω ∩
˜
E). Hence we may use the homotopy invariance property of
Brouwer degree to conclude that
d(Tk(t, )T
−1
, T(∂Ω ∩
˜
E), T(y)) = constant,
i.e.,
d(h
1
, Ω, y) = d(h
2
, Ω, y).
44
It follows from Lemma 24 that if a ﬁnite dimensional perturbation of the
identity h exists which satisﬁes (26), then we may deﬁne d(f, Ω, y) = d, where d
is the integer whose existence follows from this lemma. In order to accomplish
this we need an approximation result.
Let M be a compact subset of E. Then for every ǫ > 0 there exists a ﬁnite
covering of M by spheres of radius ǫ with centers at y
1
, , y
n
∈ M. Deﬁne
µ
i
: M →[0, ∞) by
µ
i
(y) = ǫ −y −y
i
, if y −y
i
 ≤ ǫ
= 0, otherwise
and let
λ
i
(y) =
µ
i
(y)
n
j=1
µ
j
(y)
, 1 ≤ i ≤ n.
Since not all µ
i
vanish simultaneously, λ
i
(y) is non–negative and continuous on
M and further
n
i=1
λ
i
(y) = 1. The operator P
ǫ
deﬁned by
P
ǫ
(y) =
n
i=1
λ
i
(y)y
i
(27)
is called a Schauder projection operator on M determined by ǫ, and y
1
, , y
n
.
Such an operator has the following properties:
25 Lemma • P
ǫ
: M →co¦y
1
, , y
n
¦
(the convex hull of y
1
, , y
n
) is continuous.
• P
ǫ
(M) is contained in a ﬁnite dimensional subspace of E.
• P
ǫ
y −y ≤ ǫ, y ∈ M.
26 Lemma Let f :
¯
Ω →E be a completely continuous perturbation of the identity.
Let y / ∈ f(∂Ω). Let ǫ > 0 be such that ǫ < inf
x∈∂Ω
f(x) − y. Let P
ǫ
be a Schauder projection operator determined by ǫ and points ¦y
1
, , y
n
¦ ⊂
(f −id)(
¯
Ω). Then d(id +P
ǫ
F, Ω, y) = d, where d is the integer whose existence
is established by Lemma 24.
Proof. The properties of P
ǫ
(cf Lemma 25) imply that for each x ∈
¯
Ω
P
ǫ
F(x) −F(x) ≤ ǫ inf
x∈∂Ω
f(x) −y,
and the mapping id+P
ǫ
F is a continuous ﬁnite dimensional perturbation of the
identity.
27 Deﬁnition The integer d whose existence has been established by Lemma 26
is called the LeraySchauder degree of f relative to Ω and the point y and is
denoted by
d(f, Ω, y).
4. COMPLETELY CONTINUOUS PERTURBATIONS 45
4.2 Properties of the degree
28 Proposition The Leray–Schauder degree has the solution, continuity, homo
topy invariance, additivity, and excision properties similar to the Brouwer de
gree; the Cartesian product formula also holds.
Proof. (solution property). We let Ω be a bounded open set in E and f :
¯
Ω →E be a completely continuous perturbation of the identity, y a point in E
with y / ∈ f(∂Ω) and d(f, Ω, y) ,= 0. We claim that the equation f(x) = y has
a solution in E. To see this let ¦ǫ
n
¦
∞
n=1
be a decreasing sequence of positive
numbers with lim
n→∞
ǫ
n
= 0 and ǫ
1
< inf
x∈∂Ω
f(x)−y. Let P
ǫi
be associated
Schauder projection operators. Then
d(f, Ω, y) = d(id +P
ǫn
F, Ω, y), n = 1, 2, ,
where
d(id +P
ǫn
F, Ω, y) = d(id +T
n
P
ǫn
FT
−1
n
, T
n
(Ω ∩
˜
E
n
), T
n
(y)),
and the spaces
˜
E
n
are ﬁnite dimensional. Hence the solution property of
Brouwer degree implies the existence of a solution z
n
∈ T
n
(Ω ∩
˜
E
n
) of the
equation
z +T
n
P
ǫn
FT
−1
n
(z) = T(y),
or equivalently a solution x
n
∈ Ω ∩
˜
E
n
of
x +P
ǫn
F(x) = y.
The sequence ¦x
n
¦
∞
n=1
is a bounded sequence (¦x
n
¦
∞
n=1
⊂ Ω), thus, since F is
completely continuous there exists a subsequence ¦x
ni
¦
∞
i=1
such that F(x
ni
) →
u ∈ F(
¯
Ω). We relabel the subsequence and call it again ¦x
n
¦
∞
n=1
. Then
x
n
−x
m
 = P
ǫn
F(x
n
) −P
ǫm
F(x
m
)
≤ P
ǫn
F(x
n
) −F(x
n
) +P
ǫm
F(x
m
) −F(x
m
)
+ F(x
n
) −F(x
m
) < ǫ
n
+ǫ
m
+F(x
n
) −F(x
m
).
We let ǫ > 0 be given and choose N such that n, m ≥ N imply ǫ
n
, ǫ
m
< ǫ/3 and
F(x
n
) −F(x
m
) < ǫ/3. Thus x
n
−x
m
 < ǫ. The sequence ¦x
n
¦
∞
n=1
therefore
is a Cauchy sequence, hence has a limit, say x. It now follows that x ∈
¯
Ω and
solves the equation f(x) = y, hence, since y / ∈ f(∂Ω) we have that x ∈ Ω.
4.3 Borsuk’s theorem and ﬁxed point theorems
29 Theorem (Borsuk’s theorem) Let Ω be a bounded symmetric open neigh
borhood of 0 ∈ E and let f :
¯
Ω →E be a completely continuous odd perturba
tion of the identity with 0 / ∈ f(∂Ω). Then d(f, Ω, 0) is an odd integer.
Proof. Let ǫ > 0 be such that ǫ < inf
x∈∂Ω
f(x) and let P
ǫ
be an associated
Schauder projection operator. Let f
ǫ
= id + P
ǫ
f and put h(x) = 1/2[f
ǫ
(x) −
46
f
ǫ
(−x)]. Then h is a ﬁnite dimensional perturbation of the identity which is
odd and
h(x) −f(x) ≤ ǫ.
Thus d(f, Ω, 0) = d(h, Ω ∩
˜
E, 0), but
d(h, Ω, 0) = d(ThT
−1
, T(Ω∩
˜
E), 0).
On the other hand T(Ω ∩
˜
E) is a symmetric bounded open neighborhood of
0 ∈ T(Ω ∩
˜
E) and ThT
−1
is an odd mapping, hence the result follows from
Theorem 21.
We next establish extensions of the Brouwer ﬁxed point theorem to Banach
spaces.
30 Theorem (Schauder ﬁxed point theorem) Let K be a compact convex sub
set of E and let F : K →K be continuous. Then F has a ﬁxed point in K.
Proof. Since K is a compact there exists r > 0 such that K ⊂ B
r
(0) = ¦x ∈
E : x < r¦. Using the Extension Theorem (Theorem I.20) we may continu
ously extend F to B
r
(0). Call the extension
˜
F. Then
˜
F(B
r
(0)) ⊂ coF(K) ⊂ K,
where coF(K) is the convex hull of F(K), i.e. the smallest convex set containing
F(K). Hence
˜
F is completely continuous. Consider the homotopy
k(t, x) = x −t
˜
F(x), 0 ≤ t ≤ 1.
Since t
˜
F(K) ⊂ tK ⊂ B
r
(0), 0 ≤ t ≤ 1, x ∈ B
r
(0), it follows that k(t, x) ,= 0,
0 ≤ t ≤ 1, x ∈ ∂B
r
(0). Hence by the homotopy invariance property of the
Leray–Schauder degree
constant = d(k(t, ), B
r
(0), 0) = d(id, B
r
(0), 0) = 1.
The solution property of degree therefore implies that the equation
x −
˜
F(x) = 0
has a solution in x ∈ B
r
(0), and hence in K, i.e.
x −F(x) = 0.
In many applications the mapping F is known to be completely continuous
but it is diﬃcult to ﬁnd a compact convex set K such that F : K →K, whereas
closed convex sets K having this property are more easily found. In such a case
the following result may be applied.
31 Theorem (Schauder) Let K be a closed, bounded, convex subset of E and
let F be a completely continuous mapping such that F : K → K. Then F has
a ﬁxed point in K.
Proof. Let
˜
K = coF(K). Then since F(K) is compact it follows from a
theorem of Mazur (see eg. [17], [38], and [43]) that
˜
K is compact and
˜
K ⊂ K.
Thus F :
˜
K →
˜
K and F has ﬁxed point in
˜
K by Theorem 30. Therefore F has
a ﬁxed point in K.
5. EXERCISES 47
5 Exercises
1. Let [a, b] be a compact interval in R and let f : [a, b] →R be a continuous
function such that f(a)f(b) ,= 0. Verify the following.
(i) If f(b) > 0 > f(a), then d(f, (a, b), 0) = 1
(ii) If f(b) < 0 < f(a), then d(f, (a, b), 0) = −1
(iii) If f(a)f(b) > 0, then d(f, (a, b), 0) = 0.
2. Identify R
2
with the complex plane. Let Ω be a bounded open subset of
R
2
and let f and g be functions which are analytic in Ω and continuous on
¯
Ω. Let f(z) ,= 0, z ∈ ∂Ω and assume that [f(z) −g(z)[ < [f(z)[, z ∈ ∂Ω.
Show that f and g have precisely the same number of zeros, counting
multiplicities, in Ω. This result is called Rouch´e’s theorem.
3. Let Ω be a bounded open subset of R
n
and let
f, g ∈ C(
¯
Ω, R
n
) with f(x) ,= 0 ,= g(x), x ∈ ∂Ω.
Further assume that
f(x)
[f(x)[
,=
−g(x)
[g(x)[
, x ∈ ∂Ω.
Show that d(f, Ω, 0) = d(g, Ω, 0).
4. Let Ω ⊂ R
n
be a bounded open neighborhood of 0 ∈ R
n
, let f ∈ C(Ω, R
n
)
be such that 0 / ∈ f(∂Ω) and either
f(x) ,=
x
[x[
[f(x)[, x ∈ ∂Ω
or
f(x) ,= −
x
[x[
[f(x)[, x ∈ ∂Ω.
Show that the equation f(x) = 0 has a solution in Ω.
5. Let Ω be as in Exercise 4 and let f ∈ C(
¯
Ω, R
n
) be such that 0 / ∈ f(∂Ω).
Let n be odd. Show there exists λ (λ ,= 0) ∈ R and x ∈ ∂Ω such that
f(x) = λx. (This is commonly called the hedgehog theorem.)
6. Let B
n
= ¦x ∈ R
n
: [x[ < 1¦, S
n−1
= ∂B
n
. Let f, g ∈ C(B
n
, R
n
) be such
that f(S
n−1
), g(S
n−1
) ⊂ S
n−1
and [f(x) − g(x)[ < 2, x ∈ S
n−1
. Show
that d(f, B
n
, 0) = d(g, B
n
, 0).
7. Let f be as in Exercise 6 and assume that f(S
n−1
) does not equal S
n−1
.
Show that d(f, B
n
, 0) = 0.
48
8. Let A be an n n real matrix for which 1 is not an eigenvalue. Let Ω
be a bounded open neighborhood of 0 ∈ R
n
. Show, using linear algebra
methods, that
d (id −A, Ω, 0) = (−1)
β
,
where β equals the sum of the algebraic multiplicities of all real eigenvalues
µ of A with µ > 1.
9. Let Ω ⊂ R
n
be a symmetric bounded open neighborhood of 0 ∈ R
n
and
let f ∈ C(
¯
Ω, R
n
) be such that 0 / ∈ f(∂Ω). Also assume that
f(x)
[f(x)[
,=
f(−x)
[f(−x)[
, x ∈ ∂Ω.
Show that d(f, Ω, 0) is an odd integer.
10. Let Ω be as in Exercise 9 and let f ∈ C(
¯
Ω, R
n
) be an odd function such
that f(∂Ω) ⊂ R
m
, where m < n. Show there exists x ∈ ∂Ω such that
f(x) = 0.
11. Let f and Ω be as in Exercise 10 except that f is not necessarily odd.
Show there exists x ∈ ∂Ω such that f(x) = f(−x).
12. Let K be a bounded, open, convex subset of E. Let F :
¯
K → E be
completely continuous and be such that F(∂K) ⊂ K. Then F has a ﬁxed
point in K.
13. Let Ω be a bounded open set in E with 0 ∈ Ω. Let F :
¯
Ω → E be
completely continuous and satisfy
x −F(x)
2
≥ F(x)
2
, x ∈ ∂Ω.
then F has a ﬁxed point in
¯
Ω.
14. Provide detailed proofs of the results of Section 3.
15. Provide detailed proofs of the results of Section 4.
Chapter IV
Global Solution Theorems
1 Introduction
In this chapter we shall consider a globalization of the implicit function theo
rem (see Chapter I) and provide some global bifurcation results. Our main tools
in establishing such global results will be the properties of the Leray Schauder
degree and a topological lemma concerning continua in compact metric spaces.
2 The Continuation Principle of LeraySchauder
In this section we shall extend the homotopy property of LeraySchauder
degree (Proposition III.28) to homotopy cylinders having variable cross sections
and from it deduce the LeraySchauder continuation principle. As will be seen
in later sections, this result also allows us to derive a globalization of the implicit
function theorem and results about global bifurcation in nonlinear equations.
Let O be a bounded open (in the relative topology) subset of E [a, b] ,
where E is a real Banach space, and let
F :
¯
O →E
be a completely continuous mapping. Let
f(u, λ) = u −F(u, λ) (1)
and assume that
f(u, λ) ,= 0, (u, λ) ∈ ∂O (2)
(here ∂O is the boundary of O in E [a, b]).
1 Theorem (The generalized homotopy principle) Let f be given by (1)
and satisfy (2). Then for a ≤ λ ≤ b,
d(f(, λ), O
λ
, 0) = constant,
(here O
λ
= ¦u ∈ E : (u, λ) ∈ O¦).
49
50
Proof. We may assume that O ,= ∅ and that
a = inf¦λ : O
λ
,= ∅¦, b = sup¦λ : O
λ
,= ∅¦.
We let
ˆ
O = O ∪ O
a
(a −ǫ, a] ∪ O
b
[b, b +ǫ),
where ǫ > 0 is ﬁxed. Then
ˆ
O is a bounded open subset of E R. Let
˜
F be
the extension of F to E R whose existence is guaranteed by the Dugundji
extension theorem (Theorem I.20). Let
˜
f(u, λ) = (u −
˜
F(u, λ), λ −λ
∗
),
where a ≤ λ
∗
≤ b is ﬁxed. Then
˜
f is a completely continuous perturbation of
the identity in E R.
Furthermore for any such λ
∗
˜
f(u, λ) ,= 0, (u, λ) ∈ ∂
ˆ
O,
and hence d(
˜
f,
ˆ
O, 0) is deﬁned and constant (for such λ
∗
). Let 0 ≤ t ≤ 1, and
consider the vector ﬁeld
˜
f
t
(u, λ) = (u −t
˜
F(u, λ) −(1 −t)
˜
F(u, λ
∗
), λ −λ
∗
),
then
˜
f
t
(u, λ) = 0 if and only if λ = λ
∗
and u =
˜
F(u, λ
∗
). Thus, our hypotheses
imply that
˜
f
t
(u, λ) ,= 0 for (u, λ) ∈ ∂
ˆ
O and t ∈ [0, 1]. By the homotopy
invariance principle (Proposition III.28) we therefore conclude that
d(
˜
f
1
,
ˆ
O, 0) = d(
˜
f,
ˆ
O, 0) = d(
˜
f
0
,
ˆ
O, 0).
On the other hand,
d(
˜
f
0
,
ˆ
O, 0) = d(
˜
f
0
, O
λ
∗ (a −ǫ, b +ǫ), 0),
by the excision property of degree (Proposition III.28). Using the Cartesian
product formula (Proposition III.28), we obtain
d(
˜
f
0
, O
λ
∗ (a −ǫ, b +ǫ), 0) = d(f(, λ
∗
), O
λ
∗, 0).
This completes the proof.
As an immediate consequence we obtain the continuation principle of Leray
Schauder.
2 Theorem (Leray–Schauder Continuation Theorem) Let O be a bounded
open subset of E [a, b] and let f :
¯
O → E be given by (1) and satisfy (2).
Furthermore assume that
d(f(, a), O
a
, 0) ,= 0.
Let
S = ¦(u, λ) ∈
¯
O : f(u, λ) = 0¦.
Then there exists a closed connected set C in S such that
C
a
∩ O
a
,= ∅ ,= C
b
∩ O
b
.
2. CONTINUATION PRINCIPLE 51
Proof. It follows from Theorem 1 that
d(f(, a), O
a
, 0) = d(f(, b), O
b
, 0).
Hence
S
a
¦a¦ = A ,= ∅ ,= B = S
b
¦b¦.
Using the complete continuity of F we may conclude that S is a compact metric
subspace of E [a, b]. We now apply Whyburn’s lemma (see [44]) with X = S.
If there is no such continuum (as asserted above) there will exist compact sets
X
A
, X
B
in X such that
A ⊂ X
A
, B ⊂ X
B
, X
A
∩ X
B
= ∅, X
A
∪ X
B
= X.
We hence may ﬁnd an open set U ⊂ E [a, b] such that A ⊂ U ∩ O = V and
S ∩ ∂V = ∅ = V
b
. Therefore
d(f(, λ), V
λ
, 0) = constant, λ ≥ a.
On the other hand, the excision principle implies that
d(f(, a), V
a
, 0) = d(f(, a), O
a
, 0).
Since V
b
= ∅, these equalities yield a contradiction, and there exists a con
tinuum as asserted.
In the following examples we shall develop, as an application of the above
results, some basic existence results for the existence of solutions of nonlinear
boundary value problems.
3 Example Let I = [0, 1] and let g : [0, 1] R →R be continuous. Consider the
nonlinear Dirichlet problem
_
u
′′
+g(x, u) = 0, in I
u = 0, on ∂I.
(3)
Let there exist constants a < 0 < b such that
g(x, a) > 0 > g(x, b), x ∈ Ω.
Then (3) has a solution u ∈ C
2
([0, 1], R) such that
a < u(x) < b, x ∈ I.
Proof. To see this, we consider the one parameter family of problems
_
u
′′
+λg(x, u) = 0, in I
u = 0, on ∂I.
(4)
Let G be deﬁned by
G(u)(x) = g(x, u(x)),
52
then (4) is equivalent to the operator equation
u = λLG(u), u ∈ C([0, 1], R) = E, (5)
where for each v ∈ E, w = LG(v) is the unique solution of
w
′′
+g(x, v) = 0, in I
w = 0, on ∂I.
It follows that for each v ∈ E, LG(v) ∈ C
2
(I) and since C
2
(I) is compactly
embedded in E that
LG() : E →E
is a completely continuous operator. Let O = ¦(u, λ) : u ∈ E, a < u(x) <
b, x ∈ I, 0 ≤ λ ≤ 1¦. Then O is an open and bounded set in E [0, 1]. If
(u, λ) ∈ ∂O is a solution of (4), then there will either exist x ∈ I such that
u(x) = b or there exists x ∈ I such that u(x) = a and λ > 0. In either case, (3)
yields, via elementary calculus, a contradiction. Hence (4) has no solutions in
∂O. Therefore
d(id −λLG, O
λ
, 0) = d(id, O
0
, 0) = 1,
and Theorem 2 implies the existence of a continuum C of solutions of (5), hence
of (4), such that C ∩ E ¦0¦ = ¦0¦ and C ∩ E ¦1¦ ,= ∅.
3 A Globalization of the Implicit Function The
orem
Assume that
F : E R →E
is a completely continuous mapping and consider the equation
f(u, λ) = u −F(u, λ) = 0. (6)
Let (u
0
, λ
0
) be a solution of (6) such that the condition of the implicit func
tion theorem (Theorem I.12) hold at (u
0
, λ
0
). Then there is a solution curve
¦(u(λ), λ)¦ of (6) deﬁned in a neighborhood of λ
0
, passing through (u
0
, λ
0
).
Furthermore the conditions of Theorem I.12 imply that the solution u
0
is an
isolated solution of (6) at λ = λ
0
, and if O is an isolating neighborhood, we
have that
d(f(, λ
0
), O, 0) ,= 0. (7)
We shall now show that condition (7) alone suﬃces to guarantee that equa
tion (6) has a global solution branch in the half spaces E [λ
0
, ∞) and E
(−∞, λ
0
].
3. A GLOBALIZATION OF THE IMPLICIT FUNCTION THEOREM 53
4 Theorem Let O be a bounded open subset of E and assume that for λ = λ
0
equation (6) has a unique solution in O and let (7) hold. Let
S
+
= ¦(u, λ) ∈ E [λ
0
, ∞) : (u, λ) solves (6)¦
and
S
−
= ¦(u, λ) ∈ E (−∞, λ
0
] : (u, λ) solves (6)¦.
Then there exists a continuum C
+
⊂ S
+
(C
−
⊂ S
−
) such that:
1. C
+
λ0
∩ O = ¦u
0
¦ (C
−
λ0
∩ O = ¦u
0
¦),
2. C
+
is either unbounded in E[λ
0
, ∞) (C
−
is unbounded in E(−∞, λ
0
])
or C
+
λ0
∩ (E¸O) ,= ∅ (C
−
λ0
∩ (E¸O) ,= ∅).
Proof. Let C
+
be the maximal connected subset of S
+
such that 1. above
holds. Assume that C
+
∩ (E¸O) = ∅ and that C
+
is bounded in E [λ
0
, ∞).
Then there exists a constant R > 0 such that for each (u, λ) ∈ C
+
we have that
[[u[[ +[λ[ < R. Let
S
+
2R
= ¦(u, λ) ∈ S
+
: [[u[[ +[λ[ ≤ 2R¦,
then S
+
2R
is a compact subset of E [λ
0
, ∞), and hence is a compact metric
space. There are two possibilities: Either S
+
2R
= C
+
or else there exists (u, λ) ∈
S
+
2R
such that (u, λ) / ∈ C
+
. In either case, we may ﬁnd a bounded open set
U ⊂ E [λ
0
, ∞) such that U
λ0
= O, S
+
2R
∩ ∂U = ∅, C
+
⊂ U. It therefore
follows from Theorem 1 that
d(f(, λ
0
), U
λ0
, 0) = constant, λ ≥ λ
0
,
where this constant is given by
d(f(, λ
0
), U
λ0
, 0) = d(f(, λ
0
), O, 0)
which is nonzero, because of (7). On the other hand, there exists λ
∗
> λ
0
such that U
λ
∗ contains no solutions of (6) and hence d(f(, λ
∗
), U
λ
∗ , 0) = 0,
contradicting (7). (To obtain the existence of an open set U with properties
given above, we employ again Whyburn’s lemma ([44]).)
The existence of C
−
with the above listed properties is demonstrated in a
similar manner.
5 Remark The assumption of Theorem 4 that u
0
is the unique solution of (6)
inside the set O, was made for convenience of proof. If one only assumes (7), one
may obtain the conclusion that the set of all such continua is either bounded in
the right (left) half space, or else there exists one such continuum which meets
the λ = λ
0
hyperplane outside the set O.
6 Remark If the component C
+
of Theorem 4 is bounded and
˜
O is an isolating
neighborhood of C
+
∩(E¸O) ¦λ
0
¦, then it follows from the excision property,
Whyburn’s lemma, and the generalized homotopy principle that
d(f(, λ
0
), O, 0) = −d(f(, λ
0
),
˜
O, 0).
54
This observation has the following important consequence. If equation (6)
has, for λ = λ
0
only isolated solutions and if the integer given by (7) has the
same sign with respect to isolating neighborhoods O for all such solutions where
(7) holds, then all continua C
+
must be unbounded.
7 Example Let p(z), z ∈ C, be a polynomial of degree n whose leading coeﬃ
cient is ( without loss in generality) assumed to be 1 and let q(z) =
n
i=1
(z−a
i
),
where a
1
, . . . , a
n
are distinct complex numbers. Let
f(z, λ) = λp(z) + (1 −λ)q(z).
Then f may be considered as a continuous mapping
f : R
2
R →R
2
.
Furthermore for λ ∈ [0, r], r > 0, there exists a constant R such that any
solution of
f(z, λ) = 0, (8)
satisﬁes [z[ < R. For all λ ≥ 0, (8) has only isolated solutions and for λ = 0
each such solution has the property that
d(f(, 0), O
i
, 0) = 1,
where O
i
is an isolating neighborhood of a
i
. Hence, for each i, there exists a
continuum C
+
i
of solutions of (8) which is unbounded with respect to λ, and
must therefore reach every λ−level, in particular, the level λ = 1. We conclude
that each zero of p(z) must be connected to some a
i
(apply the above argument
backwards from the λ = 1−level, if need be).
4 The Theorem of KreinRutman
In this section we shall employ Theorem 4 to prove an extension of the
PerronFrobenius theorem about eigenvalues for positive matrices. The Krein
Rutman theorem [26] is a generalization of this classical result to positive com
pact operators on a not necessarily ﬁnite dimensional Banach space.
Let E be a real Banach space and let K be a cone in E, i.e., a closed convex
subset of E with the properties:
• For all u ∈ K, t ≥ 0, tu ∈ K.
• K ∩ ¦−K¦ = ¦0¦.
It is an elementary exercise to show that a cone K induces a partial order ≤ on
E by the convention u ≤ v if and only if v−u ∈ K. A linear operator L : E →E
is called positive whenever K is an invariant set for L, i.e. L : K →K. If K is
a cone whose interior intK is nonempty, we call L a strongly positive operator,
whenever L : K¸¦0¦ →intK.
4. THE THEOREM OF KREINRUTMAN 55
8 Theorem Let E be a real Banach space with a cone K and let L : E →E be
a positive compact linear operator. Assume there exists w ∈ K, w ,= 0 and a
constant m > 0 such that
w ≤ mLw, (9)
where ≤ is the partial order induced by K. Then there exists λ
0
> 0 and
u ∈ K, [[u[[ = 1, such that
u = λ
0
Lu. (10)
Proof. Restrict the operator L to the cone K and denote by
˜
L the Dugundji
extension of this operator to E. Since L is a compact linear operator, the
operator
˜
L is a completely continuous mapping with
˜
L(E) ⊂ K. Choose ǫ > 0
and consider the equation
u −λ
˜
L(u +ǫw) = 0. (11)
For λ = 0, equation (11) has the unique solution u = 0 and we may apply
Theorem 4 to obtain an unbounded continuum C
+
ǫ
⊂ E [0, ∞) of solutions
of (11). Since
˜
L(E) ⊂ K, we have that u ∈ K, whenever (u, λ) ∈ C
+
ǫ
, and
therefore u = λL(u +ǫw). Thus
λLu ≤ u,
λǫ
m
≤ λǫLw ≤ u.
Applying L to this last inequality repeatedly, we obtain by induction that
_
λ
m
_
n
ǫw ≤ u. (12)
Since w ,= 0, by assumption, it follows from (12) that λ ≤ m. Thus, if (u, λ) ∈
C
+
ǫ
, it must be case that λ ≤ m, and hence that C
+
ǫ
⊂ K [0, m]. Since C
+
ǫ
is unbounded, we conclude that for each ǫ > 0, there exists λ
ǫ
> 0, u
ǫ
∈
K, [[u
ǫ
[[ = 1, such that
u
ǫ
= λ
ǫ
L(u
ǫ
+ǫw).
Since L is compact, the set ¦(u
ǫ
, λ
ǫ
)¦ will contain a convergent subsequence
(letting ǫ →0), converging to, say, (u, λ
0
). Since clearly [[u[[ = 1, it follows that
λ
0
> 0.
If it is the case that L is a strongly positive compact linear operator, much
more can be asserted; this will be done in the theorem of KreinRutman which
we shall establish as a corollary of Theorem 8.
9 Theorem Let E have a cone K, whose interior, intK ,= ∅. Let L be a strongly
positive compact linear operator. Then there exists a unique λ
0
> 0 with the
following properties:
1. There exists u ∈ intK, with u = λ
0
Lu.
56
2. If λ(∈ R) ,= λ
0
is such that there exists v ∈ E, v ,= 0, with v = λLv,
then v / ∈ K ∪ ¦−K¦ and λ
0
< [λ[.
Proof. Choose w ∈ K¸¦0¦, then, since Lw ∈ intK, there exists δ > 0,
small such that Lw − δw ∈ intK, i.e., in terms of the partial order δw ≤ Lw.
We therefore may apply Theorem 8 to obtain λ
0
> 0 and u ∈ K such that
u = λ
0
Lu. Since L is strongly positive, we must have that u ∈ intK. If
(v, λ) ∈ (K¸¦0¦) (0, ∞) is such that v = λLv, then v ∈ intK. Hence, for all
δ > 0, suﬃciently small, we have that u−δv ∈ intK. Consequently, there exists
a maximal δ
∗
> 0, such that u −δ
∗
v ∈ K, i.e. u −rv / ∈ K, r > δ
∗
. Now
L(u −δ
∗
v) =
1
λ
0
(u −
λ
0
λ
δ
∗
v),
which implies that u −
λ0
λ
δ
∗
v ∈ intK, unless u − δ
∗
v = 0. If the latter holds,
then λ
0
= λ, if not, then λ
0
< λ, because δ
∗
is maximal. If λ
0
< λ, we may
reverse the role of u and v and also obtain λ < λ
0
, a contradiction. Hence it
must be the case that λ = λ
0
. We have therefore proved that λ
0
is the only
characteristic value of L having an eigendirection in the cone K and further
that any other eigenvector corresponding to λ
0
must be a constant multiple of
u, i.e. λ
0
is a characteristic value of L of geometric multiplicity one, i.e the
dimension of the kernel of id −λ
0
L equals one.
Next let λ ,= λ
0
be another characteristic value of L and let v ,= 0 be such
that v / ∈ K∪¦−K¦. Again, for [δ[ small, u−δv ∈ intK and there exists δ
∗
> 0,
maximal, such that u − δ
∗
v ∈ K, and there exists δ
∗
< 0, minimal, such that
u −δ
∗
v ∈ K. Now
L(u −δ
∗
v) =
1
λ
0
(u −
λ
0
λ
δ
∗
v) ∈ K,
and
L(u −δ
∗
v) =
1
λ
0
(u −
λ
0
λ
δ
∗
v) ∈ K.
Thus, if λ > 0, we conclude that λ
0
< λ, whereas, if λ < 0, we get that
λ
0
δ
∗
< λδ
∗
, and λ
0
δ
∗
> λδ
∗
, i.e λ
2
0
< λ
2
.
As observed above we have that λ
0
is a characteristic value of geometric
multiplicity one. Before giving an application of the above result, we shall
establish that λ
0
, in fact also has algebraic multiplicity one. Recall from the
Riesz theory of compact linear operators (viz. [27], [43]) that the operator
id −λ
0
L has the following property:
There exists a minimal integer n such that
ker(id −λ
0
L)
n
= ker(id −λ
0
L)
n+1
= ker(id −λ
0
L)
n+2
= . . . ,
and the dimension of the generalized eigenspace ker(id − λ
0
L)
n
is called
the algebraic multiplicity of λ
0
.
5. GLOBAL BIFURCATION 57
With this terminology, we have the following addition to Theorem 9.
10 Theorem Assume the conditions of Theorem 9 and let λ
0
be the characteristic
value of L, whose existence is established there. Then λ
0
is a characteristic
value of L of algebraic multiplicity one.
Proof. We assume the contrary. Then, since ker(id −λ
0
L) has dimension one
(Theorem 9), it follows that there exists a smallest integer n > 1 such that the
generalized eigenspace is given by ker(id−λ
0
L)
n
. Hence, there exists a nonzero
v ∈ E such that (id − λ
0
L)
n
v = 0 and (id − λ
0
L)
n−1
v = w ,= 0. It follows
from Theorem 9 and its proof that w = ku, where u is given by Theorem 9 and
k may assumed to be positive. Let z = (id − λ
0
L)
n−2
v, then z − λ
0
Lz = ku,
and hence, by induction, we get that λ
m
0
L
m
z = z − mku, for any positive
integer m. It follows therefore that z / ∈ K, for otherwise
1
m
z −ku ∈ K, for any
integer m, implying that −ku ∈ K, a contradiction. Since u ∈ intK, there exist
α > 0 and y ∈ K such that z = αu − y. Then λ
m
0
L
m
z = αu − λ
m
0
L
m
y, or
λ
m
0
L
m
y = y + mku. Choose β > 0, such that y ≤ βu, then λ
m
0
L
m
y ≤ βu, and
by the above we see that y + mku ≤ βu. Dividing this inequality my m and
letting m →∞, we obtain that ku ∈ −K, a contradiction.
11 Remark It may be the case that, aside from real characteristic values, L also
has complex ones. If µ is such a characteristic value, then it may be shown
that [µ[ > λ
0
, where λ
0
is as in Theorem 10. We refer the interested reader to
Krasnosel’skii [24] for a veriﬁcation.
5 Global Bifurcation
As before, let E be a real Banach space and let f : E R → E have the
form
f(u, λ) = u −F(u, λ), (13)
where F : E R →E is completely continuous. We shall now assume that
F(0, λ) ≡ 0, λ ∈ R, (14)
and hence that the equation
f(u, λ) = 0, (15)
has the trivial solution for all values of λ. We shall now consider the question of
bifurcation from this trivial branch of solutions and demonstrate the existence
of global branches of nontrivial solutions bifurcating from the trivial branch.
Our main tools will again be the properties of the LeraySchauder degree and
Whyburn’s lemma.
We shall see that this result is an extension of the local bifurcation theorem,
Theorem II.6.
58
12 Theorem Let there exist a, b ∈ R with a < b, such that u = 0 is an isolated
solution of (15) for λ = a and λ = b, where a, b are not bifurcation points,
furthermore assume that
d(f(, a), B
r
(0), 0) ,= d(f(, b), B
r
(0), 0), (16)
where B
r
(0) = ¦u ∈ E : [[u[[ < r¦ is an isolating neighborhood of the trivial
solution. Let
S = ¦(u, λ) : (u, λ) solves (15) with u ,= 0¦ ∪ ¦0¦ [a, b]
and let C ⊂ S be the maximal connected subset of S which contains ¦0¦[a, b].
Then either
(i) C is unbounded in E R,
or else
(ii) C ∩ ¦0¦ (R¸[a, b]) ,= ∅.
Proof. Deﬁne a class U of subsets of E R as follows
U = ¦Ω ⊂ E R : Ω = Ω
0
∪ Ω
∞
¦,
where Ω
0
= B
r
(0) [a, b], and Ω
∞
is a bounded open subset of (E¸¦0¦) R.
We shall ﬁrst show that (15) has a nontrivial solution (u, λ) ∈ ∂Ω for any such
Ω ∈ U. To accomplish this, let us consider the following sets:
_
_
_
K = f
−1
(0) ∩ Ω,
A = ¦0¦ [a, b],
B = f
−1
(0) ∩ (∂Ω¸(B
r
(0) ¦a¦ ∪ B
r
(0) ¦b¦)).
(17)
We observe that K may be regarded as a compact metric space and A and
B are compact subsets of K. We hence may apply Whyburn’s lemma to deduce
that either there exists a continuum in K connecting A to B or else, there is
a separation K
A
, K
B
of K, with A ⊂ K
A
, B ⊂ K
B
. If the latter holds,
we may ﬁnd open sets U, V in E R such that K
A
⊂ U, K
B
⊂ V , with
U ∩ V = ∅. We let Ω
∗
= Ω ∩ (U ∪ V ) and observe that Ω
∗
∈ U. It follows, by
construction, that there are no nontrivial solutions of (15) which belong to ∂Ω
∗
;
this, however, is impossible, since, it would imply, by the generalized homotopy
and the excision principle of LeraySchauder degree, that d(f(, a), B
r
(0), 0) =
d(f(, b), B
r
(0), 0), contradicting (16). We hence have that for each Ω ∈ U there
is a continuum C of solutions of (15) which intersects ∂Ω in a nontrivial solution.
We assume now that neither of the alternatives of the theorem hold, i.e we
assume that C is bounded and C ∩ ¦0¦ (R¸[a, b]) = ∅. In this case, we may,
using the boundednes of C, construct a set Ω ∈ U, containing no nontrivial
solutions in its boundary, thus arriving once more at a contradiction.
We shall, throughout this text, apply the above theorem to several problems
for nonlinear diﬀerential equations. Here we shall, for the sake of illustration
provide two simple one dimensional examples.
5. GLOBAL BIFURCATION 59
13 Example Let f : R R →R, be given by
f(u, λ) = u(u
2
+λ
2
−1).
It is easy to see that S is given by
S = ¦(u, λ) : u
2
+λ
2
= 1¦ ∪ ¦0¦ (−∞, ∞),
and hence that (0, −1) and (0, 1) are the only bifurcation points from the trivial
solution. Furthermore, the bifurcating continuum is bounded. Also one may
quickly check that (16) holds with a, b chosen in a neighborhood of λ = −1 and
also in a neighborhood of λ = 1.
14 Example Let f : R R →R be given by
f(u, λ) = (1 −λ)u +u sin
1
u
.
In this case S is given by
S = ¦(u, λ) : λ −1 = sin
1
u
¦ ∪ ¦0¦ [0, 2],
which is an unbounded set, and we may check that (16) holds, by choosing a < 0
and b > 2.
In many interesting cases the nonlinear mapping F is of the special form
F(u, λ) = λBu +o([[u[[), as [[u[[ →0, (18)
where B is a compact linear operator. In this case bifurcation points from the
trivial solution are isolated, in fact one has the following necessary conditions
for bifurcation.
15 Proposition Assume that F has the form (18), where B is the Fr´echet deriva
tive of F. If (0, λ
0
) is a bifurcation point from the trivial solution for equation
(15), then λ
0
is a characteristic value of B.
Using this result, Theorem 12, and the LeraySchauder formula for comput
ing the degree of a compact linear perturbation of the identity (an extension to
inﬁnite dimensions of Exercise 8 of Chapter III), we obtain the following result.
16 Theorem Assume that F has the form (18) and let λ
0
be a characteristic value
of B which is of odd algebraic multiplicity. Then there exists a continuum C of
nontrivial solutions of (15) which bifurcates from the set of trivial solutions at
(0, λ
0
) and C is either unbounded in E R or else C also bifurcates from the
trivial solution set at (0, λ
1
), where λ
1
is another characteristic value of B.
60
Proof. Since λ
0
is isolated as a characteristic value, we may ﬁnd a < λ
0
< b
such that the interval [a, b] contains, besides λ
0
, no other characteristic values.
It follows that the trivial solution is an isolated solution (in E) of (15) for λ = a
and λ = b. Hence, d(f(, a), B
r
(0), 0) and d(f(, b), B
r
(0), 0) are deﬁned for
r, suﬃciently small and are, respectively, given by d(id − aB, B
r
(0), 0), and
d(id −bB, B
r
(0), 0). On the other hand,
d(id −aB, B
r
(0), 0) = (−1)
β
d(id −bB, B
r
(0), 0),
where β equals the algebraic multiplicity of λ
0
as a characteristic value of B.
Since β is odd, by assumption, the result follows from Theorem 12 and Propo
sition 15.
The following example serves to demonstrate that, in general, not every
characteristic value will yield a bifurcation point.
17 Example The system of scalar equations
_
x = λx +y
3
y = λy −x
3
(19)
has only the trivial solution x = 0 = y for all values of λ. We note, that λ
0
= 1
is a characteristic value of the Fr´echet derivative of multiplicity two.
As a further example let us consider a boundary value problem for a second
order ordinary diﬀerential equation, the pendulum equation.
18 Example Consider the boundary value problem
_
u
′′
+λsin u = 0, x ∈ [0, π]
u(0) = 0, u(π) = 0.
(20)
As already observed this problem is equivalent to an operator equation
u = λF(u),
where
F : C[0, π] →C[0, π]
is a completely continuous operator which is continuously Fr´echet diﬀerentiable
with Fr´echet derivative F
′
(0). Thus to ﬁnd the bifurcation points for (20) we
must compute the eigenvalues of F
′
(0). On the other hand, to ﬁnd these eigen
values is equivalent to ﬁnding the values of λ for which
_
u
′′
+λu = 0, x ∈ [0, π]
u(0) = 0, u(π) = 0
(21)
has nontrivial solutions. These values are given by
λ = 1, 4, k
2
, , k ∈ N.
6. EXERCISES 61
Furthermore we know from elementary diﬀerential equations that each such
eigenvalue has a onedimensonal eigenspace and one may convince oneself that
the above theorem may be applied at each such eigenvalue and conclude that
each value
(0, k
2
), k ∈ N
is a bifurcation point for (20).
6 Exercises
1. Prove Proposition 15.
2. Supply the details for the proof of Theorem 12.
3. Perform the calculations indicated in Example 13.
4. Perform the calculations indicated in Example 14.
5. Prove Proposition 15.
6. Supply the details for the proof of Theorem 16.
7. Prove the assertion of Example 17.
8. Provide the details for Example 18.
9. In Example 18 show that the second alternative of Theorem 16 cannot
hold.
62
Part II
Ordinary Diﬀerential
Equations
63
Chapter V
Existence and Uniqueness
Theorems
1 Introduction
In this chapter, we shall present the basic existence and uniqueness theo
rems for solutions of initial value problems for systems of ordinary diﬀerential
equations. To this end let D be an open connected subset of R R
N
, N ≥ 1,
and let
f : D →R
N
be a continuous mapping.
We consider the diﬀerential equation
u
′
= f(t, u),
′
=
d
dt
. (1)
and seek suﬃcient conditions for the existence of solutions of (1), where u ∈
C
1
(I, R
N
), with I an interval, I ⊂ R, is called a solution, if (t, u(t)) ∈ D, t ∈ I
and
u
′
(t) = f(t, u(t)), t ∈ I.
Simple examples tell us that a given diﬀerential equation may have a multi
tude of solutions, in general, whereas some constraints on the solutions sought
might provide existence and uniqueness of the solution. The most basic such
constraints are given by ﬁxing an initial value of a solution. By an initial value
problem we mean the following:
• Given a point (t
0
, u
0
) ∈ D we seek a solution u of (1) such that
u(t
0
) = u
0
. (2)
We have the following proposition whose proof is straightforward:
65
66
1 Proposition A function u ∈ C
1
(I, R
N
), with I ⊂ R, and I an interval con
taining t
0
is a solution of the initial value problem (1), satisfying the initial
condition (2) if and only if (t, u(t)) ∈ D, t ∈ I and
u(t) = u(t
0
) +
_
t
t0
f(s, u(s))ds. (3)
We shall now, using Proposition 1, establish some of the classical and basic
existence and existence/uniqueness theorems.
2 The PicardLindel¨of Theorem
We say that f satisﬁes a local Lipschitz condition on the domain D, provided
for every compact set K ⊂ D, there exists a constant L = L(K), such that for
all (t, u
1
), (t, u
2
) ∈ K
[f(t, u
1
) −f(t, u
2
)[ ≤ L[u
1
−u
2
[.
For such functions, one has the following existence and uniqueness theorem.
This result is usually called the PicardLindel¨of theorem
2 Theorem Assume that f : D → R
N
satisﬁes a local Lipschitz condition on
the domain D, then for every (t
0
, u
0
) ∈ D equation (1) has a unique solution
satisfying the initial condition (2) on some interval I.
We remark that the theorem as stated is a local existence and uniqueness
theorem, in the sense that the interval I, where the solution exists will depend
upon the initial condition. Global results will follow from this result, by extend
ing solutions to maximal intervals of existence, as will be seen in a subsequent
section.
Proof. Let (t
0
, u
0
) ∈ D, then, since D is open, there exist positive constants
a and b such that
Q = ¦(t, u) : [t −t
0
[ ≤ a, [u −u
0
[ ≤ b¦ ⊂ D.
Let L be the Lipschitz constant for f associated with the set Q. Further let
m = max
(t,u)∈Q
[f(t, u)[,
α = min¦a,
b
m
¦.
Let
˜
L be any constant,
˜
L > L, and deﬁne
M = ¦u ∈ C
_
[t
0
−α, t
0
+ α], R
N
_
: [u(t) −u
0
[ ≤ b, [t −t
0
[ ≤ α¦.
In C
_
[t
0
−α, t
0
+α], R
N
_
we deﬁne a new norm as follows:
u = max
t−t0≤α
e
−
˜
Lt−t0
[u(t)[.
3. THE CAUCHYPEANO THEOREM 67
And we let ρ(u, v) = u − v, then (M, ρ) is a complete metric space. Next
deﬁne the operator T on M by:
(Tu)(t) = u
0
+
_
t
t0
f(s, u(s))ds, [t −t
0
[ ≤ α. (4)
Then
[(Tu)(t) −u
0
[ ≤ [
_
t
t0
[f(s, u(s))[ds[,
and, since u ∈ M,
[(Tu)(t) −u
0
[ ≤ αm ≤ b.
Hence
T : M →M.
Computing further, we obtain, for u, v ∈ M that
[(Tu)(t) −(Tv)(t)[ ≤ [
_
t
t0
[f(s, u(s)) −f(s, v(s))[ds[
≤ L[
_
t
t0
[u(s) −v(s)[ds[,
and hence
e
−
˜
Lt−t0
[(Tu)(t) −(Tv)(t)[ ≤ e
−
˜
Lt−t0
L[
_
t
t0
[u(s) −v(s)[ds[
≤
L
˜
L
u −v,
and hence
ρ(Tu, Tv) ≤
L
˜
L
ρ(u, v),
proving that T is a contraction mapping. The result therefore follows from the
contraction mapping principle, Theorem I.6.
We remark that, since T is a contraction mapping, the contraction mapping
theorem gives a constructive means for the solution of the initial value prob
lem in Theorem 2 and the solution may in fact be obtained via an iteration
procedure. This procedure is known as Picard iteration.
In the next section, we shall show, that without the assumption of a local
Lipschitz condition, we still get the existence of solutions.
3 The CauchyPeano Theorem
The following result, called the CauchyPeano theorem provides the local
solvability of initial value problems.
3 Theorem Assume that f : D →R
N
is continuous. Then for every (t
0
, u
0
) ∈ D
the initial value problem (1), (2) has a solution on some interval I, t
0
∈ I.
68
Proof. Let (t
0
, u
0
) ∈ D, and let Q, α, m be as in the proof of Theorem 2. Con
sider the space E = C([t
0
−α, t
0
+α], R
N
) with norm u = max
t−t0≤α
[u(t)[.
Then E is a Banach space. We let M be as deﬁned in the proof of Theorem
2 and note that M is a closed, bounded convex subset of E and further that
T : M →M. We hence may apply the Schauder ﬁxed point theorem (Theorem
III.30) once we verify that T is completely continuous on M. To see this we
note, that, since f is continuous, it follows that T is continuous. On the other
hand, if ¦u
n
¦ ⊂ M, then
[(Tu
n
)(t) −(Tu
n
)(
¯
t)[ ≤ [
_
t
¯
t
[f(s, u
n
(s))[ds[
≤ m[t −
¯
t[.
Hence ¦Tu
n
¦ ⊂ M, is a uniformly bounded and equicontinuous family in E.
It therefore has a uniformly convergent subsequence (as follows from the theo
rem of AscoliArz`ela [36]), showing that ¦Tu
n
¦ is precompact and hence T is
completely continuous. This completes the proof.
We note from the above proofs (of Theorems 2 and 3) that for a solution
u thus obtained, both (t
0
± α, u(t
0
± α)) ∈ D. We hence may reapply these
theorems with initial conditions given at t
0
± α and conditions u(t
0
± α) and
thus continue solutions to larger intervals (in the case of Theorem 2 uniquely
and in the case of Theorem 3 not necessarily so.) We shall prove below that this
continuation process leads to maximal intervals of existence and also describes
the behavior of solutions as one approaches the endpoints of such maximal
intervals.
3.1 Carath´eodory equations
In many situations the nonlinear term f is not continuous as assumed above
but satisﬁes the socalled Carath´eodory conditions on any parallelepiped Q ⊂ D,
where Q is as given in the proof of Theorem 2, i.e.,
• f is measurable in t for each ﬁxed u and continuous in u for almost all t,
• for each Q there exists a function m ∈ L
1
(t
0
−a, t
0
+a) such that
[f(t, u)[ ≤ m(t), (t, u) ∈ Q.
Under such assumptions we have the following extension of the CauchyPeano
theorem, Theorem 3:
4 Theorem Let f satisfy the Carath´eodory conditions on D. Then for every
(t
0
, u
0
) ∈ D the initial value problem (1), (2) has a solution on some interval I,
t
0
∈ I, in the sense that there exists an absolutely continuous function u : I →
R
N
which satisﬁes the initial condition (2) and the diﬀerential equation (1) a.e.
in I.
4. EXTENSION THEOREMS 69
Proof. Let (t
0
, u
0
) ∈ D, and choose
Q = ¦(t, u) : [t −t
0
[ ≤ a, [u −u
0
[ ≤ b¦ ⊂ D.
Let
M = ¦u ∈ C
_
[t
0
−α, t
0
+α], R
N
_
: [u(t) −u
0
[ ≤ b, [t −t
0
[ ≤ α¦,
where α ≤ a is to be determined.
Next deﬁne the operator T on M by:
(Tu)(t) = u
0
+
_
t
t0
f(s, u(s))ds, [t −t
0
[ ≤ α.
Then, because of the Carath´eodory conditions, Tu is a a continuous function
and
[(Tu)(t) −u
0
[ ≤ [
_
t
t0
[f(s, u(s))[ds[.
Further, since u ∈ M,
[(Tu)(t) −u
0
[ ≤
_
t0+α
t0−α
m(s)ds ≤ m
L
1
[t0−α,t0+α]
≤ b,
for α small enough. Hence
T : M →M.
One next shows (see the Exercise 14 below) that T is a completely continuous
mapping, hence will have a ﬁxed point in M by the Schauder Fixed Point
Theorem. That ﬁxed points of T are solutions of the initial value problem (1),
(2), in the sense given in the theorem, is immediate.
4 Extension Theorems
In this section we establish a basic result about maximal intervals of existence
of solutions of initial value problems. We ﬁrst prove the following lemma.
5 Lemma Assume that f : D → R
N
is continuous and let
˜
D be a subdomain
of D, with f bounded on
˜
D. Further let u be a solution of (1) deﬁned on a
bounded interval (a, b) with (t, u(t)) ∈
˜
D, t ∈ (a, b). Then the limits
lim
t→a+
u(t), lim
t→b−
u(t)
exist.
70
Proof. Let t
0
∈ (a, b), then u(t) = u(t
0
) +
_
t
t0
f(s, u(s))ds. Hence for t
1
, t
2
∈
(a, b), we obtain
[u(t
1
) −u(t
2
)[ ≤ m[t
1
−t
2
[,
where m is a bound on f on
˜
D. Hence the above limits exist.
We may therefore, as indicated above, continue the solution beyond the
interval (a, b), to the left of a and the right of b.
We say that a solution u of (1) has maximal interval of existence (ω
−
, ω
+
),
provided u cannot be continued as a solution of (1) to the right of ω
+
nor to
the left of ω
−
.
The following theorem holds.
6 Theorem Assume that f : D → R
N
is continuous and let u be a solution of
(1) deﬁned on some interval I. Then u may be extended as a solution of (1) to
a maximal interval of existence (ω
−
, ω
+
) and (t, u(t)) →∂D as t →ω
±
.
Proof. We establish the existence of a right maximal interval of existence; a
similar argument will yield the existence of a left maximal one and together the
two will imply the existence of a maximal interval of existence.
Let u be a solution of (1) with u(t
0
) = u
0
deﬁned on an interval I = [t
0
, a
u
).
We say that two solutions v, w of (1), (2) satisfy
v _ w, (5)
if and only if:
• u ≡ v ≡ w on [t
0
, a
u
),
•
v is deﬁned on I
v
= [t
0
, a
v
), a
v
≥ a
u
,
w is deﬁned on I
w
= [t
0
, a
w
), a
w
≥ a
u
,
• a
w
≥ a
v
,
• v ≡ w on I
v
.
We see that _ is a partial order on the set of all solutions S of (1), (2) which agree
with u on I. One next veriﬁes that the conditions of the Hausdorﬀ maximum
principle (see [37]) hold and hence that S contains a maximal element, ˜ u. This
maximal element ˜ u cannot be further extended to the right.
Next let u be a solution of (1), (2) with right maximal interval of existence
[t
0
, ω
+
). We must show that (t, u(t)) →∂D as t →ω
+
, i.e., given any compact
set K ⊂ D, there exists t
K
, such that (t, u(t)) / ∈ K, for t > t
K
. If ω
+
= ∞, the
conclusion clearly holds. On the other hand, if ω
+
< ∞, we proceed indirectly.
In which case there exists a compact set K ⊂ D, such that for every n = 1, 2,
there exists t
n
, 0 < ω
+
−t
n
<
1
n
, and (t
n
, u(t
n
)) ∈ K. Since K is compact, there
will be a subsequence, call it again ¦(t
n
, u(t
n
))¦ such that ¦(t
n
, u(t
n
))¦ converges
to, say, (ω
+
, u
∗
) ∈ K. Since (ω
+
, u
∗
) ∈ K, it is an interior point of D. We may
therefore choose a constant a > 0, such that Q = ¦(t, u) : [ω
+
−t[ ≤ a, [u−u
∗
[ ≤
4. EXTENSION THEOREMS 71
a¦ ⊂ D, and thus for n large (t
n
, u(t
n
)) ∈ Q. Let m = max
(t,u)∈Q
[f(t, u)[, and
let n be so large that
0 < ω
+
−t
n
≤
a
2m
, [u(t
n
) − u
∗
[ ≤
a
2
.
Then
[u(t
n
) −u(t)[ < m(ω
+
−t
n
) ≤
a
2
, for t
≤
t < ω
+
,
as an easy indirect argument shows.
It therefore follows that
lim
t→ω+
u(t) = u
∗
,
and we may extend u to the right of ω
+
contradicting the maximality of u.
7 Corollary Assume that f : [t
0
, t
0
+a] R
N
→R
N
is continuous and let u be a
solution of (1) deﬁned on some right maximal interval of existence I ⊂ [t
0
, t
0
+a].
Then, either I = [t
0
, t
0
+a], or else I = [t
0
, ω
+
), ω
+
< t
0
+a, and
lim
t→ω+
[u(t)[ = ∞.
We also consider the following corollary, which is of importance for diﬀeren
tial equations whose right hand side have at most linear growth. I.e., we assume
that the following growth condition holds:
[f(t, u)[ ≤ α(t)[u[ +β(t). (6)
8 Corollary Assume that f : (a, b)R
N
→R
N
is continuous and let f satisfy (6),
where α, β ∈ L
1
(a, b) are nonnegative continuous functions. Then the maximal
interval of existence (ω
−
, ω
+
) is (a, b) for any solution of (1).
Proof. If u is a solution of (1), then u satisﬁes the integral equation (3) and
hence, because of (6), we obtain
[u(t)[ ≤ [u(t
0
)[ +[
_
t
t0
[[α(s)[u(s)[ +β(s)]ds. (7)
Considering the case t ≥ t
0
, the other case being similar, we let
v(t) =
_
t
t0
[[α(s)[[u(s)[ +[β(s)[]ds,
and c = [u(t
0
)[. Then an easy calculation yields
v
′
−α(t)v ≤ α(t)c +β(t),
and hence
v(t) ≤ e
R
t
t
0
α(τ)dτ
_
t
t0
e
R
s
t
0
α(τ)dτ
[α(s)c +β(s)]ds,
from which, using Corollary 7, follows that ω
+
= b.
72
5 Dependence upon Initial Conditions
Let again D be an open connected subset of R R
N
, N ≥ 1, and let
f : D →R
N
be a continuous mapping.
We consider the initial value problem
u
′
= f(t, u), u(t
0
) = u
0
, (8)
and assume we have conditions which guarantee that (8) has a unique solution
u(t) = u(t, t
0
, u
0
), (9)
for every (t
0
, u
0
) ∈ D. We shall now present conditions which guarantee that
u(t, t
0
, u
0
) depends either continuously or smoothly on the initial condition
(t
0
, u
0
).
A somewhat more general situation occurs frequently, where the function f
also depends upon parameters, λ ∈ R
m
, i.e. (8) is replaced by the parameter
dependent problem
u
′
= f(t, u, λ), u(t
0
) = u
0
, (10)
and solutions u then are functions of the type
u(t) = u(t, t
0
, u
0
, λ), (11)
provided (10) is uniquely solvable. This situation is a special case of the above,
as we may augment the original system (10) as
u
′
= f(t, u, λ), u(t
0
) = u
0
,
λ
′
= 0, λ(t
0
) = λ,
(12)
and obtain an initial value problem for a system of equations of higher dimension
which does not depend upon parameters.
5.1 Continuous dependence
We ﬁrst prove the following proposition.
9 Theorem Assume that f : D →R
N
is a continuous mapping and that (8) has
a unique solution u(t) = u(t, t
0
, u
0
), for every (t
0
, u
0
) ∈ D. Then the solution
depends continuously on (t
0
, u
0
), in the following sense: If ¦(t
n
, u
n
)¦ ⊂ D con
verges to (t
0
, u
0
) ∈ D, then given ǫ > 0, there exists n
ǫ
and an interval I
ǫ
such
that for all n ≥ n
ǫ
, the solution u
n
(t) = u(t, t
n
, u
n
), exists on I
ǫ
and
max
t∈Iǫ
[u(t) −u
n
(t)[ ≤ ǫ.
5. DEPENDENCE UPON INITIAL CONDITIONS 73
Proof. We rely on the proof of Theorem 3 and ﬁnd that for given ǫ > 0, there
exists n
ǫ
such that ¦(t
n
, u
n
)¦ ⊂
˜
Q, where
˜
Q = ¦(t, u) : [t −t
0
[ ≤
α
2
, [u −u
0
[ ≤
b
2
¦ ⊂ Q,
and Q is the set given in the proof of Theorems 2 and 3. Using the proof of The
orem 3 we obtain a common compact interval I
ǫ
of existence of the sequence
¦u
n
¦ and ¦(t, u
n
(t))¦ ⊂ Q, for t ∈ I
ǫ
. The sequence ¦u
n
¦ hence will be uni
formly bounded and equicontinuous on I
ǫ
and will therefore have a subsequence
converging uniformly on I
ǫ
. Employing the integral equation (3) we see that
the limit must be a solution of (8) and hence, by the uniqueness assumption
must equal u. Since this is true for every subsequence, the whole sequence must
converge to u, completing the proof.
We have the following corollary, which asserts continuity of solutions with
respect to the diﬀerential equation. The proof is similar to the above and will
hence be omitted.
10 Corollary Assume that f
n
: D → R
N
, n = 1, 2, , are continuous mappings
and that (8) (with f = f
n
) has a unique solution u
n
(t) = u(t, t
n
, u
n
), for
every (t
n
, u
n
) ∈ D. Then the solution depends continuously on (t
0
, u
0
), in the
following sense: If ¦(t
n
, u
n
)¦ ⊂ D converges to (t
0
, u
0
) ∈ D, and f
n
converges
to f, uniformly on compact subsets of D, then given ǫ > 0, there exists n
ǫ
and
an interval I
ǫ
such that for all n ≥ n
ǫ
, the solution u
n
(t) = u(t, t
n
, u
n
), exists
on I
ǫ
and
max
t∈Iǫ
[u(t) −u
n
(t)[ ≤ ǫ.
5.2 Diﬀerentiability with respect to initial conditions
In the following we shall employ the convention u = (u
1
, u
2
, , u
N
). We
have the following theorem.
11 Theorem Assume that f : D → R
N
is a continuous mapping and that the
partial derivatives
∂f
∂u
i
, 1 ≤ i ≤ N are continuous also. Then the solution
u(t) = u(t, t
0
, u
0
), of (8) is of class C
1
in the variable u
0
. Further, if J(t) is the
Jacobian matrix
J(t) = J(t, t
0
, u
0
) =
_
∂f
∂u
_
u=u(t,t0,u0)
,
then
y(t) =
∂u
∂u
i
(t, t
0
, u
0
)
is the solution of the initial value problem
y
′
= J(t)y, y(t
0
) = e
i
, 1 ≤ i ≤ N,
where e
i
∈ R
n
is given by e
k
i
= δ
ik
, with δ
ik
the Kronecker delta.
74
Proof. Let e
i
be given as above and let u(t) = u(t, t
0
, u
0
), u
h
(t) = u(t, t
0
, u
0
+
he
i
), where [h[ is suﬃciently small so that u
h
exists. We note that u and u
h
will have a common interval of existence, whenever [h[ is suﬃciently small. We
compute
(u
h
(t) −u(t))
′
= f(t, u
h
(t)) −f(t, u(t)).
Letting
y
h
(t) =
u
h
(t) −u(t)
h
,
we get y
h
(t
0
) = e
i
. If we let
G(t, y
1
, y
2
) =
_
1
0
∂f
∂u
(t, sy
1
+ (1 −s)y
2
)ds,
we obtain that y
h
is the unique solution of
y
′
= G(t, u(t), u
h
(t))y, y(t
0
) = e
i
.
Since G(t, u(t), u
h
(t)) →J(t) as h →0, we may apply Corollary 10 to conclude
that y
h
→y uniformly on a neighborhood of t
0
.
6 Diﬀerential Inequalities
We consider in R
N
the following partial orders:
x ≤ y ⇔ x
i
≤ y
i
, 1 ≤ i ≤ N,
x < y ⇔ x
i
< y
i
, 1 ≤ i ≤ N.
For a function u : I →R
N
, where I is an open interval, we consider the Dini
derivatives
D
+
u(t) = limsup
h→0+
u(t+h)−u(t)
h
,
D
+
u(t) = liminf
h→0+
u(t+h)−u(t)
h
,
D
−
u(t) = limsup
h→0−
u(t+h)−u(t)
h
,
D
−
u(t) = liminf
h→0−
u(t+h)−u(t)
h
,
where limsup and liminf are taken componentwise.
12 Deﬁnition A function f : R
N
→R
N
is said to be of type K (after Kamke [23])
on a set S ⊂ R
N
, whenever
f
i
(x) ≤ f
i
(y), ∀x, y ∈ S, x ≤ y, x
i
= y
i
.
The following theorem on diﬀerential inequalities is of use in obtaining esti
mates on solutions.
6. DIFFERENTIAL INEQUALITIES 75
13 Theorem Assume that f : [a, b] R
N
→ R
N
is a continuous mapping which
is of type K for each ﬁxed t. Let u : [a, b] → R
N
be a solution of (1) and let
v : [a, b] →R
N
be continuous and satisfy
D
−
v(t) > f(t, v(t)), a < t ≤ b,
v(a) > u(a),
(13)
then v(t) > u(t), a ≤ t ≤ b.
If z : [a, b] →R
N
is continuous and satisﬁes
D
−
z(t) < f(t, z(t)), a < t ≤ b,
z(a) < u(a),
(14)
then z(t) < u(t), a ≤ t ≤ b.
Proof. We prove the ﬁrst part of the theorem. The second part follows along
the same line of reasoning. By continuity of u and v, there exists δ > 0, such
that
v(t) > u(t), a ≤ t ≤ a +δ.
If the inequality does not hold throughout [a, b], there will exist a ﬁrst point c
and an index i such that
v(t) > u(t), a ≤ t < c, v(c) ≥ u(c), v
i
(c) = u
i
(c).
Hence (since f is of type K)
D
−
v
i
(c) > f
i
(c, v(c)) ≥ f
i
(c, u(c)) = u
i
′
(c).
On the other hand,
D
−
v
i
(c) = limsup
h→0−
v
i
(c+h)−v
i
(c)
h
≤ limsup
h→0−
u
i
(c+h)−u
i
(c)
h
= u
i
′
(c),
a contradiction.
14 Deﬁnition A solution u
∗
of (1) is called a right maximal solution on an interval
I, if for every t
0
∈ I and any solution u of (1) such that
u(t
0
) ≤ u
∗
(t
0
),
it follows that
u(t) ≤ u
∗
(t), t
0
≤ t ∈ I.
Right minimal solutions are deﬁned similarly.
15 Theorem Assume that f : D →R
N
is a continuous mapping which is of type
K for each t. Then the initial value problem (8) has a unique right maximal
(minimal) solution for each (t
0
, u
0
) ∈ D.
76
Proof. That maximal and minimal solutions are unique follows from the deﬁ
nition. Choose 0 < ǫ ∈ R
N
and let v
n
be any solution of
v
′
= f(t, v, λ) +
ǫ
n
, v(t
0
) = u
0
+
ǫ
n
. (15)
Then, given a neighborhood U of (t
0
, u
0
) ∈ D, there exists an interval [t
0
, t
1
] of
positive length such that all v
n
are deﬁned on this interval with ¦(t, v
n
(t))¦ ⊂
U, t
0
≤ t ≤ t
1
, for all n suﬃciently large. On the other hand it follows from
Theorem 13 that
v
n
(t) > v
m
(t), t
0
≤ t ≤ t
1
, n < m.
The sequence ¦v
n
¦ is therefore unformly bounded and equicontinuous on [t
0
, t
1
],
hence will have a subsequence which converges uniformly to a solution u
∗
of (8).
Since the sequence is monotone, the whole sequence will, in fact, converge to u
∗
.
Applying Theorem 13 once more, we obtain that u
∗
is right maximal on [t
0
, t
1
],
and extending this solution to a right maximal interval of existence as a right
maximal solution, completes the proof.
We next prove an existence theorem for initial value problems which allows
for estimates of the solution in terms of given solutions of related diﬀerential
inequalities.
16 Theorem Assume that f : [a, b] R
N
→R
N
is a continuous mapping which is
of type K for each ﬁxed t. Let v, z : [a, b] →R
N
be continuous and satisfy
D
+
v(t) ≥ f(t, v(t)), a ≤ t < b,
D
+
z(t) ≤ f(t, z(t)), a ≤ t < b,
z(t) ≤ v(t), a ≤ t ≤ b.
(16)
Then for every u
0
, z(a) ≤ u
0
≤ v(a), there exists a solution u of (8) (with
t
0
= a) such that
z(t) ≤ u(t) ≤ v(t), a ≤ t ≤ b.
The functions z and v are called, respectively, sub and supersolutions of (8).
Proof. Deﬁne
¯
f(t, x) = f(t, ¯ x) −(x − ¯ x), where for 1 ≤ i ≤ N,
¯ x
i
=
_
_
_
v
i
(t) , if x
i
> v
i
(t),
x
i
, if z
i
(t) ≤ x
i
≤ v
i
(t),
z
i
(t) , if x
i
< z
i
(t).
Then
¯
f is continuous and has at most linear growth, hence the initial value
problem (8), with f replaced by
¯
f has a solution u that extends to [a, b] (see
Corollary 8). We show that z(t) ≤ u(t) ≤ v(t), a ≤ t ≤ b, and hence may
conclude that u solves the original initial value problem (8). To see this, we
shall prove that for any ǫ > 0 we have that
z
i
(t) −ǫ ≤ u
i
(t) ≤ v
i
(t) +ǫ, t ∈ [a, b], i = 1, , N.
6. DIFFERENTIAL INEQUALITIES 77
This we argue indirectly and suppose there exists ǫ > 0 a value c ∈ (a, b) and i
such that
u
i
(c) = v
i
(c) +ǫ, u
i
(t) ≥ v
i
(t) +ǫ, c < t ≤ t
1
≤ b.
Since ¯ u
i
(t) = v
i
(t), c < t ≤ t
1
and ¯ u
k
(t) ≤ v
k
(t), c < t ≤ t
1
, k ,= i, we get
that
¯
f
i
(t, u(t)) = f(t, ¯ u(t)) −(u
i
(t) −v
i
(t)) ≤ f
i
(t, v(t)) −ǫ < D
+
v
i
(t), c < t ≤ t
1
.
On the other hand we have, for h > 0, small, that
u
i
(t +h) −u
i
(c) ≥ v
i
(t +h) + ǫ −(v
i
(c) +ǫ,
and hence
u
i
′
(c) ≥ D
+
v
i
(c),
contradicting what has been concluded above. Since ǫ > 0 was arbitrary, we
conclude that
u(t) ≤ v(t).
That
u(t) ≥ vz(t)
may be proved in a similar way.
17 Corollary Assume the hypotheses of Theorem 16 and that f satisﬁes a local
Lipschitz condition. Assume furthermore that
z(a) ≤ z(b), v(a) ≥ v(b).
Then the problem
u
′
= f(t, u), u(a) = u(b) (17)
has a solution u with
z(t) ≤ u(t) ≤ v(t), a ≤ t ≤ b.
Proof. Since f satisﬁes a local Lipschitz condition, initial value problems are
uniquely solvable. Hence for every u
0
, z(a) ≤ u
0
≤ v(a), there exists a unique
solution u(t, u
0
) of (8) (with t
0
= a) such that
z(t) ≤ u(t) ≤ v(t), a ≤ t ≤ b,
as follows from Theorem 16. Deﬁne the mapping
T : ¦x : z(a) ≤ x ≤ v(a)¦ →¦x : z(b) ≤ x ≤ v(b)¦
by
Tx = u(b, x),
78
then, since by hypothesis ¦x : z(a) ≤ x ≤ v(a)¦ ⊃ ¦x : z(b) ≤ x ≤ v(b)¦
and since ¦x : z(a) ≤ x ≤ v(a)¦ is convex, it follows by Brouwer’s ﬁxed point
theorem (Theorem III.22) and the fact that T is continuous (Theorem 9) that
T has a ﬁxed point, completing the proof.
An important consequence of this corollary is that if in addition f is a
function which is periodic in t with period b − a, then Corollary 17 asserts the
existence of a periodic solution (of period b −a).
The following results use comparison and diﬀerential inequality arguments
to provide a priori bounds and extendability results.
18 Theorem Assume that F : [a, b] R
+
→R
+
is a continuous mapping and that
f : [a, b] R
N
→R
N
is continuous also and
[f(t, x)[ ≤ F(t, [x[), a ≤ t ≤ b, x ∈ R
N
.
Let u : [a, b] →R
N
be a solution of (1) and let v : [a, b] →R
+
be the continuous
and right maximal solution of
v
′
(t) = F(t, v(t)), a ≤ t ≤ b,
v(a) ≥ [u(a)[,
(18)
then v(t) ≥ [u(t)[, a ≤ t ≤ b.
Proof. Let z(t) = [u(t)[, then z is continuous and D
−
z(t) = D
−
z(t). Further
D
−
z(t) = liminf
h→0+
u(t)−u(t−h)
h
≤ lim
h→0+
u(t−h)−u(t)
h
= [f(t, u(t))[ ≤ F(t, z(t)).
Hence, by Theorem 15 (actually its corollary (Exercise 6)), we conclude that
v(t) ≥ [u(t)[, a ≤ t ≤ b.
7 Uniqueness Theorems
In this section we provide supplementary conditions which guarantee the
uniqueness of solutions of ivp’s.
19 Theorem Assume that F : (a, b) R
+
→ R
+
is a continuous mapping and
that f : (a, b) R
N
→R
N
is a continuous also and
[f(t, x) −f(t, y)[ ≤ F(t, [x −y[), a ≤ t ≤ b, x, y ∈ R
N
.
Let F(t, 0) ≡ 0 and let, for any c ∈ (a, b), w ≡ 0 be the only solution of
w
′
= F(t, w) on (a, c) such that w(t) = 0(µ(t)), t → a where µ is a given
positive and continuous function. Then (1) cannot have distinct solutions such
that [u(t) −v(t)[ = 0(µ(t)), t →a.
8. EXERCISES 79
Proof. Let u, v be distinct solutions of (1) such that [u(t)−v(t)[ = 0(µ(t)), t →
a. Let z(t) = [u(t) −v(t)[. Then z is continuous and
D
+
z(t) ≤ [f(t, u(t)) −f(t, v(t))[ ≤ F(t, z(t)).
The proof is completed by employing arguments like those used in the proof of
Theorem 18.
20 Remark Theorem 19 does not require that f be deﬁned for t = a. The advan
tage of this may be that a = −∞ or that f may be singular there. A similar
result, of course, holds for t →b −.
8 Exercises
1. Prove Proposition 1.
2. Prove Corollary 10.
3. Verify that the space (M, ρ) in the proof of Theorem 2 is a complete metric
space.
4. Complete the details in the proof of Theorem 11.
5. State and prove a theorem similar to Theorem 11, providing a diﬀerential
equation for
∂u
∂λ
whenever f = f(t, u, λ) also depends upon a parameter λ.
6. Prove the following result: Assume that f : [a, b] R
N
→R
N
is a contin
uous mapping which is of type K for each ﬁxed t. Let u : [a, b] → R
N
be
a right maximal solution of (1) and let z : [a, b] →R
N
be continuous and
satisfy
D
−
z(t) ≤ f(t, z(t)), a < t ≤ b,
z(a) ≤ u(a),
(19)
then z(t) ≤ u(t), a ≤ t ≤ b.
7. Show that a real valued continuous function z(t) is nonincreasing on an
interval [a, b] if and only if D
−
z ≤ 0 on (a, b].
8. Assume that f : [a, ∞) R
N
→R
N
is a continuous mapping such that
[f(t, x)[ ≤ M(t)L([x[), a ≤ t < ∞, x ∈ R
N
,
where M and L > 0 are continuous functions on their respective domains
and
_
∞
ds
L(s)
= ∞.
Prove that ω
+
= ∞ for all solutions of (1).
80
9. Give the details of the proof of Theorem 18.
10. Assume that f : [a, b) R
N
→ R
N
is a continuous mapping and assume
the conditions of Theorem 19 with µ ≡ 1. Then the initial value problem
u
′
= f(t, u), u(a) = u
0
(20)
has at most one solution.
11. Assume that f : [a, b) R
N
→R
N
is a continuous mapping and that
[f(t, x) −f(t, y)[ ≤ c
[x −y[
t −a
, t > a, x, y ∈ R
N
, 0 < c < 1 (21)
Then the initial value problem (20) has at most one solution.
12. The previous exercise remains valid if (21) is replaced by
(f(t, x) −f(t, y)) (x −y) ≤ c
[x −y[
2
t −a
, t > a, x, y ∈ R
N
, 0 ≤ c <
1
2
.
13. Assume that f : [a, b) R
N
→R
N
is a continuous mapping and that
(f(t, x) −f(t, y)) (x −y) ≤ 0, t ≥ a, x, y ∈ R
N
.
Then every initial value problem is uniquely solvable to the right. This
exercise, of course follows from the previous one. Give a more elementary
and direct proof. Note that unique solvability to the left of an initial
point is not guaranteed. How must the above condition be modiﬁed to
guarantee uniqueness to the left of an initial point?
14. Provide the details in the proof of Theorem 4.
15. Establish a result similar to Theorem6 assuming that f satisﬁes Carath´eodory
conditions.
Chapter VI
Linear Ordinary Diﬀerential
Equations
1 Introduction
In this chapter we shall employ what has been developed to give a brief
overview of the theory of linear ordinary diﬀerential equations. The results ob
tained will be useful in the study of stability of solutions of nonlinear diﬀerential
equations as well as bifurcation theory for periodic orbits and many other facets
where linearization techniques are of importance. The results are also of interest
in their own right.
2 Preliminaries
Let I ⊂ R be a real interval and let
A : I →L(R
N
, R
N
)
f : I →R
N
be continuous functions. We consider here the system of ordinary diﬀerential
equations
u
′
= A(t)u +f(t), t ∈ I, (1)
and
u
′
= A(t)u, t ∈ I. (2)
Using earlier results we may establish the following basic proposition (see Ex
ercise 1).
1 Proposition For any given f, initial value problems for (1) are uniquely solv
able and solutions are deﬁned on all of I.
81
82
2 Remark More generally we may assume that A and f are measurable on I
and locally integrable there, in which case the conclusion of Proposition 1 still
holds. We shall not go into details for this more general situation, but leave it
to the reader to present a parallel development.
3 Proposition The set of solutions of (2) is a vector space of dimension N.
Proof. That the solution set forms a vector space is left as an exercise (Exercise
2, below). To show that the dimension of this space is N, we employ the
uniqueness principle above. Thus let t
0
∈ I, and let u
k
(t), k = 1, , N be the
solution of (2) such that
u
k
(t
0
) = e
k
, e
i
k
= δ
ki
(Kronecker delta). (3)
It follows that for any set of constants a
1
, , a
N
,
u(t) =
N
1
a
i
u
i
(t) (4)
is a solution of (2). Further, for given ξ ∈ R
N
, the solution u of (2) such that
u(t
0
) = ξ is given by (4) with a
i
= ξ
i
, i = 1, , N.
Let the N N matrix function Φ be deﬁned by
Φ(t) = (u
i
j
(t)), 1 ≤ i, j ≤ N, (5)
i.e., the columns of Φ are solutions of (2). Then (4) takes the form
u(t) = Φ(t)a, a = (a
1
, , a
N
)
T
. (6)
Hence for given ξ ∈ R
N
, the solution u of (2) such that u(t
1
) = ξ, t
1
∈ I,
is given by (6) provided that Φ(t
1
) is a nonsingular matrix, in which case a
may be uniquely determined. That this matrix is never singular, provided it is
nonsingular at some point, is known as the AbelLiouville lemma, whose proof
is left as an exercise below.
4 Lemma If g(t) = detΦ(t), then g satisﬁes
g(t) = g(t
0
)e
R
t
t
0
traceA(s)
ds. (7)
Hence, if Φ is deﬁned by (5), where u
1
, , u
N
are solutions of (2), then Φ(t) is
nonsingular for all t ∈ I if and only if Φ(t
0
) is nonsingular for some t
0
∈ I.
2.1 Fundamental solutions
A nonsingular N N matrix function Ψ whose columns are solutions of (2)
is called a fundamental matrix solution or a fundamental system of (2). Such a
matrix is a nonsingular solution of the matrix diﬀerential equation
Ψ
′
= A(t)Ψ. (8)
The following proposition characterizes the set of fundamental solutions; its
proof is again left as an exercise.
3. CONSTANT COEFFICIENT SYSTEMS 83
5 Proposition Let Φ be a given fundamental matrix solution of (2). Then every
other fundamental matrix solution has the form Ψ = ΦC, where C is a constant
nonsingular N N matrix. Furthermore the set of all solutions of (2) is given
by
¦Φc : c ∈ R
N
¦,
where Φ is a fundamental system.
2.2 Variation of constants
It follows from Propositions 3 and 5 that all solutions of (1) are given by
¦Φ(t)c +u
p
(t) : c ∈ R
N
¦,
where Φ is a fundamental system of (2) and u
p
is some particular solution of (1).
Hence the problem of ﬁnding all solutions of (1) is solved once a fundamental
system of (2) is known and some particular solution of (1) has been found. The
following formula, known as the variation of constants formula, shows that a
particular solution of (1) may be obtained from a fundamental system.
6 Proposition Let Φ be a fundamental matrix solution of (2) and let t
0
∈ I.
Then
u
p
(t) = Φ(t)
_
t
t0
Φ
−1
(s)f(s)ds (9)
is a solution of (1). Hence the set of all solutions of (1) is given by
¦Φ(t)
_
c +
_
t
t0
Φ
−1
(s)f(s)ds
_
: c ∈ R
N
¦,
where Φ is a fundamental system of (2).
3 Constant Coeﬃcient Systems
In this section we shall assume that the matrix A is a constant matrix and
thus have that solutions of (2) are deﬁned for all t ∈ R. In this case a fundamental
matrix solution Φ is given by
Φ(t) = e
tA
C, (10)
where C is a nonsingular constant N N matrix and
e
tA
=
∞
0
t
n
A
n
n!
.
Thus the solution u of (2) with u(t
0
) = ξ is given by
u(t) = e
(t−t0)A
ξ.
84
To compute e
tA
we use the (complex) Jordan canonical form J of A. Since A
and J are similar, there exists a nonsingular matrix P such that A = PJP
−1
and hence e
tA
= Pe
tJ
P
−1
. We therefore compute e
tJ
. On the other hand J has
the form
J =
_
_
_
_
_
J
0
J
1
.
.
.
J
s
_
_
_
_
_
,
where
J
0
=
_
_
_
_
_
λ
1
λ
2
.
.
.
λ
q
_
_
_
_
_
,
is a q q diagonal matrix whose entries are the simple (algebraically) and
semisimple eigenvalues of A, repeated according to their multiplicities, and for
1 ≤ i ≤ s,
J
i
=
_
_
_
_
_
_
_
λ
q+i
1
λ
q+i
1
.
.
.
.
.
.
1
λ
q+i
_
_
_
_
_
_
_
is a q
i
q
i
matrix, with
q +
s
1
q
i
= N.
By the laws of matrix multiplication it follows that
e
tJ
=
_
_
_
_
_
e
tJ0
e
tJ1
.
.
.
e
tJs
_
_
_
_
_
,
and
e
tJ0
=
_
_
_
_
_
e
λ1t
e
λ2t
.
.
.
e
λqt
_
_
_
_
_
.
4. FLOQUET THEORY 85
Further, since J
i
= λ
q+i
I
ri
+Z
i
, where I
ri
is the r
i
r
i
identity matrix and Z
i
is given by
Z
i
=
_
_
_
_
_
_
_
0 1
0 1
.
.
.
.
.
.
1
0
_
_
_
_
_
_
_
,
we obtain that
e
tJi
= e
tλq+iIr
i
e
tZi
= e
tλq+i
e
tZi
.
An easy computation now shows that
e
tZi
=
_
_
_
_
_
_
1 t
t
2
2!
t
r
i−1
ri−1!
0 1 t
t
r
i−2
ri−2!
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 1
_
_
_
_
_
_
.
Since P is a nonsingular matrix e
tA
P = Pe
tJ
is a fundamental matrix solution
as well. Also, since J and P may be complex we obtain the set of all real
solutions as
¦RePe
tJ
c, ImPe
tJ
c : c ∈ C
N
¦.
The above considerations have the following proposition as a consequence.
7 Proposition Let A be an N N constant matrix and consider the diﬀerential
equation
u
′
= Au. (11)
Then:
1. All solutions u of (11) satisfy u(t) →0, as t →∞, if and only if Reλ < 0,
for all eigenvalues λ of A.
2. All solutions u of (11) are bounded on [0, ∞), if and only if Reλ ≤ 0, for
all eigenvalues λ of A and those with zero real part are semisimple.
4 Floquet Theory
Let A(t), t ∈ R be an N N continuous matrix which is periodic with
respect to t of period T, i.e., A(t +T) = A(t), −∞< t < ∞, and consider the
diﬀerential equation
u
′
= A(t)u. (12)
We shall associate to (12) a constant coeﬃcient system which determines the
asymptotic behavior of solutions of (12). To this end we ﬁrst establish some
facts about fundamental solutions of (12).
86
8 Proposition Let Φ(t) be a fundamental matrix solution of (12), then so is
Ψ(t) = Φ(t +T).
Proof. Since Φ is a fundamental matrix it is nonsingular for all t, hence Ψ is
nonsingular. Further
Ψ
′
(t) = Φ
′
(t +T) = A(t +T)Φ(t +T)
= A(t)Φ(t +T)
= A(t)Ψ(t).
It follows by our earlier considerations that there exists a nonsingular con
stant matrix Q such that
Φ(t +T) = Φ(t)Q.
Since Q is nonsingular, there exists a matrix R (see exercises at the end of this
chapter) such that
Q = e
TR
.
Letting C(t) = Φ(t)e
−tR
we compute
C(t +T) = Φ(t +T)e
−(t+T)R
= Φ(t)Qe
−TR
e
−tR
= Φ(t)e
−tR
= C(t).
We have proved the following proposition.
9 Proposition Let Φ(t) be a fundamental matrix solution of (12), then there
exists a nonsingular periodic (of period T) matrix C and a constant matrix R
such that
Φ(t) = C(t)e
tR
. (13)
From this representation we may immediately deduce conditions which guaran
tee the existence of nontrivial T− periodic and mT− periodic (subharmonics)
solutions of (12).
10 Corollary For any positive integer m (12) has a nontrivial mT−periodic so
lution if and only if Φ
−1
(0)Φ(T) has an m−th root of unity as an eigenvalue,
where Φ is a fundamental matrix solution of (12).
Proof. The properties of fundamental matrix solutions guarantee that the
matrix Φ
−1
(0)Φ(T) is uniquely determined by the equation and Proposition 9
implies that
Φ
−1
(0)Φ(T) = e
TR
.
On the other hand a solution u of (12) is given by
u(t) = C(t)e
tR
d,
4. FLOQUET THEORY 87
where u(0) = C(0)d. Hence u is periodic of period mT if and only if
u(mT) = C(mT)e
mTR
d = C(0)e
mTR
d = C(0)d.
Which is the case if and only if e
mTR
=
_
e
TR
_
m
has 1 as an eigenvalue.
Let us apply these results to the second order scalar equation
y
′′
+p(t)y = 0, (14)
(Hill’s equation) where p : R →R is a Tperiodic function. Equation (14) may
be rewritten as the system
u
′
=
_
0 1
−p(t) 0
_
u. (15)
Let y
1
be the solution of (14) such that
y
1
(0) = 1, y
′
1
(0) = 0,
and y
2
the solution of (14) such that
y
2
(0) = 0, y
′
2
(0) = 1.
Then
Φ(t) =
_
y
1
(t) y
2
(t)
y
′
1
(t) y
′
2
(t)
_
will be a fundamental solution of (15)and
detΦ(t) = e
R
t
0
traceA(s)ds
= 1,
by the AbelLiouville formula (7). Hence (14), or equivalently (15), will have a
mT−periodic solution if and only if Φ(T) has an eigenvalue λ which is an m−th
root of unity. The eigenvalues of Φ(T) are solutions of the equation
det
_
y
1
(T) −λ y
2
(T)
y
′
1
(T) y
′
2
(T) −λ
_
= 0,
or
λ
2
−aλ + 1 = 0,
where
a = y
1
(T) +y
′
2
(T).
Therefore
λ =
a ±
√
a
2
−4
2
.
88
5 Exercises
1. Prove Proposition 1.
2. Prove that the solution set of (2) forms a vector space over either the real
ﬁeld or the ﬁeld of complex numbers.
3. Verify the AbelLiouville formula (7).
4. Prove Proposition 5. Also give an example to show that for a nonsingular
N N constant matrix C, and a fundamental solution Φ, Ψ = CΦ need
not be a fundamental solution.
5. Let [0, ∞) ⊂ I and assume that all solutions of (2) are bounded on [0, ∞).
Let Φ be a fundamental matrix solution of (2). Show that Φ
−1
(t) is
bounded on [0, ∞) if and only if
_
t
0
A(s)ds is bounded from below. If this
is the case, prove that no solution u of (2) may satisfy u(t) →0 as t →∞
unless u ≡ 0.
6. Let [0, ∞) ⊂ I and assume that all solutions of (2) are bounded on [0, ∞).
Further assume that the matrix Φ
−1
(t) is bounded on [0, ∞). Let B :
[0, ∞) →R
N×N
be continuous and such that
_
∞
0
[A(s) −B(s)[ds < ∞.
Then all solutions of
u
′
= B(t)u (16)
are bounded on [0, ∞).
7. Assume the conditions of the previous exercise. Show that corresponding
to every solution u of (2) there exists a unique solution v of (16) such that
lim
t→∞
[u(t) −v(t)[ = 0.
8. Assume that
_
∞
0
[B(s)[ds < ∞.
Show that any solution, not the trivial solution, of (16) tends to a nonzero
limit as t →∞ and for any c ∈ R
N
, there exists a solution v of (16) such
that
lim
t→∞
v(t) = c.
5. EXERCISES 89
9. Prove Proposition 7 using what has been developed about the matrix
exponential e
tA
.
10. Give an alternate veriﬁcation of Proposition 7 using the following facts
from Linear Algebra for a given matrix A.
(a) Let λ be an eigenvalue of A of algebraic multiplicity m, then there
exists a minimal integer q such that
ker(A −λI)
q
= ker(A −λI)
q+1
= ker(A −λI)
q+2
=
and the dimension
dim ker(A −λI)
q
= m.
The space ker(A−λI)
q
is called the generalized eigenspace associated
with λ.
(b) If λ and µ are diﬀerent eigenvalues of A their generalized eigenspaces
only have the zero vector in common.
Using the above and the fact that e
tA
may be written as
e
tA
= e
tλ
e
t(A−λI)
,
deduce the proposition.
11. Let Q be a nonsingular square matrix. Use the Jordan normal form and
block multiplication rules for matrices to deduce that there exists a square
matrix X such that
Q = e
X
.
12. Give necessary and suﬃcient conditions in order that all solutions u of
(12) satisfy
lim
t→∞
u(t) = 0.
13. Show that there exists a nonsingular C
1
matrix L(t) such that the substi
tution u = L(t)v reduces (12) to a constant coeﬃcient system v
′
= Bv.
14. Provide conditions on a = y
1
(T) +y
′
2
(T) in order that (14) have a
T, 2T, , mT −periodic
solution, where the period should be the minimal period.
15. Consider equation (1), where both A and f are T−periodic. Use the vari
ation of constants formula to discuss the existence of T−periodic solutions
of (1).
90
Chapter VII
Periodic Solutions
1 Introduction
In this chapter we shall develop, using the linear theory developed in the
previous chapter together with the implicit function theorem and degree theory,
some of the basic existence results about periodic solutions of periodic nonlinear
systems of ordinary diﬀerential equations. In particular, we shall mainly be
concerned with systems of the form
u
′
= A(t)u +f(t, u), (1)
where
A : R →R
N
R
N
f : R R
N
→R
N
are continuous and T−periodic with respect to t. We call the equation nonres
onant provided the linear system
u
′
= A(t)u (2)
has as its only T− periodic solution the trivial one; we call it resonant, other
wise.
2 Preliminaries
We recall from Chapter 6 that the set of all solutions of the equation
u
′
= A(t)u +g(t), (3)
is given by
u(t) = Φ(t)c + Φ(t)
_
t
t0
Φ
−1
(s)g(s)ds, c ∈ R
N
, (4)
where Φ is a fundamental matrix solution of the linear system (2). On the
other hand it follows from Floquet theory (Section VI.4) that Φ has the form
91
92
Φ(t) = C(t)e
tR
, where C is a continuous nonsingular periodic matrix of period
T and R is a constant matrix (viz. Proposition VI.9). As we may choose Φ such
that Φ(0) = I (the N N identity matrix), it follows that (with this choice)
C(0) = C(T) = I
and
u(T) = e
TR
_
c +
_
T
0
Φ
−1
(s)g(s)ds, c ∈ R
N
_
,
hence u(0) = u(T) = c if and only if
c = e
TR
_
c +
_
T
0
Φ
−1
(s)g(s)ds, c ∈ R
N
_
. (5)
We note that equation (5) is uniquely solvable for every g, if and only if
I −e
TR
is a nonsingular matrix. That is, we have the following result:
1 Proposition Equation (3) has a unique T−periodic solution for every T−periodic
forcing term g if and only if e
TR
−I is a nonsingular matrix. If this is the case,
the periodic solution u is given by the following formula:
u(t) = Φ(t)
_
(I −e
TR
)
−1
e
TR
_
T
0
Φ
−1
(s)g(s)ds +
_
t
0
Φ
−1
(s)g(s)ds
_
.(6)
This proposition allows us to formulate a ﬁxed point equation whose solution
will determine T− periodic solutions of equation (1). The following section is
devoted to results of this type.
3 Perturbations of Nonresonant Equations
In the following let
E = ¦u ∈ C([0, T], R
N
) : u(0) = u(T)¦
with u = max
t∈[0,T]
[u(t)[ and let S : E →E be given by
(Su)(t) = Φ(t)
_
(I −e
TR
)
−1
e
TR
_
T
0
Φ
−1
(s)f(s, u(s))ds
+
_
t
0
Φ
−1
(s)f(s, u(s))ds
_
.
(7)
2 Proposition Assume that I −e
TR
is nonsingular, then (1) has a T− periodic
solution u, whenever the operator S given by equation (7) has a ﬁxed point in
the space E.
3. PERTURBATIONS OF NONRESONANT EQUATIONS 93
For f as given above let us deﬁne
P(r) = max¦[f(t, u)[ : 0 ≤ t ≤ T, [u[ ≤ r¦. (8)
We have the following theorem.
3 Theorem Assume that A and f are as above and that I −e
TR
is nonsingular,
then (1) has a T− periodic solution u, whenever
liminf
r→∞
P(r)
r
= 0, (9)
where P is deﬁned by (8).
Proof. Let us deﬁne
B(r) = ¦u ∈ E : u ≤ r¦,
then for u ∈ B(r) we obtain
Su ≤ KP(r),
where S is the operator deﬁned by equation (7) and K is a constant that depends
only on the matrix A. Hence
S : B(r) →B(r),
provided that
KP(r) ≤ r,
which holds, by condition (9), for some r suﬃciently large. Since S is completely
continuous the result follows from the Schauder ﬁxed point theorem (Theorem
III.30).
As a corollary we immediately obtain:
4 Corollary Assume that A and f are as above and that I −e
TR
is nonsingular,
then
u
′
= A(t)u +ǫf(t, u) (10)
has a T− periodic solution u, provided that ǫ is suﬃciently small.
Proof. Using the operator S associated with equation (10) we obtain for
u ∈ B(r)
Su ≤ [ǫ[KP(r),
thus for given r > 0, there exists ǫ ,= 0 such that
[ǫ[KP(r) ≤ r,
and S will have a ﬁxed point in B(r).
The above corollary may be considerably extended using the global contin
uation theorem Theorem IV.4. Namely we have the following result.
94
5 Theorem Assume that A and f are as above and that I −e
TR
is nonsingular.
Let
S
+
= ¦(u, ǫ) ∈ E [0, ∞) : (u, ǫ) solves (10)¦
and
S
−
= ¦(u, ǫ) ∈ E (−∞, 0] : (u, ǫ) solves (10)¦.
Then there exists a continuum C
+
⊂ S
+
(C
−
⊂ S
−
) such that:
1. C
+
0
∩ E = ¦0¦ (C
−
0
∩ E = ¦0¦),
2. C
+
is unbounded in E [0, ∞) (C
−
is unbounded in E (−∞, 0]).
Proof. The proof follows immediately from Theorem IV.4 by noting that the
existence of T−periodic solutions of equation (10) is equivalent to the existence
of solutions of the operator equation
u −S(ǫ, u) = 0,
where
S(ǫ, u)(t) = Φ(t)
_
(I −e
TR
)
−1
e
TR
_
T
0
Φ
−1
(s)ǫf(s, u(s))ds
+
_
t
0
Φ
−1
(s)ǫf(s, u(s))ds
_
.
4 Resonant Equations
4.1 Preliminaries
We shall now consider the equation subject to constraint
u
′
= f(t, u),
u(0) = u(T),
(11)
where
f : R R
N
→R
N
is continuous. Should f be T−periodic with respect to t, then a T−periodic
extension of a solution of (11) will be a T− periodic solution of the equation.
We view (11) as a perturbation of the equation u
′
= 0, i.e. we are in the case
of resonance.
We now let
E = ¦u ∈ C([0, T], R
N
)¦
with u = max
t∈[0,T]
[u(t)[ and let S : E →E be given by
(Su)(t) = u(T) +
_
t
0
f(s, u(s))ds, (12)
4. RESONANT EQUATIONS 95
then clearly S : E → E is a completely continuous mapping because of the
continuity assumption on f.
The following lemma holds.
6 Lemma An element u ∈ E is a solution of (11) if and only if it is a ﬁxed point
of the operator S given by (12).
This lemma, whose proof is immediate, gives us an operator equation in the
space E whose solutions are solutions of the problem (11).
4.2 Homotopy methods
We shall next impose conditions on the ﬁnite dimensional vector ﬁeld
x ∈ R
N
→ g(x)
g(x) = −
_
T
0
f(s, x)ds,
(13)
which will guarantee the existence of solutions of an associated problem
u
′
= ǫf(t, u),
u(0) = u(T),
(14)
where ǫ is a small parameter. We have the following theorem.
7 Theorem Assume that f is continuous and there exists a bounded open set
Ω ⊂ R
N
such that the mapping g deﬁned by (13) does not vanish on ∂Ω. Further
assume that
d(g, Ω, 0) ,= 0, (15)
where d(g, Ω, 0) is the Brouwer degree. Then problem (14) has a solution for all
suﬃciently small ǫ.
Proof. We deﬁne the bounded open set G ⊂ E by
G = ¦u ∈ E : u : [0, T] →Ω¦. (16)
For u ∈
¯
G deﬁne
u(t, λ) = λu(t) + (1 −λ)u(T), 0 ≤ λ ≤ 1, (17)
and let
a(t, λ) = λt + (1 −λ)T, 0 ≤ λ ≤ 1. (18)
For 0 ≤ λ ≤ 1, 0 ≤ ǫ ≤ 1, deﬁne S : E [0, 1] [0, 1] →E by
S(u, λ, ǫ)(t) = u(T) +ǫ
_
a(t,λ)
0
f(s, u(s, λ))ds. (19)
96
Then S is a completely continuous mapping and the theorem will be proved
once we show that
d(id −S(, 1, ǫ), G, 0) ,= 0, (20)
for ǫ suﬃciently small, for if this is the case, S(, 1, ǫ) has a ﬁxed point in G
which is equivalent to the assertion of the theorem.
To show that (20) holds we ﬁrst show that S(, λ, ǫ) has no zeros on ∂G for
all λ ∈ [0, 1] and ǫ suﬃciently small. This we argue indirectly and hence obtain
sequences ¦u
n
¦ ⊂ ∂G, ¦λ
n
¦ ⊂ [0, 1], and ¦ǫ
n
¦, ǫ
n
→0, such that
u
n
(T) +ǫ
n
_
a(t,λn)
0
f(s, u
n
(s, λ
n
))ds ≡ u
n
(t), 0 ≤ t ≤ T,
and hence
_
T
0
f(s, u
n
(s, λ
n
))ds = 0, n = 1, 2, .
Without loss in generality, we may assume that the sequences mentioned con
verge to, say, u and λ
0
and the following must hold:
u(t) ≡ u(T) = a ∈ ∂Ω.
Hence also
u
n
(t, λ
n
) →u(t, λ
0
) ≡ u(T),
which further implies that
_
T
0
f(s, a)ds = 0,
where a ∈ ∂Ω, in contradiction to the assumptions of the theorem. Thus
d(id −S(, 0, ǫ), G, 0) = d(id −S(, 1, ǫ), G, 0)
by the homotopy invariance property of LeraySchauder degree, for all ǫ suﬃ
ciently small. On the other hand
d(id −S(, 0, ǫ), G, 0) = d(id −S(, 0, ǫ), G∩ R
N
, 0)
= d(id −S(, 0, ǫ), Ω, 0)
= d(g, Ω, 0) ,= 0,
if ǫ > 0 and (−1)
N
d(g, Ω, 0), if ǫ < 0, completing the proof.
8 Corollary Assume the hypotheses of Theorem 7 and assume that all possible
solutions u, for 0 < ǫ ≤ 1, of equation (14) are such that u / ∈ ∂G, where G is
given by (16). Then (11) has a T− periodic solution.
4. RESONANT EQUATIONS 97
4.3 A Li´enard type equation
In this section we apply Corollary 8 to prove the existence of periodic solu
tions of Li´enard type oscillators of the form
x
′′
+h(x)x
′
+x = e(t), (21)
where
e : R →R
is a continuous T− periodic forcing term and
h : R →R
is a continuous mapping. We shall prove the following result.
9 Theorem Assume that T < 2π. Then for every continuous T− periodic forcing
term e, equation (21) has a T− periodic response x.
We note that, since aside from the continuity assumption, nothing else is
assumed about h, we may, without loss in generality, assume that
_
T
0
e(s)ds = 0,
as follows from the substitution
y = x −
_
T
0
e(s)ds.
We hence shall make that assumption. In order to apply our earlier results, we
convert (21) into a system as follows:
x
′
= y
y
′
= −h(x)y −x −e(t),
(22)
and put
u =
_
x
y
_
, f(t, u) =
_
y
−h(x)y −x −e(t)
_
. (23)
We next shall show that the hypotheses of Theorem 7 and Corollary 8 may be
satisﬁed by choosing
Ω =
_
u =
_
x
y
_
: [x[ < R, [y[ < R
_
, (24)
where R is a suﬃciently large constant. We note that (15) holds for such choices
of Ω for any R > 0. Hence, if we are able to provide a priori bounds for solutions
of equation (14) for 0 < ǫ ≤ 1 for f given as above, the result will follow. Now
u =
_
x
y
_
,
is a solution of (14) whenever x satisﬁes
x
′′
+ǫh(x)x
′
+ǫ
2
x = ǫ
2
e(t). (25)
98
Integrating (25) from 0 to T, we ﬁnd that
_
T
0
x(s)ds = 0.
Multiplying (25) by x and integrating we obtain
−x
′

2
L
2 +ǫ
2
x
2
L
2 = ǫ
2
¸x, e)
L
2 , (26)
where ¸x, e)
L
2 =
_
T
0
x(s)e(s)ds. Now, since
x
2
L
2 ≤
T
2
4π
2
x
′

2
L
2, (27)
we obtain from (26)
_
1 −
T
2
4π
2
_
x
′

2
L
2 ≤ −ǫ
2
¸x, e)
L
2 , (28)
from which follows that
x
′

L
2 ≤
_
2πT
4π
2
−T
2
_
e
L
2, (29)
from which, in turn, we obtain
x
∞
≤
√
T
_
2πT
4π
2
−T
2
_
e
L
2, (30)
providing an a priori bound on x
∞
. We let
_
2πT
4π
2
−T
2
_
e
L
2 = M,
q = max
x≤M
[h(x)[, p = e
∞
.
Then
x
′′

∞
≤ ǫqx
′

∞
+ǫ
2
(M +p).
Hence, by Landau’ s inequality (Exercise 6, below), we obtain
x
′

2
∞
≤ 4M(ǫqx
′

∞
+ǫ
2
(M +p)),
from which follows a bound on x
′

∞
which is independent of ǫ, for 0 ≤ ǫ ≤ 1.
These considerations complete the proof Theorem 9.
4. RESONANT EQUATIONS 99
4.4 Partial resonance
This section is a continuation of of what has been discussed in Subsection
4.2. We shall impose conditions on the ﬁnite dimensional vector ﬁeld
x ∈ R
p
→ g(x)
g(x) = −
_
T
0
f(s, x, 0)ds,
(31)
which will guarantee the existence of solutions of an associated problem
u
′
= ǫf(t, u, v),
v
′
= Bv +ǫh(t, u, v)
u(0) = u(T), v(0) = v(T),
(32)
where ǫ is a small parameter and
f : R R
p
R
q
→R
p
h : R R
p
R
q
→R
q
are continuous and T−periodic with respect to t, p+q = N. Further B is a q q
constant matrix with the property that the system v
′
= Bv is nonresonant, i.e.
only has the trivial solution as a T−periodic solution. We have the following
theorem.
10 Theorem Assume the above and there exists a bounded open set Ω ⊂ R
p
such
that the mapping g deﬁned by (31) does not vanish on ∂Ω. Further assume that
d(g, Ω, 0) ,= 0, (33)
where d(g, Ω, 0) is the Brouwer degree. Then problem (32) has a solution for all
suﬃciently small ǫ.
Proof. To prove the existence of a T− periodic solution (u, v) of equation (32)
is equivalent to establishing the existence of a solution of
u(t) = u(T) +ǫ
_
t
0
f(s, u(s), v(s))ds
v(t) = e
Bt
v(T) +ǫe
Bt
_
t
0
e
−Bs
h(s, u(s), v(s))ds.
(34)
We consider equation (34) as an equation in the Banach space
E = C([0, T], R
p
) C([0, T], R
q
).
Let Λ be a bounded neighborhood of 0 ∈ R
q
. We deﬁne the bounded open set
G ⊂ E by
G = ¦(u, v) ∈ E : u : [0, T] →Ω, v : [0, T] →Λ¦. (35)
For (u, v) ∈
¯
G, deﬁne as in Subsection 4.2,
u(t, λ) = λu(t) + (1 −λ)u(T), 0 ≤ λ ≤ 1, (36)
and let
a(t, λ) = λt + (1 −λ)T, 0 ≤ λ ≤ 1. (37)
100
For 0 ≤ λ ≤ 1, 0 ≤ ǫ ≤ 1 deﬁne S = (S
1
, S
2
) : E [0, 1] [0, 1] →E by
S
1
(u, v, λ, ǫ)(t) = u(T) +ǫ
_
a(t,λ)
0
f(s, u(s, λ), λv(s))ds
S
2
(u, v, λ, ǫ)(t) = e
Bt
v(T) +λǫe
Bt
_
t
0
e
−Bs
h(s, u(s), v(s))ds
. (38)
Then S is a completely continuous mapping and the theorem will be proved
once we show that
d(id −S(, , 1, ǫ), G, 0) ,= 0, (39)
for ǫ suﬃciently small, for if this is the case, S(, 1, ǫ) has a ﬁxed point in G
which is equivalent to the assertion of the theorem.
To show that (39) holds we ﬁrst show that S(, λ, ǫ) has no zeros on ∂G for
all λ ∈ [0, 1] and ǫ suﬃciently small. This we argue in a manner similar to the
proof of Theorem 7. Hence
d(id −S(, , 0, ǫ), G, 0) = d(id −S(, , 1, ǫ), G, 0)
by the homotopy invariance property of LeraySchauder degree, for all ǫ suﬃ
ciently small. On the other hand
d(id −S(, , 0, ǫ), G, 0)
= d((−ǫ
_
T
0
f(s, , 0)ds, id −S
2
(, 0, ǫ)), G, (0, 0))
= d(−ǫ
_
T
0
f(s, , 0)ds, Ω, 0)d(id −S
2
(, 0, ǫ),
˜
G, 0)
= sgn det(I −e
BT
)d(−ǫ
_
T
0
f(s, , 0)ds, Ω, 0) ,= 0,
where
˜
G = C([0, T], Λ), completing the proof.
As before, we obtain the following corollary.
11 Corollary Assume the hypotheses of Theorem 10 and assume that all possible
solutions u, v, for 0 < ǫ ≤ 1, of equation (32) are such that u, v / ∈ ∂G, where G
is given by (35). Then (32) has a T− periodic solution for ǫ = 1.
5 Exercises
1. Consider equation (1) with A a constant matrix. Give conditions that
I − e
TR
be nonsingular, where I − e
TR
is given as in Theorem 3. Use
Theorem 5 to show that equation (1) has a T−periodic solution provided
the set of T−periodic solutions of (10) is a priori bounded for 0 ≤ ǫ < 1.
2. Prove Corollary 8.
3. Let f satisfy for some R > 0
f(t, x) x ,= 0, [x[ = R, 0 ≤ t ≤ T.
Prove that (11) has a T−periodic solution u with [u(t)[ < R, 0 ≤ t ≤ T.
5. EXERCISES 101
4. Let Ω ⊂ R
N
be an open convex set with 0 ∈ Ω and let f satisfy
f(t, x) n(x) ,= 0, x ∈ ∂Ω, 0 ≤ t ≤ T,
where for each x ∈ ∂Ω, n(x) is an outer normal vector to Ω at x. Prove
that (11) has a T−periodic solution u : [0, T] →Ω.
5. Verify inequality (27).
6. Let x ∈ C
2
[0, ∞). Use Taylor expansions to prove Landau’s inequality
x
′

2
∞
≤ 4x
∞
x
′′

∞
.
7. Complete the details in the proof of Theorem 10.
8. Assume that the unforced Li´enard equation (i.e. equation (21) with e ≡ 0)
has a nontrivial T−periodic solution x. Show that T ≥ 2π.
9. Prove Corollary 11.
102
Chapter VIII
Stability Theory
1 Introduction
In Chapter VI we studied in detail linear and perturbed linear systems of
diﬀerential equations. In the case of constant or periodic coeﬃcients we found
criteria which describe the asymptotic behavior of solutions (viz. Proposition 7
and Exercise 11 of Chapter 6.) In this chapter we shall consider similar problems
for general systems of the form
u
′
= f(t, u), (1)
where
f : R R
N
→R
N
is a continuous function.
If u : [t
0
, ∞) → R
N
is a given solution of (1), then discussing the behavior
of another solution v of this equation relative to the solution u, i.e. discussing
the behavior of the diﬀerence v −u is equivalent to studying the behavior of the
solution z = v −u of the equation
z
′
= f(t, z +u(t)) −f(t, u(t)), (2)
relative to the trivial solution z ≡ 0. Thus we may, without loss in generality,
assume that (1) has the trivial solution as a reference solution, i.e.
f(t, 0) ≡ 0,
an assumption we shall henceforth make.
2 Stability Concepts
There are various stability concepts which are important in the asymptotic
behavior of systems of diﬀerential equations. We shall discuss here some of them
and their interrelationships.
103
104
1 Deﬁnition We say that the trivial solution of (1) is:
(i) stable (s) on [t
0
, ∞), if for every ǫ > 0 there exists δ > 0 such that any
solution v of (1) with [v(t
0
)[ < δ exists on [t
0
, ∞) and satisﬁes [v(t)[ < ǫ, t
0
≤
t < ∞;
(ii) asymptotically stable (a.s) on [t
0
, ∞), if it is stable and lim
t→∞
v(t) = 0,
where v is as in (i);
(iii) unstable (us), if it is not stable;
(iv) uniformly stable (u.s) on [t
0
, ∞), if for every ǫ > 0 there exists δ > 0
such that any solution v of (1) with [v(t
1
)[ < δ, t
1
≥ t
0
exists on [t
1
, ∞) and
satisﬁes [v(t)[ < ǫ, t
1
≤ t < ∞;
(v) uniformly asymptotically stable (u.a.s), if it is uniformly stable and there
exists δ > 0 such that for all ǫ > 0 there exists T = T(ǫ) such that any solution v
of (1) with [v(t
1
)[ < δ, t
1
≥ t
0
exists on [t
1
, ∞) and satisﬁes [v(t)[ < ǫ, t
1
+T ≤
t < ∞;
(v) strongly stable (s.s) on [t
0
, ∞), if for every ǫ > 0 there exists δ > 0
such that any solution v of (1) with [v(t
1
)[ < δ exists on [t
0
, ∞) and satisﬁes
[v(t)[ < ǫ, t
0
≤ t < ∞.
2 Proposition The following implications are valid:
u.a.s ⇒ a.s
⇓ ⇓
s.s ⇒ u.s ⇒ s.
If the equation (1) is autonomous, i.e. f is independent of t, then the above
implications take the from
u.a.s ⇔ a.s
⇓ ⇓
u.s ⇔ s.
The following examples of scalar diﬀerential equations will serve to illustrate
the various concepts.
3 Example 1. The zero solution of u
′
= 0 is stable but not asymptotically
stable.
2. The zero solution of u
′
= u
2
is unstable.
3. The zero solution of u
′
= −u is uniformly asymptotically stable.
4. The zero solution of u
′
= a(t)u is asymptotically stable if and only if
lim
t→∞
_
t
t0
a(s)ds = −∞. It is uniformly stable if and only if
_
t
t1
a(s)ds is
bounded above for t ≥ t
1
≥ t
0
. Letting a(t) = sinlog t + cos log t − α one
sees that asymptotic stability holds but uniform stability does not.
3. STABILITY OF LINEAR EQUATIONS 105
3 Stability of Linear Equations
In the case of a linear system (A ∈ C(R →L(R
N
, R
N
)))
u
′
= A(t)u, (3)
a particular stability property of any solution is equivalent to that stability
property of the trivial solution. Thus one may ascribe that property to the
equation and talk about the equation (3) being stable, uniformly stable, etc.
The stability concepts may be expressed in terms of conditions imposed on a
fundamental matrix Φ.
4 Theorem Let Φ be a fundamental matrix solution of (3). Then equation (3)
is :
(i) stable if and only if there exists K > 0 such that
[Φ(t)[ ≤ K, t
0
≤ t < ∞; (4)
(ii) uniformly stable if and only if there exists K > 0 such that
[Φ(t)Φ
−1
(s)[ ≤ K, t
0
≤ s ≤ t < ∞; (5)
(iii) strongly stable if and only if there exists K > 0 such that
[Φ(t)[ ≤ K, [Φ
−1
(t)[ ≤ K, t
0
≤ t < ∞; (6)
(iv) asymptotically stable if and only if
lim
t→∞
[Φ(t)[ = 0; (7)
(v) uniformly asymptotically stable if and only if there exist K > 0, α > 0
such that
[Φ(t)Φ
−1
(s)[ ≤ Ke
−α(t−s)
, t
0
≤ s ≤ t < ∞. (8)
Proof. We shall demonstrate the last part of the theorem and leave the demon
stration of the remaining parts as an exercise. We may assume without loss in
generality that Φ(t
0
) = I, the N N identity matrix, since conditions (4) 
(8) will hold for any fundamental matrix if and only if they hold for a par
ticular one. Thus, let us assume that (8) holds. Then the second part of the
theorem guarantees that the equation is uniformly stable. Moreover, for each
ǫ, 0 < ǫ < K, if we put T = −
1
α
log
ǫ
K
, then for ξ ∈ R
N
, [ξ[ ≤ 1, we have
[Φ(t)Φ
−1
(t
1
)ξ[ ≤ Ke
−α(t−t1)
, if t
1
+ T < t. Thus we have uniform asymptotic
stability.
Conversely, if the equation is uniformly asymptotically stable, then there
exists δ > 0 and for all ǫ, 0 < ǫ < δ, there exists T = T(ǫ) > 0 such that if
[ξ[ < δ, then
[Φ(t)Φ
−1
(t
1
)ξ[ < ǫ, t ≥ t
1
+T, t
1
≥ t
0
.
106
In particular
[Φ(t +T)Φ
−1
(t)ξ[ < ǫ, [ξ[ < δ,
or
[Φ(t +T)Φ
−1
(t)
ξ
δ
[ <
ǫ
δ
and thus
[Φ(t +T)Φ
−1
(t)[ ≤
ǫ
δ
< 1, t ≥ t
0
.
Furthermore, since we have uniform stability,
[Φ(t +h)Φ
−1
(t)[ ≤ K, t
0
≤ t, 0 ≤ h ≤ T.
If t ≥ t
1
, we obtain for some integer n, that t
1
+nT ≤ t < t
1
+ (n + 1)T, and
[Φ(t)Φ
−1
(t
1
)[ ≤ [Φ(t)Φ
−1
(t
1
+nT)[[Φ(t
1
+nT)Φ
−1
(t
1
)[
≤ K[Φ(t
1
+nT)Φ
−1
(t
1
+ (n −1)T)[
[Φ(t
1
+T)Φ
−1
(t
1
)[
≤ K
_
ǫ
δ
_
n
.
Letting α = −
1
T
log
ǫ
δ
, we get
[Φ(t)Φ
−1
(t
1
)[ ≤ Ke
−nαT
= Ke
−nαT
e
−αT δ
ǫ
< K
δ
ǫ
e
−α(t−t1)
, t ≥ t
1
≥ t
0
.
In the case that the matrix A is independent of t one obtains the following
corollary.
5 Corollary Equation (3) is stable if and only if every eigenvalue of A has non
positive real part and those with zero real part are semisimple. It is strongly
stable if and only if all eigenvalues of A have zero real part and are semisimple.
It is asymptotically stable if and only if all eigenvalues have negative real part.
Using the AbelLiouville formula, we obtain the following result.
6 Theorem Equation (3) is unstable whenever
limsup
t→∞
_
t
t0
traceA(s)ds = ∞. (9)
If (3) is stable, then it is strongly stable if and only if
liminf
t→∞
_
t
t0
traceA(s)ds > −∞. (10)
Additional stability criteria for linear systems abound. The following concept of
the measure of a matrix due to Lozinskii and Dahlquist (see [7]) is particularly
useful in numerical computations. We provide a brief discussion.
3. STABILITY OF LINEAR EQUATIONS 107
7 Deﬁnition For an N N matrix A we deﬁne
µ(A) = lim
h→0+
[I +hA[ −[I[
h
, (11)
where [ [ is a matrix norm induced by a norm [ [ in R
N
and I is the N N
identity matrix.
We have the following proposition:
8 Proposition For any N N matrix A, µ(A) exists and satisﬁes:
1. µ(αA) = αµ(A), α ≥ 0;
2. [µ(A)[ ≤ [A[;
3. µ(A +B) ≤ µ(A) + µ(B);
4. [µ(A) −µ(B)[ ≤ [A −B[.
If u is a solution of equation (3), then the function r(t) = [u(t)[ has a right
derivative r
′
+
(t) at every point t for every norm [ [ in R
N
and r
′
+
(t) satisﬁes
r
′
+
(t) −µ(A(t))r(t) ≤ 0. (12)
Using this inequality we obtain the following proposition.
9 Proposition Let A : [t
0
, ∞) → R
N×N
be a continuous matrix and let u be a
solution of (3) on [t
0
, ∞). Then
[u(t)[e
−
R
t
t
0
µ(A(s))ds
, t
0
≤ t < ∞ (13)
is a nonincreasing function of t and
[u(t)[e
R
t
t
0
µ(−A(s))ds
, t
0
≤ t < ∞ (14)
is a nondecreasing function of t. Furthermore
[u(t
0
)[e
−
R
t
t
0
µ(−A(s))ds
≤ [u(t)[ ≤ [u(t
0
)[e
R
t
t
0
µ(A(s))ds
. (15)
This proposition has, for constant matrices A, the immediate corollary.
10 Corollary For any N N constant matrix A the following inequalities hold.
e
−tµ(−A)
≤
¸
¸
e
tA
¸
¸
≤ e
tµ(A)
.
The following theorem provides stability criteria for the system (3) in terms of
conditions on the measure of the coeﬃcient matrix. Its proof is again left as an
exercise.
11 Theorem The system (3) is:
108
1. unstable, if
liminf
t→∞
_
t
t0
µ(−A(s))ds = −∞;
2. stable, if
limsup
t→∞
_
t
t0
µ(A(s))ds < ∞;
3. asymptotically stable, if
lim
t→∞
_
t
t0
µ(A(s))ds = −∞;
4. uniformly stable, if
µ(A(t)) ≤ 0, t ≥ t
0
;
5. uniformly asymptotically stable, if
µ(A(t)) ≤ −α < 0, t ≥ t
0
.
4 Stability of Nonlinear Equations
In this section we shall consider stability properties of nonlinear equations
of the form
u
′
= A(t)u +f(t, u), (16)
where
f : R R
N
→R
N
is a continuous function with
[f(t, x)[ ≤ γ(t)[x[, x ∈ R
N
, (17)
where γ is some positive continuous function and A is a continuous NN matrix
deﬁned on R. The results to be discussed are consequences of the variation of
constants formula (Proposition VI.6) and the stability theorem Theorem 4. We
shall only present a sample of results. We refer to [7], [5], [20], and [21], where
further results are given. See also the exercises below.
Throughout this section Φ(t) will denote a fundamental matrix solution of
the homogeneous (unperturbed) linear problem (3). We hence know that if u is
a solution of equation (16), then u satisﬁes the integral equation
u(t) = Φ(t)
_
Φ
−1
(t
0
)u(t
0
) +
_
t
t0
Φ
−1
(s)f(s, u(s))ds
_
. (18)
The following theorem will have uniform and strong stability as a conse
quence.
4. STABILITY OF NONLINEAR EQUATIONS 109
12 Theorem Let f satisfy (17) with
_
∞
γ(s)ds < ∞. Further assume that
[Φ(t)Φ
−1
(s)[ ≤ K, t
0
≤ s ≤ t < ∞.
Then there exists a positive constant L = L(t
0
) such that any solution u of (16)
is deﬁned for t ≥ t
0
and satisﬁes
[u(t)[ ≤ L[u(t
1
)[, t ≥ t
1
≥ t
0
.
If in addition Φ(t) →0 as t →∞, then u(t) →0 as t →∞.
Proof. If u is a solution of of (16) it also satisﬁes (18). Hence
[u(t)[ ≤ K[u(t
1
)[ +K
_
t
t1
γ(s)[u(s)[ds, t ≥ t
1
,
and
[u(t)[ ≤ K[u(t
1
)[e
K
R
t
t
1
γ(s)ds
≤ L[u(t
1
)[, t ≥ t
1
,
where
L = Ke
K
R
∞
t
0
γ(s)ds
.
To prove the remaining part of the theorem, we note from (1) that
[u(t)[ ≤ [Φ(t)[
_
[Φ
−1
(t
0
)u(t
0
)[ +[
_
t1
t0
Φ
−1
(s)f(s, u(s))ds[
_
+ [
_
t
t1
Φ(t)Φ
−1
(s)f(s, u(s))ds[
Now
¸
¸
¸
¸
_
t
t1
Φ(t)Φ
−1
(s)f(s, u(s))ds
¸
¸
¸
¸
≤ KL[u(t
0
)[
_
∞
t1
γ(s)ds.
This together with the fact that Φ(t) →0 as t →∞ completes the proof.
13 Remark It follows from Theorem 4 that the conditions of the above theo
rem imply that the perturbed system (16) is uniformly (asymptotically) stable
whenever the unperturbed system is uniformly (asymptotically) stable. The
conditions of Theorem 12 also imply that the zero solution of the perturbed
system is strongly stable, whenever the unperturbed system is strongly stable.
A further stability result is the following. Its proof is delegated to the exer
cises.
14 Theorem Assume that there exist K > 0, α > 0 such that
[Φ(t)Φ
−1
(s)[ ≤ Ke
−α(t−s)
, t
0
≤ s ≤ t < ∞. (19)
and f satisﬁes (17) with γ <
α
K
a constant. Then any solution u of (16) is
deﬁned for t ≥ t
0
and satisﬁes
[u(t)[ ≤ Ke
−β(t−t1)
[u(t
1
)[, t ≥ t
1
≥ t
0
,
where β = α −γK > 0.
110
15 Remark It again follows from Theorem 4 that the conditions of the above
theorem imply that the perturbed system (16) is uniformly asymptotically stable
whenever the unperturbed system is uniformly asymptotically stable.
5 Lyapunov Stability
5.1 Introduction
In this section we shall introduce some geometric ideas, which were ﬁrst
formulated by Poincar´e and Lyapunov, about the stability of constant solutions
of systems of the form
u
′
= f(t, u), (20)
where
f : R R
N
→R
N
is a continuous function. As observed earlier, we may assume that f(t, 0) =
0 and we discuss the stability properties of the trivial solution u ≡ 0. The
geometric ideas amount to constructing level surfaces in R
N
which shrink to 0
and which have the property that orbits associated with (20) cross these level
surfaces transversally toward the origin, thus being guided to the origin. To
illustrate, let us consider the following example.
x
′
= ax −y +kx
_
x
2
+y
2
_
y
′
= x −ay +ky
_
x
2
+y
2
_
,
(21)
where a is a constant [a[ < 1. Obviously x = 0 = y is a stationary solution of
(21). Now consider the family of curves in R
2
given by
v(x, y) = x
2
−2axy +y
2
= constant, (22)
a family of ellipses which share the origin as a common center. If we consider an
orbit ¦(x(t), y(t)) : t ≥ t
0
¦ associated with (21), then as t varies, the orbit will
cross members of the above family of ellipses. Computing the scalar product of
the tangent vector of an orbit with the gradient to an ellipse we ﬁnd:
∇v (x
′
, y
′
) = 2k
_
x
2
+y
2
_ _
x
2
+y
2
−2axy
_
,
which is clearly negative, whenever k is (and positive if k is). Of course, we
may view v(x(t), y(t)) as a norm of the point (x(t), y(t)) and ∇v (x
′
, y
′
) =
d
dt
v(x(t), y(t)). Thus, if k < 0, v(x(t), y(t)) will be a strictly decreasing function
and hence should approach a limit as t → ∞. That this limit must be zero
follows by an indirect argument. That is, the orbit tends to the origin and the
zero solution attracts all orbits, and hence appears asymptotically stable.
5. LYAPUNOV STABILITY 111
5.2 Lyapunov functions
Let
v : R R
N
→R, v(t, 0) = 0, t ∈ R
be a continuous functional. We shall introduce the following terminology.
16 Deﬁnition The functional v is called:
1. positive deﬁnite, if there exists a continuous nondecreasing function φ :
[0, ∞) →[0, ∞), with φ(0) = 0, φ(r) ,= 0, r ,= 0 and
φ([x[) ≤ v(t, x), x ∈ R
N
, t ≥ t
0
;
2. radially unbounded, if it is positive deﬁnite and
lim
r→∞
φ(r) = ∞;
3. decrescent, if it is positive deﬁnite and there exists a continuous increasing
function ψ : [0, ∞) →[0, ∞), with ψ(0) = 0, and
ψ([x[) ≥ v(t, x), x ∈ R
N
, t ≥ t
0
.
We have the following stability criteria.
17 Theorem Let there exist a positive deﬁnite functional v and δ
0
> 0 such that
for every solution u of (20) with [u(t
0
)[ ≤ δ
0
, the function v
∗
(t) = v(t, u(t)) is
nonincreasing with respect to t, then the trivial solution of (20) is stable.
Proof. Let 0 < ǫ ≤ δ
0
be given and choose δ = δ(ǫ) such that v(t
0
, u
0
) < φ(ǫ),
for [u
0
[ < δ. Let u be a solution of (20) with u(t
0
) = u
0
. Then v
∗
(t) = v(t, u(t))
is nonincreasing with respect to t, and hence
v
∗
(t) ≤ v
∗
(t
0
) = v(t
0
, u
0
).
Therefore
φ([u(t)[) ≤ v(t, u(t)) = v
∗
(t) ≤ v
∗
(t
0
) = v(t
0
, u
0
) < φ(ǫ).
Since φ is nondecreasing, the result follows.
If v is also decrescent we obtain the stronger result.
18 Theorem Let there exist a positive deﬁnite functional v which is decrescent
and δ
0
> 0 such that for every solution u of (20) with [u(t
1
)[ ≤ δ
0
, t
1
≥ t
0
the
function v
∗
(t) = v(t, u(t)) is nonincreasing with respect to t, then the trivial
solution of (20) is uniformly stable.
112
Proof. We have already shown that the trivial solution is stable. Let 0 < ǫ ≤ δ
0
be given and let δ = ψ
−1
(φ(ǫ)). Let u be a solution of (20) with u(t
1
) = u
0
with
[u
0
[ < δ. Then
φ([u(t)[) ≤ v(t, u(t)) = v
∗
(t) ≤ v
∗
(t
1
)
= v(t
1
, u
0
) ≤ ψ([u
0
[)
< ψ(δ) = φ(ǫ).
The result now follows from the monotonicity assumption on φ.
As an example we consider the two dimensional system
x
′
= a(t)y +b(t)x
_
x
2
+y
2
_
y
′
= −a(t)x +b(t)y
_
x
2
+y
2
_
,
(23)
where the coeﬃcient functions a and b are continuous with b ≤ 0. We choose
v(x, y) = x
2
+y
2
, then, since
dv
∗
dt
=
∂v
∂t
+∇v f(t, u).
we obtain
dv
∗
dt
= 2b(t)
_
x
2
+y
2
_
2
≤ 0.
Since v is positive deﬁnite and decrescent, it follows from Theorem 18 that the
trivial solution of (23) is uniformly stable.
We next prove an instability theorem.
19 Theorem Assume there exists a continuous functional
v : R R
N
→R
with the properties:
1. there exists a continuous increasing function ψ : [0, ∞) → [0, ∞), such
that ψ(0) = 0 and
[v(t, x)[ ≤ ψ([x[);
2. for all δ > 0, and t
1
≥ t
0
, there exists x
0
, [x
0
[ < δ such that v(t
1
, x
0
) < 0;
3. if u is any solution of (23) with u(t) = x, then
lim
h→0+
v(t +h, u(t +h)) −v(t, x)
h
≤ −c([x[),
where c is a continuous increasing function with c(0) = 0.
Then the trivial solution of (23) is unstable.
5. LYAPUNOV STABILITY 113
Proof. Assume the trivial solution is stable. Then for every ǫ > 0 there exists
δ > 0 such that any solution u of (23) with [u(t
0
)[ < δ exists on [t
0
, ∞) and
satisﬁes [u(t)[ < ǫ, t
0
≤ t < ∞. Choose x
0
, [x
0
[ < δ such that v(t
0
, x
0
) < 0 and
let u be a solution with u(t
0
) = x
0
. Then
[v(t, u(t))[ ≤ ψ([u(t)[) ≤ ψ(ǫ). (24)
The third property above implies that v(t, u(t)) is nonincreasing and hence for
t ≥ t
0
,
v(t, u(t)) ≤ v(t
0
, x
0
) < 0.
Thus
[v(t
0
, x
0
)[ ≤ ψ([u(t)[)
and
ψ
−1
([v(t
0
, x
0
)[) ≤ [u(t)[.
We therefore have
v(t, u(t)) ≤ v(t
0
, x
0
) −
_
t
t0
c([u(s)[)ds,
which by (24) implies that
v(t, u(t)) ≤ v(t
0
, x
0
) −(t −t
0
)c(ψ
−1
([v(t
0
, x
0
)[),
and thus
lim
t→∞
v(t, u(t)) = −∞,
contradicting (24).
We ﬁnally develop some asymptotic stability criteria.
20 Theorem Let there exist a positive deﬁnite functional v(t, x) such that
dv
∗
dt
=
dv(t, u(t))
dt
≤ −c(v(t, u(t)),
for every solution u of (20) with [u(t
0
)[ ≤ δ
0
, with c a continuous increasing
function and c(0) = 0. Then the trivial solution is asymptotically stable. If v is
also decrescent. Then the asymptotic stability is uniform.
Proof. The hypotheses imply that the trivial solution is stable, as was proved
earlier. Hence, if u is a solution of (20) with [u(t)[ ≤ δ
0
, then
v
0
= lim
t→∞
v(t, u(t))
exists. An easy indirect argument shows that v
0
= 0. Hence, since v is positive
deﬁnite,
lim
t→∞
φ([u(t)[) = 0,
114
implying that
lim
t→∞
[u(t)[ = 0,
since φ is increasing. The proof that the trivial solution is uniformly asymptot
ically stable, whenever v is also decrescent, is left as an exercise.
In the next section we shall employ the stability criteria just derived, to,
once more study perturbed linear systems, where the associated linear system
is a constant coeﬃcient system. This we do by showing how appropriate Lya
punov functionals may be constructed. The construction is an exercise in linear
algebra.
5.3 Perturbed linear systems
Let A be a constant N N matrix and let g : [t
0
, ∞) R
N
→ R
N
be a
continuous function such that
g(t, x) = o([x[),
uniformly with respect to t ∈ [t
0
, ∞). We shall consider the equation
u
′
= Au +g(t, u) (25)
and show how to construct Lyapunov functionals to test the stability of the
trivial solution of this system.
The type of Lyapunov functional we shall be looking for are of the form
v(x) = x
T
Bx,
where B is a constant N N matrix, i.e. we are looking for v as a quadratic
form. If u is a solution of (25) then
dv
∗
dt
=
dv(t, u(t))
dt
= u
T
_
A
T
B +BA
_
u +g
T
(t, u)Bu +u
T
Bg(t, u).
(26)
Hence, given A, if B can be found so that C = A
T
B+BA has certain deﬁniteness
properties, then the results of the previous section may be applied to determine
stability or instability of the trivial solution. To proceed along these lines we
need some linear algebra results.
21 Proposition Let A be a constant N N matrix having the property that for
any eigenvalue λ of A, −λ is not an eigenvalue. Then for any N N matrix C,
there exists a unique N N matrix B such that C = A
T
B +BA.
Proof. On the space of N N matrices deﬁne the bounded linear operator L
by
L(B) = A
T
B +BA.
5. LYAPUNOV STABILITY 115
Then L may be viewed as a bounded linear operator of R
N×N
to itself, hence it
will be a bijection provided it does not have 0 as an eigenvalue. Once we show
the latter, the result follows. Thus let µ be an eigenvalue of L, i.e., there exists
a nonzero matrix B such that
L(B) = A
T
B +BA = µB.
Hence
A
T
B +B(A −µI) = 0.
From this follows
B(A −µI)
n
=
_
−A
T
_
n
B,
for any integer n ≥ 1, hence for any polynomial p
Bp(A −µI) = p(−A
T
)B. (27)
Since, on the other hand, if F and G are two matrices with no common eigen
values, there exists a polynomial p such that p(F) = I, p(G) = 0, (27) implies
that A−µI and −A
T
must have a common eigenvalue. From which follows that
µ is the sum of two eigenvalues of A, which, by hypothesis cannot equal 0.
This proposition has the following corollary.
22 Corollary Let A be a constant N N matrix. Then for any N N matrix C,
there exists µ > 0 and a unique NN matrix B such that 2µB+C = A
T
B+BA.
Proof. Let
S = ¦λ ∈ C : λ = λ
1
+λ
2
¦,
where λ
1
and λ
2
are eigenvalues of A. Since S is a ﬁnite set, there exists r
0
> 0,
such that λ(,= 0) ∈ S implies that [λ[ > r
0
. Choose 0 < µ ≤ r
0
and consider the
matrix A
1
= A − µI. We may now apply Proposition 21 to the matrix A
1
and
ﬁnd for a given matrix C, a unique matrix B such that C = A
T
1
B + BA
1
, i.e.
2µBC = A
T
B +BA.
23 Corollary Let A be a constant N N matrix having the property that all
eigenvalues λ of A have negative real parts. Then for any negative deﬁnite
N N matrix C, there exists a unique positive deﬁnite N N matrix B such
that C = A
T
B +BA.
Proof. Let C be a negative deﬁnite matrix and let B be given by Proposition
21, which may be applied since all eigenvalues of A have negative real part. Let
v(x) = x
T
Bx, and let u be a solution of u
′
= Au, u(0) = x
0
,= 0. Then
dv
∗
dt
=
dv(u(t))
dt
= u
T
_
A
T
B +BA
_
u = u
T
Cu ≤ −µ[u[
2
,
116
since C is negative deﬁnite. Since lim
t→∞
u(t) = 0 (all eigenvalues of A have
negative real part!), it follows that lim
t→∞
v(u(t)) = 0. We also have
v(u(t)) ≤ v(x
0
) −
_
t
0
µ[u(s)[
2
ds,
from which follows that v(x
0
) > 0. Hence B is positive deﬁnite.
The next corollary follows from stability theory for linear equations and what
has just been discussed.
24 Corollary A necessary and suﬃcient condition that an N N matrix A have
all of its eigenvalues with negative real part is that there exists a unique positive
deﬁnite matrix B such that
A
T
B +BA = −I.
We next consider the nonlinear problem (25) with
g(t, x) = o([x[), (28)
uniformly with respect to t ∈ [t
0
, ∞), and show that for certain types of matrices
A the trivial solution of the perturbed system has the same stability property
as that of the unperturbed problem. The class of matrices we shall consider is
the following.
25 Deﬁnition We call an N N matrix A critical if all its eigenvalues have non
positive real part and there exists at least one eigenvalue with zero real part.
We call it noncritical otherwise.
26 Theorem Assume A is a noncritical N N matrix and let g satisfy (28).
Then the stability behavior of the trivial solution of (25) is the same as that
of the trivial solution of u
′
= Au, i.e the trivial solution of (25) is uniformly
asymptotically stable if all eigenvalues of A have negative real part and it is
unstable if A has an eigenvalue with positive real part.
Proof. (i) Assume all eigenvalues of A have negative real part. By the above
exists a unique positive deﬁnite matrix B such that
A
T
B +BA = −I.
Let v(x) = x
T
Bx. Then v is positive deﬁnite and if u is a solution of (25) it
satisﬁes
dv
∗
dt
=
dv(u(t))
dt
= −[u(t)[
2
+g
T
(t, u)Bu +u
T
Bg(t, u).
(29)
Now
[g
T
(t, u)Bu +u
T
Bg(t, u)[ ≤ 2[g(t, u)[[B[[u[.
5. LYAPUNOV STABILITY 117
Choose r > 0 such that [x[ ≤ r implies
[g(t, u)[ ≤
1
4
[B[
−1
[u[,
then
dv
∗
dt
≤ −
1
2
[u(t)[
2
,
as long as [u(t)[ ≤ r. The result now follows from Theorem 20.
(ii) Let A have an eigenvalue with positive real part. Then there exists µ > 0
such that 2µ < [λ
i
+λ
j
[, for all eigenvalues λ
i
, λ
j
of A, and a unique matrix B
such that
A
T
B +BA = 2µB −I,
as follows from Corollary 22. We note that B cannot be positive deﬁnite nor
positive semideﬁnite for otherwise we must have, letting v(x) = x
T
Bx,
dv
∗
dt
= 2µv
∗
(t) −[u(t)[
2
,
or
e
−2µt
v
∗
(t) −v
∗
(0) = −
_
t
0
e
−2µs
[u(s)[
2
ds
i.e
0 ≤ v
∗
(0) −
_
t
0
e
−2µs
[u(s)[
2
ds
for any solution u, contradicting the fact that solutions u exist for which
_
t
0
e
−2µs
[u(s)[
2
ds
becomes unbounded as t → ∞ (see Chapter ??). Hence there exists x
0
,= 0,
of arbitrarily small norm, so that v(x
0
) < 0. Let u be a solution of (25). If
the trivial solution were stable, then [u(t)[ ≤ r for some r > 0. Again letting
v(x) = x
T
Bx we obtain
dv
∗
dt
= 2µv
∗
(t) −[u(t)[
2
+g
T
(t, u)Bu +u
T
Bg(t, u).
We can choose r so small that
2[g(t, u)[[B[[u[ ≤
1
2
[u[
2
, [u[ ≤ r,
hence
e
−2µt
v
∗
(t) −v
∗
(0) ≤ −
1
2
_
t
0
e
−2µs
[u(s)[
2
ds,
118
i.e.
v
∗
(t) ≤ e
2µt
v
∗
(0) →−∞,
contradicting that v
∗
(t) is bounded for bounded u. Hence u cannot stay bounded,
and we have instability.
If it is the case that the matrix A is a critical matrix, the trivial solution of
the linear system may still be stable or it may be unstable. In either case, one
may construct examples, where the trivial solution of the perturbed problem
has either the same or opposite stability behavior as the unperturbed system.
6 Exercises
1. Prove Proposition 2.
2. Verify the last part of Example 3.
3. Prove Theorem 4.
4. Establish a result similar to Corollary 5 for linear periodic systems.
5. Prove Theorem 6.
6. Show that the zero solution of the scalar equation
x
′′
+a(t)x = 0
is strongly stable, whenever it is stable.
7. Prove Proposition 8.
8. Establish inequality (12).
9. Prove Proposition 9.
10. Prove Corollary 10.
11. Prove Theorem 11.
12. Show that if
_
∞
µ(A(s))ds exists, then any nonzero solution u of (3)
satisﬁes
limsup
t→∞
[u(s)[ < ∞.
On the other hand, if
_
∞
µ(−A(s))ds exists, then
0 ,= liminf
t→∞
[u(s)[ ≤ ∞.
13. If A is a constant N N matrix show that µ(A) is an upper bound for
the real parts of the eigenvalues of A.
6. EXERCISES 119
14. Verify the following table:
[x[ [A[ µ(A)
max
i
[x
i
[ max
i
k
[a
ik
[ max
i
_
a
ii
+
k=i
[a
ik
[
_
i
[x
i
[ max
k
i
[a
ik
[ max
k
_
a
kk
+
i=k
[a
ik
[
_
_
i
[x
i
[
2
λ
∗
λ
∗
where λ
∗
, λ
∗
are, respectively, the square root of the largest eigenvalue of
A
T
A and the largest eigenvalue of
1
2
_
A
T
+A
_
.
15. Prove that if [A[
2
=
i
j
a
2
ij
, then
µ(A) =
trace(A)
√
N
.
16. If the linear system (3) is uniformly (asymptotically) stable and
B : [t
0
, ∞) →R
N×N
is continuous and satisﬁes
_
∞
[B(s)[ds < ∞,
then the system
u
′
= (A(t) +B(t))u
is uniformly (asymptotically) stable.
17. Prove Theorem 14.
18. If the linear system (3) is uniformly asymptotically stable and B : [t
0
, ∞) →
R
N×N
is continuous and satisﬁes
lim
t→∞
[B(t)[ = 0,
then the system
u
′
= (A(t) +B(t))u
is also uniformly asymptotically stable.
19. Complete the proof of Theorem 20.
120
20. Consider the Li´enard oscillator
x
′′
+f(x)x
′
+ g(x) = 0,
where f and g are continuous functions with g(0) = 0. Assume that there
exist α > 0, β > 0 such that
_
x
0
g(s)ds < β ⇒ [x[ < α,
and
0 < [x[ < α ⇒ g(x)F(x) > 0,
where F(x) =
_
x
0
f(s)ds. The equation is equivalent to the system
x
′
= y −F(x), y
′
= −g(x).
Using the functional v(x, y) =
1
2
y
2
+
_
x
0
g(s)ds prove that the trivial solu
tion is asymptotically stable.
21. Let a be a positive constant. Use the functional v(x, y) =
1
2
_
y
2
+x
2
_
to
show that the trivial solution of
x
′′
+a
_
1 −x
2
_
x
′
+x = 0,
is asymptotically stable. What can one say about the zero solution of
x
′′
+a
_
x
2
−1
_
x
′
+x = 0,
where again a is a positive constant?
22. Prove Corollary 23.
23. Consider the situation of Proposition 21 and assume all eigenvalues of
A have negative real part. Show that for given C the matrix B of the
Proposition is given by
B =
_
∞
0
e
A
T
t
Ce
At
dt.
Hint: Find a diﬀerential equation that is satisﬁed by the matrix
P(t) =
_
∞
t
e
A
T
(τ−t)
Ce
A(τ−t)
dτ
and show that it is a constant matrix.
6. EXERCISES 121
24. Consider the system
x
′
= y +ax
_
x
2
+y
2
_
y
′
= −x +ay
_
x
2
+y
2
_
.
Show that the trivial solution is stable if a > 0 and unstable if a < 0.
Contrast this with Theorem 26.
25. Show that the trivial solution of
x
′
= −2y
3
y
′
= x
is stable. Contrast this with Theorem 26.
122
Chapter IX
Invariant Sets
1 Introduction
In this chapter, we shall present some of the basic results about invariant
sets for solutions of initial value problems for systems of autonomous (i.e. time
independent) ordinary diﬀerential equations. We let D be an open connected
subset of R
N
, N ≥ 1, and let
f : D →R
N
be a locally Lipschitz continuous mapping.
We consider the initial value problem
u
′
= f(u)
u(0) = u
0
∈ D
(1)
and seek suﬃcient conditions on subsets M ⊂ D for solutions of (1) to have the
property that ¦u(t)¦ ⊂ M, t ∈ I, whenever u
0
∈ M, where I is the maximal
interval of existence of the solution u. We note that the initial value problem (1)
is uniquely solvable since f satisﬁes a local Lipschitz condition (viz. Chapter
V). Under these assumptions, we have for each u
0
∈ D a maximal interval of
existence I
u0
and a function
u : I
u0
→ D
t → u(t, u
0
),
(2)
i.e., we obtain a mapping
u : I
u0
D →D
(t, u
0
) →u(t, u
0
).
(3)
Thus if we let
U = ∪
v∈D
I
v
¦v¦ ⊂ R R
N
,
we obtain a mapping
u : U →D,
which has the following properties:
123
124
1 Lemma 1. u is continuous,
2. u(0, u
0
) = u
0
, ∀u
0
∈ D
3. for each u
0
∈ D, s ∈ I
u0
and t ∈ I
u(s,u0)
we have s + t ∈ I
u0
and
u(s +t, u
0
) = u(t, u(s, u
0
)).
A mapping having the above properties is called a ﬂow on D and we shall
henceforth, in this chapter, use this term freely and call u the ﬂow determined
by f.
2 Orbits and Flows
Let u be the ﬂow determined by f via the initial value problem (1). We shall
use the following (standard) convention. If S ⊂ I
u0
= (t
−
u0
, t
+
u0
),
u(S, u
0
) = ¦v : v = u(t, u
0
), t ∈ S ⊂ I
u0
¦.
We shall call
γ(v) = u(I
v
, v)
the orbit of v,
γ
+
(v) = u([0, t
+
v
), v)
the positive semiorbit of v and
γ
−
(v) = u((t
−
v
, 0]), v)
the negative semiorbit of v.
Furthermore, we call v ∈ D a stationary or critical point of the ﬂow, when
ever f(v) = 0. It is, of course immediate, that if v ∈ D is a stationary point,
then I
v
= R and
γ(v) = γ
+
(v) = γ
−
(v) = ¦v¦.
We call v ∈ D a periodic point of period T of the ﬂow, whenever there exists
T > 0, such that u(0, v) = u(T, v). If v is a periodic point which is not a critical
point, one calls T > 0 its minimal period, provided u(0, v) ,= u(t, v), 0 < t < T.
We have the following proposition.
2 Proposition Let u be the ﬂow determined by f and let v ∈ D. Then either:
1. v is a stationary point, or
2. v is a periodic point having a minimal positive period, or
3. the ﬂow u(, v) is injective.
If γ
+
(v) is relatively compact, then t
+
(v) = +∞, if γ
−
(v) is relatively com
pact, then t
−
(v) = −∞, whereas if γ(v) is relatively compact, then I
v
= R.
3. INVARIANT SETS 125
3 Invariant Sets
A subset M ⊂ D is called positively invariant with respect to the the ﬂow u
determined by f, whenever
γ
+
(v) ⊂ M, ∀v ∈ M,
i.e.,
γ
+
(M) ⊂ M.
We similarly call M ⊂ D negatively invariant provided
γ
−
(M) ⊂ M,
and invariant provided
γ(M) ⊂ M.
We note that a set M is invariant if and only if it is both positively and
negatively invariant.
We have the following proposition:
3 Proposition Let u be the ﬂow determined by f and let V ⊂ D. Then there
exists a smallest positively invariant subset M, V ⊂ M ⊂ D and there exists a
largest invariant set
˜
M,
˜
M ⊂ V. Also there exists a largest negatively invariant
subset M, V ⊃ M and there exists a smallest invariant set
˜
M,
˜
M ⊃ V. As a
consequence V contains a largest invariant subset and is contained in a smallest
invariant set.
As a corollary we obtain:
4 Corollary (i) If a set M is positively invariant with respect to the ﬂow u, then
so are M and intM.
(ii) A closed set M is positively invariant with respect to the ﬂow u if and only
if for every v ∈ ∂M there exists ǫ > 0 such that u([0, ǫ), v) ⊂ M.
(iii) A set M is positively invariant if and only if compM (the complement of
M) is negatively invariant.
(iv) If a set M is invariant, then so is ∂M. If ∂M is invariant, then so are M,
R ¸ M, and intM.
We now provide a geometric condition on ∂M which will guarantee the
invariance of a set M and, in particular will aid us in ﬁnding invariant sets.
We have the following theorem, which provides a relationship between the
vector ﬁeld f and the set M (usually called the subtangent condition) in order
that invariance holds.
5 Theorem Let M ⊂ D be a closed set. Then M is positively invariant with
respect to the ﬂow u determined by f if and only if for every v ∈ M
liminf
t→0+
dist(v +tf(v), M)
t
= 0. (4)
126
Proof. Let v ∈ M, then a Taylor expansion yields,
u(t, v) = v +tf(v) +o(t), t > 0.
Hence, if M is positively invariant,
dist(v +tf(v), M) ≤ [u(t, v) −v −tf(v)[ = o(t),
proving the necessity of (4).
Next, let v ∈ M, and assume (4) holds. Choose a compact neighborhood
B of v such that B ⊂ D and choose ǫ > 0 so that u([0, ǫ], v) ⊂ B. Let w(t) =
dist(u(t, v), M), then for each t ∈ [0, ǫ] there exists v
t
∈ M such that w(t) =
[u(t, v) −v
t
)[ and lim
t→0+
v
t
= v. It follows that for some constant L,
w(t +s) = [u(t +s, v) −v
t+s
[
≤ w(t) +[u(t +s, v) −u(t, v) −sf(u(t, v))[
+s[f(u(t, v)) −f(v
t
)[ +[v
t
+sf(v
t
) −v
t+s
[
≤ w(t) +sLw(t) + dist(v
t
+tf(v
t
), M),
because f satisﬁes a local Lipschitz condition. Hence
D
+
w(t) ≤ Lw(t), 0 ≤ t < ǫ, w(0) = 0,
or
D
+
_
e
−Lt
w(t)
_
≤ 0, 0 ≤ t < ǫ, w(0) = 0,
which implies w(t) ≤ 0, 0 ≤ t < ǫ, i.e. w(t) ≡ 0.
We remark here that condition (4) only needs to be checked for points v ∈
∂M since it obviously holds for interior points.
We next consider some special cases where the set M is given as a smooth
manifold. We have the following theorem.
6 Theorem Let φ ∈ C
1
(D, R) be a function which is such that every value v ∈
φ
−1
(0) is a regular value, i.e. ∇φ(v) ,= 0. Let M = φ
−1
(−∞, 0], then M is
positively invariant with respect to the ﬂow determined by f if and only if
∇φ(v) f(v) ≤ 0, ∀v ∈ ∂M = φ
−1
(0). (5)
We leave the proof as an exercise. We remark that the set M given in the
previous theorem will be negatively invariant provided the reverse inequality
holds and hence invariant if and only if
∇φ(v) f(v) = 0, ∀v ∈ ∂M = φ
−1
(0),
in which case φ(u(t, v)) ≡ 0, ∀v ∈ ∂M, i.e. φ is a ﬁrst integral for the diﬀerential
equation.
4. LIMIT SETS 127
4 Limit Sets
In this section we shall consider semiorbits and study their limit sets.
Let γ
+
(v), v ∈ D be the positive semiorbit associated with v ∈ D. We deﬁne
the set Γ
+
(v) as follows:
Γ
+
(v) = ¦w : ∃¦t
n
¦, t
n
րt
+
v
, u(t
n
, v) →w¦. (6)
The set Γ
+
(v) is the set of limit points of γ
+
(v) and is called the positive limit
set of v. (Note that if t
+
v
< ∞, then Γ
+
(v) ⊂ ∂D.) In a similar vein one deﬁnes
limit sets for negative semiorbits γ
−
(v), v ∈ D. Namely we deﬁne the set Γ
−
(v)
as follows:
Γ
−
(v) = ¦w : ∃¦t
n
¦, t
n
ցt
−
v
, u(t
n
, v) →w¦. (7)
The set Γ
−
(v) is the set of limit points of γ
−
(v) and is called the negative limit
set of v. (Again note that if t
−
v
> −∞, then Γ
−
(v) ⊂ ∂D.)
For all the results discussed below for positive limit sets there is an analagous
result for negative limit sets; we shall not state these results.
We have the following proposition:
7 Proposition (i) γ
+
(v) = γ
+
(v) ∪ Γ
+
(v).
(ii) Γ
+
(v) = ∩
w∈γ
+
(v)
γ
+
(w).
(iii) If γ
+
(v) is bounded, then Γ
+
(v) ,= ∅ and compact.
(iv) If Γ
+
(v) ,= ∅ and bounded, then
lim
t→t
+
v
dist
_
u(t, v), Γ
+
(v)
_
= 0.
(v) Γ
+
(v) ∩ D is an invariant set.
Proof. We leave most of the proof to the exercises and only discuss the veriﬁ
cation of the last part of the proposition. Thus let us assume that Γ
+
(v)∩D ,= ∅.
It then follows that t
+
v
= ∞. Let w ∈ Γ
+
(v) ∩ D. Then there exists a sequence
¦t
n
¦
∞
n=1
, t
n
ր∞, such that u(t
n
, v) →w. For each n ≥ 1, the function
u
n
(t) = u(t +t
n
, v)
is the unique solution of
u
′
= f(u), u(0) = u(t
n
, v),
and hence the maximal interval of existence of u
n
will contain the interval
[−t
n
, ∞). Since u(t
n
, v) → w, there will exist a subsequence of ¦u
n
(t)¦, which
we relabel as ¦u
n
(t)¦ converging to the solution, call it y, of
u
′
= f(u), u(0) = w.
128
We note that given any compact interval [a, b] the sequence ¦u
n
(t)¦ will be
deﬁned on [a, b] for n suﬃciently large and hence y will be deﬁned on [a, b]. Since
this interval is arbitrary it follows that y is deﬁned on (−∞, ∞). Furthermore
for any t
0
y(t
0
) = lim
n→∞
u
n
(t
0
) = lim
n→∞
u(t
0
+t
n
, v),
and hence y(t
0
) ∈ Γ
+
(v). Hence solutions through points of Γ
+
(v) are deﬁned
for all time and their orbit remains in Γ
+
(v), showing that Γ
+
(v) ∩ D is an
invariant set.
8 Theorem If γ
+
(v) is contained in a compact subset K ⊂ D, then Γ
+
(v) ,= ∅
is a compact connected set, i.e a continuum.
Proof. We already know (cf. Proposition 7) that Γ
+
(v) is a compact set. Thus
we need to show it is connected. Suppose it is not. Then there exist nonempty
disjoint compact sets M and N such that
Γ
+
(v) = M ∪ N.
Let δ = inf¦[v − w[ : v ∈ M, w ∈ N¦ > 0. Since M ⊂ Γ
+
(v) and N ⊂ Γ
+
(v),
there exist values of t arbitrarily large such that dist(u(t, v), M) <
δ
2
and values
of t arbitrarily large such that dist(u(t, v), N) <
δ
2
and hence there exists a
sequence ¦t
n
→ ∞¦ such that dist(u(t
n
, v), M) =
δ
2
. The sequence ¦u(t
n
, v)¦
must have a convergent subsequence and hence has a limit point which is in
neither M nor N, a contradiction.
4.1 LaSalle’s theorem
In this section we shall return again to invariant sets and consider Lyapunov
type functions and their use in determining invariant sets.
Thus let φ : D →R be a C
1
function. We shall use the following notation
φ
′
(v) := ∇φ(v) f(v).
9 Lemma Assume that φ
′
(v) ≤ 0, ∀v ∈ D. Then for all v ∈ D, φ is constant on
the set Γ
+
(v) ∩ D.
10 Theorem Let there exist a compact set K ⊂ D such that φ
′
(v) ≤ 0, ∀v ∈ K.
Let
˜
K = ¦v ∈ K : φ
′
(v) = 0¦
and let M be the largest invariant set contained in
˜
K. Then for all v ∈ D such
that γ
+
(v) ⊂ K
lim
t→∞
dist (u(t, v), M) = 0.
5. TWO DIMENSIONAL SYSTEMS 129
Proof. Let v ∈ D such that γ
+
(v) ⊂ K, then, using the previous lemma, we
have that φ is constant on Γ
+
(v), which is an invariant set and hence contained
in M.
11 Theorem (LaSalle’s Theorem) Assume that D = R
N
and let φ
′
(x) ≤ 0, ∀x ∈
R
N
. Furthermore suppose that φ is bounded below and that φ(x) → ∞ as
[x[ →∞. Let E = ¦v : φ
′
(v) = 0¦, then
lim
t→∞
dist (u(t, v), M) = 0,
for all v ∈ R
N
, where M is the largest invariant set contained in E.
As an example to illustrate Theorem 11 consider the oscillator
mx
′′
+hx
′
+kx = 0,
where m, h, k are positive constants. This equation may be written as
x
′
= y
y
′
= −
k
m
x −
h
m
y.
We choose
φ(x, y) =
1
2
_
my
2
+kx
2
_
and obtain that
φ
′
(x, y) = −hy
2
.
Hence E = ¦(x, y) : y = 0¦. The largest invariant set contained in E is the
origin, hence all solution orbits tend to the origin.
5 Two Dimensional Systems
In this section we analyze limit sets for two dimensional systems in somewhat
more detail and prove a classical theorem (the Theorem of Poincar´eBendixson)
about periodic orbits of such systems. Thus we shall assume throughout this
section that N = 2.
Let v ∈ D be a regular (i.e. not critical) point of f. We call a compact
straight line segment l ⊂ D through v a transversal through v, provided it
contains only regular points and if for all w ∈ l, f(w) is not parallel to the
direction of l.
We shall need the following observation.
12 Lemma Let v ∈ D be a regular point of f. Then there exists a transversal l
containing v in its relative interior. Also every orbit associated with f which
crosses l, crosses l in the same direction.
130
Proof. Let v be a regular point of f. Choose a neighborhood V of v consisting
of regular points only. Let η ∈ R
2
be any direction not parallel to f(v), i.e.
η f(v) ,= 0, (here is the cross product in R
3
). We may restrict V further
such that η f(w) ,= 0, ∀w ∈ V, and is bounded away from 0 on V. We then
may take l to be the intersection of the straight line through v with direction η
and V . The proof is completed by observing that
η f(w) = (0, 0, [η[[f(w)[ sin θ),
where θ is the angle between η and f(w).
13 Lemma Let v be an interior point of some transversal l. Then for every ǫ > 0
there exists a circular disc D
ǫ
with center at v such that for every w ∈ D
ǫ
,
u(t, w) will cross l in time t, [t[ < ǫ.
Proof. Let v ∈ int l and let
l = ¦z : z = v +sη, s
0
≤ s ≤ s
1
¦.
Let B be a disc centered at v containing only regular points of f. Let L(t, w) =
au
1
(t, w) +bu
2
(t, w) +c, u = (u
1
, u
2
), where u(t, w) is the solution with initial
condition w and au
1
+bu
2
+c = 0 is the equation of the straight line containing
l. Then L(0, v) = 0, and
∂L
∂t
(0, v) = (a, b) f(v) ,= 0. We hence may apply the
implicit function theorem to complete the proof.
14 Lemma Let l be a transversal and let Γ = ¦w = u(t, v) : a ≤ t ≤ b¦ be a
closed arc of an orbit u associated with f which has the property that u(t
1
, v) ,=
u(t
2
, v), a ≤ t
1
< t
2
≤ b. Then if Γ intersects l it does so at a ﬁnite number of
points whose order on Γ is the same as the order on l. If the orbit is periodic it
intersects l at most once.
The proof relies on the Jordan curve theorem and is left as an exercise. The
next lemma follows immediately from the previous two.
15 Lemma Let γ
+
(v) be a semiorbit which does not intersect itself and let w ∈
Γ
+
(v) be a regular point of f. Then any transversal containing w in its interior
contains no other points of Γ
+
(v) in its interior.
16 Lemma Let γ
+
(v) be a semiorbit which does not intersect itself and which is
contained in a compact set K ⊂ D and let all points in Γ
+
(v) be regular points
of f. Then Γ
+
(v) contains a periodic orbit.
Proof. Let w ∈ Γ
+
(v). it follows from Proposition 7 that Γ
+
(v) is an invariant
set and hence that γ
+
(w) ⊂ Γ
+
(v), and thus also Γ
+
(w) ⊂ Γ
+
(v). Let z ∈
Γ
+
(w), and let l be a transversal containing z in its relative interior. It follows
that the semiorbit γ
+
(w) must intersect l and by the above for an inﬁnite
number of values of t. On the other hand, the previous lemma implies that all
these point of intersection must be the same.
6. EXERCISES 131
It therefore follows from Lemma 16 that every point in Γ
+
(v) is a point on
some periodic orbit of a minimal positive period. On the other hand, it also
follows from earlier results that Γ
+
(v) is a compact connected set. Hence, if for
some w ∈ Γ
+
(v), γ(w) ,= Γ
+
(v), then Γ
+
(v) ¸γ(w) must be a relatively open set
with A = Γ
+
(v) ¸ γ(w) ∩ γ(w) ,= ∅. One now easily obtains a contradiction by
examining transversals through points of A. One hence concludes that in fact
under the hypotheses of Lemma 16, the limit set Γ
+
(v) is a periodic orbit. We
summarize the above in the following theorem.
17 Theorem (Poincar´eBendixson) Let γ
+
(v) be a semiorbit which does not
intersect itself and which is contained in a compact set K ⊂ D and let all points
in Γ
+
(v) be regular points of f. Then Γ
+
(v) is the orbit of a periodic solution
u
T
with smallest positive period T.
18 Theorem Let Γ be a periodic orbit of (1) which together with its interior is
contained in a compact set K ⊂ D. Then there exists at least one singular point
of f in the interior of D.
Proof. Let
Ω = intΓ.
Then f is continuous on Ω and does not vanish on Γ = ∂Ω. Let us assume
that f has no stationary points in Ω. Then for each w ∈ Ω, Γ
+
(w) is a periodic
orbit. We partially order the collection ¦Γ
α
¦
α∈I
, where I is an index set, of all
periodic orbits which are contained in Ω, by saying that
Γ
α
≤ Γ
β
⇔intΓ
α
⊂ intΓ
β
.
One now employs the Hausdorﬀ minimum principle together with Theorem 17
to arrive at a contradiction.
6 Exercises
1. Prove Lemma 1. Also provide conditions in order that the mapping u be
smooth, say of class C
1
.
2. Let u be the ﬂow determined by f (see Lemma 1). Show that if I
u0
=
(t
−
u0
, t
+
u0
), then −t
−
u0
: D →[0, ∞] and t
+
u0
: D →[0, ∞] are lower semicon
tinuous functions of u
0
.
3. Prove Proposition 2.
4. Prove Proposition 3 and Corollary 4.
5. Prove Theorem 6.
132
6. Consider the three dimensional system of equations (the Lorenz system)
x
′
= −σx +σy
y
′
= rx −y −xz
z
′
= −bz +xy,
where σ, r, b are positive parameters.
Find a family of ellipsoids which are positively invariant sets for the ﬂow
determined by the system.
7. Prove Theorem 6.
8. Extend Theorem 6 to the case where
M = ∩
m
i=1
v
−1
i
(−∞, 0].
Consider the special case where each φ
i
is aﬃne linear, i.e M is a paral
lelepiped.
9. Prove Proposition 7.
10. Prove Lemma 9.
11. Prove Theorem 17.
12. Prove Theorem 18.
13. Complete the proof of Lemma 12.
14. Prove Lemma 14.
15. Let u be a solution whose interval of existence is R which is not periodic
and let γ
+
(u(0)) and γ
−
(u(0)) be contained in a compact set K ⊂ D.
Prove that Γ
+
(u(0)) and Γ
−
(u(0)) are distinct periodic orbits provided
they contain regular points only.
16. Prove that rest points and periodic orbits are invariant sets.
17. Consider the system
x
′
= y −x
3
+ µx
y
′
= −x,
where µ is a real parameter. Show that periodic orbits exist for all µ > 0
and that these orbits collapse to the origin as µ → 0. Also show that no
periodic orbits exist for µ ≤ 0.
Chapter X
Hopf Bifurcation
1 Introduction
This chapter is devoted to a version of the classical Hopf bifurcation theo
rem which establishes the existence of nontrivial periodic orbits of autonomous
systems of diﬀerential equations which depend upon a parameter and for which
the stability properties of the trivial solution changes as the parameter is var
ied. The proof we give is base on the method of LyapunovSchmidt presented
in Chapter II.
2 A Hopf Bifurcation Theorem
Let
f : R
n
R →R
n
,
be a C
2
mapping which is such that
f(0, α) = 0, all α ∈ R.
We consider the system of ordinary diﬀerential equations depending on a pa
rameter α
du
dt
+f(u, α) = 0, (1)
and prove a theorem about the existence of nontrivial periodic solutions of this
system. Results of the type proved here are referred to as Hopf bifurcation
theorems.
We establish the following theorem.
1 Theorem Assume that f satisﬁes the following conditions:
1. For some given value of α = α
0
, i =
√
−1 and −i are eigenvalues of
f
u
(0, α
0
) and ±ni, n = 0, 2, 3, , are not eigenvalues of f
u
(0, α
0
);
133
134
2. in a neighborhood of α
0
there is a curve of eigenvalues and eigenvectors
f
u
(0, α)a(α) = β(α)a(α)
a(α
0
) ,= 0, β(α
0
) = i, Re
dβ
dα
[
α0
,= 0.
(2)
Then there exist positive numbers ǫ and η and a C
1
function
(u, ρ, α) : (−η, η) →C
1
2π
R R,
where C
1
2π
is the space of 2π periodic C
1
R
n
− valued functions, such that
(u(s), ρ(s), α(s)) solves the equation
du
dτ
+ρf(u, α) = 0, (3)
nontrivially, i.e. u(s) ,= 0, s ,= 0 and
ρ(0) = 1, α(0) = α
0
, u(0) = 0. (4)
Furthermore, if (u
1
, α
1
) is a nontrivial solution of (1) of period 2πρ
1
, with
[ρ
1
−1[ < ǫ, [α
1
−α
0
[ < ǫ, [u
1
(t)[ < ǫ, t ∈ [0, 2πρ
1
], then there exists s ∈ (−η, η)
such that ρ
1
= ρ(s), α
1
= α(s) and u
1
(ρ
1
t) = u(s)(τ +θ), τ = ρ
1
t ∈ [0, 2π], θ ∈
[0, 2π).
Proof. We note that u(t) will be a solution of (1) of period 2πρ, whenever
u(τ), τ = ρt is a solution of period 2π of (3). We thus let X = C
1
2π
and Y =
C
2π
be Banach spaces of C
1
, respectively, continuous 2π− periodic functions
endowed with the usual norms and deﬁne an operator
F : X R R →Y
(u, ρ, α) →u
′
+ρf(u, α),
′
=
d
dτ
.
(5)
Then F belongs to class C
2
and we seek nontrivial solutions of the equation
F(u, ρ, α) = 0, (6)
with values of ρ close to 1, α close to α
0
, and u ,= 0.
We note that
F(0, ρ, α) = 0, for all ρ ∈ R, α ∈ R.
Thus the claim is that the value (1, α
0
) of the two dimensional parameter (ρ, α)
is a bifurcation value. Theorem II.1 tells us that
u
′
+f
u
(0, α
0
)u
cannot be a linear homeomorphism of X onto Y. This is guaranteed by the
assumptions, since the functions
φ
0
= Re(e
iτ
a(α
0
)), φ
1
= Im(e
iτ
a(α
0
))
are 2π− periodic solutions of
u
′
+f
u
(0, α
0
)u = 0, (7)
2. A HOPF BIFURCATION THEOREM 135
and they span the the kernel of F
u
(0, 1, α
0
),
kerF
u
(0, 1, α
0
) = ¦φ
0
, φ
1
= −φ
′
0
¦.
It follows from the theory of linear diﬀerential equations that the image, imF
u
(0, 1, α
0
),
is closed in Y and that
imF
u
(0, 1, α
0
) = ¦f ∈ Y : ¸f, ψ
i
) = 0, i = 0, 1¦,
where ¸, ) denotes the L
2
inner product and ¦ψ
0
, ψ
1
¦ forms a basis for ker¦−u
′
+
f
T
u
(0, α
0
)u¦, the diﬀerential equation adjoint operator of u
′
+f
u
(0, α
0
)u. In fact
ψ
1
= ψ
′
0
and ¸φ
i
, ψ
j
) = δ
ij
, the Kronecker delta. Thus F
u
(0, 1, α
0
) is a linear
Fredholm mapping from X to Y having a two dimensional kernel as well as a
two dimensional cokernel. We now write, as in the beginning of Chapter II,
X = V ⊕W
Y = Z ⊕T.
We let U be a neighborhood of (0, 1, α
0
, 0) in V R R R and deﬁne G
on U as follows
G(v, ρ, α, s) =
_
1
s
F(s(φ
0
+v), ρ, α), s ,= 0
F
u
(0, ρ, α)(φ
0
+v), s = 0.
We now want to solve the equation
G(v, ρ, α, s) = 0,
for v, ρ, α in terms of s in a neighborhood of 0 ∈ R. We note that G is C
1
and
G(v, ρ, α, 0) = (φ
0
+v)
′
+ ρf
u
(0, α)(φ
0
+v).
Hence
G(0, ρ, α, 0) = φ
′
0
+ρf
u
(0, α)φ
0
.
Thus, in order to be able to apply the implicit function theorem, we need to
diﬀerentiate the map
(v, ρ, α) →G(v, ρ, α, s)
evaluate the result at (0, 1, α
0
, 0) and show that this derivative is a linear home
omorphism of V R R onto Y.
Computing the Taylor expansion, we obtain
G(v, ρ, α, s) = G(0, 1, α
0
, s) +G
ρ
(0, 1, α
0
, s)(ρ −1)
G
α
(0, 1, α
0
, s)(α −α
0
) +G
v
(0, 1, α
0
, s)v + ,
(8)
and evaluating at s = 0 we get
G
v,ρ,α
(0, 1, α
0
, 0)(v, ρ −1, α −α
0
) = f
v
(0, α
0
)φ
0
(ρ −1)
+f
vα
(0, α
0
)φ
0
(α −α
0
)
+(v
′
+f
v
(0, α
0
)v).
(9)
136
Note that the mapping
v →v
′
+f
v
(0, α
0
)v
is a linear homeomorphism of V onto T. Thus we need to show that the map
(ρ, α) →f
v
(0, α
0
)φ
0
(ρ −1) +f
vα
(0, α
0
)φ
0
(α −α
0
)
only belongs to T if ρ = 1 and α = α
0
and for all ψ ∈ Z there exists a unique
(ρ, α) such that
f
v
(0, α
0
)φ
0
(ρ −1) +f
vα
(0, α
0
)φ
0
(α −α
0
) = ψ.
By the characterization of T, we have that
f
v
(0, α
0
)φ
0
(ρ −1) +f
vα
(0, α
0
)φ
0
(α −α
0
) ∈ T
if and only if
< f
v
(0, α
0
)φ
0
, ψ
i
> (ρ −1)+ < f
vα
(0, α
0
)φ
0
, ψ
i
> (α −α
0
) = 0,
i = 1, 2.
(10)
Since
f
vα
(0, α
0
)φ
0
= −φ
′
0
= φ
1
,
we may write equation (10) as two equations in the two unknowns ρ − 1 and
α −α
0
,
< f
vα
(0, α
0
)φ
0
, ψ
0
> (α −α
0
) = 0
(ρ −1)+ < f
vα
(0, α
0
)φ
0
, ψ
1
> (α −α
0
) = 0,
(11)
which has only the trivial solution if and only if
< f
vα
(0, α
0
)φ
0
, ψ
0
>, = 0.
Computing this latter expression, one obtains
< f
vα
(0, α
0
)φ
0
, ψ
0
>= Reβ
′
(0),
which by hypothesis is not zero. The uniqueness assertion we leave as an exer
cise.
For much further discussion of Hopf bifurcation we refer to [19].
The following example of the classical Van der Pol oscillator from nonlinear
electrical circuit theory (see [19]) will serve to illustrate the applicability of the
theorem.
2 Example Consider the nonlinear oscillator
x
′′
+x −α(1 −x
2
)x
′
= 0. (12)
This equation has for for certain small values of α nontrivial periodic solu
tions with periods close to 2π.
2. A HOPF BIFURCATION THEOREM 137
We ﬁrst transform (12) into a system by setting
u =
_
u
1
u
2
_
=
_
x
x
′
_
and obtain
u
′
+
_
0 −1
1 −α
_
+
_
0
u
2
1
u
2
_
=
_
0
0
_
(13)
We hence obtain that
f(u, α) =
_
0 −1
1 −α
_
+
_
0
u
2
1
u
2
_
and
f
u
(0, α) =
_
0 −1
1 −α
_
,
whose eigenvalues satisfy
β(α +β) + 1 = 0.
Letting α
0
= 0, we get β(0) = ±i, and computing
dβ
dα
= β
′
we obtain 2ββ
′
+β
′
α+
β = 0, or β
′
=
−β
α+2β
= −
1
2
, for α = 0. Thus by Theorem II.1 there exists η > 0
and continuous functions α(s), ρ(s), s ∈ (−η, η) such that α(0) = 0, ρ(0) = 1
and (12) has for s ,= 0 a nontrivial solution x(s) with period 2πρ(s). This
solution is unique up to phase shift.
138
Chapter XI
SturmLiouville Boundary
Value Problems
1 Introduction
In this chapter we shall study a very classical problem in the theory of
ordinary diﬀerential equations, namely linear second order diﬀerential equations
which are parameter dependent and are subject to boundary conditions. While
the existence of eigenvalues (parameter values for which nontrivial solutions
exist) and eigenfunctions (corresponding nontrivial solutions) follows easily from
the abstract Riesz spectral theory for compact linear operators, it is instructive
to deduce the same conclusions using some of the results we have developed up
to now for ordinary diﬀerential equations. While the theory presented below is
for some rather speciﬁc cases, much more general problems and various other
cases may be considered and similar theorems may be established. We refer to
the books [5], [6] and [21] where the subject is studied in some more detail.
2 Linear Boundary Value Problems
Let I = [a, b] be a compact interval and let p, q, r ∈ C(I, R), with p, r positive
on I. Consider the linear diﬀerential equation
(p(t)x
′
)
′
+ (λr(t) +q(t))x = 0, t ∈ I, (1)
where λ is a complex parameter.
We seek parameter values (eigenvalues) for which (1) has nontrivial solutions
(eigensolutions or eigenfunctions) when it is subject to the set of boundary
conditions
x(a) cos α −p(a)x
′
(a) sin α = 0
x(b) cos β −p(b)x
′
(b) sin β = 0,
(2)
where α and β are given constants and (without loss in generality, 0 ≤ α <
π, 0 < β ≤ π). Such a boundary value problem is called a SturmLiouville
boundary value problem.
139
140
We note that (2) is equivalent to the requirement
c
1
x(a) +c
2
x
′
(a) = 0, [c
1
[ +[c
2
[ , = 0
c
3
x(b) +c
4
x
′
(b) = 0, [c
3
[ +[c
4
[ , = 0.
(3)
We have the following lemma.
1 Lemma Every eigenvalue of equation (1) subject to the boundary conditions
(2) is real.
Proof. Let the diﬀerential operator L be deﬁned by
L(x) = (px
′
)
′
+qx.
Then, if λ is an eigenvalue
L(x) = −λrx,
for some nontrivial x which satisﬁes the boundary conditions. Hence also
L(¯ x) = −λrx = −
¯
λr¯ x.
Therefore
¯ xL(x) −xL(¯ x) = −(λ −
¯
λ)rx¯ x.
Hence
_
b
a
(¯ xL(x) −xL(¯ x)) dt = −(λ −
¯
λ)
_
b
a
rx¯ xdt.
Integrating the latter expression and using the fact that both x and ¯ x satisfy
the boundary conditions we obtain the value 0 for this expression and hence
λ =
¯
λ.
We next let u(t, λ) = u(t) be the solution of (1) which satisﬁes the (initial)
conditions
u(a) = sin α, p(a)u
′
(a) = cos α,
then u ,= 0 and satisﬁes the ﬁrst set of boundary conditions. We introduce the
following transformation (Pr¨ ufer transformation)
ρ =
_
u
2
+p
2
(u
′
)
2
, φ = arctan
u
pu
′
.
Then ρ and φ are solutions of the diﬀerential equations
ρ
′
= −
_
(λr +q) −
1
p
_
ρ sin φcos φ (4)
and
φ
′
=
1
p
cos
2
φ + (λr +q) sin
2
φ. (5)
2. LINEAR BOUNDARY VALUE PROBLEMS 141
Further φ(a) = α. (Note that the second equation depends upon φ only, hence,
once φ is known, ρ may be determined by integrating a linear equation and
hence u is determined.
We have the following lemma describing the
dependence of φ upon λ.
2 Lemma Let φ be the solution of (5) such that φ(a) = α. Then φ satisﬁes the
following conditions:
1. φ(b, λ) is a continuous strictly increasing function of λ;
2. lim
λ→∞
φ(b, λ) = ∞;
3. lim
λ→−∞
φ(b, λ) = 0.
Proof. The ﬁrst part follows immediately from the discussion in Sections V.5
and V.6. To prove the other parts of the lemma, we ﬁnd it convenient to make
the change of independent variable
s =
_
t
a
dτ
p(τ)
,
which transforms equation (1) to
x
′′
+p(λr +q)x = 0,
′
=
d
ds
. (6)
We now apply the Pr¨ ufer transformation to (6) and use the comparison theorems
in Section V.6 to deduce the remaining parts of the lemma.
Using the above lemma we obtain the existence of eigenvalues, namely we
have the following theorem.
3 Theorem The boundary value problem (1)(2) has an unbounded inﬁnite se
quence of eigenvalues
λ
0
< λ
1
< λ
2
<
with
lim
n→∞
λ
n
= ∞.
The eigenspace associated with each eigenvalue is one dimensional and the eigen
functions associated with λ
k
have precisely k simple zeros in (a, b).
Proof. The equation
β +kπ = φ(b, λ)
has a unique solution λ
k
, for k = 0, 1, . This set ¦λ
k
¦
∞
k=0
has the desired
properties.
We also have the following lemma.
142
4 Lemma Let u
i
, i = j, k, j ,= k be eigenfunctions of the boundary value prob
lem (1)(2) corresponding to the eigenvalues λ
j
and λ
k
. Then u
j
and u
k
are
orthogonal with respect to the weight function r, i.e.
¸u
j
, u
k
) =
_
b
a
ru
j
u
k
= 0. (7)
In what is to follow we denote by ¦u
i
¦
∞
i=0
the set of eigenfunctions whose
existence is guaranteed by Theorem 3 with u
i
an eigenfunction corresponding
to λ
i
, i = 0, 1, which has been normalized so that
_
b
a
ru
2
i
= 1. (8)
We also consider the nonhomogeneous boundary value problem
(p(t)x
′
)
′
+ (λr(t) +q(t))x = rh, t ∈ I, (9)
where h ∈ L
2
(a, b) is a given function, the equation being subject to the bound
ary conditions (2) and solutions being interpreted in the Carath´eodory sense.
We have the following result.
5 Lemma For λ = λ
k
equation (9) has a solution subject to the boundary con
ditions (2) if and only if
_
b
a
ru
k
h = 0.
If this is the case, and w is a particular solution of (9)(2), then any other
solution has the form w +cu
k
, where c is an arbitrary constant.
Proof. Let v be a solution of (p(t)x
′
)
′
+(λ
k
r(t) +q(t))x = 0, which is linearly
independent of u
k
, then
(u
k
v
′
−u
′
k
v) =
c
p
,
where c is a nonzero constant. One veriﬁes that
w(t) =
1
c
_
v(t)
_
t
a
ru
k
h +u
k
_
b
t
rvh
_
is a solution of (9)(2) (for λ = λ
k
), whenever
_
b
a
ru
k
h = 0 holds.
3 Completeness of Eigenfunctions
We note that it follows from Lemma 5 that (9)(2) has a solution for every
λ
k
, k = 0, 1, 2, if and only if
_
b
a
ru
k
h = 0, for k = 0, 1, 2, . Hence, since
¦u
i
¦
∞
i=0
forms an orthonormal system for the Hilbert space L
2
r
(a, b) (i.e. L
2
(a, b)
with weight function r deﬁning the inner product), ¦u
i
¦
∞
i=0
will be a complete
3. COMPLETENESS OF EIGENFUNCTIONS 143
orthonormal system, once we can show that
_
b
a
u
k
h = 0, for k = 0, 1, 2, ,
implies that h = 0 (see [37]). The aim of this section is to prove completeness.
The following lemma will be needed in this discussion.
6 Lemma If λ ,= λ
k
, k = 0, 1, (9) has a solution subject to the boundary
conditions (2) for every h ∈ L
2
(a, b).
Proof. For λ ,= λ
k
, k = 0, 1, we let u be a nontrivial solution of (1) which
satisﬁes the ﬁrst boundary condition of (2) and let v be a nontrivial solution of
(1) which satisﬁes the second boundary condition of (2). Then
uv
′
−u
′
v =
c
p
with c a nonzero constant. Deﬁne the Green’s function
G(t, s) =
1
c
_
v(t)u(s), a ≤ s ≤ t
v(s)u(t), t ≤ s ≤ b.
(10)
Then
w(t) =
_
b
a
G(t, s)r(s)h(s)ds
is the unique solution of (9)  (2).
We have the following corollary.
7 Corollary Lemma 6 deﬁnes a continuous mapping
G : L
2
(a, b) →C
1
[a, b], (11)
by
h →G(h) = w.
Further
¸Gh, w) = ¸h, Gw).
Proof. We merely need to examine the deﬁnition of G(t, s) as given by equation
(10).
Let us now let
S = ¦w ∈ L
2
(a, b) : ¸u
i
, h) = 0, i = 0, 1, 2, ¦. (12)
Using the deﬁnition of G we obtain the lemma.
8 Lemma G : S →S.
We note that S is a linear manifold in L
2
(a, b) which is weakly closed, i.e. if
¦x
n
¦ ⊂ S is a sequence such that
¸x
n
, h) →¸x, h), ∀h ∈ L
2
(a, b),
then x ∈ S.
144
9 Lemma If S ,= ¦0¦, then there exists x ∈ S such that
¸G(x), x) , = 0.
Proof. If ¸G(x), x) = 0 for all x ∈ S, then, since S is a linear manifold, we
have for all x, y ∈ S and α ∈ R
0 = ¸G(x +αy), x +αy)
= 2α¸G(y), x),
in particular, choosing x = G(y) we obtain a contradiction, since for y ,= 0
G(y) ,= 0.
10 Lemma If S ,= ¦0¦, then there exists x ∈ S¸¦0¦ and µ ,= 0 such that
G(x) = µx.
Proof. Since there exists x ∈ S such that ¸G(x), x) , = 0 we set
µ =
_
inf¦¸G(x), x) : x ∈ S, x = 1, if ¸G(x), x) ≤ 0, ∀x ∈ S¦
sup¦¸G(x), x) : x ∈ S, x = 1, if ¸G(x), x) > 0, for some x ∈ S¦.
We easily see that there exists x
0
∈ S, x
0
 = 1 such that ¸G(x
0
), x
0
) = µ ,= 0.
If S is one dimensional, then G(x
0
) = µx
0
. If S has dimension greater than 1,
then there exists 0 ,= y ∈ S such that ¸y, x
0
) = 0. Letting z =
x0+ǫy
√
1+ǫ
2
we ﬁnd
that ¸G(z), z) has an extremum at ǫ = 0 and thus obtain that ¸G(x
0
), y) = 0,
for any y ∈ S with ¸y, x
0
) = 0. Hence since ¸G(x
0
) −µx
0
, x
0
) = 0 it follows that
¸G(x
0
), G(x
0
) − µx
0
) = 0 and thus ¸G(x
0
) − µx
0
, G(x
0
) − µx
0
) = 0, proving
that µ is an eigenvalue.
Combining the above results we obtain the following completeness theorem.
11 Theorem The set of eigenfunctions ¦u
i
¦
∞
i=0
forms a complete orthonormal sys
tem for the Hilbert space L
2
r
(a, b).
Proof. Following the above reasoning, we merely need to show that S =
¦0¦. If this is not the case, we obtain, by Lemma 10, a nonzero element h ∈
S and a nonzero number µ such that G(h) = µh. On the other hand w =
G(h) satisﬁes the boundary conditions (2) and solves (9); hence h satisﬁes the
boundary conditions and solves the equation
(p(t)h
′
)
′
+ (λr(t) +q(t))h =
r
µ
h, t ∈ I, (13)
i.e. λ −
1
µ
= λ
j
for some j. Hence h = cu
j
for some nonzero constant c,
contradicting that h ∈ S.
4. EXERCISES 145
4 Exercises
1. Find the set of eigenvalues and eigenfunctions for the boundary value
problem
x
′′
+λx = 0
x(0) = 0 = x
′
(1).
2. Supply the details for the proof of Lemma 2.
3. Prove Lemma 4.
4. Prove that the Green’s function given by (10) is continuous on the square
[a, b]
2
and that
∂G(t,s)
∂t
is continuous for t ,= s. Discuss the behavior of this
derivative as t →s.
5. Provide the details of the proof of Corollary 7. Also prove that G :
L
2
(a, b) →L
2
(a, b) is a compact mapping.
6. Let G(t, s) be deﬁned by equation (10). Show that
G(t, s) =
∞
i=0
u
i
(t)u
i
(s)
λ −λ
i
,
where the convergence is in the L
2
norm.
7. Replace the boundary conditions (2) by the periodic boundary conditions
x(a) = x(b), x
′
(a) = x
′
(b).
Prove that the existence and completeness part of the above theory may be
established provided the functions satisfy p(a) = p(b), q(a) = q(b), r(a) =
r(b).
8. Apply the previous exercise to ﬁnd the eigenvalues and eigenfunctions for
the boundary value problem
x
′′
+λx = 0,
x(0) = x(2π)
x
′
(0) = x
′
(2π).
9. Let the diﬀerential operator L be given by
L(x) = (tx
′
)
′
+
m
2
t
x, 0 < t < 1.
and consider the eigenvalue problem
L(x) = −λtx.
146
In this case the hypotheses imposed earlier are not applicable and other
types of boundary conditions than those given by (3) must be sought in
order that a development parallel to that given in Section 2 may be made.
Establish such a theory for this singular problem. Extend this to more
general singular problems.
Bibliography
[1] R. Adams, Sobolev Spaces, Academic Press, New York, 1975.
[2] E. Allgower and K. Georg, Numerical Continuation Methods, An In
troduction, SpringerVerlag, New York, 1990.
[3] J. Bebernes and D. Eberly, Mathematical Problems from Combustion
Theory, Springer Verlag, New York, 1989.
[4] G. Birkhoff and G. Rota, Ordinary Diﬀerential Equations, Blaisdell,
Waltham, 1969.
[5] E. Coddington and N. Levinson, Theory of Ordinary Diﬀerential
Equations, McGraw Hill, New York, 1955.
[6] R. Cole, Theory of Ordinary Diﬀerential Equations, AppletonCentury
Crofts, New York, 1968.
[7] W. Coppel, Stability and Asymptotic Behavior of Diﬀerential Equations,
Heath, Boston, 1965.
[8] M. Crandall and P. Rabinowitz, Nonlinear SturmLiouville problems
and topological degree, J. Math. Mech., 19 (1970), pp. 1083–1102.
[9] J. Cronin, Fixed Points and Topological Degree in Nonlinear Analysis,
Math. Surveys No. 11, Amer. Math. Soc., Providence, 1964.
[10] E. N. Dancer, Global structure of the solutions on nonlinear real analytic
eigenvalue problems, Proc. London Math. Soc, 27 (1973), pp. 747–765.
[11] , On the structure of solutions of nonlinear eigenvalue problems, In
diana Math. J., 23 (1974), pp. 1069–1076.
[12] , Solution branches for mappings in cones, and applications, J. Aus
tral. Math. Soc., 11 (1974), pp. 131–143.
[13] H. Dang and K. Schmitt, Existence and uniqueness theorems for non
linear boundary value problems, Rocky Mtn. J. Math., 24 (1994), pp. 77–91.
[14] J. Dieudonn´e, Foundations of Modern Analysis, Academic Press, New
York, 1964.
147
148 BIBLIOGRAPHY
[15] P. Dr´ abek, Solvability and Bifurcations of Nonlinear Equations, Longman
Scientiﬁc and Technical Series, Essex, 1992.
[16] J. Dugundji, Topology, Allyn and Bacon, Boston, 1966.
[17] N. Dunford and J. Schwartz, Linear Operators: Part I, Wiley, New
York, 1964.
[18] D. Gilbarg and N. Trudinger, Elliptic Partial Diﬀerential Equations
of Second Order, Sringer, Berlin, 1983.
[19] J. Guckenheimer and P. Holmes, Nonlinear Oscillations, Dynamical
Systems and Bifurcations of Vector Fields, Springer Verlag, New York,,
1983.
[20] W. Hahn, Stability of Motion, Springer, Berlin, 1967.
[21] P. Hartman, Ordinary Diﬀerential Equations, Birkh¨auser, Boston, 1982.
[22] E. Heinz, An elementary analytic theory of the degree of mapping in n
dimensional Euclidean space, J. Math. Mech., 8 (1959), pp. 231–247.
[23] E. Kamke, Zur Theorie der Systeme gew¨ ohnlicher Diﬀerentialgleichungen
II, Acta. Math, 58 (1932), pp. 57–85.
[24] M. A. Krasnosels’kii, Positive Solutions of Operator Equations, Noord
hoﬀ, Groningen, 1964.
[25] M. A. Krasnosel’skii, A. I. Perov, A. I. Povolotskii, and P. P.
Zabreiko, Plane Vector Fields, Academic Press, New York, 1966.
[26] M. Krein and M. Rutman, Linear operators leaving invariant a cone in
a Banach space, AMS Translations, 10 (1962), pp. 199–325.
[27] E. Kreyszig, Introductory Functional Analysis, Wiley, New York, 1978.
[28] N. G. Lloyd, Degree theory, Cambridge University Press, Cambridge,
1978.
[29] M. Nagumo, A theory of degree of mapping based on inﬁnitesimal analysis,
Amer. J. Math., 73 (1951), pp. 485–496.
[30] R. D. Nussbaum, A global bifurcation theorem with applications to func
tional diﬀerential equations, J. Functional Analysis, 19 (1975), pp. 319–338.
[31] C. Olech, A simple proof of a result of Z. Opial, Ann. Polon. Math., 8
(1960), pp. 61–63.
[32] H. Peitgen, D. Saupe, and K. Schmitt, Nonlinear eigenvalue prob
lems versus their ﬁnite diﬀerence approximations: Numerically irrelevant
solutions, J. reine angew. Mathematik, 322 (1981), pp. 74–117.
BIBLIOGRAPHY 149
[33] H. O. Peitgen and K. Schmitt, Global analysis of twoparameter elliptic
eigenvalue problems, Trans. Amer. Math. Soc., 283 (1984), pp. 57–95.
[34] P. Rabinowitz, Some global results for nonlinear eigenvalue problems, J.
Func. Anal., 7 (1971), pp. 487–513.
[35] , On bifurcation from inﬁnity, J. Diﬀ. Equa., 14 (1973), pp. 462–475.
[36] H. L. Royden, Real Analysis, 3rd ed., Macmillan Publishing Co., New
York, 1988.
[37] W. Rudin, Real and Complex Analysis, McGraw Hill, New York, 1966.
[38] M. Schechter, Principles of Functional Analysis, Academic Press, New
York, 1971.
[39] K. Schmitt and Z. Wang, On bifurcation from inﬁnity for potential
operators, Diﬀerential and Integral equations, 4 (1991), pp. 933–943.
[40] J. Schwartz, Nonlinear Functional Analysis, Gordon and Breach, New
York, 1969.
[41] R. E. Showalter, Hilbert Space Methods for Partial Diﬀerential Equa
tions, Pitman, London, 1977.
[42] H. Siegberg, Some historical remarks concerning degree theory, Amer.
Math. Monthly, 88 (1981), pp. 125–139.
[43] A. Taylor, Introduction to Functional Analysis, Wiley and Sons, New
York, 1972.
[44] G. Whyburn, Analytic Topolgy, Amer. Math. Soc., Providence, 1942.
150 BIBLIOGRAPHY
ii
···
Copyright c 1998 by K. Schmitt and R. Thompson
iii Preface The subject of Diﬀerential Equations is a well established part of mathematics and its systematic development goes back to the early days of the development of Calculus. Many recent advances in mathematics, paralleled by a renewed and ﬂourishing interaction between mathematics, the sciences, and engineering, have again shown that many phenomena in the applied sciences, modelled by diﬀerential equations will yield some mathematical explanation of these phenomena (at least in some approximate sense). The intent of this set of notes is to present several of the important existence theorems for solutions of various types of problems associated with diﬀerential equations and provide qualitative and quantitative descriptions of solutions. At the same time, we develop methods of analysis which may be applied to carry out the above and which have applications in many other areas of mathematics, as well. As methods and theories are developed, we shall also pay particular attention to illustrate how these ﬁndings may be used and shall throughout consider examples from areas where the theory may be applied. As diﬀerential equations are equations which involve functions and their derivatives as unknowns, we shall adopt throughout the view that diﬀerential equations are equations in spaces of functions. We therefore shall, as we progress, develop existence theories for equations deﬁned in various types of function spaces, which usually will be function spaces which are in some sense natural for the given problem.
iv .
. 27 3 Bifurcation at a Simple Eigenvalue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . II Ordinary Diﬀerential Equations 63 65 Chapter V. . . 5 Inverse Function Theorems . . . . . . . . . . . . . . . . . . . . . . . 2 Deﬁnition . . . . . . . . . . . . . . . . . . . . . . . . 1 3 3 3 8 11 20 22 25 Chapter I. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Diﬀerentiability. . . . . 4 Completely Continuous Perturbations 5 Exercises . . . . . . . . . . . . . . 27 2 Splitting Equations . . . . . . . . . . . . . . . . . . . . . . . . . Degree Theory 1 Introduction . . . 2 Continuation Principle . . . . . . . . . . . . . . . . . . . . . . . . Theorem . . . . . . . . . . . 2 Banach Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Taylor’s Theorem 4 Some Special Mappings . . Existence and Uniqueness Theorems v . . . . . . . . . . . . . . . . . 5 Global Bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 The Dugundji Extension Theorem 7 Exercises . 30 Chapter III. . . . . . Global Solution Theorems 1 Introduction . . . . . . . . . . . . . . . . The Method of LyapunovSchmidt 27 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 3 A Globalization of the Implicit Function 4 The Theorem of KreinRutman . . . . . . . . . . . . . . . . . . Chapter II. . . . . . . . . . . . . . . . . . . . . . . . . Analysis In Banach Spaces 1 Introduction . . . . . . . . . . . 6 Exercises . . . . . . . . . . 33 33 33 38 42 47 49 49 49 52 54 57 61 Chapter IV. . . . . . . . . . .Table of Contents I Nonlinear Analysis . . . . . . 3 Properties of the Brouwer Degree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . TABLE OF CONTENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Exercises . . . . . . . . . . . . . . . . Dependence upon Initial Conditions Diﬀerential Inequalities . . . . Uniqueness Theorems . . . . . . . . . . . . 2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . Invariant Sets 1 Introduction . . . . . . . . . 3 Stability of Linear Equations . . . . . . 3 Perturbations of Nonresonant Equations 4 Resonant Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 Limit Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Linear Ordinary Diﬀerential 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Orbits and Flows . . . . . o The CauchyPeano Theorem . . . . . . . . 3 Invariant Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chapter VIII. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 Exercises . . . . . . . . . . . . . . Chapter IX. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 2 A Hopf Bifurcation Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . Periodic Solutions 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 Stability of Nonlinear Equations 5 Lyapunov Stability . . The PicardLindel¨f Theorem . . . . . . . . . . . . . . . . Chapter X. . . . . . . . 2 Preliminaries . . . . . . . . . . . . . 2 Stability Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 66 67 69 72 74 78 79 81 81 81 83 85 88 91 91 91 92 94 100 103 103 103 105 108 110 118 123 123 124 125 127 129 131 Chapter VI. 4 Floquet Theory . . . .vi 1 2 3 4 5 6 7 8 Introduction . . . . . Extension Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Constant Coeﬃcient Systems . . . . . . . . . . . . . . Hopf Bifurcation 133 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chapter VII. . . . . . . . . Equations . . . . . . . . . . . . . . . 5 Two Dimensional Systems 6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 . . . . . Stability Theory 1 Introduction . . . . . . . . . . . . . . . . . Exercises . . . .
. . . . . . . . . . . . . . . . . . . SturmLiouville Boundary Value Problems 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 2 Linear Boundary Value Problems . . 3 Completeness of Eigenfunctions . . . . . . . . . . . . .TABLE OF CONTENTS Chapter XI. . . . . . . . . . . . . . . . . . . . . . . . . 4 Exercises . . . . . . . . . . Bibliography Index vii 139 139 139 142 145 147 149 . . . . . .
viii TABLE OF CONTENTS .
Part I Nonlinear Analysis 1 .
.
suitable for our purposes is stated and proved. for every scalar λ and every u ∈ E. results about linear operators which are needed in these notes will be quoted as needed. a function · : E → R+ having the properties: i) ii) iii) iv) u ≥ 0. We then discuss the notion of diﬀerentiability of Banach–space valued functions and state an inﬁnite dimensional version of Taylor’s theorem. v. for all u.Chapter I Analysis In Banach Spaces 1 Introduction This chapter is devoted to developing some tools from Banach space valued function theory which will be needed in the following chapters. v) = u − v and (E. A norm · deﬁnes a metric d : E × E → R+ by d(u. λu = λ u . In this chapter we shall mainly be concerned with results for not necessarily linear functions. As we shall see. i. u = 0 is equivalent to u = 0 ∈ E. for every u ∈ E. As a consequence we derive the Inverse Function theorem in Banach spaces and close this chapter with an extension theorem for functions deﬁned on proper subsets of the domain space (the Dugundji extension theorem). u + v ≤ u + v . We ﬁrst deﬁne the concept of a Banach space and introduce a number of examples of such which will be used later. ∈ E (triangle inequality). a crucial result is the implicit function theorem in Banach spaces. 2 Banach Spaces Let E be a real (or complex) vector space which is equipped with a norm · . · ) or simply E (if it is understood which norm is being used) is called a 3 . a version of this important result.e.
then b is called a bilinear form. ¯ then f 0 < +∞. The veriﬁcation that the spaces deﬁned are Banach spaces may be found in the standard literature on analysis. w + v. u . i. Rm ). d). v . x∈Ω (1) where  ·  is a norm in Rm . λ ∈ C. a mapping b : E × E → C having the properties (i)–(iv) above (with ·. v. v = λ u. v ∈ E u + v. deﬁne Let f 0 = sup f (x). u ∈ E. then E is called a Hilbert space. it follows that the space E = {f ∈ C 0 (Ω. Hence C 0 (Ω. ¯ ¯ If Ω is as above and Ω′ is an open set with Ω ⊂ Ω′ . all Cauchy sequences have limits in E). Rm ) : f 0 < ∞} is a Banach space. Rm ) = {the ¯ ¯ restriction to Ω of f ∈ C 0 (Ω′ . ∈ E u. They will frequently be employed in the applications presented later. is complete (i. w = u. i. u. · : E × E → R (or C (the complex numbers)) satisfying i) ii) iii) iv) u. The following collection of spaces are examples of Banach spaces. u.4 Banach space if the metric space (E. If Ω is bounded and f ∈ C 0 (Ω. u ≥ 0. w . Rm ) is a Banach space. u . w ∈ E λu. · replaced by b(·. . Rm ) = {f : Ω → Rm such that f is continuous on Ω}. then E is a normed space with the norm deﬁned by u = If E is complete with respect to this norm. and u. we let C 0 (Ω. Let Ω be an open subset of Rn . Since the uniform limit of a sequence of continuous functions is again continuous. a mapping ·. u ∈ E. ·)). Rm )}. v. u = 0 if and only if u = 0. d deﬁned as above.1 Spaces of continuous functions C 0 (Ω.e. u.e. If E is a real (or complex) vector space which is equipped with an inner product. in case E is a real vector space. v = v.e. u. 2. An inner product is a special case of what is known as a conjugate linear form.
Rm ) j ¯ m and if Ω is bounded C (Ω. R ) is a Banach space. x − yα y=x and H¨lder continuous with exponent α. in ) be a multiindex. if sup f (x) − f (y) < ∞. is given by Dβ f (x) = ∂ i1 x1 ∂ β f (x) . 0 < α ≤ 1. Let f : Ω → Rm . 0 < α ≤ 1. Let f j. i. β ≤ j}. Rm ) : f j < +∞} is a Banach space.α (Ω. 2. x − yα (3) If f ∈ C j (Ω. is H¨lder continuous with exponent α on o Ω. we say f ∈ C j. Let β = (i1 . BANACH SPACES 5 2. Let j f j = k=0 β≤k max Dβ f 0. We let β = k=1 ik . Rm ) = {f : Ω → Rm such that Dβ f is continuous for all β. β = j. xn ). β=j then the space E = {f ∈ C j.α < ∞} .2 Spaces of diﬀerentiable functions Let Ω be an open subset of Rn .y∈Ω f (x) − f (y) . · · · . (2) Then. Dβ f (x). on Ω if it is H¨lder contino o uous with the same exponent α at every x ∈ Ω. Rm ) : f is a Banach space.2. 1 ≤ k ≤ n.3 H¨lder spaces o Let Ω be an open set in Rn .α (Ω. then the partial derivative of f of order β. ¯ ¯ The space C j (Ω. at a point x ∈ Ω. Rm ) and Dβ f. j. · · · . Rm ). Rm ) is deﬁned in a manner similar to the space C 0 (Ω.e. n ik ∈ Z (the nonnegative integers). · · · ∂ in xn where x = (x1 . using further convergence results for families of diﬀerentiable functions it follows that the space E = {f ∈ C j (Ω. For such f we deﬁne α HΩ (f ) = sup x=y x.α = f j α + max HΩ (Dβ f ). A function f : Ω → Rm is called H¨lder o continuous with exponent α. Deﬁne C j (Ω.
α (Ω. 1/p f Lp = Ω f (x)p dx .α (Ω. Rm ) is a Hilbert space with inner product deﬁned by f. Then Lp (Ω. Rm ). Rm ) and ¯ ¯ C j. for 1 ≤ p < ∞. We let j. let f L∞ = essupx∈Ω f (x). Rm ) = C j (Ω. We shall also employ the following convention j. A function f : Ω → Rm is said to have compact support in Ω if the set supp f = closure{x ∈ Ω : f (x) = 0} = {x ∈ Ω : f (x) = 0} is compact. one may deﬁne the space C j. The space L2 (Ω. j. x ∈ ∂Ω}. Rm ) = {f ∈ C j. Rm ) is a Banach space. C (Ω.α (Ω.4 Functions with compact support Let Ω be an open subset of Rn . Rm ) similarly. again. For 1 ≤ p ≤ ∞.5 Lp spaces Let Ω be a Lebesgue measurable subset of Rn and let f : Ω → Rm be a measurable function. where f (x) · g(x) is the inner product of f (x) and g(x) in the Hilbert space (Euclidean space) Rn .α ¯ ¯ C0 (Ω. And again.6 As above. and for p = ∞. where essup denotes the essential supremum.0 (Ω. Let. Rm ) = C j (Ω.α ¯ Then. 2. Rm ) = {f : f Lp < +∞}.0 (Ω. let Lp (Ω. Rm ). Rm ) is a Banach space for 1 ≤ p ≤ ∞. Rm ) : supp f is a compact subset of Ω} j. g = Ω f (x) · g(x)dx. if Ω is bounded. 2.α C0 (Ω.α ¯ and deﬁne C0 (Ω. Rm ) = {f ∈ C j. if Ω is bounded. .α C j. Rm ) is a Banach space and j. Rm ) : f (x) = 0. the space C0 (Ω.
H0 (Ω. g = Ω α≤k Dα f · Dα gdx. Rm ). and W k. 2. Rm ). if f has weak derivatives up to order k. is a Banach subspace which. Rm ). up to a set of measure zero. [18]).p (Ω. respectively. if for every compact subset Ω′ ⊂ Ω. they may be identiﬁed with the following spaces. Rm ). it’s k. R ). denoted by W0 (Ω.p m closure in W (Ω. .2. It’s completion is denoted by H k. (6) These spaces play a special role in the linear theory of partial diﬀerential equations.2 For p = 2. the spaces W k. Then the vector space W k. .6 Weak derivatives Let Ω be an open subset of Rn . A function f : Ω → Rm is said to belong to class Lp (Ω. Rm ) : Dβ f ∈ Lp (Ω. for all φ ∈ C0 (Ω). v is uniquely determined. Rm ). . ¯ Consider the space C k (Ω.7 set Sobolev spaces We say that f ∈ W k (Ω. The space C0 (Ω. Then a locally integrable function v ∈ L1 (Ω. Rm ) is the completion of C0 (Ω. Rm ) are Hilbert spaces with inner product f. BANACH SPACES 7 2.g. If p = 2 it is a Hilbert space with k.p (Ω. The concept of weak derivative extends the classical concept of derivative and has many similar properties (see e. .8 Let Spaces of linear operators E Let E and X be normed linear spaces with norms · and · X.2 (Ω.p k is a Banach space.p k. Rm ) loc is called the β th weak derivative of f if it satisﬁes vφdx = (−1)β Ω Ω ∞ f Dβ φdx. Rm ).p (Ω. in general. is a proper subspace. β ≤ k}. Rm ). L(E. 2. Rm ) = {f ∈ W k (Ω. Rm ) and W0 (Ω. X) = {f : E → X such that f is linear and continuous}. βn ) be a multiindex. g given by = Ω β≤k Dβ f p dx (5) f.p ∞ inner product given by (6). . k. Rm ) equipped with the norm 1/p f W k. Rm ) is a subspace of W k. Rm ) in H k. f ∈ Lp (Ω′ .p (Ω. Rm ).p (Ω. Rm ) as a normed space using the · W k. (4) We write v = Dβ f and note that. Let β = loc (β1 .p norm. and in case Ω satisﬁes suﬃcient regularity conditions (see [1]. [41]).
e. .g. . L(E. . . . . there exists an isomorphism between these spaces which is norm preserving (see e. . X) = L(E. 3 3. . whenever X is. xn ) X : x1 E1 ≤ 1. (8) then L(E1 . L1 (E. Let x0 ∈ U . . Taylor’s Theorem Gˆteaux and Fr´chet diﬀerentiability a e f :U →X Let E and X be Banach spaces and let U be an open subset of E. This space is a Banach space. . . f is linear in each variable separately) and continuous}. let f L = sup x E ≤1 f (x) X. then f is said to be Gˆteaux diﬀerentiable (Ga diﬀerentiable) at x0 in direction h. if 1 lim {f (x0 + th) − f (x0 )} t→0 t (9) exists. We leave it as an exercise to show that the spaces L(E. . h (10) h →0 . Let f = sup{ f (x1 . If E and X are normed spaces. . if there exists e T ∈ L(E.1 Diﬀerentiability. . E. . X) with the norm deﬁned by (8) is a normed linear space. . one may deﬁne the spaces L1 (E. Ln (E. (7) Then · L is a norm for L(E. .e. . whenever X is. X) may be identiﬁed (i. It again is a Banach space. X)) . X)). X) L(E. . Ln−1 (E. let L(E1 . . . n ≥ 2. . X). .8 For f ∈ L(E. X) (E repeated n times) and Ln (E. = = . En . . . . xn En ≤ 1}. . X) L2 (E. X) such that f (x0 + h) − f (x0 ) = T (h) + o( h ) for h small. . It said to be Fr´chet diﬀerentiable (Fdiﬀerentiable) at x0 . here o( h ) means that lim o( h ) = 0. X) = {f : E1 × · · · × En → X such that f is multilinear (i. . . Let be a function. X) . [43]). En and X be n + 1 normed linear spaces. En . X). Let E1 .
if it e exists. Df (x0 ). i.) We note that Fr´chet diﬀerentiability is a more restrictive concept. since it will be clear from the context in which space we are working. X) is Fr´chet–diﬀerentiable at x0 ∈ U . TAYLOR’S THEOREM 9 (We shall use the symbol · to denote both the norm of E and the norm of X. X).8). If f : Rn → Rn is Fr´chet diﬀerentiable at x0 . X). h) as Dn f (x0 )hn . Df (x0 )(h) = ∇f (x0 ) · h. where “·” is the dot product in Rn . We shall use the following symbols interchangeably for the Fr´chet–derivative of f at x0 . where the latter is usually e used in case X = R. If h ∈ E. . DIFFERENTIABILITY. and the second derivative D2 f (x0 ) is given by the Hessian matrix ∂2f ∂xi ∂xj . or f ′′ (x0 ). or d2 f (x0 ). (see Subsection 2. we say that f is twice Fr´chet–diﬀerentiable e e and we denote the second F–derivative by D2 f (x0 ). e It follows from this deﬁnition that the Fr´chet–derivative of f at x0 . i. is unique. where hT is the transpose of the vector h. Thus D2 f (x0 ) ∈ L2 (E. · · · . then Df (x0 ) is represented by the e gradient vector ∇f (x0 ).3.e. We say that f is of class C 1 in a neighborhood of x0 if f is Fr´chet diﬀerentiable there and if the mapping e Df : x → Df (x) is a continuous mapping into the Banach space L(E. f ′ (x0 ).e. df (x0 ). If the mapping Df : U → L(E. then Df (x0 ) is given by the e Jacobian matrix Df (x0 ) = ∂fi ∂xj . x=x0 D2 f (x0 )h2 = hT ∂2f ∂xi ∂xj h. we shall write Dn f (x0 )(h. x=x0 and if f : Rn → R is Fr´chet diﬀerentiable. In an analogous way one deﬁnes higher order diﬀerentiability.
10
3.2
Taylor’s formula
1 Theorem Let f : E → X and all of its Fr´chet–derivatives Df, · · · , Dm f be e continuous on an open set U . Let x and x + h be such that the line segment connecting these points lies in U . Then
m−1
f (x + h) − f (x) =
k=1
1 m 1 k D f (x)hk + D f (z)hm , k! m!
(11)
where z is a point on the line segment connecting x to x + h. The remainder 1 m m is also given by m! D f (z)h 1 (m − 1)!
1 0
(1 − s)m−1 Dm f (x0 + sh)hm ds.
(12)
We shall not give a proof of this result here, since the proof is similar to the one for functions f : Rn → Rm (see e.g. [14]).
3.3
EulerLagrange equations
In this example we shall discuss a fundamental problem of variational calculus to illustrate the concepts of diﬀerentiation just introduced; speciﬁcally we shall derive the so called EulerLagrange diﬀerential equations. The equations derived give necessary conditions for the existence of minima (or maxima) of certain functionals. 2 Let g : [a, b] × R2 → R be twice continuously diﬀerentiable. Let E = C0 [a, b] and let T : E → R be given by
b
T (u) =
a
g(t, u(t), u′ (t))dt.
It then follows from elementary properties of the integral, that T is of class C 1 . Let u0 ∈ E be such that there exists an open neighborhood U of u0 such that for all u ∈ U (u0 is called an extremal of T ). Since T is of class C 1 we obtain that for u ∈ U T (u) = T (u0 ) + DT (u0 )(u − u0 ) + o( u − u0 ). Hence for ﬁxed v ∈ E and ǫ ∈ R small, T (u0 + ǫv) = T (u0 ) + DT (u0 )(ǫv) + o(ǫ v ). It follows from (13) that 0 ≤ DT (u0 )(ǫv) + o(ǫ v ) and hence, dividing by ǫv , 0 ≤ DT (u0 ) ± v v + o( ǫv ) , ǫv T (u0 ) ≤ T (u) (13)
4. SOME SPECIAL MAPPINGS
11
where · is the norm in E. It therefore follows, letting ǫ → 0, that for every v ∈ E, DT (u0 )(v) = 0. To derive the Euler–Lagrange equation, we must compute DT (u0 ). For arbitrary h ∈ E we have T (u0 + h) = = g(t, u0 (t) + h(t), u′ (t) + h′ (t))dt 0 g(t, u0 (t), u′ (t)dt 0 b + a ∂g (t, u0 (t), u′ (t))h(t)dt 0 ∂p +
b ∂g (t, u0 (t), u′ (t))h′ (t)dt 0 a ∂q b a b a
+ o( h ),
where p and q denote generic second, respectively, third variables in g. Thus
b
DT (u0 )(h) =
a
∂g (t, u0 (t), u′ (t))h(t)dt + 0 ∂p
b a
∂g (t, u0 (t), u′ (t))h′ (t)dt. 0 ∂q
For notation’s sake we shall now drop the arguments in g and its partial derivatives. We compute
b a
∂g ′ ∂g h dt = h ∂q ∂q
b a
b
−
h
a
d dt
∂g ∂q
dt,
and since h ∈ E, it follows that
b
DT (u0 )(h) =
a
∂g d ∂g hdt. − ∂p dt ∂q
(14)
Since DT (u0 )(h) = 0 for all h ∈ E, it follows that d ∂g ∂g (t, u0 (t), u′ (t)) − (t, u0 (t), u′ (t)) = 0, 0 0 ∂p dt ∂q (15)
for a ≤ t ≤ b (this fact is often referred to as the fundamental lemma of the calculus of variations). Equation (15) is called the EulerLagrange equation. If g is twice continuously diﬀerentiable (15) becomes ∂ 2g ∂2g ′ ∂2g ∂g − − u0 − 2 u′′ = 0, ∂p ∂t∂q ∂p∂q ∂q 0 (16)
where it again is understood that all partial derivatives are to be evaluated at (t, u0 (t), u′ (t)). We hence conclude that an extremal u0 ∈ E must solve the 0 nonlinear diﬀerential equation (16).
4
Some Special Mappings
Throughout our text we shall have occasion to study equations deﬁned by mappings which enjoy special kinds of properties. We shall brieﬂy review some such properties and refer the reader for more detailed discussions to standard texts on analysis and functional analysis (e.g. [14]).
12
4.1
Completely continuous mappings
Let E and X be Banach spaces and let Ω be an open subset of E, let f :Ω→X be a mapping. Then f is called compact, whenever f (Ω′ ) is precompact in X for every bounded subset Ω′ of Ω (i.e. f (Ω′ ) is compact in X). We call f completely continuous whenever f is compact and continuous. We note that if f is linear and compact, then f is completely continuous. 2 Lemma Let Ω be an open set in E and let f : Ω → X be completely continuous, let f be Fdiﬀerentiable at a point x0 ∈ Ω. Then the linear mapping T = Df (x0 ) is compact, hence completely continuous. Proof. Since T is linear it suﬃces to show that T ({x : x ≤ 1}) is precompact in X. (We again shall use the symbol · to denote both the norm in E and in X.) If this were not the case, there exists ǫ > 0 and a sequence {xn }∞ ⊂ n=1 E, xn ≤ 1, n = 1, 2, 3, · · · such that T xn − T xm ≥ ǫ, n = m. Choose 0 < δ < 1 such that f (x0 + h) − f (x0 ) − T h < for h ∈ E, h ≤ δ. Then for n = m f (x0 + δxn ) − f (x0 + δxm ) ≥ δ T xn − T xm − ≥ f (x0 + δxn ) − f (x0 ) − δT xn − f (x0 + δxm ) − f (x0 ) − δT xm δǫ −
δǫ 3
ǫ h , 3
−
δǫ 3
=
δǫ 3 .
Hence the sequence {f (x0 + δxn )}∞ has no convergent subsequence. On the n=1 other hand, for δ > 0, small, the set {x0 + δxn }∞ ⊂ Ω, and is bounded, n=1 implying by the complete continuity of f that {f (x0 + δxn )}∞ is precompact. n=1 We have hence arrived at a contradiction.
4.2
Proper mappings
Let M ⊂ E, Y ⊂ X and let f : M → Y be continuous, then f is called a proper mapping if for every compact subset K of Y , f −1 (K) is compact in M . (Here we consider M and Y as metric spaces with metrics induced by the norms of E and X, respectively.)
then f has a unique ﬁxed point in M . provided that f is coercive. (17) Proof. and let f = id − h be coercive. the Banach ﬁxed point theorem Let M be a subset of a Banach space E. SOME SPECIAL MAPPINGS 13 3 Lemma Let h : E → X be completely continuous and let g : E → X be proper. That N is also closed follows from the fact that g and h are continuous.e. it follows that {xn }∞ converges also. f (x) → ∞ as x → ∞. {g(xn )}∞ and {h(xn )}∞ are convergent. further. we have: 5 Lemma Let f : Rn → Rm be continuous. for all x.e. 4. the sequence {yn }∞ has a convergent subsequence. then f is proper (here id is the identity mapping). and n=1 since f is coercive the sequence {xn }∞ must be bounded. Proof. if necessary. A function f : M → E is called a contraction mapping if there exists a constant k. because n=1 h is completely continuous. (20) . Since n=1 n=1 n=1 g(xn ) = yn + h(xn ) and g is proper. In ﬁnite dimensional spaces the concepts of coercivity and properness are equivalent. Relabeling. Then there exists {yn }∞ ⊂ K such that n=1 yn = g(xn ) − h(xn ). i. say xn → x. i. then f is proper if and only if f is coercive.e. the sequence {h(xn )}∞ must have a convergent n=1 ∞ subsequence. hence N n=1 is precompact. (18) Since K is compact. then f = g − h is a proper mapping. 0 ≤ k < 1 such that f (x) − f (y) ≤ k x − y . we may assume that all three sequences {yn }∞ . Let {xn }∞ n=1 be a sequence in N . 4 Corollary Let h : E → E be a completely continuous mapping.4. (19) 6 Theorem Let M be a closed subset of E and f : M → M be a contraction mapping. y ∈ M.3 Contraction mappings. i. Let K be a compact subset of X and let N = f −1 (K). It follows that the sequence {g(xn )}n=1 has a convergent subsequence. there exists a unique x ∈ M such that f (x) = x . We note that id : E → E is a proper mapping.
7 Remark We note that the above theorem. y ∈ M both satisfy (20). hence. T ). . 1−k (22) = = xm − xm−1 + xm−1 − . u. and hence if m > n xm − xn and therefore xm − xn ≤ x1 − x0 (k n + . We consider the Dirichlet problem. i. n ≥ 1. + xn+1 − xn f (xm−1 ) − f (xm−2 ) + . To proof existence. 0 < t < T. In the following example we provide an elementary approach to the existence and uniqueness of a solution of a nonlinear boundary value problem (see [13]). . . Theorem 6. + k m−1 ) = kn − km x1 − x0 . T ] × R × R → R be a mapping satisfying Carath´odory conditions. T }. u′ ) is continuous e in (u. . Let T > 0 be given and let f : [0.e. f (t. also holds if E is a complete metric space with metric d. x must equal y. We have the following results: . u. x ∈ M . hence n=0 n→∞ lim xn = x exists and since M is closed. (21) implies that for any j ≥ 1 f (xj ) − f (xj−1 ) ≤ k j x1 − x0 . t ∈ {0. If x. The approach is based on the Lp theory of certain linear diﬀerential operators subject to boundary constraints. we shall employ the notation that  ·  stands for absolute value in R and · 2 the norm in L2 (0. the problem of ﬁnding a function u satisfying the following diﬀerential equation subject to boundary conditions u′′ u = f (t. .e. + f (xn ) − f (xn−1 ) (21) It follows from (22) that {xn }∞ is a Cauchy sequence in E. u′ ). i. establishing uniqueness of a ﬁxed point. u′ ). (23) In what is to follow. = 0. we deﬁne a sequence {xn }∞ ⊂ M inductively as follows: n=0 Choose x0 ∈ M and let xn = f (xn−1 ). since k < 1. . This is easily seen by replacing x − y by d(x. u′ ) for almost all t and measurable in t for ﬁxed (u. Using (21) we obtain that (20) holds. y) in the proof. then x − y = f (x) − f (y) ≤ k x − y .14 Proof.
0 < t < T. λ1 λ1 (25) and λ1 is the principal eigenvalue of −u′′ subject to the Dirichlet boundary conditions u(0) = 0 = u(T ) (i. T ). Then problem (23) has a unique solution u ∈ 1 C0 ([0. T ) → u ∈ L1 (0. is a continuous mapping. w. T ]). T ). π2 λ1 = T 2 . (28) where G(t. T ). q ≥ 1.(24) ˜ ˜ ˜ ˜ ˜ ˜ where a. the smallest number λ such that the problem −u′′ u = λu. Proof.e. T ) ֒→ L1 (0. To prove the theorem. with u′ absolutely continuous and the equation (23) being satisﬁed almost everywhere. (26) has a nontrivial solution). since u L1 q ≤ T q−1 u Lq .4. T ) to any Lq (0. for v ∈ L1 (0. u. q ≥ 1. On the other hand we have that the imbedding Lq (0. s) = − 1 T (T − t)s. s)v(s)ds. Results from elementary diﬀerential equations tell us that λ1 is the ﬁrst positive number λ such that the problem (26) has a nontrivial solution. (29) It follows from (24) that the operator A is a mapping of L1 (0.e. u. . T ). if 0 ≤ s ≤ t t(T − s). b are nonegative constants such that b a + √ < 1. which. v)−f (t. t ∈ {0. u. if t ≤ s ≤ T. where w(t) = − t T T 0 0 τ t τ (27) v(s)dsdτ + 0 0 v(s)dsdτ. SOME SPECIAL MAPPINGS 8 Theorem Let f satisfy 15 f (t. 0 < t < T. v ) ≤ au− u+bv − v. let us. T }. v ∈ R. in turn may be rewritten as T w(t) = 0 G(t. ∀u. = 0. i. w′ ). put Av = f (·. u ∈ Lq (0. v.
16 We hence may consider A : Lq (0. T ]. from which easily follows. T ) and u′′ ∈ L2 (0. their proofs may be obtained using Fourier series T methods. if v ∈ L2 (0. s)v(s)ds 1 is in C0 (0. λ1 = 1 and condition (25) becomes a + b < 1. s)v(s)ds that w L2 ≤ 1 v λ1 L2 . On the other hand. and will be left as an exercise. whereas a classical result of Picard requires a π π2 + b < 1. then T u(t) = 0 G(t. 8 2 (see [21] where also other results are cited). . T ) is a ﬁxed point of A. (30) The derivation of such a statement is left as an exercise. 10 Remark In case T = π. for any q ≥ 1. the following inequalities will be used. We have for w(t) = 0 G(t. 9 Remark It is clear from the proof that in the above the real line R may be replaced by Rm thus obtaining a result for systems of boundary value problems. via an integration by parts. T ) and u solves (23). T ) → Lq (0. T ). with u′ absolutely continuous. 11 Remark Theorem 8 may be somewhat extended using a result of Opial [31] which says that for u ∈ C0 [0. we have that T 0 u(x)u′ (x)dx ≤ T 4 T 0 u′ (x)2 dx. that w′ L2 1 v ≤ √ λ1 L2 Using these facts in the computations one obtains the result that A is a contraction mapping. In carrying out the computations in the case q = 2.
iii) Du G(u0 . Proof. λ0 )) = G(u. V is open in Λ) be a continuous mapping satisfying the following condition: • For each λ ∈ V the map f (·. Let us consider the equation f (u. (where U is open in E. The mapping G has the following properties: i) G(u0 . λ). λ0 )]−1 (f (u. ii) G and Du G are continuous in (u. λ0 )]−1 ∈ L(X. λ0 ). λ) = f (u0 . Then there exist δ > 0 and r > 0 and unique mapping u : Bδ (λ0 ) = {λ : λ − λ0 ≤ δ} → E such that f (u(λ). X) and [Du f (u0 . def (32) (33) (34) . X. E)). λ0 ) = 0. λ0 ) which is equivalent to [Du f (u0 .4 The implicit function theorem Let us now assume we have Banach spaces E. or u = u − [Du f (u0 . λ0 ) ∈ U × V such that Du f (u0 . λ) = f (u0 . λ0 )) = 0. 12 Theorem (Implicit Function Theorem) Let f satisfy (31) and let there exist (u0 . λ) − f (u0 . λ). X). Du f (u0 . λ) : U → X is Fr´chetdiﬀerentiable on U e with Fr´chet derivative e Du f (u. Λ and let f : U × V → X. λ0 ) is a linear homeomorphism of E onto X (i. λ0 ) = u0 . λ) (31) and the mapping (u. and u(λ) − u0 ≤ r. λ) is a continuous mapping from U × V to L(E. SOME SPECIAL MAPPINGS 17 4.4. λ) − f (u0 . u(λ0 ) = u0 .e. λ) → Du f (u. λ0 )]−1 (f (u. λ0 ) ∈ L(E.
λ) − G(u0 . u(λ0 ) = u0 . λ) sup0≤t≤1 Du G(u1 + t(u2 − u1 ). λ) in M . 2 hence g has a unique ﬁxed point by the contraction mapping principle (Theorem 6). that for λ ∈ R. and u 0 = supλ∈Bδ (λ0 ) u(λ) < +∞}. by an application of Theorem 12. 0 < t < π. Then M is a closed subset of a Banach space and (35) deﬁnes an equation u(λ) = G(u(λ). λ) − G(u0 . λ) − G(u0 . in a neighborhood of 0. (37) This is a one spacedimensional mathematical model from the theory of combustion (cf [3]) and u represents a dimensionless temperature. λ) − G(u2 . 14 Example As an example let us consider the nonlinear boundary value problem u′′ + λeu = 0. λ). λ) − u0 = G(u. 13 Remark If in the implicit function theorem f is k times continuously diﬀerentiable. λ0 ) ≤ 2 r. λ0 ) ≤ ≤ 1 1 r + r. (37) has a unique solution of small norm in C 2 ([0. π]. Now G(u. λ) 1 2 u1 − u2 . where r is small enough. u1 − u2 (35) provided u1 − u0 ≤ r. Deﬁne g by (here we think of u as an element of M ) g(u)(λ) = G(u(λ). Let Bδ (λ0 ) = {λ : λ − λ0 ≤ δ} and deﬁne M = {u : Bδ (λ0 ) → E such that u is continuous. λ) − G(u0 . We shall show. . R). u(λ) − u0 0 ≤ r. then g : M → M and it follows by (36) that g(u) − g(v) 0 (36) ≤ 1 u − v 0. 2 2 1 u − u0 + G(u0 . λ0 ) 2 1 provided λ − λ0 ≤ δ is small enough so that G(u0 . u(0) = 0 = u(π). λ) + G(u0 . then the mapping λ → u(λ) inherits this property. λ) − G(u0 . λ0 ) ≤ G(u. u2 − u0 ≤ r.18 Hence ≤ ≤ G(u1 .
λ) is given by (the reader should carry out the veriﬁcation) Du f (u0 . Let us consider the linear mapping T = Du f (0. λ)v = v ′′ + λeu0 (x) v. SOME SPECIAL MAPPINGS To this end we deﬁne 2 E = C0 ([0. λ) → Du f (u. λ) = u′′ + λeu . To see this we note that for every h ∈ X. λ) is continuous. We must show that this mapping is a linear homeomorphism. 0 < t < π. R) 19 X = C 0 [0. v(0) = 0 = v(π).4. 0) = 0. 0) : E → X. Du f (u0 . and hence the mapping (u. s)h(s)ds. π] Λ = R. Let f :E×Λ→ X be given by f (u. From the representation (38) we may conclude that there exists a constant c such that v 2 = T −1h 2 ≤ c h 0. these spaces being equipped with their usual norms (see earlier examples). 0 ≤ s ≤ t 1 − π t(π − s). (38) where G(x. . (When λ = 0 (no heat generation) the unique solution is u ≡ 0. t ≤ s ≤ π.) Furthermore. the unique solution of v ′′ = h(t). Then f is continuous and f (0. s) = 1 − π (π − t)s. for u0 ∈ E. π]. is given by (see also (28)) π v(t) = 0 G(t.
Hence 0 = u′′ + λeu > u′′ + λu. uvdx. ii) f is one to one on U ′ .e. T −1 is one to one and continuous. Hence all conditions of the implicit function theorem are satisﬁed and we may conclude that for each λ. v(0) = 0 = v(π). From (39) and (40) we obtain π (39) (40) 0> 0 (u′′ v − v ′′ u)dt + (λ − 1) π uvdt. Let v(t) = sin t. Then there exist open sets U ′ and V . (37) has a unique small solution u ∈ C 2 ([0. 0 and hence. then the corresponding solution u must be positive. 15 Theorem Let E and X be Banach spaces and let U be an open neighborhood of a ∈ E. π]. We later shall show that this ‘solution branch’ (λ. We observe that if λ > 0 is such that (37) has a solution. f (a) ∈ V and a uniquely determined function g such that: i) V = f (U ′ ). To this end we note here that the set {λ > 0 : (37) has a solution } is bounded above. for every u ∈ U ′ . R). Since the ﬁrst result is proved in exactly the same way as its ﬁnite dimensional analogue (it is an immediate consequence of the implicit function theorem) we shall not prove it here (see again [14]). iii) g : V → U ′ . then v satisﬁes v ′′ + v = 0. λ suﬃciently small. g(f (u)) = u. u(x) > 0. 0 < t < π. R). u(λ)) may be globally continued. . g(V ) = U ′ . integrating by parts. 0 < x < π. a ∈ U ′ . π].20 i. π 0 > (λ − 1) implying that λ < 1. furthermore the map λ → u(λ) is continuous from a neighborhood of 0 ∈ R to C 2 ([0. 0 5 Inverse Function Theorems We next proceed to the study of the inverse of a given mapping and provide two inverse function theorems. iv) g is a C 1 function on V and Dg(f (a)) = [Df (a)]−1 . Let f : U → X be a C 1 mapping with Df (a) a linear homeomorphism of E onto X.
Then Df (u) is deﬁned by (Df (u))(v) = v ′′ + λv + 2uv. . Hence Df (0) is a linear homeomorphism of E onto X. and that v 2 ≤ C h 0 for some constant C (only depending upon λ). (32) has a unique solution u ∈ E of small norm for every g ∈ X of small norm.. i. Then for certain values of λ. 2. u′ (0) = u′ (2π)}. Let M and Y metric spaces (e. . f is a C 1 mapping.periodic response of small norm for every 2π . . 2π]. u(0) = u(2π). has a unique 2π–periodic solution for every 2π–periodic h as long as λ = n2 . n = 1. [4]) that the problem v ′′ + λv = h.e. . We note that the above example is prototypical for forced nonlinear oscillators. and hence the mapping u → Df (u) is a continuous mapping of E to L(E.g. INVERSE FUNCTION THEOREMS 21 16 Example Consider the forced nonlinear oscillator (periodic boundary value problem) u′′ + λu + u2 = g.5. Let f :E→X be given by f (u) = u′′ + λu + u2 . X). (41) has a unique solution for all forcing terms g of small norm. u′ (0) = u′ (2π) (41) where g is a continuous 2π − periodic function and λ ∈ R. where both spaces are equipped with the norms discussed earlier. . . Virtually the same arguments can be applied (the reader might carry out the necessary calculations) to conclude that the forced pendulum equation u′′ + λ sin u = g (42) has a 2π. n = 1. and X = C 0 ([0. The following result describes a class of problems where the precise number of solutions (for every given forcing term) may be obtained by simply knowing the number of solutions for some ﬁxed forcing term. .periodic forcing term g of small norm. as long as λ = n2 . is a parameter. We thus conclude that for given λ = n2 . subsets of Banach spaces with metric induced by the norms). R) ∩ {u : u(0) = u(2π). In many physical situations (see the example below) it is of interest to know the number of solutions of the equation describing this situation. It follows from elementary diﬀerential equations theory (see eg. 2. Let E = C 2 ([0. . R). 2π].
because f is a proper mapping. Since f is locally invertible. We ﬁrst give some terminology. where Λ is an index set. hence ﬁnite. Proof. For y ∈ Y let N (y) = cardinal number of {f −1 (y)} = #{f −1 (y)}. n for some i. . Then the mapping y → N (y) is ﬁnite and locally constant. Hence ξ = ui . Let y ∈ Y and let {f −1 (y)} = {u1 . . And since f is continuous. v ∈ I!). un }. The result which guarantees the existence of extensions having the desired properties is the Dugundji extension theorem ([16]) which will be established in this section. N (ym ) > N (y) (note that for any v ∈ I. there exists. then N (Y ) is constant. N (y) is ﬁnite. Theorem 15 is applicable at each point). Since f −1 ({yn } ∪ {y}) is compact. For if not. for each u ∈ {f −1 (y)} a neighborhood Ou such that Ou ∩ ({f −1 (y)}\{u}) = ∅. such that f (ξm ) = ym . Then {Oλ }λ∈Λ is called locally ﬁnite if every point u ∈ M has a neighborhood U such that U intersects at most ﬁnitely many elements of {Oλ }λ∈Λ .g. . . 18 Corollary Assume Y is connected. and thus {f −1 (y)} is a discrete and compact set. f (ξ) = y. say ξnj → ξ.22 17 Theorem Let f : M → Y be continuous. 6 The Dugundji Extension Theorem In the course of developing the Brouwer and Leray–Schauder degree and in proving some of the classical ﬁxed point theorems we need to extend mappings deﬁned on proper subsets of a Banach space to the whole space in a suitable manner. v has a preimage in each Vi which implies N (v) ≥ N (y). n Hence there exists a sequence {ξm }. which will imply that N is constant–valued. the sequence {ξn } will have a convergent subsequence. there will exist a sequence {ym }. Since {y} is compact {f −1 (y)} is compact also. ξm ∈ i=1 Vi . We next show that N is a continuous mapping to the nonnegative integers. be an open cover of M .g [16]). such that as m → ∞. We next claim that there exists a neighborhood W ⊂ I of y such that N is constant on W . Examples illustrating this result will be given later in the text. Then there exist open sets Vi . with ym → y. We choose disjoint open neighborhoods Oi of ui . 1 ≤ i ≤ n and n let I = i=1 f (Oi ). a contradiction to ξ ∈ i=1 Vi . . Let M be a metric space and let {Oλ }λ∈Λ . proper and locally invertible (e. We ﬁrst show that for each y ∈ Y . ui ∈ Vi such that f is a homeomorphism from Vi to I. In proving the theorem we need a result from general topology which we state here for convenience (see e.
The collection {Bu }u∈E\C is an open cover of the metric space E\C and hence has a locally ﬁnite reﬁnement {Oλ }λ∈Λ . (43) The sum in the right hand side of (43) contains only ﬁnitely many terms. since {Oλ }λ∈Λ is locally ﬁnite. E\Oλ ).e. C) 3 Oλ ⊃ E\C. λ ∈ Λ. ρλ (u) = 1. q(u) It follows for λ ∈ Λ and u ∈ E\C that 0 ≤ ρλ (u) ≤ 1. This also implies that q is a continuous function. Deﬁne ρλ (u) = dist(u. Then there exists a continuous mapping ˜ f :E→K such that ˜ f (u) = f (u). ∞) by q(u) = λ∈Λ dist(u. where C is closed in E and K is convex in X. C). i. i) λ∈Λ 1 dist(u. Deﬁne q : E\C → (0. E\Oλ ) . u ∈ C. λ∈Λ . For each u ∈ E\C let ru = and Bu = {v ∈ E : v − u < ru }. Proof. u ∈ E\C. 20 Theorem Let E and X be Banach spaces and let f : C → K be a continuous mapping.6. THE DUGUNDJI EXTENSION THEOREM 23 19 Lemma Let M be a metric space. ii) for each λ ∈ Λ there exists Bu such that Oλ ⊂ Bu iii) {Oλ }λ∈Λ is locally ﬁnite. Then diamBu ≤ dist(Bu . Then every open cover of M has a locally ﬁnite reﬁnement.
To show that f is ˜ continuous on E it suﬃces therefore to show that f is continuous on ∂C. ﬁnd 0 < δ = δ(u. and therefore ˜ ˜ f (u) − f (v) ≤ ǫ ρλ (v) = ǫ. λ ∈ Λ. v ∈ Oλ . Now Oλ ⊂ Bu1 for some u1 ∈ E\C. If ρλ (v) = 0. Since v − w ≤ diamOλ we may take the inﬁmum for w ∈ Oλ and obtain v − uλ ≤ diamOλ + dist(uλ . Oλ ) ≤ 2dist(C.24 For each λ ∈ Λ choose uλ ∈ C such that dist(uλ . Let u ∈ ∂C. u ∈ C λ∈Λ ρλ (u)f (uλ ). ˜ iii) f is continuous on E\C. Therefore if u − v ≤ δ/4. since diamOλ ≤ diamBu1 ≤ dist(Bu1 . then since f is continuous we may.e. ˜ ˜ These properties follow immediately from the deﬁnition of f . Thus for λ such that ρλ (v) = 0 we get u − uλ ≤ v − u + v − uλ ≤ 4 u − v . then u − uλ ≤ δ. then dist(v. ˜ Then f has the following properties: ˜ i) f is deﬁned on E and is an extension of f . v ∈ C. i. ρλ (v)f (uλ ) ≤ λ∈Λ ρλ (v) f (u) − f (uλ ) . Hence v − uλ ≤ v − w + w − uλ for any w ∈ Oλ . ˜ ii) f is continuous on the interior of C. Oλ ) ≤ 3 v − u . E\Oλ ) > 0. Oλ ) and deﬁne ˜ f (u) = f (u). ǫ) such that f (u) − f (v) ≤ ǫ. we get v − uλ ≤ 3dist(C. and f (u) − f (uλ ) ≤ ǫ. Hence. if Now for v ∈ E\C ˜ ˜ f (u) − f (v) = f (u) − λ∈Λ u − v ≤ δ. Oλ ). Oλ ). λ∈Λ . u ∈ C. for given ǫ > 0. C) ≤ dist(C.
where ˜ C is closed in E. u ∈ K. Since K is closed and convex we may apply Corollary 21 to obtain the desired conclusion.7. Supply all the details for the proof of Theorem 8. where cof (C) is the convex hull of f (C). 7 Exercises 1. Compare the requirements discussed in Remark 10. i. Let id : K → K be the identity mapping. . K is a continuous retract of E. 5. 2.e. Then there exists a continuous mapping f : E → K such that f (E) = K and f (u) = u. Proof. Derive an improved result as suggested by Remark 11. 22 Corollary Let K be a closed convex subset of a Banach space E. 7. This map is continuous. Carry out the program laid out by Example 16 to discuss the nonlinear oscillator given by (42). 4. Establish the assertion of Remark 13. Prove Theorem 15. 6. 3. X be Banach spaces and let f : C → X be continuous. Then f has a continuous extension f to E such that ˜ f (E) ⊂ cof (C). EXERCISES 25 21 Corollary Let E. Supply the details of the proof of Example 14.
26 .
Chapter II The Method of LyapunovSchmidt 1 Introduction In this chapter we shall develop an approach to bifurcation theory which. 27 . We ﬁrst develop the method of Liapunov–Schmidt and then use it to obtain a local bifurcation result. will be of a local nature. ∀λ ∈ R. We then use this result in several examples. λ) = 0 (4) We call λ0 a bifurcation value or (0. u = 0) of the equation F (u. since the implicit function theorem plays an important role. λ0 ) a bifurcation point for (4) provided every neighborhood of (0. is one of the original approaches to the theory. λ0 ) in X × R contains solutions of (4) with u = 0. λ) = 0. The results obtained.e. Later in the text it will be used to derive a Hopf bifurcation theorem. (3) (2) We shall be interested in obtaining existence of nontrivial solutions (i. It then follows from the implicit function theorem that the following holds. and F is C 2 in a neighborhood of {0} × R. 2 Splitting Equations Let X and Y be real Banach spaces and let F be a mapping F :X ×R→Y (1) and let F satisfy the following conditions: F (0.
The types of linear operators Fu (0. F = F1 + F2 . Then there exists a closed subspace W of X and a closed subspace T of Y such that X Y = V ⊕W = Z ⊕ T. λ) = 0. • The range of L. F2 : X → T. λ) = 0. Hence Fu (0. u2 ∈ W. 2 Deﬁnition A linear operator L : X → Y is called a Fredholm operator provided: • The kernel of L. • The cokernel of L. see e. λ0 ) and using a Taylor expansion we may write F (u. is bijective and since T is closed it has a continuous inverse. λ0 ) is a bifurcation point for the equation F (u. F1 : X → Z. 0.28 1 Theorem If the point (0. The following lemma which is a basic result in functional analysis will be important for the development to follow. (9) (8) . λ) = F2 (u1 . λ0 ) we shall consider are socalled Fredholm operators. λ) = (7) We next let L = Fu (0. imL. λ) = F (0. We recall that W and Z are not uniquely given. λ0 )u + N (u. is closed in Y. is ﬁnite dimensional. u2 . λ0 ) restricted to W. Using Lemma 3 we may now decompose every u ∈ X and F uniquely as follows: u = u1 + u2 . λ). λ0 ) cannot be a linear homeomorphism of X e to Y. kerL. The operator Fu (0. λ0 )W is a linear homeomorphism of W onto T. 0. [38]. (5) then the Fr´chet derivative Fu (0. (6) Hence equation (5) is equivalent to the system of equations F1 (u1 .g. λ0 )W : W → T. Fu (0. λ0 ) + Fu (0. u2 . 3 Lemma Let Fu (0. λ0 ) be a Fredholm operator from X to Y with kernel V and cokernel Z. cokerL. or Lu + N (u. is ﬁnite dimensional. its proof may be found in any standard text. u1 ∈ V.
λ) = 0. Hence. (13) We note. that since Z is ﬁnite dimensional. SPLITTING EQUATIONS where N : X × R → Y. usually referred to as the set of bifurcation equations. which is also assumed ﬁnite dimensional!) Concerning equation (12) we have the following result. . is. This latter set of equations. That this may be done follows from the fact that at u1 = 0 and λ = λ0 equation (13) has the unique solution u2 = 0 and the Fr´chet derivative at this point e with respect to u2 is simply the identity mapping on W. λ) of equation (13) deﬁned for λ − λ0  + u1 < ǫ with u2 (u1 . the more diﬃcult part in the solution of equation (5). 4 Lemma Assume that Fu (0. λ) = 0 (14) for u1 . λ) = QF (u1 + u2 (u1 . The next sections present situations where these equations may be solved. whenever λ−λ0 + u1 < ǫ. Then there exist ǫ > 0. δ > 0 and a unique solution u2 (u1 . equation (12) is an equation in a ﬁnite dimensional space. u2 (u1 . LW : W → T has an inverse L−1 : T → W we obtain from equation (11) the equivalent system u2 + L−1 (I − Q)N (u1 + u2 . this equation will be a ﬁnite set of equations in ﬁnitely many variables (u1 ∈ V. λ) < δ. λ) = 0. u2 (u1 . Using the decomposition of X we may write equation (9) as Lu2 + N (u1 + u2 . λ) = 0. Proof. λ). hence if u2 can be determined as a function of u1 and λ. 29 (10) (11) Let Q : Y → Z and I − Q : Y → T be projections determined by the decomposition. We employ the implicit function theorem to analyze equation (13). λ). 5 Remark We note that in the above considerations at no point was it required that λ be a one dimensional parameter. then equation (10) implies that QN (u.2. (12) Since by Lemma 3. λ0 ) is a Fredholm operator with W nontrivial. we will have nontrivial solutions of equation (5) once we can solve F1 (u1 . using Lemma 4. This function solves the equation F2 (u1 . even though a ﬁnite set of equations in ﬁnitely many unknowns. λ) = 0. λ).
Since V is one dimensional u1 = αφ. λ0 )(φ. λ).(16) 2 (15) where R contains higher order remainder terms and all Fr´chet derivatives above e are evaluated at (0. λ) = 0. Using Taylor’s theorem we may write F (u. We let µ = λ − λ0 and deﬁne g(α. λ = λ(α). u2 (α. λ). 6 Theorem In the notation of the previous section assume that the kernel V and the cokernel Z of Fu (0. We hence need to solve for λ = λ(α) in the equation QF (αφ + u2 (α. u) + 2Fuλ (u. λ) + Fλλ (µ. Proof. We have the following theorem. u(α) = 0. Hence for α small and λ near λ0 there exists a unique u2 (α. λ0 ) both have dimension 1. Furthermore assume that the second Fr´chet e derivative Fuλ satisﬁes QFuλ (0. λ0 ) is a bifurcation point and there exists a unique curve u = u(α). λ) = 1 {QFuu (u. λ). µ) = QF (αφ + u2 (α. hence. deﬁned for α ∈ R in a neighborhood of 0 so that u(0) = 0. u) + 2QFuλ (u. µ)} + R. λ). Then (0. λ(0) = λ0 and F (u(α). by applying Q to (16) we obtain QF (u.30 3 Bifurcation at a Simple Eigenvalue In this section we shall consider the analysis of the bifurcation equation (14) in the particular case that the kernel V and the cokernel Z of Fu (0. λ(α)) = 0. λ) = Fu u + Fλ µ + 1 {Fuu (u. λ) such that F2 (αφ. α = 0. 2 (17) . 1) = 0. Then g maps a neighborhood of the origin of R2 into R. Let V = span{φ} and let Q be a projection of Y onto Z. λ) = 0. Because of (2) we have that Fλ and Fλλ in the above are the zero operators. λ0 ). µ)} + QR. λ0 ) both have dimension 1.
0) = QFuλ (0. λ) → u′′ + λ(u + u3 ). We next set u(α) = αφ + u2 (α. µ) = g(α. we choose X = C 2 [0. This proves the theorem. 0) is a bifurcation point for the ordinary diﬀerential equation u′′ + λ(u + u3 ) = 0 subject to the periodic boundary conditions u(0) = u(2π). Then F belongs to class C 2 with Fr´chet derivative e Fu (0.µ) α 31 = 1 2 QFuu (φ + + 2QFuλ (φ + u2 (α. µ) α + u2 (α. µ(α)) = 0. λ0 )(φ. λ0 + µ(α)). both equipped with the usual norms.3.λ) is bounded for α in a neighα borhood of 0. 7 Example The point (0. Y = C[0. 2π] ∩ {u : u(0) = u(2π). u′ (0) = u′ (2π). µ) = O(α). u′′ (0) = u′′ (2π)}. αφ α u2 (α. and ∂h(0. 1) = 0. Hence h(α. (21) (20) (19) . 2π] ∩ {u : u(0) = u(2π)}. The following example will serve to illustrate the theorem just established. λ = λ0 + µ(α).λ) . (18) It follows from Lemma 4 that the term u2 (α. λ)) 1 + α QR. λ0 )u = u′′ + λ0 u. ∂µ We hence conclude by the implicit function theorem that there exists a unique function µ = µ(α) deﬁned in a neighborhood of 0 such that h(α. 0) = 0. α We note that in fact h(0. as α → 0. To see this. The remainder formula of Taylor’s theorem implies a similar 1 statement for the term α QR. u′ (0) = u′ (2π). and F :X ×R→Y (u.λ) . BIFURCATION AT A SIMPLE EIGENVALUE and for α = 0 g(α.
0)(u. and hence the cokernel will have dimension 1 also. and hence. 0) if and only if 0 h(s)ds = 0. Applying Q we get Q1 = 1. · · · . . Computing further. 1) = 1. A projection Q : Y → 2π 1 Z then is given by Qh = 2π 0 h(s)ds. we ﬁnd that Fuλ (0. 1. 0)(1. 2π We see that h belongs to the range of Fu (0. Fuλ (0. λ) = λu. The kernel being one dimensional if and only if λ0 = 0. λ(α)). n = 0.32 This linear operator has a nontrivial kernel whenever λ0 = n2 . since we may choose φ = 1. We may therefore conclude by Theorem 6 that equation (19) has a solution u satisfying the boundary conditions (20) which is of the form u(α) = α + u2 (α.
(2). Rn ) ∩ C(Ω. Deﬁne k (1) • y ∈ Rn is such that (2) (3) (4) d(f. Rn ). We shall mainly follow the analytic development commenced by Heinz in [22] and Nagumo in [29]. then the equation f (x) = y has at most a ﬁnite number of solutions in Ω. (2). For a brief historical account we refer to [42].Chapter III Degree Theory 1 Introduction In this chapter we shall introduce an important tool for the study of nonlinear equations. 2 Deﬁnition of the Degree of a Mapping ¯ Let Ω be a bounded open set in Rn and let f : Ω → Rn be a mapping which satisﬁes • ¯ f ∈ C 1 (Ω. the degree of a mapping. y) = i=1 sgn det f ′ (xi ) 33 (5) . (3). 1 Proposition If f satisﬁes (1). 2 Deﬁnition Let f satisfy (1). y ∈ f (∂Ω). (3). / • if x ∈ Ω is such that f (x) = y then f ′ (x) = Df (x) is nonsingular. Ω.
t ≥ r > 0. s = 0.34 where x1 . φ(t) ≡ 0. In order to give this deﬁnition in the more general case we need a sequence of auxiliary results. φ(x)dx = 1. (2). s (10) . Ω. · · · . if det f ′ (xi ) > 0 ′ sgn det f (xi ) = −1. We note ﬁrst that it suﬃces to proof the lemma for functions f which are are of class C ∞ and for functions φ that vanish in a neighborhood of 0. ¯ The Brouwer degree d(f. ∞) → R be continuous and satisfy: φ(s) = 0. Then d(f. We also note that the function φ(f (x) − y) det f ′ (x) vanishes in a neighborhood of ∂Ω. hence we may extend that function to be identically zero outside Ω and φ(f (x) − y) det f ′ (x)dx = φ(f (x) − y) det f ′ (x)dx. which follows readily by making suitable changes of variables. x ∈ ∂Ω. s = 0. y) to be deﬁned for mappings f ∈ C(Ω. Ω. k. Let φ : [0. (3). y) = 0. xk are the solutions of (4) in Ω and +1. r ≤ s. (2). · · · . We let ψ(s) = s−n 0 ρn−1 φ(ρ)dρ. (3). The proof of the ﬁrst result. Rn ) which satisfy (2) will coincide with the number just deﬁned in case f satisﬁes (1). y) = Ω φ(f (x) − y)det f ′ (x)dx. will be left as an exercise. if det f ′ (xi ) < 0. 0 < s < ∞ 0. Ω ′ Ω′ where Ω is any domain with smooth boundary containing Ω. ∞) → R be continuous and satisfy φ(0) = 0. (8) φ(f (x) − y) det f ′ (x)dx = 0. 3 Lemma Let φ : [0. i = 1. If (4) has no solutions in Ω we let d(f. Ω. (9) Proof. and Then Ω ∞ 0 sn−1 φ(s)ds = 0. Rn (6) Let f satisfy the conditions (1). 4 Lemma Let f satisfy (1) and (2) and let r > 0 be such that f (x) − y > r. (7) provided r is suﬃciently small.
Rn (12) Ω have a common value. ∞). n. It follows that the functions g j (x) = ψ(x)xj . · · · . If we denote by aji (x) the cofactor of ∂f the element ∂xi in the Jacobian matrix f ′ (x). v2 . Put Lφ = 0 ∞ sn−1 φ(s)ds φ(x)dx Mφ = Rn Nφ = Ω φ(f (x) − y) det f ′ (x)dx. Proof. it follows that j div (aj1 (x). ∞). it vanishes in a neighborhood of 0 and in the interval [r.2. DEFINITION 35 Then ψ. φ(f (x) − y) det f ′ (x)dx φ(x)dx = 1. s ≥ r}. the integrals 0. We next deﬁne for i = 1. φ(s) ≡ 0 for s ≥ r. j = 1. aj2 (x). j = 1. vn ) has the property that divv = φ(f (x) − y) det f ′ (x). ajn (x)) = 0. n n (11) vi (x) = j=1 aji (x)g j (f (x) − y) and show that the function v = (v1 . · · · . φ(0) = Then for all such φ. where 0 < r ≤ minx∈∂Ω f (x) − y. Let Φ = {φ ∈ C([0. R) : φ(0) = 0. . so deﬁned is a C 1 function. 5 Lemma Let f satisfy (1) and (2) and let φ : [0. n the functions g j (f (x)−y) are C 1 functions which vanish in a neighborhood of ∂Ω. · · · . · · · . · · · . n belong to class C 1 and g j (x) = 0. and hence the result follows from the divergence theorem. x ≥ r. ∞) → R be continuous. φ(s) ≡ 0. and furthermore that for j = 1. Further ψ satisﬁes the diﬀerential equation sψ ′ (s) + nψ(s) = φ(s). · · · .
(2). (13) (14) . Rn ) ∩ C(Ω. Rn ) and ¯ fi (x) − fk (x) < ǫ. k = 1. N φ2 = N φ1 . 0 ≤ s ≤ 2ǫ 0 ≤ g(r) ≤ 1. Ω. 2ǫ ≤ r < 3ǫ g(r) = 0.e. x ∈ ∂Ω. 0). φ2 ∈ Φ with M φ1 = M φ2 = 1. f1 (x) − f2 (x) < ǫ. then L((Lφ2 )φ1 − (Lφ1 )φ2 ) = 0. assume that y = 0. 3ǫ ≤ r < ∞. 2. (3) and let ǫ > 0 be such that fi (x) − y > 7ǫ.36 Then L. N are linear functionals. i. Ω. 2. ∞) be such that g(s) = 1. We may. without loss. Proof. y) = d(f − y. x ∈ Ω (15) ¯ x ∈ Ω. It follows from Lemma 4 that M φ = 0 and N φ = 0. y). M. Consider f3 (x) = [1 − g(f1 (x))]f1 (x) + g(f1 (x))f2 (x). and N (φ2 − φ1 ) = 0. 3. since by Deﬁnition 2 d(f. let g ∈ C 1 [0. 6 Lemma Let f1 and f2 satisfy (1). It follows that (Lφ2 )(M φ1 ) − (Lφ1 )(M φ2 ) = 0. Ω. then ¯ f3 ∈ C 1 (Ω. y) = d(f2 . i. Ω. Let φ1 . then d(f1 . whenever Lφ = 0. i = 1. Lφ2 − Lφ1 = L(φ2 − φ1 ) = 0.
2. DEFINITION fi (x) > 6ǫ, x ∈ ∂Ω, i = 1, 2, 3.
37
Let φi ∈ C[0, ∞), i = 1, 2 be continuous and be such that φ1 (t) = 0, 0 ≤ t ≤ 4ǫ, 5ǫ ≤ t ≤ ∞ φ2 (t) = 0, ǫ ≤ t < ∞, φ2 (0) = 0 φi (x)dx = 1,
Rn
i = 1, 2.
We note that f3 ≡ f1 , if f1  > 3ǫ f3 ≡ f2 , if f1  < 2ǫ. Therefore
′ ′ φ1 (f3 (x))det f3 (x) = φ1 (f1 (x))det f1 (x)
(16)
′ ′ φ2 (f3 (x))det f3 (x) = φ2 (f2 (x))det f2 (x).
Integrating both sides of (16) over Ω and using Lemmas 4 and 5 we obtain the desired conclusion. 7 Corollary Let f satisfy conditions (1), (2), (3), then for ǫ > 0 suﬃciently small any function g which also satisﬁes these conditions and which is such that ¯ f (x) − g(x) < ǫ, x ∈ Ω, has the property that d(f, Ω, y) = d(g, Ω, y). Up to now we have shown that if f and g satisfy conditions (1), (2), (3) and if they are suﬃciently “close” then they have the same degree. In order to extend this deﬁnition to a broader class of functions, namely those which do not satisfy (3) we need a version of Sard’s Theorem (Lemma 8) (an important lemma of Diﬀerential Topology) whose proof may be found in [40], see also [44]. 8 Lemma If Ω is a bounded open set in Rn , f satisﬁes (1), (2), and E = {x ∈ Ω : det f ′ (x) = 0}. Then f (E) does not contain a sphere of the form {z : z − y < r}. This lemma has as a consequence the obvious corollary: 9 Corollary Let F = {h ∈ Rn : y + h ∈ f (E)}, f (x) = y + h, x ∈ Ω, h ∈ F (18) (17)
where E is given by (17), then F is dense in a neighborhood of 0 ∈ Rn and implies that f ′ (x) is nonsingular.
38 We thus conclude that for all ǫ > 0, suﬃciently small, there exists h ∈ F , 0 < h < ǫ, such that d(f, Ω, y + h) = d(f − h, Ω, y) is deﬁned by Deﬁnition 2. It also follows from Lemma 6 that for such h, d(f, Ω, y + h) is constant. This justiﬁes the following deﬁnition. 10 Deﬁnition Let f satisfy (1) and (2). We deﬁne d(f, Ω, y) = lim d(f − h, Ω, y). h→0 h∈F (19)
Where F is given by (18) and d(f − h, Ω, y) is deﬁned by Deﬁnition 2. ¯ We next assume that f ∈ C(Ω, Rn ) and satisﬁes (2). Then for ǫ > 0 suﬃ¯ ciently small there exists g ∈ C 1 (Ω, Rn ) C(Ω, Rn ) such that y ∈ g(∂Ω) and f − g = max f (x) − g(x) < ǫ/4
¯ x∈Ω
and there exists, by Lemma 8, g satisfying (1), (2), (3) such that g − g < ˜ ˜ ˜ h ǫ/4 and if h satisﬁes (1), (2), (3) and g − ˜ < ǫ/4, g − h < ǫ/2, then ˜ ˜ ˜ d(˜, Ω, y) = d(h, Ω, y) provided ǫ is small enough. Thus f may be approximated g by functions g satisfying (1), (2), (3) and d(˜, Ω, y) = constant provided f − g ˜ g ˜ is small enough. We therefore may deﬁne d(f, Ω, y) as follows. ¯ 11 Deﬁnition Let f ∈ C(Ω, Rn ) be such that y ∈ f (∂Ω). Let d(f, Ω, y) = lim d(g, Ω, y)
g→f
(20)
where g satisﬁes (1), (2), (3). The number deﬁned by (20) is called the Brouwer degree of f at y relative to Ω. It follows from our considerations above that d(f, Ω, y) is also given by formula (7), for any φ which satisﬁes: φ ∈ C([0, ∞), R), φ(0) = 0, φ(s) ≡ 0, s ≥ r > 0, where r < min f (x) − y.
x∈∂Ω
φ(x)dx = 1,
Rn
3
Properties of the Brouwer Degree
We next proceed to establish some properties of the Brouwer degree of a mapping which will be of use in computing the degree and also in extending the deﬁnition to mappings deﬁned in inﬁnite dimensional spaces and in establishing global solution results for parameter dependent equations.
3. PROPERTIES OF THE BROUWER DEGREE
39
¯ 12 Proposition (Solution property) Let f ∈ C(Ω, Rn ) be such that y ∈ f (∂Ω) and assume that d(f, Ω, y) = 0. Then the equation f (x) = y has a solution in Ω. The proof is a straightforward consequence of Deﬁnition 11 and is left as an exercise. ¯ 13 Proposition (Continuity property) Let f ∈ C(Ω, Rn ) and y ∈ Rn be such ¯ that d(f, Ω, y) is deﬁned. Then there exists ǫ > 0 such that for all g ∈ C(Ω, Rn ) and y ∈ R with f − g + y − y  < ǫ ˆ ˆ d(f, Ω, y) = d(g, Ω, y ). ˆ The proof again is left as an exercise. The proposition has the following important interpretation. ¯ 14 Remark If we let C = {f ∈ C(Ω, Rn ) : y ∈ f (∂Ω)} then C is a metric space / with metric ρ deﬁned by ρ(f, g) = f − g . If we deﬁne the mapping d : C → N (integers) by d(f ) = d(f, Ω, y), then the theorem asserts that d is a continuous function from C to N (equipped with the discrete topology). Thus d will be constant on connected components of C. Using this remark one may establish the following result. ¯ 15 Proposition (Homotopy invariance property) Let f, g ∈ C(Ω, Rn ) with ¯ → Rn be continuous such f (x) and g(x) = y for x ∈ ∂Ω and let h : [a, b] × Ω that h(t, x) = y, (t, x) ∈ [a, b] × ∂Ω. Further let h(a, x) = f (x), h(b, x) = g(x), ¯ x ∈ Ω. Then d(f, Ω, y) = d(g, Ω, y); more generally, d(h(t, ·), Ω, y) = constant for a ≤ t ≤ b. The next corollary may be viewed as an extension of Rouch´’s theorem concerne ing the equal number of zeros of certain analytic functions. This extension will be the content of one of the exercises at the end of this chapter. ¯ 16 Corollary Let f ∈ C(Ω, Rn ) be such that d(f, Ω, y) is deﬁned. Let g ∈ ¯ Rn ) be such that f (x) − g(x) < f (x) − y, x ∈ ∂Ω. Then d(f, Ω, y) = C(Ω, d(g, Ω, y). Proof. For 0 ≤ t ≤ 1 and x ∈ ∂Ω we have that y − tg(x) − (1 − t)f (x) = (y − f (x)) − t(g(x) − f (x)) ≥ y − f (x) − tg(x) − f (x) (21)
Rn ) and let K be a closed ¯ subset of Ω such that y ∈ f (∂Ω ∪ K). Then d(f. Ω2 . y) = d(f1 . f1 and f2 satisfy also (1) and (3) (interpreted appropriately). · · · . x2 ∈ Rq . p + q = n. y) = i=1 d(f. Ω2 . 18 Proposition (Additivity property) Let Ω be a bounded open set which is ¯ the union of m disjoint open sets Ω1 . (22) Proof. y) if the degree is deﬁned. Using an approximation argument. ¯ 19 Proposition (Excision property) Let f ∈ C(Ω. y2 ). i = 1. Then / d(f. Suppose that f (x) = ¯ ¯ (f1 (x1 ). Rn ) and y ∈ Rn be such that y ∈ f (∂Ωi ).40 > 0 since 0 ≤ t ≤ 1. f2 (x2 )) where f1 : Ω1 → Rp . Suppose n y = (y1 . f2 : Ω2 → Rq are continuous. Ω. Ωm . Ω. y1 )d(f2 . For such functions we have d(f. 2 2 = i=1 xi ∈f −1 (yi ) i sgn det fi′ (xi ) = d(f1 . Ω. the degree only depends on the boundary data. . y). we may assume that f. y) = x∈f −1 (y) sgn det f ′ (x) ′ f1 (x1 ) 0 ′ f2 (x2 ) = x∈f −1 (y) = sgn det 0 ′ ′ sgn det f1 (x1 ) sgn det f2 (x2 ) xi ∈ f −1 (yi ) i = 1. i = 1. Ω.e. 1] × Ω → Rn given by h(t. x2 ). Ω \ K. 20 Proposition (Cartesian product formula) Assume that Ω = Ω1 × Ω2 is a bounded open set in Rn with Ω1 open in Rp and Ω2 open in Rq . m. Ω. For x ∈ Rn write x = (x1 . To give an example to show how the above properties may be used we prove Borsuk’s theorem and the Brouwer ﬁxed point theorem. i. Ω1 . x1 ∈ Rp . Then m d(f. · · · . y) = d(f. y1 )d(f2 . Ωi . y) = d(g. then d(f. x ∈ ∂Ω. x) = tg(x) + (1 − t)f (x) satisﬁes the conditions of Proposition 15 and the conclusion follows from that proposition. Ω. 2. ¯ hence h : [0. Ω1 . As an immediate corollary we have the following: 17 Corollary Assume that f and g are mappings such that f (x) = g(x). y2 ). and let f ∈ C(Ω. y2 ) ∈ R is such that yi ∈ fi (∂Ωi ). y).
Let α : Rn → R be a continuous function such that α(x) ≡ 1. x ∈ ∂Ω. h(x) = h(x). 0 ∈ Θ.3. Θ. and hence conclude that ˜ arguments to conclude that d(h. i. Ω. 0) = d(h. Ω. 0) = d(h. x ∈ Rn and put g(x) = α(x)x + (1 − α(x))f (x) h(x) = 1 [g(x) − g(−x)] . 0) + d(h. be such that f : Ω → Ω. 0) is an ¯ even integer. 0). where ∂Bǫ (0) has been excised. ¯ 22 Theorem (Brouwer ﬁxed point theorem) Let f ∈ C(Ω. / Proof. x ≤ ǫ. Rn ) is odd and h(x) = f (x). Rn ). if x ∈ Ω. f (x) = −f (−x)). Bǫ (0). Θ2 . 0). Θ. x ∈ ∂Ω.e. 0). then by the additivity property ˜ ˜ ˜ d(h. Thus by Corollary 16 and the remark following it d(f. Since Θ is symmetric. 2 ¯ then h ∈ C(Ω. for those x ∈ Θ with xn = 0. 0) = d(h. 0) = d(h. 0) where we have used the excision property. 0 ≤ α(x) ≤ 1. x ∈ ∂Θ ˜ and is such that h(x) = 0. (23) . and h is odd one may / / ˜ ˜ ¯ show (see [40]) that there exists h ∈ C(Ω. Ω. Bǫ (0). Ω = {x ∈ ¯ ¯ Rn : x < 1}. Θ \ {x : xn = 0}. Let 0 ∈ f (∂Ω). 0) = d(h. Θ. Rn ) be an odd mapping (i. PROPERTIES OF THE BROUWER DEGREE 41 3. x < ǫ.e. Choose ǫ > 0 such that Bǫ (0) = {x ∈ Rn : x < ǫ} ⊂ Ω. 0) = d(h. α(x) = 0. h(x) = x. Hence ˜ ˜ d(h. Θ1 . 0). 0 ∈ h(∂Θ). Then f has a ﬁxed point in Ω. Rn ) which is odd. Ω. then d(f. the integer given by (23) is even. It follows from Deﬁnition 2 that d(h. Θ2 . 0) = 1. Ω \ Bǫ (0). ˜ Since Θ2 = {x : −x ∈ Θ1 } and h is odd one may now employ approximation ˜ Θ1 .e. then −x ∈ Ω) and let f ∈ C(Ω. Θ \ {x : xn = 0}. On the other hand. there ¯ exists x ∈ Ω such that f (x) = x. 0) + d(h. We let Θ1 = {x ∈ Θ : xn > 0}. the excision and additivity property imply that d(h. Θ2 = {x ∈ Θ : xn < 0}. and it therefore suﬃces to show that (letting Θ = Ω \ Bǫ (0)) d(h. 0) is an odd integer.1 The theorems of Borsuk and Brouwer 21 Theorem (Borsuk) Let Ω be a symmetric bounded open neighborhood of ¯ 0 ∈ Rn (i.
since y ∈ f (∂Ω). Consider the mapping ¯ ˜ T F T −1 : T (Ω ∩ E) → Rn . 0) by the homotopy property. Then d(f . / (25) ˜ ˜ Select a basis e1 . 0). 0) = 1 it follows from the solution property that the equation x − f (x) = 0 has a solution in Ω. Ω0 . Let h(t. Assume f has no ﬁxed points in ∂Ω. Ω. 0 ≤ t ≤ 1. then. Ω. · · · . x ∈ ∂Ω and thus d(h(t. / ˜ ˜ Let Ω0 denote the bounded open set T (Ω ∩ E) in Rn and let f = T f T −1. y0 = ˜ T (y).20). . 0). 0 ≤ t ≤ 1. · · · . Theorem 22 remains valid if the unit ball of Rn is replaced by any set homeomorphic to the unit ball (replace f by g −1 f g where g is the homeomorphism). cn ) ∈ Rn . y0 ) is deﬁned. en of E and deﬁne the linear homeomorphism T : E → Rn by n T i=1 ci e i = (c1 . The mapping f (x) = x + F (x) = (id + F )(x) (24) is called a ﬁnite dimensional perturbation of the identity in E. That the Theorem also remains valid if the unit ball is replaced in arbitrary compact convex set (or a set homeomorphic to it) may be proved using the extension theorem of Dugundji (Theorem I. Then h(t. ˜ Let y be a point in E and let E be a ﬁnite dimensional subspace of E ¯ and assume containing y and F (Ω) y ∈ f (∂Ω). Ω. 0) = d(h(0. It is an easy exercise in linear algebra to show that the following lemma holds. it follows that / ˜ T (y) ∈ T f T −1(T (∂Ω ∩ E)). x) = 0. Since d(id. x) = x − tf (x). Let F : Ω → E be continuous and let F (Ω) be contained in a ﬁnite dimensional subspace of E.1 Completely Continuous Perturbations of the Identity in a Banach Space Deﬁnition of the degree Let E be a real Banach space with norm · and let Ω ⊂ E be a bounded ¯ ¯ open set.42 Proof. 4 4.
x) = y it follows that = ≤ < f (x) − y = f (x) − th1 (x) − (1 − t)h2 (x) t(f − (x) − h1 (x)) + (1 − t)(f (x) − h2 (x)) t f (x) − h1 (x) + (1 − t) f (x) − h2 (x) inf x∈∂Ω f (x) − y (see (26)). Let E be a ﬁnite dimensional subspace contain/ ¯ ˜ ing y and (hi − id)(Ω). Ω. 0 ≤ t ≤ 1. Let k(t. Ω.. Ω0 . Hence we may use the homotopy invariance property of Brouwer degree to conclude that ˜ d(T k(t. More generally if D is any subset of E and F ∈ C(D. x ∈ T (∂Ω ∩ E). ¯ 24 Lemma Let f : Ω → E be a completely continuous perturbation of the identity and let y ∈ f (∂Ω). T (y)) = constant. then an integer valued function d(f. Ω0 . i. / / Proof. Then there exists an integer d with the following property: / ¯ If h : Ω → E is a ﬁnite dimensional continuous perturbation of the identity such that x∈Ω sup f (x) − h(x) < inf x∈∂Ω f (x) − y . x) = th1 (x) + (1 − t)h2 (x). y0 ). x ∈ Ω. y0 ) calculated above is independent of the choice ˜ ¯ the ﬁnite dimensional space E containing y and F (Ω) and the choice of basis of ˜ E.e. Ω. ·) − id)(Ω) is contained in E and T k(t. We shall now demonstrate that if f = id + F . Ω. y). E). ˜ from which follows that x ∈ ∂Ω.e. In order to accomplish this we need the following lemma. COMPLETELY CONTINUOUS PERTURBATIONS 43 ˜ 23 Lemma The integer d(f . with F completely continuous (f is called a completely continuous perturbation of the identity) and y ∈ f (∂Ω). y0 are as above. ¯ Recall that a mapping F : Ω → E is called completely continuous if F is ¯ continuous and F (Ω) is precompact in E (i. then F is called completely continuous if F (V ) is precompact for any bounded subset V of D. (26) then y ∈ h(∂Ω) and d(h. Then (k(t. then if ¯ t ∈ (0. y) = d(h2 . y) = d(T hi T −1 . then d(hi . ·)T −1(x) = ˜ T (y). ˜ where f . Ω0 . Let h1 and h2 be any two such ¯ mappings. d(h1 . T (∂Ω ∩ E). y) = d. y) (the Leray Schauder degree of f at y relative to Ω) may be deﬁned having much the same properties as the Brouwer degree. F (Ω) is compact). ·)T −1 . 1) and x ∈ Ω are such that k(t. T (Ω ∩ E). Ω. That y ∈ h(∂Ω) follows from (26). y) = d(f . Ω.4. We hence may deﬁne ˜ d(f. where T ¯ ˜ is given as above. T (y)). .
y) = d. 27 Deﬁnition The integer d whose existence has been established by Lemma 26 is called the LeraySchauder degree of f relative to Ω and the point y and is denoted by d(f. λi (y) is non–negative and continuous on n M and further i=1 λi (y) = 1. Deﬁne µi : M → [0. y ∈ M. yn } ⊂ ¯ (f − id)(Ω). and the mapping id + Pǫ F is a continuous ﬁnite dimensional perturbation of the identity. · · · .44 It follows from Lemma 24 that if a ﬁnite dimensional perturbation of the identity h exists which satisﬁes (26). Ω. Ω. Let Pǫ / be a Schauder projection operator determined by ǫ and points {y1 . The properties of Pǫ (cf Lemma 25) imply that for each x ∈ Ω Pǫ F (x) − F (x) ≤ ǫ inf x∈∂Ω f (x) − y . ¯ Proof. · · · . Such an operator has the following properties: 25 Lemma • Pǫ : M → co{y1 . otherwise y − yi ≤ ǫ Since not all µi vanish simultaneously. yn ) is continuous. y). where d is the integer whose existence follows from this lemma. if 0. In order to accomplish this we need an approximation result. Then for every ǫ > 0 there exists a ﬁnite covering of M by spheres of radius ǫ with centers at y1 . n j=1 µj (y) 1 ≤ i ≤ n. ǫ − y − yi . yn } (the convex hull of y1 . y) = d. Then d(id + Pǫ F. • Pǫ y − y ≤ ǫ. Let y ∈ f (∂Ω). Ω. yn ∈ M . and y1 . . ∞) by µi (y) = = and let λi (y) = µi (y) . yn . where d is the integer whose existence is established by Lemma 24. · · · . · · · . Let ǫ > 0 be such that ǫ < inf x∈∂Ω f (x) − y . Let M be a compact subset of E. The operator Pǫ deﬁned by n Pǫ (y) = i=1 λi (y)yi (27) is called a Schauder projection operator on M determined by ǫ. · · · . ¯ 26 Lemma Let f : Ω → E be a completely continuous perturbation of the identity. then we may deﬁne d(f. • Pǫ (M ) is contained in a ﬁnite dimensional subspace of E.
We relabel the subsequence and call it again {xn }∞ . y) = d(id + Tn Pǫn F Tn . since y ∈ f (∂Ω) we have that x ∈ Ω. Proof.4. 0) is an odd integer. and excision properties similar to the Brouwer degree. the Cartesian product formula also holds. We let ǫ > 0 be given and choose N such that n. Ω. Let ǫ > 0 be such that ǫ < inf x∈∂Ω f (x) and let Pǫ be an associated Schauder projection operator. / 4. y a point in E with y ∈ f (∂Ω) and d(f.3 Borsuk’s theorem and ﬁxed point theorems 29 Theorem (Borsuk’s theorem) Let Ω be a bounded symmetric open neigh¯ borhood of 0 ∈ E and let f : Ω → E be a completely continuous odd perturbation of the identity with 0 ∈ f (∂Ω). Tn (y)). We let Ω be a bounded open set in E and f : ¯ Ω → E be a completely continuous perturbation of the identity. To see this let {ǫn }∞ be a decreasing sequence of positive n=1 numbers with limn→∞ ǫn = 0 and ǫ1 < inf x∈∂Ω f (x)−y . continuity. Ω. · · · . It now follows that x ∈ Ω and solves the equation f (x) = y. Ω. Ω. y). hence has a limit. since F is n=1 n=1 completely continuous there exists a subsequence {xni }∞ such that F (xni ) → i=1 ¯ u ∈ F (Ω). / Proof. COMPLETELY CONTINUOUS PERTURBATIONS 45 4. ˜ and the spaces En are ﬁnite dimensional. We claim that the equation f (x) = y has / a solution in E. Let fǫ = id + Pǫ f and put h(x) = 1/2[fǫ(x) − .2 Properties of the degree 28 Proposition The Leray–Schauder degree has the solution. ˜ or equivalently a solution xn ∈ Ω ∩ En of x + Pǫn F (x) = y. n = 1. The sequence {xn }∞ therefore n=1 ¯ is a Cauchy sequence. Thus xn − xm < ǫ. homotopy invariance. Then n=1 ≤ + xn − xm = Pǫn F (xn ) − Pǫm F (xm ) Pǫn F (xn ) − F (xn ) + Pǫm F (xm ) − F (xm ) F (xn ) − F (xm ) < ǫn + ǫm + F (xn ) − F (xm ) . Ω. y) = 0. thus. hence. Then d(f. The sequence {xn }∞ is a bounded sequence ({xn }∞ ⊂ Ω). Hence the solution property of ˜ Brouwer degree implies the existence of a solution zn ∈ Tn (Ω ∩ En ) of the equation −1 z + Tn Pǫn F Tn (z) = T (y). Then d(f. additivity. say x. (solution property). Tn (Ω ∩ En ). ǫm < ǫ/3 and F (xn ) − F (xm ) < ǫ/3. where −1 ˜ d(id + Pǫn F. y) = d(id + Pǫn F. 2. m ≥ N imply ǫn . Let Pǫi be associated Schauder projection operators.
Br (0).20) we may continu˜ ˜ ously extend F to Br (0). Hence F is completely continuous. Br (0). x ∈ Br (0). i. Call the extension F . 0) = d(T hT −1 . [38]. Then F has a ﬁxed point in K. Since K is a compact there exists r > 0 such that K ⊂ Br (0) = {x ∈ E : x < r}. 0 ≤ t ≤ 1. T (Ω ∩ E). bounded. Therefore F has a ﬁxed point in K. 0 ≤ t ≤ 1. 0). Then h is a ﬁnite dimensional perturbation of the identity which is odd and ˜ Thus d(f. Using the Extension Theorem (Theorem I. ˜ Since tF (K) ⊂ tK ⊂ Br (0). ·). the smallest convex set containing ˜ F (K). 0) = 1. but h(x) − f (x) ≤ ǫ. i. 30 Theorem (Schauder ﬁxed point theorem) Let K be a compact convex subset of E and let F : K → K be continuous. 31 Theorem (Schauder) Let K be a closed. Hence by the homotopy invariance property of the Leray–Schauder degree constant = d(k(t. In such a case the following result may be applied. 0 ≤ t ≤ 1. Then since F (K) is compact it follows from a ˜ ˜ theorem of Mazur (see eg. [17]. Ω.46 fǫ (−x)]. Ω. In many applications the mapping F is known to be completely continuous but it is diﬃcult to ﬁnd a compact convex set K such that F : K → K. ˜ On the other hand T (Ω ∩ E) is a symmetric bounded open neighborhood of ˜ and T hT −1 is an odd mapping. it follows that k(t. and hence in K. hence the result follows from 0 ∈ T (Ω ∩ E) Theorem 21. where coF (K) is the convex hull of F (K).e. Proof. x) = 0. Consider the homotopy ˜ k(t. 0). whereas closed convex sets K having this property are more easily found. x ∈ ∂Br (0). Then F (Br (0)) ⊂ coF (K) ⊂ K. Let K = coF (K). The solution property of degree therefore implies that the equation ˜ x − F (x) = 0 x − F (x) = 0. Then F has a ﬁxed point in K. We next establish extensions of the Brouwer ﬁxed point theorem to Banach spaces. and [43]) that K is compact and K ⊂ K. ˜ ˜ ˜ Thus F : K → K and F has ﬁxed point in K by Theorem 30. ˜ Proof. 0) = d(h. has a solution in x ∈ Br (0).e. . convex subset of E and let F be a completely continuous mapping such that F : K → K. ˜ d(h. Ω ∩ E. x) = x − tF (x). 0) = d(id.
Let f (z) = 0. Rn ) with f (x) = 0 = g(x). g(S n−1 ) ⊂ S n−1 and f (x) − g(x) < 2. g ∈ C(B n . Show there exists λ (λ = 0) ∈ R and x ∈ ∂Ω such that f (x) = λx. Ω. x x f (x). b). . (This is commonly called the hedgehog theorem. Show that d(f. counting multiplicities. Let Ω be as in Exercise 4 and let f ∈ C(Ω. This result is called Rouch´’s theorem. Let Ω be a bounded open subset of Rn and let ¯ f. B n .5. x ∈ ∂Ω. B n . 0) = d(g. Let B n = {x ∈ Rn : x < 1}. (a. e 3. 7. Let Ω ⊂ Rn be a bounded open neighborhood of 0 ∈ Rn . Let f. Let f be as in Exercise 6 and assume that f (S n−1 ) does not equal S n−1 . then d(f. let f ∈ C(Ω. x ∈ ∂Ω x Show that the equation f (x) = 0 has a solution in Ω. EXERCISES 47 5 Exercises 1. Identify R2 with the complex plane. ¯ 5. z ∈ ∂Ω. Let Ω be a bounded open subset of R2 and let f and g be functions which are analytic in Ω and continuous on ¯ Ω. Ω. Rn ) be such that 0 ∈ f (∂Ω) and either / f (x) = or f (x) = − x f (x). (i) If f (b) > 0 > f (a). 4. Show that f and g have precisely the same number of zeros. x ∈ ∂Ω. 0) = 1 (iii) If f (a)f (b) > 0. 0). 0) = 0. Further assume that −g(x) f (x) = . then d(f. z ∈ ∂Ω and assume that f (z) − g(z) < f (z). Let [a. 0) = d(g. in Ω. g ∈ C(Ω. Rn ) be such that f (S n−1 ). S n−1 = ∂B n . B n . then d(f. b] → R be a continuous function such that f (a)f (b) = 0. b). x ∈ ∂Ω.) 6. 0) = −1 2. Verify the following. b). Rn ) be such that 0 ∈ f (∂Ω). b] be a compact interval in R and let f : [a. f (x) g(x) Show that d(f. (a. Show that d(f. 0). x ∈ S n−1 . 0) = 0. / Let n be odd. (a. (ii) If f (b) < 0 < f (a).
where m < n. Also assume that / f (−x) f (x) = . using linear algebra methods. ¯ 12. Show. Rn ) be such that 0 ∈ f (∂Ω). 15. Provide detailed proofs of the results of Section 3. convex subset of E. Let Ω be a bounded open set in E with 0 ∈ Ω. f (x) f (−x) Show that d(f. Let F : Ω → E be completely continuous and satisfy x − F (x) 2 ≥ F (x) 2 . ¯ 13. where β equals the sum of the algebraic multiplicities of all real eigenvalues µ of A with µ > 1. that d (id − A. . Let K be a bounded. Let Ω be as in Exercise 9 and let f ∈ C(Ω. Then F has a ﬁxed point in K. Let Ω ⊂ Rn be a symmetric bounded open neighborhood of 0 ∈ Rn and ¯ let f ∈ C(Ω. 9. Ω. open. Let F : K → E be completely continuous and be such that F (∂K) ⊂ K. Let f and Ω be as in Exercise 10 except that f is not necessarily odd. Provide detailed proofs of the results of Section 4. 0) is an odd integer.48 8. Show there exists x ∈ ∂Ω such that f (x) = 0. x ∈ ∂Ω. Ω. 14. Show there exists x ∈ ∂Ω such that f (x) = f (−x). Rn ) be an odd function such m that f (∂Ω) ⊂ R . 0) = (−1)β . x ∈ ∂Ω. Let Ω be a bounded open neighborhood of 0 ∈ Rn . Let A be an n × n real matrix for which 1 is not an eigenvalue. 11. ¯ 10. ¯ then F has a ﬁxed point in Ω.
λ). d(f (·. λ) = u − F (u. Oλ . b]). Then for a ≤ λ ≤ b.Chapter IV Global Solution Theorems 1 Introduction In this chapter we shall consider a globalization of the implicit function theorem (see Chapter I) and provide some global bifurcation results. (here Oλ = {u ∈ E : (u. Let O be a bounded open (in the relative topology) subset of E × [a. As will be seen in later sections. and let ¯ F :O→E be a completely continuous mapping. 0) = constant. this result also allows us to derive a globalization of the implicit function theorem and results about global bifurcation in nonlinear equations. λ) = 0. Let f (u. b] . 2 The Continuation Principle of LeraySchauder In this section we shall extend the homotopy property of LeraySchauder degree (Proposition III. 49 (2) (1) . 1 Theorem (The generalized homotopy principle) Let f be given by (1) and satisfy (2). λ) ∈ ∂O (here ∂O is the boundary of O in E × [a.28) to homotopy cylinders having variable cross sections and from it deduce the LeraySchauder continuation principle. λ) and assume that f (u. Our main tools in establishing such global results will be the properties of the Leray Schauder degree and a topological lemma concerning continua in compact metric spaces. λ) ∈ O}). (u. where E is a real Banach space.
λ) = (u − F (u. We may assume that O = ∅ and that a = inf{λ : Oλ = ∅}.28) we therefore conclude that ˜ ˆ ˜ ˆ ˜ ˆ d(f1 . 0) = d(f (·. a] ∪ Ob × [b.28). λ) ∈ O : f (u. b + ǫ). Let F be the extension of F to E × R whose existence is guaranteed by the Dugundji extension theorem (Theorem I. O. 0). Then f is a completely continuous perturbation of the identity in E × R. 0) is deﬁned and constant (for such λ∗ ). b + ǫ). O. 0) = 0. We let ˆ ˜ where ǫ > 0 is ﬁxed. b] and let f : O → E be given by (1) and satisfy (2). O. ˜ ˆ f (u. 1]. Let ¯ S = {(u. λ) = 0. λ − λ∗ ). By the homotopy ˆ imply that ft invariance principle (Proposition III. ˜ ˆ ˜ d(f0 . a). 0) = d(f0 . we obtain ˜ d(f0 . our hypotheses ˜ (u. by the excision property of degree (Proposition III. 0) = d(f . λ) = 0}. λ∗ ). Let ˜ where a ≤ λ∗ ≤ b is ﬁxed. b = sup{λ : Oλ = ∅}. Furthermore for any such λ∗ ˜ ˆ and hence d(f . λ − λ∗ ). 2 Theorem (Leray–Schauder Continuation Theorem) Let O be a bounded ¯ open subset of E × [a. λ) = (u − tF (u. Oa . λ) = 0 for (u. b + ǫ). λ∗ ). On the other hand. λ) ∈ ∂ O. Then there exists a closed connected set C in S such that Ca ∩ Oa = ∅ = Cb ∩ Ob . and consider the vector ﬁeld ˜ ˜ then ft (u. O. As an immediate consequence we obtain the continuation principle of LeraySchauder. Thus. .28). λ) ∈ ∂ O and t ∈ [0. λ) − (1 − t)F (u. Oλ∗ . Then O is a bounded open subset of E × R. 0). (u. Furthermore assume that d(f (·.50 Proof. This completes the proof. ˆ O = O ∪ Oa × (a − ǫ. λ). Using the Cartesian product formula (Proposition III. Oλ∗ × (a − ǫ. ˜ ˜ f (u. Let 0 ≤ t ≤ 1. λ) = 0 if and only if λ = λ∗ and u = F (u. ˜ ˜ ˜ ft (u. 0) = d(f0 . 0).20). Oλ∗ × (a − ǫ. λ∗ ). O.
We now apply Whyburn’s lemma (see [44]) with X = S. b] such that A ⊂ U ∩ O = V and S ∩ ∂V = ∅ = Vb . a). Consider the nonlinear Dirichlet problem u′′ + g(x. λ). XA ∩ XB = ∅. Oa . Therefore d(f (·. b). We hence may ﬁnd an open set U ⊂ E × [a. CONTINUATION PRINCIPLE Proof. some basic existence results for the existence of solutions of nonlinear boundary value problems. 3 Example Let I = [0. XB in X such that A ⊂ XA . u) = u = Let G be deﬁned by G(u)(x) = g(x. the excision principle implies that d(f (·. Proof. x ∈ I. 0) = d(f (·. 1] × R → R be continuous. (4) . as an application of the above results. 0. Since Vb = ∅. Ob . In the following examples we shall develop. Oa . b). we consider the one parameter family of problems u′′ + λg(x. 0) = d(f (·. If there is no such continuum (as asserted above) there will exist compact sets XA . Va . 1]. a). x ∈ Ω.2. u) = 0. λ ≥ a. b]. R) such that a < u(x) < b. On the other hand. 0). 1] and let g : [0. u(x)). a). 0) = constant. 51 Using the complete continuity of F we may conclude that S is a compact metric subspace of E × [a. these equalities yield a contradiction. Hence Sa × {a} = A = ∅ = B = Sb × {b}. u = 0. B ⊂ XB . It follows from Theorem 1 that d(f (·. in I on ∂I. 0. Then (3) has a solution u ∈ C 2 ([0. and there exists a continuum as asserted. Vλ . a) > 0 > g(x. XA ∪ XB = X. in I on ∂I. To see this. 0). (3) Let there exist constants a < 0 < b such that g(x.
u ∈ C([0. λ) = 0. = 0. 1]. R) = E. (7) We shall now show that condition (7) alone suﬃces to guarantee that equation (6) has a global solution branch in the half spaces E × [λ0 . λ0 ]. 0) = d(id. and if O is an isolating neighborhood. x ∈ I.52 then (4) is equivalent to the operator equation u = λLG(u). 1]. LG(v) ∈ C 2 (I) and since C 2 (I) is compactly embedded in E that LG(·) : E → E is a completely continuous operator. Then there is a solution curve {(u(λ).12) hold at (u0 . via elementary calculus. λ) = u − F (u. and Theorem 2 implies the existence of a continuum C of solutions of (5). O0 . then there will either exist x ∈ I such that u(x) = b or there exists x ∈ I such that u(x) = a and λ > 0. In either case. hence of (4). λ0 ). 3 A Globalization of the Implicit Function Theorem Assume that F :E×R→ E is a completely continuous mapping and consider the equation f (u. . Then O is an open and bounded set in E × [0. 0 ≤ λ ≤ 1}. a < u(x) < b. we have that d(f (·. λ)} of (6) deﬁned in a neighborhood of λ0 . λ0 ) be a solution of (6) such that the condition of the implicit function theorem (Theorem I. λ0 ). 0) = 0. passing through (u0 .12 imply that the solution u0 is an isolated solution of (6) at λ = λ0 . Oλ . where for each v ∈ E. ∞) and E × (−∞. w = LG(v) is the unique solution of w′′ + g(x. v) w = 0. such that C ∩ E × {0} = {0} and C ∩ E × {1} = ∅. λ) : u ∈ E. in I on ∂I. O. Therefore d(id − λLG. a contradiction. 0) = 1. λ) ∈ ∂O is a solution of (4). (5) It follows that for each v ∈ E. (6) Let (u0 . If (u. Let O = {(u. λ0 ). Hence (4) has no solutions in ∂O. Furthermore the conditions of Theorem I. (3) yields.
. λ0 ). and hence is a compact metric 2R space. Let S+ = {(u. λ0 ). ∞) such that Uλ0 = O. or else there exists one such continuum which meets the λ = λ0 hyperplane outside the set O. 0) which is nonzero. In either case. λ) ∈ E × [λ0 . Proof. λ) ∈ C + we have that u + λ < R. Uλ0 . Let then S+ is a compact subset of E × [λ0 . If one only assumes (7). we may ﬁnd a bounded open set / 2R U ⊂ E × [λ0 . Uλ0 . ∞) (C − is unbounded in E ×(−∞. was made for convenience of proof. 0) = 0. λ) solves (6)}. λ) solves (6)} and Then there exists a continuum C + ⊂ S+ (C − ⊂ S− ) such that: S− = {(u. λ0 ). O. − + 1. 0) = constant. and the generalized homotopy principle that ˜ d(f (·. λ0 ). 5 Remark The assumption of Theorem 4 that u0 is the unique solution of (6) inside the set O. λ) ∈ S+ : u + λ ≤ 2R}. where this constant is given by d(f (·. there exists λ∗ > λ0 such that Uλ∗ contains no solutions of (6) and hence d(f (·. Then there exists a constant R > 0 such that for each (u. C + ⊂ U . λ) ∈ C + . we employ again Whyburn’s lemma ([44]). λ0 ]) − + or Cλ0 ∩ (E\O) = ∅ (Cλ0 ∩ (E\O) = ∅). λ0 ] : (u. λ) ∈ E × (−∞. 0) = −d(f (·.3. ∞). then it follows from the excision property. Whyburn’s lemma. A GLOBALIZATION OF THE IMPLICIT FUNCTION THEOREM 53 4 Theorem Let O be a bounded open subset of E and assume that for λ = λ0 equation (6) has a unique solution in O and let (7) hold. C + is either unbounded in E ×[λ0 . Assume that C + ∩ (E\O) = ∅ and that C + is bounded in E × [λ0 . contradicting (7). λ∗ ). λ0 ). ∞) : (u. (To obtain the existence of an open set U with properties given above. λ ≥ λ0 . O. S+ = {(u. S+ ∩ ∂U = ∅. It therefore 2R follows from Theorem 1 that d(f (·. Let C + be the maximal connected subset of S+ such that 1. O. 0) = d(f (·. There are two possibilities: Either S+ = C + or else there exists (u. Uλ∗ . Cλ0 ∩ O = {u0 } (Cλ0 ∩ O = {u0 }). above holds. 0). On the other hand. because of (7). one may obtain the conclusion that the set of all such continua is either bounded in the right (left) half space.) The existence of C − with the above listed properties is demonstrated in a similar manner. λ) ∈ 2R S+ such that (u. ˜ 6 Remark If the component C + of Theorem 4 is bounded and O is an isolating neighborhood of C + ∩ (E\O) × {λ0 }. ∞). 2R 2.
e. tu ∈ K.. be a polynomial of degree n whose leading coeﬃn cient is ( without loss in generality) assumed to be 1 and let q(z) = i=1 (z −ai ). . an are distinct complex numbers. For all λ ≥ 0. L : K → K. . where a1 . for each i. and must therefore reach every λ−level. then all continua C + must be unbounded. r > 0. for λ = λ0 only isolated solutions and if the integer given by (7) has the same sign with respect to isolating neighborhoods O for all such solutions where (7) holds. z ∈ C. we call L a strongly positive operator. Let E be a real Banach space and let K be a cone in E. It is an elementary exercise to show that a cone K induces a partial order ≤ on E by the convention u ≤ v if and only if v − u ∈ K. i. The KreinRutman theorem [26] is a generalization of this classical result to positive compact operators on a not necessarily ﬁnite dimensional Banach space. i.e. (8) satisﬁes z < R. Hence. whenever L : K\{0} → intK. . t ≥ 0.54 This observation has the following important consequence. We conclude that each zero of p(z) must be connected to some ai (apply the above argument backwards from the λ = 1−level. • K ∩ {−K} = {0}. the level λ = 1. there exists a + continuum Ci of solutions of (8) which is unbounded with respect to λ. Oi . (8) has only isolated solutions and for λ = 0 each such solution has the property that d(f (·. Furthermore for λ ∈ [0. 7 Example Let p(z). If equation (6) has. in particular. A linear operator L : E → E is called positive whenever K is an invariant set for L. Let f (z. . . 4 The Theorem of KreinRutman In this section we shall employ Theorem 4 to prove an extension of the PerronFrobenius theorem about eigenvalues for positive matrices. λ) = λp(z) + (1 − λ)q(z). Then f may be considered as a continuous mapping f : R2 × R → R2 . where Oi is an isolating neighborhood of ai . there exists a constant R such that any solution of f (z. a closed convex subset of E with the properties: • For all u ∈ K. if need be). λ) = 0. 0). r]. 0) = 1. If K is a cone whose interior intK is nonempty.
Since clearly u = 1. Since L is compact. we have that u ∈ K. we obtain by induction that λ m ǫw ≤ u. n λǫ ≤ λǫLw ≤ u. If it is the case that L is a strongly positive compact linear operator. 9 Theorem Let E have a cone K. such that u = λ0 Lu. it must be case that λ ≤ m. whose interior. Then there exists a unique λ0 > 0 with the following properties: 1. (9) where ≤ is the partial order induced by K. Restrict the operator L to the cone K and denote by L the Dugundji extension of this operator to E. (10) ˜ Proof. and therefore u = λL(u + ǫw). (u. whenever (u. this will be done in the theorem of KreinRutman which we shall establish as a corollary of Theorem 8. Thus. THE THEOREM OF KREINRUTMAN 55 8 Theorem Let E be a real Banach space with a cone K and let L : E → E be a positive compact linear operator. with u = λ0 Lu. λǫ )} will contain a convergent subsequence (letting ǫ → 0). we conclude that for each ǫ > 0. Then there exists λ0 > 0 and u ∈ K. uǫ  = 1. Assume there exists w ∈ K. say. such that uǫ = λǫ L(uǫ + ǫw). m]. Let L be a strongly positive compact linear operator. the ˜ ˜ operator L is a completely continuous mapping with L(E) ⊂ K. and hence that Cǫ ⊂ K × [0. (11) For λ = 0. converging to. by assumption. λ) ∈ Cǫ . much more can be asserted. . intK = ∅. it follows from (12) that λ ≤ m. Since Cǫ is unbounded. u = 1. w = 0 and a constant m > 0 such that w ≤ mLw. equation (11) has the unique solution u = 0 and we may apply + Theorem 4 to obtain an unbounded continuum Cǫ ⊂ E × [0.4. Choose ǫ > 0 and consider the equation ˜ u − λL(u + ǫw) = 0. Since L is a compact linear operator. uǫ ∈ K. (12) Since w = 0. there exists λǫ > 0. Since L(E) ⊂ K. the set {(uǫ . λ0 ). m Applying L to this last inequality repeatedly. ∞) of solutions + ˜ of (11). Thus λLu ≤ u. λ) ∈ + + + Cǫ . There exists u ∈ intK. if (u. it follows that λ0 > 0.
λ0 is a characteristic value of L of geometric multiplicity one. we must have that u ∈ intK. Consequently. such that u − δ ∗ v ∈ K. a contradiction. Next let λ = λ0 be another characteristic value of L and let v = 0 be such that v ∈ K ∪ {−K}. Hence. because δ ∗ is maximal. [27]. i. We have therefore proved that λ0 is the only characteristic value of L having an eigendirection in the cone K and further that any other eigenvector corresponding to λ0 must be a constant multiple of u. Choose w ∈ K\{0}. λ) ∈ (K\{0}) × (0. We therefore may apply Theorem 8 to obtain λ0 > 0 and u ∈ K such that u = λ0 Lu. i. and there exists δ∗ < 0. . such that u − δ ∗ v ∈ K. λ0 λ 1 λ0 (u − δ ∗ v) ∈ K. Since L is strongly positive. and λ0 δ∗ > λδ ∗ . there exists a maximal δ ∗ > 0. r > δ ∗ . in fact also has algebraic multiplicity one. . unless u − δ ∗ v = 0. λ0 λ which implies that u − λ0 δ ∗ v ∈ intK. Hence it must be the case that λ = λ0 . then. If λ0 < λ. / maximal. [43]) that the operator id − λ0 L has the following property: There exists a minimal integer n such that ker(id − λ0 L)n = ker(id − λ0 L)n+1 = ker(id − λ0 L)n+2 = . minimal. for all δ > 0. for δ small. Recall from the Riesz theory of compact linear operators (viz. u − δv ∈ intK and there exists δ ∗ > 0. Now / L(u − δ ∗ v) = λ0 1 (u − δ ∗ v). small such that Lw − δw ∈ intK. we conclude that λ0 < λ. 0 As observed above we have that λ0 is a characteristic value of geometric multiplicity one.e. ∞) is such that v = λLv. and the dimension of the generalized eigenspace ker(id − λ0 L)n is called the algebraic multiplicity of λ0 .e. with v = λLv. i. we get that λ0 δ ∗ < λδ∗ . Now L(u − δ ∗ v) = and L(u − δ∗ v) = λ0 1 (u − δ∗ v) ∈ K. If the latter holds. i. we have that u − δv ∈ intK. if λ < 0. v = 0. / Proof. in terms of the partial order δw ≤ Lw. if not. since Lw ∈ intK. we shall establish that λ0 . i. then v ∈ intK. .56 2. then v ∈ K ∪ {−K} and λ0 < λ. λ0 λ Thus. If λ(∈ R) = λ0 is such that there exists v ∈ E. then λ0 < λ. there exists δ > 0. whereas. we may reverse the role of u and v and also obtain λ < λ0 . λ then λ0 = λ. Again.e the dimension of the kernel of id − λ0 L equals one. If (v. Before giving an application of the above result.. .e λ2 < λ2 . suﬃciently small. if λ > 0. u − rv ∈ K.e. such that u − δ∗ v ∈ K.
We shall now consider the question of bifurcation from this trivial branch of solutions and demonstrate the existence of global branches of nontrivial solutions bifurcating from the trivial branch. We refer the interested reader to Krasnosel’skii [24] for a veriﬁcation. for any positive 0 1 integer m. Hence. (15) (13) where F : E × R → E is completely continuous. or 0 0 λm Lm y = y + mku. Then λm Lm z = αu − λm Lm y. for any / integer m. we obtain that ku ∈ −K. Since u ∈ intK. 57 10 Theorem Assume the conditions of Theorem 9 and let λ0 be the characteristic value of L. λ) = 0. F (0. then it may be shown that µ > λ0 . there exist α > 0 and y ∈ K such that z = αu − y. and hence that the equation f (u. It follows therefore that z ∈ K. for otherwise m z − ku ∈ K. GLOBAL BIFURCATION With this terminology. whose existence is established there. If µ is such a characteristic value. We shall now assume that (14) has the trivial solution for all values of λ. Let z = (id − λ0 L)n−2 v. Our main tools will again be the properties of the LeraySchauder degree and Whyburn’s lemma. 5 Global Bifurcation As before. . λ) ≡ 0. implying that −ku ∈ K. where λ0 is as in Theorem 10. it follows that there exists a smallest integer n > 1 such that the generalized eigenspace is given by ker(id − λ0 L)n . We assume the contrary. a contradiction.6. and 0 0 by the above we see that y + mku ≤ βu. We shall see that this result is an extension of the local bifurcation theorem. then z − λ0 Lz = ku. 11 Remark It may be the case that. then λm Lm y ≤ βu.5. since ker(id − λ0 L) has dimension one (Theorem 9). Theorem II. λ). L also has complex ones. λ) = u − F (u. we have the following addition to Theorem 9. Proof. It follows from Theorem 9 and its proof that w = ku. aside from real characteristic values. Then. Dividing this inequality my m and letting m → ∞. Then λ0 is a characteristic value of L of algebraic multiplicity one. λ ∈ R. and hence. such that y ≤ βu. we get that λm Lm z = z − mku. Choose β > 0. a contradiction. let E be a real Banach space and let f : E × R → E have the form f (u. where u is given by Theorem 9 and k may assumed to be positive. by induction. there exists a nonzero v ∈ E such that (id − λ0 L)n v = 0 and (id − λ0 L)n−1 v = w = 0.
. for the sake of illustration provide two simple one dimensional examples. it would imply. b]. We shall. (16) where Br (0) = {u ∈ E : u < r} is an isolating neighborhood of the trivial solution. that d(f (·. where a. b]. containing no nontrivial solutions in its boundary. b]. We shall ﬁrst show that (15) has a nontrivial solution (u. B ⊂ KB . with U ∩ V = ∅. a). Here we shall. We hence have that for each Ω ∈ U there is a continuum C of solutions of (15) which intersects ∂Ω in a nontrivial solution. construct a set Ω ∈ U. Br (0). b). λ) : (u. we may. b]) = ∅. λ) ∈ ∂Ω for any such Ω ∈ U. a). Br (0). If the latter holds. where Ω0 = Br (0) × [a. i. let us consider the following sets: K = f −1 (0) ∩ Ω. using the boundednes of C. (17) A = {0} × [a. Proof. Br (0). is impossible. b). by the generalized homotopy and the excision principle of LeraySchauder degree. KB of K. and Ω∞ is a bounded open subset of (E\{0}) × R. V in E × R such that KA ⊂ U. B = f −1 (0) ∩ (∂Ω\(Br (0) × {a} ∪ Br (0) × {b})). 0). or else (ii) C ∩ {0} × (R\[a. throughout this text. that there are no nontrivial solutions of (15) which belong to ∂Ω∗ . Deﬁne a class U of subsets of E × R as follows U = {Ω ⊂ E × R : Ω = Ω0 ∪ Ω∞ }. such that u = 0 is an isolated solution of (15) for λ = a and λ = b. KB ⊂ V .58 12 Theorem Let there exist a. It follows. 0).e we assume that C is bounded and C ∩ {0} × (R\[a. We assume now that neither of the alternatives of the theorem hold. 0) = d(f (·. there is a separation KA . To accomplish this. furthermore assume that d(f (·. apply the above theorem to several problems for nonlinear diﬀerential equations. with A ⊂ KA . In this case. We observe that K may be regarded as a compact metric space and A and B are compact subsets of K. we may ﬁnd open sets U. however. λ) solves (15) with u = 0} ∪ {0} × [a. We hence may apply Whyburn’s lemma to deduce that either there exists a continuum in K connecting A to B or else. thus arriving once more at a contradiction. this. We let Ω∗ = Ω ∩ (U ∪ V ) and observe that Ω∗ ∈ U. Then either (i) C is unbounded in E × R. b] and let C ⊂ S be the maximal connected subset of S which contains {0} × [a. b]) = ∅. b are not bifurcation points. contradicting (16). Let S = {(u. by construction. 0) = d(f (·. b ∈ R with a < b. Br (0). since.
in fact one has the following necessary conditions for bifurcation. In this case bifurcation points from the trivial solution are isolated. 1) are the only bifurcation points from the trivial solution.5. Then there exists a continuum C of nontrivial solutions of (15) which bifurcates from the set of trivial solutions at (0. Theorem 12. where B is the Fr´chet derivae tive of F . be given by f (u. λ) : λ − 1 = sin } ∪ {0} × [0. 14 Example Let f : R × R → R be given by 1 f (u. then λ0 is a characteristic value of B. (18) where B is a compact linear operator. u In this case S is given by 1 S = {(u. 16 Theorem Assume that F has the form (18) and let λ0 be a characteristic value of B which is of odd algebraic multiplicity. λ0 ) and C is either unbounded in E × R or else C also bifurcates from the trivial solution set at (0. λ) = (1 − λ)u + u sin . . Using this result. b chosen in a neighborhood of λ = −1 and also in a neighborhood of λ = 1. and the LeraySchauder formula for computing the degree of a compact linear perturbation of the identity (an extension to inﬁnite dimensions of Exercise 8 of Chapter III). we obtain the following result. λ1 ). GLOBAL BIFURCATION 13 Example Let f : R × R → R. Also one may quickly check that (16) holds with a. It is easy to see that S is given by S = {(u. ∞). 59 and hence that (0. 15 Proposition Assume that F has the form (18). where λ1 is another characteristic value of B. as u → 0. −1) and (0. Furthermore. λ) = λBu + o(u). λ0 ) is a bifurcation point from the trivial solution for equation (15). by choosing a < 0 and b > 2. In many interesting cases the nonlinear mapping F is of the special form F (u. the bifurcating continuum is bounded. u which is an unbounded set. λ) : u2 + λ2 = 1} ∪ {0} × (−∞. If (0. and we may check that (16) holds. 2]. λ) = u(u2 + λ2 − 1).
17 Example The system of scalar equations x = λx + y 3 y = λy − x3 (19) has only the trivial solution x = 0 = y for all values of λ. Hence. Br (0). b] contains. We note. As already observed this problem is equivalent to an operator equation u = λF (u). 0). Br (0). not every characteristic value will yield a bifurcation point. that λ0 = 1 is a characteristic value of the Fr´chet derivative of multiplicity two. x ∈ [0. It follows that the trivial solution is an isolated solution (in E) of (15) for λ = a and λ = b. (21) (20) . · · · . where β equals the algebraic multiplicity of λ0 as a characteristic value of B. On the other hand. and d(id − bB. 0) are deﬁned for r. 0). given by d(id − aB. The following example serves to demonstrate that. d(id − aB. π] → C[0. no other characteristic values. the result follows from Theorem 12 and Proposition 15. by assumption.60 Proof. Br (0). Since β is odd. we may ﬁnd a < λ0 < b such that the interval [a. These values are given by λ = 1. 18 Example Consider the boundary value problem u′′ + λ sin u = 0. Since λ0 is isolated as a characteristic value. d(f (·. Br (0). · · · k 2 . On the other hand. π] u(0) = 0. in general. to ﬁnd these eigenvalues is equivalent to ﬁnding the values of λ for which u′′ + λu = 0. 0) and d(f (·. 4. b). k ∈ N. a). suﬃciently small and are. e As a further example let us consider a boundary value problem for a second order ordinary diﬀerential equation. the pendulum equation. Br (0). Br (0). u(π) = 0. 0) = (−1)β d(id − bB. respectively. Thus to ﬁnd the bifurcation points for (20) we e must compute the eigenvalues of F ′ (0). π] is a completely continuous operator which is continuously Fr´chet diﬀerentiable e with Fr´chet derivative F ′ (0). u(π) = 0 has nontrivial solutions. 0). x ∈ [0. besides λ0 . π] u(0) = 0. where F : C[0.
k 2 ). 2. 6.6. Prove Proposition 15. . 3. 4. 7. Supply the details for the proof of Theorem 16. In Example 18 show that the second alternative of Theorem 16 cannot hold. 6 Exercises 1. 5. Perform the calculations indicated in Example 14. Supply the details for the proof of Theorem 12. Prove Proposition 15. 9. Provide the details for Example 18. Perform the calculations indicated in Example 13. k ∈ N is a bifurcation point for (20). Prove the assertion of Example 17. EXERCISES 61 Furthermore we know from elementary diﬀerential equations that each such eigenvalue has a onedimensonal eigenspace and one may convince oneself that the above theorem may be applied at each such eigenvalue and conclude that each value (0. 8.
62 .
Part II Ordinary Diﬀerential Equations 63 .
.
t ∈ I and u′ (t) = f (t. u(t)) ∈ D. N ≥ 1. u). The most basic such constraints are given by ﬁxing an initial value of a solution. We consider the diﬀerential equation u′ = f (t. where u ∈ C 1 (I. RN ). To this end let D be an open connected subset of R × RN . in general. u(t)). whereas some constraints on the solutions sought might provide existence and uniqueness of the solution. We have the following proposition whose proof is straightforward: 65 (2) . and let f : D → RN be a continuous mapping. t ∈ I.Chapter V Existence and Uniqueness Theorems 1 Introduction In this chapter. with I an interval. ′ = d . if (t. Simple examples tell us that a given diﬀerential equation may have a multitude of solutions. I ⊂ R. u0 ) ∈ D we seek a solution u of (1) such that u(t0 ) = u0 . dt (1) and seek suﬃcient conditions for the existence of solutions of (1). we shall present the basic existence and uniqueness theorems for solutions of initial value problems for systems of ordinary diﬀerential equations. is called a solution. By an initial value problem we mean the following: • Given a point (t0 .
t−t0 ≤α ˜ . satisfying the initial condition (2) if and only if (t. This result is usually called the PicardLindel¨f theorem o 2 Theorem Assume that f : D → RN satisﬁes a local Lipschitz condition on the domain D. one has the following existence and uniqueness theorem. b min{a. We remark that the theorem as stated is a local existence and uniqueness theorem. and I an interval containing t0 is a solution of the initial value problem (1). where the solution exists will depend upon the initial condition. as will be seen in a subsequent section. with I ⊂ R. t ∈ I and t u(t) = u(t0 ) + t0 f (s. there exists a constant L = L(K). Proof. t0 + α]. (t.66 1 Proposition A function u ∈ C 1 (I. Further let m = α = max(t. u(s))ds. using Proposition 1. RN we deﬁne a new norm as follows: u = max e−Lt−t0  u(t). there exist positive constants a and b such that Q = {(t. since D is open. ˜ ˜ Let L be any constant. RN : u(t) − u0  ≤ b. then for every (t0 . u). then. u0 ) ∈ D equation (1) has a unique solution satisfying the initial condition (2) on some interval I. Global results will follow from this result. Let (t0 . For such functions. m }. u0 ) ∈ D. establish some of the classical and basic existence and existence/uniqueness theorems. u) : t − t0  ≤ a. in the sense that the interval I. t0 + α]. such that for all (t. Let L be the Lipschitz constant for f associated with the set Q. L > L. and deﬁne M = {u ∈ C [t0 − α. t − t0  ≤ α}. u1 ) − f (t. RN ). u − u0  ≤ b} ⊂ D. In C [t0 − α. (3) We shall now. provided for every compact set K ⊂ D.u)∈Q f (t. u1 ). u2 ) ∈ K f (t. u(t)) ∈ D. by extending solutions to maximal intervals of existence. 2 The PicardLindel¨f Theorem o We say that f satisﬁes a local Lipschitz condition on the domain D. u2 ) ≤ Lu1 − u2 .
We remark that. ˜ L ˜ t0 f (s. Then for every (t0 . for u. v(s))ds t ≤ L t0 u(s) − v(s)ds. then (M. Next deﬁne the operator T on M by: t (T u)(t) = u0 + t0 f (s. u(s))ds. The result therefore follows from the contraction mapping principle.6. the contraction mapping theorem gives a constructive means for the solution of the initial value problem in Theorem 2 and the solution may in fact be obtained via an iteration procedure. we still get the existence of solutions. v) = u − v . t0 ∈ I. Computing further. ≤  t0 f (s.3. v). v ∈ M that (T u)(t) − (T v)(t) and hence e−Lt−t0  (T u)(t) − (T v)(t) and hence ρ(T u. we shall show. since T is a contraction mapping. ρ) is a complete metric space. u0 ) ∈ D the initial value problem (1). Hence T : M → M. called the CauchyPeano theorem provides the local solvability of initial value problems. (2) has a solution on some interval I. ≤ e−Lt−t0  L ≤ L u−v . ˜ L ˜ t t0 t u(s) − v(s)ds proving that T is a contraction mapping. This procedure is known as Picard iteration. In the next section. (T u)(t) − u0  ≤ αm ≤ b. THE CAUCHYPEANO THEOREM 67 And we let ρ(u. . t − t0  ≤ α. u(s))ds. 3 Theorem Assume that f : D → RN is continuous. 3 The CauchyPeano Theorem The following result. we obtain. since u ∈ M. T v) ≤ L ρ(u. Theorem I. u(s)) − f (s. t (4) Then (T u)(t) − u0  ≤  and. that without the assumption of a local Lipschitz condition.
) We shall prove below that this continuation process leads to maximal intervals of existence and also describes the behavior of solutions as one approaches the endpoints of such maximal intervals. Let (t0 . t0 + a) such that f (t. Under such assumptions we have the following extension of the CauchyPeano theorem.e. • for each Q there exists a function m ∈ L1 (t0 − a.1 Carath´odory equations e In many situations the nonlinear term f is not continuous as assumed above but satisﬁes the socalled Carath´odory conditions on any parallelepiped Q ⊂ D. e where Q is as given in the proof of Theorem 2.e. u(t0 ± α)) ∈ D. showing that {T un} is precompact and hence T is e completely continuous. RN ) with norm u = maxt−t0 ≤α u(t). Then for every e (t0 . • f is measurable in t for each ﬁxed u and continuous in u for almost all t.30) once we verify that T is completely continuous on M. We hence may reapply these theorems with initial conditions given at t0 ± α and conditions u(t0 ± α) and thus continue solutions to larger intervals (in the case of Theorem 2 uniquely and in the case of Theorem 3 not necessarily so. t0 ∈ I. This completes the proof. We let M be as deﬁned in the proof of Theorem 2 and note that M is a closed. both (t0 ± α. We hence may apply the Schauder ﬁxed point theorem (Theorem III. un (s))ds ¯ ¯ ≤ mt − t. and let Q. (2) has a solution on some interval I. It therefore has a uniformly convergent subsequence (as follows from the theorem of AscoliArz`la [36]). To see this we note. α.. m be as in the proof of Theorem 2. Theorem 3: 4 Theorem Let f satisfy the Carath´odory conditions on D. is a uniformly bounded and equicontinuous family in E. in the sense that there exists an absolutely continuous function u : I → RN which satisﬁes the initial condition (2) and the diﬀerential equation (1) a.68 Proof. t0 + α]. then ¯ (T un )(t) − (T un )(t) ≤  t f (s. bounded convex subset of E and further that T : M → M. On the other hand. u0 ) ∈ D the initial value problem (1). since f is continuous. if {un } ⊂ M. We note from the above proofs (of Theorems 2 and 3) that for a solution u thus obtained. u0 ) ∈ D. u) ∈ Q. t Hence {T un} ⊂ M. that. (t. 3. Consider the space E = C([t0 − α. in I. . Then E is a Banach space. it follows that T is continuous. u) ≤ m(t). i.
4. b). 4 Extension Theorems In this section we establish a basic result about maximal intervals of existence of solutions of initial value problems. is immediate. t ∈ (a. b) with (t. in the sense given in the theorem. u(s))ds. t − t0  ≤ α. and choose Q = {(t. u) : t − t0  ≤ a.t0 +α] ≤ b. That ﬁxed points of T are solutions of the initial value problem (1). t − t0  ≤ α}. Then. for α small enough. T u is a a continuous function e and t (T u)(t) − u0  ≤  Further. hence will have a ﬁxed point in M by the Schauder Fixed Point Theorem. u(t)) ∈ D. Further let u be a solution of (1) deﬁned on a ˜ bounded interval (a. t0 + α]. u(s))ds. t0 +α (T u)(t) − u0  ≤ t0 −α m(s)ds ≤ m L1 [t0 −α. Next deﬁne the operator T on M by: t 69 (T u)(t) = u0 + t0 f (s. We ﬁrst prove the following lemma. Hence T : M → M. Let M = {u ∈ C [t0 − α. u0 ) ∈ D. RN : u(t) − u0  ≤ b. ˜ 5 Lemma Assume that f : D → RN is continuous and let D be a subdomain ˜ of D. because of the Carath´odory conditions. since u ∈ M. Let (t0 . where α ≤ a is to be determined. u − u0  ≤ b} ⊂ D. t0 f (s. Then the limits t→a+ lim u(t). EXTENSION THEOREMS Proof. . One next shows (see the Exercise 14 below) that T is a completely continuous mapping. t→b− lim u(t) exist. with f bounded on D. (2).
then u(t) = u(t0 ) + (a. Then u may be extended as a solution of (1) to a maximal interval of existence (ω− . u(s))ds. ˜ Next let u be a solution of (1). The following theorem holds. av ≥ au . aw ).e. We must show that (t. Proof. u(tn )) ∈ K. We may therefore choose a constant a > 0. b). ω+ ). u.70 Proof. a similar argument will yield the existence of a left maximal one and together the two will imply the existence of a maximal interval of existence. · · · 1 there exists tn . u(t)) → ∂D as t → ω+ . (5) t t0 f (s. We establish the existence of a right maximal interval of existence. u(t)) → ∂D as t → ω± . Let t0 ∈ (a. u−u∗  ≤ v w is deﬁned on Iv = [t0 . Hence the above limits exist. there exists tK . u(t)) ∈ K. u∗ ) ∈ K. as indicated above. • • aw ≥ av . u(tn ))} such that {(tn . u(tn ))} converges to. On the other hand. ω+ ) and (t. for t > tK . t2 ∈ if and only if: • u ≡ v ≡ w on [t0 . we proceed indirectly. This ˜ maximal element u cannot be further extended to the right. 2. such that for every n = 1. 6 Theorem Assume that f : D → RN is continuous and let u be a solution of (1) deﬁned on some interval I. w of (1).. b). Let u be a solution of (1) with u(t0 ) = u0 deﬁned on an interval I = [t0 . it is an interior point of D. Since (ω+ . we obtain u(t1 ) − u(t2 ) ≤ mt1 − t2 . the / conclusion clearly holds. av ). and (tn . In which case there exists a compact set K ⊂ D. au ). call it again {(tn . continue the solution beyond the interval (a. We say that a solution u of (1) has maximal interval of existence (ω− . (2) satisfy v w. Hence for t1 . i. • v ≡ w on Iv . aw ≥ au . if ω+ < ∞. If ω+ = ∞. provided u cannot be continued as a solution of (1) to the right of ω+ nor to the left of ω− . u) : ω+ −t ≤ a. (2) with right maximal interval of existence [t0 . b). there will be a subsequence. Since K is compact. ω+ ). say. We see that is a partial order on the set of all solutions S of (1). . One next veriﬁes that the conditions of the Hausdorﬀ maximum principle (see [37]) hold and hence that S contains a maximal element. to the left of a and the right of b. such that (t. (ω+ . 0 < ω+ − tn < n . au ). We say that two solutions v. We may therefore. given any compact set K ⊂ D. such that Q = {(t. (2) which agree with u on I. u∗ ) ∈ K. ˜ where m is a bound on f on D. is deﬁned on Iw = [t0 .
Then the maximal interval of existence (ω− . t0 + a] × RN → RN is continuous and let u be a solution of (1) deﬁned on some right maximal interval of existence I ⊂ [t0 . and let n be so large that a a . because of (6). u(tn ) − u∗  ≤ . Let m = max(t. t0 from which. and we may extend u to the right of ω+ contradicting the maximality of u. u) ≤ α(t)u + β(t).e. (6) 8 Corollary Assume that f : (a. We also consider the following corollary. the other case being similar. It therefore follows that t→ω+ lim u(t) = u∗ . Then an easy calculation yields v ′ − α(t)v ≤ α(t)c + β(t). we obtain t u(t) ≤ u(t0 ) +  t0 [α(s)u(s) + β(s)]ds.4. β ∈ L1 (a.. ω+ ) is (a. ω+ < t0 + a. b) are nonnegative continuous functions. Rt α(τ )dτ t Rs and hence v(t) ≤ e t0 e t0 α(τ )dτ [α(s)c + β(s)]ds. If u is a solution of (1). using Corollary 7. we let v(t) = t0 [α(s)u(s) + β(s)]ds. either I = [t0 . which is of importance for diﬀerential equations whose right hand side have at most linear growth. then u satisﬁes the integral equation (3) and hence. t (7) Considering the case t ≥ t0 . follows that ω+ = b. Then. u). and t→ω+ lim u(t) = ∞. EXTENSION THEOREMS 71 a} ⊂ D. 7 Corollary Assume that f : [t0 . 0 < ω + − tn ≤ 2m 2 Then a u(tn ) − u(t) < m(ω+ − tn ) ≤ . b) for any solution of (1). t0 +a].u)∈Q f (t. for t≤ t < ω+ . I. or else I = [t0 . and c = u(t0 ). we assume that the following growth condition holds: f (t. Proof. . and thus for n large (tn . where α. u(tn )) ∈ Q. 2 as an easy indirect argument shows. b)×RN → RN is continuous and let f satisfy (6). ω+ ). t0 + a].
where the function f also depends upon parameters. u(t0 ) = u0 . λ ∈ Rm . λ). (8) is replaced by the parameter dependent problem u′ = f (t. t0 . 9 Theorem Assume that f : D → RN is a continuous mapping and that (8) has a unique solution u(t) = u(t. u. λ(t0 ) = λ. We consider the initial value problem u′ = f (t. Then the solution depends continuously on (t0 . u0 ). u. u0 ). 0. N ≥ 1. u0 ) ∈ D. and let f : D → RN be a continuous mapping. u0 . i. λ). u0 ) ∈ D. u0 ) depends either continuously or smoothly on the initial condition (t0 .1 Continuous dependence We ﬁrst prove the following proposition. the solution un (t) = u(t. tn . un )} ⊂ D converges to (t0 . t∈Iǫ . (11) (10) provided (10) is uniquely solvable. u(t0 ) = u0 . u0 ). in the following sense: If {(tn . for every (t0 . u(t0 ) = u0 . u0 ). un ). there exists nǫ and an interval Iǫ such that for all n ≥ nǫ . 5. then given ǫ > 0. This situation is a special case of the above. (9) for every (t0 . (12) and obtain an initial value problem for a system of equations of higher dimension which does not depend upon parameters. t0 . (8) and assume we have conditions which guarantee that (8) has a unique solution u(t) = u(t. A somewhat more general situation occurs frequently. and solutions u then are functions of the type u(t) = u(t.e. as we may augment the original system (10) as u′ λ′ = = f (t. t0 . u0 ) ∈ D. exists on Iǫ and max u(t) − un (t) ≤ ǫ. We shall now present conditions which guarantee that u(t. u).72 5 Dependence upon Initial Conditions Let again D be an open connected subset of R × RN . t0 . λ).
1 ≤ i ≤ N. We have the following corollary. · · · . We rely on the proof of Theorem 3 and ﬁnd that for given ǫ > 0. DEPENDENCE UPON INITIAL CONDITIONS 73 Proof.2 Diﬀerentiability with respect to initial conditions In the following we shall employ the convention u = (u1 . The sequence {un } hence will be uniformly bounded and equicontinuous on Iǫ and will therefore have a subsequence converging uniformly on Iǫ . there ˜ exists nǫ such that {(tn . by the uniqueness assumption must equal u. with δik the Kronecker delta. Further. Then the solution u(t) = u(t. the solution un (t) = u(t. u0 ). t0 . un )} ⊂ Q. tn . t0 . i ∂f ∂u . completing the proof. t∈Iǫ 5. the whole sequence must converge to u. u0 ). u0 ) ∂ui is the solution of the initial value problem y(t) = y ′ = J(t)y. which asserts continuity of solutions with respect to the diﬀerential equation. n = 1. of (8) is of class C 1 in the variable u0 . Employing the integral equation (3) we see that the limit must be a solution of (8) and hence. for every (tn . Since this is true for every subsequence. 10 Corollary Assume that fn : D → RN . 1 ≤ i ≤ N are continuous also. there exists nǫ and an interval Iǫ such that for all n ≥ nǫ . uN ). Then the solution depends continuously on (t0 . and fn converges to f. u − u0  ≤ } ⊂ Q. un ). in the following sense: If {(tn . un ). 11 Theorem Assume that f : D → RN is a continuous mapping and that the ∂f partial derivatives ∂ui . u) : t − t0  ≤ . · · · . uniformly on compact subsets of D. un )} ⊂ D converges to (t0 . for t ∈ Iǫ . The proof is similar to the above and will hence be omitted. then given ǫ > 0.5. where ei ∈ Rn is given by ek = δik . tn . if J(t) is the Jacobian matrix J(t) = J(t. 2. Using the proof of Theorem 3 we obtain a common compact interval Iǫ of existence of the sequence {un } and {(t.t0 . un ) ∈ D. y(t0 ) = ei . are continuous mappings and that (8) (with f = fn ) has a unique solution un (t) = u(t. t0 . where b α ˜ Q = {(t. u=u(t. 2 2 and Q is the set given in the proof of Theorems 2 and 3. We have the following theorem. u0 ) = then ∂u (t. u0 ) ∈ D. un (t))} ⊂ Q. exists on Iǫ and max u(t) − un (t) ≤ ǫ. u2 .u0 ) .
∀x. h where lim sup and lim inf are taken componentwise. 1 ≤ i ≤ N. u(t). t0 . h lim suph→0− u(t+h)−u(t) . sy1 + (1 − s)y2 )ds. . we consider the Dini derivatives D+ u(t) D+ u(t) D− u(t) D− u(t) = = = = lim suph→0+ u(t+h)−u(t) . ⇔ xi < y i . Letting yh (t) = uh (t) − u(t) . uh (t))y. Since G(t. uh (t)) − f (t. ∂u we obtain that yh is the unique solution of y ′ = G(t. Let ei be given as above and let u(t) = u(t. uh (t)) → J(t) as h → 0. xi = y i . 12 Deﬁnition A function f : RN → RN is said to be of type K (after Kamke [23]) on a set S ⊂ RN . we may apply Corollary 10 to conclude that yh → y uniformly on a neighborhood of t0 . u0 + hei ). whenever f i (x) ≤ f i (y). h lim inf h→0− u(t+h)−u(t) . 1 ≤ i ≤ N. The following theorem on diﬀerential inequalities is of use in obtaining estimates on solutions. y1 . x ≤ y. y2 ) = 0 ∂f (t. We compute (uh (t) − u(t)) = f (t. uh (t) = u(t. h ′ we get yh (t0 ) = ei . y ∈ S. For a function u : I → RN . whenever h is suﬃciently small. u(t)). We note that u and uh will have a common interval of existence. 6 Diﬀerential Inequalities We consider in RN the following partial orders: x≤y x<y ⇔ xi ≤ y i . where I is an open interval. h lim inf h→0+ u(t+h)−u(t) . where h is suﬃciently small so that uh exists. u0 ). If we let 1 G(t.74 Proof. t0 . y(t0 ) = ei . u(t).
then z(t) < u(t). b] × RN → RN is a continuous mapping which is of type K for each ﬁxed t. v(c) ≥ u(c). a < t ≤ b. u0 ) ∈ D. Hence (since f is of type K) D− v i (c) > f i (c. such that v(t) > u(t). Let u : [a. z(t)). b] → RN is continuous and satisﬁes D− z(t) < f (t. b] → RN be continuous and satisfy D− v(t) v(a) > f (t. On the other hand. b] → RN be a solution of (1) and let v : [a. If z : [a. > u(a). Then the initial value problem (8) has a unique right maximal (minimal) solution for each (t0 . Right minimal solutions are deﬁned similarly. a < t ≤ b. D− v i (c) = ≤ a contradiction. (13) then v(t) > u(t). t0 ≤ t ∈ I. a ≤ t ≤ a + δ. 14 Deﬁnition A solution u∗ of (1) is called a right maximal solution on an interval I. there exists δ > 0. v i (c) = ui (c). there will exist a ﬁrst point c and an index i such that v(t) > u(t). DIFFERENTIAL INEQUALITIES 75 13 Theorem Assume that f : [a. a ≤ t ≤ b. b]. 15 Theorem Assume that f : D → RN is a continuous mapping which is of type K for each t. z(a) < u(a). if for every t0 ∈ I and any solution u of (1) such that u(t0 ) ≤ u∗ (t0 ). Proof. a ≤ t < c. The second part follows along the same line of reasoning. a ≤ t ≤ b. v(c)) ≥ f i (c.6. it follows that u(t) ≤ u∗ (t). u(c)) = ui (c). . lim suph→0− lim suph→0− v i (c+h)−v i (c) h ui (c+h)−ui (c) h ′ ′ (14) = ui (c). We prove the ﬁrst part of the theorem. v(t)). If the inequality does not hold throughout [a. By continuity of u and v.
the whole sequence will. hence the initial value ¯ problem (8). t1 ]. v(t0 ) = u0 + . with f replaced by f has a solution u that extends to [a.and supersolutions of (8). n < m. b] × RN → RN is a continuous mapping which is of type K for each ﬁxed t. x) − (x − x). where for 1 ≤ i ≤ N. t0 ≤ t ≤ t1 . That maximal and minimal solutions are unique follows from the deﬁnition. sub. On the other hand it follows from Theorem 13 that vn (t) > vm (t). v(t). a ≤ t ≤ b. completes the proof. ¯ ¯ i v (t) . · · · . vn (t))} ⊂ U. i z (t) . Deﬁne f (t. We show that z(t) ≤ u(t) ≤ v(t). Let v. for all n suﬃciently large. 16 Theorem Assume that f : [a. ¯ Then f is continuous and has at most linear growth. (16) Then for every u0 . and hence may conclude that u solves the original initial value problem (8). a ≤ t < b. u0 ) ∈ D. if xi < z i (t). t ∈ [a. z : [a. Since the sequence is monotone. v. λ) + ǫ ǫ . b] → RN be continuous and satisfy D+ v(t) ≥ D+ z(t) ≤ z(t) ≤ f (t. . ¯ Proof. hence will have a subsequence which converges uniformly to a solution u∗ of (8). respectively. t1 ] of positive length such that all vn are deﬁned on this interval with {(t. f (t. i = 1. z(t)). there exists a solution u of (8) (with t0 = a) such that z(t) ≤ u(t) ≤ v(t). a ≤ t ≤ b. The functions z and v are called. a ≤ t < b. n n (15) Then. i x = ¯ xi . b] (see Corollary 8). a ≤ t ≤ b. there exists an interval [t0 . t1 ]. v(t)). To see this. N.76 Proof. we shall prove that for any ǫ > 0 we have that z i (t) − ǫ ≤ ui (t) ≤ v i (t) + ǫ. The sequence {vn } is therefore unformly bounded and equicontinuous on [t0 . We next prove an existence theorem for initial value problems which allows for estimates of the solution in terms of given solutions of related diﬀerential inequalities. and extending this solution to a right maximal interval of existence as a right maximal solution. if xi > v i (t). converge to u∗ . t0 ≤ t ≤ t1 . Choose 0 < ǫ ∈ RN and let vn be any solution of v ′ = f (t. x) = f (t. b]. we obtain that u∗ is right maximal on [t0 . Applying Theorem 13 once more. in fact. z(a) ≤ u0 ≤ v(a). if z i (t) ≤ xi ≤ v i (t). given a neighborhood U of (t0 .
c < t ≤ t1 . DIFFERENTIAL INEQUALITIES 77 This we argue indirectly and suppose there exists ǫ > 0 a value c ∈ (a.6. that ui (t + h) − ui (c) ≥ v i (t + h) + ǫ − (v i (c) + ǫ. ui (t) ≥ v i (t) + ǫ. as follows from Theorem 16. Then the problem u′ = f (t. k = i. we get ¯ ¯ that ¯ f i (t. v(t)) − ǫ < D+ v i (t). Deﬁne the mapping T : {x : z(a) ≤ x ≤ v(a)} → {x : z(b) ≤ x ≤ v(b)} by T x = u(b. ¯ On the other hand we have. u(a) = u(b) has a solution u with z(t) ≤ u(t) ≤ v(t). b) and i such that ui (c) = v i (c) + ǫ. Since ǫ > 0 was arbitrary. c < t ≤ t1 ≤ b. c < t ≤ t1 . a ≤ t ≤ b. Proof. Since f satisﬁes a local Lipschitz condition. v(a) ≥ v(b). there exists a unique solution u(t. for h > 0. and hence ui (c) ≥ D+ v i (c). u(t)) = f (t. x). Since ui (t) = v i (t). z(a) ≤ u0 ≤ v(a). contradicting what has been concluded above. Assume furthermore that z(a) ≤ z(b). small. a ≤ t ≤ b. u0 ) of (8) (with t0 = a) such that z(t) ≤ u(t) ≤ v(t). initial value problems are uniquely solvable. 17 Corollary Assume the hypotheses of Theorem 16 and that f satisﬁes a local Lipschitz condition. we conclude that u(t) ≤ v(t). u(t)) − (ui (t) − v i (t)) ≤ f i (t. That u(t) ≥ vz(t) may be proved in a similar way. (17) ′ . u). c < t ≤ t1 and uk (t) ≤ v k (t). Hence for every u0 .
Let z(t) = u(t). t → a. y) ≤ F (t. y ∈ RN . b). Hence. . Let u : [a. x) − f (t. The following results use comparison and diﬀerential inequality arguments to provide a priori bounds and extendability results. c) such that w(t) = 0(µ(t)). v(t)). x ∈ RN . Then (1) cannot have distinct solutions such that u(t) − v(t) = 0(µ(t)). 7 Uniqueness Theorems In this section we provide supplementary conditions which guarantee the uniqueness of solutions of ivp’s. completing the proof. then z is continuous and D− z(t) = D− z(t). w ≡ 0 be the only solution of w′ = F (t. x − y). a ≤ t ≤ b. z(t)). a ≤ t ≤ b. x) ≤ F (t. b] × R+ → R+ is a continuous mapping and that f : [a. for any c ∈ (a. u(t)) ≤ F (t. 0) ≡ 0 and let. we conclude that v(t) ≥ u(t). An important consequence of this corollary is that if in addition f is a function which is periodic in t with period b − a. a ≤ t ≤ b. it follows by Brouwer’s ﬁxed point theorem (Theorem III. 18 Theorem Assume that F : [a. b] → R+ be the continuous and right maximal solution of v ′ (t) = v(a) ≥ F (t. u(a). b] → RN be a solution of (1) and let v : [a.22) and the fact that T is continuous (Theorem 9) that T has a ﬁxed point. t → a where µ is a given positive and continuous function. b] × RN → RN is continuous also and f (t. by Theorem 15 (actually its corollary (Exercise 6)). b) × R+ → R+ is a continuous mapping and that f : (a. then Corollary 17 asserts the existence of a periodic solution (of period b − a). Let F (t. x. Proof. a ≤ t ≤ b.78 then. a ≤ t ≤ b. w) on (a. 19 Theorem Assume that F : (a. since by hypothesis {x : z(a) ≤ x ≤ v(a)} ⊃ {x : z(b) ≤ x ≤ v(b)} and since {x : z(a) ≤ x ≤ v(a)} is convex. Further D− z(t) = lim inf h→0+ u(t)−u(t−h) h ≤ limh→0+ u(t−h)−u(t) h = f (t. (18) then v(t) ≥ u(t). b) × RN → RN is a continuous also and f (t. x).
where M and L > 0 are continuous functions on their respective domains and ∞ ds = ∞. 5. v be distinct solutions of (1) such that u(t)−v(t) = 0(µ(t)). z(t)). Prove Proposition 1. a ≤ t < ∞. Prove Corollary 10. 2. Complete the details in the proof of Theorem 11. A similar result. 6. Show that a real valued continuous function z(t) is nonincreasing on an interval [a. x ∈ RN . Prove the following result: Assume that f : [a. Verify that the space (M. of course. The proof is completed by employing arguments like those used in the proof of Theorem 18. . holds for t → b − . Let u : [a. a < t ≤ b. L(s) Prove that ω+ = ∞ for all solutions of (1). x) ≤ M (t)L(x). 20 Remark Theorem 19 does not require that f be deﬁned for t = a. 8.8. ρ) in the proof of Theorem 2 is a complete metric space. EXERCISES 79 Proof. The advantage of this may be that a = −∞ or that f may be singular there. b] if and only if D− z ≤ 0 on (a. z(t)). b] → RN be continuous and satisfy D− z(t) ≤ z(a) ≤ f (t. t → a. v(t)) ≤ F (t. b] × RN → RN is a continuous mapping which is of type K for each ﬁxed t. (19) then z(t) ≤ u(t). Let u. ∞) × RN → RN is a continuous mapping such that f (t. b]. Assume that f : [a. λ) also depends upon a parameter λ. u(t)) − f (t. 3. a ≤ t ≤ b. 4. 7. b] → RN be a right maximal solution of (1) and let z : [a. u. u(a). 8 Exercises 1. Let z(t) = u(t) − v(t). Then z is continuous and D+ z(t) ≤ f (t. providing a diﬀerential ∂u equation for ∂λ whenever f = f (t. State and prove a theorem similar to Theorem 11.
10. x. Then the initial value problem u′ = f (t. y ∈ RN . y) ≤ c x − y . b) × RN → RN is a continuous mapping and that (f (t. Give the details of the proof of Theorem 18. t > a. 0 ≤ c < . y ∈ RN . x) − f (t. t−a 2 13. t > a. Assume that f : [a. x. This exercise. Then every initial value problem is uniquely solvable to the right. 0 < c < 1 t−a (21) (20) Then the initial value problem (20) has at most one solution. of course follows from the previous one. Note that unique solvability to the left of an initial point is not guaranteed. t ≥ a. Provide the details in the proof of Theorem 4. b) × RN → RN is a continuous mapping and assume the conditions of Theorem 19 with µ ≡ 1. x) − f (t. 11. . Give a more elementary and direct proof. y ∈ RN .80 9. 12. u(a) = u0 has at most one solution. Establish a result similar to Theorem 6 assuming that f satisﬁes Carath´odory e conditions. 15. Assume that f : [a. x. y)) · (x − y) ≤ c x − y2 1 . y)) · (x − y) ≤ 0. The previous exercise remains valid if (21) is replaced by (f (t. b) × RN → RN is a continuous mapping and that f (t. x) − f (t. Assume that f : [a. How must the above condition be modiﬁed to guarantee uniqueness to the left of an initial point? 14. u).
81 . The results obtained will be useful in the study of stability of solutions of nonlinear diﬀerential equations as well as bifurcation theory for periodic orbits and many other facets where linearization techniques are of importance.Chapter VI Linear Ordinary Diﬀerential Equations 1 Introduction In this chapter we shall employ what has been developed to give a brief overview of the theory of linear ordinary diﬀerential equations. 2 Preliminaries Let I ⊂ R be a real interval and let A : I → L(RN . We consider here the system of ordinary diﬀerential equations u′ = A(t)u + f (t). The results are also of interest in their own right. RN ) f : I → RN be continuous functions. initial value problems for (1) are uniquely solvable and solutions are deﬁned on all of I. t ∈ I. (2) (1) Using earlier results we may establish the following basic proposition (see Exercise 1). 1 Proposition For any given f . and u′ = A(t)u. t ∈ I.
(7) Hence. · · · . 1 ≤ i. That the solution set forms a vector space is left as an exercise (Exercise 2. whose proof is left as an exercise below. (8) The following proposition characterizes the set of fundamental solutions. · · · . That this matrix is never singular. · · · . the solution u of (2) such that u(t0 ) = ξ is given by (4) with ai = ξ i . k = 1. is known as the AbelLiouville lemma. the columns of Φ are solutions of (2). aN . ei = δki (Kronecker delta). k It follows that for any set of constants a1 . uN are solutions of (2). if Φ is deﬁned by (5).82 2 Remark More generally we may assume that A and f are measurable on I and locally integrable there. Let the N × N matrix function Φ be deﬁned by Φ(t) = (ui (t)). · · · .1 Fundamental solutions A nonsingular N × N matrix function Ψ whose columns are solutions of (2) is called a fundamental matrix solution or a fundamental system of (2). its proof is again left as an exercise. i = 1. aN )T .. We shall not go into details for this more general situation. below). and let uk (t). we employ the uniqueness principle above.e. a = (a1 . j (5) i. 4 Lemma If g(t) = detΦ(t). provided it is nonsingular at some point. 2. in which case the conclusion of Proposition 1 still holds. N be the solution of (2) such that uk (t0 ) = ek . (6) traceA(s) ds. 3 Proposition The set of solutions of (2) is a vector space of dimension N . Further. the solution u of (2) such that u(t1 ) = ξ. Thus let t0 ∈ I. in which case a may be uniquely determined. is given by (6) provided that Φ(t1 ) is a nonsingular matrix. . Then (4) takes the form Hence for given ξ ∈ RN . then g satisﬁes g(t) = g(t0 )e Rt t0 u(t) = Φ(t)a. Such a matrix is a nonsingular solution of the matrix diﬀerential equation Ψ′ = A(t)Ψ. but leave it to the reader to present a parallel development. where u1 . · · · . for given ξ ∈ RN . To show that the dimension of this space is N. Proof. t1 ∈ I. N. j ≤ N. N 1 (3) u(t) = ai ui (t) (4) is a solution of (2). then Φ(t) is nonsingular for all t ∈ I if and only if Φ(t0 ) is nonsingular for some t0 ∈ I.
CONSTANT COEFFICIENT SYSTEMS 83 5 Proposition Let Φ be a given fundamental matrix solution of (2).3. 6 Proposition Let Φ be a fundamental matrix solution of (2) and let t0 ∈ I. Then t up (t) = Φ(t) t0 Φ−1 (s)f (s)ds (9) is a solution of (1). . In this case a fundamental matrix solution Φ is given by Φ(t) = etA C. where C is a nonsingular constant N × N matrix and etA = ∞ 0 (10) tn An . Furthermore the set of all solutions of (2) is given by {Φc : c ∈ RN }. where Φ is a fundamental system. The following formula. 3 Constant Coeﬃcient Systems In this section we shall assume that the matrix A is a constant matrix and thus have that solutions of (2) are deﬁned for all t ∈ R.2 Variation of constants {Φ(t)c + up (t) : c ∈ RN }. where Φ is a fundamental system of (2). Hence the set of all solutions of (1) is given by t {Φ(t) c + Φ−1 (s)f (s)ds t0 : c ∈ RN }. where C is a constant nonsingular N × N matrix. Hence the problem of ﬁnding all solutions of (1) is solved once a fundamental system of (2) is known and some particular solution of (1) has been found. n! Thus the solution u of (2) with u(t0 ) = ξ is given by u(t) = e(t−t0 )A ξ. 2. Then every other fundamental matrix solution has the form Ψ = ΦC. It follows from Propositions 3 and 5 that all solutions of (1) are given by where Φ is a fundamental system of (2) and up is some particular solution of (1). shows that a particular solution of (1) may be obtained from a fundamental system. known as the variation of constants formula.
and for 1 ≤ i ≤ s.84 To compute etA we use the (complex) Jordan canonical form J of A. On the other hand J has the form J0 J1 J = . .. . q+ 1 qi = N. . .. eλq t etJ0 = . . λq is a qi × qi matrix. there exists a nonsingular matrix P such that A = P JP −1 and hence etA = P etJ P −1 . tJs e eλ1 t eλ2 t . . repeated according to their multiplicities. . λq+i 1 λq+i 1 . 1 λq+i J0 = .. with s is a q × q diagonal matrix whose entries are the simple (algebraically) and semisimple eigenvalues of A.. and By the laws of matrix multiplication it follows that tJ e 0 etJ1 etJ = . Since A and J are similar. Js where λ1 λ2 . Ji = .. . We therefore compute etJ .. .
All solutions u of (11) are bounded on [0. . − ∞ < t < ∞. . as t → ∞. The above considerations have the following proposition as a consequence. Zi = .. and consider the diﬀerential equation u′ = A(t)u. . An easy computation now shows that ri−1 2 · · · t i−1 ! 1 t t 2! r ri−2 0 1 t · · · t i−2 ! r tZi e = . . 0 0 0 ··· 1 Since P is a nonsingular matrix etA P = P etJ is a fundamental matrix solution as well. . . ··· . . where Iri is the ri × ri identity matrix and Zi is given by 0 1 0 1 . . ∞). All solutions u of (11) satisfy u(t) → 0. A(t + T ) = A(t). 7 Proposition Let A be an N × N constant matrix and consider the diﬀerential equation u′ = Au. since Ji = λq+i Iri + Zi . . if and only if Reλ ≤ 0. . if and only if Reλ < 0. i.4. . 1 0 we obtain that etJi = etλq+i Iri etZi = etλq+i etZi .. . for all eigenvalues λ of A and those with zero real part are semisimple. Then: 1. for all eigenvalues λ of A. Also. 4 Floquet Theory Let A(t). (11) . (12) We shall associate to (12) a constant coeﬃcient system which determines the asymptotic behavior of solutions of (12). . FLOQUET THEORY 85 Further. ImP etJ c : c ∈ CN }.e. . t ∈ R be an N × N continuous matrix which is periodic with respect to t of period T.. since J and P may be complex we obtain the set of all real solutions as {ReP etJ c. 2. To this end we ﬁrst establish some facts about fundamental solutions of (12).
then so is Ψ(t) = Φ(t + T ). On the other hand a solution u of (12) is given by u(t) = C(t)etR d. Since Q is nonsingular. hence Ψ is nonsingular. Letting C(t) = Φ(t)e−tR we compute C(t + T ) = Φ(t + T )e−(t+T )R = Φ(t)Qe−T R e−tR = Φ(t)e−tR = C(t). then there exists a nonsingular periodic (of period T ) matrix C and a constant matrix R such that Φ(t) = C(t)etR . It follows by our earlier considerations that there exists a nonsingular constant matrix Q such that Φ(t + T ) = Φ(t)Q. there exists a matrix R (see exercises at the end of this chapter) such that Q = eT R . Further Ψ′ (t) = Φ′ (t + T ) = A(t + T )Φ(t + T ) = A(t)Φ(t + T ) = A(t)Ψ(t). We have proved the following proposition. (13) From this representation we may immediately deduce conditions which guarantee the existence of nontrivial T − periodic and mT − periodic (subharmonics) solutions of (12). Proof. Since Φ is a fundamental matrix it is nonsingular for all t. Proof. where Φ is a fundamental matrix solution of (12).86 8 Proposition Let Φ(t) be a fundamental matrix solution of (12). The properties of fundamental matrix solutions guarantee that the matrix Φ−1 (0)Φ(T ) is uniquely determined by the equation and Proposition 9 implies that Φ−1 (0)Φ(T ) = eT R . . 10 Corollary For any positive integer m (12) has a nontrivial mT −periodic solution if and only if Φ−1 (0)Φ(T ) has an m−th root of unity as an eigenvalue. 9 Proposition Let Φ(t) be a fundamental matrix solution of (12).
will have a mT −periodic solution if and only if Φ(T ) has an eigenvalue λ which is an m−th root of unity. Let us apply these results to the second order scalar equation y ′′ + p(t)y = 0. by the AbelLiouville formula (7). (15) Let y1 be the solution of (14) such that ′ y1 (0) = 1. FLOQUET THEORY where u(0) = C(0)d. Equation (14) may be rewritten as the system u′ = 0 1 −p(t) 0 u. 2 . Hence (14). where ′ a = y1 (T ) + y2 (T ). m 87 (14) (Hill’s equation) where p : R → R is a T periodic function. or equivalently (15). y2 (0) = 1. y1 (T ) − λ ′ y1 (T ) y2 (T ) ′ y2 (T ) − λ = 0. Therefore λ= a± √ a2 − 4 . and y2 the solution of (14) such that ′ y2 (0) = 0. The eigenvalues of Φ(T ) are solutions of the equation det or λ2 − aλ + 1 = 0. Hence u is periodic of period mT if and only if u(mT ) = C(mT )emT R d = C(0)emT R d = C(0)d. Which is the case if and only if emT R = eT R has 1 as an eigenvalue. Then Φ(t) = y1 (t) ′ y1 (t) y2 (t) ′ y2 (t) will be a fundamental solution of (15)and detΦ(t) = e Rt 0 traceA(s)ds = 1.4. y1 (0) = 0.
Show that corresponding to every solution u of (2) there exists a unique solution v of (16) such that t→∞ (16) lim u(t) − v(t) = 0. Verify the AbelLiouville formula (7). 8. there exists a solution v of (16) such that t→∞ lim v(t) = c.88 5 Exercises 1. of (16) tends to a nonzero limit as t → ∞ and for any c ∈ RN . 4. 2. 3. ∞) if and only if 0 A(s)ds is bounded from below. ∞). Show that Φ−1 (t) is t bounded on [0. ∞). Prove that the solution set of (2) forms a vector space over either the real ﬁeld or the ﬁeld of complex numbers. Let Φ be a fundamental matrix solution of (2). 5. and a fundamental solution Φ. ∞) ⊂ I and assume that all solutions of (2) are bounded on [0. Let [0. Assume that ∞ 0 B(s)ds < ∞. Let B : [0. Show that any solution. If this is the case. Then all solutions of u′ = B(t)u are bounded on [0. ∞) ⊂ I and assume that all solutions of (2) are bounded on [0. Prove Proposition 5. Further assume that the matrix Φ−1 (t) is bounded on [0. prove that no solution u of (2) may satisfy u(t) → 0 as t → ∞ unless u ≡ 0. ∞). 7. Ψ = CΦ need not be a fundamental solution. Let [0. 6. ∞). . not the trivial solution. Also give an example to show that for a nonsingular N × N constant matrix C. ∞) → RN ×N be continuous and such that ∞ 0 A(s) − B(s)ds < ∞. Prove Proposition 1. Assume the conditions of the previous exercise.
2T. Show that there exists a nonsingular C 1 matrix L(t) such that the substitution u = L(t)v reduces (12) to a constant coeﬃcient system v ′ = Bv. Use the Jordan normal form and block multiplication rules for matrices to deduce that there exists a square matrix X such that Q = eX . Let Q be a nonsingular square matrix. 10. where the period should be the minimal period. Prove Proposition 7 using what has been developed about the matrix exponential etA . where both A and f are T −periodic. 12. . Using the above and the fact that etA may be written as etA = etλ et(A−λI) . ′ 14. 15. 11. Provide conditions on a = y1 (T ) + y2 (T ) in order that (14) have a T. Give an alternate veriﬁcation of Proposition 7 using the following facts from Linear Algebra for a given matrix A. · · · . Use the variation of constants formula to discuss the existence of T −periodic solutions of (1). Consider equation (1). (b) If λ and µ are diﬀerent eigenvalues of A their generalized eigenspaces only have the zero vector in common. then there exists a minimal integer q such that ker(A − λI)q = ker(A − λI)q+1 = ker(A − λI)q+2 = · · · and the dimension The space ker(A−λI)q is called the generalized eigenspace associated with λ. 13.5. (a) Let λ be an eigenvalue of A of algebraic multiplicity m. mT − periodic solution. Give necessary and suﬃcient conditions in order that all solutions u of (12) satisfy t→∞ dim ker(A − λI)q = m. EXERCISES 89 9. lim u(t) = 0. deduce the proposition.
90 .
u). using the linear theory developed in the previous chapter together with the implicit function theorem and degree theory. c ∈ RN . where A : R → RN × RN f : R × RN → RN are continuous and T −periodic with respect to t. some of the basic existence results about periodic solutions of periodic nonlinear systems of ordinary diﬀerential equations. we call it resonant.4) that Φ has the form 91 . otherwise. On the other hand it follows from Floquet theory (Section VI. 2 Preliminaries We recall from Chapter 6 that the set of all solutions of the equation u′ = A(t)u + g(t). We call the equation nonresonant provided the linear system u′ = A(t)u (2) (1) has as its only T − periodic solution the trivial one. In particular. (4) where Φ is a fundamental matrix solution of the linear system (2). (3) is given by t u(t) = Φ(t)c + Φ(t) t0 Φ−1 (s)g(s)ds. we shall mainly be concerned with systems of the form u′ = A(t)u + f (t.Chapter VII Periodic Solutions 1 Introduction In this chapter we shall develop.
(5) We note that equation (5) is uniquely solvable for every g. The following section is devoted to results of this type. That is. As we may choose Φ such that Φ(0) = I (the N × N identity matrix).92 Φ(t) = C(t)etR . . Proposition VI.(6) This proposition allows us to formulate a ﬁxed point equation whose solution will determine T − periodic solutions of equation (1). it follows that (with this choice) C(0) = C(T ) = I and u(T ) = eT R c + 0 T Φ−1 (s)g(s)ds. where C is a continuous nonsingular periodic matrix of period T and R is a constant matrix (viz. whenever the operator S given by equation (7) has a ﬁxed point in the space E. the periodic solution u is given by the following formula: u(t) = Φ(t) (I − eT R )−1 eT R T 0 Φ−1 (s)g(s)ds + 0 t Φ−1 (s)g(s)ds . c ∈ RN . (7) 2 Proposition Assume that I − eT R is nonsingular. T ]. hence u(0) = u(T ) = c if and only if T c = eT R c+ 0 Φ−1 (s)g(s)ds. u(s))ds . we have the following result: 1 Proposition Equation (3) has a unique T −periodic solution for every T −periodic forcing term g if and only if eT R − I is a nonsingular matrix. u(s))ds Φ−1 (s)f (s.9).T ] u(t) and let S : E → E be given by (Su)(t) = Φ(t) (I − eT R )−1 eT R + t 0 T 0 Φ−1 (s)f (s. If this is the case. RN ) : u(0) = u(T )} with u = maxt∈[0. 3 Perturbations of Nonresonant Equations In the following let E = {u ∈ C([0. if and only if I − eT R is a nonsingular matrix. then (1) has a T − periodic solution u. c ∈ RN .
PERTURBATIONS OF NONRESONANT EQUATIONS For f as given above let us deﬁne P (r) = max{f (t. provided that KP (r) ≤ r. u) : 0 ≤ t ≤ T. We have the following theorem. u) has a T − periodic solution u. for some r suﬃciently large. Since S is completely continuous the result follows from the Schauder ﬁxed point theorem (Theorem III. u ≤ r}. and S will have a ﬁxed point in B(r). then (1) has a T − periodic solution u.4. which holds. As a corollary we immediately obtain: 4 Corollary Assume that A and f are as above and that I − eT R is nonsingular.3. The above corollary may be considerably extended using the global continuation theorem Theorem IV. then u′ = A(t)u + ǫf (t. Using the operator S associated with equation (10) we obtain for u ∈ B(r) Su ≤ ǫKP (r). 93 (8) 3 Theorem Assume that A and f are as above and that I − eT R is nonsingular. Let us deﬁne B(r) = {u ∈ E : u ≤ r}. by condition (9). where S is the operator deﬁned by equation (7) and K is a constant that depends only on the matrix A. Hence S : B(r) → B(r). whenever lim inf r→∞ P (r) = 0.30). Proof. Namely we have the following result. provided that ǫ is suﬃciently small. there exists ǫ = 0 such that ǫKP (r) ≤ r. r (9) where P is deﬁned by (8). (10) . thus for given r > 0. Proof. then for u ∈ B(r) we obtain Su ≤ KP (r).
e. u)(t) = Φ(t) (I − eT R )−1 eT R + t 0 T 0 Φ−1 (s)ǫf (s. T ]. Then there exists a continuum C + ⊂ S+ (C − ⊂ S− ) such that: + − 1. ∞) : (u. then a T −periodic extension of a solution of (11) will be a T − periodic solution of the equation. Proof. 4 4. u). where S(ǫ. We view (11) as a perturbation of the equation u′ = 0. C0 ∩ E = {0} (C0 ∩ E = {0}). ǫ) solves (10)} and S− = {(u. u) = 0. we are in the case of resonance. ǫ) ∈ E × (−∞. We now let E = {u ∈ C([0. We shall now consider the equation subject to constraint (11) where f : R × RN → RN is continuous. RN )} with u = maxt∈[0. 0]). The proof follows immediately from Theorem IV. ǫ) solves (10)}. 0] : (u. 2.T ] u(t) and let S : E → E be given by t (Su)(t) = u(T ) + 0 f (s. Let S+ = {(u. u(s))ds Φ−1 (s)ǫf (s. i.1 Resonant Equations Preliminaries u′ = f (t. u(0) = u(T ). ǫ) ∈ E × [0. C + is unbounded in E × [0. (12) . ∞) (C − is unbounded in E × (−∞. u(s))ds. Should f be T −periodic with respect to t.94 5 Theorem Assume that A and f are as above and that I − eT R is nonsingular.4 by noting that the existence of T −periodic solutions of equation (10) is equivalent to the existence of solutions of the operator equation u − S(ǫ. u(s))ds .
(15) (14) where d(g. whose proof is immediate.2 Homotopy methods x ∈ RN g(x) → g(x) T = − 0 f (s. 0) is the Brouwer degree. ǫ)(t) = u(T ) + ǫ 0 f (s. For 0 ≤ λ ≤ 1. Further assume that d(g. u(s. We have the following theorem. u). λ))ds. λ) = λu(t) + (1 − λ)u(T ). T ] → Ω}. We deﬁne the bounded open set G ⊂ E by G = {u ∈ E : u : [0. 0 ≤ ǫ ≤ 1. 0 ≤ λ ≤ 1. Then problem (14) has a solution for all suﬃciently small ǫ. λ. λ) = λt + (1 − λ)T. gives us an operator equation in the space E whose solutions are solutions of the problem (11). 7 Theorem Assume that f is continuous and there exists a bounded open set Ω ⊂ RN such that the mapping g deﬁned by (13) does not vanish on ∂Ω. Proof. 1] → E by a(t. ¯ For u ∈ G deﬁne u(t.λ) (16) (17) (18) S(u. This lemma. u(0) = u(T ). and let a(t. 0) = 0. 0 ≤ λ ≤ 1. RESONANT EQUATIONS 95 then clearly S : E → E is a completely continuous mapping because of the continuity assumption on f. 4. Ω. 1] × [0. We shall next impose conditions on the ﬁnite dimensional vector ﬁeld (13) which will guarantee the existence of solutions of an associated problem u′ = ǫf (t. x)ds.4. (19) . Ω. The following lemma holds. 6 Lemma An element u ∈ E is a solution of (11) if and only if it is a ﬁxed point of the operator S given by (12). where ǫ is a small parameter. deﬁne S : E × [0.
2. ǫ). 0) = 0. 1] and ǫ suﬃciently small. 0). 8 Corollary Assume the hypotheses of Theorem 7 and assume that all possible solutions u. ǫ). λn ))ds ≡ un (t). 0.λn ) un (T ) + ǫn 0 f (s. say. Ω. 0) = d(id − S(·. and {ǫn }. ǫ). ǫn → 0. un (s. 0) = d(id − S(·. for 0 < ǫ ≤ 1. a)ds = 0. This we argue indirectly and hence obtain sequences {un } ⊂ ∂G. Ω. 0.96 Then S is a completely continuous mapping and the theorem will be proved once we show that d(id − S(·. ǫ). n = 1. ǫ) has no zeros on ∂G for all λ ∈ [0. ǫ) has a ﬁxed point in G which is equivalent to the assertion of the theorem. Ω. u and λ0 and the following must hold: u(t) ≡ u(T ) = a ∈ ∂Ω. 1]. if ǫ > 0 and (−1)N d(g. 0. Without loss in generality. in contradiction to the assumptions of the theorem. λn ) → u(t. λ. Then (11) has a T − periodic solution. G. 0) = d(id − S(·. of equation (14) are such that u ∈ ∂G. ǫ). . To show that (20) holds we ﬁrst show that S(·. 1. G. which further implies that T f (s. G. 1. such that a(t. 0 ≤ t ≤ T. Thus d(id − S(·. Hence also un (t. λ0 ) ≡ u(T ). 0) = d(g. for all ǫ suﬃciently small. we may assume that the sequences mentioned converge to. ǫ). {λn } ⊂ [0. λn ))ds = 0. 0 where a ∈ ∂Ω. · · · . for if this is the case. where G is / given by (16). S(·. 0) by the homotopy invariance property of LeraySchauder degree. completing the proof. On the other hand d(id − S(·. G. (20) for ǫ suﬃciently small. un (s. G ∩ RN . and hence T 0 f (s. 0. if ǫ < 0. 1. 0) = 0.
We note that (15) holds for such choices of Ω for any R > 0. (23) (22) We next shall show that the hypotheses of Theorem 7 and Corollary 8 may be satisﬁed by choosing Ω= u= x y : x < R. as follows from the substitution T (21) y =x− e(s)ds. 9 Theorem Assume that T < 2π. the result will follow. We note that. Hence. u) = y −h(x)y − x − e(t) .4. 0 We hence shall make that assumption. Then for every continuous T − periodic forcing term e. (25) . is a solution of (14) whenever x satisﬁes x′′ + ǫh(x)x′ + ǫ2 x = ǫ2 e(t). nothing else is T assumed about h. RESONANT EQUATIONS 97 4. Now u= x y . assume that 0 e(s)ds = 0. and put u= x y .3 A Li´nard type equation e In this section we apply Corollary 8 to prove the existence of periodic solutions of Li´nard type oscillators of the form e x′′ + h(x)x′ + x = e(t). without loss in generality. In order to apply our earlier results. y < R . we may. We shall prove the following result. if we are able to provide a priori bounds for solutions of equation (14) for 0 < ǫ ≤ 1 for f given as above. equation (21) has a T − periodic response x. since aside from the continuity assumption. we convert (21) into a system as follows: x′ = y y ′ = −h(x)y − x − e(t). f (t. (24) where R is a suﬃciently large constant. where e:R→R is a continuous T − periodic forcing term and h:R→R is a continuous mapping.
e L2 . from which follows a bound on x′ ∞ which is independent of ǫ. p = e Then x′′ ∞ ≤ ǫq x′ ∞ + ǫ2 (M + p). 2 L2 T2 ′ x 4π 2 (27) we obtain from (26) 1− T2 4π 2 x′ 2 L2 ≤ −ǫ2 x. we ﬁnd that T x(s)ds = 0. (29) from which. e x L2 2 L2 + ǫ2 x T 0 2 L2 = ǫ2 x. These considerations complete the proof Theorem 9.98 Integrating (25) from 0 to T. (30) providing an a priori bound on x 2πT 4π 2 − T 2 x≤M ∞. (26) = ≤ x(s)e(s)ds. by Landau’ s inequality (Exercise 6. ∞. for 0 ≤ ǫ ≤ 1. q = max h(x). we obtain x′ 2 ∞ ≤ 4M (ǫq x′ ∞ + ǫ2 (M + p)). 0 Multiplying (25) by x and integrating we obtain − x′ where x. since 2 L2 . e L2 . Hence. We let e L2 = M. (28) from which follows that x′ L2 ≤ 2πT 4π 2 − T 2 e L2 . below). we obtain x ∞ ≤ √ T 2πT 4π 2 − T 2 e L2 . Now. in turn. .
v : [0. only has the trivial solution as a T −periodic solution. Ω. To prove the existence of a T − periodic solution (u. v) u(0) = u(T ). 0 ≤ λ ≤ 1. Proof. T ].2. 0) is the Brouwer degree. deﬁne as in Subsection 4. v(s))ds. 0 ≤ λ ≤ 1. p + q = N. We shall impose conditions on the ﬁnite dimensional vector ﬁeld x ∈ Rp → g(x) T g(x) = − 0 f (s. Further B is a q × q constant matrix with the property that the system v ′ = Bv is nonresonant. Rp ) × C([0. u(s). (37) G = {(u. Then problem (32) has a solution for all suﬃciently small ǫ. u. (33) (32) (31) where d(g. v). 0) = 0. T ] → Ω. 10 Theorem Assume the above and there exists a bounded open set Ω ⊂ Rp such that the mapping g deﬁned by (31) does not vanish on ∂Ω. We have the following theorem. x. and let a(t. Rq ). (35) t (34) (36) . which will guarantee the existence of solutions of an associated problem u′ = ǫf (t. i.2. where ǫ is a small parameter and f : R × Rp × Rq → Rp h : R × Rp × Rq → Rq are continuous and T −periodic with respect to t. Further assume that d(g. Ω. λ) = λu(t) + (1 − λ)u(T ). We consider equation (34) as an equation in the Banach space E = C([0. Let Λ be a bounded neighborhood of 0 ∈ Rq . u(t. v(0) = v(T ). v) ∈ G. u(s). v) of equation (32) is equivalent to establishing the existence of a solution of u(t) = u(T ) + ǫ 0 f (s. RESONANT EQUATIONS 99 4. v) ∈ E : u : [0. T ]. v ′ = Bv + ǫh(t. T ] → Λ}. v(s))ds t v(t) = eBt v(T ) + ǫeBt 0 e−Bs h(s.4 Partial resonance This section is a continuation of of what has been discussed in Subsection 4.4.e. λ) = λt + (1 − λ)T. We deﬁne the bounded open set G ⊂ E by ¯ For (u. 0)ds. u.
This we argue in a manner similar to the proof of Theorem 7. 1] and ǫ suﬃciently small. λ). of equation (32) are such that u. 0) = 0. λ. Let f satisfy for some R > 0 f (t. ǫ)). x) · x = 0. 11 Corollary Assume the hypotheses of Theorem 10 and assume that all possible solutions u. ·.λ) (38) Then S is a completely continuous mapping and the theorem will be proved once we show that d(id − S(·. v ∈ ∂G. ǫ). Use Theorem 5 to show that equation (1) has a T −periodic solution provided the set of T −periodic solutions of (10) is a priori bounded for 0 ≤ ǫ < 1. t S2 (u. G. u(s). 2. 1] × [0.100 For 0 ≤ λ ≤ 1. 0. x = R. ǫ). where G / is given by (35). Prove that (11) has a T −periodic solution u with u(t) < R. ǫ)(t) = u(T ) + ǫ 0 f (s. λ. 0) T BT sgn det(I − e )d(−ǫ 0 f (s. ˜ where G = C([0. 1. 1. ǫ) has a ﬁxed point in G which is equivalent to the assertion of the theorem. Consider equation (1) with A a constant matrix. T ]. ǫ). ·. 0) = 0. 0 ≤ ǫ ≤ 1 deﬁne S = (S1 . Prove Corollary 8. for all ǫ suﬃciently small. On the other hand = = = d(id − S(·. 0 ≤ t ≤ T. (39) for ǫ suﬃciently small. As before. Ω. 0) = d(id − S(·. 0)d(id − S2 (·. Give conditions that I − eT R be nonsingular. ·0. ·. 0) T d((−ǫ 0 f (s. S2 ) : E × [0. id − S2 (·. ·. 0)ds. 0 ≤ t ≤ T. Ω. 0)) T ˜ d(−ǫ 0 f (s. ·. v. λv(s))ds . S(·. To show that (39) holds we ﬁrst show that S(·. G. for 0 < ǫ ≤ 1. completing the proof. ·. 3. G. 1. . u(s. for if this is the case. G. ·. 0) by the homotopy invariance property of LeraySchauder degree. G. ǫ). G. Then (32) has a T − periodic solution for ǫ = 1. Λ). (0. ǫ) has no zeros on ∂G for all λ ∈ [0. ·0. we obtain the following corollary. where I − eT R is given as in Theorem 3. v. ǫ). 1] → E by S1 (u. v(s))ds a(t. v. 5 Exercises 1. ǫ)(t) = eBt v(T ) + λǫeBt 0 e−Bs h(s. 0. 0)ds. λ. Hence d(id − S(·. 0)ds.
Let x ∈ C 2 [0. Verify inequality (27). Show that T ≥ 2π. ∞). x) · n(x) = 0. T ] → Ω. Use Taylor expansions to prove Landau’s inequality x′ 2 ∞ ≤4 x ∞ x′′ ∞. Prove Corollary 11. 6. EXERCISES 4.5. 101 where for each x ∈ ∂Ω. 0 ≤ t ≤ T. 9. Let Ω ⊂ RN be an open convex set with 0 ∈ Ω and let f satisfy f (t. equation (21) with e ≡ 0) e has a nontrivial T −periodic solution x. Assume that the unforced Li´nard equation (i. x ∈ ∂Ω.e. 5. Complete the details in the proof of Theorem 10. n(x) is an outer normal vector to Ω at x. Prove that (11) has a T −periodic solution u : [0. 8. . 7.
102 .
103 . then discussing the behavior of another solution v of this equation relative to the solution u. Thus we may. an assumption we shall henceforth make. (1) 2 Stability Concepts There are various stability concepts which are important in the asymptotic behavior of systems of diﬀerential equations. 0) ≡ 0. If u : [t0 . u). ∞) → RN is a given solution of (1). assume that (1) has the trivial solution as a reference solution. z + u(t)) − f (t. without loss in generality. In the case of constant or periodic coeﬃcients we found criteria which describe the asymptotic behavior of solutions (viz. Proposition 7 and Exercise 11 of Chapter 6.e. where f : R × RN → RN is a continuous function. i. We shall discuss here some of them and their interrelationships. f (t.Chapter VIII Stability Theory 1 Introduction In Chapter VI we studied in detail linear and perturbed linear systems of diﬀerential equations.e. i. u(t)). discussing the behavior of the diﬀerence v − u is equivalent to studying the behavior of the solution z = v − u of the equation z ′ = f (t. (2) relative to the trivial solution z ≡ 0.) In this chapter we shall consider similar problems for general systems of the form u′ = f (t.
t1 ≤ t < ∞. 4. 3 Example 1. i. ∞) and satisﬁes v(t) < ǫ. ∞). (v) strongly stable (s.s) on [t0 .a.s).s ⇔ a. t1 ≥ t0 exists on [t1 .e.s) on [t0 . ∞). if for every ǫ > 0 there exists δ > 0 such that any solution v of (1) with v(t0 ) < δ exists on [t0 . The zero solution of u′ = −u is uniformly asymptotically stable. The zero solution of u′ = u2 is unstable.a. The following examples of scalar diﬀerential equations will serve to illustrate the various concepts. ∞) and satisﬁes v(t) < ǫ. if it is stable and limt→∞ v(t) = 0.104 1 Deﬁnition We say that the trivial solution of (1) is: (i) stable (s) on [t0 . The zero solution of u′ = 0 is stable but not asymptotically stable. (iii) unstable (us). t0 ≤ t < ∞.a.s ⇒ a. (v) uniformly asymptotically stable (u. then the above implications take the from u. 3. (ii) asymptotically stable (a.s ⇓ ⇔ s. ∞). where v is as in (i). f is independent of t. if for every ǫ > 0 there exists δ > 0 such that any solution v of (1) with v(t1 ) < δ exists on [t0 .s ⇓ u. t0 ≤ t < ∞. ∞) and satisﬁes v(t) < ǫ. ∞) and satisﬁes v(t) < ǫ.s ⇒ u. 2.s) on [t0 . .s ⇓ s. The zero solution of u′ = a(t)u is asymptotically stable if and only if t t limt→∞ t0 a(s)ds = −∞. if for every ǫ > 0 there exists δ > 0 such that any solution v of (1) with v(t1 ) < δ. Letting a(t) = sin log t + cos log t − α one sees that asymptotic stability holds but uniform stability does not. if it is uniformly stable and there exists δ > 0 such that for all ǫ > 0 there exists T = T (ǫ) such that any solution v of (1) with v(t1 ) < δ.s ⇒ ⇓ s. If the equation (1) is autonomous. 2 Proposition The following implications are valid: u. It is uniformly stable if and only if t1 a(s)ds is bounded above for t ≥ t1 ≥ t0 . t1 + T ≤ t < ∞. t1 ≥ t0 exists on [t1 . if it is not stable. ∞). (iv) uniformly stable (u.
Then the second part of the theorem guarantees that the equation is uniformly stable. t0 ≤ t < ∞. there exists T = T (ǫ) > 0 such that if ξ < δ. t0 ≤ s ≤ t < ∞. 0 < ǫ < δ. (iii) strongly stable if and only if there exists K > 0 such that Φ(t) ≤ K. if we put T = − α log K . Moreover. t1 ≥ t0 . uniformly stable. t0 ≤ t < ∞. . RN ))) u′ = A(t)u. Φ−1 (t) ≤ K. 4 Theorem Let Φ be a fundamental matrix solution of (3). for each 1 ǫ ǫ. then for ξ ∈ RN . (3) a particular stability property of any solution is equivalent to that stability property of the trivial solution. (iv) asymptotically stable if and only if t→∞ (4) (5) (6) lim Φ(t) = 0. We shall demonstrate the last part of the theorem and leave the demonstration of the remaining parts as an exercise. t ≥ t1 + T. (7) (v) uniformly asymptotically stable if and only if there exist K > 0. the N × N identity matrix. if t1 + T < t. t0 ≤ s ≤ t < ∞. then there exists δ > 0 and for all ǫ. STABILITY OF LINEAR EQUATIONS 105 3 Stability of Linear Equations In the case of a linear system (A ∈ C(R → L(RN . if the equation is uniformly asymptotically stable. Then equation (3) is : (i) stable if and only if there exists K > 0 such that Φ(t) ≤ K. Thus we have uniform asymptotic stability. (ii) uniformly stable if and only if there exists K > 0 such that Φ(t)Φ−1 (s) ≤ K. We may assume without loss in generality that Φ(t0 ) = I. Conversely. The stability concepts may be expressed in terms of conditions imposed on a fundamental matrix Φ. Thus. α > 0 such that Φ(t)Φ−1 (s) ≤ Ke−α(t−s) . 0 < ǫ < K. since conditions (4) (8) will hold for any fundamental matrix if and only if they hold for a particular one. (8) Proof.3. ξ ≤ 1. Thus one may ascribe that property to the equation and talk about the equation (3) being stable. etc. let us assume that (8) holds. we have −1 −α(t−t1 ) Φ(t)Φ (t1 )ξ ≤ Ke . then Φ(t)Φ−1 (t1 )ξ < ǫ.
Φ(t + h)Φ−1 (t) ≤ K. The following concept of the measure of a matrix due to Lozinskii and Dahlquist (see [7]) is particularly useful in numerical computations. we obtain for some integer n. t ≥ t0 . δ Furthermore. = Ke−nαT e−αT δ ǫ < K δ e−α(t−t1 ) . Using the AbelLiouville formula. t ≥ t1 ≥ t0 . We provide a brief discussion. It is strongly stable if and only if all eigenvalues of A have zero real part and are semisimple. then it is strongly stable if and only if t lim inf t→∞ t0 traceA(s)ds > −∞. 6 Theorem Equation (3) is unstable whenever t lim sup t→∞ t0 traceA(s)ds = ∞.106 In particular Φ(t + T )Φ−1 (t)ξ < ǫ. t0 ≤ t. we get Φ(t)Φ−1 (t1 ) ≤ Ke−nαT In the case that the matrix A is independent of t one obtains the following corollary. 5 Corollary Equation (3) is stable if and only if every eigenvalue of A has nonpositive real part and those with zero real part are semisimple. (10) Additional stability criteria for linear systems abound. (9) If (3) is stable. ǫ ǫ 1 Letting α = − T log δ . we obtain the following result. . It is asymptotically stable if and only if all eigenvalues have negative real part. ξ < δ. since we have uniform stability. and Φ(t)Φ−1 (t1 ) ≤ Φ(t)Φ−1 (t1 + nT )Φ(t1 + nT )Φ−1 (t1 ) ≤ KΦ(t1 + nT )Φ−1 (t1 + (n − 1)T ) · · · Φ(t1 + T )Φ−1 (t1 ) ǫ n ≤K δ . 0 ≤ h ≤ T. or ǫ ξ Φ(t + T )Φ−1 (t)  < δ δ and thus Φ(t + T )Φ−1 (t) ≤ ǫ < 1. If t ≥ t1 . that t1 + nT ≤ t < t1 + (n + 1)T.
µ(A) ≤ A. Its proof is again left as an exercise. 3. 11 Theorem The system (3) is: . e−tµ(−A) ≤ etA ≤ etµ(A) . µ(αA) = αµ(A). Furthermore u(t0 )e − Rt t0 µ(−A(s))ds ≤ u(t) ≤ u(t0 )e t0 . 10 Corollary For any N × N constant matrix A the following inequalities hold. STABILITY OF LINEAR EQUATIONS 7 Deﬁnition For an N × N matrix A we deﬁne µ(A) = lim I + hA − I .3. for constant matrices A. 9 Proposition Let A : [t0 . then the function r(t) = u(t) has a right ′ ′ derivative r+ (t) at every point t for every norm  ·  in RN and r+ (t) satisﬁes ′ r+ (t) − µ(A(t))r(t) ≤ 0. ∞) → RN ×N be a continuous matrix and let u be a solution of (3) on [t0 . Then u(t)e − Rt t0 µ(A(s))ds . h→0+ h 107 (11) where  ·  is a matrix norm induced by a norm  ·  in RN and I is the N × N identity matrix. (15) This proposition has. If u is a solution of equation (3). 4. ∞). (12) Using this inequality we obtain the following proposition. the immediate corollary. t0 ≤ t < ∞ (13) is a nonincreasing function of t and u(t)e Rt t0 µ(−A(s))ds . α ≥ 0. 2. The following theorem provides stability criteria for the system (3) in terms of conditions on the measure of the coeﬃcient matrix. µ(A + B) ≤ µ(A) + µ(B). We have the following proposition: 8 Proposition For any N × N matrix A. t0 ≤ t < ∞ Rt µ(A(s))ds (14) is a nondecreasing function of t. µ(A) exists and satisﬁes: 1. µ(A) − µ(B) ≤ A − B.
uniformly stable. where f : R × RN → RN is a continuous function with f (t. u(s))ds . then u satisﬁes the integral equation u(t) = Φ(t) Φ−1 (t0 )u(t0 ) + t t0 (16) Φ−1 (s)f (s. 2. 5. where further results are given. if t lim inf t→∞ t0 µ(−A(s))ds = −∞. (17) where γ is some positive continuous function and A is a continuous N ×N matrix deﬁned on R. if µ(A(t)) ≤ 0. if t t→∞ lim t0 µ(A(s))ds = −∞. t ≥ t0 . The results to be discussed are consequences of the variation of constants formula (Proposition VI. Throughout this section Φ(t) will denote a fundamental matrix solution of the homogeneous (unperturbed) linear problem (3). 4 Stability of Nonlinear Equations In this section we shall consider stability properties of nonlinear equations of the form u′ = A(t)u + f (t.108 1. We refer to [7]. [20]. unstable. [5]. x ∈ RN . if µ(A(t)) ≤ −α < 0. stable. asymptotically stable. 4. See also the exercises below. (18) The following theorem will have uniform and strong stability as a consequence. x) ≤ γ(t)x. u). uniformly asymptotically stable. t ≥ t0 . and [21]. . if t lim sup t→∞ t0 µ(A(s))ds < ∞. We hence know that if u is a solution of equation (16). 3. We shall only present a sample of results.6) and the stability theorem Theorem 4.
α > 0 such that Φ(t)Φ−1 (s) ≤ Ke−α(t−s) . α K (19) a constant. t ≥ t1 ≥ t0 . t ≥ t1 ≥ t0 . A further stability result is the following. Its proof is delegated to the exercises. K Rt t1 γ(s)ds ≤ Lu(t1 ). Hence t u(t) ≤ Ku(t1 ) + K and u(t) ≤ Ku(t1 )e where L = Ke K R∞ t0 t1 γ(s)u(s)ds. If u is a solution of of (16) it also satisﬁes (18). Then there exists a positive constant L = L(t0 ) such that any solution u of (16) is deﬁned for t ≥ t0 and satisﬁes u(t) ≤ Lu(t1 ). To prove the remaining part of the theorem. If in addition Φ(t) → 0 as t → ∞. u(s))ds ∞ t1 Now t t1 Φ(t)Φ−1 (s)f (s. .4. 14 Theorem Assume that there exist K > 0. Then any solution u of (16) is u(t) ≤ Ke−β(t−t1 ) u(t1 ). and f satisﬁes (17) with γ < deﬁned for t ≥ t0 and satisﬁes where β = α − γK > 0. u(s))ds Φ(t)Φ−1 (s)f (s. whenever the unperturbed system is strongly stable. t ≥ t1 . The conditions of Theorem 12 also imply that the zero solution of the perturbed system is strongly stable. γ(s)ds . STABILITY OF NONLINEAR EQUATIONS 12 Theorem Let f satisfy (17) with Φ(t)Φ −1 ∞ 109 γ(s)ds < ∞. t0 ≤ s ≤ t < ∞. 13 Remark It follows from Theorem 4 that the conditions of the above theorem imply that the perturbed system (16) is uniformly (asymptotically) stable whenever the unperturbed system is uniformly (asymptotically) stable. t0 ≤ s ≤ t < ∞. Proof. then u(t) → 0 as t → ∞. t ≥ t1 . u(s))ds ≤ KLu(t0 ) γ(s)ds. Further assume that (s) ≤ K. This together with the fact that Φ(t) → 0 as t → ∞ completes the proof. we note from (1) that u(t) ≤ Φ(t) Φ−1 (t0 )u(t0 ) +  + t t1 t1 t0 Φ−1 (s)f (s.
then as t varies. Obviously x = 0 = y is a stationary solution of (21). about the stability of constant solutions e of systems of the form u′ = f (t. y(t)) and ∇v · (x′ . the orbit tends to the origin and the zero solution attracts all orbits. To illustrate. y) = x2 − 2axy + y 2 = constant. whenever k is (and positive if k is). y(t)). u). where f : R × RN → RN is a continuous function. That this limit must be zero follows by an indirect argument. 5 5. . Computing the scalar product of the tangent vector of an orbit with the gradient to an ellipse we ﬁnd: ∇v · (x′ .1 Lyapunov Stability Introduction In this section we shall introduce some geometric ideas. the orbit will cross members of the above family of ellipses. thus being guided to the origin. That is. y ′ ) = d dt v(x(t). x′ = ax − y + kx x2 + y 2 y ′ = x − ay + ky x2 + y 2 . y ′ ) = 2k x2 + y 2 x2 + y 2 − 2axy . (22) a family of ellipses which share the origin as a common center. Of course. (21) (20) where a is a constant a < 1. 0) = 0 and we discuss the stability properties of the trivial solution u ≡ 0. v(x(t). y(t)) will be a strictly decreasing function and hence should approach a limit as t → ∞. y(t)) : t ≥ t0 } associated with (21). Now consider the family of curves in R2 given by v(x. which is clearly negative. and hence appears asymptotically stable.110 15 Remark It again follows from Theorem 4 that the conditions of the above theorem imply that the perturbed system (16) is uniformly asymptotically stable whenever the unperturbed system is uniformly asymptotically stable. Thus. we may assume that f (t. The geometric ideas amount to constructing level surfaces in RN which shrink to 0 and which have the property that orbits associated with (20) cross these level surfaces transversally toward the origin. let us consider the following example. If we consider an orbit {(x(t). which were ﬁrst formulated by Poincar´ and Lyapunov. we may view v(x(t). As observed earlier. if k < 0. y(t)) as a norm of the point (x(t).
Proof. if it is positive deﬁnite and there exists a continuous increasing function ψ : [0. 16 Deﬁnition The functional v is called: 1. 3. u(t)) is nonincreasing with respect to t.2 Let Lyapunov functions v : R × RN → R. decrescent. t ∈ R be a continuous functional. t1 ≥ t0 the function v ∗ (t) = v(t. with ψ(0) = 0. Then v ∗ (t) = v(t. If v is also decrescent we obtain the stronger result. Since φ is nondecreasing. We have the following stability criteria. 0) = 0. u(t)) = v ∗ (t) ≤ v ∗ (t0 ) = v(t0 . for u0  < δ. v(t. then the trivial solution of (20) is stable. x ∈ RN . LYAPUNOV STABILITY 111 5. u(t)) is nonincreasing with respect to t. then the trivial solution of (20) is uniformly stable. ∞). t ≥ t0 . We shall introduce the following terminology. 17 Theorem Let there exist a positive deﬁnite functional v and δ0 > 0 such that for every solution u of (20) with u(t0 ) ≤ δ0 . if it is positive deﬁnite and r→∞ lim φ(r) = ∞.5. 2. ∞). t ≥ t0 . u0 ) < φ(ǫ). and ψ(x) ≥ v(t. and hence v ∗ (t) ≤ v ∗ (t0 ) = v(t0 . φ(r) = 0. u0 ). x). the result follows. Let u be a solution of (20) with u(t0 ) = u0 . the function v ∗ (t) = v(t. ∞) → [0. if there exists a continuous nondecreasing function φ : [0. x). Let 0 < ǫ ≤ δ0 be given and choose δ = δ(ǫ) such that v(t0 . r = 0 and φ(x) ≤ v(t. u(t)) is nonincreasing with respect to t. Therefore φ(u(t)) ≤ v(t. x ∈ RN . positive deﬁnite. ∞) → [0. with φ(0) = 0. . u0 ) < φ(ǫ). 18 Theorem Let there exist a positive deﬁnite functional v which is decrescent and δ0 > 0 such that for every solution u of (20) with u(t1 ) ≤ δ0 . radially unbounded.
Let u be a solution of (20) with u(t1 ) = u0 with u0  < δ. if u is any solution of (23) with u(t) = x. u(t)) = v ∗ (t) ≤ v ∗ (t1 ) = v(t1 .112 Proof. then lim v(t + h. As an example we consider the two dimensional system x′ = a(t)y + b(t)x x2 + y 2 y ′ = −a(t)x + b(t)y x2 + y 2 . 3. x) ≤ ψ(x). there exists a continuous increasing function ψ : [0. x0 ) < 0. then. dt ∂t we obtain dv ∗ = 2b(t) x2 + y 2 dt 2 ≤ 0. 19 Theorem Assume there exists a continuous functional v : R × RN → R with the properties: 1. x0  < δ such that v(t1 . u(t + h)) − v(t. y) = x2 + y 2 . it follows from Theorem 18 that the trivial solution of (23) is uniformly stable. since dv ∗ ∂v = + ∇v · f (t. h h→0+ where c is a continuous increasing function with c(0) = 0. there exists x0 . . u). Then the trivial solution of (23) is unstable. 2. such that ψ(0) = 0 and v(t. ∞). We choose v(x. u0 ) ≤ ψ(u0 ) < ψ(δ) = φ(ǫ). for all δ > 0. (23) where the coeﬃcient functions a and b are continuous with b ≤ 0. Since v is positive deﬁnite and decrescent. Then φ(u(t)) ≤ v(t. and t1 ≥ t0 . x) ≤ −c(x). The result now follows from the monotonicity assumption on φ. We next prove an instability theorem. We have already shown that the trivial solution is stable. Let 0 < ǫ ≤ δ0 be given and let δ = ψ −1 (φ(ǫ)). ∞) → [0.
u(t)) ≤ v(t0 . We ﬁnally develop some asymptotic stability criteria. u(t)). u(t)) t→∞ exists. Thus v(t0 . x0 ) < 0. Then v(t. ∞) and satisﬁes u(t) < ǫ. x0 ) ≤ ψ(u(t)) and ψ −1 (v(t0 . then v0 = lim v(t. with c a continuous increasing function and c(0) = 0. x) such that dv(t. dt dt for every solution u of (20) with u(t0 ) ≤ δ0 . 20 Theorem Let there exist a positive deﬁnite functional v(t. t→∞ lim φ(u(t)) = 0. u(t)) = −∞. . u(t)) ≤ ψ(u(t)) ≤ ψ(ǫ). Hence. u(t)) is nonincreasing and hence for t ≥ t0 . We therefore have t v(t. u(t)) ≤ v(t0 . An easy indirect argument shows that v0 = 0. since v is positive deﬁnite. u(t)) dv ∗ = ≤ −c(v(t. (24) The third property above implies that v(t. If v is also decrescent. Assume the trivial solution is stable. Then the asymptotic stability is uniform. The hypotheses imply that the trivial solution is stable. t0 v(t. LYAPUNOV STABILITY 113 Proof. as was proved earlier. v(t. and thus t→∞ lim v(t. Proof. contradicting (24).5. if u is a solution of (20) with u(t) ≤ δ0 . x0 )) ≤ u(t). t0 ≤ t < ∞. x0 )). Choose x0 . x0 ) < 0 and let u be a solution with u(t0 ) = x0 . u(t)) ≤ v(t0 . Then for every ǫ > 0 there exists δ > 0 such that any solution u of (23) with u(t0 ) < δ exists on [t0 . Hence. Then the trivial solution is asymptotically stable. x0 ) − (t − t0 )c(ψ −1 (v(t0 . x0  < δ such that v(t0 . x0 ) − which by (24) implies that c(u(s))ds.
The type of Lyapunov functional we shall be looking for are of the form v(x) = xT Bx. ∞) × RN → RN be a continuous function such that g(t. This we do by showing how appropriate Lyapunov functionals may be constructed. uniformly with respect to t ∈ [t0 . −λ is not an eigenvalue. The proof that the trivial solution is uniformly asymptotically stable. to. We shall consider the equation u′ = Au + g(t. To proceed along these lines we need some linear algebra results. u) (25) and show how to construct Lyapunov functionals to test the stability of the trivial solution of this system. 21 Proposition Let A be a constant N × N matrix having the property that for any eigenvalue λ of A.e. if B can be found so that C = AT B+BA has certain deﬁniteness properties. On the space of N × N matrices deﬁne the bounded linear operator L by L(B) = AT B + BA. then the results of the previous section may be applied to determine stability or instability of the trivial solution. Proof. u(t)) dt T (26) T T =u A B + BA u + g (t. there exists a unique N × N matrix B such that C = AT B + BA. Then for any N × N matrix C. once more study perturbed linear systems. x) = o(x). we are looking for v as a quadratic form.3 Perturbed linear systems Let A be a constant N × N matrix and let g : [t0 . u). T Hence. i. is left as an exercise. 5. since φ is increasing. where B is a constant N × N matrix. u)Bu + u Bg(t. where the associated linear system is a constant coeﬃcient system. ∞). If u is a solution of (25) then dv ∗ dt = dv(t. given A. whenever v is also decrescent. In the next section we shall employ the stability criteria just derived. The construction is an exercise in linear algebra.114 implying that t→∞ lim u(t) = 0. .
22 Corollary Let A be a constant N × N matrix. i. Let S = {λ ∈ C : λ = λ1 + λ2 }. Let v(x) = xT Bx. Choose 0 < µ ≤ r0 and consider the matrix A1 = A − µI. Then for any negative deﬁnite N × N matrix C. there exists a polynomial p such that p(F ) = I. such that λ(= 0) ∈ S implies that λ > r0 . for any integer n ≥ 1. i. there exists r0 > 0. on the other hand.e. there exists µ > 0 and a unique N ×N matrix B such that 2µB+C = AT B+BA. Once we show the latter. and let u be a solution of u′ = Au. Then dv(u(t)) dv ∗ = = uT AT B + BA u = uT Cu ≤ −µu2 . 23 Corollary Let A be a constant N × N matrix having the property that all eigenvalues λ of A have negative real parts. hence for any polynomial p Bp(A − µI) = p(−AT )B. (27) Since. which. a unique matrix B such that C = AT B + BA1 . From this follows B(A − µI)n = −AT n B. Thus let µ be an eigenvalue of L. From which follows that µ is the sum of two eigenvalues of A. Since S is a ﬁnite set. where λ1 and λ2 are eigenvalues of A. u(0) = x0 = 0. which may be applied since all eigenvalues of A have negative real part. (27) implies that A − µI and −AT must have a common eigenvalue. there exists a unique positive deﬁnite N × N matrix B such that C = AT B + BA. dt dt . 1 2µBC = AT B + BA. the result follows. hence it will be a bijection provided it does not have 0 as an eigenvalue. there exists a nonzero matrix B such that L(B) = AT B + BA = µB. We may now apply Proposition 21 to the matrix A1 and ﬁnd for a given matrix C. Hence AT B + B(A − µI) = 0. This proposition has the following corollary.. Then for any N × N matrix C.e. Proof. by hypothesis cannot equal 0. LYAPUNOV STABILITY 115 Then L may be viewed as a bounded linear operator of RN ×N to itself. Let C be a negative deﬁnite matrix and let B be given by Proposition 21. Proof. if F and G are two matrices with no common eigenvalues.5. p(G) = 0.
116 since C is negative deﬁnite. Since limt→∞ u(t) = 0 (all eigenvalues of A have negative real part!), it follows that limt→∞ v(u(t)) = 0. We also have
t
v(u(t)) ≤ v(x0 ) −
µu(s)2 ds,
0
from which follows that v(x0 ) > 0. Hence B is positive deﬁnite. The next corollary follows from stability theory for linear equations and what has just been discussed. 24 Corollary A necessary and suﬃcient condition that an N × N matrix A have all of its eigenvalues with negative real part is that there exists a unique positive deﬁnite matrix B such that AT B + BA = −I. We next consider the nonlinear problem (25) with g(t, x) = o(x), (28)
uniformly with respect to t ∈ [t0 , ∞), and show that for certain types of matrices A the trivial solution of the perturbed system has the same stability property as that of the unperturbed problem. The class of matrices we shall consider is the following. 25 Deﬁnition We call an N × N matrix A critical if all its eigenvalues have nonpositive real part and there exists at least one eigenvalue with zero real part. We call it noncritical otherwise. 26 Theorem Assume A is a noncritical N × N matrix and let g satisfy (28). Then the stability behavior of the trivial solution of (25) is the same as that of the trivial solution of u′ = Au, i.e the trivial solution of (25) is uniformly asymptotically stable if all eigenvalues of A have negative real part and it is unstable if A has an eigenvalue with positive real part. Proof. (i) Assume all eigenvalues of A have negative real part. By the above exists a unique positive deﬁnite matrix B such that AT B + BA = −I. Let v(x) = xT Bx. Then v is positive deﬁnite and if u is a solution of (25) it satisﬁes dv ∗ dt = dv(u(t)) dt
2 T T
(29)
= −u(t) + g (t, u)Bu + u Bg(t, u). Now g T (t, u)Bu + uT Bg(t, u) ≤ 2g(t, u)Bu.
5. LYAPUNOV STABILITY Choose r > 0 such that x ≤ r implies g(t, u) ≤ then dv ∗ 1 ≤ − u(t)2 , dt 2 1 −1 B u, 4
117
as long as u(t) ≤ r. The result now follows from Theorem 20. (ii) Let A have an eigenvalue with positive real part. Then there exists µ > 0 such that 2µ < λi + λj , for all eigenvalues λi , λj of A, and a unique matrix B such that AT B + BA = 2µB − I, as follows from Corollary 22. We note that B cannot be positive deﬁnite nor positive semideﬁnite for otherwise we must have, letting v(x) = xT Bx, dv ∗ = 2µv ∗ (t) − u(t)2 , dt or
t
e−2µt v ∗ (t) − v ∗ (0) = − i.e 0 ≤ v ∗ (0) −
t 0
0
e−2µs u(s)2 ds
e−2µs u(s)2 ds
for any solution u, contradicting the fact that solutions u exist for which
t 0
e−2µs u(s)2 ds
becomes unbounded as t → ∞ (see Chapter ??). Hence there exists x0 = 0, of arbitrarily small norm, so that v(x0 ) < 0. Let u be a solution of (25). If the trivial solution were stable, then u(t) ≤ r for some r > 0. Again letting v(x) = xT Bx we obtain dv ∗ = 2µv ∗ (t) − u(t)2 + g T (t, u)Bu + uT Bg(t, u). dt We can choose r so small that 2g(t, u)Bu ≤ hence e−2µt v ∗ (t) − v ∗ (0) ≤ − 1 2
t 0
1 2 u , u ≤ r, 2
e−2µs u(s)2 ds,
118 i.e. v ∗ (t) ≤ e2µt v ∗ (0) → −∞, contradicting that v ∗ (t) is bounded for bounded u. Hence u cannot stay bounded, and we have instability. If it is the case that the matrix A is a critical matrix, the trivial solution of the linear system may still be stable or it may be unstable. In either case, one may construct examples, where the trivial solution of the perturbed problem has either the same or opposite stability behavior as the unperturbed system.
6
Exercises
1. Prove Proposition 2. 2. Verify the last part of Example 3. 3. Prove Theorem 4. 4. Establish a result similar to Corollary 5 for linear periodic systems. 5. Prove Theorem 6. 6. Show that the zero solution of the scalar equation x′′ + a(t)x = 0 is strongly stable, whenever it is stable. 7. Prove Proposition 8. 8. Establish inequality (12). 9. Prove Proposition 9.
10. Prove Corollary 10. 11. Prove Theorem 11. 12. Show that if satisﬁes
t→∞ ∞
µ(A(s))ds exists, then any nonzero solution u of (3)
lim sup u(s) < ∞. On the other hand, if
t→∞ ∞
µ(−A(s))ds exists, then
0 = lim inf u(s) ≤ ∞. 13. If A is a constant N × N matrix show that µ(A) is an upper bound for the real parts of the eigenvalues of A.
respectively. Complete the proof of Theorem 20. then the system u′ = (A(t) + B(t))u is also uniformly asymptotically stable. Prove that if A2 = µ(A) = i j a2 . then the system u′ = (A(t) + B(t))u is uniformly (asymptotically) stable. ∞) → RN ×N is continuous and satisﬁes ∞ B(s)ds < ∞. N 16. the square root of the largest eigenvalue of AT A and the largest eigenvalue of 1 AT + A . 19. ∞) → RN ×N is continuous and satisﬁes t→∞ lim B(t) = 0. If the linear system (3) is uniformly asymptotically stable and B : [t0 . Verify the following table: x i 119 maxi x  xi  i i maxi maxk A k i µ(A) aik  aik  maxi aii + maxk akk + λ∗ k=i aik  aik  i=k xi 2 λ ∗ where λ∗ . 2 15. 17. λ∗ are. Prove Theorem 14. then ij trace(A) √ . . If the linear system (3) is uniformly (asymptotically) stable and B : [t0 .6. EXERCISES 14. 18.
y) = show that the trivial solution of x′′ + a 1 − x2 x′ + x = 0. where f and g are continuous functions with g(0) = 0. 0 The equation is equivalent to the system x′ = y − F (x).120 20. 1 Using the functional v(x. where F (x) = x f (s)ds. Use the functional v(x. Let a be a positive constant. Prove Corollary 23. . β > 0 such that x 0 g(s)ds < β ⇒ x < α. Consider the situation of Proposition 21 and assume all eigenvalues of A have negative real part. x 0 g(s)ds prove that the trivial solu1 2 21. y ′ = −g(x). where again a is a positive constant? 22. Assume that there exist α > 0. y 2 + x2 to is asymptotically stable. y) = 2 y 2 + tion is asymptotically stable. 23. Consider the Li´nard oscillator e x′′ + f (x)x′ + g(x) = 0. What can one say about the zero solution of x′′ + a x2 − 1 x′ + x = 0. Show that for given C the matrix B of the Proposition is given by B= 0 ∞ eA t CeAt dt. T Hint: Find a diﬀerential equation that is satisﬁed by the matrix P (t) = t ∞ eA T (τ −t) CeA(τ −t) dτ and show that it is a constant matrix. and 0 < x < α ⇒ g(x)F (x) > 0.
25.6. EXERCISES 24. . Consider the system x′ = y + ax x2 + y 2 y ′ = −x + ay x2 + y 2 . Show that the trivial solution of x′ = −2y 3 y′ = x is stable. Contrast this with Theorem 26. Contrast this with Theorem 26. 121 Show that the trivial solution is stable if a > 0 and unstable if a < 0.
122 .
and let f : D → RN be a locally Lipschitz continuous mapping. which has the following properties: 123 . We let D be an open connected subset of RN . (3) U = ∪v∈D Iv × {v} ⊂ R × RN . we obtain a mapping Thus if we let u : Iu0 × D → D (t. we have for each u0 ∈ D a maximal interval of existence Iu0 and a function u : Iu0 t → D → u(t. we obtain a mapping u : U → D. where I is the maximal interval of existence of the solution u. t ∈ I. u0 ). We consider the initial value problem u′ = f (u) u(0) = u0 ∈ D (1) and seek suﬃcient conditions on subsets M ⊂ D for solutions of (1) to have the property that {u(t)} ⊂ M. u0 ). Chapter V). u0 ) → u(t. we shall present some of the basic results about invariant sets for solutions of initial value problems for systems of autonomous (i. We note that the initial value problem (1) is uniquely solvable since f satisﬁes a local Lipschitz condition (viz.Chapter IX Invariant Sets 1 Introduction In this chapter. N ≥ 1. (2) i. Under these assumptions. whenever u0 ∈ M.e.e.. time independent) ordinary diﬀerential equations.
u0 ). v) v the negative semiorbit of v. in this chapter. of course immediate. We have the following proposition. u u(S. v) the orbit of v. u(0. we call v ∈ D a stationary or critical point of the ﬂow. then Iv = R and γ(v) = γ + (v) = γ − (v) = {v}. 0 < t < T. We call v ∈ D a periodic point of period T of the ﬂow. the ﬂow u(·. that if v ∈ D is a stationary point. or 2. v) = u(t. If γ + (v) is relatively compact. v) v the positive semiorbit of v and γ − (v) = u((t− . for each u0 ∈ D. It is. We shall − use the following (standard) convention. ∀u0 ∈ D 3. t ∈ S ⊂ Iu0 }. one calls T > 0 its minimal period. Furthermore. s ∈ Iu0 and t ∈ Iu(s.124 1 Lemma 2. 2 Orbits and Flows Let u be the ﬂow determined by f via the initial value problem (1). v) = u(T. u is continuous. v) is injective. u0 ) = u0 . γ + (v) = u([0. u0 ) = u(t. if γ − (v) is relatively compact.u0 ) we have s + t ∈ Iu0 and u(s + t. 1. u(s. 0]). t+0 ). A mapping having the above properties is called a ﬂow on D and we shall henceforth. u0 ) = {v : v = u(t. . 2 Proposition Let u be the ﬂow determined by f and let v ∈ D. then Iv = R. v). whereas if γ(v) is relatively compact. If v is a periodic point which is not a critical point. whenever f (v) = 0. then t− (v) = −∞. whenever there exists T > 0. u0 )). t+ ). v is a stationary point. or 3. v is a periodic point having a minimal positive period. use this term freely and call u the ﬂow determined by f. then t+ (v) = +∞. We shall call γ(v) = u(Iv . v). Then either: 1. such that u(0. If S ⊂ Iu0 = (tu0 . provided u(0.
V ⊂ M ⊂ D and there exists a ˜ ˜ largest invariant set M . V ⊃ M and there exists a smallest invariant set M . ǫ). INVARIANT SETS 125 3 Invariant Sets A subset M ⊂ D is called positively invariant with respect to the the ﬂow u determined by f. (iii) A set M is positively invariant if and only if compM (the complement of M ) is negatively invariant. ∀v ∈ M. As a corollary we obtain: 4 Corollary (i) If a set M is positively invariant with respect to the ﬂow u. i.. v) ⊂ M. We similarly call M ⊂ D negatively invariant provided γ − (M ) ⊂ M. γ + (M ) ⊂ M. and invariant provided γ(M ) ⊂ M. then so are M and intM. If ∂M is invariant. then so is ∂M. R \ M. As a consequence V contains a largest invariant subset and is contained in a smallest invariant set. We have the following proposition: 3 Proposition Let u be the ﬂow determined by f and let V ⊂ D. Then M is positively invariant with respect to the ﬂow u determined by f if and only if for every v ∈ M lim inf t→0+ dist(v + tf (v). M ⊃ V.3. whenever γ + (v) ⊂ M. in particular will aid us in ﬁnding invariant sets.e. M ) = 0. Also there exists a largest negatively invariant ˜ ˜ subset M. (ii) A closed set M is positively invariant with respect to the ﬂow u if and only if for every v ∈ ∂M there exists ǫ > 0 such that u([0. then so are M . (iv) If a set M is invariant. which provides a relationship between the vector ﬁeld f and the set M (usually called the subtangent condition) in order that invariance holds. t (4) . and intM. M ⊂ V. We note that a set M is invariant if and only if it is both positively and negatively invariant. We have the following theorem. We now provide a geometric condition on ∂M which will guarantee the invariance of a set M and. 5 Theorem Let M ⊂ D be a closed set. Then there exists a smallest positively invariant subset M.
then a Taylor expansion yields. v). 0 ≤ t < ǫ. (5) We leave the proof as an exercise. w(0) = 0. v) ⊂ B. ǫ] there exists vt ∈ M such that w(t) = u(t. ∀v ∈ ∂M = φ−1 (0). ǫ]. because f satisﬁes a local Lipschitz condition. dist(v + tf (v). v)) − f (vt ) + vt + sf (vt ) − vt+s  ≤ w(t) + sLw(t) + dist(vt + tf (vt ). ∇φ(v) = 0. M ) ≤ u(t. if M is positively invariant. w(0) = 0. w(t) ≡ 0. 0 ≤ t < ǫ. We next consider some special cases where the set M is given as a smooth manifold. We remark that the set M given in the previous theorem will be negatively invariant provided the reverse inequality holds and hence invariant if and only if ∇φ(v) · f (v) = 0. 0 ≤ t < ǫ. Let v ∈ M. v) = v + tf (v) + o(t). . in which case φ(u(t. which implies w(t) ≤ 0. v) − vt ) and limt→0+ vt = v. We have the following theorem.e. Next. v) − vt+s  ≤ w(t) + u(t + s. or D+ e−Lt w(t) ≤ 0. Let M = φ−1 (−∞. u(t. i. Hence D+ w(t) ≤ Lw(t). then for each t ∈ [0. 6 Theorem Let φ ∈ C 1 (D. w(t + s) = u(t + s. Hence. ∀v ∈ ∂M. and assume (4) holds.e. t > 0. M ).e. It follows that for some constant L. We remark here that condition (4) only needs to be checked for points v ∈ ∂M since it obviously holds for interior points. proving the necessity of (4). then M is positively invariant with respect to the ﬂow determined by f if and only if ∇φ(v) · f (v) ≤ 0. 0]. Let w(t) = dist(u(t. M ). i. v) − u(t. ∀v ∈ ∂M = φ−1 (0). v)) ≡ 0. φ is a ﬁrst integral for the diﬀerential equation. i. R) be a function which is such that every value v ∈ φ−1 (0) is a regular value. let v ∈ M. Choose a compact neighborhood B of v such that B ⊂ D and choose ǫ > 0 so that u([0. v) − sf (u(t. v)) +sf (u(t.126 Proof. v) − v − tf (v) = o(t).
tn ց t− . then Γ− (v) ⊂ ∂D. then Γ+ (v) ⊂ ∂D. call it y. Proof. (iv) If Γ+ (v) = ∅ and bounded. tn ր ∞. u(0) = w. LIMIT SETS 127 4 Limit Sets In this section we shall consider semiorbits and study their limit sets. Since u(tn . and hence the maximal interval of existence of un will contain the interval [−tn . of u′ = f (u). ∞). For each n ≥ 1. v (7) The set Γ− (v) is the set of limit points of γ − (v) and is called the negative limit set of v. u(0) = u(tn . It then follows that t+ = ∞. which we relabel as {un (t)} converging to the solution. (iii) If γ + (v) is bounded. Then there exists a sequence v {tn }∞ . Let w ∈ Γ+ (v) ∩ D. v) is the unique solution of u′ = f (u). then t→t+ v lim dist u(t. We leave most of the proof to the exercises and only discuss the veriﬁcation of the last part of the proposition. v) → w. v). Thus let us assume that Γ+ (v)∩D = ∅. then Γ+ (v) = ∅ and compact. Namely we deﬁne the set Γ− (v) as follows: Γ− (v) = {w : ∃{tn }.) v For all the results discussed below for positive limit sets there is an analagous result for negative limit sets. v) → w}. u(tn . tn ր t+ . We deﬁne the set Γ+ (v) as follows: Γ+ (v) = {w : ∃{tn }. Γ+ (v) = ∩w∈γ +(v) γ + (w).) In a similar vein one deﬁnes limit sets for negative semiorbits γ − (v). we shall not state these results.4. v ∈ D. Γ+ (v) = 0. there will exist a subsequence of {un (t)}. u(tn . (v) Γ+ (v) ∩ D is an invariant set. (Again note that if t− > −∞. v (6) The set Γ+ (v) is the set of limit points of γ + (v) and is called the positive limit + set of v. such that u(tn . v ∈ D be the positive semiorbit associated with v ∈ D. (Note that if tv < ∞. We have the following proposition: 7 Proposition (ii) (i) γ + (v) = γ + (v) ∪ Γ+ (v). v). Let γ + (v). the function n=1 un (t) = u(t + tn . . v) → w}. v) → w.
b] the sequence {un (t)} will be deﬁned on [a. b]. v). Proof. Then for all v ∈ D. We already know (cf. Hence solutions through points of Γ+ (v) are deﬁned for all time and their orbit remains in Γ+ (v). Thus let φ : D → R be a C 1 function. w ∈ N } > 0. N ) < δ and hence there exists a 2 sequence {tn → ∞} such that dist(u(tn . δ there exist values of t arbitrarily large such that dist(u(t. M ) < 2 and values of t arbitrarily large such that dist(u(t. 8 Theorem If γ + (v) is contained in a compact subset K ⊂ D. Let δ = inf{v − w : v ∈ M. ∞). Proposition 7) that Γ+ (v) is a compact set. Furthermore for any t0 y(t0 ) = lim un (t0 ) = lim u(t0 + tn . ∀v ∈ K.e a continuum. . v). Thus we need to show it is connected. n→∞ n→∞ and hence y(t0 ) ∈ Γ (v). v)} 2 must have a convergent subsequence and hence has a limit point which is in neither M nor N. showing that Γ+ (v) ∩ D is an invariant set. M ) = δ . v). Since M ⊂ Γ+ (v) and N ⊂ Γ+ (v). Suppose it is not. Let ˜ K = {v ∈ K : φ′ (v) = 0} ˜ and let M be the largest invariant set contained in K. Then there exist nonempty disjoint compact sets M and N such that Γ+ (v) = M ∪ N. 9 Lemma Assume that φ′ (v) ≤ 0.1 LaSalle’s theorem In this section we shall return again to invariant sets and consider Lyapunov type functions and their use in determining invariant sets. M ) = 0. i. then Γ+ (v) = ∅ is a compact connected set. v). 10 Theorem Let there exist a compact set K ⊂ D such that φ′ (v) ≤ 0. The sequence {u(tn . We shall use the following notation φ′ (v) := ∇φ(v) · f (v). ∀v ∈ D. Then for all v ∈ D such that γ + (v) ⊂ K t→∞ lim dist (u(t. Since this interval is arbitrary it follows that y is deﬁned on (−∞. b] for n suﬃciently large and hence y will be deﬁned on [a. v). a contradiction.128 We note that given any compact interval [a. + 4. φ is constant on the set Γ+ (v) ∩ D.
Furthermore suppose that φ is bounded below and that φ(x) → ∞ as x → ∞. hence all solution orbits tend to the origin. k are positive constants. Also every orbit associated with f which crosses l. y) : y = 0}. This equation may be written as x′ = y k y′ = − m x − We choose φ(x.e. then. We shall need the following observation. ∀x ∈ RN . f (w) is not parallel to the direction of l. Thus we shall assume throughout this section that N = 2. not critical) point of f. As an example to illustrate Theorem 11 consider the oscillator mx′′ + hx′ + kx = 0. using the previous lemma. for all v ∈ RN .5. where m. y) = −hy 2 . Let v ∈ D such that γ + (v) ⊂ K. Then there exists a transversal l containing v in its relative interior. we have that φ is constant on Γ+ (v). v). which is an invariant set and hence contained in M. 11 Theorem (LaSalle’s Theorem) Assume that D = RN and let φ′ (x) ≤ 0. crosses l in the same direction. h. provided it contains only regular points and if for all w ∈ l. 1 my 2 + kx2 2 h m y. Hence E = {(x. 5 Two Dimensional Systems In this section we analyze limit sets for two dimensional systems in somewhat more detail and prove a classical theorem (the Theorem of Poincar´Bendixson) e about periodic orbits of such systems. TWO DIMENSIONAL SYSTEMS 129 Proof. . Let E = {v : φ′ (v) = 0}. where M is the largest invariant set contained in E. The largest invariant set contained in E is the origin. Let v ∈ D be a regular (i. M ) = 0. 12 Lemma Let v ∈ D be a regular point of f. We call a compact straight line segment l ⊂ D through v a transversal through v. y) = and obtain that φ′ (x. then t→∞ lim dist (u(t.
Let η ∈ R2 be any direction not parallel to f (v). We then may take l to be the intersection of the straight line through v with direction η and V . and thus also Γ+ (w) ⊂ Γ+ (v).e. Then any transversal containing w in its interior contains no other points of Γ+ (v) in its interior. and is bounded away from 0 on V. a ≤ t1 < t2 ≤ b. v) = u(t2 . We hence may apply the ∂t implicit function theorem to complete the proof. Then if Γ intersects l it does so at a ﬁnite number of points whose order on Γ is the same as the order on l. w) will cross l in time t. Proof. η × f (v) = 0. i. On the other hand. 16 Lemma Let γ + (v) be a semiorbit which does not intersect itself and which is contained in a compact set K ⊂ D and let all points in Γ+ (v) be regular points of f. Let w ∈ Γ+ (v). The next lemma follows immediately from the previous two. Then Γ+ (v) contains a periodic orbit. ηf (w) sin θ). Let z ∈ Γ+ (w). . Let L(t. t < ǫ. v) = 0.130 Proof. where θ is the angle between η and f (w). w) = au1 (t. The proof relies on the Jordan curve theorem and is left as an exercise. It follows that the semiorbit γ + (w) must intersect l and by the above for an inﬁnite number of values of t. Let v be a regular point of f. it follows from Proposition 7 that Γ+ (v) is an invariant set and hence that γ + (w) ⊂ Γ+ (v). Proof. u(t. s0 ≤ s ≤ s1 }. w) + bu2 (t. and ∂L (0. b) · f (v) = 0. and let l be a transversal containing z in its relative interior. Let B be a disc centered at v containing only regular points of f. v). 13 Lemma Let v be an interior point of some transversal l. v) : a ≤ t ≤ b} be a closed arc of an orbit u associated with f which has the property that u(t1 . w) + c. v) = (a. u = (u1 . the previous lemma implies that all these point of intersection must be the same. Choose a neighborhood V of v consisting of regular points only. u2 ). 15 Lemma Let γ + (v) be a semiorbit which does not intersect itself and let w ∈ Γ+ (v) be a regular point of f. Let v ∈ int l and let l = {z : z = v + sη. 14 Lemma Let l be a transversal and let Γ = {w = u(t. where u(t. 0. The proof is completed by observing that η × f (w) = (0. ∀w ∈ V. Then L(0. We may restrict V further such that η × f (w) = 0. Then for every ǫ > 0 there exists a circular disc Dǫ with center at v such that for every w ∈ Dǫ . If the orbit is periodic it intersects l at most once. (here × is the cross product in R3 ). w) is the solution with initial condition w and au1 + bu2 + c = 0 is the equation of the straight line containing l.
Let u be the ﬂow determined by f (see Lemma 1). Prove Proposition 3 and Corollary 4. One hence concludes that in fact under the hypotheses of Lemma 16. Let Ω = intΓ. Then for each w ∈ Ω. t+0 ). ∞] are lower semiconu u u u tinuous functions of u0 . EXERCISES 131 It therefore follows from Lemma 16 that every point in Γ+ (v) is a point on some periodic orbit of a minimal positive period. γ(w) = Γ+ (v). We partially order the collection {Γα }α∈I . Prove Theorem 6. Show that if Iu0 = (t−0 . Then there exists at least one singular point of f in the interior of D. if for some w ∈ Γ+ (v). On the other hand. then −t−0 : D → [0. 17 Theorem (Poincar´Bendixson) Let γ + (v) be a semiorbit which does not e intersect itself and which is contained in a compact set K ⊂ D and let all points in Γ+ (v) be regular points of f. by saying that Γα ≤ Γβ ⇔ intΓα ⊂ intΓβ . 4. 6 Exercises 1. Also provide conditions in order that the mapping u be smooth. 2. Then Γ+ (v) is the orbit of a periodic solution uT with smallest positive period T. One now employs the Hausdorﬀ minimum principle together with Theorem 17 to arrive at a contradiction. ∞] and t+0 : D → [0. then Γ+ (v) \ γ(w) must be a relatively open set with A = Γ+ (v) \ γ(w) ∩ γ(w) = ∅. 3. We summarize the above in the following theorem. of all periodic orbits which are contained in Ω. Then f is continuous on Ω and does not vanish on Γ = ∂Ω. the limit set Γ+ (v) is a periodic orbit. it also follows from earlier results that Γ+ (v) is a compact connected set. say of class C 1 . Prove Lemma 1. 18 Theorem Let Γ be a periodic orbit of (1) which together with its interior is contained in a compact set K ⊂ D. . where I is an index set. Hence. Γ+ (w) is a periodic orbit. Let us assume that f has no stationary points in Ω. 5. Proof. One now easily obtains a contradiction by examining transversals through points of A.6. Prove Proposition 2.
0]. 12.e M is a parallelepiped. 13. Prove Theorem 18. Show that periodic orbits exist for all µ > 0 and that these orbits collapse to the origin as µ → 0. . Prove that Γ+ (u(0)) and Γ− (u(0)) are distinct periodic orbits provided they contain regular points only. Complete the proof of Lemma 12. where σ. Prove Lemma 14. Prove that rest points and periodic orbits are invariant sets. Also show that no periodic orbits exist for µ ≤ 0. 14. Let u be a solution whose interval of existence is R which is not periodic and let γ + (u(0)) and γ − (u(0)) be contained in a compact set K ⊂ D. where µ is a real parameter. 15. 8. 17. 16.132 6. Prove Theorem 6. Consider the three dimensional system of equations (the Lorenz system) x′ = −σx + σy y ′ = rx − y − xz z ′ = −bz + xy. Consider the system x′ y′ = y − x3 + µx = −x. Prove Proposition 7. i. Consider the special case where each φi is aﬃne linear. Prove Lemma 9. Extend Theorem 6 to the case where −1 m M = ∩i=1 vi (−∞. Prove Theorem 17. b are positive parameters. 11. 7. Find a family of ellipsoids which are positively invariant sets for the ﬂow determined by the system. 9. r. 10.
3. · · · . Results of the type proved here are referred to as Hopf bifurcation theorems. α) = 0. We establish the following theorem. α0 ). α0 ) and ±ni. be a C 2 mapping which is such that f (0. all α ∈ R. dt (1) and prove a theorem about the existence of nontrivial periodic solutions of this system. For some given value of α = α0 . n = 0. 2 A Hopf Bifurcation Theorem Let f : Rn × R → Rn . 1 Theorem Assume that f satisﬁes the following conditions: √ 1. The proof we give is base on the method of LyapunovSchmidt presented in Chapter II. i = −1 and −i are eigenvalues of fu (0. 2.Chapter X Hopf Bifurcation 1 Introduction This chapter is devoted to a version of the classical Hopf bifurcation theorem which establishes the existence of nontrivial periodic orbits of autonomous systems of diﬀerential equations which depend upon a parameter and for which the stability properties of the trivial solution changes as the parameter is varied. α) = 0. are not eigenvalues of fu (0. 133 . We consider the system of ordinary diﬀerential equations depending on a parameter α du + f (u.
Theorem II. This is guaranteed by the assumptions. with values of ρ close to 1. α(s)) solves the equation (2) du + ρf (u. i. We note that F (0. ρ. ρ. α1 ) is a nontrivial solution of (1) of period 2πρ1 . 2πρ1 ]. α0 ) of the two dimensional parameter (ρ. whenever 1 u(τ ). for all ρ ∈ R. u1 (t) < ǫ. ρ. α0 )u cannot be a linear homeomorphism of X onto Y.134 2. α ∈ R. 2π].e. α) = 0. α1 = α(s) and u1 (ρ1 t) = u(s)(τ + θ). continuous 2π− periodic functions endowed with the usual norms and deﬁne an operator F :X ×R×R (u. β(α0 ) = i. α) = 0. α) = 0. ρ. Then there exist positive numbers ǫ and η and a C 1 function 1 (u. α(0) = α0 . Re dα α0 = 0. u(s) = 0. ′ = d dτ . s = 0 and ρ(0) = 1. θ ∈ [0. We thus let X = C2π and Y = 1 C2π be Banach spaces of C . u(0) = 0. φ1 = Im(eiτ a(α0 )) are 2π− periodic solutions of u′ + fu (0. α) →Y → u′ + ρf (u. (7) (6) . α0 )u = 0. Proof. then there exists s ∈ (−η. (3) (4) Furthermore. 2π). ρ(s). dτ nontrivially. Thus the claim is that the value (1. 1 where C2π is the space of 2π periodic C 1 Rn − valued functions. τ = ρ1 t ∈ [0. such that (u(s). α) is a bifurcation value. and u = 0. in a neighborhood of α0 there is a curve of eigenvalues and eigenvectors fu (0. α) : (−η. τ = ρt is a solution of period 2π of (3). (5) Then F belongs to class C 2 and we seek nontrivial solutions of the equation F (u. since the functions φ0 = Re(eiτ a(α0 )). We note that u(t) will be a solution of (1) of period 2πρ. α). α close to α0 . η) such that ρ1 = ρ(s). respectively. η) → C2π × R × R.1 tells us that u′ + fu (0. if (u1 . with ρ1 − 1 < ǫ. α)a(α) = β(α)a(α) dβ a(α0 ) = 0. α1 − α0  < ǫ. t ∈ [0.
α. where ·. 1. α0 )v). 1. α − α0 ) (9) . s) = 0. α0 ). ρ. α0 . α)φ0 . We let U be a neighborhood of (0. (8) and evaluating at s = 0 we get Gv. Fu (0. A HOPF BIFURCATION THEOREM and they span the the kernel of Fu (0. s)v + · · · . ρ. = fv (0. α0 ) = {φ0 . α). α. s) + Gρ (0. ρ.α (0. as in the beginning of Chapter II. 0) and show that this derivative is a linear homeomorphism of V × R × R onto Y. φ1 = −φ′ }. 0 Thus. 0) in V × R × R × R and deﬁne G on U as follows G(v. 1. We now write. 1. ψ1 } forms a basis for ker{−u′ + T fu (0. we obtain G(v. s) evaluate the result at (0. α. α0 ) = {f ∈ Y : f. 0) = φ′ + ρfu (0. s=0 s = 0. α0 )u. α in terms of s in a neighborhood of 0 ∈ R. α) → G(v. X Y = = V ⊕W Z ⊕ T. 0) = (φ0 + v)′ + ρfu (0. 1. ρ. α0 . α0 . the diﬀerential equation adjoint operator of u′ + fu (0. ρ. the Kronecker delta. s)(α − α0 ) + Gv (0. We note that G is C 1 and G(v. in order to be able to apply the implicit function theorem. imFu (0. 1. α)(φ0 + v). α. α0 ) is a linear Fredholm mapping from X to Y having a two dimensional kernel as well as a two dimensional cokernel. for v. α0 . is closed in Y and that imFu (0. α0 . ρ. s)(ρ − 1) Gα (0. Hence G(0. Thus Fu (0. kerFu (0. 1. ρ − 1. 1. α0 )φ0 (α − α0 ) +(v ′ + fv (0. 0)(v. 1. Computing the Taylor expansion. ρ. we need to diﬀerentiate the map (v. In fact ′ ψ1 = ψ0 and φi . α. α0 . α)(φ0 + v). α0 .2. ψi = 0. · denotes the L2 inner product and {ψ0 . 0 135 It follows from the theory of linear diﬀerential equations that the image. 1. α. 1}.ρ. 1. α0 )φ0 (ρ − 1) +fvα (0. α0 ). We now want to solve the equation G(v. ρ. ψj = δij . 1. i = 0. α0 )u}. ρ. s) = G(0. ρ. s) = 1 s F (s(φ0 + v).
By the characterization of T. α0 )v is a linear homeomorphism of V onto T. Thus we need to show that the map (ρ. α0 )φ0 . α0 )φ0 . α0 )φ0 (ρ − 1) + fvα (0. α0 )φ0 . α0 )φ0 = −φ′ = φ1 . 2 Example Consider the nonlinear oscillator x′′ + x − α(1 − x2 )x′ = 0. ψi > (ρ − 1)+ < fvα (0. ψ0 >= 0.136 Note that the mapping v → v ′ + fv (0. Computing this latter expression. which by hypothesis is not zero. ψ0 >= Reβ ′ (0). α0 )φ0 (α − α0 ) ∈ T if and only if < fv (0. ψi > (α − α0 ) = 0. ψ0 > (α − α0 ) = (ρ − 1)+ < fvα (0. (11) This equation has for for certain small values of α nontrivial periodic solutions with periods close to 2π. 2. α0 )φ0 . (10) i = 1. α) such that fv (0. α0 )φ0 (α − α0 ) only belongs to T if ρ = 1 and α = α0 and for all ψ ∈ Z there exists a unique (ρ. α0 )φ0 . α0 )φ0 (ρ − 1) + fvα (0. α0 )φ0 (α − α0 ) = ψ. The following example of the classical Van der Pol oscillator from nonlinear electrical circuit theory (see [19]) will serve to illustrate the applicability of the theorem. α0 )φ0 (ρ − 1) + fvα (0. ψ1 > (α − α0 ) = which has only the trivial solution if and only if < fvα (0. we have that fv (0. (12) 0 0. Since fvα (0. The uniqueness assertion we leave as an exercise. α) → fv (0. . α0 )φ0 . < fvα (0. one obtains < fvα (0. For much further discussion of Hopf bifurcation we refer to [19]. 0 we may write equation (10) as two equations in the two unknowns ρ − 1 and α − α0 .
This solution is unique up to phase shift. for α = 0. or β ′ = α+2β = − 2 . s ∈ (−η.1 there exists η > 0 and continuous functions α(s). A HOPF BIFURCATION THEOREM We ﬁrst transform (12) into a system by setting u= and obtain u′ + 0 1 −1 −α + 0 u2 u2 1 = 0 0 u1 u2 = x x′ 137 (13) We hence obtain that f (u. dβ Letting α0 = 0. we get β(0) = ±i.2. ρ(s). Thus by Theorem II. and computing dα = β ′ we obtain 2ββ ′ +β ′ α+ −β 1 β = 0. η) such that α(0) = 0. α) = 0 1 −1 −α . . α) = and fu (0. 0 −1 1 −α + 0 u2 u2 1 whose eigenvalues satisfy β(α + β) + 1 = 0. ρ(0) = 1 and (12) has for s = 0 a nontrivial solution x(s) with period 2πρ(s).
138 .
q. ′ (1) where α and β are given constants and (without loss in generality. r positive on I. r ∈ C(I.Chapter XI SturmLiouville Boundary Value Problems 1 Introduction In this chapter we shall study a very classical problem in the theory of ordinary diﬀerential equations. namely linear second order diﬀerential equations which are parameter dependent and are subject to boundary conditions. Such a boundary value problem is called a SturmLiouville boundary value problem. R). 2 Linear Boundary Value Problems Let I = [a. We refer to the books [5]. 0 < β ≤ π). much more general problems and various other cases may be considered and similar theorems may be established. 139 . While the existence of eigenvalues (parameter values for which nontrivial solutions exist) and eigenfunctions (corresponding nontrivial solutions) follows easily from the abstract Riesz spectral theory for compact linear operators. with p. 0 ≤ α < π. Consider the linear diﬀerential equation where λ is a complex parameter. it is instructive to deduce the same conclusions using some of the results we have developed up to now for ordinary diﬀerential equations. We seek parameter values (eigenvalues) for which (1) has nontrivial solutions (eigensolutions or eigenfunctions) when it is subject to the set of boundary conditions x(a) cos α − p(a)x′ (a) sin α = 0 (2) x(b) cos β − p(b)x′ (b) sin β = 0. (p(t)x′ ) + (λr(t) + q(t))x = 0. b] be a compact interval and let p. While the theory presented below is for some rather speciﬁc cases. [6] and [21] where the subject is studied in some more detail. t ∈ I.
Let the diﬀerential operator L be deﬁned by L(x) = (px′ ) + qx. We next let u(t. Then.140 We note that (2) is equivalent to the requirement c1 x(a) + c2 x′ (a) = 0. if λ is an eigenvalue L(x) = −λrx. 1 Lemma Every eigenvalue of equation (1) subject to the boundary conditions (2) is real. ¯ x Hence b a b ′ (3) ¯ (¯L(x) − xL(¯)) dt = −(λ − λ) x x rx¯ dt. We introduce the following transformation (Pr¨fer transformation) u ρ= u2 + p2 (u′ )2 . φ = arctan u . p(a)u′ (a) = cos α. x Therefore ¯ x xL(x) − xL(¯) = −(λ − λ)rx¯ . Hence also ¯ x L(¯) = −λrx = −λr¯. pu′ Then ρ and φ are solutions of the diﬀerential equations ρ′ = − (λr + q) − and φ′ = 1 cos2 φ + (λr + q) sin2 φ. x a Integrating the latter expression and using the fact that both x and x satisfy ¯ the boundary conditions we obtain the value 0 for this expression and hence ¯ λ = λ. then u = 0 and satisﬁes the ﬁrst set of boundary conditions. Proof. p (5) 1 ρ sin φ cos φ p (4) . c1  + c2  = 0 c3 x(b) + c4 x′ (b) = 0. c3  + c4  = 0. for some nontrivial x which satisﬁes the boundary conditions. λ) = u(t) be the solution of (1) which satisﬁes the (initial) conditions u(a) = sin α. We have the following lemma.
ρ may be determined by integrating a linear equation and hence u is determined. The eigenspace associated with each eigenvalue is one dimensional and the eigenfunctions associated with λk have precisely k simple zeros in (a. limλ→∞ φ(b. Then φ satisﬁes the following conditions: 1. λ) is a continuous strictly increasing function of λ. we ﬁnd it convenient to make the change of independent variable t s= a dτ . Proof.2. 3. 2 Lemma Let φ be the solution of (5) such that φ(a) = α. Using the above lemma we obtain the existence of eigenvalues. Proof. The ﬁrst part follows immediately from the discussion in Sections V. The equation β + kπ = φ(b. ′ = d . p(τ ) which transforms equation (1) to x′′ + p(λr + q)x = 0. 3 Theorem The boundary value problem (1)(2) has an unbounded inﬁnite sequence of eigenvalues λ0 < λ1 < λ2 < · · · with n→∞ lim λn = ∞. We have the following lemma describing the dependence of φ upon λ. limλ→−∞ φ(b. 1. .6. 2.6 to deduce the remaining parts of the lemma. We also have the following lemma. λ) = 0. for k = 0. λ) = ∞. namely we have the following theorem. · · · . λ) has a unique solution λk . φ(b. This set {λk }∞ has the desired k=0 properties. LINEAR BOUNDARY VALUE PROBLEMS 141 Further φ(a) = α. To prove the other parts of the lemma. once φ is known. (Note that the second equation depends upon φ only. ds (6) We now apply the Pr¨ fer transformation to (6) and use the comparison theorems u in Section V. b).5 and V. hence.
142 4 Lemma Let ui , i = j, k, j = k be eigenfunctions of the boundary value problem (1)(2) corresponding to the eigenvalues λj and λk . Then uj and uk are orthogonal with respect to the weight function r, i.e.
b
uj , uk =
a
ruj uk = 0.
(7)
In what is to follow we denote by {ui }∞ the set of eigenfunctions whose i=0 existence is guaranteed by Theorem 3 with ui an eigenfunction corresponding to λi , i = 0, 1, · · · which has been normalized so that
b a
ru2 = 1. i
(8)
We also consider the nonhomogeneous boundary value problem (p(t)x′ ) + (λr(t) + q(t))x = rh, t ∈ I,
′
(9)
where h ∈ L2 (a, b) is a given function, the equation being subject to the boundary conditions (2) and solutions being interpreted in the Carath´odory sense. e We have the following result. 5 Lemma For λ = λk equation (9) has a solution subject to the boundary conditions (2) if and only if
b
ruk h = 0.
a
If this is the case, and w is a particular solution of (9)(2), then any other solution has the form w + cuk , where c is an arbitrary constant. Proof. Let v be a solution of (p(t)x′ )′ + (λk r(t) + q(t))x = 0, which is linearly independent of uk , then c (uk v ′ − u′ v) = , k p where c is a nonzero constant. One veriﬁes that w(t) = 1 c
t b
v(t)
a
ruk h + uk
t
rvh
b a
is a solution of (9)(2) (for λ = λk ), whenever
ruk h = 0 holds.
3
Completeness of Eigenfunctions
We note that it follows from Lemma 5 that (9)(2) has a solution for every b λk , k = 0, 1, 2, · · · if and only if a ruk h = 0, for k = 0, 1, 2, · · · . Hence, since {ui }∞ forms an orthonormal system for the Hilbert space L2 (a, b) (i.e. L2 (a, b) r i=0 with weight function r deﬁning the inner product), {ui }∞ will be a complete i=0
3. COMPLETENESS OF EIGENFUNCTIONS
b
143
orthonormal system, once we can show that a uk h = 0, for k = 0, 1, 2, · · · , implies that h = 0 (see [37]). The aim of this section is to prove completeness. The following lemma will be needed in this discussion. 6 Lemma If λ = λk , k = 0, 1, · · · (9) has a solution subject to the boundary conditions (2) for every h ∈ L2 (a, b). Proof. For λ = λk , k = 0, 1, · · · we let u be a nontrivial solution of (1) which satisﬁes the ﬁrst boundary condition of (2) and let v be a nontrivial solution of (1) which satisﬁes the second boundary condition of (2). Then c uv ′ − u′ v = p with c a nonzero constant. Deﬁne the Green’s function G(t, s) = Then
b
1 c
v(t)u(s), a ≤ s ≤ t v(s)u(t), t ≤ s ≤ b.
(10)
w(t) =
a
G(t, s)r(s)h(s)ds
is the unique solution of (9)  (2). We have the following corollary. 7 Corollary Lemma 6 deﬁnes a continuous mapping G : L2 (a, b) → C 1 [a, b], by h → G(h) = w. Further Gh, w = h, Gw . Proof. We merely need to examine the deﬁnition of G(t, s) as given by equation (10). Let us now let S = {w ∈ L2 (a, b) : ui , h = 0, i = 0, 1, 2, · · ·}. Using the deﬁnition of G we obtain the lemma. 8 Lemma G : S → S. (12) (11)
We note that S is a linear manifold in L2 (a, b) which is weakly closed, i.e. if {xn } ⊂ S is a sequence such that xn , h → x, h , ∀h ∈ L2 (a, b), then x ∈ S.
144 9 Lemma If S = {0}, then there exists x ∈ S such that G(x), x = 0. Proof. If G(x), x = 0 for all x ∈ S, then, since S is a linear manifold, we have for all x, y ∈ S and α ∈ R 0 = G(x + αy), x + αy = 2α G(y), x ,
in particular, choosing x = G(y) we obtain a contradiction, since for y = 0 G(y) = 0. 10 Lemma If S = {0}, then there exists x ∈ S\{0} and µ = 0 such that G(x) = µx. Proof. Since there exists x ∈ S such that G(x), x = 0 we set µ= inf{ G(x), x : x ∈ S, x = 1, if G(x), x ≤ 0, ∀x ∈ S} sup{ G(x), x : x ∈ S, x = 1, if G(x), x > 0, for some x ∈ S}.
We easily see that there exists x0 ∈ S, x0 = 1 such that G(x0 ), x0 = µ = 0. If S is one dimensional, then G(x0 ) = µx0 . If S has dimension greater than 1, x then there exists 0 = y ∈ S such that y, x0 = 0. Letting z = √0 +ǫy we ﬁnd 1+ǫ2 that G(z), z has an extremum at ǫ = 0 and thus obtain that G(x0 ), y = 0, for any y ∈ S with y, x0 = 0. Hence since G(x0 ) − µx0 , x0 = 0 it follows that G(x0 ), G(x0 ) − µx0 = 0 and thus G(x0 ) − µx0 , G(x0 ) − µx0 = 0, proving that µ is an eigenvalue. Combining the above results we obtain the following completeness theorem. 11 Theorem The set of eigenfunctions {ui }∞ forms a complete orthonormal sysi=0 tem for the Hilbert space L2 (a, b). r Proof. Following the above reasoning, we merely need to show that S = {0}. If this is not the case, we obtain, by Lemma 10, a nonzero element h ∈ S and a nonzero number µ such that G(h) = µh. On the other hand w = G(h) satisﬁes the boundary conditions (2) and solves (9); hence h satisﬁes the boundary conditions and solves the equation (p(t)h′ ) + (λr(t) + q(t))h =
′
r h, t ∈ I, µ
(13)
1 i.e. λ − µ = λj for some j. Hence h = cuj for some nonzero constant c, contradicting that h ∈ S.
Let the diﬀerential operator L be given by L(x) = (tx′ )′ + m2 x. Also prove that G : L2 (a. 3. b) is a compact mapping. Prove that the existence and completeness part of the above theory may be established provided the functions satisfy p(a) = p(b). 8. 0 < t < 1. r(a) = r(b). 7. s) = ∞ i=0 ui (t)ui (s) . Supply the details for the proof of Lemma 2. q(a) = q(b).4. Find the set of eigenvalues and eigenfunctions for the boundary value problem x′′ + λx = 0 x(0) = 0 = x′ (1). b]2 and that ∂G(t. λ − λi where the convergence is in the L2 norm. x(0) = x(2π) x′ (0) = x′ (2π). Discuss the behavior of this ∂t derivative as t → s. 6. b) → L2 (a. Let G(t.s) is continuous for t = s. Show that G(t. Provide the details of the proof of Corollary 7. 2. Prove Lemma 4. Apply the previous exercise to ﬁnd the eigenvalues and eigenfunctions for the boundary value problem x′′ + λx = 0. 5. t and consider the eigenvalue problem L(x) = −λtx. EXERCISES 145 4 Exercises 1. s) be deﬁned by equation (10). Prove that the Green’s function given by (10) is continuous on the square [a. Replace the boundary conditions (2) by the periodic boundary conditions x(a) = x(b). . x′ (a) = x′ (b). 9. 4.
146 In this case the hypotheses imposed earlier are not applicable and other types of boundary conditions than those given by (3) must be sought in order that a development parallel to that given in Section 2 may be made. Establish such a theory for this singular problem. . Extend this to more general singular problems.
Numerical Continuation Methods. [9] J. Solution branches for mappings in cones. 131–143. . 77–91. Foundations of Modern Analysis. Stability and Asymptotic Behavior of Diﬀerential Equations. New e York.. SpringerVerlag. Allgower and K.. Math. 1965. Theory of Ordinary Diﬀerential Equations. New York. An Introduction. Soc. On the structure of solutions of nonlinear eigenvalue problems. Birkhoff and G. [6] R. Rabinowitz. Academic Press. 1969. Nonlinear SturmLiouville problems and topological degree. Surveys No. Eberly. Austral.Bibliography [1] R. Existence and uniqueness theorems for nonlinear boundary value problems. Georg. New York. pp. Mech. Amer. Waltham. Indiana Math.. [3] J. J. Proc. Cole. Levinson. J. Cronin. Coppel. [14] J. Math. 1975. Blaisdell. [5] E. Soc. 11. and applications. 24 (1994). Coddington and N. Math. Boston. Adams. Heath.. London Math. [11] [12] . Springer Verlag. [8] M. Bebernes and D. pp. J. New York. 1964. 1083–1102. N. Math. Rota. [2] E. pp. McGraw Hill. 1964. Schmitt. Crandall and P. 27 (1973). 147 . Theory of Ordinary Diﬀerential Equations. pp. New York. 11 (1974). Ordinary Diﬀerential Equations. Dang and K. [4] G. Fixed Points and Topological Degree in Nonlinear Analysis. pp. AppletonCenturyCrofts. Mathematical Problems from Combustion Theory.. Dancer. 1968. Global structure of the solutions on nonlinear real analytic eigenvalue problems. [7] W. Soc. 1069–1076. 1955. 1989. Sobolev Spaces. New York. Dieudonn´. [10] E. Academic Press. Math. Rocky Mtn. 1990. 19 (1970). J. Providence. 747–765. [13] H. 23 (1974).
8 (1960). pp. Elliptic Partial Diﬀerential Equations of Second Order.. Krasnosels’kii. 73 (1951). Academic Press. Ordinary Diﬀerential Equations. 1992. [17] N. Polon. A global bifurcation theorem with applications to functional diﬀerential equations. Nonlinear Oscillations. J. Springer Verlag. Solvability and Bifurcations of Nonlinear Equations. [23] E. D. 1967. A theory of degree of mapping based on inﬁnitesimal analysis. Math. J. Perov. 319–338. Povolotskii. Functional Analysis. pp. Ann. 1978. AMS Translations. 1966. [21] P. pp. Degree theory. Heinz. 19 (1975). Lloyd. New York.. Longman Scientiﬁc and Technical Series. A. I. Zabreiko. [25] M. Opial. [18] D. G. 1983. Trudinger. Positive Solutions of Operator Equations. Wiley. Introductory Functional Analysis. Cambridge. Springer. 1978. Nagumo. Dunford and J. A. J. New York. Guckenheimer and P. [28] N. [29] M. New York. Dynamical Systems and Bifurcations of Vector Fields. Boston. Topology. Nussbaum. Kreyszig. Holmes. A. [30] R. Math. Kamke. Krasnosel’skii. Dugundji. P. [32] H. Schwartz. Olech. I. 1982. Zur Theorie der Systeme gew¨hnlicher Diﬀerentialgleichungen o II. Acta. A simple proof of a result of Z. . [19] J. Groningen. Math. 74–117. 322 (1981). Schmitt. 57–85. Essex. Mech. Peitgen. 199–325. Berlin. 1966.. A. Rutman. Gilbarg and N. [27] E. Saupe. Amer. Math. Hahn. pp. New York. pp. 1964. pp. 1964. 8 (1959). Allyn and Bacon. 1983. 10 (1962).148 BIBLIOGRAPHY ´ [15] P.. Berlin. Stability of Motion. Drabek. Nonlinear eigenvalue problems versus their ﬁnite diﬀerence approximations: Numerically irrelevant solutions. J. 61–63. pp. and P. Plane Vector Fields. Birkh¨user. 485–496. Mathematik. [26] M. 231–247. [24] M. Linear operators leaving invariant a cone in a Banach space. 58 (1932). An elementary analytic theory of the degree of mapping in ndimensional Euclidean space. Linear Operators: Part I. D. a [22] E. [16] J. Wiley. [20] W. Cambridge University Press. Krein and M. Noordhoﬀ. Hartman. reine angew. Sringer. Boston. [31] C. and K.
Anal. . 88 (1981). Introduction to Functional Analysis. 1977. Monthly. Whyburn. L. 462–475. [43] A. On bifurcation from inﬁnity. 1966. London. Royden. Schechter. Func. J. 1971. [36] H.. pp. Peitgen and K. Math. Equa. 1972. Rabinowitz.. McGraw Hill. 1969. Rudin. New York. New York. New York. [39] K. 487–513.. Taylor. Showalter. Hilbert Space Methods for Partial Diﬀerential Equations. Diﬀ. pp. Principles of Functional Analysis. pp. Pitman. On bifurcation from inﬁnity for potential operators. Schmitt and Z. Real Analysis. [35] . Amer.BIBLIOGRAPHY 149 [33] H. Schmitt. pp. New York. [41] R. [44] G. 7 (1971). J. Soc. Analytic Topolgy. Wiley and Sons. Diﬀerential and Integral equations. Amer. 1942. Providence. Academic Press. [38] M. 283 (1984). Gordon and Breach. Math. 933–943. O. Siegberg. [42] H. Schwartz. [34] P.. Wang. Global analysis of twoparameter elliptic eigenvalue problems. Nonlinear Functional Analysis. Macmillan Publishing Co.. [40] J. Amer. Soc. 14 (1973). 1988. Trans. Some global results for nonlinear eigenvalue problems. Real and Complex Analysis. Math. 3rd ed. 4 (1991). [37] W. Some historical remarks concerning degree theory. New York. pp. 125–139. E.. 57–95.
150 BIBLIOGRAPHY .
This action might not be possible to undo. Are you sure you want to continue?