Topics in Linear and Nonlinear

Functional Analysis
Gerald Teschl

Graduate Studies
in Mathematics
Volume (to appear)

American Mathematical Society

Providence, Rhode Island
Gerald Teschl
Fakultät für Mathematik
Oskar-Mogenstern-Platz 1
Universität Wien
1090 Wien, Austria

E-mail: Gerald.Teschl@univie.ac.at
URL: http://www.mat.univie.ac.at/~gerald/

2010 Mathematics subject classification. 46-01, 46E30, 47H10, 47H11, 58Exx,

76D05

Abstract. This manuscript provides a brief introduction to linear and non-

linear Functional Analysis. It covers basic Hilbert and Banach space theory
as well as some advanced topics like operator semigroups, mapping degrees
and fixed point theorems.

Keywords and phrases. Functional Analysis, Banach space, Hilbert space,

operator semigroup, mapping degree, fixed point theorem, differential equa-
tion, Navier–Stokes equation.

Typeset by AMS-LATEX and Makeindex.

Version: December 14, 2020
Copyright © 1998–2020 by Gerald Teschl
Contents

Preface vii

Part 1. Functional Analysis

Chapter 1. A first look at Banach and Hilbert spaces 3
§1.1. Introduction: Linear partial differential equations 3
§1.2. The Banach space of continuous functions 7
§1.3. The geometry of Hilbert spaces 17
§1.4. Completeness 23
§1.5. Compactness 25
§1.6. Bounded operators 27
§1.7. Sums and quotients of Banach spaces 33
§1.8. Spaces of continuous and differentiable functions 37
Chapter 2. Hilbert spaces 41
§2.1. Orthonormal bases 41
§2.2. The projection theorem and the Riesz lemma 48
§2.3. Operators defined via forms 50
§2.4. Orthogonal sums and tensor products 56
§2.5. Applications to Fourier series 58
Chapter 3. Compact operators 65
§3.1. Compact operators 65
§3.2. The spectral theorem for compact symmetric operators 68
§3.3. Applications to Sturm–Liouville operators 74

iii
iv Contents

§3.4. Estimating eigenvalues 82

§3.5. Singular value decomposition of compact operators 85
§3.6. Hilbert–Schmidt and trace class operators 89
Chapter 4. The main theorems about Banach spaces 97
§4.1. The Baire theorem and its consequences 97
§4.2. The Hahn–Banach theorem and its consequences 108
§4.3. The adjoint operator 116
§4.4. Weak convergence 122
Chapter 5. Further topics on Banach spaces 131
§5.1. The geometric Hahn–Banach theorem 131
§5.2. Convex sets and the Krein–Milman theorem 135
§5.3. Weak topologies 140
§5.4. Beyond Banach spaces: Locally convex spaces 144
§5.5. Uniformly convex spaces 151
Chapter 6. Bounded linear operators 155
§6.1. Banach algebras 155
§6.2. The C ∗ algebra of operators and the spectral theorem 164
§6.3. Spectral theory for bounded operators 167
§6.4. Spectral measures 173
§6.5. The Gelfand representation theorem 177
§6.6. Fredholm operators 184

Part 2. Nonlinear Functional Analysis

Chapter 7. Analysis in Banach spaces 193
§7.1. Single variable calculus in Banach spaces 193
§7.2. Multivariable calculus in Banach spaces 196
§7.3. Minimizing nonlinear functionals via calculus 208
§7.4. Minimizing nonlinear functionals via compactness 214
§7.5. Contraction principles 220
§7.6. Ordinary differential equations 224
§7.7. Bifurcation theory 230
Chapter 8. Operator semigroups 237
§8.1. Uniformly continuous operator groups 237
§8.2. Strongly continuous semigroups 239
Contents v

§8.3. Generator theorems 246

§8.4. Semilinear equations 258
Chapter 9. The nonlinear Schrödinger equation 261
§9.1. Local well-posedness in Hr for r > n
2 261
§9.2. Strichartz estimates 264
§9.3. Well-posedness in L2 and H1 268
§9.4. Blowup in H1 274
§9.5. Standing waves 276
Chapter 10. The Brouwer mapping degree 279
§10.1. Introduction 279
§10.2. Definition of the mapping degree and the determinant formula281
§10.3. Extension of the determinant formula 285
§10.4. The Brouwer fixed point theorem 292
§10.5. Kakutani’s fixed point theorem and applications to game
theory 294
§10.6. Further properties of the degree 297
§10.7. The Jordan curve theorem 300
Chapter 11. The Leray–Schauder mapping degree 301
§11.1. The mapping degree on finite dimensional Banach spaces 301
§11.2. Compact maps 302
§11.3. The Leray–Schauder mapping degree 304
§11.4. The Leray–Schauder principle and the Schauder fixed point
theorem 306
§11.5. Applications to integral and differential equations 308
Chapter 12. Monotone maps 315
§12.1. Monotone maps 315
§12.2. The nonlinear Lax–Milgram theorem 317
§12.3. The main theorem of monotone maps 318
Appendix A. Some set theory 321
Appendix B. Metric and topological spaces 329
§B.1. Basics 329
§B.2. Convergence and completeness 335
§B.3. Functions 339
vi Contents

§B.4. Product topologies 341

§B.5. Compactness 344
§B.6. Separation 350
§B.7. Connectedness 353
§B.8. Continuous functions on metric spaces 357
Bibliography 365
Glossary of notation 369
Index 373
Preface

The present manuscript was written for my course Functional Analysis given
at the University of Vienna in winter 2004 and 2009. The second part are
the notes for my course Nonlinear Functional Analysis held at the University
of Vienna in Summer 1998, 2001, and 2018. The two parts are essentially in-
dependent. In particular, the first part does not assume any knowledge from
measure theory (at the expense of hardly mentioning Lp spaces). However,
there is an accompanying part on Real Analysis [47], where these topics are
covered.
It is updated whenever I find some errors and extended from time to
time. Hence you might want to make sure that you have the most recent
version, which is available from
http://www.mat.univie.ac.at/~gerald/ftp/book-fa/
Please do not redistribute this file or put a copy on your personal
webpage but link to the page above.

Goals

The main goal of the present book is to give students a concise introduc-
tion which gets to some interesting results without much ado while using a
sufficiently general approach suitable for further studies. Still I have tried
to always start with some interesting special cases and then work my way
up to the general theory. While this unavoidably leads to some duplications,
it usually provides much better motivation and implies that the core ma-
terial always comes first (while the more general results are then optional).
Moreover, this book is not written under the assumption that it will be

vii
viii Preface

read linearly starting with the first chapter and ending with the last. Con-
sequently, I have tried to separate core and optional materials as much as
possible while keeping the optional parts as independent as possible.
Furthermore, my aim is not to present an encyclopedic treatment but to
provide the reader with a versatile toolbox for further study. Moreover, in
contradistinction to many other books, I do not have a particular direction
in mind and hence I am trying to give a broad introduction which should
prepare you for diverse fields such as spectral theory, partial differential
equations, or probability theory. This is related to the fact that I am working
in mathematical physics, an area where you never know what mathematical
theory you will need next.
I have tried to keep a balance between verbosity and clarity in the sense
that I have tried to provide sufficient detail for being able to follow the argu-
ments but without drowning the key ideas in boring details. In particular,
you will find a show this from time to time encouraging the reader to check
the claims made (these tasks typically involve only simple routine calcula-
tions). Moreover, to make the presentation student friendly, I have tried
to include many worked out examples within the main text. Some of them
are standard counterexamples pointing out the limitations of theorems (and
explaining why the assumptions are important). Others show how to use the
theory in the investigation of practical examples.

Preliminaries

The present manuscript is intended to be gentle when it comes to required

background. Of course I assume basic familiarity with analysis (real and
complex numbers, limits, differentiation, basic (Riemann) integration, open
sets) and linear algebra (finite dimensional vector spaces, matrices).
Apart from this natural assumptions I also expect some familiarity with
metric spaces and point set topology. However, only a few basic things are
required to begin with. This and much more is collected in the Appendix
and I will refer you there from time to time such that you can refresh your
memory should need arise. Moreover, you can always go there if you are
unsure about a certain term (using the extensive index) or if there should
be a need to clarify notation or conventions. I prefer this over referring
you to several other books which might not alway be readily available. For
convenience the Appendix contains full proofs in case one needs to fill some
gaps. As some things are only outlined (or outsourced to exercises) it will
require extra effort in case you see all this for the first time.
On the other hand I do not assume familiarity with Lebesgue integration
and consequently Lp spaces will only be briefly mentioned as the completion
Preface ix

of continuous functions with respect to the corresponding integral norms in

the first part. At a few places I also assume some basic results from complex
analysis but it will be sufficient to just believe them.
The second part of course requires basic familiarity with functional anal-
ysis and measure theory (Lebesgue and Sobolev spaces). But apart from this
it is again independent form the first two parts.

Content

Below follows a short description of each chapter together with some

hints which parts can be skipped.
Chapter 1. The first part starts with Fourier’s treatment of the heat
equation which led to the theory of Fourier analysis as well as the develop-
ment of spectral theory which drove much of the development of functional
analysis around the turn of the last century. In particular, the first chap-
ter tries to introduce and motivate some of the key concepts and should be
covered in detail except for Section 1.8 which introduces some interesting
examples for later use.
Chapter 2 discusses basic Hilbert space theory and should be considered
core material except for the last section discussing applications to Fourier
series. They will only be used in some examples and could be skipped in
case they are covered in a different course.
Chapter 3 develops basic spectral theory for compact self-adjoint op-
erators. The first core result is the spectral theorem for compact symmetric
(self-adjoint) operators which is then applied to Sturm–Liouville problems.
Of course this application could be skipped, but this would reduce the didac-
tical concept to absurdity. Nevertheless it is clearly possible to shorten the
material as non of it (including the follow-up section which touches upon
some more tools from spectral theory) will be required in later chapters.
The last two sections on singular value decompositions as well as Hilbert–
Schmidt and trace class operators cover important topics for applications,
but will again not be required later on.
Chapter 4 discusses what is typically considered as the core results
from Banach space theory. In order to keep the topological requirements
to a minimum some advanced topics are shifted to the following chapters.
Except for possibly the last section, which discusses some application to
minimizing nonlinear functionals, nothing should be skipped here.
Chapter 5 contains selected more advanced stuff. Except for the geo-
metric Hahn–Banach theorem, which is a prerequisite for the other sections,
the remaining sections are independent of each other.
x Preface

Chapter 6 develops spectral theory for bounded self-adjoint operators

via the framework of C ∗ algebras. Again a bottom-up approach is used such
that the core results are in the first two sections and the rest is optional.
Again the remaining four sections are independent of each other except for
the fact that Section 6.3, which contains the spectral theorem for compact
operators, and hence the Fredholm alternative for compact perturbations
of the identity, is of course used to identify compact perturbations of the
identity as premier examples of Fredholm operators.
Chapter 7 finally gives a brief introduction to operator semigroups,
which can be considered optional.
Finally, there is a part on nonlinear functional analysis.
Chapter 16 discusses analysis in Banach spaces (with a view towards
applications in the calculus of variations and infinite dimensional dynamical
systems).
Chapter 17 and 18 cover degree theory and fixed point theorems in
finite and infinite dimensional spaces. Several applications to integral equa-
tions, ordinary differential equations and to the stationary Navier–Stokes
equation are given.
Chapter 19 provides some basics about monotone maps.
Sometimes also the historic development of the subject is of interest. This
is however not covered in the present book and we reefer to [27, 41, 42] as
good starting points.

To the teacher

There are a couple of courses to be taught from this book. First of

all there is of course a basic functional analysis course: Chapters 1 to 4
(skipping some optional material as discussed above) and perhaps adding
some material from Chapter 5 or 6. If one wants to cover Lebesgue spaces,
this can be easily done by including Chapters 1, 2, and 3 from [47]. In this
case one could cover Section 1.2 (Section 1.1 contains just motivation), give
an outline of Section 1.3 (by covering Dynkin’s π-λ theorem, the uniqueness
theorem for measures ,and then quoting the existence theorem for Lebesgue
measure), cover Section 1.5. The core material from Chapter 2 are the
first two sections and from Chapter 3 the first three sections. I think that
this gives a well-balanced introduction to functional analysis which contains
several optional topics to choose from depending on personal preferences and
time constraints.
The remaining material from the first part could then be used for a course
on advanced functional analysis. Typically one could also add some further
Preface xi

topics from the second part or some material from unbounded operators in
Hilbert spaces following [46] (where one can start with Chapter 2) or from
unbounded operators in Banach spaces following the book by Kato [24] (e.g.
Sections 3.4, 3.5 and 4.1).
The third part gives a short basis for a course on nonlinear functional
analysis.
Problems relevant for the main text are marked with a "*". A Solutions
Manual will be available electronically for instructors only.

Acknowledgments
I wish to thank my readers, Kerstin Ammann, Phillip Bachler, Batuhan
Bayır, Alexander Beigl, Mikhail Botchkarev, Ho Boon Suan, Peng Du, Chris-
tian Ekstrand, Damir Ferizović, Michael Fischer, Raffaello Giulietti, Melanie
Graf, Josef Greilhuber, Julian Grüber, Matthias Hammerl, Jona Marie Has-
senbach, Nobuya Kakehashi, Iryna, Karpenko, Jerzy Knopik, Nikolas Knotz,
Florian Kogelbauer, Helge Krüger, Reinhold Küstner, Oliver Leingang, Juho
Leppäkangas, Joris Mestdagh, Alice Mikikits-Leitner, Claudiu Mîndrilǎ, Jakob
Möller, Caroline Moosmüller, Matthias Ostermann, Piotr Owczarek, Mar-
tina Pflegpeter, Mateusz Piorkowski, Tobias Preinerstorfer, Maximilian H.
Ruep, Tidhar Sariel, Christian Schmid, Laura Shou, Bertram Tschiderer,
Liam Urban, Vincent Valmorin, David Wallauch, Richard Welke, David
Wimmesberger, Song Xiaojun, Markus Youssef, Rudolf Zeidler, and col-
leagues Pierre-Antoine Absil, Nils C. Framstad, Fritz Gesztesy, Heinz Hanß-
mann, Günther Hörmann, Aleksey Kostenko, Wallace Lam, Daniel Lenz,
Johanna Michor, Viktor Qvarfordt, Alex Strohmaier, David C. Ullrich, Hen-
drik Vogt, Marko Stautz, Maxim Zinchenko who have pointed out several
typos and made useful suggestions for improvements. I am also grateful to
Volker Enß for making his lecture notes on nonlinear Functional Analysis
available to me.

Finally, no book is free of errors. So if you find one, or if you

have comments or suggestions (no matter how small), please let
me know.

Gerald Teschl

Vienna, Austria
January, 2019
Part 1

Functional Analysis
 Chapter 1

A first look at Banach

and Hilbert spaces

Functional analysis is an important tool in the investigation of all kind of

problems in pure mathematics, physics, biology, economics, etc.. In fact, it
is hard to find a branch in science where functional analysis is not used.
The main objects are (infinite dimensional) vector spaces with different
concepts of convergence. The classical theory focuses on linear operators
(i.e., functions) between these spaces but nonlinear operators are of course
equally important. However, since one of the most important tools in investi-
gating nonlinear mappings is linearization (differentiation), linear functional
analysis will be our first topic in any case.

1.1. Introduction: Linear partial differential equations

Rather than listing an overwhelming number of classical examples I want to
focus on one: linear partial differential equations. We will use this example
as a guide throughout our first three chapters and will develop all necessary
tools for a successful treatment of our particular problem.
In his investigation of heat conduction Fourier was led to the (one di-
mensional) heat or diffusion equation
∂ ∂2
u(t, x) = u(t, x). (1.1)
∂t ∂x2
Here u : R × [0, 1] → R is the temperature distribution in a thin rod at time
t ∈ R at the point x ∈ [0, 1]. It is usually assumed, that the temperature
at x = 0 and x = 1 is fixed, say u(t, 0) = a and u(t, 1) = b. By considering
u(t, x) → u(t, x)−a−(b−a)x it is clearly no restriction to assume a = b = 0.

3
4 1. A first look at Banach and Hilbert spaces

Moreover, the initial temperature distribution u(0, x) = u0 (x) is assumed to

be known as well.
Since finding the solution seems at first sight unfeasable, we could try to
find at least some solutions of (1.1). For example, we could make an ansatz
for u(t, x) as a product of two functions, each of which depends on only one
variable, that is,
u(t, x) := w(t)y(x). (1.2)
Plugging this ansatz into the heat equation we arrive at
ẇ(t)y(x) = y 00 (x)w(t), (1.3)
where the dot refers to differentiation with respect to t and the prime to
differentiation with respect to x. Bringing all t, x dependent terms to the
left, right side, respectively, we obtain
ẇ(t) y 00 (x)
= . (1.4)
w(t) y(x)
Accordingly, this ansatz is called separation of variables.
Now if this equation should hold for all t and x, the quotients must be
equal to a constant −λ (we choose −λ instead of λ for convenience later on).
That is, we are led to the equations
− ẇ(t) = λw(t) (1.5)
and
− y 00 (x) = λy(x), y(0) = y(1) = 0, (1.6)
which can easily be solved. The first one gives
w(t) = c1 e−λt (1.7)
and the second one
√ √
y(x) = c2 cos( λx) + c3 sin( λx). (1.8)
However, y(x) must also satisfy the boundary conditions y(0) = y(1) = 0.
The first one y(0) = 0 is satisfied if c2 = 0 and the second one yields (c3 can
be absorbed by w(t)) √
sin( λ) = 0, (1.9)
√
which holds if λ = (πn) , n ∈ N (in the case λ < 0 we get sinh( −λ) = 0,
2

which cannot be satisfied and explains our choice of sign above). In summary,
we obtain the solutions
2
un (t, x) := cn e−(πn) t sin(nπx), n ∈ N. (1.10)
So we have found a large number of solutions, but we still have not dealt
with our initial condition u(0, x) = u0 (x). This can be done using the
superposition principle which holds since our equation is linear: Any finite
linear combination of the above solutions will again be a solution. Moreover,
1.1. Introduction: Linear partial differential equations 5

under suitable conditions on the coefficients we can even consider infinite

linear combinations. In fact, choosing
∞
2
X
u(t, x) := cn e−(πn) t sin(nπx), (1.11)
n=1

where the coefficients cn decay sufficiently fast (e.g. absolutely summable),

we obtain further solutions of our equation. Moreover, these solutions satisfy
∞
X
u(0, x) = cn sin(nπx) (1.12)
n=1

and expanding the initial conditions into a Fourier sine series

∞
X
u0 (x) = û0,n sin(nπx), (1.13)
n=1

we see that the solution of our original problem is given by (1.11) if we choose
cn = û0,n (cf. Problem 1.2).
Of course for this last statement to hold we need to ensure that the series
in (1.11) converges and that we can interchange summation and differentia-
tion. You are asked to do so in Problem 1.1.
In fact, many equations in physics can be solved in a similar way:
• Reaction-Diffusion equation:
∂ ∂2
u(t, x) − 2 u(t, x) + q(x)u(t, x) = 0,
∂t ∂x
u(0, x) = u0 (x),
u(t, 0) = u(t, 1) = 0. (1.14)

Here u(t, x) could be the density of some gas in a pipe and q(x) > 0 describes
that a certain amount per time is removed (e.g., by a chemical reaction).
• Wave equation:
∂2 ∂2
u(t, x) − u(t, x) = 0,
∂t2 ∂x2
∂u
u(0, x) = u0 (x), (0, x) = v0 (x)
∂t
u(t, 0) = u(t, 1) = 0. (1.15)

Here u(t, x) is the displacement of a vibrating string which is fixed at x = 0

and x = 1. Since the equation is of second order in time, both the initial
displacement u0 (x) and the initial velocity v0 (x) of the string need to be
known.
6 1. A first look at Banach and Hilbert spaces

• Schrödinger equation:
∂ ∂2
i u(t, x) = − 2 u(t, x) + q(x)u(t, x),
∂t ∂x
u(0, x) = u0 (x),
u(t, 0) = u(t, 1) = 0. (1.16)
Here |u(t, x)|2 is the probability distribution of a particle trapped in a box
x ∈ [0, 1] and q(x) is a given external potential which describes the forces
acting on the particle.
All these problems (and many others) lead to the investigation of the
following problem
d2
Ly(x) = λy(x), L := − + q(x), (1.17)
dx2
subject to the boundary conditions
y(a) = y(b) = 0. (1.18)
Such a problem is called a Sturm–Liouville boundary value problem.
Our example shows that we should prove the following facts about Sturm–
Liouville problems:
(i) The Sturm–Liouville problem has a countable number of eigenval-
ues En with corresponding eigenfunctions un , that is, un satisfies
the boundary conditions and Lun = En un .
(ii) The eigenfunctions un are complete, that is, any nice function u
can be expanded into a generalized Fourier series
∞
X
u(x) = cn un (x).
n=1
This problem is very similar to the eigenvalue problem of a matrix and we
are looking for a generalization of the well-known fact that every symmetric
matrix has an orthonormal basis of eigenvectors. However, our linear opera-
tor L is now acting on some space of functions which is not finite dimensional
and it is not at all clear what (e.g.) orthogonal should mean in this context.
Moreover, since we need to handle infinite series, we need convergence and
hence we need to define the distance of two functions as well.
Hence our program looks as follows:
• What is the distance of two functions? This automatically leads
us to the problem of convergence and completeness.
• If we additionally require the concept of orthogonality, we are led
to Hilbert spaces which are the proper setting for our eigenvalue
problem.
1.2. The Banach space of continuous functions 7

• Finally, the spectral theorem for compact symmetric operators will

be the solution of our above problem.
Problem 1.1. Suppose ∞ n=1 |cn | < ∞. Show that (1.11) is continuous for
P
(t, x) ∈ [0, ∞) × [0, 1] and solves the heat equation for (t, x) ∈ (0, ∞) × [0, 1].
(Hint: Weierstraß M-test. When can you interchange the order of summation
and differentiation?)
Problem 1.2. Show that for n, m ∈ N we have
Z 1 (
1 , n = m,
2 sin(nπx) sin(mπx)dx =
0 0, n 6= m.
Conclude that the Fourier sine coefficients are given by
Z 1
û0,n = 2 sin(nπx)u0 (x)dx
0
provided the sum in (1.13) converges uniformly. Conclude that in this case
the solution can be expressed as
Z 1
u(t, x) = K(t, x, y)u0 (y)dy, t > 0,
0
where
∞
2
X
K(t, x, y) := 2 e−(πn) t sin(nπx) sin(nπy)
n=1
1 x−y x+y

= ϑ( , iπt) − ϑ( , iπt) .
2 2 2
Here
2 τ +2πinz 2τ
X X
ϑ(z, τ ) := eiπn =1+2 eiπn cos(2πnz), Im(τ ) > 0,
n∈Z n∈N
is the Jacobi theta function.

1.2. The Banach space of continuous functions

Our point of departure will be the set of continuous functions C(I) on a
compact interval I := [a, b] ⊂ R. Since we want to handle both real and
complex models, we will formulate most results for the more general complex
case only. In fact, most of the time there will be no difference but we will
add a remark in the rare case where the real and complex case do indeed
differ.
One way of declaring a distance, well-known from calculus, is the max-
imum norm of a function f ∈ C(I):
kf k∞ := max |f (x)|. (1.19)
x∈I
8 1. A first look at Banach and Hilbert spaces

It is not hard to see that with this definition C(I) becomes a normed vector
space:
A normed vector space X is a vector space X over C (or R) with a
nonnegative function (the norm) k.k : X → [0, ∞) such that
• kf k > 0 for f 6= 0 (positive definiteness),
• kα f k = |α| kf k for all α ∈ C, f ∈ X (positive homogeneity),
and
• kf + gk ≤ kf k + kgk for all f, g ∈ X (triangle inequality).
If positive definiteness is dropped from the requirements, one calls k.k a
seminorm.
From the triangle inequality we also get the inverse triangle inequal-
ity (Problem 1.3)
|kf k − kgk| ≤ kf − gk, (1.20)
which shows that the norm is continuous.
Also note that norms are closely related to convexity. To this end recall
that a subset C ⊆ X is called convex if for every x, y ∈ C we also have
λx + (1 − λ)y ∈ C whenever λ ∈ (0, 1). Moreover, a mapping f : C → R is
called convex if f (λx + (1 − λ)y) ≤ λf (x) + (1 − λ)f (y) whenever λ ∈ (0, 1)
and in our case the triangle inequality plus homogeneity imply that every
norm is convex:
kλx + (1 − λ)yk ≤ λkxk + (1 − λ)kyk, λ ∈ [0, 1]. (1.21)
Moreover, choosing λ = 21 we get back the triangle inequality upon using
homogeneity. In particular, the triangle inequality could be replaced by
convexity in the definition.
Once we have a norm, we have a distance d(f, g) := kf − gk and hence
we know when a sequence of vectors fn converges to a vector f (namely
if d(fn , f ) → 0). We will write fn → f or limn→∞ fn = f , as usual, in
this case. Moreover, a mapping F : X → Y between two normed spaces
is called continuous if for every convergent sequence fn → f from X we
have F (fn ) → F (f ) (with respect to the norm of X and Y , respectively). In
fact, the norm, vector addition, and multiplication by scalars are continuous
(Problem 1.4).
In addition to the concept of convergence, we also have the concept of a
Cauchy sequence and hence the concept of completeness: A normed space
is called complete if every Cauchy sequence has a limit. A complete normed
space is called a Banach space.
Example 1.1. By completeness of the real numbers, R as well as C with
the absolute value as norm are Banach spaces. 
1.2. The Banach space of continuous functions 9

Example 1.2. The space `1 (N) of all complex-valued sequences a = (aj )∞

j=1
for which the norm
X∞
kak1 := |aj | (1.22)
j=1
is finite is a Banach space.
To show this, we need to verify three things: (i) `1 (N) is a vector space,
that is, closed under addition and scalar multiplication, (ii) k.k1 satisfies the
three requirements for a norm, and (iii) `1 (N) is complete.
First of all, observe
k
X k
X k
X
|aj + bj | ≤ |aj | + |bj | ≤ kak1 + kbk1
j=1 j=1 j=1

for every finite k. Letting k → ∞, we conclude that `1 (N) is closed under

addition and that the triangle inequality holds. That `1 (N) is closed under
scalar multiplication together with homogeneity as well as definiteness are
straightforward. It remains to show that `1 (N) is complete. Let an = (anj )∞
j=1
be a Cauchy sequence; that is, for given ε > 0 we can find an Nε such that
kam − an k1 ≤ ε for m, n ≥ Nε . This implies, in particular, |am
j − aj | ≤ ε for
n

every fixed j. Thus aj is a Cauchy sequence for fixed j and, by completeness

n

of C, it has a limit: aj := limn→∞ anj . Now consider kj=1 |amj − aj | ≤ ε and

n
P
take m → ∞:
Xk
|aj − anj | ≤ ε.
j=1
Since this holds for all finite k, we even have ka − an k1 ≤ ε. Hence (a − an ) ∈
`1 (N) and since an ∈ `1 (N), we finally conclude a = an + (a − an ) ∈ `1 (N).
By our estimate ka − an k1 ≤ ε, our candidate a is indeed the limit of an . 
Example 1.3. The previous example can be generalized by considering the
j=1 for which the norm
space `p (N) of all complex-valued sequences a = (aj )∞
 1/p
∞
X
kakp :=  |aj |p  , p ∈ [1, ∞), (1.23)
j=1

is finite. By |aj + bj |p ≤ 2p max(|aj |, |bj |)p = 2p max(|aj |p , |bj |p ) ≤ 2p (|aj |p +

|bj |p ) it is a vector space, but the triangle inequality is only easy to see in the
case p = 1. (It is also not hard to see that it fails for p < 1, which explains
our requirement p ≥ 1. See also Problem 1.15.)
To prove the triangle inequality we need Young’s inequality (Problem 1.8)
1 1 1 1
α1/p β 1/q ≤ α + β, + = 1, α, β ≥ 0, (1.24)
p q p q
10 1. A first look at Banach and Hilbert spaces

p=∞
1
p=4

p=2

p=1

1
p= 2

−1 1

−1

Figure 1.1. Unit balls for k.kp in R2

which implies Hölder’s inequality

kabk1 ≤ kakp kbkq (1.25)
for a ∈ `p (N), b ∈ `q (N). In fact, by homogeneity of the norm it suffices to
prove the case kakp = kbkq = 1. But this case follows by choosing α = |aj |p
and β = |bj |q in (1.24) and summing over all j.
Now using |aj + bj |p ≤ |aj | |aj + bj |p−1 + |bj | |aj + bj |p−1 , we obtain from
Hölder’s inequality (note (p − 1)q = p)
ka + bkpp ≤ kakp k(a + b)p−1 kq + kbkp k(a + b)p−1 kq
= (kakp + kbkp )ka + bkp−1
p .

Hence `p (N) is a normed space. That it is complete can be shown as in the

case p = 1 (Problem 1.9).
The unit ball with respect to these norms in R2 is depicted in Figure 1.1.
One sees that for p < 1 the unit ball is not convex (explaining once more our
restriction p ≥ 1). Moreover, for 1 < p < ∞ it is even strictly convex (that
is, the line segment joining two distinct points is always in the interior). This
is related to the question of equality in the triangle inequality and will be
discussed in Problems 1.12 and 1.13. 
Example 1.4. The space `∞ (N) of all complex-valued bounded sequences
j=1 together with the norm
a = (aj )∞
kak∞ := sup |aj | (1.26)
j∈N
1.2. The Banach space of continuous functions 11

is a Banach space (Problem 1.10). Note that with this definition, Hölder’s
inequality (1.25) remains true for the cases p = 1, q = ∞ and p = ∞, q = 1.
The reason for the notation is explained in Problem 1.14. 
Example 1.5. Every closed subspace of a Banach space is again a Banach
space. For example, the space c0 (N) ⊂ `∞ (N) of all sequences converging to
zero is a closed subspace. In fact, if a ∈ `∞ (N)\c0 (N), then lim supj→∞ |aj | =
ε > 0 and thus ka − bk∞ ≥ ε for every b ∈ c0 (N). 
Now what about completeness of C(I)? A sequence of functions fn
converges to f if and only if
lim kf − fn k∞ = lim max |f (x) − fn (x)| = 0. (1.27)
n→∞ n→∞ x∈I

That is, in the language of real analysis, fn converges uniformly to f . Now

let us look at the case where fn is only a Cauchy sequence. Then fn (x) is
clearly a Cauchy sequence of complex numbers for every fixed x ∈ I. In
particular, by completeness of C, there is a limit f (x) for each x. Thus we
get a limiting function f (x). Moreover, letting m → ∞ in
|fm (x) − fn (x)| ≤ ε ∀m, n > Nε , x ∈ I, (1.28)
we see
|f (x) − fn (x)| ≤ ε ∀n > Nε , x ∈ I; (1.29)
that is, fn (x) converges uniformly to f (x). However, up to this point we
do not know whether f is in our vector space C(I), that is, whether it is
continuous. Fortunately, there is a well-known result from real analysis which
tells us that the uniform limit of continuous functions is again continuous:
Fix x ∈ I and ε > 0. To show that f is continuous we need to find a δ such
that |x − y| < δ implies |f (x) − f (y)| < ε. Pick n so that kfn − f k∞ < ε/3
and δ so that |x − y| < δ implies |fn (x) − fn (y)| < ε/3. Then |x − y| < δ
implies
ε ε ε
|f (x)−f (y)| ≤ |f (x)−fn (x)|+|fn (x)−fn (y)|+|fn (y)−f (y)| < + + = ε
3 3 3
as required. Hence f ∈ C(I) and thus every Cauchy sequence in C(I)
converges. Or, in other words,
Theorem 1.1. Let I ⊂ R be a compact interval, then the continuous func-
tions C(I) with the maximum norm is a Banach space.

For finite dimensional vector spaces the concept of a basis plays a crucial
role. In the case of infinite dimensional vector spaces one could define a
basis as a maximal set of linearly independent vectors (known as a Hamel
basis; Problem 1.7). Such a basis has the advantage that it only requires
finite linear combinations. However, the price one has to pay is that such
a basis will be way too large (typically uncountable, cf. Problems 1.6 and
12 1. A first look at Banach and Hilbert spaces

4.2). Since we have the notion of convergence, we can handle countable

linear combinations and try to look for countable bases. We start with a few
definitions.
The set of all finite linear combinations of a set of vectors {un }n∈N ⊂ X
is called the span of {un }n∈N and denoted by
m
X
span{un }n∈N := { αj unj |nj ∈ N , αj ∈ C, m ∈ N}. (1.30)
j=1

A set of vectors {un }n∈N ⊂ X is called linearly independent if every finite

subset is. If {un }N
n=1 ⊂ X, N ∈ N ∪ {∞}, is countable, we can throw away
all elements which can be expressed as linear combinations of the previous
ones to obtain a subset of linearly independent vectors which have the same
span.
We will call a countable sequence of vectors (un )N
n=1 ⊂ X a Schauder
basis if every element f ∈ X can be uniquely written as a countable linear
combination of the basis elements:
XN
f= αn un , αn = αn (f ) ∈ C, (1.31)
n=1
where the sum has to be understood as a limit if N = ∞ (the sum is not
required to converge unconditionally and hence the order of the basis el-
ements is important). Since we have assumed the coefficients αn (f ) to be
uniquely determined, the vectors are necessarily linearly independent. More-
over, one can show that the coordinate functionals f 7→ αn (f ) are continuous
(cf. Problem 4.7). A Schauder basis and its corresponding coordinate func-
tionals u∗n : X → C, f 7→ αn (f ) form a so-called biorthogonal system:
u∗m (un ) = δm,n , where (
1, n = m,
δn,m := (1.32)
0, n 6= m,
is the Kronecker delta.
Example 1.6. The sequence of vectors δ n = (δm n =δ
n,m )m∈N is a Schauder
basis for the Banach space ` (N), 1 ≤ p < ∞.
p
Pm
Let a = (aj )∞
j=1 ∈ ` (N) be given and set a :=
p m
n=1 an δ . Then
n

 1/p
X∞
ka − am kp =  |aj |p  → 0
j=m+1

j = aj for 1 ≤ j ≤ m and aj = 0 for j > m. Hence

since am m

∞
X
a= an δ n
n=1
1.2. The Banach space of continuous functions 13

and (δ n )∞
n=1 is a Schauder basis (uniqueness of the coefficients is left as an
exercise).
Note that (δ n )∞
n=1 is also Schauder basis for c0 (N) but not for ` (N) (try
∞

to approximate a constant sequence). 

A set whose span is dense is called total, and if we have a countable total
set, we also have a countable dense set (consider only linear combinations
with rational coefficients — show this). A normed vector space containing a
countable dense set is called separable.
Warning: Some authors use the term total in a slightly different way —
see the warning on page 119.
Example 1.7. Every Schauder basis is total and thus every Banach space
with a Schauder basis is separable (the converse puzzled mathematicians for
quite some time and was eventually shown to be false by Per Enflo). In
particular, the Banach space `p (N) is separable for 1 ≤ p < ∞.
However, `∞ (N) is not separable (Problem 1.11)! 
While we will not give a Schauder basis for C(I) (Problem 1.16), we will
at least show that C(I) is separable. We will do this by showing that every
continuous function can be approximated by polynomials, a result which is
of independent interest. But first we need a lemma.

Lemma 1.2 (Smoothing). Let un be a sequence of nonnegative continuous

functions on [−1, 1] such that
Z Z
un (x)dx = 1 and un (x)dx → 0, δ > 0. (1.33)
|x|≤1 δ≤|x|≤1

(In other words, un has mass one and concentrates near x = 0 as n → ∞.)
Then for every f ∈ C[− 21 , 12 ] which vanishes at the endpoints, f (− 12 ) =
f ( 21 )
= 0, we have that
Z 1/2
fn (x) := un (x − y)f (y)dy (1.34)
−1/2

converges uniformly to f (x).

Proof. Since f is uniformly continuous, for given ε we can find a δ < 1/2
(independent of x) such that |f (x) −R f (y)| ≤ ε whenever |x − y| ≤ δ. More-
over, we can choose n such that δ≤|y|≤1 un (y)dy ≤ ε. Now abbreviate
M := maxx∈[−1/2,1/2] {1, |f (x)|} and note
Z 1/2 Z 1/2
|f (x) − un (x − y)f (x)dy| = |f (x)| |1 − un (x − y)dy| ≤ M ε.
−1/2 −1/2
14 1. A first look at Banach and Hilbert spaces

In fact, either the distance of x to one of the boundary points ± 12 is smaller

than δ and hence |f (x)| ≤ ε or otherwise [−δ, δ] ⊂ [x − 1/2, x + 1/2] and the
difference between one and the integral is smaller than ε.
Using this, we have
Z 1/2
|fn (x) − f (x)| ≤ un (x − y)|f (y) − f (x)|dy + M ε
−1/2
Z
= un (x − y)|f (y) − f (x)|dy
|y|≤1/2,|x−y|≤δ
Z
+ un (x − y)|f (y) − f (x)|dy + M ε
|y|≤1/2,|x−y|≥δ
≤ε + 2M ε + M ε = (1 + 3M )ε,
which proves the claim. 

Note that fn will be as smooth as un , hence the title smoothing lemma.

Moreover, fn will be a polynomial if un is. The same idea is used to approx-
imate noncontinuous functions by smooth ones (of course the convergence
will no longer be uniform in this case).
Now we are ready to show:
Theorem 1.3 (Weierstraß). Let I ⊂ R be a compact interval. Then the set
of polynomials is dense in C(I).

Proof. Let f ∈ C(I) be given. By considering f (x) − f (a) − f (b)−f

b−a
(a)
(x − a)
it is no loss to assume that f vanishes at the boundary points. Moreover,
without restriction, we only consider I = [− 21 , 12 ] (why?).
Now the claim follows from Lemma 1.2 using the Landau kernel
1
un (x) := (1 − x2 )n ,
In
where (using integration by parts)
Z 1 Z 1
2 n n
In := (1 − x ) dx = (1 − x)n−1 (1 + x)n+1 dx
−1 n + 1 −1
n! (n!)2 22n+1 n!
= ··· = 22n+1 = = 1 1 1 .
(n + 1) · · · (2n + 1) (2n + 1)! 2 ( 2 + 1) · · · ( 2 + n)
Indeed, the first part of (1.33) holds by construction, and the second part
follows from the elementary estimate
1
1 < In < 2,
2 +n
which shows δ≤|x|≤1 un (x)dx ≤ 2un (δ) < (2n + 1)(1 − δ 2 )n → 0.
R

1.2. The Banach space of continuous functions 15

Corollary 1.4. The monomials are total and hence C(I) is separable.

Note that while the proof of Theorem 1.3 provides an explicit way of
constructing a sequence of polynomials fn (x) which will converge uniformly
to f (x), this method still has a few drawbacks from a practical point of
view: Suppose we have approximated f by a polynomial of degree n but our
approximation turns out to be insufficient for the intended purpose. First
of all, since our polynomial will not be optimal in general, we could try to
find another polynomial of the same degree giving a better approximation.
However, as this is by no means straightforward, it seems more feasible to
simply increase the degree. However, if we do this, all coefficients will change
and we need to start from scratch. This is in contradistinction to a Schauder
basis where we could just add one new element from the basis (and where it
suffices to compute one new coefficient).
In particular, note that this shows that the monomials are no Schauder
basis for C(I) since adding monomials incrementally to the expansion gives
a uniformly convergent power series whose limit must be analytic. This
observation emphasizes that a Schauder basis is more than a set of linearly
independent vectors whose span is dense.
We will see in the next section that the concept of orthogonality resolves
these problems.

Problem* 1.3. Show that |kf k − kgk| ≤ kf − gk.

Problem* 1.4. Let X be a Banach space. Show that the norm, vector
addition, and multiplication by scalars are continuous. That is, if fn → f ,
gn → g, and αn → α, then kfn k → kf k, fn + gn → f + g, and αn gn → αg.

Problem 1.5. Let X be a Banach space. Show that ∞

P
j=1 kfj k < ∞ implies
that
X∞ X n
fj = lim fj
n→∞
j=1 j=1

exists. The series is called absolutely convergent in this case. Conversely,

show that a normed space is complete if every absolutely convergent series
converges.

Problem 1.6. While `1 (N) is separable, it still has room for an uncountable
set of linearly independent vectors. Show this by considering vectors of the
form
aα = (1, α, α2 , . . . ), α ∈ (0, 1).
(Hint: Recall the Vandermonde determinant. See Problem 4.2 for a gener-
alization.)
16 1. A first look at Banach and Hilbert spaces

Problem 1.7. A Hamel basis is a maximal set of linearly independent

vectors. Show that every vector space X has a Hamel basis {uα }α∈A . Show
P basis, every x ∈ X can be written as a finite linear
that given a Hamel
combination x = nj=1 cj uαj , where the vectors uαj and the constants cj are
uniquely determined. (Hint: Use Zorn’s lemma, see Appendix A, to show
existence.)
Problem* 1.8. Prove Young’s inequality (1.24). Show that equality occurs
precisely if α = β. (Hint: Take logarithms on both sides.)
Problem* 1.9. Show that `p (N), 1 ≤ p < ∞, is complete.
Problem* 1.10. Show that `∞ (N) is a Banach space.
Problem* 1.11. Show that `∞ (N) is not separable. (Hint: Consider se-
quences which take only the value one and zero. How many are there? What
is the distance between two such sequences?)
Problem* 1.12. Show that there is equality in the Hölder inequality (1.25)
for 1 < p < ∞ if and only if either a = 0 or |bj |p = α|aj |q for all j ∈ N.
Show that we have equality in the triangle inequality for `1 (N) if and only if
aj b∗j ≥ 0 for all j ∈ N (here the ‘∗’ denotes complex conjugation). Show that
we have equality in the triangle inequality for `p (N) with 1 < p < ∞ if and
only if a = 0 or b = αa with α ≥ 0.
Problem* 1.13. Let X be a normed space. Show that the following condi-
tions are equivalent.
(i) If kx + yk = kxk + kyk then y = αx for some α ≥ 0 or x = 0.
(ii) If kxk = kyk = 1 and x 6= y then kλx + (1 − λ)yk < 1 for all
0 < λ < 1.
(iii) If kxk = kyk = 1 and x 6= y then 21 kx + yk < 1.
(iv) The function x 7→ kxk2 is strictly convex.
A norm satisfying one of them is called strictly convex.
Show that `p (N) is strictly convex for 1 < p < ∞ but not for p = 1, ∞.
Problem 1.14. Show that p0 ≤ p implies `p0 (N) ⊂ `p (N) and kakp ≤ kakp0 .
Moreover, show
lim kakp = kak∞ .
p→∞

Problem 1.15. Formally extend the definition of `p (N) to p ∈ (0, 1). Show
that k.kp does not satisfy the triangle inequality. However, show that it is a
quasinormed space, that is, it satisfies all requirements for a normed space
except for the triangle inequality which is replaced by
ka + bk ≤ K(kak + kbk)
1.3. The geometry of Hilbert spaces 17

with some constant K ≥ 1. Show, in fact,

ka + bkp ≤ 21/p−1 (kakp + kbkp ), p ∈ (0, 1).
Moreover, show that k.kpp satisfies the triangle inequality in this case, but
of course it is no longer homogeneous (but at least you can get an honest
metric d(a, b) = ka − bkpp which gives rise to the same topology). (Hint:
Show α + β ≤ (αp + β p )1/p ≤ 21/p−1 (α + β) for 0 < p < 1 and α, β ≥ 0.)
Problem 1.16. Show that the following set of functions is a Schauder ba-
sis for C[0, 1]: We start with u1 (t) = t, u2 (t) = 1 − t and then split
[0, 1] into 2n intervals of equal length and let u2n +k+1 (t), for 1 ≤ k ≤ 2n ,
be a piecewise linear peak of height 1 supported in the k’th subinterval:
u2n +k+1 (t) = max(0, 1 − |2n+1 t − 2k + 1|) for n ∈ N0 and 1 ≤ k ≤ 2n .

1.3. The geometry of Hilbert spaces

So far it looks like C(I) has all the properties we want. However, there is
still one thing missing: How should we define orthogonality in C(I)? In
Euclidean space, two vectors are called orthogonal if their scalar product
vanishes, so we would need a scalar product:
Suppose H is a vector space. A map h., ..i : H × H → C is called a
sesquilinear form if it is conjugate linear in the first argument and linear
in the second; that is,
hα1 f1 + α2 f2 , gi = α1∗ hf1 , gi + α2∗ hf2 , gi,
α1 , α2 ∈ C, (1.35)
hf, α1 g1 + α2 g2 i = α1 hf, g1 i + α2 hf, g2 i,
where ‘∗’ denotes complex conjugation. A symmetric
hf, gi = hg, f i∗ (symmetry)
sesquilinear form is also called a Hermitian form and a positive definite
hf, f i > 0 for f 6= 0 (positive definite),
Hermitian form is called an inner product or scalar product. Note that
positivity already implies symmetry in the complex case (Problem 1.20).
Associated with every scalar product is a norm
(1.36)
p
kf k := hf, f i.
Only the triangle inequality is nontrivial. It will follow from the Cauchy–
Schwarz inequality below. Until then, just regard (1.36) as a convenient
short hand notation.
Warning: There is no common agreement whether a sesquilinear form
(scalar product) should be linear in the first or in the second argument and
different authors use different conventions.
18 1. A first look at Banach and Hilbert spaces

The pair (H, h., ..i) is called an inner product space. If H is complete
(with respect to the norm (1.36)), it is called a Hilbert space.
Example 1.8. Clearly, Cn with the usual scalar product
n
X
ha, bi := a∗j bj (1.37)
j=1

is a (finite dimensional) Hilbert space. 

Example 1.9. A somewhat more interesting example is the Hilbert space
`2 (N), that is, the set of all complex-valued sequences
n X∞ o
(aj )∞ |a |2
< ∞ (1.38)

j=1  j
j=1

with scalar product

∞
X
ha, bi := a∗j bj . (1.39)
j=1

That this sum is (absolutely) convergent (and thus well-defined) for a, b ∈

`2 (N) follows from Hölder’s inequality (1.25) in the case p = q = 2.
Observe that the norm kak = ha, ai is identical to the norm kak2
p

defined in the previous section. In particular, `2 (N) is complete and thus

indeed a Hilbert space. 
A vector f ∈ H is called normalized or a unit vector if kf k = 1.
Two vectors f, g ∈ H are called orthogonal or perpendicular (f ⊥ g) if
hf, gi = 0 and parallel if one is a multiple of the other.
If f and g are orthogonal, we have the Pythagorean theorem:

kf + gk2 = kf k2 + kgk2 , f ⊥ g, (1.40)

which is one line of computation (do it!).

Suppose u is a unit vector. Then the projection of f in the direction of
u is given by
fk := hu, f iu, (1.41)

and f⊥ , defined via

f⊥ := f − hu, f iu, (1.42)

is perpendicular to u since hu, f⊥ i = hu, f − hu, f iui = hu, f i − hu, f ihu, ui =

0.
1.3. The geometry of Hilbert spaces 19

BMB
f B f⊥
B
1
 B
 
 fk
 u

1


Taking any other vector parallel to u, we obtain from (1.40)

kf − αuk2 = kf⊥ + (fk − αu)k2 = kf⊥ k2 + |hu, f i − α|2 (1.43)
and hence fk is the unique vector parallel to u which is closest to f .
As a first consequence we obtain the Cauchy–Bunyakovsky–Schwarz
inequality:

Theorem 1.5 (Cauchy–Bunyakovsky–Schwarz). Let H0 be an inner product

space. Then for every f, g ∈ H0 we have
|hf, gi| ≤ kf k kgk (1.44)
with equality if and only if f and g are parallel.

Proof. It suffices to prove the case kgk = 1. But then the claim follows
from kf k2 = |hg, f i|2 + kf⊥ k2 . 

We will follow common practice and refer to (1.44) simply as Cauchy–

Schwarz inequality. Note that the Cauchy–Schwarz inequality implies that
the scalar product is continuous in both variables; that is, if fn → f and
gn → g, we have hfn , gn i → hf, gi.
As another consequence we infer that the map k.k is indeed a norm. In
fact,
kf + gk2 = kf k2 + hf, gi + hg, f i + kgk2 ≤ (kf k + kgk)2 . (1.45)
But let us return to C(I). Can we find a scalar product which has the
maximum norm as associated norm? Unfortunately the answer is no! The
reason is that the maximum norm does not satisfy the parallelogram law
(Problem 1.19).

Theorem 1.6 (Jordan–von Neumann). A norm is associated with a scalar

product if and only if the parallelogram law
kf + gk2 + kf − gk2 = 2kf k2 + 2kgk2 (1.46)
holds.
20 1. A first look at Banach and Hilbert spaces

In this case the scalar product can be recovered from its norm by virtue
of the polarization identity
1
kf + gk2 − kf − gk2 + ikf − igk2 − ikf + igk2 . (1.47)


hf, gi =
4
Proof. If an inner product space is given, verification of the parallelogram
law and the polarization identity is straightforward (Problem 1.20).
To show the converse, we define
1
kf + gk2 − kf − gk2 + ikf − igk2 − ikf + igk2 .


s(f, g) :=
4
Then s(f, f ) = kf k2 and s(f, g) = s(g, f )∗ are straightforward to check.
Moreover, another straightforward computation using the parallelogram law
shows
g+h
s(f, g) + s(f, h) = 2s(f, ).
2
Now choosing h = 0 (and using s(f, 0) = 0) shows s(f, g) = 2s(f, g2 ) and thus
s(f, g)+s(f, h) = s(f, g +h). Furthermore, by induction we infer 2mn s(f, g) =
s(f, 2mn g); that is, α s(f, g) = s(f, αg) for a dense set of positive rational
numbers α. By continuity (which follows from continuity of the norm) this
holds for all α ≥ 0 and s(f, −g) = −s(f, g), respectively, s(f, ig) = i s(f, g),
finishes the proof. 

In the case of a real Hilbert space, the polarization identity of course

simplifies to hf, gi = 14 (kf + gk2 − kf − gk2 ).
Note that the parallelogram law and the polarization identity even hold
for sesquilinear forms (Problem 1.20).
But how do we define a scalar product on C(I)? One possibility is
Z b
hf, gi := f ∗ (x)g(x)dx. (1.48)
a

The corresponding inner product space is denoted by L2cont (I). Note that we
have
(1.49)
p
kf k ≤ |b − a|kf k∞
and hence the maximum norm is stronger than the L2cont norm.
Suppose we have two norms k.k1 and k.k2 on a vector space X. Then
k.k2 is said to be stronger than k.k1 if there is a constant m > 0 such that
kf k1 ≤ mkf k2 . (1.50)
It is straightforward to check the following.
Lemma 1.7. If k.k2 is stronger than k.k1 , then every k.k2 Cauchy sequence
is also a k.k1 Cauchy sequence.
1.3. The geometry of Hilbert spaces 21

Hence if a function F : X → Y is continuous in (X, k.k1 ), it is also

continuous in (X, k.k2 ), and if a set is dense in (X, k.k2 ), it is also dense in
(X, k.k1 ).
In particular, L2cont is separable since the polynomials are dense. But is
it also complete? Unfortunately the answer is no:
Example 1.10. Take I = [0, 2] and define

0,
 0 ≤ x ≤ 1 − n1 ,
fn (x) := 1 + n(x − 1), 1 − n1 ≤ x ≤ 1,

1, 1 ≤ x ≤ 2.


Then fn (x) is a Cauchy sequence in L2cont , but there is no limit in L2cont !

Clearly, the limit should be the step function which is 0 for 0 ≤ x < 1 and
1 for 1 ≤ x ≤ 2, but this step function is discontinuous (Problem 1.23)! 
Example 1.11. The previous example indicates that we should consider
(1.48) on a larger class of functions, for example on the class of Riemann
integrable functions
R(I) := {f : I → C|f is Riemann integrable}
such that the integral makes sense. While this seems natural it implies
another problem: Any function which vanishes outside a set which is neg-
ligible for Rthe integral (e.g. finitely many points) has norm zero! That is,
kf k2 := ( I |f (x)|2 dx)1/2 is only a seminorm on R(I) (Problem 1.22). To
get a norm we consider N (I) := {f ∈ R(I)| kf k2 = 0}. By homogeneity and
the triangle inequality N (I) is a subspace and we can consider equivalence
classes of functions which differ by a negligible function from N (I):
L2Ri := R(I)/N (I).
Since kf k2 = kgk2 for f − g ∈ N (I) we have a norm on L2Ri . Moreover, since
this norm inherits the parallelogram law we even have an inner product space.
However, this space will not be complete unless we replace the Riemann by
the Lebesgue integral. Hence we will not pursue this further until we have
the Lebesgue integral at our disposal. 
This shows that in infinite dimensional vector spaces, different norms
will give rise to different convergent sequences. In fact, the key to solving
problems in infinite dimensional spaces is often finding the right norm! This
is something which cannot happen in the finite dimensional case.
Theorem 1.8. If X is a finite dimensional vector space, then all norms are
equivalent. That is, for any two given norms k.k1 and k.k2 , there are positive
constants m1 and m2 such that
1
kf k1 ≤ kf k2 ≤ m1 kf k1 . (1.51)
m2
22 1. A first look at Banach and Hilbert spaces

Proof. ChooseP a basis {uj }1≤j≤n such that every f ∈ X can be writ-
ten as f = j αj uj . Since equivalence of norms is an equivalence rela-
tion (check this!), we can assume that k.k2 is the usual Euclidean norm:
kf k2 := k j αj uj k2 = ( j |αj |2 )1/2 . Then by the triangle and Cauchy–
P P
Schwarz inequalities,
X sX
kf k1 ≤ |αj |kuj k1 ≤ kuj k21 kf k2
j j
qP
and we can choose m2 = j kuj k21 .
In particular, if fn is convergent with respect to k.k2 , it is also convergent
with respect to k.k1 . Thus k.k1 is continuous with respect to k.k2 and attains
its minimum m > 0 on the unit sphere S := {u|kuk2 = 1} (which is compact
by the Heine–Borel theorem, Theorem B.22). Now choose m1 = 1/m. 

Finally, I remark that a real Hilbert space can always be embedded into
a complex Hilbert space. In fact, if H is a real Hilbert space, then H × H is
a complex Hilbert space if we define
(f1 , f2 )+(g1 , g2 ) = (f1 +g1 , f2 +g2 ), (α+iβ)(f1 , f2 ) = (αf1 −βf2 , αf2 +βf1 )
(1.52)
and
h(f1 , f2 ), (g1 , g2 )i = hf1 , g1 i + hf2 , g2 i + i(hf1 , g2 i − hf2 , g1 i). (1.53)
Here you should think of (f1 , f2 ) as f1 + if2 . Note that we have a conjugate
linear map C : H × H → H × H, (f1 , f2 ) 7→ (f1 , −f2 ) which satisfies C 2 = I
and hCf, Cgi = hg, f i. In particular, we can get our original Hilbert space
back if we consider Re(f ) = 21 (f + Cf ) = (f1 , 0).
Problem 1.17. Show that the norm in a Hilbert space satisfies kf + gk =
kf k + kgk if and only if f = αg, α ≥ 0, or g = 0. Hence Hilbert spaces are
strictly convex (cf. Problem 1.13).
Problem 1.18 (Generalized parallelogram law). Show that, in a Hilbert
space, X X X
kxj − xk k2 + k xj k2 = n kxj k2
1≤j<k≤n 1≤j≤n 1≤j≤n
for every n ∈ N. The case n = 2 is (1.46).
Problem 1.19. Show that the maximum norm on C[0, 1] does not satisfy
the parallelogram law.
Problem* 1.20. Suppose Q is a complex vector space. Let s(f, g) be a
sesquilinear form on Q and q(f ) := s(f, f ) the associated quadratic form.
Prove the parallelogram law
q(f + g) + q(f − g) = 2q(f ) + 2q(g) (1.54)
1.4. Completeness 23

and the polarization identity

1
s(f, g) = (q(f + g) − q(f − g) + i q(f − ig) − i q(f + ig)) . (1.55)
4
Show that s(f, g) is symmetric if and only if q(f ) is real-valued.
Note, that if Q is a real vector space, then the parallelogram law is un-
changed but the polarization identity in the form s(f, g) = 41 (q(f + g) − q(f −
g)) will only hold if s(f, g) is symmetric.
Problem 1.21. A sesquilinear form on a complex inner product space is
called bounded if
ksk := sup |s(f, g)|
kf k=kgk=1
is finite. Similarly, the associated quadratic form q is bounded if
kqk := sup |q(f )|
kf k=1

is finite. Show
kqk ≤ ksk ≤ 2kqk
with kqk = ksk if s is symmetric. (Hint: Use the polarization identity from
the previous problem. For the symmetric case look at the real part.)
Problem* 1.22. Suppose Q is a vector space. Let s(f, g) be a sesquilinear
form on Q and q(f ) := s(f, f ) the associated quadratic form. Show that the
Cauchy–Schwarz inequality
|s(f, g)| ≤ q(f )1/2 q(g)1/2
holds if q(f ) ≥ 0. In this case q(.)1/2 satisfies the triangle inequality and
hence is a seminorm.
(Hint: Consider 0 ≤ q(f + αg) = q(f ) + 2Re(α s(f, g)) + |α|2 q(g) and
choose α = t s(f, g)∗ /|s(f, g)| with t ∈ R.)
Problem* 1.23. Prove the claims made about fn in Example 1.10.

1.4. Completeness
Since L2cont is not complete, how can we obtain a Hilbert space from it? Well,
the answer is simple: take the completion.
If X is an (incomplete) normed space, consider the set of all Cauchy
sequences X . Call two Cauchy sequences equivalent if their difference con-
verges to zero and denote by X̄ the set of all equivalence classes. It is easy
to see that X̄ (and X ) inherit the vector space structure from X. Moreover,
Lemma 1.9. If xn is a Cauchy sequence in X, then kxn k is also a Cauchy
sequence and thus converges.
24 1. A first look at Banach and Hilbert spaces

Consequently, the norm of an equivalence class [(xn )∞

n=1 ] can be defined
by k[(xn )∞
n=1 ]k := lim n→∞ kx n k and is independent of the representative
(show this!). Thus X̄ is a normed space.

Theorem 1.10. X̄ is a Banach space containing X as a dense subspace if

we identify x ∈ X with the equivalence class of all sequences converging to
x.

Proof. (Outline) It remains to show that X̄ is complete. Let ξn = [(xn,j )∞ j=1 ]

be a Cauchy sequence in X̄. Without loss of generality (by dropping terms)
we can choose the representatives xn,j such that |xn,j −xn,k | ≤ n1 for j, k ≥ n.
Then it is not hard to see that ξ = [(xj,j )∞
j=1 ] is its limit. 

Notice that the completion X̄ is unique. More precisely, every other

complete space which contains X as a dense subset is isomorphic to X̄. This
can for example be seen by showing that the identity map on X has a unique
extension to X̄ (compare Theorem 1.16 below).
In particular, it is no restriction to assume that a normed vector space
or an inner product space is complete (note that by continuity of the norm
the parallelogram law holds for X̄ if it holds for X).
Example 1.12. The completion of the space L2cont (I) is denoted by L2 (I).
While this defines L2 (I) uniquely (up to isomorphisms) it is often inconve-
nient to work with equivalence classes of Cauchy sequences. A much more
convenient characterization can be given with the help of the Lebesgue inte-
gral (see Chapter 3 from [47] if you are familiar with basic Lebesgue integra-
tion; Theorem 3.18 from [47] will establish equivalence of both approaches).
Similarly, we define Lp (I), 1 ≤ p < ∞, as the completion of C(I) with
respect to the norm
Z b 1/p
kf kp := |f (x)|p dx .
a

The only requirement for a norm which is not immediate is the triangle
inequality (except for p = 1, 2) but this can be shown as for `p (cf. Prob-
lem 1.26). 

Problem 1.24. Provide a detailed proof of Theorem 1.10.

Problem 1.25. For every f ∈ L1 (I) we can define its integral

Z d
f (x)dx
c
1.5. Compactness 25

as the (unique) extension of the corresponding linear functional from C(I) to

L1 (I) (by Theorem 1.16 below). Show that this integral is linear and satisfies
Z e Z d Z e Z d  Z d
 
f (x)dx = f (x)dx + f (x)dx, 
 f (x)dx ≤ |f (x)|dx.
c c d c c

Problem* 1.26. Show the Hölder inequality

1 1
kf gk1 ≤ kf kp kgkq , + = 1, 1 ≤ p, q ≤ ∞,
p q
for f ∈ Lp (I), g ∈ Lq (I) and conclude that k.kp is a norm on C(I). Also
conclude that Lp (I) ⊆ L1 (I).

1.5. Compactness
In finite dimensions relatively compact sets are easily identified as they are
precisely the bounded sets by the Heine–Borel theorem (Theorem B.22).
In the infinite dimensional case the situation is more complicated. Before
we look into this, please recall that for a subset U of a Banach space the
following are equivalent (see Corollary B.20 and Lemma B.26):
• U is relatively compact (i.e. its closure is compact)
• every sequence from U has a convergent subsequence
• U is totally bounded (i.e. it has a finite ε-cover for every ε > 0)
Example 1.13. Consider the bounded sequence (δ n )∞
n=1 in ` (N). Since
p
n m
kδ − δ kp = 2 1/p for n 6= m, there is no way to extract a convergent
subsequence. 
In particular, the Heine–Borel theorem fails for `p (N). In fact, it turns
out that it fails in any infinite dimensional space as we will see in Theo-
rem 4.30 below. Hence one needs criteria when a given subset is relatively
compact. Our strategy will be based on total boundedness and can be out-
lined as follows: Project the original set to some finite dimensional space
such that the information loss can be made arbitrarily small (by increasing
the dimension of the finite dimensional space) and apply Heine–Borel to the
finite dimensional space. This idea is formalized in the following lemma.
Lemma 1.11. Let X be a metric space and K some subset. Assume that
for every ε > 0 there is a metric space Yn , a surjective map Pn : X → Yn ,
and some δ > 0 such that Pn (K) is totally bounded and d(x, y) < ε whenever
x, y ∈ K with d(Pn (x), Pn (y)) < δ. Then K is totally bounded.
In particular, if X is a Banach space the claim holds if Pn can be chosen
a linear map onto a finite dimensional subspace Yn such that kPn k ≤ C,
Pn K is bounded, and k(1 − Pn )xk ≤ ε for x ∈ K.
26 1. A first look at Banach and Hilbert spaces

Proof. Fix ε > 0. Then by total boundedness of Pn (K) we can find a δ-

cover {Bδ (yj )}mj=1 for Pn (K). Now if we choose xj ∈ Pn ({yj }) ∩ K, then
−1

{Bε (xj )}nj=1 is an ε-cover for K since Pn−1 (Bδ (yj )) ∩ K ⊆ Bε (xj ).
For the last claim take Pn corresponding to ε/3 and note that kx − yk ≤
k(1 − Pn )xk + kPn (x − y)k + k(1 − Pn )yk < ε for δ := ε/3. 

The first application will be to `p (N).

Theorem 1.12 (Fréchet). Consider `p (N), 1 ≤ p < ∞, and let Pn a =
(a1 , . . . , an , 0, . . . ) be the projection onto the first n components. A subset
K ⊆ `p (N) is relatively compact if and only if
(i) it is pointwise bounded, supa∈K |aj | ≤ Mj for all j ∈ N, and
(ii) for every ε > 0 there is some n such that k(1 − Pn )akp ≤ ε for all
a ∈ K.
In the case p = ∞ conditions (i) and (ii) still imply that K is relatively
compact, but the converse only holds for K ⊆ c0 (N).

Proof. Clearly (i) and (ii) is what is needed for Lemma 1.11.
Conversely, if K is relatively compact it is bounded. Moreover, given
δ we can choose a finite δ-cover {Bδ (aj )}m
j=1 for K and some n such that
k(1 − Pn )a kp ≤ δ for all 1 ≤ j ≤ m. Now given a ∈ K we have a ∈ Bδ (aj )
j

for some j and hence k(1−Pn )akp ≤ k(1−Pn )(a−aj )kp +k(1−Pn )aj kp ≤ 2δ
as required. 
Example 1.14. Fix a ∈ `p (N) if 1 ≤ p < ∞ or a ∈ c0 (N) else. Then
K := {b| |bj | ≤ |aj |} ⊂ `p (N) is compact. 
The second application will be to C(I). A family of functions F ⊂ C(I)
is called (pointwise) equicontinuous if for every ε > 0 and every x ∈ I
there is a δ > 0 such that
|f (y) − f (x)| ≤ ε whenever |y − x| < δ, ∀f ∈ F. (1.56)
That is, in this case δ is required to be independent of the function f ∈ F .
Theorem 1.13 (Arzelà–Ascoli). Let F ⊂ C(I) be a family of continuous
functions. Then every sequence from F has a uniformly convergent sub-
sequence if and only if F is equicontinuous and the set {f (x0 )|f ∈ F } is
bounded for one x0 ∈ I. In this case F is even bounded.

Proof. Suppose F is equicontinuous and pointwise bounded. Fix ε > 0.

By compactness of I there are finitely many points x1 , . . . , xn ∈ I such
that the balls Bδj (xj ) cover I, where δj is the δ corresponding to xj as
in the definition of equicontinuity. Now first of all note that, since I is
1.6. Bounded operators 27

connected and since x0 ∈ Bδj (xj ) for some j, we see that F is bounded:
|f (x)| ≤ supf ∈F |f (x0 )| + nε.
Next consider P : C[0, 1] → Cn , P (f ) = (f (x1 ), . . . , f (xn )). Then P (F )
is bounded and kf − gk∞ ≤ 3ε whenever kP (f ) − P (g)k∞ < ε. Indeed,
just note that for every x there is some j such that x ∈ Bδj (xj ) and thus
|f (x) − g(x)| ≤ |f (x) − f (xj )| + |f (xj ) − g(xj )| + |g(xj ) − g(x)| ≤ 3ε. Hence
F is relatively compact by Lemma 1.11.
Conversely, suppose F is relatively compact. Then F is totally bounded
and hence bounded. To see equicontinuity fix x ∈ I, ε > 0 and choose a
corresponding ε-cover {Bε (fj )}nj=1 for F . Pick δ > 0 such that y ∈ Bδ (x)
implies |fj (y) − fj (x)| < ε for all 1 ≤ j ≤ n. Then f ∈ Bε (fj ) for some j and
hence |f (y) − f (x)| ≤ |f (y) − fj (y)| + |fj (y) − fj (x)| + |fj (x) − f (x)| ≤ 3ε,
proving equicontinuity. 
Example 1.15. Consider the solution yn (x) of the initial value problem
y 0 = sin(ny), y(0) = 1.
(Assuming this solution exists — it can in principle be found using separation
of variables.) Then |yn0 (x)| ≤ 1 and hence the mean value theorem shows
that the family {yn } ⊆ C([0, 1]) is equicontinuous. Hence there is a uniformly
convergent subsequence. 
Problem 1.27. Show that a subset F ⊂ c0 (N) is relatively compact if and
only if there is a nonnegative sequence a ∈ c0 (N) such that |bn | ≤ an for all
n ∈ N and all b ∈ F .
Problem 1.28. Find a family in C[0, 1] that is equicontinuous but not
bounded.
Problem 1.29. Which of the following families are relatively compact in
C[0, 1]?
(i) F = {f ∈ C 1 [0, 1]| kf k∞ ≤ 1}
(ii) F = {f ∈ C 1 [0, 1]| kf 0 k∞ ≤ 1}
(iii) F = {f ∈ C 1 [0, 1]| kf k∞ ≤ 1, kf 0 k2 ≤ 1}

1.6. Bounded operators

Given two normed spaces X and Y A linear map
A : D(A) ⊆ X → Y (1.57)
will be called a (linear) operator. The linear subspace D(A) on which A
is defined is called the domain of A and is frequently required to be dense.
The kernel (also null space)
Ker(A) := {f ∈ D(A)|Af = 0} ⊆ X (1.58)
28 1. A first look at Banach and Hilbert spaces

and range
Ran(A) := {Af |f ∈ D(A)} = AD(A) ⊆ Y (1.59)
are again linear subspaces. Note that a linear map A will be continuous if
and only if it is continuous at 0, that is, xn ∈ D(A) → 0 implies Axn → 0.
The operator A is called bounded if the operator norm
kAk := sup kAf kY (1.60)
f ∈D(A),kf kX =1

is finite. This says that A is bounded if the image of the closed unit ball
B̄1 (0) ⊂ X is contained in some closed ball B̄r (0) ⊂ Y of finite radius r
(with the smallest radius being the operator norm). Hence A is bounded if
and only if it maps bounded sets to bounded sets.
Note that if you replace the norm on X or Y then the operator norm
will of course also change in general. However, if the norms are equivalent
so will be the operator norms.
By construction, a bounded operator satisfies
kAf kY ≤ kAkkf kX , f ∈ D(A), (1.61)
and hence is Lipschitz continuous, that is, kAf − AgkY ≤ kAkkf − gkX for
f, g ∈ D(A). In particular, it is continuous. The converse is also true:
Theorem 1.14. A linear operator A is bounded if and only if it is continu-
ous.

Proof. Suppose A is continuous but not bounded. Then there is a sequence

of unit vectors un ∈ D(A) such that kAun kY ≥ n. Then fn := n1 un converges
to 0 but kAfn kY ≥ 1 does not converge to 0. 

Of course it suffices to check continuity at one point in X, say at 0, since

continuity at all other points will then follow by a simple translation.
If X is finite dimensional, then every operator is bounded.
Lemma 1.15. Let X, Y be normed spaces with X finite dimensional. Then
every linear operator A : X → Y is bounded.

Proof. Choose a basis {xj }nj=1 for X such that every x ∈ X can be written
as x = nj=1 αj xj . By Theorem 1.8 we can assume kxkX = ( nj=1 |αj |2 )1/2
P P
without loss of generality. Then
v
n
u n
X uX
kAxkY ≤ |αj |kAxj kY ≤ t kAxj k2Y kxk
j=1 j=1

and thus kAk ≤ ( nj=1 kAxj k2Y )1/2 .

P

1.6. Bounded operators 29

In the infinite dimensional case an operator can be unbounded. More-

over, one and the same operation might be bounded (i.e. continuous) or
unbounded, depending on the norm chosen.
Example 1.16. Let X := `p (N) and a ∈ `∞ (N). Consider the multiplication
operator A : X → X defined by
(Ab)j := aj bj .
Then |(Ab)j | ≤ kak∞ |bj | shows kAk ≤ kak∞ . In fact, we even have kAk =
kak∞ (show this). If a is unbounded we need a domain D(A) := {b ∈
`p (N)|(aj bj )j∈N ∈ `p (N)} and A will be unbounded (show this). 
Example 1.17. Consider the vector space of differentiable functions X :=
C 1 [0, 1] and equip it with the norm (cf. Problem 1.32)
kf k∞,1 := max |f (x)| + max |f 0 (x)|.
x∈[0,1] x∈[0,1]

Let Y := C[0, 1] and observe that the differential operator A = d

dx :X→Y
is bounded since
kAf k∞ = max |f 0 (x)| ≤ max |f (x)| + max |f 0 (x)| = kf k∞,1 .
x∈[0,1] x∈[0,1] x∈[0,1]

However, if we consider A = dx d
: D(A) ⊆ Y → Y defined on D(A) =
C [0, 1], then we have an unbounded operator. Indeed, choose un (x) :=
1

sin(nπx) which is normalized, kun k∞ = 1, and observe that

Aun (x) = u0n (x) = nπ cos(nπx)
is unbounded, kAun k∞ = nπ. Note that D(A) contains the set of polyno-
mials and thus is dense by the Weierstraß approximation theorem (Theo-
rem 1.3). 
If A is bounded and densely defined, it is no restriction to assume that
it is defined on all of X.
Theorem 1.16. Let A : D(A) ⊆ X → Y be a bounded linear operator
between a normed space X and a Banach space Y . If D(A) is dense, there
is a unique (continuous) extension of A to X which has the same operator
norm.

Proof. Since a bounded operator maps Cauchy sequences to Cauchy se-

quences, this extension can only be given by
Af := lim Afn , fn ∈ D(A), f ∈ X.
n→∞
To show that this definition is independent of the sequence fn → f , let
gn → f be a second sequence and observe
kAfn − Agn k = kA(fn − gn )k ≤ kAkkfn − gn k → 0.
30 1. A first look at Banach and Hilbert spaces

Since for f ∈ D(A) we can choose fn = f , we see that Af = Af in this case,

that is, A is indeed an extension. From continuity of vector addition and
scalar multiplication it follows that A is linear. Finally, from continuity of
the norm we conclude that the operator norm does not increase. 

The set of all bounded linear operators from X to Y is denoted by

L (X, Y ). If X = Y , we write L (X) := L (X, X). An operator in L (X, C)
is called a bounded linear functional, and the space X ∗ := L (X, C) is
called the dual space of X. The dual space takes the role of coordinate
functions in a Banach space.
Example 1.18. Let X be a finite dimensional space and {xj }nj=1 a basis.
Then every x ∈ X can be uniquely written as x = nj=1 αj xj and we can
P

define linear functionals via `j (x) := αj for 1 ≤ j ≤ n. The functionals

{`j }nj=1 are called a dual basis since `k (xj ) = δj,k and since any other
linear functional ` ∈ X ∗ can be written as ` = nj=1 `(xj )`j . In particular,
P
X and X ∗ have the same dimension. 
Example 1.19. Let X := `p (N). Then the coordinate functions

`j (a) := aj

are bounded linear functionals: |`j (a)| = |aj | ≤ kakp and hence k`j k = 1.
More general, let b ∈ `q (N) where p1 + 1q = 1. Then

n
X
`b (a) := bj aj
j=1

is a bounded linear functional satisfying k`b k ≤ kbkq by Hölder’s inequality.

In fact, we even have k`b k = kbkq (Problem 4.20). Note that the first example
is a special case of the second one upon choosing b = δ j . 
Example 1.20. Consider X := C(I). Then for every x0 ∈ I the point
evaluation `x0 (f ) := f (x0 ) is a bounded linear functional. In fact, k`x0 k = 1
(show this).
However,
q note that `x0 is unbounded on Lcont (I)! To see this take
2

2 max(0, 1 − n|x − x0 |) which is a triangle shaped peak sup-

3n
fn (x) :=
ported on [x0 − n−1 , x0 + n−1 ] and normalized according to kfn k2 = 1 for
n sufficiently
q large such that the support is contained in I. Then `x0 (f ) =
2 → ∞. This implies that `x0 cannot be extended to the com-
3n
fn (x0 ) =
pletion of Lcont (I) in a natural way and reflects the
2 fact that the integral
cannot see individual points (changing the value of a function at one point
does not change its integral). 
1.6. Bounded operators 31

Example 1.21. Consider X := C(I) and let g be some continuous function.

Then Z b
`g (f ) := g(x)f (x)dx
a
is a linear functional with norm k`g k = kgk1 . Indeed, first of all note that
Z b Z b
|`g (f )| ≤ |g(x)f (x)|dx ≤ kf k∞ |g(x)|dx
a a
shows k`g k ≤ kgk1 . To see that we have equality consider fε = g ∗ /(|g| + ε)
and note
Z b Z b
|g(x)|2 |g(x)|2 − ε2
|`g (fε )| = dx ≥ dx = kgk1 − (b − a)ε.
a |g(x)| + ε a |g(x)| + ε
Since kfε k ≤ 1 and ε > 0 is arbitrary this establishes the claim. 
Theorem 1.17. The space L (X, Y ) together with the operator norm (1.60)
is a normed space. It is a Banach space if Y is.

Proof. That (1.60) is indeed a norm is straightforward. If Y is complete

and An is a Cauchy sequence of operators, then An f converges to an element
g for every f . Define a new operator A via Af = g. By continuity of
the vector operations, A is linear and by continuity of the norm kAf k =
limn→∞ kAn f k ≤ (limn→∞ kAn k)kf k, it is bounded. Furthermore, given
ε > 0, there is some N such that kAn − Am k ≤ ε for n, m ≥ N and thus
kAn f −Am f k ≤ εkf k. Taking the limit m → ∞, we see kAn f −Af k ≤ εkf k;
that is, An → A. 

The Banach space of bounded linear operators L (X) even has a multi-
plication given by composition. Clearly, this multiplication is distributive
(A + B)C = AC + BC, A(B + C) = AB + BC, A, B, C ∈ L (X) (1.62)
and associative
(AB)C = A(BC), α (AB) = (αA)B = A (αB), α ∈ C. (1.63)
Moreover, it is easy to see that we have
kABk ≤ kAkkBk. (1.64)
In other words, L (X) is a so-called Banach algebra. However, note that
our multiplication is not commutative (unless X is one-dimensional). We
even have an identity, the identity operator I, satisfying kIk = 1.
Problem 1.30. Consider X = Cn and let A ∈ L (X) be a matrix. Equip
X with the norm (show that this is a norm)
kxk∞ := max |xj |
1≤j≤n
32 1. A first look at Banach and Hilbert spaces

and compute the operator norm kAk with respect to this norm in terms of
the matrix entries. Do the same with respect to the norm
X
kxk1 := |xj |.
1≤j≤n

Problem 1.31. Show that the integral operator

Z 1
(Kf )(x) := K(x, y)f (y)dy,
0
where K(x, y) ∈ C([0, 1] × [0, 1]), defined on D(K) := C[0, 1], is a bounded
operator both in X := C[0, 1] (max norm) and X := L2cont (0, 1). Show that
the norm in the X = C[0, 1] case is given by
Z 1
kKk = max |K(x, y)|dy.
x∈[0,1] 0

Problem* 1.32. Let I be a compact interval. Show that the set of dif-
ferentiable functions C 1 (I) becomes a Banach space if we set kf k∞,1 :=
maxx∈I |f (x)| + maxx∈I |f 0 (x)|.
Problem* 1.33. Show that kABk ≤ kAkkBk for every A, B ∈ L (X).
Conclude that the multiplication is continuous: An → A and Bn → B imply
An Bn → AB.
Problem 1.34. Let A ∈ L (X) be a bijection. Show
kA−1 k−1 = inf kAf k.
f ∈X,kf k=1

Problem* 1.35. Suppose B ∈ L (X) with kBk < 1. Then I+B is invertible
with
X∞
−1
(I + B) = (−1)n B n .
n=0
Consequently for A, B ∈ L (X, Y ), A + B is invertible if A is invertible and
kBk < kA−1 k−1 .
Problem* 1.36. Let
∞
X
f (z) := fj z j , |z| < R,
j=0

be a convergent power series with radius of convergence R > 0. Suppose X is

a Banach space and A ∈ L (X) is a bounded operator with lim supn kAn k1/n <
R (note that by kAn k ≤ kAkn the limsup is finite). Show that
∞
X
f (A) := fj Aj
j=0
1.7. Sums and quotients of Banach spaces 33

exists and defines a bounded linear operator. Moreover, if f and g are two
such functions and α ∈ C, then
(f + g)(A) = f (A) + g(A), (αf )(A) = αf (a), (f g)(A) = f (A)g(A).
(Hint: Problem 1.5.)
Problem* 1.37. Show that a linear map ` : X → C is continuous if and
only if its kernel is closed. (Hint: If ` is not continuous, we can find a
sequence of normalized vectors xn with |`(xn )| → ∞ and a vector y with
`(y) = 1.)

1.7. Sums and quotients of Banach spaces

Given two Banach spaces X1 and X2 we can define their (direct) sum
X := X1 ⊕ X2 as the Cartesian product X1 × X2 together with the norm
k(x1 , x2 )k := kx1 k + kx2 k. Clearly X is again a Banach space and a se-
quence in X converges if and only if the components converge in X1 and X2 ,
respectively. In fact, since all norms on R2 are equivalent (Theorem 1.8),
we could as well take k(x1 , x2 )kp := (kx1 kp + kx2 kp )1/p or k(x1 , x2 )k∞ :=
max(kx1 k, kx2 k). We will write X1 ⊕p X2 if we want to emphasize the norm
used. In particular, in the case of Hilbert spaces the choice p = 2 will ensure
that X is again a Hilbert space.
Note that X1 and X2 can be regarded as closed subspaces of X1 × X2
by virtue of the obvious embeddings x1 ,→ (x1 , 0) and x2 ,→ (0, x2 ). It
is straightforward to generalize this concept to finitely many spaces (Prob-
lem 1.38).
If Aj : D(Aj ) ⊆ Xj → Yj , j = 1, 2, are linear operators, then A1 ⊕ A2 :
D(A1 ) × D(A2 ) ⊆ X1 ⊕ X2 → Y1 ⊕ Y2 is defined as A1 ⊕ A2 (x1 , x2 ) =
(A1 x1 , A2 x2 ). Clearly A1 ⊕ A2 will be bounded if and only if both A1 and
A2 are bounded and kA1 ⊕ A2 k = max(kA1 k, kA2 k).
Note that if Aj : Xj → Y , j = 1, 2, there is another natural way of
defining an associated operator X1 ⊕ X2 → Y given by A1 ⊕A ˆ 2 (x1 , x2 ) :=
A1 x1 + A2 x2 . In particular, in the case Y = C we get that (X1 ⊕p X2 )∗ ∼ =
X1∗ ⊕q X2∗ for p1 + 1q = 1 via the identification (`1 , `2 ) ∈ X1∗ ⊕q X2∗ 7→ `1 ⊕`
ˆ 2∈
(X1 ⊕p X2 )∗ . Clearly this identification is bijective and preserves the norm
(to see this relate it to Hölder’s inequality in C2 and note that equality is
attained).
Given two subspaces M, N ⊆ X of a vector space, we can define their sum
as usual: M + N := {x + y|x ∈ M, y ∈ N }. In particular, the decomposition
x + y with x ∈ M , y ∈ N is unique iff M ∩ N = {0} and we will write M u N
in this case. It is important to observe, that M u N is in general different
from M ⊕ N since both have different norms. In fact, M u N might not even
34 1. A first look at Banach and Hilbert spaces

be closed (no problems occur if one of the spaces is finite dimensional — see
Corollary 1.19 below).
Example 1.22. Consider X := `p (N). Let M = {a ∈ X|a2n = 0} and N =
{a ∈ X|a2n+1 = n3 a2n }. Then both subspaces are closed and M ∩ N = {0}.
Moreover, M u N is dense since it contains all sequences with finite support.
However, it is not all of X since an = n12 6∈ M u N . Indeed, if we could
write a = b + c ∈ M u N , then c2n = 4n1 2 and hence c2n+1 = n4 contradicting
c ∈ N ⊆ X. 
A closed subspace M is called complemented if we can find another
closed subspace N with M ∩ N = {0} and M u N = X. In this case every
x ∈ X can be uniquely written as x = x1 + x2 with x1 ∈ M , x2 ∈ N and
we can define a projection P : X → M , x 7→ x1 . By definition P 2 = P
and we have a complementary projection Q := I − P with Q : X → N ,
x 7→ x2 . Moreover, it is straightforward to check M = Ker(Q) = Ran(P )
and N = Ker(P ) = Ran(Q). Of course one would like P (and hence also
Q) to be continuous. If we consider the linear operator φ : M ⊕ N → X,
(x1 , x2 ) → x1 +x2 then this is equivalent to the question if φ−1 is continuous.
By the triangle inequality φ is continuous with kφk ≤ 1 and the inverse
mapping theorem (Theorem 4.6) will answer this question affirmative.
It is important to emphasize, that it is precisely the requirement that N
is closed which makes P continuous (conversely observe that N = Ker(P )
is closed if P is continuous). Without this requirement we can always find
N by a simple application of Zorn’s lemma (order the subspaces which have
trivial intersection with M by inclusion and note that a maximal element has
the required properties). Moreover, the question which closed subspaces can
be complemented is a highly nontrivial one. If M is finite (co)dimensional,
then it can be complemented (see Problems 1.44 and 4.26).
Given a subspace M of a linear space X we can define the quotient
space X/M as the set of all equivalence classes [x] = x + M with respect
to the equivalence relation x ≡ y if x − y ∈ M . It is straightforward to see
that X/M is a vector space when defining [x] + [y] = [x + y] and α[x] = [αx]
(show that these definitions are independent of the representative of the
equivalence class). In particular, for a linear operator A : X → Y the linear
space Coker(A) := Y / Ran(A) is know as the cokernel of A. The dimension
of X/M is known as the codimension of M .

Lemma 1.18. Let M be a closed subspace of a Banach space X. Then X/M

together with the norm

k[x]k := dist(x, M ) = inf kx + yk (1.65)

y∈M

is a Banach space.
1.7. Sums and quotients of Banach spaces 35

Proof. First of all we need to show that (1.65) is indeed a norm. If k[x]k = 0
we must have a sequence yj ∈ M with yj → −x and since M is closed we
conclude x ∈ M , that is [x] = [0] as required. To see kα[x]k = |α|k[x]k we
use again the definition
kα[x]k = k[αx]k = inf kαx + yk = inf kαx + αyk
y∈M y∈M
= |α| inf kx + yk = |α|k[x]k.
y∈M

The triangle inequality follows with a similar argument and is left as an

exercise.
Thus (1.65) is a norm and it remains to show that X/M is complete. To
this end let [xn ] be a Cauchy sequence. Since it suffices to show that some
subsequence has a limit, we can assume k[xn+1 ]−[xn ]k < 2−n without loss of
generality. Moreover, by definition of (1.65) we can chose the representatives
xn such that kxn+1 − xn k < 2−n (start with x1 and then chose the remaining
ones inductively). By construction xn is a Cauchy sequence which has a limit
x ∈ X since X is complete. Moreover, by k[xn ]−[x]k = k[xn −x]k ≤ kxn −xk
we see that [x] is the limit of [xn ]. 

Observe that k[x]k = dist(x, M ) = 0 whenever x ∈ M and hence we only

get a semi-norm if M is not closed.
Example 1.23. If X := C[0, 1] and M := {f ∈ X|f (0) = 0} then X/M ∼ =
C. 
Example 1.24. If X := c(N), the convergent sequences and M := c0 (N) the
sequences converging to 0, then X/M ∼= C. In fact, note that every sequence
x ∈ c(N) can be written as x = y + αe with y ∈ c0 (N), e = (1, 1, 1, . . . ), and
α ∈ C its limit. 
Note that by k[x]k ≤ kxk the quotient map π : X → X/M , x 7→ [x] is
bounded with norm at most one. As a small application we note:
Corollary 1.19. Let X be a Banach space and let M, N ⊆ X be two closed
subspaces with one of them, say N , finite dimensional. Then M + N is also
closed.

Proof. If π : X → X/M denotes the quotient map, then M + N =

π −1 (π(N )). Moreover, since π(N ) is finite dimensional it is closed and hence
π −1 (π(N )) is closed by continuity. 
Ln
Problem* 1.38. Let Xj , j = 1, . . . , n, be Banach spaces. Let X := p,j=1 Xj
be the Cartesian product X1 × · · · × Xn together with the norm
 1/p
 Pn p
k(x1 , . . . , xn )kp := j=1 kxj k , 1 ≤ p < ∞,
j=1,...,n kxj k, p = ∞.
max
36 1. A first look at Banach and Hilbert spaces

Show that X is a Banach space. Show that all norms are equivalent and that
this sum is associative (X1 ⊕p X2 ) ⊕p X3 = X1 ⊕p (X2 ⊕p X3 ).
L
Problem 1.39. Let Xj , j ∈ N, be Banach spaces. Let X := p,j∈N Xj be
the set of all elements x = (xj )j∈N of the Cartesian product for which the
norm  1/p
kxj kp
 P
j∈N , 1 ≤ p < ∞,
kxkp :=
j∈N kxj k, p = ∞,
max

is finite. Show that X is a Banach space. Show that for 1 ≤ p < ∞ the
elements with finitely many nonzero terms are dense and conclude that X is
separable if all Xj are.

Problem 1.40. Let ` be a nontrivial linear functional. Then its kernel has
codimension one.

Problem 1.41. Compute k[e]k in `∞ (N)/c0 (N), where e = (1, 1, 1, . . . ).

Problem 1.42 (Complexification). Given a real normed space X its com-

plexification is given by XC := X × X together with the (complex) scalar
multiplication α(x, y) = (Re(α)x − Im(α)y, Re(α)y + Im(α)x). By virtue of
the embedding x ,→ (x, 0) you should of course think of (x, y) as x + iy.
Show that
kx + iykC := max k cos(t)x + sin(t)yk,
0≤t≤π

defines a norm on XC which satisfies kxkC = kxk and

max(kxk, kyk) ≤ kx + iykC ≤ (kxk2 + ky 2 k)1/2

In particular, this norm is equivalent to the product norm on X ⊕ X.
If X is a Hilbert space, then the above norm will in general not give
rise to a scalar product. However, any bilinear form s : X × X → R gives
rise to a sesquilinear
form sC (x1 + iy1 , x2 + iy2 ) := s(x1 , x2 ) + s(y1 , y2 ) +
i s(x1 , y2 ) − s(y1 , x2 ) . If s is symmetric or positive definite, so will be sC .
The corresponding norm satisfies hx + iy, x + iyiC = kxk2 + kyk2 and is
equivalent to the above one since 21 (kxk2 + kyk2 ) ≤ kx + iyk2C ≤ kxk2 + kyk2 .
Given two real normed spaces X, Y , every linear operator A : X → Y
gives rise to a linear operator AC : XC → YC via AC (x + iy) = Ax + iAy.
Show kAC k = kAk.

Problem* 1.43. Suppose A ∈ L (X, Y ). Show that Ker(A) is closed.

Suppose M ⊆ Ker(A) is a closed subspace. Show that the induced map
à : X/M → Y , [x] 7→ Ax is a well-defined operator satisfying kÃk = kAk
and Ker(Ã) = Ker(A)/M . In particular, Ã is injective for M = Ker(A).
1.8. Spaces of continuous and differentiable functions 37

Problem* 1.44. Show that if a closed subspace M of a Banach space X has

finite codimension, then it can be complemented. (Hint: Start with a basis
{[xj ]} for X/M and choose a corresponding dual basis {`k } with `k ([xj ]) =
δj,k .)

1.8. Spaces of continuous and differentiable functions

In this section we introduce a few further sets of continuous and differen-
tiable functions which are of interest in applications. Let I be some compact
interval, then we can make C 1 (I) into a Banach space by (Problem 1.32)
by introducing the norm kf k1,∞ := kf k∞ + kf 0 k∞ . By a straightforward
extension we can even get (cf. Problem 1.46)
Theorem 1.20. Let I ⊆ R be some interval. The space Cbk (I) of all func-
tions whose partial derivatives up to order k are bounded and continuous
form a Banach space with norm
X
kf kk,∞ := sup |f (k) (x)|. (1.66)
x∈I
|α|≤k

Note that the space Cbk (I) could be further refined by requiring the
highest derivatives to be Hölder continuous. Recall that a function f : I → C
is called uniformly Hölder continuous with exponent γ ∈ (0, 1] if
|f (x) − f (y)|
[f ]γ := sup (1.67)
x6=y∈I |x − y|γ
is finite. Clearly, any Hölder continuous function is uniformly continuous
and, in the special case γ = 1, we obtain the Lipschitz continuous func-
tions. Note that for γ = 0 the Hölder condition boils down to boundedness
and also the case γ > 1 is not very interesting (Problem 1.45).
Example 1.25. By the mean value theorem every function f ∈ Cb1 (I) is
Lipschitz continuous with [f ]γ ≤ kf 0 k∞ . 
Example 1.26. The prototypical example of a Hölder continuous function
is of course f (x) = xγ on [0, ∞) with γ ∈ (0, 1]. In fact, without loss of
generality we can assume 0 ≤ x < y and set t = xy ∈ [0, 1). Then we have
y γ − xγ 1 − tγ 1−t
γ
≤ γ
≤ = 1.
(y − x) (1 − t) 1−t
From this one easily gets further examples since the composition of two
Hölder continuous functions is again Hölder continuous (the exponent being
the product). 
It is easy to verify that this is a seminorm and that the corresponding
space is complete.
38 1. A first look at Banach and Hilbert spaces

Theorem 1.21. Let I ⊆ R be an interval. The space Cbk,γ (I) of all functions
whose partial derivatives up to order k are bounded and Hölder continuous
with exponent γ ∈ (0, 1] form a Banach space with norm

kf kk,γ,∞ := kf kk,∞ + [f (k) ]γ . (1.68)

As already noted before, in the case γ = 0 we get a norm which is equiv-

alent to kf k∞,k and we will set Cbk,0 (I) := Cbk (I) for notational convenience
later on.
Note that by the mean value theorem all derivatives up to order lower
than k are automatically Lipschitz continuous. Moreover, every Hölder con-
tinuous function is uniformly continuous and hence has a unique extension
to the closure I (cf. Theorem B.38). In this sense, the spaces Cb0,γ (I) and
Cb0,γ (I) are naturally isomorphic. Finally, since Hölder continuous functions
on a bounded domain are automatically bounded, we can drop the subscript
b in this situation.

Theorem 1.22. Suppose I ⊂ R is a compact interval. Then C 0,γ2 (I) ⊆

C 0,γ1 (I) ⊆ C(I) for 0 < γ1 < γ2 ≤ 1 with the embeddings being compact.

Proof. That we have continuous embeddings follows since |x − y|−γ1 =

|x−y|−γ2 +(γ2 −γ1 ) ≤ (2r)γ2 −γ1 |x−y|−γ2 if r denotes the length of I. Moreover,
that the embedding C 0,γ1 (I) ⊆ C(I) is compact follows from the Arzelà–
Ascoli theorem (Theorem B.39). To see the remaining claim let fm be a
bounded sequence in C 0,γ1 (I), explicitly kfm k∞ ≤ C and [f ]γ1 ≤ C. Hence
by the Arzelà–Ascoli theorem we can assume that fm converges uniformly to
some f ∈ C(I). Moreover, taking the limit in |fm (x) − fm (y)| ≤ C|x − y|γ1
we see that we even have f ∈ C 0,γ1 (I). To see that f is the limit of fm in
C 0,γ2 (I) we need to show [gm ]γ2 → 0, where gm := fm − f . Now observe
that
|gm (x) − gm (y)| |gm (x) − gm (y)|
[gm ]γ2 = sup γ
+ sup
x6=y∈I:|x−y|≥ε |x − y| 2
x6=y∈I:|x−y|<ε |x − y|γ2
≤ 2kgm k∞ ε−γ2 + [gm ]γ1 εγ1 −γ2 ≤ 2kgm k∞ ε−γ2 + 2Cεγ1 −γ2 ,

implying lim supm→∞ [gm ]γ2 ≤ 2Cεγ1 −γ2 and since ε > 0 is arbitrary this
establishes the claim. 

As pointed out in the example before, the embedding Cb1 (I) ⊆ Cb0,1 (I) is
continuous and combining this with the previous result immediately gives

Corollary 1.23. Suppose I ⊂ R is a compact interval, k1 , k2 ∈ N0 , and

0 ≤ γ1 , γ2 ≤ 1. Then C k2 ,γ2 (I) ⊆ C k1 ,γ1 (I) for k1 + γ1 ≤ k2 + γ2 with the
embeddings being compact if the inequality is strict.
1.8. Spaces of continuous and differentiable functions 39

For now continuous functions on intervals will be sufficient for our pur-
pose. However, once we delve deeper into the subject we will also need
continuous functions on topological spaces X. Luckily most of the results
extend to this case in a more or less straightforward way. If you are not
familiar with these extensions you can find them in Section B.8.
Problem 1.45. Let I be an interval. Suppose f : I → C is Hölder continu-
ous with exponent γ > 1. Show that f is constant.
Problem* 1.46. Suppose X is a vector space Pand k.kj , 1 ≤ j ≤ n, is a
finite family of seminorms. Show that kxk := nj=1 kxkj is a seminorm. It
is a norm if and only if kxkj = 0 for all j implies x = 0.
Problem 1.47. Let I. Show that the product of two bounded Hölder contin-
uous functions is again Hölder continuous with
[f g]γ ≤ kf k∞ [g]γ + [f ]γ kgk∞ .
 Chapter 2

Hilbert spaces

The additional geometric structure of Hilbert spaces allows for an intuitive

geometric solution of many problems. In fact, in many situations, e.g. in
Quantum Mechanics, Hilbert spaces occur naturally. This makes them the
weapon of choice whenever possible. Throughout this chapter H will be a
(complex) Hilbert space.

2.1. Orthonormal bases

In this section we will investigate orthonormal series and you will notice
hardly any difference between the finite and infinite dimensional cases. As
our first task, let us generalize the projection into the direction of one vector.
A set of vectors {uj } is called an orthonormal set if huj , uk i = 0
for j 6= k and huj , uj i = 1. Note that every orthonormal set is linearly
independent (show this).
Lemma 2.1. Suppose {uj }nj=1 is a finite orthonormal set in a Hilbert space
H. Then every f ∈ H can be written as
n
X
f = fk + f⊥ , fk := huj , f iuj , (2.1)
j=1

where fk and f⊥ are orthogonal. Moreover, huj , f⊥ i = 0 for all 1 ≤ j ≤ n.

In particular,
Xn
2
kf k = |huj , f i|2 + kf⊥ k2 . (2.2)
j=1

Moreover, every fˆ in the span of {uj }nj=1 satisfies

kf − fˆk ≥ kf⊥ k (2.3)

41
42 2. Hilbert spaces

with equality holding if and only if fˆ = fk . In other words, fk is uniquely

characterized as the vector in the span of {uj }nj=1 closest to f .

Proof. A straightforward calculation shows huj , f − fk i = 0 and hence fk

and f⊥ := f − fk are orthogonal. The formula for the norm follows by
applying (1.40) iteratively.
Pn
Now, fix a vector fˆ := j=1 αj uj in the span of {uj }j=1 . Then one
n

computes
kf − fˆk2 = kfk + f⊥ − fˆk2 = kf⊥ k2 + kfk − fˆk2
n
X
2
= kf⊥ k + |αj − huj , f i|2
j=1

from which the last claim follows. 

From (2.2) we obtain Bessel’s inequality

n
X
|huj , f i|2 ≤ kf k2 (2.4)
j=1

with equality holding if and only if f lies in the span of {uj }nj=1 .
Of course, since we cannot assume H to be a finite dimensional vec-
tor space, we need to generalize Lemma 2.1 to arbitrary orthonormal sets
{uj }j∈J . We start by assuming that J is countable. Then Bessel’s inequality
(2.4) shows that X
|huj , f i|2 (2.5)
j∈J
converges absolutely. Moreover, for any finite subset K ⊂ J we have
X X
k huj , f iuj k2 = |huj , f i|2 (2.6)
j∈K j∈K

by the Pythagorean theorem and thus j∈J huj , f iuj is a Cauchy sequence
P

if and only if j∈J |huj , f i|2 is. Now let J be arbitrary. Again, Bessel’s
P
inequality shows that for any given ε > 0 there are at most finitely many
j for which |huj , f i| ≥ ε (namely at most kf k/ε). Hence there are at most
countably many j for which |huj , f i| > 0. Thus it follows that
X
|huj , f i|2 (2.7)
j∈J

is well defined (as a countable sum over the nonzero terms) and (by com-
pleteness) so is X
huj , f iuj . (2.8)
j∈J
Furthermore, it is also independent of the order of summation.
2.1. Orthonormal bases 43

In particular, by continuity of the scalar product we see that Lemma 2.1

can be generalized to arbitrary orthonormal sets.
Theorem 2.2. Suppose {uj }j∈J is an orthonormal set in a Hilbert space H.
Then every f ∈ H can be written as
X
f = fk + f⊥ , fk := huj , f iuj , (2.9)
j∈J

where fk and f⊥ are orthogonal. Moreover, huj , f⊥ i = 0 for all j ∈ J. In

particular,
X
kf k2 = |huj , f i|2 + kf⊥ k2 . (2.10)
j∈J

Furthermore, every fˆ ∈ span{uj }j∈J satisfies

kf − fˆk ≥ kf⊥ k (2.11)
with equality holding if and only if fˆ = fk . In other words, fk is uniquely
characterized as the vector in span{uj }j∈J closest to f .

Proof. The first part follows as in Lemma 2.1 using continuity of the scalar
product. The same is true for the lastP part except for the fact that every
f ∈ span{uj }j∈J can be written as f = j∈J αj uj (i.e., f = fk ). To see this,
let fn ∈ span{uj }j∈J converge to f . Then kf −fn k2 = kfk −fn k2 +kf⊥ k2 → 0
implies fn → fk and f⊥ = 0. 

Note that from Bessel’s inequality (which of course still holds), it follows
that the map f → fk is continuous.
Of course we are
P particularly interested in the case where every f ∈ H
can be written as j∈J huj , f iuj . In this case we will call the orthonormal
set {uj }j∈J an orthonormal basis (ONB).
If H is separable it is easy to construct an orthonormal basis. In fact, if
H is separable, then there exists a countable total set {fj }N
j=1 . Here N ∈ N
if H is finite dimensional and N = ∞ otherwise. After throwing away some
vectors, we can assume that fn+1 cannot be expressed as a linear combination
of the vectors f1 , . . . , fn . Now we can construct an orthonormal set as
follows: We begin by normalizing f1 :
f1
u1 := . (2.12)
kf1 k
Next we take f2 and remove the component parallel to u1 and normalize
again:
f2 − hu1 , f2 iu1
u2 := . (2.13)
kf2 − hu1 , f2 iu1 k
44 2. Hilbert spaces

Proceeding like this, we define recursively

Pn−1
fn − j=1 huj , fn iuj
un := Pn−1 . (2.14)
kfn − j=1 huj , fn iuj k
This procedure is known as Gram–Schmidt orthogonalization. Hence
we obtain an orthonormal set {uj }N j=1 such that span{uj }j=1 = span{fj }j=1
n n

for any finite n and thus also for n = N (if N = ∞). Since {fj }N j=1 is total,
so is {uj }j=1 . Now suppose there is some f = fk + f⊥ ∈ H for which f⊥ 6= 0.
N

Since {uj }N ˆ ˆ
j=1 is total, we can find a f in its span such that kf − f k < kf⊥ k,
contradicting (2.11). Hence we infer that {uj }N j=1 is an orthonormal basis.

Theorem 2.3. Every separable Hilbert space has a countable orthonormal

basis.
Example 2.1. The vectors {δ n }n∈N form an orthonormal basis for `2 (N). 
Example 2.2. In L2cont (−1, 1), we can orthogonalize the monomials fn (x) =
xn (which are total by the Weierstraß approximation theorem — Theo-
rem 1.3). The resulting polynomials are up to a normalization known as
Legendre polynomials
3 x2 − 1
P0 (x) = 1, P1 (x) = x, P2 (x) = , ... (2.15)
2
(which are normalized such that Pn (1) = 1). 
Example 2.3. The set of functions
1
un (x) = √ einx , n ∈ Z, (2.16)
2π
forms an orthonormal basis for H = L2cont (0, 2π). The corresponding or-
thogonal expansion is just the ordinary Fourier series. We will discuss this
example in detail in Section 2.5. 
The following equivalent properties also characterize a basis.
Theorem 2.4. For an orthonormal set {uj }j∈J in a Hilbert space H, the
following conditions are equivalent:
(i) {uj }j∈J is a maximal orthogonal set.
(ii) For every vector f ∈ H we have
X
f= huj , f iuj . (2.17)
j∈J

(iii) For every vector f ∈ H we have Parseval’s relation

X
kf k2 = |huj , f i|2 . (2.18)
j∈J
2.1. Orthonormal bases 45

(iv) huj , f i = 0 for all j ∈ J implies f = 0.

Proof. We will use the notation from Theorem 2.2.

(i) ⇒ (ii): If f⊥ 6= 0, then we can normalize f⊥ to obtain a unit vector f˜⊥
which is orthogonal to all vectors uj . But then {uj }j∈J ∪ {f˜⊥ } would be a
larger orthonormal set, contradicting the maximality of {uj }j∈J .
(ii) ⇒ (iii): This follows since (ii) implies f⊥ = 0.
(iii) ⇒ (iv): If hf, uj i = 0 for all j ∈ J, we conclude kf k2 = 0 and hence
f = 0.
(iv) ⇒ (i): If {uj }j∈J were not maximal, there would be a unit vector g such
that {uj }j∈J ∪ {g} is a larger orthonormal set. But huj , gi = 0 for all j ∈ J
implies g = 0 by (iv), a contradiction. 

By continuity of the norm it suffices to check (iii), and hence also (ii),
for f in a dense set. In fact, by the inverse triangle inequality for `2 (N) and
the Bessel inequality we have
 
  s sX
X 2
X
2
 X
 |huj , f i| − |huj , gi|  ≤ |huj , f − gi|2 |huj , f + gi|2

 j∈J j∈J  j∈J j∈J

≤ kf − gkkf + gk (2.19)
implying |huj , fn i|2 → |huj , f i|2 if fn → f .
P P
j∈J j∈J
It is not surprising that if there is one countable basis, then it follows
that every other basis is countable as well.
Theorem 2.5. In a Hilbert space H every orthonormal basis has the same
cardinality.

Proof. Let {uj }j∈J and {vk }k∈K be two orthonormal bases. We first look
at the case where one of them, say the first, is finite dimensional: J =
{1, . . . , n}. Suppose
Pn the other basis has at least n elements {1, . . . , n} ⊆
K. Then vk = j=1 Uk,j uj , where Uk,j = huj , vk i. By δj,k = hvj , vk i =
Pn Pn
l=1 Uj,l Uk,l we see uj = k=1 Uk,j vk showing that K cannot have more
∗ ∗

than n elements.
Now let us turn to the case where both J and K are infinite. Set Kj =
6 0}. Since these are the expansion coefficients of uj with
{k ∈ K|hvk , uj i =
respect to {vk }k∈K , this set is countable (and nonempty). Hence the set
K̃ = j∈J Kj satisfies |K̃| ≤ |J × N| = |J| (Theorem A.9) But k ∈ K \ K̃
S

implies vk = 0 and hence K̃ = K. So |K| ≤ |J| and reversing the roles of J

and K shows |K| = |J|. 

The cardinality of an orthonormal basis is also called the Hilbert space

dimension of H.
46 2. Hilbert spaces

It even turns out that, up to unitary equivalence, there is only one sep-
arable infinite dimensional Hilbert space:
A bijective linear operator U ∈ L (H1 , H2 ) is called unitary if U pre-
serves scalar products:
hU g, U f i2 = hg, f i1 , g, f ∈ H1 . (2.20)
By the polarization identity, (1.47) this is the case if and only if U preserves
norms: kU f k2 = kf k1 for all f ∈ H1 (note a norm preserving linear operator
is automatically injective). The two Hilbert spaces H1 and H2 are called
unitarily equivalent in this case.
Let H be a separable infinite dimensional Hilbert space and let {uj }j∈N
be any orthogonal basis. Then the map U : H → `2 (N), f 7→ (huj , f i)j∈N is
unitary. Indeed by Theorem 2.4 (iii) it is norm preserving andPhence injective.
To see that it is onto, let a ∈ `2 (N) and observe that by k nj=m aj uj k2 =
Pn
j=m |aj | the vector f := j∈N aj uj is well defined and satisfies aj =
2
P

huj , f i. In particular,
Theorem 2.6. Any separable infinite dimensional Hilbert space is unitarily
equivalent to `2 (N).

Of course the same argument shows that every finite dimensional Hilbert
space of dimension n is unitarily equivalent to Cn with the usual scalar
product.
Finally we briefly turn to the case where H is not separable.
Theorem 2.7. Every Hilbert space has an orthonormal basis.

Proof. To prove this we need to resort to Zorn’s lemma (see Appendix A):
The collection of all orthonormal sets in H can be partially ordered by in-
clusion. Moreover, every linearly ordered chain has an upper bound (the
union of all sets in the chain). Hence Zorn’s lemma implies the existence of
a maximal element, that is, an orthonormal set which is not a proper subset
of every other orthonormal set. 

Hence, if {uj }j∈J is an orthogonal basis, we can show that H is unitarily

equivalent to `2 (J) and, by prescribing J, we can find a Hilbert space of any
given dimension. Here `2 (J) is the set of all complex-valued
P functions (aj )j∈J
where at most countably many values are nonzero and j∈J |aj |2 < ∞.
Example 2.4. Define the set of almost periodic functions AP (R) as the
closure of the set of trigonometric polynomials
n
X
f (t) = αk eiθk t , αk ∈ C, θk ∈ R,
k=1
2.1. Orthonormal bases 47

with respect to the sup norm. In particular AP (R) ⊂ Cb (R) is a Banach

space when equipped with the sup norm. Since the trigonometric polynomi-
als form an algebra, it is even a Banach algebra. Using the Stone–Weierstraß
theorem one can verify that every periodic function is almost periodic (make
the approximation on one period and note that you get the √ rest of R for free
from periodicity) but the converse is not true (e.g. e + e 2t is not periodic).
it i

It is not difficult to show that every almost periodic function has a mean
value Z T
1
M (f ) := lim f (t)dt
T →∞ 2T −T
and one can show that
hf, gi := M (f ∗ g)
defines a scalar product on AP (R) (only positivity is nontrivial and it will
not be shown here). Note that kf k ≤ kf k∞ . Abbreviating eθ (t) = eiθt one
computes M (eθ ) = 0 if θ 6= 0 and M (e0 ) = 1. In particular, {eθ }θ∈R is an
uncountable orthonormal set and
f (t) 7→ fˆ(θ) := heθ , f i = M (e−θ f )
maps AP (R) isometrically (with respect to k.k) into `2 (R). This map is
however not surjective (take e.g. a Fourier series which converges in mean
square but not uniformly — see later). 
Problem* 2.1. Given some vectors f1 , . . . , fn we define their Gram de-
terminant as
Γ(f1 , . . . , fn ) := det (hfj , fk i)1≤j,k≤n .
Show that the Gram determinant is nonzero if and only if the vectors are
linearly independent. Moreover, show that in this case
Γ(f1 , . . . , fn , g)
dist(g, span{f1 , . . . , fn })2 =
Γ(f1 , . . . , fn )
and
n
Y
Γ(f1 , . . . , fn ) ≤ kfj k2 .
j=1
with equality if the vectors are orthogonal. (Hint: How does Γ change when
you apply the Gram–Schmidt procedure?)
Problem 2.2. Let {uj } be some orthonormal basis. Show that a bounded
linear operator A is uniquely determined by its matrix elements Ajk :=
huj , Auk i with respect to this basis.
Problem 2.3. Give an example of a nonempty closed bounded subset of a
Hilbert space which does not contain an element with minimal norm. Can
this happen in finite dimensions? (Hint: Look for a discrete set.)
48 2. Hilbert spaces

2.2. The projection theorem and the Riesz lemma

Let M ⊆ H be a subset. Then M ⊥ := {f |hg, f i = 0, ∀g ∈ M } is called
the orthogonal complement of M . By continuity of the scalar prod-
uct it follows that M ⊥ is a closed linear subspace and by linearity that
(span(M ))⊥ = M ⊥ . For example, we have H⊥ = {0} since any vector in H⊥
must be in particular orthogonal to all vectors in some orthonormal basis.
Theorem 2.8 (Projection theorem). Let M be a closed linear subspace of a
Hilbert space H. Then every f ∈ H can be uniquely written as f = fk + f⊥
with fk ∈ M and f⊥ ∈ M ⊥ , where fk is uniquely characterized as the vector
in M closest to f . One writes
M ⊕ M⊥ = H (2.21)
in this situation.

Proof. Since M is closed, it is a Hilbert space and has an orthonormal

basis {uj }j∈J . Hence the existence part follows from Theorem 2.2. To see
uniqueness, suppose there is another decomposition f = f˜k + f˜⊥ . Then
fk − f˜k = f˜⊥ − f⊥ ∈ M ∩ M ⊥ = {0} (since g ∈ M ∩ M ⊥ implies kgk2 =
hg, gi = 0). 
Corollary 2.9. Every orthogonal set {uj }j∈J can be extended to an orthog-
onal basis.

Proof. Just add an orthogonal basis for ({uj }j∈J )⊥ . 

The operator PM f := fk is called the orthogonal projection corre-

sponding to M . Note that we have
2
PM = PM and hPM g, f i = hg, PM f i (2.22)
since hPM g, f i = hgk , fk i = hg, PM f i. Clearly we have PM ⊥ f = f −
PM f = f⊥ . Furthermore, (2.22) uniquely characterizes orthogonal projec-
tions (Problem 2.6).
Moreover, if M is a closed subspace, we have PM ⊥⊥ = I − PM ⊥ =
I − (I − PM ) = PM ; that is, M ⊥⊥ = M . If M is an arbitrary subset, we have
at least
M ⊥⊥ = span(M ). (2.23)
Note that by H⊥ = {0} we see that M ⊥ = {0} if and only if M is total.
Finally we turn to linear functionals, that is, to operators ` : H → C.
By the Cauchy–Schwarz inequality we know that `g : f 7→ hg, f i is a bounded
linear functional (with norm kgk). In turns out that, in a Hilbert space, every
bounded linear functional can be written in this way.
2.2. The projection theorem and the Riesz lemma 49

Theorem 2.10 (Riesz lemma). Suppose ` is a bounded linear functional on

a Hilbert space H. Then there is a unique vector g ∈ H such that `(f ) = hg, f i
for all f ∈ H.
In other words, a Hilbert space is equivalent to its own dual space H∗ ∼
=H
via the map f 7→ hf, .i which is a conjugate linear isometric bijection between
H and H∗ .

Proof. If ` ≡ 0, we can choose g = 0. Otherwise Ker(`) = {f |`(f ) = 0} is a

proper subspace and we can find a unit vector g̃ ∈ Ker(`)⊥ . For every f ∈ H
we have `(f )g̃ − `(g̃)f ∈ Ker(`) and hence
0 = hg̃, `(f )g̃ − `(g̃)f i = `(f ) − `(g̃)hg̃, f i.
In other words, we can choose g = `(g̃)∗ g̃. To see uniqueness, let g1 , g2 be
two such vectors. Then hg1 − g2 , f i = hg1 , f i − hg2 , f i = `(f ) − `(f ) = 0 for
every f ∈ H, which shows g1 − g2 ∈ H⊥ = {0}. 

In particular, this shows that H∗ is again a Hilbert space whose scalar

product (in terms of the above identification) is given by hhf, .i, hg, .iiH∗ =
hf, gi∗ .
We can even get a unitary map between H and H∗ but such a map is
not unique. To this end note that every Hilbert space has a conjugation C
which generalizes taking the complex conjugate of every coordinate. In fact,
choosing an orthonormal basis (and different choices will produce different
maps in general) we can set
X X
Cf := huj , f i∗ uj = hf, uj iuj .
j∈J j∈J

Then C is conjugate linear, isometric kCf k = kf k, and idempotent C 2 = I.

Note also hCf, Cgi = hf, gi∗ . As promised, the map f → hCf, .i is a unitary
map from H to H∗ .
Finally, we remark that projections can not only be defined for subspaces
but also for closed convex sets (of course they will no longer be linear in this
case).

Theorem 2.11 (Hilbert projection theorem). Let H be a Hilbert space and

K a nonempty closed convex subset. Then for every f ∈ H \ K there is
a unique PK (f ) such that kPK (f ) − f k = inf g∈K kf − gk. If we extend
PK : H → K by setting PK (g) = g for g ∈ K then PK will be Lipschitz
continuous: kPK (f ) − PK (g)k ≤ kf − gk, f, g ∈ H.

Proof. Fix f ∈ H \ K and choose a sequence fn ∈ K with kfn − f k → d :=

inf g∈K kf − gk. Then applying the parallelogram law to the vectors fn − f
50 2. Hilbert spaces

and fm − f we obtain
kfn − fm k2 = 2(kf − fn k2 + kf − fm k2 ) − 4kf − 21 (fn + fm )k2
≤ 2(kf − fn k2 + kf − fm k2 ) − 4d2 ,
which shows that fn is Cauchy and hence converges to some point which
we call P (f ). By construction kP (f ) − f k = d. If there would be another
point P̃ (f ) with the same property, we could apply the parallelogram law
to P (f ) − f and P̃ (f ) − f giving kP (f ) − P̃ (f )k2 ≤ 0 and hence P (f ) is
uniquely defined.
Next, let f ∈ H, g ∈ K and consider g̃ = (1 − t)P (f ) + t g ∈ K, t ∈ [0, 1].
Then
0 ≥ kf − P (f )k2 − kf − g̃k2 = 2tRe(hf − P (f ), g − P (f )i) − t2 kg − P (f )k2
for arbitrary t ∈ [0, 1] shows Re(hf − P (f ), P (f ) − gi) ≥ 0. Consequently
we have Re(hf − P (f ), P (f ) − P (g)i) ≥ 0 for all f, g ∈ H. Now reverse
to roles of f, g and add the two inequalities to obtain kP (f ) − P (g)k2 ≤
Rehf − g, P (f ) − P (g)i ≤ kf − gkkP (f ) − P (g)k. Hence Lipschitz continuity
follows. 

If K is a closed subspace then this projection will of course coincide with

the orthogonal projection defined before. By inspection of the proof, note
that PK (f ) is alternatively characterized by Re(hf − PK (f ), g − PK (f )i) ≤ 0
for all g ∈ K.
Problem 2.4. Suppose U : H → H is unitary and M ⊆ H. Show that
U M ⊥ = (U M )⊥ .
Problem 2.5. Show that an orthogonal projection PM 6= 0 has norm one.
Problem* 2.6. Suppose P ∈ L (H) satisfies
P2 = P and hP f, gi = hf, P gi
and set M = Ran(P ). Show
• P f = f for f ∈ M and M is closed,
• g ∈ M ⊥ implies P g ∈ M ⊥ and thus P g = 0,
and conclude P = PM . In particular
H = Ker(P ) ⊕ Ran(P ), Ker(P ) = (I − P )H, Ran(P ) = P H.

2.3. Operators defined via forms

One of the key results about linear maps is that they are uniquely deter-
mined once we know the images of some basis vectors. In fact, the matrix
elements with respect to some basis uniquely determine a linear map. Clearly
2.3. Operators defined via forms 51

this raises the question how this results extends to the infinite dimensional
setting. As a first result we show that the Riesz lemma, Theorem 2.10, im-
plies that a bounded operator A is uniquely determined by its associated
sesquilinear form hg, Af i. In fact, there is a one-to-one correspondence be-
tween bounded operators and bounded sesquilinear forms:
Lemma 2.12. Suppose s : H2 × H1 → C is a bounded sesquilinear form; that
is,
|s(g, f )| ≤ CkgkH2 kf kH1 . (2.24)
Then there is a unique bounded operator A ∈ L (H1 , H2 ) such that
s(g, f ) = hg, Af iH2 . (2.25)
Moreover, the norm of A is given by
kAk = sup |hg, Af iH2 | ≤ C. (2.26)
kgkH2 =kf kH1 =1

Proof. For every f ∈ H1 we have an associated bounded linear functional

`f (g) := s(g, f )∗ on H2 . By Theorem 2.10 there is a corresponding h ∈ H2
(depending on f ) such that `f (g) = hh, giH2 , that is s(g, f ) = hg, hiH2 and
we can define A via Af := h. It is not hard to check that A is linear and
from
kAf k2H2 = hAf, Af iH2 = s(Af, f ) ≤ CkAf kH2 kf kH1
we infer kAf kH2 ≤ Ckf kH1 , which shows that A is bounded with kAk ≤ C.
Equation (2.26) is left as an exercise (Problem 2.9). 

Note that if {uk }k∈K ⊆ H1 and {vj }j∈J ⊆ H2 are some orthogonal bases,
then the matrix elements Aj,k := hvj , Auk iH2 for all (j, k) ∈ J × K uniquely
determine hg, Af iH2 for arbitrary f ∈ H1 , g ∈ H2 (just expand f, g with
respect to these bases) and thus A by our theorem.
Example 2.5. Consider `2 (N) and let A ∈ L (`2 (N)) be some bounded
operator. Let Ajk = hδ j , Aδ k i be its matrix elements such that
∞
X
(Aa)j = Ajk ak .
k=1

Here the sum converges in `2 (N) and hence, in particular, for every fixed
j. Moreover, choosing ank = αn Ajk for k ≤ n and ank = 0 for k > n with
αn = ( nj=1 |Ajk |2 )1/2 we see αn = |(Aan )j | ≤ kAkkan k = kAk. Thus
P
P∞
j=1 |Ajk | ≤ kAk and the sum is even absolutely convergent.
2 2 
Moreover, for A ∈ L (H) the polarization identity (Problem 1.20) implies
that A is already uniquely determined by its quadratic form qA (f ) := hf, Af i.
As a first application we introduce the adjoint operator via Lemma 2.12
as the operator associated with the sesquilinear form s(f, g) := hAf, giH2 .
52 2. Hilbert spaces

Theorem 2.13. For every bounded operator A ∈ L (H1 , H2 ) there is a

unique bounded operator A∗ ∈ L (H2 , H1 ) defined via
hf, A∗ giH1 = hAf, giH2 . (2.27)

A bounded operator A ∈ L (H) satisfying A∗ = A is called self-adjoint.

Note that qA∗ (f ) = hAf, f i = qA (f )∗ and hence a bounded operator is
self-adjoint if and only if its quadratic form is real-valued.
Example 2.6. If H := Cn and A := (ajk )1≤j,k≤n , then A∗ = (a∗kj )1≤j,k≤n .

Example 2.7. If I ∈ L (H) is the identity, then I∗ = I. 
Example 2.8. Consider the linear functional ` : H → C, f → 7 hg, f i. Then
by the definition hf, ` αi = `(f ) α = hf, αgi we obtain ` : C → H, α 7→
∗ ∗ ∗

αg. 
Example 2.9. Let H := `2 (N), a ∈ `∞ (N) and consider the multiplication
operator
(Ab)j := aj bj .
Then
∞
X ∞
X
∗
hAb, ci = (aj bj ) cj = b∗j (a∗j cj ) = hb, A∗ ci
j=1 j=1
with (A∗ c)j = a∗j cj , that is, A∗ is the multiplication operator with a∗ . 
Example 2.10. Let H := `2 (N) and consider the shift operators defined via
(S ± a)j := aj±1
with the convention that a0 = 0. That is, S − shifts a sequence to the right
and fills up the left most place by zero and S + shifts a sequence to the left
dropping the left most place:
S − (a1 , a2 , a3 , · · · ) = (0, a1 , a2 , · · · ), S + (a1 , a2 , a3 , · · · ) = (a2 , a3 , a4 , · · · ).
Then
∞
X ∞
X
hS − a, bi = a∗j−1 bj = a∗j bj+1 = ha, S + bi,
j=2 j=1
which shows that (S − )∗
= S+.
Using symmetry of the scalar product we
also get hb, S − ai = hS + b, ai, that is, (S + )∗ = S − .
Note that S + is a left inverse of S − , S + S − = I, but not a right inverse
as S − S + 6= I. This is different from the finite dimensional case, where a left
inverse is also a right inverse and vice versa. 
Example 2.11. Suppose U ∈ L (H1 , H2 ) is unitary. Then U ∗ = U −1 . This
follows from Lemma 2.12 since hf, giH1 = hU f, U giH2 = hf, U ∗ U giH1 implies
U ∗ U = IH1 . Since U is bijective we can multiply this last equation from the
right with U −1 to obtain the claim. Of course this calculation shows that
2.3. Operators defined via forms 53

the converse is also true, that is U ∈ L (H1 , H2 ) is unitary if and only if

U ∗ = U −1 . 
A few simple properties of taking adjoints are listed below.
Lemma 2.14. Let A, B ∈ L (H1 , H2 ), C ∈ L (H2 , H3 ), and α ∈ C. Then
(i) (A + B)∗ = A∗ + B ∗ , (αA)∗ = α∗ A∗ ,
(ii) A∗∗ = A,
(iii) (CA)∗ = A∗ C ∗ ,
(iv) kA∗ k = kAk and kAk2 = kA∗ Ak = kAA∗ k.

Proof. (i) is obvious. (ii) follows from hg, A∗∗ f iH2 = hA∗ g, f iH1 = hg, Af iH2 .
(iii) follows from hg, (CA)f iH3 = hC ∗ g, Af iH2 = hA∗ C ∗ g, f iH1 . (iv) follows
using (2.26) from
kA∗ k = sup |hf, A∗ giH1 | = sup |hAf, giH2 |
kf kH1 =kgkH2 =1 kf kH1 =kgkH2 =1

= sup |hg, Af iH2 | = kAk

kf kH1 =kgkH2 =1

and
kA∗ Ak = sup |hf, A∗ AgiH1 | = sup |hAf, AgiH2 |
kf kH1 =kgkH2 =1 kf kH1 =kgkH2 =1

= sup kAf k2 = kAk2 ,

kf kH1 =1

where we have used that |hAf, AgiH2 | attains its maximum when Af and Ag
are parallel (compare Theorem 1.5). 

Note that kAk = kA∗ k implies that taking adjoints is a continuous op-
eration. For later use also note that (Problem 2.11)
Ker(A∗ ) = Ran(A)⊥ . (2.28)

For the remainder of this section we restrict to the case of one Hilbert
space. A sesquilinear form s : H × H → C is called nonnegative if s(f, f ) ≥ 0
and we will call A ∈ L (H) nonnegative, A ≥ 0, if its associated sesquilinear
form is. We will write A ≥ B if A − B ≥ 0. Observe that nonnegative
operators are self-adjoint (as their quadratic forms are real-valued — here it
is important that the underlying space is complex; in case of a real space a
nonnegative form is required to be symmetric).
Example 2.12. For any operator A the operators A∗ A and AA∗ are both
nonnegative. In fact hf, A∗ Af i = hAf, Af i = kAf k2 ≥ 0 and similarly
hf, AA∗ f i = kA∗ f k2 ≥ 0. 
54 2. Hilbert spaces

Lemma 2.15. Suppose A ∈ L (H) satisfies |hf, Af i| ≥ εkf k2 for some

ε > 0. Then A is a bijection with bounded inverse, kA−1 k ≤ 1ε .

Proof. By assumption εkf k2 ≤ |hf, Af i| ≤ kf kkAf k and thus εkf k ≤

kAf k. In particular, Af = 0 implies f = 0 and thus for every g ∈ Ran(A)
there is a unique f = A−1 g. Moreover, by kA−1 gk = kf k ≤ ε−1 kAf k =
ε−1 kgk the operator A−1 is bounded. So if gn ∈ Ran(A) converges to some
g ∈ H, then fn = A−1 gn converges to some f . Taking limits in gn = Afn
shows that g = Af is in the range of A, that is, the range of A is closed. To
show that Ran(A) = H we pick h ∈ Ran(A)⊥ . Then 0 = hh, Ahi ≥ εkhk2
shows h = 0 and thus Ran(A)⊥ = {0}. 

Combining the last two results we obtain the famous Lax–Milgram the-
orem which plays an important role in theory of elliptic partial differential
equations.
Theorem 2.16 (Lax–Milgram). Let s : H × H → C be a sesquilinear form
which is
• bounded, |s(f, g)| ≤ Ckf k kgk, and
• coercive, |s(f, f )| ≥ εkf k2 for some ε > 0.
Then for every g ∈ H there is a unique f ∈ H such that
s(h, f ) = hh, gi, ∀h ∈ H. (2.29)
Moreover, kf k ≤ 1ε kgk.

Proof. Let A be the operator associated with s. Then A is a bijection and

f = A−1 g. 

Note that (2.29) can also be phrased as a minimizing problem if s is

nonnegative — Problem 2.13.
Example 2.13. Consider H = `2 (N) and introduce the operator
(Aa)j := −aj+1 + 2aj − aj−1
which is a discrete version of a second derivative (discrete one-dimensional
Laplace operator). Here we use the convention a0 = 0, that is, (Aa)1 =
−a2 + 2a1 . In terms of the shift operators S ± we can write
A = −S + + 2 − S − = (S + − 1)(S − − 1)
and using (S ± )∗ = S ∓ we obtain
∞
X
sA (a, b) = h(S − − 1)a, (S − − 1)bi = (aj−1 − aj )∗ (bj−1 − bj ).
j=1
2.3. Operators defined via forms 55

In particular, this shows A ≥ 0. Moreover, we have |sA (a, b)| ≤ 4kak2 kbk2
or equivalently kAk ≤ 4.
Next, let
(Qa)j = qj aj
for some sequence q ∈ `∞ (N). Then
∞
X
sQ (a, b) = qj a∗j bj
j=1

and |sQ (a, b)| ≤ kqk∞ kak2 kbk2 or equivalently kQk ≤ kqk∞ . If in addition
qj ≥ ε > 0, then sA+Q (a, b) = sA (a, b) + sQ (a, b) satisfies the assumptions of
the Lax–Milgram theorem and
(A + Q)a = b
has a unique solution a = (A + Q)−1 b for every given b ∈ `2 (Z). Moreover,
since (A + Q)−1 is bounded, this solution depends continuously on b. 
Problem* 2.7. Let H1 , H2 be Hilbert spaces and let u ∈ H1 , v ∈ H2 . Show
that the operator
Af := hu, f iv
is bounded and compute its norm. Compute the adjoint of A.
Problem 2.8. Show that under the assumptions of Problem 1.36 one has
f (A)∗ = f # (A∗ ) where f # (z) = f (z ∗ )∗ .
Problem* 2.9. Prove (2.26). (Hint: Use kf k = supkgk=1 |hg, f i| — com-
pare Theorem 1.5.)
Problem 2.10. Suppose A ∈ L (H1 , H2 ) has a bounded inverse A−1 ∈
L (H2 , H1 ). Show (A−1 )∗ = (A∗ )−1 .
Problem* 2.11. Show (2.28).
Problem* 2.12. Show that every operator A ∈ L (H) can be written as the
linear combination of two self-adjoint operators Re(A) := 12 (A + A∗ ) and
Im(A) := 2i1 (A − A∗ ). Moreover, every self-adjoint operator can be written
as a linear combination√ of two unitary operators. (Hint: For the last part
consider f± (z) = z ± i 1 − z 2 and Problems 1.36, 2.8.)
Problem 2.13 (Abstract Dirichlet problem). Show that the solution of
(2.29) is also the unique minimizer of
1 
h 7→ Re s(h, h) − hh, gi .
2
if s is nonnegative with s(w, w) ≥ εkwk2 for all w ∈ H.
56 2. Hilbert spaces

2.4. Orthogonal sums and tensor products

Given two Hilbert spaces H1 and H2 , we define their orthogonal sum H1 ⊕
H2 to be the set of all pairs (f1 , f2 ) ∈ H1 × H2 together with the scalar
product
h(g1 , g2 ), (f1 , f2 )i := hg1 , f1 iH1 + hg2 , f2 iH2 . (2.30)
It is left as an exercise to verify that H1 ⊕ H2 is again a Hilbert space.
Moreover, H1 can be identified with {(f1 , 0)|f1 ∈ H1 }, and we can regard H1
as a subspace of H1 ⊕ H2 , and similarly for H2 . With this convention we have
H⊥1 = H2 . It is also customary to write f1 ⊕L f2 instead of (f1 , f2 ). In the
same way we can define the orthogonal sum nj=1 Hj of any finite number
of Hilbert spaces.
Example 2.14. For example we have nj=1 C = Cn and hence we will write
L
Ln
j=1 H =: H .
n 
More generally, let Hj , j ∈ N, be a countable collection of Hilbert spaces
and define
M∞ M∞ ∞
X
Hj := { fj | fj ∈ Hj , kfj k2Hj < ∞}, (2.31)
j=1 j=1 j=1

which becomes a Hilbert space with the scalar product

M∞ ∞
M ∞
X
h gj , fj i := hgj , fj iHj . (2.32)
j=1 j=1 j=1
L∞
Example 2.15. j=1 C = `2 (N). 

Similarly, if H and H̃ are two Hilbert spaces, we define their tensor prod-
uct as follows: The elements should be products f ⊗ f˜ of elements f ∈ H
and f˜ ∈ H̃. Hence we start with the set of all finite linear combinations of
elements of H × H̃
Xn
F(H, H̃) := { αj (fj , f˜j )|(fj , f˜j ) ∈ H × H̃, αj ∈ C}. (2.33)
j=1

Since we want (f1 + f2 ) ⊗ f˜ = f1 ⊗ f˜+ f2 ⊗ f˜, f ⊗ (f˜1 + f˜2 ) = f ⊗ f˜1 + f ⊗ f˜2 ,

and (αf ) ⊗ f˜ = f ⊗ (αf˜) = α(f ⊗ f˜) we consider F(H, H̃)/N (H, H̃), where
n
X n
X n
X
N (H, H̃) := span{ αj βk (fj , f˜k ) − ( αj fj , βk f˜k )} (2.34)
j,k=1 j=1 k=1

and write f ⊗ f˜ for the equivalence class of (f, f˜). By construction, every
element in this quotient space is a linear combination of elements of the type
f ⊗ f˜.
2.4. Orthogonal sums and tensor products 57

Next, we want to define a scalar product such that

hf ⊗ f˜, g ⊗ g̃i = hf, giH hf˜, g̃i H̃ (2.35)
holds. To this end we set
Xn n
X n
X
s( αj (fj , f˜j ), βk (gk , g̃k )) = αj∗ βk hfj , gk iH hf˜j , g̃k iH̃ , (2.36)
j=1 k=1 j,k=1

which is a symmetric sesquilinear form on F(H, H̃). Moreover, one verifies

that s(f, g) = 0 for arbitrary f ∈ F(H, H̃) and g ∈ N (H, H̃) and thus
n
X n
X n
X
h αj fj ⊗ f˜j , βk gk ⊗ g̃k i = αj∗ βk hfj , gk iH hf˜j , g̃k iH̃ (2.37)
j=1 k=1 j,k=1

is a symmetric sesquilinear form on F(H, H̃)/N (H, H̃). To show that this is in
fact a scalar product, we need to ensure positivity. Let f = i αi fi ⊗ f˜i 6= 0
P

and pick orthonormal bases uj , ũk for span{fi }, span{f˜i }, respectively. Then
X X
f= αjk uj ⊗ ũk , αjk = αi huj , fi iH hũk , f˜i iH̃ (2.38)
j,k i

and we compute X
hf, f i = |αjk |2 > 0. (2.39)
j,k

The completion of F(H, H̃)/N (H, H̃) with respect to the induced norm is
called the tensor product H ⊗ H̃ of H and H̃.
Lemma 2.17. If uj , ũk are orthonormal bases for H, H̃, respectively, then
uj ⊗ ũk is an orthonormal basis for H ⊗ H̃.

Proof. That uj ⊗ ũk is an orthonormal set is immediate from (2.35). More-

over, since span{uj }, span{ũk } are dense in H, H̃, respectively, it is easy to
see that uj ⊗ ũk is dense in F(H, H̃)/N (H, H̃). But the latter is dense in
H ⊗ H̃. 

Note that this in particular implies dim(H ⊗ H̃) = dim(H) dim(H̃).

Example 2.16. We have H ⊗ Cn = Hn . 
Example 2.17. P We have `2 (N) ⊗ `2 (N)
= by virtue of the identi-
`2 (N × N)
fication (ajk ) 7→ jk ajk δ j ⊗ δ k where δ j is the standard basis for `2 (N). In
fact, this follows from the previous lemma as in the proof of Theorem 2.6. 
It is straightforward to extend the tensor product to any finite number
of Hilbert spaces. We even note
M∞ M∞
( Hj ) ⊗ H = (Hj ⊗ H), (2.40)
j=1 j=1
58 2. Hilbert spaces

where equality has to be understood in the sense that both spaces are uni-
tarily equivalent by virtue of the identification
∞
X ∞
X
( fj ) ⊗ f = fj ⊗ f. (2.41)
j=1 j=1

Problem 2.14. Show that f ⊗ f˜ = 0 if and only if f = 0 or f˜ = 0.

Problem 2.15. We have f ⊗ f˜ = g ⊗ g̃ 6= 0 if and only if there is some
α ∈ C \ {0} such that f = αg and f˜ = α−1 g̃.
Problem* 2.16. Show (2.40).

2.5. Applications to Fourier series

We have already encountered the Fourier sine series during our treatment
of the heat equation in Section 1.1. Given an integrable function f we can
define its Fourier series
a0 X
(2.42)


S(f )(x) := + ak cos(kx) + bk sin(kx) ,
2
k∈N

where the corresponding Fourier coefficients are given by

1 π 1 π
Z Z
ak := cos(kx)f (x)dx, bk := sin(kx)f (x)dx. (2.43)
π −π π −π
At this point (2.42) is just a formal expression and it was (and to some extend
still is) a fundamental question in mathematics to understand in what sense
the above series converges. For example, does it converge at a given point
(e.g. at every point of continuity of f ) or when does it converge uniformly?
We will give some first answers in the present section and then come back
later to this when we have further tools at our disposal.
For our purpose the complex form
Z π
X 1
S(f )(x) = fˆk eikx , fˆk := e−iky f (y)dy (2.44)
2π −π
k∈Z

will be more convenient. The connection is given via fˆ±k = ak ∓ib2

k
, k ∈ N0
(with the convention b0 = 0). In this case the n’th partial sum can be written
as
n Z π
X 1
Sn (f )(x) := fˆk eikx = Dn (x − y)f (y)dy, (2.45)
2π −π
k=−n
where
n
X sin((n + 1/2)x)
Dn (x) = eikx = (2.46)
sin(x/2)
k=−n
2.5. Applications to Fourier series 59

3
D3 (x)

2
D2 (x)

1
D1 (x)

−π π

Figure 2.1. The Dirichlet kernels D1 , D2 , and D3

is known as the Dirichlet kernel (to obtain the second form observe that
the left-hand side is a geometric series). Note that Dn (−x) = Dn (x) and
that |Dn (x)| has a global
R π maximum Dn (0) = 2n + 1 at x = 0. Moreover, by
Sn (1) = 1 we see that −π Dn (x)dx = 1.
Since Z π
e−ikx eilx dx = 2πδk,l (2.47)
−π
the functions ek (x) :=(2π)−1/2 eikxare orthonormal in L2 (−π, π) and hence
the Fourier series is just the expansion with respect to this orthogonal set.
Hence we obtain
Theorem 2.18. For every square integrable function f ∈ L2 (−π, π), the
Fourier coefficients fˆk are square summable
Z π
X
ˆ 1
2
|fk | = |f (x)|2 dx (2.48)
2π −π
k∈Z

and the Fourier series converges to f in the sense of L2 . Moreover, this is a

continuous bijection between L2 (−π, π) and `2 (Z).

Proof. To show this theorem it suffices to show that the functions ek form
a basis. This will follow from Theorem 2.21 below (see the discussion after
this theorem). It will also follow as a special case of Theorem 3.11 below
(see the examples after this theorem) as well as from the Stone–Weierstraß
theorem — Problem 2.21. 

This gives a satisfactory answer in the Hilbert space L2 (−π, π) but does
not answer the question about pointwise or uniform convergence. The latter
60 2. Hilbert spaces

will be the case if the Fourier coefficients are summable. First of all we note
that for integrable functions the Fourier coefficients will at least tend to zero.
Lemma 2.19 (Riemann–Lebesgue lemma). Suppose f ∈ L1 (−π, π), then
the Fourier coefficients fˆk converge to zero as |k| → ∞.

Proof. By our previous theorem this holds for continuous functions. But the
map f → fˆ is bounded from C[−π, π] ⊂ L1 (−π, π) to c0 (Z) (the sequences
vanishing as |k| → ∞) since |fˆk | ≤ (2π)−1 kf k1 and there is a unique exten-
sion to all of L1 (−π, π). 

It turns out that this result is best possible in general and we cannot say
more without additional assumptions on f . For example, if f is periodic of
period 2π and continuously differentiable, then integration by parts shows
Z π
1
ˆ
fk = e−ikx f 0 (x)dx. (2.49)
2πik −π
Then, since both k −1 and the Fourier coefficients of f 0 are square summa-
ble, we conclude that fˆ is absolutely summable and hence the Fourier series
converges uniformly. So we have a simple sufficient criterion for summa-
bility of the Fourier coefficients, but can we do better? Of course conti-
nuity of f is a necessary condition for absolute summability but this alone
will not even be enough for pointwise convergence as we will see in Exam-
ple 4.3. Moreover, continuity will not tell us more about the decay of the
Fourier coefficients than what we already know in the integrable case from
the Riemann–Lebesgue lemma (see Example 4.4).
A few improvements are easy: (2.49) holds for any class of functions
for which integration by parts holds, e.g., piecewise continuously differen-
tiable functions or, slightly more general, absolutely continuous functions
(cf. Lemma 4.30 from [47]) provided one assumes that the derivative is
square integrable. However, for an arbitrary absolutely continuous func-
tion the Fourier coefficients might not be absolutely summable: For an
absolutely continuous function f we have a derivative which is integrable
(Theorem 4.29 from [47]) and hence the above formula combined with the
Riemann–Lebesgue lemma implies fˆk = o( k1 ). But on the other hand we
can choose an absolutely summable sequence ck which does not obey this
asymptotic requirement, say ck = k1 for k = l2 and ck = 0 else. Then
X X 1 2
f (x) := ck eikx = eil x (2.50)
l2
k∈Z l∈N

is a function with absolutely summable Fourier coefficients fˆk = ck (by

uniform convergence we can interchange summation and integration) but
which is not absolutely continuous. There are further criteria for absolute
2.5. Applications to Fourier series 61

summability of the Fourier coefficients, but no simple necessary and sufficient

one. a particularly simple sufficient one is:
0,γ
Theorem 2.20 (Bernstein). Suppose that f ∈ Cper [−π, π] is Hölder contin-
uous (cf. (1.67)) of exponent γ > 12 , then
X
|fˆk | ≤ Cγ kf k0,γ .
k∈Z

Proof. The proof starts with the observation that the Fourier coefficients of
fδ (x) := f (x−δ) are fˆk = e−ikδ fˆk . Now for δ := 2π
3 2
−m and 2m ≤ |k| < 2m−1

we have |e − 1| ≥ 3 implying
ikδ 2
Z π
X 1 X ikδ 1
|fˆk |2 ≤ |e − 1|2 |fˆk |2 = |fδ (x) − f (x)|2 dx
m m−1
3 2π −π
2 ≤|k|<2 k

≤ [f ]2γ δ 2γ
Now the sum on the left has 2(2m − 1) terms and hence Cauchy–Schwarz
implies
 γ
X √ 2π
ˆ
|fk | ≤ 2 (m+1)/2 γ
[f ]γ δ = 2 2(1/2−γ)m [f ]γ .
m m−1
3
2 ≤|k|<2

Summing over m shows

X
|fˆk | ≤ Cγ [f ]γ
k6=0

provided γ > 1
2 and establishes the claim since |fˆ0 | ≤ kf k∞ . 

Note however, that the situation looks much brighter if one looks at mean
values
n−1 Z π
1X 1
S̄n (f )(x) := Sk (f )(x) = Fn (x − y)f (y)dy, (2.51)
n 2π −π
k=0
where
n−1  2
1X 1 sin(nx/2)
Fn (x) = Dk (x) = (2.52)
n n sin(x/2)
k=0
is the Fejér kernel. To see the second form we use the closed form for the
Dirichlet kernel to obtain
n−1 n−1
X sin((k + 1/2)x) 1 X
nFn (x) = = Im ei(k+1/2)x
sin(x/2) sin(x/2)
k=0 k=0
inx sin(nx/2) 2
   
1 ix/2 e −1 1 − cos(nx)
= Im e = = .
sin(x/2) eix − 1 2 sin(x/2)2 sin(x/2)
62 2. Hilbert spaces

F3 (x)

F2 (x)
F1 (x)
1

−π π

Figure 2.2. The Fejér kernels F1 , F2 , and F3

The main differenceR to the Dirichlet kernel is positivity: Fn (x) ≥ 0. Of

π
course the property −π Fn (x)dx = 1 is inherited from the Dirichlet kernel.

Theorem 2.21 (Fejér). Suppose f is continuous and periodic with period

2π. Then S̄n (f ) → f uniformly.

Proof. Let us set Fn = 0 outside [−π, π]. Then Fn (x) ≤ n sin(δ/2)

1
2 for

δ ≤ |x| ≤ π implies that a straightforward adaption of Lemma 1.2 to the

periodic case is applicable. 

In particular, this shows that the functions {ek }k∈Z are total in Cper [−π, π]
(continuous periodic functions) and hence also in Lp (−π, π) for 1 ≤ p < ∞
(Problem 2.20).
Note that for a given continuous function f this result shows that if
Sn (f )(x) converges, then it must converge to S̄n (f )(x) = f (x). We also
remark that one can extend this result (see Lemma 3.21 from [47]) to show
that for f ∈ Lp (−π, π), 1 ≤ p < ∞, one has S̄n (f ) → f in the sense of Lp .
As a consequence note that the Fourier coefficients uniquely determine f for
integrable f (for square integrable f this follows from Theorem 2.18).
Finally, we look at pointwise convergence.

Theorem 2.22. Suppose

f (x) − f (x0 )
(2.53)
x − x0
2.5. Applications to Fourier series 63

is integrable (e.g. f is Hölder continuous), then

Xn
lim fˆk eikx0 = f (x0 ). (2.54)
m,n→∞
k=−m

Proof. Without loss of generality we can assume x0 = 0 (by shifting x →

x − x0 modulo 2π implying fˆk → e−ikx0 fˆk ) and f (x0 ) = 0 (by linearity since
the claim is trivial for constant functions). Then by assumption
f (x)
g(x) :=
eix −1
is integrable and f (x) = (eix − 1)g(x) implies fˆk = ĝk−1 − ĝk and hence
Xn
fˆk = ĝ−m−1 − ĝn .
k=−m
Now the claim follows from the Riemann–Lebesgue lemma. 

If we look at symmetric partial sums Sn (f ) we can do even better.

Corollary 2.23 (Dirichlet–Dini criterion). Suppose there is some α such
that
f (x0 + x) + f (x0 − x) − 2α
x
is integrable. Then Sn (f )(x0 ) → α.

Proof. Without loss of generality we can assume x0 = 0. Now observe

(since Dn (−x) = Dn (x)) Sn (f )(0) = α + Sn (g)(0), where g(x) := 21 (f (x) +
f (−x)) − α and apply the previous result. 
Problem 2.17. Compute the Fourier series of Dn and Fn .
π π 2
Problem 2.18. Show |Dn (x)| ≤ min(2n + 1, |x| ) and Fn (x) ≤ min(n, nx2 ).

0,γ
Problem 2.19. Show that if f ∈ Cper [−π, π] is Hölder continuous (cf.
(1.67)), then
[f ]γ π γ
 
ˆ
|fk | ≤ , k 6= 0.
2 |k|
(Hint: What changes if you replace e−iky by e−ik(y+π/k) in (2.44)? Now make
a change of variables y → y − π/k in the integral.)
Problem 2.20. Show that Cper [−π, π] is dense in Lp (−π, π) for 1 ≤ p < ∞.
Problem 2.21. Show that the functions ek (x) := √12π eikx , k ∈ Z, form an
orthonormal basis for H = L2 (−π, π). (Hint: Start with K = [−π, π] where
−π and π are identified and use the Stone–Weierstraß theorem.)
 Chapter 3

Compact operators

Typically, linear operators are much more difficult to analyze than matrices
and many new phenomena appear which are not present in the finite dimen-
sional case. So we have to be modest and slowly work our way up. A class
of operators which still preserves some of the nice properties of matrices is
the class of compact operators to be discussed in this chapter.

3.1. Compact operators

A linear operator A : X → Y defined between normed spaces X, Y is called
compact if every sequence Afn has a convergent subsequence whenever fn is
bounded. Equivalently (cf. Corollary B.20), A is compact if it maps bounded
sets to relatively compact ones. The set of all compact operators is denoted
by C (X, Y ). If X = Y we will just write C (X) := C (X, X) as usual.
Example 3.1. Every linear map between finite dimensional spaces is com-
pact by the Bolzano–Weierstraß theorem. Slightly more general, a bounded
operator is compact if its range is finite dimensional. 
The following elementary properties of compact operators are left as an
exercise (Problem 3.1):

Theorem 3.1. Let X, Y , and Z be normed spaces. Every compact linear

operator is bounded, C (X, Y ) ⊆ L (X, Y ). Linear combinations of compact
operators are compact, that is, C (X, Y ) is a subspace of L (X, Y ). Moreover,
the product of a bounded and a compact operator is again compact, that
is, A ∈ L (X, Y ), B ∈ C (Y, Z) or A ∈ C (X, Y ), B ∈ L (Y, Z) implies
BA ∈ C (X, Z).

65
66 3. Compact operators

In particular, the set of compact operators C (X) is an ideal of the set

of bounded operators. Moreover, if X is a Banach space this ideal is even
closed:
Theorem 3.2. Suppose X is a normed and Y a Banach space. Let An ∈
C (X, Y ) be a convergent sequence of compact operators. Then the limit A is
again compact.

Proof. Let fj0 be a bounded sequence. Choose a subsequence fj1 such that
A1 fj1 converges. From fj1 choose another subsequence fj2 such that A2 fj2
converges and so on. Since there might be nothing left from fjn as n → ∞, we
consider the diagonal sequence fj := fjj . By construction, fj is a subsequence
of fjn for j ≥ n and hence An fj is Cauchy for every fixed n. Now
kAfj − Afk k = k(A − An )(fj − fk ) + An (fj − fk )k
≤ kA − An kkfj − fk k + kAn fj − An fk k
shows that Afj is Cauchy since the first term can be made arbitrary small
by choosing n large and the second by the Cauchy property of An fj . 
Example 3.2. Let X := `p (N) and consider the operator
(Qa)j := qj aj
for some sequence q = (qj )∞ j=1 ∈ c0 (N) converging to zero. Let Qn be
associated with qj = qj for j ≤ n and qjn = 0 for j > n. Then the range of
n

Qn is finite dimensional and hence Qn is compact. Moreover, by kQn −Qk =

supj>n |qj | we see Qn → Q and thus Q is also compact by the previous
theorem. 
Example 3.3. Let X := C 1 [0, 1], Y := C[0, 1] (cf. Problem 1.32) then the
embedding X ,→ Y is compact. Indeed, a bounded sequence in X has
both the functions and the derivatives uniformly bounded. Hence by the
mean value theorem the functions are equicontinuous and hence there is
a uniformly convergent subsequence by the Arzelà–Ascoli theorem (Theo-
rem 1.13). Of course the same conclusion holds if we take X := C 0,γ [0, 1] to
be Hölder continuous functions (cf. Theorem 1.21). 
If A : X → Y is a bounded operator there is a unique extension A :
X → Y to the completion by Theorem 1.16. Moreover, if A ∈ C (X, Y ),
then A ∈ C (X, Y ) is immediate. That we also have A ∈ C (X, Y ) will follow
from the next lemma. In particular, it suffices to verify compactness on a
dense set.
Lemma 3.3. Let X, Y be normed spaces and A ∈ C (X, Y ). Let X, Y be
the completion of X, Y , respectively. Then A ∈ C (X, Y ), where A is the
unique extension of A (cf. Theorem 1.16).
3.1. Compact operators 67

Proof. Let fn ∈ X be a given bounded sequence. We need to show that

Afn has a convergent subsequence. Pick fnj ∈ X such that kfnj −fn k ≤ 1j and
by compactness of A we can assume that Afnn → g. But then kAfn − gk ≤
kAkkfn − fnn k + kAfnn − gk shows that Afn → g. 

One of the most important examples of compact operators are integral

operators. The proof will be based on the Arzelà–Ascoli theorem (Theo-
rem 1.13).
Lemma 3.4. Let X := C([a, b]) or X := L2cont (a, b). The integral operator
K : X → X defined by
Z b
(Kf )(x) := K(x, y)f (y)dy, (3.1)
a
where K(x, y) ∈ C([a, b] × [a, b]), is compact.

Proof. First of all note that K(., ..) is continuous on [a, b] × [a, b] and hence
uniformly continuous. In particular, for every ε > 0 we can find a δ > 0 such
that |K(y, t) − K(x, t)| ≤ ε for any t ∈ [a, b] whenever |y − x| ≤ δ. Moreover,
kKk∞ = supx,y∈[a,b] |K(x, y)| < ∞.
We begin with the case X := L2cont (a, b). Let g := Kf . Then
Z b Z b
|g(x)| ≤ |K(x, t)| |f (t)|dt ≤ kKk∞ |f (t)|dt ≤ kKk∞ k1k kf k,
a a
where
√ we have used Cauchy–Schwarz in the last step (note that k1k =
b − a). Similarly,
Z b
|g(x) − g(y)| ≤ |K(y, t) − K(x, t)| |f (t)|dt
a
Z b
≤ε |f (t)|dt ≤ εk1k kf k,
a

whenever |y − x| ≤ δ. Hence, if fn (x) is a bounded sequence in L2cont (a, b),

then gn := Kfn is bounded and equicontinuous and hence has a uniformly
convergent subsequence by the Arzelà–Ascoli theorem (Theorem 1.13). But
a uniformly convergent sequence is also convergent in the norm induced by
the scalar product. Therefore K is compact.
The case X := C([a, b]) follows by the same argument upon observing
Rb
a |f (t)|dt ≤ (b − a)kf k∞ . 

Compact operators share many similarities with (finite) matrices as we

will see in the next section.
Problem* 3.1. Show Theorem 3.1.
68 3. Compact operators

Problem* 3.2. Show that the adjoint of the integral operator K on L2cont (a, b)
from Lemma 3.4 is the integral operator with kernel K(y, x)∗ :
Z b
∗
(K f )(x) = K(y, x)∗ f (y)dy.
a
(Hint: Fubini.)
d
Problem 3.3. Show that the operator dx : C 2 [a, b] → C[a, b] is compact.
(Hint: Arzelà–Ascoli.)

3.2. The spectral theorem for compact symmetric operators

Let H be an inner product space. A linear operator A : D(A) ⊆ H → H is
called symmetric if its domain is dense and if
hg, Af i = hAg, f i f, g ∈ D(A). (3.2)
If A is bounded (with D(A) = H), then A is symmetric precisely if A = A∗ ,
that is, if A is self-adjoint. However, for unbounded operators there is a
subtle but important difference between symmetry and self-adjointness.
A number z ∈ C is called eigenvalue of A if there is a nonzero vector
u ∈ D(A) such that
Au = zu. (3.3)
The vector u is called a corresponding eigenvector in this case. The set of
all eigenvectors corresponding to z is called the eigenspace
Ker(A − z) (3.4)
corresponding to z. Here we have used the shorthand notation A − z for
A − zI. An eigenvalue is called (geometrically) simple if there is only one
linearly independent eigenvector.
Example 3.4. Let H := `2 (N) and consider the shift operators (S ± a)j :=
aj±1 (with a0 := 0). Suppose z ∈ C is an eigenvalue, then the corresponding
eigenvector u must satisfy uj±1 = zuj . For S − the special case j = 1 gives
0 = u0 = zu1 . So either z = 0 and u = 0 or z 6= 0 and again u = 0. Hence
there are no eigenvalues. For S + we get uj = z j u1 and this will give an
element in `2 (N) if and only if |z| < 1. Hence z with |z| < 1 is an eigenvalue.
All these eigenvalues are simple. 
Example 3.5. Let H := `2 (N) and consider the multiplication operator
(Qa)j := qj aj with a bounded sequence q ∈ `∞ (N). Suppose z ∈ C is an
eigenvalue, then the corresponding eigenvector u must satisfy (qj − z)uj = 0.
Hence every value qj is an eigenvalue with corresponding eigenvector u = δ j .
If there is only one j with z = qj the eigenvalue is simple (otherwise the
numbers of independent eigenvectors equals the number of times z appears
3.2. The spectral theorem for compact symmetric operators 69

in the sequence q). If z is different from all entries of the sequence then u = 0
and z is no eigenvalue. 
Note that in the last example Q will be self-adjoint if and only if q is real-
valued and hence if and only if all eigenvalues are real-valued. Moreover, the
corresponding eigenfunctions are orthogonal. This has nothing to do with
the simple structure of our operator and is in fact always true.
Theorem 3.5. Let A be symmetric. Then all eigenvalues are real and eigen-
vectors corresponding to different eigenvalues are orthogonal.

Proof. Suppose λ is an eigenvalue with corresponding normalized eigen-

vector u. Then λ = hu, Aui = hAu, ui = λ∗ , which shows that λ is real.
Furthermore, if Auj = λj uj , j = 1, 2, we have
(λ1 − λ2 )hu1 , u2 i = hAu1 , u2 i − hu1 , Au2 i = 0
finishing the proof. 

Note that while eigenvectors corresponding to the same eigenvalue λ will

in general not automatically be orthogonal, we can of course replace each
set of eigenvectors corresponding to λ by an set of orthonormal eigenvectors
having the same linear span (e.g. using Gram–Schmidt orthogonalization).
Example 3.6. Let H := `2 (N) and consider the Jacobi operator J := 21 (S + +
S − ):
1
(Jc)j := (cj+1 + cj−1 )
2
with the convention c0 = 0. Recall that J ∗ = J. If we look for an eigenvalue
Ju = zu, we need to solve the corresponding recursion uj+1 = 2zuj − uj−1
starting from u0 = 0 (our convention) and u1 = 1 (normalization). Like
an ordinary differential equation, a linear recursion relation with constant
coefficients can be solved by an exponential ansatz uj = k j which leads to the
characteristic polynomial k 2 = 2zk − 1. This gives two linearly independent
solutions and our requirements lead us to
k j − k −j p
uj (z) = −1
, k = z − z 2 − 1.
k−k
√
Note that k = z+ z 2 − 1 and in the case k = z = ±1 the above expression
−1

has to be understood as its limit uj (±1) = (±1)j+1 j. In fact, Uj (z) :=

uj+1 (z) are polynomials of degree j known as Chebyshev polynomials of
the second kind.
Now for z ∈ R \ [−1, 1] we have |k| < 1 and uj explodes exponentially.
For z ∈ [−1, 1] we have |k| = 1 and hence we can write k = eiκ with κ ∈ R.
Thus uj = sin(κj)
sin(κ) is oscillating. So for no value of z ∈ R our potential
eigenvector u is square summable and thus J has no eigenvalues. 
70 3. Compact operators

The previous example shows that in the infinite dimensional case sym-
metry is not enough to guarantee existence of even a single eigenvalue. In
order to always get this, we will need an extra condition. In fact, we will
see that compactness provides a suitable extra condition to obtain an or-
thonormal basis of eigenfunctions. The crucial step is to prove existence of
one eigenvalue, the rest then follows as in the finite dimensional case.
Theorem 3.6. Let H be an inner product space. A symmetric compact
operator A has an eigenvalue α1 which satisfies |α1 | = kAk.

Proof. We set α := kAk and assume α 6= 0 (i.e., A 6= 0) without loss of

generality. Since
kAk2 = sup kAf k2 = sup hAf, Af i = sup hf, A2 f i
f :kf k=1 f :kf k=1 f :kf k=1

there exists a normalized sequence un such that

lim hun , A2 un i = α2 .
n→∞

Since A is compact, it is no restriction to assume that A2 un converges, say

limn→∞ A2 un = α2 u. Now
k(A2 − α2 )un k2 = kA2 un k2 − 2α2 hun , A2 un i + α4
≤ 2α2 (α2 − hun , A2 un i)
(where we have used kA2 un k ≤ kAkkAun k ≤ kAk2 kun k = α2 ) implies
limn→∞ (A2 un − α2 un ) = 0 and hence limn→∞ un = u. In addition, u is
a normalized eigenvector of A2 since (A2 − α2 )u = 0. Factorizing this last
equation according to (A − α)u = v and (A + α)v = 0 shows that either
v 6= 0 is an eigenvector corresponding to −α or v = 0 and hence u 6= 0 is an
eigenvector corresponding to α. 

Note that for a bounded operator A, there cannot be an eigenvalue with

absolute value larger than kAk, that is, the set of eigenvalues is bounded by
kAk (Problem 3.4).
Now consider a symmetric compact operator A with eigenvalue α1 (as
above) and corresponding normalized eigenvector u1 . Setting
H1 := {u1 }⊥ = {f ∈ H|hu1 , f i = 0} (3.5)
we can restrict A to H1 since f ∈ H1 implies
hu1 , Af i = hAu1 , f i = α1 hu1 , f i = 0 (3.6)
and hence Af ∈ H1 . Denoting this restriction by A1 , it is not hard to see
that A1 is again a symmetric compact operator. Hence we can apply Theo-
rem 3.6 iteratively to obtain a sequence of eigenvalues αj with corresponding
normalized eigenvectors uj . Moreover, by construction, uj is orthogonal to
3.2. The spectral theorem for compact symmetric operators 71

all uk with k < j and hence the eigenvectors {uj } form an orthonormal set.
By construction we also have |αj | = kAj k ≤ kAj−1 k = |αj−1 |. This proce-
dure will not stop unless H is finite dimensional. However, note that αj = 0
for j ≥ n might happen if An = 0.
Theorem 3.7 (Hilbert–Schmidt; Spectral theorem for compact symmetric
operators). Suppose H is an infinite dimensional Hilbert space and A : H →
H is a compact symmetric operator. Then there exists a sequence of real
eigenvalues αj converging to 0. The corresponding normalized eigenvectors
uj form an orthonormal set and every f ∈ H can be written as
∞
X
f= huj , f iuj + h, (3.7)
j=1

where h is in the kernel of A, that is, Ah = 0.

In particular, if 0 is not an eigenvalue, then the eigenvectors form an
orthonormal basis (in addition, H need not be complete in this case).

Proof. Existence of the eigenvalues αj and the corresponding eigenvectors

uj has already been established. Since the sequence |αj | is decreasing it has a
limit ε ≥ 0 and we have |αj | ≥ ε. If this limit is nonzero, then vj = αj−1 uj is
a bounded sequence (kvj k ≤ 1ε ) for which Avj has no convergent subsequence
since kAvj − Avk k2 = kuj − uk k2 = 2, a contradiction.
Next, setting
n
X
fn := huj , f iuj ,
j=1

we have
kA(f − fn )k ≤ |αn+1 |kf − fn k ≤ |αn+1 |kf k
since f − fn ∈ Hn and kAn k = |αn+1 |. Letting n → ∞ shows A(f∞ − f ) = 0
proving (3.7). Finally, note that without completeness f∞ might not be
well-defined unless h = 0. 

By applying A to (3.7) we obtain the following canonical form of compact

symmetric operators.
Corollary 3.8. Every compact symmetric operator A can be written as
N
X
Af = αj huj , f iuj , (3.8)
j=1

where αj are the nonzero eigenvalues with corresponding eigenvectors uj from

the previous theorem.
72 3. Compact operators

Remark: There are two cases where our procedure might fail to construct
an orthonormal basis of eigenvectors. One case is where there is an infinite
number of nonzero eigenvalues. In this case αn never reaches 0 and all eigen-
vectors corresponding to 0 are missed. In the other case, 0 is reached, but
there might not be a countable basis and hence again some of the eigen-
vectors corresponding to 0 are missed. In any case, by adding vectors from
the kernel (which are automatically eigenvectors), one can always extend the
eigenvectors uj to an orthonormal basis of eigenvectors.
Corollary 3.9. Every compact symmetric operator A has an associated or-
thonormal basis of eigenvectors {uj }j∈J . The corresponding unitary map
U : H → `2 (J), f 7→ {huj , f i}j∈J diagonalizes A in the sense that U AU −1 is
the operator which multiplies each basis vector δ j = U uj by the corresponding
eigenvalue αj .
Example 3.7. Let a, b ∈ c0 (N) be real-valued sequences and consider the
operator
(Jc)j := aj cj+1 + bj cj + aj−1 cj−1 .
If A, B denote the multiplication operators by the sequences a, b, respec-
tively, then we already know that A and B are compact. Moreover, using
the shift operators S ± we can write
J = AS + + B + S − A,
which shows that J is self-adjoint since A∗ = A, B ∗ = B, and (S ± )∗ =
S ∓ . Hence we can conclude that J has a countable number of eigenvalues
converging to zero and a corresponding orthonormal basis of eigenvectors. 
In particular, in the new picture it is easy to define functions of our
operator (thus extending the functional calculus from Problem 1.36). To this
end set Σ := {αj }j∈J and denote by B(K) the Banach algebra of bounded
functions F : K → C together with the sup norm.
Corollary 3.10 (Functional calculus). Let A be a compact symmetric op-
erator with associated orthonormal basis of eigenvectors {uj }j∈J and corre-
sponding eigenvalues {αj }j∈J . Suppose F ∈ B(Σ), then
X
F (A)f = F (αj )huj , f iuj (3.9)
j∈J

defines a continuous algebra homomorphism from the Banach algebra B(Σ)

to the algebra L (H) with 1(A) = I and I(A) = A. Moreover F (A)∗ = F ∗ (A),
where F ∗ is the function which takes complex conjugate values.

Proof. This is straightforward to check for multiplication operators in `2 (J)

and hence the result follows by the previous corollary. 
3.2. The spectral theorem for compact symmetric operators 73

In many applications F will be given by a function on R (or at least on

[−kAk, kAk]) and, since only the values F (αj ) are used, two functions which
agree on all eigenvalues will give the same result.
As a brief application we will say a few words about general spectral
theory for bounded operators A ∈ L (X) in a Banach space X. In the finite
dimensional case, the spectrum is precisely the set of eigenvalues. In the
infinite dimensional case one defines the spectrum as

σ(A) := C \ {z ∈ C|∃(A − z)−1 ∈ L (X)}. (3.10)

It is important to emphasize that the inverse is required to exist as a bounded

operator. Hence there are several ways in which this can fail: First of all,
A − z could not be injective. In this case z is an eigenvalue and thus all
eigenvalues belong to the spectrum. Secondly, it could not be surjective.
And finally, even if it is bijective it could be unbounded. However, it will
follow form the open mapping theorem that this last case cannot happen
for a bounded operator. The inverse of A − z for z ∈ C \ σ(A) is known
as the resolvent of A and plays a crucial role in spectral theory. Using
Problem 1.35 one can show that the complement of the spectrum is open,
and hence the spectrum is closed. Since we will discuss this in detail in
Chapter 6, we will not pursue this here but only look at our special case of
symmetric compact operators.
To compute the inverse of A − z we will use the functional calculus and
consider F (α) = α−z1
. Of course this function is unbounded on R but if z
is neither an eigenvalue nor zero it is bounded on Σ and hence satisfies our
requirements. Then

X 1
RA (z)f := huj , f iuj (3.11)
αj − z
j∈J

satisfies (A − z)RA (z) = RA (z)(A − z) = I, that is, RA (z) = (A − z)−1 ∈

L (H). Of course, if z is an eigenvalue, then the above formula breaks down.
However, in the infinite dimensional case it also breaks down if z = 0 even
if 0 is not an eigenvalue! In this case the above definition will still give an
operator which is the inverse of A − z, however, since the sequence αj−1 is
unbounded, so will be the corresponding multiplication operator in `2 (J) and
the sum in (3.11) will only converge if {αj−1 huj , f i}j∈J ∈ `2 (J). So in the
infinite dimensional case 0 is in the spectrum even if it is not an eigenvalue.
In particular,

σ(A) = {αj }j∈J . (3.12)

74 3. Compact operators

αj
Moreover, if we use 1
αj −z = z(αj −z) − 1
z we can rewrite this as
 
N
1 X αj
RA (z)f =  huj , f iuj − f 
z αj − z
j=1

where it suffices to take the sum over all nonzero eigenvalues.

This is all we need and it remains to apply these results to Sturm–
Liouville operators.
Problem 3.4. Show that if A ∈ L (H), then every eigenvalue α satisfies
|α| ≤ kAk.
Problem 3.5. Find the eigenvalues and eigenfunctions of the integral oper-
ator K ∈ L (L2cont (0, 1)) given by
Z 1
(Kf )(x) := u(x)v(y)f (y)dy,
0
where u, v ∈ C([0, 1]) are some given continuous functions.
Problem 3.6. Find the eigenvalues and eigenfunctions of the integral oper-
ator K ∈ L (L2cont (0, 1)) given by
Z 1
(Kf )(x) := 2 (2xy − x − y + 1)f (y)dy.
0

3.3. Applications to Sturm–Liouville operators

Now, after all this hard work, we can show that our Sturm–Liouville operator
d2
L := −+ q(x), (3.13)
dx2
where q is continuous and real, defined on
D(L) := {f ∈ C 2 [0, 1]|f (0) = f (1) = 0} ⊂ L2cont (0, 1), (3.14)
has an orthonormal basis of eigenfunctions.
The corresponding eigenvalue equation Lu = zu explicitly reads
− u00 (x) + q(x)u(x) = zu(x). (3.15)
It is a second order homogeneous linear ordinary differential equation and
hence has two linearly independent solutions. In particular, specifying two
initial conditions, e.g. u(0) = 0, u0 (0) = 1 determines the solution uniquely.
Hence, if we require u(0) = 0, the solution is determined up to a multiple
and consequently the additional requirement u(1) = 0 cannot be satisfied by
a nontrivial solution in general. However, there might be some z ∈ C for
which the solution corresponding to the initial conditions u(0) = 0, u0 (0) = 1
3.3. Applications to Sturm–Liouville operators 75

happens to satisfy u(1) = 0 and these are precisely the eigenvalues we are
looking for.
Note that the fact that L2cont (0, 1) is not complete causes no problems
since we can always replace it by its completion H = L2 (0, 1). A thorough
investigation of this completion will be given later, at this point this is not
essential.
We first verify that L is symmetric:
Z 1
hf, Lgi = f (x)∗ (−g 00 (x) + q(x)g(x))dx
0
Z 1 Z 1
0 ∗ 0
= f (x) g (x)dx + f (x)∗ q(x)g(x)dx
0 0
Z 1 Z 1
00 ∗
= −f (x) g(x)dx + f (x)∗ q(x)g(x)dx (3.16)
0 0
= hLf, gi.
Here we have used integration by parts twice (the boundary terms vanish
due to our boundary conditions f (0) = f (1) = 0 and g(0) = g(1) = 0).
Of course we want to apply Theorem 3.7 and for this we would need to
show that L is compact. But this task is bound to fail, since L is not even
bounded (see Example 1.17)!
So here comes the trick: If L is unbounded its inverse L−1 might still
be bounded. Moreover, L−1 might even be compact and this is the case
here! Since L might not be injective (0 might be an eigenvalue), we consider
RL (z) := (L − z)−1 , z ∈ C, which is also known as the resolvent of L.
In order to compute the resolvent, we need to solve the inhomogeneous
equation (L − z)f = g. This can be done using the variation of constants
formula from ordinary differential equations which determines the solution
up to an arbitrary solution of the homogeneous equation. This homogeneous
equation has to be chosen such that f ∈ D(L), that is, such that f (0) =
f (1) = 0.
Define
Z x
u+ (z, x)  
f (x) := u− (z, t)g(t)dt
W (z) 0
Z 1
u− (z, x)  
+ u+ (z, t)g(t)dt , (3.17)
W (z) x
where u± (z, x) are the solutions of the homogeneous differential equation
−u00± (z, x)+(q(x)−z)u± (z, x) = 0 satisfying the initial conditions u− (z, 0) =
0, u0− (z, 0) = 1 respectively u+ (z, 1) = 0, u0+ (z, 1) = 1 and
W (z) := W (u+ (z), u− (z)) = u0− (z, x)u+ (z, x) − u− (z, x)u0+ (z, x) (3.18)
76 3. Compact operators

is the Wronski determinant, which is independent of x (check this!).

Then clearly f (0) = 0 since u− (z, 0) = 0 and similarly f (1) = 0 since
u+ (z, 1) = 0. Furthermore, f is differentiable and a straightforward compu-
tation verifies
u0 (z, x)  x
Z 
f 0 (x) = + u− (z, t)g(t)dt
W (z) 0
u0− (z, x)  1
Z 
+ u+ (z, t)g(t)dt . (3.19)
W (z) x

Thus we can differentiate once more giving

u00 (z, x)  x
Z 
f 00 (x) = + u− (z, t)g(t)dt
W (z) 0
u00− (z, x)  1
Z 
+ u+ (z, t)g(t)dt − g(x)
W (z) x
=(q(x) − z)f (x) − g(x). (3.20)

In summary, f is in the domain of L and satisfies (L − z)f = g.

Note that z is an eigenvalue if and only if W (z) = 0. In fact, in this case
u+ (z, x) and u− (z, x) are linearly dependent and hence u+ (z, x) = c u− (z, x)
with c = u0+ (z, 0). Evaluating this identity at x = 0 shows u+ (z, 0) =
c u− (z, 0) = 0 that u+ (z, x) satisfies both boundary conditions and is thus
an eigenfunction.
Introducing the Green function

1 u+ (z, x)u− (z, t), x ≥ t,
G(z, x, t) := (3.21)
W (u+ (z), u− (z)) u+ (z, t)u− (z, x), x ≤ t,

we see that (L − z)−1 is given by

Z 1
(L − z)−1 g(x) = G(z, x, t)g(t)dt. (3.22)
0

Moreover, from G(z, x, t) = G(z, t, x) it follows that (L − z)−1 is symmetric

for z ∈ R (Problem 3.7) and from Lemma 3.4 it follows that it is compact.
Hence Theorem 3.7 applies to (L − z)−1 once we show that we can find a
real z which is not an eigenvalue.

Theorem 3.11. The Sturm–Liouville operator L has a countable number of

discrete and simple eigenvalues En which accumulate only at ∞. They are
bounded from below and can hence be ordered as follows:

min q(x) < E0 < E1 < · · · . (3.23)

x∈[0,1]
3.3. Applications to Sturm–Liouville operators 77

The corresponding normalized eigenfunctions un form an orthonormal basis

for L2cont (0, 1), that is, every f ∈ L2cont (0, 1) can be written as
∞
X
f (x) = hun , f iun (x). (3.24)
n=0

Moreover, for f ∈ D(L) this series is absolutely uniformly convergent.

Proof. If Ej is an eigenvalue with corresponding normalized eigenfunction

uj we have
Z 1
|u0j (x)|2 + q(x)|uj (x)|2 dx < min q(x) (3.25)


Ej = huj , Luj i =
0 x∈[0,1]

where we have used integration by parts as in (3.16). Note that equality

could only occur if uj is constant, which is incompatible with our boundary
conditions. Hence the eigenvalues are bounded from below.
Now pick a value λ ∈ R such that RL (λ) exists (λ < minx∈[0,1] q(x)
say). By Lemma 3.4 RL (λ) is compact and by Lemma 3.3 this remains
true if we replace L2cont (0, 1) by its completion. By Theorem 3.7 there are
eigenvalues αn of RL (λ) with corresponding eigenfunctions un . Moreover,
RL (λ)un = αn un is equivalent to Lun = (λ + α1n )un , which shows that
En = λ+ α1n are eigenvalues of L with corresponding eigenfunctions un . Now
everything follows from Theorem 3.7 except that the eigenvalues are simple.
To show this, observe that if un and vn are two different eigenfunctions
corresponding to En , then un (0) = vn (0) = 0 implies W (un , vn ) = 0 and
hence un and vn are linearly dependent.
To show that (3.24) converges uniformly if f ∈ D(L) we begin by writing
f = RL (λ)g, g ∈ L2cont (0, 1), implying
∞
X ∞
X ∞
X
hun , f iun (x) = hRL (λ)un , giun (x) = αn hun , giun (x).
n=0 n=0 n=0

Moreover, the Cauchy–Schwarz inequality shows

 2
 n  n n
X X X
2
|αj uj (x)|2 .
 

 |αj huj , giuj (x)|  ≤
 |huj , gi|
j=m  j=m j=m

Now, by (2.18), ∞ j=0 |huj , gi| = kgk and hence the first term is part of a
2 2
P
convergent series. Similarly, the second term can be estimated independent
of x since
Z 1
αn un (x) = RL (λ)un (x) = G(λ, x, t)un (t)dt = hun , G(λ, x, .)i
0
78 3. Compact operators

implies

n
X ∞
X Z 1
2 2
|αj uj (x)| ≤ |huj , G(λ, x, .)i| = |G(λ, x, t)|2 dt ≤ M (λ)2 ,
j=m j=0 0

where M (λ) := maxx,t∈[0,1] |G(λ, x, t)|, again by (2.18). 

Moreover, it is even possible to weaken our assumptions for uniform

convergence. To this end we consider the sequilinear form associated with
L:
Z 1
f 0 (x)∗ g 0 (x) + q(x)f (x)∗ g(x) dx (3.26)


sL (f, g) := hf, Lgi =
0

for f, g ∈ D(L), where we have used integration by parts as in (3.16). In

fact, the above formula continues to hold for f in a slightly larger class of
functions,

Q(L) := {f ∈ Cp1 [0, 1]|f (0) = f (1) = 0} ⊇ D(L), (3.27)

which we call the form domain of L. Here Cp1 [a, b] denotes the set of
piecewise continuously differentiable functions f in the sense that f is con-
tinuously differentiable except for a finite number of points at which it is
continuous and the derivative has limits from the left and right. In fact, any
class of functions for which the partial integration needed to obtain (3.26)
can be justified would be good enough (e.g. the set of absolutely continuous
functions to be discussed in Section 4.4 from [47]).

Lemma 3.12. For a regular Sturm–Liouville problem (3.24) converges ab-

solutely uniformly provided f ∈ Q(L).

Proof. By replacing L → L − q0 for q0 < minx∈[0,1] q(x) we can assume

q(x) > 0 without loss of generality. (This will shift the eigenvalues En →
En − q0 and leave the eigenvectors unchanged.) In particular, we have
qL (f ) := sL (f, f ) > 0 after this change. By (3.26) we also have Ej =
huj , Luj i = qL (uj ) > 0.
Now let f ∈ Q(L) and consider (3.24). Then, observing that sL (f, g) is
a symmetric sesquilinear form (after our shift it is even a scalar product) as
3.3. Applications to Sturm–Liouville operators 79

well as sL (f, uj ) = Ej hf, uj i one obtains

n
X n
X


0 ≤qL f − huj , f iuj = qL (f ) − huj , f isL (f, uj )
j=m j=m
n
X n
X
− huj , f i∗ sL (uj , f ) + huj , f i∗ huk , f isL (uj , uk )
j=m j,k=m
n
X
=qL (f ) − Ej |huj , f i|2
j=m

which implies
n
X
Ej |huj , f i|2 ≤ qL (f ).
j=m

In particular, note that this estimate applies to f (y) = G(λ, x, y). Now
we can proceed as in the proof of the previous theorem (with λ = 0 and
αj = Ej−1 )
n
X n
X
|huj , f iuj (x)| = Ej |huj , f ihuj , G(0, x, .)i|
j=m j=m
 1/2
n
X n
X
≤ Ej |huj , f i|2 Ej |huj , G(0, x, .)i|2 
j=m j=m
 1/2
n
X
≤ Ej |huj , f i|2  qL (G(0, x, .))1/2 ,
j=m

where we have usedPthe Cauchy–Schwarz inequality for the weighted scalar

product (fj , gj ) 7→ j fj∗ gj Ej . Finally note that qL (G(0, x, .)) is continuous
with respect to x and hence can be estimated by its maximum over [0, 1].
This shows that the sum (3.24) is absolutely convergent, uniformly with
respect to x. 

Another consequence of the computations in the previous proof is also

worthwhile noting:

Corollary 3.13. We have

∞
X 1
G(z, x, y) = uj (x)uj (y), (3.28)
Ej − z
j=0
80 3. Compact operators

where the sum is uniformly convergent. Moreover, we have the following

trace formula
Z 1 ∞
X 1
G(z, x, x)dx = . (3.29)
0 Ej −z
j=0

Proof. Using the conventions from the proof of the previous lemma we have
huj , G(0, x, .)i = Ej−1 uj (x) and since G(0, x, .) ∈ Q(L) for fixed x ∈ [a, b] we
have
∞
X 1
uj (x)uj (y) = G(0, x, y),
Ej
j=0

where the convergence is uniformly with respect to y (and x fixed). Moreover,

for x = y Dini’s theorem (cf. Problem B.35) shows that the convergence is
uniform with respect to x = y and this also proves uniform convergence of
our sum since
 1/2 1/2
n n n
X 1 X 1 X 1
|uj (x)uj (y)| ≤ C(z)  uj (x)2   uj (y)2  ,
|Ej − z| Ej Ej
j=0 j=0 j=0

Ej
where C(z) := supj |Ej −z| .
Finally, the last claim follows upon computing the integral using (3.28)
and observing kuj k = 1. 

Example 3.8. Let us look at the Sturm–Liouville problem with q = 0. Then

the underlying differential equation is
−u00 (x) = z u(x)
√ √
whose solution is given by u(x) = c1 sin( zx) + c2 cos( zx). The solution
√
satisfying the boundary condition at the left endpoint is u− (z, x) = sin( zx)
√
and it will be an eigenfunction if and only if u− (z, 1) = sin( z) = 0. Hence
the corresponding eigenvalues and normalized eigenfunctions are
√
En = π 2 n2 , un (x) = 2 sin(nπx), n ∈ N.
Moreover, every function f ∈ L2cont (0, 1) can be expanded into a Fourier
sine series
∞
X Z 1
f (x) = fn un (x), fn := un (x)f (x)dx,
n=1 0

which is convergent with respect to our scalar product. If f ∈ Cp1 [0, 1] with
f (0) = f (1) = 0 the series will converge uniformly. For an application of the
trace formula see Problem 3.10. 
3.3. Applications to Sturm–Liouville operators 81

Example 3.9. We could also look at the same equation as in the previous
problem but with different boundary conditions
u0 (0) = u0 (1) = 0.
Then (
2 2 1, n = 0,
En = π n , un (x) = √
2 cos(nπx), n ∈ N.
Moreover, every function f ∈ L2cont (0, 1) can be expanded into a Fourier
cosine series
∞
X Z 1
f (x) = fn un (x), fn := un (x)f (x)dx,
n=1 0

which is convergent with respect to our scalar product. 

Example 3.10. Combining the last two examples we see that every symmet-
ric function on [−1, 1] can be expanded into a Fourier cosine series and every
anti-symmetric function into a Fourier sine series. Moreover, since every
function f (x) can be written as the sum of a symmetric function f (x)+f 2
(−x)

and an anti-symmetric function f (x)−f

2
(−x)
, it can be expanded into a Fourier
series. Hence we recover Theorem 2.18. 
Problem* 3.7. Show that for our Sturm–Liouville operator u± (z, x)∗ =
u± (z ∗ , x). Conclude RL (z)∗ = RL (z ∗ ). (Hint: Problem 3.2.)
Problem 3.8. Show that the resolvent RA (z) = (A − z)−1 (provided it exists
and is densely defined) of a symmetric operator A is again symmetric for
z ∈ R. (Hint: g ∈ D(RA (z)) if and only if g = (A−z)f for some f ∈ D(A).)
Problem 3.9. Suppose E0 > 0 and equip Q(L) with the scalar product sL .
Show that
f (x) = sL (G(0, x, .), f ).
In other words, point evaluations are continuous functionals associated with
the vectors G(0, x, .) ∈ Q(L). In this context, G(0, x, y) is called a repro-
ducing kernel.
Problem 3.10. Show that
∞ √ √
X 1 1 − π z cot(π z)
= , z ∈ C \ N.
n2 − z 2z
n=1

In particular, for z = 0 this gives Euler’s solution of the Basel problem:

∞
X 1 π2
= .
n2 6
n=1
82 3. Compact operators

In fact, comparing the power series of both sides at z = 0 gives

∞
X 1 (−1)k+1 (2π)2k B2k
= , k ∈ N,
n2k 2(2k)!
n=1
x P∞ Bk k
where Bk are the Bernoulli numbers defined via ex −1 = k=0 k! z .
(Hint: Use the trace formula (3.29).)
Problem 3.11. Consider the Sturm–Liouville problem on a compact interval
[a, b] with domain
D(L) = {f ∈ C 2 [a, b]|f 0 (a) − αf (a) = f 0 (b) − βf (b) = 0}
for some real constants α, β ∈ R. Show that Theorem 3.11 continues to hold
except for the lower bound on the eigenvalues.

3.4. Estimating eigenvalues

In general, there is no way of computing eigenvalues and their corresponding
eigenfunctions explicitly. Hence it is important to be able to determine the
eigenvalues at least approximately.
Let A be a symmetric operator which has a lowest eigenvalue α1 (e.g.,
A is a Sturm–Liouville operator). Suppose we have a vector f which is an
approximation for the eigenvector u1 of this lowest eigenvalue α1 . Moreover,
suppose we can write
∞
X ∞
X
A := αj huj , .iuj , D(A) := {f ∈ H| |αj huj , f i|2 < ∞}, (3.30)
j=1 j=1

where {uj }j∈N is an orthonormal basis of eigenvectors. Since α1 is supposed

to be the lowest eigenvalue we have αj ≥ α1 for all j ∈ N.
Writing f = j γj uj , γj = huj , f i, one computes
P

∞
X ∞
X
hf, Af i = hf, αj γj uj i = αj |γj |2 , f ∈ D(A), (3.31)
j=1 j=1

and we clearly have

hf, Af i
α1 ≤ , f ∈ D(A), (3.32)
kf k2
with equality for f = u1 . In particular, any f will provide an upper bound
and if we add some free parameters to f , one can optimize them and obtain
quite good upper bounds for the first eigenvalue. For example we could
take some orthogonal basis, take a finite number of coefficients and optimize
them. This is known as the Rayleigh–Ritz method.
3.4. Estimating eigenvalues 83

Example 3.11. Consider the Sturm–Liouville operator L with potential

q(x) = x and Dirichlet boundary conditions f (0) = f (1) = 0 on the in-
terval [0, 1]. Our starting point is the quadratic form
Z 1
|f 0 (x)|2 + q(x)|f (x)|2 dx


qL (f ) := hf, Lf i =
0
which gives us the lower bound
hf, Lf i ≥ min q(x) = 0.
0≤x≤1

While the corresponding differential equation can in principle be solved in

terms of Airy functions, there is no closed form for the eigenvalues.
First of all we can improve the above bound upon observing 0 ≤ q(x) ≤ 1
which implies
hf, L0 f i ≤ hf, Lf i ≤ hf, (L0 + 1)f i, f ∈ D(L) = D(L0 ),
where L0 is the Sturm–Liouville operator corresponding to q(x) = 0. Since
the lowest eigenvalue of L0 is π 2 we obtain
π 2 ≤ E1 ≤ π 2 + 1
for the lowest eigenvalue E1 of L.
√
Moreover, using the lowest eigenfunction f1 (x) = 2 sin(πx) of L0 one
obtains the improved upper bound
1
E1 ≤ hf1 , Lf1 i = π 2 + ≈ 10.3696.
2
√
Taking the second eigenfunction f2 (x) = 2 sin(2πx) of L0 we can make the
ansatz f (x) = (1 + γ 2 )−1/2 (f1 (x) + γf2 (x)) which gives
1 γ 32

hf, Lf i = π 2 + + 2
3π 2 γ − 2 .
2 1+γ 9π
The right-hand side has a unique minimum at γ = 27π4 +√1024+729π
32
8
giving
the bound √
5 2 1 1024 + 729π 8
E1 ≤ π + − ≈ 10.3685
2 2 18π 2
which coincides with the exact eigenvalue up to five digits. 
But is there also something one can say about the next eigenvalues?
Suppose we know the first eigenfunction u1 . Then we can restrict A to
the orthogonal complement of u1 and proceed as before: E2 will be the
minimum of hf, Af i over all f restricted to this subspace. If we restrict to
the orthogonal complement of an approximating eigenfunction f1 , there will
still be a component in the direction of u1 left and hence the infimum of the
expectations will be lower than E2 . Thus the optimal choice f1 = u1 will
give the maximal value E2 .
84 3. Compact operators

Theorem 3.14 (Max-min). Let A be a symetric operator and let α1 ≤ α2 ≤

· · · ≤ αN be eigenvalues of A with corresponding orthonormal eigenvectors
u1 , u2 , . . . , uN . Suppose
N
X
A= αj huj , .iuj + Ã (3.33)
j=1

with hf, Ãf i ≥ αN kf k2 for all f ∈ D(A) and u1 , . . . , uN ∈ Ker(Ã). Then

αj = sup inf hf, Af i, 1 ≤ j ≤ N, (3.34)
f1 ,...,fj−1 f ∈U (f1 ,...,fj−1 )

where
U (f1 , . . . , fj ) := {f ∈ D(A)| kf k = 1, f ∈ span{f1 , . . . , fj }⊥ }. (3.35)

Proof. We have
inf hf, Af i ≤ αj .
f ∈U (f1 ,...,fj−1 )
Pj
In fact, set f = k=1 γk uk and choose γk such that f ∈ U (f1 , . . . , fj−1 ).
Then
j
X
hf, Af i = |γk |2 αk ≤ αj
k=1
and the claim follows.
Conversely, let γk = huk , f i and write f = jk=1 γk uk + f˜. Then
P
 
XN
inf hf, Af i = inf  |γk |2 αk + hf˜, Ãf˜i = αj . 
f ∈U (u1 ,...,uj−1 ) f ∈U (u1 ,...,uj−1 )
k=j

Of course if we are interested in the largest eigenvalues all we have to do

is consider −A.
Note that this immediately gives an estimate for eigenvalues if we have
a corresponding estimate for the operators. To this end we will write
A≤B ⇔ hf, Af i ≤ hf, Bf i, f ∈ D(A) ∩ D(B). (3.36)
Corollary 3.15. Suppose A and B are symmetric operators with corre-
sponding eigenvalues αj and βj as in the previous theorem. If A ≤ B and
D(B) ⊆ D(A) then αj ≤ βj .

Proof. By assumption we have hf, Af i ≤ hf, Bf i for f ∈ D(B) implying

inf hf, Af i ≤ inf hf, Af i ≤ inf hf, Bf i,
f ∈UA (f1 ,...,fj−1 ) f ∈UB (f1 ,...,fj−1 ) f ∈UB (f1 ,...,fj−1 )

where we have indicated the dependence of U on the operator via a subscript.

Taking the sup on both sides the claim follows. 
3.5. Singular value decomposition of compact operators 85

Example 3.12. Let L be again our Sturm–Liouville operator and L0 the

corresponding operator with q(x) = 0. Set q− = min0≤x≤1 q(x) and q+ =
max0≤x≤1 q(x). Then L0 + q− ≤ L ≤ L0 + q+ implies
π 2 n2 + q− ≤ En ≤ π 2 n2 + q+ .
In particular, we have proven the famous Weyl asymptotic
En = π 2 n2 + O(1)
for the eigenvalues. 
There is also an alternative version which can be proven similar (Prob-
lem 3.12):
Theorem 3.16 (Min-max). Let A be as in the previous theorem. Then
αj = inf sup hf, Af i, (3.37)
Vj ⊂D(A),dim(Vj )=j f ∈Vj ,kf k=1

where the inf is taken over subspaces with the indicated properties.
Problem* 3.12. Prove Theorem 3.16.
Problem 3.13. Suppose A, An are self-adjoint, bounded and An → A.
Then αk (An ) → αk (A). (Hint: For B self-adjoint kBk ≤ ε is equivalent to
−ε ≤ B ≤ ε.)

3.5. Singular value decomposition of compact operators

Our first aim is to find a generalization of Corollary 3.8 for general com-
pact operators between Hilbert spaces. The key observation is that if K ∈
C (H1 , H2 ) is compact, then K ∗ K ∈ C (H1 ) is compact and symmetric and
thus, by Corollary 3.8, there is a countable orthonormal set {uj } ⊂ H1 and
nonzero real numbers s2j 6= 0 such that
X
K ∗ Kf = s2j huj , f iuj . (3.38)
j

Moreover, kKuj k2 = huj , K ∗ Kuj i = huj , s2j uj i = s2j shows that we can set
sj := kKuj k > 0. (3.39)
The numbers sj = sj (K) are called singular values of K. There are either
finitely many singular values or they converge to zero.
Theorem 3.17 (Singular value decomposition of compact operators). Let
K ∈ C (H1 , H2 ) be compact and let sj be the singular values of K and {uj } ⊂
H1 corresponding orthonormal eigenvectors of K ∗ K. Then
X
K= sj huj , .ivj , (3.40)
j
86 3. Compact operators

where vj = s−1
j Kuj . The norm of K is given by the largest singular value

kKk = max sj (K). (3.41)

j

Moreover, the vectors {vj } ⊂ H2 are again orthonormal and satisfy K ∗ vj =

sj uj . In particular, vj are eigenvectors of KK ∗ corresponding to the eigen-
values s2j .

Proof. For any f ∈ H1 we can write

X
f= huj , f iuj + f⊥
j

with f⊥ ∈ Ker(K ∗ K) = Ker(K) (Problem 3.14). Then

X X
Kf = huj , f iKuj = sj huj , f ivj
j j

as required. Furthermore,
hvj , vk i = (sj sk )−1 hKuj , Kuk i = (sj sk )−1 hK ∗ Kuj , uk i = sj s−1
k huj , uk i

shows that {vj } are orthonormal. By definition K ∗ vj = s−1 ∗

j K Kuj = sj uj
which also shows KK vj = sj Kuj = sj vj .
∗ 2

Finally, (3.41) follows using Bessel’s inequality

X X  
kKf k2 = k sj huj , f ivj k2 = s2j |huj , f i|2 ≤ max sj (K)2 kf k2 ,
j
j j

where equality holds for f = uj0 if sj0 = maxj sj (K). 

If K ∈ C (H) is self-adjoint, then uj = σj vj , σj2 = 1, are the eigenvectors

of K and σj sj are the corresponding eigenvalues. In particular, for a self-
adjoint operators the singular values are the absolute values of the nonzero
eigenvalues.
The above theorem also gives rise to the polar decomposition
K = U |K| = |K ∗ |U, (3.42)
where
√ X √ X
|K| := K ∗K = sj huj , .iuj , |K ∗ | = KK ∗ = sj hvj , .ivj (3.43)
j j

are self-adjoint (in fact nonnegative) and

X
U := huj , .ivj (3.44)
j

is an isometry from Ran(K ∗ ) = span{uj } onto Ran(K) = span{vj }.

From the min-max theorem (Theorem 3.16) we obtain:
3.5. Singular value decomposition of compact operators 87

Lemma 3.18. Let K ∈ C (H1 , H2 ) be compact; then

sj (K) = min max kKf k, (3.45)
f1 ,...,fj−1 f ∈U (f1 ,...,fj−1 )

where U (f1 , . . . , fj ) := {f ∈ H1 | kf k = 1, f ∈ span{f1 , . . . , fj }⊥ }.

In particular, note
sj (AK) ≤ kAksj (K), sj (KA) ≤ kAksj (K) (3.46)
whenever K is compact and A is bounded (the second estimate follows from
the first by taking adjoints).
An operator K ∈ L (H1 , H2 ) is called a finite rank operator if its
range is finite dimensional. The dimension
rank(K) := dim Ran(K)
is called the rank of K. Since for a compact operator
Ran(K) = span{vj } (3.47)
we see that a compact operator is finite rank if and only if the sum in (3.40)
is finite. Note that the finite rank operators form an ideal in L (H) just as
the compact operators do. Moreover, every finite rank operator is compact
by the Heine–Borel theorem (Theorem B.22).
Now truncating the sum in the canonical form gives us a simple way to
approximate compact operators by finite rank ones. Moreover, this is in fact
the best approximation within the class of finite rank operators:
Lemma 3.19. Let K ∈ C (H1 , H2 ) be compact and let its singular values be
ordered. Then
sj (K) = min kK − F k, (3.48)
rank(F )<j
where the minimum is attained for
j−1
X
Fj−1 := sk huk , .ivk . (3.49)
k=1
In particular, the closure of the ideal of finite rank operators in L (H) is the
ideal of compact operators.

Proof. That there is equality for F = Fj−1 follows from (3.41). In general,
the restriction of F to span{u1 , . . . , uj } will have a nontrivial kernel. Let
f = jk=1 αj uj be a normalized element of this kernel, then k(K − F )f k2 =
P

kKf k2 = jk=1 |αk sk |2 ≥ s2j .

P

In particular, every compact operator can be approximated by finite rank

ones and since the limit of compact operators is compact, we cannot get more
than the compact operators. 
88 3. Compact operators

Two more consequences are worthwhile noting.

Corollary 3.20. An operator K ∈ L (H1 , H2 ) is compact if and only if K ∗ K
is.

Proof. Just observe that K ∗ K compact is all that was used to show Theo-
rem 3.17. 
Corollary 3.21. An operator K ∈ L (H1 , H2 ) is compact (finite rank) if
and only K ∗ ∈ L (H2 , H1 ) is. In fact, sj (K) = sj (K ∗ ) and
X
K∗ = sj hvj , .iuj . (3.50)
j

Proof. First of all note that (3.50) follows from (3.40) since taking ad-
joints is continuous and (huj , .ivj )∗ = hvj , .iuj (cf. Problem 2.7). The rest is
straightforward. 

From this last lemma one easily gets a number of useful inequalities for
the singular values:
Corollary 3.22. Let K1 and K2 be compact and let sj (K1 ) and sj (K2 ) be
ordered. Then
(i) sj+k−1 (K1 + K2 ) ≤ sj (K1 ) + sk (K2 ),
(ii) sj+k−1 (K1 K2 ) ≤ sj (K1 )sk (K2 ),
(iii) |sj (K1 ) − sj (K2 )| ≤ kK1 − K2 k.

Proof. Let F1 be of rank j − 1 and F2 of rank k − 1 such that kK1 − F1 k =

sj (K1 ) and kK2 − F2 k = sk (K2 ). Then sj+k−1 (K1 + K2 ) ≤ k(K1 + K2 ) −
(F1 + F2 )k = kK1 − F1 k + kK2 − F2 k = sj (K1 ) + sk (K2 ) since F1 + F2 is of
rank at most j + k − 2.
Similarly F = F1 (K2 − F2 ) + K1 F2 is of rank at most j + k − 2 and hence
sj+k−1 (K1 K2 ) ≤ kK1 K2 − F k = k(K1 − F1 )(K2 − F2 )k ≤ kK1 − F1 kkK2 −
F2 k = sj (K1 )sk (K2 ).
Next, choosing k = 1 and replacing K2 → K2 − K1 in (i) gives sj (K2 ) ≤
sj (K1 ) + kK2 − K1 k. Reversing the roles gives sj (K1 ) ≤ sj (K2 ) + kK1 − K2 k
and proves (iii). 
Example 3.13. On might hope that item (i) from the previous corollary
can be improved to sj (K1 + K2 ) ≤ sj (K1 ) + sj (K2 ). However, this is not
the case as the following example shows:
   
1 0 0 0
K1 := , K2 := .
0 0 0 1
Then 1 = s2 (K1 + K2 ) 6≤ s2 (K1 ) + s2 (K2 ) = 0. 
3.6. Hilbert–Schmidt and trace class operators 89

Problem* 3.14. Show that Ker(A∗ A) = Ker(A) for any A ∈ L (H1 , H2 ).

Problem 3.15. Let K be multiplication by a sequence k ∈ c0 (N) in the
Hilbert space `2 (N). What are the singular values of K?
Problem 3.16. Let K be multiplication by a sequence k ∈ c0 (N) in the
Hilbert space `2 (N) and consider L = KS − . What are the singular values of
L? Does L have any eigenvalues?
Problem 3.17. Let K ∈ C (H1 , H2 ) be compact and let its singular values be
ordered. Let M ⊆ H1 , N ⊆ H1 be subspaces whith corresponding orthogonal
projections PM , PN , respectively. Then
sj (K) = min kK − KPM k = min kK − PN Kk,
dim(M )<j dim(N )<j

where the minimum is taken over all subspaces with the indicated dimension.
Moreover, the minimum is attained for
M = span{uk }j−1
k=1 , N = span{vk }j−1
k=1 .

3.6. Hilbert–Schmidt and trace class operators

We can further subdivide the class of compact operators C (H) according to
the decay of their singular values. We define
X 1/p
kKkp := sj (K)p (3.51)
j

plus corresponding spaces

Jp (H) = {K ∈ C (H)|kKkp < ∞}, (3.52)
which are known as Schatten p-classes. Even though our notation hints
at the fact that k.kp is a norm, we will only prove this here for p = 1, 2 (the
only nontrivial part is the triangle inequality). Note that by (3.41)
kKk ≤ kKkp (3.53)
and that by sj (K) = sj (K ∗ ) we have
kKkp = kK ∗ kp . (3.54)

The two most important cases are p = 1 and p = 2: J2 (H) is the space
of Hilbert–Schmidt operators and J1 (H) is the space of trace class
operators.
Example 3.14. Any multiplication operator by a sequence from `p (N) is in
the Schatten p-class of H = `2 (N). 
Example 3.15. By virtue of the Weyl asymptotics (see Example 3.12) the
resolvent of our Sturm–Liouville operator is trace class. 
90 3. Compact operators

Example 3.16. Let k be a periodic function which is square integrable over

[−π, π]. Then the integral operator
Z π
1
(Kf )(x) = k(y − x)f (y)dy
2π −π
has the eigenfunctions uj (x) = (2π)−1/2 e−ijx with corresponding eigenvalues
k̂j , j ∈ Z, where k̂j are the Fourier coefficients of k. Since {uj }j∈Z is an ONB
we have found all eigenvalues. In particular, the Fourier transform maps K
to the multiplication operator with the sequence of its eigenvalues k̂j . Hence
the singular values are the absolute values of the nonzero eigenvalues and
(3.40) reads X
K= k̂j huj , .iuj .
j∈Z
Moreover, since the eigenvalues are in `2 (Z) we see that K is a Hilbert–
Schmidt operator. If k is continuous with summable Fourier coefficients
(e.g. k ∈ Cper
2 [−π, π]), then K is trace class. 
We first prove an alternate definition for the Hilbert–Schmidt norm.
Lemma 3.23. A bounded operator K is Hilbert–Schmidt if and only if
X
kKwj k2 < ∞ (3.55)
j∈J

for some orthonormal basis and

X 1/2
kKk2 = kKwj k2 , (3.56)
j∈J

for every orthonormal basis in this case.

Proof. First of all note that (3.55) implies that K is compact. To see this,
let Pn be the projection onto the space spanned by the first n elements of
the orthonormal basis {wj }. Then Kn = KPn is finite rank and converges
to K since
X X X 1/2
k(K − Kn )f k = k cj Kwj k ≤ |cj |kKwj k ≤ kKwj k2 kf k,
j>n j>n j>n

where f = j cj wj .
P

The rest follows from (3.40) and

X X X X
kKwj k2 = |hvk , Kwj i|2 = |hK ∗ vk , wj i|2 = kK ∗ vk k2
j k,j k,j k
X
= sk (K)2 = kKk22 .
k

Here we have used span{vk } = Ker(K ∗ )⊥ = Ran(K) in the first step. 

3.6. Hilbert–Schmidt and trace class operators 91

Corollary 3.24. The Hilbert–Schmidt norm satisfies the triangle inequality

and hence is indeed a norm.

Proof. This follows from (3.56) upon using the triangle inequality for H and
for `2 (J). 

Now we can show

Lemma 3.25. The set of Hilbert–Schmidt operators forms an ideal in L (H)
and
kKAk2 ≤ kAkkKk2 , respectively, kAKk2 ≤ kAkkKk2 . (3.57)

Proof. If K1 and K2 are Hilbert–Schmidt operators, then so is their sum

since
X 1/2  X 1/2
kK1 + K2 k2 = k(K1 + K2 )wj k2 ≤ (kK1 wj k + kK2 wj k)2
j∈J j∈J
≤ kK1 k2 + kK2 k2 ,
where we have used the triangle inequality for `2 (J).
Let K be Hilbert–Schmidt and A bounded. Then AK is compact and
X X
kAKk22 = kAKwj k2 ≤ kAk2 kKwj k2 = kAk2 kKk22 .
j j

For KA just consider adjoints. 

Example 3.17. Consider `2 (N) and let K be some compact operator. Let
Kjk = hδ j , Kδ k i = (Kδ j )k be its matrix elements such that
∞
X
(Ka)j = Kjk ak .
k=1

Then, choosing wj = δj in (3.56) we get

∞
X 1/2 ∞ X
X ∞ 1/2
kKk2 = kKδ j k2 = |Kjk |2 .
j=1 j=1 k=1

Hence K is Hilbert–Schmidt if and only if its matrix elements are in `2 (N ×

N) and the Hilbert–Schmidt norm coincides with the `2 (N × N) norm of
the matrix elements. Especially in the finite dimensional case the Hilbert–
Schmidt norm is also known as Frobenius norm.
Of course the same calculation shows that a bounded operator is Hilbert–
Schmidt if and only if its matrix elements hwj , Kwk i with respect to some
orthonormal basis {wj }j∈J are in `2 (J × J) and the Hilbert–Schmidt norm
coincides with the `2 (J × J) norm of the matrix elements. 
92 3. Compact operators

Example 3.18. Let I = [a, b] be a compact interval. Suppose K : L2 (I) →

C(I) is continuous, then√K : L2 (I) → L2 (I) is Hilbert–Schmidt with Hilbert–
Schmidt norm kKk2 ≤ b − aM , where M := kKkL2 (I)→C(I) .
To see this start by observing that point evaluations are continuous func-
tionals on C(I) and hence f 7→ (Kf )(x) is a continuous linear functional on
L2 (I) satisfying |(Kf )(x)| ≤ M kf k. By the Riesz lemma there is some
Kx ∈ L2 (I) with kKx k ≤ M such that
(Kf )(x) = hKx , f i
and hence for any orthonormal basis {wj }j∈N we have
X X
|(Kwj )(x)|2 = |hKx , wj i|2 = kKx k2 ≤ M 2 .
j∈N j∈N

But then
X XZ b Z bX 
2 2
kKwj k = |(Kwj )(x)| dx = |(Kwj )(x)|2 dx
j∈N j∈N a a j∈N
2
≤ (b − a)M
as claimed. 
Since Hilbert–Schmidt operators turn out easy to identify (cf. also Sec-
tion 3.5 from [47]), it is important to relate J1 (H) with J2 (H):
Lemma 3.26. An operator is trace class if and only if it can be written as
the product of two Hilbert–Schmidt operators, K = K1 K2 , and in this case
we have
kKk1 ≤ kK1 k2 kK2 k2 . (3.58)
In fact, K1 , K2 can be chosen such that kKk1 = kK1 k2 kK2 k2 .

Proof. Using (3.40) (where we can extend un and vn to orthonormal bases

if necessary) and Cauchy–Schwarz we have
X X
kKk1 = hvn , Kun i = |hK1∗ vn , K2 un i|
n n
X X 1/2
≤ kK1∗ vn k2 kK2 un k2 = kK1 k2 kK2 k2
n n

and hence K = K1 K2 is trace class if both K1 and K2 are Hilbert–Schmidt

operators. To see the converse, let K be given by (3.40) and choose K1 =
, respectively, j (K)huj , .iuj . Note that in
P p P p
j s j (K)huj , .iv j K2 = j s
this case kKk1 = kK1 k2 = kK2 k2 .
2 2 

Now we can also explain the name trace class:

3.6. Hilbert–Schmidt and trace class operators 93

Lemma 3.27. If K is trace class, then for every orthonormal basis {wn }
the trace X
tr(K) = hwn , Kwn i (3.59)
n
is finite,
| tr(K)| ≤ kKk1 , (3.60)
and independent of the orthonormal basis.

Proof. If we write K = K1 K2 with K1 , K2 Hilbert–Schmidt such that

kKk1 = kK1 k2 kK2 k2 , then the Cauchy–Schwarz inequality implies | tr(K)| ≤
kK1∗ k2 kK2 k2 = kKk1 . Moreover, if {w̃n } is another orthonormal basis, we
have
X X X
hwn , K1 K2 wn i = hK1∗ wn , K2 wn i = hK1∗ wn , w̃m ihw̃m , K2 wn i
n n n,m
X X
= hK2∗ w̃m , wn ihwn , K1 w̃m i = hK2∗ w̃m , K1 w̃m i
m,n m
X
= hw̃m , K2 K1 w̃m i.
m
In the special case w = w̃ we see tr(K1 K2 ) = tr(K2 K1 ) and the general case
now shows that the trace is independent of the orthonormal basis. 

Clearly for self-adjoint trace class operators, the trace is the sum over
all eigenvalues (counted with their multiplicity). To see this, one just has to
choose the orthonormal basis to consist of eigenfunctions. This is even true
for all trace class operators and is known as Lidskij trace theorem (see [36]
for an easy to read introduction).
Example 3.19. We already mentioned that the resolvent of our Sturm–
Liouville operator is trace class. Choosing a basis of eigenfunctions we see
that the trace of the resolvent is the sum over its eigenvalues and combining
this with our trace formula (3.29) gives
∞ Z 1
X 1
tr(RL (z)) = = G(z, x, x)dx
Ej − z 0
j=0

for z ∈ C no eigenvalue. 
Example 3.20. For our integral operator K from Example 3.16 we have in
the trace class case X
tr(K) = k̂j = k(0).
j∈Z
Note that this can again be interpreted as the integral over the diagonal
(2π)−1 k(x − x) = (2π)−1 k(0) of the kernel. 
We also note the following elementary properties of the trace:
94 3. Compact operators

Lemma 3.28. Suppose K, K1 , K2 are trace class and A is bounded.

(i) The trace is linear.
(ii) tr(K ∗ ) = tr(K)∗ .
(iii) If K1 ≤ K2 , then tr(K1 ) ≤ tr(K2 ).
(iv) tr(AK) = tr(KA).

Proof. (i) and (ii) are straightforward. (iii) follows from K1 ≤ K2 if and
only if hf, K1 f i ≤ hf, K2 f i for every f ∈ H. (iv) By Problem 2.12 and (i),
it is no restriction to assume that A is unitary. Let {wn } be some ONB and
note that {w̃n = Awn } is also an ONB. Then
X X
tr(AK) = hw̃n , AK w̃n i = hAwn , AKAwn i
n n
X
= hwn , KAwn i = tr(KA)
n

and the claim follows. 

We also mention a useful criterion for K to be trace class.

Lemma 3.29. An operator K is trace class if and only if it can be written

as
X
K= hfj , .igj (3.61)
j

for some sequences fj , gj satisfying

X
kfj kkgj k < ∞. (3.62)
j

Moreover, in this case

X
tr(K) = hfj , gj i (3.63)
j

and
X
kKk1 = min kfj kkgj k, (3.64)
j

where the minimum is taken over all representations as in (3.61).

Proof. To see that a trace class operator (3.40) can be written in such a
way choose fj = uj , gj = sj vj . This also shows that the minimum in (3.64)
is attained. Conversely note that the sum converges in the operator norm
3.6. Hilbert–Schmidt and trace class operators 95

and hence K is compact. Moreover, for every finite N we have

N
X N
X N X
X N
XX
sk = hvk , Kuk i = hvk , gj ihfj , uk i = hvk , gj ihfj , uk i
k=1 k=1 k=1 j j k=1

N
!1/2 N
!1/2
X X X X
2
≤ |hvk , gj i| |hfj , uk i|2 ≤ kfj kkgj k.
j k=1 k=1 j

This also shows that the right-hand side in (3.64) cannot exceed kKk1 . To
see the last claim we choose an ONB {wk } to compute the trace
X XX XX
tr(K) = hwk , Kwk i = hwk , hfj , wk igj i = hhwk , fj iwk , gj i
k k j j k
X
= hfj , gj i. 
j

An immediate consequence of (3.64) is:

Corollary 3.30. The trace norm satisfies the triangle inequality and hence
is indeed a norm.

Finally, note that

1/2
kKk2 = tr(K ∗ K) (3.65)
which shows that J2 (H) is in fact a Hilbert space with scalar product given
by
hK1 , K2 i = tr(K1∗ K2 ). (3.66)
Problem 3.18. Let H := `2 (N) and let A be multiplication by a sequence
a = (aj )∞ 2
j=1 . Show that A is Hilbert–Schmidt if and only if a ∈ ` (N).
Furthermore, show that kAk2 = kak in this case.
Problem 3.19. An operator of the form K : `2 (N) → `2 (N), fn 7→ j∈N kn+j fj
P
is called Hankel operator.
• Show that K is Hilbert–Schmidt if and only if j∈N j|kj+1 |2 < ∞
P

and this number equals kKk2 .

• Show that K is Hilbert–Schmidt with kKk2 ≤ kck1 if |kj | ≤ cj ,
where cj is decreasing and summable.
(Hint: For the first item use summation by parts.)
 Chapter 4

The main theorems

about Banach spaces

Despite the many advantages of Hilbert spaces, there are also situations
where a non-Hilbert space is better suited (in fact the choice of the right
space is typically crucial for many problems). Hence we will devote our
attention to Banach spaces next.

4.1. The Baire theorem and its consequences

Recall that the interior of a set is the largest open subset (that is, the union
of all open subsets). A set is called nowhere dense if its closure has empty
interior. The key to several important theorems about Banach spaces is the
observation that a Banach space cannot be the countable union of nowhere
dense sets.

Theorem 4.1 (Baire category theorem). Let X be a (nonempty) complete

metric space. Then X cannot be the countable union of nowhere dense sets.

Proof. Suppose X = ∞ n=1 Xn . We can assume that the sets Xn are closed
S
and none of them contains a ball; that is, X \ Xn is open and nonempty for
every n. We will construct a Cauchy sequence xn which stays away from all
Xn .
Since X \ X1 is open and nonempty, there is a ball Br1 (x1 ) ⊆ X \ X1 .
Reducing r1 a little, we can even assume Br1 (x1 ) ⊆ X \ X1 . Moreover,
since X2 cannot contain Br1 (x1 ), there is some x2 ∈ Br1 (x1 ) that is not
in X2 . Since Br1 (x1 ) ∩ (X \ X2 ) is open, there is a closed ball Br2 (x2 ) ⊆
Br1 (x1 ) ∩ (X \ X2 ). Proceeding recursively, we obtain a sequence (here we

97
98 4. The main theorems about Banach spaces

use the axion of choice) of balls such that

Brn (xn ) ⊆ Brn−1 (xn−1 ) ∩ (X \ Xn ).
Now observe that in every step we can choose rn as small as we please; hence
without loss of generality rn → 0. Since by construction xn ∈ BrN (xN ) for
n ≥ N , we conclude that xn is Cauchy and converges to some point x ∈ X.
But x ∈ Brn (xn ) ⊆ X \ Xn for every n, contradicting our assumption that
the Xn cover X. 

In other words, if Xn ⊆ X is a sequence of closed subsets which cover

X, at least one Xn contains a ball of radius ε > 0.
Example 4.1. The set of rational numbers Q can be written as a count-
able union of its elements. This shows that the completeness assumption is
crucial. 
Remark: Sets which can be written as the countable union of nowhere
dense sets are said to be of first category or meager (also meagre). All
other sets are second category or fat (also residual). Hence explaining
the name category theorem.
Since a closed set is nowhere dense if and only if its complement is open
and dense (cf. Problem B.7), there is a reformulation which is also worthwhile
noting:
Corollary 4.2. Let X be a complete metric space. Then any countable
intersection of open dense sets is again dense.

Proof. Let {On } be a family of open dense sets whose intersection is not
dense. ThenS this intersection must be missing some closed ball Bε . This ball
will lie in n Xn , where Xn := X \ On are closed and nowhere dense. Now
note that X̃n := Xn ∩ Bε are closed nowhere dense sets in Bε . But Bε is a
complete metric space, a contradiction. 

Countable intersections of open sets are in some sense the next general
sets after open sets and are called Gδ sets (here G and δ stand for the German
words Gebiet and Durchschnitt, respectively). The complement of a Gδ set is
a countable union of closed sets also known as an Fσ set (here F and σ stand
for the French words fermé and somme, respectively). The complement of
a dense Gδ set will be a countable union of nowhere dense sets and hence
by definition meager. Consequently properties which hold on a dense Gδ are
considered generic in this context.
Example 4.2. The irrational numbers are a dense Gδ set in R. To see
this, let xn be an enumeration of the rational numbers and consider the
intersection of the open sets On := R \ {xn }. The rational numbers are
hence an Fσ set. 
4.1. The Baire theorem and its consequences 99

Now we are ready for the first important consequence:

Theorem 4.3 (Banach–Steinhaus). Let X be a Banach space and Y some

normed vector space. Let {Aα } ⊆ L (X, Y ) be a family of bounded operators.
Then
• either {Aα } is uniformly bounded, kAα k ≤ C,
• or the set {x ∈ X| supα kAα xk = ∞} is a dense Gδ .

Proof. Consider the sets

[
On := {x| kAα xk > n for some α} = {x| kAα xk > n}, n ∈ N.
α

By continuity of Aα and the norm, each On is a union of open sets and hence
open. Now either all of these sets are dense and hence their intersection
\
On = {x| sup kAα xk = ∞}
α
n∈N

is a dense Gδ by Corollary 4.2. Otherwise, X \ On is nonempty and open

for one n and we can find a ball of positive radius Bε (x0 ) ⊂ X \ On . Now
observe
kAα yk = kAα (y + x0 − x0 )k ≤ kAα (y + x0 )k + kAα x0 k ≤ 2n
for kyk ≤ ε. Setting y = ε kxk
x
, we obtain

2n
kAα xk ≤ kxk
ε
for every x. 

Note that there is also a variant of the Banach–Steinhaus theorem for

pointwise limits of bounded operators which will be discussed in Lemma 4.33.
Hence there are two ways to use this theorem by excluding one of the two
possible options. Showing that the pointwise bound holds on a sufficiently
large set (e.g. a ball), thereby ruling out the second option, implies a uniform
bound and is known as the uniform boundedness principle.

Corollary 4.4. Let X be a Banach space and Y some normed vector space.
Let {Aα } ⊆ L (X, Y ) be a family of bounded operators. Suppose kAα xk ≤
C(x) is bounded for every fixed x ∈ X. Then {Aα } is uniformly bounded,
kAα k ≤ C.

Conversely, if there is no uniform bound, the pointwise bound must fail

on a dense Gδ . This is illustrated in the following example.
100 4. The main theorems about Banach spaces

Example 4.3. Consider the Fourier series (2.44) of a continuous periodic

function f ∈ Cper [−π, π] = {f ∈ C[−π, π]|f (−π) = f (π)}. (Note that this
is a closed subspace of C[−π, π] and hence a Banach space — it is the kernel
of the linear functional `(f ) = f (−π) − f (π).) We want to show that for
every fixed x ∈ [−π, π] there is a dense Gδ set of functions in Cper [−π, π] for
which the Fourier series will diverge at x (it will even be unbounded).
Without loss of generality we fix x = 0 as our point. Then the n’th
partial sum gives rise to the linear functional
Z π
1
`n (f ) := Sn (f )(0) = Dn (x)f (x)dx
2π −π
and it suffices to show that the family {`n }n∈N is not uniformly bounded.
By Example 1.21 (adapted to our present periodic setting) we have
1
k`n k = kDn k1 .
2π
Now we estimate
Z π π
| sin((n + 1/2)x)|
Z
kDn k1 = 2 |Dn (x)|dx ≥ 2 dx
0 0 x/2
Z (n+1/2)π n Z kπ n
dy X dy 8X1
=4 | sin(y)| ≥4 | sin(y)| =
0 y (k−1)π kπ π k
k=1 k=1

and note that the harmonic series diverges.

In fact, we can even do better. Let G(x) ⊂ Cper [−π, π] be the dense
Gδ of functions whose Fourier series divergesTat x. Then, given countably
many points {xj }j∈N ⊂ [−π, π], the set G = j∈N G(xj ) is still a dense Gδ
by Corollary 4.2. Hence there is a dense Gδ of functions whose Fourier series
diverges on a given countable set of points. 
Example 4.4. Recall that the Fourier coefficients of an absolutely continu-
ous function satisfy the estimate
(
kf k∞ , k = 0,
|fˆk | ≤ kf 0 k∞
|k| , k 6= 0.

This raises the question if a similar estimate can be true for continuous
functions. More precisely, can we find a sequence ck > 0 such that
|fˆk | ≤ Cf ck ,
where Cf is some constant depending on f . If this were true, the linear
functionals
fˆk
`k (f ) := , k ∈ Z,
ck
4.1. The Baire theorem and its consequences 101

satisfy the assumptions of the uniform boundedness principle implying k`k k ≤

C. In other words, we must have an estimate of the type
|fˆk | ≤ Ckf k∞ ck
which implies 1 ≤ C ck upon choosing f (x) = eikx . Hence our assumption
cannot hold for any sequence ck converging to zero and there is no universal
decay rate for the Fourier coefficients of continuous functions beyond the fact
that they must converge to zero by the Riemann–Lebesgue lemma. 
The next application is

Theorem 4.5 (Open mapping). Let A ∈ L (X, Y ) be a continuous linear

operator between Banach spaces. Then A is open (i.e., maps open sets to
open sets) if and only if it is onto.

Proof. Set BrX := BrX (0) and similarly for BrY (0). By translating balls
(using linearity of A), it suffices to prove that for every ε > 0 there is a δ > 0
such that BδY ⊆ A(BεX ).
So let ε > 0 be given. Since A is surjective we have
∞
[ ∞
[ ∞
[
Y = AX = A nBεX = A(nBεX ) = nA(BεX )
n=1 n=1 n=1

and the Baire theorem implies that for some n, nA(BεX ) contains a ball.
Since multiplication by n is a homeomorphism, the same must be true for
n = 1, that is, BδY (y) ⊂ A(BεX ). Consequently

BδY ⊆ −y + A(BεX ) ⊂ A(BεX ) + A(BεX ) ⊆ A(BεX ) + A(BεX ) ⊆ A(B2ε

X ).

So it remains to get rid of the closure. To this end choose εn > 0 such that
P∞
n=1 εn < ε and corresponding δn → 0 such that Bδn ⊂ A(Bεn ). Now
Y X

for y ∈ Bδ1 ⊂ A(BεX1 ) we have x1 ∈ Bε1 such that Ax1 is arbitrarily close
Y X

to y, say y − Ax1 ∈ BδY2 ⊂ A(BεX2 ). Hence we can find x2 ∈ BεX2 such

that (y − Ax1 ) − Ax2 ∈ BδY3 ⊂ A(BεX3 ) and proceeding like this a sequence
xn ∈ BεXn such that
X n
y− Axk ∈ BδYn+1 .
k=1

By construction the limit x := ∞ k=1 xk exists and satisfies x ∈ BεX as well

P
as y = Ax ∈ A(Bε ). That is, Bδ1 ⊆ A(BεX ) as desired.
X Y

Conversely, if A is open, then the image of the unit ball contains again
some ball BεY ⊆ A(B1X ). Hence by scaling Brε
Y ⊆ A(B X ) and letting r → ∞
r
we see that A is onto: Y = A(X). 
102 4. The main theorems about Banach spaces

Example 4.5. However, note that, under the assumptions of the open map-
ping theorem, the image of a closed set might not be closed. For example,
consider the bounded linear operator A : `2 (N) → `2 (N), x 7→ (x2 , x4 , . . . )
which is clearly surjective. Then the image of the closed set U = {x ∈
`2 (N)|x2n = x2n−1 /n} is dense (it contains all sequences with finite sup-
port) but not all of `2 (N) (e.g. yn = n1 is missing since this would imply
x2n = 1). 
As an immediate consequence we get the inverse mapping theorem:
Theorem 4.6 (Inverse mapping). Let A ∈ L (X, Y ) be a continuous linear
bijection between Banach spaces. Then A−1 is continuous.
Example 4.6. Consider the operator (Aa)nj=1 = ( 1j aj )nj=1 in `2 (N). Then its
inverse (A−1 a)nj=1 = (j aj )nj=1 is unbounded (show this!). This is in agree-
ment with our theorem since its range is dense (why?) but not all of `2 (N):
For example, (bj = 1j )∞
j=1 6∈ Ran(A) since b = Aa gives the contradiction
∞
X ∞
X ∞
X
∞= 1= |jbj |2 = |aj |2 < ∞.
j=1 j=1 j=1

This should also be compared with Corollary 4.9 below. 

Example 4.7. Consider the Fourier series (2.44) of an integrable function.
Using the inverse function theorem we can show that not every sequence
tending to 0 (which is a necessary condition according to the Riemann–
Lebesgue lemma) arises as the Fourier coefficients of an integrable function:
By the elementary estimate
1
kfˆk∞ ≤ kf k1
2π
we see that that the mapping F (f ) := fˆ continuously maps F : L1 (−π, π) →
c0 (Z) (the Banach space of sequences converging to 0). In fact, this estimate
holds for continuous functions and hence there is a unique continuous exten-
sion of F to all of L1 (−π, π) by Theorem 1.16. Moreover, it can be shown
that F is injective (for f ∈ L2 this follows from Theorem 2.18, the general
case f ∈ L1 see Example 3.8 from [47]). Now if F were onto, the inverse
mapping theorem would show that the inverse is also continuous, that is, we
would have an estimate kfˆk∞ ≥ Ckf k1 for some C > 0. However, consider-
ing the Dirichlet kernel Dn we have kD̂n k∞ = 1 but kDn k1 → ∞ as shown
in Example 4.3. 
Another important consequence is the closed graph theorem. The graph
of an operator A is just
Γ(A) := {(x, Ax)|x ∈ D(A)}. (4.1)
4.1. The Baire theorem and its consequences 103

If A is linear, the graph is a subspace of the Banach space X ⊕ Y (provided

X and Y are Banach spaces), which is just the Cartesian product together
with the norm
k(x, y)kX⊕Y := kxkX + kykY . (4.2)
Note that (xn , yn ) → (x, y) if and only if xn → x and yn → y. We say that
A has a closed graph if Γ(A) is a closed subset of X ⊕ Y .

Theorem 4.7 (Closed graph). Let A : X → Y be a linear map from a

Banach space X to another Banach space Y . Then A is continuous if and
only if its graph is closed.

Proof. If Γ(A) is closed, then it is again a Banach space. Now the projec-
tion π1 (x, Ax) = x onto the first component is a continuous bijection onto
X. So by the inverse mapping theorem its inverse π1−1 is again continuous.
Moreover, the projection π2 (x, Ax) = Ax onto the second component is also
continuous and consequently so is A = π2 ◦ π1−1 . The converse is easy. 

Remark: The crucial fact here is that A is defined on all of X!

Operators whose graphs are closed are called closed operators. Warn-
ing: By Example 4.5 a closed operator will not map closed sets to closed
sets in general. In particular, the concept of a closed operator should not be
confused with the concept of a closed map in topology!
Being closed is the next option you have once an operator turns out to
be unbounded. If A is closed, then xn → x does not guarantee you that
Axn converges (like continuity would), but it at least guarantees that if Axn
converges, it converges to the right thing, namely Ax:
• A bounded (with D(A) = X): xn → x implies Axn → Ax.
• A closed (with D(A) ⊆ X): xn → x, xn ∈ D(A), and Axn → y
implies x ∈ D(A) and y = Ax.
If an operator is not closed, you can try to take the closure of its graph,
to obtain a closed operator. If A is bounded this always works (which is
just the content of Theorem 1.16). However, in general, the closure of the
graph might not be the graph of an operator as we might pick up points
(x, y1 ), (x, y2 ) ∈ Γ(A) with y1 6= y2 . Since Γ(A) is a subspace, we also have
(x, y2 ) − (x, y1 ) = (0, y2 − y1 ) ∈ Γ(A) in this case and thus Γ(A) is the graph
of some operator if and only if

Γ(A) ∩ {(0, y)|y ∈ Y } = {(0, 0)}. (4.3)

If this is the case, A is called closable and the operator A associated with
Γ(A) is called the closure of A.
104 4. The main theorems about Banach spaces

In particular, A is closable if and only if xn → 0 and Axn → y implies

y = 0. In this case
D(A) = {x ∈ X|∃xn ∈ D(A), y ∈ Y : xn → x and Axn → y},
Ax = y. (4.4)
For yet another way of defining the closure see Problem 4.13.
Example 4.8. Consider the operator A in `p (N) defined by Aaj := jaj on
D(A) = {a ∈ `p (N)|aj 6= 0 for finitely many j}.
(i). A is closable. In fact, if an → 0 and Aan → b then we have anj → 0
and thus janj → 0 = bj for any j ∈ N.
(ii). The closure of A is given by
(
{a ∈ `p (N)|(jaj )∞ p
j=1 ∈ ` (N)}, 1 ≤ p < ∞,
D(A) =
{a ∈ c0 (N)|(jaj )∞
j=1 ∈ c0 (N)}, p = ∞,

and Aaj = jaj . In fact, if an → a and Aan → b then we have anj → aj and
janj → bj for any j ∈ N and thus bj = jaj for any j ∈ N. In particular,
j=1 = (bj )j=1 ∈ ` (N) (c0 (N) if p = ∞). Conversely, suppose (jaj )j=1 ∈
(jaj )∞ ∞ p ∞

` (N) (c0 (N) if p = ∞) and consider

p
(
n aj , j ≤ n,
aj :=
0, j > n.
Then an → a and Aan → (jaj )∞
j=1 .
(iii). Extending the basis vectors {δ n }n∈N to a Hamel basis (Problem 1.7)
and setting Aa = 0 for every other element from this Hamel basis we obtain
a (still unbounded) operator which is everywhere defined. However, this
extension cannot be closed! 
Example 4.9. Here is a simple P example of a nonclosable operator: Let
X := `2 (N) and consider Ba := ( ∞ j=1 aj )δ defined on ` (N) ⊂ ` (N). Let
1 1 2

anj := n1 for 1 ≤ j ≤ n and anj := 0 for j > n. Then kan k2 = √1n implying
an → 0 but Ban = δ 1 6→ 0. 
Example 4.10. Another example are point evaluations in L2 (0, 1): Let x0 ∈
[0, 1] and consider `x0 : D(`x0 ) → C, f 7→ f (x0 ) defined on D(`x0 ) :=
C[0, 1] ⊆ L2 (0, 1). Then fn (x) := max(0, 1 − n|x − x0 |) satisfies fn → 0 but
`x0 (fn ) = 1. In fact, a linear functional is closable if and only if it is bounded
(Problem 4.9). 
−1
Lemma 4.8. Suppose A is closable and A is injective. Then A = A−1 .

Proof. If we set
Γ−1 = {(y, x)|(x, y) ∈ Γ}
4.1. The Baire theorem and its consequences 105

then Γ(A−1 ) = Γ−1 (A) and

−1 −1
Γ(A−1 ) = Γ(A)−1 = Γ(A) = Γ(A)−1 = Γ(A ). 

Note that A injective does not imply A injective in general.

Example 4.11. Let PM be the projection in `2 (N) on M := {b}⊥ , where
j=1 . Explicitly we have PM a = a − hb, aib. Then PM restricted
b := (2−j/2 )∞
to the space of sequences with finitely many nonzero terms is injective, but
its closure is not. 
As a consequence of the closed graph theorem we obtain:
Corollary 4.9. Suppose A : D(A) ⊆ X → Y is closed and injective. Then
A−1 defined on D(A−1 ) = Ran(A) is closed. Moreover, in this case Ran(A)
is closed if and only if A−1 is bounded.

The question when Ran(A) is closed plays an important role when in-
vestigating solvability of the equation Ax = y and the last part gives us a
convenient criterion. Moreover, note that A−1 is bounded if and only if there
is some c > 0 such that
kAxk ≥ ckxk, x ∈ D(A). (4.5)
Indeed, this follows upon setting x = A−1 y in the above inequality which
also shows that c = kA−1 k−1 is the best possible constant. Factoring out
the kernel we even get a criterion for the general case:
Corollary 4.10. Suppose A : D(A) ⊆ X → Y is closed. Then Ran(A) is
closed if and only if
kAxk ≥ c dist(x, Ker(A)), x ∈ D(A), (4.6)
for some c > 0. The sup over all possible c is known as the (reduced) mini-
mum modulus of A.

Proof. Consider the quotient space X̃ := X/ Ker(A) and the induced op-
erator à : D(Ã) → Y where D(Ã) = D(A)/ Ker(A) ⊆ X̃. By construction
Ã[x] = 0 iff x ∈ Ker(A) and hence à is injective. To see that à is closed we
use π̃ : X × Y → X̃ × Y , (x, y) 7→ ([x], y) which is bounded, surjective and
hence open. Moreover, π̃(Γ(A)) = Γ(Ã). In fact, we even have (x, y) ∈ Γ(A)
iff ([x], y) ∈ Γ(Ã) and thus π̃(X × Y \ Γ(A)) = X̃ × Y \ Γ(Ã) implying that
Y \ Γ(Ã) is open. Finally, observing Ran(A) = Ran(Ã) we have reduced it
to the previous corollary. 

There is also another criterion which does not involve the distance to the
kernel.
106 4. The main theorems about Banach spaces

Corollary 4.11. Suppose A : D(A) ⊆ X → Y is closed. Then Ran(A)

is closed if for some given ε > 0 and 0 ≤ δ < 1 we can find for every
y ∈ Ran(A) a corresponding x ∈ D(X) such that
εkxk + ky − Axk ≤ δkyk. (4.7)
Conversely, if Ran(A) is closed this can be done whenever ε < cδ with c from
the previous corollary.

Proof. If Ran(A) is closed and ε < cδ there is some x ∈ D(A) with y = Ax

and kAxk ≥ δε kxk after maybe adding an element from the kernel to x. This
x satisfies εkxk + ky − Axk = εkxk ≤ δkyk as required.
Conversely, fix y ∈ Ran(A) and recursively choose a sequence xn such
that
X
εkxn k + k(y − Ax̃n−1 ) − Axn k ≤ δky − Ax̃n−1 k, x̃n := xm .
m≤n

In particular, ky − Ax̃n k ≤ as well as εkxn k ≤

δ n kyk which shows
δ n kyk,
x̃n → x and Ax̃n → y. Hence x ∈ D(A) and y = T x ∈ Ran(A). 

The closed graph theorem tells us that closed linear operators can be
defined on all of X if and only if they are bounded. So if we have an
unbounded operator we cannot have both! That is, if we want our operator
to be at least closed, we have to live with domains. This is the reason why
in quantum mechanics most operators are defined on domains. In fact, there
is another important property which does not allow unbounded operators to
be defined on the entire space:
Theorem 4.12 (Hellinger–Toeplitz). Let A : H → H be a linear operator on
some Hilbert space H. If A is symmetric, that is hg, Af i = hAg, f i, f, g ∈ H,
then A is bounded.

Proof. It suffices to prove that A is closed. In fact, fn → f and Afn → g

implies
hh, gi = lim hh, Afn i = lim hAh, fn i = hAh, f i = hh, Af i
n→∞ n→∞
for every h ∈ H. Hence Af = g. 
Problem 4.1. Every subset of a meager set is again meager.
Problem 4.2. An infinite dimensional Banach space cannot have a count-
able Hamel basis (see Problem 1.7). (Hint: Apply Baire’s theorem to Xn :=
span{uj }nj=1 .)
Problem 4.3. Let X := C[0, 1]. Show that the set of functions which are
nowhere differentiable contains a dense Gδ . (Hint: Consider Fk := {f ∈
X| ∃x ∈ [0, 1] : |f (x) − f (y)| ≤ k|x − y|, ∀y ∈ [0, 1]}. Show that this set is
4.1. The Baire theorem and its consequences 107

closed and nowhere dense. For the first property Bolzano–Weierstraß might
be useful, for the latter property show that the set of piecewise linear functions
whose slopes are bounded below by some fixed number in absolute value are
dense.)
Problem 4.4. Let X be a complete metric space without isolated points.
Show that a dense Gδ set cannot be countable. (Hint: A single point is
nowhere dense.)
Problem 4.5. Let X be the space of sequences with finitely many nonzero
terms together with the sup norm. Consider the family of operators {An }n∈N
given by (An a)j := jaj , j ≤ n and (An a)j := 0, j > n. Then this family
is pointwise bounded but not uniformly bounded. Does this contradict the
Banach–Steinhaus theorem?
Problem 4.6. Show that a bilinear map B : X × Y → Z is bounded,
kB(x, y)k ≤ Ckxkkyk, if and only if it is separately continuous with respect
to both arguments. (Hint: Uniform boundedness principle.)
Problem 4.7. Consider a Schauder basis as in (1.31). Show that the co-
ordinate functionals αn are continuous. (Hint: Denote the set of all pos-
of Schauder coefficients by A and equip it with the norm
sible sequences P
kαk := supn k nk=1 αk uk k; note that A is precisely the set of sequences
P this norm is finite. By construction the operator A : A → X,
for which
α 7→ k αk uk has norm one. Now show that A is complete and apply the
inverse mapping theorem.)
Problem 4.8. Show that a compact symmetric operator in an infinite-dimensional
Hilbert space cannot be surjective.
Problem* 4.9. A linear functional defined on a dense subspace is closable
if and only if it is bounded.
Problem 4.10. Show that if A is a closed and B a bounded operator, then
A + B is closed. Show that this in general fails if B is not bounded. (Here
A + B is defined on D(A + B) = D(A) ∩ D(B).)
Problem 4.11. Suppose B ∈ L (X, Y ) is bounded and A : D(A) ⊆ Y → Z
is closed. Then AB : D(AB) ⊆ X → Z is closed, where D(AB) := {x ∈
D(B)|Ax ∈ D(A)}.
d
Problem 4.12. Show that the differential operator A = dx defined on
1
D(A) = C [0, 1] ⊂ C[0, 1] (sup norm) is a closed operator. (Compare the
example in Section 1.6.)
Problem* 4.13. Consider a linear operator A : D(A) ⊆ X → Y , where X
and Y are Banach spaces. Define the graph norm associated with A by
kxkA := kxkX + kAxkY , x ∈ D(A). (4.8)
108 4. The main theorems about Banach spaces

Show that A : D(A) → Y is bounded if we equip D(A) with the graph norm.
Show that the completion XA of (D(A), k.kA ) can be regarded as a subset of
X if and only if A is closable. Show that in this case the completion can
be identified with D(A) and that the closure of A in X coincides with the
extension from Theorem 1.16 of A in XA . In particular, A is closed if and
only if (D(A), k.kA ) is complete.
Problem 4.14. Let X := `2 (N) P and (Aa)j := j aj with D(A) := {a ∈ `2 (N)|
(jaj )j∈N ∈ ` (N)} and Ba := ( j∈N aj )δ 1 . Then we have seen that A is
2

closed while B is not closable. Show that A+B, D(A+B) = D(A)∩D(B) =

D(A) is closed.

4.2. The Hahn–Banach theorem and its consequences

Let X be a Banach space. Recall that we have called the set of all bounded
linear functionals the dual space X ∗ (which is again a Banach space by
Theorem 1.17).
Example 4.12. Consider the Banach space `p (N), 1 ≤ p < ∞. Taking
the Kronecker deltas δ n as a Schauder basis the n’th term xn of a sequence
x ∈ `p (N) can also be considered as the n’th coordinate of x with respect to
this basis. Moreover, the map ln (x) = xn is a bounded linear functional, that
is, ln ∈ `p (N)∗ , since |ln (x)| = |xn | ≤ kxkp . It is a special case of the following
more general example (in fact, we have ln = lδn ). Since the coordinates of
a vector carry all the information this explains why understanding linear
functionals if of key importance. 
Example 4.13. Consider the Banach space `p (N), 1 ≤ p < ∞. We have
already seen that by Hölder’s inequality (1.25) every y ∈ `q (N) gives rise to
a bounded linear functional
X
ly (x) := yn xn (4.9)
n∈N
whose norm is kly k = kykq (Problem 4.20). But can every element of `p (N)∗
be written in this form?
Suppose p := 1 and choose l ∈ `1 (N)∗ . Now define
yn := l(δ n ).
Then
|yn | = |l(δ n )| ≤ klk kδ n k1 = klk
shows kyk∞ ≤ klk, that is, y ∈ `∞ (N). By construction l(x) = ly (x) for every
x ∈ span{δ n }. By continuity of l it even holds for x ∈ span{δ n } = `1 (N).
Hence the map y 7→ ly is an isomorphism, that is, `1 (N)∗ ∼ = `∞ (N). A similar
p ∗ ∼
argument shows ` (N) = ` (N), 1 ≤ p < ∞ (Problem 4.21). One usually
q

identifies `p (N)∗ with `q (N) using this canonical isomorphism and simply
4.2. The Hahn–Banach theorem and its consequences 109

writes `p (N)∗ = `q (N). In the case p = ∞ this is not true, as we will see
soon. 
It turns out that many questions are easier to handle after applying a
linear functional ` ∈ X ∗ . For example, suppose x(t) is a function R → X
(or C → X), then `(x(t)) is a function R → C (respectively C → C) for
any ` ∈ X ∗ . So to investigate `(x(t)) we have all tools from real/complex
analysis at our disposal. But how do we translate this information back to
x(t)? Suppose we have `(x(t)) = `(y(t)) for all ` ∈ X ∗ . Can we conclude
x(t) = y(t)? The answer is yes and will follow from the Hahn–Banach
theorem.
We first prove the real version from which the complex one then follows
easily.
Theorem 4.13 (Hahn–Banach, real version). Let X be a real vector space
and ϕ : X → R a convex function (i.e., ϕ(λx+(1−λ)y) ≤ λϕ(x)+(1−λ)ϕ(y)
for λ ∈ (0, 1)).
If ` is a linear functional defined on some subspace Y ⊂ X which satisfies
`(y) ≤ ϕ(y), y ∈ Y , then there is an extension ` to all of X satisfying
`(x) ≤ ϕ(x), x ∈ X.

Proof. Let us first try to extend ` in just one direction: Take x 6∈ Y and
set Ỹ = span{x, Y }. If there is an extension `˜ to Ỹ it must clearly satisfy
˜ + αx) = `(y) + α`(x).
`(y ˜
˜
So all we need to do is to choose `(x) ˜ + αx) ≤ ϕ(y + αx). But
such that `(y
this is equivalent to
ϕ(y − αx) − `(y) ˜ ≤ inf ϕ(y + αx) − `(y)
sup ≤ `(x)
α>0,y∈Y −α α>0,y∈Y α
and is hence only possible if
ϕ(y1 − α1 x) − `(y1 ) ϕ(y2 + α2 x) − `(y2 )
≤
−α1 α2
for every α1 , α2 > 0 and y1 , y2 ∈ Y . Rearranging this last equations we see
that we need to show
α2 `(y1 ) + α1 `(y2 ) ≤ α2 ϕ(y1 − α1 x) + α1 ϕ(y2 + α2 x).
Starting with the left-hand side we have
α2 `(y1 ) + α1 `(y2 ) = (α1 + α2 )` (λy1 + (1 − λ)y2 )
≤ (α1 + α2 )ϕ (λy1 + (1 − λ)y2 )
= (α1 + α2 )ϕ (λ(y1 − α1 x) + (1 − λ)(y2 + α2 x))
≤ α2 ϕ(y1 − α1 x) + α1 ϕ(y2 + α2 x),
110 4. The main theorems about Banach spaces

where λ = α1 +α2 .
α2
Hence one dimension works.
To finish the proof we appeal to Zorn’s lemma (see Appendix A): Let E
be the collection of all extensions `˜ satisfying `(x)
˜ ≤ ϕ(x). Then E can be
partially ordered by inclusion (with respect to the domain) and every linear
chain has an upper bound (defined on the union of all domains). Hence there
is a maximal element ` by Zorn’s lemma. This element is defined on X, since
if it were not, we could extend it as before contradicting maximality. 

Note that linearity gives us a corresponding lower bound −ϕ(−x) ≤ `(x),

x ∈ X, for free. In particular, if ϕ(x) = ϕ(−x) then |`(x)| ≤ ϕ(x).
Theorem 4.14 (Hahn–Banach, complex version). Let X be a complex vector
space and ϕ : X → R a convex function satisfying ϕ(αx) ≤ ϕ(x) if |α| = 1.
If ` is a linear functional defined on some subspace Y ⊂ X which satisfies
|`(y)| ≤ ϕ(y), y ∈ Y , then there is an extension ` to all of X satisfying
|`(x)| ≤ ϕ(x), x ∈ X.

Proof. Set `r = Re(`) and observe

`(x) = `r (x) − i`r (ix).
By our previous theorem, there is a real linear extension `r satisfying `r (x) ≤
ϕ(x). Now set `(x) = `r (x) − i`r (ix). Then `(x) is real linear and by
`(ix) = `r (ix) + i`r (x) = i`(x) also complex linear. To show |`(x)| ≤ ϕ(x)
`(x)∗
we abbreviate α = |`(x)|
and use

|`(x)| = α`(x) = `(αx) = `r (αx) ≤ ϕ(αx) ≤ ϕ(x),

which finishes the proof. 

Note that ϕ(αx) ≤ ϕ(x), |α| = 1 is in fact equivalent to ϕ(αx) = ϕ(x),

|α| = 1.
If ` is a bounded linear functional defined on some subspace, the choice
ϕ(x) = k`kkxk implies:
Corollary 4.15. Let X be a normed space and let ` be a bounded linear
functional defined on some subspace Y ⊆ X. Then there is an extension
` ∈ X ∗ preserving the norm.
Example 4.14. Note that in a Hilbert space this result is trivial: First of
all there is a unique extension to Y by Theorem 1.16. Now set `¯ = 0 on Y ⊥ .
Moreover, any other extension is of the form `¯+ `1 , where `1 vanishes on Y .
Then k`¯ + `1 k2 = k`k2 + k`1 k2 and the norm will increase unless `1 = 0. 
Example 4.15. In a Banach space this extension will in general not be
unique: Consider X := `1 (N) and `(x) := x1 on Y := span{δ 1 }. Then by
4.2. The Hahn–Banach theorem and its consequences 111

Example 4.13 any extension is of the form `¯ = ly with y ∈ `∞ (N) and y1 = 1,

kyk∞ ≤ 1. (Sometimes it still might be unique: Problems 4.15 and 4.16). 
Moreover, we can now easily prove our anticipated result

Corollary 4.16. Let X be a normed space and x ∈ X fixed. Suppose `(x) =

0 for all ` in some total subset Y ⊆ X ∗ . Then x = 0.

Proof. Clearly, if `(x) = 0 holds for all ` in some total subset, this holds
for all ` ∈ X ∗ . If x 6= 0 we can construct a bounded linear functional on
span{x} by setting `(αx) = α and extending it to X ∗ using the previous
corollary. But this contradicts our assumption. 

Example 4.16. Let us return to our example `∞ (N). Let c(N) ⊂ `∞ (N) be
the subspace of convergent sequences. Set
l(x) = lim xn , x ∈ c(N), (4.10)
n→∞

then l is bounded since

|l(x)| = lim |xn | ≤ kxk∞ . (4.11)
n→∞

Hence we can extend it to `∞ (N) by Corollary 4.15. Then l(x) cannot be

written as l(x) = ly (x) for some y ∈ `1 (N) (as in (4.9)) since yn = l(δ n ) = 0
shows y = 0 and hence `y = 0. The problem is that span{δ n } = c0 (N) 6=
`∞ (N), where c0 (N) is the subspace of sequences converging to 0.
Moreover, there is also no other way to identify `∞ (N)∗ with `1 (N), since
is separable whereas `∞ (N) is not. This will follow from Lemma 4.21 (iii)
`1 (N)
below. 
Another useful consequence is

Corollary 4.17. Let Y ⊆ X be a subspace of a normed vector space and let

x0 ∈ X \ Y . Then there exists an ` ∈ X ∗ such that (i) `(y) = 0, y ∈ Y , (ii)
`(x0 ) = dist(x0 , Y ), and (iii) k`k = 1.

Proof. Replacing Y by Y we see that it is no restriction to assume that

Y is closed. (Note that x0 ∈ X \ Y if and only if dist(x0 , Y ) > 0.) Let
Ỹ = span{x0 , Y }. Since every element of Ỹ can be uniquely written as
y + αx0 we can define
`(y + αx0 ) = α dist(x0 , Y ).
By construction ` is linear on Ỹ and satisfies (i) and (ii). Moreover, by
dist(x0 , Y ) ≤ kx0 − −y
α k for every y ∈ Y we have

|`(y + αx0 )| = |α| dist(x0 , Y ) ≤ ky + αx0 k, y ∈ Y.

112 4. The main theorems about Banach spaces

Hence k`k ≤ 1 and there is an extension to X by Corollary 4.15. To see

that the norm is in fact equal to one, take a sequence yn ∈ Y such that
dist(x0 , Y ) ≥ (1 − n1 )kx0 + yn k. Then
1
|`(yn + x0 )| = dist(x0 , Y ) ≥ (1 − )kyn + x0 k
n
establishing (iii). 

Two more straightforward consequences of the last corollary are also

worthwhile noting:

Corollary 4.18. Let Y ⊆ X be a subspace of a normed vector space. Then

x ∈ Y if and only if `(x) = 0 for every ` ∈ X ∗ which vanishes on Y .

Corollary 4.19. Let Y be a closed subspace and let {xj }nj=1 be a linearly
independent subset of X. If Y ∩ span{xj }nj=1 = {0}, then there exists a
biorthogonal system {`j }nj=1 ⊂ X ∗ such that `j (xk ) = 0 for j 6= k,
`j (xj ) = 1 and `(y) = 0 for y ∈ Y .

Proof. Fix j0 . Since Yj0 = Y uspan{xj }1≤j≤n;j6=j0 is closed (Corollary 1.19),

xj0 6∈ Yj0 implies dist(xj0 , Yj0 ) > 0 and existence of `j0 follows from Corol-
lary 4.17. 

If we take the bidual (or double dual) X ∗∗ of a normed space X,

then the Hahn–Banach theorem tells us, that X can be identified with a
subspace of X ∗∗ . In fact, consider the linear map J : X → X ∗∗ defined by
J(x)(`) = `(x) (i.e., J(x) is evaluation at x). Then

Theorem 4.20. Let X be a normed space. Then J : X → X ∗∗ is isometric

(norm preserving).

Proof. Fix x0 ∈ X. By |J(x0 )(`)| = |`(x0 )| ≤ k`k∗ kx0 k we have at least

kJ(x0 )k∗∗ ≤ kx0 k. Next, by Hahn–Banach there is a linear functional `0 with
norm k`0 k∗ = 1 such that `0 (x0 ) = kx0 k. Hence |J(x0 )(`0 )| = |`0 (x0 )| =
kx0 k shows kJ(x0 )k∗∗ = kx0 k. 

Example 4.17. This gives another quick way of showing that a normed
space has a completion: Take X̄ = J(X) ⊆ X ∗∗ and recall that a dual space
is always complete (Theorem 1.17). 
Thus J : X → X ∗∗ is an isometric embedding. In many cases we even
have J(X) = X ∗∗ and X is called reflexive in this case.
Example 4.18. The Banach spaces `p (N) with 1 < p < ∞ are reflexive:
Identify `p (N)∗ with `q (N) (cf. Problem 4.21) and choose z ∈ `p (N)∗∗ . Then
4.2. The Hahn–Banach theorem and its consequences 113

there is some x ∈ `p (N) such that

y ∈ `q (N) ∼
X
z(y) = yj xj , = `p (N)∗ .
j∈N

But this implies z(y) = y(x), that is, z = J(x), and thus J is surjective.
(Warning: It does not suffice to just argue `p (N)∗∗ ∼
= `q (N)∗ ∼
= `p (N).)
However, `1 is not reflexive since `1 (N)∗ ∼= `∞ (N) but `∞ (N)∗ ∼ 6 `1 (N)
=
as noted earlier. Things get even a bit more explicit if we look at c0 (N),
where we can identify (cf. Problem 4.22) c0 (N)∗ with `1 (N) and c0 (N)∗∗ with
`∞ (N). Under this identification J(c0 (N)) = c0 (N) ⊆ `∞ (N). 
Example 4.19. By the same argument, every Hilbert space is reflexive. In
fact, by the Riesz lemma we can identify H∗ with H via the (conjugate linear)
map x 7→ hx, .i. Taking z ∈ H∗∗ we have, again by the Riesz lemma, that
z(y) = hhx, .i, hy, .iiH∗ = hx, yi∗ = hy, xi = J(x)(y). 
Lemma 4.21. Let X be a Banach space.
(i) If X is reflexive, so is every closed subspace.
(ii) X is reflexive if and only if X ∗ is.
(iii) If X ∗ is separable, so is X.

Proof. (i) Let Y be a closed subspace. Denote by j : Y ,→ X the natural

inclusion and define j∗∗ : Y ∗∗ → X ∗∗ via (j∗∗ (y 00 ))(`) = y 00 (`|Y ) for y 00 ∈ Y ∗∗
and ` ∈ X ∗ . Note that j∗∗ is isometric by Corollary 4.15. Then
J
X
X −→ X ∗∗
j↑ ↑ j∗∗
Y −→ Y ∗∗
JY

commutes. In fact, we have j∗∗ (JY (y))(`) = JY (y)(`|Y ) = `(y) = JX (y)(`).

Moreover, since JX is surjective, for every y 00 ∈ Y ∗∗ there is an x ∈ X such
that j∗∗ (y 00 ) = JX (x). Since j∗∗ (y 00 )(`) = y 00 (`|Y ) vanishes on all ` ∈ X ∗
which vanish on Y , so does `(x) = JX (x)(`) = j∗∗ (y 00 )(`) and thus x ∈ Y
by Corollary 4.18. That is, j∗∗ (Y ∗∗ ) = JX (Y ) and JY = j ◦ JX ◦ j∗∗ −1 is

surjective.
(ii) Suppose X is reflexive. Then the two maps
(JX )∗ : X ∗ → X ∗∗∗ (JX )∗ : X ∗∗∗ → X ∗
−1
x0 7→ x0 ◦ JX x000 7→ x000 ◦ JX
−1 00
are inverse of each other. Moreover, fix x00 ∈ X ∗∗ and let x = JX (x ).
−1 00
Then JX ∗ (x )(x ) = x (x ) = J(x)(x ) = x (x) = x (JX (x )), that is JX ∗ =
0 00 00 0 0 0 0

(JX )∗ respectively (JX ∗ )−1 = (JX )∗ , which shows X ∗ reflexive if X reflexive.

To see the converse, observe that X ∗ reflexive implies X ∗∗ reflexive and hence
JX (X) ∼= X is reflexive by (i).
114 4. The main theorems about Banach spaces

(iii) Let {`n }∞

n=1 be a dense set in X . Then we can choose xn ∈ X such
∗

that kxn k = 1 and `n (xn ) ≥ k`n k/2. We will show that {xn }∞ n=1 is total in
X. If it were not, we could find some x ∈ X \ span{xn }n=1 and hence there
∞

is a functional ` ∈ X ∗ as in Corollary 4.17. Choose a subsequence `nk → `.

Then
k` − `nk k ≥ |(` − `nk )(xnk )| = |`nk (xnk )| ≥ k`nk k/2,
which implies `nk → 0 and contradicts k`k = 1. 

If X is reflexive, then the converse of (iii) is also true (since X ∼

= X ∗∗
separable implies X separable), but in general this fails as the example
∗

`1 (N)∗ ∼= `∞ (N) shows. In fact, this can be used to show that a separable
space is not reflexive, by showing that its dual is not separable.
Example 4.20. The space C(I) is not reflexive. To see this observe that
the dual space contains point evaluations `x0 (f ) := f (x0 ), x0 ∈ I. Moreover,
for x0 6= x1 we have k`x0 − `x1 k = 2 and hence C(I)∗ is not separable. You
should appreciate the fact that it was not necessary to know the full dual
space which is quite intricate (see Theorem 6.5 from [47]). 
Note that the product of two reflexive spaces is also reflexive. In fact,
this even holds for countable products — Problem 4.24.
Problem 4.15. Let X := C3 equipped with the norm |(x, y, z)|1 := |x| +
|y| + |z| and Y := {(x, y, z)|x + y = 0, z = 0}. Find at least two extensions
of `(x, y, z) := x from Y to X which preserve the norm. What if we take
Y := {(x, y, z)|x + y = 0}?
Problem 4.16. Show that the extension from Corollary 4.15 is unique if X ∗
is strictly convex. (Hint: Problem 1.13.)
Problem* 4.17. Let X be some normed space. Show that
kxk = sup |`(x)|, (4.12)
`∈V, k`k=1

where V ⊂ X ∗ is some dense subspace. Show that equality is attained if

V = X ∗.
Problem 4.18. Let X be some normed space. By definition we have
k`k = sup |`(x)|
x∈X,kxk=1

for every ` ∈ X ∗ . One calls ` ∈ X ∗ norm-attaining, if the supremum is

attained, that is, there is some x ∈ X such that k`k = |`(x)|.
Show that in a reflexive Banach space every linear functional is norm-
attaining. Give an example of a linear functional which is not norm-attaining.
For uniqueness see Problem 5.35. (Hint: For the first part apply the previous
problem to X ∗ . For the second part consider Problem 4.22 below.)
4.2. The Hahn–Banach theorem and its consequences 115

Problem 4.19. Let X, Y be some normed spaces and A : D(A) ⊆ X → Y .

Show
kAk = sup |`(Ax)|, (4.13)
x∈X, kxk=1; `∈V, k`k=1

where V ⊂ Y ∗ is a dense subspace.

Problem* 4.20. Show that kly k = kykq , where ly ∈ `p (N)∗ as defined in
(4.9). (Hint: Choose x ∈ `p such that xn yn = |yn |q .)
Problem* 4.21. Show that every l ∈ `p (N)∗ , 1 ≤ p < ∞, can be written as
X
l(x) = yn x n
n∈N

with some y ∈ `q (N). (Hint: To see y ∈ `q (N) consider xN defined such

that xN q N
n = |yn | /yn for n ≤ N with yn 6= 0 and xn = 0 else. Now look at
|l(xN )| ≤ klkkxN kp .)
Problem* 4.22. Let c0 (N) ⊂ `∞ (N) be the subspace of sequences which
converge to 0, and c(N) ⊂ `∞ (N) the subspace of convergent sequences.
(i) Show that c0 (N), c(N) are both Banach spaces and that c(N) =
span{c0 (N), e}, where e = (1, 1, 1, . . . ) ∈ c(N).
(ii) Show that every l ∈ c0 (N)∗ can be written as
X
l(a) = bn an
n∈N

with some b ∈ `1 (N) which satisfies kbk1 = k`k.

(iii) Show that every l ∈ c(N)∗ can be written as
X
l(a) = bn an + b0 lim an
n→∞
n∈N

with some b ∈ `1 (N) which satisfies |b0 | + kbk1 = k`k.

Problem 4.23. Let un ∈ X be a Schauder basis and suppose the complex
numbers cn satisfy |cn | ≤ ckun k. Is there a bounded linear functional ` ∈ X ∗
with `(un ) = cn ? (Hint: Consider e.g. X = `2 (Z).)
L
Problem* 4.24. Let X := p,j∈N Xj be defined as in Problem 1.39 and let
1 1 ∗ ∼ ∗
L
p + q = 1. Show that for 1 ≤ p < ∞ we have X = q,j∈N Xj , where the
identification is given by
X M M
y(x) = yj (xj ), x = (xj )j∈N ∈ Xj , y = (yj )j∈N ∈ Xj∗ .
p,j∈N q,j∈N
j∈N

Moreover, if all Xj are reflexive, so is X for 1 < p < ∞.

116 4. The main theorems about Banach spaces

Problem 4.25 (Banach limit). Let c(N) ⊂ `∞ (N) be the subspace of all
bounded sequences for which the limit of the Cesàro means
n
1X
L(x) := lim xk
n→∞ n
k=1
exists. Note that c(N) ⊆ c(N) and L(x) = limn→∞ xn for x ∈ c(N).
Show that L can be extended to all of `∞ (N) such that
(i) L is linear,
(ii) |L(x)| ≤ kxk∞ ,
(iii) L(Sx) = L(x) where (Sx)n = xn+1 is the shift operator,
(iv) L(x) ≥ 0 when xn ≥ 0 for all n,
(v) lim inf n xn ≤ L(x) ≤ lim sup xn for all real-valued sequences.
(Hint: Of course existence follows from Hahn–Banach and (i), (ii) will come
for free. Also (iii) will be inherited from the construction. For (iv) note
that the extension can assumed to be real-valued and investigate L(e − x) for
x ≥ 0 with kxk∞ = 1 where e = (1, 1, 1, . . . ). (v) then follows from (iv).)
Problem* 4.26. Show that a finite dimensional subspace M of a Banach
space X can be complemented. (Hint: Start with a basis {xj } for M and
choose a corresponding dual basis {`k } with `k (xj ) = δj,k which can be ex-
tended to X ∗ .)

4.3. The adjoint operator

Given two normed spaces X and Y and a bounded operator A ∈ L (X, Y )
we can define its adjoint A0 : Y ∗ → X ∗ via A0 y 0 = y 0 ◦ A, y 0 ∈ Y ∗ . It is
immediate that A0 is linear and boundedness follows from
!
kA0 k = sup kA0 y 0 k = sup sup |(A0 y 0 )(x)|
y 0 ∈Y ∗ : ky 0 k=1 y 0 ∈Y ∗ : ky 0 k=1 x∈X: kxk=1
!
= sup sup |y 0 (Ax)| = sup kAxk = kAk,
y 0 ∈Y ∗ : ky 0 k=1 x∈X: kxk=1 x∈X: kxk=1

where we have used Problem 4.17 to obtain the fourth equality. In summary,
Theorem 4.22. Let A ∈ L (X, Y ), then A0 ∈ L (Y ∗ , X ∗ ) with kAk = kA0 k.

Note that for A, B ∈ L (X, Y ) and α, β ∈ C we have

(αA + βB)0 = αA0 + βB 0 (4.14)
and for A ∈ L (X, Y ) and B ∈ L (Y, Z) we have
(BA)0 = A0 B 0 (4.15)
4.3. The adjoint operator 117

which is immediate from the definition. Moreover, note that (IX )0 = IX ∗

which shows that if A is invertible then so is A0 is with
(A−1 )0 = (A0 )−1 . (4.16)
That A is invertible if A0 is will follow from Theorem 4.26 below.
Example 4.21. Given a Hilbert space H we have the conjugate linear isom-
etry C : H → H∗ , f 7→ hf, ·i. Hence for given A ∈ L (H1 , H2 ) we have
A0 C2 f = hf, A ·i = hA∗ f, ·i which shows A0 = C1 A∗ C2−1 . 
Example 4.22. Let X := Y := `p (N), 1 ≤ p < ∞, such that X ∗ = `q (N),
p + q = 1. Consider the right shift R ∈ L (` (N)) given by
1 1 p

Rx := (0, x1 , x2 , . . . ).
Then for y 0 ∈ `q (N)
∞
X ∞
X ∞
X
y 0 (Rx) = yj0 (Rx)j = yj0 xj−1 = 0
yj+1 xj
j=1 j=2 j=1

which shows (R0 y 0 )k = 0

yk+1 upon choosing x = δk . Hence R0 = L is the left
shift: Ly := (y2 , y3 , . . . ). Similarly, L0 = R. 
Example 4.23. Recall that an operator K ∈ L (X, Y ) is called a finite
rank operator if its range is finite dimensional. The dimension of its
range rank(K) := dim Ran(K) is called the rank of K. Choosing a ba-
sis {yj = Kxj }nj=1 for Ran(K) and a corresponding dual basis {yj0 }nj=1 (cf.
Problem 4.26), then x0j := K 0 yj0 is a dual basis for xj and
n
X n
X n
X
Kx = yj0 (Kx)yj = x0j (x)yj , 0 0
Ky = y 0 (yj )x0j .
j=1 j=1 j=1

In particular, rank(K) = rank(K 0 ). 

Of course we can also consider the doubly adjoint operator A00 . Then a
simple computation
A00 (JX (x))(y 0 ) = JX (x)(A0 y 0 ) = (A0 y 0 )(x) = y 0 (Ax) = JY (Ax)(y 0 ) (4.17)
shows that the following diagram commutes
A
X −→ Y
JX ↓ ↓ JY
X ∗∗ −→
00
Y ∗∗
A
Consequently
−1
A00 Ran(JX ) = JY AJX , A = JY−1 A00 JX . (4.18)
Hence, regarding X as a subspace JX (X) ⊆ X ∗∗ and Y as a subspace
JY (Y ) ⊆ Y ∗∗ , then A00 is an extension of A to X ∗∗ but with values in
118 4. The main theorems about Banach spaces

Y ∗∗ . In particular, note that B ∈ L (Y ∗ , X ∗ ) is the adjoint of some other

operator B = A0 if and only if B 0 (JX (X)) = A00 (JX (X)) ⊆ JY (Y ) (for the
converse note that A := JY−1 B 0 JX will do the trick). This can be used to
show that not every operator is an adjoint (Problem 4.27).
Theorem 4.23 (Schauder). Suppose X, Y are Banach spaces and A ∈
L (X, Y ). Then A is compact if and only if A0 is.

Proof. If A is compact, then A(B1X (0)) is relatively compact and hence

K = A(B1X (0)) is a compact metric space. Let yn0 ∈ Y ∗ be a bounded
sequence and consider the family of functions fn := yn0 |K ∈ C(K). Then
this family is bounded and equicontinuous since
|fn (y1 ) − fn (y2 )| ≤ kyn0 kky1 − y2 k ≤ Cky1 − y2 k.
Hence the Arzelà–Ascoli theorem (Theorem B.39) implies existence of a uni-
formly converging subsequence fnj . For this subsequence we have
kA0 yn0 j − A0 yn0 k k ≤ sup |yn0 j (Ax) − yn0 k (Ax)| = kfnj − fnk k∞
x∈B1X (0)

since A(B1X (0)) ⊆ K is dense. Thus yn0 j is the required subsequence and A0
is compact.
To see the converse note that if A0 is compact then so is A00 by the first
part and hence also A = JY−1 A00 JX . 

Finally we discuss the relation between solvability of Ax = y and the

corresponding adjoint equation A0 y 0 = x0 . To this end we need the analog
of the orthogonal complement of a set. Given subsets M ⊆ X and N ⊆ X ∗
we define their annihilator as
M ⊥ := {` ∈ X ∗ |`(x) = 0 ∀x ∈ M } = {` ∈ X ∗ |M ⊆ Ker(`)}
\ \
= {` ∈ X ∗ |`(x) = 0} = {x}⊥ ,
x∈M x∈M
\ \
N⊥ := {x ∈ X|`(x) = 0 ∀` ∈ N } = Ker(`) = {`}⊥ . (4.19)
`∈N `∈N

In particular, {`}⊥ = Ker(`) while {x}⊥ = Ker(J(x)) (with J : X ,→ X ∗∗

the canonical embedding). Note {0}⊥ = X ∗ and {0}⊥ = X.
Example 4.24. In a Hilbert space the annihilator is simply the orthogonal
complement. 
The following properties are immediate from the definition (by linearity
and continuity):
• M ⊥ is a closed subspace of X ∗ and M ⊥ = (span(M ))⊥ .
• N⊥ is a closed subspace of X and N⊥ = (span(N ))⊥ .
4.3. The adjoint operator 119

Note also that

span(M ) = X ⇔ M ⊥ = {0},
span(N ) = X ∗ ⇒ N⊥ = {0} (4.20)
by Corollary 4.17 and Corollary 4.16, respectively. The converse of the last
statement is wrong in general.
Example 4.25. Consider X := `1 (N) and N := {δ n }n∈N ⊂ `∞ (N) ∼ = X ∗.
Then span(N ) = c0 (N) but N⊥ = {0}. 
Lemma 4.24. We have (M ⊥ )⊥ = span(M ) and (N⊥ )⊥ ⊇ span(N ).

Proof. By the preceding remarks we can assume M , N to be closed sub-

spaces. The first part
(M ⊥ )⊥ = {x ∈ X|`(x) = 0 ∀` ∈ X ∗ with M ⊆ Ker(`)} = span(M )
is Corollary 4.18 and for the second part one just has to spell out the defi-
nition: \
(N⊥ )⊥ = {` ∈ X ∗ | ˜ ⊆ Ker(`)} ⊇ span(N ).
Ker(`) 
˜
`∈N

Note that we have equality in the preceding lemma if N is finite di-

mensional (Problem 4.33). Moreover, with a little more machinery one can
show equality if X is reflexive (Problem 5.16). For non-reflexive spaces the
inclusion can be strict as the previous example shows.
Warning: Some authors call a set N ⊆ X ∗ total if N⊥ = {0}. By the
preceding discussion this is equivalent to our definition if X is reflexive, but
otherwise might differ.
Furthermore, we have the following analog of (2.28).
Lemma 4.25. Suppose X, Y are normed spaces and A ∈ L (X, Y ). Then
Ran(A0 )⊥ = Ker(A) and Ran(A)⊥ = Ker(A0 ).

Proof. For the first claim observe: x ∈ Ker(A) ⇔ Ax = 0 ⇔ `(Ax) = 0,

∀` ∈ X ∗ ⇔ (A0 `)(x) = 0, ∀` ∈ X ∗ ⇔ x ∈ Ran(A0 )⊥ .
For the second claim observe: ` ∈ Ker(A0 ) ⇔ A0 ` = 0 ⇔ (A0 `)(x) = 0,
∀x ∈ X ⇔ `(Ax) = 0, ∀x ∈ X ⇔ ` ∈ Ran(A)⊥ . 

Taking annihilators in these formulas we obtain

Ker(A0 )⊥ = (Ran(A)⊥ )⊥ = Ran(A) (4.21)
and
Ker(A)⊥ = (Ran(A0 )⊥ )⊥ ⊇ Ran(A0 ) (4.22)
which raises the question of equality in the latter.
120 4. The main theorems about Banach spaces

Theorem 4.26 (Closed range). Suppose X, Y are Banach spaces and A ∈

L (X, Y ). Then the following items are equivlaent:
(i) Ran(A) is closed.
(ii) Ker(A)⊥ = Ran(A0 ).
(iii) Ran(A0 ) is closed.
(iv) Ker(A0 )⊥ = Ran(A).

Proof. (i) ⇔ (vi): Immediate from (4.21).

˜ = `˜ ◦ A vanishes on
(i) ⇒ (ii): Note that if ` ∈ Ran(A0 ) then ` = A0 (`)
Ker(A) and hence ` ∈ Ker(A) . Conversely, if ` ∈ Ker(A)⊥ we can set `(y)
⊥ ˜ =
`(Ã y) for y ∈ Ran(A) and extend it to all of Y using Corollary 4.15. Here
−1

à : X/ Ker(A) → Ran(A) is the induced map (cf. Problem 1.43) which has
a bounded inverse by Theorem 4.6. By construction ` = A0 (`) ˜ ∈ Ran(A0 ).
(ii) ⇒ (iii): Clear since annihilators are closed.
(iii) ⇒ (i): Let Z = Ran(A) and let à : X → Z be the range restriction
of A. Then Ã0 is injective (since Ker(Ã0 ) = Ran(Ã)⊥ = {0}) and has the
same range Ran(Ã0 ) = Ran(A0 ) (since every linear functional in Z ∗ can be
extended to one in Y ∗ by Corollary 4.15). Hence we can assume Z = Y and
hence A0 injective without loss of generality.
Suppose Ran(A) were not closed. Then, given ε > 0 and 0 ≤ δ < 1, by
Corollary 4.11 there is some y ∈ Y such that εkxk + ky − Axk > δkyk for
all x ∈ X. Hence there is a linear functional ` ∈ Y ∗ such that δ ≤ k`k ≤ 1
and kA0 `k ≤ ε. Indeed consider X ⊕ Y and use Corollary 4.17 to choose
`¯ ∈ (X ⊕ Y )∗ such that `¯ vanishes on the closed set V := {(εx, Ax)|x ∈ X},
¯ = 1, and `(0,
k`k ¯ y) = dist((0, y), V ) (note that (0, y) 6∈ V since y 6= 0). Then
¯
`(.) = `(0, .) is the functional we are looking for since dist((0, y), V ) ≥ δkyk
and (A0 `)(x) = `(0,¯ Ax) = `(−εx,
¯ ¯ 0). Now this allows us to
0) = −ε`(x,
choose `n with k`n k → 1 and kA `n k → 0 such that Corollary 4.10 implies
0

that Ran(A0 ) is not closed. 

With the help of annihilators we can also describe the dual spaces of
subspaces.

Theorem 4.27. Let M be a closed subspace of a normed space X. Then

there are canonical isometries
(X/M )∗ ∼
= M ⊥, M∗ ∼
= X ∗ /M ⊥ . (4.23)

Proof. In the first case the isometry is given by ` 7→ ` ◦ j, where j : X →

X/M is the quotient map. In the second case x0 + M ⊥ 7→ x0 |M . The details
are easy to check. 
4.3. The adjoint operator 121

Problem* 4.27. Let X := Y := c0 (N) and recall that X ∗ ∼ = `1 (N) and

X ∼
∗∗
= ` (N). Consider the operator A ∈ L (` (N)) given by
∞ 1
X
Ax := ( xn , 0, . . . ).
n∈N

Show that
A0 x0 = (x01 , x01 , . . . ).
Conclude that A is not the adjoint of an operator from L (c0 (N)).
Problem* 4.28. Show
Ker(A0 ) ∼
= Coker(A)∗ , Coker(A0 ) ∼
= Ker(A)∗
for A ∈ L (X, Y ) with Ran(A) closed.
Problem 4.29. Let Xj be Banach spaces. A sequence of operators Aj ∈
L (Xj , Xj+1 )
A
1 2 A n A
X1 −→ X2 −→ X3 · · · Xn −→ Xn+1
is said to be exact if Ran(Aj ) = Ker(Aj+1 ) for 1 ≤ j ≤ n. Show that a
sequence is exact if and only if the corresponding dual sequence
A0 A0 A0
X1∗ ←−
1
X2∗ ←−
2
X3∗ · · · Xn∗ ←−
n ∗
Xn+1
is exact.
Problem 4.30. Suppose X is separable. Show that there exists a countable
set N ⊂ X ∗ with N⊥ = {0}.
Problem 4.31. Show that for A ∈ L (X, Y ) we have
rank(A) = rank(A0 ).
Problem* 4.32 (Riesz lemma). Let X be a normed vector space and Y ⊂ X
some subspace. Show that if Y 6= X, then for every ε ∈ (0, 1) there exists an
xε with kxε k = 1 and
inf kxε − yk ≥ 1 − ε. (4.24)
y∈Y
Note: In a Hilbert space the claim holds with ε = 0 for any normalized x in
the orthogonal complement of Y and hence xε can be thought of a replacement
of an orthogonal vector. (Hint: Choose a yε ∈ Y which is close to x and look
at x − yε .)
Problem* 4.33. Suppose X is a vector space and `, `P
1 , . . . , `n are linear
functionals such that nj=1 Ker(`j ) ⊆ Ker(`). Then ` = nj=1 αj `j for some
T

constants αj ∈ C.
P (Hint: Find a dual basis xk ∈ X such that `j (xk ) = δj,k
and look at x − nj=1 `j (x)xj .)
122 4. The main theorems about Banach spaces

Problem 4.34. Suppose M1 , M2 are closed subspaces of X. Show

M1 ∩ M2 = (M1⊥ + M2⊥ )⊥ , M1⊥ ∩ M2⊥ = (M1 + M2 )⊥
and
(M1 ∩ M2 )⊥ ⊇ (M1⊥ + M2⊥ ), (M1⊥ ∩ M2⊥ )⊥ = (M1 + M2 ).
∗
Problem 4.35. Let us write `n * ` provided the sequence converges point-
∗
wise, that is, `n (x) → `(x) for all x ∈ X. Let N ⊆ X ∗ and suppose `n * `
with `n ∈ N . Show that ` ∈ (N⊥ )⊥ .

4.4. Weak convergence

In Section 4.2 we have seen that `(x) = 0 for all ` ∈ X ∗ implies x = 0.
Now what about convergence? Does `(xn ) → `(x) for every ` ∈ X ∗ imply
xn → x? In fact, in a finite dimensional space component-wise convergence
is equivalent to convergence. Unfortunately in the infinite dimensional this
is no longer true in general:
Example 4.26. Let un be an infinite orthonormal set in some Hilbert space.
Then hg, un i → 0 for every g since these are just the expansion coefficients
of g which are in `2 (N) by Bessel’s inequality. Since by the Riesz lemma
(Theorem 2.10), every bounded linear functional is of this form, we have
`(un ) → 0 for every bounded linear functional. (Clearly un does not converge
to 0, since kun k = 1.) 
If `(xn ) → `(x) for every ` ∈ X ∗ we say that xn converges weakly to
x and write
w-lim xn = x or xn * x. (4.25)
n→∞
Clearly, xn → x implies xn * x and hence this notion of convergence is
indeed weaker. Moreover, the weak limit is unique, since `(xn ) → `(x) and
`(xn ) → `(x̃) imply `(x − x̃) = 0. A sequence xn is called a weak Cauchy
sequence if `(xn ) is Cauchy (i.e. converges) for every ` ∈ X ∗ .

Lemma 4.28. Let X be a Banach space.

(i) xn * x, yn * y and αn → α implies xn + yn * x + y and
αn xn * αx.
(ii) xn * x implies kxk ≤ lim inf kxn k.
(iii) Every weak Cauchy sequence xn is bounded: kxn k ≤ C.
(iv) If X is reflexive, then every weak Cauchy sequence converges weakly.
(v) A sequence xn is Cauchy if and only if `(xn ) is Cauchy, uniformly
for ` ∈ X ∗ with k`k = 1.
4.4. Weak convergence 123

Proof. (i) Follows from `(αn xn + yn ) = αn `(xn ) + `(yn ) → α`(x) + `(y). (ii)
Choose ` ∈ X ∗ such that `(x) = kxk (for the limit x) and k`k = 1. Then
kxk = `(x) = lim inf `(xn ) ≤ lim inf kxn k.
(iii) For every ` we have that |J(xn )(`)| = |`(xn )| ≤ C(`) is bounded. Hence
by the uniform boundedness principle we have kxn k = kJ(xn )k ≤ C.
(iv) If xn is a weak Cauchy sequence, then `(xn ) converges and we can define
j(`) = lim `(xn ). By construction j is a linear functional on X ∗ . Moreover,
by (iii) we have |j(`)| ≤ sup |`(xn )| ≤ k`k sup kxn k ≤ Ck`k which shows
j ∈ X ∗∗ . Since X is reflexive, j = J(x) for some x ∈ X and by construction
`(xn ) → J(x)(`) = `(x), that is, xn * x.
(v) This follows from
kxn − xm k = sup |`(xn − xm )|
k`k=1

(cf. Problem 4.17). 

Item (ii) says that the norm is sequentially weakly lower semicontinuous
(cf. Problem B.18) while the previous example shows that it is not sequen-
tially weakly continuous (this will in fact be true for any convex function
as we will see later). However, bounded linear operators turn out to be
sequentially weakly continuous (Problem 4.37).
Example 4.27. Consider L2 (0, 1) and recall (see Example 3.8) that
√
un (x) = 2 sin(nπx), n ∈ N,
form an ONB and hence un * 0. However, vn = u2n * 1. In fact, one easily
computes
√ √
2(1 − (−1)m ) 4n2 2(1 − (−1)m )
hum , vn i = → = hum , 1i
mπ (4n2 − m2 ) mπ
q
and the claim follows from Problem 4.40 since kvn k = 32 . 
Remark: One can equip X with the weakest topology for which all ` ∈ X ∗
remain continuous. This topology is called the weak topology and it is
given by taking all finite intersections of inverse images of open sets as a
base. By construction, a sequence will converge in the weak topology if and
only if it converges weakly. By Corollary 4.17 the weak topology is Hausdorff,
but it will not be metrizable in general. In particular, sequences do not suffice
to describe this topology. Nevertheless we will stick with sequences for now
and come back to this more general point of view in Section 5.3.
In a Hilbert space there is also a simple criterion for a weakly convergent
sequence to converge in norm (see Theorem 5.19 for a generalization).
124 4. The main theorems about Banach spaces

Lemma 4.29. Let H be a Hilbert space and let fn * f . Then fn → f if

and only if lim sup kfn k ≤ kf k.

Proof. By (ii) of the previous lemma we have lim kfn k = kf k and hence

kf − fn k2 = kf k2 − 2Re(hf, fn i) + kfn k2 → 0.

The converse is straightforward. 

Now we come to the main reason why weakly convergent sequences are of
interest: A typical approach for solving a given equation in a Banach space
is as follows:
(i) Construct a (bounded) sequence xn of approximating solutions
(e.g. by solving the equation restricted to a finite dimensional sub-
space and increasing this subspace).
(ii) Use a compactness argument to extract a convergent subsequence.
(iii) Show that the limit solves the equation.
Our aim here is to provide some results for the step (ii). In a finite di-
mensional vector space the most important compactness criterion is bound-
edness (Heine–Borel theorem, Theorem B.22). In infinite dimensions this
breaks down as we have already seen in Section 1.5. We even have

Theorem 4.30. The closed unit ball in a Banach space X is compact if and
only if X is finite dimensional.

Proof. If X is finite dimensional, then by Theorem 1.8 we can assume X =

Cn and the closed unit ball is compact by the Heine–Borel theorem.
Conversely, suppose S = {x ∈ X| kxk = 1} is compact. Then {X \
Ker(`)}`∈X ∗ is an open cover since for every x ∈ S there is some ` ∈ X ∗
with `(x) 6= 0 by CorollaryTn 4.15. This cover T has a finite subcover, S ⊂
n n
Hence j=1 Ker(`j ) = {0} and the
S
j=1 (X \ Ker(`j )) = X \ j=1 Ker(` j ).
map X → C , x 7→ (`1 (x), · · · , `n (x)) is injective, that is, dim(X) ≤ n. 
n

However, if we are willing to treat convergence for weak convergence, the

situation looks much brighter!

Theorem 4.31. Let X be a reflexive Banach space. Then every bounded

sequence has a weakly convergent subsequence.

Proof. Let xn be some bounded sequence and consider Y = span{xn }. Then

Y is reflexive by Lemma 4.21 (i). Moreover, by construction Y is separable
and so is Y ∗ by the remark after Lemma 4.21.
4.4. Weak convergence 125

Let `k be a dense set in Y ∗ . Then by the usual diagonal sequence argu-

ment we can find a subsequence xnm such that `k (xnm ) converges for every
k. Denote this subsequence again by xn for notational simplicity. Then,
|`(xn ) − `(xm )| ≤|`(xn ) − `k (xn )| + |`k (xn ) − `k (xm )|
+ |`k (xm ) − `(xm )|
≤2Ck` − `k k + |`k (xn ) − `k (xm )|

shows that `(xn ) converges for every ` ∈ span{`k } = Y ∗ . Thus there is a

limit by Lemma 4.28 (iv). 

Note that this theorem breaks down if X is not reflexive.

Example 4.28. Consider the sequence of vectors δ n (with δnn = 1 and δm
n =

0, n 6= m) in ` (N), 1 ≤ p < ∞. Then δ * 0 for 1 < p < ∞. In fact,

p n

since every l ∈ `p (N)∗ is of the form l = ly for some y ∈ `q (N) we have

ly (δ n ) = yn → 0.
If we consider the same sequence in `1 (N) there is no weakly convergent
subsequence. In fact, since ly (δ n ) → 0 for every sequence y ∈ `∞ (N) with
finitely many nonzero entries, the only possible weak limit is zero. On the
other hand choosing the constant sequence y = (1)∞ j=1 we see ly (δ ) = 1 6→ 0,
n

a contradiction. 
Example 4.29. Let X := L1 [−1, 1]. Every bounded integrable ϕ gives rise
to a linear functional Z
`ϕ (f ) := f (x)ϕ(x) dx

in L1 [−1, 1]∗ . Take some nonnegative u1 with compact support, ku1 k1 = 1,

and set uk (x) = ku1 (k x) (implying kuk k1 = 1). Then we have
Z
uk (x)ϕ(x) dx → ϕ(0)

(see Problem 3.29 from [47]) for every continuous ϕ. Furthermore, if ukj * u
we conclude Z
u(x)ϕ(x) dx = ϕ(0).

In particular, choosing ϕk (x) = max(0, 1−k|x|) we infer from the dominated

convergence theorem
Z Z
1 = u(x)ϕk (x) dx → u(x)χ{0} (x) dx = 0,

a contradiction.
In fact, uk converges to the Dirac measure centered at 0, which is not in
L1 [−1, 1]. 
126 4. The main theorems about Banach spaces

Note that the above theorem also shows that in an infinite dimensional
reflexive Banach space weak convergence is always weaker than strong con-
vergence since otherwise every bounded sequence had a weakly, and thus by
assumption also norm, convergent subsequence contradicting Theorem 4.30.
In a non-reflexive space this situation can however occur.
Example 4.30. In `1 (N) every weakly convergent sequence is in fact (norm)
convergent (such Banach spaces are said to have the Schur property). First
of all recall that `1 (N)∗ ∼
= `∞ (N) and an * 0 implies
∞
X
lb (an ) = bk ank → 0, ∀b ∈ `∞ (N).
k=1
Now suppose we could find a sequence an * 0 for which lim supn kan k1 ≥
ε > 0. After passing to a subsequence we can assume kan k1 ≥ ε/2 and
after rescaling the vector even kan k1 = 1. Now weak convergence an * 0
implies anj = lδj (an ) → 0 for every fixed j ∈ N. Hence the main contri-
bution to the norm of an must move towards ∞ and we can find a subse-
quence nj and a corresponding increasing sequence of integers kj such that
nj
kj ≤k<kj+1 |ak | ≥ 3 . Now set
P 2

n
bk = sign(ak j ), kj ≤ k < kj+1 .
Then
 
 
nj
X nj  X nj  2 1 1
|lb (a )| ≥ |ak | − 
 bk ak  ≥ − = ,
kj ≤k<kj+1 1≤k<kj ; kj+1 ≤k  3 3 3

contradicting anj * 0. 
It is also useful to observe that compact operators will turn weakly con-
vergent into (norm) convergent sequences.
Theorem 4.32. Let A ∈ C (X, Y ) be compact. Then xn * x implies Axn →
Ax. If X is reflexive the converse is also true.

Proof. If xn * x we have supn kxn k ≤ C by Lemma 4.28 (ii). Consequently

Axn is bounded and we can pass to a subsequence such that Axnk → y.
Moreover, by Problem 4.37 we even have y = Ax and Lemma B.5 shows
Axn → Ax.
Conversely, if X is reflexive, then by Theorem 4.31 every bounded se-
quence xn has a subsequence xnk * x and by assumption Axnk → x. Hence
A is compact. 

Operators which map weakly convergent sequences to convergent se-

quences are also called completely continuous. However, be warned that
some authors use completely continuous for compact operators. By the above
4.4. Weak convergence 127

theorem every compact operator is completely continuous and the converse

also holds in reflexive spaces. However, the last example shows that the
identity map in `1 (N) is completely continuous but it is clearly not compact
by Theorem 4.30.
Similar concepts can be introduced for operators. This is of particular
importance for the case of unbounded operators, where convergence in the
operator norm makes no sense at all.
A sequence of operators An is said to converge strongly to A,
s-lim An = A :⇔ An x → Ax ∀x ∈ D(A) ⊆ D(An ). (4.26)
n→∞

It is said to converge weakly to A,

w-lim An = A :⇔ An x * Ax ∀x ∈ D(A) ⊆ D(An ). (4.27)
n→∞

Clearly norm convergence implies strong convergence and strong convergence

implies weak convergence. If Y is finite dimensional strong and weak con-
vergence will be the same and this is in particular the case for Y = C.
Example 4.31. Consider the operator Sn ∈ L (`2 (N)) which shifts a se-
quence n places to the left, that is,
Sn (x1 , x2 , . . . ) = (xn+1 , xn+2 , . . . ) (4.28)
and the operator Sn∗ ∈ L (`2 (N)) which shifts a sequence n places to the
right and fills up the first n places with zeros, that is,
Sn∗ (x1 , x2 , . . . ) = (0, . . . , 0, x1 , x2 , . . . ). (4.29)
| {z }
n places

Then Sn converges to zero strongly but not in norm (since kSn k = 1) and Sn∗
converges weakly to zero (since hx, Sn∗ yi = hSn x, yi) but not strongly (since
kSn∗ xk = kxk) . 
Lemma 4.33. Suppose An , Bn ∈ L (X, Y ) are sequences of bounded opera-
tors.
(i) s-lim An = A, s-lim Bn = B, and αn → α implies s-lim(An +Bn ) =
n→∞ n→∞ n→∞
A + B and s-lim αn An = αA.
n→∞
(ii) s-lim An = A implies kAk ≤ lim inf kAn k.
n→∞ n→∞
(iii) If An x converges for all x ∈ X then kAn k ≤ C and there is an
operator A ∈ L (X, Y ) such that s-lim An = A.
n→∞
(iv) If An y converges for y in a total set and kAn k ≤ C, then there is
an operator A ∈ L (X, Y ) such that s-lim An = A.
n→∞
The same result holds if strong convergence is replaced by weak convergence.
128 4. The main theorems about Banach spaces

Proof. (i) limn→∞ (αn An + Bn )x = limn→∞ (αn An x + Bn x) = αAx + Bx.

(ii) follows from
kAxk = lim kAn xk ≤ lim inf kAn k
n→∞ n→∞
for every x ∈ X with kxk = 1.
(iii) by linearity of the limit, Ax := limn→∞ An x is a linear operator. More-
over, since convergent sequences are bounded, kAn xk ≤ C(x), the uniform
boundedness principle implies kAn k ≤ C. Hence kAxk = limn→∞ kAn xk ≤
Ckxk.
(iv) By taking linear combinations we can replace the total set by a dense
one. Moreover, we can define a linear operator A on this dense set via
Ay := limn→∞ An y. By kAn k ≤ C we see kAk ≤ C and there is a unique
extension to all of X. Now just use
kAn x − Axk ≤ kAn x − An yk + kAn y − Ayk + kAy − Axk
≤ 2Ckx − yk + kAn y − Ayk
and choose y in the dense subspace such that kx − yk ≤ ε
4C and n large such
that kAn y − Ayk ≤ 2ε .
The case of weak convergence is left as an exercise (Problem 4.17 might
be useful). 

Item (iii) of this lemma is sometimes also known as Banach–Steinhaus

theorem. For an application of this lemma see Lemma 3.21 from [47].
Example 4.32. Let X be a Banach space of functions f : [−π, π] → C
such the functions {ek (x) := eikx }k∈Z are total. E.g. X = Cper [−π, π] or
X = Lp [−π, π] for 1 ≤ p < ∞. Then the Fourier series (2.44) converges on
a total set and hence it will converge on all of X if and only if kSn k ≤ C.
For example, if X = Cper [−π, π] then
1
kSn k = sup kSn (f )k = sup |Sn (f )(0)| = kDn k1
kf k∞ =1 kf k∞ =1 2π
which is unbounded as we have seen in Example 4.3. In fact, in this example
we have even shown failure of pointwise convergence and hence this is nothing
new. However, if we consider X = L1 [−π, π] we have (recall the Fejér kernel
which satisfies kFn k1 = 1 and use (2.52) together with Sn (Dm ) = Dmin(m,n) )
kSn k = sup kSn (f )k ≥ lim kSn (Fm )k1 = kDn k1
kf k1 =1 m→∞

and we get that the Fourier series does not converge for some L1 function. 
Lemma 4.34. Suppose An ∈ L (Y, Z), Bn ∈ L (X, Y ) are two sequences of
bounded operators.
(i) s-lim An = A and s-lim Bn = B implies s-lim An Bn = AB.
n→∞ n→∞ n→∞
4.4. Weak convergence 129

(ii) w-lim An = A and s-lim Bn = B implies w-lim An Bn = AB.

n→∞ n→∞ n→∞
(iii) lim An = A and w-lim Bn = B implies w-lim An Bn = AB.
n→∞ n→∞ n→∞

Proof. For the first case just observe

k(An Bn − AB)xk ≤ k(An − A)Bxk + kAn kk(Bn − B)xk → 0.
The remaining cases are similar and again left as an exercise. 
Example 4.33. Consider again the last example. Then
Sn∗ Sn (x1 , x2 , . . . ) = (0, . . . , 0, xn+1 , xn+2 , . . . )
| {z }
n places

converges to 0 weakly (in fact even strongly) but

Sn Sn∗ (x1 , x2 , . . . ) = (x1 , x2 , . . . )
does not! Hence the order in the second claim is important. 
For a sequence of linear functionals `n , strong convergence is also called
weak-∗ convergence. That is, the weak-∗ limit of `n is ` if `n (x) → `(x) for
all x ∈ X and we will write
∗
w*-lim `n = ` or `n * ` (4.30)
n→∞
in this case. Note that this is not the same as weak convergence on X ∗ unless
X is reflexive: ` is the weak limit of `n if
j(`n ) → j(`) ∀j ∈ X ∗∗ , (4.31)
whereas for the weak-∗ limit this is only required for j ∈ J(X) ⊆ X ∗∗ (recall
J(x)(`) = `(x)).
Example 4.34. In a Hilbert space weak-∗ convergence of the linear func-
tionals hxn , .i is the same as weak convergence of the vectors xn . 
∗ ∼
Example 4.35. Consider X = c0 (N), X = ` (N), and X = ` (N) with
1 ∗∗ ∼ ∞

J corresponding to the inclusion c0 (N) ,→ `∞ (N). Then weak convergence

on X ∗ implies
∞
X
lb (an − a) = bk (ank − ak ) → 0
k=1
for all b ∈`∞ (N) and weak-* convergence implies that this holds for all b ∈
c0 (N). Whereas we already have seen that weak convergence is equivalent to
norm convergence, it is not hard to see that weak-* convergence is equivalent
to the fact that the sequence is bounded and each component converges (cf.
Problem 4.41). 
With this notation it is also possible to slightly generalize Theorem 4.31
(Problem 4.42):
130 4. The main theorems about Banach spaces

Lemma 4.35 (Helly). Suppose X is a separable Banach space. Then every

bounded sequence `n ∈ X ∗ has a weak-∗ convergent subsequence.
Example 4.36. Let us return to the example after Theorem 4.31. Consider
Rthe Banach space of continuous functions X = C[−1, 1]. Using `f (ϕ) =
ϕf dx we can regard L1 [−1, 1] as a subspace of X ∗ . Then the Dirac measure
centered at 0 is also in X ∗ and it is the weak-∗ limit of the sequence uk . 
Problem 4.36. Suppose `n → ` in X ∗ and xn * x in X. Then `n (xn ) →
`(x). Similarly, suppose s-lim `n → ` and xn → x. Then `n (xn ) → `(x).
Does this still hold if s-lim `n → ` and xn * x?
Problem* 4.37. Show that xn * x implies Axn * Ax for A ∈ L (X, Y ).
Conversely, show that if xn → 0 implies Axn * 0 then A ∈ L (X, Y ).
Problem 4.38. Let X := X1 ⊕ X2 show that (x1,n , x2,n ) * (x1 , x2 ) if and
only if xj,n * xj for j = 1, 2.
Problem 4.39. Suppose An , A ∈ L (X, Y ). Show that s-lim An = A and
lim xn = x implies lim An xn = Ax.
Problem* 4.40. Show that if {`j } ⊆ X ∗ is some total set, then xn * x if
and only if xn is bounded and `j (xn ) → `j (x) for all j. Show that this is
wrong without the boundedness assumption (Hint: Take e.g. X = `2 (N)).
∗
Problem* 4.41. Show that if {xj } ⊆ X is some total set, then `n * ` if
and only if `n ∈ X ∗ is bounded and `n (xj ) → `(xj ) for all j.
Problem* 4.42. Prove Lemma 4.35.
 Chapter 5

Further topics on
Banach spaces

5.1. The geometric Hahn–Banach theorem

The Hahn–Banach theorem is about constructing (continuous) linear func-
tionals with given properties. In our original version this was done by show-
ing that a functional defined on a subspace can be extended to the entire
space. In this section we will establish a geometric version which establishes
existence of functionals separating given convex sets. The key ingredient will
be an association between convex sets and convex functions such that we can
apply our original version of the Hahn–Banach theorem.
Let X be a vector space. For every subset U ⊂ X we define its Minkowski
functional (or gauge)

pU (x) = inf{t > 0|x ∈ t U }. (5.1)

Here t U = {tx|x ∈ U }. Note that 0 ∈ U implies pU (0) = 0 and pU (x) will

be finite for all x when U is absorbing, that is, for every x ∈ X there is
some r such that x ∈ αU for every |α| ≥ r. Note that every absorbing set
contains 0 and every neighborhood of 0 in a Banach space is absorbing.
Example 5.1. Let X be a Banach space and U = B1 (0), then pU (x) = kxk.
If X = R2 and U = (−1, 1) × R then pU (x) = |x1 |. If X = R2 and U =
(−1, 1) × {0} then pU (x) = |x1 | if x2 = 0 and pU (x) = ∞ else. 
We will only need minimal requirements and it will suffice if X is a
topological vector space, that is, a vector space which carries a topology
such that both vector addition X × X → X and scalar multiplication C ×
X → X are continuous mappings. Of course every normed vector space is

131
132 5. Further topics on Banach spaces

`(x) = c

Figure 5.1. Separation of convex sets via a hyperplane

a topological vector space with the usual topology generated by open balls.
As in the case of normed linear spaces, X ∗ will denote the vector space of
all continuous linear functionals on X.
Lemma 5.1. Let X be a vector space and U a convex subset containing 0.
Then
pU (x + y) ≤ pU (x) + pU (y), pU (λ x) = λpU (x), λ ≥ 0. (5.2)
Moreover, {x|pU (x) < 1} ⊆ U ⊆ {x|pU (x) ≤ 1}. If, in addition, X is a
topological vector space and U is open, then U = {x|pU (x) < 1}.

Proof. The homogeneity condition p(λ x) = λp(x) for λ > 0 is straight-

forward. To see the sublinearity Let t, s > 0 with x ∈ t U and y ∈ s U ,
then
t x s y x+y
+ =
t+s t t+ss t+s
is in U by convexity. Moreover, pU (x + y) ≤ s + t and taking the infimum
over all t and s we find pU (x + y) ≤ pU (x) + pU (y).
Suppose pU (x) < 1, then t−1 x ∈ U for some t < 1 and thus x ∈ U by
convexity. Similarly, if x ∈ U then t−1 x ∈ U for t ≥ 1 by convexity and thus
pU (x) ≤ 1. Finally, let U be open and x ∈ U , then (1 + ε)x ∈ U for some
ε > 0 and thus p(x) ≤ (1 + ε)−1 . 

Note that (5.2) implies convexity

pU (λx + (1 − λ)y) ≤ λpU (x) + (1 − λ)pU (y), λ ∈ [0, 1]. (5.3)

Theorem 5.2 (geometric Hahn–Banach, real version). Let U , V be disjoint

nonempty convex subsets of a real topological vector space X and let U be
open. Then there is a linear functional ` ∈ X ∗ and some c ∈ R such that
`(x) < c ≤ `(y), x ∈ U, y ∈ V. (5.4)
If V is also open, then the second inequality is also strict.
5.1. The geometric Hahn–Banach theorem 133

Proof. Choose x0 ∈ U and y0 ∈ V , then

W = (U − x0 ) − (V − y0 ) = {(x − x0 ) − (y − y0 )|x ∈ U, y ∈ V }
is open (since U is), convex (since U and V are) and contains 0. Moreover,
since U and V are disjoint we have z0 = y0 −x0 6∈ W . By the previous lemma,
the associated Minkowski functional pW is convex and by the Hahn–Banach
theorem there is a linear functional satisfying
`(tz0 ) = t, |`(x)| ≤ pW (x).
Note that since z0 6∈ W we have pW (z0 ) ≥ 1. Moreover, W = {x|pW (x) <
1} ⊆ {x||`(x)| < 1} which shows that ` is continuous at 0 by scaling. By
translations ` is continuous everywhere.
Finally we again use pW (z) < 1 for z ∈ W implying
`(x) − `(y) + 1 = `(x − y + z0 ) ≤ pW (x − y + z0 ) < 1
and hence `(x) < `(y) for x ∈ U and y ∈ V . Therefore `(U ) and `(V ) are
disjoint convex subsets of R. Finally, let us suppose that there is some x1
for which `(x1 ) = sup `(U ). Then, by continuity of the map t 7→ x1 + tz0
there is some ε > 0 such that x1 + εz0 ∈ U . But this gives a contradiction
`(x1 )+ε = `(x1 +εz0 ) ≤ `(x1 ). Thus the claim holds with c = sup `(U ). If V
is also open an analogous argument shows inf `(V ) < `(y) for all y ∈ V . 

Of course there is also a complex version.

Theorem 5.3 (geometric Hahn–Banach, complex version). Let U , V be
disjoint nonempty convex subsets of a topological vector space X and let U
be open. Then there is a linear functional ` ∈ X ∗ and some c ∈ R such that
Re(`(x)) < c ≤ Re(`(y)), x ∈ U, y ∈ V. (5.5)
If V is also open, then the second inequality is also strict.

Proof. Consider X as a real Banach space. Then there is a continuous

real-linear functional `r : X → R by the real version of the geometric Hahn–
Banach theorem. Then `(x) = `r (x) − i`r (ix) is the functional we are looking
for (check this). 
Example 5.2. The assumption that one set is open is crucial as the following
example shows. Let X = c0 (N), U = {a ∈ c0 (N)|∃N : aN > 0 and an =
0, n > N } and V = {0}. Note that U is convex but not open and that
U ∩ V = ∅. Suppose we could find a linear functional ` as in the geometric
Hahn–Banach theorem (of course we can choose α = `(0) = 0 in P this case).
Then by Problem 4.22 there is some bj ∈ `∞ (N) such that `(a) = ∞ j=1 bj aj .
Moreover, we must have bj = `(δ ) < 0. But then a = (b2 , −b1 , 0, . . . ) ∈ U
j

and `(a) = 0 6< 0. 

134 5. Further topics on Banach spaces

Note that two disjoint closed convex sets can be separated strictly if
one of them is compact. However, this will require that every point has
a neighborhood base of convex open sets. Such topological vector spaces
are called locally convex spaces and they will be discussed further in
Section 5.4. For now we just remark that every normed vector space is
locally convex since balls are convex.
Corollary 5.4. Let U , V be disjoint nonempty closed convex subsets of a
locally convex space X and let U be compact. Then there is a linear functional
` ∈ X ∗ and some c, d ∈ R such that
Re(`(x)) ≤ d < c ≤ Re(`(y)), x ∈ U, y ∈ V. (5.6)

Proof. Since V is closed, for every x ∈ U there is a convex open neighbor-

hood Nx of 0 such that x + Nx does not intersect V . By compactness of U
there are x1 , . . . , xT
n such that the corresponding neighborhoods xj + 2 Nxj
1
n
cover U . Set N = j=1 Nxj which is a convex open neighborhood of 0. Then
n n n
1 [ 1 1 [ 1 1 [
Ũ = U + N ⊆ (xj + Nxj )+ N ⊆ (xj + Nxj + Nxj ) = (xj +Nxj )
2 2 2 2 2
j=1 j=1 j=1

is a convex open set which is disjoint from V . Hence by the previous theorem
we can find some ` such that Re(`(x)) < c ≤ Re(`(y)) for all x ∈ Ũ and y ∈
V . Moreover, since `(U ) is a compact interval [e, d], the claim follows. 

Note that if U and V are absolutely convex, that is, αU + βU ⊆ U

for |α| + |β| ≤ 1, then we can write the previous condition equivalently as
|`(x)| ≤ d < c ≤ |`(y)|, x ∈ U, y ∈ V, (5.7)
since x ∈ U implies θx ∈ U for θ = sign(`(x)) and thus |`(x)| = θ`(x) =
`(θx) = Re(`(θx)).
From the last corollary we can also obtain versions of Corollaries 4.17
and 4.15 for locally convex vector spaces.
Corollary 5.5. Let Y ⊆ X be a subspace of a locally convex space and let
x0 ∈ X \ Y . Then there exists an ` ∈ X ∗ such that (i) `(y) = 0, y ∈ Y and
(ii) `(x0 ) = 1.

Proof. Consider ` from Corollary 5.4 applied to U = {x0 } and V = Y . Now

observe that `(Y ) must be a subspace of C and hence `(Y ) = {0} implying
Re(`(x0 )) < 0. Finally `(x0 )−1 ` is the required functional. 
Corollary 5.6. Let Y ⊆ X be a subspace of a locally convex space and let
` : Y → C be a continuous linear functional. Then there exists a continuous
extension `¯ ∈ X ∗ .
5.2. Convex sets and the Krein–Milman theorem 135

Proof. Without loss of generality we can assume that ` is nonzero such

that we can find x0 ∈ y with `(x0 ) = 1. Since Y has the subset topology
x0 6∈ Y0 := Ker(`), where the closure is taken in X. Now Corollary 5.5 gives
a functional `¯ with `(x
¯ 0 ) = 1 and Y0 ⊆ Ker(`).
¯ Moreover,
¯ − `(x) = `(x)
`(x) ¯ − `(x)`(x
¯ 0 ) = `(x
¯ − `(x)x0 ) = 0, x ∈ Y,
since x − `(x)x0 ∈ Ker(`). 
Problem 5.1. Show that in a Banach space Br (0) ⊆ U implies pU (x) ≤
1
r kxk.

Problem* 5.2. Let X be a topological vector space. Show that U + V is

open if one of the sets is open.
Problem 5.3. Show that Corollary 5.4 fails even in R2 unless one set is
compact.
Problem 5.4. Let X be a topological vector space and M ⊆ X, N ⊆ X ∗ .
Then the corresponding polar, prepolar sets are
M ◦ = {` ∈ X ∗ ||`(x)| ≤ 1 ∀x ∈ M }, N◦ = {x ∈ X||`(x)| ≤ 1 ∀` ∈ N },
respectively. Show
(i) M ◦ is closed and absolutely convex.
(ii) M1 ⊆ M2 implies M2◦ ⊆ M1◦ .
(iii) For α 6= 0 we have (αM )◦ = |α|−1 M ◦ .
(iv) If M is a subspace we have M ◦ = M ⊥ .
The same claims hold for prepolar sets.
Problem 5.5 (Bipolar theorem). Let X be a locally convex space and sup-
pose M ⊆ X is absolutely convex we have αM + βM ⊆ M . Show (M ◦ )◦ =
M . (Hint: Use Corollary 5.4 to show that for every y 6∈ M there is some
` ∈ X ∗ with Re(`(x)) ≤ 1 < `(y), x ∈ M .)

5.2. Convex sets and the Krein–Milman theorem

Let X be a locally convex vector space. Since the intersection of arbitrary
convex sets is again convex we can define the convex hull of a set U as the
smallest convex set containing U , that is, the intersection of all convex sets
containing U . It is straightforward to show (Problem 5.6) that the convex
hull is given by
n
X n
X
conv(U ) := { λj xj |n ∈ N, xj ∈ U, λj = 1, λj ≥ 0}. (5.8)
j=1 j=1
136 5. Further topics on Banach spaces

A line segment is convex and can be generated as the convex hull of its
endpoints. Similarly, a full triangle is convex and can be generated as the
convex hull of its vertices. However, if we look at a ball, then we need its
entire boundary to recover it as the convex hull. So how can we characterize
those points which determine a convex sets via the convex hull?
Let K be a set and M ⊆ K a nonempty subset. Then M is called
an extremal subset of K if no point of M can be written as a convex
combination of two points unless both are in M : For given x, y ∈ K and
λ ∈ (0, 1) we have that
λx + (1 − λ)y ∈ M ⇒ x, y ∈ M. (5.9)
If M = {x} is extremal, then x is called an extremal point of K. Hence
an extremal point cannot be written as a convex combination of two other
points from K.
Note that we did not require K to be convex. If K is convex, then M is
extremal if and only if K \M is convex. Note that the nonempty intersection
of extremal sets is extremal. Moreover, if L ⊆ M is extremal and M ⊆ K is
extremal, then L ⊆ K is extremal as well (Problem 5.7).
Example 5.3. Consider R2 with the norms k.kp . Then the extremal points
of the closed unit ball (cf. Figure 1.1) are the boundary points for 1 < p < ∞
and the vertices for p = 1, ∞. In any case the boundary is an extremal set.
Slightly more general, in a strictly convex space, (ii) of Problem 1.13 says
that the extremal points of the unit ball are precisely its boundary points. 
Example 5.4. Consider R3 and let C = {(x1 , x2 , 0) ∈ R3 |x21 + x22 = 1}.
Take two more points x± = (0, 0, ±1) and consider the convex hull K of
M = C ∪ {x+ , x− }. Then M is extremal in K and, moreover, every point
from M is an extremal point. However, if we change the two extra points to
be x± = (1, 0, ±1), then the point (1, 0, 0) is no longer extremal. Hence the
extremal points are now M \ {(1, 0, 0)}. Note in particular that the set of
extremal points is not closed in this case. 
Extremal sets arise naturally when minimizing linear functionals.
Lemma 5.7. Suppose K ⊆ X and ` ∈ X ∗ . If
K` := {x ∈ K|`(x) = inf Re(`(y))}
y∈K

is nonempty (e.g. if K is compact), then it is extremal in K. If K is closed

and convex, then K` is closed and convex.

Proof. Set m = inf y∈K Re(`(y)). Let x, y ∈ K, λ ∈ (0, 1) and suppose

λx + (1 − λ)y ∈ K` . Then
m = Re(`(λx+(1−λ)y)) = λRe(`(x))+(1−λ)Re(`(y)) ≥ λm+(1−λ)m = m
5.2. Convex sets and the Krein–Milman theorem 137

with strict inequality if Re(`(x)) > m or Re(`(y)) > m. Hence we must

have x, y ∈ K` . Finally by linearity K` is convex and by continuity it is
closed. 

If K is a closed convex set, then nonempty subsets of the type K` are

called faces of K and H` := {x ∈ X|`(x) = inf y∈K Re(`(y))} is called a
support hyperplane of K.
Conversely, if K is convex with nonempty interior, then every point x
on the boundary has a supporting hyperplane (observe that the interior is
convex and apply the geometric Hahn–Banach theorem with U = K ◦ and
V = {x}).
Next we want to look into existence of extremal points.
Example 5.5. Note that an interior point can never be extremal as it can
be written as convex combination of some neighboring points. In particular,
an open convex set will not have any extremal points (e.g. X, which is also
closed, has no extremal points). Conversely, if K is closed and convex, then
the boundary is extremal since K \ ∂K = K ◦ is convex (Problem 5.8). 
Example 5.6. Suppose X is a strictly convex Banach space. Then every
nonempty compact subset K has an extremal point. Indeed, let x ∈ K be
such that kxk = supy∈K kyk, then x is extremal: If x = λy + (1 − λ)z then
kxk ≤ λkyk + (1 − λ)kzk ≤ kxk shows that we have equality in the triangle
inequality and hence x = y = z by Problem 1.13 (i). 
Example 5.7. In a not strictly convex space the situation is quite different.
For example, consider the closed unit ball in `∞ (N). Let a ∈ `∞ (N). If
there is some index j such that λ := |aj | < 1 then a = 21 b + 21 c where
b = a+εδ j and c = a−εδ j with ε ≤ 1−|aj |. Hence the only possible extremal
points are those with |aj | = 1 for all j ∈ N. If we have such an a, then if
a = λb + (1 − λ)c we must have 1 = |λbn + (1 − λ)cn | ≤ λ|bn | + (1 − λ)|cn | ≤ 1
and hence an = bn = cn by strict convexity of the absolute value. Hence all
such sequences are extremal.
However, if we consider c0 (N) the same argument shows that the closed
unit ball contains no extremal points. In particular, the following lemma
implies that there is no locally convex topology for which the closed unit
ball in c0 (N) is compact. Together with the Banach–Alaoglu theorem (The-
orem 5.10) this will show that c0 (N) is not the dual of any Banach space. 
Lemma 5.8 (Krein–Milman). Let X be a locally convex space. Suppose
K ⊆ X is compact and nonempty. Then it contains at least one extremal
point.

Proof. We want to apply Zorn’s lemma. To this end consider the family
M = {M ⊆ K|compact and extremal in K}
138 5. Further topics on Banach spaces

with the partial order given by reversed inclusion. Since K ∈ M this family
is nonempty. Moreover, given a linear chain C ⊂ M we consider M := C.
T
Then M ⊆ K is nonempty by the finite intersection property and since it
is closed also compact. Moreover, as the nonempty intersection of extremal
sets it is also extremal. Hence M ∈ M and thus M has a maximal element.
Denote this maximal element by M .
We will show that M contains precisely one point (which is then ex-
tremal by construction). Indeed, suppose x, y ∈ M . If x 6= y we can, by
Corollary 5.4, choose a linear functional ` ∈ X ∗ with Re(`(x)) 6= Re(`(y)).
Then by Lemma 5.7 M` ⊂ M is extremal in M and hence also in K. But
by Re(`(x)) 6= Re(`(y)) it cannot contain both x and y contradicting maxi-
mality of M . 

Finally, we want to recover a convex set as the convex hull of its ex-
tremal points. In our infinite dimensional setting an additional closure will
be necessary in general.
Since the intersection of arbitrary closed convex sets is again closed and
convex we can define the closed convex hull of a set U as the smallest closed
convex set containing U , that is, the intersection of all closed convex sets
containing U . Since the closure of a convex set is again convex (Problem 5.8)
the closed convex hull is simply the closure of the convex hull.
Theorem 5.9 (Krein–Milman). Let X be a locally convex space. Suppose
K ⊆ X is convex and compact. Then it is the closed convex hull of its
extremal points.

Proof. Let E be the extremal points and M := conv(E) ⊆ K be its closed

convex hull. Suppose x ∈ K \ M and use Corollary 5.4 to choose a linear
functional ` ∈ X ∗ with
min Re(`(y)) > Re(`(x)) ≥ min Re(`(y)).
y∈M y∈K

Now consider K` from Lemma 5.7 which is nonempty and hence contains an
extremal point y ∈ E. But y 6∈ M , a contradiction. 

While in the finite dimensional case the closure is not necessary (Prob-
lem 5.9), it is important in general as the following example shows.
Example 5.8. Consider the closed unit ball in `1 (N). Then the extremal
points are {eiθ δ n |n ∈ N, θ ∈ R}. Indeed, suppose kak1 = 1 with λ :=
|aj | ∈ (0, 1) for some j ∈ N. Then a = λb + (1 − λ)c where b := λ−1 aj δ j
and c := (1 − λ)−1 (a − aj δ j ). Hence the only possible extremal points
are of the form eiθ δ n . Moreover, if eiθ δ n = λb + (1 − λ)c we must have
1 = |λbn +(1−λ)cn | ≤ λ|bn |+(1−λ)|cn | ≤ 1 and hence an = bn = cn by strict
convexity of the absolute value. Thus the convex hull of the extremal points
5.2. Convex sets and the Krein–Milman theorem 139

are the sequences from the unit ball which have finitely many terms nonzero.
While the closed unit ball is not compact in the norm topology it will be in
the weak-∗ topology by the Banach–Alaoglu theorem (Theorem 5.10). To
this end note that `1 (N) ∼
= c0 (N)∗ . 
Also note that in the infinite dimensional case the extremal points can
be dense.
Example 5.9. Let X = C([0, 1], R) and consider the convex set K = {f ∈
C 1 ([0, 1], R)|f (0) = 0, kf 0 k∞ ≤ 1}. Note that the functions f± (x) = ±x are
extremal. For example, assume
x = λf (x) + (1 − λ)g(x)
then
1 = λf 0 (x) + (1 − λ)g 0 (x)
which implies f 0 (x) = g 0 (x) = 1 and hence f (x) = g(x) = x.
To see that there are no other extremal functions, suppose |f 0 (x)| ≤ 1−ε
on some interval I. Choose a nontrivial continuous function R x g which is 0
outside I and has integral 0 over I and kgk∞ ≤ ε. Let G = 0 g(t)dt. Then
f = 21 (f + G) + 12 (f − G) and hence f is not extremal. Thus f± are the
only extremal points and their (closed) convex is given by fλ (x) = λx for
λ ∈ [−1, 1].
Of course the problem is that K is not closed. Hence we consider the
Lipschitz continuous functions K̄ := {f ∈ C 0,1 ([0, 1], R)|f (0) = 0, [f ]1 ≤ 1}
(this is in fact the closure of K, but this is a bit tricky to see and we won’t
need this here). By the Arzelà–Ascoli theorem (Theorem 1.13) K̄ is relatively
compact and since the Lipschitz estimate clearly is preserved under uniform
limits it is even compact.
Now note that piecewise linear functions with f 0 (x) ∈ {±1} away from
the kinks are extremal in K̄. Moreover, these functions are dense: Split
[0, 1] into n pieces of equal length using xj = nj . Set fn (x0 ) = 0 and
fn (x) = fn (xj ) ± (x − xj ) for x ∈ [xj , xj+1 ] where the sign is chosen such
that |f (xj+1 ) − fn (xj+1 )| gets minimal. Then kf − fn k∞ ≤ n1 . 
Problem* 5.6. Show that the convex hull is given by (5.8).
Problem* 5.7. Show that the nonempty intersection of extremal sets is
extremal. Show that if L ⊆ M is extremal and M ⊆ K is extremal, then
L ⊆ K is extremal as well.
Problem 5.8. Let X be a topological vector space. Show that the closure
and the interior of a convex set is convex. (Hint: One way of showing the
first claim is to consider the continuous map f : X × X → X given by
(x, y) 7→ λx + (1 − λ)y and use Problem B.14.)
140 5. Further topics on Banach spaces

Problem 5.9 (Carathéodory). Show that for a compact convex set K ⊆ Rn

every point can be written as convex combination of n + 1 extremal points.
(Hint: Induction on n. Without loss assume that 0 is an extremal point. If
K is contained in an n − 1 dimensional subspace we are done. Otherwise K
has an open interior. Now for a given point the line through this point and
0 intersects the boundary where we have a corresponding face.)

5.3. Weak topologies

In Section 4.4 we have defined weak convergence for sequences and this raises
the question about a natural topology associated with this convergence. To
this end we define the weak topology on X as the weakest topology for
which all ` ∈ X ∗ remain continuous. Recall that a base for this topology is
given by sets of the form
n
\
|`j |−1 [0, εj ) = {x̃ ∈ X||`j (x − x̃)| < εj , 1 ≤ j ≤ n},


x+
j=1
x ∈ X, `j ∈ X ∗ , εj > 0. (5.10)
In particular, it is straightforward to check that a sequence converges with
respect to this topology if and only if it converges weakly. Since the linear
functionals separate points (cf. Corollary 4.16) the weak topology is Haus-
dorff.
Note that, if X ∗ is separable, given a total set {`n }n∈N ⊂ X ∗ of (w.l.o.g.)
normalized linear functionals
∞
X 1
d(x, x̃) = |`n (x − x̃)| (5.11)
2n
n=1

defines a metric on the unit ball B1 (0) ⊂ X which can be shown to generate
the weak topology (Problem 5.13). However, on all of X the weak topology
is not first countable unless X is finite dimensional (Problem 5.14).
Similarly, we define the weak-∗ topology on X ∗ as the weakest topology
for which all j ∈ J(X) ⊆ X ∗∗ remain continuous. In particular, the weak-∗
topology is weaker than the weak topology on X ∗ and both are equal if X
is reflexive. Like the weak topology it is Hausdorff (since different linear
functionals must differ at least at one point) and not first countable unless
X is finite dimensional (Problem 5.14). A base for the weak-∗ topology is
given by sets of the form
n
\
|J(xj )|−1 [0, εj ) = {`˜ ∈ X ∗ ||(` − `)(x
˜ j )| < εj , 1 ≤ j ≤ n},


`+
j=1
` ∈ X ∗ , xj ∈ X, εj > 0. (5.12)
5.3. Weak topologies 141

Note that, if X is separable, given a total set {xn }n∈N ⊂ X of (w.l.o.g.)

normalized vectors
∞
˜
X 1 ˜ n )|
d(`, `) = |(` − `)(x (5.13)
2n
n=1
defines a metric on the unit ball ⊂ X ∗ which can be shown to generate
B1∗ (0)
the weak-∗ topology (Problem 5.13). Hence Lemma 4.35 could also be stated
as B̄1∗ (0) ⊂ X ∗ being weak-∗ compact. This is in fact true without assuming
X to be separable and is known as Banach–Alaoglu theorem.
Theorem 5.10 (Banach–Alaoglu). Let X be a Banach space. Then B̄1∗ (0) ⊂
X ∗ is compact in the weak-∗ topology.
∗
Proof. Abbreviate B = B̄1X (0), B ∗ = B̄1X (0), and Bx = B̄kxk C (0). Consider

the (injective) map Φ : X ∗ → CX given by |Φ(`)(x)| = `(x) and identify X ∗

with Φ(X ∗ ). Then the weak-∗ topology on X ∗ coincides with the relative
topology on Φ(X ∗ ) ⊆ CX (recall that the product topology on CX is the
weakest topology which makes all point evaluations continuous). Moreover,

Φ(`) ≤ k`kkxk implies Φ(B ∗ ) ⊂ x∈X Bx where the last product is compact
by Tychonoff’s theorem. Hence it suffices to show that Φ(B ∗ ) is closed. To
this end let l ∈ Φ(B ∗ ). We need to show that l is linear and bounded. Fix
x1 , x2 ∈ X, α ∈ C, and consider the open neighborhood
n   |h(x + x ) − l(x + αx )| < ε, o
1 2 1 2
U (l) = h ∈ Bx 
|h(x1 ) − l(x1 )| < ε, |α||h(x2 ) − l(x2 )| < ε
x∈B

of l. Since U (l) ∩ Φ(X ∗ ) is nonempty we can choose an element h from

this intersection to show |l(x1 + αx2 ) − l(x1 ) − αl(x2 )| < 3ε. Since ε > 0
is arbitrary we conclude l(x1 + αx2 ) = l(x1 ) − αl(x2 ). Moreover, |l(x1 )| ≤
|h(x1 )| + ε ≤ kx1 k + ε shows klk ≤ 1 and thus l ∈ Φ(B ∗ ). 

If X is a reflexive space and we apply this to X ∗ , we get that the closed

unit ball is compact in the weak topology and we recover Theorem 4.31. In
fact, the converse is also true.
Theorem 5.11 (Kakutani). A Banach space X is reflexive if and only if
the closed unit ball B̄1 (0) is weakly compact.

Proof. Suppose X is not reflexive and choose x00 ∈ B̄1∗∗ (0) \ J(B̄1 (0)) with
kx00 k = 1. Then, if B̄1 (0) is weakly compact, J(B̄1 (0)) is weak-∗ compact
and by Corollary 5.4 we can find some ` ∈ X ∗ with k`k = 1 and
Re(x00 (`)) < inf Re(y 00 (`)) = inf Re(`(y)) = −1.
y 00 ∈J(B̄1 (0)) y∈B̄1 (0)

But this contradicts |x00 (`)| ≤ 1. 

142 5. Further topics on Banach spaces

Since the weak topology is weaker than the norm topology every weakly
closed set is also (norm) closed. Moreover, the weak closure of a set will in
general be larger than the norm closure. However, for convex sets both will
coincide. In fact, we have the following characterization in terms of closed
(affine) half-spaces, that is, sets of the form {x ∈ X|Re(`(x)) ≤ α} for
some ` ∈ X ∗ and some α ∈ R.
Theorem 5.12 (Mazur). The weak as well as the norm closure of a convex
set K is the intersection of all half-spaces containing K. In particular, a
convex set K ⊆ X is weakly closed if and only if it is closed.

Proof. Since the intersection of closed-half spaces is (weakly) closed, it suf-

fices to show that for every x not in the (weak) closure there is a closed half-
plane not containing x. Moreover, if x is not in the weak closure it is also
not in the norm closure (the norm closure is contained in the weak closure)
and by Theorem 5.3 with U = Bdist(x,K) (x) and V = K there is a functional
` ∈ X ∗ such that K ⊆ Re(`)−1 ([c, ∞)) and x 6∈ Re(`)−1 ([c, ∞)). 
w
Example 5.10. Suppose X is infinite dimensional. The weak closure S
of S = {x ∈ X| kxk = 1} is the closed unit ball B̄1 (0). Indeed, since
w
B̄1 (0) is convex the previous lemma shows S ⊆ B̄1 (0). Conversely, if x ∈
Snis not −1
B1 (0) in the weak closure, then there must be an open neighborhood
x + j=1 |`j | ([0, ε)) not contained in the weak closure. Since X is infinite
dimensional we can find a nonzero element x0 ∈ nj=1 Ker(`j ) such that the
T
w
affine line x + tx0 is in this neighborhood and hence also avoids S . But this
is impossible since by the intermediate value theorem there is some t0 > 0
w
such that kx + t0 x0 k = 1. Hence B̄1 (0) ⊆ S . 
Note that this example also shows that in an infinite dimensional space
the weak and norm topologies are always different! In a finite dimensional
space both topologies of course agree.
Corollary 5.13 (Mazur
P k lemma). SupposeP k xk * x, then there are convex
combinations yk = nj=1 λk,j xj (with nj=1 λk,j = 1 and λk,j ≥ 0) such that
yk → x.

Proof. Let K = { nj=1 λj xj |n ∈ N, nj=1 λj = 1, λj ≥ 0} be the convex

P P

hull of the points {xn }. Then by the previous result x ∈ K. 

Example 5.11. Let H be a Hilbert space and {ϕj } some infinite ONS.
Then we already know ϕj * 0. Moreover, the convex combination ψj =
1 Pj
k=1 ϕk → 0 since kψj k = j
−1/2 . 
j

Finally, we note two more important results. For the first note that
since X ∗∗ is the dual of X ∗ it has a corresponding weak-∗ topology and by
5.3. Weak topologies 143

the Banach–Alaoglu theorem B̄1∗∗ (0) is weak-∗ compact and hence weak-∗
closed.
Theorem 5.14 (Goldstine). The image of the closed unit ball B̄1 (0) under
the canonical embedding J into the closed unit ball B̄1∗∗ (0) is weak-∗ dense.

Proof. Let j ∈ B̄1∗∗ (0) be given. Since sets of the form j + nk=1 |`k |−1 ([0, ε))
T
provide a neighborhood base (where we can assume the `k ∈ X ∗ to be
linearly independent without loss of generality) it suffices to find some x ∈
B̄1+ε (0) with `k (x) = j(`k ) for 1 ≤ k ≤ n since then (1 + ε)−1 J(x) will
be in the above neighborhood. Without the requirement kxk ≤ 1 + ε this
follows from surjectivity of the map F : X → Cn , x 7→ (`1 (x), . . . , `n (x)).
Moreover, given
T one such x the same is true for every element from x + Y ,
where Y = k Ker(`k ). So if (x + Y ) ∩ B̄1+ε (0) were empty, we would have
dist(x, Y ) ≥ 1 + ε and by Corollary 4.17 we could find some normalized
` ∈ X ∗ which vanishes on Y and satisfies `(x) ≥ 1 + ε. But by Problem 4.33
we have ` ∈ span(`1 , . . . , `n ) implying
1 + ε ≤ `(x) = j(`) ≤ kjkk`k ≤ 1
a contradiction. 

Note that if B̄1 (0) ⊂ X is weakly compact, then J(B̄1 (0)) is compact
(and thus closed) in the weak-∗ topology on X ∗∗ . Hence Glodstine’s theorem
implies J(B̄1 (0)) = B̄1∗∗ (0) and we get an alternative proof of Kakutani’s
theorem.
Example 5.12. Consider X = c0 (N), X ∗ ∼ = `1 (N), and X ∗∗ ∼
= `∞ (N) with
J corresponding to the inclusion c0 (N) ,→ `∞ (N). Then we can consider the
linear functionals `j (x) = xj which are total in X ∗ and a sequence in X ∗∗
will be weak-∗ convergent if and only if it is bounded and converges when
composed with any of the `j (in other words, when the sequence converges
componentwise — cf. Problem 4.41). So for example, cutting off a sequence
in B̄1∗∗ (0) after n terms (setting the remaining terms equal to 0) we get a
sequence from B̄1 (0) ,→ B̄1∗∗ (0) which is weak-∗ convergent (but of course
not norm convergent). 
Problem 5.10. Show that in an infinite dimensional space, a weakly open
neighborhood of 0 contains a nontrivial subspace. Show the analogue state-
ment for weak-∗ open neighborhoods of 0.
Problem 5.11. Show that a weakly sequentially compact set is bounded.
Problem 5.12. Show that a convex set K ⊆ X is weakly closed if and only
if it is weakly sequentially closed.
Problem 5.13. Show that (5.11) generates the weak topology on B1 (0) ⊂ X.
Show that (5.13) generates the weak topology on B1∗ (0) ⊂ X ∗ .
144 5. Further topics on Banach spaces

Problem 5.14. Show that the neither the weak nor the weak-∗ topology
is first countable if X is infinite dimensional. (Hint: If there is a countable
neighborhood base, you can find, using Problem 5.10, a sequence of unbounded
vectors which converge weakly to zero.)

Problem 5.15. Show that A : D(A) ⊆ X → Y is closed if and only if for

xn ∈ D(A) with xn * x and Axn * y we have x ∈ D(A) and Ax = y.

Problem* 5.16. Show that the annihilator M ⊥ of a set M ⊆ X is weak-∗

weak-∗
closed. Moreover show that (N⊥ )⊥ = span(N ) . In particular, (N⊥ )⊥ =
span(N ) if X is reflexive. (Hint: The first part and hence one inclusion of the
second part are straightforward. For the other inclusion use Corollary 4.19.)

Problem 5.17. Suppose K ⊆ X is convex and x is a boundary point of

K. Then there is a supporting hyperplane at x. That is, there is some
` ∈ X ∗ such that `(x) = 0 and K is contained in the closed half-plane
{y|Re(`(y − x)) ≤ 0}.

5.4. Beyond Banach spaces: Locally convex spaces

We have already seen that it is often important to weaken the notion of
convergence (i.e., to weaken the underlying topology) to get a larger class of
converging sequences. It turns out that all cases considered so far fit within
a general framework which we want to discuss in this section. We start with
an alternate definition of a locally convex vector space which we already
briefly encountered in Corollary 5.4 (equivalence of both definitions will be
established below).
A vector space X together with a topology is called a locally convex
vector space if there exists a family of seminorms {qα }α∈A which generates
the topology in the sense that the topology is the weakest topology for which
the family of functions {qα (.−x)}α∈A,x∈X is continuous. Hence the topology
is generated by sets of the form x + qα−1 (I), where I ⊆ [0, ∞) is open (in the
relative topology). Moreover, sets of the form
n
\
x+ qα−1
j
([0, εj )) (5.14)
j=1

are a neighborhood base at x and hence it is straightforward to check that a

locally convex vector space is a topological vector space, that is, both vector
addition and scalar multiplication are continuous. T For example, if z = x + y
then the preimage of the open neighborhood z + nj=1 qα−1 ([0, ε )) contains
Tn Tn j −1 j
the open neighborhood (x + j=1 qαj ([0, εj /2)), y + j=1 qαj ([0, εj /2))) by
−1

virtue of the triangle inequality. Similarly, if z = γx then the preimage of

5.4. Beyond Banach spaces: Locally convex spaces 145

the open neighborhood z + nj=1 qα−1 ([0, εj )) contains the open neighborhood
T
j
Tn εj ε
(Bε (γ), x + j=1 qαj ([0, 2(|γ|+ε) )) with ε < 2qα j(x) .
−1
j

Moreover, note that a sequence xn will converge to x in this topology if

and only if qα (xn − x) → 0 for all α.
Example 5.13. Of course every Banach space equipped with the norm
topology is a locally convex vector space if we choose the single seminorm
q(x) = kxk. 
Example 5.14. A Banach space X equipped with the weak topology is
a locally convex vector space. In this case we have used the continuous
linear functionals ` ∈ X ∗ to generate the topology. However, note that the
corresponding seminorms q` (x) := |`(x)| generate the same topology since
x + q`−1 ([0, ε)) = `−1 (Bε (x)) in this case. The same is true for X ∗ equipped
with the weak or the weak-∗ topology. 
Example 5.15. The bounded linear operators L (X, Y ) together with the
seminorms qx (A) := kAxk for all x ∈ X (strong convergence) or the semi-
norms q`,x (A) := |`(Ax)| for all x ∈ X, ` ∈ Y ∗ (weak convergence) are locally
convex vector spaces. 
Example 5.16. The continuous functions C(I) together with the pointwise
topology generated by the seminorms qx (f ) := |f (x)| for all x ∈ I is a locally
convex vector space. 
In all these examples we have one additional property which is often
required as part of the definition: The seminorms are called separated if
for every x ∈ X there is a seminorm with qα (x) 6= 0. In this case the
corresponding locally convex space is Hausdorff, since for x 6= y the neigh-
borhoods U (x) = x + qα−1 ([0, ε)) and U (y) = y + qα−1 ([0, ε)) will be disjoint
for ε = 12 qα (x − y) > 0 (the converse is also true; Problem 5.24).
It turns out crucial to understand when a seminorm is continuous.

Lemma 5.15. Let X be a locally convex vector space with corresponding

family of seminorms {qα }α∈A . Then a seminorm q is continuous if and only
if therePare seminorms qαj and constants cj > 0, 1 ≤ j ≤ n, such that
q(x) ≤ nj=1 cj qαj (x).

Proof. If q is continuous, then q −1 (B1 (0)) contains an open neighborhood

of 0 of the form j=1 qαj ([0, εj )) and choosing cj = max1≤j≤n ε−1
Tn −1
j we ob-
Pn
tain that j=1 cj qαj (x) < 1 implies q(x) < 1 and the claim follows from
Problem 5.19. ConverselyTnote that if q(x) = r then Pn q (Bε (r)) con-
−1
n
tains the set U (x) = x + j=1 qαj ([0, εj )) provided
−1
j=1 cj εj ≤ ε since
Pn
|q(y) − q(x)| ≤ q(y − x) ≤ j=1 cj qαj (x − y) < ε for y ∈ U (x). 
146 5. Further topics on Banach spaces

Example 5.17. The weak topology on an infinite dimensional space cannot

be generated by a norm. Indeed,
Tn let q be a continuous seminorm and qαj =
|`αj | as in the lemma. Then j=1 Ker(`αj ) has codimension at most n and
hence contains some x 6= 0 implying that q(x) ≤ nj=1 cj qαj (x) = 0. Thus
P
q is no norm. Similarly, the other examples cannot be generated by a norm
except in finite dimensional cases. 
Moreover, note that the topology is translation invariant in the sense that
U (x) is a neighborhood of x if and only if U (x) − x = {y − x|y ∈ U (x)} is
a neighborhood of 0. Hence we can restrict our attention to neighborhoods
of 0 (this is of course true for any topological vector space). Hence if X
and Y are topological vector spaces, then a linear map A : X → Y will be
continuous if and only if it is continuous at 0. Moreover, if Y is a locally
convex space with respect to some seminorms pβ , then A will be continuous
if and only if pβ ◦ A is continuous for every β (Lemma B.11). Finally, since
pβ ◦ A is a seminorm, the previous lemma implies:

Corollary 5.16. Let (X, {qα }) and (Y, {pβ }) be locally convex vector spaces.
Then a linear map A : X → Y is continuous if and only if for every β
P seminorms qαj and constants cj > 0, 1 ≤ j ≤ n, such that
there are some
pβ (Ax) ≤ nj=1 cj qαj (x).
Pn
It will shorten notation when sums of the type j=1 cj qαj (x), which
appeared in the last two results, can be replaced by a single expression c qα .
This can be done if the family of seminorms {qα }α∈A is directed, that is,
for given α, β ∈ A there is a γ ∈ A such that qα (x) + qβ (x) ≤ Cqγ (x)
for some C P > 0. Moreover, if F(A) is the set of all finite subsets of A,
then {q̃F = α∈F qα }F ∈F (A) is a directed family which generates the same
topology (since every q̃F is continuous with respect to the original family we
do not get any new open sets).
While the family of seminorms is in most cases more convenient to work
with, it is important to observe that different families can give rise to the
same topology and it is only the topology which matters for us. In fact, it
is possible to characterize locally convex vector spaces as topological vector
spaces which have a neighborhood basis at 0 of absolutely convex sets. Here a
set U is called absolutely convex, if for |α|+|β| ≤ 1 we have αU +βU ⊆ U .
Since the sets qα−1 ([0, ε)) are absolutely convex we always have such a basis
in our case. To see the converse note that such a neighborhood U of 0 is
also absorbing (Problem 5.18) und hence the corresponding Minkowski func-
tional (5.1) is a seminorm (Problem 5.23).
T By construction, these seminorms
generate the topology since if U0 = nj=1 qα−1 j
([0, εj )) ⊆ U we have for the
corresponding Minkowski functionals pU (x) ≤ pU0 (x) ≤ ε−1 nj=1 qαj (x),
P

where ε = min εj . With a little more work (Problem 5.22), one can even
5.4. Beyond Banach spaces: Locally convex spaces 147

show that it suffices to assume to have a neighborhood basis at 0 of convex

open sets.
Given a topological vector space X we can define its dual space X ∗ as
the set of all continuous linear functionals. However, while it can happen in
general that the dual space is empty, X ∗ will always be nontrivial for a lo-
cally convex space since the Hahn–Banach theorem can be used to construct
linear functionals (using a continuous seminorm for ϕ in Theorem 4.14) and
also the geometric Hahn–Banach theorem (Theorem 5.3) holds (see also its
corollaries). In this respect note that for every continuous linear functional `
in a topological vector space |`|−1 ([0, ε)) is an absolutely convex open neigh-
borhoods of 0 and hence existence of such sets is necessary for the existence
of nontrivial continuous functionals. As a natural topology on X ∗ we could
use the weak-∗ topology defined to be the weakest topology generated by the
family of all point evaluations qx (`) = |`(x)| for all x ∈ X. Since different
linear functionals must differ at least at one point the weak-∗ topology is
Hausdorff. Given a continuous linear operator A : X → Y between locally
convex spaces we can define its adjoint A0 : Y ∗ → X ∗ as before,
(A0 y ∗ )(x) := y ∗ (Ax). (5.15)
A brief calculation
qx (A0 y ∗ ) = |(A0 y ∗ )(x)| = |y ∗ (Ax)| = qAx (y ∗ ) (5.16)
verifies that A0 is continuous in the weak-∗ topology by virtue of Corol-
lary 5.16.
The remaining theorems we have established for Banach spaces were
consequences of the Baire theorem (which requires a complete metric space)
and this leads us to the question when a locally convex space is a metric
space. From our above analysis we see that a locally convex vector space
will be first countable if and only if countably many seminorms suffice to
determine the topology. In this case X turns out to be metrizable.
Theorem 5.17. A locally convex Hausdorff space is metrizable if and only
if it is first countable. In this case there is a countable family of separated
seminorms {qn }n∈N generating the topology and a metric is given by
1 qn (x − y)
d(x, y) := max . (5.17)
n∈N 2n 1 + qn (x − y)
Proof. If X is first countable there is a countable neighborhood base at 0
and hence also a countable neighborhood base of absolutely convex sets. The
Minkowski functionals corresponding to the latter base are seminorms of the
required type.
Now in this case it is straightforward to check thatT(5.17) defines a metric
(see also Problem B.3). Moreover, the balls Brm (x) = n:2−n >r {y|qn (y−x) <
148 5. Further topics on Banach spaces

r
2−n −r
} are clearly open and convex (note that the intersection is finite). Con-
ε
versely, for every set of the form (5.14) we can choose ε = min{2−αj 1+εj j |1 ≤
j ≤ n} such that Bε (x) will be contained in this set. Hence both topologies
are equivalent (cf. Lemma B.2). 

In general, a locally convex vector space X which has a separated count-

able family of seminorms is called a Fréchet space if it is complete with
respect to the metric (5.17). Note that the metric (5.17) is translation
invariant
d(f, g) = d(f − h, g − h). (5.18)
Example 5.18. The continuous functions C(R) together with local uniform
convergence are a Fréchet space. A countable family of seminorms is for
example
kf kj = sup |f (x)|, j ∈ N. (5.19)
|x|≤j

Then fk → f if and only if kfk − f kj → 0 for all j ∈ N and it follows that

C(R) is complete. 
Example 5.19. The space C ∞ (Rm ) together with the seminorms
X
kf kj,k = sup |∂α f (x)|, j ∈ N, k ∈ N0 , (5.20)
|α|≤k |x|≤j

is a Fréchet space.
Note that ∂α : C ∞ (Rm ) → C ∞ (Rm ) is continuous. Indeed by Corol-
lary 5.16 it suffices to observe that k∂α f kj,k ≤ kf kj,k+|α| . 
Example 5.20. The Schwartz space
S(Rm ) = {f ∈ C ∞ (Rm )| sup |xα (∂β f )(x)| < ∞, ∀α, β ∈ Nm
0 } (5.21)
x

together with the seminorms

qα,β (f ) = kxα (∂β f )(x)k∞ , α, β ∈ Nm
0 . (5.22)
To see completeness note that a Cauchy sequence fn is in particular a
Cauchy sequence in C ∞ (Rm ). Hence there is a limit f ∈ C ∞ (Rm ) such
that all derivatives converge uniformly. Moreover, since Cauchy sequences
are bounded kxα (∂β fn )(x)k∞ ≤ Cα,β we obtain f ∈ S(Rm ).
Again ∂γ : S(Rm ) → S(Rm ) is continuous since qα,β (∂γ f ) ≤ qα,β+γ (f )

γ!
and so is xγ : S(Rm ) → S(Rm ) since qα,β (xγ f ) ≤ η≤β βη (γ−η)!
P
qα+γ−η,β−η (f ).
The dual space S ∗ (Rm ) is known as the space of tempered distribu-
tions. 
Example 5.21. The space of all entire functions f (z) (i.e. functions which
are holomorphic on all of C) together with the seminorms kf kj = sup|z|≤j |f (z)|,
5.4. Beyond Banach spaces: Locally convex spaces 149

j ∈ N, is a Fréchet space. Completeness follows from the Weierstraß conver-

gence theorem which states that a limit of holomorphic functions which is
uniform on every compact subset is again holomorphic. 
Example 5.22. In all of the previous examples the topology cannot be gen-
erated by a norm. For example, if q is a norm for C(R), then by Lemma 5.15
that there is some index j such that q(f ) ≤ Ckf kj . Now choose a nonzero
function which vanishes on [−j, j] to get a contradiction. 
There is another useful criterion when the topology can be described by a
single norm. To this end we call a set B ⊆ X bounded if supx∈B qα (x) < ∞
for every α. By Corollary 5.16 this will then be true for any continuous
seminorm on X.

Theorem 5.18 (Kolmogorov). A locally convex vector space can be gener-

ated from a single seminorm if and only if it contains a bounded open set.

Proof. In a Banach space every open ball is bounded and hence only the
converse direction is nontrivial. So let U be a bounded open set. By shifting
and decreasing U if necessary we can assume U to be an absolutely convex
open neighborhood of 0 and consider the associated Minkowski functional
q = pU . Then since U = {x|q(x) < 1} and supx∈U qα (x) = Cα < ∞ we infer
qα (x) ≤ Cα q(x) (Problem 5.19) and thus the single seminorm q generates
the topology. 

Finally, we mention that, since the Baire category theorem holds for
arbitrary complete metric spaces, the open mapping theorem (Theorem 4.5),
the inverse mapping theorem (Theorem 4.6) and the closed graph theorem
(Theorem 4.7) hold for Fréchet spaces without modifications. In fact, they
are formulated such that it suffices to replace Banach by Fréchet in these
theorems as well as their proofs (concerning the proof of Theorem 4.5 take
into account Problems 5.18 and 5.25).

Problem* 5.18. In a topological vector space every neighborhood U of 0 is

absorbing.

Problem* 5.19. Let p, q be two seminorms. Then p(x) ≤ Cq(x) if and

only if q(x) < 1 implies p(x) < C.

Problem 5.20. Let X be a vector space. We call a set U balanced if

αU ⊆ U for every |α| ≤ 1. Show that a set is balanced and convex if and
only if it is absolutely convex.

Problem* 5.21. The intersection of arbitrary absolutely convex/balanced

sets is again absolutely convex/balanced convex. Hence we can define the
150 5. Further topics on Banach spaces

absolutely convex/balanced hull of a set U as the smallest absolutely con-

vex/balanced set containing U , that is, the intersection of all absolutely con-
vex/balanced sets containing U . Show that the absolutely convex hull is given
by
n
X Xn
ahull(U ) := { λj xj |n ∈ N, xj ∈ U, |λj | ≤ 1}
j=1 j=1
and the balanced hull by
bhull(U ) := {αx|x ∈ U, |α| ≤ 1}.
Show that ahull(U ) = conv(bhull(U )).
Problem* 5.22. In a topological vector space every convex open neighbor-
hood U of zero contains an absolutely convex open neighborhood of zero.
(Hint: By continuity of the scalar multiplication U contains a set of the
form BεC (0) · V , where V is an open neighborhood of zero.)
Problem* 5.23. Let X be a vector space. Show that the Minkowski func-
tional of a balanced, convex, absorbing set is a seminorm.
Problem* 5.24. If a locally convex space is Hausdorff then any correspond-
ing family of seminorms is separated.
Problem* 5.25. Suppose X is Pa complete vector space with a translation
invariant metric d. Show that ∞
j=1 d(0, xj ) < ∞ implies that
∞
X n
X
xj = lim xj
n→∞
j=1 j=1
exists and
∞
X ∞
X
d(0, xj ) ≤ d(0, xj )
j=1 j=1
in this case (compare also Problem 1.5).
Problem 5.26. Instead of (5.17) one frequently uses
X 1 qn (x − y)
˜ y) :=
d(x, .
2n 1 + qn (x − y)
n∈N
Show that this metric generates the same topology.
Consider the Fréchet space C(R) with qn (f ) = sup[−n,n] |f |. Show that
the metric balls with respect to d˜ are not convex.
Problem 5.27. Suppose X is a metric vector space. Then balls are convex
if and only if the metric is quasiconvex:
d(λx + (1 − λ)y, z) ≤ max{d(x, z), d(y, z)}, λ ∈ (0, 1).
(See also Problem 7.14.)
5.5. Uniformly convex spaces 151

Problem 5.28. Consider `p (N) for p ∈ (0, 1) — compare Problem 1.15.

Show that k.kp is not convex. Show that every convex open set is unbounded.
Conclude that it is not a locally convex vector space. (Hint: Consider BR (0).
Then for r < R all vectors which have one entry equal to r and all other
entries zero are in this ball. By taking convex combinations all vectors which
have n entries equal to r/n are in the convex hull. The quasinorm of such a
vector is n1/p−1 r.)

Problem 5.29. Show that Cc∞ (Rm ) is dense in S(Rm ).

Problem 5.30. Let X be a topological vector space and M a closed subspace.

Show that the quotient space X/M is again a topological vector space and that
π : X → X/M is linear, continuous, and open. Show that points in X/M
are closed.

5.5. Uniformly convex spaces

In a Banach space X, the unit ball is convex by the triangle inequality.
Moreover, X is called strictly convex if the unit ball is a strictly convex
set, that is, if for any two points on the unit sphere their average is inside the
unit ball. See Problem 1.13 for some equivalent definitions. This is illustrated
in Figure 1.1 which shows that in R2 this is only true for 1 < p < ∞.
Example 5.23. By Problem 1.13 it follows that `p (N) is strictly convex for
1 < p < ∞ but not for p = 1, ∞. 
A more qualitative notion is to require that if two unit vectors x, y satisfy
kx−yk ≥ ε for some ε > 0, then there is some δ > 0 such that k x+y 2 k ≤ 1−δ.
In this case one calls X uniformly convex and
n  o
δ(ε) := inf 1 − k x+y k kxk = kyk = 1, kx − yk ≥ ε , 0 ≤ ε ≤ 2, (5.23)

2 

is called the modulus of convexity. Of course every uniformly convex space is

strictly convex. In finite dimensions the converse is also true (Problem 5.34).
Note that δ is nondecreasing and
ε
k x+y
2 k = kx −
x−y
2 k ≥1−
2
shows 0 ≤ δ(ε) ≤ 2ε . Moreover, δ(2) = 1 implies X strictly convex. In fact
in this case 1 = δ(2) ≤ 1 − k x+y
2 k ≤ 1 for 2 ≤ kx − yk ≤ 2. That is, x = −y
whenever kx − yk = 2 = kxk + kyk.
Example 5.24. Every q Hilbert space is uniformly convex with modulus of
2
convexity δ(ε) = 1 − 1 − ε4 (Problem 5.32). 
152 5. Further topics on Banach spaces

Example 5.25. Consider C[0, 1] with the norm

Z 1 −1
2
kxk := kxk∞ + kxk2 = max |x(t)| + |x(t)| dt .
t∈[0,1] 0
Note that by kxk2 ≤ kxk∞ this norm is equivalent to the usual one: kxk∞ ≤
kxk ≤ 2kxk∞ . While with the usual norm k.k∞ this space is not strictly
convex, it is with the new one. To see this we use (i) from Problem 1.13.
Then if kx + yk = kxk + kyk we must have both kx + yk∞ = kxk∞ + kyk∞
and kx + yk2 = kxk2 + kyk2 . Hence strict convexity of k.k2 implies strict
convexity of k.k.
Note however, that k.k is not uniformly convex. In fact, since by the
Milman–Pettis theorem below, every uniformly convex space is reflexive,
there cannot be an equivalent norm on C[0, 1] which is uniformly convex (cf.
Example 4.20). 
Example 5.26. It can be shown that `p (N) is uniformly convex for 1 < p <
∞ (see Theorem 3.11 from [47]). 
Equivalently, uniform convexity implies that if the average of two unit
vectors is close to the boundary, then they must be close to each other.
Specifically, if kxk = kyk = 1 and k x+y
2 k > 1 − δ(ε) then kx − yk < ε. The
following result (which generalizes Lemma 4.29) uses this observation:
Theorem 5.19 (Radon–Riesz theorem). Let X be a uniformly convex Ba-
nach space and let xn * x. Then xn → x if and only if lim sup kxn k ≤ kxk.

Proof. If x = 0 there is nothing to prove. Hence we can assume xn 6= 0 for

all n and consider yn := kxxnn k . Then yn * y := kxk
x
and it suffices to show
yn → y. Next choose a linear functional ` ∈ X with k`k = 1 and `(y) = 1.
∗

Then
yn + y
yn + y
` ≤ ≤1
2 2
and letting n → ∞ shows k yn2+y k → 1. Finally uniform convexity shows
yn → y. 

For the proof of the next result we need to following equivalent condition.
Lemma 5.20. Let X be a Banach space. Then
n  o
x+y 
δ(ε) = inf 1 − k 2 k kxk ≤ 1, kyk ≤ 1, kx − yk ≥ ε (5.24)
for 0 ≤ ε ≤ 2.

Proof. It suffices to show that for given x and y which are not both on
the unit sphere there is a better pair in the real subspace spanned by these
vectors. By scaling we could get a better pair if both were strictly inside the
5.5. Uniformly convex spaces 153

unit ball and hence we can assume at least one vector to have norm one, say
kxk = 1. Moreover, consider
cos(t)x + sin(t)y
u(t) := , v(t) := u(t) + (y − x).
k cos(t)x + sin(t)yk
Then kv(0)k = kyk < 1. Moreover, let t0 ∈ ( π2 , 3π 4 ) be the value such that
the line from x to u(t0 ) passes through y. Then, by convexity we must have
kv(t0 )k > 1 and by the intermediate value theorem there is some 0 < t1 < t0
with kv(t1 )k = 1. Let u := u(t1 ), v := v(t1 ). The line through u and x is
not parallel to the line through 0 and x + y and hence there are α, λ ≥ 0
such that
α
(x + y) = λu + (1 − λ)x.
2
Moreover, since the line from x to u is above the line from x to y (since
t1 < t0 ) we have α ≥ 1. Rearranging this equation we get
α
(u + v) = (α + λ)u + (1 − α − λ)x.
2
Now, by convexity of the norm, if λ ≤ 1 we have λ + α > 1 and thus
kλu + (1 − λ)xk ≤ 1 < k(α + λ)u + (1 − α − λ)xk. Similarly, if λ > 1 we
have kλu + (1 − λ)xk < k(α + λ)u + (1 − α − λ)xk again by convexity of the
norm. Hence k 12 (x + y)k ≤ k 12 (u + v)k and u, v is a better pair. 

Now we can prove:

Theorem 5.21 (Milman–Pettis). A uniformly convex Banach space is re-
flexive.

Proof. Pick some x00 ∈ X ∗∗ with kx00 k = 1. It suffices to find some x ∈ B̄1 (0)
with kx00 − J(x)k ≤ ε. So fix ε > 0 and δ := δ(ε), where δ(ε) is the modulus
of convexity. Then kx00 k = 1 implies that we can find some ` ∈ X ∗ with
k`k = 1 and |x00 (`)| > 1 − 2δ . Consider the weak-∗ neighborhood
U := {y 00 ∈ X ∗∗ | |(y 00 − x00 )(`)| < 2δ }
of x00 . By Goldstine’s theorem (Theorem 5.14) there is some x ∈ B̄1 (0) with
J(x) ∈ U and this is the x we are looking for. In fact, suppose this were not
the case. Then the set V := X ∗∗ \ B̄ε∗∗ (J(x)) is another weak-∗ neighborhood
of x00 (since B̄ε∗∗ (J(x)) is weak-∗ compact by the Banach-Alaoglu theorem)
and appealing again to Goldstine’s theorem there is some y ∈ B̄1 (0) with
J(y) ∈ U ∩ V . Since x, y ∈ U we obtain
1− δ
2 < |x00 (`)| ≤ |`( x+y
2 )| +
δ
2 ⇒ 1 − δ < |`( x+y x+y
2 )| ≤ k 2 k,
a contradiction to uniform convexity since kx − yk ≥ ε. 
Problem 5.31. Find an equivalent norm for `1 (N) such that it becomes
strictly convex (cf. Problems 1.13 and 1.17).
154 5. Further topics on Banach spaces

Problem* 5.32. Show that a Hilbert space is uniformly convex. (Hint: Use
the parallelogram law.)
Problem 5.33. A Banach space X is uniformly convex if and only if kxn k =
kyn k = 1 and k xn +y
2 k → 1 implies kxn − yn k → 0.
n

Problem* 5.34. Show that a finite dimensional space is uniformly convex

if and only if it is strictly convex.
Problem 5.35. Let X be strictly convex. Show that every nonzero linear
functional attains its norm for at most one unit vector (cf. Problem 4.18).
 Chapter 6

Bounded linear operators

We have started out our study by looking at eigenvalue problems which, from
a historic view point, were one of the key problems driving the development
of functional analysis. In Chapter 3 we have investigated compact operators
in Hilbert space and we have seen that they allow a treatment similar to
what is known from matrices. However, more sophisticated problems will
lead to operators whose spectra consist of more than just eigenvalues. Hence
we want to go one step further and look at spectral theory for bounded
operators. Here one of the driving forces was the development of quantum
mechanics (there even the boundedness assumption is too much — but first
things first). A crucial role is played by the algebraic structure, namely recall
from Section 1.6 that the bounded linear operators on X form a Banach
space which has a (non-commutative) multiplication given by composition.
In order to emphasize that it is only this algebraic structure which matters,
we will develop the theory from this abstract point of view. While the reader
should always remember that bounded operators on a Hilbert space is what
we have in mind as the prime application, examples will apply these ideas
also to other cases thereby justifying the abstract approach.
To begin with, the operators could be on a Banach space (note that even
if X is a Hilbert space, L (X) will only be a Banach space) but eventually
again self-adjointness will be needed. Hence we will need the additional
operation of taking adjoints.

6.1. Banach algebras

A Banach space X together with a multiplication satisfying

(x + y)z = xz + yz, x(y + z) = xy + xz, x, y, z ∈ X, (6.1)

155
156 6. Bounded linear operators

and
(xy)z = x(yz), α (xy) = (αx)y = x (αy), α ∈ C, (6.2)
and
kxyk ≤ kxkkyk. (6.3)
is called a Banach algebra. In particular, note that (6.3) ensures that
multiplication is continuous (Problem 6.1). In fact, one can show that (sep-
arate) continuity of multiplication implies existence of an equivalent norm
satisfying (6.3) (Problem 6.2).
An element e ∈ X satisfying
ex = xe = x, ∀x ∈ X (6.4)
is called identity (show that e is unique) and we will assume kek = 1 in this
case (by Problem 6.2 this can be done without loss of generality).
Example 6.1. The continuous functions C(I) over some compact interval
form a commutative Banach algebra with identity 1. 
Example 6.2. The differentiable functions C n (I) over some compact inter-
val do not form a commutative Banach algebra since (6.3) fails for n ≥ 1.
However, the equivalent norm
n
X kf (k) k∞
kf k∞,n :=
k!
k=0

remedies this problem. 

Example 6.3. The bounded linear operators L (X) form a Banach algebra
with identity I. 
Example 6.4. The bounded sequences `∞ (N) together with the component-
wise product form a commutative Banach algebra with identity 1. 
Example 6.5. The space of all periodic continuous functions which have an
absolutely convergent Fourier series A together with the norm
X
kf kA := |fˆk |
k∈Z

and the usual product is known as the Wiener algebra. Of course as a

Banach space it is isomorphic to `1 (Z) via the Fourier transform. To see
that it is a Banach algebra note that
X X X
f (x)g(x) = fˆk eikx ĝj eijx = fˆk ĝj ei(k+j)x
k∈Z j∈Z k,j∈Z
XX 
= fˆj ĝk−j e ikx
.
k∈Z j∈Z
6.1. Banach algebras 157

Moreover, interchanging the order of summation

X  X  XX
kf gkA = fˆj ĝk−j  ≤ |fˆj ||ĝk−j | = kf kA kgkA


k∈Z j∈Z j∈Z k∈Z

shows that A is a Banach algebra. The identity is of course given by e(x) ≡ 1.

Moreover, note that A ⊆ Cper [−π, π] and kf k∞ ≤ kf kA . 
Example 6.6. The space L1 (Rn ) together with the convolution
Z Z
(g ∗ f )(x) := g(x − y)f (y)dy = g(y)f (x − y)dy (6.5)
Rn Rn
is a commutative Banach algebra (Problem 6.10) without identity. 
A Banach algebra with identity is also known as unital and we will
assume X to be a Banach algebra with identity e throughout the rest of this
section. Note that an identity can always be added if needed (Problem 6.3).
An element x ∈ X is called invertible if there is some y ∈ X such that
xy = yx = e. (6.6)
In this case y is called the inverse of x and is denoted by x−1 . It is straight-
forward to show that the inverse is unique (if one exists at all) and that
(xy)−1 = y −1 x−1 , (x−1 )−1 = x. (6.7)
In particular, the set of invertible elements G(X) forms a group under mul-
tiplication.
Example 6.7. If X = L (Cn ) is the set of n by n matrices, then G(X) =
GL(n) is the general linear group. 
Example 6.8. Let X = L (`p (N)) and recall the shift operators S ± defined
via (S ± a)j = aj±1 with the convention that a0 = 0. Then S + S − = I but
S − S + 6= I. Moreover, note that S + S − is invertible while S − S + is not. So
you really need to check both xy = e and yx = e in general. 
If x is invertible, then the same will be true all elements in a neigh-
borhood. This will be a consequence from the following straightforward
generalization of the geometric series to our abstract setting.
Lemma 6.1. Let X be a Banach algebra with identity e. Suppose kxk < 1.
Then e − x is invertible and
∞
X
−1
(e − x) = xn . (6.8)
n=0

Proof. Since kxk < 1 the series converges and

∞
X ∞
X ∞
X
n n
(e − x) x = x − xn = e
n=0 n=0 n=1
158 6. Bounded linear operators

respectively
∞ ∞ ∞
!
X X X
n n
x (e − x) = x − xn = e. 
n=0 n=0 n=1

Corollary 6.2. Suppose x is invertible and kx−1 yk < 1 or kyx−1 k < 1.

Then (x − y) is invertible as well and
∞
X ∞
X
(x − y)−1 = (x−1 y)n x−1 or (x − y)−1 = x−1 (yx−1 )n . (6.9)
n=0 n=0

In particular, both conditions are satisfied if kyk < kx−1 k−1 and the set of
invertible elements G(X) is open and taking the inverse is continuous:
kykkx−1 k2
k(x − y)−1 − x−1 k ≤ . (6.10)
1 − kx−1 yk

Proof. Just observe x − y = x(e − x−1 y) = (e − yx−1 )x. 

The resolvent set is defined as

ρ(x) := {α ∈ C|(x − α) is invertible in X} ⊆ C, (6.11)
where we have used the shorthand notation x − α := x − αe. Its complement
is called the spectrum
σ(x) := C \ ρ(x). (6.12)
It is important to observe that the inverse has to exist as an element of
X. That is, if the elements of X are bounded linear operators, it does
not suffice that x − α is injective, as it might not be surjective. If it is
bijective, boundedness of the inverse will come for free from the inverse
mapping theorem.
Example 6.9. If X := L (Cn ) is the space of n by n matrices, then the
spectrum is just the set of eigenvalues. More general, if X are the bounded
linear operators on an infinite-dimensional Hilbert or Banach space, then
every eigenvalue will be in the spectrum but the converse is not true in
general as an injective operator might not be surjective. In fact, this already
can happen for compact operators where 0 could be in the spectrum without
being an eigenvalue. 
Example 6.10. If X := C(I), then the spectrum of a function x ∈ C(I) is
just its range, σ(x) = x(I). Indeed, if α 6∈ Ran(x) then t 7→ (x(t) − α)−1 is
the inverse of x−α (note that Ran(x) is compact). Conversely, if α ∈ Ran(x)
and y were an inverse, then y(t0 )(x(t0 ) − α) = 1 gives a contradiction for
any t0 ∈ I with f (t0 ) = α. 
Example 6.11. If X = A is the Wiener algebra, then, as in the previous
example, every function which vanishes at some point cannot be inverted.
If it does not vanish anywhere, it can be inverted and the inverse will be a
6.1. Banach algebras 159

continuous function. But will it again have a convergent Fourier series, that
is, will it be in the Wiener Algebra? The affirmative answer of this question
is a famous theorem of Wiener, which will be given later in Theorem 6.24. 
The map α 7→ (x − α)−1 is called the resolvent of x ∈ X. If α0 ∈ ρ(x)
we can choose x → x − α0 and y → α − α0 in (6.9) which implies
∞
X
−1
(x − α) = (α − α0 )n (x − α0 )−n−1 , |α − α0 | < k(x − α0 )−1 k−1 . (6.13)
n=0

In particular, since the radius of convergence cannot exceed the distance to

the spectrum (since everything within the radius of convergent must belong
to the resolvent set), we see that the norm of the resolvent must diverge
1
k(x − α)−1 k ≥ (6.14)
dist(α, σ(x))

as α approaches the spectrum. Moreover, this shows that (x − α)−1 has a

convergent power series with coefficients in X around every point α0 ∈ ρ(x).
As in the case of coefficients in C, such functions will be called analytic.
In particular, `((x − α)−1 ) is a complex-valued analytic function for every
` ∈ X ∗ and we can apply well-known results from complex analysis:

Theorem 6.3. For every x ∈ X, the spectrum σ(x) is compact, nonempty

and satisfies
σ(x) ⊆ {α| |α| ≤ kxk}. (6.15)

Proof. Equation (6.13) already shows that ρ(x) is open. Hence σ(x) is
closed. Moreover, x − α = −α(e − α1 x) together with Lemma 6.1 shows
∞
1 X  x n
(x − α)−1 = − , |α| > kxk,
α α
n=0

which implies σ(x) ⊆ {α| |α| ≤ kxk} is bounded and thus compact. More-
over, taking norms shows
∞
−1 1 X kxkn 1
k(x − α) k≤ n
= , |α| > kxk,
|α| |α| |α| − kxk
n=0

which implies (x − α)−1 → 0 as α → ∞. In particular, if σ(x) is empty,

then `((x − α)−1 ) is an entire analytic function which vanishes at infinity.
By Liouville’s theorem we must have `((x − α)−1 ) = 0 for all ` ∈ X ∗ in this
case, and so (x − α)−1 = 0, which is impossible. 
160 6. Bounded linear operators

Example 6.12. The spectrum of the matrix

 
0 1
 0 1 
. .. ...
 
A := 
 

 
 0 1 
−c0 −c1 · · · · · · −cn−1
is given by the zeros of the polynomial (show this)
det(zI − A) = z n + cn−1 z n−1 + · · · + c1 z + c0 .
Hence the fact that σ(A) is nonempty implies the fundamental theorem
of algebra, that every non-constant polynomial has at least one zero. 
As another simple consequence we obtain:
Theorem 6.4 (Gelfand–Mazur). Suppose X is a Banach algebra in which
every element except 0 is invertible. Then X is isomorphic to C.

Proof. Pick x ∈ X and α ∈ σ(x). Then x − α is not invertible and hence

x−α = 0, that is x = α. Thus every element is a multiple of the identity. 

Now we look at functions of x. Given a polynomial p(α) = nj=0 pj αj

P
we of course set
X n
p(x) := pj xj . (6.16)
j=0
In fact, we could easily extend this definition to arbitrary convergent power
series whose radius of convergence is larger than kxk (cf. Problem 1.36).
While this will give a nice functional calculus sufficient for many applications
our aim is the spectral theorem which will allow us to handle arbitrary
continuous functions. Since continuous functions can be approximated by
polynomials by the Weierstraß theorem, polynomials will be sufficient for
now. Moreover, the following result will be one of the two key ingredients
for the proof of the spectral theorem.
Theorem 6.5 (Spectral mapping). For every polynomial p and x ∈ X we
have
σ(p(x)) = p(σ(x)), (6.17)
where p(σ(x)) := {p(α)|α ∈ σ(x)}.

Proof. Let α ∈ σ(x) and observe

p(x) − p(α) = (x − α)q(x).
But since (x − α) is not invertible, the same is true for (x − α)q(x) =
q(x)(x − α) by Problem 6.6 and hence p(α) ∈ p(σ(x)).
6.1. Banach algebras 161

Conversely, let β ∈ σ(p(x)). Then

p(x) − β = a(x − λ1 ) · · · (x − λn )
and at least one λj ∈ σ(x) since otherwise the right-hand side would be
invertible. But then β = p(λj ) ∈ p(σ(x)). 

The second key ingredient for the proof of the spectral theorem is the
spectral radius
r(x) := sup |α| (6.18)
α∈σ(x)
of x. Note that by (6.15) we have
r(x) ≤ kxk. (6.19)
As our next theorem shows, it is related to the radius of convergence of the
Neumann series for the resolvent
∞
−1 1 X  x n
(x − α) = − (6.20)
α α
n=0
encountered in the proof of Theorem 6.3 (which is just the Laurent expansion
around infinity).
Theorem 6.6 (Beurling–Gelfand). The spectral radius satisfies
r(x) = inf kxn k1/n = lim kxn k1/n . (6.21)
n∈N n→∞

Proof. By spectral mapping we have r(x)n = r(xn ) ≤ kxn k and hence

r(x) ≤ inf kxn k1/n .
Conversely, fix ` ∈ X ∗ , and consider
∞
−1 1X 1
`((x − α) )=− `(xn ). (6.22)
α αn
n=0

Then `((x − α)−1 ) is analytic in |α| > r(x) and hence (6.22) converges
absolutely for |α| > r(x) by Cauchy’s integral formula for derivatives. Hence
for fixed α with |α| > r(x), `(xn /αn ) converges to zero for every ` ∈ X ∗ .
Since every weakly convergent sequence is bounded we have
kxn k
≤ C(α)
|α|n
and thus
lim sup kxn k1/n ≤ lim sup C(α)1/n |α| = |α|.
n→∞ n→∞
Since this holds for every |α| > r(x) we have
r(x) ≤ inf kxn k1/n ≤ lim inf kxn k1/n ≤ lim sup kxn k1/n ≤ r(x),
n→∞ n→∞
162 6. Bounded linear operators

which finishes the proof. 

Note that it might be tempting to conjecture that the sequence kxn k1/n
is monotone, however this is false in general – see Problem 6.7. To end this
section let us look at some examples illustrating these ideas.
Example 6.13. In X := C(I) we have σ(x) = x(I) and hence r(x) = kxk∞
for all x. 
Example 6.14. If X := L (C2 ) and x := ( 00 10 ) such that x2 = 0 and
consequently r(x) = 0. This is not surprising, since x has the only eigenvalue
0. In particular, the spectral radius can be strictly smaller then the norm
(note that kxk = 1 in our example). The same is true for any nilpotent
matrix. In general x will be called nilpotent if xn = 0 for some n ∈ N and
any nilpotent element will satisfy r(x) = 0. 
Example 6.15. Consider the linear Volterra integral operator
Z t
K(x)(t) := k(t, s)x(s)ds, x ∈ C([0, 1]), (6.23)
0
then, using induction, it is not hard to verify (Problem 6.9)
kkkn∞ tn
|K n (x)(t)| ≤ kxk∞ . (6.24)
n!
Consequently
kkkn∞
kK n xk∞ ≤ kxk∞ ,
n!
kkkn
that is kK n k ≤ n! ,
∞
which shows
kkk∞
r(K) ≤ lim = 0.
n→∞ (n!)1/n

Hence r(K) = 0 and for every λ ∈ C and every y ∈ C(I) the equation
x − λK x = y (6.25)
has a unique solution given by
∞
X
x = (I − λK)−1 y = λn K n y. (6.26)
n=0
Note that σ(K) = {0} but 0 is in general not an eigenvalue (consider
e.g. k(t, s) = 1). Elements of a Banach algebra with r(x) = 0 are called
quasinilpotent. 
In the last two examples we have seen a strict inequality in (6.19). If we
regard r(x) as a spectral norm for x, then the spectral norm does not control
the algebraic norm in such a situation. On the other hand, if we had equal-
ity for some x, and moreover, this were also true for any polynomial p(x),
then spectral mapping would imply that the spectral norm supα∈σ(x) |p(α)|
6.1. Banach algebras 163

equals the algebraic norm kp(x)k and convergence on one side would imply
convergence on the other side. So by taking limits we could get an isometric
identification of elements of the form f (x) with functions f ∈ C(σ(x)). But
this is nothing but the content of the spectral theorem and self-adjointness
will be the property which will make all this work.
Problem* 6.1. Show that the multiplication in a Banach algebra X is con-
tinuous: xn → x and yn → y imply xn yn → xy.
Problem* 6.2. Suppose that X satisfies all requirements for a Banach al-
gebra except that (6.3) is replaced by
kxyk ≤ Ckxkkyk, C > 0.
Of course one can rescale the norm to reduce it to the case C = 1. However,
this might have undesirable side effects in case there is a unit. Show that if
X has a unit e, then kek ≥ C −1 and there is an equivalent norm k.k0 which
satisfies (6.3) and kek0 = 1.
Finally, note that for this construction to work it suffices to assume that
multiplication is separately continuous by Problem 4.6.
(Hint: Identify x ∈ X with the operator Lx : X → X, y 7→ xy in L (X).
For the last part use the uniform boundedness principle.)
Problem* 6.3 (Unitization). Show that if X is a Banach algebra then
C ⊕ X is a unital Banach algebra, where we set k(α, x)k = |α| + kxk and
(α, x)(β, y) = (αβ, αy + βx + xy).
Problem 6.4. Show σ(x−1 ) = σ(x)−1 if x is invertible.
Problem* 6.5. Suppose x has both a right inverse y (i.e., xy = e) and a
left inverse z (i.e., zx = e). Show that y = z = x−1 .
Problem* 6.6. Suppose xy and yx are both invertible, then so are x and y:
y −1 = (xy)−1 x = x(yx)−1 , x−1 = (yx)−1 y = y(xy)−1 .
(Hint: Previous problem.)
Problem* 6.7. Let X := L (C2 ) and compute kxn k1/n for x := 0 α


β 0 .
Conclude that this sequence is not monotone in general.
Problem 6.8. Let X := `∞ (N). Show σ(x) = {xn }n∈N . Also show that
r(x) = kxk for all x ∈ X.
Problem* 6.9. Show (6.24).
Problem 6.10. Show that L1 (Rn ) with convolution as multiplication is a
commutative Banach algebra without identity (Hint: Lemma 3.20 from [47]).
164 6. Bounded linear operators

Problem 6.11. Show the first resolvent identity

(x − α)−1 − (x − β)−1 = (α − β)(x − α)−1 (x − β)−1
= (α − β)(x − β)−1 (x − α)−1 , (6.27)
for α, β ∈ ρ(x).
Problem 6.12. Show σ(xy) \ {0} = σ(yx) \ {0}. (Hint: Find a relation
between (xy − α)−1 and (yx − α)−1 .)

6.2. The C ∗ algebra of operators and the spectral theorem

We begin by recalling that if H is some Hilbert space, then for every A ∈
L (H) we can define its adjoint A∗ ∈ L (H). Hence the Banach algebra
L (H) has an additional operation in this case which will also give us self-
adjointness, a property which has already turned out crucial for the spectral
theorem in the case of compact operators. Even though this is not imme-
diately evident, in some sense this additional structure adds the convenient
geometric properties of Hilbert spaces to the picture.
A Banach algebra X together with an involution satisfying
(x + y)∗ = x∗ + y ∗ , (αx)∗ = α∗ x∗ , x∗∗ = x, (xy)∗ = y ∗ x∗ , (6.28)
and
kxk2 = kx∗ xk (6.29)
is called a C∗algebra. Any subalgebra (we do not require a subalgebra
to contain the identity) which is also closed under involution, is called a
∗-subalgebra.
The condition (6.29) might look a bit artificial at this point. Maybe
a requirement like kx∗ k = kxk might seem more natural. In fact, at this
point the only justification is that it holds for our guiding example L (H)
(cf. Lemma 2.14). Furthermore, it is important to emphasize that (6.29)
is a rather strong condition as it implies that the norm is already uniquely
determined by the algebraic structure. More precisely, Lemma 6.7 below
implies that the norm of x can be computed from the spectral radius of x∗ x
via kxk = r(x∗ x)1/2 . So while there might be several norms which turn X
into a Banach algebra, there is at most one which will give a C ∗ algebra.
Note that (6.29) implies kxk2 = kx∗ xk ≤ kxkkx∗ k and hence kxk ≤ kx∗ k.
By x∗∗ = x this also implies kx∗ k ≤ kx∗∗ k = kxk and hence
kxk = kx∗ k, kxk2 = kx∗ xk = kxx∗ k. (6.30)
Example 6.16. The continuous functions C(I) together with complex con-
jugation form a commutative C ∗ algebra. 
6.2. The C ∗ algebra of operators and the spectral theorem 165

Example 6.17. The Banach algebra L (H) is a C ∗ algebra by Lemma 2.14.

The compact operators C (H) are a ∗-subalgebra. 
Example 6.18. The bounded sequences `∞ (N) together with complex con-
jugation form a commutative C ∗ algebra. The set c0 (N) of sequences con-
verging to 0 are a ∗-subalgebra. 
If X has an identity e, we clearly have e∗ = e, kek = 1, (x−1 )∗ = (x∗ )−1
(show this), and
σ(x∗ ) = σ(x)∗ . (6.31)
We will always assume that we have an identity and we note that it is always
possible to add an identity (Problem 6.13).
If X is a C ∗ algebra, then x ∈ X is called normal if x∗ x = xx∗ , self-
adjoint if x∗ = x, and unitary if x∗ = x−1 . Moreover, x is called positive if
x = y 2 for some y = y ∗ ∈ X. Clearly both self-adjoint and unitary elements
are normal and positive elements are self-adjoint. If x is normal, then so is
any polynomial p(x) (it will be self-adjoint if x is and p is real-valued).
As already pointed out in the previous section, it is crucial to identify
elements for which the spectral radius equals the norm. The key ingredient
will be (6.29) which implies 2
p kx k = kxk p if x is self-adjoint. For unitary
2

elements we have kxk = kx∗ xk = kek = 1. Moreover, for normal

elements we get
Lemma 6.7. If x ∈ X is normal, then kx2 k = kxk2 and r(x) = kxk.

Proof. Using (6.29) three times we have

kx2 k = k(x2 )∗ (x2 )k1/2 = k(x∗ x)∗ (x∗ x)k1/2 = kx∗ xk = kxk2
k k
and hence r(x) = limk→∞ kx2 k1/2 = kxk. 

The next result generalizes the fact that self-adjoint operators have only
real eigenvalues.
Lemma 6.8. If x is self-adjoint, then σ(x) ⊆ R. If x is positive, then
σ(x) ⊆ [0, ∞).

Proof. Suppose α + iβ ∈ σ(x), λ ∈ R. Then α + i(β + λ) ∈ σ(x + iλ) and

α2 + (β + λ)2 ≤ kx + iλk2 = k(x + iλ)(x − iλ)k = kx2 + λ2 k ≤ kxk2 + λ2 .
Hence α2 + β 2 + 2βλ ≤ kxk2 which gives a contradiction if we let |λ| → ∞
unless β = 0.
The second claim follows from the first using spectral mapping (Theo-
rem 6.5). 
Example 6.19. If X := L (C2 ) and x := ( 00 10 ) then σ(x) = {0}. Hence the
converse of the above lemma is not true in general. 
166 6. Bounded linear operators

Given x ∈ X we can consider the C ∗ algebra C ∗ (x) (with identity)

generated by x (i.e., the smallest closed ∗-subalgebra containing e and x). If
x is normal we explicitly have
C ∗ (x) = {p(x, x∗ )|p : C2 → C polynomial}, xx∗ = x∗ x, (6.32)
and, in particular, C ∗ (x) is commutative (Problem 6.14). In the self-adjoint
case this simplifies to
C ∗ (x) := {p(x)|p : C → C polynomial}, x = x∗ . (6.33)
Moreover, in this case C ∗ (x) is isomorphic to C(σ(x)) (the continuous func-
tions on the spectrum):
Theorem 6.9 (Spectral theorem). If X is a C ∗ algebra and x ∈ X is self-
adjoint, then there is an isometric isomorphism Φ : C(σ(x)) → C ∗ (x) such
that f (t) = t maps to Φ(t) = x and f (t) = 1 maps to Φ(1) = e.
Moreover, for every f ∈ C(σ(x)) we have
σ(f (x)) = f (σ(x)), (6.34)
where f (x) = Φ(f ).

Proof. First of all, Φ is well defined for polynomials p and given by Φ(p) =
p(x). Moreover, since p(x) is normal spectral mapping implies
kp(x)k = r(p(x)) = sup |α| = sup |p(α)| = kpk∞
α∈σ(p(x)) α∈σ(x)

for every polynomial p. Hence Φ is isometric. Next we use that the poly-
nomials are dense in C(σ(x)). In fact, to see this one can either consider
a compact interval I containing σ(x) and use the Tietze extension theo-
rem (Theorem B.29 to extend f to I and then approximate the extension
using polynomials (Theorem 1.3) or use the Stone–Weierstraß theorem (The-
orem B.41). Thus Φ uniquely extends to a map on all of C(σ(x)) by Theo-
rem 1.16. By continuity of the norm this extension is again isometric. Sim-
ilarly, we have Φ(f g) = Φ(f )Φ(g) and Φ(f )∗ = Φ(f ∗ ) since both relations
hold for polynomials.
To show σ(f (x)) = f (σ(x)) fix some α ∈ C. If α 6∈ f (σ(x)), then
1
g(t) = f (t)−α ∈ C(σ(x)) and Φ(g) = (f (x) − α)−1 ∈ X shows α 6∈ σ(f (x)).
Conversely, if α 6∈ σ(f (x)) then g = Φ−1 ((f (x) − α)−1 ) = f −α
1
is continuous,
which shows α 6∈ f (σ(x)). 

In particular, this last theorem tells us that we have a functional calculus

for self-adjoint operators, that is, if A ∈ L (H) is self-adjoint, then f (A) is
well defined for every f ∈ C(σ(A)). Specifically, we can compute f (A) by
choosing a sequence of polynomials pn which converge to f uniformly on
σ(A), then we have pn (A) → f (A) in the operator norm. In particular, if
6.3. Spectral theory for bounded operators 167

f is given by a power series, then f (A) defined via Φ coincides with f (A)
defined via its power series (cf. Problem 1.36).
Problem* 6.13 (Unitization). Show that if X is a non-unital C ∗ algebra
then C⊕X is a unital C ∗ algebra, where we set k(α, x)k := sup{kαy+xyk|y ∈
X, kyk ≤ 1}, (α, x)(β, y) = (αβ, αy + βx + xy) and (α, x)∗ = (α∗ , x∗ ). (Hint:
It might be helpful to identify x ∈ X with the operator Lx : X → X, y 7→ xy
in L (X). Moreover, note kLx k = kxk.)
Problem* 6.14. Let X be a C ∗ algebra and Y a ∗-subalgebra. Show that if
Y is commutative, then so is Y .
Problem 6.15. Show that the map Φ from the spectral theorem is positivity
preserving, that is, f ≥ 0 if and only if Φ(f ) is positive.
Problem 6.16. Let x be self-adjoint. Show that the following are equivalent:
(i) σ(x) ⊆ [0, ∞).
(ii) x is positive.
(iii) kλ − xk ≤ λ for all λ ≥ kxk.
(iv) kλ − xk ≤ λ for one λ ≥ kxk.
Problem 6.17. Let A ∈ L (H). Show that A is normal if and only if
kAuk = kA∗ uk, ∀u ∈ H.
In particular, Ker(A) = Ker(A∗ ). (Hint: Problem 1.20.)
Problem 6.18. Show that the Cayley transform of a self-adjoint element
x,
y = (x − i)(x + i)−1
is unitary. Show that 1 6∈ σ(y) and
x = i(1 + y)(1 − y)−1 .
Problem 6.19. Show if x is unitary then σ(x) ⊆ {α ∈ C||α| = 1}.
Problem 6.20. Suppose x is self-adjoint. Show that
1
k(x − α)−1 k = .
dist(α, σ(x))

6.3. Spectral theory for bounded operators

So far we have developed spectral theory on an algebraic level based on the
fact that bounded operators form a Banach algebra. In this section we want
to take a more operator centered view and consider bounded linear operators
L (X), where X is some Banach space. Now we can make a finer subdivision
of the spectrum based on why our operator fails to have a bounded inverse.
168 6. Bounded linear operators

Since in the bijective case boundedness of the inverse comes for free from
the inverse mapping theorem (Theorem 4.6), there are basically two things
which can go wrong: Either our map is not injective or it is not surjective.
Moreover, in the latter case one can also ask how far it is away from being
surjective, that is, if the range is dense or not. Accordingly one defines the
point spectrum
σp (A) := {α ∈ σ(A)| Ker(A − α) 6= {0}} (6.35)
as the set of all eigenvalues, the continuous spectrum
σc (A) := {α ∈ σ(A) \ σp (A)|Ran(A − α) = X} (6.36)
and finally the residual spectrum
σr (A) := {α ∈ σ(A) \ σp (A)|Ran(A − α) 6= X}. (6.37)
Clearly we have
σ(A) = σp (A) ∪· σc (A) ∪· σr (A). (6.38)
Here the dot indicates that the union is disjoint. Note that in a Hilbert space
σx (A∗ ) = σx (A0 )∗ for x ∈ {p, c, r}.
Example 6.20. Suppose H is a Hilbert space and A = A∗ is self-adjoint.
Then by (2.28), σr (A) = ∅. 
Example 6.21. Suppose X := `p (N) and L is the left shift. Then σ(L) =
B̄1 (0). Indeed, a simple calculation shows that Ker(L − α) = span{(αj )j∈N }
for |α| < 1 if 1 ≤ p < ∞ and for |α| ≤ 1 if p = ∞. Hence σp (L) = B1 (0) for
1 ≤ p < ∞ and σp (L) = B̄1 (0) if p = ∞. In particular, since the spectrum
is closed and kLk P= 1 we have σ(L) = B̄1 (0). Moreover, for y ∈ `c (N)
we set xj := − ∞ k=j α j−k−1 y such that (L − α)x = y. In particular,
k
`c (N) ⊂ Ran(L − α) and hence Ran(S − α) is dense for 1 ≤ p < ∞. Thus
σc (L) = ∂B1 (0) for 1 ≤ p < ∞. Consequently, σr (L) = ∅. 
Since A is invertible if and only if A0 is by Theorem 4.26 we obtain:
Lemma 6.10. Suppose A ∈ L (X). Then
σ(A) = σ(A0 ). (6.39)
Moreover,
σp (A0 ) ⊆ σp (A) ∪· σr (A), σp (A) ⊆ σp (A0 ) ∪· σr (A0 ),
σr (A0 ) ⊆ σp (A) ∪· σc (A), σr (A) ⊆ σp (A0 ), (6.40)
0 0 0
σc (A ) ⊆ σc (A), σc (A) ⊆ σr (A ) ∪· σc (A ).
If in addition, X is reflexive we have σr (A0 ) ⊆ σp (A) as well as σc (A0 ) =
σc (A).

Proof. This follows from Lemma 4.25 and (4.20). In the reflexive case use
A∼= A00 . 
6.3. Spectral theory for bounded operators 169

Example 6.22. Consider L0 from the previous example, which is just the
right shift in `q (N) if 1 ≤ p < ∞. Then σ(L0 ) = σ(L) = B̄1 (0). Moreover,
it is easy to see that σp (L0 ) = ∅. Thus in the reflexive case 1 < p < ∞ we
have σc (L0 ) = σc (L) = ∂B1 (0) as well as σr (L0 ) = σ(L0 ) \ σc (L0 ) = B1 (0).
Otherwise, if p = 1, we only get B1 (0) ⊆ σr (L0 ) and σc (L0 ) ⊆ σc (L) =
∂B1 (0). Hence it remains to investigate Ran(L0 − α) for |α| = 1: If we have
(L0 − α)x = y with some y ∈ `p (N) we must have xj := −α−j−1 jk=1 αk yk .
P
Thus y = (αn )n∈N is clearly not in Ran(L0 − α). Moreover, if ky − ỹk∞ ≤ ε
we have |x̃j | = | jk=1 αk ỹk | ≥ (1 − ε)j and hence ỹ 6∈ Ran(L0 − α), which
P
shows that the range is not dense and hence σr (L0 ) = B̄1 (0), σc (L0 ) = ∅. 
Moreover, for compact operators the spectrum is particularly simple (cf.
also Theorem 3.7). We start with the following observation:

Lemma 6.11. Suppose that K ∈ C (X) and α ∈ C \ {0}. Then Ker(K − α)

is finite dimensional and the range Ran(K − α) is closed.

Proof. For α 6= 0 we can consider I − α−1 K and assume α = 1 without

loss of generality. First of all note that K restricted to Ker(I − K) is the
identity and since the identity is compact the corresponding space must
be finite dimensional by Theorem 4.30. In particular, it is complemented
(Problem 4.26), that is, there exists a closed subspace X0 ⊆ X such that
X = Ker(I − K) u X0 .
To see that Ran(I−K) is closed we consider I+K restricted to X0 which
is injective and has the same range. Hence if Ran(I − K) were not closed,
Corollary 4.10 would imply that there is a sequence xn ∈ X0 with kxn k = 1
and xn − Kxn → 0. By compactness of K we can pass to a subsequence
such that Kxn → y implying xn → y ∈ X0 and hence y ∈ Ker(I − K)
contradicting y ∈ X0 with kyk = 1. 

Next, we want to have a closer look at eigenvalues. Note that eigenvec-

tors corresponding to different eigenvalues are always linearly independent
(Problem 6.22). In Theorem 3.7 we have seen that for a symmetric compact
operator in a Hilbert space we can choose an orthonormal basis of eigenfunc-
tions. Without the symmetry assumption we know that even in the finite
dimensional case we can in general no longer find a basis of eigenfunctions
and that the Jordan canonical form is the best one can do. There the gen-
eralized eigenspaces Ker((A − α)k ) play an important role. In this respect
one looks at the following ascending and descending chains of subspaces
associated to A ∈ L (X) (where we have assumed α = 0 without loss of
generality):
{0} ⊆ Ker(A) ⊆ Ker(A2 ) ⊆ Ker(A3 ) ⊆ · · · (6.41)
170 6. Bounded linear operators

and
X ⊇ Ran(A) ⊇ Ran(A2 ) ⊇ Ran(A3 ) ⊇ · · · (6.42)

We will say that the kernel chain stabilizes at n if Ker(An+1 ) = Ker(An ).

In this case the number n is also called the ascent of A. Substituting x = Ay
in the equivalence An x = 0 ⇔ An+1 x = 0 gives An+1 y = 0 ⇔ An+2 y = 0
and hence by induction we have Ker(An+k ) = Ker(An ) for all k ∈ N0 in this
case. Similarly, will say that the range chain stabilizes at m if Ran(Am+1 ) =
Ran(Am ) and call m the decent of A. Again, if x = Am+2 y ∈ Ran(Am+2 ) we
can write Am+1 y = Am z for some z which shows x = Am+1 z ∈ Ran(Am+1 )
and thus Ran(Am+k ) = Ran(Am ) for all k ∈ N0 in this case. While in a
finite dimensional space both chains eventually have to stabilize, there is no
reason why the same should happen in an infinite dimensional space.
Example 6.23. For the left shift operator L we have Ran(Ln ) = `p (N)
for all n ∈ N while the kernel chain does not stabilize as Ker(Ln ) = {a ∈
`p (N)|aj = 0, j > n}. Similarly, for the right shift operator R we have
Ker(Rn ) = {0} while the range chain does not stabilize as Ran(Rn ) = {a ∈
`p (N)|aj = 0, 1 ≤ j ≤ n}. 

Lemma 6.12. Suppose A : X → X is a linear operator.

(i) The kernel chain stabilizes at n if Ran(An ) ∩ Ker(A) = {0}. Con-

versely, if the kernel chain stabilizes at n, then Ran(An )∩Ker(An ) =
{0}.
(ii) The range chain stabilizes at m if Ker(Am ) + Ran(A) = X. Con-
versely, if the range chain stabilizes at m, then Ker(Am )+Ran(Am ) =
X.
(iii) If both chains stabilize, then m = n and Ker(Am )uRan(Am ) = X.

Proof. (i). If Ran(An ) ∩ Ker(A) = {0} then x ∈ Ker(An+1 ) implies An x ∈

Ran(An ) ∩ Ker(A) = {0} and the kernel chain stabilizes at n. Conversely,
let x ∈ Ran(An ) ∩ Ker(An ), then x = An y and An x = A2n y = 0 implying
y ∈ Ker(A2n ) = Ker(An ), that is, x = An y = 0.
(ii). If Ker(Am ) + Ran(A) = X, then for any x = z + Ay we have
Am x = Am+1 y and hence Ran(Am ) = Ran(Am+1 ). Conversely, if the range
chain stabilizes at m, then Am x = A2m y and x = Am y + (x − Am y).
(iii). Suppose Ran(Am+1 ) = Ran(Am ) but Ker(Am ) ( Ker(Am+1 ). Let
x ∈ Ker(Am+1 ) \ Ker(Am ) and observe that by 0 6= Am x = Am+1 y there
is an x ∈ Ker(Am+2 ) \ Ker(Am+1 ). Iterating this argument would shows
that the kernel chain does not stabilize contradiction our assumption. Hence
n ≤ m.
6.3. Spectral theory for bounded operators 171

Conversely, suppose Ker(An+1 ) = Ker(An ) and Ran(Am+1 ) = Ran(Am )

for m ≥ n. Then
Am x = Am+1 y ⇒ x − Ay ∈ Ker(Am ) = Ker(An ) ⇒ An x = An+1 y
shows Ran(An+1 ) = Ran(An ), that is, m ≤ n. The rest follows by combining
(i) and (ii). 

Note that in case (iii) A is bijective when restricted to Ran(An ) →

Ran(An ) and nilpotent when restricted to Ker(An ) → Ker(An ).
Example 6.24. In a finite dimensional space we are of course always in case
(iii). 
Example 6.25. Let A ∈ L (H) be a self-adjoint operator in a Hilbert space.
Then by (2.28) the kernel chain always stabilizes at n = 1 and the range
chain stabilizes at n = 1 if Ran(A) is closed. 
Now we can apply this to our situation.
Lemma 6.13. Suppose that K ∈ C (X) and α ∈ C\{0}. Then there is some
n = n(α) ∈ N such that Ker(K − α)n = Ker(K − α)n+k and Ran(K − α)n =
Ran(K − α)n+k for every k ≥ 0 and
X = Ker(K − α)n u Ran(K − α)n . (6.43)
Moreover, the space Ker(K−α)m is finite dimensional and the space Ran(K−
m
α) is closed for every m ∈ N.

Proof. Since α 6= 0 we can consider I−α−1 K and assume α = 1 without loss

of generality. Moreover, since (I−K)n −I ∈ C (X) we see that Ker(I−K)n is
finite dimensional and Ran(I − K)n is closed for every n ∈ N. Next suppose
the kernel chain does not stabilize. Abbreviate Kn := Ker(I − K)n . Then,
by Problem 4.32, we can choose xn ∈ Kn+1 \ Kn such that kxn k = 1 and
dist(xn , Kn ) ≥ 21 . But since (I − K)xn ∈ Kn and Kxn ∈ Kn+1 , we see that
1
kKxn − Kxm k = kxn − (I − K)xn − Axm k ≥ dist(xn , Kn ) ≥
2
for n > m and hence the bounded sequence Kxn has no convergent subse-
quence, a contradiction. Consequently the kernel sequence for K 0 also sta-
bilizes, thus by Problem 4.28 Coker(I − K)∗ ∼ = Ker(I − K 0 ) the sequence of
cokernels stabilizes, which finally implies that the range sequence stabilizes.
The rest follows from the previous lemma. 

For an eigenvalue α ∈ C the dimension dim(Ker(A − α)) is called the

geometric multiplicity of α and if the kernel chain stabilizes at n, then
n is called the index of α and dim(Ker(A − α)n ) is called the algebraic
multiplicity of α. Otherwise, if the kernel chain does not stabilize, both
the index and the algebraic multiplicity are set to infinity. The order of a
172 6. Bounded linear operators

generalized eigenvector u corresponding to an eigenvalue α is the smallest n

such that (A − α)n u = 0.
Example 6.26. Consider X := `p (N) and K ∈ C (X) given by (Ka)n =
n an+1 . Then Ka = α a implies an = α (n − 1)!a1 for α 6= 0 (and hence
1 n

a1 = 0) and a = a1 δ for α = 0. Hence σ(K) = {0} with 0 being an

1

eigenvalue of geometric multiplicity one. Since K n a = 0 implies aj = 0 for

1 ≤ j ≤ n we see that its index as well as its algebraic multiplicity is ∞.
Moreover, δ n is a generalized eigenvalue of order n. 

Theorem 6.14 (Spectral theorem for compact operators; Riesz). Suppose

that K ∈ C (X). Then every α ∈ σ(K) \ {0} is an eigenvalue of finite alge-
braic multiplicity and X can de decomposed into invariant closed subspaces
according to (6.43), where n is the index of α. Furthermore, there are at
most countably many eigenvalues which can only accumulate at 0. If X is
infinite dimensional, we have 0 ∈ σ(K). In this case either 0 ∈ σp (K) with
dim(Ker(K)) = ∞ or Ran(K) is not closed.

Proof. That every eigenvalue α 6= 0 has finite algebraic multiplicity follows

from the previous two lemmas. Moreover if Ker(A−α) = {0}, then the kernel
chain stabilizes as n = 1 and hence Ran(A − α) = X, that is α 6∈ σ(K).
Let αn be a sequence of different eigenvalues with |αn | ≥ ε. Let xn
be corresponding normalized eigenvectors and let Xn := span{xj }nj=1 . The
sequence of spaces Xn is increasing and by Problem 4.32 we can choose
normalized vectors x̃n ∈ Xn such that dist(x̃n , Xn−1 ) ≥ 21 . Now kK x̃n −
K x̃m k = kαn x̃n + ((K − αn )x̃n − K x̃m )k ≥ |α2n | ≥ 2ε for m < n and hence
there is no convergent subsequence, a contradiction. Moreover, if 0 ∈ ρ(K)
then K −1 is bounded and hence I = K −1 K is compact, implying that X is
finite dimensional.
Finally, if Ran(K) is closed we can consider the bijective operator K̃ :
X/ Ker(A) → Ran(A) (cf. Problem 1.43) which is again compact. Hence
IX/ Ker(A) = K̃ −1 K̃ is compact and thus X/ Ker(A) is finite dimensional. 

Example 6.27. Note that in contradistinction to the symmetric case, there

might be no eigenvalues at all, as the Volterra integral operator from Exam-
ple 6.15 shows. 
As an immediate consequence we get the famous Fredholm alternative:

Theorem 6.15 (Fredholm alternative). Suppose that K ∈ C (X) and α ∈

C \ {0}. Then, either the inhomogeneous equation

(K − α)x = y (6.44)
6.4. Spectral measures 173

has a unique solution for every y ∈ X or the corresponding homogeneous

equation
(K − α)x = 0 (6.45)
has a nontrivial solution.

In particular, this applies to the case where K is a compact integral

operator (cf. Lemma 3.4), which was the case originally studied by Fredholm.
Problem 6.21. Discuss the spectrum of the right shift R on `1 (N). Show
σ(R) = σr (R) = B̄1 (0) and σp (R) = σc (R) = ∅.
Problem* 6.22. Suppose A ∈ L (X). Show that generalized eigenvectors
corresponding to different eigenvalues or with different order are linearly in-
dependent.
Problem 6.23. Suppose H is a Hilbert space and A ∈ L (H) is normal.
Then σp (A) = σp (A∗ )∗ , σc (A) = σc (A∗ )∗ , and σr (A) = σr (A∗ ) = ∅. (Hint:
Problem 6.17.)
Problem 6.24. Suppose Aj ∈ L (Xj ), j = 1, 2. Then A1 ⊕ A2 ∈ L (X1 ⊕
X2 ) and σ(A1 ⊕ A2 ) = σ(A1 ) ∪ σ(A2 ).
Problem 6.25. Let A : X → Y , B : Y → Z. Show dim(Ker(BA)) ≤
dim(Ker(A)) + dim(Ker(B)) and hence dim(Ker(An )) ≤ n dim(Ker(A)) if
A : X → X.

6.4. Spectral measures

The purpose of this section is to derive another formulation of the spectral
theorem which is important in quantum mechanics. This reformulation re-
quires familiarity with measure theory and can be skipped as the results will
not be needed in the sequel.
Using the Riesz representation theorem we get a formulation in terms of
spectral measures:
Theorem 6.16. Let H be a Hilbert space, and let A ∈ L (H) be self-adjoint.
For every u, v ∈ H there is a corresponding complex Borel measure µu,v
supported on σ(A) (the spectral measure) such that
Z
hu, f (A)vi = f (t)dµu,v (t), f ∈ C(σ(A)). (6.46)
σ(A)
We have
µu,v1 +v2 = µu,v1 + µu,v2 , µu,αv = αµu,v , µv,u = µ∗u,v (6.47)
and |µu,v |(σ(A)) ≤ kukkvk. Furthermore, µu = µu,u is a positive Borel
measure with µu (σ(A)) = kuk2 .
174 6. Bounded linear operators

Proof. Consider the continuous functions on I = [−kAk, kAk] and note that
every f ∈ C(I) gives rise to some f ∈ C(σ(A)) by restricting its domain.
Clearly `u,v (f ) = hu, f (A)vi is a bounded linear functional and the existence
of a corresponding measure µu,v with |µu,v |(I) = k`u,v k ≤ kukkvk follows
from the Riesz representation theorem (Theorem 6.5 from [47]). Since `u,v (f )
depends only on the value of f on σ(A) ⊆ I, µu,v is supported on σ(A).
Moreover, if f ≥ 0 we have `u (f ) = hu, f (A)ui = hf (A)1/2 u, f (A)1/2 ui =
kf (A)1/2 uk2 ≥ 0 and hence `u is positive and the corresponding measure µu
is positive. The rest follows from the properties of the scalar product. 

It is often convenient to regard µu,v as a complex measure on R by using

µu,v (Ω) = µu,v (Ω ∩ σ(A)). If we do this, we can also consider f as a function
on R. However, note that f (A) depends only on the values of f on σ(A)!
Moreover, it suffices to consider µu since using the polarization identity (1.55)
we have
1
µu,v (Ω) = (µu+v (Ω) − µu−v (Ω) + iµu−iv (Ω) − iµu+iv (Ω)). (6.48)
4
Now the last theorem can be used to define f (A) for every bounded mea-
surable function f ∈ B(σ(A)) via Lemma 2.12 and extend the functional
calculus from continuous to measurable functions:
Theorem 6.17 (Spectral theorem). If H is a Hilbert space and A ∈ L (H)
is self-adjoint, then there is an homomorphism Φ : B(σ(A)) → L (H) given
by Z
hu, f (A)vi = f (t)dµu,v (t), f ∈ B(σ(A)). (6.49)
σ(A)
Moreover, if fn (t) → f (t) pointwise and supn kfn k∞ is bounded, then fn (A)u →
f (A)u for every u ∈ H.

Proof. The map Φ is a well-defined linear operator by Lemma 2.12 since we

have Z 
f (t)dµu,v (t) ≤ kf k∞ |µu,v |(σ(A)) ≤ kf k∞ kukkvk
 

σ(A)
and (6.47). Next, observe that Φ(f )∗ = Φ(f ∗ ) and Φ(f g) = Φ(f )Φ(g)
holds at least for continuous functions. To obtain it for arbitrary bounded
functions, choose a (bounded) sequence fn converging to f in L2 (σ(A), dµu )
and observe
Z
k(fn (A) − f (A))uk2 = |fn (t) − f (t)|2 dµu (t)

(use kh(A)uk2 = hh(A)u, h(A)ui = hu, h(A)∗ h(A)ui). Thus fn (A)u →

f (A)u and for bounded g we also have that (gfn )(A)u → (gf )(A)u and
g(A)fn (A)u → g(A)f (A)u. This establishes the case where f is bounded
6.4. Spectral measures 175

and g is continuous. Similarly, approximating g removes the continuity re-

quirement from g.
The last claim follows since fn → f in L2 by dominated convergence in
this case. 

Our final aim is to generalize Corollary 3.8 to bounded self-adjoint op-

erators. Since the spectrum of an arbitrary self-adjoint might contain more
than just eigenvalues we need to replace the sum by an integral. To this end
we recall the family of Borel sets B(R) and begin by defining the spectral
projections
PA (Ω) = χΩ (A), Ω ∈ B(R), (6.50)
such that
µu,v (Ω) = hu, PA (Ω)vi. (6.51)
By χ2Ω = χΩ and = χΩ they are orthogonal projections, that is
χ∗Ω
P = P and P = P . Recall that any orthogonal projection P decomposes
2 ∗

H into an orthogonal sum

H = Ker(P ) ⊕ Ran(P ), (6.52)
where Ker(P ) = (I − P )H, Ran(P ) = P H.
In addition, the spectral projections satisfy
[∞ ∞
X
PA (R) = I, PA ( · Ωn )u = PA (Ωn )u, Ωn ∩ Ωm = ∅, m 6= n, (6.53)
n=1 n=1

for every u ∈ H. Here the dot inside the union just emphasizes that the sets
are mutually disjoint. Such a family of projections is called a projection-
valued measure. Indeed the first claim follows since χR = 1 and by
χΩ1 ∪· Ω2 = χΩ1 + χΩ2 if Ω1 ∩ Ω2 = ∅ the second claim follows at least for
finite unions. The case of
Pcountable unions follows from the last part of the
previous theorem since N χ
n=1 Ωn = χ S
·N → χ · n=1 Ωn pointwise (note
S ∞
n=1 Ωn
that the limit will not be uniform unless the Ωn are eventually empty and
hence there is no chance that this series will converge in the operator norm).
Moreover, since all spectral measures are supported on σ(A) the same is true
for PA in the sense that
PA (σ(A)) = I. (6.54)
I also remark that in this connection the corresponding distribution function
PA (t) := PA ((−∞, t]) (6.55)
is called a resolution of the identity.
Using our projection-valued measure we can define an operator-valued
integral as follows: For every simple function f = nj=1 αj χΩj (where Ωj =
P
176 6. Bounded linear operators

f −1 (αj )), we set

Z n
X
f (t)dPA (t) := αj PA (Ωj ). (6.56)
R j=1

By (6.51) we conclude that this definition agrees with f (A) from Theo-
rem 6.17: Z
f (t)dPA (t) = f (A). (6.57)
R
Extending this integral to functions from B(σ(A)) by approximating such
functions with simple functions we get an alternative way of defining f (A)
for such functions. This can in fact be done by just using the definition of
a projection-valued measure and hence there is a one-to-one correspondence
between projection-valued measures (with bounded support) and (bounded)
self-adjoint operators such that
Z
A = t dPA (t). (6.58)

If PA ({α}) 6= 0, then α is an eigenvalue and Ran(PA ({α})) is the corre-

sponding eigenspace (Problem 6.27). The fact that eigenspaces to different
eigenvalues are orthogonal now generalizes to
Lemma 6.18. Suppose Ω1 ∩ Ω2 = ∅. Then
Ran(PA (Ω1 )) ⊥ Ran(PA (Ω2 )). (6.59)

Proof. Clearly χΩ1 χΩ2 = χΩ1 ∩Ω2 and hence

PA (Ω1 )PA (Ω2 ) = PA (Ω1 ∩ Ω2 ).
Now if Ω1 ∩ Ω2 = ∅, then
hPA (Ω1 )u, PA (Ω2 )vi = hu, PA (Ω1 )PA (Ω2 )vi = hu, PA (∅)vi = 0,
which shows that the ranges are orthogonal to each other. 

Example 6.28. Let A ∈ L (Cn ) be some symmetric matrix and let α1 , . . . , αm

be its (distinct) eigenvalues. Then
m
X
A= αj PA ({αj }),
j=1

where PA ({αj }) is the projection onto the eigenspace Ker(A − αj ) corre-

sponding to the eigenvalue αj by Problem 6.27. In fact, using that PA is
supported on the spectrum, PA (σ(A)) = I, we see
X
PA (Ω) = PA (σ(A))PA (Ω) = PA (σ(A) ∩ Ω) = PA ({αj }).
αj ∈Ω
6.5. The Gelfand representation theorem 177

Hence
Pm using that any f ∈ B(σ(A)) is given as a simple function f =
j=1 f (αj )χ{αj } we obtain
Z m
X
f (A) = f (t)dPA (t) = f (αj )PA ({αj }).
j=1

In particular, for f (t) = t we recover the above representation for A. 

Example 6.29. Let A ∈ L (Cn ) be self-adjoint and let α be an eigenvalue.
Let P = PA ({α}) be the projection onto the corresponding eigenspace and
consider the restriction à = A H̃ onto the orthogonal complement of this


eigenspace H̃ = (1 − P )H. Then by Lemma 6.18 we have µu,v ({α}) = 0 for
u, v ∈ H̃. Hence the integral in (6.49) does not see the point α in the sense
that
Z Z
hu, f (A)vi = f (t)dµu,v (t) = f (t)dµu,v (t), u, v ∈ H̃.
σ(A) σ(A)\{α}

Hence Φ extends to a homomorphism Φ̃ : B(σ(A) \ {α}) → L (H̃). In

particular, if α is an isolated eigenvalue, that is (α − ε, α + ε) ∩ σ(A) = {α}
for ε > 0 sufficiently small, we have (. − α)−1 ∈ B(σ(A) \ {α}) and hence
α ∈ ρ(Ã). 
Problem 6.26. Suppose A is self-adjoint. Let R α be an eigenvalue and u a
corresponding normalized eigenvector. Show f (t)dµu (t) = f (α), that is,
µu is the Dirac delta measure (with mass one) centered at α.
Problem* 6.27. Suppose A is self-adjoint. Show
Ran(PA ({α})) = Ker(A − α).
(Hint: Start by verifying Ran(PA ({α})) ⊆ Ker(A − α). To see the converse,
let u ∈ Ker(A − α) and use the previous example.)

6.5. The Gelfand representation theorem

In this section we look at an alternative approach to the spectral theorem
by trying to find a canonical representation for a Banach algebra. The idea
is as follows: Given the Banach algebra C[a, b] we have a one-to-one corre-
spondence between points x0 ∈ [a, b] and point evaluations mx0 (f ) = f (x0 ).
These point evaluations are linear functionals which at the same time pre-
serve multiplication. In fact, we will see that these are the only (non-
trivial) multiplicative functionals and hence we also have a one-to-one cor-
respondence between points in [a, b] and multiplicative functionals. Now
mx0 (f ) = f (x0 ) says that the action of a multiplicative functional on a func-
tion is the same as the action of the function on a point. So for a general
algebra X we can try to go the other way: Consider the multiplicative func-
tionals m als points and the elements x ∈ X as functions acting on these
178 6. Bounded linear operators

points (the value of this function being m(x)). This gives a map, the Gelfand
representation, from X into an algebra of functions.
A nonzero algebra homeomorphism m : X → C will be called a multi-
plicative linear functional or character:
m(xy) = m(x)m(y), m(e) = 1. (6.60)
Note that the last equation comes for free from multiplicativity since m is
nontrivial. Moreover, there is no need to require that m is continuous as this
will also follow automatically (cf. Lemma 6.20 below).
As we will see, they are closely related to ideals, that is linear subspaces
I of X for which a ∈ I, x ∈ X implies ax ∈ I and xa ∈ I. An ideal is called
proper if it is not equal to X and it is called maximal if it is not contained
in any other proper ideal.
Example 6.30. Let X := C([a, b]) be the continuous functions over some
compact interval. Then for fixed x0 ∈ [a, b], the linear functional mx0 (f ) :=
f (x0 ) is multiplicative. Moreover, its kernel Ker(mx0 ) = {f ∈ C([a, b])|f (x0 ) =
0} is a maximal ideal (we will prove this in more generality below). 
Example 6.31. Let X be a Banach space. Then the compact operators are
a closed ideal in L (X) (cf. Theorem 3.1). 
We first collect a few elementary properties of ideals.
Lemma 6.19. Let X be a unital Banach algebra.
(i) A proper ideal can never contain an invertible element.
(ii) If X is commutative every non-invertible element is contained in
a proper ideal.
(iii) The closure of a (proper) ideal is again a (proper) ideal.
(iv) Maximal ideals are closed.
(v) Every proper ideal is contained in a maximal ideal.

Proof. (i). Let I be a proper ideal. If x ∈ I is invertible then y = x(x−1 y) ∈

I shows I = X. (ii). Consider the ideal xX = {x y|y ∈ X}. Then xX = X
if and only if there is some y ∈ X with xy = e, that is , y = x−1 . (iii) and
(iv). That the closure of an ideal is again an ideal follows from continuity
of the product. Indeed, for a ∈ I choose a sequence an ∈ I converging to a.
Then xan ∈ I → xa ∈ I as well as an x ∈ I → ax ∈ I. Moreover, note that
by Lemma 6.1 all elements in the ball B1 (e) are invertible and hence every
proper ideal must be contained in the closed set X\B1 (e). So the closure
of a proper ideal is proper and any maximal ideal must be closed. (v). To
see that every ideal I is contained in a maximal ideal consider the family
of proper ideals containing I ordered by inclusion. Then, since any union
of a chain of proper ideals is again a proper ideal (that the union is again
6.5. The Gelfand representation theorem 179

an ideal is straightforward, to see that it is proper note that it does not

contain B1 (e)). Consequently Zorn’s lemma implies existence of a maximal
element. 

Note that if I is a closed ideal, then the quotient space X/I (cf. Lemma 1.18)
is again a Banach algebra if we define
[x][y] = [xy]. (6.61)
Indeed (x + I)(y + I) = xy + I and hence the multiplication is well-defined
and inherits the distributive and associative laws from X. Also [e] is an
identity. Finally,
k[xy]k = inf kxy + ak = inf k(x + b)(y + c)k ≤ inf kx + bk inf ky + ck
a∈I b,c∈I b∈I c∈I
= k[x]kk[y]k. (6.62)
In particular, the projection map π : X → X/I is a Banach algebra homo-
morphism.
Example 6.32. Consider the Banach algebra L (X) together with the ideal
of compact operators C (X). Then the Banach algebra L (X)/C (X) is
known as the Calkin algebra. Atkinson’s theorem (Theorem 6.34) says
that the invertible elements in the Calkin algebra are precisely the images of
the Fredholm operators. 
Lemma 6.20. Let X be a unital Banach algebra and m a character. Then
Ker(m) is a maximal ideal and m is continuous with kmk = m(e) = 1.

Proof. It is straightforward to check that Ker(m) is an ideal. Moreover,

every x can be written as
x = m(x)e + y, y ∈ Ker(m).
Let I be an ideal containing Ker(m). If there is some x ∈ I \ Ker(m)
then m(x)−1 x = e + m(x)−1 y ∈ I and hence also e = (e + m(x)−1 y) −
m(x)−1 y ∈ I. Thus Ker(m) is maximal. Since maximal ideals are closed
by the previous lemma, we conclude that m is continuous by Problem 1.37.
Clearly kmk ≥ m(e) = 1. Conversely, suppose we can find some x ∈ X
with kxk < 1 and m(x) = 1. Consequently kxn k ≤ kxkn → 0 contradicting
m(xn ) = m(x)n = 1. 

In a commutative algebra the other direction is also true.

Lemma 6.21. In a commutative unital Banach algebra the characters and
maximal ideals are in one-to-one correspondence.

Proof. We have already seen that for a character m there is corresponding

maximal ideal Ker(m). Conversely, let I be a maximal ideal and consider
180 6. Bounded linear operators

the projection π : X → X/I. We first claim that every nontrivial element in

X/I is invertible. To this end suppose [x0 ] 6= [0] were not invertible. Then
J = [x0 ]X/I is a proper ideal (if it would contain the identity, X/I would
contain an inverse of [x0 ]). Moreover, I 0 = {y ∈ X|[y] ∈ J} is a proper ideal
of X (since e ∈ I 0 would imply [e] ∈ J) which contains I (since [x] = [0] ∈ J
for x ∈ I) but is strictly larger as x0 ∈ I 0 \ I. This contradicts maximality
and hence by the Gelfand–Mazur theorem (Theorem 6.4), every element of
X/I is of the form α[e]. If h : X/I → C, h(α[e]) 7→ α is the corresponding
algebra isomorphism, then m = h ◦ π is a character with Ker(m) = I. 

Now we continue with the following observation: For fixed x ∈ X we

get a map X ∗ → C, ` 7→ `(x). Moreover, if we equip X ∗ with the weak-
∗ topology then this map will be continuous (by the very definition of the
weak-∗ topology). So we have a map X → C(X ∗ ) and restricting this map
to the set of all characters M ⊆ X ∗ (equipped with the relative topology of
the weak-∗ topology) it is known as the Gelfand transform:
Γ : X → C(M), x 7→ x̂(m) = m(x). (6.63)
Theorem 6.22 (Gelfand representation theorem). Let X be a unital Banach
algebra. Then the set of all characters M ⊆ X ∗ is a compact Hausdorff space
with respect to the weak-∗ topology and the Gelfand transform is a continuous
algebra homomorphism with ê = 1.
Moreover, x̂(M) ⊆ σ(x) and hence kx̂k∞ ≤ r(x) ≤ kxk where r(x) is
the spectral radius of x. If X is commutative then x̂(M) = σ(x) and hence
kx̂k∞ = r(x).

Proof. As pointed out before, for fixed x, y ∈ X the map X ∗ → C3 , ` 7→

(`(x), `(y), `(xy)) is continuous and so is the map X ∗ → C, ` 7→ `(x)`(y) −
`(xy) as a composition of continuous maps. Hence the kernel of this map
∗
T x,y = {` ∈ X |`(x)`(y) = `(xy)}
M is weak-∗ closed and so is M = M0 ∩
x,y∈X M x,y where M 0 = {` ∈ X ∗ |`(e) = 1}. So M is a weak-∗ closed subset

of the unit ball in X ∗ and the first claim follows form the Banach–Alaoglu
theorem (Theorem 5.10).
Next (x+y)∧ (m) = m(x+y) = m(x)+m(y) = x̂(m)+ ŷ(m), (xy)∧ (m) =
m(xy) = m(x)m(y) = x̂(m)ŷ(m), and ê(m) = m(e) = 1 shows that the
Gelfand transform is an algebra homomorphism.
Moreover, if m(x) = α then x − α ∈ Ker(m) implying that x − α is
not invertible (as maximal ideals cannot contain invertible elements), that
is α ∈ σ(x). Conversely, if X is commutative and α ∈ σ(x), then x − α is
not invertible and hence contained in some maximal ideal, which in turn is
the kernel of some character m. Whence m(x − α) = 0, that is m(x) = α
for some m. 
6.5. The Gelfand representation theorem 181

Of course this raises the question about injectivity or surjectivity of the

Gelfand transform. Clearly
\
x ∈ Ker(Γ) ⇔ x ∈ Ker(m) (6.64)
m∈M

and it can only be injective if X is commutative. In this case

\
x ∈ Ker(Γ) ⇔ x ∈ Rad(X) := I, (6.65)
I maximal ideal

where Rad(X) is known as the Jacobson radical of X and a Banach algebra

is called semi-simple if the Jacobson radical is zero. So to put this result
to use one needs to understand the set of characters, or equivalently, the set
of maximal ideals. Two examples where this can be done are given below.
The first one is not very surprising.
Example 6.33. If we start with a compact Hausdorff space K and consider
C(K) we get nothing new. Indeed, first of all notice that the map K →
M, x0 7→ mx0 which assigns each x0 the corresponding point evaluation
mx0 (f ) = f (x0 ) is injective and continuous. Hence, by compactness of K, it
will be a homeomorphism once we establish surjectivity (Corollary B.17). To
this end we will show that all maximal ideals are of the form I = Ker(mx0 )
for some x0 ∈ K. So let I be an ideal and suppose there is no point where
all functions vanish. Then for every x ∈ K there is a ball Br(x) (x) and a
function fx ∈ C(K) such that |fx (y)| ≥ 1 for y ∈ Br(x) (x). ByPcompactness
finitely many of these balls will cover K. Now consider f = j fx∗j fxj ∈ I.
Then f ≥ 1 and hence f is invertible, that is I = C(K). Thus maximal
ideals are of the form Ix0 = {f ∈ C(K)|f (x0 ) = 0} which are precisely the
kernels of the characters mx0 (f ) = f (x0 ). Thus M ∼= K as well as fˆ ∼
= f. 
Example 6.34. Consider the Wiener algebra A of all periodic continuous
functions which have an absolutely convergent Fourier series. As in the
previous example it suffices to show that all maximal ideals are of the form
Ix0 = {f ∈ A|f (x0 ) = 0}. To see this set ek (x) = eikx and note kek kA = 1.
Hence for every character m(ek ) = m(e1 )k and |m(ek )| ≤ 1. Since the
last claim holds for both positive and negative k, we conclude |m(ek )| = 1
and thus there is some x0 ∈ [−π, π] with m(ek ) = eikx0 . Consequently
m(f ) = f (x0 ) and point evaluations are the only characters. Equivalently,
every maximal ideal is of the form Ker(mx0 ) = Ix0 .
So, as in the previous example, M ∼= [−π, π] (with −π and π identified)
ˆ ∼
as well hat f = f . Moreover, the Gelfand transform is injective but not
surjective since there are continuous functions whose Fourier series are not
absolutely convergent. Incidentally this also shows that the Wiener algebra
is no C ∗ algebra (despite the fact that we have a natural conjugation which
satisfies kf ∗ kA = kf kA — this again underlines the special role of (6.29)) as
182 6. Bounded linear operators

the Gelfand–Naimark theorem below will show that the Gelfand transform
is bijective for commutative C ∗ algebras. 
Since 0 6∈ σ(x) implies that x is invertible the Gelfand representation
theorem also contains a useful criterion for invertibility.
Corollary 6.23. In a commutative unital Banach algebra an element x is
invertible if and only if m(x) 6= 0 for all characters m.

And applying this to the last example we get the following famous the-
orem of Wiener:
Theorem 6.24 (Wiener). Suppose f ∈ Cper [−π, π] has an absolutely con-
vergent Fourier series and does not vanish on [−π, π]. Then the function f1
also has an absolutely convergent Fourier series.

If we turn to a commutative C ∗ algebra the situation further simplifies.

First of all note that characters respect the additional operation automati-
cally.
Lemma 6.25. If X is a unital C ∗ algebra, then every character satisfies
m(x∗ ) = m(x)∗ . In particular, the Gelfand transform is a continuous ∗-
algebra homomorphism with ê = 1 in this case.

Proof. If x is self-adjoint then σ(x) ⊆ R (Lemma 6.8) and hence m(x) ∈ R

by the Gelfand representation theorem. Now for general x we can write
∗ ∗
x = a + ib with a = x+x2 and b = x−x
2i self-adjoint implying
m(x∗ ) = m(a − ib) = m(a) − im(b) = (m(a) + im(b))∗ = m(x)∗ .
Consequently the Gelfand transform preserves the involution: (x∗ )∧ (m) =
m(x∗ ) = m(x)∗ = x̂∗ (m). 
Theorem 6.26 (Gelfand–Naimark). Suppose X is a unital commutative C ∗
algebra. Then the Gelfand transform is an isometric isomorphism between
C ∗ algebras.

Proof. As in a commutative C ∗ algebra all elements are normal, Lemma 6.7

implies that the Gelfand transform is isometric. Moreover, by the previous
lemma the image of X under the Gelfand transform is a closed ∗-subalgebra
which contains ê ≡ 1 and separates points (if x̂(m1 ) = x̂(m2 ) for all x ∈ X
we have m1 = m2 ). Hence it must be all of C(M) by the Stone–Weierstraß
theorem (Theorem B.41). 

The first moral from this theorem is that from an abstract point of view
there is only one commutative C ∗ algebra, namely C(K) with K some com-
pact Hausdorff space. Moreover, the formulation also very much reassembles
6.5. The Gelfand representation theorem 183

the spectral theorem and in fact, we can derive the spectral theorem by ap-
plying it to C ∗ (x), the C ∗ algebra generated by x (cf. (6.32)). This will even
give us the more general version for normal elements. As a preparation we
show that it makes no difference whether we compute the spectrum in X or
in C ∗ (x).
Lemma 6.27 (Spectral permanence). Let X be a C ∗ algebra and Y ⊆ X
a closed ∗-subalgebra containing the identity. Then σ(y) = σY (y) for every
y ∈ Y , where σY (y) denotes the spectrum computed in Y .

Proof. Clearly we have σ(y) ⊆ σY (y) and it remains to establish the reverse
inclusion. If (y−α) has an inverse in X, then the same is true for (y−α)∗ (y−
α) and (y − α)(y − α)∗ . But the last two operators are self-adjoint and hence
have real spectrum in Y . Thus ((y − α)∗ (y − α) + ni )−1 ∈ Y and letting
n → ∞ shows ((y − α)∗ (y − α))−1 ∈ Y since taking the inverse is continuous
and Y is closed. Similarly ((y −α)(y −α)∗ )−1 ∈ Y and whence (y −α)−1 ∈ Y
by Problem 6.6. 

Now we can show

Theorem 6.28 (Spectral theorem). If X is a C ∗ algebra and x is normal,
then there is an isometric isomorphism Φ : C(σ(x)) → C ∗ (x) such that
f (t) = t maps to Φ(t) = x and f (t) = 1 maps to Φ(1) = e.
Moreover, for every f ∈ C(σ(x)) we have
σ(f (x)) = f (σ(x)), (6.66)
where f (x) = Φ(f ).

Proof. Given a normal element x ∈ X we want to apply the Gelfand–

Naimark theorem in C ∗ (x). By our lemma we have σ(x) = σC ∗ (x) (x). We
first show that we can identify M with σ(x). By the Gelfand representation
theorem (applied in C ∗ (x)), x̂ : M → σ(x) is continuous and surjective.
Moreover, if for given m1 , m2 ∈ M we have x̂(m1 ) = m1 (x) = m2 (x) =
x̂(m2 ) then
m1 (p(x, x∗ )) = p(m1 (x), m1 (x)∗ ) = p(m2 (x), m2 (x)∗ ) = m2 (p(x, x∗ ))
for any polynomial p : C2 → C and hence m1 (y) = m2 (y) for every y ∈ C ∗ (x)
implying m1 = m2 . Thus x̂ is injective and hence a homeomorphism as M
is compact. Thus we have an isometric isomorphism
Ψ : C(σ(x)) → C(M), f 7→ f ◦ x̂,
and the isometric isomorphism we are looking for is Φ = Γ−1 ◦ Ψ. Finally,
σ(f (x)) = σC ∗ (x) (Φ(f )) = σC(σ(x)) (f ) = f (σ(x)). 
184 6. Bounded linear operators

Example 6.35. Let X be a C ∗ algebra and x ∈ X normal. By the spectral

theorem C ∗ (x) is isomorphic to C(σ(x)). Hence every y ∈ C ∗ (x) can be
written as y = f (x) for some f ∈ C(σ(x)) and every character is of the form
m(y) = m(f (x)) = f (α) for some α ∈ σ(x). 
Problem 6.28. Show that C 1 [a, b] is a Banach algebra. What are the char-
acters? Is it semi-simple?
Problem 6.29. Consider the subalgebra of the Wiener algebra consisting
of all functions whose negative Fourier coefficients vanish. What are the
characters? (Hint: Observe that these functions can be identified with holo-
morphicPfunctions inside the unit disc with summable Taylor coefficients via
f (z) = ∞ ˆ k 1
k=0 fk z known as the Hardy space H of the disc.)

6.6. Fredholm operators

In this section we want to investigate solvability of the equation
Ax = y (6.67)
for A ∈ L (X, Y ) given y ∈ Y . Clearly there exists a solution if y ∈ Ran(A)
and this solution is unique if Ker(A) = {0}. Hence these subspaces play
a crucial role. Moreover, if the underlying Banach spaces are finite dimen-
sional, the kernel has a complement X = Ker(A) u X0 and after factor-
ing out the kernel this complement is isomorphic to the range of A. As a
consequence, the dimensions of these spaces are connected by the famous
rank-nullity theorem
dim Ker(A) + dim Ran(A) = dim X (6.68)
from linear algebra. In our infinite dimensional setting (apart from the tech-
nical difficulties that the kernel might not be complemented and the range
might not be closed) this formula does not contain much information, but if
we rewrite it in terms of the index,
ind(A) := dim Ker(A) − dim Coker(A) = dim(X) − dim(Y ), (6.69)
at least the left-hand side will be finite if we assume both Ker(A) and
Coker(A) to be finite dimensional. One of the most useful consequences
of the rank-nullity theorem is that in the case X = Y the index will van-
ish and hence uniqueness of solutions for Ax = y will automatically give
you existence for free (and vice versa). Indeed, for equations of the form
x + Kx = y with K compact originally studied by Fredholm this is still
true by the famous Fredholm alternative (Theorem 6.15). It took a while
until Noether found an example of singular integral equations which have a
nonzero index and started investigating the general case.
We first note that in this case Ran(A) will be automatically closed.
6.6. Fredholm operators 185

Lemma 6.29. Suppose A ∈ L (X, Y ) with finite dimensional cokernel.

Then Ran(A) is closed.

Proof. First of all note that the induced map à : X/ Ker(A) → Y is in-
jective (Problem 1.43). Moreover, the assumption that the cokernel is finite
says that there is a finite subspace Y0 ⊂ Y such that Y = Y0 u Ran(A).
Then
 : X/ Ker(A) ⊕ Y0 → Y, Â(x, y) = Ãx + y
is bijective and hence a homeomorphism by Theorem 4.6. Since X is a closed
subspace of X ⊕ Y0 we see that Ran(A) = Â(X) is closed in Y . 

Hence we call an operator A ∈ L (X, Y ) a Fredholm operator (also

Noether operator) if both its kernel and cokernel are finite dimensional.
In this case we define its index as
ind(A) := dim Ker(A) − dim Coker(A). (6.70)
The set of Fredholm operators will be denoted by Φ(X, Y ) and the set of
Fredholm operators with index zero will be denoted by Φ0 (X, Y ). An im-
mediate consequence of Theorem 4.26 is:
Theorem 6.30 (Riesz). A bounded operator A is Fredholm if and only if A0
is and in this case
ind(A0 ) = − ind(A). (6.71)

Note that by Problem 4.28 we have

ind(A) = dim Ker(A) − dim Ker(A0 ) = dim Coker(A0 ) − dim Coker(A)
(6.72)
since for a finite dimensional space the dual space has the same dimension.
Example 6.36. The right shift operator S in X = Y = `p (N), 1 ≤ p < 1
is Fredholm. In fact, recall that S 0 is the right shift and Ker(S) = {0},
Ker(S 0 ) = span{δ 1 }. In particular, ind(S) = 1 and ind(S 0 ) = −1. 
In the case of Hilbert spaces Ran(A) closed implies H = Ran(A) ⊕
Ran(A)⊥ and thus Coker(A) ∼ = Ran(A)⊥ . Hence an operator is Fredholm if
Ran(A) is closed and Ker(A) and Ran(A)⊥ are both finite dimensional. In
this case
ind(A) = dim Ker(A) − dim Ran(A)⊥ (6.73)
and ind(A∗ ) = − ind(A) as is immediate from (2.28).
Example 6.37. Suppose H is a Hilbert space and A = A∗ is a self-adjoint
Fredholm operator, then (2.28) shows that ind(A) = 0. In particular, a self-
adjoint operator is Fredholm if dim Ker(A) < ∞ and Ran(A) is closed. For
example, according to Example 6.29, A − λ is Fredholm if λ is an eigenvalue
186 6. Bounded linear operators

of finite multiplicity (in fact, inspecting this example shows that the converse
is also true).
It is however important to notice that Ran(A)⊥ finite dimensional does
not imply Ran(A) closed! For example consider (Ax)n = n1 xn in `2 (N) whose
range is dense but not closed. 
Another useful formula concerns the product of two Fredholm operators.
For its proof it will be convenient to use the notion of an exact sequence:
Let Xj be Banach spaces. A sequence of operators Aj ∈ L (Xj , Xj+1 )
A
1 2 A n A
X1 −→ X2 −→ X3 · · · Xn −→ Xn+1

is said to be exact if Ran(Aj ) = Ker(Aj+1 ) for 0 ≤ j < n. We will also

need the following two easily checked facts: If Xj−1 and Xj+1 are finite
dimensional, so is Xj (Problem 6.30) and if the sequence of finite dimensional
spaces starts with X0 = {0} and ends with Xn+1 = {0}, then the alternating
sum over the dimensions vanishes (Problem 6.31).

Lemma 6.31 (Atkinson). Suppose A ∈ L (X, Y ), B ∈ L (Y, Z). If two of

the operators A, B, AB are Fredholm, so is the third and we have

ind(AB) = ind(A) + ind(B). (6.74)

Proof. It is straightforward to check that the sequence

A
0 −→ Ker(A) −→ Ker(BA) −→ Ker(B) −→ Coker(A)
B
−→ Coker(BA) −→ Coker(B) −→ 0

is exact. Here the maps which are not explicitly stated are canonical inclu-
sions/projections. Hence by Problem 6.30, if two operators are Fredholm,
so is the third. Moreover, the formula for the index follows from Prob-
lem 6.31. 

Next we want to look a bit further into the structure of Fredholm op-
erators. First of all, since Ker(A) is finite dimensional it is complemented
(Problem 4.26), that is, there exists a closed subspace X0 ⊆ X such that
X = Ker(A)uX0 and a corresponding projection P ∈ L (X) with Ran(P ) =
Ker(A). Similarly, Ran(A) is complemented (Problem 1.44) and there exists
a closed subspace Y0 ⊆ Y such that Y = Y0 u Ran(A) and a corresponding
projection Q ∈ L (Y ) with Ran(Q) = Y0 . With respect to the decomposition
Ker(A) ⊕ X0 → Y0 ⊕ Ran(A) our Fredholm operator is given by
 
0 0
A= , (6.75)
0 A0
6.6. Fredholm operators 187

where A0 is the restriction of A to X0 → Ran(A). By construction A0 is

bijective and hence a homeomorphism (Theorem 4.6). Defining
 
0 0
B := (6.76)
0 A−1
0
we get
AB = I − Q, BA = I − P (6.77)
and hence A is invertible up to finite rank operators. Now we are ready for
showing that the index is stable under small perturbations.
Theorem 6.32 (Dieudonné). The set of Fredholm operators Φ(X, Y ) is open
in L (X, Y ) and the index is locally constant.

Proof. Let C ∈ L (X, Y ) and write it as

 
C11 C12
C=
C21 C22
with respect to the above splitting. Then if kC22 k < kA−1
0 k
−1 we have that

A0 + C22 is still invertible (Problem 1.35). Now introduce

I −C12 (A0 + C22 )−1
   
I 0
D1 = , D2 =
0 I −(A0 + C22 )−1 C21 I
and observe
C11 − C12 (A0 + C22 )−1 C21
 
0
D1 (A + C)D2 = .
0 A0 + C22
Since D1 , D2 are homeomorphisms we see that A + C is Fredholm since the
right-hand side obviously is. Moreover, ind(A + C) = ind(C11 − C12 (A0 +
C22 )−1 C21 ) = dim(Ker(A)) − dim(Y0 ) = ind(A) since the second operator
is between finite dimensional spaces and the index can be evaluated using
(6.69). 

Since the index is locally constant, it is constant on every connected

component of Φ(X, Y ) which often is useful for computing the index. The
next result identifies an important class of Fredholm operators and uses this
observation for computing the index.
Theorem 6.33 (Riesz). For every K ∈ C (X) we have I − K ∈ Φ0 (X).

Proof. That I−K is Fredholm follows from Lemma 6.11 since K 0 is compact
as well and Coker(I − K)∗ ∼= Ker(I − K 0 ) by Problem 4.28. Furthermore, the
index is constant along [0, 1] → Φ(X), α 7→ I − αK and hence ind(I − K) =
ind(I) = 0. 

Next we show that an operator is Fredholm if and only if it has a left/right

inverse up to compact operators.
188 6. Bounded linear operators

Theorem 6.34 (Atkinson). An operator A ∈ L (X, Y ) is Fredholm if and

only if there exist B1 , B2 ∈ L (Y, X) such that B1 A−I ∈ C (X) and AB2 −I ∈
C (Y ).

Proof. If A is Fredholm we have already given an operator B in (6.76)

such that BA − I and AB − I are finite rank. Conversely, according to
Theorem 6.33 B1 A and AB2 are Fredholm. Since Ker(A) ⊆ Ker(B2 A) and
Ran(AB2 ) ⊆ Ran(A) this shows that A is Fredholm. 

Operators B1 and B2 as in the previous theorem are also known as a

left and right parametrix, respectively. As a consequence we can now
strengthen Theorem 6.33:
Corollary 6.35 (Yood). For every A ∈ Φ(X, Y ) and K ∈ C (X, Y ) we have
A + K ∈ Φ(X, Y ) with ind(A + K) = ind(A).

Proof. Using (6.76) we see that B(A + K) − I = −P + BK ∈ C (Y ) and

(A + K)B − I = −Q + KB ∈ C (X) and hence A + K is Fredholm. In
fact, A + αK is Fredholm for α ∈ [0, 1] and hence ind(A + K) = ind(A) by
continuity of the index. 

Fredholm operators are also used to split the spectrum. For A ∈ L (X)
one defines the essential spectrum
σess (A) := {α ∈ C|A − α 6∈ Φ0 (X)} ⊆ σ(A) (6.78)
and the Fredholm spectrum
σΦ (A) := {α ∈ C|A − α 6∈ Φ(X)} ⊆ σess (A). (6.79)
By Dieudonné’s theorem both σess (A) and σΦ (A) are closed. Warning:
These definitions are not universally accepted and several variants can be
found in the literature.
Example 6.38. let X be infinite dimensional and K ∈ C (X). Then σess (K) =
σΦ (K) = {0}. 
By Corollary 6.35 both the Fredholm spectrum and the essential spec-
trum are invariant under compact perturbations:
Theorem 6.36 (Weyl). Let A ∈ L (X), then
σΦ (A + K) = σΦ (A), σess (A + K) = σess (A), K ∈ C (X). (6.80)

Note that if X is a Hilbert space and A is self-adjoint, then σ(A) ⊆ R

and hence ind(A − α) = − ind((A − α)∗ ) = ind(A − α∗ ) shows that the
index is always zero for α 6∈ σΦ (A). Thus σess (A) = σΦ (A) for self-adjoint
operators.
6.6. Fredholm operators 189

Moreover, if α ∈ σess (A), using the notation from (6.75) for A − α, we

can find an invertible operator K0 : Ker(A) → Y0 and extend it to a finite
rank operator K : X → X by setting it equal to 0 on X0 . Then (A + K) − α
has a bounded inverse implying α ∈ ρ(A + K). Thus the essential spectrum
is precisely the part which is stable under compact perturbations
\
σess (A) = σ(A + K). (6.81)
K∈C (X)

The complement is called the discrete spectrum

σd (A) := σ(A) \ σess (A) = {α ∈ σ(A)|A − α ∈ Φ0 (X)}. (6.82)
Clearly points in the discrete spectrum are eigenvalues with finite geometric
multiplicity. Moreover, if the algebraic multiplicity of an eigenvalue in σd (A)
is finite, then by Problem 6.33 and Lemma 6.12 there is an n such that X =
Ker((A−α)n )uRan((A−α)n ). These spaces are invariant and Ker((A−α)n )
is still finite dimensional (since (A − α)n is also Fredholm). With respect
to this decomposition A has a simple block structure with the first block
A0 : Ker((A − α)n ) → Ker((A − α)n ) such that A0 − α is nilpotent and the
second block A1 : Ran((A − α)n ) → Ran((A − α)n ) such that α ∈ ρ(A0 ).
Hence for sufficiently small ε > 0 we will have α+ε ∈ ρ(A0 ) and α+ε ∈ ρ(A1 )
implying α + ε ∈ ρ(A) such that α is an isolated point of the spectrum. This
happens for example if A is self-adjoint (in which case n = 1). However, in
the general case, σd could contain much more than just isolated eigenvalues
with finite algebraic multiplicity as the following example shows.
Example 6.39. Let X = `2 (N) ⊕ `2 (N) with A = L ⊕ R, where L, R are the
left, right shift, respectively. Explicitly, A(x, y) = ((x2 , x3 , . . . ), (0, y1 , y2 , . . . )).
Then σ(A) = σ(L) ∪ σ(R) = B̄1 (0) and σp (A) = σp (L) ∪ σp (R) = B1 (0).
Moreover, note that A ∈ Φ0 and that 0 is an eigenvalue of infinite algebraic
multiplicity.
Now consider the rank-one operator K(x, y) := ((0, 0, . . . ), (x1 , 0, . . . ))
such that (A + K)(x, y) = ((x2 , x3 , . . . ), (x1 , y1 , y2 , . . . )). Then A + K is
unitary (note that this is essentially a two-sided shift) and hence σ(A+K) ⊆
∂B1 (0). Thus σd (A) = B1 (0) and σess (A) = ∂B1 (0). 
A B
Problem* 6.30. Suppose X −→ Y −→ Z is exact. Show that if X and Z
are finite dimensional, so is Y .
Problem* 6.31. Let Xj be finite dimensional vector spaces and suppose
1 A 2 A An−1
0 −→ X1 −→ X2 −→ X3 · · · Xn−1 −→ Xn −→ 0
is exact. Show that
n
X
(−1)j dim(Xj ) = 0.
j=1
190 6. Bounded linear operators

(Hint: Rank-nullity theorem.)

Problem 6.32. Let A ∈ Φ(X, Y ) with a corresponding parametrix B1 , B2 ∈
L (Y, X). Set K1 := I − B1 A ∈ C (X), K2 := I − AB2 ∈ C (Y ) and show
B1 − B = B1 Q − K1 B ∈ C (Y, X), B2 − B = P B2 − BK2 ∈ C (Y, X).
Hence a parametrix is unique up to compact operators. Moreover, B1 , B2 ∈
Φ(Y, X).
Problem* 6.33. Suppose A ∈ Φ(X). If the kernel chain stabilizes then
ind(A) ≤ 0. If the range chain stabilizes then ind(A) ≥ 0. So if A ∈ Φ0 (X),
then the kernel chain stabilizes if and only if the range chain stabilizes.
Problem 6.34. The definition of a Fredholm operator can be literally ex-
tended to unbounded operators, however, in this case A is additionally as-
sumed to be densely defined and closed. Let us denote A regarded as an
operator from D(A) equipped with the graph norm k.kA to X by Â. Recall
that  is bounded (cf. Problem 4.13). Show that a densely defined closed
operator A is Fredholm if and only if  is.
Part 2

Nonlinear Functional
Analysis
 Chapter 7

Analysis in Banach
spaces

7.1. Single variable calculus in Banach spaces

As a warmup we will look at mappings from an interval to a Banach space.
This case is somewhat simpler then the case of mappings between Banach
spaces but nevertheless is sufficient for many applications.
Let X be a Banach space. Let I ⊆ R be some interval and denote by
C(I, X) the set of continuous functions from I to X. Given t ∈ I we call
f : I → X differentiable at t if the limit

f (t + ε) − f (t)
f˙(t) := lim (7.1)
ε→0 ε
exists. If t is a boundary point, the limit/derivative is understood as the
corresponding onesided limit/derivative.
The set of functions f : I → X which are differentiable at all t ∈ I and
for which f˙ ∈ C(I, X) is denoted by C 1 (I, X). Clearly C 1 (I, X) ⊂ C(I, X).
As usual we set C k+1 (I, X) := {f ∈ C 1 (I, X)|f˙ ∈ C k (I, X)}. Note that if
U ∈ L (X, Y ) and f ∈ C k (I, X), then U f ∈ C k (I, Y ) and dt
d
U f = U f˙.
The following version of the mean value theorem will be crucial.

Theorem 7.1 (Mean value theorem). Suppose f ∈ C 1 (I, X). Then

kf (t) − f (s)k ≤ M |t − s|, M := sup kf˙(τ )k, (7.2)

τ ∈[s,t]

for s ≤ t ∈ I.

193
194 7. Analysis in Banach spaces

Proof. Fix M̃ > M and consider d(τ ) := kf (τ ) − f (s)k − M̃ (τ − s) for

τ ∈ [s, t]. Suppose τ0 is the largest τ for which d(τ ) ≤ 0 holds. So there is a
sequence εn ↓ 0 such that for sufficiently large n
0 < d(τ0 + εn ) = kf (τ0 + εn ) − f (τ0 ) + f (τ0 ) − f (s)k − M̃ (τ0 + εn − s)
≤ kf (τ0 + εn ) − f (τ0 )k − M̃ εn = kf˙(τ0 )εn + o(εn )k − M̃ εn
≤ (M − M̃ )εn + o(εn ) < 0.
This contradicts our assumption. 

In particular,
Corollary 7.2. For f ∈ C 1 (I, X) we have f˙ = 0 if and only if f is constant.

Next we turn to integration. Let I := [a, b] be compact. A function

f : I → X is called a step function provided there are numbers
t0 = a < t1 < t2 < · · · < tn−1 < tn = b (7.3)
such that f (t) is constant on each of the open intervals (tj−1 , tj ). The set
of all step functions S(I, X) forms a linear space and can be equipped with
the sup norm. The corresponding Banach space obtained after completion is
called the set of regulated functions R(I, X). In other words, a regulated
function is the uniform limit of a step function.
Observe that C(I, X) ⊂ R(I, X). In fact, consider the functions fn :=
Pn−1
j=0 f (tj )χ[tj ,tj+1 ) ∈ S(I, X), where tj = a + j n and χ is the character-
b−a

istic function. Since f ∈ C(I, X) is uniformly continuous, we infer that fn

converges uniformly to f . Slightly more general, note that piecewise contin-
uous functions are regulated since every piecewise continuous function is the
sum of a continuous function and a step function.
For a step function f ∈ S(I, X) we can define a linear map : S(I, X) →
R

X by
Z b n
X
f (t)dt := xj (tj − tj−1 ), (7.4)
a j=1

where xi is the value of f on (tj−1 , tj ). This map satisfies

Z b
(7.5)

f (t)dt ≤ kf k∞ (b − a).

a

and hence it can be extended uniquely to a linear map : R(I, X) → X

R

with the same norm (b − a). We even have

Z b Z b
(7.6)

f (t)dt ≤ |f (t)|dt

a a
7.1. Single variable calculus in Banach spaces 195

since this holds for simple functions by the triangle inequality and hence for
all functions by approximation.
We remark that it is possible to extend the integral to a larger class of
functions in various ways. The first generalization is to replace step functions
by simple functions (and at the same time one could also replace the Lebesgue
measure on I by an arbitrary finite measure). Then the same approach
defines the integral for uniform limits of simple functions. However, things
only get interestingR when you also replace the sup norm by an L1 type
seminorm: kf k1 := kf (x)k dµ(x). As before the integral can be extended
to all functions which can be approximated by simple functions with respect
to this seminorm. This is know as the Bochner integral and we refer to
Section 5.5 from [47] for details.
In addition, if A ∈ L (X, Y ), then f ∈ R(I, X) implies Af ∈ R(I, Y )
and
Z b Z b
A f (t)dt = A f (t)dt. (7.7)
a a
Again this holds for step functions and thus extends to all regulated functions
by continuity. In particular, if ` ∈ X ∗ is a continuous linear functional, then
Z b Z b
`( f (t)dt) = `(f (t))dt, f ∈ R(I, X). (7.8)
a a
Rt
Moreover, we will use the usual conventions t12 f (s)ds := I χ(t1 ,t2 ) (s)f (s)ds
R
Rt Rt
and t21 f (s)ds := − t12 f (s)ds. Note that we could replace (t1 , t2 ) by a
closed or half-open interval with the same endpoints (why?) and hence
R t3 R t2 R t3
t1 f (s)ds = t1 f (s)ds + t2 f (s)ds.

Theorem 7.3 (Fundamental theorem of calculus). Suppose F ∈ C 1 (I, X),

then
Z t
F (t) = F (a) + Ḟ (s)ds. (7.9)
a
Rt
Conversely, if f ∈ C(I, X), then F (t) = a f (s)ds ∈ C 1 (I, X) and Ḟ (t) =
f (t).
Rt
Proof. Let f ∈ C(I, X) and set G(t) := a f (s)ds. Then G ∈ C 1 (I, X)
with Ġ(t) = f (t) as can be seen from
Z t+ε Z t Z t+ε
k f (s)ds − f (s)ds − f (t)εk = k (f (s) − f (t))dsk
a a t
≤ |ε| sup kf (s) − f (t)k.
s∈[t,t+ε]
196 7. Analysis in Banach spaces

Rt
Hence if F ∈ C 1 (I, X) then G(t) := a (Ḟ (s))ds satisfies Ġ = Ḟ and hence
F (t) = C + G(t) by Corollary 7.2. Choosing t = a finally shows F (a) =
C. 
Problem* 7.1 (Product rule). Let X be a Banach algebra. Show that if
d
f, g ∈ C 1 (I, X) then f g ∈ C 1 (I, X) and dt f g = f˙g + f ġ.
Problem* 7.2. Let f ∈ R(I, X) and I˜ := I + t0 . then f (t − t0 ) ∈ R(I,
˜ X)
and Z Z
f (t)dt = f (t − t0 )dt.
I I˜

Problem* 7.3. Let A : D(A) ⊆ X → X be a closed operator. Show that

(7.7) holds for f ∈ C(I, X) with Ran(f ) ⊆ D(A) and Af ∈ C(I, X).
Problem 7.4. Let I = [a, b] and J = [c, d] be two compact intervals. Suppose
f (s, t) : I × J → X is regulated in the sense that it is a uniform limit of step
functions being constant on disjoint open rectangles (sj−1 , sj ) × (tk−1 , tk )
whose closure cover I × J. Show that
Z Z  Z Z 
f (s, t)ds dt = f (s, t)dt ds.
J I I J
(Hint: One way is to use linear functionals and reduce it to the classical
Fubini theorem.)

7.2. Multivariable calculus in Banach spaces

We now turn to calculus in Banach spaces. Most facts will be similar to
the situation of multivariable calculus for functions from Rn to Rm . To
emphasize this we will use |.| for the norm in this section.
Let X and Y be two Banach spaces and let U be an open subset of X.
Denote by C(U, Y ) the set of continuous functions from U ⊆ X to Y and by
L (X, Y ) ⊂ C(X, Y ) the Banach space of bounded linear functions equipped
with the operator norm
kLk := sup |Lu|. (7.10)
|u|=1
Then a function F : U → Y is called differentiable at x ∈ U if there exists a
linear function dF (x) ∈ L (X, Y ) such that
F (x + u) = F (x) + dF (x) u + o(u), (7.11)
where o, O are the Landau symbols. Explicitly
|F (x + u) − F (x) − dF (x) u|
lim = 0. (7.12)
u→0 |u|
The linear map dF (x) is called the Fréchet derivative of F at x. It is
uniquely defined since if dG(x) were another derivative we had (dF (x) −
7.2. Multivariable calculus in Banach spaces 197

dG(x))u = o(u) implying that for every ε > 0 we can find a δ > 0 such that
|(dF (x)−dG(x))u| ≤ ε|u| whenever |u| ≤ δ. By homogeneity of the norm we
conclude kdF (x) − dG(x)k ≤ ε and since ε > 0 is arbitrary dF (x) = dG(x).
Note that for this argument to work it is crucial that we can approach x
from arbitrary directions u, which explains our requirement that U should
be open.
If I ⊆ R, we have an isomorphism L (I, X) ≡ X and if F : I → X
we will write Ḟ (t) instead of dF (t) if we regard dF (t) as an element of X.
Clearly this is consistent with the definition (7.1) from the previous section.
Example 7.1. Let X be a Hilbert space and consider F : X → R given by
F (x) := |x|2 . Then
F (x + u) = hx + u, x + ui = |x|2 + 2Rehx, ui + |u|2 = F (x) + 2Rehx, ui + o(u).
Hence if X is a real Hilbert space, then F is differentiable with dF (x)u =
2hx, ui. However, if X is a complex Hilbert space, then F is not differentiable.

The previous example emphasizes that for F : U ⊆ X → Y it makes a big
difference whether X is a real or a complex Banach space. In fact, in case of
a complex Banach space X, we obtain a version of complex differentiability
which of course is much stronger than real differentiability. Note that in this
respect it makes no difference whether Y is real or complex.
Example 7.2. Suppose f ∈ C 1 (R) with f (0) = 0. Let X := `pR (N), then
F : X → X, (xn )n∈N 7→ (f (xn ))n∈N
is differentiable for every x ∈ X with derivative given by the multiplication
operator
(dF (x)u)n = f 0 (xn )un .
First of all note that the mean value theorem implies |f (t)| ≤ MR |t| for
|t| ≤ R with MR := sup|t|≤R |f 0 (t)|. Hence, since kxk∞ ≤ kxkp , we have
kF (x)kp ≤ Mkxk∞ kxkp and F is well defined. This also shows that multipli-
cation by f 0 (xn ) is a bounded linear map. To establish differentiability we
use Z 1
0
f 0 (t + sτ ) − f 0 (t) dτ


f (t + s) − f (t) − f (t)s = s
0
and since f 0 is uniformly continuous on every compact interval, we can find
a δ > 0 for every given R > 0 and ε > 0 such that
|f 0 (t + s) − f 0 (t)| < ε if |s| < δ, |t| < R.
Now for x, u ∈ X with kxk∞ < R and kuk∞ < δ we have |f (xn + un ) −
f (xn ) − f 0 (xn )un | < ε|un | and hence
kF (x + u) − F (x) − dF (x)ukp < εkukp
198 7. Analysis in Banach spaces

which establishes differentiability. Moreover, using uniform continuity of

f on compact sets a similar argument shows that dF : X → L (X, X) is
continuous (observe that the operator norm of a multiplication operator by
a sequence is the sup norm of the sequence) and hence one writes F ∈
C 1 (X, X) as usual. 
Differentiability implies existence of directional derivatives
F (x + εu) − F (x)
δF (x, u) := lim , ε ∈ R \ {0}, (7.13)
ε→0 ε
which are also known as Gâteaux derivative or variational derivative.
Indeed, if F is differentiable at x, then (7.11) implies
δF (x, u) = dF (x)u. (7.14)
In particular, we call F Gâteaux differentiable at x ∈ U if the limit on the
right-hand side in (7.13) exists for all u ∈ X. However, note that Gâteaux
differentiability does not imply differentiability. In fact, the Gâteaux deriva-
tive might be unbounded or it might even fail to be linear in u. Some authors
require the Gâteaux derivative to be a bounded linear operator and in this
case we will write δF (x, u) = δF (x)u but even this additional requirement
does not imply differentiability in general. Note that in any case the Gâteaux
derivative is homogenous, that is, if δF (x, u) exists, then δF (x, λu) exists
for every λ ∈ R and
δF (x, λu) = λ δF (x, u), λ ∈ R. (7.15)
3
Example 7.3. The function F : R2 → R given by F (x, y) := x2x+y2 for
(x, y) 6= 0 and F (0, 0) = 0 is Gâteaux differentiable at 0 with Gâteaux
derivative
F (εu, εv)
δF (0, (u, v)) = lim = F (u, v),
ε→0 ε
which is clearly nonlinear.
The function F : R2 → R given by F (x, y) = x for y = x2 and F (x, y) :=
0 else is Gâteaux differentiable at 0 with Gâteaux derivative δF (0) = 0,
which is clearly linear. However, F is not differentiable.
If you take a linear function L : X → Y which is unbounded, then L
is everywhere Gâteaux differentiable with derivative equal to Lu, which is
linear but, by construction, not bounded. 
Example 7.4. Let X := L2R (0, 1) and consider
F : X → X, x 7→ sin(x).
First of all note that by | sin(t)| ≤ |t| our map is indeed from X to X and since
sine is Lipschitz continuous we get the same for F : kF (x)−F (y)k2 ≤ kx−yk2 .
7.2. Multivariable calculus in Banach spaces 199

Moreover, F is Gâteaux differentiable at x = 0 with derivative given by

δF (0) = I

but it is not differentiable at x = 0.

To see that the Gâteaux derivative is the identity note that
sin(εu(t))
lim = u(t)
ε→0 ε
pointwise and hence

sin(εu(.))
lim
− u(.)
=0
ε→0 ε 2

by dominated convergence since | sin(εu(t))

ε | ≤ |u(t)|.
To see that F is not differentiable let
π
un = πχ[0,1/n] , kun k2 = √
n
and observe that F (un ) = 0, implying that
kF (un ) − un k2
=1
kun k2
does not converge to 0. Note that this problem does not occur in X := C[0, 1]
(Problem 7.6). 
Example 7.5. Consider Lp (X, dµ), 1 ≤ p < ∞ and let G : C → R be (real)
differentiable with
q
|G(z)| ≤ C|z|p , |∂x G(z)|2 + |∂y G(z)|2 ≤ C|z|p−1 , z = x + iy,

or, if µ is finite,
q
p
|G(z)| ≤ C(1 + |z| ), |∂x G(z)|2 + |∂y G(z)|2 ≤ C(1 + |z|p−1 ).

Note that the first condition comes for free from the second in the finite case
and also in the general case if G(0) = 0. We only write down the estimates in
the first case and leave the easy adaptions for the second case as an exercise.
Then Z
N (f ) := G(f )dµ
X
is Gâteaux differentiable and we have
Z


δN (f )g = (∂x G)(f )Re(g) + (∂y G)(f )Im(g) dµ.
X
200 7. Analysis in Banach spaces

In fact, by the chain rule h(ε) := G(f + εg) is differentiable with h0 (0) =
(∂x G)(f )Re(g) + (∂y G)(f )Im(g). Moreover, by the mean value theorem
 
 h(ε) − h(0)  q 
  ≤ sup (∂x G)(f + τ g)2 + (∂y G)(f + τ g)2 |g|
 ε  0≤τ ≤ε

≤ C2p−1 (|f |p−1 + |g|p−1 )|g|

and hence we can invoke dominated convergence to interchange differentia-

tion and integration. Note that using Hölder’s inequality this last estimate
also shows local Lipschitz continuity on bounded domains:

|N (f ) − N (g)| ≤ C(kf kp + kgkp )p−1 kf − gkp .

In particular, for 1 < p < ∞ the norm

Z
N (f ) := |f |p dµ
X

is Gâteaux differentiable with

Z
δN (f )g = p |f |p−2 Re(f g ∗ )dµ. 
X

We will mainly consider Fréchet derivatives in the remainder of this chap-

ter as it will allow a theory quite close to the usual one for multivariable
functions. First of all we of course have linearity (which is easy to check):

Lemma 7.4. Suppose F, G : U → Y are differentiable at x ∈ U and α, β ∈

C. Then αF + βG is differentiable at x with d(αF + βG)(x) = αdF (x) +
βdG(x). Similarly, if the Gâteaux derivatives δF (x, u) and δG(x, u) exist,
then so does δ(F + G)(x, u) = δF (x, u) + δG(x, u).

Next, Fréchet differentiability implies continuity:

Lemma 7.5. Suppose F : U → Y is differentiable at x ∈ U . Then F is

continuous at x. Moreover, we can find constants M, δ > 0 such that

|F (x + u) − F (x)| ≤ M |u|, |u| ≤ δ. (7.16)

Proof. For every ε > 0 we can find a δ > 0 such that |F (x + u) − F (x) −
dF (x) u| ≤ ε|u| for |u| ≤ δ. Now choose M = kdF (x)k + ε. 

Example 7.6. Note that this lemma fails for the Gâteaux derivative as the
example of an unbounded linear function shows. In fact, it already fails in
R2 as the function F : R2 → R given by F (x, y) = 1 for y = x2 6= 0 and
F (x, y) = 0 else shows: It is Gâteaux differentiable at 0 with δF (0) = 0 but
it is not continuous since limε→0 F (ε, ε2 ) = 1 6= 0 = F (0, 0). 
7.2. Multivariable calculus in Banach spaces 201

If F is differentiable for all x ∈ U we call F differentiable. In this case

we get a map
dF : U → L (X, Y )
. (7.17)
x 7→ dF (x)

If dF : U → L (X, Y ) is continuous, we call F continuously differentiable

and write F ∈ C 1 (U, Y ).
If X or Y has a (finite) product structure, then the computation of the
derivatives can be reduced as usual. The following facts are simple and can
be shown as in the case of X = Rn and Y = Rm .

Let Y := m j=1 Yj and let F : X → Y be given by F = (F1 , . . . , Fm ) with
Fj : X → Yj . Then F ∈ C 1 (X, Y ) if and only if Fj ∈ C 1 (X, Yj ), 1 ≤ j ≤ m,

and in this case dF = (dF1 , . . . , dFm ). Similarly, if X = ni=1 Xi , then one
can define the partial derivative ∂i F ∈ L (Xi , Y ), which is the derivative
of F considered as a function of the Pni-th variable alone (the other variables
being fixed). We have dF u = i=1 ∂i F ui , u = (u1 , . . . , un ) ∈ X, and
F ∈ C 1 (X, Y ) if and only if all partial derivatives exist and are continuous.
Example 7.7. In the case of X = Rn and Y = Rm , the matrix representa-
tion of dF with respect to the canonical basis in Rn and Rm is given by the
partial derivatives ∂i Fj (x) and is called Jacobi matrix of F at x. 
Given F ∈ C 1 (U, Y ) we have dF ∈ C(U, L (X, Y )) and we can define
the second derivative (provided it exists) via

dF (x + v) = dF (x) + d2 F (x)v + o(v). (7.18)

In this case d2 F : U → L (X, L (X, Y )) which maps x to the linear map

v 7→ d2 F (x)v which for fixed v is a linear map u 7→ (d2 F (x)v)u. Equivalently,
we could regard d2 F (x) as a map d2 F (x) : X 2 → Y , (u, v) 7→ (d2 F (x)v)u
which is linear in both arguments. That is, d2 F (x) is a bilinear map X 2 → Y .
The corresponding norm on L (X, L (X, Y )) explicitly spelled out reads

kd2 F (x)k = sup kd2 F (x)vk = sup k(d2 F (x)v)uk. (7.19)

|v|=1 |u|=|v|=1

Example 7.8. Note that if F ∈ L (X, Y ), then dF (x) = F (independent of

x) and d2 F (x) = 0. 
Example 7.9. Let X be a real Hilbert space and F (x) = |x|2 . Then we
have already seen dF (x)u = 2hx, ui and hence

dF (x + v)u = 2hx + v, ui = 2hx, ui + 2hv, ui = dF (x)u + 2hv, ui

which shows (d2 F (x)v)u = 2hv, ui. 

202 7. Analysis in Banach spaces

Example 7.10. Suppose f ∈ C 2 (R) with f (0) = 0 and continue Exam-

ple 7.2. Then we have F ∈ C 2 (X, X) with d2 F (x)v the multiplication oper-
ator by the sequence f 00 (xn )vn , that is,
((d2 F (x)v)u)n = f 00 (xn )vn un .
Indeed, arguing in a similar fashion we can find a δ1 such that |f 0 (xn + vn ) −
f 0 (xn ) − f 00 (xn )vn | ≤ ε|vn | whenever kxk∞ < R and kvk∞ < δ1 . Hence
kdF (x + v) − dF (x) − d2 F (x)vk < εkvkp
which shows differentiability. Moreover, since kd2 F (x)k = kf 00 (x)k∞ one
also easily verifies that F ∈ C 2 (X, X) using uniform continuity of f 00 on
compact sets. 
We can iterate the procedure of differentiation and write F ∈ C r (U, Y ),
r ≥ 1, if the r-th derivative of F , dr F (i.e., the derivative of the (r − 1)-
th derivative of F ), exists and is continuous. Note that dr F (x) will be
a multilinear T
map in r arguments as we will show below. Finally, we set
C ∞ (U, Y ) = r∈N C r (U, Y ) and, for notational convenience, C 0 (U, Y ) =
C(U, Y ) and d0 F = F .
Example 7.11. Let X be a Banach algebra. Consider the multiplication
M : X × X → X. Then
∂1 M (x, y)u = uy, ∂2 M (x, y)u = xu
and hence
dM (x, y)(u1 , u2 ) = u1 y + xu2 .
Consequently dM is linear in (x, y) and hence
(d2 M (x, y)(v1 , v2 ))(u1 , u2 ) = u1 v2 + v1 u2
Consequently all differentials of order higher than two will vanish and in
particular M ∈ C ∞ (X × X, X). 
If F is bijective and F , F −1 are both of class C r , r ≥ 1, then F is called
a diffeomorphism of class C r .
For the composition of mappings we have the usual chain rule.
Lemma 7.6 (Chain rule). Let U ⊆ X, V ⊆ Y and F ∈ C r (U, V ) and
G ∈ C r (V, Z), r ≥ 1. Then G ◦ F ∈ C r (U, Z) and
d(G ◦ F )(x) = dG(F (x)) ◦ dF (x), x ∈ X. (7.20)

Proof. Fix x ∈ U , y = F (x) ∈ V and let u ∈ X such that v = dF (x)u with

x+u ∈ U and y+v ∈ V for |u| sufficiently small. Then F (x+u) = y+v+o(u)
and, with ṽ = v + o(u),
G(F (x + u)) = G(y + ṽ) = G(y) + dG(y)ṽ + o(ṽ).
7.2. Multivariable calculus in Banach spaces 203

Using |ṽ| ≤ kdF (x)k|u| + |o(u)| we see that o(ṽ) = o(u) and hence
G(F (x + u)) = G(y) + dG(y)v + o(u) = G(F (x)) + dG(F (x)) ◦ dF (x)u + o(u)
as required. This establishes the case r = 1. The general case follows from
induction. 

In particular, if ` ∈ Y ∗ is a bounded linear functional, then d(` ◦ F ) =

d` ◦ dF = ` ◦ dF . As an application of this result we obtain
Theorem 7.7 (Schwarz). Suppose F ∈ C 2 (Rn , Y ). Then
∂i ∂j F = ∂j ∂i F
for any 1 ≤ i, j ≤ n.

Proof. First of all note that ∂j F (x) ∈ L (R, Y ) and thus it can be regarded
as an element of Y . Clearly the same applies to ∂i ∂j F (x). Let ` ∈ Y ∗
be a bounded linear functional, then ` ◦ F ∈ C 2 (R2 , R) and hence ∂i ∂j (` ◦
F ) = ∂j ∂i (` ◦ F ) by the classical theorem of Schwarz. Moreover, by our
remark preceding this lemma ∂i ∂j (` ◦ F ) = ∂i `(∂j F ) = `(∂i ∂j F ) and hence
`(∂i ∂j F ) = `(∂j ∂i F ) for every ` ∈ Y ∗ implying the claim. 

Now we let F ∈ C 2 (X, Y ) and look at the function G : R2 → Y , (t, s) 7→

G(t, s) = F (x + tu + sv). Then one computes

∂t G(t, s) = dF (x + sv)u

t=0
and hence
 
∂s ∂t G(t, s) = ∂s dF (x + sv)u = (d2 F (x)u)v.
 
(s,t)=0 s=0

Since by the previous theorem the oder of the derivatives is irrelevant, we

obtain
d2 F (u, v) = d2 F (v, u), (7.21)
that is, d2 F is a symmetric bilinear form. This result easily generalizes to
higher derivatives. To this end we introduce some notation first.

A function L : nj=1 Xj → Y is called multilinear if it is linear with
respect to each argument. It is not hard to see that L is continuous if and
only if
kLk = sup |L(x1 , . . . , xn )| < ∞. (7.22)
x:|x1 |=···=|xn |=1
If we take n copies of the same space, the set of multilinear functions
L : X n → Y will be denoted by L n (X, Y ). A multilinear function is
called symmetric provided its value remains unchanged if any two ar-
guments are switched. With the norm from above it is a Banach space
and in fact there is a canonical isometric isomorphism between L n (X, Y )
204 7. Analysis in Banach spaces

and L (X, L n−1 (X, Y )) given by L : (x1 , . . . , xn ) 7→ L(x1 , . . . , xn ) maps to

x1 7→ L(x1 , .).
Lemma 7.8. Suppose F ∈ C r (X, Y ). Then for every x ∈ X we have that
r
X
dr F (x)(u1 , . . . , ur ) = ∂t1 · · · ∂tr F (x + ti ui )|t1 =···=tr =0 . (7.23)
i=1

Moreover, dr F (x) ∈ L r (X, Y ) is a bounded symmetric multilinear form.

Proof. The representation (7.23) follows using induction as before. Sym-

metry follows since the order of the partial derivatives can be interchanged
by Lemma 7.7. 

Finally, note that to each L ∈ L n (X, Y ) we can assign its polar form
L ∈ C(X, Y ) using L(x) = L(x, . . . , x), x ∈ X. If L is symmetric it can be
reconstructed using polarization (Problem 7.8):
n
1 X
L(u1 , . . . , un ) = ∂t1 · · · ∂tn L( ti ui ). (7.24)
n!
i=1

We also have the following version of the product rule: Suppose L ∈

L 2 (X, Y ), then L ∈ C 1 (X 2 , Y ) with
dL(x)u = L(u1 , x2 ) + L(x1 , u2 ) (7.25)
since
L(x1 + u1 , x2 + u2 ) − L(x1 , x2 ) = L(u1 , x2 ) + L(x1 , u2 ) + L(u1 , u2 )
= L(u1 , x2 ) + L(x1 , u2 ) + O(|u|2 ) (7.26)

as |L(u1 , u2 )| ≤ kLk|u1 ||u2 | = O(|u|2 ). If X is a Banach algebra and

L(x1 , x2 ) = x1 x2 we obtain the usual form of the product rule.
Next we have the following mean value theorem.
Theorem 7.9 (Mean value). Suppose U ⊆ X and F : U → Y is Gâteaux
differentiable at every x ∈ U . If U is convex, then
x−y
|F (x)−F (y)| ≤ M |x−y|, M := sup |δF ((1−t)x+ty, )|. (7.27)
0≤t≤1 |x − y|
Conversely, (for any open U ) if
|F (x) − F (y)| ≤ M |x − y|, x, y ∈ U, (7.28)
then
sup |δF (x, e)| ≤ M. (7.29)
x∈U,|e|=1
7.2. Multivariable calculus in Banach spaces 205

Proof. Abbreviate f (t) = F ((1 − t)x + ty), 0 ≤ t ≤ 1, and hence df (t) =

δF ((1 − t)x + ty, y − x) implying |df (t)| ≤ M̃ := M |x − y| by (7.15). For the
first part it suffices to show
φ(t) = |f (t) − f (0)| − (M̃ + δ)t ≤ 0
for any δ > 0. Let t0 = max{t ∈ [0, 1]|φ(t) ≤ 0}. If t0 < 1 then
φ(t0 + ε) = |f (t0 + ε) − f (t0 ) + f (t0 ) − f (0)| − (M̃ + δ)(t0 + ε)
≤ |f (t0 + ε) − f (t0 )| − (M̃ + δ)ε + φ(t0 )
≤ |df (t0 )ε + o(ε)| − (M̃ + δ)ε
≤ (M̃ + o(1) − M̃ − δ)ε = (−δ + o(1))ε ≤ 0,
for ε ≥ 0, small enough. Thus t0 = 1.
To prove the second claim suppose we can find an e ∈ X, |e| = 1 such
that |δF (x0 , e)| = M + δ for some δ > 0 and hence
M ε ≥ |F (x0 + εe) − F (x0 )| = |δF (x0 , e)ε + o(ε)|
≥ (M + δ)ε − |o(ε)| > M ε
since we can assume |o(ε)| < εδ for ε > 0 small enough, a contradiction. 
Corollary 7.10. Suppose U ⊆ X and F ∈ C 1 (U, Y ). Then F is locally
Lipschitz continuous.

Proof. Fix x0 ∈ U and note that by continuity there is a neighborhood U0

of x0 such that kdF (x)k ≤ M for x ∈ U0 . Hence the claim follows from the
mean value theorem. 

Note, however, that a C 1 function is in general not Lipschitz on arbitrary

bounded sets since in the infinite dimensional case continuity of dF does not
suffice to conclude boundedness on bounded closed sets.
Example 7.12. Let X be an infinite Hilbert space and {un }n∈N some or-
thonormal set. Then the functions Fn (x) := max(0, 1−2kx−un k) are contin-
uous with disjoint supports. Hence F (x) := n∈N nFn (x) is also continuous
P
(show this). But F is not bounded on the unit ball since F (un ) = n. 
As an immediate consequence we obtain
Corollary 7.11. Suppose U is a connected subset of a Banach space X. A
Gâtaux differentiable mapping F : U → Y is constant if and only if δF = 0.
In addition, if F1,2 : U → Y and δF1 = δF2 , then F1 and F2 differ only by a
constant.

As an application of the fundamental theorem of calculus (Theorem 7.3)

we obtain a generalization of the well-known fact that continuity of the
directional derivatives implies continuous differentiability.
206 7. Analysis in Banach spaces

Lemma 7.12. Suppose F : U ⊆ X → Y is Gâteaux differentiable such that

the Gâteaux derivative is linear and continuous, δF ∈ C(U, L (X, Y )). Then
F ∈ C 1 (U, Y ) and dF = δF .

Proof. By assumption f (t) := F (x + tu) is in C 1 ([0, 1], Y ) for u with suf-

ficiently small norm. Moreover, by definition we have f˙ = δF (x + tu)u and
using the fundamental theorem of calculus we obtain
Z 1 Z 1
F (x + u) − F (x) = f (1) − f (0) = ˙
f (t)dt = δF (x + tu)u dt
0 0
Z 1 
= δF (x + tu)dt u,
0
where the last equality follows from continuity of the integral since it clearly
holds for simple functions. Consequently
Z 1  


|F (x + u) − F (x) − δF (x)u| = 
 δF (x + tu) − δF (x) dt u
0
Z 1 
≤ kδF (x + tu) − δF (x)kdt |u|
0
≤ max kδF (x + tu) − δF (x)k |u|.
t∈[0,1]

By the continuity assumption on δF , the right-hand side is o(u) as required.



As another consequence we obtain Taylor’s theorem.

Theorem 7.13 (Taylor). Suppose U ⊆ X and F ∈ C r+1 (U, Y ). Then
1 1
F (x + u) =F (x) + dF (x)u + d2 F (x)u2 + · · · + dr F (x)ur
2 r!
 Z 1 
1
+ (1 − t)r dr+1 F (x + tu)dt ur+1 , (7.30)
r! 0
where uk := (u, . . . , u) ∈ X k .

Proof. As in the proof of the previous lemma, the case r = 0 is just the
fundamental theorem of calculus applied to f (t) := F (x + tu). For the in-
duction step we use integration by parts. To this end let fj ∈ C 1 ([0, 1], Xj ),
L ∈ L 2 (X1 × X2 , Y ) bilinear. Then the product rule (7.25) and the funda-
mental theorem of calculus imply
Z 1 Z 1
˙
L(f1 (t), f2 (t))dt = L(f1 (1), f2 (1))−L(f1 (0), f2 (0))− L(f1 (t), f˙2 (t))dt.
0 0
Hence applying integration by parts with L(y, t) = ty, f1 (t) = dr+1 F (x+ut),
r+1
and f2 (t) = (1−t)
(r+1)! establishes the induction step. 
7.2. Multivariable calculus in Banach spaces 207

Of course this also gives the Peano form for the remainder:
Corollary 7.14. Suppose U ⊆ X and F ∈ C r (U, Y ). Then
1 1
F (x+u) = F (x)+dF (x)u+ d2 F (x)u2 +· · ·+ dr F (x)ur +o(|u|r ). (7.31)
2 r!
Proof. Just estimate
 Z 1  
1 1
r−1 r
(1 − t) d F (x + tu)dt − d F (x) ur 
r
 

 (r − 1)! r!
0
r Z 1
|u|
≤ (1 − t)r−1 kdr F (x + tu) − dr F (x)kdt
(r − 1)! 0
|u|r
≤ sup kdr F (x + tu) − dr F (x)k. 
r! 0≤t≤1

Finally we remark that it is often necessary to equip C r (U, Y ) with a

norm. A suitable choice is
X
kF k = sup kdj F (x)k. (7.32)
0≤j≤r x∈U

The set of all r times continuously differentiable functions for which this
norm is finite forms a Banach space which is denoted by Cbr (U, Y ).
In the definition of differentiability we have required U to be open. Of
course there is no stringent reason for this and (7.12) could simply be required
for all sequences from U \ {x} converging to x. However, note that the
derivative might not be unique in case you miss some directions (the ultimate
problem occurring at an isolated point). Our requirement avoids all these
issues. Moreover, there is usually another way of defining differentiability
at a boundary point: By C r (U , Y ) we denote the set of all functions in
C r (U, Y ) all whose derivatives of order up to r have a continuous extension
to U . Note that if you can approach a boundary point along a half-line then
the fundamental theorem of calculus shows that the extension coincides with
the Gâteaux derivative.
Problem 7.5. Let X be a real Hilbert space, A ∈ L (X) and F (x) :=
hx, Axi. Compute dn F .
Problem* 7.6. Let X := C([0, 1], R) and suppose f ∈ C 1 (R). Show that
F : X → X, x 7→ f ◦ x
is differentiable for every x ∈ X with derivative given by
(dF (x)y)(t) = f 0 (x(t))y(t).
Problem 7.7. Let X := `2 (N), Y := `1 (N) and F : X → Y given by
F (x)j := x2j . Show F ∈ C ∞ (X, Y ) and compute all derivatives.
208 7. Analysis in Banach spaces

Problem* 7.8. Show (7.24).

Problem 7.9. Let X be a Banach algebra, I ⊆ R an open interval, and
x, y ∈ C 1 (I, X). Show that xy ∈ C 1 (I, X) with (xy)˙ = ẋy + xẏ.
Problem 7.10. Let X be a Banach algebra and G(X) the group of invertible
elements. Show that I : G(X) → G(X), x 7→ x−1 is differentiable with
dI(x)y = −x−1 yx.
(Hint: (6.9))

7.3. Minimizing nonlinear functionals via calculus

Many problems in applications lead to finding the minimum (or maximum)
of a given (nonlinear) functional F : X → R. For example, many physical
problems can be described by an energy functional and one seeks a solution
which minimizes this energy. Since the minima of −F are the maxima of F
and vice versa, we will restrict our attention to minima only. Of course if
X = R (or Rn ) we can find the local extrema by searching for the zeros of
the derivative and then checking the second derivative to determine if it is a
minimum or maximum. In fact, by virtue of our version of Taylor’s theorem
(7.31) we see that F will take values above and below F (x) in a vicinity of
x if we can find some u such that dF (x)u 6= 0. Hence dF (x) = 0 is clearly a
necessary condition for a local extremum. Moreover, if dF (x) = 0 we can go
one step further and conclude that all values in a vicinity of x will lie above
F (x) provided the second derivative d2 F (x) is positive in the sense that there
is some c > 0 such that d2 F (x)u2 > c for all directions u ∈ ∂B1 (0). While
this gives a viable solution to the problem of finding local extrema, we can
easily do a bit better. To this end we look at the variations of f along lines
through x, that is, we look at the behavior of the function
f (t) := F (x + tu) (7.33)
for a fixed direction u ∈ B1 (0). Hence this approach is also known as cal-
culus of variations. Then, if F has a local extremum at x the same will
be true for f and hence a necessary condition for an extremum is that the
Gâteaux derivative vanishes in every direction: δF (x, u) = 0 for all unit vec-
tors u. Similarly, a necessary condition for a local minimum at x is that f
has a local minimum at 0 for all unit vectors u. For example δ 2 F (x, u) > 0
for all unit vectors u. Here the higher order Gâteaux derivatives are defined
as  n
d 
δ n F (x, u) := F (x + tu) (7.34)

dt t=0
with the derivative defined as a limit as in (7.13). That is we have the
recursive definition δ n F (x, u) = limε→0 ε−1 δ n−1 F (x+εu, u)−δ n−1 F (x, u) .


7.3. Minimizing nonlinear functionals via calculus 209

Note that if δ n F (x, u) exists, then δ n F (x, λu) exists for every λ ∈ R and
δ n F (x, λu) = λn δ n F (x, u), λ ∈ R. (7.35)
However, the condition > 0 for all unit vectors u is not sufficient as
δ 2 F (x, u)
there are certain features you might miss when you only look at the function
along rays through a fixed point. This is demonstrated by the following
example:
Example 7.13. Let X = R2 and consider the points (xn , yn ) := ( n1 , n12 ).
For each point choose a radius rn such that the balls Bn := Brn (xn , yn )
2
are disjoint and lie between two parabolas Bn ⊂ {(x, y)|x ≥ 0, x2 ≤ y ≤
2x2 }. Moreover, choose a smooth nonnegative bump function φ(r2 ) with
support in [−1, 1] and maximum 1 at 0. Now consider F (x, y) = x2 + y 2 −
2 2
2 n∈N ρn φ( (x−xn ) r+(y−y n)
), where ρn = x2n + yn2 . By construction F is
P
2
n
smooth away from zero. Moreover, at zero F is continuous and Gâteaux
differentiable of arbitrary order with F (0, 0) = 0, δF ((0, 0), (u, v)) = 0,
δ 2 F ((0, 0), (u, v)) = 2(u2 + v 2 ), and δ k F ((0, 0), (u, v)) = 0 for k ≥ 3.
In particular, F (ut, vt) has a strict local minimum at t = 0 for every
(u, v) ∈ R2 \{0}, but F has no local minimum at (0, 0) since F (xn , yn ) = −ρn .
Cleary F is not differentiable at 0. In fact, note that the Gâteaux derivatives
are not continuous at 0 (the derivatives in Bn grow like rn−2 ). 
Lemma 7.15. Suppose F : U → R has Gâteaux derivatives up to the order
of two. A necessary condition for x ∈ U to be a local minimum is that
δF (x, u) = 0 and δ 2 F (x, u) ≥ 0 for all u ∈ X. A sufficient condition for a
strict local minimum is if in addition δ 2 F (x, u) ≥ c > 0 for all u ∈ ∂B1 (0)
and δ 2 F is continuous at x uniformly with respect to u ∈ ∂B1 (0).

Proof. The necessary conditions have already been established. To see the
sufficient conditions note that the assumptions on δ 2 F imply that there is
some ε > 0 such that δ 2 F (y, u) ≥ 2c for all y ∈ Bε (x) and all u ∈ ∂B1 (0).
Equivalently, δ 2 F (y, u) ≥ 2c |u|2 for all y ∈ Bε (x) and all u ∈ X. Hence
applying Taylor’s theorem to f (t) using f¨(t) = δ 2 F (x + tu, u) gives
Z 1
c
F (x + u) = f (1) = f (0) + (1 − s)f¨(s)ds ≥ F (x) + |u|2
0 4
for u ∈ Bε (0). 

Note that if F ∈ C 2 (U, R) then δ 2 F (x, u) = d2 F (x)u2 and we obtain

Corollary 7.16. Suppose F ∈ C 2 (U, R). A sufficient condition for x ∈ U to
be a strict local minimum is dF (x) = 0 and d2 F (x)u2 ≥ c|u|2 for all u ∈ X.

Proof. Observe that by |δ 2 F (x, u) − δ 2 F (y, u)| ≤ kd2 F (x) − d2 F (y)k|u|2

the continuity requirement from the previous lemma is satisfied. 
210 7. Analysis in Banach spaces

Example 7.14. If X is a real Hilbert space, then the symmetric bilinear

form d2 F has a corresponding self-adjoint operator A ∈ L (X) such that
d2 F (u, v) = hu, Avi and the condition d2 F (x)u2 ≥ c|u|2 is equivalent to
the spectral condition σ(A) ⊂ [c, ∞). In the finite dimensional case A is of
course the Jacobi matrix and the spectral conditions says that all eigenvalues
must be positive. 
Example 7.15. Let X := `2 (N) and consider

X  x2 
n 4
F (x) := − xn .
2n2
n∈N

Then F ∈ C 2 (X, R) with dF (x)u = n∈N ( xnn2 − 4x3n )un and d2 F (x)(u, v) =
P

( 1 − 12x2n )vn un . In particular, F (0) = 0, dF (0) = 0 and d2 F (0)u2 =

P
Pn∈N−2n22
n n un > 0 for u 6= 0. However, F (δ /m) < 0 shows that 0 is no
m

local minimum. So the condition d2 F (x)u2 > 0 is not sufficient in infinite

dimensions. It is however, sufficient in finite dimensions since compactness
of the unit ball leads to the stronger condition d2 F (x, u) ≥ c > 0 for all
u ∈ ∂B1 (0). 
Example 7.16. Consider a classical particle whose location at time t is given
by q(t). Then the least action principle states that, if the particle moves
from q(a) to q(b), the path of the particle will make the action functional

Z b
S(q) := L(t, q(t), q̇(t))dt
a

stationary, that is

δS(q) = 0.

Here L : R × Rn × Rn → R is the Lagrangian of the system. The name

suggests that the action should attain a minimum, but this is not always the
case and hence it is also referred to as stationary action principle.
More precisely, let L ∈ C 2 (R2n+1 , R) and in order to incorporate the
requirement that the initial and end points are fixed, we take X = {x ∈
C 2 ([a, b], Rn )|x(a) = x(b) = 0} and consider

t−a
q(t) := q(a) + (q(b) − q(a)) + x(t), x ∈ X.
b−a
7.3. Minimizing nonlinear functionals via calculus 211

Hence we want to compute the Gâteaux derivative of F (x) := S(q), where

x and q are related as above with q(a), q(b) fixed. Then
Z b
d 
δF (x, u) =  L(s, q(s) + t u(s), q̇(s) + t u̇(s))ds
dt t=0 a
Z b


= Lq (s, q(s), q̇(s))u(s) + Lq̇ (s, q(s), q̇(s))u̇(s) ds
a
Z b
d

= Lq (s, q(s), q̇(s))u(s) − Lq̇ (s, q(s), q̇(s)) u(s)ds,
a ds
where we have used integration by parts (including the boundary conditions)
to obtain the last equality. Here Lq , Lq̇ are the gradients with respect to q,
q̇, respectively, and products are understood as scalar products in Rn .
If we want this to vanish for all u ∈ X we obtain the corresponding
Euler–Lagrange equation
d
Lq̇ (s, q(s), q̇(s)) = Lq (s, q(s), q̇(s)).
ds
For example, for a classical particle of mass m > 0 moving in a conservative
force field described by a potential V ∈ C 1 (Rn , R) the Lagrangian is given
by the difference between kinetic and potential energy
m
L(t, q, q̇) := q̇ 2 − V (q)
2
and the Euler–Lagrange equations read
mq̈ = −Vq (q),
which are just Newton’s equations of motion. 
Finally we note that the situation simplifies a lot when F is convex. Our
first observation is that a local minimum is automatically a global one.
Lemma 7.17. Suppose C ⊆ X is convex and F : C → R is convex. Every
local minimum is a global minimum. Moreover, if F is strictly convex then
the minimum is unique.

Proof. Suppose x is a local minimum and F (y) < F (x). Then F (λy + (1 −
λ)x) ≤ λF (y) + (1 − λ)F (x) < F (x) for λ ∈ (0, 1) contradicts the fact that x
is a local minimum. If x, y are two global minima, then F (λy + (1 − λ)x) <
F (y) = F (x) yielding a contradiction unless x = y. 

Moreover, to find the global minimum it suffices to find a point where

the Gâteaux derivative vanishes.
Lemma 7.18. Suppose C ⊆ X is convex and F : C → R is convex. If the
Gâteaux derivative exists at an interior point x ∈ C and satisfies δF (x, u) =
0 for all u ∈ X, then x is a global minimum.
212 7. Analysis in Banach spaces

Proof. By assumption f (t) := F (x + tu) is a convex function defined on an

interval containing 0 with f 0 (0) = 0. If y is another point we can choose u =
y − x and Lemma 3.2 from [47] (iii) implies F (y) = f (1) ≥ f (0) = F (x). 

As in the one-dimensional case, convexity can be read off from the second
derivative.

Lemma 7.19. Suppose C ⊆ X is open and convex and F : C → R

has Gâteaux derivatives up to order two. Then F is convex if and only
if δ 2 F (x, u) ≥ 0 for all x ∈ C and u ∈ X. Moreover, F is strictly convex if
δ 2 F (x, u) > 0 for all x ∈ C and u ∈ X \ {0}.

Proof. We consider f (t) := F (x + tu) as before such that f 0 (t) = δF (x +

tu, u), f 00 (t) = δ 2 F (x + tu, u). Moreover, note that f is (strictly) convex for
all x ∈ C and u ∈ X \ {0} if and only if F is (strictly) convex. Indeed, if F
is (strictly) convex so is f as is easy to check. To see the converse note
F (λy + (1 − λ)x) = f (λ) ≤ λf (1) − (1 − λ)f (0) = λF (y) − (1 − λ)F (x)
with strict inequality if f is strictly convex. The rest follows from Problem 3.6
from [47]. 

There is also a version using only first derivatives plus the concept of a
monotone operator. A map F : U ⊆ X → X ∗ is monotone if
(F (x) − F (y))(x − y) ≥ 0, x, y ∈ U.
It is called strictly monotone if we have strict inequality for x 6= y. Mono-
tone operators will be the topic of Chapter 12.

Lemma 7.20. Suppose C ⊆ X is open and convex and F : C → R has

Gâteaux derivatives δF (x) ∈ X ∗ for every x ∈ C. Then F is (strictly)
convex if and only if δF is (strictly) monotone.

Proof. Note that by assumption δF : C → X ∗ and the claim follows as in

the previous lemma from Problem 3.6 from [47] since f 0 (t) = δF (x + tu)u
which shows that δF is (strictly) monotone if and only if f 0 is (strictly)
increasing. 

Example 7.17. The length of a curve q : [a, b] → Rn is given by

Z b
|q 0 (s)|ds.
a

Of course we know that the shortest curve between two given points q0 and q1
is a straight line. Notwithstanding that this is evident, defining the length as
7.3. Minimizing nonlinear functionals via calculus 213

the total variation, let us show this by seeking the minimum of the following
functional
Z b
t−a
F (x) := |q 0 (s)|ds, q(t) = x(t) + q0 + (q1 − q0 )
a b−a
for x ∈ X := {x ∈ C 1 ([a, b], Rn )|x(a) = x(b) = 0}. Unfortunately our inte-
grand will not be differentiable unless |q 0 | ≥ c. However, since the absolute
value is convex, so is F and it will suffice to search for a local minimum
within the convex open set C := {x ∈ X||x0 | < |q2(b−a)1 −q0 |
}. We compute
Z b 0
q (s)u0 (s)
δF (x, u) = ds
a |q 0 (s)|
which shows (Lemma 3.24 from [47]) that q 0 /|q 0 | must be constant. Hence
the local minimum in C is indeed a straight line and this must also be a
global minimum in X. However, since the length of a curve is independent
of its parametrization, this minimum is not unique! 
Example 7.18. Let us try to find a curve y(x) from y(0) = 0 to y(x1 ) = y1
which minimizes
Z x1 r
1 + y 0 (x)2
F (y) := dx.
0 x
√
Note that since the function t 7→ 1 + t2 is convex, we obtain that F is
convex. Hence it suffices to find a zero of
y 0 (x)u0 (x)
Z x1
δF (y, u) = p dx,
0 x(1 + y 0 (x)2 )
y0
which shows (Lemma 3.24 from [47]) that √ = C −1/2 is constant or
x(1+y 02 )
equivalently r
0 x
y (x) =
C −x
and hence r 
x p
y(x) = C arctan − x(C − x).
C −x
The constant C has to be chosen such that y(x1 ) matches the given value
y1 . Note that C 7→ y(x1 ) decreases from πx2 1 to 0 and hence there will be a
unique C > x1 for 0 < y1 < πx2 1 . 
Problem 7.11. Consider the least action principle for a classical one-dimensional
particle. Show that
Z b
2
m u̇(s)2 − V 00 (q(s))u(s)2 ds.


δ F (x, u) =
a
Moreover, show that we have indeed a minimum if V 00 ≤ 0.
214 7. Analysis in Banach spaces

7.4. Minimizing nonlinear functionals via compactness

Another approach for minimizing a nonlinear functional F : M ⊆ X → R is
based on compactness. If M is compact and F is continuous, then we can
proceed as in the finite-dimensional case to show that there is a minimizer:
Start with a sequence xn such that F (xn ) → inf M F . By compactness we
can assume that xn → x0 after passing to a subsequence and by continuity
F (xn ) → F (x0 ) = inf M F . Now in the infinite dimensional case we will
use weak convergence to get compactness and hence we will also need weak
(sequential) continuity of F . However, since there are more weakly than
strongly convergent subsequences, weak (sequential) continuity is in fact a
stronger property than just continuity!
Example 7.19. By Lemma 4.28 (ii) the norm is weakly sequentially lower
semicontinuous but it is in general not weakly sequentially continuous as any
infinite orthonormal set in a Hilbert space converges weakly to 0. However,
note that this problem does not occur for linear maps. This is an immediate
consequence of the very definition of weak convergence (Problem 4.37). 
Hence weak continuity might be too much to hope for in concrete appli-
cations. In this respect note that, for our argument to work lower semicon-
tinuity (cf. Problem B.18) will already be sufficient:
Theorem 7.21 (Variational principle). Let X be a reflexive Banach space
and let F : M ⊆ X → (−∞, ∞]. Suppose M is nonempty, weakly se-
quentially closed and that either F is weakly coercive, that is F (x) → ∞
whenever kxk → ∞, or that M is bounded. Then, if F is weakly sequentially
lower semicontinuous, there exists some x0 ∈ M with F (x0 ) = inf M F .
If F is Gâteaux differentiable, then
δF (x0 ) = 0.

Proof. Without loss of generality we can assume F (x) < ∞ for some x ∈ M .
As above we start with a sequence xn ∈ M such that F (xn ) → inf M F < ∞.
If M = X then the fact that F is coercive implies that xn is bounded.
Otherwise, it is bounded since we assumed M to be bounded. Hence by
Theorem 4.31 we can pass to a subsequence such that xn * x0 with x0 ∈
M since M is assumed sequentially closed. Now since F is weakly se-
quentially lower semicontinuous we finally get inf M F = limn→∞ F (xn ) =
lim inf n→∞ F (xn ) ≥ F (x0 ). 

Of course in a metric space the definition of closedness in terms of se-

quences agrees with the corresponding topological definition. In the present
situation sequentially weakly closed implies (sequentially) closed and the
converse holds at least for convex sets.
7.4. Minimizing nonlinear functionals via compactness 215

Lemma 7.22. Suppose M ⊆ X is convex. Then M is closed if and only if

it is sequentially weakly closed.

Proof. Suppose M is closed and let x be in the weak sequential closure of M ,

that is, there is a sequence xn * x. If x 6∈ M , then by Corollary 5.4 we can
find a linear functional ` which separates {x} and M . But this contradicts
`(x) = d < c ≤ `(xn ) → `(x). 

Similarly, the same is true with lower semicontinuity. In fact, a slightly

weaker assumption suffices. Let X be a vector space and M ⊆ X a convex
subset. A function F : M → R is called quasiconvex if
F (λx + (1 − λ)y) ≤ max{F (x), F (y)}, λ ∈ (0, 1), x, y ∈ M. (7.36)
It is called strictly quasiconvex if the inequality is strict for x 6= y. By
λF (x) + (1 − λ)F (y) ≤ max{F (x), F (y)} every (strictly) convex function
is (strictly) quasiconvex. The converse is not true as the following example
shows.
Example 7.20. Every (strictly) monotone function on R is (strictly) qua-
siconvex. Moreover, the same is true for symmetric functions
p which are
(strictly) monotone on [0, ∞). Hence the function F (x) = |x| is strictly
quasiconvex. But it is clearly not convex on M = R. 
Note that we can extend a (quasi-)convex function F : M → R to all of
X by setting F (x) = ∞ for x ∈ X \ M and the resulting function will still
be (quasi-)convex and will have the same infimum.
Now we are ready for the next
Lemma 7.23. Suppose M ⊆ X is a closed convex set and suppose F : M →
R is quasiconvex. Then F is weakly sequentially lower semicontinuous if and
only if it is (sequentially) lower semicontinuous.

Proof. Suppose F is lower semicontinuous. If it were not weakly sequen-

tially lower semicontinuous we could find a sequence xn * x0 with F (xn ) →
a < F (x0 ). But then xn ∈ F −1 ((−∞, a]) for n sufficiently large implying
x0 ∈ F −1 ((−∞, a]) as this set is convex (Problem 7.14) and closed (Prob-
lem B.18). But this gives the contradiction a < F (x0 ) ≤ a. 
Example 7.21. Let U ⊆ Rn and K : U ×C → [0, ∞). Suppose u 7→ K(x, u)
is quasi-convex and continuous for fixed x ∈ U . Then
Z
F (u) := K(x, u(x))dn x
U
is weakly sequentially lower semicontinuous on Lp (U ) for 1 ≤ p ≤ ∞. Since
F is obviously quasiconvex, it suffices to show lower semicontinuity. As-
sume the contrary, then we can find some u ∈ Lp and a sequence un → u
216 7. Analysis in Banach spaces

such that F (u) > lim inf F (un ). After passing to a subsequence we can
assume that un (x) → u(x) a.e. and hence K(x, un (x)) → K(x, u(x)) a.e.
Finally applying Fatou’s lemma (Theorem 2.4 from [47]) gives the contra-
diction F (u) ≤ lim inf F (un ). Note that this result generalizes to Cn -valued
functions in a straightforward manner. 
Moreover, in this case our variational principle reads as follows:
Corollary 7.24. Let X be a reflexive Banach space and let M be a nonempty
closed convex subset. If F : M ⊆ X → R is quasiconvex, lower semicontinu-
ous, and, if M is unbounded, weakly coercive, then there exists some x0 ∈ M
with F (x0 ) = inf M F . If F is strictly quasiconvex then x0 is unique.

Proof. It remains to show uniqueness. Let x0 and x1 be two different min-

ima. Then F (λx0 + (1 − λ)x1 ) < max{F (x0 ), F (x1 )} = inf M F , a contradic-
tion. 
Example 7.22. Let X be a reflexive Banach space. Suppose M ⊆ X is a
nonempty closed convex set. Then for every x ∈ X there is a point x0 ∈ M
with minimal distance, kx − x0 k = dist(x, M ). Indeed, F (z) = dist(x, z) is
convex, continuous and, if M is unbounded weakly coercive. Hence the claim
follows from Corollary 7.24. Note that the assumption that X is reflexive
is crucial (Problem 7.12). Moreover, we also get that x0 is unique if X is
strictly convex (see Problem 1.13). 
Example 7.23. Let H be a Hilbert space and ` ∈ H∗ a linear functional.
We will give a variational proof of the Riesz lemma (Theorem 2.10). Since
we already need to know that Hilbert spaces are reflexive, it should not be
taken too serious. To this end consider
1
F (x) = kxk2 − Re `(x) ,


x ∈ H.
2
Then F is convex, continuous, and weakly coercive. Hence there is some
x0 ∈ H with F (x0 ) = inf x∈H F (x). Moreover, for fixed x ∈ H,

ε2
R → R, ε 7→ F (x0 + εx) = F (x0 ) + εRe hx0 , xi − `(x) + kxk2
2
is a smooth map which has a minimum at ε = 0. Hence its derivative at
ε = 0 must vanish: Re hx0 , xi
− `(x) = 0 for all x ∈ H. Replacing x → −ix

we also get Im hx0 , xi − `(x) = 0 and hence `(x) = hx0 , xi. 

Example 7.24. Let H be a Hilbert space and let us consider the problem
of finding the lowest eigenvalue of a positive operator A ≥ 0. Of course
this is bound to fail since the eigenvalues could accumulate at 0 without 0
being an eigenvalue (e.g. the multiplication operator with the sequence n1 in
`2 (N)). Nevertheless it is instructive to see how things can go wrong (and it
underlines the importance of our various assumptions).
7.4. Minimizing nonlinear functionals via compactness 217

To this end consider its quadratic form qA (f ) = hf, Af i. Then, since

1/2
qA is a seminorm (Problem 1.22) and taking squares is convex, qA is con-
vex. If we consider it on M = B̄1 (0) we get existence of a minimum from
Theorem 7.21. However this minimum is just qA (0) = 0 which is not very
interesting. In order to obtain a minimal eigenvalue we would need to take
M = S1 = {f | kf k = 1}, however, this set is not weakly closed (its weak
closure is B̄1 (0) as we will see in the next section). In fact, as pointed out
before, the minimum is in general not attained on M in this case.
Note that our problem with the trivial minimum at 0 would also dis-
appear if we would search for a maximum instead. However, our lemma
above only guarantees us weak sequential lower semicontinuity but not weak
sequential upper semicontinuity. In fact, note that not even the norm (the
quadratic form of the identity) is weakly sequentially upper continuous (cf.
Lemma 4.28 (ii) versus Lemma 4.29). If we make the additional assumption
that A is compact, then qA is weakly sequentially continuous as can be seen
from Theorem 4.32. Hence for compact operators the maximum is attained
at some vector f0 . Of course we will have kf0 k = 1 but is it an eigenvalue?
To see this we resort to a small ruse: Consider the real function
qA (f0 + tf ) α0 + 2tRehf, Af0 i + t2 qA (f )
φ(t) = = , α0 = qA (f0 ),
kf0 + tf k2 1 + 2tRehf, f0 i + t2 kf k2
which has a maximum at t = 0 for any f ∈ H. Hence we must have φ0 (0) =
2Rehf, (A − α0 )f0 i = 0 for all f ∈ H. Replacing f → if we get 2Imhf, (A −
α0 )f0 i = 0 and hence hf, (A − α0 )f0 i = 0 for all f , that is Af0 = α0 f . So
we have recovered Theorem 3.6. 
Example 7.25. Let us consider the following nonlinear elliptic problem
−∆u + u|u| + u = f
in L2R (Rn ) for a given function f ∈ L2R (Rn ). We are going to look for weak
solutions, that is, solutions u ∈ HR1 (Rn ) satisfying
Z
(∂u · ∂φ + (|u|u + u − f )φ) dn x = 0, φ ∈ Cc∞ (Rn ).
Rn
We start by introducing the functional
Z  
1 1 3 1 2
F (u) := |∂u| + |u| + u − uf dn x
2
R n 2 3 2
on L2R (Rn ) and set F (u) = ∞ if u 6∈ HR1 (Rn ) ∩ L3R (Rn ). One checks that
for u ∈ HR1 (Rn ) ∩ L3R (Rn ) and φ ∈ Cc∞ (Rn ) this functional has a variational
derivative
Z
δF (u, φ) = (∂u · ∂φ + (|u|u + u − f )φ) dn x = 0
Rn
218 7. Analysis in Banach spaces

which coincides with the weak formulation of our problem. Hence a mini-
mizer (which is necessarily in HR1 (Rn ) ∩ L3R (Rn )) is a weak solution of our
nonlinear elliptic problem and it remains to show existence of a minimizer.
First of all note that
1 1
F (u) ≥ kuk22 − kuk2 kf k2 ≥ kuk22 − kf k22
2 4
and hence F is coercive. To see that it is weakly sequentially lower contin-
uous, observe that for the second terms this follows from Example 7.21 and
the last two are easy. For the first term let un * u in L2 and observe

k∂uk2 = sup |hu, ∂φi| = sup lim |hun , ∂φi|

kφk2 =1,φ∈Cc∞ kφk2 =1,φ∈Cc∞ n
≤ lim inf sup |hun , ∂φi|
n kφk2 =1,φ∈Cc∞
= lim inf k∂un k2 .
n

Hence the claim follows. 

If we look at the previous problem in the case f = 0, our approach will
only give us the trivial solution. In fact, for a linear problem one expects
nontrivial solutions for the homogenous problem only at an eigenvalue. Since
the Laplace operator has no eigenvalues on Rn (as is not hard to see using
the Fourier transform), we look at a bounded domain U instead. To avoid
the trivial solution we will add a constraint. As a preparation we note

Lemma 7.25. Let X, Y be Banach spaces such that X is compactly embedded

into Y and let N : Y → R be continuous. Then M := {x ∈ X|N (x) = N0 }
is weakly sequentially closed for any N0 ∈ R. The same holds for M := {x ∈
X|N (x) ≤ N0 }.

Proof. This follows from Theorem 4.32 since every weakly convergent se-
quence in X is convergent in Y . 

Theorem 7.26 (Variational principle with constraints). Let X be a reflexive

Banach space and let F : X → R be weakly sequentially lower semicontinuous
and weakly coercive. Let Y be another Banach space such that X is compactly
embedded into Y and let N : Y → R be continuous. Fix N0 ∈ R and suppose
that M := {x ∈ X|N (x) = N0 } is nonempty. Then there exists some x0 ∈ M
with F (x0 ) = inf M F .
If in addition F is differentiable, N is Gâteaux differentiable and δN does
not vanish on M , then there is a constant λ ∈ R (the Lagrange multiplier)
such that
dF (x0 ) = λδN (x0 ). (7.37)
7.4. Minimizing nonlinear functionals via compactness 219

Proof. Existence follows from Theorem 7.21 which is applicable thanks to

our previous lemma. Now choose some x1 ∈ X such that δN (x0 )x1 6= 0 and
x ∈ X arbitrary. Then the function
f (t, s) := N (x0 + t x + x1 s)
is C 2 (R2 ) and satisfies
∂t f (t, s) = δN (x0 + t x + s x1 )x, ∂s f (t, s) = δN (x0 + t x + x1 s)x1
and since ∂s f (0, 0) 6= 0 the implicit function theorem implies existence of a
function σ ∈ C 1 (−ε, ε) such that σ(0) = 0 and f (t, σ(t)) = f (0, 0), that is,
x(t) := x0 + t x + σ(t)x1 ∈ M for |t| < ε. Moreover,
∂t f (0, 0) δN (x0 )x
σ 0 (0) = − =− .
∂s f (0, 0) δN (x0 )x1
Hence by the chain rule
d
F (x0 +t x+σ(t)x1 )|t=0 = dF (x0 )(x+σ 0 (0)x1 ) = dF (x0 )x−λδN (x0 )x = 0,
dt
where
dF (x0 )x1
λ := . 
δN (x0 )x1
Example 7.26. Let U ⊂ Rn be a bounded domain and consider
Z
1
F (u) := |∂u|2 dn x, u ∈ H01 (U, R)
2 U
subject to the constraint
Z
N (u) := G(u)dn x = N0 ,
U
where G : R → R is differentiable and satisfies
|G0 (x)| ≤ C(1 + |x|).
This condition implies
|G(x)| ≤ C̃(1 + |x|2 )
and ensures that N (u) is well-defined for all u ∈ L2 (U ).
In order to apply the theorem we set X = H01 (U ) and Y = L2 (U ). That
X is compactly embedded into Y is the Rellich–Kondrachov theorem (The-
orem 7.27 from [47]). Moreover, by the Poincaré inequality (Theorem 7.30
from [47]) we can choose kxk := F (x) as an equivalent norm on X. In
particular, F satisfies the requirements of our theorem and so does N by
Example 7.5. Consequently if N0 is such that
Z
1
M := {f ∈ H0 (U, R)| G(u)dn x = N0 }
U
220 7. Analysis in Banach spaces

is nonempty, there is a minimizer u0 . By Example 7.5

Z
δN (u0 )u = G0 (u0 )u dn x
U

and if we can find some u ∈ H01 (U )such that this derivative is nonzero, then
u0 satisfies
Z
∂u0 · ∂u − λG0 (u0 )u dn x = 0, u ∈ H01 (R),


U
and hence is a weak solution of the nonlinear eigenvalue problem
−∆u0 = λG0 (u0 ).
Note that is last condition is for example satisfied if G(0) 0
R =0 0, G (x)x >0
for x 6= 0, and N0 > 0. Indeed, in this case δN (u0 )u0 = U G (u0 )u0 dn x > 0
since otherwise we would have u0 = 0 contradicting 0 < N0 = N (u0 ) =
N (0) = 0.
Of course in the case G(x) = 12 |x|2 and N0 = 1 this gives us the lowest
eigenvalue of the Laplacian on U with Dirichlet boundary conditions.
Note that using continuous embeddings L2 ,→ Lp with 2 ≤ p ≤ ∞ for
n = 1, 2 ≤ p < ∞ for n = 2, and 2 ≤ p ≤ n−2
2n
for n ≥ 3 one can improve
this result to the case
|G0 (x)| ≤ C(1 + |x|p−1 ). 
R1
Problem 7.12. Consider X = C[0, 1] and M = {f | 0 f (x)dx = 1, f (0) =
0}. Show that M is closed and convex. Show that d(0, M ) = 1 but there is
no minimizer. If we replace the boundary condition by f (0) = 1 there is a
unique minimizer and for f (0) = 2 there are infinitely many minimizers.
Problem 7.13. Show that F : M → R is convex if and only if its epigraph
epi F := {(x, a) ∈ M × R|F (x) ≤ a} ⊂ X × R is convex.
Problem* 7.14. Show that F : M → R is quasiconvex if and only if the
sublevel sets F −1 ((−∞, a]) are convex for every a ∈ R.

7.5. Contraction principles

Let X be a Banach space. A fixed point of a mapping F : C ⊆ X → C is
an element x ∈ C such that F (x) = x. Moreover, F is called a contraction
if there is a contraction constant θ ∈ [0, 1) such that
|F (x) − F (x̃)| ≤ θ|x − x̃|, x, x̃ ∈ C. (7.38)
Note that a contraction is continuous. We also recall the notation F n (x) =
F (F n−1 (x)), F 0 (x) = x.
7.5. Contraction principles 221

Theorem 7.27 (Contraction principle). Let C be a nonempty closed subset

of a Banach space X and let F : C → C be a contraction, then F has a
unique fixed point x ∈ C such that
θn
|F n (x) − x| ≤ |F (x) − x|, x ∈ C. (7.39)
1−θ
Proof. If x = F (x) and x̃ = F (x̃), then |x − x̃| = |F (x) − F (x̃)| ≤ θ|x − x̃|
shows that there can be at most one fixed point.
Concerning existence, fix x0 ∈ C and consider the sequence xn = F n (x0 ).
We have
|xn+1 − xn | ≤ θ|xn − xn−1 | ≤ · · · ≤ θn |x1 − x0 |
and hence by the triangle inequality (for n > m)
n
X n−m−1
X
|xn − xm | ≤ |xj − xj−1 | ≤ θm θj |x1 − x0 |
j=m+1 j=0
θm
≤ |x1 − x0 |. (7.40)
1−θ
Thus xn is Cauchy and tends to a limit x. Moreover,
|F (x) − x| = lim |xn+1 − xn | = 0
n→∞

shows that x is a fixed point and the estimate (7.39) follows after taking the
limit m → ∞ in (7.40). 

Note that we can replace θn by any other summable sequence θn (Prob-

lem 7.16):
Theorem 7.28 (Weissinger). Let C be a nonempty closed subset of a Banach
space X. Suppose F : C → C satisfies
|F n (x) − F n (y)| ≤ θn |x − y|, x, y ∈ C, (7.41)
P∞
with n=1 θn < ∞. Then F has a unique fixed point x such that
 
∞
X
|F n (x) − x| ≤  θj  |F (x) − x|, x ∈ C. (7.42)
j=n

Next, we want to investigate how fixed points of contractions vary with

respect to a parameter. Let X, Y be Banach spaces, U ⊆ X, V ⊆ Y be
open and consider F : U × V → U . The mapping F is called a uniform
contraction if there is a θ ∈ [0, 1) such that
|F (x, y) − F (x̃, y)| ≤ θ|x − x̃|, x, x̃ ∈ U , y ∈ V, (7.43)
that is, the contraction constant θ is independent of y.
222 7. Analysis in Banach spaces

Theorem 7.29 (Uniform contraction principle). Let U , V be nonempty

open subsets of Banach spaces X, Y , respectively. Let F : U × V → U be a
uniform contraction and denote by x(y) ∈ U the unique fixed point of F (., y).
If F ∈ C r (U × V, U ), r ≥ 0, then x(.) ∈ C r (V, U ). If F is Lipschitz with
respect to the paramter, so is the fixed point.

Proof. Let us first show that x(y) is continuous. From

|x(y + v) − x(y)| = |F (x(y + v), y + v) − F (x(y), y + v)
+ F (x(y), y + v) − F (x(y), y)|
≤ θ|x(y + v) − x(y)| + |F (x(y), y + v) − F (x(y), y)|
we infer
1
|x(y + v) − x(y)| ≤ |F (x(y), y + v) − F (x(y), y)| (7.44)
1−θ
and hence x(y) ∈ C(V, U ). Now let r := 1 and let us formally differentiate
x(y) = F (x(y), y) with respect to y,
d x(y) = ∂x F (x(y), y)d x(y) + ∂y F (x(y), y). (7.45)
Considering this as a fixed point equation x0 ,
where T (., y) :
T (x0 , y)
=
L (Y, X) → L (Y, X), x 7→ ∂x F (x(y), y)x + ∂y F (x(y), y) is a uniform con-
0 0

traction since we have k∂x F (x(y), y)k ≤ θ by Theorem 7.9. Hence we get a
unique continuous solution x0 (y). It remains to show
x(y + v) − x(y) − x0 (y)v = o(v).
Let us abbreviate u := x(y + v) − x(y), then using (7.45) and the fixed point
property of x(y) we see
(1 − ∂x F (x(y), y))(u − x0 (y)v) =
= F (x(y) + u, y + v) − F (x(y), y) − ∂x F (x(y), y)u − ∂y F (x(y), y)v
= o(u) + o(v)
since F ∈ C 1 (U × V, U ) by assumption. Moreover, k(1 − ∂x F (x(y), y))−1 k ≤
(1 − θ)−1 and u = O(v) (by (7.44)) implying u − x0 (y)v = o(v) as desired.
Finally, suppose that the result holds for some r − 1 ≥ 1. Thus, if F is
Cr, then x(y) is at least C r−1 and the fact that d x(y) satisfies (7.45) shows
d x(y) ∈ C r−1 (V, U ) and hence x(y) ∈ C r (V, U ). 

As an important consequence we obtain the implicit function theorem.

Theorem 7.30 (Implicit function). Let X, Y , and Z be Banach spaces and
let U , V be open subsets of X, Y , respectively. Let F ∈ C r (U × V, Z), r ≥ 0,
and fix (x0 , y0 ) ∈ U ×V . Suppose ∂x F ∈ C(U ×V, L (X, Z)) exists (if r = 0)
and ∂x F (x0 , y0 ) ∈ L (X, Z) is an isomorphism. Then there exists an open
neighborhood U1 × V1 ⊆ U × V of (x0 , y0 ) such that for each y ∈ V1 there
7.5. Contraction principles 223

exists a unique point (ξ(y), y) ∈ U1 × V1 satisfying F (ξ(y), y) = F (x0 , y0 ).

Moreover, ξ is in C r (V1 , Z) and fulfills (for r ≥ 1)

dξ(y) = −(∂x F (ξ(y), y))−1 ◦ ∂y F (ξ(y), y). (7.46)

Proof. Using the shift F → F − F (x0 , y0 ) we can assume F (x0 , y0 ) = 0.

Next, the fixed points of G(x, y) = x − (∂x F (x0 , y0 ))−1 F (x, y) are the solu-
tions of F (x, y) = 0. The function G has the same smoothness properties
as F and since ∂x G(x0 , y0 ) = 0, we can find balls U1 and V1 around x0
and y0 such that k∂x G(x, y)k ≤ θ < 1 for (x, y) ∈ U1 × V1 . Thus by the
mean value theorem (Theorem 7.9) G(., y) is a uniform contraction on U1
for y ∈ V1 . Moreover, choosing the radius of V1 sufficiently small such that
|G(x0 , y) − G(x0 , y0 )| < (1 − θ)r for y ∈ V1 , where r is the radius of U1 , we
get

|G(x, y) − x0 | = |G(x, y) − G(x0 , y0 )| ≤ θ|x − x0 | + (1 − θ)r < r

for (x, y) ∈ U1 × V1 , that is, G : U1 × V1 → U1 . The rest follows from the

uniform contraction principle. Formula (7.46) follows from differentiating
F (ξ(y), y) = 0 using the chain rule. 

Note that our proof is constructive, since it shows that the solution ξ(y)
can be obtained by iterating x − (∂x F (x0 , y0 ))−1 (F (x, y) − F (x0 , y0 )).
Moreover, as a corollary of the implicit function theorem we also obtain
the inverse function theorem.

Theorem 7.31 (Inverse function). Suppose F ∈ C r (U, Y ), r ≥ 1, U ⊆

X, and let dF (x0 ) be an isomorphism for some x0 ∈ U . Then there are
neighborhoods U1 , V1 of x0 , F (x0 ), respectively, such that F ∈ C r (U1 , V1 ) is
a diffeomorphism.

Proof. Apply the implicit function theorem to G(x, y) = y − F (x). 

Example 7.27. It is important to emphasize that invertibility of dF on all of

U does not imply injectivity on U as the following example in X := R2 shows:
F (x, y) = (e2x − y 2 + 3, 4e2x y − y 3 ). Note that det ∂(x,y)
∂F
= 8e4x + 10e2x y 2
and F (0, 2) = (0, 0) = F (0, −2). 
Example 7.28. Let X be a Banach algebra and G(X) the group of invertible
elements. We have seen that multiplication is C ∞ (X × X, X) and hence
taking the inverse is also C ∞ (G(X), G(X)). Consequently, G(X) is an (in
general infinite-dimensional) Lie group. 
Further applications will be given in the next section.
224 7. Analysis in Banach spaces

Problem 7.15. Derive Newton’s method for finding the zeros of a twice
continuously differentiable function f (x),
f (x)
xn+1 = F (xn ), F (x) = x − ,
f 0 (x)
from the contraction principle by showing that if x is a zero with f 0 (x) 6=
0, then there is a corresponding closed interval C around x such that the
assumptions of Theorem 7.27 are satisfied.

Problem* 7.16. Prove Theorem 7.28. Moreover, suppose F : C → C and

that F n is a contraction. Show that the fixed point of F n is also one of
F . Hence Theorem 7.28 (except for the estimate) can also be considered as
a special case of Theorem 7.27 since the assumption implies that F n is a
contraction for n sufficiently large.

7.6. Ordinary differential equations

As a first application of the implicit function theorem, we prove (local) ex-
istence and uniqueness for solutions of ordinary differential equations in Ba-
nach spaces. Let X be a Banach space, U ⊆ X a (nonempty) open subset,
and I ⊆ R a compact interval. Denote by C(I, U ) the Banach space of
bounded continuous functions equipped with the sup norm.
The following lemma, known as omega lemma, will be needed in the
proof of the next theorem.

Lemma 7.32. Suppose I ⊆ R is a compact interval and f ∈ C r (U, Y ). Then

f∗ ∈ C r (C(I, U ), C(I, Y )), where
(f∗ x)(t) := f (x(t)). (7.47)

Proof. Fix x0 ∈ C(I, U ) and ε > 0. For each t ∈ I we have a δ(t) > 0
such that B2δ(t) (x0 (t)) ⊂ U and |f (x) − f (x0 (t))| ≤ ε/2 for all x with
|x − x0 (t)| ≤ 2δ(t). The balls Bδ(t) (x0 (t)), t ∈ I, cover the set {x0 (t)}t∈I
and since I is compact, there is a finite subcover Bδ(tj ) (x0 (tj )), 1 ≤ j ≤
n. Let kx − x0 k ≤ δ := min1≤j≤n δ(tj ). Then for each t ∈ I there is
a tj such that |x0 (t) − x0 (tj )| ≤ δ(tj ) and hence |f (x(t)) − f (x0 (t))| ≤
|f (x(t)) − f (x0 (tj ))| + |f (x0 (tj )) − f (x0 (t))| ≤ ε since |x(t) − x0 (tj )| ≤
|x(t) − x0 (t)| + |x0 (t) − x0 (tj )| ≤ 2δ(tj ). This settles the case r = 0.
Next let us turn to r = 1. We claim that df∗ is given by (df∗ (x0 )x)(t) :=
df (x0 (t))x(t). To show this we use Taylor’s theorem (cf. the proof of Corol-
lary 7.14) to conclude that
|f (x0 (t) + x) − f (x0 (t)) − df (x0 (t))x| ≤ |x| sup kdf (x0 (t) + sx) − df (x0 (t))k.
0≤s≤1
7.6. Ordinary differential equations 225

By the first part (df )∗ is continuous and hence for a given ε we can find a
corresponding δ such that |x(t) − y(t)| ≤ δ implies kdf (x(t)) − df (y(t))k ≤ ε
and hence kdf (x0 (t) + sx) − df (x0 (t))k ≤ ε for |x0 (t) + sx − x0 (t)| ≤ |x| ≤ δ.
But this shows differentiability of f∗ as required and it remains to show that
df∗ is continuous. To see this we use the linear map
λ : C(I, L (X, Y )) → L (C(I, X), C(I, Y )) ,
T 7→ T∗
where (T∗ x)(t) := T (t)x(t). Since we have
kT∗ xk = sup |T (t)x(t)| ≤ sup kT (t)k|x(t)| ≤ kT kkxk,
t∈I t∈I
we infer |λ| ≤ 1 and hence λ is continuous. Now observe df∗ = λ ◦ (df )∗ .
The general case r > 1 follows from induction. 

Now we come to our existence and uniqueness result for the initial value
problem in Banach spaces.
Theorem 7.33. Let I be an open interval, U an open subset of a Banach
space X and Λ an open subset of another Banach space. Suppose F ∈ C r (I ×
U × Λ, X), r ≥ 1, then the initial value problem
ẋ = F (t, x, λ), x(t0 ) = x0 , (t0 , x0 , λ) ∈ I × U × Λ, (7.48)
has a unique solution x(t, t0 , x0 , λ) ∈ C r (I1
× I2 × U1 × Λ1 , U ), where I1,2 ,
U1 , and Λ1 are open subsets of I, U , and Λ, respectively. The sets I2 , U1 ,
and Λ1 can be chosen to contain any point t0 ∈ I, x0 ∈ U , and λ0 ∈ Λ,
respectively.

Proof. Adding t and λ to the dependent variables x, that is considering

(τ, x, λ) ∈ R × X × Λ and augmenting the differential equation according to
(τ̇ , ẋ, λ̇) = (1, F (τ, x, λ), 0), we can assume that F is independent of t and
λ. Moreover, by a translation we can even assume t0 = 0.
Our goal is to invoke the implicit function theorem. In order to do this
we introduce an additional parameter ε ∈ R and consider
ẋ = εF (x0 + x), x ∈ D1 := {x ∈ C 1 ([−1, 1], Bδ (0))|x(0) = 0}, (7.49)
such that we know the solution for ε = 0. The implicit function theorem will
show that solutions still exist as long as ε remains small. At first sight this
doesn’t seem to be good enough for us since our original problem corresponds
to ε = 1. But since ε corresponds to a scaling t → εt, the solution for one
ε > 0 suffices. Now let us turn to the details.
Our problem (7.49) is equivalent to looking for zeros of the function
G : D1 × U0 × R → C([−1, 1], X),
(7.50)
(x, x0 , ε) 7→ ẋ − εF (x0 + x),
226 7. Analysis in Banach spaces

where U0 is a neighborhood of x0 and δ sufficiently small such that U0 +

Bδ (0) ⊆ U . Lemma 7.32 ensures that this function is C 1 . Now fix x0 , then
G(0, x0 , 0) = 0 and ∂x G(0, x0 , 0) = T , where T x := ẋ. Since (T −1 x)(t) =
Rt
0 x(s)ds we can apply the implicit function theorem to conclude that there
is a unique solution x(x0 , ε) ∈ C 1 (U1 × (−ε0 , ε0 ), D1 ) ,→ C 1 ([−1, 1] × U1 ×
(−ε0 , ε0 ), X). In particular, the map (t, x0 ) 7→ x0 + x(x0 , ε)(t/ε) is in
C 1 ((−ε, ε) × U1 , X). Hence it is the desired solution of our original problem.
This settles the case r = 1.
For r > 1 we use induction. Suppose F ∈ C r+1 and let x(t, x0 ) be the
solution which is at least C r . Moreover, y(t, x0 ) := ∂x0 x(t, x0 ) satisfies
ẏ = ∂x F (x(t, x0 ))y, y(0) = I,
and hence y(t, x0 ) ∈ C r . Moreover, the differential equation shows ∂t x(t, x0 ) =
F (x(t, x0 )) ∈ C r which shows x(t, x0 ) ∈ C r+1 . 
Example 7.29. The simplest example is a linear equation
ẋ = Ax, x(0) = x0 ,
where A ∈ L (X). Then it is easy to verify that the solution is given by
x(t) = exp(tA)x0 ,
where
∞ k
X t
exp(tA) := Ak .
k!
k=0
It is easy to check that the last series converges absolutely (cf. also Prob-
lem 1.36) and solves the differential equation (Problem 7.17). 
Example 7.30. The classical example ẋ = x2 , x(0) = x0 , in X := R with
solution 
(−∞, x10 ), x0 > 0,
x0 
x(t) = , t ∈ R, x0 = 0,
1 − x0 t 
 1
( x0 , ∞), x0 < 0.
shows that solutions might not exist for all t ∈ R even though the differential
equation is defined for all t ∈ R. 
This raises the question about the maximal interval on which a solution
of the initial value problem (7.48) can be defined.
Suppose that solutions of the initial value problem (7.48) exist locally and
are unique (as guaranteed by Theorem 7.33). Let φ1 , φ2 be two solutions
of (7.48) defined on the open intervals I1 , I2 , respectively. Let I := I1 ∩
I2 = (T− , T+ ) and let (t− , t+ ) be the maximal open interval on which both
solutions coincide. I claim that (t− , t+ ) = (T− , T+ ). In fact, if t+ < T+ ,
both solutions would also coincide at t+ by continuity. Next, considering the
7.6. Ordinary differential equations 227

initial value problem with initial condition x(t+ ) = φ1 (t+ ) = φ2 (t+ ) shows
that both solutions coincide in a neighborhood of t+ by local uniqueness.
This contradicts maximality of t+ and hence t+ = T+ . Similarly, t− = T− .
Moreover, we get a solution
(
φ1 (t), t ∈ I1 ,
φ(t) := (7.51)
φ2 (t), t ∈ I2 ,
defined on I1 ∪ I2 . In fact, this even extends to an arbitrary number of
solutions and in this way we get a (unique) solution defined on some maximal
interval.
Theorem 7.34. Suppose the initial value problem (7.48) has a unique local
solution (e.g. the conditions of Theorem 7.33 are satisfied). Then there ex-
ists a unique maximal solution defined on some maximal interval I(t0 ,x0 ) =
(T− (t0 , x0 ), T+ (t0 , x0 )).

Proof. Let S be the set of all Ssolutions φ of (7.48) which are defined on
an open interval Iφ . Let I := φ∈S Iφ , which is again open. Moreover, if
t1 > t0 ∈ I, then t1 ∈ Iφ for some φ and thus [t0 , t1 ] ⊆ Iφ ⊆ I. Similarly for
t1 < t0 and thus I is an open interval containing t0 . In particular, it is of the
form I = (T− , T+ ). Now define φmax (t) on I by φmax (t) := φ(t) for some
φ ∈ S with t ∈ Iφ . By our above considerations any two φ will give the same
value, and thus φmax (t) is well-defined. Moreover, for every t1 > t0 there is
some φ ∈ S such that t1 ∈ Iφ and φmax (t) = φ(t) for t ∈ (t0 − ε, t1 + ε) which
shows that φmax is a solution. By construction there cannot be a solution
defined on a larger interval. 

The solution found in the previous theorem is called the maximal so-
lution. A solution defined for all t ∈ R is called a global solution. Clearly
every global solution is maximal.
The next result gives a simple criterion for a solution to be global.
Lemma 7.35. Suppose F ∈ C 1 (R×X, X) and let x(t) be a maximal solution
of the initial value problem (7.48). Suppose |F (t, x(t))| is bounded on finite
t-intervals. Then x(t) is a global solution.

Proof. Let (T− , T+ ) be the domain of x(t) and suppose T+ < ∞. Then
|F (t, x(t))| ≤ C for t ∈ (t0 , T+ ) and for t0 < s < t < T+ we have
Z t Z t
|x(t) − x(s)| ≤ |ẋ(τ )|dτ = |F (τ, x(τ ))|dτ ≤ C|t − s|.
s s

Thus x(tn ) is Cauchy whenever tn is and hence limt→T+ x(t) = x+ exists.

Now let y(t) be the solution satisfying the initial condition y(T+ ) = x+ .
228 7. Analysis in Banach spaces

Then (
x(t), t < T+ ,
x̃(t) =
y(t), t ≥ T+ ,
is a larger solution contradicting maximality of T+ . 
Example 7.31. Finally, we want to apply this to a famous example, the so-
called FPU lattices (after Enrico Fermi, John Pasta, and Stanislaw Ulam
who investigated such systems numerically). This is a simple model of a
linear chain of particles coupled via nearest neighbor interactions. Let us
assume for simplicity that all particles are identical and that the interaction
is described by a potential V ∈ C 2 (R). Then the equation of motions are
given by
q̈n (t) = V 0 (qn+1 − qn ) − V 0 (qn − qn−1 ), n ∈ Z,
where qn (t) ∈ R denotes the position of the n’th particle at time t ∈ R and
the particle index n runs through all integers. If the potential is quadratic,
V (r) = k2 r2 , then we get the discrete linear wave equation


q̈n (t) = k qn+1 (t) − 2qn (t) + qn−1 (t) .
If we use the fact that the Jacobi operator Aqn = −k(qn+1 − 2qn + qn−1 ) is
a bounded operator in X = `pR (Z) we can easily solve this system as in the
case of ordinary differential equations. In fact, if q 0 = q(0) and p0 = q̇(0)
are the initial conditions then one can easily check (cf. Problem 7.17) that
the solution is given by
sin(tA1/2 ) 0
q(t) = cos(tA1/2 )q 0 + p .
A1/2
In the Hilbert space case p = 2 these functions of our operator A could
be defined via the spectral theorem but here we just use the more direct
definition
∞ ∞
1/2
X t2k k sin(tA1/2 ) X t2k+1
cos(tA ) := A , := Ak .
(2k)!
k=0
A1/2 (2k + 1)!
k=0
In the general case an explicit solution is no longer possible but we are still
able to show global existence under appropriate conditions. To this end
we will assume that V has a global minimum at 0 and hence looks like
V (r) = V (0) + k2 r2 + o(r2 ). As V (0) does not enter our differential equation
we will assume V (0) = 0 without loss of generality. Moreover, we will also
introduce pn := q̇n to have a first order system
q̇n = pn , ṗn = V 0 (qn+1 − qn ) − V 0 (qn − qn−1 ).
Since V 0 ∈ C 1 (R) with V 0 (0) = 0 it gives rise to a C 1 map on `pR (N) (see
Example 7.2). Since the same is true for shifts, the chain rule implies that
the right-hand side of our system is a C 1 map and hence Theorem 7.33 gives
7.6. Ordinary differential equations 229

us existence of a local solution. To get global solutions we will need a bound

on solutions. This will follow from the fact that the energy of the system
X  p2 
n
H(p, q) := + V (qn+1 − qn )
2
n∈Z

is conserved. To ensure that the above sum is finite we will choose X :=

`2R (Z) ⊕ `2R (Z) as our underlying Banach (in this case even Hilbert) space.
Recall that since we assume V to have a minimum at 0 we have |V (r)| ≤
CR r2 for |r| < R and hence H(p, q) < ∞ for (p, q) ∈ X. Under these
assumptions it is easy to check that H ∈ C 1 (X, R) and that
d X

H(p(t), q(t)) = ṗn (t)pn (t) + V 0 (qn+1 (t) − qn (t)) q̇n+1 (t) − q̇n (t)
dt
n∈Z
X
V 0 (qn+1 − qn ) − V 0 (qn − qn−1 ) pn (t)


=
n∈Z


+ V 0 (qn+1 (t) − qn (t)) pn+1 (t) − pn (t)
X 
= − V 0 (qn − qn−1 )pn (t) + V 0 (qn+1 (t) − qn (t))pn+1 (t)
n∈Z
=0
provided (p(t), q(t)) solves our equation. Consequently, since V ≥ 0,
kp(t)k22 ≤ 2H(p(t), q(t)) = 2H(p(0), q(0)).
Rt
Moreover, qn (t) = qn (0) + 0 pn (s)ds (note that since the `2 norm is stronger
than the `∞ norm, qn (t) is differentiable for fixed n) implies
Z t p
kq(t)k2 ≤ kq(0)k2 + kpn (s)k2 ds ≤ kq(0)k2 + 2H(p(0), q(0))t.
0
So Lemma 7.35 ensures that solutions are global in X. Of course every
solution from X is also a solution from Y = `pR (Z) ⊕ `pR (Z) for all p ≥ 2
(since the k.k2 norm is stronger than the k.kp norm for p ≥ 2).
Examples include the original FPU β-model Vβ (r) := 12 r2 + β4 r4 , β > 0,
and the famous Toda lattice V (r) := e−r + r − 1. 
Example 7.32. Consider the discrete nonlinear Schrödinger equation
(dNLS)
iu̇(t) = Hu(t) ± |u(t)|2p u(t), t ∈ R,
where Hun = un+1 + un−1 + qn un is the Jacobi operator, in X = `2 (Z) ∼
=
`2R (Z) ⊕ `2R (Z). Here q ∈ `∞ (Z) is a real-valued sequence corresponding to
an external potential and q = 0 (or q = −2, depending on your preferences)
is the free Jacobi operator. Clearly the right-hand side is C 1 for p ≥ 0 and
hence there is a unique local solution by Theorem 7.33. Please note that
230 7. Analysis in Banach spaces

even though X is a complex Banach space we consider it as a real Banach

space
Moreover, for a solution we have
d
ku(t)k22 = 2Rehu̇(t), u(t)i = 2Im hHu, ui ± h|u(t)|2p u(t), u(t)i = 0


dt
and hence the dNLS has a unique global norm preserving solution u ∈
C 1 (R, `2 (Z)). Note that this in fact works for any self-adjoint H ∈ L (X). 
It should be mentioned that the above theory does not suffice to cover
partial differential equations. In fact, if we replace the difference operator
by a differential operator we run into the problem that differentiation is not
a continuous process!
Problem* 7.17. Let
∞
X
f (z) := fj z j , |z| < R,
j=0
be a convergent power series with convergence radius R > 0. Suppose X is
a Banach space and A ∈ L (X) is a bounded operator with kAk < R. Show
that
∞
X
f (tA) := fj tj Aj
j=0
is in C ∞ (I, L (X)),
I= (−RkAk , RkAk−1 )
−1 and
dn
f (tA) = An f (n) (tA), n ∈ N0 .
dtn
(Compare also Problem 1.36.)
Problem 7.18. Consider the FPU α-model Vα (r) := 12 r2 + α3 r3 . Show that
1
solutions satisfying kqn+1 (0) − qn (0)k∞ < |α| and H(p(0), q(0)) < 6α1 2 are
global in X := `2 (Z)⊕`2 (Z). (Hint: Of course local solutions follow from our
considerations above. Moreover, note that Vα (r) has a maximum at r = − α1 .
Now use conservation of energy to conclude that the solution cannot escape
1
the region |r| < |α| .)

7.7. Bifurcation theory

One of the most basic tasks is finding the zeros of a given function F ∈
C k (U, Y ), where U ⊆ X and X, Y are Banach spaces. Frequently the
equation will depend on a parameter µ ∈ R (of course we could also consider
the case where the parameter is again in some Banach space, but we will
only consider this basic case here). That is, we are looking for solutions x(µ)
of the equation
F (µ, x) = 0, (7.52)
7.7. Bifurcation theory 231

where F ∈ C k (I × U, Y ) for some suitable k ∈ N and some open interval

I ⊆ R. Moreover, we are interested in the case of values µ0 ∈ I, where there
is a change in the number of solutions (i.e. where new solutions appear or
old solutions disappear as µ increases). Such points µ0 ∈ I will be called
bifurcation points. Clearly this cannot happen at a point where the im-
plicit function theorem is applicable and hence a necessary condition for a
bifurcation point is that

∂x F (µ0 , x0 ) (7.53)

is not invertible at some point x0 satisfying the equation F (µ0 , x0 ) = 0.

Example 7.33. Consider f (µ, x) = x2 −µ in C ∞ (R×R, R). Then f (µ, x) =
x2 − µ = 0, ∂x f (µ, x) = 2x = 0 shows that µ0 = 0 is the only possible
bifurcation point. Since there are no solutions for µ < 0 and two solutions
√
x0 (µ) = ± µ for µ > 0, we see that µ0 = 0 is a bifurcation point. Consider
f (µ, x) = x3 − µ in C ∞ (R × R, R). Then f (µ, x) = x3 − µ = 0, ∂x f (µ, x) =
3x2 = 0 shows that again µ0 = 0 is the only possible bifurcation point.
However, this time there is only one solution x(µ) = sign(µ)|µ|1/3 for all
µ ∈ R and hence there is no bifurcation occurring at µ0 = 0. 
So the derivative ∂x F tells us where to look for bifurcation points while
further derivatives can be used to determine what kind of bifurcation (if
any) occurs. Here we want to show how this can be done in the infinite
dimensional case.
Suppose we have an abstract problem F (µ, x) = 0 with µ ∈ R and x ∈ X
some Banach space. We assume that F ∈ C 1 (R × X, X) and that there is a
trivial solution x = 0, that is, F (µ, 0) = 0.
The first step is to split off the trivial solution and reduce it effectively
to a finite-dimensional problem. To this end we assume that we have found
a point µ0 ∈ R such that the derivative A := ∂x F (µ0 , 0) is not invertible.
Moreover, we will assume that A is a Fredholm operator such that there
exists (cf. Section 6.6) continuous linear projections

P : X = Ker(A) u X0 → Ker(A), Q : X = X1 u Ran(A) → X1 . (7.54)

Now split our equation into a system of two equations according to the above
splitting of the underlying Banach space:

F (µ, x) = 0 ⇔ F1 (µ, u, v) = 0, F2 (µ, u, v) = 0, (7.55)

where x = u+v with u = P x ∈ Ker(A), v = (1−P )x ∈ X0 and F1 (µ, u, v) =

Q F (µ, u + v), F2 (µ, u, v) = (1 − Q)F (µ, u + v).
232 7. Analysis in Banach spaces

Since P, Q are bounded, this system is still C 1 and the derivatives are
given by (recall the block structure of A from (6.75))
∂u F1 (µ0 , 0, 0) = 0, ∂v F1 (µ0 , 0, 0) = 0,
∂u F2 (µ0 , 0, 0) = 0, ∂v F2 (µ0 , 0, 0) = A0 . (7.56)
Moreover, since A0 is an isomorphism, the implicit function theorem tells us
that we can (locally) solve F2 for v. That is, there exists a neighborhood U
of (µ0 , 0) ∈ R × Ker(A) and a unique function ψ ∈ C 1 (U, X0 ) such that
F2 (µ, u, ψ(µ, u)) = 0, (µ, u) ∈ U. (7.57)
In particular, by the uniqueness part we have ψ(µ, 0) = 0. Moreover,
∂u ψ(µ0 , 0) = −A−1
0 ∂u F2 (µ0 , 0, 0) = 0.
Plugging this into the first equation reduces to the original system to the
finite dimensional system
F̃1 (µ, u) = F1 (µ, u, ψ(µ, u)) = 0. (7.58)
Of course the chain rule tells us that F̃ ∈ C 1 . Moreover, we still have
F̃1 (µ, 0) = F1 (µ, 0, ψ(µ, 0)) = QF (µ, 0) = 0 as well as
∂u F̃1 (µ0 , 0) = ∂u F1 (µ0 , 0, 0) + ∂v F1 (µ0 , 0, 0)∂u ψ(µ0 , 0) = 0. (7.59)
This is known as Lyapunov–Schmidt reduction.
Now that we have reduced the problem to a finite-dimensional system,
it remains to find conditions such that the finite dimensional system has a
nontrivial solution. For simplicity we make the requirement
dim Ker(A) = dim Coker(A) = 1 (7.60)
such that we actually have a problem in R × R → R.
Explicitly, let u0 span Ker(A) and let u1 span X1 . Then we can write
F̃1 (µ, λu0 ) = f (µ, λ)u1 , (7.61)
where f ∈ C 1 (V, R) with V = {(µ, λ)|(µ, λu0 ) ∈ U } ⊆ R2 a neighborhood of
(µ0 , 0). Of course we still have f (µ, 0) = 0 for (µ, 0) ∈ V as well as
∂λ f (µ0 , 0)u1 = ∂u F̃1 (µ0 , 0)u0 = 0. (7.62)
It remains to investigate f . To split off the trivial solution it suggests itself
to write
f (µ, λ) = λ g(µ, λ) (7.63)
We already have
g(µ0 , 0) = ∂λ f (µ0 , 0) = 0 (7.64)
and hence if
0 6= ∂µ g(µ0 , 0) = ∂µ ∂λ f (µ0 , 0) 6= 0 (7.65)
7.7. Bifurcation theory 233

the implicit function theorem implies existence of a function µ(λ) with µ(0) =
∂ 2 f (µ ,0)
µ0 and g(µ(λ), λ) = 0. Moreover, µ0 (0) = − ∂∂µλ g(µ0 ,0)
g(µ0 ,0) = − 2∂µ ∂λ f (µ0 ,0) .
λ 0

Of course this last condition is a bit problematic since up to this point

we only have f ∈ C 1 and hence g ∈ C 0 . However, if we change our original
assumption to F ∈ C 2 we get f ∈ C 2 and thus g ∈ C 1 .
So all we need to do is to trace back our definitions and compute
 
∂λ2 f (µ0 , 0)u1 = ∂λ2 F̃1 (µ0 , λu0 ) = ∂λ2 F1 (µ0 , λu0 , ψ(µ0 , λu0 ))
 
λ=0 λ=0
2
= ∂u F1 (µ0 , 0, 0)(u0 , u0 ) = Q∂x2 F (µ0 , 0)(u0 , u0 )
(recall ∂u ψ(µ0 , 0) = 0) and

∂µ ∂λ f (µ0 , 0)u1 = ∂µ ∂λ F̃1 (µ, λu0 )

λ=0,µ=µ0

= ∂µ ∂λ F1 (µ, λu0 , ψ(µ, λu0 ))

λ=0,µ=µ0
= Q∂µ ∂x F (µ0 , 0)u0 .
Theorem 7.36 (Crandall–Rabinowitz). Assume F ∈ C 2 (R × X, X) with
F (µ, 0) = 0 for all µ ∈ R. Suppose that for some µ0 ∈ R we have that
∂x F (µ0 , 0) is a Fredholm operator of index zero with a one-dimensional kernel
spanned by u0 ∈ X. Then, if
∂µ ∂x F (µ0 , 0)u0 6∈ Ran(∂x F (µ0 , 0)) (7.66)
there are open neighborhoods I ⊆ R of 0, J ⊆ R of µ0 , and U ⊆ span{u0 }
of 0 plus corresponding functions µ ∈ C 1 (I, J) and ψ ∈ C 2 (J × U, X0 ) such
that every nontrivial solution of F (µ, x) = 0 in a neighborhood of (µ0 , 0) is
given by
x(λ) = λu0 + ψ(µ(λ), λu0 ). (7.67)
Moreover,
`1 (∂x2 F (µ0 , 0)(u0 , u0 ))
µ(λ) = µ0 − λ + o(λ), x(λ) = λu0 + o(λ). (7.68)
2`1 (∂µ ∂x F (µ0 , 0)u0 )
where `1 is any nontrivial functional which vanishes on Ran(∂x F (µ0 , 0)).

Note that if Q∂x2 F (µ0 , 0)(u0 , u0 ) 6= 0 we could have also solved for λ
obtaining a function λ(µ) with λ(µ0 ) = 0. However, in this case it is not
obvious that λ(µ) 6= 0 for µ 6= µ0 , and hence that we get a nontrivial
solution, unless we also require Q∂µ ∂x F (µ0 , 0)u0 6= 0 which brings us back
to our previous condition. If both conditions are met, then µ0 (0) 6= 0 and
there is a unique nontrivial solution x(µ) which crosses the trivial solution
non transversally at µ0 . This is known as transcritical bifurcation. If
µ0 (0) = 0 but µ00 (0) 6= 0 (assuming this derivative exists), then two solutions
will branch off (either for µ > µ0 or for µ < µ0 depending on the sign of
234 7. Analysis in Banach spaces

the second derivative). This is known as a pitchfork bifurcation and is

typical in case of an odd function F (µ, −x) = −F (µ, x).
Example 7.34. Now we can establish existence of a stationary solution of
the dNLS of the form
un (t) = e−itω φn (ω)
Plugging this ansatz into the dNLS we get the stationary dNLS
Hφ − ωφ ± |φ|2p φ = 0.
Of course we always have the trivial solution φ = 0.
Applying our analysis to
1
F (ω, φ) = (H − ω)φ ± |φ|2p φ, p> ,
2
we have (with respect to φ = φr + iφi ∼
= (φr , φi ))
 2 
2(p−1) φr φr φi
∂φ F (ω, φ)u = (H − ω)u ± 2p|φ| u ± |φ|2p u
φr φi φ2i
and in particular ∂φ F (ω, 0) = H −ω and hence ω must be an eigenvalue of H.
In fact, if ω0 is a discrete eigenvalue, then self-adjointness implies that H −ω0
is Fredholm of index zero. Moreover, if there are two eigenfunction u and v,
then one checks that the Wronskian W (u, v) = u(n)v(n+1)−u(n+1)v(n) is
constant. But square summability implies that the Wronskian must vanish
and hence u and v must be linearly dependent (note that a solution of Hu =
ω0 u vanishing at two consecutive points must vanish everywhere). Hence
eigenvalues are always simple for our Jacobi operator H. Finally, if u0 is the
eigenfunction corresponding to ω0 we have
∂ω ∂φ F (ω0 , 0)u0 = −u0 6∈ Ran(H − ω0 ) = Ker(H − ω0 )⊥
and the Crandall–Rabinowitz theorem ensures existence of a stationary so-
lution φ for ω in a neighborhood of ω0 . Note that
∂φ2 F (ω, φ)(u, v) = ±2p(2p + 1)|φ|2p−1 sign(φ)uv
and hence ∂φ2 F (ω, 0) = 0. This is of course not surprising and related to
the symmetry F (ω, −φ) = −F (ω, φ) which implies that zeros branch off in
symmetric pairs.
Of course this leaves the problem of finding a discrete eigenvalue open.
One can show that for the free operator H0 (with q = 0) the spectrum
is σ(H0 ) = [−2, 2] and that there are no eigenvalues (in fact, the discrete
Fourier transform will map H0 to a multiplication operator in L2 [−π, π]). If
q ∈ c0 (Z), then the corresponding multiplication operator is compact and
σess (H) = σ(H0 ) by Weyl’s theorem (Theorem 6.36). Hence every point in
σ(H) \ [−2, 2] will be an isolated eigenvalue. 
7.7. Bifurcation theory 235

Problem 7.19. Show that if F (µ, −x) = −F (µ, x), then ψ(µ, −u) = −ψ(µ, u)
and µ(−λ) = µ(λ).
 Chapter 8

Operator semigroups

In this chapter we want to look at (semi)linear ordinary linear differential

equations in Banach spaces. We will need a few relevant facts about differ-
entiation and integration for Banach space valued functions. Section 7.1 will
be sufficient.

8.1. Uniformly continuous operator groups

Our aim is to investigate the abstract Cauchy problem
u̇ = Au, u(0) = u0 (8.1)
in some Banach space X. Here A is some linear operator and we will assume
that A ∈ L (X) to begin with. Note that in the simplest case X = Rn this
is just a linear first order system with constant coefficient matrix A. In this
case the solution is given by
u(t) = T (t)u0 , (8.2)
where
∞ j
X t
T (t) := exp(tA) := Aj (8.3)
j!
j=0
is the exponential of tA. It is not difficult to see that this also gives the
solution in our Banach space setting.
Theorem 8.1. Let A ∈ L (X). Then the series in (8.3) converges and
defines a uniformly continuous operator group:
(i) The map t 7→ T (t) is continuous, T ∈ C(R, L (X)).
(ii) T (0) = I and T (t + s) = T (t)T (s) for all t, s ∈ R.

237
238 8. Operator semigroups

Moreover, T ∈ C 1 (R, L (X)) is the unique solution of Ṫ (t) = AT (t) with

T (0) = I and it commutes with A, AT (t) = T (t)A.

Proof. Set
n
X tj
Tn (t) := Aj .
j!
j=0
Then (for m ≤ n)

n n
X tj j |t|j |t|m+1

X
kTn (t)−Tm (t)k =
A ≤
kAkj ≤ kAkm+1 e|t|kAk .
j=m+1 j! j=m+1 j! (m + 1)!

In particular,
kT (t)k ≤ e|t| kAk
and AT (t) = limn→∞ ATn (t) = limn→∞ Tn (t)A = T (t)A. Furthermore we
have Ṫn+1 = ATn and thus
Z t
Tn+1 (t) = I + ATn (s)ds.
0
Taking limits shows
Z t
T (t) = I + AT (s)ds
0
or equivalently T (t) ∈ C 1 (R, L (X)) and Ṫ (t) = AT (t), T (0) = I.
Suppose S(t) is another solution, Ṡ = AS, S(0) = I. Then, by the
product rule (Problem 7.1), dtd
T (−t)S(t) = T (−t)AS(t) − AT (−t)S(t) = 0
implying T (−t)S(t) = T (0)S(0) = I. In the special case T = S this shows
T (−t) = T −1 (t) and in the general case it hence proves uniqueness S = T .
Finally, T (t + s) and T (t)T (s) both satisfy our differential equation and
coincide at t = 0. Hence they coincide for all t by uniqueness. 

Clearly A is uniquely determined by T (t) via A = Ṫ (0). Moreover, from

this we also easily get uniqueness for our original Cauchy problem. We will
in fact be slightly more general and consider the inhomogeneous problem
u̇ = Au + g, u(0) = u0 , (8.4)
where g ∈ C(I, X). A solution necessarily satisfies
d
T (−t)u(t) = −AT (−t)u(t) + T (−t)u̇(t) = T (−t)g(t)
dt
and integrating this equation (fundamental theorem of calculus) shows the
Duhamel formula
 Z t  Z t
u(t) = T (t) u0 + T (−s)g(s)ds = T (t)u0 + T (t − s)g(s)ds. (8.5)
0 0
8.2. Strongly continuous semigroups 239

Using Problem 8.1 it is straightforward to verify that this is indeed a solution

for any given g ∈ C(I, X).

Lemma 8.2. Let A ∈ L (X) and g ∈ C(I, X). Then (8.4) has a unique
solution given by (8.5).

Example 8.1. For example look at the discrete linear wave equation


q̈n (t) = k qn+1 (t) − 2qn (t) + qn−1 (t) , n ∈ Z.
Factorizing this equation according to


q̇n (t) = pn (t), ṗn (t) = k qn+1 (t) − 2qn (t) + qn−1 (t) ,
we can write this as a first order system
    
d qn 0 1 qn
=
dt pn k A0 0 pn
with the Jacobi operator A0 qn = qn+1 −2qn +qn−1 . Since A0 is a bounded op-
erator on X = `p (Z) we obtain a well-defined uniformly continuous operator
group in `p (Z) ⊕ `p (Z). 

Problem* 8.1 (Product rule). Suppose f ∈ C 1 (I, X) and T ∈ C 1 (I, L (X, Y )).
d
Show that T f ∈ C 1 (I, Y ) and dt T f = Ṫ f + T f˙.

Problem 8.2. Let A ∈ L (X). Show that in fact T ∈ C ∞ (R, L (X)).

8.2. Strongly continuous semigroups

In the previous section we have found a quite complete solution of the ab-
stract Cauchy problem (8.4) in the case when A is bounded. However, since
differential operators are typically unbounded, this assumption is too strong
for applications to partial differential equations. Since it is unclear what the
conditions on A should be, we will go the other way and impose conditions
on T . First of all, even rather simple equations like the heat equation are
only solvable for positive times and hence we will only assume that the solu-
tions give rise to a semigroup. Moreover, continuity in the operator topology
is too much to ask for (in fact, it is equivalent to boundedness of A — Prob-
lem 8.3) and hence we go for the next best option, namely strong continuity.
In this sense, our problem is still well-posed.
A strongly continuous operator semigroup (also C0 -semigoup) is a
family of operators T (t) ∈ L (X), t ≥ 0, such that
(i) T (t)g ∈ C([0, ∞), X) for every g ∈ X (strong continuity) and
(ii) T (0) = I, T (t+s) = T (t)T (s) for all t, s ≥ 0 (semigroup property).
240 8. Operator semigroups

If item (ii) holds for all t, s ∈ R it is called a strongly continuous operator

group.
We first note that kT (t)k is uniformly bounded on compact time inter-
vals.
Lemma 8.3. Let T (t) be a C0 -semigroup. Then there are constants M ≥ 1,
ω ≥ 0 such that
kT (t)k ≤ M eωt , t ≥ 0. (8.6)
In case of a C0 -group we have kT (t)k ≤ M eω|t| , t ∈ R.

Proof. Since kT (.)gk ∈ C[0, 1] for every g ∈ X we have supt∈[0,1] kT (t)gk ≤

Mg . Hence by the uniform boundedness principle supt∈[0,1] kT (t)k ≤ M for
some M ≥ 1. Setting ω = log(M ) the claim follows by induction using the
semigroup property. For the group case apply the semigroup case to both
T (t) and S(t) := T (−t). 

Inspired by the previous section we define the generator A of a strongly

continuous semigroup as the linear operator
1
(8.7)


Af := lim T (t)f − f ,
t↓0 t

where the domain D(A) is precisely the set of all f ∈ X for which the above
limit exists. By linearity of limits D(A) is a linear subspace of X (and A
is a linear operator) but at this point it is unclear whether it contains any
nontrivial elements. We will however postpone this issue and begin with
the observation that a C0 -semigroup is the solution of the abstract Cauchy
problem associated with its generator A:
Lemma 8.4. Let T (t) be a C0 -semigroup with generator A. If f ∈ D(A)
then T (t)f ∈ D(A) and AT (t)f = T (t)Af . Moreover, suppose g ∈ X with
u(t) := T (t)g ∈ D(A) for t > 0. Then u(t) ∈ C([0, ∞), X) ∩ C 1 ((0, ∞), X)
and u(t) is the unique solution of the abstract Cauchy problem
u̇(t) = Au(t), u(0) = g. (8.8)
This is, for example, the case if g ∈ D(A) in which case we even have
u(t) ∈ C 1 ([0, ∞), X).
Similarly, if T (t) is a C0 -group and g ∈ D(A), then u(t) := T (t)g ∈
C 1 (R, X) is the unique solution of (8.8) for all t ∈ R.

Proof. Let f ∈ D(A) and t > 0 (respectively t ∈ R for a group), then

1
1

lim u(t + ε) − u(t) = lim T (t) T (ε)f − f = T (t)Af.
ε↓0 ε ε↓0 ε
8.2. Strongly continuous semigroups 241

This shows the first part. To show that u(t) is differentiable it remains to
compute
1
1

lim u(t − ε) − u(t) = lim T (t − ε) T (ε)f − f
ε↓0 −ε ε↓0 ε


= lim T (t − ε) Af + o(1) = T (t)Af
ε↓0

since kT (t)k is bounded on compact t intervals. Hence u(t) ∈ C 1 ([0, ∞), X)

(respectively u(t) ∈ C 1 (R, X) for a group) solves (8.8). In the general case
f = T (t0 )g ∈ D(A) and u(t) = T (t)g = T (t − t0 )f solves our differential
equation for every t > t0 . Since t0 > 0 is arbitrary it follows that u(t) solves
(8.8) by the first part. To see that it is the only solution, let v(t) be a solution
corresponding to the initial condition v(0) = 0. For s < t we have
d 1

T (t − s)v(s) = lim T (t − s − ε)v(s + ε) − T (t − s)v(s)
ds ε→0 ε
1

= lim T (t − s − ε) v(s + ε) − v(s)
ε→0 ε
1

− lim T (t − s − ε) T (ε)v(s) − v(s)
ε→0 ε
=T (t − s)Av(s) − T (t − s)Av(s) = 0.
Whence, v(t) = T (t − t)v(t) = T (t − s)v(s) = T (t)v(0) = 0. 

Note that our proof in fact even shows a bit more: If g ∈ D(A) we have
u ∈ C 1 ([0, ∞), X) and hence not only u ∈ C([0, ∞), X) but also Au = u̇ ∈
C([0, ∞), X). Hence, if we regard D(A) as a normed space equipped with the
graph norm kf kA := kf k + kAf k, in which case we will write [D(A)], then
g ∈ D(A) implies u ∈ C([0, ∞), [D(A)]). Similarly, u(t) = T (t)g ∈ D(A) for
t > 0 implies u ∈ C((0, ∞), [D(A)]). Moreover, recall that [D(A)] will be a
Banach space if and only if A is a closed operator (cf. Problem 4.13) and the
latter fact will be established in Corollary 8.6 below.
Before turning to some examples, we establish a useful criterion for a
semigroup to be strongly continuous.
Lemma 8.5. A (semi)group of bounded operators is strongly continuous if
and only if lim supε↓0 kT (ε)gk < ∞ for every g ∈ X and limε↓0 T (ε)f = f
for f in a dense subset.

Proof. We first show that lim supε↓0 kT (ε)gk < ∞ for every g ∈ X implies
that T (t) is bounded in a small interval [0, δ]. Otherwise there would exist
a sequence εn ↓ 0 with kT (εn )k → ∞. Hence kT (εn )gk → ∞ for some g by
the uniform boundedness principle, a contradiction. Thus there exists some
M such that supt∈[0,δ] kT (t)k ≤ M . Setting ω = log(M
δ
)
we even obtain (8.6).
242 8. Operator semigroups

Moreover, boundedness of T (t) shows that limε↓0 T (ε)f = f for all f ∈ X

by a simple approximation argument (Lemma 4.33 (iv)).
In case of a group this also shows kT (−t)k ≤ kT (δ − t)kkT (−δ)k ≤
M kT (−δ)k for 0 ≤ t ≤ δ. Choosing M̃ = max(M, M kT (−δ)k) we conclude
kT (t)k ≤ M̃ exp(ω̃|t|).
Finally, right continuity is implied by the semigroup property: limε↓0 T (t+
ε)g = T (ε)T (t)g = T (t)g. Left continuity follows from kT (t − ε)g − T (t)gk =
kT (t − ε)(T (ε)g − g)k ≤ kT (t − ε)kkT (ε)g − gk. 

Example 8.2. Let X := C0 (R) be the continuous functions vanishing as

|x| → ∞. Then it is straightforward to check that

(T (t)f )(x) := f (x + t)

defines a group of continuous operators on X. Since shifting a function does

not alter its supremum we have kT (t)f k∞ = kf k∞ and hence kT (t)k = 1.
Moreover, strong continuity is immediate for uniformly continuous functions.
Since every function with compact support is uniformly continuous and since
such functions are dense, we get that T is strongly continuous. Moreover,
for f ∈ D(A) we have
f (t + ε) − f (t)
lim = (Af )(t)
ε→0 ε
uniformly. In particular, f ∈ C 1 (R) with f, f 0 ∈ C0 (R). Conversely, for
f ∈ C 1 (R) with f, f 0 ∈ C0 (R) we have
f (t + ε) − f (t) − εf 0 (t) 1 ε 0
Z
f (t + s) − f 0 (t) ds ≤ sup kT (s)f 0 − f 0 k∞


=
ε ε 0 0≤s≤ε

which converges to zero as ε ↓ 0 by strong continuity of T . Whence

d
A= , D(A) = {f ∈ C 1 (R) ∩ C0 (R)|f 0 ∈ C0 (R)}.
dx
It is not hard to see that T is not uniformly continuous or, equivalently, that
A is not bounded (cf. Problem 8.3).
Note that this group is not strongly continuous when considered
√ on X :=
Cb (R). Indeed for f (x) = cos(x 2 ) we can choose x =
n 2πn and tn =
√ q 1 √
2π( n + 4 − n) = 4 2n +O(n−3/2 ) such that kT (tn )f −f k∞ ≥ |f (xn +
1
pπ

tn ) − f (xn )| = 1. 
Next consider
Z t
u(t) = T (t)g, v(t) := u(s)ds, g ∈ X. (8.9)
0
8.2. Strongly continuous semigroups 243

Then v ∈ C 1 ([0, ∞), X) with v̇(t) = u(t) and (Problem 7.2)

1
 1 1

lim T (ε)v(t) − v(t) = lim − v(ε) + v(t + ε) − v(t) = −g + u(t).
ε↓0 ε ε↓0 ε ε
(8.10)
Consequently v(t) ∈ D(A) and Av(t) = −g + u(t) implying that u(t) solves
the following integral version of our abstract Cauchy problem
Z t
u(t) = g + A u(s)ds. (8.11)
0
Note that while in the case of a bounded generator both versions are equiva-
lent, this will not be the case in general. So while u(t) = T (t)g always solves
the integral version, it will only solve the differential version if u(t) ∈ D(A)
for t > 0 (which is clearly also necessary for the differential version to make
sense). In the latter case u(t) is sometimes called a strong solution (also
classical solution), while otherwise it is called a mild solution.
Two further consequences of these considerations are also worth while
noticing:
Corollary 8.6. Let T (t) be a C0 -semigroup with generator A. Then A is a
densely defined and closed operator.

Proof. Since v(t) ∈ D(A) and limt↓0 1t v(t) = g for arbitrary g, we see that
D(A) is dense. Moreover, if fn ∈ D(A) and fn → f , Afn → g then
Z t
T (t)fn − fn = T (s)Afn ds.
0
Taking n → ∞ and dividing by t we obtain

1 t
Z
1
T (t)f − f = T (s)g ds.
t t 0
Taking t ↓ 0 finally shows f ∈ D(A) and Af = g. 

Note that by the closed graph theorem we have D(A) = X if and only if
A is bounded. Moreover, since a C0 -semigroup provides the unique solution
of the abstract Cauchy problem for A, we obtain
Corollary 8.7. A C0 -semigroup is uniquely determined by its generator.

Proof. Suppose T and S have the same generator A. Then by uniqueness

for (8.8) we have T (t)g = S(t)g for all g ∈ D(A). Since D(A) is dense this
implies T (t) = S(t) as both operators are continuous. 

Finally, as in the uniformly continuous case, the inhomogeneous problem

can be solved by Duhamel’s formula. However, now it is not so clear when
this will actually be a solution.
244 8. Operator semigroups

Lemma 8.8. Let A be the generator of a C0 -semigroup and f ∈ C([0, ∞), X).
If the inhomogeneous problem
u̇ = Au + f, u(0) = g, (8.12)
has a solution it is necessarily given by Duhamel’s formula
Z t
u(t) = T (t)g + T (t − s)f (s)ds. (8.13)
0
Conversely, this formula gives a solution if either one of the following con-
ditions is satisfied:
• g ∈ D(A) and f ∈ C([0, ∞), [D(A)]).
• g ∈ D(A) and f ∈ C 1 ([0, ∞), X).

Proof. Set v(s) := T (t − s)u(s), 0 ≤ s ≤ t, then one shows as in the proof

of Lemma 8.4 that
v̇(s) = −AT (t − s)u(s) + T (t − s)u̇(s)
= −AT (t − s)u(s) + T (t − s)(Au(s) + f (s))
= T (t − s)f (s), 0 < s < t.
Hence the fundamental theorem of calculus (taking limits towards the bound-
ary points) gives (8.13).
For the converse observe that T (t)g is a solution of the homogenous
equation if g ∈ D(A). Hence it remains to investigate the integral, which we
will denote by u(t). Then, if the first condition holds,
1 ε
Z
1 1
(u(t + ε) − u(t)) = T (ε − s)f (t + s)ds + (T (ε) − I)u(t),
ε ε 0 ε
where the first term converges to f (t) and the second term converges to
Au(t) since u(t) ∈ D(A) by Problem 7.3.
In case of the second condition we note that
Z t
u(t) = T (s)f (t − s)ds
0

by a change of variables (Problem 7.1) and hence

1 t
Z
1
(u(t + ε) − u(t)) = T (s)(f (t + ε − s) − f (t − s))ds
ε ε 0
1 t+ε
Z
+ T (s)f (t − s)ds
ε t
Z t
→ T (s)f˙(t − s)ds + T (t)f (0)
ε→0 0
8.2. Strongly continuous semigroups 245

since f ∈ C 1 . Using this we see

Z ε
1 1
1
(T (ε) − I)u(t) = u(t + ε) − u(t) − T (ε − s)f (t + s)ds
ε ε ε 0
→ u̇(t) − f (t).
ε→0
This shows u(t) ∈ D(A) with Au(t) = u̇(t) − f (t). 

The function u(t) defined by (8.13) is called the mild solution of the
inhomogeneous problem. In general a mild solution is not a solution:
Example 8.3. Let T (t) be a strongly continuous group with an unbounded
generator A (e.g. the one form Example 8.2). Choose f0 ∈ X \ D(A) and set
g := 0, f (t) := T (t)f0 . Then f ∈ C(R, X) and the mild solution is given by
Z t Z t
u(t) = T (t) T (−s)f (s)ds = T (t) f0 ds = t T (t)f0 .
0 0
Since T (t) leaves D(A) invariant, we have u(t) 6∈ D(A) for all t ∈ R and
hence u(t) is not a solution. 
Problem* 8.3. Show that a uniformly continuous semigroup has a bounded
generator. (Hint: Write T (t) = V (t0 )−1 V (t0 )T (t) = . . . with V (t) :=
Rt 1
0 T (s)ds and conclude that it is C .)

Problem 8.4. Let T (t) be a C0 -semigroup. Show that if T (t0 ) has a bounded
inverse for one t0 > 0 then it extends to a strongly continuous group.
Problem 8.5. Define a semigroup on L1 (−1, 1) via
(
2f (s − t), 0 < s ≤ t,
(T (t)f )(s) =
f (s − t), else,
where we set f (s) = 0 for s < 0. Show that the estimate from Lemma 8.3
does not hold with M < 2.
Problem 8.6. Let A be the generator of a C0 -semigroup T (t). Show
Z t
T (t)f = f + tAf + (t − s)T (s)A2 f dt, f ∈ D(A2 ).
0

Problem 8.7. Let T (t) be a C0 -semigroup with generator A. Show that

k ) is dense. Moreover, show that T (t)g ∈ D(Ak ) and T (t)g ∈
T
k∈N D(A
C k ([0, ∞), X) if g ∈ D(Ak ) with
 k
d
T (t)g = Aj T (t)Ak−j u, t ≥ 0,
dt
R1
for any j = 0, . . . , k. (Hint: Set gm := m 0 φ(ms)T (s)g ds, where φ ∈
R1
Cc∞ (0, 1) with 0 φ(s)ds = 1.)
246 8. Operator semigroups

Problem 8.8. Let T (t) be a C0 -semigroup with generator A. Suppose T (t)g ∈

D(A) for all g ∈ X and t > 0. Show that T (t)g ∈ D(Ak ) for all k ∈ N, t > 0
and T ∈ C ∞ ((0, ∞), L (X)). (Hint: Start with the first claim and use the
closed graph theorem to conclude that Ak T (t) is a bounded operator. Then
use Problem 8.6.)
Problem 8.9 (Landau inequality). Let A be the generator of a C0 -semigroup
T (t) satisfying kT (t)k ≤ M . Derive the abstract Landau inequality
kAf k ≤ 2M kA2 f k1/2 kf k1/2 .
(Hint: Problem 8.6.)
Problem 8.10. Let A be the generator of a C0 -semigroup. Consider the
integral version of our inhomogeneous problem (8.12):
Z t Z t
u(t) = g + A u(s)ds + f (s)ds
0 0
for given g ∈ X, f ∈ C([0, 1), X). Show that this problem has a unique
Rt
solution u ∈ C([0, 1), X) such that 0 u(s)ds ∈ D(A) for t ≥ 0 which is
given by Duhamel’s formula (8.13). (Hint: Problem 7.4.)

8.3. Generator theorems

Of course in practice the abstract Cauchy problem, that is the operator A,
is given and the question is if A generates a corresponding C0 -semigroup.
Corollary 8.6 already gives us some necessary conditions but this alone is
not enough.
It turns out that it is crucial to understand the resolvent of A. As in the
case of bounded operators (cf. Section 6.1) we define the resolvent set via
ρ(A) := {z ∈ C|A − z is bijective with a bounded inverse} (8.14)
and call
RA (z) := (A − z)−1 , z ∈ ρ(A) (8.15)
the resolvent of A. The complement σ(A) = C \ ρ(A) is called the spec-
trum of A. As in the case of Banach algebras it follows that the resolvent
is analytic and that the resolvent set is open (Problem 8.12). However, the
spectrum will no longer be bounded in general and both σ(A) = ∅ as well
as σ(A) = C are possible (cf. the example in Problem 8.14). Note that if
A is closed, then bijectivity implies boundedness of the inverse (see Corol-
lary 4.9). Moreover, by Lemma 4.8 an operator with nonempty resolvent set
must be closed.
R∞
Using an operator-valued version of the elementary integral 0 et(a−z) dt =
−(a − z)−1 (for Re(a − z) < 0) we can make the connection between the
resolvent and the semigroup.
8.3. Generator theorems 247

Lemma 8.9. Let T be a C0 -semigroup with generator A satisfying (8.6).

Then {z|Re(z) > ω} ⊆ ρ(A) and
Z ∞
RA (z) = − e−zt T (t)dt, Re(z) > ω, (8.16)
0
were the right-hand side is defined as
Z ∞  Z s
−zt
e T (t)dt f := lim e−zt T (t)f dt. (8.17)
0 s→∞ 0

Moreover,
M
kRA (z)k ≤ , Re(z) > ω. (8.18)
Re(z) − ω
Rs
Proof. Let us abbreviate Rs (z)f := − 0 e−zt T (t)f dt. Then, by virtue
of (8.6), ke−zt T (t)f k ≤ M e(ω−Re(z))t kf k shows that Rs (z) is a bounded
operator satisfying kRs (z)k ≤ M (Re(z) − ω)−1 . Moreover, this estimates
also shows that the limit R(z) := lims→∞ Rs (z) exists (and still satisfies
kR(z)k ≤ M (Re(z) − ω)−1 ). Next note that S(t) = e−zt T (t) is a semigroup
with generator A − z (Problem 8.11) and hence for f ∈ D(A) we have
Z s Z s
Rs (z)(A − z)f = − S(t)(A − z)f dt = − Ṡ(t)f dt = f − S(s)f.
0 0
In particular, taking the limit s → ∞, we obtain R(z)(A − z)f = f for
f ∈ D(A). Similarly, still for f ∈ D(A), by Problem 7.3
Z s Z s
(A − z)Rs (z)f = − (A − z)S(t)f dt = − Ṡ(t)f dt = f − S(s)f
0 0
and taking limits, using closedness of A, implies (A − z)R(z)f = f for
f ∈ D(A). Finally, if f ∈ X choose fn ∈ D(A) with fn → f . Then
R(z)fn → R(z)f and (A − z)R(z)fn = fn → f proving R(z)f ∈ D(A) and
(A − z)R(z)f = f for f ∈ X. 
Corollary 8.10. Let T be a C0 -semigroup with generator A satisfying (8.6).
Then
(−1)n+1 ∞ n −zt
Z
RA (z) n+1
= t e T (t)dt, Re(z) > ω, (8.19)
n! 0
and
M
kRA (z)n k ≤ , Re(z) > ω, n ∈ N. (8.20)
(Re(z) − ω)n
R∞
Proof. Abbreviate Rn (z) := 0 tn e−zt T (t)dt and note that
Z ∞
Rn (z + ε) − Rn (z)
= −Rn+1 (z) + ε tn+2 φ(εt)e−zt T (t)dt
ε 0
248 8. Operator semigroups

where |φ(ε)| ≤ 12 e|ε| from which we see dz d

Rn (z) = −Rn+1 (z) and hence
dn dn
dz n RA (z) = − dznn R0 (z) = (−1)
n+1 Rn (z). Now the first claim follows using
RA (z)n+1 = n!1 d
R
dz n A (z) (Problem 8.13). Estimating the integral using (8.6)
establishes the second claim. 

Given these preparations we can now try to answer the question when
A generates a semigroup. In fact, we will be constructive and obtain the
corresponding semigroup by approximation. To this end we introduce the
Yosida approximation
An := −nARA (ω + n) = −n − n(ω + n)RA (ω + n) ∈ L (X). (8.21)
Of course this is motivated by the fact that this is a valid approximation for
numbers limn→∞ a−ω−n−n
= 1. That we also get a valid approximation for
operators is the content of the next lemma.

Lemma 8.11. Suppose A is a densely defined closed operator with (ω, ∞) ⊂

ρ(A) satisfying
M
kRA (ω + n)k ≤ . (8.22)
n
Then
lim −nRA (ω + n)f = f, f ∈ X, lim An f = Af, f ∈ D(A).
n→∞ n→∞
(8.23)

Proof. If f ∈ D(A) we have −nRA (ω + n)f = f − RA (ω + n)(A − ω)f

which shows −nRA (ω + n)f → f if f ∈ D(A). Since D(A) is dense and
knRA (ω + n)k ≤ M this even holds for all f ∈ X. Moreover, for f ∈ D(A)
we have An f = −nARA (ω + n)f = −nRA (ω + n)(Af ) → Af by the first
part. 

Moreover, An can also be used to approximate the corresponding semi-

group under suitable assumptions.

Theorem 8.12 (Feller–Miyadera–Phillips). A linear operator A is the gen-

erator of a C0 -semigroup satisfying (8.6) if and only if it is densely defined,
closed, (ω, ∞) ⊆ ρ(A), and
M
kRA (λ)n k ≤ , λ > ω, n ∈ N. (8.24)
(λ − ω)n

Proof. Necessity has already been established in Corollaries 8.6 and 8.10.
For the converse we use the semigroups
Tn (t) := exp(tAn )
8.3. Generator theorems 249

corresponding to the Yosida approximation (8.21). We note (using eA+B =

eA eB for commuting operators A, B)
∞
X (tn(ω + n))j
kTn (t)k ≤ e−tn kRA (ω + n)j k ≤ M e−tn et(ω+n) = M eωt .
j!
j=0

Moreover, since RA (ω + m) and RA (ω + n) commute by the first resolvent

identity (Problem 8.13), we conclude that the same is true for Am , An as well
as for Tm (t), Tn (t) (by the very definition as a power series). Consequently
Z 1
d
kTn (t)f − Tm (t)f k = Tn (st)Tm ((1 − s)t)f ds
0 ds

Z 1
≤t kTn (st)Tm ((1 − s)t)(An − Am )f kds
0
2 ωt
≤ tM e k(An − Am )f k.
Thus, from f ∈ D(A) we have a Cauchy sequence and can define a linear
operator by T (t)f := limn→∞ Tn (t)f . Since kT (t)f k = limn→∞ kTn (t)f k ≤
M eωt kf k, we see that T (t) is bounded and has a unique extension to all of
X. Moreover, T (0) = I and
kTn (t)Tn (s)f − T (t)T (s)f k ≤
M eωt kTn (s)f − T (s)f k + kTn (t)T (s)f − T (t)T (s)f k
implies T (t + s)f = limn→∞ Tn (t + s)f = limn→∞ Tn (t)Tn (s)f = T (t)T (s)f ,
that is, the semigroup property holds. Finally, by
kT (ε)f − f k ≤ kT (ε)f − Tn (ε)f k + kTn (ε)f − f k
≤ εM 2 eωε k(A − An )f k + kTn (ε)f − f k
we see limε↓0 T (ε)f = f for f ∈ D(A) and Lemma 8.5 shows that T is a
C0 -semigroup. It remains to show that A is its generator. To this end let
f ∈ D(A), then
Z t
T (t)f − f = lim Tn (t)f − f = lim Tn (s)An f ds
n→∞ n→∞ 0
Z t Z t 
= lim Tn (s)Af ds + Tn (s)(An − A)f ds
n→∞ 0 0
Z t
= T (s)Af ds
0

which shows limt↓0 1t (T (t)f − f ) = Af for f ∈ D(A). Finally, note that

the domain of the generator cannot be larger, since A − ω − 1 is bijective
and adding a vector to its domain would destroy injectivity. But then ω + 1
would not be in the resolvent set contradicting Lemma 8.9. 
250 8. Operator semigroups

Note that in combination with the following lemma this also answers the
question when A generates a C0 -group.

Lemma 8.13. An operator A generates a C0 -group if and only if both A

and −A generate C0 -semigroups.

Proof. Clearly, if A generates a C0 -group T (t), then S(t) := T (−t) is a C0 -

group with generator −A. Conversely, let T (t), S(t) be the C0 -semigroups
generated by A, −A, respectively. Then a short calculation shows
d
T (t)S(t)g = −T (t)AS(t)g + T (t)AS(t)g = 0, t ≥ 0.
dt
Consequently, T (t)S(t) = T (0)S(0) = I and similarly S(t)T (t) = I, that is,
S(t) = T (t)−1 . Hence it is straightforward to check that T extends to a
group via T (−t) := S(t), t ≥ 0. 

The following examples show that the spectral conditions are indeed cru-
cial. Moreover, they also show that an operator might give rise to a Cauchy
problem which is uniquely solvable for a dense set of initial conditions, with-
out generating a strongly continuous semigroup.
Example 8.4. Let
 
0 A0
A= , D(A) = X × D(A0 ).
0 0

Then u(t) = 10 tA1 0 ff01 = f0 +tA 0 f1

is the unique solution of the corre-




f1
sponding abstract Cauchy problem for given f ∈ D(A). Nevertheless, if A0
is unbounded, the corresponding semigroup is not strongly continuous.
Note that in this case we have σ(A) = {0} if A0 is bounded and σ(A) = C
else. In fact, since A is not injective we must have {0} ⊆ σ(A). For z 6= 0
the inverse of A − z is given by
1 1 z1 A0
 
(A − z)−1 = − , D((A − z)−1 ) = Ran(A − z) = X × D(A0 ),
z 0 1
which is bounded if and only if A is bounded. 
Example 8.5. Let X0 = C0 (R) and m(x) = ix. Then we can regard m as
a multiplication operator on X0 when defined maximally, that is, f 7→ m f
with D(m) = {f ∈ X0 |m f ∈ X0 }. Note that since Cc (R) ⊆ D(m) we see
that m is densely defined. Moreover, it is easy to check that m is closed.
Now consider X = X0 ⊕ X0 with kf k = max(kf0 k, kf1 k) and note that
 
m m
A= , D(A) = D(m) ⊕ D(m),
0 m
8.3. Generator theorems 251

is closed. Moreover, for z 6∈ iR the resolvent is given by the multiplication

operator
m
 
1 1 − m−z
RA (z) = − .
m−z 0 1
For λ > 0 we compute
 
1 |x| 3
kRA (λ)f k ≤ sup + sup kf k = kf k
x∈R |ix − λ| x∈R |ix − λ|2 2λ
and hence A satisfies (8.24) with M = 23 , ω = 0 and n = 1. However, by
 
 infn (n) 
kRA (λ + in)k ≥ kRA (λ + in)(0, fn )k ≥  = n,
(λ − in + in)2  λ2
where fn is chosen such that fn (n) = 1 and kfn k∞ = 1, it does not satisfy
(8.20). Hence A does not generate a C0 -semigroup. Indeed, the solution of
the corresponding Cauchy problem is
 
tm 1 tm
T (t) = e , D(T ) = X0 ⊕ D(m),
0 1
which is unbounded. 
When it comes to applying this theorem, the main difficulty will be es-
tablishing the resolvent estimate (8.24). Moreover, while it might be already
difficult to estimate the resolvent, it will in general be even more challeng-
ing to get estimates on its powers. In this connection note that the trivial
estimate kRA (z)n k ≤ kRA (z)kn will do the job if and only if M = 1. Hence
we finally look at the special case of contraction semigroups satisfying
kT (t)k ≤ 1. (8.25)
By a simple transform the case M = 1 in Lemma 8.3 can always be reduced
to this case (Problem 8.11). Moreover, as already anticipated, in the case
M = 1 the estimate (8.18) immediately implies the general estimate (8.20)
and it suffices to establish (8.24) for n = 1:
Corollary 8.14 (Hille–Yosida). A linear operator A is the generator of a
contraction semigroup if and only if it is densely defined, closed, (0, ∞) ⊆
ρ(A), and
1
kRA (λ)k ≤ , λ > 0. (8.26)
λ
Example 8.6. If A is the generator of a contraction, then clearly all eigen-
values z must satisfy Re(z) ≤ 0. Moreover, for
 
0 1
A=
0 0
252 8. Operator semigroups

we have    
1 1 1/z 1 t
RA (z) = − , T (t) = ,
z 0 1 0 1
which shows that the bound on the resolvent is crucial. 
However, for a given operator even the simple estimate (8.26) might be
difficult to establish directly. Hence we outline another criterion.
Example 8.7. Let X be a Hilbert space and observe that for a contraction
semigroup the expression kT (t)f k must be nonincreasing. Consequently, for
f ∈ D(A) we must have
d 
2


kT (t)f k  = 2Re hf, Af i ≤ 0.
dt t=0
Operators satisfying Re(hf, Af i) ≤ 0 are called dissipative and this clearly
suggests to replace the resolvent estimate by dissipativity. 
To formulate this condition for Banach spaces, we first introduce the
duality set
J (x) := {x0 ∈ X ∗ |x0 (x) = kxk2 = kx0 k2 } (8.27)
of a given vector x ∈ X. In other words, the elements from J (x) are those
linear functionals which attain their norm at x and are normalized to have
the same norm as x. As a consequence of the Hahn–Banach theorem (Corol-
lary 4.15) note that J (x) is nonempty. Moreover, it is also easy to see that
J (x) is convex and weak-∗ closed.
Example 8.8. Let X be a Hilbert space and identify X with X ∗ via x 7→
hx, .i as usual. Then J (x) = {x}. Indeed since we have equality hx0 , xi =
kx0 kkxk in the Cauchy–Schwarz inequality, we must have x0 = αx for some
α ∈ C with |α| = 1 and α∗ kxk2 = hx0 , xi = kxk2 shows α = 1. 
Example 8.9. If X ∗ is strictly convex (cf. Problem 1.13), then the duality
set contains only one point. In fact, suppose x0 , y 0 ∈ J (x), then z 0 = 21 (x0 +
y 0 ) ∈ J (x) and kxk 0 kxk
2 kx + y k = z (x) = 2 (kx k + ky k) implying x = y
0 0 0 0 0 0

by strict convexity. Note that the converse is also true: If x , y ∈ J (x) for
0 0

some x ∈ B1X (0), then x0 (x) + y 0 (x) = 2 implies kx0 + y 0 k = 2 contradicting

strict convexity.
This applies for example to X := `p (N) if 1 < p < ∞ (cf. Problem 1.13)
in which case X ∗ ∼= `q (N) with q = p−1p
. In fact, for a ∈ X we have
J (a) = {a0 } with a0j = kak2−p
p sign(a∗j )|aj |p−1 . 
Example 8.10. Let X be a measurable space with a σ-finite measure µ.
The previous example can be generalized to Lp (X, dµ) if 1 < p < ∞ (which
are strictly, in fact even uniformly, convex by Theorem 3.11 from [47]). In
this case we have Lp (X, dµ)∗ ∼ = Lq (X, dµ) and for f ∈ Lp (X, dµ) we have
2−p
J (f ) = {g} with g = kf kp sign(f ∗ )|f |p−1 . 
8.3. Generator theorems 253

Example 8.11. Let X := C[0, 1] and choose x ∈ X. If t0 is chosen such

that |x(t0 )| = kxk∞ , then the functional y 7→ x0 (y) := x(t0 )∗ y(t0 ) satisfies
x0 ∈ J (x). Clearly J (x) will contain more than one element in general.
Note that for X = Cb (R) the situation is more complicated since the
supremum might not be attained. However, we can choose a sequence tn ∈ R
such that x(tn ) → x0 with |x0 | = kxk∞ and set x0 (y) = x∗0 L(y(tn )), where
L is the Banach limit from Problem 4.25. 
Now a given operator D(A) ⊆ X → X is called dissipative if
Re x0 (Ax) ≤ 0 for one x0 ∈ J (x) and all x ∈ D(A). (8.28)

Lemma 8.15. Let x, y ∈ X. Then kxk ≤ kx − αyk for all α > 0 if and only
if there is an x0 ∈ J (x) such that Re(x0 (y)) ≤ 0.

Proof. Without loss of generality we can assume x 6= 0. If Re(x0 (y)) ≤ 0

for some x0 ∈ J (x), then for α > 0 we have
kxk2 = x0 (x) ≤ Re x0 (x − αy) ≤ kx0 kkx − αyk

implying kxk ≤ kx − αyk.

Conversely, if kxk ≤ kx − αyk for all α > 0, let x0α ∈ J (x − αy) and set
yα0 = kx0α k−1 x0α . Then
kxk ≤ kx − αyk = yα0 (x − αy) = Re yα0 (x) − αRe yα0 (y)

≤ kxk − αRe yα0 (y) .

Now by the Banach–Alaoglu theorem we can choose a subsequence y1/n 0

j
→
y0 in the weak-∗ sense. (Note that the use of the Banach–Alaoglu theorem
0

could be avoided by restricting yα0 to the two dimensional subspace spanned

by x , y, passing to the limit in this subspace and then

extending the limit
to X using Hahn–Banach.) Consequently
Re y0 (y) ≤ 0 and y0 (x) = kxk.
∗ 0 0

Whence x00 = kxky00 ∈ J (x) and Re x00 (y) ≤ 0. 

As a straightforward consequence we obtain:

Corollary 8.16. A linear operator is dissipative if and only if

k(A − λ)f k ≥ λkf k, λ > 0, f ∈ D(A). (8.29)

In particular, for a dissipative operator A − λ is injective for λ > 0 and

(A − λ)−1 is bounded with k(A − λ)−1 k ≤ λ−1 . However, this does not imply
that λ is in the resolvent set of A since D((A − λ)−1 ) = Ran(A − λ) might
not be all of X.
Now we are ready to show
254 8. Operator semigroups

Theorem 8.17 (Lumer–Phillips). A linear operator A is the generator of

a contraction semigroup if and only if it is densely defined, dissipative, and
A − λ0 is surjective for one λ0 > 0. Moreover, in this case (8.28) holds for
all x0 ∈ J (x).

Proof. Let A generate a contraction semigroup T (t) and let x ∈ D(A),

x0 ∈ J (x). Then
Re x0 (T (t)x − x) ≤ |x0 (T (t)x)| − kxk2 ≤ kx0 kkxk − kxk2 = 0

and dividing by t and letting t ↓ 0 shows Re x0 (Ax) ≤ 0. Hence A is

dissipative and by Corollary 8.14 (0, ∞) ⊆ ρ(A), that is A − λ is bijective

for λ > 0.
Conversely, by Corollary 8.16 A − λ has a bounded inverse satisfying
k(A − λ)−1 k ≤ λ−1 for all λ > 0. In particular, for λ0 the inverse is defined
on all of X and hence closed. Thus A is also closed and λ0 ∈ ρ(A). Moreover,
from kRA (λ0 )k ≤ λ−1
0 (cf. Problem 8.12) we even get (0, 2λ0 ) ⊆ ρ(A) and
iterating this argument shows (0, ∞) ⊆ ρ(A) as well as kRA (λ)k ≤ λ−1 ,
λ > 0. Hence the requirements from Corollary 8.14 are satisfied. 

Note that generators of contraction semigroups are maximal dissipative

in the sense that they do not have any dissipative extensions. In fact, if we
extend A to a larger domain we must destroy injectivity of A − λ and thus
the extension cannot be dissipative.
Example 8.12. Let X := C0 [0, 1] = {f ∈ C[0, 1]|f (0) = f (1) = 0} and
consider the one-dimensional heat equation
∂ ∂2
u(t, x) = u(t, x)
∂t ∂x2
on a finite interval x ∈ [0, 1] with Dirichlet boundary conditions u(0) =
u(1) = 0 and the initial condition u(0, x) = u0 (x). The corresponding
operator is
Af = f 00 , D(A) = {f ∈ C0 [0, 1]|f ∈ C 2 [0, 1]} ⊂ X.
Note that D(A) is dense. For ` ∈ J (f ) we can choose `(g) = f (x0 )∗ g(x0 )
where x0 is chosen such that |f (x0 )| = kf k∞ . Then Re(f (x0 )∗ f (x)) has a
global maximum at x = x0 and if f ∈ D(A) we must have Re(f (x0 )∗ f 00 (x0 )) ≤
0 provided this maximum is in the interior of (0, 1). If x0 is at the boundary
this holds trivially and consequently A is dissipative. That A−λ is surjective
follows using the Green’s function as in Section 3.3: For g ∈ X the function
Z 1
f (x) := (RA (λ)g)(x) = G(λ, x, y)g(y)dy,
0
8.3. Generator theorems 255

where
( √ √
−1 sinh( λ(1 − x)) sinh( λy), y ≤ x,
G(λ, x, y) := √ √ √
λ sinh(λ) sinh( λ(1 − y)) sinh( λx), x ≤ y,
is in D(A) and satisfies (A − λ)f = g. Note that alternatively one could
compute the norm of the resolvent
 
1 1
kRA (λ)k = 1− √
λ cosh( λ/2)
(equality is attained for constant functions; while these are not in X, you can
approximate them by choosing functions which are constant on [ε, 1−ε]). 
Example 8.13. Another neat example is the following linear delay differ-
ential equation
Z t
u̇(t) = u(s)dν(s), t > 0, u(s) = g(s), −1 ≤ s ≤ 0,
t−1
where ν is a complex measure. To this end we introduce the following oper-
ator
Z 0
0 1 0
Af := f , D(A) := {f ∈ C [−1, 0]|f (0) = f (s)dν(s)} ⊂ C[0, 1].
−1
Suppose that we can show that it generates a semigroup T on X = C[0, 1] and
set u(t) := (T (t)f )(0) for f ∈ D(A). Then, since T leaves D(A) invariant,
the function r 7→ (T (t + r)f )(s − r) is differentiable with
d
(T (t + r)f )(s − r) = (T (t + r)Af )(s − r) − (T (t + r)f 0 )(s − r) = 0
dr
and we conclude (T (t + r)f )(s − r) = (T (t)f )(s) for −1 + r ≤ s ≤ 0. In
particular, for r = s we obtain u(t + s) = (T (t)f )(s). Hence we obtain
Z 0
d
u̇(t) = (T (t)f )(0) = (AT (t)f )(0) = (T (t)f )(s)dν(s)
dt −1
Z 0
= u(t + s)dν(s)
−1
and u solves our delay differential equation. Now if g ∈ C[0, 1] is given we
can approximate it by a sequence fn ∈ D(A). Then un (t) := (Tn (t)fn )(0)
will converge uniformly on compact sets to u(t) := (T (t)g)(0) and taking the
limit in the differential equation shows that u is differentiable and satisfies
the differential equation.
Hence it remains to show that A generates a semigroup. First of all
we claim that à := A − kνk is dissipative, where kνk is the total vari-
ation of ν. As in the previous example, for ` ∈ J (f ) we can choose
`(g) = f (x0 )∗ g(x0 ) where x0 is chosen such that |f (x0 )| = kf k∞ . Then
256 8. Operator semigroups

Re(f (x0 )∗ f (x)) has a global maximum at x = x0 and if f ∈ D(A) we must

have Re(f (x0 )∗ f 0 (x0 )) = 0 provided x0 is in the interior. If x0 = −1 we still
must have Re(f (x0 )∗ f 0 (x0 )) ≤ 0. In both cases Re(`(Ãf )) ≤ −kνk|f (x0 )|2 ≤
0. If x0 = 0 we compute
 Z 0 
∗
Re(`(Ãf )) = Re f (0) f (s)dν(s) − kνk|f (0)|2 ≤ 0
−1

since |f (s)| ≤ |f (0)|. Thus à is dissipative. Moreover, it is straightforward

to verify that the differential equation (Ã − λ)f = g has a unique solution
R0
f ∈ D(A) for λ > 0 since | −1 e(λ+kνk)s dν(s)| ≤ kνk. 
Finally, we note that the condition that A − λ0 is surjective can be
weakened to the condition that Ran(A − λ0 ) is dense. To this end we need:

Lemma 8.18. Suppose A is a densely defined dissipative operator. Then A

is closable and the closure A is again dissipative.

Proof. Recall that A is closable if and only if for every xn ∈ D(A) with
xn → 0 and Axn → y we have y = 0. So let xn be such a sequence and chose
another sequence yn ∈ D(A) such that yn → y (which is possible since D(A)
is assumed dense). Then by dissipativity (specifically Corollary 8.16)
k(A − λ)(λxn + ym )k ≥ λkλxn + ym k, λ>0
and letting n → ∞ and dividing by λ shows
ky + (λ−1 A − 1)ym k ≥ kym k.
Finally λ → ∞ implies ky − ym k ≥ kym k and m → ∞ yields 0 ≥ kyk, that
is, y = 0 and A is closable. To see that A is dissipative choose x ∈ D(A) and
xn ∈ D(A) with xn → x and Axn → Ax. Then (again using Corollary 8.16)
taking the limit in k(A − λ)xn k ≥ λkxn k shows k(A − λ)xk ≥ λkxk as
required. 

Consequently:

Corollary 8.19. Suppose the linear operator A is densely defined, dissipa-

tive, and Ran(A − λ0 ) is dense for one λ0 > 0. Then A is closable and A is
the generator of a contraction semigroup.

Proof. By the previous lemma A is closable with A again dissipative. In

particular, A is injective and by Lemma 4.8 we have (A−λ0 )−1 = (A − λ0 )−1 .
Since (A−λ0 )−1 is bounded its closure is defined on the closure of its domain,
that is, Ran(A − λ0 ) = Ran(A − λ0 ) = X. The rest follows from the Lumer–
Phillips theorem. 
8.3. Generator theorems 257

Problem* 8.11. Let T (t) be a C0 -semigroup and α > 0, λ ∈ C. Show that

S(t) := eλt T (αt) is a C0 -semigroup with generator B = αA + λ, D(B) =
D(A).

Problem* 8.12. Let A be a closed operator. Show that if z0 ∈ ρ(A), then

∞
X
RA (z) = (z − z0 )n RA (z0 )n+1 , |z − z0 | < kRA (z0 )k−1 .
n=0

In particular, the resolvent is analytic and

1
k(A − z)−1 k ≥ .
dist(z, σ(A))

Problem* 8.13. Let A be a closed operator. Show the first resolvent

identity

RA (z0 ) − RA (z1 ) = (z0 − z1 )RA (z0 )RA (z1 )

= (z0 − z1 )RA (z1 )RA (z0 ),

for z0 , z1 ∈ ρ(A). Moreover, conclude

dn d
RA (z) = n!RA (z)n+1 , RA (z)n = nRA (z)n+1 .
dz n dz
d
Problem* 8.14. Consider X = C[0, 1] and A = dx with D(A) = C 1 [0, 1].
d
Compute σ(A). Do the same for A0 = dx with D(A) = {x ∈ C 1 [0, 1]|x(0) =
0}.

Problem 8.15. Suppose z0 ∈ ρ(A) (in particular A is closed; also note that
0 ∈ σ(RA (z0 )) if and only if A is unbounded). Show that for z 6= 0 we have
σ(A) = z0 + (σ(RA (z0 )) \ {0})−1 and RRA (z0 ) (z) = − z1 − z12 RA (z0 + z1 ) for
z ∈ ρ(RA (z0 )) \ {0}. Moreover, Ker(RA (z0 ) − z)n = Ker(A − z0 − z1 )n and
Ran(RA (z0 ) − z)n = Ran(A − z0 − z1 )n for every n ∈ N0 .

Problem 8.16. Show that RA (z) ∈ C (X) for one z ∈ ρ(A) if and only
this holds for all z ∈ ρ(A). Moreover, in this case the spectrum of A consists
only of discrete eigenvalues with finite (geometric and algebraic) multiplicity.
(Hint: Use the previous problem to reduce it to Theorem 6.14.)

Problem 8.17. Consider the heat equation (Example 8.12) on [0, 1] with
Neumann boundary conditions u0 (0) = u0 (1) = 0.

Problem 8.18. Consider the heat equation (Example 8.12) on Cb (R) and
C0 (R).
258 8. Operator semigroups

8.4. Semilinear equations

Linear problems are often only a first approximation and adding a nonlinear
perturbation leads to the following semilinear problem
u̇ = Au + F (u), u(0) = g, (8.30)
where A is supposed to generate a semigroup T (t) and F ∈ C(X, X) such
that we can recast this problem as
Z t
u(t) = T (t)g + T (t − s)F (u(s))ds. (8.31)
0

In fact, if we have a solution u ∈ C([0, t+ ), [D(A)]) ∩ C 1 ([0, t+ ), X) of (8.30)

then Duhamel’s formula shows that (8.31) holds. In the other direction you
need a stronger assumption on F , by Lemma 8.8 F ∈ C([D(A)], [D(A)]) will
do. However, it will be more convenient to work with (8.31) and we will call
a solution a mild solution of (8.30). In fact, (8.31) is of fixed point type and
hence begs us to apply the contraction principle. As always with nonlinear
equations, we expect the solution to be only defined on a finite time interval
[0, t+ ) in general.
Theorem 8.20. Suppose F is Lipschitz continuous on bounded sets. Then
for every g ∈ X there is a t0 = t0 (kgk) > 0, such that there is a unique
mild solution u ∈ C([0, t0 ], X). Moreover, the solution map g 7→ u(t) will be
Lipschitz continuous from every ball kgk ≤ ρ to C([0, t0 (ρ)], X).

Proof. We will consider 0 ≤ t ≤ 1 and set M := sup0≤t≤1 kT (t)k. Let

r := 1 + M kgk and consider the closed ball B̄r (0) ⊂ X. Let L = L(r) be the
Lipschitz constant of F on B̄r (0). Set
Z t
K(u)(t) := T (t)g + T (t − s)F (u(s))ds
0

and note that

Z t


kK(u)(t)k ≤ M kgk + M kF (0)k + Lku(s)k ds
0
≤ M kgk + M kF (0)kt + M Lt sup ku(s)k
0≤s≤t

and
Z t

kK(u)(t) − K(v)(t)k ≤ M L ku(s) − v(s)k ds ≤ M Lt sup ku(s) − v(s)k
0 0≤s≤t

Hence if we choose t0 ≤ 1 such that

M (kF (0)k + Lr)t0 < 1
8.4. Semilinear equations 259

then θ := M Lt0 < 1 and K will be a contraction on B̄r (0) ⊂ C([0, t0 ], X).
In particular, for two solutions uj corresponding to gj with kgj k ≤ kgk we
will have ku1 − u2 k∞ ≤ 1−θ1
kg1 − g2 k by (7.44).
This establishes the theorem except for the fact that it only shows unique-
ness for solutions which stay within B̄r (0). However, since K maps from
B̄r (0) to its interior Br (0), a potential different solution starting at g ∈ Br (0)
would need to branch off at the boundary, which is impossible since our so-
lution does not reach the boundary. 

Corollary 8.21. If F is globally Lipschitz, then solutions are global.

Proof. In this case we can consider K on all of C([0, t0 ], X) and set M :=

sup0≤t≤t0 kT (t)k. By induction we get for the iterates
(M Lt)n
kK n (u)(t) − K n (v)(t)k ≤ sup ku(s) − v(s)k
n! 0≤s≤t

and Weissinger’s fixed point theorem (Theorem 7.28) gives a solution on

C([0, t0 ], X). Since t0 > 0 is arbitrary, the claim follows. 

If solutions are not global, there is still a unique maximal solution: Fix
g ∈ X and let uj be two solutions on [0, tj ) with 0 < t1 < t2 . By the
uniqueness part of our theorem, we will have u1 (t) = u2 (t) for 0 ≤ t < τ
for some τ > 0. Suppose τ < t1 and τ is chosen maximal. Let r :=
max0≤t≤τ ku1 (t)k and 0 < ε < min(τ, t0 (r)/2) with t0 (r) from our theorem.
Then there is a solution v starting with initial condition u1 (τ − ε) which is
defined on [0, 2ε]. Moreover, again by the uniqueness part of our theorem
u1 (t) = v(t − (τ − ε)) = u2 (t) for τ − ε ≤ t ≤ τ + ε) contradiction our
assumption that τ is maximal. Hence taking the union (with respect to
their domain) over all mild solutions starting at g, we get a unique solution
defined on a maximal domain [0, t+ (g)). Note that if t+ (g) < ∞, then ku(t)k
must blow up as t → t+ (g):

Lemma 8.22. Let t+ (g) be the maximal time of existence for the mild so-
lution starting at g. If t+ (g) < ∞, then lim inf t→t+ (g) ku(t)k = ∞.

Proof. Assume that ρ := sup0≤t≤t+ (g) ku(t)k < ∞. As above, choose 0 <
ε < min(t0 , t+ (ρ)) with t+ (ρ) from our theorem. Then the solution v starting
with initial condition u(t+ (g) − ε) extends u to the interval [0, t+ (g) + ε),
contradicting maximality. 

So the key to proving global existence of solutions is an a priori bound

on the norm of the solution. Typically such a bound will come from a
conservation law.
260 8. Operator semigroups

Example 8.14. Let X = C[0, 1] and consider the one-dimensional reaction-

diffusion equation
∂ ∂2
u(t, x) = u(t, x) + F (u(t, x))
∂t ∂x2
on a finite interval x ∈ [0, 1] with the boundary conditions u(0) = u(1) = 0
and the initial condition u(0, x) = u0 (x). From Example 8.12 we know
that the corresponding linear operator generates a C0 -semigroup. Hence if
F : R → R is Lipschitz, then our theorem applies and we get existence of
local mild solutions. If we even have that F ∈ C 2 (R, R) with F (0) = 0,
then F : [D(A)] → [D(A)] and the mild solutions will in fact be strong
solutions. 
Let me close with a few remarks: First of all, it is straightforward to
extend these results to the situation where F depends on t or to the case
where T is a group. Details are left to the reader. Moreover, if A is bounded,
then it is Lipschitz continuous and could be absorbed in F . In fact, in this
case our theorem just gives the Picard–Lindelöf theorem for ordinary differ-
ential equations in Banach spaces (in particular, in this case the differential
equation (8.30) and the integral equation (8.31) are equivalent).
Problem 8.19. Consider the reaction-diffusion equation (Example 8.14) on
Cb (R) and on C0 (R). (Hint: Problem 8.18.)
Problem 8.20. Show that solutions are global if kF (x)k ≤ C(1 + kxk) for
some constant C. (Hint: Use Gronwall’s inequality to bound ku(t)k.)
 Chapter 9

The nonlinear
Schrödinger equation

The purpose of this chapter is to investigate a prototypical example, the

initial value problem for the nonlinear Schrödinger equation (NLS)
iut + ∆u = ±|u|α−1 u, u(0) = g. (9.1)
The two cases − and + are known as focusing and defocusing, respectively.
Of particular importance in applications are the cubic (α = 3) and quintic
(α = 5) case. Note that if u is a solution, then so will be v(t, x) = u(−t, x)∗
and hence it suffices to look at positive times only.

n
9.1. Local well-posedness in H r for r > 2

Equation (9.1) is a semilinear equation of the type considered in Section 8.4

and hence we need to look at the linear Schrödinger equation
iut + ∆u = 0, u(0) = g (9.2)
first. We recall that the solution for g ∈ H 2 (Rn ) can be obtained using the
Fourier transform and is given by
2
u(t) = TS (t)g, TS (t) = F −1 e−i|p| t F. (9.3)
Note that TS (t) : L2 (Rn ) → L2 (Rn ) is a unitary operator (since |e −i|p|2 t |=
1):
ku(t)k2 = kgk2 . (9.4)
In fact, we even have have that TS (t) : H r (Rn ) → H r (Rn ) is unitary.
Theorem 9.1. The family TS (t) is a C0 -group in H r (Rn ) whose generator
is i∆, D(i∆) = H r+2 (Rn ).

261
262 9. The nonlinear Schrödinger equation

Note that we have

(T (t)g)∗ = T (−t)g ∗ , g ∈ L2 (Rn ) (9.5)
and
∂j T (t)g = T (t)∂j g, 1 ≤ j ≤ n, g ∈ H 1 (Rn ). (9.6)
Next we turn to the nonlinear Schrödinger equation. If we assume that
u, |u|α−1 u ∈ C([0, T ], L2 (Rn )) we can use Duhamel’s formula to rewrite the
nonlinear Schrödinger equation as
Z t
u(t) = TS (t)g ∓ i TS (t − s)|u(s)|α−1 u(s)ds (9.7)
0
just as we did in Section 8.4. In order to apply our theory, we need that the
nonlinearity F (u) = ∓i|u|α−1 u is Lipschitz on X. Clearly for X = L2 (Rn )
this will not be the case, as the image of a square integrable function will
not be square integrable. However, the key observation is that for r > n2 the
space H r (Rn ) is a Banach algebra (Lemma 8.33 from [47]) and hence, if we
assume our nonlinearity to be of the form F (u) = ∓i|u|α−1 u with α − 1 = 2k
where k ∈ N, then F : H r (Rn ) → H r (Rn ) is Lipschitz on bounded sets since
F (u) − F (v) = uk+1 Qk−1 (u∗ , v ∗ )(u − v)∗ + (v ∗ )k Qk (u, v)(u − v), (9.8)
where Qk (x, y) = ∓i kj=0 xk−j y j . Another algebra which is natural in this
P
context is the Wiener algebra.
A(Rn ) := {f |fˆ ∈ L1 (Rn )}, kf kA := kfˆk1 . (9.9)
Just as with H r (Rn ), the Schrödinger group TS leaves A(Rn ) invariant and
preserves its norm. Note that we have H r (Rn ) ⊂ A(Rn ) for r > n2 since
(1 + |p|2 )−r ∈ L2 (Rn ) for such r. The embedding being continuous, kf kA ≤
k(1 + |.|2 )−r k2 kf kH r .
Hence Theorem 8.20 applies and we get:
Theorem 9.2. Let α = 2k + 1 be an odd integer and X = H r (Rn ) for r > n2
or X = A(Rn ). Then for every g ∈ X there is a t0 = t0 (kgk) > 0, such
that there is a unique solution u ∈ C([−t0 , t0 ], X) of (9.7). Moreover, the
solution map g 7→ u(t) will be Lipschitz continuous from every ball kgk ≤ ρ
to C([−t0 (ρ), t0 (ρ)], X).

Note that the mild solution will be a strong solution for g ∈ H r+2 since
F : H r+2 → H r+2 is continuous. Moreover, for each initial condition there
is a maximal solution and Lemma 8.22 implies:
Lemma 9.3. This solution exists on a maximal time interval (t− (g), t+ (g))
and if |t± (g)| < ∞ we must have lim inf t→t± (g) ku(t)k = ∞.

An interesting observation is that the maximal existence time does not

depend on r. This is known as persistence of regularity:
9.1. Local well-posedness in H r for r > n
2 263

Lemma 9.4. Let g ∈ H r (Rn ) with r > n2 or g ∈ A(Rn ). Let t+,r (g),
t+,A (g) be the maximal existence time of the solution with initial condition
g with respect to these cases. Then t+,r (g) = t+,A (g).

Proof. Note that by (Lemma 8.33 from [47])

kf gkH r ≤ Cn,r kf kH r kgkA + kf kA kgkH r
the Lipschitz constant depends only on the Wiener norm of the functions.
In particular,
Z t
ku(t)kH ≤ ku(0)kH +
r r kF (u(s))kH r ds
0
Z t
≤ ku(0)kH r + C̃n,r k(u(s))kα−1
A k(u(s))kH ds
r
0
and hence Gronwall’s inequality shows that the H r norm can be controlled
by the Wiener norm. In particular, the H r norm cannot blow up before the
A norm. 

In order to get global solutions the following conservation laws will be

crucial: Momentum
1
M (t) := ku(t)k22
2
and energy
1 1
E(t) := k∂u(t)k22 ± ku(t)kα+1
α+1 .
2 α+1
Lemma 9.5. Let r > n2 and g ∈ H r (Rn ). Then M (t) = M (0) for all
t ∈ (t− (g), t+ (g)). If in addition, r ≥ 1 then also E(t) = E(0) for all
t ∈ (t− (g), t+ (g)).

Proof. If u is a sufficiently smooth solution this can be verified directly

(Problem 9.1). For the general case approximate by smooth solutions (using
local Lipschitz continuity of the solution map) and conclude that M (t) is
locally constant and hence constant on its interval of existence. Similarly for
E(t). 

So in the focusing case we get existence of global solutions in H 1 if n = 1

such that our local results holds for r = 1. In the defocusing case the energy
is not positive and we cannot immediately control the H 1 norm using E and
M.
Finally we remark that this approach still applies if we replace the non-
linearity by an arbitrary polynomial in u and u∗ except for the conservation
laws, which depend crucially on the form of the nonlinearity.
264 9. The nonlinear Schrödinger equation

Problem 9.1. Let u ∈ C([−t0 , t0 ], H r+2 (Rn )) ∩ C 1 ([−t0 , t0 ], H r (Rn )) be a

strong solution of the NLS equation (with r > n2 ). Show that momentum and
energy are independent of t ∈ [−t0 , t0 ].

9.2. Strichartz estimates

In order to improve the results from the previous section we need a better
understanding of the linear Schrödinger equation. Unlike for example the
heat equation, the Schrödinger equation does only preserve but not improve
the regularity of the initial condition. For example, choosing f ∈ L2 \ Lp
(for some p 6= 2) and considering g = TS (−t0 )f shows that there are initial
conditions in L2 which are not in Lp at a given later time t0 . However, our
aim in this section is to show that we still have T (t)g ∈ Lp most of the time.
To this end we first need an explicit expression for the solution. As
in the case of the heat equation, we would like to express our solution as a
convolution with the initial condition. However, here we run into the problem
2
that e−i|p| t is not integrable. To overcome this problem we consider
2
fε (p) = e−(it+ε)p , ε > 0. (9.10)
Then using the well-known formula (Lemma 8.6 from [47])
2 1 2
F(e−z|x| /2 )(p) = n/2 e−|p| /(2z) , Re(z) > 0, (9.11)
z
where z n/2 is the standard branch with branch cut along the negative real
axis, together with the fact that the Fourier transform maps convolutions
into products (Corollary 8.15 from [47]) we obtain
|x−y|2
Z
∨ 1 − 4(it+ε)
(fε ĝ) (x) = e g(y)dn y. (9.12)
(4π(it + ε))n/2 Rn
Taking the limit ε ↓ 0 we finally arrive at
Z
1 |x−y|2
TS (t)g(x) = ei 4t
g(y)dn y (9.13)
(4πit)n/2 Rn
for t 6= 0 and g ∈ L2 (Rn ) ∩ L1 (Rn ). In fact, the left-hand side converges
to TS (t)g in L2 and the limit of the right-hand side exists pointwise by
dominated convergence and its pointwise limit must thus be equal to its L2
limit.
Using this explicit form, we can again draw some further consequences.
For example, if g ∈ L2 (Rn ) ∩ L1 (Rn ), then u(t) := TS (t)g ∈ C0 (Rn ) for t 6= 0
(Problem 9.3) and satisfies
1
ku(t)k∞ ≤ kgk1 . (9.14)
|4πt|n/2
Moreover, we even have u ∈ C(R \ {0}, C0 (Rn )) (Problem 9.3).
9.2. Strichartz estimates 265

Thus we have spreading of wave functions in this case. In fact, invoking

the Riesz–Thorin interpolation theorem (Theorem 9.2 from [47]) we even
get
1
ku(t)kp ≤ n/2−n/p
kgkp0 . (9.15)
|4πt|
for any p ∈ [2, ∞] with 1
p + 1
p0 = 1. This also gives u ∈ C(R \ {0}, Lp (Rn )).
Next we look at average decay in an Lp sense instead of pointwise es-
timates with respect to t. To this end we will consider functions f ∈
Lr (R, Lp (Rn )) and we will denote the corresponding norm by
( R
1/r
kf (t)krp dt , r < ∞,
kf kLr (Lp ) := R (9.16)
supt∈R kf (t)kp , r = ∞.
Please recall that Lr (R, Lp (Rn )) is a Banach space defined with the help of
the Bochner integral (cf. Theorem 5.32 from [47]). It consists of (equivalence
classes with respect to equality a.e. of) strongly measurable functions f (t) for
which kf (t)kp is in Lr . Here strongly measurable means, that f (t) is a limit
of simple functions sn (t). It turns out that a function is strongly measurable
if and only if it is measurable and its range is separable. In our situation
this latter condition will come for free in the case p < ∞ and similarly in
the case p = ∞ if the range is contained in C0 (Rn ). We will also need the
following variational characterization of our space-time norms (Problem 5.25
from [47]) for a given strongly measurable function f :
Z Z 
n
(9.17)
 
kf kLr (Lp ) = sup 
 f (x, t)g(x, t)d x dt.

kgk 0 0 =1 R Rn
Lr (Lp )

Moreover, it suffices to take the sup over functions which have support in a
compact rectangle.
We call a pair (p, r) admissible if
(
2 ≤ p ≤ ∞, n=1 2 n n
2n
, = − . (9.18)
2 ≤ p < n−2 , n ≥ 2 r 2 p

Note r ∈ [4, ∞] for n = 1 and r ∈ ( n−1

2
, ∞] for n ≥ 2.

Lemma 9.6. Let TS be the Schrödinger group and let (p, r) be admissible
with p > 2. Then we have
Z Z r 1/r
kTS (t − s)g(s)kp ds dt ≤ CkgkLr0 (Lp0 ) , (9.19)
R R

where a prime denotes the corresponding dual index. Moreover, s 7→ T (t −

s)g(s) ∈ Lp (Rn ) is integrable for a.e. t ∈ R.
266 9. The nonlinear Schrödinger equation

Proof. Of course TS (t − s)g(s) is measurable. Applying our interpolation

estimate we obtain
Z Z
1
kTS (t − s)g(s)kp ds ≤ C kg(s)kp0 ds,
|t − s|1−α

where α = 1 − n(1/2 − 1/p) ∈ (0, 1) by our restriction on p.

Furthermore, our choice for r implies α = 1 − 2r = r10 − 1r with r0 = 1+α
2
∈
(1, α ). So taking the k.kLr norm on both sides and using the Hardy–
−1

Littlewood–Sobolev inequality (Theorem 9.10 from [47]) gives the estimate.

Hence the claim about integrability follows from Minkowski’s integral
inequality (Theorem 5.30 from [47]). 

Note that the case p = 2 (and r = ∞) in the above lemma holds by

unitarity and does not provide much new insight.

Theorem 9.7 (Strichartz estimates). Let TS be the Schrödinger group and

0 0
let (p, r) be admissible. Suppose g ∈ Lr (R, Lp (Rn )) and f ∈ L2 (Rn ). Then
we have the following estimates:

kTS (t)f kLr (Lp ) ≤ Ckf k2 , (9.20)

Z
(9.21)

TS (s)g(s)ds ≤ Ckgk r0 p0 ,
L (L )
R 2
Z
(9.22)

TS (t − s)g(s)ds
r p ≤ CkgkLr0 (Lp0 ) ,
R L (L )

where a prime denotes the corresponding dual index.

Here s 7→ TS (t − s)g(s) ∈ Lp (Rn ) is integrable for a.e. t ∈ R and the
integral in (9.21) has to be understood as a limit in L2 when taking an ap-
proximating sequence of functions g with support in compact rectangles.

Proof. Since the case p = 2 follows from unitarity, we can assume p > 2.
The claims about integrability and the last estimate follow from the lemma.
Using unitarity of TS and Fubini we get
Z Z Z Z
(TS (t)f )(x)g(t, x)dn x dt = f (x) (TS (t)g(t))(x)dt dn x,
R Rn Rn R
0 0
for g ∈ Lr (R, Lp (Rn )) with support in a compact rectangle. Note that in
this case we have g(t) ∈ L2 (Rn ) since p0 ≤ 2. This shows that the first and
second estimate are equivalent upon using the above characterization (9.17)
as well as the analogous characterization for the L2 norm.
9.2. Strichartz estimates 267

Similarly, using again unitarity of TS and Fubini

Z 2

TS (t)g(t)dt

2
Z Z Z
= (TS (t)g(t))(x)dt (TS (s)g(s))(x)∗ ds dn x
n
ZR ZR Z R

= g(t, x) TS (t − s)g(s, x)∗ ds dt dn x,

Rn R R

which shows that the second and the third estimate are equivalent with a
similar argument as before. 

Note that using the scaling f (x) → f (λx) for λ > 0 shows that the left-
hand side of (9.20) scales like λn/p+2/r while the right-hand side scales like
λn/2 . So (9.20) can only hold if np + 2r = n2 .
In connection with the Duhamel formula the following easy consequence
is also worth while noticing:

Corollary 9.8. We also have

Z t
(9.23)

TS (t − s)g(s)ds ≤ Ckgk r0 p0 ,
L (L )
0 2
Z t
(9.24)

TS (t − s)g(s)ds
r p ≤ CkgkLr0 (Lp0 ) .
0 L (L )

Proof. The second estimate is immediate from the lemma and the first
estimate follows from (9.21) upon restricting to functions g supported in
Rt Rt
[0, t] and using a simple change of variables 0 T (t − s)g(s)ds = 0 T (s)g(t −
s)ds. 

Note that apart from unitarity of TS only (9.14) was used to derive these
estimates. Moreover, since TS commutes with derivatives, we can also get
analogous estimates for derivatives:

Corollary 9.9. We have the following estimates for k ∈ N0 :

kTS (t)f kLr (W k,p ) ≤ Ckf kH k , (9.25)

Z
(9.26)

TS (s)g(s)ds ≤ Ckgk r0 k,p0 ,
k L (W )
R H
Z
(9.27)

TS (t − s)g(s)ds
r k,p ≤ CkgkLr0 (W k,p0 ) ,
R L (W )
268 9. The nonlinear Schrödinger equation

as well as
Z t
(9.28)

TS (t − s)g(s)ds ≤ Ckgk r0 k,p0 ,
k L (W )
0 H
t
Z
(9.29)


TS (t − s)g(s)ds
≤ CkgkLr0 (W k,p0 ) .
0 Lr (W k,p )

Proof. Consider dense sets f ∈ S(Rn ) and g ∈ Cc (R, S(Rn )). Then we have
for example
k∂j TS (t)f kLr (Lp ) = kTS (t)∂j f kLr (Lp ) ≤ Ck∂j f k2
by applying (9.20) to ∂j f . Combining the estimates for f and its derivatives
gives (9.25). Similarly for the other estimates. 
Problem 9.2. Does the translation group T (t)g(x) := g(x−t) satisfy (9.14)?
Problem 9.3. Let u(t) := TS (t)g for some g ∈ L1 (Rn ). Show that u ∈
C(R \ {0}, C0 (Rn )). (Hint: Lemma 4.33 (iv).)
Problem 9.4. Show that (9.20) can only hold if 2r = n2 − np . (Hint: Look
for a scaling which leaves the Schrödinger equation invariant.)
Problem 9.5. Prove that there is no triple p, q, t with 1 ≤ q < p < ∞, t ∈ R
such that
kTS (t)gkq ≤ Ckgkp .
(Hint: The translation operator Ta f (x) := f (x − a) commutes with TS (t).
Moreover, we have
lim kf + Ta f kp = 21/p kf kp , 1 ≤ p < ∞.
|a|→∞

Now apply this to the claimed estimate.)

9.3. Well-posedness in L2 and H 1

The main obstacle to proving a local existence result in L2 is the fact that
our nonlinearity does not map L2 to L2 (and this was precisely the reason for
choosing H r in the previous section). On the other hand, the time evolution
conserves the L2 norm and hence we expect global solutions in this case.
So let us make two observations: First of all our nonlinearity F (u) =
|u|α−1 u maps Lp to Lp/α , so the only chance is that the linear time evolution
improves this behavior. Now we know, since our evolution is unitary, there
is no hope to get this for fixed t, but this is true in some averaged sense
by the Strichartz estimate (9.20). Hence, if we add such a space-time norm
to the L2 norm, we might be able to control our singularity. In fact, the
estimates (9.23) and (9.24) allow us to control the Duhamel part in (9.7)
both in the L2 and the space-time norm, respectively (the linear part being
9.3. Well-posedness in L2 and H 1 269

taken care of by unitarity and (9.20)). Since the spatial parts of the space-
time norms must match up, we need p0 α = p, that is, p = 1 + α. For the
0
time part an inequality r0 α ≤ r is sufficient since in this case Lr α ⊆ Lr by
Hölder’s inequality. This imposes the restriction α ≤ 1 + n4 . In fact, we will
impose a strict inequality since we will use the contribution from Hölder’s
inequality to get a contraction. Moreover, note that the dependence on the
initial condition g is controlled by the L2 norm alone and this will imply
that our contraction is uniform (in fact Lipschitz on bounded domains) with
respect to the initial condition in L2 , and so will be the solution.
4
Theorem 9.10. Suppose 1 < α < 1 + n and consider the Banach space
4(α + 1)
X := C([−t0 , t0 ], L2 (Rn )) ∩ Lr ([−t0 , t0 ], Lα+1 (Rn )), r= , (9.30)
n(α − 1)
with norm
Z t0 1/r
kf k := sup kf (t)k2 + kf (t)krα+1 dt . (9.31)
t∈[−t0 ,t0 ] −t0

Then for every g ∈ L2 (Rn ) there is a t0 = t0 (kgk2 ) > 0, such that there is a
unique solution u ∈ X of (9.7). Moreover, the solution map g 7→ u(t) will
be Lipschitz continuous from every ball kgk2 ≤ ρ to X defined with t0 (ρ).

Proof. We take [0, t0 ] as an interval for notational simplicity. We will show

that (9.7) gives rise to a contraction on the closed ball B̄a (0) ⊂ X provided
a and t0 are chosen accordingly. Denote the right-hand side of (9.7) by
K(u) ≡ Kg (u). We will fist show that K : B̄a (0) → B̄a (0) for a suitable
a depending on kgk2 . To this end we first invoke (9.20), (9.23), and (9.24)
with p = α + 1 (p0 = α+1 α ) to obtain
Z t0 1/r0
α r0
kK(u)k ≤ (1 + C)kgk2 + 2C k|u| (t)k(α+1)/α dt
0
Z t0 1/r0
0
≤ (1 + C)kgk2 + 2C ku(t)kαr
α+1 dt .
0
n(α−1)
Next, since r10 = θ + α−1
r + r , where θ = 1 − r = 1 −
1 α+1
4 > 0 we can
use the generalized Hölder inequality in the form
k1 · f1α−1 f2 kr0 ≤ k1k1/θ kf1α−1 kr/(α−1) kf2 kr = tθ0 kf1 kα−1
r kf2 kr
(with f1 (t) = f2 (t) = ku(t)kα+1 ) to obtain
Z t0 α/r
θ
kK(u)k ≤ (1 + C)kgk2 + 2Ct0 ku(t)krα+1 dt
0
≤ (1 + C)kgk2 + 2Ctθ0 aα
270 9. The nonlinear Schrödinger equation

for u ∈ B̄a (0). Now we choose a = (2 + C)kgk2 and 2C(2 + C)tθ0 aα−1 < 1
such that
kK(u)k ≤ (1 + C)kgk2 + 2Ctθ0 (2 + C)α kgkα2 < (2 + C)kgk2 = a.
Similarly we can show that K is a contraction. Invoking (9.23) and (9.24)
we have
Z t0 1/r0
α−1 α−1 r0
kK(u) − K(v)k ≤ 2C k|u(t)| u(t) − |v(t)| v(t)k(α+1)/α dt
0
Now using (Problem 9.6)
 α−1
u − |v|α−1 v  ≤ α |u|α−1 + |v|α−1 |u − v|,


|u| u, v ∈ C,
and invoking the generalized Hölder inequality in the form
k|u|α−1 |u − v|k(α+1)/α ≤ k|u|α−1 k(α+1)/(α−1) ku − vkα+1 = kukα−1
α+1 ku − vkα+1
and then in the previous form with f1 = kukα+1 , f2 = ku − vkα+1 , we obtain
Z t0 1/r0
α−1 α−1

r 0
kK(u) − K(v)k ≤ 2αC (kukα+1 + kvkα+1 )ku − vkα+1 dt
0
Z t0 1/r
≤ 2αCtθ0 2aα−1 ku − vkrα+1 dt
0
≤ 4αCtθ0 aα−1 ku − vk.
Hence, decreasing t0 further (if necessary), such that we also have 4αCtθ aα−1 <
1, we get a contraction. Moreover, since kKg (u) − Kf (u)k = kKg−f (0)k ≤
(1+C)kg−f k2 the uniform contraction principle establishes the theorem. 

By interpolation (Problem 9.8) we also have:

Corollary 9.11. The solution u is also in
Lr/θ ([−t0 , t0 ], L2(α+1)/(α+1−θ(α−1)) (Rn ))
for any θ ∈ (0, 1).

Moreover, as in the previous section we obtain:

Corollary 9.12. The maximal solution u is global in C(R, L2 (Rn )) and
preserves the L2 norm: ku(t)k2 = kgk2 . In addition, it has the properties
stated in the theorem for any t0 > 0.

Let me remark that it is possible to cover the case α = 1 + n4 . The main

difference is that the Hölder-type estimate in terms of tθ for the integral in
(9.7) is useless since θ = 0. However, the integral still tends to zero as t → 0.
This will be true locally in a sufficiently small neighborhood, but we cannot
control this neighborhood in terms of kgk2 .
However, we will turn to the case of initial conditions in H 1 instead.
9.3. Well-posedness in L2 and H 1 271

n+2
Theorem 9.13. Suppose n ≥ 3 and 2 ≤ α < n−2 . Consider the Banach
space
X := C([−t0 , t0 ], H 1 (Rn )) ∩ Lr ([−t0 , t0 ], W 1,p (Rn )), (9.32)
where
n(α + 1) 4(α + 1)
p= , r= , (9.33)
n+α−1 (n − 2)(α − 1)
with norm
Z t0 1/r
kf k := sup kf (t)k1,2 + kf (t)kr1,p dt . (9.34)
t∈[−t0 ,t0 ] −t0

Then for every g ∈ H 1 (Rn ) there is a t0 = t0 (kgk1,2 ) > 0, such that there is
a unique solution u ∈ X of (9.7). Moreover, the solution map g 7→ u(t) will
be Lipschitz continuous from every ball kgk1,2 ≤ ρ to X defined with t0 (ρ).

Proof. We begin with estimating the nonlinearity. For u, v ∈ W 1,p and

w ∈ Lp we obtain

k|u|α−2 vwkp0 ≤ kukα−2

q kvkq kwkp ≤ Ck∂ukα−2
p k∂vkp kwkp ,

where we have applied the generalized Hölder inequality with p10 = α−2 1 1
q +q+p
in the first step (requiring α ≥ 2) and the Gagliardo–Nierenberg–Sobolev
inequality (Theorem 7.18 from [47] – since we need p < n, we need to
require n > 2) with 1q = p1 − n1 in the second step. In particular, this imposes

2 α−1 α−1 α−1

1− = = −
p q p n
and explains our choice for p. The choice of r is of course dictated by (9.18)
such that we can apply our Strichartz estimates. At this point a weaker
upper bound (namely α < n−4 n
for n ≥ 4) is still sufficient.
Now using this estimate we see (cf. Problem 9.6)

k|u|α−1 ukp0 ≤ Ck∂ukpα−1 kukp , k∂|u|α−1 ukp0 ≤ αk|u|α−1 ∂ukp0 ≤ αCk∂ukαp

and hence
k|u|α−1 uk1,p0 ≤ C̃kukα1,p .
Similarly we obtain

k|u|α−1 u − |v|α−1 vkp0 ≤ αk |u|α−1 + |v|α−1 |u − v|kp0

≤ αC k∂ukα−1 + k∂vkα−1


p p ku − vkp
272 9. The nonlinear Schrödinger equation

and

k∂|u|α−1 u − ∂|v|α−1 vkp0

≤ (α − 1)(α + 2)k |u|α−2 + |v|α−2 |u − v| |∂u|kp0

+ αk|v|α−1 |∂u − ∂v|kp0

≤ (α − 1)(α + 2)C k∂ukα−2 + k∂vkα−2


p p k∂(u − v)kp k∂ukp
+ αCk∂vkpα−1 k∂u − ∂vkp .

In summary,

k|u|α−1 u − |v|α−1 vk1,p0 ≤ C̄ kukα−1 α−1

1,p + kvk1,p ku − vk1,p .

Now the rest follows as in the proof of Theorem 9.10. Note that in this case
2+n+(2−n)α
θ = 1 − α+1
r = 4 explaining our upper limit for α. 

Note that since we have H 1 (Rn ) ⊆ Lα+1 (Rn ) for n ≥ 3 and α < n−2
n+2

by the Gagliardo–Nierenberg–Sobolev inequality (Theorem 7.18 from [47]),

both the momentum and the energy are finite and preserved by our solutions.
Moreover, in the defocusing case the momentum and the energy control the
H 1 norm and hence we obtain:

Corollary 9.14. In the defocusing case the maximal solution u is global in

C(R, H 1 (Rn )) and preserves both momentum and energy. In addition, it has
the properties stated in the theorem for any t0 > 0.

In the focusing case we need to control the Lα+1 norm in terms of the
H1 norm using the Gagliardo–Nierenberg–Sobolev inequality.

Corollary 9.15. In the focusing case the maximal solution u is global in

C(R, H 1 (Rn )) and preserves both momentum and energy if one of the fol-
lowing conditions hold:
(i) α < 1 + n4 .
(ii) α = 1 + 4
n
2C
and kgk2 < ( 2+4/n )n/4 .
(iii) α > 1 + 4
n and kgk1,2 is sufficiently small such that 2E(0) +
2C α+1−n(α−1)/2
α+1 kgk2 < 1.

Proof. Using the Gagliardo–Nierenberg–Sobolev inequality and the Lya-

punov inequality (Problem 3.12 from [47]) with 1+α
1
2 (i.e.
= θ( 21 − n1 ) + 1−θ
n(α−1)
θ= 2(α+1) ) we obtain
α+1−n(α−1)/2 n(α−1)/2
ku(t)kα+1
α+1 ≤ Cku(t)k2 k∂u(t)k2 . (9.35)
9.3. Well-posedness in L2 and H 1 273

Thus
2
k∂u(t)k22 = 2E(0) + ku(t)kα+1
α+1
α+1
2C α+1−n(α−1)/2 n(α−1)/2
≤ 2E(0) + kgk2 k∂u(t)k2 . (9.36)
α+1
n(α−1)
(i). Now if α < 1 + n4 , then 2 < 2 and k∂u(t)k2 remains bounded.
4/n
(ii). In the case α = 1 + 4
n this remains still true if 2C
2+4/n kgk2 < 1.
(iii). If α >
1+ n4we can choose kgk1,2 so small such that the given conditions
hold. Note that this is possible since our above calculation shows
1 2C α+1−n(α−1)/2 n(α−1)/2
E(0) ≤ k∂gk22 + kgk2 k∂gk2 .
2 α+1
Now if we start with k∂u(0)k22 ≤ 1 and assume k∂u(t)k22 = 1 we get the
α+1−n(α−1)/2
contradiction 1 = k∂u(t)k22 ≤ 2E(0) + α+1
2C
kgk2 < 1. Hence
k∂u(t)k22 < 1 as desired. 

Problem 9.6. Show that the real derivative (with respect to the identification
C∼
= R2 ) of F (u) = |u|α−1 u is given by

F 0 (u)v = |u|α−1 v + (α − 1)|u|α−3 uRe(u∗ v).

Conclude in particular,

|F 0 (u)v| ≤ α|u|α−1 |v|, |F (u) − F (v)| ≤ α(|u|α−1 + |v|α−1 )|u − v|.

Moreover, the second derivative is given by

vF 00 (u)w = (α − 1)|u|α−5 u (α + 1)Re(u∗ v)Re(u∗ w) − u2 v ∗ w∗ .

and hence
|vF 00 (u)w| ≤ (α − 1)(α + 2)|u|α−2 |v||w|.

Problem 9.7. Show that (9.32) is a Banach space. (Hint: Work with test
functions from Cc∞ .)

Problem 9.8. Suppose f ∈ Lp0 (I, Lq0 (U )) ∩ Lp1 (I, Lq1 (U )). Show that
f ∈ Lpθ (I, Lqθ (U )) for θ ∈ [0, 1], where
1 1−θ θ 1 1−θ θ
= + , = + .
pθ p0 p1 qθ q0 q1
(Hint: Lyapunov and generalized Hölder inequality — Problem 3.12 from
[47] and Problem 3.9 from [47].)
274 9. The nonlinear Schrödinger equation

9.4. Blowup in H 1
In this section we will show that solutions are not always global in the focus-
ing case. For simplicity we will only consider the one-dimensional case. We
first complement Theorem 9.13 with a result for the one-dimensional case.
Theorem 9.16. Let n = 1 and α ≥ 2. For every g ∈ H 1 (R) there is a t0 =
t0 (kgk1,2 ) > 0, such that there is a unique solution u ∈ C([−t0 , t0 ], H 1 (R))
of (9.7). Moreover, the solution map g 7→ u(t) will be Lipschitz continuous
from every ball kgk1,2 ≤ ρ to C([−t0 (ρ), t0 (ρ)], H 1 (R)).

Proof. It suffices to verify that F : H 1 (R) → H 1 (R) is locally Lipschitz on

bounded sets. But this follows using Problem 9.6 since
kF (u) − F (v)k2 ≤ α(kukα−1 α−1
∞ + kvk∞ )ku − vk2

and
k∂(F (u) − F (v))k2 ≤ (α − 1)(α + 2)(kukα−2 α−2
∞ + kvk∞ )k∂uk2 ku − vk2
+ αkvkα−1
∞ k∂(u − v)k2

together with kf k∞ ≤ kf k1,2 (Problem 9.9). 

In the one-dimensional case we did not need a space-time norm to estab-

lish the above theorem. However, it is worthwhile mentioning that they still
hold.
Corollary 9.17. Let u be the solution for given g ∈ H 1 . Then
u ∈ Lr ([−t0 , t0 ], Lp (Rn )) (9.37)
whenever (p, r) is an admissible pair satisfying (9.18).

Proof. Just observe that the solution satisfies

0 α− p20 2
0 0
kF (u)kLr0 (Lp0 ) ≤ (2t0 )1/r sup kuk∞ kuk2p ≤ (2t0 )1/r sup kukα1,2
|t|≤t0 |t|≤t0

and apply the Strichartz estimates (9.20) and (9.24). 

As in the previous section one gets global existence in the defocusing

case and, with a little more effort, also in the focusing case if α < 5 (Prob-
lem 9.10).
Lemma 9.18. Let n = 1 and α ≥ 2. Suppose g ∈ H 1 (R) satisfies kxg(x)k2 <
∞ and let u ∈ C((t− , t+ ), H 1 (R)) be the maximal solution of (9.7). Then
Z
M1 (t) := x2 |u(t, x)|2 dx (9.38)
R
9.4. Blowup in H 1 275

remains finite as long as u exists and satisfies

Z
Ṁ1 (t) = 4Im x u(t, x)∗ ∂u(t, x)dx, (9.39)
R
4(α − 5)
Z
M̈1 (t) = 16E(t) ± |u(t, x)|α+1 dx, (9.40)
α+1 R
known as virial and Morawetz identity, respectively.

Proof. Consider H 1,1 (R) := H 1 (R) ∩ L2 (R, x2 dx) together with the norm
kf k2 = kf k22 + kf 0 k22 + kxf (x)k22 . Then TS (t) is a C 0 group satisfying
kTS (t)f k ≤ (1 + 2|t|)kf k. Moreover, as in the previous theorem one ver-
ifies that F : H 1,1 (R) → H 1,1 (R) is locally Lipschitz on bounded sets. In
fact, note that by
kx(F (u)(x) − F (v)(x))k2 ≤ α(kukα−1 α−1
∞ + kvk∞ )kx(u(x) − v(x))k2

the Lipschitz constant depends only on the H 1 norm. Hence we get existence
of local solutions. Moreover, using (9.7) we obtain
Z t
ku(t)k ≤ (1 + 2t)kgk + α (1 + 2(t − s))ku(s)kα−1
1,2 ku(s)kds,
0
and Gronwall’s inequality
Z t 
ku(t)k ≤ (1 + 2t)kgk exp (1 + 2(t − s))ku(s)kα−1
1,2 ds
0

shows that our norm cannot blow up before the H 1 norm.

The formulas for Ṁ1 and M̈1 are straightforward computations. 

Now we are ready to establish blowup for the focusing NLS equation.
Theorem 9.19. Consider the one-dimensional focusing NLS equation with
α ≥ 5. Let g ∈ H 1 (R) ∩ L2 (R, x2 dx) with negative energy E < 0. Then the
corresponding maximal mild solution u satisfies t+ (g) < ∞.

Proof. Due to our assumption α ≥ 5 we obtain M̈1 (t) ≤ 16E implying

M1 (t) ≤ 8Et2 + Ṁ1 (0)t + M1 (0). Hence
 
−1
q
t+ (g) < Ṁ1 (0) + Ṁ1 (0) − 32EM1 (0)
16E
since M1 (t) must remain positive. Note that this also shows Ṁ1 (0) > 0 since
otherwise M1 (t) would be decreasing and hence would remain bounded. 

Notice that there are initial conditions with negative energy, since the
two contributions to the energy scale differently. In particular, the energy
will become negative if we scale g with a sufficiently large factor.
276 9. The nonlinear Schrödinger equation

Problem 9.9. Let f ∈ H 1 (R). Show kf k2∞ ≤ 2kf k2 kf 0 k2 and hence

kf k∞ ≤ kf k1,2 .

Problem 9.10. Show that the one-dimensional focusing NLS equation has
global solutions in H 1 (R) if either α < 5 or α = 5 and kgk2 ≤ ( 34 )1/4 or α > 5
and kgk1,2 sufficiently small. (Hint: Use the estimate from Problem 9.9.)

9.5. Standing waves

A solution of the form
u(x, t) = ϕω (x)eiωt , ω > 0, (9.41)
of the focusing NLS equation is called a standing wave. Inserting this
ansatz into the equation shows that ϕω must be a solution of the following
nonlinear elliptic problem
− ∆ϕω + ωϕω = |ϕω |α−1 ϕω . (9.42)
Note that one can choose ω = 1 without loss of generality since if ϕ is a
solution for ω = 1 then
1
ϕω (x) = ω α−1 ϕ(ω 1/2 x) (9.43)
is a solution for ω > 0. Moreover, if ϕ is a solution, so is eiθ ϕ(. − a) for any
θ ∈ R and a ∈ Rn .
If one multiplies (9.42) with a test function v ∈ H 1 (Rn ) and integrates
over Rn one obtains the weak formulation
Z
∂ϕ · ∂v + ϕv − |ϕ|α−1 ϕv dn x = 0, v ∈ H 1 (Rn ). (9.44)


Rn

In particular, choosing v = ϕ∗ we obtain

Z
|∂ϕ|2 + |ϕ|2 − |ϕ|α+1 dn x = 0, (9.45)


Rn

which shows that, if we flip the sign in front of the nonlinearity (defocusing
case), there is only the trivial solution.
In one-dimension one has the explicit solution
 √ 1/β
1+β α−1
ϕ(x) = , β= . (9.46)
cosh(βx) 2
In higher dimensions we can apply Theorem 7.26 to get existence of solutions:
n+2
Theorem 9.20. Suppose n ≥ 2 and 1 < α < n−2 . Then the nonlinear
elliptic problem (9.46) has a weak positive radial solution in H 1 (Rn ).
9.5. Standing waves 277

Proof. To apply Theorem 7.26 we choose X = Hrad 1 (Rn , R) and Y =

Lrad (R , R) and note that the Strauss lemma (Problem 7.31 from [47])
α+1 n

implies compactness of the embedding X ,→ Y for the range of α under

consideration. Hence minimizing
Z
1
|∂u|2 + |u|2 dn x


F (u) =
2 Rn
under the constraint (cf. Example 7.5)
Z
1
N (u) = |u|α+1 dn x = 1
α + 1 Rn
gives a weak radial solution u0 of the problem
−∆u + u = λ|u|α−1 u.
In particular, choosing u0 as a test function for the weak formulation shows
λ > 0. Moreover, by Problem 7.16 from [47] we have |u0 | ∈ Hrad 1 (Rn ) with

F (|u0 |) = F (u0 ) and hence |u0 | is also a minimizer. Rescaling this solution
according to ϕ(x) = λ1/(α−1) |u0 (x)| establishes the claim. 

Note that for α ≤ n−2

n
we have |u|α−1 u ∈ L2 (Rn ) for u ∈ H 1 (Rn ) and
hence (−∆ + 1)ϕ ∈ L2 (Rn ) implying ϕ ∈ H 2 (Rn ).
n+2
Problem 9.11. Let 1 < α < n−2 be an odd integer (i.e. n = 2 and α =
3, 5, 6, . . . or n = 3 and α = 3). Show that ϕ ∈ H k (Rn ) for any k ∈ N.
(Hint: As already pointed out we have ϕ ∈ H 2 .)
 Chapter 10

The Brouwer mapping

degree

10.1. Introduction
Many applications lead to the problem of finding zeros of a mapping f : U ⊆
X → X, where X is some (real) Banach space. That is, we are interested in
the solutions of
f (x) = 0, x ∈ U. (10.1)
In most cases it turns out that this is too much to ask for, since determining
the zeros analytically is in general impossible.
Hence one has to ask some weaker questions and hope to find answers
for them. One such question would be “Are there any solutions, respectively,
how many are there?”. Luckily, these questions allow some progress.
To see how, lets consider the case f ∈ H(C), where H(U ) denotes the set
of holomorphic functions on a domain U ⊂ C. Recall the concept of the
winding number from complex analysis. The winding number of a path
γ : [0, 1] → C \ {z0 } around a point z0 ∈ C is defined by
Z
1 dz
n(γ, z0 ) := ∈ Z. (10.2)
2πi γ z − z0
It gives the number of times γ encircles z0 taking orientation into account.
That is, encirclings in opposite directions are counted with opposite signs.
In particular, if we pick f ∈ H(C) one computes (assuming 0 6∈ f (γ))
Z 0
1 f (z) X
n(f (γ), 0) = dz = n(γ, zk )αk , (10.3)
2πi γ f (z)
k

279
280 10. The Brouwer mapping degree

where zk denote the zeros of f and αk their respective multiplicity. Moreover,

if γ is a Jordan curve encircling a simply connected domain U ⊂ C, then
n(γ, zk ) = 0 if zk 6∈ U and n(γ, zk ) = 1 if zk ∈ U . Hence n(f (γ), 0) counts
the number of zeros inside U .
However, this result is useless unless we have an efficient way of comput-
ing n(f (γ), 0) (which does not involve the knowledge of the zeros zk ). This
is our next task.
Now, lets recall how one would compute complex integrals along com-
plicated paths. Clearly, one would use homotopy invariance and look for a
simpler path along which the integral can be computed and which is homo-
topic to the original one. In particular, if f : γ → C \ {0} and g : γ → C \ {0}
are homotopic, we have n(f (γ), 0) = n(g(γ), 0) (which is known as Rouché’s
theorem).
More explicitly, we need to find a mapping g for which n(g(γ), 0) can be
computed and a homotopy H : [0, 1]×γ → C\{0} such that H(0, z) = f (z)
and H(1, z) = g(z) for z ∈ γ. For example, how many zeros of f (z) =
2 z + z − 3 lie inside the unit circle? Consider g(z) = z, then H(t, z) =
1 6 1

(1 − t)f (z) + t g(z) is the required homotopy since |f (z) − g(z)| < |g(z)|,
|z| = 1, implying H(t, z) 6= 0 on [0, 1] × γ. Hence f (z) has one zero inside
the unit circle.
Summarizing, given a (sufficiently smooth) domain U with enclosing Jor-
dan curve ∂U , we have defined a degree deg(f, U, z0 ) = n(f (∂U ), z0 ) =
n(f (∂U ) − z0 , 0) ∈ Z which counts the number of solutions of f (z) = z0
inside U . The invariance of this degree with respect to certain deformations
of f allowed us to explicitly compute deg(f, U, z0 ) even in nontrivial cases.
Our ultimate goal is to extend this approach to continuous functions
f : Rn → Rn . However, such a generalization runs into several problems.
First of all, it is unclear how one should define the multiplicity of a zero. But
even more severe is the fact, that the number of zeros is unstable with respect
to small perturbations. For example, consider fε : [−1, 2] → R, x 7→ x2 − ε.
Then fε has no √ zeros for ε < 0, one zero
√ for ε = 0, two zeros for 0 < ε ≤ 1,
one for 1 < ε ≤ 2, and none for ε > 2. This shows the following facts.
(i) Zeros with f 0 6= 0 are stable under small perturbations.
(ii) The number of zeros can change if two zeros with opposite sign
change (i.e., opposite signs of f 0 ) run into each other.
(iii) The number of zeros can change if a zero drops over the boundary.
Hence we see that we cannot expect too much from our degree. In addition,
since it is unclear how it should be defined, we will first require some basic
properties a degree should have and then we will look for functions satisfying
these properties.
10.2. Definition of the mapping degree and the determinant formula 281

10.2. Definition of the mapping degree and the determinant

formula
To begin with, let us introduce some useful notation. Throughout this section
U will be a bounded open subset of Rn . For f ∈ C 1 (U, Rn ) the Jacobi matrix
of f at x ∈ U is df (x) = (∂xj fi (x))1≤i,j≤n and the Jacobi determinant of f
at x ∈ U is
Jf (x) := det df (x). (10.4)
The set of regular values is
RV(f ) := {y ∈ Rn |∀x ∈ f −1 (y) : Jf (x) 6= 0}. (10.5)
Its complement CV(f ) := Rn \ RV(f ) is called the set of critical values.
We will also need the spaces
C̄ k (U, Rn ) := C k (U, Rn ) ∩ C(U , Rn ) (10.6)
and regard them as subspaces of the Banach space C(U , Rn ) (cf. Section 7.2).
Note that C̄ ∞ (U, Rn ) is dense in C(U , Rn ). To see this you can either ap-
ply Stone–Weierstraß or use the Tietze extension theorem to extend f ∈
C(U , Rn ) to all of Rn and then mollify. If you use mollification and f ∈
C̄ k (U, Rn ) then all derivatives up to order k will converge uniformly on com-
pact subsets of U . Finally, for y ∈ Rn we set
C̄yk (U , Rn ) := {f ∈ C̄ k (U, Rn )|y 6∈ f (∂U )} (10.7)
and C̄y (U, Rn ) := C̄y0 (U, Rn ).
Note that, since U is bounded, ∂U is compact and so is f (∂U ) if f ∈
C(U , Rn ). In particular,
dist(y, f (∂U )) = min |y − f (x)| (10.8)
x∈∂U

is positive for f ∈ C̄y (U, Rn ) and thus C̄y (U, Rn ) is an open subset of
C(U , Rn ).
Now that these things are out of the way, we come to the formulation of
the requirements for our degree.
A function deg which assigns each f ∈ C̄y (U, Rn ), y ∈ Rn , a real number
deg(f, U, y) will be called degree if it satisfies the following conditions.
(D1). deg(f, U, y) = deg(f − y, U, 0) (translation invariance).
(D2). deg(I, U, y) = 1 if y ∈ U (normalization).
(D3). If U1,2 are open, disjoint subsets of U such that y 6∈ f (U \(U1 ∪U2 )),
then deg(f, U, y) = deg(f, U1 , y) + deg(f, U2 , y) (additivity).
(D4). If H(t) = (1 − t)f + tg ∈ C̄y (U, Rn ), t ∈ [0, 1], then deg(f, U, y) =
deg(g, U, y) (homotopy invariance).
 282 10. The Brouwer mapping degree

Before we draw some first conclusions form this definition, let us discuss
the properties (D1)–(D4) first. (D1) is natural since deg(f, U, y) should
have something to do with the solutions of f (x) = y, x ∈ U , which is the
same as the solutions of f (x) − y = 0, x ∈ U . (D2) is a normalization
since any multiple of deg would also satisfy the other requirements. (D3)
is also quite natural since it requires deg to be additive with respect to
components. In addition, it implies that sets where f 6= y do not contribute.
(D4) is not that natural since it already rules out the case where deg is the
cardinality of f −1 ({y}). On the other hand it will give us the ability to
compute deg(f, U, y) in several cases.
Theorem 10.1. Suppose deg satisfies (D1)–(D4) and let f, g ∈ C̄y (U, Rn ),
then the following statements hold.
(i). We have deg(f, ∅, y) = 0. Moreover, if Ui , 1 ≤ i ≤ N , are disjoint
open subsets of U such that y 6∈ f (U \ N
S
PN i=1 Ui ), then deg(f, U, y) =
i=1 deg(f, Ui , y).
(ii). If y 6∈ f (U ), then deg(f, U, y) = 0 (but not the other way round).
Equivalently, if deg(f, U, y) 6= 0, then y ∈ f (U ).
(iii). If |f (x)−g(x)| < |f (x)−y|, x ∈ ∂U , then deg(f, U, y) = deg(g, U, y).
In particular, this is true if f (x) = g(x) for x ∈ ∂U .

Proof. For the first part of (i) use (D3) with U1 = U and U2 = ∅. For
the second part use U2 = ∅ in (D3) if N = 1 and the rest follows from
induction. For (ii) use N = 1 and U1 = ∅ in (i). For (iii) note that H(t, x) =
(1 − t)f (x) + t g(x) satisfies |H(t, x) − y| ≥ dist(y, f (∂U )) − |f (x) − g(x)| for
x on the boundary. 

Item (iii) is a version of Rouché’s theorem for our degree. Next we

show that (D4) implies several at first sight stronger looking facts.
Theorem 10.2. We have that deg(., U, y) and deg(f, U, .) are both continu-
ous. In fact, we even have
(i). deg(., U, y) is constant on each component of C̄y (U, Rn ).
(ii). deg(f, U, .) is constant on each component of Rn \ f (∂U ).
Moreover, if H : [0, 1] × U → Rn and y : [0, 1] → Rn are both con-
tinuous such that H(t) ∈ C̄y(t) (U, Rn ), t ∈ [0, 1], then deg(H(0), U, y(0)) =
deg(H(1), U, y(1)).

Proof. For (i) it suffices to show that deg(., U, y) is locally constant. But
if |g − f | < dist(y, f (∂U )), then deg(f, U, y) = deg(g, U, y) by (D4) since
|H(t) − y| ≥ |f − y| − |g − f | > 0, H(t) = (1 − t)f + t g. The proof of (ii) is
similar.
10.2. Definition of the mapping degree and the determinant formula 283

For the remaining part observe, that if H : [0, 1] × U → Rn , (t, x) 7→

H(t, x), is continuous, then so is H : [0, 1] → C(U , Rn ), t 7→ H(t), since U
is compact. Hence, if in addition H(t) ∈ C̄y (U, Rn ), then deg(H(t), U, y) is
independent of t and if y = y(t) we can use deg(H(0), U, y(0)) = deg(H(t) −
y(t), U, 0) = deg(H(1), U, y(1)). 

In this context note that a Banach space X is locally path-connected

and hence the components of any open subset are open (in the topology of
X) and path-connected (see Lemma B.32 (vi)).
Moreover, note that this result also shows why deg(f, U, y) cannot be de-
fined meaningful for y ∈ f (∂U ). Indeed, approaching y from within different
components of Rn \ f (∂U ) will result in different limits in general!
Now let us try to compute deg using its properties. If you are not in-
terested in how to derive the determinant formula for the degree from its
properties you can of course take it as a definition and skip to the next
section.
Let’s start with a simple case and suppose f ∈ C̄y1 (U, Rn ) and y 6∈ CV(f ).
Without restriction we consider y = 0. In addition, we avoid the trivial
case f −1 ({0}) = ∅. Since the points of f −1 ({0}) inside U are isolated (use
Jf (x) 6= 0 and the inverse function theorem) they can only cluster at the
boundary ∂U . But this is also impossible since f would equal 0 at the limit
point on the boundary by continuity. Hence f −1 ({0}) = {xi }N i=1 . Picking
sufficiently small neighborhoods U (x ) around x we consequently get
i i

N
X
deg(f, U, 0) = deg(f, U (xi ), 0). (10.9)
i=1

It suffices to consider one of the zeros, say x1 . Moreover, we can even assume
x1 = 0 and U (x1 ) = Bδ (0). Next we replace f by its linear approximation
around 0. By the definition of the derivative we have
f (x) = df (0)x + |x|r(x), r ∈ C(Bδ (0), Rn ), r(0) = 0. (10.10)
Now consider the homotopy H(t, x) = df (0)x + (1 − t)|x|r(x). In order
to conclude deg(f, Bδ (0), 0) = deg(df (0), Bδ (0), 0) we need to show 0 6∈
H(t, ∂Bδ (0)). Since Jf (0) 6= 0 we can find a constant λ such that |df (0)x| ≥
λ|x| and since r(0) = 0 we can decrease δ such that |r| < λ. This implies
|H(t, x)| ≥ ||df (0)x| − (1 − t)|x||r(x)|| ≥ λδ − δ|r| > 0 for x ∈ ∂Bδ (0) as
desired.
In summary we have
N
X
deg(f, U, 0) = deg(df (xi ), Bδ (0), 0) (10.11)
i=1
284 10. The Brouwer mapping degree

and it remains to compute the degree of a nonsingular matrix. To this end

we need the following lemma.
Lemma 10.3. Two nonsingular matrices M1,2 ∈ GL(n) are homotopic in
GL(n) if and only if sign det M1 = sign det M2 .

Proof. We will show that any given nonsingular matrix M is homotopic to

diag(sign det M, 1, . . . , 1), where diag(m1 , . . . , mn ) denotes a diagonal matrix
with diagonal entries mi .
In fact, note that adding one row to another and multiplying a row by
a positive constant can be realized by continuous deformations such that all
intermediate matrices are nonsingular. Hence we can reduce M to a diagonal
matrix diag(m1 , . . . , mn ) with (mi )2 = 1. Next,
 
± cos(πt) ∓ sin(πt)
,
sin(πt) cos(πt)
shows that diag(±1, 1) and diag(∓1, −1) are homotopic. Now we apply this
result to all two by two subblocks as follows. For each i starting from n
and going down to 2 transform the subblock diag(mi−1 , mi ) into diag(1, 1)
respectively diag(−1, 1). The result is the desired form for M .
To conclude the proof note that a continuous deformation within GL(n)
cannot change the sign of the determinant since otherwise the determinant
would have to vanish somewhere in between (i.e., we would leave GL(n)). 

Using this lemma we can now show the main result of this section.
Theorem 10.4. Suppose f ∈ C̄y1 (U, Rn ) and y 6∈ CV(f ), then a degree
satisfying (D1)–(D4) satisfies
X
deg(f, U, y) = sign Jf (x), (10.12)
x∈f −1 ({y})
P
where the sum is finite and we agree to set x∈∅ = 0.

Proof. By the previous lemma we obtain

deg(df (0), Bδ (0), 0) = deg(diag(sign Jf (0), 1, . . . , 1), Bδ (0), 0)
since det M 6= 0 is equivalent to M x 6= 0 for x ∈ ∂Bδ (0). Hence it remains to
show deg(M± , Bδ (0), 0) = ±1, where M± := diag(±1, 1, . . . , 1). For M+ this
is true by (D2) and for M− we note that we can replace Bδ (0) by any neigh-
borhood U of 0. Now abbreviate U1 := {x ∈ Rn ||xi | < 1, 1 ≤ i ≤ n}, U2 :=
{x ∈ Rn |1 < x1 < 3, |xi | < 1, 2 ≤ i ≤ n}, U := {x ∈ Rn | − 1 < x1 < 3, |xi | <
1, 2 ≤ i ≤ n}, and g(r) = 2−|r−1|, h(r) = 1−r2 . Consider the two functions
f1 (x) = (1 − g(x1 )h(x2 ) · · · h(xn ), x2 , . . . , xn ) and f2 (x) = (1, x2 , . . . , xn ).
Clearly f1−1 ({0}) = {x1 , x2 } with x1 = 0, x2 = (2, 0, . . . , 0) and f2−1 (0) = ∅.
Since f1 (x) = f2 (x) for x ∈ ∂U we infer deg(f1 , U, 0) = deg(f2 , U, 0) = 0.
10.3. Extension of the determinant formula 285

Moreover, we have deg(f1 , U, 0) = deg(f1 , U1 , 0) + deg(f1 , U2 , 0) and hence

deg(M− , U1 , 0) = deg(df1 (x1 ), U1 , 0) = deg(f1 , U1 , 0) = − deg(f1 , U2 , 0) =
− deg(df1 (x2 ), U1 , 0) = − deg(I, U1 , 0) = −1 as claimed. 

Up to this point we have only shown that a degree (provided there is one
at all) necessarily satisfies (10.12). Once we have shown that regular values
are dense, it will follow that the degree is uniquely determined by (10.12)
since the remaining values follow from point (iii) of Theorem 10.1. On the
other hand, we don’t even know whether a degree exists since it is unclear
whether (10.12) satisfies (D4). Hence we need to show that (10.12) can be
extended to f ∈ C̄y (U, Rn ) and that this extension satisfies our requirements
(D1)–(D4).

10.3. Extension of the determinant formula

Our present objective is to show that the determinant formula (10.12) can
be extended to all f ∈ C̄y (U, Rn ). As a preparation we prove that the set
of regular values is dense. This is a consequence of a special case of Sard’s
theorem which says that CV(f ) has zero measure.

Lemma 10.5 (Sard). Suppose f ∈ C 1 (U, Rn ), then the Lebesgue measure of

CV(f ) is zero.

Proof. Since the claim is easy for linear mappings our strategy is as follows.
We divide U into sufficiently small subsets. Then we replace f by its linear
approximation in each subset and estimate the error.
Let CP(f ) := {x ∈ U |Jf (x) = 0} be the set of critical points of f . We
first pass to cubes which are easier to divide. Let {Qi }i∈N be a countable
cover for U consisting of open cubes such that Qi ⊂ U . Then it suffices
to prove that f (CP(f ) ∩ Qi ) has zero measure since CV(f ) = f (CP(f )) =
i f (CP(f ) ∩ Qi ) (the Qi ’s are a cover).
S

Let Q be anyone of these cubes and denote by ρ the length of its edges.
Fix ε > 0 and divide Q into N n cubes Qi of length ρ/N . These cubes don’t
have to be open and hence we can assume that they cover Q. Since df (x) is
uniformly continuous on Q we can find an N (independent of i) such that
Z 1
ερ
|f (x) − f (x̃) − df (x̃)(x − x̃)| ≤ |df (x̃ + t(x − x̃)) − df (x̃)||x̃ − x|dt ≤
0 N
(10.13)
for x̃, x ∈ Qi . Now pick a Qi which contains a critical point x̃i ∈ CP(f ).
Without restriction we assume x̃i = 0, f (x̃i ) = 0 and set M := df (x̃i ). By
det M = 0 there is an orthonormal basis {bi }1≤i≤n of Rn such that bn is
286 10. The Brouwer mapping degree

orthogonal to the image of M . In addition,

v
n u n
X
i t
uX √ ρ
Qi ⊆ { λi b | |λi |2 ≤ n }
N
i=1 i=1

and hence there is a constant (again independent of i) such that

n−1
X √ ρ
M Qi ⊆ { λi bi | |λi | ≤ C n }
N
i=1

(e.g., C := maxx∈Q |df (x)|). Next, by our estimate (10.13) we even have
n
X √ ρ √ ρ
f (Qi ) ⊆ { λi bi | |λi | ≤ (C + ε) n , |λn | ≤ ε n }
N N
i=1

and hence the measure of f (Qi ) is smaller than N n . Since there are at most
C̃ε

N such Qi ’s, we see that the measure of f (CP(f ) ∩ Q) is smaller than

n

C̃ε. 

By (ii) of Theorem 10.2, deg(f, U, y) should be constant on each com-

ponent of Rn \ f (∂U ). Unfortunately, if we connect y and a nearby regular
value ỹ by a path, then there might be some critical values in between.
Example 10.1. The function f (x) := x2 sin( 2x π
) is in C̄01 ([−1, 1], R). It has
0 as a critical value and the critical values accumulate at 0. 
To overcome this problem we need a definition for deg which works for
critical values as well. Let us try to look for an integral representation. For-
mally (10.12) can be written as deg(f, U, y) = U δy (f (x))Jf (x)dn x, where
R

δy (.) is the Dirac distribution at y. But since we don’t want to mess with
distributions, we replace δy (.) by φε (. − y), where {φε }ε>0 is a family of
functions such R that φε is supported on the ball Bε (0) of radius ε around 0
and satisfies Rn φε (x)dn x = 1.
Lemma 10.6 (Heinz). Suppose f ∈ C̄y1 (U, Rn ) and y 6∈ CV(f ). Then the
degree defined as in (10.12) satisfies
Z
deg(f, U, y) = φε (f (x) − y)Jf (x)dn x (10.14)
U
for all positive ε smaller than a certain ε0 depending on f and y. Moreover,
supp(φε (f (.) − y)) ⊂ U for ε < dist(y, f (∂U )).

Proof. If f −1 ({y}) = ∅, we can set ε0 = dist(y, f (U )), implying φε (f (x) −

y) = 0 for x ∈ U .
If f −1 ({y}) = {xi }1≤i≤N , the inverse function theorem ensures that we
can find an ε0 > 0 such that f −1 (Bε0 (y)) is a union of disjoint neighborhoods
10.3. Extension of the determinant formula 287

U (xi ) of xi with f |U (xi ) a bijection and Jf (x) nonzero on U (xi ). Again

φε (f (x) − y) = 0 for x ∈ U \ N i=1 U (x ) and hence
i
S

Z N Z
X
φε (f (x) − y)Jf (x)dn x = φε (f (x) − y)Jf (x)dn x
U i=1 U (xi )
N
X Z
= sign(Jf (xi )) φε (x̃)dn x̃ = deg(f, U, y),
i=1 Bε0 (0)

where we have used the change of variables x̃ = f (x) − y in the second

step. 

Our new integral representation makes sense even for critical values. But
since ε0 depends on f and y, continuity is not clear. This will be tackled
next.
The key idea is to show that the integral representation is independent
of ε as long as ε < dist(y, f (∂U )). To this end we will rewrite the difference
as an integral over a divergence supported in U and then apply the Gauss–
Green theorem. For this purpose the following result will be used.

Lemma 10.7. Suppose f ∈ C 2 (U, Rn ) and u ∈ C 1 (Rn , Rn ), then

(div u)(f )Jf = div Df (u), (10.15)

where Df (u)j is the determinant of the matrix

Pn obtained from df by replacing
the j-th column by u(f ). Here div u = j=1 ∂j uj is the divergence of a
vector field.

Proof. We compute
n
X n
X
div Df (u) = ∂xj Df (u)j = Df (u)j,k ,
j=1 j,k=1

where Df (u)j,k is the determinant of the matrix obtained from the matrix
associated with Df (u)j by applying ∂xj to the k-th column. Since ∂xj ∂xk f =
∂xk ∂xj f we infer Df (u)j,k = −Df (u)k,j , j 6= k, by exchanging the k-th and
the j-th column. Hence
n
X
div Df (u) = Df (u)i,i .
i=1

(i,j)
Now let Jf (x) denote the (i, j) cofactor of df (x) and recall the cofactor
(i,j)
expansion of the determinant ni=1 Jf ∂xi fk = δj,k Jf . Using this to expand
P
288 10. The Brouwer mapping degree

the determinant Df (u)i,i along the i-th column shows

n n n
(i,j) (i,j)
X X X
div Df (u) = Jf ∂xi uj (f ) = Jf (∂xk uj )(f )∂xi fk
i,j=1 i,j=1 k=1
n n n
(i,j)
X X X
= (∂xk uj )(f ) Jf ∂xi fk = (∂xj uj )(f )Jf
j,k=1 i=1 j=1

as required. 

Now we can prove

Theorem 10.8. There is a unique degree deg satisfying (D1)–(D4). More-
over, deg(., U, y) : C̄y (U, Rn ) → Z is constant on each component and given
f ∈ C̄y (U, Rn ) we have
X
deg(f, U, y) = sign Jf˜(x) (10.16)
x∈f˜−1 (y)

where f˜ ∈ C̄y1 (U, Rn ) is in the same component of C̄y (U, Rn ), say kf − f˜k∞ <
dist(y, f (∂U )), such that y ∈ RV(f˜).

Proof. We will first show that our integral formula works in fact for all
ε < ρ := dist(y, f (∂U )). For this we will make some additional assumptions:
Let f ∈ C̄ 2 (U, Rn ) and choose Ra family of functions φε ∈ C ∞ ((0, ∞)) with
ε
supp(φε ) ⊂ (0, ε) such that Sn 0 φ(r)rn−1 dr = 1. Consider
Z
Iε (f, U, y) := φε (|f (x) − y|)Jf (x)dn x.
U
Then I := Iε1 − Iε2 will be of the same form but with φε replaced
Rρ by ϕ :=
φε1 −φε2 , where ϕ ∈ C ∞ ((0, ∞)) with supp(ϕ) ⊂ (0, ρ) and 0 ϕ(r)rn−1 dr =
0. To show that I = 0 we will use our previous lemma with u chosen such
that div(u(x)) = ϕ(|x|). To this end we make the ansatz u(x) = ψ(|x|)x
such that div(u(x)) = |x|ψ 0 (|x|) + n ψ(|x|). Our requirement now leads to
an ordinary differential equation whose solution is
Z r
1
ψ(r) = n sn−1 ϕ(s)ds.
r 0
Moreover, one checks ψ ∈ C ∞ ((0, ∞)) with supp(ψ) ⊂ (0, ρ). Thus our
lemma shows Z
I= div Df −y (u)dn x
U
and since the integrand vanishes in a neighborhood of ∂U we can extend it
to all of Rn by setting it zero outside U and choose a cube Q ⊃ U . Then
elementary coordinatewise integration gives I = Q div Df −y (u)dn x = 0.
R
10.3. Extension of the determinant formula 289

Now fix δ < ρ and look at Iε (f + g, U, y) for g ∈ Bδ (f ) ∩ C̄ 2 (U, Rn ) ⊂

C(U , Rn ) and ε < ρ − δ < dist((f + g)(∂U ), y) fixed. Then t 7→ Iε (f +
t g, U, y), t ∈ [0, 1], is continuous and it is integer valued (since it is equal to
our determinant formula) on a dense set. Consequently it must be constant
and we can extend Iε to a function I¯ε on all of Bδ (f ) and hence on all of
C̄y (U, Rn ) . Note that by mollifying f ∈ C̄ 1 (U, Rn ) we get a sequence of
smooth functions for which both f and df converge uniformly on compact
subsets of U and hence Iε converges, such that for such f we still have
I¯ε (f, U, y) = Iε (f, U, y).
Now we set
deg(f, U, y) := Iε (f˜, U, y),
where f˜ ∈ C̄ 1 (U, Rn ) with ε < ρ and |f˜ − f | < ρ − ε. Then (D1) holds since
it holds for Iε , (D2) holds since Iε extends the determinant formula, (D3)
holds since the integrand of Iε vanishes on U \ (U1 ∪ U2 ), and (D4) holds
since we can choose ε < mint∈[0,1] dist(H(t)(∂U ), y) such that Iε (H(t), U, y)
is continuous and hence constant for t ∈ [0, 1]. 

To conclude this section, let us give a few simple examples illustrating

the use of the Brouwer degree.
Example 10.2. First, let’s investigate the zeros of
f (x1 , x2 ) := (x1 − 2x2 + cos(x1 + x2 ), x2 + 2x1 + sin(x1 + x2 )).
Denote the linear part by
g(x1 , x2 ) := (x1 − 2x2 , x2 + 2x1 ).
√
Then we have |g(x)| = 5|x| and |f (x) − g(x)| = 1 and hence h(t)√=
(1 − t)g + t f = g + t(f − g) satisfies |h(t)| ≥ |g| − t|f − g| > 0 for |x| > 1/ 5
implying
√
deg(f, Br (0), 0) = deg(g, Br (0), 0) = 1, r > 1/ 5.
Moreover, since Jf (x) = 5+3 cos(x1 +x2 )+sin(x1 +x2 ) > 1 the determinant
formula (10.12) for the degree implies that f (x) = 0 has√a unique solution
in R2 . This solution even has to lie on
√ the circle |x| = 1/ 5 since f (x) = 0
implies 1 = |f (x) − g(x)| = |g(x)| = 5|x|. 
Next let us prove the following result which implies the hairy ball (or
hedgehog) theorem.
Theorem 10.9. Suppose U is open, bounded and contains the origin and let
f : ∂U → Rn \ {0} be continuous. If n is odd, then there exists an x ∈ ∂U
and a λ 6= 0 such that f (x) = λx.

Proof. By Theorem 11.10 we can assume f ∈ C(U , Rn ) and since n is odd

we have deg(−I, U, 0) = −1. Now if deg(f, U, 0) 6= −1, then H(t, x) =
290 10. The Brouwer mapping degree

(1 − t)f (x) − tx must have a zero (t0 , x0 ) ∈ (0, 1) × ∂U and hence f (x0 ) =
1−t0 x0 . Otherwise, if deg(f, U, 0) = −1 we can apply the same argument to
t0

H(t, x) = (1 − t)f (x) + tx. 

In particular, this result implies that a continuous tangent vector field

on the unit sphere f : S n−1 → Rn (with f (x)x = 0 for all x ∈ S n−1 ) must
vanish somewhere if n is odd. Or, for n = 3, you cannot smoothly comb
a hedgehog without leaving a bald spot or making a parting. It is however
possible to comb the hair smoothly on a torus and that is why the magnetic
containers in nuclear fusion are toroidal.
Example 10.3. The result fails in even dimensions as the example n = 2,
U = B1 (0), f (x1 , x2 ) = (−x2 , x1 ) shows. 
Another illustration is the fact that a vector field on Rn , which points
outwards (or inwards) on a sphere, must vanish somewhere inside the sphere
(Problem 10.2).
One more useful observation is that odd functions have odd degree:
Theorem 10.10 (Borsuk). Let 0 ∈ U ⊆ Rn be open, bounded and symmetric
with respect to the origin (i.e., U = −U ). Let f ∈ C̄0 (U, Rn ) be odd (i.e.,
f (−x) = −f (x)). Then deg(f, U, 0) is odd.

Proof. If f ∈ C̄01 (U ) and 0 ∈ RV(f ), then the claim is straightforward since

X
deg(f, U, 0) = sign Jf (0) + sign Jf (x),
x∈f −1 (0)\{0}

where the sum is even since for every x ∈ f −1 (0) \ {0} we also have −x ∈
f −1 (0) \ {0} as well as Jf (x) = Jf (−x).
Hence we need to reduce the general case to this one. Clearly if f ∈
C̄0 (U, Rn ) we can choose an approximating f0 ∈ C̄01 (U, Rn ) and replacing f0
by its odd part 12 (f0 (x) − f0 (−x)) we can assume f0 to be odd. Moreover, if
Jf0 (0) = 0 we can replace f0 by f0 (x) + δx such that 0 is regular. However,
if we choose a nearby regular value y and consider f0 (x) − y we have the
problem that constant functions are even. Hence we will try the next best
thing and perturb by a function which is constant in all except one direction.
To this end we choose an odd function ϕ ∈ C 1 (R) such that ϕ0 (0) = 0 (since
we don’t want to alter the behavior at 0) and ϕ(t) 6= 0 for t 6= 0. Now we
consider f1 (x) = f0 (x) − ϕ(x1 )y 1 and note
f0 (x) f0 (x)

df1 (x) = df0 (x) − dϕ(x1 )y 1 = df0 (x) − dϕ(x1 ) = ϕ(x1 )d
ϕ(x1 ) ϕ(x1 )
for every x ∈ U1 := {x ∈ U |x1 6= 0} with f1 (x) = 0. Hence if y 1 is chosen
f0 (x)
such that y 1 ∈ RV(h1 ), where h1 : U1 → Rn , x 7→ ϕ(x 1)
, then 0 will be
10.3. Extension of the determinant formula 291

a regular value for f1 when restricted to V1 := U1 . Now we repeat this

procedure and consider f2 (x) = f1 (x) − ϕ(x2 )y 2 with y 2 ∈ RV(h2 ) as before.
Then every point x ∈ V2 := U1 ∪ U2 with f2 (x) = 0 either satisfies x2 6= 0
and thus is regular by our choice of y 2 or satisfies x2 = 0 and thus is regular
since it is in V1 and df2 (x) = df1 (x) by our assumption φ0 (0) = 0. After
n steps we reach Vn = U \ {0} and fn is the approximation we are looking
for. 

At first sight the obvious conclusion that an odd function has a zero
does not seem too spectacular since the fact that f is odd already implies
f (0) = 0. However, the result gets more interesting upon observing that it
suffices when the boundary values are odd. Moreover, local constancy of the
degree implies that f does not only attain 0 but also any y in a neighborhood
of 0. The next two important consequences are based on this observation:

Theorem 10.11 (Borsuk–Ulam). Let 0 ∈ U ⊆ Rn be open, bounded and

symmetric with respect to the origin. Let f ∈ C(∂U, Rm ) with m < n. Then
there is some x ∈ ∂U with f (x) = f (−x).

Proof. Consider g(x) = f (x) − f (−x) and extend it to a continuous odd

function U → Rn (extend the domain by Tietze and then take the odd part,
finally fill up the missing coordinates by setting them equal to 0). If g does
not vanish on ∂U , we get that deg(g, U, y) = deg(g, U, 0) 6= 0 for y in a
neighborhood of 0 and thus the image of g contains a neighborhood of 0 (in
Rn ), which contradicts the fact that the image is in Rm × {0} ⊂ Rn . 

This theorem is often illustrated by the fact that there are always two
opposite points on the earth which have the same weather (in the sense that
they have the same temperature and the same pressure). In a similar manner
one can also derive the invariance of domain theorem.

Theorem 10.12 (Brouwer). Let U ⊆ Rn be open and let f : U → Rn be

continuous and locally injective. Then f (U ) is also open.

Proof. It suffices to show that every point x ∈ U contains a neighborhood

Br (x) such that the image f (Br (x)) contains a ball centered at f (x). By
simple translations we can assume x = 0 as well as f (x) = 0. Now choose
r sufficiently small such that f restricted to B̄r (0) is injective and consider
1
H(t, x) := f ( 1+t t
x) − f (− 1+t x) for t ∈ [0, 1] and x ∈ B̄r (0). Moreover,
if H(t, x) = 0 then by injectivity 1+t 1
x = − 1+tt
x, that is, x = 0. Thus
deg(f, Br (0), 0) = deg(H(1), Br (0), 0) 6= 0 since H(1) = f ( 12 x) − f (− 21 x)
is odd. But then we also have deg(f, Br (0), y) 6= 0 for y ∈ Bε (0) and thus
Bε (0) ⊆ f (Br (0)). 
292 10. The Brouwer mapping degree

An easy consequence worth while noting is the topological invariance of

dimension:
Corollary 10.13. If m < n and U is a nonempty open subset of Rn , then
there is no continuous injective mapping from U to Rm .

Proof. Suppose there where such a map and extend it to a map from U to
Rn by setting the additional coordinates equal to zero. The resulting map
contradicts the invariance of domain theorem. 

In particular, Rm and Rn are not homeomorphic for m 6= n.

Problem 10.1. Suppose U = (a, b) ⊂ R1 . Show
1

deg(f, (a, b), y) = sign(f (b) − y) − sign(f (a) − y) .
2
In particular, our degree reduces to the intermediate value theorem in this
case.
Problem* 10.2. Suppose f : B̄r (0) → Rn is continuous and satisfies
f (x)x > 0, |x| = r.
Then f (x) vanishes somewhere inside Br (0).
Problem 10.3. Show that in Borsuk’s theorem the condition f is odd can
be replaced by f (x) 6= tf (−x) for all x ∈ ∂U and t ∈ (0, 1]. Note that this
condition will hold if sign(f (x)) 6= sign(f (−x)), x ∈ ∂U (where sign(f (x)) :=
f (x)
|f (x)| ).

10.4. The Brouwer fixed point theorem

Now we can show that the famous Brouwer fixed point theorem is a simple
consequence of the properties of our degree.
Theorem 10.14 (Brouwer fixed point). Let K be a topological space home-
omorphic to a compact, convex subset of Rn and let f ∈ C(K, K), then f
has at least one fixed point.

Proof. Clearly we can assume K ⊂ Rn since homeomorphisms preserve

fixed points. Now lets assume K = B̄r (0). If there is a fixed point on
the boundary ∂Br (0) we are done. Otherwise H(t, x) = x − t f (x) satisfies
0 6∈ H(t, ∂Br (0)) since |H(t, x)| ≥ |x| − t|f (x)| ≥ (1 − t)r > 0, 0 ≤ t < 1.
And the claim follows from deg(x − f (x), Br (0), 0) = deg(x, Br (0), 0) = 1.
Now let K be convex. Then K ⊆ Bρ (0) and, by the Hilbert pro-
jection theorem (Theorem 2.11) (or alternatively by the Tietze extension
theorem or its variant Theorem 11.10 below), we can find a continuous
retraction R : Rn → K (i.e., R(x) = x for x ∈ K) and consider f˜ =
10.4. The Brouwer fixed point theorem 293

f ◦ R ∈ C(B̄ρ (0), B̄ρ (0)). By our previous analysis, there is a fixed point
x = f˜(x) ∈ conv(f (K)) ⊆ K. 

Note that any compact, convex subset of a finite dimensional Banach

space (complex or real) is isomorphic to a compact, convex subset of Rn since
linear transformations preserve both properties. In addition, observe that all
assumptions are needed. For example, the map f : R → R, x 7→ x+1, has no
fixed point (R is homeomorphic to a bounded set but not to a compact one).
The same is true for the map f : ∂B1 (0) → ∂B1 (0), x 7→ −x (∂B1 (0) ⊂ Rn
is simply connected for n ≥ 3 but not homeomorphic to a convex set).
As an easy example of how to use the Brouwer fixed point theorem we
show the famous Perron–Frobenius theorem.

Theorem 10.15 (Perron–Frobenius). Let A be an n × n matrix all whose

entries are nonnegative and there is an m such the entries of Am are all
positive. Then A has a positive eigenvalue and the corresponding eigenvector
can be chosen to have positive components.

Proof. We equip Rn with the norm |x|1 := nj=1 |xj | and set ∆ := {x ∈
P

Rn |xj ≥ 0, |x|1 = 1}. For x ∈ ∆ we have Ax 6= 0 (since Am x 6= 0) and hence

Ax
f : ∆ → ∆, x 7→
|Ax|1

has a fixed point x0 by the Brouwer fixed point theorem. Then Ax0 =
|Ax0 |1 x0 and x0 has positive components since Am x0 = |Ax0 |m
1 x0 has. 

Let me remark that the Brouwer fixed point theorem is equivalent to

the fact that there is no continuous retraction R : B1 (0) → ∂B1 (0) (with
R(x) = x for x ∈ ∂B1 (0)) from the unit ball to the unit sphere in Rn .
In fact, if R would be such a retraction, −R would have a fixed point
x0 ∈ ∂B1 (0) by Brouwer’s theorem. But then x0 = −R(x0 ) = −x0 which is
impossible. Conversely, if a continuous function f : B1 (0) → B1 (0) has no
fixed point we can define a retraction R(x) = f (x) + t(x)(x − f (x)), where
t(x) ≥ 0 is chosen such that |R(x)|2 = 1 (i.e., R(x) lies on the intersection
of the line spanned by x, f (x) with the unit sphere).
Using this equivalence the Brouwer fixed point theorem can also be de-
rived easily by showing that the homology groups of the unit ball B1 (0) and
its boundary (the unit sphere) differ (see, e.g., [37] for details).
294 10. The Brouwer mapping degree

10.5. Kakutani’s fixed point theorem and applications to

game theory
In this section we want to apply Brouwer’s fixed point theorem to show the
existence of Nash equilibria for n-person games. As a preparation we extend
Brouwer’s fixed point theorem to set-valued functions.
Denote by CS(K) the set of all nonempty convex subsets of K.
Theorem 10.16 (Kakutani). Suppose K is a compact convex subset of Rn
and f : K → CS(K). If the set
Γ := {(x, y)|y ∈ f (x)} ⊆ K 2 (10.17)
is closed, then there is a point x ∈ K such that x ∈ f (x).

Proof. Our strategy is to apply Brouwer’s theorem, hence we need a func-

tion related to f . For this purpose it is convenient to assume that K is a
simplex
K = conv(v1 , . . . , vm ), m ≤ n + 1,
where vi are the vertices. Recall that each point x ∈ K can Pmbe uniquely
represented
P by its barycentric coordinates λi (x) (i.e., λi ≥ 0, i=1 λi (x) = 1
and x = m i=1 λi vi ). Now if we pick yi ∈ f (vi ) we could set
m
X
f 1 (x) = λi (x)yi .
i=1

By construction, f 1 ∈ C(K, K) and there is a fixed point x1 . But unless x1

is one of the vertices, this doesn’t help us too much. So lets choose a bet-
ter function as follows. Consider the k-th barycentric subdivision, that is,
for every permutation vσ1 , . . . , vσm of the vertices you consider the simplex
conv(vσ1 , 21 (vσ1 + vσ2 ), . . . , m
1
(vσ1 + · · · + vσm )). This gives you m! smaller
simplices (note that the maximal distance between vertices of the subsim-
plices decreases by a factor m−1 m during the subdivision) whose union is the
simplex you have started with. Now repeat this construction k times.
v3

v1 v2
For each vertex vi in this subdivision pick an element yi ∈ f (vi ). Now de-
fine f k (vi ) = yi and extend f k to the interior of each subsimplex as before.
10.5. Kakutani’s fixed point theorem and applications to game theory 295

Hence f k ∈ C(K, K) and there is a fixed point xk in one of the subsimplices.

Denote this subsimplex by conv(v1k , . . . , vm
k ) such that

m
X m
X
xk = λki vik = λki yik , yik = f k (vik ). (10.18)
i=1 i=1

Since ∈ K ×[0, 1]m ×K m we can assume that this

(xk , λk1 , . . . , λkm , y1k , . . . , ym
k)

sequence converges to some limit (x0 , λ01 , . . . , λ0m , y10 , . . . , ym

0 ) after passing to

a subsequence. Since the subsimplices shrink to a point, this implies vik → x0

and hence yi0 ∈ f (x0 ) since (vik , yik ) ∈ Γ → (vi0 , yi0 ) ∈ Γ by the closedness
assumption. Now (10.18) tells us
m
X
0
x = λ0i yi0 ∈ f (x0 )
i=1

since f (x0 ) is convex and the claim holds if K is a simplex.

If K is not a simplex, we can pick a simplex S containing K and proceed
as in the proof of the Brouwer theorem. 

If f (x) contains precisely one point for all x, then Kakutani’s theorem
reduces to the Brouwer’s theorem (show that the closedness of Γ is equivalent
to continuity of f ).
Now we want to see how this applies to game theory.
An n-person game consists of n players who have mi possible actions to
choose from. The set of all possible actions for the i-th player will be denoted
by Φi = {1, . . . , mi }. An element ϕi ∈ Φi is also called a pure strategy for
reasons to become clear in a moment. Once all players have chosen their
move ϕi , the payoff for each player is given by the payoff function

n
Ri (ϕ) ∈ R, ϕ = (ϕ1 , . . . , ϕn ) ∈ Φ = Φi (10.19)
i=1

of the i-th player. We will consider the case where the game is repeated a
large number of times and where in each step the players choose their action
according to a fixed strategy. Here a strategy si for the i-th player is a
probability distribution on Φi , that is, si = (s1i , . . . , sm
i ) such that si ≥ 0
i k
Pmi k
and k=1 si = 1. The set of all possible strategies for the i-th player is
denoted by Si . The number ski is the probability for the k-th pure strategy

to be chosen. Consequently, if s = (s1 , . . . , sn ) ∈ S = ni=1 Si is a collection
of strategies, then the probability that a given collection of pure strategies
gets chosen is
n
Y
s(ϕ) = si (ϕ), si (ϕ) = ski i , ϕ = (k1 , . . . , kn ) ∈ Φ (10.20)
i=1
296 10. The Brouwer mapping degree

(assuming all players make their choice independently) and the expected
payoff for player i is X
Ri (s) = s(ϕ)Ri (ϕ). (10.21)
ϕ∈Φ
By construction, Ri : S → R is polynomial and hence in particular continu-
ous.
The question is of course, what is an optimal strategy for a player? If
the other strategies are known, a best reply of player i against s would be
a strategy si satisfying
Ri (s \ si ) = max Ri (s \ s̃i ) (10.22)
s̃i ∈Si

Here s \ s̃i denotes the strategy combination obtained from s by replacing si

by s̃i . The set of all best replies against s for the i-th player is denoted by
Bi (s). Since
mi
X
Ri (s) = ski Ri (s/k) (10.23)
k=1
we have si ∈ Bi (s) if and only if = 0 whenever Ri (s\k) < max1≤l≤mi Ri (s\
ski
l). In particular, since there are no restrictions on the other entries, Bi (s) is
a nonempty convex set.
Let s, s ∈ S, we call s a best reply against s if si is a best reply against

s for all i. The set of all best replies against s is B(s) = ni=1 Bi (s).
A strategy combination s ∈ S is a Nash equilibrium for the game if it
is a best reply against itself, that is,
s ∈ B(s). (10.24)
Or, put differently, s is a Nash equilibrium if no player can increase his
payoff by changing his strategy as long as all others stick to their respective
strategies. In addition, if a player sticks to his equilibrium strategy, he is
assured that his payoff will not decrease no matter what the others do.
To illustrate these concepts, let us consider the famous prisoner’s dilemma.
Here we have two players which can choose to defect or to cooperate. The
payoff is symmetric for both players and given by the following diagram
R1 d2 c2 R2 d2 c2
d1 0 2 d1 0 −1 (10.25)
c1 −1 1 c1 2 1
where ci or di means that player i cooperates or defects, respectively. You
should think of two prisoners who are offered a reduced sentence if they
testify against the other.
It is easy to see that the (pure) strategy pair (d1 , d2 ) is the only Nash
equilibrium for this game and that the expected payoff is 0 for both players.
10.6. Further properties of the degree 297

Of course, both players could get the payoff 1 if they both agree to cooperate.
But if one would break this agreement in order to increase his payoff, the
other one would get less. Hence it might be safer to defect.
Now that we have seen that Nash equilibria are a useful concept, we
want to know when such an equilibrium exists. Luckily we have the following
result.
Theorem 10.17 (Nash). Every n-person game has at least one Nash equi-
librium.

Proof. The definition of a Nash equilibrium begs us to apply Kakutani’s

theorem to the set-valued function s 7→ B(s). First of all, S is compact and
convex and so are the sets B(s). Next, observe that the closedness condition
of Kakutani’s theorem is satisfied since if sm ∈ S and sm ∈ B(sn ) both
converge to s and s, respectively, then (10.22) for sm , sm
Ri (sm \ s̃i ) ≤ Ri (sm \ sm
i ), s̃i ∈ Si , 1 ≤ i ≤ n,
implies (10.22) for the limits s, s
Ri (s \ s̃i ) ≤ Ri (s \ si ), s̃i ∈ Si , 1 ≤ i ≤ n,
by continuity of Ri (s). 

10.6. Further properties of the degree

We now prove some additional properties of the mapping degree. The first
one will relate the degree in Rn with the degree in Rm . It will be needed
later on to extend the definition of degree to infinite dimensional spaces. By
virtue of the canonical embedding Rm ,→ Rm × {0} ⊂ Rn we can consider
Rm as a subspace of Rn . We can project Rn to Rm by setting the last n − m
coordinates equal to zero.
Theorem 10.18 (Reduction property). Let U ⊆ Rn be open and bounded,
f ∈ C(U , Rm ) and y ∈ Rm \ (I + f )(∂U ), then
deg(I + f, U, y) = deg(I + fm , Um , y), (10.26)
where fm = f |Um , where Um is the projection of U to Rm .

Proof. After perturbing f a little, we can assume f ∈ C 1 (U, Rm ) without

loss of generality. Let x ∈ (I + f )−1 ({y}), then x = y − f (x) ∈ Rm implies
(I + f )−1 ({y}) = (I + fm )−1 ({y}). Moreover,
 
δ + ∂j fi (x) ∂j fj (x)
JI+f (x) = det(I + df )(x) = det ij
0 δij
= det(δij + ∂j fi ) = JI+fm (x)
298 10. The Brouwer mapping degree

So if y ∈ RV(I+fm ) we immediately get deg(I+f, U, y) = deg(I+fm , Um , y)

as desired. Otherwise, if y ∈ CV(I + fm ) we can choose some ỹ ∈ RV(I +
fm ) (and hence also ỹ ∈ RV(I + f ) by the first part) with |y − ỹ| <
min(dist(y, f (∂U )), dist(y, fm (∂Um ))) and the claim follows from the reg-
ular case. 

Let U ⊆ Rn and f ∈ C(U , Rn ) be as usual. By Theorem 10.2 we

know that deg(f, U, y) is the same for every y in a connected component of
Rn \ f (∂U ). Since Rn \ f (∂U ) is open and locally path connected, these
components are open. We will denote these components by Gj and write
deg(f, U, Gj ) := deg(f, U, y) if y ∈ Gj . In this context observe that since
f (∂U ) is compact any unbounded component (there will be two for n = 1
and one for n > 1) will have degree zero.

Theorem 10.19 (Product formula). Let U ⊆ Rn be a bounded and open

set and denote by Gj the connected components of Rn \ f (∂U ). If g ◦ f ∈
C̄y (U, Rn ), then

X
deg(g ◦ f, U, y) = deg(f, U, Gj ) deg(g, Gj , y), (10.27)
j

where only finitely many terms in the sum are nonzero (and in particu-
lar, summands corresponding to unbounded components are considered to
be zero).

Proof. Since y 6∈ (g ◦ f )(∂U ) we have g −1 ({y}) ∩ f (∂U ) = ∅, that is,

g −1 ({y}) ⊂ j Gj . Moreover, since f (U ) is compact, we can find an r > 0
S

such that f (U ) ⊆ Br (0). Moreover, since g −1 ({y}) is closed, g −1 ({y})∩Br (0)

is compact andS hence can be covered by finitely many components, say
g −1 ({y}) ⊂ m j=1 Gj . In particular, the others will be either unbounded
or have deg(f, Gk , y) = 0 and hence only finitely many terms in the above
sum are nonzero.
We begin by computing deg(g ◦ f, U, y) in the case where f, g ∈ C 1
and y 6∈ CV(g ◦ f ). Since d(g ◦ f )(x) = dg(f (x)) ◦ df (x) the claim is a
10.6. Further properties of the degree 299

straightforward calculation
X
deg(g ◦ f, U, y) = sign(Jg◦f (x))
x∈(g◦f )−1 ({y})
X
= sign(Jg (f (x))) sign(Jf (x))
x∈(g◦f )−1 ({y})
X X
= sign(Jg (z)) sign(Jf (x))
z∈g −1 ({y}) x∈f −1 ({z}))
X
= sign(Jg (z)) deg(f, U, z)
z∈g −1 ({y})

and, using our cover {Gj }m

j=1 ,
m
X X
deg(g ◦ f, U, y) = sign(Jg (z)) deg(f, U, z)
j=1 z∈g −1 ({y})∩Gj
m
X X
= deg(f, U, Gj ) sign(Jg (z))
j=1 z∈g −1 ({y})∩Gj
m
X
= deg(f, U, Gj ) deg(g, Gj , y).
j=1

Moreover, this formula still holds for y ∈ CV(g ◦ f ) and for g ∈ C by

construction of the Brouwer degree. However, the case f ∈ C will need a
closer investigation since the components Gj depend on f . To overcome this
problem we will introduce the sets
Ll := {z ∈ Rn \ f (∂U )| deg(f, U, z) = l}.
Observe that Ll , l 6= 0, must be a union of some sets from {Gj }m
j=1 , that is,
Ll = m
Sm
and .
S l S
G
k=1 j l L
l6=0 l = G
j=1 j
k

Now choose f˜ ∈ C 1 such that |f (x) − f˜(x)| < 2−1 dist(g −1 ({y}), f (∂U ))
for x ∈ U and define G̃j , L̃l accordingly. Then we have Ll ∩ g −1 ({y}) =
L̃l ∩ g −1 ({y}) by Theorem 10.1 (iii) and hence deg(g, L̃l , y) = deg(g, Ll , y)
by Theorem 10.1 (i) implying
m̃
X
deg(g ◦ f, U, y) = deg(g ◦ f˜, U, y) = deg(f˜, U, G̃j ) deg(g, G̃j , y)
j=1
X X
= l deg(g, L̃l , y) = l deg(g, Ll , y)
l6=0 l6=0
ml
XX m
X
= l deg(g, Gj l , y) = deg(f, U, Gj ) deg(g, Gj , y)
k
l6=0 k=1 j=1
300 10. The Brouwer mapping degree

which proves the claim. 

10.7. The Jordan curve theorem

In this section we want to show how the product formula (10.27) for the
Brouwer degree can be used to prove the famous Jordan curve theo-
rem which states that a homeomorphic image of the circle dissects R2 into
two components (which necessarily have the image of the circle as common
boundary). In fact, we will even prove a slightly more general result.
Theorem 10.20. Let Cj ⊂ Rn , j = 1, 2, be homeomorphic compact sets.
Then Rn \ C1 and Rn \ C2 have the same number of connected components.

Proof. Denote the components of Rn \C1 by Hj and those of Rn \C2 by Kj .

Since our sets are closed these components are open. Moreover, ∂Hj ⊆ C1
since a sequence from Hj cannot converge to a (necessarily interior) point of
Hk for some k 6= j. Let h : C1 → C2 be a homeomorphism with inverse k :
C2 → C1 . By Theorem 11.10 we can extend both to Rn . Then Theorem 10.1
(iii) and the product formula imply
X
1 = deg(k ◦ h, Hj , y) = deg(h, Hj , Gl ) deg(k, Gl , y)
l
for any y ∈ Hj , where Gl are the components of Rn \ h(∂Hj ). Now we have
[ [
Ki = Rn \ C2 ⊆ Rn \ h(∂Hj ) = Gl
i l
and hence for every i we have Ki ⊆ Gl for some l since components are
connected. Let Nl := {i|Ki ⊆ Gl } and observe that we have deg(k, Gl , y) =
by Theorem 10.1 (i) since −1 ({y}) ⊆ and of
P S
i∈Nl deg(k, K i , y) k j Kj
course deg(h, Hj , Ki ) = deg(h, Hj , Gl ) for every i ∈ Nl . Therefore,
XX X
1= deg(h, Hj , Ki ) deg(k, Ki , y) = deg(h, Hj , Ki ) deg(k, Ki , Hj )
l i∈Nl i
By reversing the role of C1 and C2 , the same formula holds with Hj and Ki
interchanged.
Hence
X XX X
1= deg(h, Hj , Ki ) deg(k, Ki , Hj ) = 1
i i j j

shows that if either the number of components of Rn \ C1 or the number

of components of Rn \ C2 is finite, then so is the other and both are equal.
Otherwise there is nothing to prove. 
 Chapter 11

The Leray–Schauder
mapping degree

11.1. The mapping degree on finite dimensional Banach

spaces
The objective of this section is to extend the mapping degree from Rn to
general Banach spaces. Naturally, we will first consider the finite dimensional
case.
Let X be a (real) Banach space of dimension n and let φ be any isomor-
phism between X and Rn . Then, for f ∈ C̄y (U, X), U ⊂ X open, y ∈ X, we
can define
deg(f, U, y) := deg(φ ◦ f ◦ φ−1 , φ(U ), φ(y)) (11.1)
provided this definition is independent of the isomorphism chosen. To see
this let ψ be a second isomorphism. Then A = ψ ◦ φ−1 ∈ GL(n). Abbreviate
f ∗ = φ ◦ f ◦ φ−1 , y ∗ = φ(y) and pick f˜∗ ∈ C̄y1∗ (φ(U ), Rn ) in the same
component of C̄y∗ (φ(U ), Rn ) as f ∗ such that y ∗ ∈ RV(f ∗ ). Then A ◦ f˜∗ ◦
A−1 ∈ C̄y1 (ψ(U ), Rn ) is in the same component of C̄y (ψ(U ), Rn ) as A ◦ f ∗ ◦
A−1 = ψ ◦ f ◦ ψ −1 (since A is also a homeomorphism) and

JA◦f˜∗ ◦A−1 (Ay ∗ ) = det(A)Jf˜∗ (y ∗ ) det(A−1 ) = Jf˜∗ (y ∗ ) (11.2)

by the chain rule. Thus we have deg(ψ ◦ f ◦ ψ −1 , ψ(U ), ψ(y)) = deg(φ ◦ f ◦

φ−1 , φ(U ), φ(y)) and our definition is independent of the basis chosen. In
addition, it inherits all properties from the mapping degree in Rn . Note also
that the reduction property holds if Rm is replaced by an arbitrary subspace
X1 since we can always choose φ : X → Rn such that φ(X1 ) = Rm .

301
302 11. The Leray–Schauder mapping degree

Our next aim is to tackle the infinite dimensional case. The following
example due to Kakutani shows that the Brouwer fixed point theorem (and
hence also the Brouwer degree) does not generalize to infinite dimensions
directly.
Example 11.1. Let X be the Hilbert space `2 (N) and let R be the right
shift
p given by Rx := (0, px1 , x2 , . . . ). Define f : B̄1 (0) → B̄1 (0), x 7→
1 − kxk δ1 + Rx = ( 1 − kxk2 , x1 , x2 , . . . ). Then a short calculation
2

shows kf (x)k2 p
= (1 − kxk2 ) + kxk2 = 1 and any fixed point must satisfy
kxk = 1, x1 = 1 − kxk2 = 0 and xj+1 = xj , j ∈ N giving the contradiction
xj = 0, j ∈ N. 
However, by the reduction property we expect that the degree should
hold for functions of the type I + F , where F has finite dimensional range.
In fact, it should work for functions which can be approximated by such
functions. Hence as a preparation we will investigate this class of functions.

11.2. Compact maps

Let X, Y be Banach spaces and U ⊆ X. A map F : U ⊂ X → Y is called
finite dimensional if its range is finite dimensional. In addition, it is called
compact if it is continuous and maps bounded sets into relatively compact
ones. The set of all compact maps is denoted by C(U, Y ) and the set of
all compact, finite dimensional maps is denoted by F(U, Y ). Both sets are
normed linear spaces and we have C(U, Y ) ⊆ Cb (U, Y ) if U is bounded (recall
that compact sets are automatically bounded).
If K is compact, then C(K, Y ) = C(K, Y ) (since the continuous image
of a compact set is compact) and if dim(Y ) < ∞, then F(U, Y ) = C(U, Y ).
In particular, if U ⊂ Rn is bounded, then F(U , Rn ) = C(U , Rn ) = C(U , Rn ).
Example 11.2. Note that for nonliner functions it is important to include
continuity in the definition of compactness. Indeed, if X is a Hilbert space
with an orthonormal basis {xj }j∈N , then
(
j(1 − 2|x − xj |), x ∈ B1/2 (xj ),
φ(x) =
0, else,

is in C(B1 (0), R) but not bounded. Hence F (x) = φ(x)x1 is one-dimensional

but not compact. Choosing φ(x) = 1 for x ∈ B1/2 (0) and φ(x) = 0 else gives
a map F which maps bounded sets to relatively compact ones but which is
not continuous. 
Now let us collect some results needed in the sequel.
11.2. Compact maps 303

Lemma 11.1. If K ⊂ X is compact, then for every ε > 0 there is a finite

dimensional subspace Xε ⊆ X and a continuous map Pε : K → Xε such that
|Pε (x) − x| ≤ ε for all x ∈ K.

Proof. Pick {xi }ni=1 ⊆ K such that ni=1 Bε (xi ) covers K. Let {φi }ni=1 be
S
a partition of unity (restricted to K) subordinate
Pn to {Bε (xi )}i=1 , that is,
n

φi ∈ C(K, [0, 1]) with supp(φi ) ⊂ Bε (xi ) and i=1 φi (x) = 1, x ∈ K. Set
n
X
Pε (x) = φi (x)xi ,
i=1
then
n
X n
X n
X
|Pε (x) − x| = | φi (x)x − φi (x)xi | ≤ φi (x)|x − xi | ≤ ε. 
i=1 i=1 i=1

This lemma enables us to prove the following important result.

Theorem 11.2. Let U be bounded, then the closure of F(U, Y ) in Cb (U, Y )
is C(U, Y ).

Proof. Suppose FN ∈ C(U, Y ) converges to F . If F 6∈ C(U, Y ) then we can

find a sequence xn ∈ U such that |F (xn ) − F (xm )| ≥ ρ > 0 for n 6= m. If N
is so large that |F − FN | ≤ ρ/4, then
|FN (xn ) − FN (xm )| ≥ |F (xn ) − F (xm )| − |FN (xn ) − F (xn )|
− |FN (xm ) − F (xm )|
ρ ρ
≥ρ−2 =
4 2
This contradiction shows F(U, Y ) ⊆ C(U, Y ). Conversely, let F ∈ C(U, Y ),
set K := F (U ) and choose Pε according to Lemma 11.1. Then Fε = Pε ◦ F ∈
F(U, Y ) converges to F . Hence C(U, Y ) ⊆ F(U, Y ) and we are done. 

Finally, let us show some interesting properties of mappings I + F , where

F ∈ C(U, Y ).
Lemma 11.3. Let U ⊆ X be bounded and closed. Suppose F ∈ C(U , X),
then I + F is proper (i.e., inverse images of compact sets are compact) and
maps closed subsets to closed subsets.

Proof. Let A ⊆ U be closed and suppose yn = (I + F )(xn ) ∈ (I + F )(A)

converges to some point y. Since yn −xn = F (xn ) ∈ F (U ) we can assume that
yn − xn → z after passing to a subsequence and hence xn → x = y − z ∈ A.
Since y = x + F (x) ∈ (I + F )(A), (I + F )(A) is closed.
Next, let U be closed and K ⊂ X be compact. Let {xn } ⊆ (I+F )−1 (K).
Then yn := xn + F (xn ) ∈ K and we can pass to a subsequence ynm =
304 11. The Leray–Schauder mapping degree

xnm + F (xnm ) such that ynm → y. As before this implies xnm → x and thus
(I + F )−1 (K) is compact. 

Finally note that if F ∈ C(U , Y ) and G ∈ C(Y, Z), then G ◦ F ∈ C(U , Z)

and similarly, if G ∈ Cb (V , U ), then F ◦ G ∈ C(V , Y ).
Now we are all set for the definition of the Leray–Schauder degree, that
is, for the extension of our degree to infinite dimensional Banach spaces.

11.3. The Leray–Schauder mapping degree

For an open set U ⊂ X we set
C¯y (U, X) := {F ∈ C(U , X)|y 6∈ (I + F )(∂U )} (11.3)
and F̄y (U, X) := {F ∈ F(U , X)|y 6∈ (I + F )(∂U )}. Note that for F ∈
C¯y (U, X) we have dist(y, (I + F )(∂U )) > 0 since I + F maps closed sets to
closed sets (cf. Problem B.27).
Abbreviate ρ := dist(y, (I + F )(∂U )) and pick F1 ∈ F(U , X) such that
|F − F1 | < ρ implying F1 ∈ F̄y (U, X). Next, let X1 be a finite dimensional
subspace of X such that F1 (U ) ⊂ X1 , y ∈ X1 and set U1 := U ∩ X1 . Then
we have F1 ∈ F̄y (U1 , X1 ) and might define
deg(I + F, U, y) := deg(I + F1 , U1 , y) (11.4)
provided we show that this definition is independent of F1 and X1 (as above).
Pick another map F2 ∈ F(U , X) such that |F − F2 | < ρ and let X2 be a
corresponding finite dimensional subspace as above. Consider X0 := X1 +
X2 , U0 = U ∩ X0 , then Fi ∈ F̄y (U0 , X0 ), i = 1, 2, and
deg(I + Fi , U0 , y) = deg(I + Fi , Ui , y), i = 1, 2, (11.5)
by the reduction property. Moreover, set H(t) = I + (1 − t)F1 + t F2 implying
H(t) ∈ C¯y (U0 , X0 ), t ∈ [0, 1], since |H(t) − (I + F )| < ρ for t ∈ [0, 1]. Hence
homotopy invariance
deg(I + F1 , U0 , y) = deg(I + F2 , U0 , y) (11.6)
shows that (11.4) is independent of F1 , X1 .
Theorem 11.4. Let U be a bounded open subset of a (real) Banach space
X and let F ∈ C¯y (U, X), y ∈ X. Then the following hold true.
(i). deg(I + F, U, y) = deg(I + F − y, U, 0).
(ii). deg(I, U, y) = 1 if y ∈ U .
(iii). If U1,2 are open, disjoint subsets of U such that y 6∈ f (U \(U1 ∪U2 )),
then deg(I + F, U, y) = deg(I + F, U1 , y) + deg(I + F, U2 , y).
11.3. The Leray–Schauder mapping degree 305

(iv). If H : [0, 1] × U → X and y : [0, 1] → X are both continuous such

that H(t) ∈ C¯y(t) (U, X), t ∈ [0, 1], then deg(I + H(0), U, y(0)) =
deg(I + H(1), U, y(1)).

Proof. Except for (iv) all statements follow easily from the definition of the
degree and the corresponding property for the degree in finite dimensional
spaces. Considering H(t, x) − y(t), we can assume y(t) = 0 by (i). Since
H([0, 1], ∂U ) is compact, we have ρ = dist(y, H([0, 1], ∂U )) > 0. By Theo-
rem 11.2 we can pick H1 ∈ F([0, 1] × U, X) such that |H(t) − H1 (t)| < ρ,
t ∈ [0, 1]. This implies deg(I + H(t), U, 0) = deg(I + H1 (t), U, 0) and the rest
follows from Theorem 10.2. 

In addition, Theorem 10.1 and Theorem 10.2 hold for the new situation
as well (no changes are needed in the proofs).

Theorem 11.5. Let F, G ∈ C¯y (U, X), then the following statements hold.
(i). We have deg(I + F, ∅, y) = 0. Moreover, if Ui , 1 ≤ i ≤ N , are
disjoint open subsets of U such that y 6∈ (I + F )(U \ N
S
PN i=1 Ui ), then
deg(I + F, U, y) = i=1 deg(I + F, Ui , y).
(ii). If y 6∈ (I + F )(U ), then deg(I + F, U, y) = 0 (but not the other way
round). Equivalently, if deg(I + F, U, y) 6= 0, then y ∈ (I + F )(U ).
(iii). If |F (x) − G(x)| < dist(y, (I + F )(∂U )), x ∈ ∂U , then deg(I +
F, U, y) = deg(I + G, U, y). In particular, this is true if F (x) =
G(x) for x ∈ ∂U .
(iv). deg(I + ., U, y) is constant on each component of C¯y (U, X).
(v). deg(I + F, U, .) is constant on each component of X \ (I + F )(∂U ).

Note that it is easy to generalize Borsuk’s theorem.

Theorem 11.6 (Borsuk). Let U ⊆ X be open, bounded and symmetric

with respect to the origin (i.e., U = −U ). Let F ∈ C¯0 (U, X) be odd (i.e.,
F (−x) = −F (x)). Then deg(I + F, U, 0) is odd.

Proof. Choose F1 and U1 as in the definition of the degree. Then U1 is

symmetric and F1 can be chosen to be odd by replacing it by its odd part.
Hence the claim follows from the finite dimensional version. 

In the same way as in the finite dimensional case we also obtain the
invariance of domain theorem.

Theorem 11.7 (Brouwer). Let U ⊆ X be open and let F ∈ C(U, X) be

compact with I + F locally injective. Then I + F is also open.
306 11. The Leray–Schauder mapping degree

11.4. The Leray–Schauder principle and the Schauder fixed

point theorem
As a first consequence we note the Leray–Schauder principle which says that
a priori estimates yield existence.
Theorem 11.8 (Schaefer fixed point or Leray–Schauder principle). Suppose
F ∈ C(X, X) and any solution x of x = tF (x), t ∈ [0, 1] satisfies the a priori
bound |x| ≤ M for some M > 0, then F has a fixed point.

Proof. Pick ρ > M and observe deg(I − F, Bρ (0), 0) = deg(I, Bρ (0), 0) = 1

using the compact homotopy H(t, x) := −tF (x). Here H(t) ∈ C¯0 (Bρ (0), X)
due to the a priori bound. 

Now we can extend the Brouwer fixed point theorem to infinite dimen-
sional spaces as well.
Theorem 11.9 (Schauder fixed point). Let K be a closed, convex, and
bounded subset of a Banach space X. If F ∈ C(K, K), then F has at least
one fixed point. The result remains valid if K is only homeomorphic to a
closed, convex, and bounded subset.

Proof. Since K is bounded, there is a ρ > 0 such that K ⊆ Bρ (0). By

Theorem 11.10 below we can find a continuous retraction R : X → K (i.e.,
R(x) = x for x ∈ K) and consider F̃ = F ◦ R ∈ C(Bρ (0), Bρ (0)). Now
either tF̃ has a fixed point on the boundary ∂Bρ (0) or the compact homo-
topy H(t, x) := −tF̃ (x) satisfies 0 6∈ (I − tF̃ )(∂Bρ (0)) and thus deg(I −
F̃ , Bρ (0), 0) = deg(I, Bρ (0), 0) = 1. Hence there is a point x0 = F̃ (x0 ) ∈ K.
Since F̃ (x0 ) = F (x0 ) for x0 ∈ K we are done. 

It remains to prove the following variant of the Tietze extension the-

orem needed in the proof.
Theorem 11.10. Let X be a metric space, Y a normed space and let K
be a closed subset of X. Then F ∈ C(K, Y ) has a continuous extension
F̄ ∈ C(X, Y ) such that F̄ (X) ⊆ conv(F (K)).

Proof. Consider the open cover {Bρ(x) (x)}x∈X\K for X \ K, where ρ(x) =
dist(x, K)/2. Choose a (locally finite) partition of unity {φλ }λ∈Λ subordinate
to this cover (cf. Lemma B.30) and set
X
F̄ (x) := φλ (x)F (xλ ) for x ∈ X \ K,
λ∈Λ
where xλ ∈ K satisfies dist(xλ , supp φλ ) ≤ 2 dist(K, supp φλ ). By con-
struction, F̄ is continuous except for possibly at the boundary of K. Fix
x0 ∈ ∂K, ε > 0 and choose δ > 0 such that |F (x) − F (x0 )| ≤ ε for all
11.4. The Leray–Schauder principle 307

x ∈ K with |x − x0 | < 4δ. We will show that |F̄ (x) − F (x0 )| ≤ ε for
all
P x ∈ X with |x − x0 | < δ. Suppose x 6∈ K, then |F̄ (x) − F (x0 )| ≤
λ∈Λ φλ (x)|F (xλ ) − F (x0 )|. By our construction, xλ should be close to x
for all λ with x ∈ supp φλ since x is close to K. In fact, if x ∈ supp φλ we
have
|x − xλ | ≤ dist(xλ , supp φλ ) + diam(supp φλ )
≤ 2 dist(K, supp φλ ) + diam(supp φλ ),
where diam(supp φλ ) := supx,y∈supp φλ |x − y|. Since our partition of unity is
subordinate to the cover {Bρ(x) (x)}x∈X\K we can find a x̃ ∈ X \ K such that
supp φλ ⊂ Bρ(x̃) (x̃) and hence diam(supp φλ ) ≤ 2ρ(x̃) ≤ dist(K, Bρ(x̃) (x̃)) ≤
dist(K, supp φλ ). Putting it all together implies that we have |x − xλ | ≤
3 dist(K, supp φλ ) ≤ 3|x0 − x| whenever x ∈ supp φλ and thus
|x0 − xλ | ≤ |x0 − x| + |x − xλ | ≤ 4|x0 − x| ≤ 4δ
as expected. By our choice of δ we have |F (xλ ) − F (x0 )| ≤ ε for all λ with
φλ (x) 6= 0. Hence |F (x) − F (x0 )| ≤ ε whenever |x − x0 | ≤ δ and we are
done. 
Example 11.3. Consider the nonlinear integral equation
Z 1
x = F (x), F (x)(t) := e−ts cos(λx(s))ds
0
in X := C[0, 1] with λ > 0. Then one checks that F ∈ C(X, X) since
Z 1
|F (x)(t) − F (y)(t)| ≤ e−ts | cos(λx(s)) − cos(λy(s))|ds
0
Z 1
≤ e−ts λ|x(s) − y(s)|ds ≤ λkx − yk∞ .
0
In particular, for λ < 1 we have a contraction and the contraction principle
gives us existence of a unique fixed point. Moreover, proceeding similarly,
one obtains estimates for the norm of F (x) and its derivative:
kF (x)k∞ ≤ 1, kF (x)0 k∞ ≤ 1.
Hence the Arzelà–Ascoli theorem (Theorem B.39) implies that the image of
F is a compact subset of the unit ball and hence F ∈ C(B̄1 (0), B̄1 (0)). Thus
the Schauder fixed point theorem guarantees a fixed point for all λ > 0. 
Finally, let us prove another fixed point theorem which covers several
others as special cases.
Theorem 11.11. Let U ⊂ X be open and bounded and let F ∈ C(U , X).
Suppose there is an x0 ∈ U such that
F (x) − x0 6= α(x − x0 ), x ∈ ∂U, α ∈ (1, ∞). (11.7)
308 11. The Leray–Schauder mapping degree

Then F has a fixed point.

Proof. Consider H(t, x) := x − x0 − t(F (x) − x0 ), then we have H(t, x) 6= 0

for x ∈ ∂U and t ∈ [0, 1] by assumption. If H(1, x) = 0 for some x ∈ ∂U ,
then x is a fixed point and we are done. Otherwise we have deg(I−F, U, 0) =
deg(I − x0 , U, 0) = deg(I, U, x0 ) = 1 and hence F has a fixed point. 

Now we come to the anticipated corollaries.

Corollary 11.12. Let F ∈ C(B̄ρ (0), X). Then F has a fixed point if one of
the following conditions holds.
(i) F (∂Bρ (0)) ⊆ B̄ρ (0) (Rothe).
(ii) |F (x) − x|2 ≥ |F (x)|2 − |x|2 for x ∈ ∂Bρ (0) (Altman).
(iii) X is a Hilbert space and hF (x), xi ≤ |x|2 for x ∈ ∂Bρ (0) (Kras-
nosel’skii).

Proof. Our strategy is to verify (11.7) with x0 = 0. (i). F (∂Bρ (0)) ⊆ B̄ρ (0)
and F (x) = αx for |x| = ρ implies |α|ρ ≤ ρ and hence (11.7) holds. (ii).
F (x) = αx for |x| = ρ implies (α − 1)2 ρ2 ≥ (α2 − 1)ρ2 and hence α ≤ 1.
(iii). Special case of (ii) since |F (x) − x|2 = |F (x)|2 − 2hF (x), xi + |x|2 . 

11.5. Applications to integral and differential equations

In this section we want to show how our results can be applied to integral
and differential equations. To be able to apply our results we will need to
know that certain integral operators are compact.
Lemma 11.13. Suppose I = [a, b] ⊂ R and f ∈ C(I × I × Rn , Rn ), τ ∈
C(I, I), then
F : C(I, Rn ) → C(I, Rn ) (11.8)
R τ (t)
x(t) 7→ F (x)(t) = a f (t, s, x(s))ds
is compact.

Proof. We first need to prove that F is continuous. Fix x0 ∈ C(I, Rn ) and

ε > 0. Set ρ := kx0 k∞ + 1 and abbreviate B̄ = B̄ρ (0) ⊂ Rn . The function
f is uniformly continuous on Q := I × I × B̄ since Q is compact. Hence for
ε1 := ε/(b − a) we can find a δ ∈ (0, 1] such that |f (t, s, x) − f (t, s, y)| ≤ ε1
for |x − y| < δ. But this implies
Z τ (t)
kF (x) − F (x0 )k∞ ≤ sup |f (t, s, x(s)) − f (t, s, x0 (s))|ds
t∈I a
≤ sup(b − a)ε1 = ε,
t∈I
11.5. Applications to integral and differential equations 309

for kx − x0 k∞ < δ. In other words, F is continuous. Next we note that if

U ⊂ C(I, Rn ) is bounded, say U ⊂ B̄ρ (0), then
Z 
 τ (t) 
kF (x)k∞ ≤ sup  f (t, s, x(s))ds ≤ (b − a)M, x ∈ U,
 
x∈U  a 

where M := max |f (I, I, B̄)|. Moreover, the family F (U ) is equicontinuous.

Fix ε and ε1 := ε/(2(b − a)), ε2 := ε/(2M ). Since f and τ are uniformly
continuous on I × I × B̄ and I, respectively, we can find a δ > 0 such that
|f (t, s, x) − f (t0 , s, x)| ≤ ε1 and |τ (t) − τ (t0 )| ≤ ε2 for |t − t0 | < δ. Hence we
infer for |t − t0 | < δ
Z 
 τ (t) Z τ (t0 ) 
|F (x)(t) − F (x)(t0 )| =  f (t, s, x(s))ds − f (t0 , s, x(s))ds
 
 a a 
Z τ (t0 )  
Z τ (t) 
≤ |f (t, s, x(s)) − f (t0 , s, x(s))|ds +  |f (t, s, x(s))|ds
 
a  τ (t0 ) 
≤ (b − a)ε1 + ε2 M = ε.
This implies that F (U ) is relatively compact by the Arzelà–Ascoli theorem
(Theorem B.39). Thus F is compact. 

As a first application we use this result to show existence of solutions to

integral equations.
Theorem 11.14. Let F be as in the previous lemma. Then the integral
equation
x − λF (x) = y, λ ∈ R, y ∈ C(I, Rn ) (11.9)
ρ
has at least one solution x ∈ C(I, Rn ) if |λ| ≤ (b−a)M (ρ) , where M (ρ) =
max(s,t,x)∈I×I×B̄ρ (0) |f (s, t, x − y(s))| and ρ > 0 is arbitrary.

Proof. Note that, by our assumption on λ, λF + y maps B̄ρ (y) into itself.
Now apply the Schauder fixed point theorem. 

This result immediately gives the Peano theorem for ordinary differential
equations.
Theorem 11.15 (Peano). Consider the initial value problem
ẋ = f (t, x), x(t0 ) = x0 , (11.10)
where f ∈ C(I ×U, Rn ) and I ⊂ R is an interval containing t0 . Then (11.10)
has at least one local solution x ∈ C 1 ([t0 −ε, t0 +ε], Rn ), ε > 0. For example,
any ε satisfying εM (ε, ρ) ≤ ρ, ρ > 0 with M (ε, ρ) := max |f ([t0 − ε, t0 + ε] ×
B̄ρ (x0 ))| works. In addition, if M (ε, ρ) ≤ M̃ (ε)(1 + ρ), then there exists a
global solution.
310 11. The Leray–Schauder mapping degree

Proof. For notational simplicity we make the shift t → t − t0 , x → x − x0 ,

f (t, x) → f (t + t0 , x + t0 ) and assume t0 = 0, x0 = 0. In addition, it suffices
to consider t ≥ 0 since t → −t amounts to f → −f .
Now observe, that (11.10) is equivalent to
Z t
x(t) − f (s, x(s))ds = 0, x ∈ C([0, ε], Rn )
0
and the first part follows from our previous theorem. To show the second,
fix ε > 0 and assume M (ε, ρ) ≤ M̃ (ε)(1 + ρ). Then
Z t Z t
|x(t)| ≤ |f (s, x(s))|ds ≤ M̃ (ε) (1 + |x(s)|)ds
0 0

implies |x(t)| ≤ exp(M̃ (ε)ε) by Gronwall’s inequality (Problem 11.1). Hence

we have an a priori bound which implies existence by the Leary–Schauder
principle. Since ε was arbitrary we are done. 

As another example we look at the stationary Navier–Stokes equation.

Our goal is to use the Leray–Schauder principle to prove an existence and
uniqueness result for solutions.
Let U (6= ∅) be an open, bounded, and connected subset of R3 . We
assume that U is filled with an incompressible fluid described by its velocity
field v(t, x) ∈ R3 and its pressure p(t, x) ∈ R at time t ∈ R and at a point
x ∈ U . The requirement that the fluid is incompressible implies ∇ · v = 0
(here we use a dot to emphasize a scalar product in R3 ), which follows from
the Gauss theorem since the flux through any closed surface must be zero.
Moreover, the outer force density (force per volume) will be denoted by
K(x) ∈ R3 and is assumed to be known (e.g. gravity).
Then the Navier–Stokes equation governing the motion of the fluid
reads
ρ∂t v = η∆v − ρ(v · ∇)v − ∇p + K, (11.11)
where η > 0 is the viscosity constant and ρ > 0 is the density of the fluid.
In addition to the incompressibility condition ∇v = 0 we also require the
boundary condition v|∂U = 0, which follows from experimental observations.
In what follows we will only consider the stationary Navier–Stokes equa-
tion
0 = η∆v − ρ(v · ∇)v − ∇p + K. (11.12)
Our first step is to switch to a weak formulation and rewrite this equation
in integral form, which is more suitable for our further analysis. We pick as
underlying Hilbert space H01 (U, R3 ) with scalar product
X3 Z
hu, vi = (∂j ui )(∂j vi )dx. (11.13)
i,j=1 U
11.5. Applications to integral and differential equations 311

Recall that by the Poincaré inequality (Theorem 7.30 from [47]) the corre-
sponding norm is equivalent to the usual one. In order to take care of the
incompressibility condition we will choose
X := {v ∈ H01 (U, R3 )|∇ · v = 0}. (11.14)
as our configuration space (check that this is a closed subspace of H01 (U, R3 )).
Now we multiply (11.12) by w ∈ X and integrate over U
Z   Z
η∆v − ρ(v · ∇)v − K · w d x = (∇p) · w d3 x
3
U
ZU
= p(∇w)d3 x = 0, (11.15)
U

where we have used integration by parts (Lemma 7.9 from [47] (iii)) to
conclude that the pressure term drops out of our picture. Using further inte-
gration by parts we finally arrive at the weak formulation of the stationary
Navier–Stokes equation
Z
ηhv, wi − a(v, v, w) − K · w d3 x = 0, for all w ∈ X , (11.16)
U
where
3 Z
X
a(u, v, w) := uk vj (∂k wj ) d3 x. (11.17)
j,k=1 U

In other words, (11.16) represents a necessary solubility condition for the

Navier–Stokes equations and a solution of (11.16) will also be called a weak
solution. If we can show that a weak solution is in H 2 (U, R3 ), then we can
undo the integration by parts and obtain again (11.15). Since the integral
on the left-hand side vanishes for all w ∈ X , one can conclude that the
expression in parenthesis must be the gradient of some function p ∈ L2 (U )
and hence one recovers the original equation. In particular, note that p
follows from v up to a constant if U is connected.
For later use we note
XZ 1X
Z
3
a(v, v, v) = vk vj (∂k vj ) d x = vk ∂k (vj vj ) d3 x
2
j,k U j,k U
Z
1 X
=− (vj vj )∂k vk d3 x = 0, v ∈ X. (11.18)
2 U
j,k

We proceed by studying (11.16). Let K ∈ L2 (U, R3 ), then U K · w d3 x is a

R

bounded linear functional on X and hence there is a K̃ ∈ X such that

Z
K · w d3 x = hK̃, wi, w ∈ X. (11.19)
U
312 11. The Leray–Schauder mapping degree

Moreover, applying the Cauchy–Schwarz inequality twice to each summand

in a(u, v, w) we see
XZ 1/2  Z 1/2
2
|a(u, v, w)| ≤ (uk vj ) dx (∂k wj )2 dx
j,k U U
XZ 1/4  Z 1/4
4
≤ kwk (uk ) dx (vj )4 dx = kuk4 kvk4 kwk.
j,k U U
(11.20)
Since by the Gagliardo–Nierenberg–Sobolev inequality (Theorem 7.18 from
[47]) there is a continous embedding L4 (U, R3 ) ,→ H 1 (U, R3 ) (which is in
this context also known as Ladyzhenskaya inequality), the map a(u, v, .) is
a bounded linear functional in X whenever u, v ∈ X , and hence there is an
element B(u, v) ∈ X such that
a(u, v, w) = hB(u, v), wi, w ∈ X. (11.21)
In addition, the map B : X 2 → X is bilinear and bounded kB(u, v)k ≤
kuk4 kvk4 . In summary we obtain
hηv − B(v, v) − K̃, wi = 0, w ∈ X, (11.22)
and hence
ηv − B(v, v) = K̃. (11.23)
So in order to apply the theory from our previous chapter, we choose the
Banach space Y := L4 (U, R3 ) such that X ,→ Y is compact by the Rellich–
Kondrachov theorem (Theorem 7.27 from [47]).
Motivated by this analysis we formulate the following theorem which
implies existence of weak solutions and uniqueness for sufficiently small outer
forces.
Theorem 11.16. Let X be a Hilbert space, Y a Banach space, and suppose
there is a compact embedding X ,→ Y . In particular, kukY ≤ βkuk. Let
a : X 3 → R be a multilinear form such that
|a(u, v, w)| ≤ αkukY kvkY kwk (11.24)
and a(v, v, v) = 0. Then for any K̃ ∈ X , η > 0 we have a solution v ∈ X to
the problem
ηhv, wi − a(v, v, w) = hK̃, wi, w ∈ X. (11.25)
Moreover, if 2αβkK̃k < η 2 this solution is unique.

Proof. It is no loss to set η = 1. Arguing as before we see that our equation

is equivalent to
v − B(v, v) + K̃ = 0,
11.5. Applications to integral and differential equations 313

where our assumption (11.24) implies

kB(u, v)k ≤ αkukY kvkY ≤ αβ 2 kukkvk
Here the second equality follows since the embedding X ,→ Y is continuous.
Abbreviate F (v) = B(v, v). Observe that F is locally Lipschitz continu-
ous since if kuk, kvk ≤ ρ we have
kF (u) − F (v)k = kB(u − v, u) − B(v, u − v)k ≤ 2αβρku − vkY
≤ 2αβ 2 ρku − vk.
Moreover, let vn be a bounded sequence in X . After passing to a subsequence
we can assume that vn is Cauchy in Y and hence F (vn ) is Cauchy in X by
kF (u) − F (v)k ≤ 2α ρku − vkY . Thus F : X → X is compact.
Hence all we need to apply the Leray–Schauder principle is an a priori
estimate. Suppose v solves v = tF (v) + tK̃, t ∈ [0, 1], then
hv, vi = t a(v, v, v) + thK̃, vi = thK̃, vi.
Hence kvk ≤ kK̃k is the desired estimate and the Leray–Schauder principle
yields existence of a solution.
Now suppose there are two solutions vi , i = 1, 2. By our estimate they
satisfy kvi k ≤ kK̃k and hence kv1 −v2 k = kF (v1 )−F (v2 )k ≤ 2αβ 2 kK̃kkv1 −
v2 k which is a contradiction if 2αβ 2 kK̃k < 1. 
Problem* 11.1 (Gronwall’s inequality). Let α ≥ 0 and β, ϕ : [0, T ] →
[0, ∞) be integrable functions satisfying
Z t
ϕ(t) ≤ α + β(s)ϕ(s)ds.
0
Rt Rt
β(s)ds


Then ϕ(t) ≤ αe 0 . (Hint: Differentiate log α + 0 β(s)ϕ(s)ds .)
 Chapter 12

Monotone maps

12.1. Monotone maps

The Leray–Schauder theory can only be applied to compact perturbations of
the identity. If F is not compact, we need different tools. In this section we
briefly present another class of maps, namely monotone ones, which allow
some progress.
If F : R → R is continuous and we want F (x) = y to have a unique
solution for every y ∈ R, then f should clearly be strictly monotone in-
creasing (or decreasing) and satisfy limx→±∞ F (x) = ±∞. Rewriting these
conditions slightly such that they make sense for vector valued functions the
analogous result holds.

Lemma 12.1. Suppose F : Rn → Rn is continuous and satisfies

F (x)x
lim = ∞. (12.1)
|x|→∞ |x|

Then the equation

F (x) = y (12.2)
has a solution for every y ∈ Rn . If F is strictly monotone

(F (x) − F (y))(x − y) > 0, x 6= y, (12.3)

then this solution is unique.

Proof. Our first assumption implies that G(x) = F (x) − y satisfies G(x)x =
F (x)x − yx > 0 for |x| sufficiently large. Hence the first claim follows from
Problem 10.2. The second claim is trivial. 

315
316 12. Monotone maps

Now we want to generalize this result to infinite dimensional spaces.

Throughout this chapter, H will be a real Hilbert space with scalar product
h., ..i. A map F : H → H is called monotone if
hF (x) − F (y), x − yi ≥ 0, x, y ∈ H, (12.4)
strictly monotone if
hF (x) − F (y), x − yi > 0, x 6= y ∈ H, (12.5)
and finally strongly monotone if there is a constant C > 0 such that
hF (x) − F (y), x − yi ≥ Ckx − yk2 , x, y ∈ H. (12.6)
Note that the same definitions can be made for a Banach space X and
mappings F : X → X ∗ .
Observe that if F is strongly monotone, then it is coercive
hF (x), xi
lim = ∞. (12.7)
kxk→∞ kxk
(Just take y = 0 in the definition of strong monotonicity.) Hence the follow-
ing result is not surprising.
Theorem 12.2 (Zarantonello). Suppose F ∈ C(H, H) is (globally) Lipschitz
continuous and strongly monotone. Then, for each y ∈ H the equation
F (x) = y (12.8)
has a unique solution x(y) ∈ H which depends continuously on y.

Proof. Set
G(x) := x − t(F (x) − y), t > 0,
then F (x) = y is equivalent to the fixed point equation
G(x) = x.
It remains to show that G is a contraction. We compute
kG(x) − G(x̃)k2 = kx − x̃k2 − 2thF (x) − F (x̃), x − x̃i + t2 kF (x) − F (x̃)k2
C
≤ (1 − 2 (Lt) + (Lt)2 )kx − x̃k2 ,
L
where L is a Lipschitz constant for F (i.e., kF (x) − F (x̃)k ≤ Lkx − x̃k).
Thus, if t ∈ (0, 2C
L ), G is a uniform contraction and the rest follows from the
uniform contraction principle. 

Again observe that our proof is constructive. In fact, the best choice
for t is clearly t = LC2 such that the contraction constant θ = 1 − ( C
L ) is
2

minimal. Then the sequence

C
xn+1 = xn − 2 (F (xn ) − y), x0 = y, (12.9)
L
12.2. The nonlinear Lax–Milgram theorem 317

converges to the solution.

Example 12.1. Let A ∈ L (H) and consider F (x) = Ax. Then the condition

hAx, xi ≥ Ckxk2

implies that A has a bounded inverse A−1 : H → H with kA−1 k ≤ C −1 (cf.

Problem 1.34). 

12.2. The nonlinear Lax–Milgram theorem

As a consequence of the last theorem we obtain a nonlinear version of the
Lax–Milgram theorem. We want to investigate the following problem:

a(x, y) = b(y), for all y ∈ H, (12.10)

where a : H2 → R and b : H → R. For this equation the following result

holds.

Theorem 12.3 (Nonlinear Lax–Milgram theorem). Suppose b ∈ L (H, R)

and a(x, .) ∈ L (H, R), x ∈ H, are linear functionals such that there are
positive constants L and C such that for all x, y, z ∈ H we have

a(x, x − y) − a(y, x − y) ≥ Ckx − yk2 (12.11)

and
|a(x, z) − a(y, z)| ≤ Lkzkkx − yk. (12.12)
Then there is a unique x ∈ H such that (12.10) holds.

Proof. By the Riesz lemma (Theorem 2.10) there are elements F (x) ∈ H
and z ∈ H such that a(x, y) = b(y) is equivalent to hF (x) − z, yi = 0, y ∈ H,
and hence to
F (x) = z.
By (12.11) the map F is strongly monotone. Moreover, by (12.12) we infer

kF (x) − F (y)k = sup |hF (x) − F (y), x̃i| ≤ Lkx − yk

x̃∈H,kx̃k=1

that F is Lipschitz continuous. Now apply Theorem 12.2. 

The special case where a ∈ L 2 (H, R) is a bounded bilinear form which

is strongly coercive, that is,

a(x, x) ≥ Ckxk2 , x ∈ H, (12.13)

is usually known as (linear) Lax–Milgram theorem (Theorem 2.16).

318 12. Monotone maps

Example 12.2. For example, let U ⊂ Rn be a bounded domain and consider

the Dirichlet problem for the second order nonlinear elliptic problem
n
X
− ∂i Aij (x)∂j u(x) + F (x, u(x)) = w(x)
i,j=1

with Ai,j ∈ L∞ (U, R) and F : U × Rn → R. If we impose Dirichlet boundary

conditions, then it is natural to look for solutions from the class D = {u ∈
H01 (U, R)|Aij ∂j u ∈ H 1 (U, R), 1 ≤ i, j ≤ n}. Multiplying this equation with
a function v ∈ H01 (U, R) and integrating over U gives the associated weak
formulation
a(u, v) = w(v),
where
Xn Z
a(u, v) := (Aij (x)(∂j u(x))(∂i v(x)) + F (x, u(x))v(x)) dn x,
i,j=1 U
Z
w(v) := w(x)u(x)dn x.
U
Here we have assumed w ∈ L2 (U, R) but, somewhat more general, w ∈
H 1 (U, R)∗ would also suffice.
If we require
C := inf ei Aij (x)ej > 0
e∈S n ,x∈U
as well as
|F (x, u1 ) − F (x, u2 )| ≤ L|u1 − u2 | and (F (x, u1 ) − F (x, u2 ))(u1 − u2 ) ≥ 0,
then the assumption of the nonlinear Lax–Milgram theorem are satisfied on
H01 (U, R). 

12.3. The main theorem of monotone maps

Now we return to the investigation of F (x) = y and weaken the conditions
of Theorem 12.2. We will assume that H is a separable Hilbert space and
that F : H → H is a continuous, coercive monotone map. In fact, it suffices
to assume that F is demicontinuous
lim hF (xn ), yi = hF (x), yi, for all y ∈ H (12.14)
n→∞
whenever xn → x.
The idea is as follows: Start with a finite dimensional subspace Hn ⊂ H
and project the equation F (x) = y to Hn resulting in an equation
Fn (xn ) = yn , xn , yn ∈ Hn . (12.15)
More precisely, let Pn be the (linear) projection onto Hn and set Fn (xn ) =
Pn F (xn ), yn = Pn y (verify that Fn is continuous and monotone!).
12.3. The main theorem of monotone maps 319

Now Lemma 12.1 ensures that there exists a solution S

xn . Now chose the
subspaces Hn such that Hn → H (i.e., Hn ⊂ Hn+1 and ∞ n=1 Hn is dense).
Then our hope is that xn converges to a solution x.
This approach is quite common when solving equations in infinite dimen-
sional spaces and is known as Galerkin approximation. It can often be
used for numerical computations and the right choice of the spaces Hn will
have a significant impact on the quality of the approximation.
So how should we show that xn converges? First of all observe that our
construction of xn shows that xn lies in some ball with radius Rn , which is
chosen such that

hFn (x), xi > kyn kkxk, kxk ≥ Rn , x ∈ Hn . (12.16)

Since hFn (x), xi = hPn F (x), xi = hF (x), Pn xi = hF (x), xi for x ∈ Hn we can

drop all n’s to obtain a constant R (depending on kyk) which works for all
n. So the sequence xn is uniformly bounded

kxn k ≤ R. (12.17)

Now by Theorem 4.31 there exists a weakly convergent subsequence. That

is, after dropping some terms, we can assume that there is some x such that
xn * x, that is,

hxn , zi → hx, zi, for every z ∈ H. (12.18)

And it remains to show that x is indeed a solution. This follows from

Lemma 12.4. Suppose F : H → H is demicontinuous and monotone, then

hy − F (z), x − zi ≥ 0 for every z ∈ H (12.19)

implies F (x) = y.

Proof. Choose z = x ± tw, then ∓hy − F (x ± tw), wi ≥ 0 and by continuity

∓hy − F (x), wi ≥ 0. Thus hy − F (x), wi = 0 for every w implying y − F (x) =
0. 

Now we can show

Theorem 12.5 (Browder–Minty). Let H be a separable Hilbert space. Sup-

pose F : H → H is demicontinuous, coercive and monotone. Then the equa-
tion
F (x) = y (12.20)
has a solution for every y ∈ H. If F is strictly monotone then this solution
is unique.
320 12. Monotone maps

Proof. We have hy − F (z), xn − zi = hyn − Fn (z), xn − zi ≥ 0 for z ∈ Hn .

Taking theSlimit (Problem 4.36) implies hy − F (z), x − zi ≥ 0 for every
z ∈ H∞ = ∞ n=1 Hn . Since H∞ is dense, hy − F (z), x − zi ≥ 0 for every z ∈ H
by continuity and hence F (x) = y by our lemma. 

Note that in the infinite dimensional case we need monotonicity even to

show existence. Moreover, this result can be further generalized in two more
ways. First of all, the Hilbert space H can be replaced by a reflexive Banach
space X if F : X → X ∗ . The proof is similar. Secondly, it suffices if
t 7→ hF (x + ty), zi (12.21)
is continuous for t ∈ [0, 1] and all x, y, z ∈ H, since this condition together
with monotonicity can be shown to imply demicontinuity.
 Appendix A

Some set theory

At the beginning of the 20th century Russell showed with his famous paradox
that naive set theory can lead into contradictions. Hence it was replaced by
axiomatic set theory, more specific we will take the Zermelo–Fraenkel
set theory (ZF), which assumes existence of some sets (like the empty
set and the integers) and defines what operations are allowed. Somewhat
informally (i.e. without writing them using the symbolism of first order logic)
they can be stated as follows:

• Axiom of empty set. There is a set ∅ which contains no elements.

• Axiom of extensionality. Two sets A and B are equal A = B
if they contain the same elements. If a set A contains all elements
from a set B, it is called a subset A ⊆ B. In particular A ⊆ B and
B ⊆ A if and only if A = B.

The last axiom implies that the empty set is unique and that any set which
is not equal to the empty set has at least one element.

• Axiom of pairing. If A and B are sets, then there exists a set

{A, B} which contains A and B as elements. One writes {A, A} =
{A}. By the axiom of extensionality we have {A, B} = {B, A}.
• Axiom of union. S Given a set F whose elements are again sets,
there is a set A = F containing every element that is a member of
some member ofSF. In particular, given two sets A, B there exists
a set A ∪ B = {A, B} consisting of the elements of both sets.
Note that this definition ensures that the union is commutative
A∪B S = B ∪ A and associative (A ∪ B) ∪ C = A ∪ (B ∪ C). Note
also {A} = A.

321
322 A. Some set theory

• Axiom schema of specification. Given a set A and a logical

statement φ(x) depending on x ∈ A we can form the set B =
{x ∈ A|φ(x)} of all elements from A obeying φ. For example,
given two sets A and B we can define their intersection as A ∩
B = {x ∈ A ∪ B|(x ∈ A) ∧ (x ∈ B)} and their complement as
S 6∈ B}. Or the intersection of a family of sets F
A \TB = {x ∈ A|x
as F = {x ∈ F|∀F ∈ F : x ∈ F }.
• Axiom of power set. For any set A, there is a power set P(A)
that contains every subset of A.
From these axioms one can define ordered pairs as (x, y) = {{x}, {x, y}}
and the Cartesian product as A × B = {z ∈ P(A ∪ P(A ∪ B))|∃x ∈ A, y ∈
B : z = (x, y)}. Functions f : A → B are defined as single valued relations,
that is f ⊆ A × B such that (x, y) ∈ f and (x, ỹ) ∈ f implies y = ỹ.
• Axiom schema of replacement. For every function f the image
of a set A is again a set B = {f (x)|x ∈ A}.
So far the previous axioms were concerned with ensuring that the usual
set operations required in mathematics yield again sets. In particular, we can
start constructing sets with any given finite number of elements starting from
the empty set: ∅ (no elements), {∅} (one element), {∅, {∅}} (two elements),
etc. However, while existence of infinite sets (like e.g. the integers) might
seem obvious at this point, it cannot be deduced from the axioms we have
so far. Hence it has to be added as well.
• Axiom of infinity. There exists a set A which contains the empty
set and for every element x ∈ A we also have x ∪ {x} ∈ A. The
smallest such set {∅, {∅}, {∅, {∅}}, . . . } can be identified with the
integers via 0 = ∅, 1 = {∅}, 2 = {∅, {∅}}, . . .
Now we finally have the integers and thus everything we need to start
constructing the rational, real, and complex numbers in the usual way. Hence
we only add one more axiom to exclude some pathological objects which will
lead to contradictions.
• Axiom of Regularity. Every nonempty set A contains an ele-
ment x with x ∩ A = ∅. This excludes for example the possibility
that a set contains itself as an element (apply the axiom to {A}).
Similarly, we can only have A ∈ B or B ∈ A but not both (apply
it to the set {A, B}).
Hence a set is something which can be constructed from the above ax-
ioms. Of course this raises the question if these axioms are consistent but as
has been shown by Gödel this question cannot be answered: If ZF contains
a statement of its own consistency then ZF is inconsistent. In fact, the same
A. Some set theory 323

holds for any other sufficiently rich (such that one can do basic math) system
of axioms. In particular, it also holds for ZFC defined below. So we have to
live with the fact that someday someone might come and prove that ZFC is
inconsistent.
Starting from ZF one can develop basic analysis (including the construc-
tion of the real numbers). However, it turns out that several fundamental
results require yet another construction for their proof:
Given an index set A and for every α ∈ A some set M Sα the product

α∈A M α is defined to be the set of all functions ϕ : A → α∈A Mα which
assign each element α ∈ A some element mα ∈ Mα . If all sets Mα are
nonempty it seems quite reasonable that there should be such a choice func-
tion which chooses an element from Mα for every α ∈ A. However, no matter
how obvious this might seem, it cannot be deduced from the ZF axioms alone
and hence has to be added:
• Axiom of Choice: Given an index set A and nonempty sets

{Mα }α∈A their product α∈A Mα is nonempty.
ZF augmented by the axiom of choice is known as ZFC and we accept
it as the fundament upon which our functional analytic house is built.
Note that the axiom of choice is not only used to ensure that infinite
products are nonempty but also in many proofs! For example, suppose you
start with a set M1 and recursively construct some sets Mn such that in
every step you have a nonempty set. Then the axiom of choice guarantees
the existence of a sequence x = (xn )n∈N with xn ∈ Mn .
The axiom of choice has many important consequences (many of which
are in fact equivalent to the axiom of choice and it is hence a matter of taste
which one to choose as axiom).
A partial order is a binary relation "" over a set P such that for all
A, B, C ∈ P:
• A  A (reflexivity),
• if A  B and B  A then A = B (antisymmetry),
• if A  B and B  C then A  C (transitivity).
It is custom to write A ≺ B if A  B and A 6= B.
Example A.1. Let P(X) be the collections of all subsets of a set X. Then
P is partially ordered by inclusion ⊆. 
It is important to emphasize that two elements of P need not be com-
parable, that is, in general neither A  B nor B  A might hold. However,
if any two elements are comparable, P will be called totally ordered. A
set with a total order is called well-ordered if every nonempty subset has
324 A. Some set theory

a least element, that is some A ∈ P with A  B for every B ∈ P. Note

that the least element is unique by antisymmetry.
Example A.2. R with ≤ is totally ordered and N with ≤ is well-ordered. 
On every well-ordered set we have the
Theorem A.1 (Induction principle). Let K be well ordered and let S(k) be
a statement for arbitrary k ∈ K. Then, if A(l) true for all l ≺ k implies
A(k) true, then A(k) is true for all k ∈ K.

Proof. Otherwise the set of all k for which A(k) is false had a least element
k0 . But by our choice of k0 , A(l) holds for all l ≺ k0 and thus for k0
contradicting our assumption. 

The induction principle also shows that in a well-ordered set functions

f can be defined recursively, that is, by a function ϕ which computes the
value of f (k) from the values f (l) for all l ≺ k. Indeed, the induction
principle implies that on the set Mk = {l ∈ K|l ≺ k} there is at most one
such function fk . Since k is arbitrary, f is unique. In case of the integers
existence of fk is also clear provided f (1) is given. In general, one can prove
existence provided fk is given for some k but we will not need this.
If P is partially ordered, then every totally ordered subset is also called
a chain. If Q ⊆ P, then an element M ∈ P satisfying A  M for all A ∈ Q
is called an upper bound.
Example A.3. Let P(X) as before. Then a collection of subsets {An }n∈N ⊆
P(X) satisfying An ⊆ An+1 is a chain. The set n An is an upper bound. 
S

An element M ∈ P for which M  A for some A ∈ P is only possible if

M = A is called a maximal element.
Theorem A.2 (Zorn’s lemma). Every partially ordered set in which every
chain has an upper bound contains at least one maximal element.

Proof. Suppose it were false. Then to every chain C we can assign an

element m(C) such that m(C)  x for all x ∈ C (here we use the axiom of
choice). We call a chain C distinguished if it is well-ordered and if for every
segment Cx = {y ∈ C|y ≺ x} we have m(Cx ) = x. We will also regard C as
a segment of itself.
Then (since for the least element of C we have Cx = ∅) every distin-
guished chain must start like m(∅) ≺ m(m(∅)) ≺ · · · and given two segments
C, D we expect that always one must be a segment of the other.
So let us first prove this claim. Suppose D is not a segment of C. Then
we need to show C = Dz for some z. We start by showing that x ∈ C implies
x ∈ D and Cx = Dx . To see this suppose it were wrong and let x be the
A. Some set theory 325

least x ∈ C for which it fails. Then y ∈ Kx implies y ∈ L and hence Kx ⊂ L.

Then, since Cx 6= D by assumption, we can find a least z ∈ D \ Cx . In fact
we must even have z  Cx since otherwise we could find a y ∈ Cx such that
x  y  z. But then, using that it holds for y, y ∈ D and Cy = Dy so we
get the contradiction z ∈ Dy = Cy ⊂ Cx . So z  Cx and thus also Cx = Dz
which in turn shows x = m(Cx ) = m(Dz ) = z and proves that x ∈ C implies
x ∈ D and Cx = Dx . In particular C ⊂ D and as before C = Dz for the
least z ∈ D \ C. This proves the claim.
Now using this claim we see that we can take the union over all dis-
tinguished chains to get a maximal distinguished chain Cmax . But then we
could add m(Cmax ) 6∈ Cmax to Cmax to get a larger distinguished chain
Cmax ∪ {m(Cmax )} contradicting maximality. 

We will also frequently use the cardinality of sets: Two sets A and
B have the same cardinality, written as |A| = |B|, if there is a bijection
ϕ : A → B. We write |A| ≤ |B| if there is an injective map ϕ : A → B. Note
that |A| ≤ |B| and |B| ≤ |C| implies |A| ≤ |C|. A set A is called infinite if
|A| ≥ |N|, countable if |A| ≤ |N|, and countably infinite if |A| = |N|.
Theorem A.3 (Schröder–Bernstein). |A| ≤ |B| and |B| ≤ |A| implies |A| =
|B|.

Proof. Suppose ϕ : A → B and ψ : B → A are two injective maps. Now

consider sequences xn defined recursively via x2n+1 = ϕ(x2n ) and x2n+1 =
ψ(x2n ). Given a start value x0 ∈ A the sequence is uniquely defined but
might terminate at a negative integer since our maps are not surjective. In
any case, if an element appears in two sequences, the elements to the left and
to the right must also be equal (use induction) and hence the two sequences
differ only by an index shift. So the ranges of such sequences form a partition
for A ∪· B and it suffices to find a bijection between elements in one partition.
If the sequence stops at an element in A we can take ϕ. If the sequence stops
at an element in B we can take ψ −1 . If the sequence is doubly infinite either
of the previous choices will do. 
Theorem A.4 (Zermelo). Either |A| ≤ |B| or |B| ≤ |A|.

Proof. Consider the set of all bijective functions ϕα : Aα → B with Aα ⊆ A.

Then we can define a partial ordering via ϕα  ϕβ if Aα ⊆ Aβ and ϕβ |Aα =
ϕα . Then every chain has an upper bound (the unique function defined on
the union of all domains) and by Zorn’s lemma there is a maximal element
ϕm . For ϕm we have either Am = A or ϕm (Am ) = B since otherwise there
is some x ∈ A \ Am and some y ∈ B \ f (Am ) which could be used to extend
ϕm to Am ∪ {x} by setting ϕ(x) = y. But if Am = A we have |A| ≤ |B| and
if ϕm (Am ) = B we have |B| ≤ |A|. 
326 A. Some set theory

The cardinality of the power set P(A) is strictly larger than the cardi-
nality of A.
Theorem A.5 (Cantor). |A| < |P(A)|.

Proof. Suppose there were a bijection ϕ : A → P(A). Then, for B = {x ∈

A|x 6∈ ϕ(x)} there must be some y such that B = ϕ(y). But y ∈ B if and
only if y 6∈ ϕ(y) = B, a contradiction. 

This innocent looking result also caused some grief when announced by
Cantor as it clearly gives a contradiction when applied to the set of all sets
(which is fortunately not a legal object in ZFC).
The following result and its corollary will be used to determine the car-
dinality of unions and products.
Lemma A.6. Any infinite set can be written as a disjoint union of countably
infinite sets.

Proof. Consider collections of disjoint countably infinite subsets. Such col-

lections can be partially ordered by inclusion and hence there is a maximal
collection by Zorn’s lemma. If the union of such a maximal collection falls
short of the whole set the complement must be finite. Since this finite re-
minder can be added to a set of the collection we are done. 
Corollary A.7. Any infinite set can be written as a disjoint union of two
disjoint subsets having the same cardinality as the original set.

Proof. By the lemma we can write A = · Aα , where all Aα are countably

S
infinite. Now split Aα = Bα ∪· Cα into two disjoint countably infinite sets
(map Aα bijective to N and then split into even
S and odd elements). Then
the desired splitting is A = B ∪· C with B = · Bα and C = · Cα .
S

Theorem A.8. Suppose A or B is infinite. Then |A ∪ B| = max{|A|, |B|}.

Proof. Suppose A is infinite and |B| ≤ |A|. Then |A| ≤ |A ∪ B| ≤ |A ∪· B| ≤

|A∪· A| = |A| by the previous corollary. Here ∪· denotes the disjoint union. 

A standard theorem proven in every introductory course is that N × N

is countable. The generalization of this result is also true.
Theorem A.9. Suppose A is infinite and B 6= ∅. Then |A×B| = max{|A|, |B|}.

Proof. Without loss of generality we can assume |B| ≤ |A| (otherwise ex-
change both sets). Then |A| ≤ |A × B| ≤ |A × A| and it suffices to show
|A × A| = |A|.
We proceed as before and consider the set of all bijective functions ϕα :
Aα → Aα × Aα with Aα ⊆ A with the same partial ordering as before. By
A. Some set theory 327

Zorn’s lemma there is a maximal element ϕm . Let Am be its domain and

let A0m = A \ Am . We claim that |A0m | < |Am . If not, A0m had a subset
A00m with the same cardinality of Am and hence we had a bijection from
A00m → A00m × A00m which could be used to extend ϕ. So |A0m | < |Am and thus
|A| = |Am ∪ A0m | = |Am |. Since we have shown |Am × Am | = |Am | the claim
follows. 
Example A.4. Note that for A = N we have |P(N)| = |R|. Indeed, since
|R| = |Z × [0, 1)| = |[0, 1)| it suffices to show |P(N)| = |[0, 1)|. To this
end note that P(N) can be identified with the set of all sequences with val-
ues in {0, 1} (the value of the sequence at a point tells us wether it is in the
corresponding subset). Now every point in [0, 1) can be mapped to such a se-
quence via its binary expansion. This map is injective but not surjective since
a point can have different binary expansions: |[0, 1)| ≤ |P(N)|.PConversely,
given a sequence an ∈ {0, 1} we can map it to the number ∞ n=1 an 4
−n .

Since this map is again injective (note that we avoid expansions which are
eventually 1) we get |P(N)| ≤ |[0, 1)|. 
Hence we have
|N| < |P(N)| = |R| (A.1)
and the continuum hypothesis states that there are no sets whose cardi-
nality lie in between. It was shown by Gödel and Cohen that it, as well as
its negation, is consistent with ZFC and hence cannot be decided within this
framework.
Problem A.1. Show that Zorn’s lemma implies the axiom of choice. (Hint:
Consider the set of all partial choice functions defined on a subset.)
Problem A.2. Show |RN | = |R|. (Hint: Without loss we can replace R by
(0, 1) and identify each x ∈ (0, 1) with its decimal expansion. Now the digits
in a given sequence are indexed by two countable parameters.)
 Appendix B

Metric and topological

spaces

This chapter collects some basic facts from metric and topological spaces as
a reference for the main text. I presume that you are familiar with most of
these topics from your calculus course. As a general reference I can warmly
recommend Kelley’s classical book [25] or the nice book by Kaplansky [23].
As always such a brief compilation introduces a zoo of properties. While
sometimes the connection between these properties are straightforward, oth-
ertimes they might be quite tricky. So if at some point you are wondering if
there exists an infinite multi-variable sub-polynormal Woffle which does not
satisfy the lower regular Q-property, start searching in the book by Steen
and Seebach [43].

B.1. Basics
One of the key concepts in analysis is convergence. To define convergence
requires the notion of distance. Motivated by the Euclidean distance one is
lead to the following definition:
A metric space is a space X together with a distance function d :
X × X → [0, ∞) such that for arbitrary points x, y, z ∈ X we have

(i) d(x, y) = 0 if and only if x = y,

(ii) d(x, y) = d(y, x),
(iii) d(x, z) ≤ d(x, y) + d(y, z) (triangle inequality).

329
330 B. Metric and topological spaces

If (i) does not hold, d is called a pseudometric. As a straightforward

consequence we record the inverse triangle inequality (Problem B.1)
|d(x, y) − d(z, y)| ≤ d(x, z). (B.1)
Example B.1. The role model for a metric Pnspace is of 2course Euclidean
space R (or C ) together with d(x, y) := ( k=1 |xk − yk | ) .
n n 1/2 
Several notions from Rn carry over to metric spaces in a straightforward
way. The set
Br (x) := {y ∈ X|d(x, y) < r} (B.2)
is called an open ball around x with radius r > 0. We will write BrX (x) in
case we want to emphasize the corresponding space. A point x of some set
U ⊆ X is called an interior point of U if U contains some open ball around
x. If x is an interior point of U , then U is also called a neighborhood of
x. A point x is called a limit point of U (also accumulation or cluster
point) if for any open ball Br (x), there exists at least one point in Br (x) ∩ U
distinct from x. Note that a limit point x need not lie in U , but U must
contain points arbitrarily close to x. A point x is called an isolated point
of U if there exists a neighborhood of x not containing any other points of U .
A set which consists only of isolated points is called a discrete set. If any
neighborhood of x contains at least one point in U and at least one point
not in U , then x is called a boundary point of U . The set of all boundary
points of U is called the boundary of U and denoted by ∂U .
Example B.2. Consider R with the usual metric and let U := (−1, 1).
Then every point x ∈ U is an interior point of U . The points [−1, 1] are
limit points of U , and the points {−1, +1} are boundary points of U .
Let U := Q, the set of rational numbers. Then U has no interior points
and ∂U = R. 
A set all of whose points are interior points is called open. The family
of open sets O satisfies the properties
(i) ∅, X ∈ O,
(ii) O1 , O2 ∈ O implies O1 ∩ O2 ∈ O,
(iii) {Oα } ⊆ O implies α Oα ∈ O.
S

That is, O is closed under finite intersections and arbitrary unions. In-
deed, (i) is obvious, (ii) follows since the intersection of two open balls cen-
tered at x is again an open ball centered at x (explicitly Br1 (x) ∩ Br2 (x) =
Bmin(r1 ,r2) (x)), and (iii) follows since every ball contained in one of the sets
is also contained in the union.
Now it turns out that for defining convergence, a distance is slightly more
than what is actually needed. In fact, it suffices to know when a point is
B.1. Basics 331

in the neighborhood of another point. And if we adapt the definition of a

neighborhood by requiring it to contain an open set around x, then we see
that it suffices to know when a set is open. This motivates the following
definition:
A space X together with a family of sets O, the open sets, satisfying
(i)–(iii), is called a topological space. The notions of interior point, limit
point, and neighborhood carry over to topological spaces if we replace open
ball around x by open set containing x.
There are usually different choices for the topology. Two not too inter-
esting examples are the trivial topology O = {∅, X} and the discrete
topology O = P(X) (the power set of X). Given two topologies O1 and
O2 on X, O1 is called weaker (or coarser) than O2 if O1 ⊆ O2 . Conversely,
O1 is called stronger (or finer) than O2 if O2 ⊆ O1 .
Example B.3. Note that different metrics can give rise to the same topology.
For example, we can equip Rn (or Cn ) with the Euclidean distance d(x, y)
as before or we could also use
n
X
˜
d(x, y) := |xk − yk |. (B.3)
k=1
Then v
n u n n
1 X uX X
√ |xk | ≤ t |xk |2 ≤ |xk | (B.4)
n
k=1 k=1 k=1

shows Br/√n (x) ⊆ B̃r (x) ⊆ Br (x), where B, B̃ are balls computed using d,
˜ respectively. In particular, both distances will lead to the same notion of
d,
convergence. 
Example B.4. We can always replace a metric d by the bounded metric
˜ y) := d(x, y)
d(x, (B.5)
1 + d(x, y)
without changing the topology (since the family of open balls does not
change: Bδ (x) = B̃δ/(1+δ) (x)). To see that d˜ is again a metric, observe
that f (r) = 1+r r
is monotone as well as concave and hence subadditive,
f (r + s) ≤ f (r) + f (s) (cf. Problem B.3). 
Every subspace Y of a topological space X becomes a topological space
of its own if we call O ⊆ Y open if there is some open set Õ ⊆ X such that
O = Õ ∩Y . This natural topology O ∩Y is known as the relative topology
(also subspace, trace or induced topology).
Example B.5. The set (0, 1] ⊆ R is not open in the topology of X := R,
but it is open in the relative topology when considered as a subset of Y :=
[−1, 1]. 
332 B. Metric and topological spaces

A family of open sets B ⊆ O is called a base for the topology if for each
x and each neighborhood U (x), there is some set O ∈ B with x ∈ O ⊆ U (x).
Since an open set
S O is a neighborhood of every one of its points, it can be
written as O = O⊇Õ∈B Õ and we have

Lemma B.1. A family of open sets B ⊆ O is a base for the topology if and
only if every open set can be written as a union of elements from B.

Proof. To see the converse let x and U (x) be given. Then U (x) contains an
open set O containing x which can be written as a union of elements from
B. One of the elements in this union must contain x and this is the set we
are looking for. 

A family of open sets B ⊆ O is called a subbase for the topology if every

open set can be written as a union of finite intersections of elements from B.
Example B.6. The intervals form a base for the topology on R. Slightly
more general, the open balls are a base for the topology in a metric space.
Intervals of the form (α, ∞) or (−∞, α) with α ∈ R are a subbase for topology
of R. 
Note that a subbase B generates the topology in the sense that the
corresponding topology is the coarsest topology for which all of the sets
from B are open.
Example B.7. The extended real numbers R̄ = R ∪ {∞, −∞} have a
topology generated by the extended intervals of the form (α, ∞] or [−∞, α)
with α ∈ R. Note that the map f (x) := 1+|x|
x
maps R̄ → [−1, 1]. This map
becomes isometric if we choose d(x, y) = |f (x) − f (y)| as a metric on R̄. It
is not hard to verify that this metric generates our topology and hence we
can think of R̄ as [−1, 1]. 
There is also a local version of the previous notions. A neighborhood
base for a point x is a collection of neighborhoods B(x) of x such that for
each neighborhood U (x), there is some set B ∈ B(x) with B ⊆ U (x). Note
that the sets in a neighborhood base are not required to be open.
If every point has a countable neighborhood base, then X is called first
countable. If there exists a countable base, then X is called second count-
able. Note that a second countable space is in particular first countable since
for every base B the subset B(x) := {O ∈ B|x ∈ O} is a neighborhood base
for x.
Example B.8. In a metric space the open balls {B1/m (x)}m∈N are a neigh-
borhood base for x. Hence every metric space is first countable. Taking the
union over all x, we obtain a base. In the case of Rn (or Cn ) it even suffices
B.1. Basics 333

to take balls with rational center, and hence Rn (as well as Cn ) is second
countable. 
Given two topologies on X their intersection will again be a topology on
X. In fact, the intersection of an arbitrary collection of topologies is again a
topology and hence given a collection M of subsets of X we can define the
topology generated by M as the smallest topology (i.e., the intersection of all
topologies) containing M. Note that if M is closed under finite intersections
and ∅, X ∈ M, then it will be a base for the topology generated by M
(Problem B.9).
Given two bases we can use them to check if the corresponding topologies
are equal.

Lemma B.2. Let Oj , j = 1, 2 be two topologies for X with corresponding

bases Bj . Then O1 ⊆ O2 if and only if for every x ∈ X and every B1 ∈ B1
with x ∈ B1 there is some B2 ∈ B2 such that x ∈ B2 ⊆ B1 .

Proof. Suppose O1 ⊆ O2 , then B1 ∈ O2 and there is a corresponding B2 by

the very definition of a base. Conversely, let O1 ∈ O1 and pick some x ∈ O1 .
Then there is some B1 ∈ B1 with x ∈ B1 ⊆ O1 and by assumption some
B2 ∈ B2 such that x ∈ B2 ⊆ B1 ⊆ O1 which shows that x is an interior
point with respect to O2 . Since x was arbitrary we conclude O1 ∈ O2 . 

The next definition will ensure that limits are unique: A topological
space is called a Hausdorff space if for any two different points there are
always two disjoint neighborhoods.
Example B.9. Any metric space is a Hausdorff space: Given two different
points x and y, the balls Bd/2 (x) and Bd/2 (y), where d = d(x, y) > 0,
are disjoint neighborhoods. A pseudometric space will in general not be
Hausdorff since two points of distance 0 cannot be separated by open balls.

The complement of an open set is called a closed set. It follows from
De Morgan’s laws
[
\ \
[
X\ Uα = (X \ Uα ), X \ Uα = (X \ Uα ) (B.6)
α α α α

that the family of closed sets C satisfies

(i) ∅, X ∈ C,
(ii) C1 , C2 ∈ C implies C1 ∪ C2 ∈ C,
(iii) {Cα } ⊆ C implies α Cα ∈ C.
T

That is, closed sets are closed under finite unions and arbitrary intersections.
334 B. Metric and topological spaces

The smallest closed set containing a given set U is called the closure
\
U := C, (B.7)
C∈C,U ⊆C

and the largest open set contained in a given set U is called the interior
[
U ◦ := O. (B.8)
O∈O,O⊆U

It is not hard to see that the closure satisfies the following axioms (Kuratowski
closure axioms):
(i) ∅ = ∅,
(ii) U ⊆ U ,
(iii) U = U ,
(iv) U ∪ V = U ∪ V .
In fact, one can show that these axioms can equivalently be used to define the
topology by observing that the closed sets are precisely those which satisfy
U = U . Similarly, the open sets are precisely those which satisfy U ◦ = U .
Lemma B.3. Let X be a topological space. Then the interior of U is the
set of all interior points of U , and the closure of U is the union of U with
all limit points of U . Moreover, ∂U = U \ U ◦ .

Proof. The first claim is straightforward. For the second claim observe that
by Problem B.7 we have that U = (X \ (X \ U )◦ ), that is, the closure is the
set of all points which are not interior points of the complement. That is,
x 6∈ U iff there is some open set O containing x with O ⊆ X \ U . Hence,
x ∈ U iff for all open sets O containing x we have O 6⊆ X \ U , that is,
O ∩ U 6= ∅. Hence, x ∈ U iff x ∈ U or if x is a limit point of U . The last
claim is left as Problem B.8. 
Example B.10. For any x ∈ X the closed ball
B̄r (x) := {y ∈ X|d(x, y) ≤ r} (B.9)
is a closed set (check that its complement is open). But in general we have
only
Br (x) ⊆ B̄r (x) (B.10)
since an isolated point y with d(x, y) = r will not be a limit point. In Rn
(or Cn ) we have of course equality. 
Problem B.1. Show that |d(x, y) − d(z, y)| ≤ d(x, z).
Problem B.2. Show the quadrangle inequality |d(x, y) − d(x0 , y 0 )| ≤
d(x, x0 ) + d(y, y 0 ).
B.2. Convergence and completeness 335

Problem B.3. Show that if f : [0, ∞) → R is concave, f (λx + (1 − λ)y) ≥

λf (x)+(1−λ)f (y) for λ ∈ [0, 1], and satisfies f (0) = 0, then it is subadditive,
f (x+y) ≤ f (x)+f (y). If in addition f is increasing and d is a pseudometric,
then so is f (d). (Hint: Begin by showing f (λx) ≥ λf (x).)
Problem B.4. Show De Morgan’s laws.
Problem B.5. Show that in a (nonempty) Hausdorff space X singleton sets
{x} (with x ∈ X) are closed.
Problem B.6. Show that the closure satisfies the Kuratowski closure axioms.
Problem B.7. Show that the closure and interior operators are dual in the
sense that
X \ U = (X \ U )◦ and X \ U ◦ = (X \ U ).
In particular, the closure is the set of all points which are not interior points
of the complement. (Hint: De Morgan’s laws.)
Problem B.8. Show that the boundary of U is given by ∂U = U \ U ◦ .
Problem B.9. Suppose M is a collection of sets closed under finite inter-
sections containing
S ∅ and X. Then the topology generated by M is given by
O(M) := { α Mα |Mα ∈ M}.

B.2. Convergence and completeness

A sequence (xn )∞ n=1 ∈ X
N is said to converge to some point x ∈ X if

limn→∞ d(x, xn ) = 0. We write limn→∞ xn = x or xn → x as usual in

this case. Clearly the limit x is unique if it exists (this is not true for a
pseudometric). We will also frequently identify the sequence with its values
xn for simplicity of notation.
Note that convergent sequences are bounded. Here a set U ⊆ X is called
bounded if it is contained within a ball, that is, U ⊆ Br (x) for some x ∈ X
and r > 0.
Note that convergence can also be equivalently formulated in topological
terms: A sequence (xn )∞ n=1 converges to x if and only if for every neighbor-
hood U (x) of x there is some N ∈ N such that xn ∈ U (x) for n ≥ N . In a
Hausdorff space the limit is unique. However, sequences usually do not suf-
fice to describe a topology and, in general, definitions in terms of sequences
are weaker (see the example below). This could be avoided by using general-
ized sequences, so-called nets, where the index set N is replaced by arbitrary
directed sets. We will not pursue this here.
Example B.11. For example, we can call a set U sequentially closed if
every convergent sequence from U also has its limit in U . If U is closed, then
336 B. Metric and topological spaces

every point in the complement is an interior point of the complement, thus

no sequence from U can converge to such a point. Hence every closed set is
sequentially closed. In a metric space (or more generally in a first countable
space) we can find a sequence for every limit point x by choosing a point
(different from x) from every set in a neighborhood base. Hence the converse
is also true in this case. 
Note that the argument from the previous example shows that in a first
countable space sequentially closed is the same as closed. In particular, in
this case the family of closed sets is uniquely determined by the convergent
sequences:
Lemma B.4. Two first countable topologies agree if and only if they give
rise to the same convergent sequences.

Of course every subsequence of a convergent sequence will converge to

the same limit and we have the following converse:
Lemma B.5. Let X be a topological space, (xn )∞ n=1 ∈ X a sequence and
N

x ∈ X. If every subsequence has a further subsequence which converges to x,

then xn converges to x.

Proof. We argue by contradiction: If xn 6→ x we can find a neighborhood

U (x) and a subsequence xnk 6∈ U (x). But then no subsequence of xnk can
converge to x. 

This innocent observation is often useful to establish convergence in situ-

ations where one knows that the limit of a subsequence solves a given problem
together with uniqueness of solutions for this problem. It can also be used
to show that a notion of convergence does not stem from a topology (cf.
Problem 5.11 from [47]).
In summary: A metric induces a natural topology and a topology induces
a natural notion of convergence. However, a notion of convergence might
not stem form a topology (or different topologies might give rise to the same
notion of convergence) and a topology might not stem from a metric.
A sequence (xn )∞n=1 ∈ X is called a Cauchy sequence if for every
N

ε > 0 there is some N ∈ N such that

d(xn , xm ) ≤ ε, n, m ≥ N. (B.11)
Every convergent sequence is a Cauchy sequence. If the converse is also true,
that is, if every Cauchy sequence has a limit, then X is called complete.
It is easy to see that a Cauchy sequence converges if and only if it has a
convergent subsequence.
Example B.12. Both Rn and Cn are complete metric spaces. 
B.2. Convergence and completeness 337

Example B.13. The metric

d(x, y) := | arctan(x) − arctan(y)| (B.12)
gives rise to the standard topology on R (since arctan is bi-Lipschitz on every
compact interval). However, xn = n is a Cauchy sequence with respect to
this metric but not with respect to the usual metric. Moreover, any sequence
with xn → ∞ or xn → −∞ will be Cauchy with respect to this metric and
hence (show this) for the completion of R precisely the two new points −∞
and +∞ have to be added (cf. Example B.7). 
As noted before, in a metric space x is a limit point of U if and only if
there exists a sequence (xn )∞n=1 ⊆ U \ {x} with limn→∞ xn = x. Hence U
is closed if and only if for every convergent sequence the limit is in U . In
particular,

Lemma B.6. A subset of a complete metric space is again a complete metric

space if and only if it is closed.

A set U ⊆ X is called dense if its closure is all of X, that is, if U = X.

A space is called separable if it contains a countable dense set.

Lemma B.7. A metric space is separable if and only if it is second countable

as a topological space.

Proof. From every dense set we get a countable base by considering open
balls with rational radii and centers in the dense set. Conversely, from every
countable base we obtain a dense set by choosing an element from each set
in the base. 

Lemma B.8. Let X be a separable metric space. Every subset Y of X is

again separable.

Proof. Let A = {xn }n∈N be a dense set in X. The only problem is that A∩Y
might contain no elements at all. However, some elements of A must be at
least arbitrarily close to this intersection: Let J ⊆ N2 be the set of all pairs
(n, m) for which B1/m (xn )∩Y 6= ∅ and choose some yn,m ∈ B1/m (xn )∩Y for
all (n, m) ∈ J. Then B = {yn,m }(n,m)∈J ⊆ Y is countable. To see that B is
dense, choose y ∈ Y . Then there is some sequence xnk with d(xnk , y) < 1/k.
Hence (nk , k) ∈ J and d(ynk ,k , y) ≤ d(ynk ,k , xnk ) + d(xnk , y) ≤ 2/k → 0. 

If X is an (incomplete) metric space, consider the set of all Cauchy se-

quences X from X. Call two Cauchy sequences x = (xn )n∈N and y = (yn )n∈N
equivalent if dX (xn , yn ) → 0 and denote by X̄ the set of all equivalence
classes [x]. Moreover, the quadrangle inequality (Problem B.2) shows that
338 B. Metric and topological spaces

if x = (xn )n∈N and y = (yn )n∈N are Cauchy sequences, so is d(xn , yn ) and
hence we can define a metric on X̄ via

dX̄ ([x], [y]) = lim dX (xn , yn ). (B.13)

n→∞

Indeed, it is straightforward to check that dX̄ is well defined (independent of

the representative) and inherits all properties of a metric from dX . Moreover,
dX̄ agrees with dX of the limits whenever both Cauchy sequences converge
in X.

Theorem B.9. The space X̄ is a metric space containing X as a dense

subspace if we identify x ∈ X with the equivalence class of all sequences
converging to x. Moreover, this embedding is isometric.

Proof. The map J : X → X̄, x0 7→ [(x0 , x0 , . . . )] is an isometric embedding

(i.e., it is injective and preserves the metric). Moreover, for a Cauchy se-
quence x = (xn )n∈N the sequence J(xn ) converges to [x] and hence J(X) is
dense in X̄. It remains to show that X̄ is complete. Let ξn = [(xn,j )j∈N ] be
a Cauchy sequence in X̄. Without loss of generality (by dropping terms) we
can choose the representatives xn,j such that d(xn,j , xn,k ) ≤ n1 for j, k ≥ n.
Then it is not hard to see that ξ = [(xj,j )j∈N ] is its limit. 

Notice that the completion X̄ is unique. More precisely, suppose X̃ is

another complete metric space which contains X as a dense subset such that
the embedding J˜ : X ,→ X̃ is isometric. Then I = J˜ ◦ J −1 : J(X) → J(X)˜
has a unique isometric extension I : X̄ → X̃ (compare Theorem B.38 below).
¯
In particular, it is no restriction to assume that a metric space is complete.

Problem B.10. Let U ⊆ V be subsets of a metric space X. Show that if U

is dense in V and V is dense in X, then U is dense in X.

Problem B.11. Let X be a metric space and denote by B(X) the set of all
bounded functions X → C. Introduce the metric

d(f, g) = sup |f (x) − g(x)|.

x∈X

Show that B(X) is complete.

Problem B.12. Let X be a metric space and B(X) as in the previous prob-
lem. Consider the embedding J : X ,→ B(X) defind via

y 7→ J(x)(y) = d(x, y) − d(x0 , y)

for some fixed x0 ∈ X. Show that this embedding is isometric. Hence J(X)
is another (equivalent) completion of X.
B.3. Functions 339

B.3. Functions
Next, we come to functions f : X → Y , x 7→ f (x). We use the usual
conventions f (U ) := {f (x)|x ∈ U } for U ⊆ X and f −1 (V ) := {x|f (x) ∈ V }
for V ⊆ Y . Note
U ⊆ f −1 (f (U )), f (f −1 (V )) ⊆ V. (B.14)
The set Ran(f ) := f (X) is called the range of f , and X is called the
domain of f . A function f is called injective or one-to-one if for each
y ∈ Y there is at most one x ∈ X with f (x) = y (i.e., f −1 ({y}) contains at
most one point) and surjective or onto if Ran(f ) = Y . A function f which
is both injective and surjective is called bijective.
Recall that we always have
[ [ \ \
f −1 ( Vα ) = f −1 (Vα ), f −1 ( Vα ) = f −1 (Vα ),
α α α α
−1 −1
f (Y \ V ) = X \ f (V ) (B.15)
as well as
[ [ \ \
f ( Uα ) = f (Uα ), f ( Uα ) ⊆ f (Uα ),
α α α α
f (X) \ f (U ) ⊆ f (X \ U ) (B.16)
with equality if f is injective.
A function f between metric spaces X and Y is called continuous at a
point x ∈ X if for every ε > 0 we can find a δ > 0 such that
dY (f (x), f (y)) ≤ ε if dX (x, y) < δ. (B.17)
If f is continuous at every point, it is called continuous. In the case
dY (f (x), f (y)) = dX (x, y) we call f isometric and every isometry is of
course continuous.
Lemma B.10. Let X, Y be metric spaces. The following are equivalent:
(i) f is continuous at x (i.e., (B.17) holds).
(ii) f (xn ) → f (x) whenever xn → x.
(iii) For every neighborhood V of f (x) the preimage f −1 (V ) is a neigh-
borhood of x.

Proof. (i) ⇒ (ii) is obvious. (ii) ⇒ (iii): If (iii) does not hold, there is a
neighborhood V of f (x) such that Bδ (x) 6⊆ f −1 (V ) for every δ. Hence we
can choose a sequence xn ∈ B1/n (x) such that xn 6∈ f −1 (V ). Thus xn → x
but f (xn ) 6→ f (x). (iii) ⇒ (i): Choose V = Bε (f (x)) and observe that by
(iii), Bδ (x) ⊆ f −1 (V ) for some δ. 
340 B. Metric and topological spaces

In a general topological space we use (iii) as the definition of continuity

and (ii) is called sequential continuity. Then continuity will imply se-
quential continuity but the converse will not be true unless we assume (e.g.)
that X is first countable (Problem B.13).
In particular, (iii) implies that f is continuous if and only if the preimage
of every open set is again open (equivalently, the inverse image of every
closed set is closed). Note that by (B.15) it suffices to check this for sets in
a subbase. If the image of every open set is open, then f is called open. A
bijection f is called a homeomorphism if both f and its inverse f −1 are
continuous. Note that if f is a bijection, then f −1 is continuous if and only
if f is open. Two topological spaces are called homeomorphic if there is a
homeomorphism between them.
In a topological space X a function f : X → R is lower semicontin-
uous if the set f −1 ((a, ∞]) is open for every a ∈ R. Similarly, f is upper
semicontinuous if the set f −1 ([−∞, a)) is open for every a ∈ R. Clearly f
is lower semicontinuous if and only if −f is upper semicontinuous.
Finally, the support of a function f : X → Cn is the closure of all points
x for which f (x) does not vanish; that is,
supp(f ) := {x ∈ X|f (x) 6= 0}. (B.18)
Problem B.13. Let X, Y be topological spaces. Show that if f : X → Y
is continuous at x ∈ X then it is also sequential continuous. Show that the
converse holds if X is first countable.
Problem B.14. Let f : X → Y be continuous. Then f (A) ⊆ f (A).
Problem B.15. Let X, Y be topological spaces and let f : X → Y be con-
tinuous. Show that if X is separable, then so is f (X).
Problem B.16. Let X be a topological space and f : X → R. Let x0 ∈ X
and let B(x0 ) be a neighborhood base for x0 . Define
lim inf f (x) := sup inf f, lim sup f (x) := inf sup f.
x→x0 U ∈B(x0 ) U (x0 ) x→x0 U ∈B(x0 ) U (x0 )

Show that both are independent of the neighborhood base and satisfy
(i) lim inf x→x0 (−f (x)) = − lim supx→x0 f (x).
(ii) lim inf x→x0 (αf (x)) = α lim inf x→x0 f (x), α ≥ 0.
(iii) lim inf x→x0 (f (x) + g(x)) ≥ lim inf x→x0 f (x) + lim inf x→x0 g(x).
Moreover, show that
lim inf f (xn ) ≥ lim inf f (x), lim sup f (xn ) ≤ lim sup f (x)
n→∞ x→x0 n→∞ x→x0
for every sequence xn → x0 and there exists a sequence attaining equality if
X is a metric space.
B.4. Product topologies 341

Problem* B.17. Show that the supremum over lower semicontinuous func-
tions is again lower semicontinuous.
Problem* B.18. Let X be a topological space and f : X → R. Show that
f is lower semicontinuous if and only if
lim inf f (x) ≥ f (x0 ), x0 ∈ X.
x→x0

Similarly, f is upper semicontinuous if and only if

lim sup f (x) ≤ f (x0 ), x0 ∈ X.
x→x0

Show that a lower semicontinuous function is also sequentially lower semi-

continuous
lim inf f (xn ) ≥ f (x0 ), xn → x0 , x0 ∈ X.
n→∞
Show the converse if X is a metric space. (Hint: Problem B.16.)

B.4. Product topologies

If X and Y are metric spaces, then X × Y together with
d((x1 , y1 ), (x2 , y2 )) := dX (x1 , x2 ) + dY (y1 , y2 ) (B.19)
is a metric space. A sequence (xn , yn ) converges to (x, y) if and only if xn → x
and yn → y. In particular, the projections onto the first (x, y) 7→ x, respec-
tively, onto the second (x, y) 7→ y, coordinate are continuous. Moreover, if
X and Y are complete, so is X × Y .
In particular, by the inverse triangle inequality (B.1),
|d(xn , yn ) − d(x, y)| ≤ d(xn , x) + d(yn , y), (B.20)
we see that d : X × X → R is continuous.
Example B.14. If we consider R × R, we do not get the Euclidean distance
of R2 unless we modify (B.19) as follows:
˜ 1 , y1 ), (x2 , y2 )) := dX (x1 , x2 )2 + dY (y1 , y2 )2 . (B.21)
p
d((x
As noted in our previous example, the topology (and thus also conver-
gence/continuity) is independent of this choice. 
If X and Y are just topological spaces, the product topology is de-
fined by calling O ⊆ X × Y open if for every point (x, y) ∈ O there are
open neighborhoods U of x and V of y such that U × V ⊆ O. In other
words, the products of open sets form a base of the product topology. Again
the projections onto the first and second component are continuous. In the
case of metric spaces this clearly agrees with the topology defined via the
product metric (B.19). There is also another way of constructing the prod-
uct topology, namely, as the weakest topology which makes the projections
342 B. Metric and topological spaces

continuous. In fact, this topology must contain all sets which are inverse
images of open sets U ⊆ X, that is all sets of the form U × Y as well as
all inverse images of open sets V ⊆ Y , that is all sets of the form X × V .
Adding finite intersections we obtain all sets of the form U × V and hence
the same base as before. In particular, a sequence (xn , yn ) will converge if
and only if both components converge.
Note that the product topology immediately extends to the product of

an arbitrary number of spaces X := α∈A Xα by defining it as the weakest
topology which makes all projections πα : X → Xα continuous.
Example B.15. Let X be a topological space and A an index set. Then

X A = A X is the set of all functions x : A → X and a neighborhood base
at x are sets of functions which coincide with x at a given finite number
of points. Convergence with respect to the product topology corresponds to
pointwise convergence (note that the projection πα is the point evaluation at
α: πα (x) = x(α)). If A is uncountable (and X is not equipped with the trivial
topology), then there is no countable neighborhood base (if there were such a
base, it would involve only a countable number of points, now choose a point
from the complement . . . ). In particular, there is no corresponding metric
even if X has one. Moreover, this topology cannot be characterized with
sequences alone. For example, let X = {0, 1} (with the discrete topology)
and A = R. Then the set F = {x|x−1 (1) is countable} is sequentially closed
but its closure is all of {0, 1}R (every set from our neighborhood base contains
an element which vanishes except at finitely many points). 
In fact this is a special case of a more general construction which is often
used. Let {fα }α∈A be a collection of functions fα : X → Yα , where Yα are
some topological spaces. Then we can equip X with the weakest topology
(known as the initial topology) which makes all fα continuous. That is, we
take the topology generated by sets of the forms fα−1 (Oα ), where Oα ⊆ Yα
is open, known as open cylinders. Finite intersections of such sets, known
as open cylinder sets, are hence a base for the topology and a sequence xn
will converge to x if and only if fα (xn ) → fα (x) for all α ∈ A. In particular,
if the collection is countable, then X will be first (or second) countable if all
Yα are.
The initial topology has the following characteristic property:
Lemma B.11. Let X have the initial topology from a collection of functions
{fα : X → Yα }α∈A and let Z be another topological space. A function
f : Z → X is continuous (at z) if and only if fα ◦ f is continuous (at z) for
all α ∈ A.

Proof. If f is continuous at z, then so is theTcomposition fα ◦ f . Conversely,

let U ⊆ X be a neighborhood of f (z). Then nj=1 fα−1j
(Oαj ) ⊆ U for some αj
B.4. Product topologies 343

and some open neighborhoods Oαj of fαj (f (z)). But then f −1 (U ) contains
the neighborhood f −1 ( nj=1 fα−1 (Oαj )) = nj=1 (fαj ◦ f )−1 (Oαj ) of z.
T T
j


If all Yα are Hausdorff and if the collection {fα }α∈A separates points,
that is for every x 6= y there is some α with fα (x) 6= fα (y), then X will
again be Hausdorff. Indeed for x 6= y choose α such that fα (x) 6= fα (y)
and let Uα , Vα be two disjoint neighborhoods separating fα (x), fα (y). Then
fα−1 (Uα ), fα−1 (Vα ) are two disjoint neighborhoods separating x, y. In partic-

ular, X = α∈A Xα is Hausdorff if all Xα are.
Note that a similar construction works in the other direction. Let {fα }α∈A
be a collection of functions fα : Xα → Y , where Xα are some topological
spaces. Then we can equip Y with the strongest topology (known as the
final topology) which makes all fα continuous. That is, we take as open
sets those for which fα−1 (O) is open for all α ∈ A.
Example B.16. Let ∼ be an equivalence relation on X with equivalence
classes [x] = {y ∈ X|x ∼ y}. Then the quotient topology on the set of
equivalence classes X/ ∼ is the final topology of the projection map π : X →
X/ ∼. 
Example B.17. Let Xα be a collection of topological spaces. The disjoint
union [
X := · Xα
α∈A
is usually given the final topology from the canonical injections iα : Xα ,→ X
such that O ⊆ X is open if and only if O ∩ Xα is open for all α ∈ A. 
Lemma B.12. Let Y have the final topology from a collection of functions
{fα : Xα → Y }α∈A and let Z be another topological space. A function
f : Y → Z is continuous if and only if f ◦ fα is continuous for all α ∈ A.

Proof. If f is continuous, then so is the composition f ◦ fα . Conversely, let

V ⊆ Z be open. Then f ◦ fα implies (f ◦ fα )−1 (V ) = fα−1 (f −1 (V )) open for
all α and hence f −1 (V ) open. 
Problem B.19. Show that X is Hausdorff if and only if the diagonal ∆ :=
{(x, x)|x ∈ X} ⊆ X × X is closed.

Problem B.20. Let X = α∈A Xα with the product topology. Show that
the projection maps are open.

Problem B.21. Let X = α∈A Xα with the product topology. Show that

the product α∈A Cα of closed sets Cα ⊆ Xα is closed.
Problem B.22 (Gluing lemma). Suppose X, Y are topological spaces and
fα : Aα → Y are continuous functions
S defined on Aα ⊆ X. Suppose fα = fβ
on Aα ∩ Aβ such that f : A := α Aα → Y is well defined by f (x) = fα (x)
344 B. Metric and topological spaces

if x ∈ Aα . Show that f is continuous if either all sets Aα are open or if the

collection Aα is finite and all are closed.
Problem B.23. Let {(Xj , dj )}j∈N be a sequence of metric spaces. Show that
X 1 dj (xj , yj ) 1 dj (xj , yj )
d(x, y) = j
or d(x, y) = max j
2 1 + dj (xj , yj ) j∈N 2 1 + dj (xj , yj )
j∈N

is a metric on X = n∈N Xn which generates the product topology. Show
that X is complete if all Xn are.

B.5. Compactness
A cover of a set Y ⊆ X is a family of sets {Uα } such that Y ⊆ α Uα . A
S
cover is called open if all Uα are open. Any subset of {Uα } which still covers
Y is called a subcover.
Lemma B.13 (Lindelöf). If X is second countable, then every open cover
has a countable subcover.

Proof. Let {Uα } be an open cover for Y , and let B be a countable base.
Since every Uα can be written as a union of elements from B, the set of all
B ∈ B which satisfy B ⊆ Uα for some α form a countable open cover for Y .
Moreover, for every Bn in this set we can find an αn such that Bn ⊆ Uαn .
By construction, {Uαn } is a countable subcover. 

A refinement {Vβ } of a cover {Uα } is a cover such that for every β

there is some α with Vβ ⊆ Uα . A cover is called locally finite if every point
has a neighborhood that intersects only finitely many sets in the cover.
Lemma B.14 (Stone). In a metric space every countable open cover has a
locally finite open refinement.

Proof. Denote the cover by {Oj }j∈N and introduce the sets
[
Ôj,n := B2−n (x), where
x∈Aj,n
[
Aj,n := {x ∈ Oj \ (O1 ∪ · · · ∪ Oj−1 )|x 6∈ Ôk,l and B3·2−n (x) ⊆ Oj }.
k∈N,1≤l<n

Then, by construction, Ôj,n is open, Ôj,n ⊆ Oj , and it is a cover since for

every x there is a smallest j such that x ∈ Oj and a smallest n such that
B3·2−n (x) ⊆ Oj implying x ∈ Ôk,l for some l ≤ n.
To show that Ôj,n is locally finite fix some x and let j be the small-
est integer such that x ∈ Ôj,n for some n. Moreover, choose m such that
B2−m (x) ⊆ Ôj,n . It suffices to show that:
B.5. Compactness 345

(i) If i ≥ n + m then B2−n−m (x) is disjoint from Ôk,i for all k.

(ii) If i < n + m then B2−n−m (x) intersects Ôk,i for at most one k.
To show (i) observe that since i > n every ball B2−i (y) used in the
definition of Ôk,i has its center outside of Ôj,n . Hence d(x, y) ≥ 2−m and
B2−n−m (x) ∩ B2−i (x) = ∅ since i ≥ m + 1 as well as n + m ≥ m + 1.
To show (ii) let y ∈ Ôj,i and z ∈ Ôk,i with j < k. We will show
d(y, z) > 2−n−m+1 . There are points r and s such that y ∈ B2−i (r) ⊆ Ôj,i
and z ∈ B2−i (s) ⊆ Ôk,i . Then by definition B3·2−i (r) ⊆ Oj but s 6∈ Oj . So
d(r, s) ≥ 3 · 2−i and d(y, z) > 2−i ≥ 2−n−m+1 . 

A subset K ⊂ X is called compact if every open cover of K has a finite

subcover. A set is called relatively compact if its closure is compact.
Lemma B.15. A topological space is compact if and only if it has the finite
intersection property: The intersection of a family of closed sets is empty
if and only if the intersection of some finite subfamily is empty.

Proof. By taking complements, to every family of open sets there corre-

sponds a family of closed sets and vice versa. Moreover, the open sets are a
cover if and only if the corresponding closed sets have empty intersection. 
Lemma B.16. Let X be a topological space.
(i) The continuous image of a compact set is compact.
(ii) Every closed subset of a compact set is compact.
(iii) If X is Hausdorff, every compact set is closed.
(iv) The finite union of compact sets is compact.
(v) If X is Hausdorff, any intersection of compact sets is compact.

Proof. (i) Observe that if {Oα } is an open cover for f (Y ), then {f −1 (Oα )}
is one for Y .
(ii) Let {Oα } be an open cover for the closed subset Y (in the induced
topology). Then there are open sets Õα with Oα = Õα ∩Y and {Õα }∪{X\Y }
is an open cover for X which has a finite subcover. This subcover induces a
finite subcover for Y .
(iii) Let Y ⊆ X be compact. We show that X \ Y is open. Fix x ∈ X \ Y
(if Y = X there is nothing to do). By the definition of Hausdorff, for
every y ∈ Y there are disjoint neighborhoods V (y) of y and Uy (x) of x. By
compactness
Tn of Y , there are y1 , . . . , yn such that the V (yj ) cover Y . But
then j=1 Uyj (x) is a neighborhood of x which does not intersect Y .
(iv) Note that a cover of the union is a cover for each individual set and
the union of the individual subcovers is the subcover we are looking for.
346 B. Metric and topological spaces

(v) Follows from (ii) and (iii) since an intersection of closed sets is closed.


As a consequence we obtain a simple criterion when a continuous function

is a homeomorphism.
Corollary B.17. Let X and Y be topological spaces with X compact and Y
Hausdorff. Then every continuous bijection f : X → Y is a homeomorphism.

Proof. It suffices to show that f maps closed sets to closed sets. By (ii)
every closed set is compact, by (i) its image is also compact, and by (iii) it
is also closed. 

Concerning products of compact sets we have


Theorem B.18 (Tychonoff). The product α∈A Kα of an arbitrary collec-
tion of compact topological spaces {Kα }α∈A is compact with respect to the
product topology.

Proof. We say that a family F of closed subsets of K has the finite inter-
section property if the intersection of every finite subfamily has nonempty
intersection. The collection of all such families which contain F is partially
ordered by inclusion and every chain has an upper bound (the union of all
sets in the chain). Hence, by Zorn’s lemma, there is a maximal family FM
(note that this family is closed under finite intersections).
Denote by πα : K → Kα the projection onto the α component. Then
the closed sets {πα (F )}F ∈FM also have the finite intersection property and
since Kα is compact, there is some xα ∈ F ∈FM πα (F ). Consequently, if Fα
T

is a closed neighborhood of xα , then πα−1 (Fα ) ∈ FM (otherwise there would

be some F ∈ FM with F ∩ πα−1 (Fα ) = ∅ contradicting T Fα ∩−1 πα (F ) 6= ∅).
Furthermore, for every finite subset A0 ⊆ A we have α∈A0 πα (Fα ) ∈ FM
and so every neighborhood of x = (xα )α∈A intersects F . Since F is closed,
x ∈ F and hence x ∈ FM F .
T


A subset K ⊆ X is called sequentially compact if every sequence

from K has a convergent subsequence whose limit is in K. In a metric
space, compact and sequentially compact are equivalent.
Lemma B.19. Let X be a metric space. Then a subset is compact if and
only if it is sequentially compact.

Proof. Without loss of generality we can assume the subset to be all of X.

Suppose X is compact and let xn be a sequence which has no convergent
subsequence. Then K := {xn } has no limit points and is hence compact by
Lemma B.16 (ii). For every n there is a ball Bεn (xn ) which contains only
B.5. Compactness 347

finitely many elements of K. However, finitely many suffice to cover K, a

contradiction.
Conversely, suppose X is sequentially compact and let {Oα } be some
open cover which has no finite subcover. For every x ∈ X we can choose
some α(x) such that if Br (x) is the largest ball contained in Oα(x) , then
there is no β with B2r (x) ⊂ OβS(show that this is possible). Now choose
a sequence xn such that xn 6∈ m<n Oα(xm ) . Note that by construction
the distance d = d(xm , xn ) to every successor of xm satisfies that the ball
B2d (xm ) does not fit into any of the Oα .
Now let y be the limit of some convergent subsequence and fix some
r > 0 such that Br (y) ⊆ Oα(y) . Then this subsequence must eventually be in
Br/6 (y), but this is impossible since if d := d(xn1 , xn2 ) < r/3 is the distance
between two consecutive elements of this subsequence within Br/6 (y), then
B2d (xn1 ) cannot fit into Oα(y) by construction whereas on the other hand
B2d (xn1 ) ⊆ Br (y) ⊆ Oα(y) . 

If we drop the requirement that the limit must be in K, we obtain rela-

tively compact sets:
Corollary B.20. Let X be a metric space and K ⊂ X. Then K is relatively
compact if and only if every sequence from K has a convergent subsequence
(the limit need not be in K).

Proof. For any sequence xn ∈ K we can find a nearby sequence yn ∈ K

with xn − yn → 0. If we can find a convergent subsequence of yn then the
corresponding subsequence of xn will also converge (to the same limit) and
K is (sequentially) compact in this case. The converse is trivial. 

As another simple consequence observe that

Corollary B.21. A compact metric space X is complete and separable.

Proof. Completeness is immediate from the previous lemma. To see that

X is separable note that, by compactness, for every n ∈ N there
S is a finite
set Sn ⊆ X such that the balls {B1/n (x)}x∈Sn cover X. Then n∈N Sn is a
countable dense set. 

Recall that a set in a metric space is called bounded if it is contained

inside some ball. Clearly the union of two bounded sets is bounded. More-
over, compact sets are always bounded since the can be covered by finitely
many balls. In Rn (or Cn ) the converse also holds.
Theorem B.22 (Heine–Borel). In Rn (or Cn ) a set is compact if and only
if it is bounded and closed.
348 B. Metric and topological spaces

Proof. By Lemma B.16 (ii), (iii), and Tychonoff’s theorem it suffices to

show that a closed interval in I ⊆ R is compact. Moreover, by Lemma B.19,
it suffices to show that every sequence in I = [a, b] has a convergent sub-
sequence. Let xn be our sequence and divide I = [a, a+b 2 ] ∪ [ 2 , b]. Then
a+b

at least one of these two intervals, call it I1 , contains infinitely many ele-
ments of our sequence. Let y1 = xn1 be the first one. Subdivide I1 and pick
y2 = xn2 , with n2 > n1 as before. Proceeding like this, we obtain a Cauchy
sequence yn (note that by construction In+1 ⊆ In and hence |yn − ym | ≤ b−a 2n
for m ≥ n). 

By Lemma B.19 this is equivalent to

Theorem B.23 (Bolzano–Weierstraß). Every bounded infinite subset of Rn
(or Cn ) has at least one limit point.

Combining Theorem B.22 with Lemma B.16 (i) we also obtain the ex-
treme value theorem.
Theorem B.24 (Weierstraß). Let X be compact. Every continuous function
f : X → R attains its maximum and minimum.

A metric space X for which the Heine–Borel theorem holds is called

proper. Lemma B.16 (ii) shows that X is proper if and only if every closed
ball is compact. Note that a proper metric space must be complete (since
every Cauchy sequence is bounded). A topological space is called locally
compact if every point has a compact neighborhood. Clearly a proper
metric space is locally compact. A topological space is called σ-compact,
if it can be written as a countable union of compact sets. Again a proper
space is σ-compact.
Lemma B.25. For a metric space X the following are equivalent:
(i) X is separable and locally compact.
(ii) X contains a countable base consisting of relatively compact sets.
(iii) X is locally compact and σ-compact.
(iv) X can be written as the union of an increasing sequence Un of
relatively compact open sets satisfying Un ⊆ Un+1 for all n.

Proof. (i) ⇒ (ii): Let {xn } be a dense set. Then the balls Bn,m = B1/m (xn )
form a base. Moreover, for every n there is some mn such that Bn,m is
relatively compact for m ≤ mn . Since those balls are still a base we are
done. (ii) ⇒ (iii): Take
S the union over the closures of all sets in the base.
(iii) ⇒ (vi): Let X = n Kn with Kn compact. Without loss Kn ⊆ Kn+1 .
For a given compact set K we can find a relatively compact open set V (K)
such that K ⊆ V (K) (cover K by relatively compact open balls and choose
B.5. Compactness 349

a finite subcover). Now define Un = V (Un ). (vi) ⇒ (i): Each of the sets Un
has a countable dense subset by Corollary B.21. The union gives a countable
dense set for X. Since every x ∈ Un for some n, X is also locally compact. 

A sequence of compact sets Un as in (iv) is known as a compact ex-

haustion of X. Since every point of X is an interior point of some Un , every
compact set K ⊆ X is contained in Un for n sufficiently large (cover K by
taking an open neighborhood contained in some Un for every point).
Example B.18. X = (0, 1) with the usual metric is locally compact and
σ-compact but not proper. 
Example B.19. Consider `2 (N) with the standard basis δ j . Let Xj :=
{λδ j |λ ∈ [0, 1]} and note that the metric on Xj inherited from `2 (Z) is
the same as the usual metric from R. Then X := j∈N Xj is a complete
S
separable σ-compact space, which is not locally compact. In fact, consider
a ball of radius ε around zero. Then (ε/2)δ j ∈ Bε (0) is a bounded sequence
√
which has no convergent subsequence since d((ε/2)δ j , (ε/2)δ k ) = ε/ 2 for
k 6= j. 
However, under the assumptions of the previous lemma we can always
switch to a new metric which generates the same topology and for which X
is proper. To this end recall that a function f : X → Y between topological
spaces is called proper if the inverse image of a compact set is again compact.
Now given a proper function (Problem B.30) there is a new metric with the
claimed properties (Problem B.26).
A subset U of a complete metric space X is called totally bounded if
for every ε > 0 it can be covered with a finite number of balls of radius ε.
We will call such a cover an ε-cover. Clearly every totally bounded set is
bounded.
Example B.20. Of course in Rn the totally bounded sets are precisely the
bounded sets. This is in fact true for every proper metric space since the
closure of a bounded set is compact and hence has a finite cover. 

Lemma B.26. Let X be a complete metric space. Then a set is relatively

compact if and only if it is totally bounded.

Proof. Without loss of generality we can assume our set to be closed.

Clearly a compact set K is closed and totally bounded (consider the cover
by all balls of radius ε with center in the set and choose a finite subcover).
Conversely, we will show that K is sequentially compact. So start with
ε1 = 1 and choose a finite cover of balls with radius ε1 . One of these balls
contains an infinite number of elements x1n from our sequence xn . Choose
ε2 = 21 and repeat the process with the sequence x1n . The resulting diagonal
350 B. Metric and topological spaces

sequence xnn gives a subsequence which is Cauchy and hence converges by

completeness. 
Problem B.24 (Alexandroff one-point compactification). Suppose X is a
locally compact Hausdorff space which is not compact. Introduce a new point
∞, set X̂ = X ∪ {∞} and make it into a topological space by calling O ⊆ X̂
open if either ∞ 6∈ O and O is open in X or if ∞ ∈ O and X̂ \ O is compact.
Show that X̂ is a compact Hausdorff space which contains X as a dense
subset.
Problem B.25. Show that every open set O ⊆ R can be written as a count-
able union of disjoint intervals. (Hint: Consider the set {Iα } of all maximal
open subintervals of O; that is, Iα ⊆ O and there is no other subinterval of
O which contains Iα .
Problem B.26. Let (X, d) be a metric space. Show that if there is a proper
function f : X → R, then
˜ y) = d(x, y) + |f (x) − f (y)|
d(x,
˜ is proper.
is a metric which generates the same topology and for which (X, d)

B.6. Separation
The distance between a point x ∈ X and a subset Y ⊆ X is
dist(x, Y ) := inf d(x, y). (B.22)
y∈Y

Note that x is a limit point of Y if and only if dist(x, Y ) = 0 (Problem B.27).

Lemma B.27. Let X be a metric space and Y ⊆ X nonempty. Then
| dist(x, Y ) − dist(z, Y )| ≤ d(x, z). (B.23)
In particular, x 7→ dist(x, Y ) is continuous.

Proof. Taking the infimum in the triangle inequality d(x, y) ≤ d(x, z) +

d(z, y) shows dist(x, Y ) ≤ d(x, z)+dist(z, Y ). Hence dist(x, Y )−dist(z, Y ) ≤
d(x, z). Interchanging x and z shows dist(z, Y ) − dist(x, Y ) ≤ d(x, z). 

A topological space is called normal if for any two disjoint closed sets C1
and C2 , there are disjoint open sets O1 and O2 such that Cj ⊆ Oj , j = 1, 2.
Lemma B.28 (Urysohn). Let X be a topological space. Then X is normal
if and only if for every pair of disjoint closed sets C1 and C2 , there exists a
continuous function f : X → [0, 1] which is one on C1 and zero on C2 .
If in addition X is locally compact and C1 is compact, then f can be
chosen to have compact support.
B.6. Separation 351

Proof. To construct f we choose an open neighborhood O0 of C1 (e.g. O0 :=

X \ C2 ). Now we could set f equal to one on C1 , equal to zero outside O0
and equal to 21 on the layer O0 \ C1 in between. Clearly this function is not
continuous, but we can successively improve the situation by introducing
additional layers in between.
To this end observe that X is normal if and only if for every closed set
C and every open neighborhood U of C, there exists an open set O1 and
a closed set C1 such that C ⊂ O1 ⊂ C1 ⊂ O (just use the identification
C1 ↔ C, C2 ↔ X \ O and O1 ↔ O1 , O2 ↔ X \ C).
Using our observation we can find an open set O1/2 and a closed set C1/2
such that C1 ⊆ O1/2 ⊆ C1/2 ⊆ O0 . Repeating this argument we get two
more open and two more closed sets such that
C1 ⊆ O3/4 ⊆ C3/4 ⊆ O1/2 ⊆ C1/2 ⊆ O1/4 ⊆ C1/4 ⊆ O0 .
Iterating this construction we get open sets Oq and closed sets Cq for every
dyadic rational q = k/2n ∈ [0, 1] such that Oq ⊆ Cq and Cp ⊆ Oq for p < q.
Now set fns (x) := max{q = k/2n |x ∈ Uq , 0 ≤ k < 2n } for x ∈ O0 and
fns (x) := 0 else as well as fni (x) := min{q = k/2n |x 6∈ Cq , 0 ≤ k < 2n } for
x 6∈ C1 and fni (x) := 1 else. Then fns (x) % f s (x) := sup{q|x ∈ Uq } and
fni & f i (s) := inf{q|x 6∈ Cq }. Moreover, if fns (x) = q we have x ∈ Oq \Oq+2−n
and depending on x ∈ Cq+2−n or x 6∈ Cq+2−n we have fni (x) = q + 2−n+1 or
fni (x) = q + 2−nS , respectively. In particular, fTs (x) = f i (x). Finally, since
(f ) ((r, 1]) = q>r Or and (f i )−1 ([0, r)) = q<r X \ Cr are open we see
s −1

that f := f s = f i is continuous.
Conversely, given f choose O1 := f −1 ([0, 1/2)) and O2 := f −1 ((1/2, 1]).
For the second claim, observe that there is an open set O0 such that O0
is compact and C1 ⊂ O0 ⊂ O0 ⊂ X \ C2 . In fact, for every x ∈ C1 , there
is a ball Bε (x) such that Bε (x) is compact and Bε (x) ⊂ X \ C2 . Since C1
is compact, finitely many of them cover C1 and we can choose the union of
those balls to be O0 . 
dist(x,C2 )
Example B.21. In a metric space we can choose f (x) := dist(x,C1 )+dist(x,C2 )
and hence every metric space is normal. 
Another important result is the Tietze extension theorem:
Theorem B.29 (Tietze). Suppose C is a closed subset of a normal topo-
logical space X. For every continuous function f : C → [−1, 1] there is a
continuous extension f¯ : X → [−1, 1].

Proof. The idea is to construct a rough approximation using Urysohn’s

lemma and then iteratively improve this approximation. To this end we set
C1 := f −1 ([ 13 , 1]) and C2 := f −1 ([−1, − 31 ]) and let g be the function from
352 B. Metric and topological spaces

Urysohn’s lemma. Then f1 := 2g−1 3 satisfies |f (x) − f1 (x)| ≤ 32 for x ∈ C as

well as |f1 (x)| ≤ 3 for all x ∈ X. Applying this same procedure to f − f1 we
1

2 2
obtain a function such that for x ∈ C and


f2 |f (x) − f1 (x) − f2 (x)| ≤ 3
|f2 (x)| ≤ 31 23 . Continuing this process we arrive at a sequence of functions



n
n−1
fn such that |f (x) − nj=1 fj (x)| ≤ 23 for x ∈ C and |fn (x)| ≤ 31 32 .
P
By constructionPthe corresponding series converges uniformly to the desired
extension f¯ := ∞ j=1 fj . 

Note that by extending each component we can also handle functions

with values in Rn .
P A partition of unity is a collection of functions hj : X → [0, 1] such that
j hj (x) = 1. We will only consider the case where there are countably many
functions. A partition of unity is locally finite if every x has a neighborhood
where all but a finite number of the functions hj vanish. Moreover, given a
cover {Oj } of X it is called subordinate to this cover if every hj has support
contained in some set Ok from this cover. Of course the partition will be
locally finite if the cover is locally finite which we can always assume without
loss of generality for an open cover if X is a metric space by Lemma B.14.
Lemma B.30. Let X be a metric space and {Oj } a countable open cover.
Then there is a continuous partition of unity subordinate to this cover.
We can even choose the cover such that hj has support contained in Oj .

Proof. For notational simplicity we assume j ∈ N. Now introduce fn (x) :=

min(1, supj≤n d(x, X \Oj )) and gn = fn −fn−1 (with the convention f0 (x) :=
0). Since fn is increasing we have 0 ≤ gn ≤ 1. Moreover, gn (x) > 0
implies d(x, X \ On ))P > 0 and thus supp(gn ) ⊂ On . Next, by monotonicity
f∞ := limn→∞ fn = n gn exists and f∞ (x) = 0 implies d(x, X \ Oj ) = 0
for all j, that is, x 6∈ Oj for all j. Hence f∞ is everywhere positive since
{Oj } is a cover. Finally, by
|fn (x) − fn (y)| ≤ | sup d(x, X \ Oj ) − sup d(y, X \ Oj )|
j≤n j≤n
≤ sup |d(x, X \ Oj ) − d(y, X \ Oj )| ≤ d(x, y)
j≤n

we see that all fn (and hence all gn ) are continuous. Moreover, the very same
argument shows that f∞ is continuous and thus we have found the required
partition of unity hj = gj /f∞ . 

Finally, we also mention that in the case of subsets of Rn there is a

smooth partition of unity. To this end recall that for every point x ∈ Rn
there is a smooth bump function with values in [0, 1] which is positive at x
and supported in a given neighborhood of x.
B.7. Connectedness 353

Example B.22. The standard bump function is φ(x) := exp( |x|21−1 ) for
|x| < 1 and φ(x) = 0 otherwise. To show that this function is indeed smooth
it suffices to show that all left derivatives of f (r) = exp( r−1
1
) at r = 1
vanish, which can be done using l’Hôpital’s rule. By scaling and translation
r )we get a bump function which is supported in Br (x0 ) and satisfies
φ( x−x 0

φ( r ) x=x0 = φ(0) = e−1 .

x−x0 

Lemma B.31. Let X ⊆ Rn be open and {Oj } a countable open cover.
Then there is a locally finite partition of unity of functions from Cc∞ (X)
subordinate to this cover. If the cover is finite so will be the partition.

Proof. Let Uj be as in Lemma B.25 (iv). For U j choose finitely many bump
functions h̃j,k such that h̃j,1 (x) + · · · + h̃j,kj (x) > 0 for every x ∈ U j \ Uj−1
and such that supp(h̃j,k ) is contained in one of the Ok and in Uj+1 \ Uj−1 .
Then {h̃j,k }j,k is locally finite and hence h := j,k h̃j,k is a smooth function
P

which is everywhere positive. Finally, {h̃j,k /h}j,k is a partition of unity of

the required type. 
Problem B.27. Show dist(x, Y ) = dist(x, Y ). Moreover, show x ∈ Y if
and only if dist(x, Y ) = 0.
Problem B.28. Let Y, Z ⊆ X and define
dist(Y, Z) := inf d(y, z).
y∈Y,z∈Z

Show dist(Y, Z) = dist(Y , Z). Moreover, show that if K is compact, then

dist(K, Y ) > 0 if and only if K ∩ Y = ∅.
Problem B.29. Let K ⊆ U with K compact and U open. Show that there
is some ε > 0 such that Kε := {x ∈ X| dist(x, K) < ε} ⊆ U .
Problem B.30. Let (X, d) be a locally compact metric space. Then X is σ-
compact if and only if there exists a proper function f : X → [0, ∞). (Hint:
Let Un be as in item (iv) of Lemma B.25 and use Uryson’s lemma to find
functions fn : X → [0, 1] such thatPf (x) = 0 for x ∈ Un and f (x) = 1 for
x ∈ X \ Un+1 . Now consider f = ∞ n=1 fn .)

B.7. Connectedness
Roughly speaking a topological space X is disconnected if it can be split
into two (nonempty) separated sets. This of course raises the question what
should be meant by separated. Evidently it should be more than just disjoint
since otherwise we could split any space containing more than one point.
Hence we will consider two sets separated if each is disjoint form the closure
of the other. Note that if we can split X into two separated sets X = U ∪ V
then U ∩ V = ∅ implies U = U (and similarly V = V ). Hence both sets
354 B. Metric and topological spaces

must be closed and thus also open (being complements of each other). This
brings us to the following definition:
A topological space X is called disconnected if one of the following
equivalent conditions holds
• X is the union of two nonempty separated sets.
• X is the union of two nonempty disjoint open sets.
• X is the union of two nonempty disjoint closed sets.
In this case the sets from the splitting are both open and closed. A topo-
logical space X is called connected if it cannot be split as above. That
is, in a connected space X the only sets which are both open and closed
are ∅ and X. This last observation is frequently used in proofs: If the set
where a property holds is both open and closed it must either hold nowhere
or everywhere. In particular, any continuous mapping from a connected to
a discrete space must be constant since the inverse image of a point is both
open and closed.
A subset of X is called (dis-)connected if it is (dis-)connected with respect
to the relative topology. In other words, a subset A ⊆ X is disconnected if
there are disjoint nonempty open sets U and V which split A according to
A = (U ∩ A) ∪ (V ∩ A).
Example B.23. In R the nonempty connected sets are precisely the inter-
vals (Problem B.32). Consequently A = [0, 1] ∪ [2, 3] is disconnected with
[0, 1] and [2, 3] being its components (to be defined precisely below). While
you might be reluctant to consider the closed interval [0, 1] as open, it is im-
portant to observe that it is the relative topology which is relevant here. 
The maximal connected subsets (ordered by inclusion) of a nonempty
topological space X are called the connected components of X.
Example B.24. Consider Q ⊆ R. Then every rational point is its own
component (if a set of rational points contains more than one point there
would be an irrational point in between which can be used to split the set). 
In many applications one also needs the following stronger concept. A
space X is called path-connected if any two points x, y ∈ X can be joined
by a path, that is a continuous map γ : [0, 1] → X with γ(0) = x and
γ(1) = y. A space is called locally (path-)connected if for every given
point and every open set containing that point there is a smaller open set
which is (path-)connected.
Example B.25. Every normed vector space is (locally) path-connected since
every ball is path-connected (consider straight lines). In fact this also holds
for locally convex spaces. Every open subset of a locally (path-)connected
space is locally (path-)connected. 
B.7. Connectedness 355

Every path-connected space is connected. In fact, if X = U ∪ V were

disconnected but path-connected we could choose x ∈ U and y ∈ V plus a
path γ joining them. But this would give a splitting [0, 1] = γ −1 (U )∪γ −1 (V )
contradicting our assumption. The converse however is not true in general
as a space might be impassable (an example will follow).
Example B.26. The spaces R and Rn , n > 1, are not homeomorphic. In
fact, removing any point form R gives a disconnected space while removing
a point form Rn still leaves it (path-)connected. 
We collect a few simple but useful properties below.

Lemma B.32. Suppose X and Y are topological spaces.

(i) Suppose f : X → Y is continuous. Then if X is (path-)connected

so is the image f (X).
(ii) Suppose
T
S Aα ⊆ X are (path-)connected and α Aα 6= ∅. Then
α Aα is (path-)connected
(iii) A ⊆ X is (path-)connected if and only if any two points x, y ∈ A
are contained in a (path-)connected set B ⊆ A

(iv) Suppose X1 , . . . , Xn are (path-)connected then so is nj=1 Xj .
(v) Suppose A ⊆ X is connected, then A is connected.
(vi) A locally path-connected space is path-connected if and only if it is
connected.

Proof. (i). Suppose we have a splitting f (X) = U ∪ V into nonempty

disjoint sets which are open in the relative topology. Hence, there are open
sets U 0 and V 0 such that U = U 0 ∩ f (X) and V = V 0 ∩ f (X) implying that
the sets f −1 (U ) = f −1 (U 0 ) and f −1 (V ) = f −1 (V 0 ) are open. Thus we get a
corresponding splitting X = f −1 (U ) ∪ f −1 (V ) into nonempty disjoint open
sets contradicting connectedness of X.
If X is path connected, let y1 = f (x1 ) and y2 = f (x2 ) be given. If γ is
a path connecting x1 and x2 , then f ◦ γ is a path connecting y1 and y2 .
(ii). Let A = α Aα Tand suppose there is a splitting A = (U ∩A)∪(V ∩A).
S
Since there is some x ∈ α Aα we can assume x ∈ U w.l.o.g. Hence there is
a splitting Aα = (U ∩ Aα ) ∪ (V ∩ Aα ) and since Aα is connected and U ∩ Aα
is nonempty we must have V ∩Aα = ∅. Hence V ∩A = ∅ and A is connected.
If the x ∈ Aα and y ∈ Aβ then choose a point z ∈ Aα ∩ Aβ and paths γα
from x to z and γβ from z to y, then γα  γβ is a path from x to y, where
γα  γβ (t) = γα (2t) for 0 ≤ t ≤ 21 and γα  γβ (t) = γβ (2t − 1) for 21 ≤ t ≤ 1
(cf. Problem B.22).
356 B. Metric and topological spaces

(iii). If X is connected we can choose B = A. Conversely, fix some

S x∈A
and let By be the corresponding set for the pair x, y. Then A = y∈A By is
(path-)connected by the previous item.
(iv). We first consider two spaces X = X1 × X2 . Let x, y ∈ X. Then
{x1 } × X2 is homeomorphic to X2 and hence (path-)connected. Similarly
X1 × {y2 } is (path-)connected as well as {x1 } × X2 ∪ X1 × {y2 } by (ii) since
both sets contain (x1 , y2 ) ∈ X. But this last set contains both x, y and hence
the claim follows from (iii). The general case follows by iterating this result.
(v). Let x ∈ A. Then {x} and A cannot be separated and hence {x} ∪ A
is connected. The rest follows from (ii).
(vi). Consider the set U (x) of all points connected to a fixed point
x (via paths). If y ∈ U (x) then so is any path-connected neighborhood
of y by gluing paths (as in item (ii)). Hence U (x) is open. Similarly, if
y ∈ U (x) then any path-connected neighborhood of y will intersect U (y)
and hence y ∈ U (x). Thus U (x) is also closed and hence must be all of X
by connectedness. The converse is trivial. 

A few simple consequences are also worth while noting: If two different
components contain a common point, their union is again connected con-
tradicting maximality. Hence two different components are always disjoint.
Moreover, every point is contained in a component, namely the union of all
connected sets containing this point. In other words, the components of any
topological space X form a partition of X (i.e., they are disjoint, nonempty,
and their union is X). Moreover, every component is a closed subset of the
original space X. In the case where their number is finite we can take com-
plements and each component is also an open subset (the rational numbers
from our first example show that components are not open in general). In a
locally (path-)connected space, components are open and (path-)connected
by (vi) of the last lemma. Note also that in a second countable space an
open set can have at most countably many components (take those sets from
a countable base which are contained in some component, then we have a
surjective map from these sets to the components).
Example B.27. Consider the graph of the function f : (0, 1] → R, x 7→
sin( x1 ). Then Γ(f ) ⊆ R2 is path-connected and its closure Γ(f ) = Γ(f ) ∪
{0} × [−1, 1] is connected. However, Γ(f ) is not path-connected as there is
no path from (1, 0) to (0, 0). Indeed, suppose γ were such a path. Then,
since γ1 covers [0, 1] by the intermediate value theorem (see below), there is
a sequence tn → 1 such that γ1 (tn ) = (2n+1)π
2
. But then γ2 (tn ) = (−1)n 6→ 0
contradicting continuity. 
B.8. Continuous functions on metric spaces 357

Theorem B.33 (Intermediate Value Theorem). Let X be a connected topo-

logical space and f : X → R be continuous. For any x, y ∈ X the function f
attains every value between f (x) and f (y).

Proof. The image f (X) is connected and hence an interval. 

Problem B.31. A nonempty subset of R is connected if and only if it is an
interval.
Problem B.32. Let U = · j Uj ⊆ Rn be an open set with Uj its connected
S
components. ShowS∂Uj ⊆ ∂U . Show that in the case of finitely many com-
ponents we have nj=1 ∂Uj = ∂U . Show that this fails for infinitely many
components in general.

B.8. Continuous functions on metric spaces

Let X, Y be topological spaces and let C(X, Y ) be the set of all continuous
functions f : X → Y . Set C(X) := C(X, C). Moreover, if Y is a metric
space then Cb (X, Y ) will denote the set of all bounded continuous functions,
that is, those continuous functions for which supx∈X dY (f (x), y) is finite for
some (and hence for all) y ∈ Y . Note that by the extreme value theorem
Cb (X, Y ) = C(X, Y ) if X is compact. For these functions we can introduce
a metric via
d(f, g) := sup dY (f (x), g(x)). (B.24)
x∈X
In fact, the requirements for a metric are readily checked. Of course conver-
gence with respect to this metric implies pointwise convergence but not the
other way round.
Example B.28. Consider X := [0, 1], then fn (x) := max(1 − |nx − 1|, 0)
converges pointwise to 0 (in fact, fn (0) = 0 and fn (x) = 0 on [ n2 , 1]) but not
with respect to the above metric since fn ( n1 ) = 1. 
This kind of convergence is known as uniform convergence since for ev-
ery positive ε there is some index N (independent of x) such that dY (fn (x), f (x)) <
ε for n ≥ N . In contradistinction, in the case of pointwise convergence, N
is allowed to depend on x. One advantage is that continuity of the limit
function comes for free.
Theorem B.34. Let X be a topological space and Y a metric space. Suppose
fn ∈ C(X, Y ) converges uniformly to some function f : X → Y . Then f is
continuous.

Proof. Let x ∈ X be given and write y := f (x). We need to show that

f −1 (Bε (y)) is a neighborhood of x for every ε > 0. So fix ε. Then we can find
−1
an N such that d(fn , f ) < 2ε for n ≥ N implying fN (Bε/2 (y)) ⊆ f −1 (Bε (y))
358 B. Metric and topological spaces

since d(fn (z), y) < 2ε implies d(f (z), y) ≤ d(f (z), fn (z)) + d(fn (z), y) ≤
2 + 2 = ε for n ≥ N .
ε ε

Corollary B.35. Let X be a topological space and Y a complete metric
space. The space Cb (X, Y ) together with the metric d is complete.

Proof. Suppose fn is a Cauchy sequence with respect to d, then fn (x)

is a Cauchy sequence for fixed x and has a limit since Y is complete.
Call this limit f (x). Then dY (f (x), fn (x)) = limm→∞ dY (fm (x), fn (x)) ≤
supm≥n d(fm , fn ) and since this last expression goes to 0 as n → ∞, we see
that fn converges uniformly to f . Moreover, f ∈ C(X, Y ) by the previous
theorem so we are done. 

Let Y be a vector space. By Cc (X, Y ) ⊆ Cb (X, Y ) we will denote the set

of continuous functions with compact support. Its closure will be denoted
by C0 (X, Y ) := Cc (X, Y ) ⊆ Cb (X, Y ). Of course if X is compact all these
spaces agree Cc (X, Y ) = C0 (X, Y ) = Cb (X, Y ) = C(X, Y ). In the general
case one at least assumes X to be locally compact since if we take a closed
neighborhood V of f (x) 6= 0 which does not contain 0, then f −1 (U ) will be
a compact neighborhood of x. Hence without this assumption f must vanish
on every point which does not have a compact neighborhood and Cc (X, Y )
will not be sufficiently rich.
Example B.29. Let X be a separable and locally compact metric space and
Y = Cn . Then
C0 (X, Cn ) = {f ∈ Cb (X, Cn )| ∀ε > 0, ∃K ⊆ X compact : (B.25)
|f (x)| < ε, x ∈ X \ K}.
To see this denote the set on the right-hand side by C. Let Km be an
increasing sequence of compact sets with Km % X (Lemma B.25) and let ϕm
be a corresponding sequence as in Urysohn’s lemma (Lemma B.28). Then for
f ∈ C the sequence fm = ϕm f ∈ Cc (X, Cn ) will converge to f . Conversely,
if fn ∈ Cc (X, Cn ) converges to f ∈ Cb (X, Cn ), then given ε > 0 choose
K = supp(fm ) for some m with d(fm , f ) < ε.
In the case where X is an open subset of Rn this says that C0 (X, Y ) are
those which vanish at the boundary (including the case as |x| → ∞ if X is
unbounded). 
Lemma B.36. If X is a separable and locally compact space then C0 (X, Cn )
is separable.

Proof. Choose a countable base B for X and let I the collection of all balls in
Cn with rational radius and center. Given O1 , . . . , Om ∈ B and I1 , . S
. . , Im ∈
I we say that f ∈ Cc (X, Cn ) is adapted to these sets if supp(f ) ⊆ m j=1 Oj
and f (Oj ) ⊆ Ij . The set of all tuples (Oj , Ij )1≤j≤m is countable and for
B.8. Continuous functions on metric spaces 359

each tuple we choose a corresponding adapted function (if there exists one
at all). Then the set of these functions F is dense. It suffices to show that
the closure of F contains Cc (X, Cn ). So let f ∈ Cc (X, Cn ) and let ε > 0
be given. Then for every x ∈ X there is some neighborhood O(x) ∈ B such
that |f (x) − f (y)| < ε for y ∈ O(x). Since supp(f ) is compact, it can be
covered by O(x1 ), . . . , O(xm ). In particular f (O(xj )) ⊆ Bε (f (xj )) and we
can find a ball Ij of radius at most 2ε with f (O(xj )) ⊆ Ij . Now let g be
the function from F which is adapted to (O(xj ), Ij )1≤j≤m and observe that
|f (x) − g(x)| < 4ε since x ∈ O(xj ) implies f (x), g(x) ∈ Ij . 

Let X, Y be metric spaces. A function f ∈ C(X, Y ) is called uniformly

continuous if for every ε > 0 there is a δ > 0 such that
dY (f (x), f (y)) ≤ ε whenever dX (x, y) < δ. (B.26)
Note that with the usual definition of continuity on fixes x and then chooses
δ depending on x. Here δ has to be independent of x. If the domain is
compact, this extra condition comes for free.
Theorem B.37. Let X be a compact metric space and Y a metric space.
Then every f ∈ C(X, Y ) is uniformly continuous.

Proof. Suppose the claim were wrong. Fix ε > 0. Then for every δn = n1
we can find xn , yn with dX (xn , yn ) < δn but dY (f (xn ), f (yn )) ≥ ε. Since
X is compact we can assume that xn converges to some x ∈ X (after pass-
ing to a subsequence if necessary). Then we also have yn → x implying
dY (f (xn ), f (yn )) → 0, a contradiction. 

Note that a uniformly continuous function maps Cauchy sequences to

Cauchy sequences. This fact can be used to extend a uniformly continuous
function to boundary points.
Theorem B.38. Let X be a metric space and Y a complete metric space.
A uniformly continuous function f : A ⊆ X → Y has a unique continuous
extension f¯ : A → Y . This extension is again uniformly continuous.

Proof. If there is an extension it must be f¯(x) := limn→∞ f (xn ), where xn ∈

A is some sequence converging to x ∈ A. Indeed, since xn converges, f (xn )
is Cauchy and hence has a limit since Y is assumed complete. Moreover,
uniqueness of limits shows that f¯(x) is independent of the sequence chosen.
Also f¯(x) = f (x) for x ∈ A by continuity. To see that f¯ is uniformly
continuous, let ε > 0 be given and choose a δ which works for f . Then for
given x, y with dX (x, y) < 3δ we can find x̃, ỹ ∈ A with dX (x̃, x) < 3δ and
dY (f (x̃), f¯(x)) ≤ ε as well as dX (ỹ, y) < 3δ and dY (f (ỹ), f¯(y)) ≤ ε. Hence
dY (f¯(x), f¯(y)) ≤ dY (f¯(x), f (x̃)) + dY (f (x̃), f (ỹ)) + dY (f (x), f¯(y)) ≤ 3ε. 
360 B. Metric and topological spaces

Next we want to identify relatively compact subsets in C(X, Y ). A family

of functions F ⊂ C(X, Y ) is called (pointwise) equicontinuous if for every
ε > 0 and every x ∈ X there is a neighborhood U (x) of x such that
dY (f (x), f (y)) ≤ ε whenever y ∈ U (x), ∀f ∈ F. (B.27)
Theorem B.39 (Arzelà–Ascoli). Let X be a compact space and Y a proper
metric space. Let F ⊂ C(X, Y ) be a family of continuous functions. Then
every sequence from F has a uniformly convergent subsequence if and only if
F is equicontinuous and the set {f (x)|f ∈ F } is bounded for every x ∈ X.
In this case F is even bounded.

Proof. Suppose F is equicontinuous and pointwise bounded. Fix ε > 0. By

compactness of X there are finitely many points x1 , . . . , xn ∈ X such that the
neighborhoods U (xj ) (from the definition of equicontinuity) cover X. Now
first of all note that, F is bounded since dY (f (x), y) ≤ maxj supf ∈F dY (f (xj ), y)+
ε for every x ∈ X and every f ∈ F .
Next consider P : C(X, Y ) → Y n , P (f ) = (f (x1 ), . . . , f (xn )). Then
P (F ) is bounded and d(f, g) ≤ 3ε whenever dY (P (f ), P (g)) < ε. Indeed,
just note that for every x there is some j such that x ∈ U (xj ) and thus
dY (f (x), g(x)) ≤ dY (f (x), f (xj )) + dY (f (xj ), g(xj )) + dY (g(xj ), g(x)) ≤ 3ε.
Hence F is relatively compact by Lemma 1.11.
Conversely, suppose F is relatively compact. Then F is totally bounded
and hence bounded. To see equicontinuity fix x ∈ X, ε > 0 and choose a cor-
responding ε-cover {Bε (fj )}nj=1 for F . Pick a neighbrohood U (x) such that
y ∈ U (x) implies dY (fj (y), fj (x)) < ε for all 1 ≤ j ≤ n. Then f ∈ Bε (fj )
for some j and hence dY (f (y), f (x)) ≤ dY (f (y), fj (y)) + dY (fj (y), fj (x)) +
dY (fj (x), f (x)) ≤ 3ε, proving equicontinuity. 

In many situations a certain property can be seen for a class of nice

functions and then extended to a more general class of functions by approx-
imation. In this respect it is important to identify classes of functions which
allow to approximate all functions. That is, in our present situation we are
looking for functions which are dense in C(X, Y ). For example, the classical
Weierstraß approximation theorem (Theorem 1.3) says that the polynomi-
als are dense in C([a, b]) for any compact interval. Here we will present a
generalization of this result. For its formulation observe that C(X) is not
only a vector space but also comes with a natural product, given by point-
wise multiplication of functions, which turns it into an algebra over C. By
a subalgebra we will mean a subspace which is closed under multiplication
and by a ∗-subalgebra we will mean a subalgebra which is also closed under
complex conjugation. The (∗-)subalgebra generated by a set is of course the
smallest (∗-)subalgebra containing this set.
B.8. Continuous functions on metric spaces 361

The proof will use the fact that the absolute value can be approximated
by polynomials on [−1, 1]. This of course follows from the Weierstraß ap-
proximation theorem but can also be seen directly by defining the sequence
of polynomials pn via
t2 − pn (t)2
p1 (t) := 0, pn+1 (t) := pn (t) + . (B.28)
2
Then this sequence of polynomials satisfies pn (t) ≤ pn+1 (t) ≤ |t| and con-
verges pointwise to |t| for t ∈ [−1, 1]. Hence by Dini’s theorem (Prob-
lem B.35) it converges uniformly. By scaling we get the corresponding result
for arbitrary compact subsets of the real line.

Theorem B.40 (Stone–Weierstraß, real version). Suppose K is a compact

topological space and consider C(K, R). If F ⊂ C(K, R) contains the identity
1 and separates points (i.e., for every x1 6= x2 there is some function f ∈ F
such that f (x1 ) 6= f (x2 )), then the subalgebra generated by F is dense.

Proof. Denote by A the subalgebra generated by F . Note that if

 f 1∈ A,
we have |f | ∈ A: Choose a polynomial pn (t) such that |t| − pn (t) < n for


t ∈ f (K) and hence pn (f ) → |f |.
In particular, if f, g ∈ A, we also have
(f + g) + |f − g| (f + g) − |f − g|
max{f, g} = ∈ A, min{f, g} = ∈ A.
2 2
Now fix f ∈ C(K, R). We need to find some f ε ∈ A with kf − f ε k∞ < ε.
First of all, since A separates points, observe that for given y, z ∈ K
there is a function fy,z ∈ A such that fy,z (y) = f (y) and fy,z (z) = f (z)
(show this). Next, for every y ∈ K there is a neighborhood U (y) such that
fy,z (x) > f (x) − ε, x ∈ U (y),
and since K is compact, finitely many, say U (y1 ), . . . , U (yj ), cover K. Then
fz = max{fy1 ,z , . . . , fyj ,z } ∈ A
and satisfies fz > f − ε by construction. Since fz (z) = f (z) for every z ∈ K,
there is a neighborhood V (z) such that
fz (x) < f (x) + ε, x ∈ V (z),
and a corresponding finite cover V (z1 ), . . . , V (zk ). Now
f ε = min{fz1 , . . . , fzk } ∈ A
satisfies f ε < f + ε. Since f − ε < fzl for all 1 ≤ l ≤ k we have f − ε < f ε
and we have found a required function. 
362 B. Metric and topological spaces

Example B.30. The set {f ∈ C(K, R)|f (x0 ) = 0} for some x0 ∈ K is a

closed algebra which, in particular, is not dense. The same is true for the set
{f ∈ C(K, R)|f (x1 ) = f (x2 )} for some x1 , x2 ∈ K. These examples show
that the above two conditions are also necessary. 

Theorem B.41 (Stone–Weierstraß). Suppose K is a compact topological

space and consider C(K). If F ⊂ C(K) contains the identity 1 and separates
points, then the ∗-subalgebra generated by F is dense.

Proof. Just observe that F̃ = {Re(f ), Im(f )|f ∈ F } satisfies the assump-
tion of the real version. Hence every real-valued continuous function can be
approximated by elements from the subalgebra generated by F̃ ; in particular,
this holds for the real and imaginary parts for every given complex-valued
function. Finally, note that the subalgebra spanned by F̃ is contained in the
∗-subalgebra spanned by F . 

Note that the additional requirement of being closed under complex con-
jugation is crucial: The functions holomorphic on the unit disc and contin-
uous on the boundary separate points, but they are not dense (since the
uniform limit of holomorphic functions is again holomorphic).

Corollary B.42. Suppose K is a compact topological space and consider

C(K). If F ⊂ C(K) separates points, then the closure of the ∗-subalgebra
generated by F is either C(K) or {f ∈ C(K)|f (t0 ) = 0} for some t0 ∈ K.

Proof. There are two possibilities: either all f ∈ F vanish at one point
t0 ∈ K (there can be at most one such point since F separates points) or
there is no such point.
If there is no such point, then the identity can be approximated by
elements in A: First of all note that |f | ∈ A if f ∈ A, since the polynomials
pn (t) used to prove this fact can be replaced by pn (t)−pn (0) which contain no
constant term. Hence for every point y we can find a nonnegative function
in A which is positive at y and by compactness we can find a finite sum
of such functions which is positive everywhere, say m ≤ f (t) ≤ M . Now
approximate min(m−1 t, t−1 ) by polynomials qn (t) (again a constant term is
not needed) to conclude that qn (f ) → f −1 ∈ A. Hence 1 = f · f −1 ∈ A as
claimed and so A = C(K) by the Stone–Weierstraß theorem.
If there is such a t0 we have A ⊆ {f ∈ C(K)|f (t0 ) = 0} and the identity
is clearly missing from A. However, adding the identity to A we get A + C =
C(K) by the Stone–Weierstraß theorem. Moreover, if f ∈ C(K) with f (t0 ) =
0 we get f = f˜ + α with f˜ ∈ A and α ∈ C. But 0 = f (t0 ) = f˜(t0 ) + α = α
implies f = f˜ ∈ A, that is, A = {f ∈ C(K)|f (t0 ) = 0}. 
B.8. Continuous functions on metric spaces 363

Problem B.33. Suppose X is compact and connected and let F ⊂ C(X, Y )

be a family of equicontinuous functions. Then {f (x0 )|f ∈ F } bounded for
one x0 implies F bounded.
Problem B.34. Let X, Y be metric spaces. A family of functions F ⊂
C(X, Y ) is called uniformly equicontinuous if for every ε > 0 there is a
δ > 0 such that
dY (f (x), f (y)) ≤ ε whenever dX (x, y) < δ, ∀f ∈ F. (B.29)
Show that if X is compact, then a family F is pointwise equicontinuous if
and only if it is uniformly equicontinuous.
Problem* B.35 (Dini’s theorem). Suppose X is compact and let fn ∈ C(X)
be a sequence of decreasing (or increasing) functions converging pointwise
fn (x) & f (x) to some function f ∈ C(X). Then fn → f uniformly. (Hint:
Reduce it to the case fn & 0 and apply the finite intersection property to
fn−1 ([ε, ∞).)
Problem B.36. Let k ∈ N and I ⊆ R. Show that the ∗-subalgebra generated
by fz0 (t) = (t−z1 )k for one z0 ∈ C is dense in the set C0 (I) of continuous
0
functions vanishing at infinity:
• for I = R if z0 ∈ C\R and k = 1 or k = 2,
• for I = [a, ∞) if z0 ∈ (−∞, a) and k arbitrary,
• for I = (−∞, a] ∪ [b, ∞) if z0 ∈ (a, b) and k odd.
(Hint: Add ∞ to R to make it compact.)
Problem B.37. Let U ⊆ C\R be a set which has a limit point and is
symmetric under complex conjugation. Show that the span of {(t − z)−1 |z ∈
U } is dense in the set C0 (R) of continuous functions vanishing at infinity.
(Hint: The product of two such functions is in the span provided they are
different.)
Problem B.38. Let K ⊆ C be a compact set. Show that the set of all
functions f (z) = p(x, y), where p : R2 → C is polynomial and z = x + iy, is
dense in C(K).
Bibliography

[1] H. W. Alt, Lineare Funktionalanalysis, 4th ed., Springer, Berlin, 2002.

[2] H. Bauer, Measure and Integration Theory, de Gruyter, Berlin, 2001.
[3] M. Berger and M. Berger, Perspectives in Nonlinearity, Benjamin, New York,
1968.
[4] A. Bowers and N. Kalton, An Introductory Course in Functional Analysis,
Springer, New York, 2014.
[5] H. Brezis, Functional Analysis, Sobolev Spaces and Partial Differential Equa-
tions, Springer, New York, 2011.
[6] S.-N. Chow and J. K. Hale, Methods of Bifurcation Theory, Springer, New York,
1982.
[7] J. B. Conway, A Course in Functional Analysis, 2nd ed., Springer, New York,
1994.
[8] K. Deimling, Nichtlineare Gleichungen und Abbildungsgrade, Springer, Berlin,
1974.
[9] K. Deimling, Nonlinear Functional Analysis, Springer, Berlin, 1985.
[10] E. DiBenedetto, Real Analysis, Birkhäuser, Boston, 2002.
[11] L. C. Evans, Weak Convergence Methods for nonlinear Partial Differential Equa-
tions, CBMS 74, American Mathematical Society, Providence, 1990.
[12] L. C. Evans, Partial Differential Equations, 2nd ed., American Mathematical
Society, Providence, 2010.
[13] G. B. Folland, Real Analysis: Modern Techniques and Their Applications, 2nd.
ed., Wiley, Hoboken NJ, 1999.
[14] J. Franklin, Methods of Mathematical Economics, Springer, New York, 1980.
[15] I. Gohberg, S. Goldberg, and M.A. Kaashoek, Basic Classes of Linear Opeartors,
Springer, Basel, 2003.
[16] J. Goldstein, Semigroups of Linear Operators and Appications, Oxford University
Press, New York, 1985.
[17] L. Grafakos, Classical Fourier Analysis, 2nd ed., Springer, New York, 2008.

365
366 Bibliography

[18] L. Grafakos, Modern Fourier Analysis, 2nd ed., Springer, New York, 2009.
[19] G. Grubb, Distributions and Operators, Springer, New York, 2009.
[20] E. Hewitt and K. Stromberg, Real and Abstract Analysis, Springer, Berlin, 1965.
[21] D. Hundertmark, M. Meyries, L. Machinek, and R. Schnaubelt, Operator Semi-
groups and Dispersive Equations, Lecture Notes (16th Internet Seminar on Evo-
lution Equations), 2013. https://isem.math.kit.edu/images/b/b3/Isem16_
final.pdf
[22] K. Jänich, Toplogy, Springer, New York, 1995.
[23] I. Kaplansky, Set Theory and Metric Spaces, AMS Chelsea, Providence, 1972.
[24] T. Kato, Perturbation Theory for Linear Operators, Springer, New York, 1966.
[25] J. L. Kelley, General Topology, Springer, New York, 1955.
[26] O. A. Ladyzhenskaya, The Boundary Values Problems of Mathematical Physics,
Springer, New York, 1985.
[27] P. D. Lax, Functional Analysis, Wiley, New York, 2002.
[28] E. Lieb and M. Loss, Analysis, 2nd ed., Amer. Math. Soc., Providence, 2000.
[29] F. Linares and G. Ponce, Introduction to Nonlinear Dispersive Equations, 2nd
ed., Springer, New York, 2015.
[30] G. Leoni, A First Course in Sobolev Spaces, Amer. Math. Soc., Providence, 2009.
[31] N. Lloyd, Degree Theory, Cambridge University Press, London, 1978.
[32] R. Meise and D. Vogt, Introduction to Functional Analysis, Oxford University
Press, Oxford, 2007.
[33] F. W. J. Olver et al., NIST Handbook of Mathematical Functions, Cambridge
University Press, Cambridge, 2010.
[34] I. K. Rana, An Introduction to Measure and Integration, 2nd ed., Amer. Math.
Soc., Providence, 2002.
[35] M. Reed and B. Simon, Methods of Modern Mathematical Physics I. Functional
Analysis, rev. and enl. edition, Academic Press, San Diego, 1980.
[36] J. R. Retherford, Hilbert Space: Compact Operators and the Trace Theorem,
Cambridge University Press, Cambridge, 1993.
[37] J.J. Rotman, Introduction to Algebraic Topology, Springer, New York, 1988.
[38] H. Royden, Real Analysis, Prencite Hall, New Jersey, 1988.
[39] W. Rudin, Real and Complex Analysis, 3rd edition, McGraw-Hill, New York,
1987.
[40] M. Ru̇žička, Nichtlineare Funktionalanalysis, Springer, Berlin, 2004.
[41] H. Schröder, Funktionalanalysis, 2nd ed., Harri Deutsch Verlag, Frankfurt am
Main 2000.
[42] B. Simon, A Comprehensive Course in Analysis, Amer. Math. Soc., Providence,
2015.
[43] L. A. Steen and J. A. Seebach, Jr., Counterexamples in Topology, Springer, New
York, 1978.
[44] T. Tao, Nonlinear Dispersive Equations: Local and Global Analysis, Amer. Math.
Soc., Providence, 2006.
[45] M. E. Taylor, Measure Theory and Integration, Amer. Math. Soc., Providence,
2006.
Bibliography 367

[46] G. Teschl, Mathematical Methods in Quantum Mechanics; With Applications to

Schrödinger Operators, Amer. Math. Soc., Providence, 2009.
[47] G. Teschl, Topics in Real Analysis, Amer. Math. Soc., Providence, to appear.
[48] J. Weidmann, Lineare Operatoren in Hilberträumen I: Grundlagen, B.G.Teubner,
Stuttgart, 2000.
[49] D. Werner, Funktionalanalysis, 7th edition, Springer, Berlin, 2011.
[50] M. Willem, Functional Analysis, Birkhäuser, Basel, 2013.
[51] E. Zeidler, Applied Functional Analysis: Applications to Mathematical Physics,
Springer, New York 1995.
[52] E. Zeidler, Applied Functional Analysis: Main Principles and Their Applications,
Springer, New York 1995.
Glossary of notation

arg(z) . . . argument of z ∈ C; arg(z) ∈ (−π, π], arg(0) = 0

Br (x) . . . open ball of radius r around x, 330
B(X) . . . Banach space of bounded measurable functions
B = B1
Bn . . . Borel σ-algebra of Rn , see [47]
C . . . the set of complex numbers
C(U ) . . . set of continuous functions from U to C
C0 (U ) . . . set of continuous functions vanishing on the
boundary ∂U , 358
Cc (U ) . . . set of compactly supported continuous functions
Cper [a, b] . . . set of periodic continuous functions (i.e. f (a) = f (b))
C k (U ) . . . set of k times continuously differentiable functions
Cc∞ (U ) . . . set of compactly supported smooth functions
C(U, Y ) . . . set of continuous functions from U to Y , 196
C r (U, Y ) . . . set of r times continuously differentiable
functions, 202
Cbr (U, Y ) . . . functions in C r with derivatives bounded, 37, 207
Ccr (U, Y ) . . . functions in C r with compact support
c0 (N) . . . set of sequences converging to zero, 11
C (X, Y ) . . . set of compact linear operators from X to Y , 65
C(U, Y ) . . . set of compact maps from U to Y , 302
CP(f ) . . . critical points of f , 281
CS(K) . . . nonempty convex subsets of K, 294

369
370 Glossary of notation

CV(f ) . . . critical values of f , 281

χΩ (.) . . . characteristic function of the set Ω
D(.) . . . domain of an operator
δn,m . . . Kronecker delta, 12
deg(D, f, y) . . . mapping degree, 281, 288
det . . . determinant
dim . . . dimension of a linear space
div . . . divergence of a vector filed, 287
diam(U ) = sup(x,y)∈U 2 d(x, y) diameter of a set
dist(U, V ) = inf (x,y)∈U ×V d(x, y) distance of two sets
Dyr (U , Y ) . . . functions in C r (U , Y ) which do not attain y on the boundary, 281
C¯y (U, Y ) . . . functions in C(U , Y ) which do not attain y on the boundary, 304
e . . . Napier’s constant, ez = exp(z)
epi F . . . epigraph of F , 220
dF . . . derivative of F , 196
F(X, Y ) . . . set of compact finite dimensional functions, 302
Φ(X, Y ) . . . set of all linear Fredholm operators from X to Y , 185
Φ0 (X, Y ) . . . set of all linear Fredholm operators of index 0, 185
GL(n) . . . general linear group in n dimensions
Γ(f1 , . . . , fn ) . . . Gram determinant, 47
H . . . a Hilbert space
conv(.) . . . convex hull
H(U ) . . . set of holomorphic functions on a domain U ⊆ C, 279
H k (U ) = W k,2 (U ), Sobolev space
H0k (U ) = W0k,2 (U ), Sobolev space
i . . . complex unity, i2 = −1
Im(.) . . . imaginary part of a complex number
inf . . . infimum
Jf (x) = det df (x) Jacobi determinant of f at x, 281
Ker(A) . . . kernel of an operator A, 27
λn . . . Lebesgue measure in Rn , see [47]
L (X, Y ) . . . set of all bounded linear operators from X to Y , 30
L (X) = L (X, X)
Lp (X, dµ) . . . Lebesgue space of p integrable functions, see [47]
L∞ (X, dµ) . . . Lebesgue space of bounded functions, see [47]
Lploc (X, dµ) . . . locally p integrable functions, see [47]
L2cont (I) . . . space of continuous square integrable functions, 20
`p (N) . . . Banach space of p summable sequences, 9
`2 (N) . . . Hilbert space of square summable sequences, 18
`∞ (N) . . . Banach space of bounded sequences, 10
max . . . maximum
Glossary of notation 371

N . . . the set of positive integers

N0 = N ∪ {0}
n(γ, z0 ) . . . winding number
O(.) . . . Landau symbol, f = O(g) iff lim supx→x0 |f (x)/g(x)| < ∞
o(.) . . . Landau symbol, f = o(g) iff limx→x0 |f (x)/g(x)| = 0
Q . . . the set of rational numbers
R . . . the set of real numbers
ρ(A) . . . resolvent set of an operator A, 158
RV(f ) . . . regular values of f , 281
Ran(A) . . . range of an operator A, 28
Re(.) . . . real part of a complex number
R(I, X) . . . set of regulated functions, 194
σ(A) . . . spectrum of an operator A, 158 ,168
S n−1 = {x ∈ Rn | |x| = 1} unit sphere in Rn
sign(z) = z/|z| for z 6= 0 and 1 for z = 0; complex sign function
S(I, X) . . . step functions f : I → X, 194
sup . . . supremum
supp(f ) . . . support of a function f , 340
span(M ) . . . set of finite linear combinations from M , 12
W k,p (U ) . . . Sobolev space, see [47]
W0k,p (U ) . . . Sobolev space, see [47]
Z . . . the set of integers
I . . . identity operator
√
z . . . square root of z with branch cut along (−∞, 0)
z∗ . . . complex conjugation
A∗ . . . adjoint of of operators A, 52
A . . . closure of of operators A, 103
fˆ = Ff , Fourier coefficients/transform of f , 58
fˇ =F qPf , inverse Fourier transform of f
−1
n
|x| = j=1 |xj | Euclidean norm in R or C
2 n n

|Ω| . . . Lebesgue measure of a Borel set Ω

k.k . . . norm, 17
k.kp . . . norm in the Banach space `p and Lp , 9, 24
h., ..i . . . scalar product in H, 17
372 Glossary of notation

⊕ . . . direct/orthogonal sum of vector spaces or operators, 33, 56

⊕ˆ . . . direct sum of operators with the same image space, 33
⊗ . . . tensor product, 57
∪· . . . union of disjoint sets
bxc = max{n ∈ Z|n ≤ x}, floor function
dxe = min{n ∈ Z|n ≥ x}, ceiling function
∂ = (∂1 f, . . . , ∂m f ) gradient in Rm
∂α . . . partial derivative in multi-index notation
∂x F (x, y) . . . partial derivative with respect to x, 201
∂U = U \ U ◦ boundary of the set U , 330
U . . . closure of the set U , 334
U◦ . . . interior of the set U , 334
M⊥ . . . orthogonal complement, 48
(λ1 , λ2 ) = {λ ∈ R | λ1 < λ < λ2 }, open interval
[λ1 , λ2 ] = {λ ∈ R | λ1 ≤ λ ≤ λ2 }, closed interval
xn → x . . . norm convergence, 8
xn * x . . . weak convergence, 122
∗
xn * x . . . weak-∗ convergence, 129
An → A . . . norm convergence of operators
s
An → A . . . strong convergence of operators, 127
An * A . . . weak convergence of operators, 127
Index

sigma-comapct, 348 boundary condition, 6

boundary point, 330
absolute convergence, 15 boundary value problem, 6
absolutely convex, 134, 146 bounded
absorbing set, 131 operator, 28
accumulation point, 330 sesquilinear form, 23
adjoint operator, 51, 116 set, 335
Alexandroff extension, 350 Brouwer fixed point theorem, 292
almost periodic, 46
analytic, 159 calculus of variations, 208
annihilator, 118 Calkin algebra, 179
ascent, 170 Cauchy sequence, 336
Axiom of Choice, 323 weak, 122
axiomatic set theory, 321 Cauchy–Bunyakovsky–Schwarz inequality,
see Cauchy–Schwarz inequality
Baire category theorem, 97 Cauchy–Schwarz inequality, 19
balanced set, 149 Cayley transform, 167
ball Cesàro mean, 116
closed, 334 chain rule, 202
open, 330 character, 178
Banach algebra, 31, 156 Chebyshev polynomials, 69
Banach limit, 116 closed
Banach space, 8 ball, 334
Banach–Steinhaus theorem, 99 set, 333
base, 332 closure, 334
Basel problem, 81 cluster point, 330
basis, 12 codimension, 34
orthonormal, 43 coercive, 54, 316
Bernoulli numbers, 82 weakly, 214
Bessel inequality, 42 cokernel, 34
best reply, 296 compact, 345
bidual space, 112 locally, 348
bifurcation point, 231 sequentially, 346
bijective, 339 compact map, 302
biorthogonal system, 12, 112 complemented subspace, 34
Bolzano–Weierstraß theorem, 348 complete, 8, 336

373
374 Index

completion, 23 eigenspace, 68
complexification, 36 eigenvalue, 68
component, 354 algebraic multiplicity, 171
conjugate linear, 17 geometric multiplicity, 171
connected, 354 index, 171
continuous, 339 simple, 68
contraction principle, 221 eigenvector, 68
uniform, 222 order, 171
contraction semigroup, 251 elliptic problem, 318
convergence, 335 epigraph, 220
strong, 127 equicontinuous, 26, 360
weak, 122 uniformly, 363
weak-∗, 129 equilibrium
convex, 8 Nash, 296
absolutely, 134, 146 equivalent norms, 21
cover, 344 exact sequence, 121, 186
locally finite, 344 exhaustion, 349
open, 344 extended real numbers, 332
refinement, 344 extremal
critical value, 281 point, 136
C ∗ algebra, 164 subset, 136
cylinder, 342 Extreme value theorem, 348
set, 342
Fσ set, 98
face, 137
De Morgan’s laws, 333 fat set, 98
decent, 170 Fejér kernel, 61
delay differential equation, 255 final topology, 343
demicontinuous, 318 finite dimensional map, 302
dense, 337 finite intersection property, 345
derivative first category, 98
Fréchet, 196 first countable, 332
Gâteaux, 198 first resolvent identity, 164, 257
partial, 201 fixed point theorem
variational, 198 Altman, 308
diffeomorphism, 202 Brouwer, 292
differentiable, 201 contraction principle, 221
differential equations, 224 Kakutani, 294
diffusion equation, 3 Krasnosel’skii, 308
dimension, 45 Rothe, 308
direct sum, 33 Schauder, 306
directed, 146 Weissinger, 221
Dirichlet kernel, 59 form
Dirichlet problem, 55 bounded, 23
disconnected, 354 Fourier series, 44, 58
discrete set, 330 cosine, 81
discrete topology, 331 sine, 80
disjoint union topology, 343 FPU lattice, 228
dissipative, 253 Fréchet derivative, 196
distance, 329, 350 Fréchet space, 148
divergence, 287 Fredholm alternative, 172
domain, 27 Fredholm operator, 185
double dual space, 112 Frobenius norm, 91
dual basis, 30 from domain, 78
dual space, 30 function
duality set, 252 open, 340
Duhamel formula, 238, 244 fundamental theorem of algebra, 160
Index 375

fundamental theorem of calculus, 195 Jacobi operator, 69, 228, 229, 239
Jacobi theta function, 7
Gδ set, 98 Jacobson radical, 181
Gâteaux derivative, 198 Jordan curve theorem, 300
Galerkin approximation, 319
gauge, 131 Kakutani’s fixed point theorem, 294
Gelfand transform, 180 kernel, 27
global solution, 227 Kronecker delta, 12
Gram determinant, 47 Kuratowski closure axioms, 334
Gram–Schmidt orthogonalization, 44
graph, 102
Ladyzhenskaya, 312
graph norm, 107
Landau inequality, 246
Green function, 76
Landau kernel, 14
Gronwall’s inequality, 313
Landau symbols, 196
group
Lax–Milgram theorem, 54
strongly continuous, 240
nonlinear, 317
Legendre polynomials, 44
half-space, 142
Leray–Schauder principle, 306
Hamel basis, 11, 16
Lidskij trace theorem, 93
Hankel operator, 95
Lie group, 223
Hardy space, 184
liminf, 340
Hausdorff space, 333
limit, 335
heat equation, 3, 254
limit point, 330
Heine–Borel theorem, 347
limsup, 340
Hermitian form, 17
Lindelöf theorem, 344
Hilbert space, 18
linear
dimension, 45
functional, 30, 48
Hilbert–Schmidt operator, 89
operator, 27
Hölder continuous, 37
linearly independent, 12
Hölder’s inequality, 10, 25
holomorphic function, 279 Lipschitz continuous, 37
homeomorphic, 340 locally
homeomorphism, 340 (path-)connected, 354
homotopy, 280 locally convex space, 134, 144
homotopy invariance, 281 lower semicontinuous, 340
Lyapunov–Schmidt reduction, 232
ideal, 178
proper, 178 maximal solution, 227
identity, 31, 156 maximum norm, 7
implicit function theorem, 222 meager set, 98
index, 185 mean value theorem, 204
induced topology, 331 metric, 329
Induction Principle, 324 translation invariant, 148
initial topology, 342 mild solution, 245
injective, 339 minimum modulus, 105
inner product, 17 Minkowski functional, 131
inner product space, 18 monotone, 212, 316
integral, 194 map, 315
interior, 334 strictly, 212, 316
interior point, 330 strongly, 316
inverse function theorem, 223 Morawetz identity, 275
involution, 164 multilinear function, 203
isolated point, 330 symmetric , 203
isometric, 339 multiplicative linear functional, 178
multiplicity
Jacobi determinant, 281 algebraic, 171
Jacobi matrix, 201 geometric, 171
376 Index

Nash equilibrium, 296 basis, 43

Nash theorem, 297 set, 41
Navier–Stokes equation, 310
neighborhood, 330 parallel, 18
neighborhood base, 332 parallelogram law, 19
Neumann series, 161 parametrix, 188
nilpotent, 162 Parseval relation, 44
Noether operator, 185 partial order, 323
nonlinear Schrödinger equation, 229, 261 partition of unity, 352
norm, 8 path, 354
operator, 28 path-connected, 354
strictly convex, 16, 151 payoff, 295
stronger, 20 Peano theorem, 309
uniformly convex, 151 perpendicular, 18
norm-attaining, 114 polar decomposition, 86
normal polar set, 135
operator, 165, 167 polarization identity, 20
space, 350 power set, 322
normalized, 18 prisoner’s dilemma, 296
normed space, 8 product rule, 196, 204, 239
nowhere dense, 97 product topology, 341
n-person game, 295 projection-valued measure, 175
null space, 27 proper
function, 349
one-point compactification, 350 map, 303
one-to-one, 339 metric space, 348
onto, 339 pseudometric, 330
open Pythagorean theorem, 18
ball, 330
function, 340 quadrangle inequality, 334
set, 330 quasiconvex, 215
operator quasinilpotent, 162
adjoint, 51, 116 quasinorm, 16
bounded, 28 quotient space, 34
closable, 103 quotient topology, 343
closed, 103
closure, 103 range, 28
compact, 65 rank, 87, 117
completely continuous, 126 Rayleigh–Ritz method, 82
domain, 27 reaction-diffusion equation, 5, 260
finite rank, 87, 117 reduction property, 297
linear, 27 refinement, 344
nonnegative, 53 reflexive, 112
self-adjoint, 68 regular value, 281
strong convergence, 127 regulated function, 194
symmetric, 68 relative topology, 331
unitary, 46 relatively compact, 345
weak convergence, 127 reproducing kernel, 81
order residual set, 98
partial, 323 resolution of the identity, 175
total, 323 resolvent, 73, 75, 159, 246
well, 323 resolvent identity
orthogonal, 18 first, 164, 257
complement, 48 resolvent set, 158, 246
projection, 48, 175 Riemann–Lebesgue lemma, 60
sum, 56 Riesz lemma, 49, 121
orthonormal Ritz method, 82
Index 377

Rouchés theorem, 280 strictly convex space, 151

Russell’s paradox, 321 strong convergence, 127
strong solution, 243
Sard’s theorem, 285 Sturm–Liouville problem, 6
scalar product, 17 subbase, 332
Schatten p-class, 89 subcover, 344
Schauder basis, 12 subspace topology, 331
Schrödinger equation, 6, 261 support, 340
Schur property, 126 support hyperplane, 137
Schwartz space, 148 surjective, 339
Schwarz’ theorem, 203 symmetric
second category, 98 operator, 68
second countable, 332 sesquilinear form, 17
self-adjoint, 52, 165
semigroup Taylor’s theorem, 206
generator, 240 Taylors theorem, 206
strongly continuous, 239 tempered distributions, 148
uniform, 237 tensor product, 57
seminorm, 8 theorem
separable, 13, 337 Altman, 308
separation, 350 Arzelà–Ascoli, 26, 360
of convex sets, 132 Atkinson, 186, 188
of points, 343, 361 Bair, 97
of variables, 4 Banach–Alaoglu, 141
seminorms, 145 Banach–Steinhaus, 99, 128
sequentially closed, 335 Bernstein, 61
sequentially continuous, 340 Beurling–Gelfand, 161
series bipolar, 135
absolutely convergent, 15 Bolzano–Weierstraß, 348
sesquilinear form, 17 Borsuk, 290, 305
bounded, 23 Borsuk–Ulam, 291
parallelogram law, 22 Brouwer, 291, 305
polarization identity, 23 Browder–Minty, 319
shift operator, 52, 68 Carathéodory, 140
singular value decomposition, 85 closed graph, 103
singular values, 85 closed range, 120
span, 12 Crandall–Rabinowitz, 233
spectral measure, 173 Dieudonné, 187
spectral projections, 175 Dini, 363
spectral radius, 161 Fejér, 62
spectral theorem Feller–Miyadera–Phillips, 248
compact operators, 172 fundamental thm. of calculus, 195
compact self-adjoint operators, 71 Gelfand representation, 180
normal operators, 183 Gelfand–Mazur, 160
self-adjoint operators, 166, 174 Gelfand–Naimark, 182
spectrum, 73, 158, 246 Goldstine, 143
continuous, 168 Hahn–Banach, 110
discrete, 189 Hahn–Banach, geometric, 133
essential, 188 hairy ball, 289
Fredholm, 188 Heine–Borel, 347
point, 168 Hellinger–Toeplitz, 106
residual, 168 Helly, 130
∗-subalgebra, 164 Hilbert projection, 49
step function, 194 Hilbert–Schmidt, 71
Stone–Weierstraß theorem, 362 Hille–Yosida, 251
strategy, 295 implicit function, 222
strictly convex, 16 intermediate value, 356
378 Index

invariance of domain, 291, 305 trace topology, 331

inverse function, 223 transcritical bifurcation, 233
Jordan–von Neumann, 19 triangle inequality, 8, 329
Kakutani, 141 inverse, 8, 330
Kolmogorov, 149 trivial topology, 331
Krasnosel’skii, 308
Krein–Milman, 138 uniform boundedness principle, 99
Lax–Milgram, 54, 317 uniform contraction principle, 222
Leray–Schauder, 306 uniform convergence, 357
Lindelöf, 344 uniformly continuous, 359
Lumer–Phillips, 254 uniformly convex space, 151
mean value, 193 unit vector, 18
Milman–Pettis, 153 unital, 157
Nash, 297 unitarily equivalent, 46
Omega lemma, 224 unitary, 165
open mapping, 101, 102 Unitization, 163, 167
Peano, 309 upper semicontinuous, 340
Perron–Frobenius, 293 Urysohn lemma, 350
Pythagorean, 18
Radon–Riesz, 152 Vandermonde determinant, 15
Riesz, 172, 185, 187 variational derivative, 198
Rothe, 308 virial identity, 275
Rouché, 280, 282 Volterra integral operator, 162
Sard, 285
Schaefer, 306 wave equation, 5
Schauder, 118, 306 weak convergence, 122
Schröder–Bernstein, 325 weak solution, 311
Schwarz, 203 weak topology, 123, 140
spectral, 71, 166, 172, 174, 183 weak-∗ convergence, 129
spectral mapping, 160 weak-∗ topology, 140
Stone–Weierstraß, 362 weakly coercive, 214
Taylor, 206 Weierstraß approximation, 14
Tietze, 306, 351 Weierstraß theorem, 348
Tychonoff, 346 well-order, 323
Urysohn, 350 Weyl asymptotic, 85
Weierstraß, 14, 348 Wiener algebra, 156
Weissinger, 221 winding number, 279
Weyl, 188
Wiener, 182 Young inequality, 9
Yood, 188
Zarantonello, 316 Zermelo–Fraenkel set theory, 321
Zermelo, 325 ZF, 321
Zorn, 324 ZFC, 323
Toda lattice, 229 Zorn’s lemma, 324
topological space, 331
topological vector space, 131
topology
base, 332
product, 341
relative, 331
subbase, 332
total order, 323
total set, 13, 119
totally bounded, 349
trace, 93
class, 89
trace formula, 80