An Analysis Companion - Heil

Christopher Heil
An Analysis Companion
January 23, 2010
c _2010 by Christopher Heil
Contents
General Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
A Metrics, Norms, Inner Products, and Topology . . . . . . . . . . . . 3
A.1 Metrics and Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
A.2 Norms and Seminorms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
A.2.1 Innite Series in Normed Spaces . . . . . . . . . . . . . . . . . . . . . 5
A.2.2 Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
A.3 Examples of Banach Spaces:
p
, C
b
, C
0
, C
m
b
. . . . . . . . . . . . . . . . 7
A.3.1 The
p
Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
A.3.2 Some Spaces of Continuous and Dierentiable Functions 9
A.4 Inner Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
A.5 Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
A.5.1 Product Topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
A.6 Convergence and Continuity in Topological Spaces . . . . . . . . . . . 17
A.6.1 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
A.6.2 Continuity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
A.6.3 Equivalent Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
A.7 Closed and Dense Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
A.8 Compact Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
A.9 Complete Sequences and a First Look at Schauder Bases . . . . . 28
A.9.1 Span and Closed Span . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
A.9.2 Hamel Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
A.9.3 Introduction to Schauder Bases . . . . . . . . . . . . . . . . . . . . . 31
A.10 Unconditional Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
A.11 Orthogonality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
A.11.1 Orthogonality and the Pythagorean Theorem . . . . . . . . . 34
A.11.2 Orthogonal Direct Sums . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
A.11.3 Orthogonal Projections and Orthogonal Complements . 35
A.12 Orthogonality and Complete Sequences . . . . . . . . . . . . . . . . . . . . 36
A.13 Urysohns Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
vi Contents
B Lebesgue Measure and Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
B.1 Exterior Lebesgue Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
B.2 Lebesgue Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
B.2.1 Denition and Basic Properties . . . . . . . . . . . . . . . . . . . . . 43
B.2.2 Equivalent Formulations of Measurability . . . . . . . . . . . . . 44
B.2.3 Almost Everywhere . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
B.3 Measurable Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
B.4 Convergence in Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
B.5 The Lebesgue Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
B.5.1 Integration of Nonnegative Simple Functions . . . . . . . . . . 49
B.5.2 Integration of Nonnegative Functions . . . . . . . . . . . . . . . . 50
B.5.3 Integration of Real-Valued and Complex-Valued
Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
B.5.4 The Lebesgue Dominated Convergence Theorem . . . . . . 53
B.5.5 Relation to the Riemann Integral . . . . . . . . . . . . . . . . . . . . 54
B.6 The L
p
Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
B.6.1 Norm and Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
B.6.2 On Abuses of Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
B.6.3 Convergence in L
p
(E) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
B.6.4 Local Integrability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
B.7 Repeated Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
B.8 Functions of Bounded Variation . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
B.8.1 Denition and Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
B.8.2 The Jordan Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . 66
B.8.3 Dierentiability of Functions of Bounded Variation . . . . 67
B.9 Absolutely Continuous and Singular Functions . . . . . . . . . . . . . . 68
B.9.1 Singular Functions on the Real Line . . . . . . . . . . . . . . . . . 68
B.9.2 Absolutely Continuous Functions on the Real Line . . . . . 70
B.9.3 Preparation for the BanachZarecki Theorem . . . . . . . . . 73
B.9.4 The BanachZarecki Theorem and its Implications . . . . 75
B.10 H older Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
C Functional Analysis and Operator Theory . . . . . . . . . . . . . . . . . 81
C.1 Linear Operators on Normed Spaces . . . . . . . . . . . . . . . . . . . . . . . 81
C.1.1 Equivalence of Boundedness and Continuity . . . . . . . . . . 83
C.1.2 Isomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
C.1.3 Eigenvalues and Eigenvectors . . . . . . . . . . . . . . . . . . . . . . . 85
C.2 Some Useful Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
C.2.1 Orthogonal Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
C.2.2 Multiplication Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
C.2.3 Integral Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
C.2.4 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
C.3 The Space B(X, Y ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
C.4 Banach Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
C.5 Some Dual Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Contents vii
C.5.1 The Dual of a Hilbert Space . . . . . . . . . . . . . . . . . . . . . . . . 95
C.5.2 The Dual of L
p
(E) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
C.5.3 The Relation between L
p
(E) and L
p
(E)
. . . . . . . . . . . . 97
C.6 Adjoints for Operators on Hilbert Spaces . . . . . . . . . . . . . . . . . . . 99
C.6.1 Adjoints of Bounded Operators . . . . . . . . . . . . . . . . . . . . . 99
C.6.2 Adjoints of Unbounded Operators . . . . . . . . . . . . . . . . . . . 101
C.6.3 Bounded Self-Adjoint Operators on Hilbert Spaces . . . . 101
C.6.4 Positive and Positive Denite Operators on Hilbert
Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
C.7 Compact Operators on Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . 105
C.7.1 Denition and Basic Properties . . . . . . . . . . . . . . . . . . . . . 105
C.7.2 Finite-Rank Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
C.7.3 Integral Operators with Square-Integrable Kernels . . . . . 109
C.8 The Spectral Theorem for Compact Self-Adjoint Operators . . . 110
C.8.1 Existence of an Eigenvalue . . . . . . . . . . . . . . . . . . . . . . . . . 111
C.8.2 The Spectral Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
C.9 HilbertSchmidt Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
C.9.1 Denition and Basic Properties . . . . . . . . . . . . . . . . . . . . . 115
C.9.2 Singular Numbers and Schatten Classes . . . . . . . . . . . . . . 116
C.9.3 HilbertSchmidt Integral Operators . . . . . . . . . . . . . . . . . . 119
C.9.4 Trace-Class Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
C.10 The HahnBanach Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
C.10.1 Abstract Statement of the HahnBanach Theorem. . . . . 122
C.10.2 Implications of the HahnBanach Theorem . . . . . . . . . . . 123
C.10.3 Orthogonal Complements in Normed Spaces . . . . . . . . . . 125
C.10.4 X
and Reexivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

C.10.5 Adjoints of Operators on Banach Spaces . . . . . . . . . . . . . 126
C.11 The Baire Category Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
C.12 The Uniform Boundedness Principle . . . . . . . . . . . . . . . . . . . . . . . 129
C.13 The Open Mapping Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
C.14 The Closed Graph Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
C.15 Schauder Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
C.15.1 Continuity of the Coecient Functionals . . . . . . . . . . . . . 132
C.15.2 Minimal Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
C.15.3 A Characterization of Schauder Bases . . . . . . . . . . . . . . . . 135
C.15.4 Unconditional Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
C.16 Weak and Weak* Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
D Borel and Radon Measures on the Real Line . . . . . . . . . . . . . . 139
D.1 -Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
D.2 Signed Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
D.2.1 The Jordan Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . 143
D.3 Positive Measures and Integration . . . . . . . . . . . . . . . . . . . . . . . . . 145
D.3.1 Basic Properties of Positive Measures . . . . . . . . . . . . . . . . 145
D.3.2 Borel Measurable Functions . . . . . . . . . . . . . . . . . . . . . . . . 146
viii Contents
D.3.3 Integration of Nonnegative Functions . . . . . . . . . . . . . . . . 146
D.3.4 Integration of Arbitrary Functions . . . . . . . . . . . . . . . . . . . 148
D.4 Signed Measures and Integration . . . . . . . . . . . . . . . . . . . . . . . . . . 150
D.5 Complex Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
D.6 Fubini and Tonelli for Borel Measures . . . . . . . . . . . . . . . . . . . . . . 156
D.7 Radon Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
D.8 The Riesz Representation Theorem for Positive Linear
Functionals on C
c
(R) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
D.8.1 Topologies on C
c
(R) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
D.8.2 Positive Linear Functionals on C
c
(R) . . . . . . . . . . . . . . . . 162
D.9 The Relation Between Radon and Borel Measures on R . . . . . . 163
D.10 The Dual of C
0
(R) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
E Topological Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
E.1 Motivation and Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
E.2 Topological Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
E.2.1 Base for a Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
E.2.2 Topological Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 174
E.3 Topologies Induced by Families of Seminorms . . . . . . . . . . . . . . . 176
E.3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
E.3.2 The Topology Associated with a Family of Seminorms . 176
E.3.3 The Convergence Criterion . . . . . . . . . . . . . . . . . . . . . . . . . 178
E.3.4 Continuity of the Vector Space Operations . . . . . . . . . . . . 180
E.3.5 Continuity Equals Boundedness . . . . . . . . . . . . . . . . . . . . . 180
E.4 Topologies Induced by Countable Families of Seminorms . . . . . 182
E.4.1 Metrizing the Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
E.4.2 Tempered and Compactly Supported Distributions . . . . 184
E.5 C
c
(R) and its Dual Space T
(R) . . . . . . . . . . . . . . . . . . . . . . . . . . 186
E.5.1 The Topology on C
c
(R) . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
E.5.2 The Space of Distributions . . . . . . . . . . . . . . . . . . . . . . . . . 188
E.6 The Weak and Weak* Topologies on a Normed Linear Space . . 189
E.6.1 The Weak Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
E.6.2 The Weak* Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
E.7 Alaoglus Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
E.7.1 Product Topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
E.7.2 Statement and Proof of Alaoglus Theorem . . . . . . . . . . . 193
E.7.3 Implications for Separable Spaces . . . . . . . . . . . . . . . . . . . . 194
F Complex Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
F.1 Analytic Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
F.2 Power Series and Taylor Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
F.3 Dirichlet Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
F.4 Trigonometric Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
F.5 Interpolation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Contents ix
G Zorns Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Hints and Solution Sketches for Exercises and Additional
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
Index of Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
General Notation
We review here some of the notational conventions that will be used through-
out these notes.
Unless otherwise specied, all vector spaces are taken over the complex
eld C. In particular, functions whose domain is R
d
(or a subset of R
d
) are
generally allowed to take values in the complex plane C.
Integrals with unspecied limits are taken over either the real line or R
d
,
according to context. In particular, if f : R C, then we take
_
f(x) dx =
_

f(x) dx.
The extended real line is R, = [, ]. We use the conventions
that 1/0 = , 1/= 0, and 0 = 0.
If 1 p is given, then its dual index or dual exponent is the extended
real number p
that satises
1
p
+
1
p
= 1.
Explicitly,
p
=
p
p 1
.
The dual index lies in the range 1 p
, and we have 1
= , 2
= 2,
and
= 1.
The Kronecker delta is
ij
=
_
1, i = j,
0, i ,= j.
If S is a subspace of a vector space V, then the nite linear span of S is
denoted by span(S). If V is also a metric space, then the closed linear span
span(V ) is dened to be the closure of span(S) in V. We say that S is complete
if span(S) = V.
2 Contents
We use the symbol to denote the end of a proof, and the symbol to
denote the end of a denition, remark, example, or exercise, or the end of the
statement of a theorem whose proof will be omitted.
A
Metrics, Norms, Inner Products, and Topology
These notes provide mini-courses on analysis, Lebesgue measure, functional
analysis, Borel measures, topological vector spaces, and other related material.
They were originally written as background notes for my text Introduction to
Harmonic Analysis. Each appendix gives a substantial, though not exhaus-
tive, introduction to and review of its respective subject. Topics that are not
typically part of standard beginning mathematics graduate courses are given
more detailed attention, while other results are either formulated as exercises
(often with hints) or stated without proof. Sources for additional information
on the material of this and most of the other appendices include Follands real
analysis text [Fol99], Conways functional analysis text [Con90], and the the
operator theory/Hilbert space text [GG01] by Gohberg and Goldberg. Hints
and some solution sketches for exercises and problems are given at the end of
the notes.
A.1 Metrics and Convergence
A metric determines a notion of distance between points in a set.
Denition A.1 (Metric Space). Let X be a set. A metric on X is a function
d: X X R such that for all f, g, h X we have:
(a) d(f, g) 0,
(b) d(f, g) = 0 if and only if f = g,
(c) d(f, g) = d(g, f), and
(d) the Triangle Inequality: d(f, h) d(f, g) + d(g, h).
In this case, X is a called a metric space. The value d(f, g) is the distance
from f to g.
If we need to explicitly identify the metric we write let X be a metric
space with metric d or let (X, d) be a metric space.
4 A Metrics, Norms, Inner Products, and Topology
A metric space need not be a vector space, although this will be true of
most of the metric spaces encountered in this volume.
Once we have a notion of distance, we have a corresponding notion of
convergence.
Denition A.2 (Convergent and Cauchy Sequences). Let X be a met-
ric space with metric d, and let f
n
nN
be a sequence of elements of X.
(a) We say that f
n
nN
converges to f X if lim
n
d(f
n
, f) = 0, i.e., if
> 0, N > 0, n N, d(f
n
, f) < .
In this case, we write lim
n
f
n
= f or f
n
f.
(b) We say that f
n
nN
is Cauchy if
> 0, N > 0, m, n N, d(f
m
, f
n
) < .
Exercise A.3. Let X be a metric space.
(a) Every convergent sequence in X is Cauchy.
(b) The limit of a convergent sequence is unique.
In general, however, a Cauchy sequence need not be convergent (Exer-
cises A.63A.61).
Denition A.4 (Complete Metric Space). If every Cauchy sequence in a
metric space X has the property that it converges to an element of X, then
X is said to be complete.
Beware that the term complete is heavily overused and has a number of
distinct mathematical meanings.
Notation A.5. Let X be a metric space. Given f X and r > 0, the open
ball in X of radius r centered at f is
B
r
(f) =
_
g X : d(f, g) < r
_
. (A.1)
A.2 Norms and Seminorms
A norm provides a notion of the length of a vector in a vector space.
Denition A.6 (Seminorms and Norms). Let X be a vector space over
the eld C of complex scalars. A seminorm on X is a function | |: X R
such that for all f, g X and all scalars c C we have:
(a) |f| 0,
(b) |cf| = [c[ |f|, and
(c) the Triangle Inequality: |f +g| |f| +|g|.
A.2 Norms and Seminorms 5
A seminorm is a norm if we also have:
(d) |f| = 0 if and only if f = 0.
A vector space X together with a norm | | is called a normed linear space
or simply a normed space. If the norm is not clear from context, we may write
(X, | |) to denote that | | is the norm on X.
If S is a subspace of a normed space X, then S is itself a normed space
with respect to the norm on X (restricted to S).
Exercise A.7. If X is a normed space, then d(f, g) = |f g| denes a metric
on X, called the induced metric.
The Schwartz space (Example E.4) is an example of a metric space whose
metric is not induced from any norm; another is
p
with 0 < p < 1 (see
Exercise A.18).
Exercise A.8. Show that if X is a normed linear space, then the following
statements hold.
(a) Reverse Triangle Inequality:
|f| |g|
|f g|.
(b) Continuity of the norm: f
n
f = |f
n
| |f|.
(c) Continuity of vector addition: f
n
f and g
n
g = f
n
+g
n
f +g.
(d) Continuity of scalar multiplication: f
n
f and
n
=
n
f
n
f.
(e) Boundedness of convergent sequences: If f
n
nN
converges then we have
sup |f
n
| < .
(f) Boundedness of Cauchy sequences: If f
n
nN
is Cauchy then we have
sup |f
n
| < .
Denition A.9 (Banach Space). A normed linear space X is called a Ba-
nach space if it is complete, i.e., if every Cauchy sequence is convergent.
Thus, the terms Banach space and complete normed space are inter-
changeable.
An important fact that we will assume without proof is that the complex
plane C under absolute value is a Banach space.
A.2.1 Innite Series in Normed Spaces
Since a normed space has both an operation of vector addition and a notion
of convergence, we can consider innite series.
Denition A.10 (Convergent Series). Let f
n
nN
be a sequence in a
normed linear space X. Then the series
n=1
f
n
converges and equals f X
if the partial sums s
N
=
N
n=1
f
n
converge to f, i.e., if
lim
N
|f s
N
| = lim
N
_
_
_f
N
n=1
f
n
_
_
_ = 0.
Note that the ordering of a series may be important! If we reorder a series,
or in other words consider a new series
n=1
f
(n)
where : N N is a
bijection, there is no guarantee that this reordered series will still converge.
These issues are addressed in more detail in Section A.10.
Denition A.11 (Absolutely Convergent Series). Let X be a normed
space and let f
n
nN
be a sequence of elements of X. If
n=1
|f
n
| < ,
then we say that the series
n=1
f
n
is absolutely convergent in X.
The denition of absolute convergence does not require that the series
f
n
converge in X. This will always be the case if X is a Banach space, and
indeed this property is an equivalent characterization of completeness.
Exercise A.12. Let X be a normed space. Prove that X is a Banach space
if and only if every absolutely convergent series in X converges in X.
A.2.2 Convexity
Denition A.13 (Convex Set). If X is a vector space and K X, then K
is convex if
x, y K, 0 t 1 = tx + (1 t)y K.
Thus, the entire line segment between x and y is contained in K (including
the midpoint
1
2
x +
1
2
y in particular).
Every subspace of a vector space is convex by denition. The fact that
balls in a normed space are convex is an important property.
Exercise A.14. Show that if X is a normed linear space, then each open ball
B
r
(f) in X is convex.
Additional Problems
A.1. If f
n
nN
is a Cauchy sequence in a normed space X and there exists
a subsequence f
n
k
kN
that converges to f X, then f
n
f.
A.2. If f
n
nN
is a Cauchy sequence in a normed space X, then there exists
a subsequence f
n
k
kN
such that |f
n
k+1
f
n
k
| < 2
k
for all k N.
A.3. Let X be a normed space. Show that if f
n
X satisfy |f
n+1
f
n
| < 2
n
for every n, then f
n
nN
is Cauchy.
p
, C
b
, C0, C
m
b
7
A.4. Let f
n
nN
be a sequence in a normed space X, and let f X be
xed. Suppose that every subsequence g
n
nN
of f
n
nN
has a subsequence
h
n
nN
of g
n
nN
such that h
n
f. Show that f
n
f.
A.5. Let X be a normed space. Extend the denition of convergence to fam-
ilies indexed by a real parameter by declaring that if f X and f
t
X for
t R, then f
t
f as t 0 if for every > 0 there exists a > 0 such
that |f f
t
| < whenever [t[ < . Show that f
t
f as t 0 if and only if
f
t
k
f for every sequence of real numbers t
k
kN
such that t
k
0.
p
, C
b
, C
0
, C
m
b
In this section we give a few examples of Banach and other spaces.
A.3.1 The
p
Spaces
Denition A.15. Let I be a nite or countably innite index sequence.
(a) If 0 < p < , then
p
(I) consists of all sequences of scalars x = (x
k
)
kI
such that
|x|
p
= |(x
k
)
kI
|
p
=
_
kI
[x
k
[
p
_
1/p
< .
(b) For p = , the space
(I) consists of all sequences of scalars x = (x

k
)
kI
such that
|x|
= |(x
k
)
kI
|
= sup
kI
[x
k
[ < .
If I = N, then we write
p
instead of
p
(N).
If I = 1, . . . , d, then
p
(I) = C
d
, and in this case we refer to
p
(I) as C
d
under the
p
norm. The
2
norm on C
d
is called the Euclidean norm.
Now we prove a fundamental inequality for the
p
spaces.
Theorem A.16 (H olders Inequality). Let I be a nite or countable index
set. Given 1 p , if x = (x
k
)
kI

p
(I) and y = (y
k
)
kI

p
(I), then
xy = (x
k
y
k
)
kI

1
(I) and
|xy|
1
|x|
p
|y|
p
.
For 1 < p < , this inequality is
kI
[x
k
y
k
[
_
kI
[x
k
[
p
_
1/p
_
kI
[y
k
[
p
_
1/p
.
Proof. The cases p = 1 and p = are straightforward exercises. Assume
1 < p < . The key to the proof is a special case of an inequality due
to Young for continuous, strictly increasing functions. Namely, since x
p1
is
continuous and strictly increasing and its inverse function is y
1
p1
, we have
for all a, b 0 that
ab
_
a
0
x
p1
dx +
_
b
0
y
1
p1
dy =
a
p
p
+
b
p
(see the proof by picture in Figure A.1, or Problem A.10).

a
b
Fig. A.1. Illustration of Youngs Inequality. Area of the vertically hatched region
is
R
a
0
x
p1
dx; area of the horizontally hatched region is
R
b
0
y
1
p1
dy; area of the
rectangle is ab.
Consequently, if x
p
(I) and y
p
(I) satisfy |x|

p
= 1 = |y|
p
, then
|xy|
1
=
kI
[x
k
y
k
[
kI
_
[x
k
[
p
p
+
[y
k
[
p
_
=
1
p
+
1
p
= 1. (A.2)
For general nonzero x, y, we apply equation (A.2) to the normalized vectors
x/|x|
p
and y/|y|
p
to obtain
|xy|
1
|x|
p
|y|
p
=
_
_
_
x
|x|
p
y
|y|
p
_
_
_
1
1.
The next exercise shows that if p 1 then | |
p
is a norm on
p
(I).
The Triangle Inequality on
p
(often called Minkowskis Inequality) is easy to
prove for p = 1 and p = , but more dicult for 1 < p < . A hint for using
H olders Inequality to prove Minkowskis Inequality is given in the solutions
section at the end of the text.
p
, C
b
, C0, C
m
b
9
Exercise A.17. Let I be a nite or countable index set. Show that if 1 p
, then | |
p
is a norm on
p
(I), and
p
(I) is a Banach space with respect
to this norm.
On the other hand, if p < 1 then | |
p
fails the Triangle Inequality and
hence is not a norm. Still, we can modify the distance function so that
p
is a
complete metric space, though this metric is not induced from any norm.
Exercise A.18. Let I be a nite or countably innite index set. Show that if
0 < p < 1, then |x+y|
p
p
|x|
p
p
+|y|
p
p
. Consequently,
p
(I) is a vector space
and d(x, y) = |x y|
p
p
is a metric on
p
(I). Show that
p
(I) is complete with
respect to this metric. However, if I contains more than one element, then the
unit ball B
1
(0) is not convex, and hence this metric is not induced from any
norm (compare Exercise A.14).
We can also dene
p
(I) when the index set I is uncountable. In this case,
for p < we dene
p
(I) to be the space of all sequences x = (x
k
)
kI
with
at most countably many terms nonzero such that
[x
k
[
p
< . With this
denition,
p
(I) is again a Banach space.
A.3.2 Some Spaces of Continuous and Dierentiable Functions
We now give some examples of normed spaces of functions.
Denition A.19. The support of a function f : R C is the closure of the
set of points where f is nonzero:
supp(f) = x R : f(x) ,= 0.
Since the support of a function is a closed set, a function on R has compact
support if and only if it is zero outside of a nite interval.
Exercise A.20. Let C
b
(R) denote the space of continuous, bounded functions
f : R C. Show that C
b
(R) is a Banach space with respect to the uniform
norm
|f|
= sup
tR
[f(t)[.
Show that the subspace
C
0
(R) =
_
f C
b
(R) : lim
|t|
f(t) = 0
_
is also a Banach space with respect to the uniform norm, but the subspace
C
c
(R) =
_
f C
b
(R) : supp(f) is compact
_
(A.3)
is not complete with respect to the uniform norm.
Beware, some authors use the symbols C
0
to denote the space that we
refer to as C
c
.
Exercise A.20 has an extension to m-times dierentiable functions, as fol-
lows.
Exercise A.21. Let C
m
b
(R) be the space of all m-times dierentiable func-
tions on R each of whose derivatives is bounded and continuous, i.e.,
C
m
b
(R) =
_
f C
b
(R) : f, f
, . . . , f
(m)
C
b
(R)
_
.
Show that C
m
b
(R) is a Banach space with respect to the norm
|f|
C
m
b
= |f|
+|f
+ +|f
(m)
|
,
and
C
m
0
(R) =
_
f C
0
(R) : f, f
, . . . , f
(m)
C
0
(R)
_
is a subspace of C
m
b
(R) that is also a Banach space with respect to the same
norm. However,
C
m
c
(R) =
_
f C
c
(R) : f, f
, . . . , f
(m)
C
c
(R)
_
is not a Banach space with respect to this norm.
Although they are not normed spaces, it is often important to consider
the space of functions that are continuous or m-times dierentiable but not
bounded. We denote these by:
C(R) =
_
f : R C : f is continuous
_
,
C
m
(R) =
_
f C(R) : f, f
, . . . , f
(m)
C(R)
_
.
Additionally, we sometimes need to consider spaces of innitely dierentiable
functions, including the following:
C
(R) =
_
f C(R) : f, f
, . . . C(R)
_
,
C
b
(R) =
_
f C
b
(R) : f, f
, . . . C
b
(R)
_
.
C
0
(R) =
_
f C
0
(R) : f, f
, . . . C
0
(R)
_
.
C
c
(R) =
_
f C
c
(R) : f, f
, . . . C
c
(R)
_
.
The space C
c
(R) will be especially important to us in Appendix E. Although
not a normed space, it is topological vector space, and is often denoted by
T(R) = C
c
(R).
A.4 Inner Products 11
Additional Problems
A.6. Fix 0 < p . Let x
n
nN
be a sequence of vectors in
p
(I), and x a
vector in
p
(I). Write the components of x
n
and x as x
n
= (x
n
(1), x
n
(2), . . . )
and x = (x(1), x(2), . . . ), and prove the following statements.
(a) If x
n
x in
p
(I), then x
n
converges componentwise to x, i.e., for
each xed k we have lim
n
x
n
(k) = x(k).
(b) If I is nite then componentwise convergence implies convergence with
respect to the norm | |
p
.
(c) If I is innite then componentwise convergence need not imply conver-
gence in the norm of
p
(I).
A.7. Show that if 1 p < q , then
p

q
, and |x|
q
|x|
p
for all
x
p
.
A.8. Show that if x
q
(I) for some nite q then |x|
p
|x|
as p ,
but this can fail if x /
q
(I) for any nite q.
A.9. Let I be a nite or countable index set, and let w: I (0, ) be xed.
Given a sequence of scalars x = (x
k
)
kI
, set
|x|
p,w
=
_
_
_
kI
[x
k
[
p
w(k)
p
_
1/p
, 0 < p < ,
sup
kI
[x
k
[ w(k), p = ,
and dene the weighted
p
space
p
w
(I) =
_
x : |x|
p,w
<
_
. Show that
p
w
(I)
is a Banach space for each 1 p .
A.10. (a) Show that if 0 < < 1, then t
t + (1 ) for t > 0, with

equality if and only if t = 1.
(b) Suppose that 1 < p < and a, b 0. Apply part (a) with t = a
p
b
p
and = 1/p to show that ab a

p
/p + b
p
/p
, with equality if and only if

b = a
p1
.
A.11. Show that equality holds in H olders Inequality (Theorem A.16) if and
only if there exist scalars , , not both zero, such that [x
k
[
p
= [y
k
[
p
for
each k I.
A.4 Inner Products
While a norm on a vector space provides a notion of the length of a vector,
an inner product provides us with a notion of the angle between vectors.
Denition A.22 (Semi-Inner Product, Inner Product). If H is a vector
space over the complex eld C, then a semi-inner product on H is a function
, ): H H C such that for all f, g, h H and scalars , C we have:
(a) f, f) 0,
(b) f, g) = g, f),
(c) Linearity in the rst variable: f +g, h) = f, h) +g, h).
If a semi-inner product , ) also satises:
(d) f, f) = 0 if and only if f = 0,
then it is called an inner product on H. In this case, H is called a inner product
space or a pre-Hilbert space.
There are many dierent standard notations for semi-inner products, in-
cluding [f, g], (f, g), or f[g), in addition to our preferred notation f, g).
Exercise A.23. If , ) is a semi-inner product on a vector space H, show
that the following statements hold.
(a) Antilinearity in the second variable: f, g +h) = f, g) +

f, h).
(b) f, 0) = 0 = 0, f).
(c) 0, 0) = 0.
A function of two variables that is linear in the rst variable and antilinear
(also called conjugate linear) in the second variable conjugate linear function
is referred to as a sesquilinear form. Thus each semi-inner product , ) is an
example of a sesquilinear form.
1
Sometimes an inner product is required to
be antilinear in the rst variable and linear in the second; this is common in
the physics literature.
Every subspace S of an inner product space H is itself an inner product
space (using the inner product on H restricted to S).
The next exercise gives the prototypical example of an inner product.
Exercise A.24. Given an index set I and x = (x
k
)
kI
, (y
k
)
kI

2
(I), set
x, y) =
kI
x
k
y
k
.
Show that , ) is an inner product on
2
(I).
Our next goal is to show that every semi-inner product induces a seminorm
on H, and every inner product induces a norm.
Notation A.25. If , ) is a semi-inner product on a vector space H, then we
write
|f| = f, f)
1/2
, f H.
Prejudicing the issue, we refer to | | as the seminorm induced by , ). If , )
is an inner product, then we call | | the norm induced by , ).
1
The prex sesqui- means one and a half.
A.4 Inner Products 13
Before showing that | | actually is a seminorm or norm, we derive some
of its basic properties.
Exercise A.26. Given a semi-inner product , ) on a vector space H, show
that the following statements hold for all f, g H.
(a) Polar Identity: |f +g|
2
= |f|
2
+ 2 Ref, g) +|g|
2
.
(b) Parallelogram Law: |f +g|
2
+|f g|
2
= 2
_
|f|
2
+|g|
2
_
.
Now we prove an important inequality; this should be compared to
H olders Inequality (Theorem A.16) for p = 2. This inequality is variously
known as the Schwarz, CauchySchwarz, or CauchyBunyakowskiSchwarz
Inequality.
Theorem A.27 (CauchyBunyakowskiSchwarz Inequality). If , )
is a semi-inner product on a vector space H, then
f, g, H, [f, g)[ |f| |g|.
Proof. If f = 0 or g = 0 then there is nothing to prove, so suppose both are
nonzero. Write f, g) = [f, g)[ where C and [[ = 1. Then for t R we
have by the Polar Identity that
0 |f tg|
2
= |f|
2
2 Re tf, g) +t
2
|g|
2
= |f|
2
2t [f, g)[ +t
2
|g|
2
.
This is a real-valued quadratic polynomial in the variable t. In order for it
to be nonnegative, it can have at most one real root. This requires that the
discriminant be at most zero, so
_
2 [f, g)[
_
2
4 |f|
2
|g|
2
0. The desired
inequality then follows upon rearranging.
When f, g) = 0, we say that f and g are orthogonal. More details on
orthogonality appear in Section A.11.
Finally, the CauchyBunyakowskiSchwarz Inequality can be used to show
that | | is indeed a seminorm or norm on H.
Exercise A.28. Given a semi-inner product , ) on a vector space H, show
that | | is a seminorm on H, and if , ) is an inner product then | | is a
norm on H.
Thus, all inner product spaces are normed linear spaces. The following
exercise shows that there are normed space whose norm is not induced from
any inner product on the space.
Exercise A.29. Let I be an index set containing at least two elements. Show
that | |
p
does not satisfy the Parallelogram Law if 1 p and p ,= 2.
Therefore, there is no inner product on
p
(I) whose induced norm is ||
p
.
On the other hand, it is certainly possible for
p
(I) to be an inner product
space with respect to some inner product. For example, if I is nite or if p < 2,
then
p
(I)
2
(I), so in this case
p
(I) is an inner product space with respect
to the inner product on
2
(I) restricted to the subspace
p
(I). However, the
norm induced from this inner product is | |
2
and not | |
p
. Further, if I is
innite then
p
(I) is not complete with respect to the induced norm | |
2
,
whereas it is complete with respect to the norm | |
p
.
Denition A.30 (Hilbert Space). An inner product space H is called a
Hilbert space if it is complete with respect to the induced norm, i.e., if every
Cauchy sequence is convergent.
Thus, a Hilbert space is an inner product space that is a Banach space
with respect to the induced norm. In particular,
2
(I) is a Hilbert space.
Additional Problems
A.12. Show that , ) is an inner product on C
d
if and only if there exists a
positive denite matrix A such that x, y) = Ax y, where x y = x
1
y
1
+ +
x
d
y
d
denotes the usual dot product on C
d
.
A.13. Continuity of the inner product: If H is an inner product space and
f
n
f, g
n
g in H, then f
n
, g
n
) f, g).
A.14. Let H be an inner product space. Show that if the series
n=1
f
n
converges in H, then for any g H,
_

n=1
f
n
, g
_
=
n=1
f
n
, g).
Note that this is not merely a consequence of the linearity of the inner product
in the rst variable the continuity of the inner product is also needed.
A.15. Show that equality holds in the CauchyBunyakowskiSchwarz In-
equality if and only if there exist scalars , C, not both zero, such that
|f +g| = 0. In particular, if | | is a norm, then either f = cg or g = cf
for some scalar c.
A.16. Justify the following statement: The angle between two vectors f, g in
a Hilbert space H is the value of that satises Ref, g) = |f| |g| cos .
A.5 Topology
Now we consider topologies, especially on metric and normed spaces. General
background references on topology include Munkres [Mun75] and Singer and
Thorpe [ST76].
A.5 Topology 15
Denition A.31 (Topology). A topology on a set X is a family T of subsets
of X such that the following statements hold.
(a) , X T .
(b) Closure under arbitrary unions: If I is any index set and U
i
T for i I,
then
iI
U
i
T .
(c) Closure under nite intersections: If U, V T , then U V T .
If these hold then T is called a topology on X and X is called a topological
space. The elements of T are called the open subsets of X. The complements
of the open subsets are the closed subsets of X. A neighborhood of a point
x X is any set A such that there exists an open set U with x U A.
In particular, an open neighborhood of x is any open set U that contains x. If
the topology on X is not clear from context, we may write (X, T ) to denote
that T is the topology on X.
The following is a convenient criterion for testing for openness.
Exercise A.32. Let X be a topological space. Prove that V X is open if
and only if
x V, open U V such that x U.
If a space X has two topologies T
1
, T
2
and if T
1
T
2
, i.e., every set that
is open with respect to T
1
is also open with respect to T
2
, then we say that
T
1
is weaker than T
2
and T
2
is stronger than T
1
. These terms should not be
confused with the weak topology or the strong topology on a space. The strong
topology on a normed space is dened below (see Denition A.34) and the
weak topology is dened in Example E.7.
The most familiar topologies are those associated with metric spaces.
Exercise A.33. Let (X, d) be a metric space. Declare a subset U X to be
open if
f U, r > 0 such that B
r
(f) U,
where B
r
(f) is the open ball centered at f with radius r. Let T be the collec-
tion of all subsets of X that are open according to this denition. Show that
T is a topology on X.
Thus every metric space has a natural topology associated with it, and
consequently so does every normed space.
Denition A.34. (a) If (X, d) is a metric space, then the topology T dened
in Exercise A.33 is called the topology on X induced from the metric d, or
simply the induced topology on!X.
(b) If (X, | |) is a normed linear space, then the topology T induced from the
metric d(f, g) = |f g| is called the norm topology, the strong topology,
or the topology induced from | |.
(c) Let (X, T ) be a topological space. If there exists a metric d on X whose
induced topology is exactly T , then the topology T on X is said to be
metrizable.
(d) Let X be a vector space with a topology T . If there exists a norm | |
on X whose induced topology is exactly T , then the topology T on X is
said to be normable.
The Schwartz space o(R) and the space C
(R) are important examples of

topological spaces that are metrizable but not normable (see Example E.28).
The topology induced by a metric has the following special property.
Denition A.35. A topological space X is Hausdor if
x ,= y X, disjoint U, V T such that x U, y V.
Exercise A.36. Every metric space is Hausdor.
Every subset of a topological space X inherits a topology from X.
Exercise A.37. Let X be a topological space. Given Y X, show that
T
Y
= U Y : U is open in X
is a topology on Y, called the topology on Y relative to X, the topology on Y
inherited from X, the topology on X restricted to Y, etc.
One way to generate a topology on a set X is to begin with a collection
of sets that we want to be open, and then to create a topology that includes
those particular sets. There will be many such topologies in general, but the
following exercise shows that there is a smallest topology that includes those
chosen sets.
Exercise A.38. Let c be a collection of subsets of a set X. Show that
T (c) =
_
T : T is a topology on X with c T
_
is a topology on X. We call T (c) the topology generated by c. The collection
c is sometimes called a subbase for the topology T (c).
Note that if T
1
, T
2
are topologies, then T
1
T
2
is not formed by intersecting
the elements of T
1
with those of T
2
. Rather, it is the collection of all sets that
are common to both T
1
and T
2
. Thus, if T is any topology that contains c then
we will have T (c) T , which explains why T (c) is the smallest topology
that contains c. In particular, the topology induced by a metric d on a metric
space X is the topology generated by the set of open balls in X.
We can characterize the generated topology T (c) as follows.
Exercise A.39. If c is a collection of subsets of a set X whose union is X,
then T (c) is set of all unions of nite intersections of elements of c:
T (c) =
_
iI
n
j=1
E
ij
: I arbitrary, n N, E
ij
c
_
.
A.6 Convergence and Continuity in Topological Spaces 17
A.5.1 Product Topologies
As an application of generated topologies, we show how to construct a nat-
ural topology on the Cartesian product of two topological spaces. A product
topology on an innite collection of topological spaces can also be dened but
requires a little more care, see Denition E.44.
Denition A.40 (Product Topology). Let X and Y be topological spaces,
and set
B = U V : open U X, open V Y .
The product topology on X Y is the topology T (B) generated by B.
Exercise A.41. (a) Show that B as given above is a base for T (B), which
means that if W T (B), then there exist sets U
open in X and V
open
in Y such that W =
(U
).
(b) Show that if W XY is open with respect to the product topology, then
for each x X the restriction W
x
= y Y : (x, y) W is open in Y,
and likewise for each y Y the restriction W
y
= x X : (x, y) W is
open in X.
A.6 Convergence and Continuity in Topological Spaces
In large part, the importance of topologies in this volume is that they provide
notions of convergence and continuity. Indeed, a basic philosophy that we
will expand upon in this section is that topologies and convergence criteria
are equivalent. Thus, though many of the spaces that we will encounter are
dened in terms of norms or families of seminorms (i.e., convergence criteria),
this is equivalent to dening them in terms of a topology.
A.6.1 Convergence
In metric spaces, convergence is dened with respect to sequences indexed by
the natural numbers (Denition A.2). In a general topological space, conver-
gence must be formulated in terms of nets instead of countable sequences.
Denition A.42 (Directed Sets, Nets). A directed set is a set I together
with a relation on I such that:
(a) is reexive: i i for all i I,
(b) is transitive: i j and j k implies i k, and
(c) for any i, j I, there exists k I such that i k and j k.
A net in a set X is a sequence x
i
iI
of elements of X indexed by a
directed set (I, ).
Remark A.43. By denition, a sequence x
i
iI
is shorthand for the function
x: I X dened by x(i) = x
i
for i I. In particular, unlike a set, a sequence
allows repetitions of the x
i
. Technically, we should be careful to distinguish
between a sequence x
i
iI
and a set x
i
: i I, but it is usually clear from
context whether a sequence or a set is meant.
The set of natural numbers I = N under the usual ordering is one example
of a directed set, and hence every ordinary sequence indexed by the natural
numbers is a net. Another typical example is I = T(X), the power set of X,
ordered by reverse inclusion, i.e.,
U V V U.
Denition A.44 (Convergence of a Net). Let X be a topological space,
let x
i
iI
be a net in X, and let x X be given. Then we say that x
i
iI
converges to x (with respect to the directed set I), and write x
i
x, if for
any open neighborhood U of x there exists i
0
I such that
i i
0
= x
i
U.
Next, we will dene the notion of accumulation points of a subset of a
generic topological space and see how this denition can be reformulated in
terms of nets. We will also see that topologies induced from a metric have
the advantage that we only need to use convergence of ordinary sequences
indexed by N instead of general nets.
Denition A.45 (Accumulation Point). Let E be a subset of a topological
space X. Then a point x X is an accumulation point of E if every open
neighborhood of x contains a point of E other than x itself, i.e.,
U open and x U = E (Ux) ,= .
Lemma A.46. If E is a subset of a topological space X and x X, then the
following statements are equivalent.
(a) x is an accumulation point of E.
(b) There exists a net x
i
iI
contained in Ex such that x
i
x.
If X is a metric space, then these statements are also equivalent to the
following.
(c) There exists a sequence x
n
nN
contained in Ex such that x
n
x.
Proof. (a) (b). Assume that x is an accumulation point of E. Dene
I =
_
U X : U is open and x U
_
.
Exercise: Show that I is a directed set when ordered by reverse inclusion.
For each U I, by denition of accumulation point there exists a point
x
U
E (Ux). Then x
U
UI
is a net in Ex, and we claim that
x
U
x. To see this, x any open neighborhood V of x. Set U
0
= V, and
suppose that U U
0
. Then, by denition, U I and U U
0
. Hence x
U

U U
0
= V. Therefore x
U
x.
(b) (a). Suppose that x
i
iI
is a net in Ex and x
i
x. Let U be
any open neighborhood of x. Then there exists an i
0
such that x
i
U for all
i i
0
. Since x
i
,= x, this implies that x
i
E (Ux) for all i i
0
.
(a) (c), assuming X is metric. Suppose that x is an accumulation point
of E. For each n N, the open ball B
1/n
(x) is an open neighborhood of x,
and hence there must exist some x
n
E(B
1/n
(x)x). Therefore x
n
nN
is a sequence in Ex, and since d(x, x
n
) < 1/n, we have x
n
x.
(c) (b), assuming X is metric. This follows from the fact that every
countable sequence x
n
nN
is a net.
We can now give an equivalent formulation of closed sets in terms of nets
and accumulation points.
Exercise A.47. Given a subset E of a topological space, prove that the fol-
lowing statements are equivalent.
(a) E is closed, i.e., XE is open.
(b) If x is an accumulation point of E, then x E.
(c) If x
i
iI
is a net in E and x
i
x X, then x E.
If X is a metric space, show that these are also equivalent to the following
statement.
(d) If x
n
nN
is a sequence in E and x
n
x X, then x E.
Now we can quantify the philosophy that topologies and convergence cri-
teria are equivalent. For arbitrary topologies, this requires that we use con-
vergence with respect to nets, but for topologies induced from a metric we are
able to use convergence of ordinary sequences indexed by the natural numbers.
Exercise A.48. Given two topologies T
1
, T
2
on a set X, prove that the fol-
lowing statements are equivalent.
(a) T
1
T
2
, i.e.,
U is open with respect to T
1
= U is open with respect to T
2
.
(b) If x
i
iI
is a net in X and x X, then
x
i
x with respect to T
2
= x
i
x with respect to T
1
.
If T
1
is induced from a metric d
1
on X, and T
2
is induced from a metric d
2
on X, show that these statements are also equivalent to the following.
(c) If x
n
nN
is a sequence in X and x X, then
lim
n
d
2
(x
n
, x) = 0 = lim
n
d
1
(x
n
, x) = 0.
Interchanging the roles of T
1
and T
2
in Exercise A.48, we see that T
1
= T
2
if and only if T
1
and T
2
dene exactly the same convergence criterion.
Example A.49. An example of a topological space where it is important to
distinguish between convergence of ordinary sequences and convergence with
respect to nets is the sequence space
1
under the weak topology. This topology
will be dened precisely in Section E.6, but the important point for us at the
moment is that it can be shown that if x
n
nN
is a sequence in
1
and x
n
x
with respect to the weak topology, then x
n
x in norm, i.e., |x x
n
|
1
0
(see [Con90, Prop. V.5.2]). However, the weak topology on
1
is not the same
as the topology induced by the norm | |
1
. The moral is that when discussing
convergence in a topological space that is not a metric space, it is important
to consider nets instead of ordinary sequences.
A.6.2 Continuity
Our next goal is to reformulate continuity of a function in terms of convergence
of nets or sequences. Recall that if f : X Y and V Y, then the preimage
of V is f
1
(V ) = x X : f(x) V .
Denition A.50 (Continuity). Let X, Y be topological spaces. Then a
function f : X Y is continuous if
V is open in Y = f
1
(V ) is open in X.
We say that f is a topological isomorphism or a homeomorphism if f is a
bijection and both f and f
1
are continuous.
It will be convenient to restate continuity in terms of continuity at a point.
Denition A.51 (Continuity at a Point). Let X, Y be topological spaces
and let x X be given. Then a function f : X Y is continuous at x if for
each open neighborhood V of f(x) in Y, there exists an open neighborhood
U of x in X such that U f
1
(V ).
Exercise A.52. Prove that f is continuous if and only if f is continuous at
each x X.
Now we can formulate continuity in terms of preservation of convergence of
nets. For the case of a metric space, this reduces to preservation of convergence
of sequences.
Lemma A.53. If X, Y be topological spaces, f : X Y, and x X are given,
then the following statements are equivalent.
(a) f is continuous at x.
(b) For any net x
i
iI
in X,
x
i
x in X = f(x
i
) f(x) in Y.
If X is a metric space, then these are also equivalent to the following.
(c) For any sequence x
n
nN
in X,
x
n
x in X = f(x
n
) f(x) in Y.
Proof. (a) (b). Assume that f is continuous at x X. Let x
i
iI
be any
net in X such that x
i
x. Let V be any open neighborhood of f(x). Then
by denition of continuity at a point, there exists an open neighborhood U
of x that is contained in f
1
(V ). Hence, by denition of x
i
x, there exists
an i
0
I such that x
i
U f
1
(V ) for all i i
0
. Hence f(x
i
) V for all
i i
0
, which means that f(x
i
) f(x).
(b) (a). Suppose that statement (a) fails, i.e., f is not continuous at
x X. Then by denition there exists an open neighborhood V of f(x) such
that no open neighborhood U of x can be contained in f
1
(V ). Therefore
each open neighborhood U of x must contain some point x
U
Uf
1
(V ).
Now let
I =
_
U : U is an open neighborhood of x
_
.
Then I is a directed set when ordered by reverse inclusion, so x
U
UI
is a
net in X. Exercise: Show that x
U
x. However, f(x
U
) does not converge to
f(x) because V is an open neighborhood of f(x) but V contains no points
f(x
U
). Hence statement (b) fails.
The remaining implications are exercises.
A.6.3 Equivalent Norms
Next we consider the equivalence of convergence criteria and topologies for
the case of normed spaces.
Denition A.54. Suppose that X is a normed linear space with respect to
a norm | |
a
and also with respect to another norm | |
b
. Then we say that
these norms are equivalent if there exist constants C
1
, C
2
> 0 such that
f X, C
1
|f|
a
|f|
b
C
2
|f|
a
. (A.4)
We write | |
a
| |
b
to denote that | |
a
and | |
b
are equivalent norms.
Theorem A.55. Let | |
a
and | |
b
be two norms on a vector space X. Then
the following statements are equivalent.
(a) | |
a
and | |
b
are equivalent norms.
(b) | |
a
and | |
b
induce the same topologies on X.
(c) | |
a
and | |
b
dene the same convergence criterion. That is, if x
n
nN
is a sequence in X and x X, then
lim
n
|x x
n
|
a
= 0 lim
n
|x x
n
|
b
= 0.
Proof. (b) (a). Assume that statement (b) holds. Let B
a
r
(x) and B
b
r
(x)
denote the open balls of radius r centered at x X with respect to | |
a
and
| |
b
, respectively. Since B
a
1
(0) is open with respect to | |
a
, statement (b)
implies that B
a
1
(0) is open with respect to | |
b
. Therefore, since 0 B
a
1
(0),
there must exist some r > 0 such that B
b
r
(0) B
a
1
(0).
Now choose any x X and any > 0. Then
(r )
|x|
b
x B
b
r
(0) B
a
1
(0),
so
_
_
_
(r ) x
|x|
b
_
_
_
a
< 1.
Rearranging, this implies (r ) |x|
a
< |x|
b
. Since this is true for every ,
we conclude that r |x|
a
|x|
b
.
A symmetric argument, interchanging the roles of the two norms, shows
that there exists an s > 0 such that |x|
b
s |x|
a
for every x X. Hence the
two norms are equivalent.
Given any nite-dimensional vector space X, we can dene many norms
on X. In particular, the following norms are analogues of the
p
norms dened
in Section A.3.
Exercise A.56. Let B = x
1
, . . . , x
d
be any basis for a nite-dimensional
vector space X, and let x =
d
k=1
c
k
(x) x
k
denote the unique expansion of
x X with respect to this basis (the vector [x]
B
= (c
1
(x), . . . , c
d
(x)) is called
the coordinate vector of x with respect to the basis B). Show that
|x|
p
=
_
_
_
d
k=1
[c
k
(x)[
p
_
1/p
, 1 p < ,
max
k
[c
k
(x)[, p = ,
are norms on X, and X is complete with respect to each of these norms. Note
that |x|
p
is simply the
p
norm of the coordinate vector [x]
B
.
It is not dicult to see that all of the norms dened in Exercise A.56
are equivalent. Although we will not prove it, it is an important fact that all
norms on a nite-dimensional space are equivalent.
Theorem A.57. If X is a nite-dimensional vector space, then any two
norms on X are equivalent. In particular, if | | is any norm on X and | |
p
is any one of the norms constructed in Exercise A.56, then | | | |
p
.
A.7 Closed and Dense Sets 23
Additional Problems
A.17. (a) Let X be a Hausdor topological space. Show that if a net x
i
iI
converges in X, then the limit is unique.
(b) Show that if X is not Hausdor, then there exists a net x
i
iI
in X
that has two distinct limits.
A.18. Let x
i
iI
be a net in a Hausdor topological space X. Show that if
x
i
x in X, then either:
(a) there exists an open neighborhood U of x and some i
0
I such that
x
i
= x for all i i
0
, or
(b) every open neighborhood U of x contains innitely many distinct x
i
,
i.e., the set x
i
: i I and x
i
U is innite.
A.19. Show that if (X, d
1
) and (Y, d
2
) are metric spaces, then
d
_
(f
1
, g
1
), (f
2
, g
2
)
_
= d
1
(f
1
, f
2
) + d
2
(g
1
, g
2
)
denes a metric on X Y that induces the product topology. Conclude that
convergence in X Y is componentwise convergence, i.e., (f
n
, g
n
) (f, g) in
X Y if and only if f
n
f in X and g
n
g in Y.
A.20. Let X be a vector space with a metric d. We say that the metric
is translation-invariant if d(f + h, g + h) = d(f, g) for every f, g, h X.
Show that vector addition is continuous in this case, i.e., (f, g) f + g is a
continuous mapping of X X into X.
A.21. Let X, Y be topological spaces. Let (f
i
, g
i
)
iI
be any net in X Y,
and suppose (f, g) X Y. Show that (f
i
, g
i
) (f, g) with respect to the
product topology on X Y if and only if f
i
f in X and g
i
g in Y.
A.7 Closed and Dense Sets
The smallest closed set that contains a given set is called its closure, dened
precisely as follows.
Denition A.58. If E is a subset of a topological space X, then the closure
of E, denoted E, is the smallest closed set in X that contains E:
E =
F X : F is closed and F E.
If E = X, then we say that E is dense in X.
Often it is more convenient to use the following equivalent form of the
closure of a set.
Exercise A.59. Given a subset E of a topological space X, show that E is
the union of E and all the accumulation points of E.
The typical method for showing that a subset of a metric space is dense is
given in the next exercise.
Exercise A.60. Let X be a metric space, and let E X be given. Show that
E is dense in X if and only if for each f X there exist a sequence f
n
nN
in E such that f
n
f.
The next exercise characterizes the closed subspaces of a Banach space.
Exercise A.61. Let M be a subspace of a Banach space X. Then M is itself a
Banach space (using the norm inherited fromX) if and only if M is closed.
Every subspace of a nite-dimensional normed space is closed, but this
need not be the case in innite dimensions.
Exercise A.62. Show that every nite-dimensional subspace of a normed
linear space is closed.
Exercise A.63. (a) Fix 1 p . Dene
c
00
=
_
x = (x
1
, . . . , x
N
, 0, 0, . . . ) : N > 0, x
1
, . . . , x
N
C
_
.
Prove that c
00
is a proper dense subspace of
p
for each 1 p < , and
hence is not closed with respect to | |
p
. The vectors in c
00
are sometimes
called nite sequences because they contain at most nitely many nonzero
components.
(b) Dene
c
0
=
_
x = (x
k
)
k=1
: lim
k
x
k
= 0
_
.
Prove that c
0
is a closed subspace of
(N), and c
0
is the closure of c
00
in
-norm.
(c) Show that C
c
(R) is a proper dense subspace of C
0
(R), and C
0
(R) is the
closure of C
c
(R) in the uniform norm.
We now introduce a denition that in some sense distinguishes between
small and large innite-dimensional spaces.
Denition A.64. A topological space that contains a countable dense subset
is said to be separable.
Exercise A.65. (a) Show that if I is a nite or countably innite index set,
then
p
(I) is separable for 0 < p < .
(b) Show that if I is innite, then
(I) is not separable.

(c) Show that if I is uncountable, then
p
(I) is not separable for any p.
Exercise A.66. Show that c
0
and C
0
(R) are separable.
A.8 Compact Sets 25
A.8 Compact Sets
Now we briey review the denition and properties of compact sets.
Denition A.67 (Compact Set). A subset K of a topological space X is
compact if every cover of K by open sets has a nite subcover. That is, K is
compact if it is the case that whenever
K

iI
U
i
,
where U
i
iI
is any collection of open subsets of X, there exist nitely many
i
1
, . . . , i
N
I such that
K
N
k=1
U
i
k
.
In a metric space, we can reformulate compactness in several equivalent
ways. We need the following terminology.
Denition A.68. Let E be a subset of a metric space X.
(a) E is sequentially compact if every sequence f
n
nN
of points of E contains
a convergent subsequence f
n
k
kN
whose limit belongs to E.
(b) E is complete if every Cauchy sequence in E converges to a point in E.
(c) E is totally bounded if for every r > 0, we can cover E by nitely many
open balls of radius r, i.e., there exist nitely many x
1
, . . . , x
N
X such
that
E
N
k=1
B
r
(x
k
).
Similarly to Exercise A.61, if X is complete then a subset E is complete
if and only if it is closed.
Theorem A.69. If E is a subset of a metric space X, then the following
statements are equivalent.
(a) E is compact.
(b) E is sequentially compact.
(c) E is complete and totally bounded.
Proof. (a) (b). Suppose there exists a sequence f
n
nN
in E that has no
subsequence that converges to an element of E. Choose any f E. If every
open ball centered at f contained innitely many vectors f
n
, then there would
be a subsequence f
n
k
kN
that converged to f. Therefore, there must exist
some open ball B
f
centered at f that contains only nitely many f
n
. But
then B
f
fE
is an open cover of E that has no nite subcover. Hence E is
not compact.
(b) (c). Suppose that E is sequentially compact. Exercise: Show that E
is complete.
Suppose that E was not totally bounded. Then there is an r > 0 such
that E cannot be covered by nitely many open balls of radius r centered at
points of X. Choose any f
1
E. Since E cannot be covered by a single r-ball,
E cannot be a subset of B
r
(f
1
). Hence there exists a point f
2
E B
r
(f
1
).
In particular, we have d(f
2
, f
1
) r. But E cannot be covered by two r-balls,
so there must exist an f
3
E
_
B
r
(f
1
) B
r
(f
2
)
_
. In particular, we have
both d(f
3
, f
1
) r and d(f
3
, f
2
) r. Continuing in this way we obtain a
sequence of points f
n
nN
in E that has no convergent subsequence, which
is a contradiction.
(c) (b). Assume that E is complete and totally bounded, and let
f
n
nN
be any sequence of points in E. Since E is covered by nitely many
open balls of radius
1
2
, one of those balls must contain innitely many f
n
, say
f
(1)
n

nN
. Then we have
m, n N, d
_
f
(1)
m
, f
(1)
n
_
< 1.
Since E is covered by nitely many open balls of radius
1
4
, we can nd a
subsequence f
(2)
n

nN
of f
(1)
n

nN
such that
m, n N, d
_
f
(2)
m
, f
(2)
n
_
<
1
2
.
Continuing by induction, for each k > 1 we nd a subsequence f
(k)
n

nN
of
f
(k1)
n

nN
such that d
_
f
(k)
m
, f
(k)
n
_
<
1
k
for all m, n N.
Now consider the diagonal subsequence f
(k)
k

kN
. Given > 0, let N be
large enough that
1
N
< . If j k > N, then f
(j)
j
is one element of the
sequence f
(k)
n

nN
, say f
(j)
j
= f
(k)
n
. Hence
d
_
f
(j)
j
, f
(k)
k
_
= d
_
f
(k)
n
, f
(k)
k
_
<
1
k
< .
Thus f
(k)
k

nN
is a Cauchy subsequence of the original sequence. Since E is
complete, this subsequence must therefore converge to some element of E.
(b) (a). Assume that E is sequentially compact. Then by the implication
(b) (c) proved above, we also know that E is complete and totally bounded.
Choose any open cover U
J
of E. We must show that it contains a
nite subcover. The key ingredient in proving this is the following claim.
Claim. There exists a number > 0 such that if B is any open ball of
radius that intersects E, then there is an J such that B U
.
To prove the claim, suppose that U
J
was an open cover of E such that
no with the required property existed. Then for each n N, we could nd a
open ball B
n
with radius
1
n
that intersects E but is not contained in any U
.
A.8 Compact Sets 27
Choose any f
n
B
n
E. Since E is sequentially compact, there must be a
subsequence f
n
k
kN
that converges to an element of E, say f
n
k
f E.
Since U
J
is a cover of E, we must have f U
for some J, and

since U
is open there must exist some r > 0 such that B

r
(f) U
. Now
choose k large enough that we have both
1
n
k
<
r
3
and d
_
f, f
n
k
_
<
r
3
.
Then it follows that B
n
k
B
r
(f) U
, which is a contradiction. Hence the

claim holds.
To nish the proof, we use the fact that E is totally bounded, and therefore
can be covered by nitely many open balls of radius . By the claim, each
of these balls is contained in some U
, so E is covered by nitely many of

the U
.
We will state only a few of the many special properties possessed by func-
tions on compact sets.
Exercise A.70. Let X and Y be topological spaces. If f : X Y is contin-
uous and K X is compact, then f(K) is a compact subset of Y.
Denition A.71. Let (X, d
X
) and (Y, d
Y
) be metric spaces. We say that a
function f : X Y is uniformly continuous if for every > 0 there exists a
> 0 such that
x, y X, d
X
(x, y) < = d
Y
(f(x), f(y)) < .
Exercise A.72. Let (X, d
X
) and (Y, d
Y
) be metric spaces, and let K X
be compact. Then any continuous function f : K Y is uniformly continu-
ous.
Additional Problems
A.22. Let X be a topological space.
(a) Prove that every closed subset of a compact subset of X is compact.
(b) Show that if the topology on X is Hausdor, then every compact subset
of X is closed.
A.23. Show that if K is a compact subset of a topological space X, then every
innite subset of K has an accumulation point.
A.24. Show that if E is a totally bounded subset of a metric space X, then its
closure E is compact. (A set with compact closure is said to be precompact.)
A.25. Show that a function f : R
d
C is uniformly continuous if and only if
lim
a0
|f T
a
|
= 0, where T
a
f(x) = f(x a) is the translation operator.
A.26. (a) Show that every compact subset of a normed linear space X is both
closed and bounded.
(b) Show that every closed and bounded subset of R
d
or C
d
is compact.
(c) Use the fact that all norms on a nite-dimensional vector space are
equivalent to extend part (b) to arbitrary nite-dimensional vector spaces.
(d) Let H be an innite-dimensional inner product space. Show that the
closed unit ball x H : |x| 1 is closed and bounded but not compact.
A.27. This problem will extend Problem A.26(d) to an arbitrary normed
linear space X.
(a) Prove F. Rieszs Lemma: If M is a proper, closed subspace of X and
> 0, then there exists g X with |g| = 1 such that
dist(g, M) = inf
fM
|g f| > 1 .
(b) Prove that the closed unit ball x X : |x| 1 in X is compact if
and only if X is nite-dimensional.
A.9 Complete Sequences and a First Look at Schauder
Bases
In this section we dene complete sequences of vectors in normed spaces.
In nite dimensions, these are simply spanning sets. However, in innite di-
mensions there are subtle, but important, distinctions between spanning sets,
complete sets, and bases. For more details on bases in Banach spaces, we re-
fer to the texts by Singer [Sin70], Lindenstrauss and Tzafriri [LT77], or the
introductory text [Hei10].
A.9.1 Span and Closed Span
Denition A.73 (Span). Let A be a subset of a vector space X. The nite
linear span of A, denoted span(A), is the set of all nite linear combinations
of elements of A:
span(A) =
_
n
k=1
k
f
k
: n > 0, f
k
A,
k
C
_
.
We also refer to the nite linear span of A as the nite span, the linear span,
or simply the span of A.
In particular, if A is a countable sequence, say A = f
k
kN
, then
span
_
f
k
kN
_
=
_
n
k=1
k
f
k
: n > 0,
k
C
_
.
We will use the following vectors to illustrate many of the concepts in this
section.
A.9 Complete Sequences and a First Look at Schauder Bases 29
Denition A.74 (Standard Basis Vectors). Let
e
n
= (
mn
)
mN
= (0, . . . , 0, 1, 0, 0, . . . )
denote the sequence that has a 1 in the nth component and zeros elsewhere.
We call e
n
the nth standard basis vector, and refer to c = e
n
nN
as the
standard basis for
p
(p nite) or c
0
(p = ).
Of course, this denition begs the question of in what sense the standard
basis vectors form a basis. This is answered in Exercise A.81, but for now let
us consider the nite span of the standard basis.
Example A.75. The nite span of e
n
nN
is span
_
e
n
nN
_
= c
00
.
In particular, the nite span of e
n
nN
is not
p
for any p and likewise
is not c
0
. The standard basis vectors do not form a basis for
p
or c
0
in the
vector space sense. Instead, in the strict vector space sense of spanning and
being nitely independent, the standard basis vectors form a vector space
basis for c
00
.
In a generic vector space, we can form nite linear combinations of vectors,
but unless we have a notion of convergence, we cannot take limits or form
innite sums. However, once we impose some extra structure, such as the
existence of a norm, these concepts make sense.
Denition A.76 (Closed Span and Complete Sets). Let A be a subset
of a normed linear space X.
(a) The closed nite span of A, denoted span(A), is the closure of the set of
all nite linear combinations of elements of A:
span(A) = span(A) = z X : y
n
span(A) such that y
n
z.
We also refer to the closed nite span of A as the closed linear span or
the closed span of A.
(b) We say that A is complete in X if span(A) is dense in X, that is, if
span(A) = X.
There are many other terminologies in use for complete sets, for example,
they are also called total or fundamental.
Example A.77. Let c = e
n
nN
be the standard basis. Taking the closure
with respect to the norm | |
p
, we have span(c) =
p
if 1 p < , and
span(c) = c
0
if p = .
Beware: The denition of the closed span does NOT imply that
span(A) =
_

k=1
k
f
k
: f
k
A,
k
C
_
This need not hold!
In particular it is NOT true that an arbitrary element of span(A) can be
written f =
k=1
k
f
k
for some f
k
A,
k
C (see Exercise A.83). Instead,
to illustrate the meaning of the closed span, consider the case of a countable
set A = f
k
kN
. Here we have
span
_
f
k
kN
_
=
_
f X :
k,n
C such that
n
k=1
k,n
f
k
f as n
_
.
That is, an element f lies in the closed span of f
k
kN
if there exist
k,n
C
such that
n
k=1
k,n
f
k
f as n .
In contrast, to say that f =
k=1
k
f
k
means that
n
k=1
k
f
k
f as n . (A.5)
In particular, in order for equation (A.5) to hold, the scalars
k
must be
independent of n.
A.9.2 Hamel Bases
We have already dened the nite span of a set, and now we recall the de-
nition of nite linear independence.
Denition A.78 (Finite Independence). A subset A of a vector space X is
nitely linearly independent if every nite subset of A is linearly independent.
That is, if f
1
, . . . , f
n
are any choice of distinct vectors from A, then we must
have that
n
k=1
k
f
k
= 0 =
1
= =
n
= 0.
We often simply say that A is linearly independent or just independent to
mean that it is nitely linearly independent.
A Hamel basis is an ordinary vector space basis.
Denition A.79 (Hamel Basis). A subset A of a vector space X is a Hamel
basis or a vector space basis for X if it spans and is nitely linearly indepen-
dent. That is, A is a Hamel basis if span(A) = X and every nite subset of A
is linearly independent.
For example, the standard basis c = e
n
nN
is a Hamel basis for c
00
,
and the set of monomials x
n
: n 0 is a Hamel basis for the set T of all
polynomials.
A.9 Complete Sequences and a First Look at Schauder Bases 31
It is shown in Theorem G.3 that the Axiom of Choice implies that every
vector space has a basis (in fact, this statement is equivalent to the Axiom
of Choice). However, if X is an innite-dimensional Banach space, then any
Hamel basis for X must be uncountable (Exercise C.99). Typically, we cannot
explicitly exhibit a Hamel basis for X, i.e., we usually only know that one
exists because of the Axiom of Choice. Consequently, Hamel bases are usually
not very useful in the setting of normed spaces, and they make only rare
appearances in this volume. One interesting consequence of the existence of
Hamel bases is that there exist unbounded linear functionals on any innite-
dimensional normed linear space (Problem C.8).
A.9.3 Introduction to Schauder Bases
In a normed space, Hamel bases are unnecessarily restrictive, because they
require the nite linear span to be the entire space. In contrast, Schauder
bases allow innite linear combinations, and as such are much more useful
in normed spaces. We will be careful to avoid using the term basis without
qualication, because in the setting of an abstract vector space it usually
means a Hamel basis, while in the setting of a Banach space it usually means
a Schauder basis.
Denition A.80 (Schauder Basis). A sequence T = f
k
kN
in a Banach
space X is a Schauder basis for X if we can write every f X as
f =
k=1
k
(f) f
k
(A.6)
for a unique choice of scalars
k
(f), where the series converges in the norm
of X.
Note that the uniqueness requirement implies that f
k
,= 0 for every k.
Exercise A.81. Show that the standard basis c = e
n
nN
is a Schauder
basis for
p
for each 1 p < , and c is a Schauder basis for c
0
with respect
to
-norm.
Every Schauder basis is both complete and nitely linearly independent.
However, the next exercise shows that the converse fails in general. For this
example we will need the following useful theorem on approximation by poly-
nomials. Given a < b, we set
C[a, b] = f : [a, b] C : f is continuous. (A.7)
This is a Banach space under the uniform norm.
Theorem A.82 (Weierstrass Approximation Theorem). If f C[a, b]
and > 0, then there exists a polynomial p such that
|f p|
= sup
x[a,b]
[f(x) p(x)[ < .
Exercise A.83. Use the Weierstrass Approximation Theorem to show that
the set of monomials x
k
k=0
is complete and nitely linearly independent
in C[a, b]. Prove that there exist functions f C[a, b] that cannot be written
as f(x) =
k=0
k
x
k
with convergence of the series in the uniform norm.
Consequently x
k
k=0
is not a Schauder basis for C[a, b].
A complete discussion of Schauder bases requires the HahnBanach and
Uniform Boundedness Theorems, and therefore we will postpone additional
discussion of the relations and distinctions between bases and complete lin-
early independent sets until Section C.15. One interesting fact that we will
see there is that the uniqueness requirement in the denition of a Schauder
basis implies that the functionals
k
appearing in equation (A.6) must be
continuous, and hence belong to the dual space of X.
We end this section with one important exercise.
Exercise A.84. Let X be a Banach space. Show that if there exists a count-
able subset f
n
nN
in X that is complete, then X is separable.
In particular, a Banach space that has a Schauder basis must be separable.
The question of whether every separable Banach space has a basis was a
longstanding open problem known as the Basis Problem. It was nally shown
by Eno [Enf73] that there exist separable reexive Banach spaces that have
no Schauder bases!
A.10 Unconditional Convergence
Recall that a series
n=1
f
n
in a normed space X converges and equals a
vector f if the sequence of partial sums
N
n=1
f
n
converges to f as N .
However, the ordering of the f
n
can be important, for if we rearrange the
series, it may no longer converge. A series that converges regardless of ordering
possesses important stability properties that we will review in this section.
Denition A.85 (Unconditionally Convergent Series). Let X be a
normed linear space and let f
n
nN
be a sequence of elements of X. Then
the series
n=1
f
n
is said to converge unconditionally if every rearrangement
of the series converges, i.e., if for each bijection : N N the series
n=1
f
(n)
converges in X.
Thus, a series
n=1
f
n
converges unconditionally if and only if for each
bijection : N N, the partial sums
N
n=1
f
(n)
converge to some vector as
N . Note that while we do not explicitly require in this denition that
A.10 Unconditional Convergence 33
the partial sums converge to the same limit for each bijection , Exercise A.90
shows that this must in fact happen.
The method of the next example can be used to show that if X is a
nite-dimensional normed space, then absolute convergence is equivalent to
unconditional convergence. However, this equivalence fails in any innite di-
mensional normed space. That is, if X is innite-dimensional, then there will
exist series that converge unconditionally but not absolutely.
Example A.86. To illustrate the importance of unconditional convergence,
consider the Banach space X = C. The harmonic series
n=1
1
n
does not
converge, as its partial sums are unbounded. On the other hand, the alter-
nating harmonic series
n=1
(1)
n+1 1
n
does converge (in fact, its partial
sums converge to ln 2). While this latter series converges, it does not converge
absolutely.
Now consider what happens if we change the order of summation. For
n N, let p
n
=
1
2n1
and q
n
=
1
2n
, i.e., the p
n
are the positive terms from the
alternating series and the q
n
are the absolute values of the negative terms.
Each series
p
n
and
q
n
diverges to innity. Hence there must exist an
m
1
> 0 such that
p
1
+ +p
m1
> 1.
Also, there must exist an m
2
> m
1
such that
p
1
+ +p
m1
q
1
+p
m1+1
+ +p
m2
> 2.
Continuing in this way, we can create a rearrangement
p
1
+ +p
m1
q
1
+p
m1+1
+ +p
m2
q
2
+
of
(1)
n 1
n
that diverges to +. Likewise, we can construct a rearrange-
ment that diverges to , converges to any given real number r, or simply
oscillates without ever converging.
A modication of the argument above gives us the following theorem for
series of scalars.
Theorem A.87. A series of scalars converges absolutely if and only if it con-
verges unconditionally. That is, given c
n
C,
n=1
c
n
converges unconditionally
n=1
[c
n
[ < .
One direction of Theorem A.87 generalizes to all Banach spaces.
Exercise A.88. Let X be a Banach space. Show that if a series
n=1
f
n
converges absolutely in X then it converges unconditionally.
The converse of Exercise A.88 fails in general. For example, if e
n
nN
is
any orthonormal sequence in a Hilbert space H, then the series
1
n
e
n
con-
verges unconditionally but not absolutely. It can be shown that absolute and
unconditional convergence are equivalent only for nite-dimensional spaces
(this is the DvoretzkyRogers Theorem, see [LT77] or [Hei10]).
The following theorem, which we state without proof, gives several refor-
mulations of unconditional convergence (see [Sin70] or [Hei10] for details). For
this result, recall the denition of directed set given in Denition A.42, and
note that the set I consisting of all nite subsets of N forms a directed set
with respect to inclusion of sets.
Theorem A.89. Let f
n
nN
be a sequence in a Banach space X. Then the
(a)
n=1
f
n
converges unconditionally.
(b) The net
_
nF
f
n
: F N, F nite
_
converges in X. That is, there exists
an f X such that for every > 0, there is a nite F
0
N such that
F
0
F N, F nite =
_
_
_
_
f
nF
f
n
_
_
_
_
< .
(c) If (c
n
)
nN
is any bounded sequence of scalars, then
n=1
c
n
f
n
converges
in X.
Exercise A.90. Use Theorem A.89 to show that if a series
n=1
f
n
converges
unconditionally in a Banach space X, then there exists a single f X such
that for every bijection : N N we have
n=1
f
(n)
= f.
A.11 Orthogonality
In this section we review some of the special properties of inner product spaces,
especially in relation to orthogonality.
A.11.1 Orthogonality and the Pythagorean Theorem
Denition A.91. Let H be an inner product space, and let I be an arbitrary
index set.
(a) Vectors f, g H are orthogonal, denoted f g, if f, g) = 0.
(b) A collection of vectors f
i
iI
is orthogonal if f
i
, f
j
) = 0 whenever i ,= j.
(c) A collection of vectors f
i
iI
is orthonormal if it is orthogonal and each
vector is a unit vector, i.e., f
i
, f
j
) =
ij
.
For example, the standard basis is an orthonormal sequence in
2
.
A.11 Orthogonality 35
Exercise A.92 (Pythagorean Theorem). Show that if f
1
, . . . , f
n
H are
orthogonal, then
_
_
_
_
n
k=1
f
k
_
_
_
_
2
=
n
k=1
|f
k
|
2
.
The existence of a notion of orthogonality gives inner product spaces a
much simpler structure than general Banach spaces. We will derive some of
the basic properties of inner product spaces below, most of which are straight-
forward consequences of orthogonality. Some (but not all) of these results have
analogues for general Banach spaces. However, even for the results that do
have analogues, the corresponding proofs in the non-Hilbert space setting are
usually much more complicated or are nonconstructive (see the discussion of
the HahnBanach Theorem in Section C.10).
A.11.2 Orthogonal Direct Sums
Denition A.93 (Orthogonal Direct Sum). Let M, N be closed sub-
spaces of a Hilbert space H.
(a) The direct sum of M and N is M +N = f +g : f M, g N.
(b) We say that M and N are orthogonal subspaces, denoted M N, if f g
for every f M and g N.
(c) If M, N are orthogonal subspaces in H, then we call their direct sum the
orthogonal direct sum of M and N, and denote it by M N.
Exercise A.94. Show that if M, N are closed, orthogonal subspaces of H,
then M N is a closed subspace of H.
A.11.3 Orthogonal Projections and Orthogonal Complements
Denition A.95. Let X be a normed linear space and x A H. The dis-
tance from a point f H to the set A is
dist(f, A) = inf
_
|f g| : g A
_
.
Given a closed convex subset K of a Hilbert space H and given any x H,
there is a unique point in K that is closest to x.
Exercise A.96 (Closest Point Property). If H is a Hilbert space and K
is a nonempty closed, convex subset of H, then given any h H there exists
a unique point k
0
K that is closest to h. That is, there is a unique point
k
0
K such that
|h k
0
| = dist(h, K) = inf
_
|h k| : k K
_
.
Applying the closest point theorem to the particular case of closed sub-
spaces gives us the existence of orthogonal projections in a Hilbert space.
Denition A.97 (Orthogonal Projection). Let M be a closed subspace
of a Hilbert space H. For each h H, the unique point p M closest to h is
called the orthogonal projection of h onto M.
Orthogonal projections can be characterized as follows.
Exercise A.98. Let M be a closed subspace of a Hilbert space H. Given
h H, show that the following statements are equivalent.
(a) h = p +e where p is the (unique) point in M closest to h.
(b) h = p +e where p M and e M (i.e., e f for every f M).
Denition A.99 (Orthogonal Complement). Let A be a subset (not nec-
essarily a subspace) of a Hilbert space H. The orthogonal complement of A is
A
=
_
f H : f A
_
=
_
f H : f, g) = 0 for all g A
_
.
In the terminology of orthogonal complements, a vector p is the orthogonal
projection of f onto a closed subspace M if and only if f = p+e where p M
and e M
.
Exercise A.100. Show that if A is an arbitrary subset of a Hilbert space H,
then A
is a closed subspace of H (even if A is not), and

(A
= span(A).
Hence, if M is a closed subspace of a Hilbert space H then (M
= M.
Further, H = M M
, the orthogonal direct sum of M and M
.
The next exercise will allow us to give a useful equivalent formulation of
completeness of a sequence in a Hilbert space.
Exercise A.101. Let A be a subset of a Hilbert space H. Show that A is
complete if and only if A
= 0, i.e., the only vector orthogonal to every

element of A is the zero vector.
A.12 Orthogonality and Complete Sequences
In a Hilbert space, the combination of completeness and orthonormality of
a sequence e
n
nN
leads to especially nice series representations of vectors
in H in terms of the vectors e
n
. We will explore this topic in this section.
The following exercise summarizes some basic results connected to conver-
gence of series of orthonormal vectors.
Exercise A.102. Let e
n
nN
be any orthonormal sequence in a Hilbert
space H. Show that the following statements hold.
(a) Bessels Inequality:
n=1
[f, e
n
)[
2
|f|
2
.
A.12 Orthogonality and Complete Sequences 37
(b) If f =
n=1
c
n
e
n
converges, then c
n
= f, e
n
).
(c)
n=1
c
n
e
n
converges
n=1
[c
n
[
2
< .
(d) If
n=1
c
n
e
n
converges then it converges unconditionally, i.e., it converges
regardless of the ordering of the indices (see Section A.10).
(e) f span
_
e
n
nN
_
f =
n=1
f, e
n
) e
n
.
(f) If f H, then p =
n=1
f, e
n
) e
n
is the orthogonal projection of f onto
span
_
e
n
nN
_
.
Now we characterize those sequences in a Hilbert space that are both
complete and orthonormal.
Exercise A.103. Let e
n
nN
be any orthonormal sequence in a Hilbert
space H. Show that the following statements are equivalent.
(a) e
n
nN
is complete.
(b) For each f H there exist unique scalars (c
n
)
nN
such that f =
n=1
c
n
e
n
.
(c) For each f H, f =
n=1
f, e
n
) e
n
.
(d) Plancherels Equality: For each f H, |f|
2
=
n=1
[f, e
n
)[
2
.
(e) Parsevals Equality: For each f, g H, f, g) =
n=1
f, e
n
) e
n
, g).
Thus, for a countable orthonormal sequence, completeness implies the ex-
istence of series expansions of vectors. This need not be true for arbitrary
complete sequences, even in a Hilbert space (Problem A.28).
Denition A.104 (Orthonormal Basis). Let H be a Hilbert space. An
orthonormal sequence e
n
nN
which satises the equivalent conditions of
Exercise A.103 is called an orthonormal basis for H.
Remark A.105. Recall from Denition A.80 that a sequence f
n
nN
in a Ba-
nach space X is called a Schauder basis if every vector f X can be written
f =
n=1
c
n
(f)f
n
for a unique choice of scalars c
n
(f). In this terminology,
an orthonormal sequence is complete if and only if it is a Schauder basis.
However, it is important to emphasize that an arbitrary complete sequence
need not be a Schauder basis. Thus, in a Hilbert space we have:
complete orthonormal sequence
=
= /
Schauder basis.
We have been considering countable orthonormal bases. By Exercise A.84,
any Hilbert space that has a countable orthonormal basis must be separa-
ble. Conversely, every separable Hilbert space has an orthonormal basis (see
Problem A.31).
The results of this section can be extended to uncountable orthonormal se-
quences in nonseparable Hilbert spaces (such as
2
(I) when I is uncountable),
but then we must be extremely careful regarding the use of the terminology
basis, as in much of the Banach space literature the terminology basis is
reserved for a countable sequence that forms a Schauder basis. We summarize
without proof the results that hold for uncountable orthonormal sequences in
Hilbert spaces (see [Con90] for details).
Theorem A.106. Let H be a Hilbert space and let I be an index set. If e
i
iI
is an orthonormal set in H, then the following statements hold.
(a) If f H then f, e
i
) , = 0 for at most countably many i I.
(b) For each f H,
iI
[f, e
i
)[
2
|f|
2
.
(c) For each f H, the series p =
iI
f, e
i
) e
i
converges with respect to
the net of nite subsets of I, and p is the orthogonal projection of f onto
span
_
e
i
iI
_
.
Theorem A.107. Let e
i
iI
be an orthonormal set in a Hilbert space H.
Then the following statements are equivalent.
(a) e
i
iI
is complete.
(b) For each f H we have f =
iI
f, e
i
) e
i
, where the series converges
with respect to the net of nite subsets of I.
(c) For each f H, |f|
2
=
iI
[f, e
i
)[
2
.
Note that, because of orthogonality, the series that appear in both The-
orems A.106 and A.107 contain only countably many nonzero terms. If we
restrict to just those countably many nonzero terms, then, in the language of
Section A.10, the series converge unconditionally.
Additional Problems
A.28. We say that a sequence f
n
nN
in a Banach space X is -dependent
if there exist scalars c
n
, not all zero, such that
n=1
c
n
f
n
= 0, where the
series converges in the norm of X. A sequence is -independent if it is not
-dependent.
(a) Show that every Schauder basis is complete and -independent.
(b) Let , C be xed nonzero scalars such that [[ > [[. Let e
n
nN
be the standard basis for
2
, and dene x
0
= e
1
and x
n
= e
n
+ e
n+1
for
n N. Prove that x
k
k0
is complete and nitely independent in
2
, but is
not -independent and therefore is not a Schauder basis for
2
.
A.29. Let f
n
nN
be a sequence in a Hilbert space H. Prove that the fol-
lowing two statements are equivalent.
(a) For each m N we have f
m
/ spanf
n
n=m
(such a sequence is said
to be minimal ).
A.13 Urysohns Lemma 39
(b) There exists a sequence g
n
nN
in H such that f
m
, g
n
) =
mn
for
all m, n N (we say that sequences f
n
nN
and g
n
nN
satisfying this
condition are biorthogonal ).
Show further that in case statements (a), (b) hold, the sequence g
n
nN
is unique if and only if f
n
nN
is complete.
Remark: Exercise C.111 establishes an analogous result for Banach spaces.
A.30. Formulate the GramSchmidt orthogonalization procedure in an arbi-
trary inner product space, and use it to show that any innite-dimensional
inner product space contains an innite orthonormal sequence e
n
nN
.
A.31. Use the Axiom of Choice in the form of Zorns Lemma (Axiom G.2) to
show that every Hilbert space has a complete orthonormal subset. In partic-
ular, every separable Hilbert space has an orthonormal basis.
A.13 Urysohns Lemma
Urysohns Lemma is a general result that holds in a large class of topological
spaces (specically, the normal topological spaces, which include all metric
spaces). It states that if A and B are disjoint closed subsets of a normal
topological space X, then there exists a continuous function f : X [0, 1]
that is identically 0 on A and identically 1 on B. We prove here a version
of Urysohns Lemma for R
d
(and the same simple proof can be used in any
metric space).
First, we need the following lemma.
Lemma A.108. If E R
d
is nonempty, then
f(x) = dist(x, E) = inf
_
[x z[ : z E
_
is uniformly continuous on R
d
.
Proof. Fix > 0. Choose any x, y R
d
with [xy[ < /2. By denition, there
exist a, b E such that [xa[ < dist(x, E)+/2 and [yb[ < dist(y, E)+/2.
Hence
f(y) = dist(y, A) [y a[
[y x[ +[x a[
<

2
+ dist(x, E) +

2
= f(x) +.
Similarly f(x) < f(y) +, so [f(x) f(y)[ < whenever [x y[ < /2.
Theorem A.109 (Urysohns Lemma). If E, F are disjoint closed subsets
of R
d
, then there exists a continuous function : R
d
R
d
such that
(a) 0 1,
(b) = 0 on E, and
(c) = 1 on F.
Proof. Because E is closed, if x / E then we have dist(x, E) > 0. Also, by
Lemma A.108, dist(x, E) and dist(x, F) are each continuous functions of x. It
follows that the function
(x) =
dist(x, E)
dist(x, E) + dist(x, F)
has the required properties.
B
Lebesgue Measure and Integral
In much of harmonic analysis, measures make an appearance in essentially
two distinct ways. First, Lebesgue measure on R (and sometimes on R
2
or
R
d
), is used to dene the Fourier transform and the spaces that it acts upon,
such as the Lebesgue spaces L
p
(R). Second, the class M
b
(R) of bounded Borel
measures enters as another space of objects upon which the Fourier transform
acts.
Since Lebesgue measure is a particular unbounded Borel measure, if our
goal was conciseness then it would make sense to develop abstract measure
theory with Lebesgue measure as one special case of this theory. However,
since the role of Lebesgue measure is so fundamental and because there are
many facts that we will need that are specic to Lebesgue measure, we choose
to review the theory of Lebesgue measure and integration here as a topic
in itself, and to separately review the theory of general Borel measures in
Appendix D.
For more detail on this material, we refer to the texts by Wheeden and Zyg-
mund [WZ77] and Folland [Fol99]. Additionally, more details on the Banach
Zarecki Theorem can be found in the texts by Benedetto [Ben76] or Bruckner,
Bruckner, and Thomson [BBT97].
B.1 Exterior Lebesgue Measure
We begin with the familiar notion of the volume of a rectangular box in R
d
,
which for simplicity we refer to as a cube (even though we do not require
all side lengths to be equal).
Denition B.1. A cube in R
d
is a set of the form
Q = [a
1
, b
1
] [a
d
, b
d
] =
d
i=1
[a
i
, b
i
].
The volume of this cube is
42 B Lebesgue Measure and Integral
vol(Q) = (b
1
a
1
) (b
d
a
d
) =
d
i=1
(b
i
a
i
).
We extend the notion of volume to arbitrary sets by covering them with
countably many cubes in all possible ways.
Denition B.2. The exterior Lebesgue measure or outer Lebesgue measure
of a set E R
d
is
[E[
e
= inf
_
k
vol(Q
k
)
_
where the inmum is taken over all nite or countable collections of cubes Q
k
such that E
k
Q
k
.
Thus, every subset of R
d
has a uniquely dened exterior measure, which
lies in the range 0 [E[
e
.
There are many seemingly obvious results concerning Lebesgue measure
that actually require rather tedious proofs, many of which we will omit. For
example, although it is clear from the denition that [Q[
e
vol(Q), it requires
some care to show that the exterior measure of a cube actually coincides with
its volume. We state this now without proof.
Theorem B.3. (a) If Q is a cube in R
d
then [Q[
e
= vol(Q).
(b) If Q
1
, . . . , Q
n
are disjoint cubes in R
d
, then
k=1
Q
k
e
=
n
k=1
vol(Q
k
).
The next exercise gives some basic properties of exterior measure.
Exercise B.4. (a) Monotonicity: If E F R
d
, then [E[
e
[F[
e
.
(b) Countable subadditivity: If E
k
R
d
for k N, then
k=1
E
k
k=1
[E
k
[
e
.
(c) Translation invariance: If E R
d
and h R
d
, then [E+h[
e
= [E[
e
, where
E +h = x +h : x E.
Our next theorem gives a type of regularity property for exterior
Lebesgue measure: Every set E can be surrounded by an open set U whose
exterior measure is only larger than that of E (by monotonicity we also have
[E[
e
[U[
e
, so the measure of U is very close to the measure of E).
Theorem B.5. If E R
d
and > 0, then there exists an open set U E
such that [U[
e
[E[
e
+, and hence
[E[
e
= inf
_
[U[
e
: U open, U E
_
. (B.1)
B.2 Lebesgue Measure 43
Proof. If [E[
e
= , take U = R
d
. Otherwise we have [E[
e
< , so by
denition of exterior measure there must exist cubes Q
k
such that E Q
k
and
vol(Q
k
) < [E[
e
+

2
. Let Q
k
be a larger cube that contains Q
k
in its
interior, and such that vol(Q
k
) vol(Q
k
) + 2
k1
. Let U be the union of
the interiors of the cubes Q
k
. Then E U, U is open, and
[U[
e

k
vol(Q
k
)
k
vol(Q
k
) +

2
< [E[
e
+.
Since E and UE are disjoint and their union is U, we might expect that
the sum of their exterior measures is the exterior measure of U. Unfortunately,
this is false in general (although the Axiom of Choice is required to show the
existence of a counterexample. Consequently, the fact that [U[
e
[E[
e
+
does not imply that [UE[
e
! The well-behaved sets for which this is
true will be said to be measurable, and will be studied next.
B.2 Lebesgue Measure
B.2.1 Denition and Basic Properties
Denition B.6. A set E R
d
is Lebesgue measurable, or simply measurable,
if
> 0, open U E such that [UE[
e
.
If E is Lebesgue measurable, then its Lebesgue measure is [E[ = [E[
e
. We set
L = L(R
d
) = E R
d
: E is Lebesgue measurable.
Thus, there is no dierence between the numeric value of the Lebesgue
measure and the exterior Lebesgue measure of a measurable set, but when we
know that E is measurable we denote this value using the symbols [E[ instead
of [E[
e
.
Exercise B.7. Show that if [E[
e
= 0, then E L.
Consequently, if [E[
e
= 0, then not only is E measurable, but every subset
of E is measurable. In the language of abstract measure theory, the measure
space (R
d
, L, [ [) is said to be complete.
The following result, whose proof will be omitted, summarizes some of the
basic properties of measurable sets.
Theorem B.8. (a) L is a -algebra of subsets of R
d
. That is:
i. , R
d
L,
ii. if E
1
, E
2
, . . . L, then E
k
L,
iii. if E L, then R
d
E L.
(b) Every open and every closed set belongs to L.
Since L is closed under complements and countable unions, it follows that
it is closed under countable intersections as well. General -algebras are dis-
cussed in more detail in Section D.1.
Since exterior Lebesgue measure is subadditive, the same is true for
Lebesgue measurable sets, i.e., if E
k
for k N are measurable subsets of
R
d
then we have
E
k

[E
k
[. Although we will not prove it, one of the
main reasons for restricting to Lebesgue measurable sets is that for these sets
we have additivity of the Lebesgue measure of disjoint measurable sets. The
analogous statement for exterior Lebesgue measure does not hold in general!
Theorem B.9 (Countable Additivity of Lebesgue Measure). If E
1
,
E
2
, . . . are disjoint Lebesgue measurable subsets of R
d
, then
k=1
E
k
k=1
[E
k
[.
Here are some additional basic properties of Lebesgue measure.
Exercise B.10. Let E and E
k
for k N be measurable subsets of R
d
, and
prove the following facts.
(a) If E
1
E
2
and [E
1
[ < , then [E
2
E
1
[ = [E
2
[ [E
1
[.
(b) Continuity from below: If E
1
E
2
, then
E
k
= lim
k
[E
k
[.
(c) Continuity from above: If E
1
E
2
and [E
1
[ < , then
E
k
=
lim
k
[E
k
[.
(d) Translation invariance: If h R
d
, then [E + h[ = [E[, where E + h =
x +h : x E.
When a linear change of variable is made, the measure of a set is multiplied
by the absolute value of the determinant of the transformation. Linear trans-
formations are special cases of Lipschitz functions, compare Denition B.71
and Problem B.22.
Theorem B.11. If T : R
d
R
d
is linear and E R
d
is Lebesgue measurable,
then T(E) is measurable and [T(E)[ = [det(T)[ [E[.
B.2.2 Equivalent Formulations of Measurability
We will recast measurability in several equivalent ways.
Denition B.12. (a) A set H R
d
is a G
-set if there exist nitely or

countably many open sets U
k
such that H = U
k
.
(b) A set H R
d
is an F
-set if there exist nitely or countably many closed

sets F
k
such that H = F
k
.
Exercise B.13. Show that if E R
d
, then there exists a G
-set H E such
that [E[
e
= [H[.
B.2 Lebesgue Measure 45
Exercise B.14. Let E R
d
be given. Show that the following statements
are equivalent.
(a) E is Lebesgue measurable.
(b) For every > 0, there exists a closed set F E such that [EF[
e
.
(c) E = HZ where H is a G
-set and [Z[ = 0.

(d) E = H Z where H is an F
-set and [Z[ = 0.

The following exercise is a nice application of these characterizations to
prove the seemingly obvious statement that the measure of a Cartesian
product of two sets is the product of the measures of those sets. The proof is
surprisingly nontrivial, and consists of proceeding through cases (open sets,
G
sets, zero measure sets, and nally arbitrary measurable sets).

Exercise B.15. Show that if E R
m
and F R
n
are Lebesgue measurable,
then E F R
m+n
is Lebesgue measurable, and [E F[ = [E[ [F[.
Our statement of Lebesgue measurability, as given in Denition B.6, is for-
mulated in terms of the existence of surrounding open sets, and Exercise B.14
likewise interprets measurability in terms of sets with other topological prop-
erties. On the other hand, the equivalent formulation of measurability given
in the next theorem does not depend on any topological notions. As such,
this Caratheodory Criterion is the appropriate denition to use to generalize
measurability to more abstract settings, as is done in Appendix D.
Theorem B.16 (Caratheodorys Criterion). A set E R
d
is Lebesgue
measurable if and only if
A R
d
, [A[
e
= [A E[
e
+[AE[
e
. (B.2)
Proof. . Suppose that E is measurable and that A is any subset of R
d
. Since
A = (AE) (AE), we have by subadditivity that [A[
e
[AE[
e
+[AE[
e
.
By Exercise B.13, we can nd a G
-set H A such that [H[ = [A[

e
. Note that
we can write H as the disjoint union H = (H E) (HE). Since Lebesgue
measure is countably additive and H, E are measurable, we therefore have
[A[
e
= [H[ = [H E[ +[HE[ [A E[
e
+[AE[
e
,
where the nal inequality follows from monotonicity.
. Suppose that equation (B.2) holds. Assume rst that E is bounded,
and let H E be a G
-set such that [H[ = [E[

e
. Then equation (B.2) implies
that
[E[
e
= [H[ = [H E[
e
+[HE[
e
= [E[
e
+[HE[
e
.
Since [E[
e
< , we conclude that Z = HE has zero exterior measure and
hence is measurable. Since E = HZ, it is measurable as well.
Exercise: By considering the sets E
k
= x E : [x[ k, extend to
arbitrary sets E that satisfy equation (B.2).
B.2.3 Almost Everywhere
Notation B.17. A property that holds except possibly on a set of measure
zero is said to hold almost everywhere, abbreviated a.e.
For example, if C is the classical Cantor middle-thirds set, then [C[ = 0
(Problem B.1). Hence, the characteristic function

C
of C satises

C
(x) = 0
except for those x that belong to the zero measure set C. Therefore we say
that

C
(x) = 0 for almost every x, or

C
= 0 a.e. for short.
The essential supremum of a function is an example of a quantity that is
dened in terms of a property that holds almost everywhere.
Denition B.18 (Essential Supremum). The essential supremum of a
function f : E [, ] is
ess sup
xE
f(x) = infM : f(x) M a.e.. (B.3)
We say that f is essentially bounded if
ess sup
xE
[f(x)[ < .
Exercise B.19. Show that the inmum in equation (B.3) is achieved, i.e.,
if we set M = ess sup
xE
f(x) then we have f(x) M a.e. In particular, if
ess sup
xE
[f(x)[ = 0, then f = 0 a.e.
Additional Problems
B.1. Show that every countable set E R
d
satises [E[
e
= 0 and hence is
Lebesgue measurable. Show that the Cantor middle-thirds set C is an example
of an uncountable subset of R that satises [C[
e
= 0.
B.2. Dene the inner Lebesgue measure of a set E R
d
to be
[E[
i
= sup
_
[F[ : F closed, F E
_
.
Show that if [E[
e
< , then E is Lebesgue measurable if and only if [E[
e
=
[E[
i
, but this can fail if [E[
e
= .
B.3. Show that continuity from below holds for Lebesgue exterior measure,
i.e., if E
1
E
2
is any nested increasing sequence of subsets of R
d
, then
[ E
k
[
e
= lim
k
[E
k
[
e
.
B.3 Measurable Functions 47
B.3 Measurable Functions
One of the goals of this appendix is to dene the Lebesgue integral of functions
on R
d
. We will not be able to integrate every function, and, in particular, the
functions that we will integrate will need to be measurable. We rst dene
measurability for extended real-valued functions; a complex-valued function
will be measurable if its real and imaginary parts are measurable.
Denition B.20 (Real-Valued Measurable Functions). Let E R
d
and
f : E [, ] be given. Then f is a Lebesgue measurable function, or sim-
ply a measurable function, if f
1
(, ] = x E : f(x) > is a measurable
subset of R
d
for each R.
Notation B.21. For convenience of notation, we often use abbreviations such
as
f > a = x E : f(x) > = f
1
(, ]
and
f g = x E : f(x) g(x).
Every continuous function f : R
d
R is measurable. However, a measur-
able function need not be continuous.
d
be given. Show that E is a Lebesgue measurable
set if and only if

E
is a Lebesgue measurable function.
From now on, we will always implicitly assume that the domain E of a
measurable function f is a Lebesgue measurable set. Often, given a particular
measurable set E, it will be convenient to consider functions on the domain
EZ where Z is a set of measure zero. In this case we say that f is dened
almost everywhere on E.
Measurability is preserved under most of the usual operations, including
addition, multiplication, and limits. Some care does need to be taken with
compositions, but if we compose a measurable function with a continuous
function in the correct order, then measurability will be assured.
d
be Lebesgue measurable.
(a) If f : E [, ] is measurable and g = f a.e., then g is measurable.
(b) If f, g : E [, ] are measurable and nite a.e., then so is f +g.
(c) If f : E [, ] is measurable and nite a.e. and : R R is con-
tinuous, then f is measurable. Consequently, [f[, f
2
, f
+
, f
, and [f[
p
for p > 0 are all measurable.
(d) If f, g : E [, ] are measurable and nite a.e., then so is fg.
(e) If f
n
: E [, ] are measurable for n N, then so are sup f
n
, inf f
n
,
limsup f
n
, and liminf f
n
.
(f) If f
n
: E [, ] are measurable for n N and f(x) = lim
n
f
n
(x)
exists for a.e. x, then f is measurable.
Denition B.24 (Complex-Valued Measurable Functions). Given a
domain E R
d
and a complex-valued function f : E C, write f in real and
imaginary parts as f = f
r
+if
i
. Then we say that f is measurable if both f
r
and f
i
are measurable.
Although we will not prove it, Lusins Theorem states that if f is a mea-
surable function on a compact interval, then we can nd a continuous function
that equals f except on a set of arbitrarily small measure.
Theorem B.25 (Lusins Theorem). If f : [a, b] C is Lebesgue measur-
able and > 0, then there exists a continuous function g : [a, b] C such that
[f ,= g[ < .
Additional Problems
B.4. (a) Show that f : R
d
R is Lebesgue measurable if and only if f
1
(U)
is measurable for every open U R.
(b) Let f : R
d
R be Lebesgue measurable. Suppose that g : R R
is Borel measurable, i.e., g
1
(U) is a Borel set for every open U R (see
Denition D.3). Show that the composition g f is Lebesgue measurable.
Also generalize to the case of complex-valued f, g.
B.5. Show that if f : R R is continuous a.e., i.e., f is continuous at almost
every point, then f is Lebesgue measurable. Give an example of a function
f that is continuous a.e., but such that there is no continuous function g
satisfying f = g a.e.
B.6. Show that if f : R R is dierentiable at almost every point, then f
is
Lebesgue measurable.
B.4 Convergence in Measure
In this section we consider a particular notion of convergence related to mea-
sure.
Denition B.26 (Convergence in Measure). Let E R
d
be Lebesgue
measurable, and let f
n
, f be measurable functions on E that are either
complex-valued or are extended real-valued but nite a.e. Then we say that
f
n
converges in measure to f on E, and write f
n
m
f, if
> 0, lim
n
x E : [f(x) f
n
(x)[ >
= 0.
B.5 The Lebesgue Integral 49
In general, convergence in measure does not imply pointwise convergence.
Exercise B.27. Give an example of functions f
n
: R R such that f
n
m
0,
but f
n
(x) does not converge pointwise to zero as n .
On the other hand, a sequence of functions that converges in measure will
always have a subsequence that converges pointwise almost everywhere.
d
be Lebesgue measurable. Assume either that f
n
,
f : E [, ] are measurable and nite a.e., or that f
n
, f : E C are
measurable. If f
n
m
f, show there exists a subsequence f
n
k
kN
such that
f
n
k
f pointwise a.e.
There is also a Cauchy criterion for convergence in measure.
d
be Lebesgue measurable. Assume either that f
n
,
f : E [, ] are measurable and nite a.e., or that f
n
, f : E C are
measurable, and show that the following statements are equivalent.
(a) There exists a measurable f such that f
n
m
f.
(b) For each > 0 there exists an N > 0 such that
[f
m
f
n
[ >
< for
all m, n > N.
B.5 The Lebesgue Integral
To dene the Lebesgue integral of a measurable function, we rst begin with
simple functions and then extend through cases to nonnegative functions,
real-valued functions, and complex-valued functions.
B.5.1 Integration of Nonnegative Simple Functions
Denition B.30 (Simple Functions). Let E R
d
be Lebesgue measur-
able. A simple function on E is a function : E C of the form
=
N
k=1
a
k
E
k
, (B.4)
where N > 0, a
k
C, and the E
k
are Lebesgue measurable subsets of E.
Thus, a simple function is a measurable function that takes only nitely
many distinct scalar values. If a
1
, . . . , a
N
C are the distinct values assumed
by a simple function and we set E
k
= = a
k
, then has the form given
in equation (B.4) and furthermore the sets E
1
, . . . , E
N
form a partition of E.
We call this the standard representation of .
Note that if and are simple functions on E, then so are + and .
The integral of a nonnegative simple function is the measure of the region
under its graph (recall our convention that 0 = 0).
Denition B.31. If is a nonnegative simple function on E with standard
representation =
N
k=1
a
k

E
k
, then the Lebesgue integral of over E is
_
=
_
(x) dx =
N
k=1
a
k
[E
k
[.
Note that if is a simple function on E and A E is measurable, then
A
is a simple function on A, and
_
A
=
_
E

A
.
Exercise B.32. Prove the following facts for nonnegative simple functions ,
on E R
d
.
(a)
_
E
c = c
_
E
for c 0.
(b)
_
E
( +) =
_
E
+
_
E
.
(c) If then
_
E

_
E
.
(d)
_
E
= 0 if and only if = 0 a.e.
(e) If A
1
, A
2
, . . . are disjoint measurable subsets of E and A = A
k
, then
_
A
=
k=1
_
A
k
.
Remark B.33. In the language of Appendix D, the countable additivity prop-
erty given in part (e) of Exercise B.32 implies that (A) =
_
A
denes a
positive measure on the Lebesgue measurable subsets of E. Just as Exer-
cise B.10 shows that Lebesgue measure is continuous from below, the same is
true for the measure . That is, if A
1
A
2
and A = A
k
, then
_
A
= lim
k
_
A
k
. (B.5)
B.5.2 Integration of Nonnegative Functions
We obtain the integral of a nonnegative function by considering all possible
approximations from below by simple functions.
Denition B.34 (Lebesgue Integral of a Nonnegative Function). Let
E R
d
be Lebesgue measurable. If f : E [0, ] is a measurable function,
then the Lebesgue integral of f over E is
_
E
f =
_
E
f(x) dx = sup
__
E
: 0 f, simple
_
.
If A is a measurable subset of E, then we write
_
A
f to mean
_
E
f

A
.
When E = R
d
, we write simply
_
f or
_
f(x) dx to denote the integral of
f over R
d
.
Following are some of the basic properties of integrals of nonnegative func-
tions.
d
and f, g : E [0, ] be measurable, and prove
the following facts.
(a) If is a simple function on E, then the integrals of given in Deni-
tions B.31 and B.34 coincide.
(b) If f g then
_
E
f
_
E
g.
(c) Tchebyshevs Inequality: If > 0, then [x E : f(x) > [
1
_
E
f.
(d)
_
E
f = 0 if and only if f = 0 a.e.
The denition of
_
E
f given in Denition B.34 is often cumbersome to
implement. One application of the next result (which is also known as the
Beppo-Levi Theorem) is that we will be able to obtain the integral of f as
a limit instead of supremum of integrals of simple functions. This is quite
useful, since limits are linear while suprema are not in general.
We say that a sequence of extended real-valued functions f
n
nN
is mono-
tone increasing if f
1
(x) f
2
(x) for all x. We write f
n
f to mean
that f
n
nN
is monotone increasing and f
n
(x) f(x) pointwise.
Theorem B.36 (Monotone Convergence Theorem). Let E R
d
be
measurable, and assume f
n
nN
are nonnegative measurable functions on E
such that f
n
f. Then
lim
n
_
E
f
n
=
_
E
f.
Proof. Since
__
E
f
n
_
nN
is an increasing sequence of real scalars, I =
lim
n
_
E
f
n
exists as an extended real number. Further, 0 I
_
E
f
since 0 f
n
f for all n.
Choose any simple function with 0 f, and x 0 < < 1. Set
E
n
= f
n
and observe that E
1
E
2
and E
n
= E. By the
continuity from below property given in equation (B.5), we therefore have
that
_
En

_
E
. Consequently,
_
E
f
n

_
En
f
n

_
En

_
E
as n ,
and therefore I
_
E
. Since < 1 is arbitrary it follows that I
_
E
.
Taking the supremum over all such simple functions , we conclude that
I
_
E
f.
As an application of the Monotone Convergence Theorem, we can prove
additivity of the integral of nonnegative functions. In order to do this, we need
to show that we can always nd simple functions that increase pointwise to
an arbitrary nonnegative measurable function.
Exercise B.37. Suppose that E R
d
and f, g : E [0, ] are measurable.
(a) Show that
n
(x) =
_
_
j 1
2
n
, if
j 1
2
n
f(x) <
j
2
n
, j = 1, . . . , n2
n
,
n, if f(x) n.
(B.6)
is a simple function,
n
(x) f(x) for each x, and if f is bounded then
n
converges uniformly to f. Apply the Monotone Convergence Theorem
to conclude that
_
E

n

_
E
f.
(b) Show that
_
E
(f +g) =
_
E
f +
_
E
g.
Since the partial sums of a series of nonnegative functions form a monotone
increasing sequence, we can apply the Monotone Convergence Theorem to the
issue of interchanging a sum with an integral.
Corollary B.38. If f
n
nN
is a sequence of measurable, nonnegative func-
tions on a measurable set E R
d
, then
_
E
_

n=1
f
n
_
=
n=1
_
E
f
n
.
In particular, if f : E [0, ] is measurable, A
1
, A
2
, . . . are disjoint and
measurable, and A = A
k
, then
_
A
f =
k
_
A
k
f.
If we have functions f
n
that are not monotone increasing, then we may
not be able to interchange a limit with an integral. On the other hand, the
following result states that as long as the f
n
are all nonnegative, we do at
least have an inequality.
Exercise B.39 (Fatous Lemma). If f
n
nN
is a sequence of measurable,
nonnegative functions on a measurable set E R
d
, then
_
E
_
liminf
n
f
n
_
liminf
n
_
E
f
n
.
Exercise B.40. Show that strict inequality can hold in Fatous Lemma.
Since changing the value of a function on a set of zero measure does not
change the value of its integral, it suces to assume that the hypotheses
in the Monotone Convergence Theorem hold a.e. instead of everywhere, and
similarly for other theorems in this section.
B.5.3 Integration of Real-Valued and Complex-Valued Functions
We dene the integral of a general real-valued function by writing it as a
dierence of two nonnegative functions.
Denition B.41 (Lebesgue Integral of a Real-Valued Function). Let
E R
d
be measurable, and suppose that a measurable extended real-valued
function f : E [, ] is given. Dene
f
+
(x) = max
_
f(x), 0
_
, f
(x) = max
_
f(x), 0
_
.
Then f
+
, f
0, and we have f = f
+
f
and [f[ = f
+
+ f
. We dene
the Lebesgue integral of f on E to be
_
E
f =
_
E
f
+
_
E
f
,
as long as this does not have the form (in that case, the integral is
undened).
Similarly, the integral of a complex-valued function is dened by breaking
into real and imaginary parts.
Denition B.42 (Lebesgue Integral of a Complex-Valued Function).
Let E R
d
be measurable, and let f : E C be a measurable complex-
valued function on E. Write the real and imaginary parts of f as f = f
r
+if
i
.
If
_
E
f
r
and
_
E
f
i
both exist and are nite, then the Lebesgue integral of f on
E is
_
E
f =
_
E
f
r
+ i
_
E
f
i
.
Note that if f is complex-valued and
_
E
f exists then, by denition,
_
E
f
is a (nite) complex scalar.
Exercise B.43. Given an extended real-valued or a complex-valued measur-
able function f, show that
_
E
f exists and is a nite scalar if and only if
_
E
[f[ < . Further, in this case we have
_
E
f

_
E
[f[.
B.5.4 The Lebesgue Dominated Convergence Theorem
We have already seen several theorems, including the Monotone Convergence
Theorem and Fatous Lemma, that deal with the issue of interchanging a limit
and an integral. The Dominated Convergence Theorem is the workhorse of the
stable of convergence theorems.
Exercise B.44 (Lebesgue Dominated Convergence Theorem). As-
sume f
n
nN
is a sequence of measurable functions on a measurable set
E R
d
, either extended real-valued or complex-valued, such that:
(a) f(x) = lim
n
f
n
(x) exists for a.e. x E, and
(b) there exists g L
1
(E) such that [f
n
(x)[ g(x) a.e. for every n.
Then f
n
converges to f in L
1
-norm, i.e.,
lim
n
|f f
n
|
1
= lim
n
_
E
[f f
n
[ = 0,
and, consequently,
lim
n
_
E
f
n
=
_
E
f.
B.5.5 Relation to the Riemann Integral
For functions dened on nite intervals [a, b], we can ask how the Lebesgue in-
tegral of f relates to its Riemann integral. Unfortunately, not every Lebesgue
integrable function is Riemann integrable. For example, the characteristic
function

Q
of the rational numbers is Lebesgue integrable but not Riemann
integrable. We state without proof the following characterization of Riemann
integrability.
Theorem B.45. A bounded function f is Riemann integrable on [a, b] if and
only if f is continuous at almost every point in [a, b].
Note that continuity a.e. does not imply that there exists a continuous
function g such that f = g a.e.
Although not every Lebesgue integrable function is Riemann integrable,
the good news is that, on a nite interval, every Riemann integrable function
is Lebesgue integrable, and the two integrals agree.
Theorem B.46. If f is a bounded function that is Riemann integrable on a
nite interval [a, b], then it is Lebesgue integrable on [a, b] and its Riemann
integral equals its Lebesgue integral
_
b
a
f.
The situation is somewhat more complicated when dealing with improper
Riemann integrals. For example, Problem B.18 below shows that the improper
Riemann integral
_

0
sin x
x
dx = lim
a
_
a
0
sin x
x
dx =

2
exists. However, if we set f(x) =
sin x
x
then we have both
_
0
f
+
= and
_
0
f
= , so the Lebesgue integral of f on [0, ) does not exist. In essence,

improper Riemann integrals may exist because of fortunate cancellations,
while the Lebesgue integral requires absolute convergence.
B.6 The L
p
Spaces 55
Additional Problems
B.7. Let f : R C be Lebesgue measurable. Show that there exist simple
functions
k
such that
k
(x) f(x) pointwise as k , [
k
(x)[ [f(x)[
for every k and x, and the convergence is uniform on every set on which f is
bounded.
B.6 The L
p
Spaces
In this section we introduce and examine the Lebesgue spaces L
p
(E), which
are Banach spaces for 1 p and complete metric spaces for 0 < p < 1.
B.6.1 Norm and Completeness
Denition B.47 (Integrable Function). Let E R
d
be measurable. Then
a measurable function f on E (either extended real-valued or complex-valued)
is integrable on E if
_
E
[f[ < .
Note that any integrable function must be nite almost everywhere.
Denition B.48. Let E R
d
be measurable.
(a) If 0 < p < , then L
p
(E) consists of all measurable functions f : E C
such that [f[
p
is integrable, i.e.,
|f|
p
=
__
E
[f[
p
_
1/p
< .
(b) For p = , the space L
(E) consists of all those measurable functions

f : E C for which [f[ is essentially bounded, i.e.,
|f|
= ess sup
xE
[f(x)[ < .
We refer to L
p
(R) for p < as the Lebesgue space of p-integrable func-
tions, and to L
(R) as the Lebesgue space of essentially bounded functions.

The proof of H olders Inequality for
p
, given in Theorem A.16, carries
over to L
p
(E).
Exercise B.49 (H olders Inequality). Let E R be measurable, and x
1 p . Let p
denote the dual index to p. If f L

p
(E) and g L
p
(E)
then fg L
1
(E), and
|fg|
1
|f|
p
|g|
p
.
For 1 < p < , this inequality is
_
E
[fg[
__
E
[f[
p
_
1/p
__
E
[g[
p
_
1/p
.
Equality holds in H olders Inequality if and only if there exist scalars , ,
not both zero, such that [f(x)[
p
= [g(x)[
p
a.e. (Problem B.12).

We saw in Exercise A.17 that, as a consequence of H olders Inequality,
the
p
spaces are Banach spaces for p 1. Unfortunately, for the Lebesgue
spaces,
_
E
[f[
p
= 0 only implies that f = 0 a.e. Consequently, while | |
p
is a
seminorm, it is not a norm on L
p
(E).
Exercise B.50. If E R
d
is measurable and 1 p , then | |
p
is a
seminorm on L
p
(E).
The Triangle Inequality on L
p
is also known as Minkowskis Inequality.
By making some appropriate identications, we can deal with the technical
complication that | |
p
is only a seminorm, and in the end prove Banach space
properties of the L
p
spaces. The basic problem is illustrated by the fact that
if E is any set of measure zero then |
E
|
p
= 0 even though

E
is not the zero
function. However, we do have

E
= 0 a.e., which suggests that when dealing
with the L
p
spaces we may not wish to distinguish between functions that are
equal almost everywhere. Indeed, the next exercise is the standard procedure
for converting a seminorm into a norm by forming equivalence classes.
Exercise B.51. Show that the relation f g if f = g a.e. is an equivalence
relation on L
p
(E). Let

f denote the equivalence class of f in L
p
(E) under
this relation, and set [[[
f[[[
p
= |f|
p
. Show that this quantity is independent of
the choice of representative f of

f. Let the quotient space

L
p
(E) consist of
all the distinct equivalence classes of f L
p
(E), i.e.,
L
p
(E) =
_
f : f L
p
(E)
_
,
and show that

L
p
(E) is a normed space with respect to [[[ [[[
p
.
Typically we abuse notation and let the symbol f denote the equivalence
class

f of all functions equal to f a.e., and we write L
p
(E) instead of

L
p
(E).
In other words, we identify any two functions that are equal a.e. Adopting
this convention, we will shortly see that the L
p
spaces are not only normed
spaces, but are complete.
B.6.2 On Abuses of Notation
Ignoring the distinction between a function and the equivalence class of func-
tions that are equal to it a.e. is not usually a problem, but on occasion some
care needs to be taken. One such situation arises when dealing with continu-
ous functions. Every function in C
b
(R) is continuous and bounded, so we often
write C
b
(R) L
(R). However, in doing so we are really identifying C

b
(R)
with its image in L
(R) under the equivalence relation , i.e., if f C

b
(R)
then it determines an equivalence class

f of functions that are equal to it
B.6 The L
p
Spaces 57
almost everywhere, and it is this equivalence class

f that belongs to L
(R).
Conversely, if we are given f L
(R) (really an equivalence class

f of func-
tions) and there is a representative of this equivalence class that belongs to
C
b
(R), then we write f C
b
(R), meaning that there is a representative of f
that belongs to C
b
(R).
Remark B.52. The two statements f is continuous a.e. and f equals a
continuous function a.e. are distinct. The rst means that lim
yx
f(y) =
f(x) for almost every x, while the second means that there exists a continuous
function g such that f(x) = g(x) for almost every x. Only in the latter case can
we say that there is a representative of f that is a continuous function.
Exercise B.53. Show that if f C
b
(R), then
ess sup
xR
[f(x)[ = sup
xR
[f(x)[.
Consequently, for continuous bounded functions, the uniform norm dened in
Exercise A.20 coincides with the L
-norm dened above.

Another place where the fact that elements of L
p
(E) are equivalence classes
must be taken into account is when discussing the support of a function in
L
p
(E). For example,

Q
is one representative of the zero function in L
p
(R),
yet the support of

Q
is the entire real line. Unfortunately, the support of a
function depends very much on the choice of representative. Still, it is such a
convenient concept that we usually abuse notation and apply support termi-
nology to elements of L
p
(E). For example, we write f has compact support
with the understanding that this means that some representative of f has
compact support. Thus, f has compact support if f(x) = 0 for almost every x
outside of some compact set K.
When dealing with support issues, it is often enough to know that a func-
tion is zero almost everywhere outside of some particular set. However, it is
sometimes necessary to be more precise about the support of a function. In
this volume, we adopt the following denition of the support of f L
p
(R
d
).
This denition corresponds to taking the support of f to be the support of the
distribution determined by f. This denition can be adapted to elements of
L
p
(E), although some care should be taken if E is not a closed subset of R
d
.
Denition B.54. The support of f L
p
(R
d
) is
supp(f) =
_
F R
d
: F is closed and f(x) = 0 for a.e. x / F
_
.
In particular, if F R
d
is closed, then
supp(f) F f(x) = 0 for a.e. x / F.
Although we now have two denitions of support (Denition A.19 and
Denition B.54), it is usually clear from context which is meant. If a given
function f is dened everywhere then we apply Denition A.19, while if f
represents an equivalence class of functions that are equal almost everywhere,
then we use Denition B.54. If f is a continuous function (in the sense that
there is a representative of f that is continuous), then the two denitions
coincide.
B.6.3 Convergence in L
p
(E)
In the
p
spaces, convergence in
p
norm implies componentwise convergence
(see Problem A.6). The situation in L
p
(E) is a little dierent. For p = , if
f
n
f in L
(E) then it follows that f

n
(x) f(x) pointwise a.e. However,
for p nite, an L
p
-convergent sequence need not converge pointwise.
Exercise B.55. Let 0 < p < be xed. Give an example of functions f
n

L
p
(R) such that f
n
0 in L
p
-norm (i.e., |f
n
|
p
0), but f
n
(x) does not
converge pointwise to zero a.e. as n .
Fortunately, it is true that an L
p
-convergent sequence always has a subse-
quence that converges pointwise almost everywhere.
d
be measurable and x 0 < p . Show that if
f
n
, f L
p
(E) and f
n
f in L
p
(E), then f
n
m
f. Consequently, there exists
a subsequence f
n
k
kN
such that f
n
k
(x) f(x) for almost every x E.
The preceding exercise can be used to prove the completeness of the L
p
spaces.
d
be measurable, and prove the following state-
ments.
(a) If 1 p , then | |
p
is a norm on L
p
(E), and L
p
(E) is a Banach
space with respect to this norm.
(b) If 0 < p < 1, then d(f, g) = |f g|
p
p
is a metric on L
p
(E), and L
p
(E) is
complete with respect to this metric.
The space L
2
(E) is special. As before, we consider elements of L
2
(E) to
be equivalence classes of functions that are equal almost everywhere.
Exercise B.58. If E R
d
is measurable, show that f, g) =
_
E
f(x) g(x) dx
denes an inner product on L
2
(E), and L
2
(E) is a Hilbert space with respect
to this inner product.
It is often useful to know that we can approximate a given L
p
function
by functions that have some special properties. For example, combining the
Lebesgue Dominated Convergence Theorem with Exercise B.37, we see that
the L
p
simple functions are dense in L
p
(E). In fact, we can even restrict
further to simple functions with compact support when p is nite.
B.6 The L
p
Spaces 59
d
be Lebesgue measurable. Show that the set
of all simple functions on E is dense in L
p
(E) for each 1 p , and
the set of compactly supported simple functions is dense in L
p
(E) for each
1 p < .
We can then use the denseness of the simple functions to prove that the
space of continuous, compactly supported functions is dense in L
p
(R
d
) for
nite p. For p = , we have instead that C
c
(R
d
) is dense in C
0
(R
d
) in the
uniform norm.
Theorem B.60. C
c
(R
d
) is dense in L
p
(R
d
) for each 1 p < .
Proof. First consider the function f =

E
where E R
d
is bounded. If we
x > 0, then there exists a bounded open set U E such that [UE[ < .
By Problem B.2, we can also nd a compact set K E such that [EK[ < .
By Urysohns Lemma (Theorem A.109), we can nd a continuous function
: R
d
R such that 0 1, = 1 on K, and = 0 on R
d
U. Then
C
c
(R
d
), and we have
|
E
|
p
p
=
_
[
E
[
p
=
_
U\K
[
E
[
p
[UK[ < 2.
Hence

E
can be approximated arbitrarily closely in L
p
-norm by elements of
C
c
(R
d
). Exercise: Complete the proof by making use of Exercise B.59.
For p = , the space C
c
(R
d
) is not dense in L
(R
d
), but it is dense in
C
0
(R
d
) with respect to the uniform norm.
Theorem B.60 can be used to show that the space of nite linear combina-
tions of characteristic functions of intervals (sometimes called really simple
functions) is dense in L
p
(R). This provides us with another useful set of
approximating functions for L
p
(R).
Exercise B.61. Show that
[a,b]
: < a < b < is complete in L
p
(R)
when 1 p < .
B.6.4 Local Integrability
We introduce one nal space in this section. This space is not a Banach space,
but it is very useful when we only need to consider integrability at a local
level.
Denition B.62 (Locally Integrable Functions). The space of locally
integrable functions on R
d
is
L
1
loc
(R
d
) =
_
f : R
d
C : f
K
L
1
(R
d
) for every compact K R
d
_
.
As with the L
p
spaces, we regard the elements of L
1
loc
(R
d
) as being equiv-
alence classes of functions that are equal almost everywhere. L
1
loc
(R
d
) is not a
Banach space, but it is a typical example of a topological vector space whose
topology is dened by an innite family of seminorms instead of a single norm
(see Example E.5).
If we likewise dene L
p
loc
(R
d
) for p > 1, then we have
L
p
(R
d
) L
p
loc
(R
d
) L
1
loc
(R
d
), 1 p .
L
1
loc
(R
d
) contains many functions that do not belong to any L
p
(R
d
). For
example, every polynomial belongs to L
1
loc
(R), as does e
x
.
Additional Problems
B.8. Let f L
1
(R
d
) be given. Given > 0, show that there exists a > 0
such that for any measurable E R
d
satisfying [E[ < , we have
_
E
[f[ < .
In particular, if [E[ = 0, then
_
E
f = 0.
B.9. Show that if [E[ < and 0 < p q , then L
q
(E) L
p
(E). In
contrast, show that L
p
(R) is not contained in L
q
(R) for any p and q.
B.10. Prove that if [E[ < , then |f|
p
|f|
as p .
B.11. Show that L
p
(R) is separable for 1 p < , but L
(R) is not sepa-

rable.
B.12. Show that equality holds in H olders Inequality if and only if there exist
scalars , , not both zero, such that [f[
p
= [g[
p
a.e.
B.13. This problem generalizes H olders Inequality to the case of more than
two functions. Show that if
1
p1
+ +
1
p
k
= 1 and f
i
L
pi
(R) for i = 1, . . . , k,
then f
1
f
k
L
1
(R) and |f
1
f
k
|
1
|f
1
|
p1
|f
k
|
p
k
.
B.14. Given 1 p < q , show that L
p
(R) L
q
(R) is a Banach space
under the norm |f| = |f|
p
+ |f|
q
. Further, if 1 p < r < q then we
have L
p
(R) L
q
(R) L
r
(R), with
|f|
r
|f|
p
|f|
1
q
where
1
p
+

q
=
1
r
.
Show also that if r < then L
p
(R) L
q
(R) is dense in L
r
(R).
B.15. Given 1 p < q , show that
L
p
(R) +L
q
(R) =
_
f +g : f L
p
(R), g L
q
(R)
_
is a Banach space under the norm
|f| = inf
_
|g|
p
+|h|
q
: f = g +h with g L
p
(R), h L
q
(R)
_
.
Further, if 1 p < r < q then we have L
r
(R) L
p
(R) +L
q
(R).
B.7 Repeated Integration 61
B.7 Repeated Integration
Let E R
m
and F R
n
be measurable. If f is a measurable function on
E F then there are three natural integrals of f over E F. First, there is
the integral of f over the set E F R
m+n
, which we write as the double
integral
__
EF
f =
__
EF
f(x, y) (dxdy).
Second, for each xed y we can integrate f(x, y) as a function of x, and then
integrate the result in y, obtaining the iterated integral
_
F
__
E
f(x, y) dx
_
dy.
Third, we also have the iterated integral
_
E
__
F
f(x, y) dy
_
dx.
In general these three integrals need not be equal, even if they all exist.
In this section we state without proof the theorems of Fubini and Tonelli,
which give sucient conditions under which we can exchange the order of
integration. We begin with Tonellis Theorem, which states that interchange is
allowed if f is nonnegative. In particular, this suggests that a counterexample
to equality of the integrals must be related to the indeterminacy of
(see Problem B.16).
Theorem B.63 (Tonellis Theorem). Let E be a measurable subset of R
m
and F a measurable subset of R
n
. If f : E F [0, ] is measurable, then
the following statements hold.
(a) f
x
(y) = f(x, y) is measurable on F for each x E.
(b) f
y
(x) = f(x, y) is measurable on E for each y F.
(c) g(x) =
_
F
f
x
(y) dy is a measurable function on E.
(d) h(y) =
_
E
f
y
(x) dx is a measurable function on F.
(e) As extended real numbers,
__
EF
f(x, y) (dxdy) =
_
F
__
E
f(x, y) dx
_
dy
=
_
E
__
F
f(x, y) dy
_
dx. (B.7)
As a corollary, we obtain the useful fact that to test whether a given
function belongs to L
1
(E F) we can simply show that any one of three
possible integrals is nite.
Corollary B.64. Let E be a measurable subset of R
m
and F a measurable
subset of R
n
. If f is a measurable function on E F, then (as extended real
numbers):
__
EF
[f(x, y)[ (dxdy) =
_
F
__
E
[f(x, y)[ dx
_
dy =
_
E
__
F
[f(x, y)[ dy
_
dx.
Consequently, if any one of these three integrals is nite, then we have f
L
1
(E F).
Fubinis Theorem allows the interchange of integrals if f is integrable
(thereby again avoiding the ambiguity that is ).
Theorem B.65 (Fubinis Theorem). Let E be a measurable subset of R
m
and F a measurable subset of R
n
. If f L
1
(E F), then the following state-
ments hold.
(a) f
x
(y) = f(x, y) is measurable and integrable on F for almost every x E.
(b) f
y
(x) = f(x, y) is measurable and integrable on E for almost every y F.
(c) g(x) =
_
F
f
x
(y) dy is a measurable and integrable function on E.
(d) h(y) =
_
E
f
y
(x) dx is a measurable and integrable function on F.
(e) We have
__
EF
f(x, y)(dxdy) =
_
F
__
E
f(x, y) dx
_
dy
=
_
E
__
F
f(x, y) dy
_
dx.
Additional Problems
B.16. Show that the following iterated integrals have the indicated values:
_

1
__

1
x
2
y
2
(x
2
+y
2
)
2
dy
_
dx =
4
,
_

1
__

1
x
2
y
2
(x
2
+y
2
)
2
dx
_
dy =

4
,
_

1
__

1
x
2
y
2
(x
2
+y
2
)
2
dx
_
dy = .
Conclude that equality need not hold in equation (B.7) if the hypotheses of
Fubinis Theorem are not fullled.
B.17. Let f(x, y) be a measurable function on R
m+n
= R
m
R
n
, and x
1 p < . Prove Minkowskis Integral Inequality:
B.8 Functions of Bounded Variation 63
__ __
[f(x, y)[ dy
_
p
dx
_
1/p
_ __
[f(x, y)[
p
dx
_
1/p
dy. (B.8)
Remark: This equation may be more revealing if we rewrite it as
_
_
_
_
_
[f(, y)[ dy
_
_
_
_
p
_
|f(, y)|
p
dy.
Thus, Minkowskis Integral Inequality is an integral version of the Triangle
Inequality (also known as Minkowskis Inequality) on L
p
(R
m
).
B.18. If x > 0, then
_
0
e
tx
dt =
1
x
. Combine this with Fubinis Theorem to
evaluate the integral
_
a
0
sin x
x
dx. Then apply the Lebesgue Dominated Con-
vergence Theorem to show that lim
a
_
a
0
sin x
x
dx =

2
. Thus, even though
sinx
x
is not integrable, the improper Riemann integral
_
0
sin x
x
dx exists and
equals

2
(this integral can also be evaluated by using contour integration).
B.19. This problem will establish a version of Hardys Inequalities.
(a) Given 1 p < and < 1, show there exists a constant C(, p) such
that for any measurable f : (0, ) [0, ] we have
_

0
__
x
0
f(t) dt
_
p
x
dx C(, p)
_

0
f(t)
p
t
+p
dt. (B.9)
Show that if > 1 then the inequality is
_

0
__

x
f(t) dt
_
p
x
dx C(, p)
_

0
f(t)
p
t
+p
dt.
(b) For the case = p < 1, show that the optimal constant in equa-
tion (B.9) is
C(p, p) =
_
p
_
p
=
_
p
p 1
_
p
.
(c) Suppose that f L
p
(R) where 1 < p < . Dene F(x) =
1
x
_
x
0
[f(t)[ dt
for x R, and show that
|F|
p
p
|f|
p
, (B.10)
with p
being the best possible constant. Also show that equality holds in
equation (B.10) if and only if f = 0 a.e.
B.8 Functions of Bounded Variation
In this section we briey review the denition and basic properties of functions
with bounded variation on the real line.
B.8.1 Denition and Examples
We begin by dening bounded variation for functions on nite closed intervals
in R.
Denition B.66. Let f : [a, b] C be given. Given any nite partition =
a = x
0
< < x
n
= b of [a, b], set
S
=
n
j=1
[f(x
j
) f(x
j1
)[.
The variation of f over [a, b] is
V [f; a, b] = sup
_
S
: is a partition of [a, b]
_
.
The function f has bounded variation on [a, b] if V [f; a, b] < . We set
BV[a, b] =
_
f : [a, b] C : f has bounded variation
_
.
Note that in this denition we are considering f to be a function that is
dened at all points, rather than an equivalence class of functions that are
equal a.e.
The space BV[a, b] is sometimes dened to consist of real-valued functions
of bounded variation, but all the denitions and results extend to complex-
valued functions, which is the setting of most of this volume. We could just
as well have dened bounded variation for real-valued functions, and then
declared a complex-valued function to have bounded variation if its real and
imaginary parts have bounded variation.
Exercise B.67. Given f : [a, b] C, write the real and imaginary parts as
f = f
r
+if
i
. Show that f BV[a, b] if and only if f
r
, f
i
BV[a, b].
For functions on the domain R we make the following denition.
Denition B.68. The variation of a function f : R C is
V [f; R] = sup
a<b
V [f; a, b].
We say that f has bounded variation if V [f; R] < , and we dene
BV(R) =
_
f : R C : f has bounded variation
_
.
Example B.69. A function f : [a, b] R is monotone increasing if
a x y b = f(x) f(y).
In this case f has bounded variation, and V [f; a, b] = f(b) f(a).
Exercise B.70. Dene f(x) = xsin
1
x
for x ,= 0 and f(0) = 0. Show that f
is continuous, but has unbounded variation on [1, 1].
Lipschitz functions on bounded domains all have bounded variation.
Denition B.71. A function f : [a, b] C is Lipschitz on [a, b] if there exists
a constant C > 0 (called a Lipschitz constant for f) such that
[f(x) f(y)[ C [x y[, x, y [a, b].
The class of Lipschitz functions on [a, b] is denoted by
Lip[a, b] =
_
f : [a, b] C : f is Lipschitz
_
.
Notation B.72. We will often need to discuss the dierentiability properties
of functions on nite intervals. We will say that a function f is everywhere
dierentiable on [a, b] if it is dierentiable on the interior (a, b) and if the
appropriate one-sided derivatives exist at the endpoints. In other words, f is
everywhere dierentiable on [a, b] if
lim
yx,
y[a,b]
f(y) f(x)
y x
= f
(x) exists and is nite for all x [a, b].

For example, x
3/2
is dierentiable everywhere on [0, 1], and x
1/2
is dier-
entiable everywhere on [0, 1] except at x = 0.
Exercise B.73. Prove the following.
(a) If f is Lipschitz on [a, b], then f is uniformly continuous and has bounded
variation, and V [f; a, b] C (b a).
(b) A Lipschitz function need not be dierentiable everywhere on [a, b].
(c) If f is dierentiable everywhere on [a, b] and f
is bounded on [a, b], then

f is Lipschitz with C = |f
. In particular, if f, f
are both continuous

on [a, b], then f is Lipschitz.
Not every function of bounded variation need be Lipschitz, see Exer-
cise B.85 below.
Here are some basic properties of the variation function V [f; a, b].
Exercise B.74. Let f : [a, b] C be given.
(a) Show that if
is a renement of , then S
.
(b) Show that if [a
, b
] [a, b], then V [f; a
, b
] V [f; a, b].
(d) Show that if a < c < b, then V [f; a, b] = V [f; a, c] +V [f; c, b].
B.8.2 The Jordan Decomposition
Our next goal is to prove the Jordan decomposition, which characterizes real-
valued functions of bounded variation as the dierence of two monotone in-
creasing functions.
First we introduce the positive and negative variation functions.
Denition B.75. Let f : [a, b] R be given. Given a nite partition =
a = x
0
< < x
n
= b of [a, b], dene
S
+
=
n
i=1
_
f(x
i
) f(x
i1
)
+
and S
=
n
i=1
_
f(x
i
) f(x
i1
)
.
Thus S
+
is the sum of the positive terms of S
, and S
is the sum of the

negative terms. The positive variation of f on [a, b] is
V
+
[f; a, b] = sup
_
S
+
_
,
and the negative variation is
V
[f; a, b] = sup
_
S
_
.
Observe that S
+
+S
= S
and S
+
= f(b)f(a). The next exercise

extends these equalities from particular partitions to the variation functions.
Exercise B.76. Show that if f : [a, b] R, then
V
+
[f; a, b] +V
[f; a, b] = V [f; a, b].

Further, if any one of V [f; a, b], V
+
[f; a, b], or V
[f; a, b] is nite, then they

are all nite, and in this case we also have
V
+
[f; a, b] V
[f; a, b] = f(b) f(a).

Now we can prove the Jordan decomposition.
Theorem B.77 (Jordan Decomposition). If f : [a, b] R is given, then
(a) f BV[a, b].
(b) There exist monotone increasing functions f
1
, f
2
: [a, b] R such that
f = f
1
f
2
.
Proof. (a) (b). For x [a, b], the functions V
+
[f; a, x] and V
[f; a, x] are
monotonically increasing with x. Furthermore, by Exercise B.76 we have
V
+
[f; a, x] V
[f; a, x] = f(x) f(a).

Therefore f = f
1
f
2
where f
1
= V
+
[f; a, x] +f(a) and f
2
= V
[f; a, x], and

f
1
, f
2
are each monotonically increasing.
Consequently, a complex-valued function f : [a, b] C will have bounded
variation if and only if we can write f = (f
+
1
f
1
) +i(f
+
2
f
2
) where f
+
1
,
f
1
, f
+
2
, f
2
are monotone increasing.
B.8.3 Dierentiability of Functions of Bounded Variation
An important property of monotone increasing functions is that they are
dierentiable a.e.
Theorem B.78. If f : [a, b] R is monotone increasing, then f
(x) exists
for a.e. x (in fact, for all but at most countably many x), f
is measurable
and nonnegative a.e., and
0
_
b
a
f
f(b) f(a). (B.11)

Proof. Note that f is bounded by hypothesis, and extend f to R by setting
f(x) = f(a) for x < a and f(x) = f(b) for x > b.
Since f is increasing and bounded, f(x) = lim
yx
f(y) and f(x+) =
lim
yx
+ f(y) exist for each x. Hence each point of discontinuity of f must be
a jump discontinuity. Further, since f is bounded and increasing, given any
xed k N, the set of x such that
f(x+) f(x)
1
k
must be nite. Since every jump discontinuity must satisfy this inequality
for some k N, we conclude that there can be at most countably many
discontinuities.
The proof that f is dierentiable almost everywhere is somewhat more
involved and will be omitted, see [Fol99] or [WZ77] for details. Assuming this,
we have that the functions
f
k
(x) =
f(x + 1/k) f(x)
1/k
= k
_
f(x +
1
k
) f(x)
_
converge pointwise to f
(x) a.e. as k . In particular, f
0 a.e., and
applying Fatous Lemma we have
_
b
a
f
liminf
k
_
b
a
f
k
= k
_
b+
1
k
a+
1
k
f k
_
b
a
f
= k
_
b+
1
k
b
f k
_
a+
1
k
a
f
k
_
b+
1
k
b
f(b) k
_
a+
1
k
a
f(a) (since f is increasing)
= f(b) f(a).
Exercise B.81 shows that equality need not hold in equation (B.11).
Combining Theorem B.78 with the Jordan decomposition, we see that
all functions of bounded variation on [a, b] are dierentiable a.e. and have
integrable derivatives.
Corollary B.79. If f BV[a, b], then f
(x) exists for a.e. x, and f
L
1
[a, b].
The nal result stated in this section gives a useful connection between f
and the total variation function.

Theorem B.80. Given f BV[a, b], set V (x) = V [f; a, x] for x [a, b].
Then V
(x) = [f
(x)[ for almost every x [a, b].

Additional Problems
B.20. Show that if f, g BV[a, b], then f + g BV[a, b] for all , C
(so BV[a, b] is a vector space), and fg BV[a, b]. If [g(x)[ > 0 for all
x [a, b] then f/g BV[a, b].
B.21. Set f(x) = x
2
sin(1/x) and g(x) = x
2
sin(1/x
2
) for x ,= 0, and f(0) =
g(0) = 0. Show that f and g are dierentiable everywhere, and f BV[1, 1]
but g / BV[1, 1] (compare Problem B.23).
B.22. Let E R be measurable, and suppose that f : E R is Lipschitz
on E, i.e., [f(x) f(y)[ C [xy[ for all x, y E. Prove that if A E, then
[f(A)[
e
C [A[
e
(compare Lemma B.94).
B.9 Absolutely Continuous and Singular Functions
In this section we review the properties of absolutely continuous functions on
the real line (which are those functions for which the Fundamental Theorem
of Calculus holds) and singular functions on R (which are dierentiable at
almost every point but have the property that the derivative is zero a.e.).
B.9.1 Singular Functions on the Real Line
We begin with an example of a singular function.
Exercise B.81 (CantorLebesgue Function). Consider the two functions
1
,
2
pictured in Figure B.1. The function
1
takes the constant value 1/2
on the interval (1/3, 2/3) that is removed in the rst stage of the construction
of the Cantor middle-thirds set, and is linear on the remaining intervals. The
function
2
also takes the same constant 1/2 on the interval (1/3, 2/3) but
additionally is constant with values 1/4 and 3/4 on the two intervals that are
removed in the second stage of the construction of the Cantor set. Continue
this process, dening
3
,
4
, . . . , and prove the following facts.
B.9 Absolutely Continuous and Singular Functions 69
0 0.25 0.5 0.75 1
0
0.25
0.5
0.75
1
0 0.25 0.5 0.75 1
0
0.25
0.5
0.75
1
Fig. B.1. First stages in the construction of the CantorLebesgue function. Left:
The function 1. Right: The function 2.
(a) Each
k
is monotone increasing on [0, 1].
(b) [
k+1
(x)
k
(x)[ < 2
k
for every x [0, 1].
(c) (x) = lim
k
k
(x) converges uniformly on [0, 1].
The function constructed in this manner is called the CantorLebesgue
function or, more picturesquely, the Devils staircase. Prove the following
facts.
(d) is continuous and monotone increasing on [0, 1], but is not Lipschitz.
(e) is dierentiable for a.e. x [0, 1], and
(x) = 0 a.e.
(f) The Fundamental Theorem of Calculus does not apply to :
(1) (0) ,=
_
1
0
(x) dx.
If we extend to R by reecting it about the point x = 1 and declaring
it to be zero outside of [0, 2], we obtain the continuous function pictured in
Figure B.2. It is interesting that it can be shown that is an example of a
renable function, as it satises the following renement equation:
(x) =
1
2
(3x) +
1
2
(3x1) +(3x2) +
1
2
(3x3) +
1
2
(3x4). (B.12)
Thus equals a nite linear combination of compressed and translated copies
of itself, and so exhibits a type of self-similarity. Another example of a renable
function is discussed in Section B.10. Renable functions are widely studied
and play important roles in wavelet theory and in subdivision schemes in
computer-aided graphics, see [Dau92].
Exercise B.82. Renable functions are easy to plot to any desired level of
accuracy. For example, since we know the values of (k) for k integer, we
1 2
1
Fig. B.2. The reected Devils staircase (CantorLebesgue function).
can compute the values (k/3) for k Z by considering x = k/3 in equa-
tion (B.12). Iterating this, we can obtain the values (k/3
j
) for any k Z,
j N. Plot the CantorLebesgue function.
The CantorLebesgue function is the prototypical example of a singular
function.
Denition B.83 (Singular Function). A function f : [a, b] [, ] or
f : R C is singular if f is dierentiable at almost every point in its domain
and f
= 0 a.e.
B.9.2 Absolutely Continuous Functions on the Real Line
Now we turn to absolutely continuous functions. A collection of intervals in R
is said to be nonoverlapping if the interiors of the intervals are disjoint.
Denition B.84 (Absolutely Continuous Function). We say that a
function f : [a, b] C is absolutely continuous on [a, b] if for every > 0
there exists a > 0 such that for any nite or countably innite collection of
nonoverlapping subintervals
_
[a
j
, b
j
]
_
j
of [a, b], we have
j
(b
j
a
j
) < =
j
[f(b
j
) f(a
j
)[ < .
We denote the class of absolutely continuous functions on [a, b] by
AC[a, b] =
_
f : [a, b] C : f is absolutely continuous on [a, b]
_
.
The space of locally absolutely continuous functions on R is
AC
loc
(R) =
_
f : R C : f AC[a, b] for every a < b
_
.
Exercise B.85. Prove the following statements.
(a) If g AC[a, b], then g is uniformly continuous on [a, b].
(b) Lip[a, b] AC[a, b] BV[a, b].
Absolute continuity and singularity are complementary properties, in the
sense that only constant functions can be both absolutely continuous and
singular. The standard proof of this fact relies on the Vitali Covering Theorem
and will be omitted (see Problem B.26 for a dierent proof).
Theorem B.86. If f : [a, b] C is both absolutely continuous and singular,
then f is constant.
In order to give the connection between absolutely continuous functions
and the Fundamental Theorem of Calculus, we need the following important
result, whose standard proof relies on the HardyLittlewood maximal func-
tion and will be omitted. While stated for functions on R, it also applies to
functions on nite intervals [a, b], since any integrable function on [a, b] can
be extended to a locally integrable function on R, e.g., by declaring it to be
zero outside of [a, b].
Theorem B.87 (Lebesgue Dierentiation Theorem). Let f L
1
loc
(R)
be given, i.e., f is integrable on every compact subset of R. Then for almost
every x R,
lim
h0
1
h
_
x+h
x
f(y) dy = f(x).
Consequently, the indenite integral of f,
F(x) =
_
x
a
f(y) dy,
is dierentiable a.e., and F
= f a.e.
In fact, a stronger conclusion holds: If f L
1
loc
(R), then for a.e. x we have
lim
h0
1
h
_
x+h
x
[f(y) f(x)[ dy = lim
h0
1
2h
_
x+h
xh
[f(y) f(x)[ dy = 0. (B.13)
Moreover, the intervals [x, x + h] or [x h, x + h] can be replaced by any
collection of sets S
h
h>0
that shrink regularly to x, which means that
diam(S
h
) 0, and there exists a constant C > 0 such that if Q
h
is the
smallest interval centered at x that contains S
h
, then [Q
h
[ C [S
h
[.
Denition B.88 (Lebesgue Set). Given f L
1
loc
(R), a point x for which
equation (B.13) holds is called a Lebesgue point of f, and the set of Lebesgue
points is called the Lebesgue set of f.
Thus, almost every point in the domain of a locally integrable function is
a Lebesgue point.
Exercise B.89. Given f L
1
loc
(R), show that every point of continuity of f
is a Lebesgue point of f.
The next exercise is motivation for the Fundamental Theorem of Calculus
for absolutely continuous functions. This exercise shows that the antiderivative
of an integrable function is absolutely continuous.
Exercise B.90. Show that if f L
1
[a, b], then g(x) =
_
x
a
f(t) dt belongs to
AC[a, b], and furthermore g
= f a.e.
In fact, much more holds.
Theorem B.91 (Fundamental Theorem of Calculus). If g : [a, b] C,
then the following statements are equivalent.
(a) g AC[a, b].
(b) There exists f L
1
[a, b] such that
g(x) g(a) =
_
x
a
f(t) dt, x [a, b].
(c) g is dierentiable almost everywhere, g
L
1
[a, b], and
g(x) g(a) =
_
x
a
g
(t) dt, x [a, b].

Proof. (a) (c). Suppose that g is absolutely continuous on [a, b]. Then g
has bounded variation, and so by Corollary B.79 we know that g
exists a.e.
and is integrable. Therefore the function
G(x) =
_
x
a
g
is well-dened for each x [a, b]. Moreover, by the Lebesgue Dierentiation

Theorem, G
= g
a.e. Hence (Gg)
= 0 a.e., so the function Gg is singular

on [a, b]. On the other hand, both g and G are absolutely continuous on [a, b],
so Gg is absolutely continuous as well. Therefore we have by Theorem B.86
that Gg is constant. Consequently, given any x [a, b], we have
G(x) g(x) = G(a) g(a) = 0 g(a) = g(a).
Thus G(x) = g(x) g(a) for all x [a, b], so statement (c) holds.
If is the CantorLebesgue function on [0, 1], then is singular, and
hence is dierentiable almost everywhere with
L
1
[a, b], yet we have
(x) (0) ,=
_
x
0

(t) dt = 0, conrming the fact that is not absolutely

continuous.
We can use Exercise B.90 and Theorem B.86 to prove the following fun-
damental decomposition of functions of bounded variation.
Corollary B.92. If f BV[a, b], then f = g +h where g AC[a, b] and h is
singular on [a, b]. Moreover, g and h are unique up to additive constants, and
we can take
g(x) =
_
x
a
f
, x [a, b]. (B.14)

Proof. Since f has bounded variation on [a, b], we know that f
exists a.e. and

is integrable. Therefore the function g given by equation (B.14) is well-dened.
Set h = f g. By Exercise B.90, we have g AC[a, b] and g
= f
a.e., so
h
= 0 a.e. Thus h is singular.

If we also had f = g
1
+h
1
with g
1
absolutely continuous and h
1
singular,
then g g
1
= h
1
h, so g g
1
and h
1
h are each absolutely continuous and
singular, and therefore are constant by Theorem B.86.
An important fact is that integration by parts is valid for absolutely con-
tinuous functions.
Exercise B.93 (Integration by Parts). Show that if f, g AC[a, b], then
_
b
a
f(x) g
(x) dx = f(b)g(b) f(a)g(a)

_
b
a
f
(x) g(x) dx.

B.9.3 Preparation for the BanachZarecki Theorem
The BanachZarecki Theorem provides some reformulations of absolute con-
tinuity. To prove it, we will need two lemmas, and in order to motivate these
we recall Exercise B.73, which states that if f : [a, b] C is dierentiable
everywhere and f
is bounded, then f is Lipschitz and hence is absolutely

continuous on [a, b]. One of the implications of the BanachZarecki Theorem
is the much more subtle fact that if f is dierentiable everywhere on [a, b]
and we assume only that f
L
1
[a, b], then f is absolutely continuous. The
subtlety here is that while the assumptions f, f
L
1
[a, b] do imply that the
antiderivative g(x) =
_
x
a
f
(t) dt exists and is absolutely continuous, it is not

at all obvious whether g need equal f.
Our rst lemma is a renement of Problem B.22, which showed that if a
function f is Lipschitz on [a, b] and E is any subset of [a, b], then [f(E)[
e

C [E[
e
, where C is a Lipschitz constant for f. In particular, if f is dierentiable
on [a, b] and f
is bounded on [a, b], then we know that f is Lipschitz on every

subset of [a, b], and hence can apply Problem B.22 to this f. However, if we
only know that f is dierentiable and f
is bounded on a particular subset E

then we cannot apply Problem B.22. Still, by making the argument a little
more sophisticated, we can show that the same conclusion holds.
Lemma B.94. Let f : [a, b] R and E [a, b] be given. Suppose that f is
dierentiable at every point of E, and that
M = sup
xE
[f
(x)[ < .
Then
[f(E)[
e
M[E[
e
.
Proof. Fix > 0. Given x E, we have
lim
yx,
y[a,b]
[f(x) f(y)[
[x y[
= [f
(x)[ M.
Hence there exists some n
x
N such that if y [a, b] then
[x y[ <
1
n
x
= [f(x) f(y)[ (M +) [x y[.
Therefore, if for each n N we dene
E
n
=
_
x E : if y [a, b] and [xy[ <
1
n
then [f(x)f(y)[ (M+) [xy[
_
,
then we have that E = E
n
. Further, E
1
E
2
. Even though the sets
E
n
need not measurable, we have by Problem B.3 that continuity from below
holds for exterior Lebesgue measure, so
[E[
e
= lim
n
[E
n
[
e
.
As the sets f(E
n
) are also nested increasing and increase to f(E), we also
have
[f(E)[
e
= lim
n
[f(E
n
)[
e
.
Now, for each n, we can nd at most countably many intervals I
k
n
such
that
E
n

k
I
k
n
and
k
[I
k
n
[ [E
n
[
e
+.
By subdividing if necessary, we may assume that each interval I
k
n
has length
less than
1
n
. Therefore, if we take x, y E
n
I
k
n
, then we have [xy[ <
1
n
, so
[f(x) f(y)[ (M +) [x y[.
Consequently, the image f(E
n
I
k
n
) is contained in an interval of length at
most (M +) [I
k
n
[, so
[f(E
n
I
k
n
)[
e
(M +) [I
k
n
[.
Therefore
[f(E
n
)[
e

k
[f(E
n
I
k
n
)[ (M +)
k
[I
k
n
[ (M +) ([E
n
[
e
+).
Hence,
[f(E)[
e
= lim
n
[f(E
n
)[
e
(M + ) lim
n
([E
n
[
e
+)
= (M + ) ([E[
e
+).
Since is arbitrary, the result follows.
The hypothesis of dierentiability in Lemma B.94 can be weakened, see
[BBT97, Lemma 7.9].
The second lemma relates the exterior measure of f(E) to the integral
of [f
[ on E. Note that even though we now assume that E is measurable,

we cannot conclude that f(E) is measurable, and hence this result must also
be formulated in terms of the exterior Lebesgue measure of f(E) (compare
Problem B.24, which shows that an absolutely continuous function must map
measurable sets to measurable sets, but an arbitrary continuous function need
not do so).
Lemma B.95. Let f : [a, b] R be measurable. If E [a, b] is measurable
and f is dierentiable at every point of E, then
[f(E)[
e

_
E
[f
[.
Proof. By Problem B.6, the derivative f
is a measurable function, and hence

_
E
[f
[ does exist as an extended real number. For each k N, dene

E
k
= x E : (k 1) [f
(x)[ < k.
Since f is dierentiable everywhere on E, we have E = E
k
disjointly. Further,
by Lemma B.94 we have that [f(E
k
)[
e
k [E
k
[. Therefore
[f(E)[
e
=
k=1
f(E
k
)
k=1
[f(E
k
)[
e
k=1
k [E
k
[
=
k=1
(k 1) [E
k
[ +
k=1
[E
k
[
k=1
_
E
k
[f
[ + [E[
=
_
E
[f
[ + [E[.
Since is arbitrary, the result follows.
B.9.4 The BanachZarecki Theorem and its Implications
Now we prove the BanachZarecki Theorem.
Theorem B.96 (BanachZarecki Theorem). Given f : [a, b] C, the
(a) f AC[a, b].
(b) f is continuous, f BV[a, b], and [f(A)[ = 0 for every A [a, b] with
[A[ = 0.
(c) f is continuous and is dierentiable a.e., f
L
1
[a, b], and [f(A)[ = 0 for
every A [a, b] with [A[ = 0.
Proof. By breaking into real and imaginary parts, it suces to prove the result
for real-valued functions.
(a) (b). Suppose that f AC[a, b] is real-valued. Then f is continuous
and has bounded variation by Exercise B.85, so it only remains to show that
f maps zero measure sets to zero measure sets.
Suppose that A is a subset of [a, b] with measure zero. Since f(a), f(b)
is a set of measure zero, it suces to assume that A (a, b) with [A[ = 0.
Choose any > 0. Since f is absolutely continuous, there exists a > 0
such that if [a
j
, b
j
]
j
is any collection of nonoverlapping intervals such that
(b
j
a
j
) < , then
[f(b
j
) f(a
j
)[ < .
By Theorem B.5, we can nd an open set U A with measure [U[ <
[A[ + = , and by intersecting with the open interval (a, b) we may assume
that U (a, b). Then we can write U = (a
j
, b
j
) as a union of at most
countably many disjoint open intervals contained in (a, b). Since [a
j
, b
j
]
[a, b], there is a point c
j
[a
j
, b
j
] where f attains its minimum value on
[a
j
, b
j
], and likewise a point d
j
[a
j
, b
j
] where f attains its maximum. Then
we have
[d
j
c
j
[
(b
j
a
j
) < , so
[f(A)[
e
[f(U)[
e

f[a
j
, b
j
]
f(d
j
) f(c
j
)
< .
Since is arbitrary, we conclude that [f(A)[ = 0.
(c) (a). Assume that statement (c) holds for some real-valued f. Let
D be the set of points where f is dierentiable, so Z = [a, b]D has measure
zero.
Suppose that [c, d] is any subinterval of [a, b]. By the Intermediate Value
Theorem, f[c, d] contains either [f(c), f(d)] or [f(d), f(c)], depending on or-
der. Let E = [c, d] D and F = [c, d]D. Then F has measure zero, so by
hypothesis we have [f(F)[ = 0. Also, f is dierentiable everywhere on E, so
by Lemma B.95 we have
[f(d) f(c)[ [f([c, d])[
e
[f(E)[
e
+ [f(F)[
e
_
E
[f
[ + 0 =
_
d
c
[f
[, (B.15)
the nal equality following from the fact that E is a subset of [c, d] with full
measure.
Now choose any > 0. Since f
is integrable, by Problem B.8 there exists

a > 0 such that
_
E
[f
[ < for any measurable set E [a, b] with [E[ < .

Suppose that [a
j
, b
j
]
j
is a collection of nitely or countably many nonover-
lapping intervals in [a, b] such that
(b
j
a
j
) < . Dene E = [a
j
, b
j
], so
[E[ < . Then by equation (B.15), we have
f(b
j
) f(a
j
)
j
_
bj
aj
[f
[ =
_
E
[f
[ < ,
so f is absolutely continuous on [a, b].
Corollary B.97. If f : [a, b] C is everywhere dierentiable and f
L
1
[a, b], then f AC[a, b].
Proof. By splitting into real and imaginary parts, we may assume that f is
real-valued. Suppose that A [a, b] and [A[ = 0. Then since f is dierentiable
at every point of A, we have by Lemma B.95 that [f(A)[
e

_
A
[f
[ = 0.
Theorem B.96 therefore implies that f is absolutely continuous.
The hypotheses of Corollary B.97 can be relaxed somewhat. For example,
if f is dierentiable except at countably many points and f
L
1
[a, b], then f
will be absolutely continuous (Problem B.28). However, the assumptions that
f is dierentiable a.e. and f
L
1
[a, b] are by themselves not sucient to en-
sure that f is absolutely continuous (consider the CantorLebesgue function).
Additional Problems
B.23. Dene g(x) = x
2
sin(1/x
2
) for x ,= 0 and g(0) = 0. Show that g
L
1
[1, 1], g is everywhere dierentiable, g
/ L
1
[1, 1], and g / AC[1, 1]
(compare Problem B.21).
B.24. (a) Show that a continuous function need not map a measurable set to
a measurable set.
(b) Show that if f : [a, b] C is absolutely continuous and E [a, b] is
measurable, then [f(E)[ is measurable as well.
B.25. Show that AC[a, b] is closed under pointwise products, i.e., if f, g
AC[a, b] then fg AC[a, b].
B.26. Show that if f AC[a, b] and f
= 0 a.e. then f is constant.

B.27. Suppose that f AC
loc
[a, b]. In particular, f
exists a.e. Suppose there

is a continuous function g such that f
= g a.e. Show that f is dierentiable

everywhere and f
(x) = g(x) for all x [a, b].

B.28. Suppose that f : [a, b] C is dierentiable at all but countably many
points, and f
L
1
[a, b]. Show that f is absolutely continuous.
B.10 Holder Continuity
H older continuity is a generalization of Lipschitz continuity (Denition B.71).
Denition B.98. We say that a function f : R C is H older continuous
with exponent > 0 if there exists a constant K > 0 such that
x, y R, [f(x) f(y)[ K[x y[
.
Thus, Lipschitz continuity is H older continuity for the case = 1. The
reader should verify that the only functions that are H older continuous with
exponent > 1 are the constant functions.
While a Lipschitz function is almost dierentiable in some sense, the
graph of a function that is H older continuous with exponent < 1 typically
has a fractal appearance. The smaller that we must take , the more jagged
the graph of the function appears.
One example of a H older continuous function is the CantorLebesgue func-
tion.
Exercise B.99. Show that the CantorLebesgue function is H older continu-
ous precisely for exponents in the range 0 < log
3
2 0.6309 . . . .
Another example is the Daubechies D
4
function. This is the compactly
supported function that satises the four-term renement equation
D
4
(x) =
1 +
3
4
D
4
(2x) +
3 +
3
4
D
4
(2x 1)
+
3
3
4
D
4
(2x 2) +
1
3
4
D
4
(2x 3). (B.16)
Thus, D
4
exhibits a kind of self-similarity, as it equals a linear combination
of four smaller, shifted copies of itself. It can be shown that there exists a
unique (up to scale) compactly supported function that satises this rene-
ment equation, and furthermore this solution is H older continuous precisely
for those exponents that lie in the range
0 < log
2
_
1 +
3
4
_
0.5500 . . . ,
see [Dau92]. The CantorLebesgue function also satises a renement equa-
tion but based on dilation by 3 rather than 2, see equation (B.12).
Exercise B.100. For this exercise, assume that equation (B.16) has a solu-
tion that is continuous and compactly supported.
(a) Show that supp(D
4
) [0, 3].
(b) Combine part (a) with the renement equation to nd the values of D
4
(k)
for all integer k.
B.10 Holder Continuity 79
(c) Now compute the values D
4
(k/2) for k Z by considering x = k/2 in
equation (B.16). Iterating this, we can obtain the values D
4
(k/2
j
) for any
k Z, j N. Plot the Daubechies D
4
function.
When suitably normalized, D
4
has the interesting property that its integer
translates are orthonormal, i.e., D
4
(xk)
kZ
forms an orthonormal system
in L
2
(R). The Daubechies D
4
function is but one renable function that has
orthonormal integer translates. Each such orthonormal scaling function leads
to a second function, called a wavelet, which generates an orthonormal basis
for all of L
2
(R) via the operations of translation and dilation. For D
4
, the
corresponding wavelet is
W
4
(x) =
1
3
4
D
4
(2x)
3
3
4
D
4
(2x 1)
+
3 +
3
4
D
4
(2x 2)
1 +
3
4
D
4
(2x 3).
This function has the property that
_
2
n/2
W
4
(2
n
x k)
_
n,kZ
forms an or-
thonormal basis for L
2
(R), and
_
D
4
(x k)
_
kZ

_
2
n/2
W
4
(2
n
x k)
_
n0, kZ
is another orthonormal basis for L
2
(R).
Why the subscript 4? The Daubechies scaling function D
4
is the second
of an innite family of functions D
2N
NN
, each of which satises a 2N-
term renement equation, is supported in [0, 2N 1], and has orthonormal
integer translates. Moreover, the smoothness of D
2N
increases with N. For
example, D
2
=

[0,1]
is discontinuous, D
4
is continuous but not dierentiable,
and D
6
is just barely dierentiable. Each of these scaling functions has an
associated wavelet W
2N
whose integer translates and dyadic dilations form an
orthonormal basis for L
2
(R).
The rst wavelet, the function W
2
=

[0,1/2)

[1/2,1)
, was introduced
by Alfred Haar (18851933) in his 1910 Ph.D. thesis [Haa10], and is called
the Haar wavelet, see Problem B.29. It was not until much later that other
wavelets, such as the Daubechies scaling functions and wavelets, were discov-
ered. We will not be able to do justice to wavelet theory here, but only mention
that the main papers in the development of wavelet theory, including a trans-
lation of Haars original paper into English, appear in the reprint volume
[HW06]. More fundamentally, we refer to Daubechies classic text [Dau92] for
complete details on scaling functions and wavelets, and to the texts by Mallat
[Mal98] and Strang and Nguyen [SN96] for their relation to signal processing.
The text by Walnut [Wal02] is an accessible introduction to wavelet theory
and its applications.
0.5 1.0 1.5 2.0 2.5 3.0
1.0
0.5
0.5
1.0
1.5
0.5 1.0 1.5 2.0 2.5 3.0
1.0
0.5
0.5
1.0
1.5
Fig. B.3. Top: The Daubechies D4 scaling function. Bottom: The corresponding
wavelet W4.
Additional Problems
B.29. Let

=

[0,1)
be the box function, and let = W
2
=

[0,1/2)
[1/2,1)
be the Haar wavelet. Prove that
_
(x k)
_
kZ

_
2
n/2
(2
n
x k)
_
n0, kZ
forms an orthonormal basis for L
2
(R), called the Haar system. Observe that
satises the renement equation

(x) =

(2x) +

(2x 1), while is
determined from

by the equation (x) =

(2x)
(2x 1).
B.30. Given 0 < < 1, dene
C
(R) =
_
f C(R) : f is H older continuous with exponent
_
.
Show that
|f|
C
= [f(0)[ + sup
x=y
f(x) f(y)
[x y[
is a norm on C
(R), and C
(R) is complete with respect to this norm.

C
Functional Analysis and Operator Theory
C.1 Linear Operators on Normed Spaces
In this section we will reviewthe basic properties of linear operators on normed
spaces.
Denition C.1 (Notation for Operators). Let X, Y be vector spaces, and
let T : X Y be a function mapping X into Y. We write either T(f) or Tf
to denote the image under T of an element f X.
(a) T is linear if T(f +g) = T(f)+T(g) for every f, g X and , C.
(b) T is antilinear if T(f +g) = T(f) +

T(g) for f, g X and , C.
(c) T is injective if T(f) = T(g) implies f = g.
(d) The kernel or nullspace of T is ker(T) = f X : T(f) = 0.
(e) The range of T is range(T) = T(f) : f X.
(f) The rank of T is the vector space dimension of its range, i.e., rank(T) =
dim(range(T)). In particular, T is nite-rank if range(T) is nite-dimen-
sional.
(g) T is surjective if range(T) = Y.
(h) T is a bijection if it is both injective and surjective.
We use either the symbol I or I
X
to denote the identity map of a space
X onto itself.
A mapping between vector spaces is often referred to as an operator or a
transformation, especially if it is linear. We introduce the following terminol-
ogy for operators on normed spaces.
Denition C.2 (Operators on Normed Spaces). Let X, Y be normed
linear spaces, and let L: X Y be a linear operator.
82 C Functional Analysis and Operator Theory
(a) L is bounded if there exists a nite K 0 such that
f X, |Lf| K|f|.
By context, |Lf| denotes the norm of Lf in Y, while |f| denotes the
norm of f in X.
(b) The operator norm of L is
|L| = sup
f=1
|Lf|. (C.1)
On those occasions where we need to specify the spaces in question, we
will write |L|
XY
for the operator norm of L: X Y.
(c) We set
B(X, Y ) =
_
L: X Y : L is bounded and linear
_
.
If X = Y then we write B(X) = B(X, X).
(d) If Y = C then we say that L is a functional. The set of all bounded linear
functionals on X is the dual space of X, and is denoted
X
= B(X, C) =
_
L: X C : L is bounded and linear
_
.
Another common notation for the dual space is X
.
Notation C.3 (Terminology for Unbounded Operators). Unbounded
operators are often not dened on the entire space X but only on some dense
subspace. For example, the dierentiation operator Df = f
is not dened
on all of L
p
(R), but it is common to refer to the dierentiation operator D
on L
p
(R), with the understanding that D is only dened on some associated
dense subspace such as L
p
(R)C
1
(R) or o(R). Another common terminology
is to write that D: L
p
(R) L
p
(R) is densely dened, again meaning that
the domain of D is implicitly assumed to be a dense subspace of L
p
(R) and
D maps this domain into L
p
(R).
Exercise C.4. Let X, Y be normed linear spaces. Let L: X Y be a linear
operator, and prove the following statements.
(a) L is injective if and only if ker L = 0.
(b) If L is a bijection then the inverse map L
1
: Y X is also a linear
bijection.
(c) L is bounded if and only if |L| < .
(d) If L is bounded then |Lf| |L| |f| for every f X, and |L| is the
smallest K such that |Lf| K|f| for all f X.
C.1 Linear Operators on Normed Spaces 83
(e) |L| = sup
f1
|Lf| = sup
f=0
|Lf|
|f|
.
Example C.5. Consider a linear operator on a nite-dimensional space, say
L: C
n
C
m
. For simplicity, impose the Euclidean norm on both C
n
and
C
m
. If we let S = x C
n
: |x| = 1 be the unit sphere in C
n
, then
L(S) = Lx : |x| = 1 is a (possibly degenerate) ellipsoid in R
m
. The
supremum in the denition of the operator norm of L is achieved in this case,
and is the length of a semimajor axis of the ellipsoid L(S). Thus, |L| is the
maximum distortion of the unit sphere under L, illustrated for the case
m = n = 2 (with real scalars) in Figure C.1.
3 2 1 1 2 3
3
2
1
1
2
3
Fig. C.1. Image of the unit circle under a particular linear operator L: R
2
R
2
.
The operator norm L of L is the length of a semimajor axis of the ellipse.
C.1.1 Equivalence of Boundedness and Continuity
Our rst main result of this section shows that continuity is equivalent
to boundedness for linear operators on normed spaces. Recall that, by
Lemma A.53, if X and Y are normed spaces, then L: X Y is continu-
ous at a point f X if f
n
f in X implies Lf
n
Lf in Y, and L is
continuous if it is continuous at every point.
Theorem C.6 (Equivalence of Bounded and Continuous Linear Op-
erators). If X, Y are normed spaces and L: X Y is linear, then the
(a) L is continuous at some f X.
(b) L is continuous at f = 0.
(c) L is continuous.
(d) L is bounded.
Proof. (c) (d). Suppose that L is continuous but unbounded. Then we have
|L| = , so there must exist f
n
X with |f
n
| = 1 such that |Lf
n
| n.
Set g
n
= f
n
/n. Then |g
n
0| = |g
n
| = |f
n
|/n 0, so g
n
0. Since L is
continuous and linear, this implies Lg
n
L0 = 0. By the continuity of the
norm, we therefore have |Lg
n
| |0| = 0. However,
|Lg
n
| =
1
n
|Lf
n
|
1
n
n = 1
for all n, which is a contradiction. Hence L must be bounded.
Thus, if X, Y are normed and L: X Y is linear, the terms continuous
and bounded are interchangeable.
C.1.2 Isomorphisms
The notion of a topological isomorphism (or homeomorphism) between arbi-
trary topological spaces was introduced in Denition A.50. We repeat it here
for the case of normed spaces, along with additional terminology for operators
that preserve norms.
Denition C.7 (Isometries and Isomorphisms). Let X, Y be normed
spaces, and let L: X Y be linear.
(a) If L: X Y is a linear bijection such that both L and L
1
are continuous,
then L is called a topological isomorphism, or is said to be continuously
invertible.
(b) If there exists a topological isomorphism L: X Y, then we say that X
and Y are topologically isomorphic.
(c) If |Lf| = |f| for all f X then L is called an isometry or is said to be
norm-preserving.
(d) An isometry L: X Y that is a bijection is an isometric isomorphism.
(e) If there exists an isometry L: X Y then we say that X and Y are
isometrically isomorphic, and we write X

= Y in this case.
On occasion, we will deal with antilinear isometric isomorphisms, which
are entirely analogous except that the mapping L is antilinear instead of linear.
Remark C.8. The Inverse Mapping Theorem, which will be discussed in Sec-
tion C.13, states that if X and Y are Banach spaces and L: X Y is a
bounded linear bijection, then L
1
is automatically bounded and hence L is
a topological isomorphism. Thus, when X and Y are Banach spaces, every
continuous linear bijection is actually a topological isomorphism.
C.1 Linear Operators on Normed Spaces 85
We use the following special terminology for isometric isomorphisms on
Hilbert spaces.
Denition C.9 (Unitary Operator). If H, K are Hilbert spaces and
L: H K is an isometric isomorphism, then L is called a unitary opera-
tor, and in this case we say that H and K are unitarily isomorphic.
An isometry on an inner product space must preserve the inner product
as well as the norm.
Exercise C.10. Let H, K be inner product spaces, and let L: H K be a
linear mapping. Prove that L is an isometry if and only if Lf, Lg) = f, g)
for all f, g H.
C.1.3 Eigenvalues and Eigenvectors
We recall the denition of the eigenvalues and eigenvectors of a linear operator
that maps a space into itself.
Denition C.11 (Eigenvalues and Eigenvectors). Let X be a normed
space and L: X X a linear operator.
(a) A scalar is an eigenvalue of L if there exists a nonzero vector f X
such that Lf = f.
(b) A nonzero vector f X is an eigenvector of L if there exists a scalar
such that Lf = f.
If f is an eigenvector of L corresponding to the eigenvalue , then we often
say that f is a -eigenvector.
If is an eigenvalue of L, then ker(L I) is called the eigenspace corre-
sponding to , or the -eigenspace for short.
Additional Problems
C.1. Show that if X is any nite-dimensional vector space (under any norm)
and Y is any normed linear space, then every linear function L: X Y is
bounded.
C.2. (a) Dene L:
2
(N)
2
(N) by L(x) = (x
2
, x
3
, . . . ). Prove that this
left-shift operator is bounded, linear, surjective, not injective, and is not an
isometry. Find |L| and all eigenvalues and eigenvectors of L.
(b) Dene R:
2
(N)
2
(N) by R(x) = (0, x
1
, x
2
, x
3
, . . . ). Prove that
this right-shift operator is bounded, linear, injective, not surjective, and is an
isometry. Find |R| and all eigenvalues and eigenvectors of R.
(c) Compute LR and RL and show that LR ,= RL. Contrast this compu-
tation with the fact that in nite dimensions, if A, B: C
n
C
n
are linear
maps (hence correspond to multiplication by nn matrices), then AB = I if
and only if BA = I.
C.3. If X, Y are normed spaces and L: X Y is continuous, show that
ker(L) is a closed subspace of X.
C.4. Let X be a Banach space and Y a normed linear space. Suppose that
L: X Y is bounded and linear. Prove that if there exists a scalar c > 0 such
that |Lf| c|f| for all f X, then L is injective and range(L) is closed.
C.5. Show that if L: X Y is a topological isomorphism, then we have
|L
1
|
1
|f| |Lf| |L| |f| for all f X.
C.6. Show that if H, K are separable Hilbert spaces, then H and K are
unitarily isomorphic.
C.7. Let A be an mn complex matrix, which we view as a linear transfor-
mation A: C
n
C
m
. The operator norm of A depends on the choice of norm
for C
n
and C
m
. Compute an explicit formula for |A|, in terms of the entries
of A, when the norm on C
n
and C
m
is taken to be the
1
norm. Then do the
same for the
norm. Compare your formulas to the version of Schurs Test

given in Theorem C.20.
C.8. The Axiom of Choice implies that every vector space X has a Hamel
basis (Theorem G.3). Use this to show that if X is an innite-dimensional
normed linear space, then there exists a linear functional : X C that is
unbounded.
C.2 Some Useful Operators
In this section we describe several types of operators that appear often in the
main part of the text.
C.2.1 Orthogonal Projections
We begin with orthogonal projections in Hilbert spaces.
Denition C.12 (Orthogonal Projection). Let M be a closed subspace
of a Hilbert space H. Dene P : H H by Ph = p, where p is the orthogonal
projection of h onto M, (see Denition A.97). The operator P is the orthogonal
projection of H onto M.
Exercise C.13 (Properties of Orthogonal Projections). Let M ,= 0
be a closed subspace of a Hilbert space H, and let P be the orthogonal pro-
jection of H onto M. Show that the following statements hold.
(a) If h H then Ph is the unique vector in M such that h Ph M
.
(b) |h Ph| = dist(h, M) for every h H.
(c) P is linear, |Ph| |h| for every h H, and |P| = 1.
(d) P is idempotent, i.e., P
2
= P.
(e) ker(P) = M
and range(P) = M.
(f) I P is the orthogonal projection of H onto M
.
C.2 Some Useful Operators 87
C.2.2 Multiplication Operators
Next we consider two types of multiplication operators. The rst type mul-
tiplies each term in an orthonormal basis expansion by a xed scalar.
Exercise C.14. Let e
n
nN
be an orthonormal basis for a separable Hilbert
space H. Then, by Exercise A.103, we know that every f H can be written
f =
n=1
f, e
n
) e
n
. Fix any sequence of scalars = (
n
)
nN
. For those
f H for which the following series converges, dene
M
f =
n=1
n
f, e
n
) e
n
. (C.2)
Prove the following facts.
(a) The series dening M
f in (C.2) converges for every f H if and only

if
. In this case M
is a bounded linear mapping of H into itself,

and |M
| = ||
.
(b) If /
, then M
denes an unbounded linear mapping from

domain(M
) =
_
f H :
n=1
[
n
f, e
n
)[
2
<
_
(C.3)
into H. Note that domain(M
) contains the nite span of e

n
nN
and
hence is dense in H.
If H =
2
and e
n
nN
is the standard basis for
2
, then the multiplication
operator M
dened in equation (C.2) is simply componentwise multiplica-

tion: M
x = x = (
1
x
1
,
2
x
2
, . . . ) for x = (x
1
, x
2
, . . . )
2
. This is a discrete
version of the multiplication operator dened in the next exercise.
Exercise C.15. Let : R C and 1 p be given.
(a) Show that if L
(R), then M
f = f is a bounded mapping of
L
p
(R) into itself, and |M
| = ||
.
(b) Conversely, show that if f L
p
(R) for every f L
p
(R), then we
must have L
(R).
C.2.3 Integral Operators
Now we dene the important class of integral operators for the setting of the
real line.
Denition C.16 (Integral Operator). Let k be a xed measurable func-
tion on R
2
. Then the integral operator L
k
with kernel k is formally dened
by
L
k
f(x) =
_
k(x, y) f(y) dy, (C.4)
i.e., L
k
f is dened whenever this integral makes sense.
An integral operator is a generalization of ordinary matrix-vector multi-
plication. Let A be an mn matrix with entries a
ij
and let u C
n
be given.
Then Au C
m
, and its components are
(Au)
i
=
n
j=1
a
ij
u
j
, i = 1, . . . , m.
Thus, the function values k(x, y) are analogous to the entries a
ij
of the matrix
A, and the values L
k
f(x) are analogous to the entries (Au)
i
.
Example C.17 (Tensor Product Kernels). The tensor product of two functions
g, h on R is the function g h on R
2
dened by
(g h)(x, y) = g(x) h(y), x, y R.
Sometimes the complex conjugate is omitted in the denition of tensor prod-
uct, but it will be convenient for our purposes to include it.
An important special case of an integral operator is where the kernel k is
a tensor product. If we assume that g, h L
2
(R) and set k = g h, then for
f L
2
(R),
L
k
f(x) =
_
g(x) h(y) f(y) dy = f, h) g(x),
at least for all x for which g(x) is dened. If either g = 0 or h = 0 then L
k
is
the zero operator, otherwise the range of L
k
is the one-dimensional subspace
spanned by g. Thus, L
k
is a very simple operator in this case, being a
bounded, rank one operator on L
2
(R).
Notation C.18. When k = g h is a tensor product, we often identify the
operator L
k
with the kernel g h. In other words, given g and h we often let
g h denote the operator whose rule is
(g h)(f) = f, h) g, f L
2
(R).
We can extend this notion of an operator g h to arbitrary Hilbert spaces by
simply replacing L
2
(R) with H on the line above. That is, if g, h H then
we dene g h to be the rank one operator given by (g h)(f) = f, h) g
for f H. Note that if g = h and |g|
2
= 1, then g g is the orthogonal
projection of H onto the line through g.
In general, it is not obvious how to tie properties of the kernel k to prop-
erties of the corresponding integral operator L
k
. The next two theorems will
provide sucient conditions that imply L
k
is a bounded operator on L
2
(R).
First, we show that if the kernel is square-integrable, then the corresponding
integral operator is a bounded mapping of L
2
(R) into itself. The Hilbert
Schmidt operators on L
2
(R) are precisely those operators that can be written
as integral operators with kernels k L
2
(R
2
), see Theorem C.79.
Theorem C.19 (HilbertSchmidt Integral Operators). If k L
2
(R
2
),
then the integral operator L
k
given by (C.4) denes a bounded mapping of
L
2
(R) into itself, and has operator norm |L
k
| |k|
2
.
Proof. Suppose that k L
2
(R
2
), and dene k
x
(y) = k(x, y). Then by Fubinis
Theorem k
x
L
2
(R) for a.e. x. Hence, if f L
2
(R) then
L
k
f(x) = k
x
,

f ) =
_
k
x
(y) f(y) dy
exists for almost every x.
To see why L
k
f is a measurable and square-integrable function of x, con-
sider rst the case where f and k are both nonnegative. Then k(x, y) f(y) is
a measurable function on R
2
, so Tonellis Theorem tells us that L
k
f(x) =
_
k(x, y) f(y) dy is a measurable function of x. We estimate its L
2
-norm by
applying the CauchyBunyakowskiSchwarz Inequality:
|L
k
f|
2
2
=
_
[L
k
f(x)[
2
dx
=
_
_
k(x, y) f(y) dy
2
dx
_ __
[k(x, y)[
2
dy
___
[f(y)[
2
dy
_
dx
=
_ _
[k(x, y)[
2
dy |f|
2
2
dx
= |k|
2
2
|f|
2
2
< .
Hence L
k
f L
2
(R).
Now suppose that f L
2
(R) and k L
2
(R
2
) are arbitrary, and write
f = (f
+
1
f
1
)+i(f
+
2
f
2
) and k = (k
+
1
k
1
)+i(k
+
2
k
2
), where each function
f
and k
j
is nonnegative. Then, by the work above, each function L
k
j
(f
)
is measurable and belongs to L
2
(R). Since L
k
f is a nite linear combination
of the sixteen functions L
k
j
(f
), we conclude that L
k
f is measurable and
belongs to L
2
(R).
Now that we know that L
k
f is measurable, we can follow exactly the
same estimates as were used in the nonnegative case to show that |L
k
f|
2

|k|
2
|f|
2
. Hence L
k
is a bounded mapping of L
2
(R) into itself, with operator
norm |L
k
| |k|
2
.
The next result, originally formulated in [Sch11], is often called Schurs
Test (not to be confused with Schurs Lemma). Here we formulate Schurs
Test for boundedness of integral operators, but it is instructive to compare
this result to Problem C.7, which essentially is Schurs Test for nite matrices.
Theorem C.20 (Schurs Test). Assume that k is a measurable function
on R
2
that satises the mixed-norm conditions
C
1
= ess sup
xR
_
[k(x, y)[ dy < ,
C
2
= ess sup
yR
_
[k(x, y)[ dx < .
(C.5)
Then the integral operator L
k
given by (C.4) denes a bounded mapping of
L
2
(R) into itself, with operator norm |L
k
| (C
1
C
2
)
1/2
.
Proof. As in the proof of Theorem C.19, measurability of L
k
f is most easily
shown by rst considering nonnegative f, k, and then extending to the general
case. We omit the details and assume that L
k
f is measurable for all f L
2
(R).
Then, by applying the CauchyBunyakowskiSchwarz Inequality, we have
|L
k
f|
2
2
=
_
[L
k
f(x)[
2
dx
=
_
_
k(x, y) f(y) dy
2
dx
_ __
[k(x, y)[
1/2
[k(x, y)[
1/2
[f(y)[ dy
_
2
dx
_ __
[k(x, y)[ dy
___
[k(x, y)[ [f(y)[
2
dy
_
dx
_
C
1
_
[k(x, y)[ [f(y)[
2
dy dx
= C
1
_
[f(y)[
2
_
[k(x, y)[ dxdy
C
1
_
[f(y)[
2
C
2
dy = C
1
C
2
|f|
2
2
,
where we have used Tonellis Theorem to interchange the order of integration.
Thus L
k
is bounded and |L
k
| (C
1
C
2
)
1/2
.
The next exercise shows that the hypotheses of Schurs Test actually yield
boundedness on every L
p
(R), not just for p = 2.
Exercise C.21. Show that if k satises the conditions (C.5), then L
k
is a
bounded mapping of L
p
(R) into itself for every 1 p , and satises
|L
k
| C
1/p
1
C
1/p
2
.
Remark C.22. If we assume only that k is measurable and C
2
< (with no
hypothesis about C
1
), then we have that L
k
: L
1
(R) L
1
(R) is a bounded
mapping. Similarly, if C
1
< then L
k
: L
(R) L
(R) is bounded. Fur-

ther, the proofs of these two particular endpoint cases are quite simple.
Exercise C.21 says that if C
1
and C
2
are both nite, then not only do we have
boundedness for the straightforward endpoint cases L
1
(R) and L
(R), but
we can also prove the more dicult result of boundedness on L
p
(R) for each
1 p . This type of extension problem is very common, and indeed there
is an entire theory of interpolation theorems that deal with similar exten-
sion issues, see [BeL76]. One basic interpolation theorem is the RieszThorin
Theorem, which is proved in Theorem F.11.
C.2.4 Convolution
Convolution is a special type of an integral operator. In particular, the convo-
lution of f and g is (f g)(x) =
_
f(y) g(x y) dy, whenever this is dened.
With g xed, the mapping f f g is the integral operator L
k
whose kernel
is k(x, y) = g(x y).
Exercise C.23. Use Schurs Test to prove the following version of Youngs
Inequality: If 1 p , then
f L
p
(R), g L
1
(R), |f g|
p
|f|
p
|g|
1
.
As a consequence, L
1
(R) is closed under convolution and is an example of
a Banach algebra (see Denition C.28).
Additional Problems
C.9. Choose
, and set = inf

n
[
n
[. Dene M
as in Exercise C.14,
and prove the following.
(a) Each
n
is an eigenvalue for M
with corresponding eigenvector e

n
.
(b) M
is injective if and only if

n
,= 0 for every n.
(c) M
is surjective if and only if > 0.

(d) If = 0 but
n
,= 0 for every n then range(M
) is a dense but proper

subspace of H.
(e) M
is unitary if and only if [

n
[ = 1 for every n.
C.10. Let L
(R) be xed, and let M
be dened as in Exercise C.15.

Fix 1 p .
(a) (a) Determine a necessary and sucient condition on that implies that
M
: L
p
(R) L
p
(R) is injective.
(b) Determine a necessary and sucient condition on that implies that
M
: L
p
(R) L
p
(R) is surjective.
(c) Show directly that if M
is injective but not surjective then the inverse

mapping M
1
: range(M
) L
p
(R) is unbounded.
C.11. Prove the following weighted version of Schurs Test. Assume that k is
a measurable function on R
2
, and there are measurable functions u, v 0
on R such that
_
[k(x, y)[ v(y) dy C
1
u(x), a.e. x,
_
[k(x, y)[ u(x) dx C
2
v(y), a.e. y.
Show that the integral operator L
k
with kernel k denes a bounded mapping
of L
2
(R) into itself.
C.3 The Space B(X, Y )
Now we turn our attention to the space B(X, Y ) of all bounded linear maps
from X into Y, which was introduced in Denition C.2.
Exercise C.24. Let X and Y be normed spaces.
(a) B(X, Y ) is a vector space, and operator norm is a norm on B(X, Y ).
(b) If Y is a Banach space, then B(X, Y ) is a Banach space with respect to
operator norm.
Consequently, if X is a normed space, then its dual space X
= B(X, C)
is a Banach space.
In addition to operations of vector addition and scalar multiplication, there
is a third operation that we can perform with operators, namely composition.
Exercise C.25. Prove that the operator norm is submultiplicative, i.e., if A
B(X, Y ) and B B(Y, Z), then BA B(X, Z), and
|BA| |B| |A|. (C.6)
In particular, B(X) is closed under compositions, and is an example of a
noncommutative Banach algebra (see Denition C.28).
The following exercise shows that a bounded operator that is dened on a
dense subspace of a normed space can be extended to the entire space.
Exercise C.26 (Extension of Bounded Operators). Let Y be a dense
subspace of a normed space X, and let Z be a Banach space. Let L B(Y, Z)
be given.
(a) Show that there exists a unique operator

L B(X, Z) whose restriction
to Y is L. Prove that |
L| = |L|.
(b) Show that if L: Y range(L) is a topological isomorphism, then
L: X range(L) is a topological isomorphism.

C.4 Banach Algebras 93
C.4 Banach Algebras
We have seen some examples of Banach spaces that, in addition to be-
ing complete normed vector spaces, are also are closed under an additional
multiplication-like operation. These are examples of Banach algebras, the
precise denition of which is as follows.
Denition C.27 (Algebra). An algebra over a eld K is a vector space /
over K such that for each x, y / there exists a unique product xy / that
satises the following for all x, y, z / and K:
(a) (xy)z = x(yz),
(b) x(y +z) = xy +xz and (x +y)z = xz +yz, and
(c) (xy) = (x)y = x(y).
If K = R then / is a real algebra; if K = C then / is a complex algebra.
If xy = yx for all x, y / then / is commutative.
If there exists an element e / such that ex = xe = x for every x /
then / is an algebra with identity.
Denition C.28 (Banach Algebra). A normed algebra is a normed linear
space / that is an algebra and also satises
x, y /, |xy| |x| |y|.
A Banach algebra is a normed algebra that is a Banach space.
Here are the examples of Banach algebras that we have seen so far in this
appendix, plus some other examples.
Exercise C.29. (a) L
1
(R) is a commutative Banach algebra under convolu-
tion. However, it does not have an identity.
(b) If X is a Banach space, then B(X) is a noncommutative Banach algebra
with identity under composition of operators.
(c) C
b
(R) is a commutative Banach algebra with identity under the oper-
ation of pointwise products of functions, i.e., (fg)(x) = f(x)g(x).
(d) C
0
(R) is a commutative Banach algebra without identity under the
operation of pointwise products of functions.
As in abstract ring theory, the concept of an ideal plays an important role
in the theory of Banach algebras. Ideals are the black holes of the algebra,
sucking any product of an algebra element with an ideal element into the
ideal.
Denition C.30 (Ideals). Let / be a Banach algebra.
(a) A subspace I of / is a left ideal in / if xy I whenever x / and y I.
(b) A subspace I of / is a right ideal in / if yx I whenever x / and
y I.
(c) A subspace I of / is a two-sided ideal, or simply an ideal in / if xy, yx I
whenever x / and y I.
For example, the space C
c
(R) is an ideal in C
0
(R) under the operation
of pointwise multiplication of functions. By Exercise A.63, we also know that
C
c
(R) is a dense subspace of C
0
(R). However, not all ideals are dense sub-
spaces. For example, if E R, then I = f C
0
(R) : f(x) = 0 for all x E
is a proper, closed ideal in C
0
(R).
Exercise C.31. Let I be an ideal in a commutative Banach algebra /.
(a) Prove that if x /, then x/ = xy : y / is an ideal in /, called the
ideal generated by x.
(b) Give a specic example that shows that x need not belong to x/.
(c) Show that if I is an ideal in /, then so is its closure I.
Some Banach algebras also have an additional operation that has proper-
ties similar to that of conjugation of complex numbers.
Denition C.32 (Involution). An involution on a Banach algebra / is a
mapping x x
of / into itself that satises the following for all x, y /

and all scalars C:
(a) (x
= x,
(b) (xy)
= y
,
(c) (x +y)
= x
+y
, and
(d) (x)
= x
.
Exercise C.33. Given f L
1
(R), dene

f(x) = f(x). Show that f

f
denes an involution on L
1
(R) with respect to convolution.
Another example of an involution is the adjoint operation on B(H), see
Section C.6 below.
C.5 Some Dual Spaces
In this section we consider the dual space of a Hilbert space and the dual
space of the Lebesgue space L
p
(E).
C.5 Some Dual Spaces 95
C.5.1 The Dual of a Hilbert Space
If H is a Hilbert space and g H is xed, then the CauchyBunyakowski
Schwarz Inequality implies that
g
: H C given by
g
: f f, g) is
a bounded linear functional on H. The Riesz Representation Theorem for
Hilbert spaces asserts that every bounded linear functional has this form.
Consequently, every Hilbert space is self-dual.
Exercise C.34 (Riesz Representation Theorem). Given g H, dene
g
: H C by
g
: f f, g).
(a) Show that
g
H
for each g H, and that

|g| = |
g
| = sup
f=1
[f, g)[.
(b) For each H
, show there exists a unique g H such that =

g
.
(c) Dene T : H H
by T(g) =
g
. Prove that T is an antilinear isometric
bijection of H onto H
. In particular,
g+h
=
g
+

h
.
For the specic case of
2
or L
2
(E), the Riesz Representation Theorem
takes the following form.
Corollary C.35. (a) If is a bounded linear functional on
2
(I), then there
exists a unique y = (y
k
)
kI

2
(I) such that
: x
kI
x
k
y
k
= x, y), x = (x
k
)
kI

2
(I). (C.7)
(b) If is a bounded linear functional on L
2
(E), then exists a unique
g L
2
(E) such that
: f
_
E
f(x) g(x) dx = f, g), f L
2
(E). (C.8)
We usually identify the functional H
with the element g H that

satises =
g
. However, it is important to note that this identication g
g
is antilinear. The examples given in equations (C.7) and (C.8) illustrate
that this antilinearity is a natural consequence of the denition of the inner
product. For this reason, it is most convenient for us to consider the pairing
of a vector f in a normed space X with a linear functional on X to be a
generalization of the inner product on a Hilbert space, i.e., it is a sesquilinear
form that is linear as a function of f but antilinear as a function of . We
therefore adopt the following notations for denoting the action of a linear
functional on a vector.
Notation C.36 (Notation for Linear Functionals). Let X be a normed
linear space. Given a xed linear functional : X C, we use two notations
to denote the image of f under .
(a) We write
(f)
to denote the image of f under , with the understanding that this nota-
tion is linear in both f and , i.e.,
(f +g) = (f) +(g)
and
( +)(f) = (f) +(f).
(b) We write
f, )
to denote the image of f under , with the understanding that this nota-
tion is linear in f but antilinear in , i.e.,
f +g, ) = f, ) +g, )
while
f, +) = f, ) +

f, ). (C.9)
This is our preferred notation for linear functionals.
C.5.2 The Dual of L
p
(E)
The fact that the dual space of the Hilbert space L
2
(E) is (antilinearly) iso-
morphic to L
2
(E) has a generalization to other L
p
spaces. By H olders Inequal-
ity, if g L
p
(E) is xed, then f,

g
) =
_
E
f(x) g(x) dx denes a bounded
linear functional
g
on L
p
(E), and the following exercise shows that the op-
erator norm of
g
equals the L
p
-norm of the function g.

Exercise C.37. Let E be a Lebesgue measurable subset of R, and x 1
p . For each g L
p
(E), dene
g
: L
p
(E) C by
f,
g
) =
_
E
f(x) g(x) dx, f L
p
(E). (C.10)
Show that
g
L
p
(E)
and |
g
| = |g|
p
.
Although we will not prove it, the next theorem states that if 1 p <
then every bounded linear functional on L
p
(E) has the form
g
for some
g L
p
(E). Consequently, L
p
(E)
and L
p
(E) are (antilinearly) isomorphic.

The standard proof of Theorem C.38 relies on the RadonNikodym Theorem
(see Theorem D.54).
C.5 Some Dual Spaces 97
Theorem C.38 (Dual Space of L
p
(E)). Let E be a Lebesgue measurable
subset of R, and x 1 p < . For each g L
p
(E), dene
g
as in equa-
tion (C.10). Then the mapping T : L
p
(E) L
p
(E)
dened by T(g) =
g
is
an antilinear isometric isomorphism of L
p
(E) onto L
p
(E)
.
Remark C.39. Theorem C.38 generalizes to L
p
(X) for any measure space
(X, , ) when 1 < p < . It also generalizes to L
1
(X) if is -nite,
see [Fol99] for details. In particular, an analogue of Theorem C.38 holds for
the
p
spaces.
If p = , then the map T : L
1
(E) L
(E)
given by T(g) =
g
is
still an antilinear isometry, but it is not surjective. In this sense, L
1
(E) has
a canonical image within L
(E)
, but there are functionals in L
(E)
that
do not correspond to elements of L
1
(E), compare Problem E.8.
Because of Theorem C.38, we usually identify L
p
(E) with L
p
(E)
when p
is nite, and also identify L
1
(E) with its image in L
(E)
. Abusing notation,
we write
L
p
(E)
= L
p
(E) for 1 p < ,

and
L
1
(E) L
(E)
,
with the understanding that these hold in the sense of the identication of
g L
p
(E) with
g
L
p
(E)
.
C.5.3 The Relation between L
p
(E) and L
p
(E)
We have chosen to consider the relation between L

p
(E) and L
p
(E)
in a
manner that most directly generalizes the inner product on a Hilbert space
and the characterization of the dual space of a Hilbert space as given by
the Riesz Representation Theorem. With our choice, we write the action of
L
p
(E)
on f L
p
(E) as f, ), and regard this as a sesquilinear form,
linear in f but antilinear in . With this notation, the following statements
hold (we restrict our attention in this discussion to 1 p < ).
(a) L
p
(E), L
p
(E), and L
p
(E)
are linear spaces.

(b) L
p
(E)
is the space of bounded linear functionals on L

p
(E).
(c) T : L
p
(E) L
p
(E)
given by T(g) =
g
is an isometric isomorphism, but
is antilinear.
To illustrate one advantage of this approach, consider the special case p = 2.
Since L
2
(E) is both a Hilbert space and a particular L
p
space, we have in-
troduced two dierent uses of the notation , ) with regard to L
2
(E). On
the one hand, f, g) denotes the inner product of f, g L
2
(E), while, on the
other hand, f, ) denotes the action of L
2
(E)
on f L
2
(E). Fortu-
nately, f, g) = f,
g
), so our linear functional notation is not in conict with
our inner product notation. This notationally simplies certain calculations.
For example, if A: L
2
(E) L
2
(E) is unitary then we have for f, g L
2
(E)
that f, g) = Af, Ag), and also f,
g
) = Af,
Ag
). However, we do have to
accept that our identication of g with
g
is antilinear rather than linear.
There are various alternative approaches, each with their own advantages
and disadvantages. In particular, a second choice is to base our notation on
the usual convention that if is a linear functional, then the notation (f)
is linear in both f and . If we follow this convention, then we associate a
function g L
p
(E) with the functional

g
: L
p
(E) C dened by
g
(f) =
_
E
f(x) g(x) dx.
Using this notation, the following facts hold.
(a) L
p
(E), L
p
(E), and L
p
(E)
are linear spaces.

(b) L
p
(E)
is the space of bounded linear functionals on L

p
(E).
(c) U : L
p
(E) L
p
(E)
given by U(g) =
g
is an isometric isomorphism,
and is linear.
This is a natural choice except for the fact that the notation (f) is not
an extension of the inner product on L
2
(E). Specically, although we iden-
tify g L
2
(E) with
g
L
2
(E)
, the inner product f, g) does not coincide

with
g
(f). Hence for p = 2, if A: L
2
(E) L
2
(E) is unitary, then while we
do have f, g) = Af, Ag), we do not have equality of
g
(f) and
Ag
(Af).
Another consequence is that if L: L
2
(E) L
2
(E) is linear, then the ad-
joint L
of L dened by the requirement that Lf, g) = f, L
g) is dierent
than the adjoint dened by the requirement that
g
(Lf) =
L
g
(f) (adjoints
are considered in Section C.6). Essentially, we end up with dierent notions
for concepts on L
2
(E) depending on whether we regard L
2
(E) as a Hilbert
space under the inner product, or a member of the class of Banach spaces
L
p
(E) with the identication between L
p
(E)
and L
p
(E) given by U. The

isomorphism U : L
2
(E) L
2
(E)
is dierent from the one given by the Riesz

Representation Theorem (Exercise C.34).
A third possibility is to let the functionals on L
p
(E) be antilinear in-
stead of linear. For example, we can associate g L
p
(E) with the functional
g
: L
p
(E) C given by
[f,
g
] =
_
E
f(x) g(x) dx.
Then the dual space is a space of bounded antilinear functionals:
L
p
(E)
= : L
p
(E) C : is bounded and antilinear.
In this case, we have the following facts.
(a) L
p
(E), L
p
(E), and L
p
(E)
are linear spaces.

C.6 Adjoints for Operators on Hilbert Spaces 99
(b) L
p
(E)
is the linear space whose elements are the bounded antilinear

functionals on L
p
(E).
(c) V : L
p
(E) L
p
(E)
given by V (g) =
g
is an isometric isomorphism,
and is linear.
While V is linear, we again have a disagreement between the notation [, ]
and the inner product on L
2
(E).
Despite the fact that our discussion of notation has been quite lengthy,
in the end the dierence between these choices comes down to nothing more
than convenience each choice makes certain formulas pretty and others
unpleasant. Our choice is motivated by the formulas of harmonic analysis,
and in particular the Parseval formula for the Fourier transform. We choose a
notation that directly generalizes the inner product, and consequently obtain
the simplest notational representation for generalizing the Fourier transform
to distributions and measures.
Additional Problems
C.12. Let E R be measurable. Given 1 p and a measurable weight
function w: E (0, ), the weighted L
p
space L
p
w
(R) consists of all measur-
able functions f : E C for which |f|
p,w
= |fw|
p
< . Prove that L
p
w
(R)
is a Banach space with respect to the norm | |
p,w
. Assuming Theorem C.38,
prove that if 1 p < then L
p
w
(E)

= L
p
1/w
(E).
C.6 Adjoints for Operators on Hilbert Spaces
If A is an mn complex matrix and x y is the ordinary dot product on C
d
,
then
Ax y = x A
y, x C
n
, y C
m
,
where A
= A
T
= A
H
is the Hermitian, or conjugate transpose, of A. As an
application of the Riesz Representation Theorem, we will show that there is
an analogue of the Hermitian matrix for linear operators on Hilbert spaces.
In Section C.10 we will see that this extends further to operators on Banach
spaces, but in that setting we need to appeal to the HahnBanach Theorem
in order to construct the adjoint.
Throughout this section, we will let H and K denote Hilbert spaces.
C.6.1 Adjoints of Bounded Operators
Exercise C.40. Let L: H K be a bounded linear operator.
(a) For each g K, dene a functional
g
: H C by
f,
g
) = Lf, g), f H.
Note that, following Notation C.36, f,
g
) denotes the action of the func-
tional
g
on the vector f, while Lf, g) denotes the inner product of the
vectors Lf, g K. Show that
g
H
. Consequently, by the Riesz Rep-

resentation Theorem, there exists a unique element g
H such that
f,
g
) = f, g
), f H.
(b) Dene L
: K H by L
g = g
. Show that L
is a bounded linear map,

(L
= L, and |L
| = |L|.
We formalize this as a denition.
Denition C.41 (Adjoint). The adjoint of L B(H, K) is the unique op-
erator L
: K H that satises
Lf, g) = f, L
g), f H, g K.
When H = K, we use the following additional terminology.
Denition C.42 (Self-Adjoint and Normal Operators). Let L B(H)
be given.
(a) We say that L is self-adjoint or Hermitian if L = L
. Equivalently, L is
self-adjoint if
f, g H, Lf, g) = f, Lg).
(b) We say that L is normal if L commutes with its adjoint, i.e., if LL
=
L
L.
All self-adjoint operators are normal, but not all normal operators are
self-adjoint (Problem C.20).
Remark C.43. Each complex m n matrix A determines a linear map of C
n
to C
m
. The adjoint A
of this map is determined by the conjugate transpose

of the matrix A. That is, identifying matrices with maps, we have A
= A
T
=
A
H
. This conjugate transpose matrix A
H
is called the Hermitian of A. For
matrices it is customary to instead say that A is Hermitian if A = A
H
, instead
of saying that A is self-adjoint.
If L B(H, K), then L
is a linear map of H into K, but the mapping

of B(H, K) to B(K, H) given by L L
is antilinear, as (A + B)
=
A
+

B
. Exercise C.40 shows that |L
| = |L|, and hence the map L L
is an isometry.
We will need some facts about the relationship between invariant subspaces
and adjoints.
Denition C.44. Let A B(H) be given. A closed subspace M H is said
to be invariant under A if A(M) M, where A(M) = Af : f M.
Note that we do not require that A(M) equal M, it could be a proper
subspace of M.
The simplest example of an invariant subspace is M = spanf where f
is an eigenvector of A.
Exercise C.45. Show that if a closed subspace M H is invariant under
A B(H), then M
is invariant under A
.
C.6.2 Adjoints of Unbounded Operators
Adjoints can also be dened for unbounded operators, although now we must
be careful with domains. For example, consider the dierentiation operator
Df = f
. This operator is not dened on all of L

2
(R), but instead is densely
dened in the sense of Notation C.3. For example, D maps the dense subspace
S =
_
f C
1
0
(R) : f, f
L
2
(R)
_
into L
2
(R). Although D: S L
2
(R) is
unbounded, if f, g S then we have fg C
0
(R), so integration by parts
yields
Df, g) =
_
f
(x) g(x) dx =
_
f(x) g
(x) dx = f, Dg).
Hence D
= D as an operator mapping S into L

2
(R).
The important point in the preceding example is that if g S is xed,
then f Df, g) is a bounded linear functional on S, even though D is an
unbounded operator. The following exercise extends this to general operators.
Exercise C.46. Suppose that S is a dense subspace of H, and L: S K is
a linear, but not necessarily bounded, operator. Let
S
= g K : f Lf, g) is a bounded linear functional on S.

Show that there is an operator L
: S
H such that
Lf, g) = f, L
g), f S, g S
.
If necessary, we can always restrict a densely dened operator to a smaller
but still dense domain. Given an operator L mapping some dense subspace
of H into H, if we can nd some dense subspace S on which L is dened and
such that Lf, g) = f, Lg) for f, g S, then we say that L is self-adjoint.
C.6.3 Bounded Self-Adjoint Operators on Hilbert Spaces
We now focus in more detail on bounded self-adjoint operators on Hilbert
spaces, which have many useful properties. Some of the results that we de-
scribe next hold only for complex Hilbert spaces; we will indicate the changes
needed for real Hilbert spaces. Recall that, following the conventions laid out
in the General Notation section, we always assume that a vector space is over
the complex eld unless specically stated otherwise.
For complex Hilbert spaces, we have the following characterization of self-
adjoint operators. This theorem does not have an analogue for real Hilbert
spaces.
Theorem C.47. Let H be a (complex) Hilbert space, and let A B(H) be
given. Then:
A is self-adjoint Af, f) R for all f H.
Proof. . Assume that Af, f) is real for all f. Choose any f, g H. Then
A(f +g), f +g) = Af, f) +Af, g) +Ag, f) +Ag, g).
Since A(f + g), f + g), Af, f), and Ag, g) are all real, we conclude that
Af, g) +Ag, f) is real. Hence it equals its own complex conjugate:
Af, g) +Ag, f) = Af, g) +Ag, f) = g, Af) +f, Ag). (C.11)
Similarly, after examining the equation
A(f +ig), f +ig) = Af, f) iAf, g) +iAg, f) +Ag, g),
we conclude that
Af, g) Ag, f) = g, Af) +f, Ag). (C.12)
Adding (C.11) and (C.12) together, we obtain
2Af, g) = 2f, Ag) = 2A
f, g).
Since this is true for every f and g, we conclude that A = A
.
The next result gives an alternative formula for the operator norm of a
self-adjoint operator. This theorem holds for both real and complex Hilbert
spaces.
Theorem C.48. If A B(H) is self-adjoint, then
|A| = sup
f=1
[Af, f)[.
Proof. Set M = sup
f=1
[Af, f)[. By CauchyBunyakowskiSchwarz and
the denition of operator norm, it follows that M |A|.
Choose any unit vectors f, g H. Then, by expanding the inner products,
canceling terms, and using the fact that A = A
, we see that
A(f +g), f +g
_
A(f g), f g
_
= 2 Af, g) + 2 Ag, f)
= 2 Af, g) + 2 g, Af)
= 4 ReAf, g).
Therefore, applying the denition of M and using the Parallelogram Law, we
have
4 ReAf, g) [A(f +g), f +g)[ +[A(f g), f g)[
M |f +g|
2
+M|f g|
2
= 2M
_
|f|
2
+|g|
2
_
= 4M.
That is, ReAf, g) M for every choice of unit vectors f and g. Write
[Af, g)[ = Af, g) where C satises [[ = 1. Then g is another unit
vector, so
[Af, g)[ = Af, g) = Af, g) M.
Hence
|Af| = sup
g=1
[Af, g)[ M,
and therefore |A| M.
As a corollary, we obtain the following useful fact for operators on a com-
plex Hilbert space. This result also holds for operators on a real Hilbert space
if we include the assumption A = A
as a hypothesis.
Exercise C.49. Let H be a (complex) Hilbert space. If A B(H) and
Af, f) = 0 for every f, then A = 0.
C.6.4 Positive and Positive Denite Operators on Hilbert Spaces
Among the self-adjoint operators, we distinguish the special class of positive
operators.
Denition C.50 (Positive and Positive Denite Operators). Let A
B(H) be given.
(a) We say that A is positive or nonnegative, denoted A 0, if Af, f) 0
for every f H.
(b) We say that A is positive denite or strictly positive, denoted A > 0, if
Af, f) > 0 for every nonzero vector f H.
By Theorem C.47, since we are dealing with complex Hilbert spaces, all
positive and positive denite operators are self-adjoint. When dealing with
real Hilbert spaces, the assumption of self-adjointness should be added in the
denition of positive and positive denite operators.
Exercise C.51. Show that if A B(H, K), then A
A B(H) and AA
B(K) are both positive operators. Determine conditions on A that imply that
A
A or AA
is positive denite.
Additional Problems
C.13. Let H
1
, H
2
, H
3
be Hilbert spaces. Show that if A B(H
1
, H
2
) and
B B(H
2
, H
3
), then (A
= A and (BA)
= A
.
C.14. Show that if A B(H, K) is a topological isomorphism then A
B(K, H) is also a topological isomorphism, and (A

1
)
= (A
)
1
.
C.15. Show that if A B(H, K), then |A| = |A
| = |A
A|
1/2
= |AA
|
1/2
.
C.16. Show that A A
denes an involution on B(H) in the sense of

Denition C.32.
Remark: In the language of operator theory, the fact that this involution
satises |A
A| = |A|
2
means that B(H) is an example of a C
-algebra.
C.17. Given A B(H), show that A is normal if and only if |Af| = |A
f|
for every f H.
C.18. Show that if A B(H, K) then the following statements hold.
(a) ker(A) = range(A
.
(b) ker(A)
= range(A
).
(c) A is injective if and only if range(A
) is dense in H.
C.19. Given A B(H), show that
ker(A) = ker(A
A) and range(A
A) = range(A
).
Use this to show that if A is normal, then
ker(A) = ker(A
) and range(A) = range(A
).
C.20. Fix
, and let M
be the multiplication operator dened in

Exercise C.14. Find M
, and show that M
is normal. Determine when M
is self-adjoint, positive, or positive denite.

C.21. Fix L
(R), and let M

Exercise C.15. Find M
, and show that M
is normal. Determine when M
is self-adjoint, positive, or positive denite.

C.22. Given k L
2
(R
2
), let L
k
be the integral operator with kernel k. Show
that the adjoint operator (L
k
)
is the integral operator L

k
whose kernel is
k
(x, y) = k(y, x). Characterize those kernels k L

2
(R
2
) corresponding to
self-adjoint operators L
k
.
C.23. Let L and R be the left- and right-shift operators from Problem C.2.
Show that R = L
, and conclude that L and R are not normal.

C.7 Compact Operators on Hilbert Spaces 105
C.24. Let e
n
nN
be an orthonormal basis for a separable Hilbert space H.
Let T : H
2
(N) be the analysis operator Tf = f, e
n
)
nN
. Find a formula
for the synthesis operator T
:
2
(N) H.
C.25. Let M be a closed subspace of H, and let P B(H) be given. Show
that P is the orthogonal projection of H onto M if and only if P
2
= P,
P
= P, and range(P) = M.
C.26. Suppose that A, B B(H) are self-adjoint. Show that ABA and BAB
are self-adjoint, but AB is self-adjoint if and only if AB = BA. Exhibit self-
adjoint operators A, B that do not commute.
C.27. Let L: H H be a self-adjoint operator on H, either bounded or
unbounded and densely dened. Show that all eigenvalues of L are real, and
that eigenvectors of L corresponding to distinct eigenvalues are orthogonal.
C.28. Let L B(H) be normal. Show that if is an eigenvalue of L, then
is an eigenvalue of L
, and the -eigenspace of L equals the

-eigenspace
of L
. Also show that eigenvectors of L corresponding to distinct eigenvalues

are orthogonal.
C.29. Let A B(H) be given. Show that if A B(H) is a positive oper-
ator, then all eigenvalues of A are real and nonnegative, and if A B(H)
is a positive denite operator, then all eigenvalues of A are real and strictly
positive.
C.7 Compact Operators on Hilbert Spaces
In many senses, compact operators on a Hilbert space are the ones that are
most similar to linear operators on nite-dimensional spaces.
Throughout this section, H and K will denote Hilbert spaces.
C.7.1 Denition and Basic Properties
The closed unit ball makes frequent appearances in any discussion of compact
operators, so we introduce a notation for it.
Notation C.52. Ball
H
will denote the closed unit ball in a Hilbert space H:
Ball
H
=
_
f H : |f| 1
_
.
By Problem A.26, Ball
H
is compact if and only if H is nite-dimensional.
If H and K are nite-dimensional and T : H K is linear, then T will
map the closed unit ball in H to a closed ellipsoid in K, which is necessarily
compact. Even if K is innite-dimensional, if H is nite-dimensional then
T(Ball
H
) will be a nite-dimensional ellipsoid in K, and hence will still be
compact. However, this need not be the case if H is innite-dimensional. For
example, if H is innite-dimensional and I : H H is the identity map, then
I(Ball
H
) = Ball
H
is not compact. In general, even if T : H K is bounded
and linear, T(Ball
H
) need not be closed in K (see Problem C.33).
Denition C.53 (Compact Operators). A linear operator T : H K is
compact if T(Ball
H
) has compact closure in K. We dene
B
0
(H, K) =
_
T : H K : T is compact
_
,
and set B
0
(H) = B
0
(H, H).
The next exercise gives some useful reformulations of the denition of
compact operator.
Exercise C.54. Given a linear operator T : H K, show that the following
(a) T is compact.
(b) T(Ball
H
) is totally bounded.
(c) If f
n
nN
is a bounded sequence in H, then Tf
n
nN
contains a con-
vergent subsequence in K.
We show now that all compact operators are bounded.
Theorem C.55. B
0
(H, K) B(H, K).
Proof. Assume that T : H K is linear but unbounded. Then there exist
vectors f
n
H such that |f
n
| = 1 but |Tf
n
| n. Therefore every subse-
quence of Tf
n
nN
is unbounded, and hence cannot converge. Exercise C.54
therefore implies that T is not compact.
Now we consider the structure of the space of compact operators.
Theorem C.56 (Limits of Compact Operators). B
0
(H, K) is a closed
subspace of B(H, K). Specically, if T
n
: H K are compact, T : H K is
linear and bounded, and |T T
n
| 0, then T is compact.
Proof. Exercise: Show that B
0
(H, K) is a subspace of B(H, K).
We must show that B
0
(H, K) is closed. Assume that T
n
are compact op-
erators and that T
n
T in operator norm. By Exercise C.54, to show that T
is compact it suces to show that T(Ball
H
) is a totally bounded subset of K.
Choose any > 0. Then there exists an n such that |T T
n
| < . Since
T
n
is compact, we know that T
n
(Ball
H
) is totally bounded. This implies that
there exist nitely many points h
1
, . . . , h
m
Ball
H
such that
T
n
(Ball
H
)
m
j=1
B
(T
n
h
j
). (C.13)
We will show that T(Ball
H
) is totally bounded by showing that
T(Ball
H
)
m
j=1
B
3
(Th
j
). (C.14)
Choose any element of T(Ball
H
), i.e., any vector Tf with |f| 1. Then
T
n
f T
n
(Ball
H
), so by equation (C.13) there must be some j such that
|T
n
f T
n
h
j
| < . Consequently,
|Tf Th
j
| |Tf T
n
f| +|T
n
f T
n
h
j
| +|T
n
h
j
Th
j
|
< |T T
n
| |f| + +|T
n
T| |h
j
|
< 1 + + 1 = 3.
Hence equation (C.14) follows, so T is compact.
Compact operators are well-behaved with respect to compositions.
Exercise C.57. Let H
1
, H
2
, H
3
be Hilbert spaces.
(a) Show that if A: H
1
H
2
is bounded and linear and T : H
2
H
3
is
compact, then TA: H
1
H
3
is compact.
(b) Show that if T : H
1
H
2
is compact and A: H
2
H
3
is bounded and
linear, then AT : H
1
H
3
is compact.
In particular, B
0
(H) is a closed, two-sided ideal in B(H) under composition
of operators.
Note that it is entirely possible for the product of two noncompact opera-
tors to be compact. For example, let M be a closed subspace of H such that
both M and M
are innite dimensional. If P is the orthogonal projection

onto M and Q is the orthogonal projection onto M
, then neither P nor Q

is compact, but the composition PQ is the zero operator, which is compact.
Less trivial examples can be constructed by modifying this idea.
C.7.2 Finite-Rank Operators
Recall that the rank of an operator T : H K is the vector space dimension
of range(T). We say that T is a nite-rank operator if range(T) is nite-
dimensional. We set
B
00
(H, K) =
_
T B(H, K) : T is nite-rank
_
,
and let B
00
(H) = B
00
(H, H).
Problem C.8 shows that a linear, nite-rank operator need not be bounded
(that is why we include the assumption of boundedness in the denition
of B
00
(H, K) above). On the other hand, every nite-rank operator that is
bounded must be compact.
Exercise C.58. Show that if T : H K is bounded, linear, and has nite
rank, then T is compact. Thus,
B
00
(H, K) B
0
(H, K).
Combining the preceding exercise with Theorem C.56, we obtain a useful
technique for showing that a given operator is compact. Specically, if we are
given T : H K, and if we can nd bounded linear nite-rank operators T
n
such that |T T
n
| 0, then T must be compact. Moreover, the converse is
also true: Every compact operator T : H K can be realized as the operator
norm limit of a sequence of bounded nite-rank operators.
Theorem C.59. B
00
(H, K) is a dense subspace of B
0
(H, K). In particular,
if T : H K is compact, then there exist bounded linear nite-rank operators
T
n
: H K such that T
n
T in operator norm.
Proof. Assume that T is compact and set R = range(T). If R is nite-
dimensional, then T is nite-rank and we are done. So, assume that R is
innite-dimensional. Problem C.35 shows that R is separable. Therefore,
since R is closed, we can nd a countable orthonormal basis e
n
nN
for
R. For any f H we have Tf R, so
Tf =
n=1
Tf, e
n
) e
n
, f H.
Dene
T
N
f =
N
n=1
Tf, e
n
) e
n
, f H,
and note that T
N
= P
N
T where P
N
is the orthogonal projection of K onto
the closed subspace spane
1
, . . . , e
N
. By denition, T
N
converges to T in the
strong operator topology, i.e., T
N
f Tf for every f. Our goal is to show
that T
N
T in operator norm.
Choose any > 0. Since T(Ball
H
) is totally bounded, it is covered
by nitely many -balls centered at points in T(Ball
H
). Hence, there exist
h
1
, . . . , h
m
Ball
H
such that
T(Ball
H
)
m
k=1
B
(Th
k
).
Since lim
N
|Th
k
T
N
h
k
| = 0 for k = 1, . . . , m, we can nd an N
0
such
that
N > N
0
, |Th
k
T
N
h
k
| < , k = 1, . . . , m.
Choose any f with |f| = 1 and any N > N
0
. Then Tf B
(Th
k
) for
some k, i.e., |Tf Th
k
| < . Therefore
|T
N
f T
N
h
k
| = |P
N
Tf P
N
Th
k
| |P
N
| |Tf Th
k
| < 1 ,
so
|Tf T
N
f| |Tf Th
k
| +|Th
k
T
N
h
k
| +|T
N
h
k
T
N
f|
< + + = 3.
This is true for every unit vector, so |T T
N
| 3 for all N > N
0
, and thus
|T T
N
| 0.
The next exercise characterizes the nite-rank operators, and uses this
characterization to show that compactness is preserved under adjoints.
Exercise C.60. (a) Show that T B(H, K) has nite rank if and only if
there exist
1
, . . . ,
N
H and
1
, . . . ,
N
K such that
Tf =
N
k=1
f,
k
)
k
, f H. (C.15)
In case this holds, nd T
and show that T
is also nite-rank.
(b) Show that if T B(H, K), then T is compact if and only if T
is com-
pact.
Following Notation C.18, an alternative way to write the nite-rank oper-
ator given in equation (C.15) is T =
N
k=1
(
k

k
).
C.7.3 Integral Operators with Square-Integrable Kernels
The next exercise shows that integral operators that have a square-integrable
kernel (which we call the HilbertSchmidt integral operators) are compact. We
will rene this yet further in Theorem C.79 below.
Exercise C.61 (HilbertSchmidt Integral Operators). Fix k L
2
(R
2
),
and let L
k
be the integral operator with kernel k.
(a) Let e
mn
m,nN
be an orthonormal basis for L
2
(R
2
) of the form e
mn
=
e
m
e
n
, as constructed in Problem C.30. Dene
k
N
=
N
m=1
N
n=1
k, e
mn
) e
mn
,
and show that the corresponding integral operator L
kN
is bounded and
has nite-rank.
(b) Show that L
k
L
kN
is the integral operator with kernel k k
N
, and that
|L
k
L
kN
| |k k
N
|
2
. Conclude that L
k
is a compact mapping of
L
2
(R) into itself.
Additional Problems
C.30. Let e
n
nN
be any orthonormal basis for L
2
(R). Dene
e
mn
(x, y) = (e
m
e
n
)(x, y) = e
m
(x) e
n
(y), x, y R.
Show that e
mn
m,nN
is an orthonormal basis for L
2
(R
2
).
C.31. Suppose that T : H K is compact and e
n
nN
is an orthonormal
sequence in H. Show that Te
n
0.
C.32. Let H be an innite-dimensional Hilbert space. Show that if T : H K
is compact and injective, then T
1
: range(T) H is unbounded.
C.33. Fix
, and let M

Exercise C.14. Show that M
is compact if and only if c

0
(i.e.,
n
0
as n ). Show that if c
0
and
n
,= 0 for every n, then T(Ball
H
) is not
closed in H.
C.34. Fix L
(R), and let M

Exercise C.15, where we take p = 2. Show that M
is compact if and only if

= 0 a.e.
C.35. Prove that if T B
0
(H, K), then range(T) is a separable subspace
of K.
C.36. Show that if T : H H is compact and ,= 0 is an eigenvalue of T,
then the corresponding eigenspace ker(T I) is nite-dimensional.
C.8 The Spectral Theorem for Compact Self-Adjoint
Operators
Throughout this section, H will denote a Hilbert space.
The Spectral Theorem provides a fundamental decomposition for self-
adjoint operators on a Hilbert space. The version of the Spectral Theorem
that we will present asserts that if T is a compact self-adjoint operator on H,
then there is an orthonormal basis for H that consists of eigenvectors of T,
and T has a simple representation with respect to this orthonormal basis.
This is only a basic version of the Spectral Theorem. The theorem general-
izes to compact normal operators with little change, allowing eigenvalues to
be complex instead of real. Much more deeply, however, there is a Spectral
Theorem for self-adjoint operators that are not compact, and even for some
that are unbounded. We refer to texts on operator theory for such extensions,
e.g., [Con00].
C.8 The Spectral Theorem for Compact Self-Adjoint Operators 111
C.8.1 Existence of an Eigenvalue
In the nite-dimensional setting, the rst step in proving the Spectral The-
orem is to note that every n n matrix always has at least one (complex)
eigenvalue. Unfortunately, this is not true in innite dimensions, even for com-
pact operators.
Exercise C.62. (a) The Volterra operator on L
2
[0, 1] is the integral operator
V f(x) =
_
x
0
f(y) dy. Show that V is compact, not self-adjoint, and has no
eigenvalues.
(b) Show that the multiplication operator M: L
2
[0, 1] L
2
[0, 1] given by
Mf(x) = xf(x) is self-adjoint (in fact, positive denite), but is not compact
and has no eigenvalues.
On the other hand, we will show that an operator that is both compact
and self-adjoint must have an eigenvalue. In order to do this, we rst need the
following sucient condition for the existence of an eigenvalue of a compact
operator.
Lemma C.63. If T : H H is compact and ,= 0, then
inf
f=1
|Tf f| = 0 = is an eigenvalue of T.
Proof. Suppose we can nd unit vectors f
n
such that |Tf
n
f
n
| 0. Since
T is compact, Tf
n
nN
has a convergent subsequence, say Tf
n
k
g H.
Then
f
n
k
=
_
f
n
k
Tf
n
k
_
+Tf
n
k
0 +g = g.
Since the f
n
k
are unit vectors and ,= 0, we conclude that g ,= 0. Moreover,
since T is continuous it follows that Tf
n
k
Tg. But we also know that
Tf
n
k
g, so we conclude that Tg = g.
Now we show that compactness combined with self-adjointness implies the
existence of an eigenvalue. Since the eigenvalues of a self-adjoint operator are
real and are bounded in absolute value by the operator norm, we know that
any eigenvalue must lie in the range |T| |T|.
Theorem C.64. If T : H H is compact and self-adjoint, then either |T|
or |T| is an eigenvalue of T.
Proof. Since T is self-adjoint, we have |T| = sup
f=1
[Tf, f)[ by Theo-
rem C.48. Hence, there must exist unit vectors f
n
such that [Tf
n
, f
n
)[ |T|.
Since T is self-adjoint, every inner product Tf
n
, f
n
) is real, so we can nd a
subsequence Tg
n
, g
n
)
nN
that converges either to |T| or to |T|. Let
be either |T| or |T|, as appropriate. Then we have |g
n
| = 1 for every n
and Tg
n
, g
n
) . Hence, since both and Tg
n
, g
n
) are real,
|Tg
n
g
n
|
2
= |Tg
n
|
2
2Tg
n
, g
n
) +
2
|g
n
|
2
|T|
2
|g
n
|
2
2Tg
n
, g
n
) +
2
|g
n
|
2
=
2
2Tg
n
, g
n
) +
2

2
2
2
+
2
= 0.
It therefore follows from Lemma C.63 that is an eigenvalue of T.
C.8.2 The Spectral Theorem
We obtain the Spectral Theorem by iterating Theorem C.64. If T has nite
rank then the index set J in the following theorem is J = 1, . . . , N where
N = dim
_
range(T)
_
; otherwise it is J = N.
Theorem C.65 (Spectral Theorem for Compact Self-Adjoint Oper-
ators). Let T : H H be compact and self-adjoint. Then there exist:
(a) countably many nonzero real numbers
n
nJ
, either nitely many or
satisfying
n
0 if innitely many, and
(b) an orthonormal basis e
n
nJ
of range(T),
such that
Tf =
nJ
n
f, e
n
) e
n
, f H. (C.16)
Each
n
is an eigenvalue of T, and each e
n
is a corresponding eigenvector.
Proof. If T = 0 then we can take J = , so assume that T is not the zero
operator. Since T is compact, Problem C.35 implies that range(T) is separable,
and hence has either a nite or a countably innite orthonormal basis.
Dene H
1
= H and T
1
= T. By Theorem C.64, T
1
has an eigenvalue
1
that satises [
1
[ = |T
1
| > 0. Let e
1
be a corresponding eigenvector, normal-
ized so that |e
1
| = 1.
Let H
2
= e
1
and let T
2
= T[
H2
. If T
2
= 0, then we stop at this point.
Otherwise, we continue as follows. Since spane
1
is invariant under T
1
, we
know from Exercise C.45 that H
2
is invariant under T
1
= T
1
.
Exercise: Show that T
2
: H
2
H
2
is compact and self-adjoint.
Therefore T
2
has an eigenvalue
2
such that [
2
[ = |T
2
| > 0, and since T
2
is a restriction of T
1
, we have [
2
[ = |T
2
| |T
1
| = [
1
[. Let e
2
be a corre-
sponding unit eigenvector. By denition of H
2
, we have e
2
e
1
. Further,
2
is an eigenvalue of T (not just T
2
), and e
2
is the corresponding eigenvector
of T.
Now set H
3
= e
1
, e
2
and T
3
= T[
H3
. If T
3
= 0, then we stop at this
point. Otherwise, we continue as before to construct an eigenvalue
3
and unit
eigenvector e
3
(which will be orthogonal to both e
1
and e
2
). As we continue
in this process, there are two possibilities.
C.8 The Spectral Theorem for Compact Self-Adjoint Operators 113
Case 1: T
N+1
= 0 for some N. In this case, since H
N+1
= e
1
, . . . , e
N
,
we have
H = spane
1
, . . . , e
N
H
N+1
.
Therefore, each f H can be written uniquely as
f =
N
n=1
f, e
n
) e
n
+ v
f
where v
f
H
N+1
. Then since Tv
f
= T
N+1
v
f
= 0,
Tf =
N
n=1
f, e
n
) Te
n
+ Tv
f
=
N
n=1
n
f, e
n
) e
n
.
In this case T is nite-rank and has the required form.
Case 2: T
N
,= 0 for any N. In this case we obtain countably many eigenval-
ues
n
and corresponding orthonormal eigenvectors e
n
. Since T is compact,
Problem C.31 implies that Te
n
0. Since Te
n
=
n
e
n
and |e
n
| = 1, we
conclude that
n
0.
Let M = spane
n
nN
. Then e
n
nN
is an orthonormal basis for M, and
H = M M
. Hence, each f H can be written uniquely as

f =
n=1
f, e
n
) e
n
+ v
f
for some v
f
M
. Therefore
Tf =
n=1
f, e
n
) Te
n
+ Tv
f
=
n=1
n
f, e
n
) e
n
+ Tv
f
.
Note that since spane
1
, . . . , e
N
M, we have for any N > 0 that
v
f
M
spane
1
, . . . , e
N
= H
N
.
Since T[
HN
= T
N
, we therefore have
|Tv
f
| = |T
N
v
f
| |T
N
| |v
f
| = [
N
[ |v
f
| 0 as N .
Consequently Tv
f
= 0, so T has the required form.
Exercise: Complete the proof by showing that, in either case, e
n
nJ
is
an orthonormal basis for range(T).
Since the nonzero eigenvalues of a compact self-adjoint operator converge
to zero (if there are innitely many), we usually order them in decreasing
order according to absolute value:
[
1
[ [
2
[ .
The multiplicity of a given eigenvalue is the number of times it is repeated
in the list of eigenvalues given above. Equivalently, the multiplicity is the
dimension of the -eigenspace ker(T I). The multiplicity of any given
eigenvalue is nite.
The Spectral Theorem is stated above in terms of the nonzero eigenval-
ues of T. The zero eigenspace is ker(T), and the vectors in this space are
eigenvectors of T for = 0. If H is separable, then ker(T) is separable, so
by combining an orthonormal basis for ker(T) with the orthonormal basis for
range(T) given by Theorem C.65, we obtain the following reformulation of
the Spectral Theorem, which states that every compact, self-adjoint operator
on a separable Hilbert space is actually a multiplication operator M
of the
form introduced in Exercise C.14.
Exercise C.66. Let H be a separable Hilbert space. Let J = 1, . . . , dim(H)
if H is nite-dimensional, or J = N if H is innite-dimensional. Given T
B(H), show that the following statements are equivalent.
(a) T is compact and self-adjoint.
(b) There exists a real sequence =
n
nJ
, with c
0
if J = N, and an
orthonormal basis e
n
nJ
for H such that
Tf = M
f =
nJ
n
f, e
n
) e
n
, f H.
For the nite-dimensional setting, the Spectral Theorem corresponds to
the diagonalization of Hermitian matrices.
Exercise C.67. Let A be an n n complex matrix, viewed as a mapping of
C
n
into itself. Show that A is a Hermitian matrix if and only if A = UU
1
where U is unitary and is real and diagonal. Also show that, in this case,
the diagonal entries of are the eigenvalues of A, and the columns of U form
an orthonormal basis for C
n
consisting of eigenvectors of A.
As an application of the Spectral Theorem, we construct a square root of
a compact positive operator.
Exercise C.68. Assume that T : H H is both compact and positive, and
let equation (C.16) be the representation of T given by the Spectral Theorem.
Dene
T
1/2
f =
nJ
1/2
n
f, e
n
) e
n
, f H.
Show that T
1/2
is well-dened, compact, positive, and (T
1/2
)
2
= T.
More generally, it can be shown that every positive operator on a Hilbert
space space has a square root, see [Hei10] for a proof.
C.9 HilbertSchmidt Operators 115
Additional Problems
C.37. Assume T : H H is compact and ,= 0. Show that if is not an
eigenvalue of T and

is not an eigenvalue of T
, then T I is a topological
isomorphism of H onto itself.
C.38. Dene the spectral radius (T) of an mm matrix T to be
(T) = max[[ : is an eigenvalue of T.
Let A be an mn complex matrix. Show directly that if we place the
2
norm
on both C
n
and C
m
, then the operator norm of A as a mapping of C
n
into
C
m
is |A| = (A
A)
1/2
. Compare this to Problem C.7.
C.9 HilbertSchmidt Operators
The HilbertSchmidt operators are a special subclass of the compact operators
on a Hilbert space. Although the denition of a HilbertSchmidt operator
makes sense when H is nonseparable (by using a complete orthonormal system
instead of a countable orthonormal basis), we will restrict our discussion to
HilbertSchmidt operators on separable spaces. Therefore, throughout this
section, H will denote a separable Hilbert space.
C.9.1 Denition and Basic Properties
Denition C.69. An operator T B(H) is HilbertSchmidt if there exists
an orthonormal basis e
n
nN
for H such that
n=1
|Te
n
|
2
< .
The next exercise shows that the choice of orthonormal basis is irrelevant.
Exercise C.70. Fix T B(H). Given an orthonormal basis c = e
n
nN
for H, set
S(c) =
_

n=1
|Te
n
|
2
_
1/2
.
Show that S(c) is independent of the choice of orthonormal basis c. That is,
if S(c) is nite for one orthonormal basis then it is nite for all and always
takes the same value, and if S(c) is innite for one orthonormal basis then it
is innite for all.
Denition C.71. The space of HilbertSchmidt operators on H is
B
2
(H) =
_
T B(H) : T is HilbertSchmidt
_
.
The HilbertSchmidt norm of T B
2
(H) is
|T|
HS
= |T|
B2
=
_

n=1
|Te
n
|
2
_
1/2
,
where e
n
nN
is any orthonormal basis for H.
The notation B
2
(H) emphasizes that the space of HilbertSchmidt oper-
ators is the central (p = 2) space in the family of Schatten classes B
p
(H)
described below. We will use the notations | |
HS
and | |
B2
interchangeably.
Note that if we write
|T|
2
B2
=
n=1
|Te
n
|
2
=
n=1
m=1
[Te
n
, e
m
)[
2
, (C.17)
then we see that the HilbertSchmidt norm is simply the
2
-norm of the matrix
representation of T with respect to the orthonormal basis e
n
nN
.
The fact that ||
B2
is indeed a norm on B
2
(H), and some of the other basic
properties of HilbertSchmidt operators, are developed in the next exercise.
Exercise C.72. Prove the following facts.
(a) |T| |T|
B2
for all T B
2
(H).
(b) | |
B2
is a norm on B
2
(H), and B
2
(H) is complete in this norm.
(c) B
2
(H) is closed under adjoints, and |T
|
B2
= |T|
B2
for all T B
2
(H).
(d) If T B
2
(H) and A B(H), then AT, TA B
2
(H), and we have
|AT|
B2
|A| |T|
B2
, |TA|
B2
|A| |T|
B2
.
Consequently, B
2
(H) is a two-sided ideal in B(H).
(e) All nite-rank operators are HilbertSchmidt, i.e., B
00
(H) B
2
(H).
(f) All HilbertSchmidt operators are compact, i.e., B
2
(H) B
0
(H).
(g) The space B
00
(H) of nite-rank operators is dense in B
2
(H) with respect
to both the operator norm and the HilbertSchmidt norm.
C.9.2 Singular Numbers and Schatten Classes
We will give an equivalent formulation of HilbertSchmidt operators in terms
of their singular numbers. Singular numbers can be dened for any com-
pact operator T : H H. Although a compact operator need not have any
eigenvalues, the operator T
T is both compact and self-adjoint (in fact, pos-

itive). Therefore, we can apply the Spectral Theorem to T
T to deduce that
there exists an orthonormal sequence e
n
nJ
and corresponding real num-
bers
n
nJ
such that
T
Tf =
nJ
n
f, e
n
) e
n
, f H.
The scalars
n
are the nonzero eigenvalues of T
T. Since T
T is positive,
each
n
is strictly positive. There are either nitely many nonzero eigenval-
ues of T
T, or if innitely many then

n
0. The index set J is either
J = 1, . . . , N with N = dim(range(T
T)) if T
T is nite-rank (which by
Problem C.19 happens if and only if T is nite-rank), or J = N otherwise.
Denition C.73 (Singular Numbers). Let T : H H be compact, and
let
n
nJ
and e
n
nJ
be as constructed above. Then the singular numbers
or singular values of T are
s
n
(T) =
1/2
n
, n J,
taken in decreasing order:
s
1
(T) s
2
(T) > 0.
The vectors e
n
are corresponding singular vectors of T.
In particular, a bounded nite-rank operator has only nitely many
nonzero singular numbers.
The singular numbers of a self-adjoint operator can be written directly in
terms of its eigenvalues.
Exercise C.74. Show that if T is compact and self-adjoint and
n
(T)
nJ
are its nonzero eigenvalues taken in decreasing order, then s
n
(T) = [
n
(T)[.
Consequently, if T is compact and positive then s
n
(T) =
n
(T).
Exercise C.68 shows how to construct the square root of a compact positive
operator. If T is an arbitrary compact operator, then T
T is positive and
compact, and we have a special name for its square root.
Denition C.75 (Absolute Value). If T : H H is compact, then the
positive compact operator
[T[ = (T
T)
1/2
is the absolute value of T.
In this terminology, the singular numbers of a compact operator T are the
nonzero eigenvalues of [T[ = (T
T)
1/2
, so we have the relations
s
n
(T) =
n
([T[) =
n
(T
T)
1/2
.
In particular, since T
T is positive, its largest eigenvalue is its operator norm.

Therefore,
s
1
(T) =
1
(T
T)
1/2
= |T
T|
1/2
= |T|, (C.18)
the nal equality following from Problem C.15.
Remark C.76. It can be shown that every bounded operator A on H can be
written in the form A = UP where P is positive and U is a partial isometry
(meaning that |Uf| = |f| for all f (ker U)
), and this polar decomposition

is unique if we impose the condition that ker(U) = ker(P). In particular, if
A is compact then P coincides with the operator [A[ dened above. Thus,
absolute value can be dened for arbitrary bounded operators.
Now we can reformulate the denition of HilbertSchmidt operators in
terms of singular numbers.
Exercise C.77. Let T : H H be compact, and let s = s
n
(T)
nJ
be the
singular numbers of T. Show that T is HilbertSchmidt if and only if s
2
(J),
and, in this case,
|T|
B2
= |s|
2
=
_
nJ
s
n
(T)
2
_
1/2
.
More generally, the Schatten class B
p
(H) is the space of compact operators
that have
p
singular numbers.
Denition C.78 (Schatten Class). Given 1 p < , the Schatten class
B
p
(H) consists of all compact operators T : H H whose corresponding
singular numbers s
n
(T)
nJ
satisfy
|T|
Bp
=
_
nJ
s
n
(T)
p
_
1/p
< .
Note that since a bounded nite-rank operator has only nitely many
singular numbers, we have B
00
(H) B
p
(H), for each 1 p < . In fact, it
can be shown that the nite-rank operators are dense in B
p
(H) for each p.
Since the
-norm of the singular values is

|s|
= sup
nJ
s
n
(T) = s
1
(T) = |T|,
we usually dene the p = Schatten class to be B
(H) = B
0
(H), the space
of all compact operators on H.
As in the discussion leading up to Exercise C.66, it is often convenient to
extend the singular vectors so that we obtain an orthonormal basis for all of H.
The singular vectors e
n
nJ
as we have dened them form an orthonormal
basis for range(T). Since T
T is self-adjoint,
range(T)
= ker(T
T).
Hence we can form an orthonormal basis for all of H by combining e
n
nN
with an orthonormal basis f
m
mI
for ker(T
T). We can regard the vectors

f
m
as being singular vectors of T corresponding to zero singular numbers.
Each vector f H can be written
f =
nJ
f, e
n
) e
n
+
nI
f, f
m
) f
m
,
and we have
T
Tf =
nJ
f, e
n
) T
Te
n
+
nI
f, f
m
) T
Tf
m
=
nJ
s
n
(T)
2
f, e
n
) e
n
.
C.9.3 HilbertSchmidt Integral Operators
We now specialize to the case of integral operators on L
2
(R). Theorem C.19
showed that if k L
2
(R
2
), then the corresponding integral operator L
k
is a
bounded operator on L
2
(R), with operator norm |L
k
| |k|
2
. Exercise C.61
further showed that L
k
is compact. The next result shows that L
k
is actually
HilbertSchmidt when k L
2
(R
2
), and further that all HilbertSchmidt op-
erators on L
2
(R) can be written as integral operators with square-integrable
kernels.
Theorem C.79 (HilbertSchmidt Kernel Theorem).
(a) If k L
2
(R
2
), then the integral operator L
k
with kernel k is Hilbert
Schmidt, and
|L
k
|
B2
= |k|
2
.
(b) If T is a HilbertSchmidt operator on L
2
(R), then there exists k L
2
(R
2
)
such that T = L
k
.
(c) The mapping k L
k
is an isometric isomorphism of L
2
(R
2
) onto
B
2
(L
2
(R)).
Proof. (a) Assume that k L
2
(R
2
), and choose any orthonormal basis
e
n
nN
for L
2
(R). Problem C.30 shows that if we set
e
mn
(x, y) = (e
m
e
n
)(x, y) = e
m
(x) e
n
(y),
then e
mn
m,nN
is an orthonormal basis for L
2
(R
2
). Since k e
mn
L
1
(R
2
),
Fubinis Theorem allows us to interchange integrals in the following calcula-
tion:
k, e
mn
) =
__
k(x, y) e
m
(x) e
n
(y) dxdy
=
_ __
k(x, y) e
n
(y) dy
_
e
m
(x) dx
=
_
L
k
e
n
(x) e
m
(x) dx = L
k
e
n
, e
m
).
Therefore,
|L
k
|
2
B2
=
n=1
|L
k
e
n
|
2
=
n=1
_

m=1
[L
k
e
n
, e
m
)[
2
_
=
m=1
n=1
[k, e
mn
)[
2
= |k|
2
2
< ,
so L
k
is HilbertSchmidt and the mapping k L
k
is isometric.
(b) Suppose that T is HilbertSchmidt on L
2
(R). Choose any orthonormal
basis e
n
nN
for L
2
(R), and let e
mn
be as above. Equation (C.17) shows that
Te
m
, e
n
)
m,nN

2
(N
2
). Therefore, the series
k =
m=1
n=1
Te
n
, e
m
) e
mn
converges unconditionally in L
2
(R
2
). Hence k L
2
(R
2
), so the integral oper-
ator L
k
is HilbertSchmidt. Since part (a) shows that k L
k
is an isometry,
we conclude that
L
k
=
m=1
n=1
Te
n
, e
m
) L
emn
, (C.19)
with unconditional convergence of this series in HilbertSchmidt norm. Since
the HilbertSchmidt norm dominates the operator norm, the series in equa-
tion (C.19) also converges unconditionally in operator norm. Hence, given
f L
2
(R), we have that
L
k
f =
m=1
n=1
Te
n
, e
m
) L
emn
f,
where the series converges unconditionally in L
2
(R). Now, Example C.17
shows that L
emn
is the rank one operator L
emn
f = f, e
n
) e
m
. Therefore,
using the unconditionality of the convergence to reorder the summations, we
have
L
k
f =
m=1
n=1
Te
n
, e
m
) f, e
n
) e
m
=
n=1
f, e
n
)
_

m=1
Te
n
, e
m
) e
m
_
=
n=1
f, e
n
) Te
n
= T
_

n=1
f, e
n
) e
n
_
= Tf,
where the continuity of T is used to obtain the penultimate equality. Hence
T = L
k
.
Thus B
2
(L
2
(R)) is isometrically isomorphic to L
2
(R
2
), and therefore in-
herits a Hilbert space structure from L
2
(R
2
). In terms of kernels, the inner
product on B
2
(L
2
(R)) is
L
k
, L
h
) = k, h).
B
2
(L
2
(R)) is a separable Hilbert space, for if e
n
nN
is any orthonormal basis
for L
2
(R) and we set e
mn
= e
m
e
n
, then L
emn
m,nN
is an orthonormal
basis for B
2
(L
2
(R)). By Example C.17, each operator L
emn
is a bounded,
rank one operator whose range is spane
m
. As discussed in Notation C.18,
we often identify this operator with the function e
m
e
n
, and write
L
emn
= e
m
e
n
.
Thus e
m
e
n
m,nN
is an orthonormal basis for B
2
(L
2
(R)) consisting of
bounded rank one operators, and equation (C.19) is the orthonormal basis
representation of a HilbertSchmidt operator as a superposition of rank one
operators, which we can also write as
T =
m=1
n=1
Te
n
, e
m
) (e
m
e
n
).
This series converges both in HilbertSchmidt norm and in operator norm.
C.9.4 Trace-Class Operators
Denition C.80 (Trace-Class Operators). A compact operator T B(H)
is trace-class if its singular numbers s
n
(T)
nJ
satisfy
n
s
n
(T) < .
The space of trace-class operators on H is the Schatten class B
1
(H).
Since the largest singular number of a compact operator T is s
1
(T) = |T|,
the operator norm is dominated by indextrace-class norm the trace-class norm,
or indeed by any Schatten-class norm:
|T| |T|
Bp
, 1 p .
The next exercise provides an equivalent denition of the trace-class norm
reminiscent of Denition C.71.
Exercise C.81. Suppose that T : H H is compact, and prove the following
facts.
(a) The quantity
n=1
[T[e
n
, e
n
_
is independent of the choice of orthonor-
mal basis e
n
nN
.
(b) T B
1
(H) if and only if
n=1
[T[e
n
, e
n
_
< for some orthonormal
basis e
n
nN
.
(c) If T B
1
(H), then |T|
B1
=
n=1
[T[e
n
, e
n
_
.
(d) T B
1
(H) if and only if [T[
1/2
B
2
(H).
The trace class B
1
(H) is a Banach space with respect to the norm | |
B1
,
and is also a two-sided ideal in B(H). All nite-rank operators are trace-class,
and every trace-class operator is the limit in trace-class norm of a sequence
of nite-rank operators.
The trace of a trace-class operator T B
1
(H) is
trace(T) =
n=1
Te
n
, e
n
_
,
where e
n
nN
is an orthonormal basis for H. This series converges and is
independent of the choice of orthonormal basis. Thus the trace is the sum of
the diagonal elements of the matrix representation of T with respect to an
orthonormal basis.
The trace induces an inner product on the space of HilbertSchmidt op-
erators. Specically, if A, B B
2
(H), then AB
is trace-class, and
A, B) = trace(AB
)
denes an inner product on B
2
(H) whose induced norm is exactly | |
B2
.
Additional Problems
C.39. Given g, h H and 1 p < , show that the Schatten-class norm of
g h is |g h|
Bp
= |g| |h|, where (g h)(f) = f, h) g.
C.10 The HahnBanach Theorem
The HahnBanach Theorem is a powerful result dealing with extension of
linear functionals dened on subspaces of a space.
C.10.1 Abstract Statement of the HahnBanach Theorem
We concentrate in this volume on complex vector spaces. For such spaces, the
abstract version of the HahnBanach Theorem is as follows.
Theorem C.82 (HahnBanach Theorem). Let X be a complex vector
space and let be a seminorm on X. If M is a subspace of X and : M C
is a linear functional on M satisfying
[f, )[ (f), f M,
then there exists a linear functional : X C such that
[
M
= and [f, )[ (f), f X.
C.10 The HahnBanach Theorem 123
For real vector spaces, the dominating seminorm can be replaced by
a sublinear function q : X R that satises q(f + g) q(f) + q(g) and
q(cf) = q(f) for all f, g X and scalars c 0.
The proof of the HahnBanach Theorem takes some preparation, and
therefore we will omit it. The proof for arbitrary vector spaces is not con-
structive, as it relies on the Axiom of Choice in the form of Zorns Lemma.
The HahnBanach Theorem is intimately tied to convexity, and there are gen-
eralizations to more abstract settings such as locally convex topological vector
spaces.
C.10.2 Implications of the HahnBanach Theorem
In practice, it is usually not the HahnBanach Theorem itself but rather its
corollaries that are applied. Therefore we will concentrate on the implications
of the HahnBanach Theorem. Since these corollaries are so important, when
invoking any one of them it is customary to write by the HahnBanach
Theorem instead of by a corollary to the HahnBanach Theorem.
Our rst corollary states that any bounded linear functional on a sub-
space M of a normed space X has an extension to the entire space whose
operator norm on X equals the operator norm of on M. This is easy to
prove when the space is a Hilbert space, but it is far from obvious that such
an extension should be possible on non-inner product spaces. Note that we
do not require the subspace M to be closed.
Corollary C.83 (HahnBanach). Let X be a normed space and M a sub-
space of X. If M
, then there exists X
such that
[
M
= and || = ||.
Proof. Set (f) = || |f| for f X. Note that is dened on all of X, and
is a seminorm on X (in fact, it is a norm if ,= 0). Further,
f M, [f, )[ |f| || = (f).
Hence, Theorem C.82 implies that there exists a linear functional : X C
such that [
M
= (which implies || ||) and
f X, [f, )[ (f) = || |f|,
which implies that || ||.
Given a normed space X and given X
, the operator norm of is

|| = sup
fX, f=1
[f, )[.
Thus, we obtain the operator norm of by looking back at its action on X.
The next corollary provides a complementary viewpoint: The norm of f X
can be obtained by looking forward to its action on X
.
Corollary C.84 (HahnBanach). If X is a normed space and f X, then
|f| = sup
X
, =1
[f, )[. (C.20)
Further, the supremum is achieved.
Proof. Fix f X, and let denote the supremum on the right-hand side of
equation (C.20). Since [f, )[ |f| ||, we have |f|.
Let M = spanx, and dene : M C by cf, ) = c |f|. Then M
and || = 1. Corollary C.83 therefore implies that there exists some X
with [
M
= and || = || = 1. In particular, since f M, we have
[f, )[ = [f, )[ = |f|, and therefore the supremum in equation (C.20)
is achieved.
Now we can give one of the most powerful and often-used implications
of the HahnBanach Theorem. It states that we can nd a bounded linear
functional that separates a point from a closed subspace of a normed space.
This is easy to prove constructively for the case of a Hilbert space, but it is
quite amazing that we can do this in arbitrary normed spaces.
Corollary C.85 (HahnBanach). Let X be a normed linear space. Suppose
that:
(a) M is a closed subspace of X,
(b) f
0
XM, and
(c) d = dist(f
0
, M) = inf
_
|f
0
m| : m M
_
.
Then there exists X
such that
f
0
, ) = 1, [
M
= 0, and || =
1
d
.
Proof. Note that d > 0 since M is closed. Dene M
1
= spanM, x
0
. Since
f
0
/ M, each f M
1
can be written uniquely as f = m
f
+ t
f
f
0
for some
m
f
M and t
f
C. Dene : M
1
C by f, ) = t
f
. Then is linear,
[
M
= 0, and f
0
, ) = 1.
If f M
1
and t
f
,= 0, then we have m
f
/t
f
M, so
|f| = |t
f
f
0
+m
f
| = [t
f
[
_
_
_f
0
_
m
f
t
f
__
_
_ [t
f
[ d.
If t
f
= 0 (so f M), this is still true. Hence, [f, )[ = [t
f
[ |f|/d for all
f M
1
. Therefore || 1/d.
On the other hand, there exist m
n
M such that |f
0
m
n
| d. Since
vanishes on M, we therefore have
1 = f
0
, ) = f
0
m
n
, ) |f
0
m
n
| || d ||.
Therefore || 1/d.
Applying Corollary C.83, there exists a X
such that [
M1
= and
|| = ||. This functional has all of the required properties.
C.10 The HahnBanach Theorem 125
C.10.3 Orthogonal Complements in Normed Spaces
We can generalize the notion of orthogonal complements of subsets of Hilbert
spaces to normed spaces by allowing the orthogonal complement of S X to
be a subset of X
rather than X.
Denition C.86 (Orthogonal Complement). If S is a subset of a normed
space X, then its orthogonal complement S
is the subset of X
dened by
S
=
_
X
: f, ) = 0 for all f S
_
.
Exercise C.87. Show that S
is a closed subspace of X
.
The next exercise gives a useful characterization of complete sequences in
a normed space. The analogous characterization for the case of Hilbert spaces
appeared as Exercise A.101.
Exercise C.88. Let S be a subset of a normed space X. Prove that
S is complete S
= 0.
Consequently, a subspace M of X is dense in X if and only if M
= 0.
C.10.4 X
and Reexivity
Let X be a normed linear space. Following our notational conventions, if
f X and X
then we write f, ) to denote the action of on f, and

we take this notation to be linear in f and antilinear in . Holding xed,
the mapping f f, ) is a bounded linear functional on X. On the other
hand, if we hold f xed, then f, ) is an antilinear functional on X
.
We can make a linear mapping by considering f, ) instead. Thus, each
element f X determines a linear functional

f : X
C by the formula
,

f ) = f, ) for X
. If this linear functional is bounded, then it is a

bounded linear functional on X
and hence belongs to the dual space of X
,
which we denote by X
. The next result shows that

f is indeed bounded and
has operator norm |
f | = |f|. This gives us a a natural identication of X

with a subset of X
.
Theorem C.89. Let X be a normed space. Given f X, dene

f : X
C
by
,

f ) = f, ), X
.
Then

f is a bounded linear functional on X
, with operator norm |
f | = |f|.
Consequently, the mapping
T : X X
f

f,
is a linear isometry of X into X
.
Proof. By denition of the operator norm, we have
|
f | = sup
X
, =1
[,

f )[.
On the other hand, by the HahnBanach Theorem in the form of Corol-
lary C.84, we have
|f| = sup
X
, =1
[f, )[.
Since [f, )[ = [,

f )[, the result follows.
Denition C.90 (Natural Embedding of X into X
). Let X be a
normed space.
(a) The mapping T : f

f dened in Theorem C.89 is called the natural
embedding or the canonical embedding of X into X
.
(b) If the natural embedding of X into X
is surjective, then we say that X

is reexive.
Remark C.91. We emphasize that in order for X to be called reexive, the
natural embedding must be a surjective isometry. There exist Banach spaces X
such that X

= X
even though X is not reexive [Jam51].

Exercise C.92. Let E R
d
be Lebesgue measurable. Show that
p
and
L
p
(E) are reexive for each 1 < p < .
C.10.5 Adjoints of Operators on Banach Spaces
We saw in Section C.6 that if H, K are Hilbert spaces and A B(H, K),
then there exists an adjoint operator A
B(K, H) uniquely dened by the

condition
f H, g H, Af, g)
K
= f, A
g)
H
. (C.21)
Now we will consider the case where X, Y are Banach spaces and A B(X, Y ).
We will see that, as a consequence of the HahnBanach Theorem, there ex-
ists a unique adjoint A
B(Y
, X
) dened by an equation analogous to

equation (C.21). However, while we have A: X Y, the adjoint will be a
map A
: Y
. In particular, unlike the Hilbert space case, we cannot

consider compositions of A with A
.
Exercise C.93. Let X, Y be Banach spaces, and x A B(X, Y ).
(a) Fix Y
. Dene A
: X C to be the functional whose rule is

f, A
) = Af, ), f X.
Show that A
.
C.11 The Baire Category Theorem 127
(b) Show that A
: A
is bounded and linear, and |A
| = |A|.
(c) Prove that A
is the unique operator from Y
to X
that satises
f X, Y
, Af, ) = f, A
). (C.22)
Denition C.94 (Adjoint). Let X, Y be Banach spaces. Then the adjoint
of A B(X, Y ) is the unique operator A
: Y
that satises equa-

tion (C.22).
Remark C.95. Our use of a sesquilinear form notation for linear functionals
allows us to create a denition of the adjoint that directly generalizes the
denition of the adjoint of an operator on Hilbert spaces. In other words, if
H and K are Hilbert spaces and A B(H, K) then there is no dierence
between the adjoint A
dened in Section C.6, where we considering H and

K to be Hilbert spaces, and the adjoint A
dened above, where we consider

H and K to be Banach spaces. If we instead used a bilinear form notation,
there would be a distinction between these two adjoints.
Additional Problems
C.40. Let X be a normed space. Show that if X
is separable then X is
separable, but the converse can fail.
C.41. Let X, Y be Banach spaces. Show that if T : X Y is a topological
isomorphism, then the adjoint T
: Y
is a topological isomorphism,
and similarly if T is an isometric isomorphism then so is T
.
C.11 The Baire Category Theorem
It is not possible to write the Euclidean plane R
2
as the union of countably
many straight lines. This is a special case of the Baire Category Theorem,
which states that a complete metric space cannot be written as a countable
union of nowhere dense sets.
Denition C.96 (Nowhere Dense Sets). Let X be a metric space, and
let E X be given.
(a) E is nowhere dense or rare if X E is dense in X.
(b) E is meager or rst category if it can be written as a countable union of
nowhere dense sets.
(c) E is nonmeager or second category if it is not meager.
Exercise C.97. Given a nonempty subset E of a metric space X, show that
(a) E is nowhere dense.
(b) E contains no nonempty open subsets.
The Cantor set is an example of a nowhere dense subset of [0, 1]. The set of
rationals Q is not a nowhere dense subset of R, but it is meager in R. Further,
Q with the metric d(x, y) = [x y[ is an incomplete metric space that is a
meager subset of itself.
Now we can prove the Baire Category Theorem, which states that no com-
plete metric space can be a meager subset of itself. Unlike the HahnBanach
Theorem, the driving property here is completeness rather than convexity.
Theorem C.98 (Baire Category Theorem). A nonempty complete met-
ric space is nonmeager in itself. Consequently, if
X =
n=1
E
n
where each E
n
is a closed subset of X, then at least one E
n
contains a
nonempty open subset.
Proof. Suppose that X = E
n
where each E
n
is nowhere dense. Then U
n
=
X E
n
is dense and open for each n.
Choose f
1
U
1
and let r
1
> 0 be such that B
1
= B
r1
(f
1
) U
1
. Then
since U
2
is dense, we can nd an f
2
U
2
B
1
. Since U
2
and B
1
are both
open, there exists some r
2
> 0 such that B
2
= B
r2
(f
2
) U
2
B
1
. Without
loss of generality, we can take r
2
small enough that we have both r
2
< r
1
/2
and B
2
B
1
.
Continuing in this way we obtain points f
n
U
n
and balls B
n
= B
rn
(f
n
)
U
n
such that
r
n
<
r
n1
2
and B
n
B
n1
.
In particular, r
n
0, and the balls B
n
are nested.
Fix > 0, and let N be large enough that r
N
< /2. If m, n > N, then
we have f
m
, f
n
B
N
. Hence d(f
m
, f
n
) < 2r
N
< . Thus f
n
nN
is Cauchy,
and since X is complete, there exists some f X such that f
n
f.
Now x any N > 0. Then, since the B
n
are nested, we have f
n
B
N+1
for all n > N. As f
n
f, this implies that f B
N+1
B
N
. This is true for
every N, so
f
n=1
B
n

n=1
U
n
=
n=1
(X E
n
).
But then f / E
n
, which is a contradiction.
The next exercise applies the Baire Category Theorem to show that Hamel
bases for Banach spaces are generally too unweildy to be of much use.
Exercise C.99. Let X be an innite-dimensional Banach space, and prove
the following.
C.12 The Uniform Boundedness Principle 129
(a) Any Hamel basis for X must be uncountable.
(b) Any innite-dimensional subspace of X that has a countable Hamel
basis is a meager subset of X, and cannot be a closed subspace of X.
(c) C
c
(R) is a meager, dense subspace of C
0
(R) that does not have a
countable Hamel basis.
C.12 The Uniform Boundedness Principle
The Uniform Boundedness Principle states that a family of bounded linear
operators on a Banach space that are uniformly bounded at each individual
point must actually be uniformly bounded in operator norm.
Theorem C.100 (Uniform Boundedness Principle). Let X be a Ba-
nach space and Y a normed space. If A
i
iI
is any collection of operators
in B(X, Y ) such that
f X, sup
iI
|A
i
f| < ,
then
sup
iI
|A
i
| < .
Proof. Set
E
n
=
_
f X : sup
iI
|A
i
f| n
_
.
Then X = E
n
by hypothesis, and since each A
i
is continuous it follows
that E
n
is closed. Consequently, the Baire Category Theorem implies that
some E
n
must contain an open ball, say B
r
(f
0
) E
n
.
Given an arbitrary f X, if we set g = f
0
+ sf with s =
r
2f
then we
have g B
r
(f
0
) E
n
, and therefore
|A
i
f| =
_
_
_A
i
_
g f
0
s
__
_
_
1
s
_
|A
i
g| +|A
i
f
0
|
_

2 |f|
r
2n =
4n
r
|f|.
Consequently, |A
i
| 4n/r, which is a constant independent of i.
The following special case of the Uniform Bounded Principle is often use-
ful (sometimes the names Uniform Boundedness Principle and Banach
Steinhaus Theorem are used interchangeably).
Exercise C.101 (BanachSteinhaus Theorem). Let X, Y be Banach
spaces. If A
n
B(X, Y ) for each n N and Af = lim
n
A
n
f exists for
each f X, then A B(X, Y ) and |A| sup |A
n
| < .
C.13 The Open Mapping Theorem
Recall that a function is continuous if the inverse image of any open set is
open. It is often important to consider direct images of open sets as well.
Denition C.102 (Open Mapping). Let X, Y be topological spaces. A
function U : X Y is an open mapping if
U is open in X = A(U) is open in Y.
In general, a continuous function need not be an open mapping, but the
Open Mapping Theorem asserts that any continuous linear surjection of one
Banach space onto another must be open. The key to the proof is the following
lemma. For clarity, we will write B
X
r
(f) and B
Y
r
(g) to distinguish open balls
in X from open balls in Y.
Lemma C.103. Let X, Y be Banach spaces and x A B(X, Y ). If A(B
X
1
(0))
contains an open ball in Y, then A(B
X
1
(0)) contains an open ball B
Y
r
(0) for
some r > 0.
Proof. Suppose that A(B
X
1
(0)) contains an open ball B
Y
s
(g). We claim that if
we set r = s/2, then
B
Y
r
(0) A(B
X
1
(0)). (C.23)
To see this, x h B
Y
r
(0), i.e., |h|
X
< r = s/2. Then 2h + g B
Y
s
(g)
A(B
X
1
(0)). Hence there exist f
n
X with |f
n
|
X
< 1 such that Af
n
2h+g.
Also, g B
Y
s
(g) A(B
X
1
(0)), so there exist g
n
X with |g
n
|
X
< 1 such
that Ag
n
g. Then w
n
= (f
n
g
n
)/2 B
X
1
(0), and
Aw
n
=
Af
n
Ag
n
2

(2h + g) g
2
= h as n .
Hence h A(B
X
1
(0)), so equation (C.23) holds.
Now we will show that we actually have B
Y
r/2
(0) A(B
X
1
(0)). To see this,
suppose h B
Y
r/2
(0), i.e., |h| < r/2. Rescaling equation (C.23), we have
h A(B
X
1/2
(0)), so there exists some f
1
X with |f
1
| < 1/2 such that
|hAf
1
| < r/4. Then hAf
1
B
Y
r/4
(0) A(B
X
1/4
(0)), so there exists some
f
2
X with |f
2
| < 1/4 such that |(h Af
1
) Af
2
| < r/8. Continuing in
this way, we obtain vectors f
n
X with |f
n
| < 2
n
such that
|h Ag
n
| <
r
2
n+1
,
where g
n
=
n
k=1
f
k
. Hence Ag
n
h. However, g
n
nN
is Cauchy in X, so
g
n
g for some g X. Since A is continuous, it follows that h = Ag. Since
|g| < 1, we therefore have h A(B
X
1
(0)).
C.14 The Closed Graph Theorem 131
Theorem C.104 (Open Mapping Theorem). If X, Y are Banach spaces
and A: X Y is a continuous linear surjection, then A is an open mapping.
Proof. Since A is surjective we have
Y =
k=1
A(B
X
k
(0)).
The Baire Category Theorem therefore implies that some set A(B
X
k
(0)) must
contain an open ball. Therefore, by Lemma C.103, there is some r > 0 such
that
B
Y
r
(0) A(B
X
1
(0)). (C.24)
Now suppose that U X is open and y A(U). Then y = Ax for
some x U, so B
X
s
(x) U for some s > 0. Rescaling equation (C.24),
B
Y
t
(0) A(B
X
s
(0)) for some t > 0. Therefore
B
Y
t
(y) = B
Y
t
(0) +Ax A(B
X
s
(0) +x) = A(B
X
s
(x)) A(U),
so A(U) is open.
Specializing to the case of bijections yields the following result.
Theorem C.105 (Inverse Mapping Theorem). If X, Y are Banach
spaces and A: X Y is a continuous linear bijection, then A
1
: Y X
is continuous. Consequently A is a topological isomorphism.
The next exercise is a typical application of the Inverse Mapping Theorem.
Exercise C.106. Suppose X is a vector space that is complete with respect
to two norms | | and [[[ [[[. If there exists C > 0 such that |f| C [[[f[[[ for
all f X, then | | and [[[ [[[ are equivalent norms on X.
Additional Problems
C.42. Let X and Y be Banach spaces. Show that A B(X, Y ) is surjective
if and only if range(A) is not meager in Y.
C.14 The Closed Graph Theorem
The Closed Graph Theorem provides a convenient means of testing whether
a linear operator on Banach spaces is continuous.
Theorem C.107 (Closed Graph Theorem). Let X and Y be Banach
spaces. If A: X Y is linear, then the following statements are equivalent.
(a) A is continuous.
(b) graph(A) =
_
(f, Af) : f X
_
is a closed subset of X Y.
(c) If f
n
f in X and Af
n
g in Y, then g = Af.
Proof. (c) (a). Assume that statement (c) holds. For clarity, write | |
X
and | |
Y
for the norms on X and Y, and dene
[[[f[[[ = |f|
X
+|Af|
Y
, f X.
Exercise: Show that [[[ [[[ is a norm on X, and that X is complete with respect
to this norm.
Since we have |f|
X
[[[f[[[ for f X and X is complete with respect to
both norms, it follows from Exercise C.106 that there exists a constant C > 0
such that [[[f[[[ C |f|
X
for f X. Consequently, |Af|
Y
C |f|
X
, so A is
bounded.
C.15 Schauder Bases
Schauder bases were introduced and briey discussed in Section A.9. In this
section we derive some additional properties of Schauder bases and related
systems.
C.15.1 Continuity of the Coecient Functionals
Recall from Denition A.80 that a sequence T = f
k
kN
is a Schauder basis
for a Banach space X if every f X can be written as
f =
k=1
k
(f) f
k
(C.25)
for a unique choice of scalars
k
(f), where the series converges in the norm
of X.
With k xed,
k
is a mapping of X to scalars, i.e., it is a functional on X.
Further, it follows immediately from the denition that
k
is linear. We call
kN
the associated family of coecient functionals. With j xed, we have
k=1
jk
f
k
= f
j
=
k=1
a
k
(f
j
) f
k
,
so by uniqueness we must have a
k
(f
j
) =
jk
. We therefore say that f
k
kN
and
k
kN
are biorthogonal sequences.
Despite the fact that the denition of Schauder basis does not require the
functionals
k
to be bounded, we will shortly see the surprising result that
each
k
must be continuous, and hence belongs to X
. In order to prove this,

for each N N dene the partial sum operator S
N
: X X by
C.15 Schauder Bases 133
S
N
f =
N
k=1
k
(f) f
k
, f X.
By denition, for each f X we have that S
N
f f as N . Clearly,
range(S
N
) spanf
1
, . . . , f
N
, and by biorthogonality we actually have that
range(S
N
) = spanf
1
, . . . , f
N
.
Each S
N
is a linear mapping on X, but we do not yet know whether it is
continuous. This will be shown in the next exercise. The main idea is to show
that
[[[f[[[ = sup
N
|S
N
f|, f X.
denes an equivalent norm for X.
Exercise C.108. Let f
k
kN
be a Schauder basis for a Banach space X.
Using the notations given above, prove the following facts.
(a) Show that [[[ [[[ is a norm on X, and |f| [[[f[[[ for all f X.
(b) Suppose that g
n
nN
is a Cauchy sequence in X with respect to [[[ [[[.
With N xed, show that S
N
g
n
nN
is Cauchy with respect to | |.
Let h
N
be such that |h
N
S
N
g
n
| 0 as n , and observe that
h
N
spanf
1
, . . . , f
N
.
(c) Show that
lim
n
_
sup
N
|h
N
S
N
g
n
|
_
= 0,
and use this to show that h
N
NN
is Cauchy with respect to | |. Let
g X be the element such that |g h
N
| 0.
(d) Show that S
N
(h
N+1
) = h
N
, and use this to show that h
N
=
N
k=1
c
k
f
k
where c
k
is independent of N.
(e) Show that g =
k=1
c
k
f
k
, and hence h
N
= S
N
g. Use this to show that
[[[g g
n
[[[ 0, and conclude that X is complete with respect to [[[ [[[.
(f) Show that | | and [[[ [[[ are equivalent norms on X, and use this to show
that
C = sup
N
|S
N
| < .
The number C is called the basis constant for f
k
kN
.
(g) Show that 1 |
k
| |f
k
| 2C for each k N.
Notation C.109. Since the coecient functionals associated with a Schauder
basis are continuous, we adopt our standard functional notation and write
f,
k
) instead of
k
(f), with the understanding that this functional notation
is linear in the rst variable but antilinear in the second. More details on this
functional notation are given in Notation C.36.
C.15.2 Minimal Sequences
As we have seen, if f
k
kN
is a Schauder basis with coecient functionals
kN
, then we have the biorthogonality condition f
j
,
k
) =
jk
. Unfortu-
nately, the existence of a biorthogonal sequence is not sucient by itself to
guarantee that we have a Schauder basis. So, let us spend a little time con-
sidering sequences that possess a biorthogonal sequence, and then see what
extra conditions we need to impose in order to have a Schauder basis.
The next exercise will recast the existence of a biorthogonal sequence in
terms of a type of generalized linear independence property called minimal-
ity.
Denition C.110 (Minimal and Biorthogonal Sequences). Let f
k
kN
be a sequence in a Banach space X.
(a) f
k
kN
is minimal if no vector f
j
lies in the closed span of the other f
k
:
j N, f
j
/ span
_
f
k
k=j
_
.
(b) A sequence
k
kN
in X
is biorthogonal to f
k
nN
if f
j
,
k
) =
jk
.
(c) A sequence that is both minimal and complete is said to be exact.
Compare the following characterization of minimal sequences in Banach
spaces to the analogous result for Hilbert spaces given in Problem A.29.
Whereas the proof for Hilbert spaces relies on the existence of an orthog-
onal complements, the key for generic Banach spaces is the HahnBanach
Theorem.
k
nN
be a sequence in a Banach space X. Prove that
(a) f
k
kN
is minimal.
(b) There exists a sequence
k
kN
in X
that is biorthogonal to f
k
kN
.
Additionally, prove that the following statements are equivalent.
(a) f
k
kN
is exact (both minimal and complete).
(b) There exists a unique sequence
k
kN
in X
that is biorthogonal to
f
k
kN
.
Every Schauder basis is minimal since it possesses a biorthogonal sequence,
and it is complete by denition. In nite dimensions, a sequence x
1
, . . . , x
n
is minimal if and only if it is linearly independent, and in this case it is a basis

for its span. These simple facts do not extend to innite sequences. Instead,
there are several shades of grey in the meaning of independence in innite-
dimensional spaces, as illustrated in the following exercise and examples (see
Problem A.28) for the denition of an -independent sequence).
C.15 Schauder Bases 135
k
kN
be a sequence in a Banach space X, and prove
the following implications:
Schauder basis = minimal = -independent =
nitely
independent.

The following examples show that none of the converse implications hold
in general, even in a Hilbert space.
Example C.113. (a) Problem C.43 shows that xe
2inx
n=0
is both minimal
and complete in L
2
[0, 1], but is not a Schauder basis for L
2
[0, 1].
(b) Let e
n
nN
be an orthonormal basis for a Hilbert space H. Set f
1
= e
1
,
and for n > 2 dene f
n
= e
1
+ (e
n
/n). Exercise: f
n
nN
is -independent
and complete, but is not minimal.
(c) Problem A.28 constructs an example of a sequence in
2
that is nitely
linearly independent and complete, but is not -independent.
C.15.3 A Characterization of Schauder Bases
Now we will determine exactly what extra property an exact system must
possess in order to be a Schauder basis.
Suppose that f
k
kN
is a minimal sequence in a Banach space X. Then,
by denition, there exists a sequence
k
kN
in X
that is biorthogonal
to f
k
kN
. Therefore, even though we do not know whether f
k
kN
is a
Schauder basis, we can dene partial sum operators
S
N
f =
N
k=1
f,
k
) f
k
, f X,
and we do have that each S
N
is a bounded operator on X. If we have that
S
N
f f, then we will have f =
k=1
f,
k
) f
k
, and the minimality as-
sumption will ensure that this representation is unique. The question is under
what conditions we can be sure that the partial sums will converge, and this
is answered in the next exercise.
k
kN
be a minimal sequence in a Banach space X,
and let
k
kN
be its biorthogonal sequence in X
. Prove that the following

(a) f
k
kN
is a Schauder basis for X.
(b) S
N
f f for each f X.
(c) f
k
kN
is complete, and for each f X we have sup |S
N
f| < .
(d) f
k
kN
is complete, and sup |S
N
| < .
Thus, a Schauder basis is precisely an exact sequence whose basis constant
C = sup |S
N
| is nite.
C.15.4 Unconditional Bases
An unconditional basis for a Banach space X is a Schauder basis f
k
kN
which has the property that the representations f =
k=1
f,
k
) f
k
converge
unconditionally for each f X. Not every Schauder basis is unconditional.
For example, while e
2inx
nZ
is an orthonormal and hence unconditional
basis for L
2
[0, 1], if 1 < p < and p ,= 2 then it is a Schauder basis for L
p
[0, 1]
that is not unconditional. A Hilbert space example of a Schauder basis that
is not unconditional is
_
x
1
2
1/4
e
2inx
_
nZ
in L
2
[0, 1] [Bab48] (see also the discussion in [Sin70, pp. 351354]). In contrast,
to these systems, the Haar system
_
[0,1]
_

_
2
n/2
(2
n
x k)
_
n0, k=0,...,2
n
1
,
where =

[0,1/2)
[1/2,1)
, forms an unconditional basis for L
p
[0, 1] for each
1 < p < (compare Problem B.29 for the case p = 2).
Additional Problems
C.43. (a) Show that xe
2inx
n=0
is both minimal and complete in L
2
(T),
and nd its biorthogonal system g
n
n=0
in L
2
(T).
(b) Show that sup |g
n
|
2
= , and use this to prove that xe
2inx
n=0
is
not a Schauder basis for L
2
(T) with respect to any ordering of Z.
C.16 Weak and Weak* Convergence
In this section we will discuss weak and weak* convergence of sequences. A
detailed discussion of the weak and weak* topologies and the corresponding
convergence criterion for nets is given in Section E.6. For now, we take the
following as our denition of weak or weak* convergence of sequences.
Denition C.115. Let X be a normed linear space.
(a) A sequence f
n
nN
in X converges weakly to f X, denoted f
n
w
f, if
X
, lim
n
f
n
, ) = f, ).
(b) A sequence
n
nN
in X
converges weak* to X
, denoted
n
w*
,
if
f X, lim
n
f,
n
) = f, ).
C.16 Weak and Weak* Convergence 137
Thus, weak* convergence is just pointwise convergence of the opera-
tors
n
. Weak* convergence is only dened for sequences that lie in a dual
space X
. Given a sequence
n
nN
in X
, we can consider three types

of convergence: strong (norm), weak, and weak* convergence. By denition,
these are:
n
lim
n
|
n
| = 0,
n
w
T X
, lim
n
n
, T) = , T),
n
w*
x X, lim
n
x,
n
) = x, ).
Exercise C.116. Let X be a normed space and suppose that f
n
, f X and
n
, X
. Show that
f
n
f = f
n
w
f
and
n
=
n
w
=
n
w*
.
If X is reexive, then weak and weak* convergence coincide. In par-
ticular, this is the case for Hilbert spaces. In any nite-dimensional space,
strong and weak convergence coincide, but they are distinct in any innite-
dimensional space. For example, if e
n
nN
is an orthonormal sequence in a
Hilbert space H, then it follows from Bessels Inequality that e
n
, f) 0 for
each f H and hence e
n
w
0, but |e
n
0| = 1 / 0.
Here are some properties of weak and weak* convergent sequences. While
the fact that norm-convergent sequences are bounded is trivial, the corre-
sponding fact for weak and weak* convergent sequences is much more subtle
and is most easily proved by applying the Uniform Boundedness Principle.
Exercise C.117. Let X be a Banach space. Show that if
n
w*
in X
,
then is unique and sup |
n
| < .
In order to conclude boundedness of a weak* convergent sequence, the
hypothesis in Exercise C.117 that X be complete is necessary. For example, if
we take X = c
00
then we have
1
= (c
00
)
(compare Problem D.29) and the

sequence n
n
nN
in
1
converges weak* to 0, but is unbounded. In general,
given any normed linear space X there is always a unique Banach space

X,
called the completion of X, that contains X as a dense subspace. Consequently
X
= (

X)
, and we should always use

X instead of X when discussing weak*
convergence. For the space c
00
, the completion is c
0
.
On the other hand, the analogous uniqueness and boundedness results for
weakly convergent sequence hold for any normed space, because of the fact
that X
is complete.
Exercise C.118. Let X be a normed space. Show that if f
n
w
f in X, then
f is unique and sup |f
n
| < .
Additional Problems
C.44. (a) Given 1 < p < and x
n
, y
p
, show that x
n
w
y in
p
if and only
if x
n
(k) y(k) for each k (componentwise convergence) and sup |x
n
|
p
< .
Does either implication remain valid if p = 1?
(b) Problem D.29 shows that
1
= c
0
. Assuming this, given x

n
, y
1
, show that x
n
w*
y in
1
if and only if x
n
(k) y(k) for each k and
sup |x
n
|
1
< .
(c) Recall that

=
1
. Given x
n
, y
, show that x
n
w*
y in
if
and only if x
n
(k) y(k) for each k and sup |x
n
|
< .
D
Borel and Radon Measures on the Real Line
In this appendix we review the theory of signed and complex Borel and Radon
measures. These can be dened on locally compact Hausdor spaces, but here
we will only deal with Borel and Radon measures on the real line. This allows
several simplications as compared to the most general theory. We follow the
development of abstract measures, Borel measures, and Radon measures given
in Follands text [Fol99], specializing to the case of measures on the real line.
D.1 -Algebras
It can be shown that the Axiom of Choice implies that there is no way to
dene a function on all of the subsets of R so that all of the following hold:
(i) 0 (E) for every E R,
(ii) ([a, b]) = b a,
(iii) if E
1
, E
2
, . . . are nitely or countably many disjoint sets, then (
k
E
k
) =
k
(E
k
), and
(iv) (E +h) = (E) for all E R and h R.
There are several ways to address this. In Appendix B we began with
Lebesgue exterior measure [ [
e
, which does satisfy (i), (ii), and (iv), but
fails requirement (iii). Rather unsettlingly, for Lebesgue exterior measure,
E F = does not imply that [E F[
e
= [E[
e
+[F[
e
.
In order to obtain Lebesgue measure [ [, we therefore dropped require-
ment (i), with the result that not all subsets of R are Lebesgue measurable.
Although we cannot measure every set, we do have the satisfying fact that
requirements (ii), (iii), and (iv) hold for all those sets E that are Lebesgue
measurable.
On the other hand, there are good reasons for relaxing the requirements
in other ways. For example, one of the most important measures is the
measure, which assigns the size 1 or 0 to a set E depending on whether the
140 D Borel and Radon Measures on the Real Line
origin belongs to E or not. Requirements (ii) and (iv) are not satised by the
measure, but both (i) and (iii) do hold.
Additional alternatives are to allow a measure to take real or complex
values, instead of just nonnegative values. This leads to the idea of signed and
complex measures on R.
In this appendix we will review the denitions and properties of abstract
Borel and Radon measures on the real line. In order to give a useful denition
of a measure, we must rst decide on the properties that a class of sets should
possess in order to be measured.
Denition D.1 (-Algebra). A -algebra on R is a nonempty collection
of subsets of R which satises:
(a) is closed under both nite and countable unions:
E
1
, E
2
, =

k
E
k
,
(b) is closed under complements:
E = E
C
= RE .
Note that if is a -algebra then, since it is nonempty, it contains some
set E R, hence contains RE, and therefore contains both R = E (RE)
and = RR.
By Theorem B.8, the class L of all Lebesgue measurable subsets of R forms
a -algebra. The power set T(R) = E : E R is trivially another -algebra.
Given a particular class c of subsets of R, there will be many -algebras
that contain c. However, there is a unique smallest -algebra that contains c.
Exercise D.2. Let c be a collection of subsets of R. Show that
(c) =
_
: is a -algebra and c
_
is a -algebra on R, called the -algebra generated by c.
Note that if
1
,
2
are -algebras, then
1

2
is not formed by inter-
secting the elements of
1
with those of
2
. Rather, it is the collection of all
sets that are common to both
1
and
2
. Thus, if is any -algebra that
contains c, then we will have (c) , which explains why (c) is the
smallest -algebra that contains c.
The next denition introduces the particular -algebra that will concern
us in this appendix.
Denition D.3 (Borel -algebra). The Borel -algebra B
on R is the
smallest -algebra that contains all the open subsets of R. That is, if we set
| = U R : U is open, then
B
= (|).
The elements of B
are called the Borel subsets of R.

D.2 Signed Measures 141
In particular, B
includes all the open and closed subsets of R, as well

as the G
and F
sets that were introduced in Denition B.12. However, not

every subset of R is a Borel set. The Borel -algebra can be dened on R
d
or
C
d
in an analogous manner.
Although our focus will be on the Borel -algebra on R, many of the
denitions and results that we will discuss are valid on more general domains.
In particular, in this appendix it is often instructive for the reader to consider
the even simpler case of measures on the natural numbers N. Some of the
additional problems at the end of each section of this appendix deal with this
setting. The topology on N is the discrete topology, i.e., every subset of N is
open. Hence every subset of N is a Borel set, or, in other words, the Borel
-algebra on N is T(N), the power set of N. Whenever we speak of a measure
on N, we will assume that the associated -algebra is T(N).
D.2 Signed Measures
Denition D.4 (Signed Measure). A function : B
[, ] is a
signed Borel measure on R, or simply a signed measure, if
(a) () = 0,
(b) takes at most one of the values , and
(c) if E
1
, E
2
, . . . are nitely or countably many disjoint Borel sets, then
k
E
k
_
=
k
(E
k
).
If (E) 0 for each E B
, then we say that is a positive measure, and

in this case we write 0.
If [(E)[ < for each E B
, then we say that is a bounded measure

or a nite measure.
If [(K)[ < for every compact set K R, then we say that is a locally
nite measure.
Note that many texts dene a measure to be what we are calling a
positive measure.
Notation D.5. For simplicity of notation, we will use the abbreviations
(a, b) = ((a, b)), [a, b] = ([a, b]), x = (x), etc.
Example D.6. By Theorem B.8, the -algebra L consisting of all Lebesgue
measurable subsets of R contains all the open sets, and hence contains the
Borel sets. Therefore L B
, although there do exist Lebesgue measurable

sets that are not Borel sets. For example, all sets with exterior Lebesgue
measure zero are Lebesgue measurable, but need not be Borel sets. Therefore
we can consider Lebesgue measure restricted to the Borel sets (which we
simply refer to as Lebesgue measure), and this gives an example of a positive,
unbounded Borel measure. We usually denote this measure by dx, although we
will continue to write [E[ for the Lebesgue measure of E instead of dx(E).
Our rst example of a Borel measure modies Lebesgue measure by
weighting the real line.
Exercise D.7. Let f : R [, ] be Lebesgue measurable, with at least
one of
_
f
+
,
_
f
nite. Show that

(E) =
_
E
f(x) dx, E B
,
denes a signed measure. This measure is positive if and only if f 0 a.e.,
and it is bounded if and only if f L
1
(R). Note that Lebesgue measure
corresponds to the case where f is the constant function 1.
Our next example is a very dierent type of measure.
Exercise D.8 (The Measure). Given a R, for E R dene
a
(E) =
_
1, a E,
0, a / E.
Show that
a
is a positive, bounded Borel measure.
The measure
a
has many names, including the measure at a, the Dirac
measure at a, and the point mass at a. For a = 0 we use the shorthand
=
0
.
Note that
a
is not translation-invariant, i.e.,
a
(E) and
a
(E+h) will not be
equal in general.
Here is another example of an unbounded positive measure.
Exercise D.9. Dene counting measure by declaring (E) to be the car-
dinality of E if E is a nite set, and otherwise. Show that is a positive,
unbounded Borel measure that is not locally nite.
We can create signed measures from positive measures as follows.
Exercise D.10. Show that if
1
,
2
are positive measures with at least one
of
1
,
2
bounded, then
1
2
is a signed measure.
Any positive measure has the useful property that it is monotonic, i.e.,
if A, B are Borel sets and A B, then (A) (B). In particular, if is
positive and (E) = 0, then (A) = 0 for every Borel A E. However, for
signed measures it is important to distinguish between sets E that satisfy
(E) = 0, and sets that are null for in the following sense.
D.2 Signed Measures 143
Denition D.11 (Null Sets). We say that a signed measure is null on a
set E B
if (A) = 0 for every A B
with A E.
Denition D.12 (Mutually Singular Measures). Two signed measures
, are mutually singular, denoted , if there exist E, F B
such that
(a) E F = R and E F = ,
(b) is null on F, and
(c) is null on E.
Exercise D.13. Show that Lebesgue measure and the -measure are mutu-
tally singular measures, i.e., dx .
D.2.1 The Jordan Decomposition
We come now to a fundamental decomposition for signed measures. The fol-
lowing exercise motivates this by considering the special case of measures of
the form (E) =
_
E
f(x) dx.
Exercise D.14. Let (E) =
_
E
f(x) dx be a measure of the form constructed
in Exercise D.7. Set P = f 0 and N = f < 0. By changing P and N
by a set of measure zero, we may assume that P and N are Borel sets (see
Exercise B.14). For E B
dene
+
(E) = (E P) =
_
E
f
+
(x) dx,
(E) = (E N) =
_
E
f
(x) dx.
Show that
+
,
are positive measures, =

+
, and
+

.
Although we will not prove it, the next result states that this same kind
of decomposition holds for arbitrary signed measures.
Theorem D.15 (Jordan Decomposition Theorem). If is a signed
Borel measure on R, then there exist unique positive Borel measures
+
,
such that =
+
and
+

.
Thus, for any signed measure , there exist disjoint Borel sets P, N R
with P N = R such that =
+
is null on P, and
+
is null on N.
We call =
+

the Jordan decomposition of (since it is unique), and

R = P N an associated Hahn decomposition of R (it is not unique).
Denition D.16 (Positive, Negative, and Total Variation Measures).
Given a signed Borel measure , let =
+
be its Jordan decomposition.

(a) We call
+
the positive variation of , and
the negative variation of .

(b) The total variation of is the positive measure
[[ =
+
+
.
That is, [[ is dened by
[[(E) =
+
(E) +
(E), E B
. (D.1)
Observe that equation (D.1) implies that
[(E)[ [[(E), E B
.
Further, since [[ is a positive measure, it is monotonic, and hence
[[(E) [[(R), E B
.
Of course, [[(R) could be innite, but if it should be nite then it follows that
[(E)[ is bounded by the nite quantity [[(R) for every E B
, and hence
is a bounded measure. In fact, the next exercise shows that the converse is
also true, which explains the terminology bounded measure instead of just
nite measure: If (E) is nite for all Borel sets E, then there is a nite
upper bound to the values of [(E)[.
Exercise D.17. Let be a signed Borel measure. Prove that
is bounded [[(R) < ,
and in this case we have sup
EB
[(E)[ [[(R).
Here are some useful equivalent formulations of the positive, negative, and
total variation measures.
Exercise D.18. Let be a signed Borel measure. Show that if E B
, then
+
(E) = sup
_
(A) : A B
, A E
_
,
(E) = inf
_
(A) : A B
, A E
_
,
[[(E) = sup
_
n
k=1
[(E
k
)[ : n N, E
k
B
, E =
n
k=1
E
k
disjointly
_
.
Now that we have dened the total variation measure, we can dene -
nite measures.
Denition D.19 (-Finite Measures). Let be a signed Borel measure.
If we can write R = E
k
using at most countably many sets E
k
B
each
with [[(E
k
) < , then we say that is a -nite measure.
All bounded measures are -nite. Lebesgue measure is an example of an
unbounded measure that is -nite, while counting measure is an example of
a measure that is not -nite.
Exercise D.20. Show that every locally nite signed Borel measure is -
nite, but a -nite signed Borel measure need not be locally nite.
D.3 Positive Measures and Integration 145
Additional Problems
D.1. Show that E B
is a null set for a signed measure if and only if

[[(E) = 0.
D.2. Let =
+

be the Jordan decomposition of a signed measure ,

and let R = P N be an associated Hahn decomposition. Show that
+
(E) =
(EP),
(E) = (EN), and [[(E) = (EP)(EN) for E B
.
D.3. Suppose that is a signed measure, and that E
k
are disjoint Borel sets
such that [
_
E
k
_
[ < . Show that
[(E
k
)[ < .
D.4. Give the denition of a signed measure on the natural numbers N (recall
that all subsets of N are open). Show that a signed measure on N is com-
pletely determined by the values (k)
kN
, i.e., (k)
kN
is an injective
map of the signed measures on N into the space of all sequences of extended
real numbers. Identify the range of this map, and identify the sequences that
correspond to bounded measures on N.
D.3 Positive Measures and Integration
The next several sections are devoted to developing the denition of inte-
gration with respect to measures, beginning with positive measures in this
section. As the theory of integration with respect to positive measures very
closely parallels the theory of integration with respect to Lebesgue measure
that was presented in Section B.5, we shall be brief and simply state the main
results of this section without proof.
D.3.1 Basic Properties of Positive Measures
Theorem D.21. Let be a positive measure on R. Prove the following state-
ments.
(a) Monotonicity: If A, B B
and A B, then (A) (B).

(b) If A, B B
, B A, and (B) < , then (AB) = (A) (B).

(c) Countable subadditivity: If E
k
B
for k N, then
k=1
E
k
_
k=1
(E
k
).
(d) Continuity from below: If E
k
B
for k N, then
E
1
E
2
=
_

k=1
E
k
_
= lim
k
(E
k
).
(e) Continuity from above: If E
k
B
for k N, then
E
1
E
2
and (E
1
) < =
_

k=1
E
k
_
= lim
k
(E
k
).
D.3.2 Borel Measurable Functions
We need to restrict to those functions which are measurable with respect to
the Borel -algebra. These are the functions for which the inverse image of a
Borel set is a Borel set.
Denition D.22 (Borel Measurable Function).
(a) A function f : R C is Borel measurable if
E C is a Borel set = f
1
(E) R is a Borel set.
(b) A function f : R [, ] is Borel measurable if
E [, ] and E R is a Borel set
= f
1
(E) R is a Borel set.
Because the Borel -algebra is generated by the open sets, a function
f : R C is Borel measurable if f
1
(U) is a Borel set for each open set
U C. If f is real-valued, then f is Borel measurable if f
1
(a, ) is a Borel
set for every a R, and hence the denition of a Borel measurable function is
entirely analogous to the denition of a Lebesgue measurable function given
in Denition B.20. In particular, every continuous function on R is Borel
measurable.
Lemma D.23. (a) If f, g : R R are Borel measurable, then so are f + g
and fg.
(b) If f
n
: R R are Borel measurable for n N, then so are sup f
n
, inf f
n
,
limsup f
n
, and liminf f
n
. Consequently, if f(x) = lim
n
f
n
(x) exists
for each x, then f is Borel measurable.
Appropriate parts of Lemma D.23 extend to complex-valued functions,
and can also be extended to extended real-valued functions if we are careful
about .
Compositions of Borel measurable functions also behave well. If we have
two Borel measurable functions f, g : R R, then it follows directly from the
denition that g f is also Borel measurable. Generalizing the denition of
Borel measurability in the natural way to functions on C, if f : R C and
g : C C are Borel measurable, then so is g f. In particular, f
2
, [f[, [f[
p
,
etc., are all Borel measurable if f is.
D.3.3 Integration of Nonnegative Functions
Simple functions are dened just as in Denition B.30, except that now we
require our functions to be Borel measurable instead of Lebesgue measurable.
Thus, a simple function on the real line is a Borel measurable function
on R that assumes only nitely many distinct scalar values. If these distinct
values are a
1
, . . . , a
N
and we set E
k
= = a
k
, then =
N
k=1
a
k
E
k
is the
standard representation of .
Denition D.24 (Integral of Nonnegative Functions). Let be a pos-
itive Borel measure on R.
(a) If 0 is a simple function on R with standard representation =
N
k=1
a
k
E
k
, then the integral of with respect to is
_
d =
_
(x) d(x) =
N
k=1
a
k
(E
k
).
(b) If f : R [0, ] is Borel measurable, then the integral of f with respect
to is
_
f d =
_
f(x) d(x) = sup
__
d : 0 f, simple
_
.
We write
_
E
f d to mean
_
f

E
d.
If is a positive measure and a certain property holds except for a set E
with (E) = 0, then we say that this property holds -a.e.
We have the following convergence theorem for positive measures, analo-
gous to Theorem B.36 for Lebesgue measure.
Theorem D.25 (Monotone Convergence Theorem). Let be a pos-
itive Borel measure on R, and let f
n
nN
be a sequence of Borel measur-
able, nonnegative, monotone increasing functions on R. If we dene f(x) =
lim
n
f
n
(x), then
lim
n
_
f
n
d =
_
f d.
Corollary D.26. Let be a positive Borel measure on R. If f
n
nN
be a
sequence of Borel measurable, nonnegative functions on R, then
_ _

n=1
f
n
_
d =
n=1
_
f
n
d.
Just as in Exercise B.37, we can always create a sequence of simple func-
tions that increases monotonically to a given nonnegative f. Combining this
with the Monotone Convergence Theorem, we obtain the following facts.
Theorem D.27. The following properties hold for any positive Borel measure
on R and any Borel measurable functions f, g : R [0, ].
(a)
_
f d = 0 if and only if f = 0 -a.e.
(b) If f g -a.e., then
_
f d
_
g d.
(c)
_
(f +g) d =
_
f d +
_
g d.
(d) If c 0 then
_
(cf) d = c
_
f d.
(e) If A, B B
with A B, then
_
A
f d
_
B
f d.
Fatous Lemma is another useful convergence theorem.
Theorem D.28 (Fatous Lemma). Let be a positive Borel measure on R.
If f
n
nN
is a sequence of Borel measurable, nonnegative functions on R, then
_
_
liminf
n
f
n
_
d liminf
n
_
f
n
d.
D.3.4 Integration of Arbitrary Functions
Next we will extend integration with respect to a positive measure to arbi-
trary functions. As before, we write f
+
(x) = max
_
f(x), 0
_
and f
(x) =
max
_
f(x), 0
_
.
Denition D.29 (Integrable Functions). Let be a positive Borel mea-
sure on R.
(a) A Borel measurable function f : R [, ] is a (real) extended -
integrable function if at least one of
_
f
+
d,
_
f
d is nite.
(b) A Borel measurable function f : R [, ] or f : R C is -integrable
if
_
[f[ d is nite.
Denition D.30 (Integration). Let be a positive Borel measure on R.
(a) If f is a real, extended -integrable function, then we dene
_
f d =
_
f
+
d
_
f
d.
(b) If f : R C is Borel measurable and
_
Re(f) d and
_
Im(f) d both
exist and are nite, then we dene
_
f d =
_
Re(f) d +i
_
Im(f) d.
In all other cases,
_
f d is undened. Note in particular that if f is
complex-valued, then
_
f d, if it exists, will be a complex scalar. On the
other hand, if f is real-valued, then
_
f d, if it exists, can be either a nite
real scalar or .
Lemma D.31. If is a positive Borel measure on R and
_
f d exists, then
_
f d

_
[f[ d.
When dening the space L
1
() of functions that are integrable with respect
to , we have a choice between letting our functions be extended real-valued
or complex-valued. In this volume, we will take L
1
() to consist of complex-
valued -integrable functions.
Denition D.32. If is a positive Borel measure on R, then L
1
() consists
of -integrable functions f : R C. The L
1
-norm of f L
1
() is
|f|
1
=
_
[f[ d.
There are many other notations commonly used to denote L
1
(), including
L
1
(d), L
1
(R; ), or L
1
(R; d).
Note the implicit dependence of the notation |f|
1
on . When we need
to emphasize the dependence on , we will write |f|
1,
=
_
[f[ d.
Theorem D.33. If we identify functions in L
1
() that are equal -a.e., then
||
1
is a norm on L
1
(), and L
1
() is complete with respect to this norm.
The following result is one of the most useful convergence theorems.
Theorem D.34 (Dominated Convergence Theorem). Let be a pos-
itive Borel measure on R. Let f
n
nN
be a sequence of Borel measurable
functions on R such that:
(a) f
n
(x) f(x) for -a.e. x, and
(b) there exists g L
1
() such that [f
n
(x)[ g(x) -a.e. for every n.
Then f
n
converges to f in L
1
-norm, i.e.,
lim
n
|f f
n
|
1
= lim
n
_
[f f
n
[ d = 0,
and, consequently, lim
n
_
f
n
d =
_
f d.
Additional Problems
D.5. Show that if f is Borel measurable and a R, then
_
f d
a
= f(a).
Characterize L
1
(
a
).
D.6. Let be a positive Borel measure on R, and suppose that g 0 is Borel
measurable. Show that (E) =
_
E
g d is a positive Borel measure, and if
f L
1
() then
_
f d =
_
fg d.
D.7. Given a positive Borel measure on R, let S be the set of all simple
functions =
N
k=1
c
k
E
k
such that (E
k
) < for each k. Show that S is
dense in L
1
().
D.8. Given a positive Borel measure on R, show that if f
n
f in L
1
()
then there exists a subsequence such that f
n
k
f pointwise a.e.
D.9. Let be a positive measure on N. Set w(k) = k. Show that if f =
(f(k))
kN
is a nonnegative sequence of scalars, then
_
f d =
f(k) w(k).
Conclude that L
1
() =
1
w
, the weighted
1
space dened in Problem A.9.
D.4 Signed Measures and Integration
We extend integration to signed measures by making use of the Jordan de-
composition of the measure.
Denition D.35. Let be a signed Borel measure on R, and let =
+
be its Jordan decomposition. Assume that f is a Borel measurable map of R

into either [, ] or C. If
_
[f[ d
+
,
_
[f[ d
< , then we dene

_
f d =
_
f d
+
_
f d
.
Exercise D.36. Let be a signed Borel measure on R. Show that if f : R C
is Borel measurable, then
_
[f[ d
+
,
_
[f[ d
<
_
[f[ d[[ < ,
and, in this case,
_
f d

_
[f[ d[[.
Denition D.37. If is a signed Borel measure on R, then we dene L
1
() =
L
1
([[). Thus, L
1
() consists of all Borel measurable functions f : R C such
that |f|
1
=
_
[f[ d[[ < .
Note that L
1
() = L
1
(
+
) L
1
(
).
The following exercise gives an important class of signed measures related
to a given positive Borel measure .
Exercise D.38. Let be a positive Borel measure on R.
(a) Show that if f is a real, extended -integrable function, then (E) =
_
E
f d denes a signed Borel measure.
(b) Show that
+
(E) =
_
E
f
+
d,
(E) =
_
E
f
d, and [[(E) =
_
E
[f[ d.
Consequently, is bounded if f L
1
().
Notation D.39. We write d = f d to mean that is given by (E) =
_
E
f d. Sometimes it is convenient to instead write = f d.
The next exercise will show that if d = f d then is an example of a
measure that is absolutely continuous with respect to in the following sense.
Denition D.40 (Absolutely Continuous Measure). A signed Borel
measure is absolutely continuous with respect to a positive Borel measure ,
denoted , if
E B
, (E) = 0 = (E) = 0.
D.4 Signed Measures and Integration 151
Exercise D.41. Show that if the measures , are dened as in Exercise
D.38, then .
Now we come to a major structural result for -nite signed measures.
Theorem D.42 (LebesgueRadonNikodym Theorem). Let be a -
nite signed measure and let be a -nite positive measure on R.
(a) There exist unique -nite signed Borel measures , such that
= +, , .
(b) There exists a real, extended -integrable function f such that d = f d,
i.e.,
= f d +.
(c) If we also have =

f d + where

f is a real, extended -integrable
function, then

f = f -a.e.
We refer to = + the Lebesgue decomposition of with respect to .
Corollary D.43 (RadonNikodym Theorem). If is a -nite signed
measure and is a -nite positive measure such that , then there
exists a real, extended -integrable function f such that d = f d. Any two
functions which have this property are equal -a.e.
Denition D.44 (RadonNikodym Derivative). The function f given
in Corollary D.43 is called the RadonNikodym derivative of with respect
to , often denoted f =
d
d
. With this notation we have d =
d
d
d. The
RadonNikodym derivative is unique up to sets of -measure zero.
Note that if d = fd then d[[ = [f[ d by Exercise D.38.
Exercise D.45. Let be a -nite signed measure, and let , be -nite
positive measures. Show that
and = ,
and, in this case, if = f d and = g d, then = fg d.
Remark D.46. The measure is not absolutely continuous with respect to
Lebesgue measure. Assuming a willing suspension of belief for the moment,
if we did have dx (which we do not), then there would exist a function
(x) such that d = (x) dx. By Problem D.5 we know that
_
f d = f(0), so
this means that we would have
f(0) =
_
f d =
_
f(x) (x) dx.
Although there is no such function (x), it is common to abuse notation
and write
_
f(x) (x) dx = f(0), even though what is really meant with these
symbols is integration of f with respect to the measure, which in our notation
should be written as
_
f d or
_
f(x) d(x).
The following exercise suggests why the terminology absolute continuity
is used in connection with the relation .
Exercise D.47. Let be a bounded signed Borel measure and a positive
Borel measure on R. Prove that if and only if
> 0, > 0 such that E B
, (E) < = [(E)[ < .

Additional Problems
D.10. Let =
+
be the Jordan decomposition of a signed Borel measure

, and let R = P N be an associated Hahn decomposition. Show that
d
+
=

P
d, d
N
d, and d[[ = (
N
) d.
D.11. Show that if is a signed Borel measure, then
[[(E) = sup
_
_
E
f d
: [f[ 1
_
.
D.12. Show that if is a signed Borel measure and a positive Borel measure
such that and , then = 0.
D.13. Given a signed Borel measure and a positive Borel measure , show
that
[[
+
,
.
D.14. Let denote counting measure on R (see Exercise D.9).
(a) Prove that if f : R [0, ] is Borel measurable, then
_
f d = sup
_
N
j=1
f(x
j
) : N N, x
j
R
_
.
(b) Prove that dx , but dx ,= f d for any function f.
(c) Prove that has no Lebesgue decomposition with respect to dx, i.e., there
do not exist signed measures and such that = + , dx, and
dx.
D.5 Complex Measures
Next we expand the class of measures by allowing them to be complex-valued.
Denition D.48 (Complex Measure). A function : B
C is a complex
Borel measure on R, or simply a complex measure, if
(a) () = 0,
D.5 Complex Measures 153
(b) If E
1
, E
2
, . . . are nitely or countably many disjoint Borel sets, then
k
E
k
_
=
k
(E
k
).
Note that for a complex measure we have [(E)[ < for any Borel
set E, and in fact the following exercise shows that every complex measure is
bounded.
Exercise D.49. Let be a complex Borel measure. For E B
, dene
r
(E) = Re((E)) and
i
(E) = Im((E)). Show that
r
,
i
are bounded
signed measures. Show that for any E B
we have
[(E)[ [
r
(E)[ + [
i
(E)[ [
r
[(R) +[
i
[(R),
and conclude that is bounded in the sense that sup
EB
[(E)[ < .
We refer to the measures
r
,
i
as the real and imaginary parts of ,
respectively.
Denition D.50. If is a complex Borel measure on R, then we dene
L
1
() = L
1
(
r
) L
1
(
i
) = L
1
([
r
[) L
1
([
i
[). If f L
1
(), then we dene
_
f d =
_
f d
r
+i
_
f d
i
.
Denition D.51 (Absolutely Continuous Measure). A complex Borel
measure is absolutely continuous with respect to a positive Borel measure ,
denoted , if
E B
, (E) = 0 = (E) = 0.
Equivalently, if we have both
r
and
i
.
Exercise D.52. Let be a positive Borel measure on R, and suppose that
g L
1
(). Show that if we dene d = g d (i.e., (E) =
_
E
g d for E B
),
then is a complex Borel measure and .
Denition D.53 (Singular Measures). A complex Borel measure is sin-
gular with respect to another complex Borel measure , denoted , if
r

r
,
r

i
,
i

r
,
i

i
.
Next we give an important structure theorem for complex measures (com-
pare Theorem D.42 for the case of signed measures).
Theorem D.54 (LebesgueRadonNikodym Theorem). Let be a
complex Borel measure on R, and let be a -nite positive measure on R.
Then there exists an f L
1
() and a complex Borel measure such that
= f d +, . (D.2)
If we also have =

f d +

where

f L
1
(d) and

, then

= and
f = f -a.e.
We will need the following exercise in order to dene the total variation
of a complex measure.
Exercise D.55. Let be a complex Borel measure on R, and dene =
[
r
[ + [
i
[, so is a positive bounded Borel measure. Show there exists a
function f L
1
() such that d = f d.
The total variation of a complex measure is a little more awkward to
dene than it is for a signed measure. By Exercise D.55, we know that if is
a complex measure, then there exists at least one positive measure and one
function f L
1
() such that d = f d. We will dene the total variation
of to be the measure d[[ = [f[ d, but of course we need to know that
this is well-dened. The following theorem shows that this denition is indeed
independent of the choice of and f.
Theorem D.56. Let be a complex Borel measure on R. If
1
,
2
are
bounded positive measures and f
1
L
1
(
1
), f
2
L
1
(
2
) are such that
f
1
d
1
= d = f
2
d
2
, then [f
1
[ d
1
= [f
2
[ d
2
.
Proof. Dene =
1
+
2
. Then since
1
, there exists a function g
1

L
1
() such that d
1
= g
1
d. Likewise, there exists g
2
L
1
() such that
d
2
= g
2
d. Because
1
,
2
, 0, we have g
1
, g
2
0 -a.e.
Thus, we have d = f
1
d
1
and d
1
= g
1
d. Exercise: Show that Ex-
ercise D.45 generalizes to complex measures, and use this to show that
d = f
1
g
1
d, and similarly d = f
2
g
2
d (see also Problem D.6).
The uniqueness statement in the LebesgueRadonNikodym Theorem
therefore implies that f
1
g
1
= f
2
g
2
-a.e. Consequently,
[f
1
[ g
1
= [f
1
g
1
[ = [f
2
g
2
[ = [f
2
[ g
2
-a.e.,
and hence
[f
1
[ d
1
= [f
1
[ g
1
d = [f
2
[ g
2
d = [f
2
[ d
2
.
Denition D.57 (Total Variation of a Complex Measure). Let be a
complex Borel measure on R. Then the total variation [[ of is the positive
measure d[[ = [f[ d, where is any positive measure and f is any function
in L
1
() such that d = f d.
Next we give some properties of complex measures.
Exercise D.58. Let be a complex Borel measure on R. Show that the
following statements hold.
(a) [(E)[ [[(E) for all E B
.
(b) [[, and there exists g with [g[ = 1 [[-a.e. such that d = g d[[.
(c) If f L
1
(), then
_
f d

_
[f[ d[[.
D.5 Complex Measures 155
The representation d = g d[[ in part (b) of the preceding exercise is
called the polar decomposition of .
The following equivalent reformulations of the total variation measure are
often easier to employ in practice than Denition D.57.
Exercise D.59. Let be a complex Borel measure on R. Prove the following
equivalent characterizations of [[.
(a) [[(E) = sup
_
n
k=1
[(E
k
)[ : n N, E
k
B
, E =
n
k=1
E
k
disjointly
_
.
(b) [[(E) = sup
_

k=1
[(E
k
)[ : E
k
B
, E =
k=1
E
k
disjointly
_
.
(c) [[(E) = sup
_
_
E
f d
: [f[ 1 [[-a.e.
_
.
Exercise D.60. Show that if is a complex measure, then E B
is a null
set for if and only if [[(E) = 0.
For a complex measure , we say that a property holds -almost everywhere
if it holds except possibly on a null set for . Thus -almost everywhere is
the same as [[-almost everywhere.
The space M
b
(R) of all complex Borel measures is a Banach space.
Denition D.61 (Space of Complex Borel Measures). We set
M
b
(R) =
_
: is a complex Borel measure on R
_
,
and dene the norm of a complex measure to be
|| = [[(R). (D.3)
Exercise D.62. Show that | | as dened in equation (D.3) is a norm on
M
b
(R), and that M
b
(R) is a Banach space with respect to this norm.
We identify some particular subspaces of M
b
(R).
Exercise D.63. Show that if be a positive Borel measure on R and d =
f d where f L
1
(), then || = |f|
1
=
_
[f[ d.
Consequently, if is a positive measure then we have L
1
() M
b
(R).
More precisely, if we dene d
g
= g d for each g L
1
(), then Exercise D.63
shows that |
: g
g
is an isometric embedding of L
1
() into M
b
(R). If
is -nite, then range(|
) = M
b
(R) : . In particular, Lebesgue
measure is a positive Borel measure, and hence if we identify f L
1
(R) with
f dx M
b
(R) then we have L
1
(R) M
b
(R).
Additional Problems
D.15. Let be a positive Borel measure on R, and suppose that g L
1
().
Dene = g d. Show that if f L
1
() then
_
f d =
_
fg d.
D.16. Show that if is a complex Borel measure, then
[
r
[, [
i
[ [[ [
r
[ +[
i
[ =
+
r
+
r
+
+
i
+
i
,
and show that L
1
() = L
1
([[).
D.17. Show that if is a complex Borel measure and (R) = [[(R), then
= [[.
D.18. Let be a complex Borel measure and a positive Borel measure.
Show that if and only if [[ .
D.19. Let be a complex Borel measure. Dene the complex conjugate of
by (E) = (E) for E B
. Show that is a complex measure, L

1
() =
L
1
( ), and
_
f(x) d (x) =
_
f(x) d(x) for all f L
1
().
D.20. Let be a complex Borel measure. Dene the involution of by
(E) = (E) for E B
. Show that is a complex measure, f L

1
() if and
only if f(x) L
1
( ), and
_
f(x) d (x) =
_
f(x) d(x) for all f L
1
().
D.21. Extend the Dominated Convergence Theorem to complex Borel mea-
sures: Show that if M
b
(R), f
n
, f are Borel measurable, f
n
(x) f(x)
for -a.e. x, and there exists g L
1
() such that [f
n
(x)[ g(x) -a.e., then
_
(f f
n
) d

_
[f f
n
[ d[[ 0.
D.22. Given a complex measure on N, nd an explicit description of [[.
D.23. Show that (k)
kN
is an isometric isomorphism of M
b
(N) onto
1
(N). Thus M
b
(N)

=
1
(N). Compare this to Problem D.29.
D.6 Fubini and Tonelli for Borel Measures
In this section we will state Fubinis and Tonellis theorems for Borel measures.
The denition of measurability on R
2
is analogous to the denition for R.
Denition D.64. (a) The Borel -algebra on R
2
is the smallest -algebra of
subsets of R
2
that contains all the open subsets of R
2
.
(b) A function F : R
2
C is Borel measurable if
E C is a Borel set = F
1
(E) R
2
is a Borel set.
D.6 Fubini and Tonelli for Borel Measures 157
(c) A function F : R
2
[, ] is Borel measurable if
E [, ] and E R is a Borel set
= F
1
(E) R
2
is a Borel set.
Tonellis Theorem applies to all -nite positive measures.
Theorem D.65 (Tonellis Theorem). Let , be -nite positive Borel
measures on R. If F : R
2
[0, ] is Borel measurable, then the following
statements hold.
(a) F
x
(y) = F(x, y) is Borel measurable for every x R.
(b) F
y
(x) = F(x, y) is Borel measurable for every y R.
(c) g(x) =
_
F
x
(y) d(y) is Borel measurable.
(d) h(y) =
_
F
y
(x) d(x) is Borel measurable.
(e)
_ __
F(x, y) d(x)
_
d(y) =
_ __
F(x, y) d(y)
_
d(x).
Fubinis Theorem for positive measures is as follows, see [Fol99, Thm. 7.27].
Theorem D.66 (Fubinis Theorem). Let , be -nite positive measures
on R. If F : R
2
[, ] or F : R
2
C is Borel measurable and
__
[F(x, y)[ d(x) d(y) < , (D.4)
then the following statements hold.
(a) F
x
(y) = F(x, y) is Borel measurable and -integrable for -a.e. x R.
(b) F
y
(x) = F(x, y) is Borel measurable and -integrable for -a.e. y R.
(c) g(x) =
_
F
x
(y) d(y) is Borel measurable and -integrable.
(d) h(y) =
_
F
y
(x) d(x) is Borel measurable and -integrable.
(e)
_ __
F(x, y) d(x)
_
d(y) =
_ __
f(x, y) d(y)
_
d(x).
Although stated for positive measures, Fubinis Theorem extends to signed
and complex measures. Suppose that , are -nite signed measures and
F : R
2
[, ] is a Borel measurable function that satises
__
[F(x, y)[ d[[(x) d[[(y) < . (D.5)
Then by breaking and into positive and negative parts and applying
Fubinis theorem to each of those parts, we can see that the conclusions of
Fubinis Theorem hold for F. Likewise, if , are complex measures (hence
bounded) and F : R
2
C is measurable, then by breaking into real and
imaginary parts we see that the conclusions of Fubinis Theorem hold if F
satises equation (D.5).
D.7 Radon Measures
In this section we will discuss Radon measures on the real line. Radon mea-
sures can be dened on any locally compact Hausdor space, but because we
are only dealing with the real line, certain simplications occur.
Denition D.67 (Radon Measures). Let be a positive Borel measure
on R.
(a) is outer regular on E B
if (E) = inf
_
(U) : U E, U open
_
.
(b) is inner regular on E B
if (E) = sup
_
(K) : K E, K compact
_
.
(c) If is both inner and outer regular on every Borel set, then is regular.
(d) is locally nite if (K) < for every compact set K R.
(e) is a Radon measure on R if
i. is locally nite,
ii. is outer regular on all Borel sets, and
iii. is inner regular on every open set.
Although every Radon measure is positive by denition, we sometimes
emphasize this fact by writing positive Radon measure.
On the real line, the distinction between Borel and Radon measures is not
very signicant. In fact, we will see that is a Radon measure on R if and
only if is a locally nite positive Borel measure on R (counting measure is
an example of a positive Borel measure that is not locally nite). However, on
domains other than Euclidean space, the distinction becomes more important,
and we refer to [Fol99] for complete details.
The rst step in showing the equivalence of Radon measures with locally
nite Borel measures is the following result.
Theorem D.68. Every Radon measure on R is regular.
Proof. By denition, a Radon measure is outer regular, so we just have to
show that it is inner regular on every Borel set.
Suppose rst that E is a Borel set with (E) < , and choose any > 0.
Since is outer regular on E, there exists an open U E such that (E)
(U) < (E) + . As U is open and is inner regular on open sets, there
exists a compact set F U such that (F) > (U) .
Now, since E has nite measure, we have (UE) = (U) (E) < .
Also, is outer regular on UE, so there exists an open V UE such that
(V ) < .
Set K = FV. Then K is compact, K E, and
(K) = (F) (F V ) > (U) (V ) > (E) 2.
Hence is inner regular on any Borel set E with nite measure.
Now suppose that E is a Borel set with (E) = . Dene E
k
= E[k, k].
Then since is locally nite, we have (E
k
) [k, k] < . Further, E
1

D.8 The Riesz Representation Theorem for Positive Linear Functionals on Cc(R) 159
E
2
and E = E
k
, so lim
k
(E
k
) = (E) = by Theorem D.21.
Hence given R > 0, there exists a k with (E
k
) > R. Since is inner regular
on E
k
, there exists a compact K E
k
such that (K) > R. Hence
sup
_
_
= = (E),
so is inner regular on E.
Remark D.69. The fact that the real line is -compact, i.e., is a union of count-
ably many compact sets, is clearly an important ingredient of the preceding
proof. On a general space, a Radon measure will be inner regular on any
subset that is -nite, see [Fol99].
We state the following useful property of Radon measures without proof.
Theorem D.70 (Lusins Theorem). Let be a Radon measure on R. If
f : R C is Borel measurable and (supp(f)) < , then for every > 0,
there exists C
c
(R) such that
_
f ,=
_
< .
Further, if f is bounded, then can be constructed so that
sup
xR
[(x)[ sup
xR
[f(x)[.
Additional Problems
D.24. Show that if is a Radon measure then C
c
(R) L
1
().
D.25. Since every subset of N is open, any positive measure on N is a Borel
measure. Show that is locally nite if and only if k < for every k N,
and that the Radon measures on N are precisely the locally nite positive
measures on N.
D.8 The Riesz Representation Theorem for Positive
Linear Functionals on C
c
(R)
In this section we will discuss one version of the Riesz Representation Theo-
rem, which shows the equivalence between Radon measures and positive linear
functionals on C
c
(R). This version is only concerned with positive measures
and positive functionals, while in Section D.10 we will see a Riesz Represen-
tation Theorem for complex Radon measures and bounded functionals.
In this section we deal both with measures and functionals. Typically, we
will let denote a functional and a measure.
Notation D.71. In keeping with the notations introduced in Appendix C
(see Notation C.36), we write f, ) to denote the action of a linear functional
: C
c
(R) C on a vector f C
c
(R). Further, f, ) is a sesquilinear form,
linear in f but antilinear in .
Each Radon measure on R induces an associated linear functional on
C
c
(R) by the formula
f, ) =
_
f d, f C
c
(R). (D.6)
Note that, by denition, is a positive measure here. In order to ensure that
, ) is a sesquilinear form, whenever we extend our consideration to complex
measures we will need to replace d by d in equation (D.6).
This example immediately raises several questions, which we will address in
this section. First, is the functional dened in equation (D.6) continuous on
C
c
(R)? Of course, continuity is not even dened until we specify the topology
on C
c
(R), and, as it turns out, there is more than one natural choice.
Second, once we specify the topology on C
c
(R), does every continuous
linear functional on C
c
(R) have the form given in equation (D.6)? In other
words, can we characterize the dual space of C
c
(R)? This question also requires
some renement, since we have specied that Radon measures are positive
measures, whereas if we let be a complex measure then we can still dene a
functional by equation (D.6).
To address these questions, we next discuss two particular topologies on
C
c
(R).
D.8.1 Topologies on C
c
(R)
Since we wish to study the continuity of linear functionals on C
c
(R), we must
specify a topology or a convergence criterion on C
c
(R). There are two natural
choices.
(a) The uniform (or L
-norm) topology. C
c
(R) is a normed space with respect
to the topology induced by the uniform, or L
, norm. A linear functional

on C
c
(R) is continuous with respect to the uniform topology if and only
if it is bounded with respect to the L
norm. That is, is continuous if

and only if there exists a constant C > 0 such that
[f, )[ C |f|
, all f C
c
(R).
Since C
c
(R) is a dense subspace of the Banach space C
0
(R), such a has
a unique extension to a bounded linear functional on all of C
0
(R), which
we also refer to as (see Exercise C.26).
(b) The inductive limit topology. For each compact K R, dene
C(K) =
_
f C
c
(R) : supp(f) K
_
.
D.8 The Riesz Representation Theorem for Positive Linear Functionals on Cc(R) 161
Each C(K) is a Banach space with respect to the L
-norm. Further, as
a set,
C
c
(R) =
_
C(K) : K R, K compact
_
.
We can dene a topology on C
c
(R) by declaring that a function whose
domain is C
c
(R) is continuous if for each compact K its restriction to
C(K) is continuous with respect to the L
-norm on C(K). In particular, a

linear functional : C
c
(R) C is continuous with respect to this topology
if and only if for each compact K the restriction [
C(K)
: C(K) C is
continuous. Since C(K) is a normed space, this happens if and only if
each [
C(K)
is bounded with respect to the norm on C(K), which means
that for each compact K there exists a constant C
K
> 0 such that
[f, )[ C
K
|f|
, all f C(K). (D.7)

However, unlike boundedness with respect to the uniform topology, where
there is a single constant C that determines the boundedness, the con-
stants C
K
in equation (D.7) can depend on the compact set K. In technical
language, this topology corresponds to the inductive limit of the topolo-
gies (C(K), | |
) with K compact, and hence we will refer to it as the

inductive limit topology on C
c
(R). We refer to [Con90] for details on the
inductive limit of topologies. This type of topology is also discussed in
Section E.5.
The following denition of the convergence criterion on C
c
(R) corresponds
to convergence with respect to each of these two topologies.
Denition D.72. Let f
n
nN
be a sequence of functions in C
c
(R).
(a) We say that f
n
converges to f uniformly, or in L
-norm, if |f f
n
|

0. In this case, we write f
n
f uniformly.
(b) We say that f
n
converges to f in C
c
(R) if there exists a compact set K
such that supp(f
n
) K for all n, and |f f
n
|
0. In this case, we
write f
n
f in C
c
(R).
In particular, we have that
f
n
f in C
c
(R) = f
n
f uniformly . (D.8)
However, the converse implication does not hold in general, so these are two
distinct topologies on C
c
(R). Equation (D.8) implies that the uniform topol-
ogy on C
c
(R) is weaker than the inductive limit topology.
In this section we are focusing on Radon measures (which by denition
are positive but possibly unbounded) and corresponding positive linear func-
tionals on C
c
(R). For these results it is the inductive limit topology on C
c
(R)
that will be important. In contrast, in Section D.10 we will consider complex
Radon measures (which are necessarily bounded) and corresponding linear
functionals on C
c
(R), and there it will be the L
-topology on C
c
(R) that will
be important.
D.8.2 Positive Linear Functionals on C
c
(R)
The next exercise shows that every Radon measure induces a linear functional
on C
c
(R) that is continuous with respect to the inductive limit topology on
C
c
(R). Further, this functional is positive in the following sense.
Denition D.73. A functional : C
c
(R) C is positive if f, ) 0 for all
f C
c
(R) with f 0.
Exercise D.74. Let be a Radon measure on R. Dene : C
c
(R) C by
f, ) =
_
f d, f C
c
(R).
(a) Show that is a positive linear functional on C
c
(R).
(b) Show that [
C(K)
: C(K) R is continuous for every compact K R,
i.e.,
compact K R, C
K
> 0 such that
f C(K) = [f, )[ C
K
|f|
.
(D.9)
Thus, those positive linear functionals on C
c
(R) that are induced from
Radon measures are continuous with respect to the inductive limit topology
on C
c
(R). Next we will show directly every positive linear functional on C
c
(R)
is continuous with respect to the inductive limit topology on C
c
(R).
Theorem D.75. If : C
c
(R) C is a positive linear functional on C
c
(R),
then is continuous on C
c
(R) with respect to the inductive limit topology.
Specically, [
C(K)
: C(K) C is continuous for each compact K R.
Proof. Given a compact set K, by Urysohns Lemma (Theorem A.109), we
can nd
K
C
c
(R) such that
K
0 and
K
= 1 on K.
Suppose rst that f C(K) is real-valued. Then
[f(x)[ = [f(x)[
K
(x) |f|
K
(x)
for all x R. Therefore |f|
K
f 0, and so
0
|f|
K
f,
_
= |f|
K
, ) f, ).
Consequently,
[f, )[
K
, ) |f|
.
Second, given an arbitrary f C(K), we have
[f, )[ [Re(f), )[ +[Im(f), )[ 2
K
, ) |f|
.
Hence the result follows with C
K
= 2
K
, ).
D.9 The Relation Between Radon and Borel Measures on R 163
Although we will not prove it, the Riesz Representation Theorem com-
pletes the characterization of positive linear functionals on C
c
(R): Every pos-
itive linear functional on C
c
(R) is induced from a Radon measure.
Theorem D.76 (Riesz Representation Theorem I). If : C
c
(R) C
is a positive linear functional, then there exists a unique positive Radon mea-
sure on R such that
f, ) =
_
f d, f C
c
(R).
Moreover, if U R is open, then
(U) = sup
_
f, ) : f C
c
(R), 0 f 1, supp(f) U
_
,
and if K R is compact then
(K) = inf
_
f, ) : f C
c
(R), f

K
_
.
Thus, Radon measures and positive linear functionals on C
c
(R) are equiva-
lent. Therefore, we often use the same symbol to represent a Radon measure
and the positive functional f f, ) =
_
f d that it induces.
Additional Problems
D.26. This problem will show that the locally nite positive measures on N
(which by Problem D.25 are precisely the Radon measures on N) are in 1-1
correspondence with the positive linear functionals on c
00
.
(a) Give the convergence criterion corresponding to the inductive limit
topology on c
00
.
(b) Show that if is a positive locally nite measure on N, then f, ) =
f(k) k denes a positive linear functional on c

00
that is continuous with
respect to the inductive limit topology on c
00
.
(c) Show that if is a positive linear functional on c
00
then there exists
a unique sequence of nonnegative scalars w = (w
k
)
kN
such that f, ) =
f(k) w
k
for f c
00
. Show there is a unique locally nite positive measure
on N such that w
k
= k for every k.
D.9 The Relation Between Radon and Borel Measures
on R
We will use the Riesz Representation Theorem (Theorem D.76) to show that
every locally nite positive Borel measure on R is a Radon measure (the
converse holds by denition). First we need the following lemma.
Lemma D.77. If is a -nite Radon measure and E B
, then for every

> 0 there exists an open set U and a closed set F such that
F E U and (UF) < .
Proof. Since is -nite, there exist disjoint E
k
B
with (E
k
) < such
that E = E
k
. Since is outer regular, there exist open U
k
E
k
such that
(U
k
E
k
) < 2
k1
. Then U = U
k
is open, U E, and
(UE)
k
(U
k
E
k
)

2
.
Since E
C
= RE is a Borel set, we can apply the same argument to E
C
to nd an open set V E
C
with (V E
C
)

2
. Hence F = V
C
is closed,
F E, and (EF) = (V E
C
)

2
. Finally,
(UF) (UE) +(EF)

2
+

2
= .
Now we can complete our characterization of Radon measures on the real
line.
Theorem D.78. The class of Radon measures on R coincides with the class
of locally nite positive Borel measures on R.
Proof. By denition, if is a Radon measure, then it is a locally nite positive
Borel measure.
Conversely, suppose that is a locally nite positive Borel measure. We
will show that is regular, and hence is a Radon measure. Since C
c
(R)
L
1
(), we can dene f, ) =
_
f d for f C
c
(R), and this denes a pos-
itive linear functional on C
c
(R). The Riesz Representation Theorem (Theo-
rem D.76) therefore implies that there exists a Radon measure such that
f, ) = f, ) for f C
c
(R).
Suppose that U is any open subset of R. Then we can write U =
j=1
K
j
where each K
j
is compact. We claim that there exist f
n
C
c
(R) with 0
f
n
1 and supp(f
n
) U such that f
n
= 1 on
n
j=1
K
j
and on
n1
j=1
supp(f
n
).
To prove this, we proceed by induction. For n = 1, Urysohns Lemma
(Theorem A.109) implies that there exists a function f
1
C
c
(R) that satises
0 f
1
1, supp(f
1
) U, and f
1
= 1 on K
1
.
Assume now that f
1
, . . . , f
n
have been constructed satisfying the required
properties. Then since
F =
_
n+1
j=1
K
j
_

_
n
j=1
supp(f
j
)
_
is a compact subset of U, by Urysohns Lemma we can nd f
n+1
C
c
(R)
such that 0 f
n+1
1, supp(f
n+1
) U, and f
n+1
= 1 on F. This completes
the induction.
D.9 The Relation Between Radon and Borel Measures on R 165
By construction, the sequence f
n
nN
is a monotone increasing sequence
of functions, and f
n

U
pointwise. Hence by the Monotone Convergence
Theorem (Theorem D.25), which applies to both and , we have
(U) =
_

U
d = lim
n
_
f
n
d = lim
n
_
f
n
d =
_

U
d = (U).
Thus and agree on all the open sets.
Now let E be any Borel set, and choose > 0. Then Lemma D.77 implies
that there exist an open U and a closed F with F E U and (UF) < .
Since UF is open, and assign it the same measure, so we have (UF) <
as well. Therefore
(U) = (UF) +(F) +(E).
Thus (E) = inf
_
(U) : U E, U open
_
, so is outer regular on every Borel
set.
Also, we have
(E) = (F) +(EF) (F) +.
Although F need not be compact, if we dene F
k
= F [k, k] then F
k
is compact and (F
k
) (F). If (E) < , then there exists a k such
that (F
k
) (F) , and hence (F
k
) (E) 2. If (E) = , then
(F) = as well, and so (F
k
) . In either case, we conclude that
(E) = sup
_
_
, so is inner regular on every Borel
set.
Thus is regular, and hence is a Radon measure. In fact, by the uniqueness
statement in the Riesz Representation Theorem, we actually have = .
Corollary D.79. The following statements are equivalent.
(a) is a locally nite positive Borel measure on R.
(b) is a regular locally nite positive Borel measure on R.
(c) is a Radon measure on R.
The following statements are also equivalent.
(a) is a bounded positive Borel measure on R.
(b) is a regular bounded positive Borel measure on R.
(c) is a bounded Radon measure on R.
More general domains on which the class of complex Borel measures coin-
cides with the class of complex Radon measures are discussed in [Fol99].
D.10 The Dual of C
0
(R)
In this section we will see that the dual space of C
0
(R) can be identied with
the space of complex Radon measures on the real line.
Radon measures, as discussed so far, are positive by denition. We extend
the denition to signed and complex measures as follows. As usual,
+
,
denote the positive and negative parts of a signed measure , and

r
,
i
denote
the real and imaginary parts of a complex measure .
Denition D.80. A signed Borel measure on R is a signed Radon measure
on R if
+
,
are Radon measures.

A complex Borel measure on R is a complex Radon measure on R if
r
,
i
are signed Radon measures.
Because of the properties of the real line, these notions simplify as follows.
Lemma D.81. The following statements are equivalent.
(a) is a bounded signed Borel measure on R.
(b) is a bounded signed Radon measure on R.
The following statements are also equivalent.
(a) is a complex Borel measure on R.
(b) is a complex Radon measure on R.
Proof. A measure is a bounded signed Borel measure if and only if
+
,
are bounded positive Borel measures. By Theorem D.78, this happens if and
only if
+
,
are bounded Radon measures, which is equivalent to being a

bounded signed Radon measure.
A similar argument applies to complex measures, noting that all complex
measures are bounded.
Consequently, the Banach space M
b
(R) of all complex Borel measures on R
introduced in Denition D.61 coincides with the space of all complex Radon
measures on R. For other domains, the distinction between these two spaces
becomes important.
The following exercise shows that if is a complex Radon measure (which
is necessary bounded), then it induces a linear functional on C
c
(R) that is con-
tinuous with respect to the uniform topology, and hence this linear functional
extends to a continuous linear functional on C
0
(R).
Exercise D.82. Assume is a complex Radon measure, and let be the
complex conjugate measure dened in Problem D.19. Dene a functional
on C
c
(R) by
f, ) =
_
f d , f C
c
(R),
and prove the following statements.
D.10 The Dual of C0(R) 167
(a) is bounded on C
c
(R) with respect to the L
-norm, and furthermore

[f, )[ |f|
||, f C
c
(R), (D.10)
where || = [[(R) is the norm of the measure .
(b) The operator norm of is || = [[(R) = ||.
(c) extends to a bounded linear functional on C
0
(R), also dened by f, ) =
_
f d for f C
0
(R).
Although we will not prove it, the Riesz Representation Theorem states
that every bounded linear functional is induced from a complex Radon mea-
sure, as follows.
Theorem D.83 (Riesz Representation Theorem II). Given M
b
(R),
dene
: C
0
(R) C by
f,
) =
_
f d , f C
0
(R).
Then | :
is an antilinear isometry of M
b
(R) onto C
0
(R)
.
Thus, C
0
(R)

= M
b
(R). We often write C
0
(R)
= M
b
(R), meaning equal-
ity in the sense of the identication given in Theorem D.83. Since C
c
(R) is
dense in C
0
(R) in the uniform topology, this implies that C
c
(R)

= M
b
(R)
with respect to the uniform topology on C
c
(R). Therefore, a complementary
development of complex Radon measures could have started by declaring a
complex Radon measure to be an element of the dual space of C
c
(R) or C
0
(R)
with respect to the uniform topology, and then showing that every such linear
functional induces a corresponding complex measure.
Although we will not discuss this further, we close this appendix by com-
menting that the space of unbounded complex Radon measures can be dened
to be the dual space of C
c
(R) with respect to the inductive limit topology on
C
c
(R). Indeed, Theorem D.76 shows that the positive Radon measures corre-
spond exactly to the positive linear functionals on C
c
(R) that are continuous
with respect to this inductive limit topology.
Additional Problems
D.27. Show directly that if is an unbounded Radon measure on R, then
there exist f
n
C
c
(R) with f
n
0 and |f
n
|
1 such that f
n
, ) as
n .
D.28. Given f
n
, f C
0
(R), show that f
n
w
f if and only if f
n
(x) f(x)
pointwise for each x R and sup |f
n
|
< (weak convergence of sequences

is discussed in Section C.16).
D.29. Given g = (g
k
)
kN

1
, dene
g
: c
0
C by f,
g
) =
f
k
g
k
.
Show that g
g
is an antilinear isometric isomorphism of
1
onto c
0
. Thus
c
0

=
1
. Compare this to Problem D.23, which showed that M
b
(N)

=
1
.
Thus we see directly that the Riesz Representation Theorem holds for c
0
, i.e.,
c
0

= M
b
(N).
E
Topological Vector Spaces
Many of the important vector spaces in analysis have topologies that are
generated by a family of seminorms instead of a given metric or a norm.
We will consider these types of topologies in this appendix. References for
the material in this appendix include the texts by Conway [Con90], Folland
[Fol99], or Rudin [Rud91].
E.1 Motivation and Examples
If X is a metric space then every open subset of X is, by denition, a union of
open balls. The set of open balls is an example of a base for the topology on X
(see Denition E.9). If X is also a vector space, it is usually very important to
know whether these open balls are convex. This is certainly true if the metric
is induced from a norm, but it is not true in general.
Example E.1. Exercise B.57 tells us that if E R and 0 < p < 1, then L
p
(E)
is a complete metric space with respect to the metric d(f, g) = |f g|
p
p
. Be-
cause the operations of vector addition and scalar multiplication are continu-
ous on L
p
(E), we call L
p
(E) a topological vector space (see Denition E.12).
Unfortunately, the unit ball in L
p
(E) is not convex when p < 1 (compare the
illustration in Figure E.1). In fact, it can be shown that L
p
(E) contains no
nontrivial open convex subsets, so we cannot get around this issue by substi-
tuting some other open sets for the open balls. When p < 1, there is no base
for the topology on L
p
(E) that consists of convex sets.
For 0 < p < 1, the topology on L
p
(E) is generated by a metric, but this
metric is not induced from a norm or a family of seminorms. We will see that
topologies that are generated from families of seminorms do have a base that
consists of convex open sets. If the family of seminorms is nite, then we can
nd a single norm that induces the same topology. If the family of seminorms
is countable, then we can nd a single metric that induces the same topology.
170 E Topological Vector Spaces
1.0 0.5 0.5 1.0
1.0
0.5
0.5
1.0
1.0 0.5 0.5 1.0
1.0
0.5
0.5
1.0
1.0 0.5 0.5 1.0
1.0
0.5
0.5
1.0
1.0 0.5 0.5 1.0
1.0
0.5
0.5
1.0
Fig. E.1. Unit circles in R
2
with respect to various metrics. Top left:
1/2
. Top
right:
1
. Bottom left:
2
. Bottom right:
.
A signicant advantage of having a topology induced from a metric is that the
corresponding convergence criterion can be dened in terms of convergence of
ordinary sequences instead of convergence of nets (see Section A.6).
In this section we will give several examples of spaces whose topologies
are induced from families of seminorms, indicating without proof some of the
special features that we will consider in more detail in the following sections.
In these examples, we specify a convergence criterion instead of directly spec-
ifying a topology. The connection between convergence criteria and topologies
is reviewed in Section A.6 and will be expanded on below.
The simplest examples are those where the family of seminorms is nite.
Example E.2. Given m 0, Exercise A.21 shows that the space C
m
b
(R) is a
Banach space with respect to the norm
|f|
C
m
b
= |f|
+|f
+ +|f
(m)
|
.
The natural associated family of seminorms is
n
m
n=0
, where
n
(f) = |f
(n)
|
, f C
m
b
(R).
Note that
0
is a norm on C
m
b
(R), while
1
, . . . ,
m
are only seminorms. How-
ever, while C
m
b
(R) is complete with respect to the norm | |
C
m
b
, if m > 0 then
E.1 Motivation and Examples 171
it is not complete with respect to the norm
0
. We need all of the seminorms
to dene the correct topology. Since there are only nitely many, we usually
combine these seminorms to form the norm | |
C
m
b
.
The convergence criterion dened by the norm | |
C
m
b
is:
f
k
f in C
m
b
(R) |f f
k
|
C
m
b
0.
The same convergence criterion dened in terms of the family of seminorms
working in concert is:
f
k
f in C
m
b
(R)
n
(f f
k
) 0 for each n = 0, . . . , m.
Things become more interesting when there are innitely many seminorms
in the family. In this case, we usually cannot create a single norm that denes
the same topology.
Example E.3. The topology on the space C
b
(R) is dened by the countable
family of seminorms
n
(f)
n=0
, where
n
(f) = |f
(n)
|
, f C
b
(R).
The corresponding convergence criterion is joint convergence with respect to
all of the seminorms:
f
k
f in C
b
(R)
n
(f f
k
) 0 for each n = 0, 1, . . . .
Because there are only countably many seminorms, we will see that this topol-
ogy is induced from the metric
d(f, g) =
n=0
2
n
|f
(n)
g
(n)
|
1 +|f
(n)
g
(n)
|
.
While the topology on L
p
(E), p < 1, is also induced from a metric, that
topology has no base consisting of convex open sets, whereas the topology on
C
b
(R) does have a natural base consisting of convex open sets.
Another example of a space whose topology is generated by a countable
family of seminorms is the Schwartz space. This space plays an important role
in harmonic analysis.
Example E.4. The Schwartz space is
o(R) =
_
f C
(R) : x
m
f
(n)
(x) L
(R) for all m, n 0

_
.
The natural associated family of seminorms is
mn
(f)
m,n0
, where
mn
(f) = |x
m
f
(n)
(x)|
, f o(R).
Convergence in the Schwartz space is dened by
f
k
f in o(R)
mn
(f f
k
) 0 for all m, n 0.
As there are only countably many seminorms, there does exist a metric that
induces the same topology.
Sometimes, an uncountable family of seminorms can be reduced to a count-
able family.
Example E.5. The space L
1
loc
(R) consisting of all locally integrable functions
on R was introduced in Denition B.62. A natural associated family of semi-
norms is
K
: compact K R, where
K
(f) = |f
K
|
1
=
_
K
[f(x)[ dx, f L
1
loc
(R).
Since each compact set K R is contained in some interval [N, N], the
same topology is determined by the countable family of seminorms
N
NN
,
where
N
(f) = |f
[N,N]
|
1
=
_
N
N
[f(x)[ dx, f L
1
loc
(R).
Since we can reduce to countably many seminorms, this topology is also in-
duced from an associated metric.
The next example combines several of the features of the preceding ones.
Example E.6. Functions in C
(R) need not be bounded, so we cannot create

a topology on C
(R) by using the seminorms from Example E.3. Instead, the

topology is induced from the family
K,n
: n 0, compact K R, where
K,n
(f) = |f
(n)
K
|
, f C
(R).
As in Example E.5, we can generate the same topology using a countable
family of seminorms. T,his space with this topology is often denoted by
c(R) = C
(R).
Not every family of seminorms can be reduced to a countable family.
Example E.7 (The Weak Topology). Let X be any normed linear space. The
norm induces one topology on X, but it is not the only natural topology. Each
element of the dual space X
provides us with a seminorm
on X dened
by
(x) = [x, )[, x X.

The topology induced by the family of seminorms
X
is called the weak
topology on X. Since there are uncountably many seminorms (and no obvious
way to reduce to a countable collection in general), in order to relate the
topology to a convergence criterion we must use nets instead of sequences.
Writing x
i
w
x to denote convergence of a net x
i
iI
in X with respect to
the weak topology, the convergence criterion is
x
i
w
x
(x x
i
) 0 for all X
.
E.1 Motivation and Examples 173
Equivalently,
x
i
w
x x
i
, ) x, ) for all X
.
Norm convergence implies weak convergence, but the converse fails in general.
In terms of topologies, every set that is open with respect to the weak topology
is also open with respect to the norm topology, but not conversely. In essence,
it is easier to converge in the weak topology because there are fewer open
sets in that topology (hence the weak topology is weaker than the norm
topology).
Example E.8 (The Weak* Topology). Let X be any normed linear space. Then
its dual space X
is also a normed linear space, and hence has a topology

dened by its norm, as well as a weak topology as described above. However,
there is also a third natural topology associated with X
. Each element x
in X determines a seminorm
x
on X
by
x
() = [x, )[, X
.
The topology this family of seminorms
x
xX
induces is called the weak*
topology on X
, and convergence with respect to this topology is denoted by
i
w*
. Explicitly, if
i
iI
is a net in X
then the convergence criterion

corresponding to the weak* topology is
i
w*

x
(
i
) 0 for all x X.
Equivalently,
i
w*
x,
i
) x, ) for all x X.
Since X X
, the family of seminorms associated with the weak* topology

includes only some of the seminorms associated with the weak topology. Hence
weak convergence in X
implies weak* convergence in X
. Of course, if X is
reexive then X = X
and the weak and weak* topologies on X
are the
same.
Additional Problems
E.1. Let T(R) be the vector space containing all functions f : R C. For
each x R, dene a seminorm on T(R) by
x
(f) = [f(x)[.
(a) Show that convergence with respect to the family of seminorms
x
xR
corresponds to pointwise convergence of functions.
(b) Show that there is no norm on T(R) that denes the same convergence
criterion.
E.2 Topological Vector Spaces
Now we will consider vector spaces that have nice topologies, especially
those that are generated by families of seminorms.
E.2.1 Base for a Topology
A base is a set of building blocks for a topology, playing a role analogous
to the one played by the collection of open balls in a metric space.
Denition E.9 (Base for a Topology). Let T be a topology on a set X.
A base for the topology is a collection of open sets B T such that for any
open set U T and any vector x U there exists a base element B B that
contains x and is contained in U, i.e.,
U T , x U, B B such that x B U.
Consequently, if B is a base for a topology, then a set U is open if and only
if
U =
xU
_
B
x
: B
x
B and x B
x
U
_
.
In particular, if X is a metric space then the collection of open balls in X
forms a base for the topology induced by the metric.
Remark E.10. If c is an arbitrary collection of subsets of X, then the smallest
topology that contains c is called the topology generated by c, denoted T (c),
see Exercise A.38. Each element of T (c) can be written as a union of nite
intersections of elements of c. In contrast, if B is a base for a topology T , then
each element of T can be written as a union of elements of B. The topology
generated by B is T , but we do not need to go through the extra step of taking
nite intersections to form arbitrary open sets in the topology. Sometimes c
is called a subbase for the topology T (c).
Denition E.11 (Locally Convex). If X is a vector space that has a topol-
ogy T , then we say that this topology is locally convex if there exists a base B
for the topology that consists of convex sets.
For example, if X is a normed linear space, then X is locally convex since
each open ball in X is convex. However, as L
p
(E) with p < 1 illustrates, a
metric linear space need not be locally convex in general.
E.2.2 Topological Vector Spaces
A topological vector space is a vector space that has a topology such that
the operations of vector addition and scalar multiplication are continuous. In
order to dene this precisely, the reader should recall the denition of the
topology on a product space X Y given in Section A.5.
E.2 Topological Vector Spaces 175
Denition E.12 (Topological Vector Space). A topological vector space
(TVS) is a vector space X together with a topology T such that
(a) (x, y) x +y is a continuous map of X X into X, and
(b) (c, x) cx is a continuous map of C X into X.
By Exercise A.8, every normed linear space is a locally convex topological
vector space.
Remark E.13. Some authors additionally require in the denition of topolog-
ical vector space that the topology on X be Hausdor, and some further
require the topology to be locally convex.
Lemma E.14. If X is a topological vector space, then the topology on X is
translation-invariant, meaning that if U X is open, then U +x is open for
every x X.
Proof. Suppose U X is open. Then, since vector addition is continuous, the
inverse image of U under vector addition, which is
+
1
(U) =
_
(y, z) X X : y +z U
_
,
is open in X X. Exercise A.41 therefore implies that the restriction
_
y X : y +z U
_
is open in X for each z X. In particular,
U +x =
_
u +x : u U
_
=
_
y X : y + (x) U
_
is open in X.
Additional Problems
E.2. Let X be a normed vector space. Show that the topology induced from
the norm is the smallest topology with respect to which X is a topological
vector space and x |x| is continuous.
E.3. Let X be a normed vector space. For each 0 < r < s < , let A
rs
be
the open annulus A
rs
= x X : r < |x| < s centered at the origin. Let
B consist of , X, all open balls B
r
(0) centered at the origin, and all open
annuli A
rs
centered at the origin. Prove the following facts.
(a) B is a base for the topology T (B) generated by B.
(b) x |x| is continuous with respect to the topology T (B).
(c) X is not a topological vector space with respect to the topology T (B).
E.3 Topologies Induced by Families of Seminorms
Our goal in this section is to show that a family of seminorms on a vector
space X induces a natural topology on that space, and that X is a locally
convex topological vector space with respect to that topology.
E.3.1 Motivation
In order to motivate the construction of the topology associated with a family
of seminorms, let us consider the ordinary topology on the Euclidean space R
2
.
We usually consider this topology to be induced from the Euclidean norm
on R
2
. We will show that the same topology is induced from the two semi-
norms
1
and
2
on R
2
dened by
1
(x
1
, x
2
) = [x
1
[ and
2
(x
1
, x
2
) = [x
2
[.
In analogy with how we create open balls from a norm, dene
B
r
(x) = y R
2
:
(x y) < r, x R
2
, r > 0, = 1, 2.
These sets are open strips instead of open balls, see the illustration in Fig-
ure E.2. By taking nite intersections of these strips, we obtain all possible
open rectangles (a, b) (c, d), and unions of these rectangles exactly give us
all the subsets of R
2
that are open with respect to the Euclidean topology.
Thus
c =
_
B
r
(x) : x R
2
, r > 0, = 1, 2
_
generates the usual topology on R
2
. However, c is not a base for this topology,
since we cannot write an arbitrary open set as a union of open strips. Instead,
we have to take one more step: The collection of nite intersections of the
open strips forms the base. Every open set is a union of nite intersections
of open strips. Furthermore, each of these nite intersections of strips is an
open rectangle, which is convex, so our base consists of convex sets. Thus, this
topology is locally convex. The topology induced from an arbitrary family of
seminorms on a vector space will be dened in exactly the same way.
E.3.2 The Topology Associated with a Family of Seminorms
Denition E.15 (Topology Induced from Seminorms). Let
J
be
a family of seminorms on a vector space X. Then the th open strip of radius r
centered at x X is
B
r
(x) =
_
y X :
(x y) < r
_
.
Let c be the collection of all open strips in X:
c =
_
B
r
(x) : J, r > 0, x X
_
.
The topology T (c) generated by c is called the topology induced by
J
.
E.3 Topologies Induced by Families of Seminorms 177
x
4 2 2 4
1
1
2
3
4
Fig. E.2. The open strip B
2
r
(x) for x = (1, 2) and r = 1.
The fact that
is a seminorm ensures that each open strip B
r
(x) is
convex. Hence all nite intersections of open strips will also be convex.
One base for the topology generated by the open strips is the collection
of all possible nite intersections of open strips. However, it is usually more
notationally convenient to use the somewhat smaller base consisting of nite
intersections of strips that are all centered at the same point and have the
same radius.
Theorem E.16. Let
J
be a family of seminorms on a vector space X.
Then
B =
_
n
j=1
B
j
r
(x) : n N,
j
J, r > 0, x X
_
forms a base for the topology induced from these seminorms. In fact, if U is
open and x U, then there exists an r > 0 and
1
, . . . ,
n
J such that
n
j=1
B
j
r
(x) U.
Further, every element of B is convex.
Proof. Suppose that U X is open and x U. By the characterization of the
generated topology given in Exercise A.38, U is a union of nite intersections
of elements of c. Hence we have
x
n
j=1
B
j
rj
(x
j
)
for some n > 0,
j
J, r
j
> 0, and x
j
X. Since x B
j
rj
(x
j
), we have
j
(x x
j
) < r
j
for each j. Therefore, if we set
r = min
_
r
j

j
(x x
j
) : j = 1, . . . , n
_
,
then we have B
j
r
(x) B
j
rj
(x
j
) for each j = 1, . . . , n. Hence
B =
n
j=1
B
j
r
(x) B,
and we have x B U.
Note that even if there are innitely many seminorms in our family, when
constructing the base B we only intersect nitely many strips at a time.
Fortunately, the topology induced by a family of seminorms is always
locally convex. Unfortunately, it need not be Hausdor in general.
Exercise E.17. Let
J
be a family of seminorms on a vector space X.
Show that the induced topology on X is Hausdor if and only if
(x) = 0 for all J x = 0.

Hausdorness is an important property because without it the limit of
a sequence need not be unique (see Problem A.17). Hence in almost every
practical circumstance we require the topology to be Hausdor. If any one
of the seminorms in our family is a norm, then the corresponding topology
is automatically Hausdor (for example, this is the case for C
b
(R), see Ex-
ample E.3). On the other hand, the topology can be Hausdor even if no
individual seminorm is a norm (consider L
1
loc
(R) in Example E.5).
E.3.3 The Convergence Criterion
The meaning of convergence with respect to a net in an arbitrary topological
space X was given in Denition A.44. Specically, a net x
i
iI
converges to
x X if for any open neighborhood U of x there exists i
0
I such that
i i
0
= x
i
U.
In this case we write x
i
x.
When the topology is induced from a family of seminorms, we can refor-
mulate the meaning of convergence directly in terms of the seminorms instead
of open neighborhoods. Since we are dealing with arbitrary collections of semi-
norms at this point, we must still deal with convergence in terms of nets rather
than ordinary sequences, but even so the fact that seminorms are real-valued
allows a certain amount of notational simplication. Specically, given a net
x
i
iI
in X and given a seminorm on X, since the open intervals form a
base for the topology on R, we have that (x
i
) 0 in R with respect to the
directed set I if and only if for every > 0 there exists an i
0
I such that
i i
0
= (x
i
) < .
The next theorem shows that convergence with respect to the topology
induced from a family of seminorms is exactly what we expect it should be,
namely, simultaneous convergence with respect to each individual seminorm.
Theorem E.18. Let X be a vector space whose topology is induced from a
family of seminorms
J
. Then given any net x
i
iI
and any x X,
we have
x
i
x J,
(x x
i
) 0.
Proof. . Suppose that x
i
x, and x any J and > 0. Then B
(x)
is an open neighborhood of x, so by denition of convergence with respect to
a net, there exists an i
0
I such that
i i
0
= x
i
B
(x).
Therefore, for all i i
0
we have
(x x
i
) < , so
(x x
i
) 0.
. Suppose that
(x x
i
) 0 for every J, and let U be any open
neighborhood of x. Then by Theorem E.16, we can nd an r > 0 and nitely
many
1
, . . . ,
n
J such that
x
n
j=1
B
j
r
(x) U.
Now, given any j = 1, . . . , n we have
j
(x x
i
) 0. Hence, for each j we
can nd a k
j
I such that
i k
j
=
j
(x x
i
) < r.
Since I is a directed set, there exists some i
0
I such that i
0
k
j
for
j = 1, . . . , n. Thus, for all i i
0
we have
j
(xx
i
) < r for each j = 1, . . . , n,
so
x
i

n
j=1
B
j
r
(x) U, i i
0
.
Hence x
i
x.
By combining Theorem E.18 with Lemma A.53, we obtain a criterion for
continuity in terms of the seminorms.
Corollary E.19. Let X be a vector space whose topology is induced from a
family of seminorms
J
, let Y be any topological space, and x x X.
Then the following two statements are equivalent.
(a) f : X Y is continuous at x.
(b) For any net x
i
iI
in X,
(x x
i
) 0 for each J = f(x
i
) f(x) in Y.
In particular, if : X C is a linear functional, then is continuous if
and only if for each net x
i
iI
in X we have
(x
i
) 0 for each J = x
i
, ) 0.
The dual space X
of X is the space of all continuous linear functionals on X.

Remark E.20. Because of the Reverse Triangle Inequality,
(x x
i
) 0
implies
(x
i
)
(x). Hence each seminorm
is continuous with respect

to the induced topology.
E.3.4 Continuity of the Vector Space Operations
Now we can show that a vector space with a topology induced from a family
of seminorms is a topological vector space.
Theorem E.21. If X is a vector space whose topology is induced from a fam-
ily of seminorms
J
, then X is a locally convex topological vector space.
Proof. We have already seen that there is a base for the topology that consists
of convex open sets, so we just have to show that vector addition and scalar
multiplication are continuous with respect to this topology.
Suppose that (c
i
, x
i
)
iI
is any net in C X, and that (c
i
, x
i
) (c, x)
with respect to the product topology on C X. By Problem A.21, this is
equivalent to assuming that c
i
c in C and x
i
x in X. Fix any J
and any > 0. Suppose that
(x) ,= 0. Since
(x x
i
) 0, there exist i
1
,
i
2
I such that
i i
1
= [c c
i
[ < min
_

2
(x)
, 1
_
.
and
i i
2
=
(x x
i
) <

2 ([c[ + 1)
By denition of directed set, there exists some i
0
i
1
, i
2
, so both of these
inequalities hold for i i
0
. In particular, c
i
ii0
is a bounded sequence, with
[c
i
[ < [c[ + 1 for all i i
0
. Hence, for i i
0
we have
(cx c
i
x
i
)
(cx c
i
x) +
(c
i
x c
i
x
i
)
= [c c
i
[
(x) +[c
i
[
(x x
i
)
<

2
+

2
= .
If
(x) = 0 then we similarly obtain
(cx c
i
x
i
) < /2 for i i
0
. Thus we
have
(cx c
i
x
i
) 0. Since this is true for every , Corollary E.19 implies
that c
i
x
i
cx.
Exercise: Show that vector addition is continuous.
E.3.5 Continuity Equals Boundedness
For linear maps on normed vector spaces, Theorem C.6 tells us that continuity
is equivalent to boundedness. We will now prove an analogous result for oper-
ators on vector spaces whose topogies are induced from families of seminorms.
In the statement of the following result, it is perhaps surprising at rst glance
that boundedness of a given operator is completely determined by a xed
nite subcollection of the seminorms. This is a reection of the construction
of the topology, and specically of the fact that a base for the topology is
obtained by intersecting only nitely many open strips at a time.
Theorem E.22 (Continuity Equals Boundedness). Let X be a vector
space whose topology is induced from a family of seminorms
I
. Let Y be
a vector space whose topology is induced from a family of seminorms q
J
.
If L: X Y is linear, then the following statements are equivalent.
(a) L is continuous.
(b) For each J, there exist N N,
1
, . . . ,
N
I, and C > 0 (all
depending on ) such that
q
(Lx) C
N
j=1
j
(x), x X. (E.1)
Proof. (a) (b). Assume that L is continuous, and x J. Since q
is
continuous, so is q
L. Hence
(q
L)
1
(1, 1) = x X : q
(Lx) < 1
is open in X. Further, this set contains x = 0, so there must exist a base set
B =
N
j=1
B
j
r
(0) such that
B x X : q
(Lx) < 1. (E.2)

We will show that equation (E.1) is satised with C = 2/r. To see this, x
any x X, and set
=
N
j=1
j
(x).
Case 1: = 0. In this case, given any > 0 we have
j
(x) =
j
(x) = 0, j = 1, . . . , N.
Hence x B
j
r
(0) for each j = 1, . . . , N, so x B. In light of the inclusion
in equation (E.2), we therefore have for every > 0 that
q
(Lx) = q
(L(x)) < 1.
Therefore q
(Lx) = 0. Hence in this case the inequality (E.1) is trivially

satised.
Case 2: > 0. In this case we have for each j = 1, . . . , N that
j
_
rx
2
_
=
r
2

j
(x)
r
2
< r,
so
rx
2
B. Considering equation (E.2), we therefore have
r
2
q
(Lx) = q
_
L
_
rx
2
__
< 1.
Hence
q
(Lx) <
2
r
=
2
r
N
j=1
j
(x),
which is the desired inequality.
(b) (a). Suppose that statement (b) holds. Let x
i
iA
be any net in X,
and suppose that x
i
x in X. Then for each I, we have by Corollary E.19
that
(x x
i
) 0 for each I. Equation (E.1) therefore implies that
q
(Lx Lx
i
) 0 for each J. Using Corollary E.19 again, this says that
Lx
i
Lx in Y, so L is continuous.
Often, the space Y will be a normed space. In this case, the statement of
Theorem E.22 simplies as follows.
Corollary E.23. Let X be a vector space whose topology is induced from
a family of seminorms
J
, and let Y be a normed linear space. If
L: X Y is linear, then the following statements are equivalent.
(a) L is continuous.
(b) There exist N N,
1
, . . . ,
N
J, and C > 0 such that
|Lx| C
N
j=1
j
(x), x X. (E.3)
Thus, to show that a given linear operator L: X Y is continuous, we
need only nd nitely many seminorms such that the boundedness condition
in equation (E.3) holds.
E.4 Topologies Induced by Countable Families of
Seminorms
In this section, we will show that if a Hausdor topology is induced from a
countable collection of seminorms, then it is metrizable. In particular, we can
dene continuity of operators on such a space using convergence of ordinary
sequences instead of nets.
E.4.1 Metrizing the Topology
Exercise E.24. Let X be a vector space whose topology is induced from a
countable family of seminorms
n
nN
, and assume that X is Hausdor.
Prove the following statements.
E.4 Topologies Induced by Countable Families of Seminorms 183
(a) The function
d(x, y) =
n=1
2
n

n
(x y)
1 +
n
(x y)
denes a metric on X.
(b) The metric d generates the same topology as the family of seminorms
nN
.
(c) The metric d is translation-invariant, i.e.,
d(x +z, y +z) = d(x, y), x, y, z X.
Since the metric denes the same topology as the family of seminorms,
they dene the same convergence criteria, i.e., given a sequence x
k
kN
in X
and given x X, we have
d(x, x
k
) 0 n N,
n
(x x
k
) 0.
Denition E.25 (Frechet Space). Let X be a Hausdor vector space whose
topology is induced from a countable family of seminorms
n
nN
. If X is
complete with respect to the metric constructed in Exercise E.24, then we
call X a Frechet space.
Exercise E.26. Let X be as above, and suppose that a sequence x
k
kN
is Cauchy with respect to the metric d. Show that x
k
kN
is Cauchy with
respect to each individual seminorm
n
.
Exercise E.27. Show that the spaces C
b
(R), o(R), L
1
loc
(R), and C
(R)
considered in Examples E.3E.6 are all Frechet spaces.
The topology on a Frechet space need not be normable. For example, we
will show that this is the case for the space C
(R).
Example E.28. The topology on C
(R) is induced from the family of semi-

norms
mn
m,n0
, where
mn
(f) = |f
(n)
[m,m]
|
for f C
(R). Sup-
pose that there was some norm | | that induced this topology, and let
U = f C
(R) : |f| < 1 be the unit open ball. Then U must also
be open with respect to the topology induced from the seminorms, so there
exists some base element
B =
k
j=1
B
m
k
,n
k
r
(0)
=
_
f C
(R) : |f
(nj)
[mj,mj]
|
< r for j = 1, . . . , k
_
that is contained in U. Let M = maxm
1
, . . . , m
k
, and let f be any nonzero
function in C
(R) that vanishes on [M, M]. Then cf B U for every

c R, which is a contradiction since |f| ,= 0.
We remark that many of the big theorems of functional analysis that we
stated in Appendix C have extensions to the setting of Frechet spaces (and
beyond). The essence of the HahnBanach Theorem is convexity, while the
essence of the Baire Category Theorem, Open Mapping Theorem, and Closed
Graph Theorems is completeness. As Frechet spaces are both locally convex
and complete, it is not too surprising that these theorems can be extended to
Frechet spaces. We state a particular extension of the Closed Graph Theorem
here. This is a special case of [Rud91, Thm. 2.15].
Theorem E.29 (Closed Graph Theorem for Frechet Spaces). Let X
and Y be Frechet spaces. If A: X Y is linear, then the following statements
are equivalent.
(a) A is continuous.
(b) graph(A) =
_
(f, Af) : f X
_
is a closed subset of X Y.
(c) If f
n
f in X and Af
n
g in Y, then g = Af.
E.4.2 Tempered and Compactly Supported Distributions
The dual space X
of a topological vector space X is the space of all continuous

linear functionals on X. Restating our earlier continuity equals boundedness
results for the specic case of linear functionals gives us the following result.
Theorem E.30. Let X be a vector space whose topology is induced from a
family of seminorms
J
, and let : X C be linear. Then the following
(a) is continuous, i.e., X
.
(b) There exist N N,
1
, . . . ,
N
J, and C > 0 such that
[x, )[ C
N
j=1
j
(x), x X. (E.4)
We will apply this to two particular dual spaces, which are called the space
of tempered distributions and the space of compactly supported distributions.
Both are subsets of the space of distributions, which is dened in Section E.5.
Denition E.31 (Tempered Distributions). The space of tempered dis-
tributions o
(R) is the dual space of o(R):

o
(R) = o(R)
=
_
: o(R) C : is linear and continuous
_
.
Since the topology on o(R) is induced from the countably many seminorms
mn
(f) = |x
m
f
(n)
(x)|
, m, n 0,
it is metrizable. Hence we can formulate continuity of operators on the
Schwartz space in terms of convergence of ordinary sequences. Combining
this with Theorem E.30 gives us the following characterization of bounded
linear functionals on the Schwartz space.
E.4 Topologies Induced by Countable Families of Seminorms 185
Theorem E.32. If : o(R) C is linear, then the following statements are
equivalent.
(a) is continuous, i.e., o
(R).
(b) If f
k
o(R) and f
k
0 in o(R), then f
k
, ) 0.
(c) There exist C > 0 and M, N 0 such that
[f, )[ C
M
m=0
N
n=0
|x
m
f
(n)
(x)|
, f o(R).
Example E.33. Dene : o(R) C by f, ) = f(0). Given any f o(R), we
have
[f, )[ = [f(0)[ |f|
=
00
(f).
Theorem E.32 therefore implies that is continuous, and is an example of a
tempered distribution. For this particular operator, boundedness is satised
using an estimate involving only the single seminorm
00
.
Similarly, the functional
: o(R) C given by f,
) = f
(0) is contin-
uous because
[f,
)[ = [f
(0)[ |f
=
01
(f),
and we can likewise dene higher derivatives of by setting
f,
(j)
) = (1)
j
f
(j)
(0). (E.5)
In this sense, the delta functional is innitely dierentiable in o
(R).
Next we consider the dual space of c(R) = C
(R).
Denition E.34 (Compactly Supported Distributions). The space of
compactly supported distributions c
(R) is the dual space of C
(R) = c(R):
c
(R) = C
(R)
=
_
: C
(R) C : is linear and continuous

_
.
Since the topology on C
(R) is induced from the countably many semi-

norms
m,n
(f) = |f
(n)
[m,m]
|
with m, n 0, it is metrizable. Hence we

obtain the following characterization of bounded linear functionals on C
(R).
Theorem E.35. If : C
(R) C is linear, then the following statements

are equivalent.
(a) is continuous, i.e., c
(R).
(b) If f
k
C
(R) and f
k
0 in C
(R), then f
k
, ) 0.
(c) There exist C > 0 and M, N 0 such that
[f, )[ C
N
n=0
|f
(n)
[M,M]
|
, f C
(R).
The following exercise suggests why the terminology compactly sup-
ported is used in connection with the dual space of C
(R).
Exercise E.36. Show that if c
(R), then there exists a compact set

K R such that if f C
(R) and supp(f) RK, then f, ) = 0.

Although we will not pursue this further, the support of can be given
a precise meaning. For example,
(j)
as dened in equation (E.5) belongs to
c
(R) and its support is exactly supp(

(n)
) = 0.
E.5 C
c
(R) and its Dual Space D
(R)
Not every topological vector space has a topology easily described by an ex-
plicit family of seminorms. The most important such space for us is
T(R) = C
c
(R) =
_
f C
(R) : supp(f) is compact

_
.
We will sketch some facts about the topology on C
c
(R) in this section, but
will not prove every detail. For these, we refer the reader to functional analysis
text such as Conway [Con90] or Rudin [Rud91].
E.5.1 The Topology on C
c
(R)
Denition E.37. Given a compact set K R, we let C
(K) denote the col-

lection of innitely dierentiable functions on the real line that are supported
within K:
C
(K) =
_
f C
(R) : supp(f) K
_
.
With K xed, C
(K) has a topology dened by the countable family of

seminorms
K,n
n0
, where
K,n
(f) = |f
(n)
K
|
= |f
(n)
|
, f C
(K). (E.6)
Further, C
(K) is complete with respect to this topology, so it is a Frechet

space.
As a set, we have
C
c
(R) =
_
C
(K) : K R, K compact
_
.
However, we do not dene the topology on C
c
(R) by simply combining all
of the seminorms
K,n
for each n 0 and compact set K. Indeed, as we saw
in Example E.6,
K,n
: n 0, compact K R is exactly the family of
seminorms that induces the topology on C
(R). Since C
c
(R) is contained in
C
(R), this family of seminorms does induce a topology on C
c
(R), but it is
not the correct topology for C
c
(R). In particular, while C
(R) is complete
in this topology, C
c
(R) is not (see Problem E.4). Instead, the appropriate
topology on C
c
(R) is obtained by forming the inductive limit of the topologies
on C
([N, N]) for N N. To motivate this, set K

N
= [N, N], and note
the following facts.
E.5 C
c
(R) and its Dual Space D
(R) 187
(a) C
c
(R) is a vector space.
(b) C
(K
N
) is a vector space whose topology is dened by a family of semi-
norms, and this topology is Hausdor.
(c) N is a directed set, and if M N then C
(K
M
) C
(K
N
).
(d) If M N and U is an open subset of C
(K
N
), then U C
(K
M
) is an
open subset of C
(K
M
). That is, the embedding C
(K
M
) C
(K
N
)
is continuous.
(e) C
c
(R) =
_
C
(K
N
) : N N
_
.
In the terminology of [Con90],
_
C
c
(R), C
(K
N
)
NN
_
is an inductive sys-
tem, and C
c
(R) inherits its topology from this system. We will refer to this
topology as the inductive limit topology on C
c
(R). A similar topology on the
space C
c
(R) was discussed in Section D.8.
To give the precise denition of this topology, let us say that a subset
V C
c
(R) is balanced if
f V, [[ < 1 = f V.
Dene B to be the collection of all V C
c
(R) such that V is convex and
balanced, and V C
(K
N
) is open in C
(K
N
) for each N. Then a set
U C
c
(R) is dened to be open if for each f U, there exists a set V B
such that f + V U. With this topology, C
c
(R) is a locally convex topo-
logical vector space. There does exist a family of seminorms that induces this
topology, but it is not a metrizable topology.
This explicit denition of the topology is not very revealing. For us, the fol-
lowing theorem characterizing the convergence criterion in C
c
(R) is far more
important. In eect, we take the next theorem as the denition of convergence
in C
c
(R).
Theorem E.38 (Convergence in C
c
(R)). If C
c
(R) is given the inductive
limit topology described above, then the following statements hold.
(a) A sequence f
k
kN
is Cauchy in C
c
(R) if and only if there exists a
compact set K R such that f
k
C
(K) for each k and f

k
kN
is
Cauchy in C
(K).
(b) A sequence f
k
kN
converges to f in C
c
(R) if and only if there exists
a compact set K R such that f
k
C
(K) for each k and f

k
f in
C
(K).
Since each space C
(K) is a Frechet space, it follows that every Cauchy

sequence in C
c
(R) converges in C
c
(R). Thus C
c
(R) is complete with respect
to this topology.
It is useful to explicitly restate the convergence criteria in C
c
(R) as fol-
lows: A sequence f
k
kN
converges to f in C
c
(R) if and only if there exists
a single compact set K R such that supp(f
k
) K for each k, and
n 0, lim
k
|f
(n)
f
(n)
k
|
= 0.
In particular, convergence in C
c
(R) requires that all the elements of the
sequence be supported within one single compact set.
Continuity can then be dened in the usual way by using convergence of
sequences, as in the next exercise.
Exercise E.39. Prove that dierentiation is a continuous mapping of C
c
(R)
into itself, i.e., if f
k
f in C
c
(R), then f
k
f
in C
c
(R).
E.5.2 The Space of Distributions
Denition E.40. A linear functional : C
c
(R) C is continuous if
f
k
0 in C
c
(R) = f
k
, ) 0.
The space of distributions T
(R) is the dual space of T(R) = C
c
(R):
T
(R) = C
c
(R)
=
_
: C
c
(R) C : is linear and continuous
_
.
Combining Theorem E.38 with our previous continuity equals bounded-
ness results (Corollary E.23) yields the following characterization of contin-
uous linear functionals on C
c
(R).
Theorem E.41. If : C
c
(R) C is linear, then the following statements
are equivalent.
(a) is continuous, i.e., T
(R).
(b) [
C
(K)
is continuous for each compact K R. That is, if K R is
compact and f
k
C
(K) for k N, then

f
k
0 in C
(K) = f
k
, ) 0.
(c) For each compact K R, there exist C
K
> 0 and N
K
0 such that
[f, )[ C
K
NK
n=0
|f
(n)
|
, f C
(K). (E.7)
Proof. (a) (b). Suppose that T
(R), and x any compact K R.

Suppose that f
k
0 in C
(K). Then, by denition, f

k
0 in C
c
(R), and
therefore f
k
, ) 0 since is continuous.
(b) (c). Suppose that statement (b) holds, and let K R be compact.
Then restricted to C
(K) is continuous. The topology on C
(K) is de-
termined by the family of seminorms
K,n
n0
dened in equation (E.6).
Theorem E.30 therefore implies that there exist C
K
> 0 and N
K
0 such
that equation (E.7) holds.
E.6 The Weak and Weak* Topologies on a Normed Linear Space 189
(c) (a). Suppose that statement (c) holds, and that f
k
0 in C
c
(R).
Then there exists a compact K R such that supp(f
k
) K for all k. Let
C
K
, N
K
be as given in statement (c). Then by equation (E.7), we have
[f
k
, )[ C
K
NK
n=0
|f
(n)
k
|
0 as k .
Hence is continuous.
Thus, a linear functional on C
c
(R) is continuous if each restriction of
to C
(K) is continuous. Since the topology on any particular C
(K) is given
by a countable family of seminorms, for each individual compact K we have
the continuity equals boundedness characterization given in equation (E.7).
However, each compact K can give us possibly dierent constants C
K
and N
K
.
If a single integer N can be used for all compact K (with possibly dier-
ent C
K
), then the smallest such N is called the order of . For example,
(j)
dened as in equation (E.5) belongs to T
(R) and has order n.

Additional Problems
E.4. Show that C
c
(R) is not complete with respect to the metric induced
from the family of seminorms
mn
m,n0
, where
mn
(f) = |f
(n)
[m,m]
|
.
E.6 The Weak and Weak* Topologies on a Normed
Linear Space
The weak topology on a normed space and the weak* topology on the dual
of a normed space were introduced in Examples E.7 and E.8. We will study
these topologies more closely in this section. They are specic examples of
generic weak topologies determined by the requirement that a given class
of mappings f
: X Y
be continuous. The weak topology corresponding

to such a class is the weakest (smallest) topology such that each map f
is
continuous. Another example of such a weak topology is the product topology,
which is dened in Section E.7.
E.6.1 The Weak Topology
Let X be a normed space. The topology induced from the norm on X is called
the strong or norm topology on X.
For each X
, the functional
(x) = [x, )[ for x X is a seminorm

on X. The topology induced by the family of seminorms
X
is the weak
topology on X, denoted
(X, X
).
By Theorem E.16, X is a locally convex topological vector space with respect
to the weak topology. If
(x) = 0 for every X
, then, by the Hahn

Banach Theorem,
|x| = sup
=1
[x, )[ = sup
=1
(x) = 0.
Hence x = 0, so the weak topology is Hausdor (see Exercise E.17). If we let
B
r
(x) =
_
y X :
(x y) < r
_
=
_
y X : [x y, )[ < r
_
denote the open strips determined by these seminorms, then a base for the
topology (X, X
) is
B =
_
n
j=1
B
j
r
(x) : n N,
j
X
, r > 0, x X
_
.
Convergence with respect to this topology is called weak convergence in X,
denoted x
i
w
x.
Suppose that X
and x
i
iI
is a net in X such that x
i
w
0. By
Remark E.20,
is continuous with respect to the weak topology, so

[x
i
, )[ =
(x
i
) 0.
Thus is continuous with respect to the weak topology as well. The next
exercise shows that the weak topology is the smallest topology with respect
to which each X
is continuous.
Exercise E.42. Suppose that X is a normed space, and that T is a topology
on X such that each X
is continuous with respect to T . Show that

(X, X
) T .
Even though the weak topology is weaker than the norm topology, there
are a number of situations where we have the surprise that a weak prop-
erty implies a strong property. For example, Problem E.6 shows that every
weakly closed subspace of a normed space is strongly closed, and vice versa.
E.6.2 The Weak* Topology
Let X be a normed space. For each x X, the functional
x
() = [x, )[
for X
is a seminorm on X
. The topology induced by the family of

seminorms
x
xX
is the weak* topology on X
, denoted
(X
, X).
By Theorem E.16, X
is a locally convex topological vector space with respect

to the weak* topology. Further, if
x
() = 0 for every x X then, by denition
of the operator norm,
E.6 The Weak and Weak* Topologies on a Normed Linear Space 191
|| = sup
x=1
[x, )[ = sup
x=1
x
() = 0.
Hence = 0, so the weak* topology is Hausdor. If we let
B
x
r
() =
_
X
:
x
( ) < r
_
=
_
X
: [x, )[ < r
_
denote the open strips determined by these seminorms, then a base for the
topology (X
, X) is
B =
_
n
j=1
B
xj
r
() : n N, x
j
X, r > 0, X
_
.
Convergence with respect to this topology is called weak* convergence in X
,
denoted
i
w*
.
Each vector x X is identied with the functional x X
dened by
, x ) = x, ), X
.
Suppose that x X and that
i
iI
is a net in X
such that
i
w*
0. Then
[
i
, x)[ = [x,
i
)[ =
x
(
i
) 0,
so x is continuous with respect to the weak* topology. The next exercise shows
that the weak* topology is the smallest topology with respect to which x is
continuous for each x X.
Exercise E.43. Suppose that X is a normed space, and that T is a topology
on X
such that x is continuous with respect to T for each x X. Show that

(X
, X) T .
Additional Problems
E.5. Let X be a normed space, and let T be the strong topology on X.
(a) Show directly that (X, X
) T .
(b) Use the fact that each X
is continuous (by denition) in the

strong topology to show that (X, X
) T .
E.6. Let X be a normed space and S a subspace of X. Prove that the following
(a) S is strongly closed (i.e., closed with respect to the norm topology).
(b) S is weakly closed (i.e., closed with respect to the weak topology).
E.7. Let e
n
nN
be an orthonormal basis for a Hilbert space H.
(a) Show that e
n
w
0.
(b) Show that n
1/2
e
n
does not converge weakly to 0.
(c) Show that 0 belongs to the weak closure of n
1/2
e
n
nN
, i.e., 0 is an
accumulation point of this set with respect to the weak topology on H.
E.7 Alaoglus Theorem
If X is a normed space, then the closed unit ball in X or X
is compact with
respect to the norm topology if and only if X is nite-dimensional (Prob-
lem A.27). Even so, Alaoglus Theorem states that the closed unit ball in X
is compact in the weak* topology. We will prove this theorem in this section.
E.7.1 Product Topologies
For the case of two topological spaces X and Y, the product topology on
X Y was dened in Section A.5. We review here some facts about the
product topology on arbitrary products of topological spaces.
Denition E.44 (Product Topology). Let J be a nonempty index set, and
for each j J let X
j
be a nonempty topological space. Let X be the Cartesian
product of the X
j
:
X =
jJ
X
j
=
_
x
j
jJ
: x
j
X
j
for j J
_
,
where a sequence F = x
j
jJ
denotes the mapping F : J
j
X
j
given by
F(j) = x
j
. The Axiom of Choice states that X is nonempty. The product
topology on X is the topology generated by the collection
B =
_
jJ
U
j
: U
j
open in X
j
, and U
j
= X
j
except for nitely many j
_
.
Since B is closed under nite intersections, it forms a base for the product
topology. Thus, the open sets in X are unions of elements of B.
For each j, we dene the canonical projection of X onto X
j
to be the
mapping
j
: X X
j
dened by
j
_
x
i
iJ
_
= x
j
.
If U
j
is an open subset of X
j
and we dene U
i
= X
i
for all i ,= j, then
1
j
(U
j
) =
iJ
U
i
, (E.8)
which is open in X. Hence each
j
is a continuous map. Moreover, if T is
any other topology on X such that each canonical projection
j
is continu-
ous then T must contain all of the sets given by equation (E.8). Since nite
intersections of those sets form the base B, we conclude that T must contain
the product topology on X. Thus, the product topology on X is the weakest
topology such that each canonical projection
j
is continuous. This is another
example of a general kind of weak topology determined by the requirement
that a given class of mappings on X be continuous.
E.7 Alaoglus Theorem 193
Tychonos Theorem is a fundamental result on compact sets in the prod-
uct topology. The proof uses the Axiom of Choice, see [Fol99]. In fact, Kelley
proved that Tychonos Theorem is equivalent to the Axiom of Choice [Kel50].
Theorem E.45 (Tychonos Theorem). For each j J, let X
j
be a
topological space. If each X
j
is compact, then X =
jJ
X
j
is compact in the
product topology.
E.7.2 Statement and Proof of Alaoglus Theorem
Now we can prove Alaoglus Theorem (also known as the BanachAlaoglu
Theorem).
Theorem E.46 (Alaoglus Theorem). Let X be a normed linear space,
and let
B
= X
: || 1
be the closed unit ball in X
. Then B
is compact in X
with respect to the

weak* topology on X
.
Proof. For each x X, let
D
x
= z C : [z[ |x|
be the closed unit ball of radius |x| in the complex plane. Each D
x
is compact
in C, so Tychonos Theorem implies that D =
xX
D
x
is compact in the
product topology.
The elements of D are sequences =
x
xX
where
x
D
x
for each x.
More precisely, is a mapping of X into
xX
D
x
= C that satises [
x
[
|x| for all x X. Thus is a functional on X, although it need not be linear.
Since is a functional, we adopt our standard notation and write x, ) =
x
.
Then we have
[x, )[ |x|, x X.
If is linear then we have || 1, so B
. In fact B
consists exactly of
those elements of D that are linear.
Our next goal is to show that B
is closed with respect to the product

topology restricted to D. Suppose that
i
iI
is a net in B
and
i
D,
where the convergence is with respect to the product topology. Since the
canonical projections are continuous in the product topology, we have
x,
i
) =
x
(
i
)
x
() = x, ), x X. (E.9)
In particular, given x, y X and a, b C, we have
ax +by,
i
) ax +by, ).
However, each
i
is linear since it belongs to B
, so we also have
ax +by,
i
) = a x,
i
) +b y,
i
) a x, ) +b y, ).
Therefore ax+by, ) = a x, ) +b y, ), so is linear and therefore belongs
to B
. Hence B
is a closed subset of D. Since D is compact in the product

topology, we conclude that B
is also compact in the product topology (see

Problem A.22).
Now we will show that the product topology on D restricted to B
is
the same as the weak* topology on X
restricted to B
. To do this, let T
denote the product topology on D restricted to B
, and let denote the

weak* topology on X
restricted to B
. Suppose that
i
iI
is a net in B
and
i
with respect to the product topology on D. Equation (E.9) shows
that B
and
i
w*
in this case. Hence every subset of B
that is
closed with respect to is closed with respect to T , so we have T (see
Exercise A.48).
Conversely, x any x X and suppose that
i
iI
is a net in B
such that
i
w*
. Then
x
(
i
) = x,
i
) x, ) =
x
(), so the canonical projection
x
is continuous with respect to the weak* topology restricted to B
. However,
T is the weakest topology with respect to which each canonical projection is
continuous, so T .
Thus, T = . Since we know that B
is compact with respect to T , we

conclude that it is also compact with respect to . That is, B
is compact
with respect to the weak* topology on X
restricted to B
, and this implies

that it is compact with respect to the weak* topology on X
.
As a consequence, if X is reexive then the closed unit ball in X
is
weakly compact. In particular, the closed unit ball in a Hilbert space is weakly
compact. On the other hand, the space c
0
is not reexive, and its closed unit
ball is not weakly compact (Problem E.9).
E.7.3 Implications for Separable Spaces
Alaoglus Theorem has some important consequences for separable spaces.
Although we will restrict our attention here to normed spaces, many of these
results hold more generally, see [Rud91].
We will need the following lemma.
Lemma E.47. Let T
1
, T
2
be topologies on a set X such that:
(a) X is Hausdor with respect to T
1
,
(b) X is compact with respect to T
2
, and
(c) T
1
T
2
.
Then T
1
= T
2
.
Proof. Suppose that F X is T
2
-closed. Then F is T
2
-compact since X is T
2
-
compact (see Problem A.22). Suppose that U
J
is any cover of F by sets
E.7 Alaoglus Theorem 195
that are T
1
-open. Then each of these sets is also T
2
-open, so there must exist
a nite subcollection that covers F. Hence F is T
1
-compact, and therefore
is T
1
-closed since T
1
is Hausdor (again see Problem A.22). Consequently,
T
2
T
1
.
Now we can show that a weak*-compact subset of the dual space of a
separable normed space is metrizable.
Theorem E.48. Let X be a separable normed space, and x K X
. If K is
weak*-compact, then the weak* topology of X
restricted to K is metrizable.
Proof. Let x
n
nN
be a countable dense subset of X.
Each x X determines a seminorm
x
on X
given by
x
() = [x, )[
for X
. The family of seminorms

x
xX
induces the weak* topology
(X
, X) on X
. The subfamily
xn
nN
also induces a topology on X
,
which we will call T . Since this is a smaller family of seminorms, we have
T (X
, X).
Suppose that X
and
xn
() = 0 for every n N. Then we have
x
n
, ) = 0 for every n. Since x
n
nN
is dense in X and is continuous,
this implies that = 0. Consequently, by Exercise E.17, the topology T is
Hausdor. Thus T is a Hausdor topology induced from a countable fam-
ily of seminorms, so Exercise E.24 tells us that this topology is metrizable.
Specically, T is induced from the metric
d(, ) =
n=1
2
n

xn
( )
1 +
xn
( )
, , X
.
Let T [
K
and (X
, X)[
K
denote these two topologies restricted to the
subset K. Then we have that K is Hausdor with respect to T [
K
, and is
compact with respect to (X
, X)[
K
. Lemma E.47 therefore implies that
T [
K
= (X
, X)[
K
. The topology T
K
is metrizable, as it is formed by re-
stricting the metric d to K. Hence (X
, X)[
K
is metrizable as well.
In the course of the proof of Theorem E.48 we constructed a metrizable
topology T on X
, and showed that the restrictions of T and (X
, X) to
any weak*-compact set are equal. This does not show that T and (X
, X)
are equal. In fact, if X is innite-dimensional, the weak* topology on X is not
metrizable, see [Rud91, p. 70].
Now we can prove a stronger form of Alaoglus Theorem for separable
normed spaces. Specically, we show that if X is normed and separable, then
any bounded sequence in X
has a weak*-convergent subsequence.

Theorem E.49. If X is a separable normed space then the closed unit ball
in X
is sequentially weak*-compact. That is, if

n
nN
is any sequence
in X
with |
n
| 1 for n N, then there exists a subsequence
n
k
kN
and
X
such that
n
k
w*
.
Proof. By Alaoglus Theorem, the closed unit ball B
in X
is weak*-compact.
Since X is separable, Theorem E.48 implies that the weak* topology on B
is metrizable. Finally, since B
is a compact subset of a metric space, Theo-

rem A.69 implies that it is sequentially compact.
Corollary E.50. If X is a separable reexive normed space then the closed
unit ball in X
is sequentially weakly compact. In particular, the closed unit

ball in a separable Hilbert space is sequentially weakly compact.
However, even if X is separable, the closed unit ball in X need not be
metrizable in the weak topology. For example, [Con90, Prop. V.5.2] shows
that this is the case for the closed unit ball in
1
.
There are several deep results on weak compactness that we will not elab-
orate upon. For example, the Eberlein-
Smulian Theorem states that a subset

of a Banach space is weakly compact if and only if it is weakly sequentially
compact, see [Con90, Thm. V.13.1].
Additional Problems
E.8. This problem will use Alaoglus Theorem to construct an element of
(
that does not belong to

1
.
(a) For each n N, dene
n
:
C by x,
n
) =
1
n
(x
1
+ +x
n
) for
x = (x
1
, x
2
, . . . )
. Show that
n
(
and |
n
| 1.
(b) Use Alaoglus Theorem to show that there exists a (
that is
an accumulation point of
n
nN
.
(c) Show that ,= x for any x
1
, where x is the image of x under the
natural embedding of
1
into (
1
)
= (
.
E.9. Show that the closed unit ball x c
0
: |x|
1 in c
0
is not weakly
compact.
F
Complex Analysis
In this appendix, we collect some basic denitions and facts from complex
analysis. Complex analysis is a vast and beautiful subject, and there is a rich
interaction between harmonic analysis and complex analysis, some of which
can be seen in the texts by Dym and McKean [DM72] and Young [You01].
Some basic texts on complex analysis include Conway [Con78], Marsden and
Homan [MH87], and Stein and Shakarchi [SS03b].
F.1 Analytic Functions
Complex analysis deals with functions f that map a domain C into the
complex plane C. Such a function f is said to be analytic if it has a complex
derivative. This is a very strong requirement no matter what path to the
origin that we take, the limit in the denition of the derivative must exist.
Denition F.1 (Analytic Function). Let C be open. Then a function
f : C is analytic or holomorphic on if the limit
f
(z) = lim
h0
f(z +h) f(z)
h
(F.1)
exists for all z . The function f
is called the derivative or complex deriva-

tive of f. If f : C C is analytic on C, then we say that f is entire.
Note that the variable h in equation (F.1) is a complex variable. The
meaning of the limit is that for every > 0, there exists a > 0 such that
0 < [h[ < , z +h =
(z)
f(z +h) f(z)
h
< .
Any polynomial p(z) = a
0
+a
1
z + +a
n
z
n
is entire, as is the exponential
function e
z
. The function f(z) = 1/z is analytic on C0. If r > 0, then we
set r
z
= e
z ln r
.
198 F Complex Analysis
There are many equivalent formulations of analyticity. Analytic functions
are very highly constrained, and some of their many properties are given in
the following theorems.
Theorem F.2 (Properties of Analytic Functions). If f is analytic on
an open set C, then the following statements hold.
(a) f
is analytic on .
(b) f has innitely many complex derivatives f
, f
, f
, . . . on .
(c) f is uniquely determined by its values on any innite set that has an
accumulation point. Specically, if S be any innite subset of that has
an accumulation point in and f, g are analytic functions on such that
f(z) = g(z) for all z S, then f(z) = g(z) for all z .
Theorem F.3 (Liouvilles Theorem). If f is both bounded and analytic
on all of C, then f is constant.
Theorem F.4 (Maximum Modulus Principle). Let be a bounded,
open, and connected subset of C. Suppose that f is analytic on and continu-
ous on , and let M = sup
z
[f(z)[ be the maximum of [f[ on the boundary
of . Then [f(z)[ M for all z . Moreover, if [f(z)[ = M for some z
then f is constant.
F.2 Power Series and Taylor Series
There are many useful connections between analytic functions and innite
series. In particular, we can use power series to construct analytic functions.
A power series is any formal series of the form
f(z) =
n=0
a
n
z
n
, (F.2)
where the scalars a
n
C are xed. Letting D
r
(z) = w C : [w z[ < r
denote the open disk in the complex plane of radius r centered at z, we have
the following analyticity property of power series.
Theorem F.5. Let a
n
C be given for n 0. Then there exists an R [0, ]
(called the radius of convergence) such that the following facts hold.
(a) The power series in equation (F.2) converges absolutely for all complex z
with [z[ < R, and diverges whenever [z[ > R. Furthermore, it converges
uniformly on compact subsets of D
R
(0).
(b) The function f dened by equation (F.2) is analytic on D
R
(0).
F.3 Dirichlet Series 199
(c) The complex derivative of f is
f
(z) =
n=1
na
n
z
n1
,
and this power series has exactly the same radius of convergence R.
By replacing z by z z
0
, we can create power series that converge in a
disk centered at z
0
.
Conversely, every analytic function can be represented as a power series.
Indeed, an equivalent denition of analyticity can be dened in terms of the
existence of convergent power series expansions.
Theorem F.6 (Taylor Series). Suppose C is open, and f : C is
analytic. If D
r
(z
0
) , then
f(z) =
n=0
f
(n)
(z
0
)
n!
(z z
0
)
n
, z D
r
(z
0
). (F.3)
This series converges absolutely on D
r
(z
0
), and uniformly on any compact
subset of D
r
(z
0
).
The Taylor series for the exponential function is
e
z
=
n=0
z
n
n!
, z C.
F.3 Dirichlet Series
Another important type of series is a Dirichlet series, which is a formal series
of the form
f(z) =
n=1
a
n
n
z
, (F.4)
where a
n
C.
Note that n
z
= e
z ln n
, which is analytic on C. Also, [n
z
[ = [n
Re(z)
[, so if
sup [a
n
[ < , then the Dirichlet series in equation (F.4) converges uniformly
on compact subsets of the half-plane z C : Re(z) > 1, and therefore
denes an analytic function f(z) on this half-plane. The following result gives
a more subtle fact about Dirichlet series.
Theorem F.7. If a
n
C satisfy
sup
n>0, k0
[a
n
+ +a
n+k
[ < ,
then the Dirichlet series in equation (F.4) converges and is analytic on the
half-plane z C : Re(z) > 0.
The most famous Dirichlet series denes the Riemann zeta function, which
is
(s) =
n=1
1
n
s
, Re(s) > 1. (F.5)
It is traditional to use the letter s as the independent variable when dealing
with this function. Unfortunately, we cannot use Theorem F.7 to extend the
domain of , but other techniques can be used to show that the Riemann
zeta function has a unique extension to a function that is analytic on all of
C1, with a simple pole at s = 1. This extended function is still called the
Riemann zeta function, but the series denition in equation (F.5) is only valid
for Re(s) > 1.
Aside from trivial zeros at s = 2, 4, . . . , it is known that any other
zeros of are restricted to the critical strip s C : 0 < Re(s) < 1,
and furthermore they are symmetrically distributed about the critical line
s C : Re(s) = 1/2. The Riemann hypothesis, formulated by Bernhard
Riemann in 1859, states that all nontrivial zeros of (s) lie on the critical line.
It is still unknown whether this conjecture is true or false.
F.4 Trigonometric Polynomials
Denition F.8. Given R, set e
(x) = e
2ix
. A trigonometric polynomial
is any nite linear combination of the functions e
, i.e., it has the form

p(x) =
N
k=1
c
k
e
2i
k
x
, x R, (F.6)
where c
k
C and
k
R.
Some authors restrict the terminology trigonometric polynomial to the
case where the frequencies
k
are all integers, and refer to a function of the
form given in equation (F.6) with arbitrary real
k
as a nonharmonic trigono-
metric polynomial.
If we extend the domain of e
from R to C, the resulting function

e
(z) = e
2iz
, z C,
is entire. Consequently, any trigonometric polynomial p has an analytic exten-
sion to all of C. Hence, if p is not identically zero, then its zeros cannot have
an accumulation point. In particular, p can only have countably many zeros,
and so p is zero only on a set of measure zero. This gives us the following
simple but useful consequence.
Corollary F.9. Let E R be any Lebesgue measurable set with [E[ > 0.
Then the collection e
R
, where e
(x) = e
2ix
for x R, is nitely linearly
independent in L
p
(E) for each 1 p .
F.5 Interpolation 201
F.5 Interpolation
In this section we will prove the RieszThorin Interpolation Theorem. The
statement of the RieszThorin Theorem is concerned with boundedness of
operators on L
p
spaces and as such may not appear to explicitly involve com-
plex analysis. However, the standard proofs are based on properties of analytic
functions. Indeed, the RieszThorin Theorem is the prototypical example of a
complex interpolation theorem dealing with boundedness of operators on com-
plex Banach spaces. For details on real and complex interpolation methods,
we refer to the text by Bergh and L ofstrom [BeL76]. The proof of the Riesz
Thorin Interpolation Theorem that we give is based on Follands presentation
in [Fol99].
We will need the following result about functions analytic in a strip in the
complex plane.
Theorem F.10 (Three Lines Lemma). Let f be a function that is bounded
and continuous on the closed strip = z C : 0 Re(z) 1 and analytic
on the interior of this strip. Set
M
0
= sup
Re(z)=0
[f(z)[ and M
1
= sup
Re(z)=1
[f(z)[.
Then
[f(z)[ M
1
0
M
1
for Re(z) = , 0 < < 1.
Proof. Step 1. Assume rst that M
0
, M
1
1 and that
lim
b
_
sup
0a1
[f(a +ib)[
_
= 0.
Choose R large enough that [f(z)[ 1 for all z with Im(z) = R. Then
by the Maximum Modulus Principle (Theorem F.4) we have [f(z)[ 1 for
all z in the rectangle z = a + ib : 0 a 1, R b R. Letting R
we obtain [f(z)[ 1 for all z , which establishes the result for this f.
Step 2. Now let f be an arbitrary function satisfying the hypotheses of the
theorem. Given > 0, set
f
(z) = M
z1
0
M
z
1
e
z(z1)
f(z).
Exercise: Show that f
satises the extra conditions required in Step 1. Con-

sequently, [f
(z)[ 1 everywhere on . Hence if Re(z) = then

[f(z)[ M
1
0
M
1
= lim
0
[f
(z)[ 1.
Theorem F.11 (RieszThorin Interpolation Theorem). Let 1 p
0
,
p
1
, q
0
, q
1
be given. Assume that
(a) T : L
p0
(R) +L
p1
(R) L
q0
(R) +L
q1
(R) is linear,
(b) T(L
p0
(R)) L
q0
(R),
(c) T(L
p1
(R)) L
q1
(R), and
(d) the operator norms M
0
= |T|
L
p
0L
q
0 and M
1
= |T|
L
p
1L
q
1 are nite.
Given 0 < < 1, dene p, q by
1
p
=
1
p
0
+

p
1
and
1
q
=
1
q
0
+

q
1
. (F.7)
Then T maps L
p
(R) into L
q
(R), and T : L
p
(R) L
q
(R) is bounded with
operator norm
|T|
L
p
L
q M
1
0
M
1
.
Proof. Let 0 < < 1 be xed, and let p, q be as dened in equation (F.7).
Note that L
p
(R) L
p0
(R) +L
p1
(R) by Problem B.15, so T is well-dened on
L
p
(R).
Case 1: p
0
= p
1
. In this case p = p
0
= p
1
, so it follows from Problem B.14
that
|Tf|
q
|Tf|
1
q0
|Tf|
q1
M
1
0
M
1
|f|
p
.
Thus T maps L
p
(R) boundedly into L
q
(R).
Case 2: p
0
,= p
1
and 1 < q . It follows from equation (F.7) that
p < , and therefore the set S of all simple functions that have compact
support is dense in L
p
(R) by Exercise B.59. Appealing to [Fol99, Thm. 6.14],
|Tf|
q
= sup
_
_
Tf(x) g(x) dx
: g S, |g|
q
= 1
_
. (F.8)
Fix any f S with |f|
p
= 1, and suppose that g S satises |g|
q
= 1.
Since f and g are both simple functions, we can write them as
f =
M
j=1
c
j
Ej
and g =
N
k=1
d
k
F
k
.
We can assume that the scalars c
j
, d
k
are all nonzero, the sets E
j
are disjoint,
and the sets F
k
are disjoint. Write c
j
= a
j
[c
j
[ and d
k
= b
k
[d
k
[ where a
j
and
b
k
are complex numbers with unit modulus. Set
(z) =
1 z
p
0
+
z
p
1
and (z) =
1 z
q
0
+
z
q
1
,
and note that () = 1/p and () = 1/q. In particular, 0 < (), () 1.
Dene
F.5 Interpolation 203
f
z
=
M
j=1
a
j
[c
j
[
(z)/()
Ej
and g
z
=
N
k=1
b
k
[d
k
[
(1(z))/(1())
F
k
,
and set
(z) =
_
Tf
z
(x) g(x) dx =
M
j=1
N
k=1
C
jk
[c
j
[
(z)/()
[d
k
[
(1(z))/(1())
,
where
C
jk
= a
j
b
k
_
T
Ej
(x)
F
k
(x) dx.
Since and are polynomials in z, the function is entire. Exercise: Show
that is bounded on the closed strip
= z C : 0 Re(z) 1.
Suppose that Re(z) = 0, i.e., z = ib where b R. Note that
(ib) =
1 ib
p
0
+
ib
p
1
=
1
p
0
+ib
_
1
p
1
1
p
0
_
.
Assume for the moment that p
0
, q
0
< . Then since the sets E
j
are disjoint,
we have that
[f
ib
[ =
M
j=1
[c
j
[
(ib)/()

Ej
=
M
j=1
[c
j
[
p/p0
Ej
= [f[
p/p0
,
and a similar calculation shows that [g
ib
[ = [g[
q
/q
0
. Therefore
[(z)[ =
_
Tf
ib
(x) g
ib
(x) dx
|Tf
ib
|
q0
|g
ib
|
q
0
M
0
|f
ib
|
p0
|g
ib
|
q
0
= M
0
_
_
[f[
p/p0
_
_
p0
_
_
[g[
q
/q
0
_
_
q
0
= M
0
|f|
p/p0
p
|g|
q
/q
0
q
= M
0
.
A small modication of this argument shows that the conclusion [(z)[ M
0
for Re(z) = 0 still holds when p
0
= or q
0
= , and a similar calculation
shows that [(z)[ M
1
for Re(z) = 1 (exercises).
The Three Lines Lemma therefore implies that [()[ M
1
0
M
1
. How-
ever, f
= f and g
= g, so we have
_
Tf(x) g(x) dx
= [()[ M
0
M
1
1
.
Applying equation (F.8), we conclude that |Tf|
q
M
1
0
M
1
for all unit
vectors f S.
Since S is dense in L
p
(R), we know that T has a unique extension to a
bounded operator on L
p
(R), but the problem is that we do know not know if
this extension agrees with the operator T given in the hypotheses of the the-
orem. So, we still must verify that the original operator T is in fact bounded
on all of L
p
(R). To address this point, x f L
p
(R). Then we can nd simple
functions f
n
S such that f
n
f pointwise a.e. with [f
n
[ [f[ a.e. for
every n, and furthermore the convergence is uniform on any set where f is
bounded (see Problem B.7). Since p is nite, the Lebesgue Dominated Con-
vergence Theorem therefore implies that f
n
f in L
p
-norm.
Set E = x R : [f(x)[ > 1, and dene
g = f

E
, g
n
= f
n
E
, h = f g, h
n
= f
n
g
n
.
At least one of p
0
, p
1
is nite, so without loss of generality let us say that it
is p
0
. Since f is bounded on the set RE, which has nite measure, we have
|ThTh
n
|
q1
M
0
|hh
n
|
p1
M
0
[RE[
1/p1
|hh
n
|
0 as n .
Additionally, by the Lebesgue Dominated Convergence Theorem,
|Tg Tg
n
|
q0
M
0
|g g
n
|
p0
0 as n .
Since L
p
-norm convergence implies pointwise a.e. convergence of a subse-
quence, by replacing g
n
nN
and h
n
nN
with appropriate subsequences
we may assume that Th
n
h pointwise a.e. and Tg
n
g pointwise a.e.
Since T is linear, this implies that Tf
n
f pointwise a.e. Applying Fatous
Lemma, we conclude that
|Tf|
q
=
__
[Tf(x)[
q
dx
_
1/q
liminf
n
__
[Tf
n
(x)[
q
dx
_
1/q
= liminf
n
|Tf
n
|
q
liminf
n
M
1
0
M
1
|f
n
|
p
= M
1
0
M
1
|f|
p
.
Consequently T is a bounded mapping of L
p
(R) into L
q
(R), and |T|
L
p
L
q
M
1
0
M
1
.
Case 3: p
0
,= p
1
and q = 1. The proof for Case 2 can be adapted by setting
g
z
= g for all z (exercise).
G
Zorns Lemma
The Axiom of Choice is one of the axioms of the standard form of set theory
most commonly accepted in mathematics (ZermeloFraenkel set theory with
the Axiom of Choice, or ZFC). Essentially, the Axiom of Choice states that
the product
iI
A
i
of any collection A
i
iI
of nonempty sets is nonempty.
We briey review Zorns Lemma in this appendix, which is an equivalent form
of the Axiom of Choice.
Denition G.1 (Partial Order). Let S be a set, and let be a relation on
S S.
(a) The relation is a partial order on S if the following are satised for all
elements A, B, C S:
i. Reexivity: A A,
ii. Symmetry: A B and B A implies A = B, and
iii. Transitivity: A B and B C implies A C.
(b) Two elements A, B S are comparable if either A B or B A.
(c) If is a partial ordering on S such that every pair A, B of elements of S
are comparable is called a linear ordering or a total ordering of S.
(d) A nonempty subset of S that is linearly ordered by is called a chain
in S.
(e) An element A S is maximal in S if for every B S we have that
B is comparable to A = B A.
(f) An element A S is an upper bound for U S if B A for every
B U.
Note that a maximal element need not be comparable to all elements of S,
and it need not be unique.
Now we can state Zorns Lemma.
206 G Zorns Lemma
Axiom G.2 (Zorns Lemma). Let be a partial order on a set S. If every
chain in S has an upper bound in S, then S contains a maximal element.
We illustrate the use of Zorns Lemma by using it to show that every
vector space X (other than the trivial space 0) possesses a Hamel basis,
i.e., a subset that is both nitely linearly independent and whose nite linear
span is X (see Denition A.79).
Theorem G.3. If X is a nontrivial vector space, then there exists a subset B
that is a Hamel basis for X.
Proof. Let
S =
_
A X : A is nitely linearly independent
_
.
Note that S is nonempty since any singleton x with x ,= 0 is linearly inde-
pendent. The inclusion relation is a partial order on S.
Suppose that C is a chain in S, say C = A
i
iI
where I is some index set.
By denition, each set A
i
is nitely independent, and we claim that A = A
i
is also nitely independent. To see this, suppose that x
1
, . . . , x
n
A. Then
for each k = 1, . . . , n we have x
k
A
i
k
for some i
k
I. Since C is a chain,
it is linearly ordered by inclusion. Therefore there is a largest set A
i
k
, i.e.,
there is a j such that A
i
k
A
ij
for k = 1, . . . , n. Hence x
1
, . . . , x
n
belong to
A
ij
and therefore are independent. Consequently A S, and since we have
A
i
A for each i I, the set A is an upper bound for the chain C.
Therefore, Zorns Lemma implies that S contains a maximal element B.
By denition, B is nitely independent, so if its nite span is X then it is a
Hamel basis for X. Suppose that there exists some x X span(B). Then
B
= Bx is nitely independent and hence belongs to S. However, B B
,
which implies that B is not a maximal element in S. This is a contradiction,
so we must have X = span(B).
In fact, the statement every vector space has a Hamel basis is equivalent
to the Axiom of Choice and Zorns Lemma.
Hints and Solution Sketches for Exercises and
Additional Problems
Exercises from Appendix A
A.12 Hints: . Let s
N
=
N
n=1
f
n
, and show that the sequence of partial
sums s
N
NN
is Cauchy in X.
. Suppose that every absolutely convergent series is convergent. Let
f
n
nN
be a Cauchy sequence in X. Show that there exists a subsequence
f
n
k
kN
such that |f
n
k+1
f
n
k
| < 2
k
for every k (see Problem A.2). Then
k
(f
n
k+1
f
n
k
) is absolutely convergent, hence converges, say to f. Show
that f
n
nN
has a subsequence that converges (consider the partial sums of
k
(f
n
k+1
f
n
k
)). Show f
n
nN
converges (consider Problem A.1).
A.17 Hints: There are several ways to prove the Triangle Inequality for
1 < p < . One way is to begin with
|x +y|
p
p
=
kI
[x
k
+y
k
[
p
=
kI
[x
k
+y
k
[
p1
[x
k
+y
k
[
kI
[x
k
+y
k
[
p1
[x
k
[ +
kI
[x
k
+y
k
[
p1
[y
k
[.
Then apply H olders Inequality to each sum using the exponent p
on the rst
factor and p for the second (recall that p
= p/(p1)). Then divide both sides

by |x +y|
p1
p
.
To show completeness, consider Problem A.6. Show that if x
n
nN
is a
Cauchy sequence in
p
and we write x
n
= (x
n
(1), x
n
(2), . . . ), then for each
xed k we have that x
n
(k)
nN
is a Cauchy sequence of scalars, hence con-
verges. This gives a candidate sequence x for the limit of x
n
nN
. Use the
fact that x
n
nN
is Cauchy together with the componentwise convergence to
show that |x x
n
|
p
0.
A.18 Hint: Assume 0 < p < 1. Show that (1 +t)
p
1 +t
p
for t > 0, and use
this to show that (a +b)
p
a
p
+b
p
for a, b 0.
208 Solutions
A.20 Hints: To show that C
c
(R) is not complete, choose a function in C
0
(R)
that is nonzero everywhere, and then dene appropriate continuous functions
g
N
supported on [N 1, N + 1] such that g
N
= g on [N, N]. Show that
g
N
NN
is a Cauchy sequence in C
c
(R) that does not converge within C
c
(R)
with respect to the uniform norm. On the other hand, this sequence does
converge in C
0
(R).
A.28 Hint: To prove the Triangle Inequality, consider the Polar Identity and
CauchyBunyakowskiSchwarz.
A.39 Remark: We allow I to be empty, and hence the empty set implicitly
belongs to the collection of sets given in the exercise.
A.41 Hint: (a) Show that B is closed under nite intersections.
A.47 Solution. (a) (b). Suppose E is closed. If x / E then U = XE
is an open neighborhood of x that contains no points of E, so x is not an
accumulation point of E.
(b) (a). Suppose (b) holds and x / E. Then statement (b) implies that x
is not an accumulation point of E. Hence there exists some open neighborhood
U of x such that E (Ux) = . But since x / E, this implies U XE.
Hence XE is open by Exercise A.32.
(b) (c). Suppose (b) holds and x
i
iI
is a net in E such that x
i
x.
If x = x
i
for some i then x E. Otherwise Lemma A.46 implies that x is an
accumulation point of E, and therefore x E by statement (b).
(c) (b). Suppose (c) holds and x is an accumulation point of E. Then
by Lemma A.46, there exists a net x
i
iI
in Ex such that x
i
x. Hence
statement (c) implies that x E.
(d) (b), assuming X is metric. Suppose (c) holds and x is an accumu-
lation point of E. Then by Lemma A.46, since X is metric, there exists a
sequence x
n
nN
in Ex such that x
n
x. Hence statement (d) implies
that x E.
A.48 Solution. (b) (a). Suppose that statement (b) holds. Suppose that
U is open with respect to T
1
, so E = XU is closed with respect to T
1
.
Suppose that x
i
iI
is a net in E and x
i
x with respect to T
2
. Then by
statement (b), x
i
x with respect to T
1
. Since E is closed with respect to
T
1
, this implies x E by Exercise A.47. Hence E is closed with respect to T
2
,
so U = XE T
2
.
A.52 Solution. . Suppose f is continuous at each x X. Let V be any open
subset of Y. Choose any x f
1
(V ). Then V is an open neighborhood of f(x),
so there must exist an open neighborhood U
x
of x such that U
x
f
1
(V ).
Hence
f
1
(V ) =
xf
1
(V )
x

xf
1
(V )
U
x
f
1
(V ).
Thus f
1
(V ) is a union of the open sets U
x
and hence is itself open.
Solutions 209
A.59 Solution sketch. Let A be the union of E and the accumulation points
of E, and suppose that x / A. Then x is not an accumulation point of E, so
there exists an open neighborhood U of x such that E (Ux) = . Since
x / E, this implies U contains no points of E. Show that U cannot contain
any accumulation points of E either, and conclude that U XA. Therefore
XA is open, so A is closed, and consequently E A.
A.62 Solution sketch. Let M be a nite-dimensional subspace of a normed
space X. Suppose that f
n
M and f
n
g X. If g / M, dene
M
1
= M + spang = m+cg : m M, c C.
Show that if f = m+cg with m M and c C, then |f|
M1
= |m| +[c[ is a
well-dened norm on M
1
. By Theorem A.57, all norms on M
1
are equivalent,
so f
n
g in the norm of M
1
. But |g f
n
|
M1
= |f
n
| +1 1 for every n, so
this is a contradiction.
A.66 Hints: Let
M
be 1 on [M, M], zero outside [M 1, M + 1], and
linear on [M 1, M] and [M, M +1]. Use the Weierstrass Approximation
Theorem (Theorem A.82) to show that
S =
_
N
k=0
r
k
x
k
M
(x) : M N, N 0, Re(r
k
), Im(r
k
) Q
_
is countable and dense in C
0
(R).
A.72 Hint: Let B
X
r
(x) and B
Y
s
(y) denote open balls in X and Y, respectively.
For each z Y, let U
z
= f
1
(B
Y
/2
(z)). Then U
z
zY
is an open cover of K.
As in the proof of Theorem A.69, there exists a number > 0 such that if B
is any open ball of radius in X that intersects K, then there is an z Y
such that B U
z
.
A.83 Hint: A series of the form f(x) =
k=0
k
x
k
is a power series, see
Section F.2. If it converges at some point x = r, then it converges absolutely
for all [x[ < r, and the function f so dened is innitely dierentiable on
(r, r).
A.90 Solution. Suppose
f
n
is unconditionally convergent. Then by Theo-
rem A.89, the net
_
nF
f
n
: F N, F nite
_
converges in X, say to f. Let
be any permutation of N, and choose > 0. Then there is a nite set F
0
N
such that
_
_
f

nF
f
n
_
_
< for any nite set F F
0
. Let N
0
be large
enough that F
0
(1), . . . , (N
0
). If N > N
0
, set F = (1), . . . , (N).
Then F F
0
, so
_
_
f
N
n=1
f
(n)
_
_
=
_
_
f
nF
f
n
_
_
< . Hence f =
f
(n)
.
A.96 Hints: Set d = dist(h, K) = inf|h k| : k K. Then there exist
k
n
K such that |h k
n
| d. By the Parallelogram Law,
|(hk
n
) (hk
m
)|
2
+|(hk
n
) +(hk
m
)|
2
= 2
_
|hk
n
|
2
+|hk
m
|
2
_
.
210 Solutions
Use this to show that k
n
nN
is Cauchy.
A.98 Hints: (a) (b). Let p be the point in M closest to h, and let e = ph.
Given f M and C, show that |hp|
2
|h(p+f)|
2
= |hp|
2
2 Re(f, e)) +[[

2
|f|
2
. Consider = t > 0 to show that Re(f, e)) 0, and
then consider other to obtain f, e) = 0.
A.100 Hints: First show that if M = span(A), then A
= M
. Then show
that (M
= M.
A.102 Hints: (a) Set f
N
= f
N
n=1
f, e
n
) e
n
, show f
N
e
1
, . . . , e
N
and
apply the Pythagorean Theorem.
(c) Set s
N
=
N
n=1
c
n
e
n
and t
N
=
N
n=1
[c
n
[
2
. Use the Pythagorean
Theorem to show that if t
N
converges then s
N
NN
is Cauchy and hence
converges.
(f) Show that f p, e
n
) = 0 for every n. Use the linearity and continuity
of the inner product to conclude that f p span
_
e
n
nN
_
.
Additional Problems from Appendix A
A.7 Hint: To show strict inclusion, consider x
k
= (k log
2
k)
1/q
.
A.22 Hint: (b) Fix y XF. Then for each x F there exist disjoint open
sets U
x
, V
x
such that x U
x
and y V
x
. The collection U
x
xF
is an open
cover of F.
A.23 Hints: Suppose that E is a subset of K that has no accumulation points.
Then E is closed by Exercise A.59, and hence is compact by Problem A.22.
Show that each x E has an open neighborhood U
x
that contains no points
of E other than x itself, and use the compactness of E to conclude that E is
nite.
A.26 Hints: (b) Since C
d
is complete, any closed subset is also complete.
(d) Since H is innite-dimensional, Problem A.30 shows that there exists
an innite orthonormal sequence e
n
nN
in H.
A.27 Hints: (a) Choose any u XM. Since M is closed, a = dist(u, M) > 0.
Fix > 0 small enough that
a
a+
> 1 . Then there exists v M such that
a |u v| < a + . Set g = (u v)/|u v|. Given f M we have
h = v +|u v| f M. Show that
|g f| =
|u h|
|u v|
> 1 .
(b) If X is innite-dimensional, repeatedly apply part (a) to nd vectors e
n
such that |e
n
e
m
| > 1/2 for all m < n.
Exercises from Appendix B
B.4 Hint: (b) For each k, there exists a covering
_
Q
(k)
j
_
j
of E
k
by at most
countably many cubes Q
(k)
j
such that
j
vol
_
Q
(k)
j
_
< [E
k
[ + 2
k
.
B.10 Hints: (b) If [E
k
[ < for every k, write
k=1
E
k
= E
1
(E
2
E
1
)
(E
3
E
2
) . Use additivity and part (a) to show that
[E[ = [E
1
[ + lim
N
N
k=1
_
[E
k+1
[ [E
k
[
_
.
(c) Write E
1
=
_
k=1
E
k
_
(E
1
E
2
) (E
2
E
3
) , and then apply
countable additivity in a similar fashion to how it was used in part (b).
B.14 Hint: (a) (c): For each k there exists an open U
k
E such that
[U
k
E[
e
< 1/k.
B.15 Hint: First do the following special cases: (a) both E
1
, E
2
are open, (b)
both E
1
, E
2
are bounded G
-sets, and (c) either E

1
or E
2
has measure zero.
Then use Exercise B.14 to extend to bounded measurable sets, and nally to
arbitrary measurable sets.
B.23 Hints: (a) Let Z be the set of measure zero where f(x) ,= g(x), and
write g > = f > x Z : g(x) > x Z : f(x) < .
(b) By using part (a), we can assume that f and g are nite everywhere,
so, in particular, f(x)+g(x) never has the form . Now break into steps.
First, show that g is measurable. For this it may help to show rst that
g is measurable.
Second, show that f < g is measurable by writing it as a countable
union of measurable sets: f > g =
rQ
_
f < r r < g
_
.
Third, show that f +b is measurable for any real scalar b.
Fourth, show f +g is measurable by writing f +g > = f > g.
(c) By part (a), we can assume that f is nite everywhere. Then write
( f)
1
(, ) = f
1
(
1
(, )), and use the fact that is continuous
(compare Problem B.4).
(d) By part (a), we can assume that f, g are nite everywhere. Write
fg =
1
4
_
(f +g)
2
(f g)
2
_
, and use parts (b) and (c).
(e) Let f(x) = sup f
n
(x), and show that f > = f
n
> . Also,
inf f
n
= sup (f
n
), and limsup f
n
(x) = inf
m
_
sup
nm
f
n
(x)
_
.
B.27 Hint: Consider

[0,1]
,

[0,1/2]
,

[1/2,1]
,

[0,1/3]
,

[1/3,2/3]
,

[2/3,1]
, etc.
B.28 Hints: Suppose f
n
m
f. Then there exist n
1
< n
2
< such that
_
[f f
n
[ >
1
k
_
2
k
for all n n
k
. Dene E
k
=
_
[f f
n
k
[ >
1
k
_
and
H
m
=
k=m
E
k
. Observe that [E
k
[ < 2
k
and [H
m
[ 2
m1
, and therefore
Z =
m=1
H
m
has measure zero. Show that f
n
k
(x) f(x) for all x / Z.
212 Solutions
B.29 Hints: (a) (b). [f
m
f
n
[ > 2 [f f
m
[ > [f f
m
[ > .
(b) (a). Suppose that statement (b) holds, and show there exists a
subsequence g
j
jN
of f
n
nN
such that if E
j
= [g
j
g
j+1
[ 2
j
then [E
j
[ 2
j
. Set H
k
=
j=k
E
j
and Z =
k=1
H
k
. Show that if
x / Z, then the sequence of scalars g
j
(x)
jN
is Cauchy. Hence if we de-
ne f(x) = lim
j
g
j
(x) whenever the limit exists, then f is measurable and
g
j
f pointwise a.e. Show that [g
j
f[ 2
j+1
H
k
, and use this
to show that g
j
m
f. Combine this with the fact that f
n
nN
is Cauchy in
measure to show that f
n
m
f.
B.32 Hints: (b) If =
M
j=1
a
j

Ej
and =
N
k=1
a
k
E
k
are the standard
representations of and , then + =
M
j=1
N
k=1
(a
j
+b
k
)
EjF
k
.
(c) If , then we must have a
j
b
k
whenever E
j
F
k
,= .
(e) Since
A
=
M
j=1
a
j

EjA
is a nite linear combination of nonneg-
ative simple functions, part (b) implies that
_
A
=
M
j=1
a
j
[E
j
A[. Now
expand [E
j
A[ using the countable additivity of Lebesgue measure.
B.35 Hints: (c) Let F = f > . Then
_
E
f
_
F
f
_
F
= [F[.
(d) If
_
E
f = 0, use Tchebyshevs Inequality to show that f > 1/n has
measure zero for each n > 0.
B.39 Hint: Dene g
k
(x) = inf
nk
f
n
(x) and apply the Monotone Convergence
Theorem.
B.40 Hint: Consider f
n
= n
[0,1/n]
.
B.43 Hint: The inequality is easy if f is real-valued. For the complex-valued
case, write
_
E
f
=
_
E
f where [[ = 1, and write f = f
1
+ if
2
where f
1
,
f
2
are real.
B.44 Hint: Assume rst that f
n
0. Apply Fatous Lemma to f
n
nN
and
to g f
n
nN
to show that
_
E
f liminf
n
_
E
f
n
and
_
E
g
_
E
f
_
E
g limsup
n
_
E
f
n
.
B.56 Hint: For p < , apply Tchebyshevs Inequality to [f f
k
[
p
.
B.57 Hints: Assume rst that 0 p < and that f
n
nN
is a Cauchy se-
quence in L
p
(E). Use Tchebyshevs Inequality to show that f
n
nN
is Cauchy
in measure. Then by Exercise B.29, there exists a function f on E such that
f
n
m
f, and hence by Exercise B.28 there exist f
n
k
such that f
n
k
(x) f(x)
for a.e. x E. Now show that |f f
n
k
|
p
0 by xing k large enough and
applying Fatous Lemma to
_
E
[f
nj
(x) f
n
k
(x)[
p
dx as j .
For the case p = , set Z
mn
=
_
x E : [f
m
(x) f
n
(x)[ > |f
m
f
n
|
_
.
Then Z = Z
mn
has measure zero, and f
n
(x)
nN
is a Cauchy sequence of
scalars for each x EZ.
Solutions 213
B.70 Hint: Consider a partition of [1, 1] that includes the points
2
n
for
n = 1, . . . , N.
B.73 Hints: (b) Consider f(x) = [x[.
(c) Mean-Value Theorem.
B.76 Hint: Show that we can nd partitions
k
, where each
k+1
is a rene-
ment of
k
, such that lim
k
S
k
= V
[f; a, b].
B.85 Hints: (a) Consider a single subinterval [x, y] in the denition of
absolutely continuity.
(b) To nd an absolutely continuous function that is not Lipschitz, consider
Exercise B.90. To nd a function of bounded variation that is not absolutely
continuous, consider the CantorLebesgue function.
B.89 Hint: Suppose f is continuous at x. For the case h > 0, dene k
h
=
1
h
[0,h]
and write
1
h
_
x+h
x
f(y) dy f(x)
_
h
0
_
f(x +y) f(x)
_
k
h
(y) dy
.
B.93 Hint: Let E = (x, y) [a, b]
2
: x y. By Fubinis Theorem, the two
iterated integrals
__
E
f
(x) g
(y) dxdy =
_
b
a
__
y
a
f
(x) dx
_
g
(y) dy
and
__
E
f
(x) g
(y) dy dx =
_
b
a
f
(x)
__
b
x
g
(y) dy
_
dx
are equal.
B.99 Hint: If 3
k1
[x y[ 3
k
, then [(x) (y)[ 2
k
.
B.100 Hint: (b) Apply the renement equation at the points x = 1 and
x = 2. Show that this implies that the vector
_
D
4
(1)
D
4
(2)
_
is a 1-eigenvector
of an appropriate 2 2 matrix, whose entries are determined by the scalars
appearing in the renement equation.
Additional Problems from Appendix B
B.5 Hint: Suppose that f is continuous at almost every point. If f restricted
to every interval [a, b] is measurable then f is measurable on R, so it suces
to consider [a, b] to be the domain of f. For each k N, nd a partition
k
= a = x
k
0
< x
k
1
< < x
k
N
k
= b of [a, b] such that
214 Solutions
f(x)
N
k
j=1
f(x
k
j
)
[x
k
j1
,x
k
j
)
(x)
<
1
k
, x F
k
,
where F
k
is a closed subset of [a, b] such that f is continuous at each point of
F
k
and the measure of [a, b]F
k
is less than 1/k (and it may help to construct
the F
k
so they are nested).
The function

[0,1]
is an example of a function that is continuous a.e. but
does not equal any continuous function a.e.
B.6 Hint: If f
n
(x) = n
_
f(x +
1
n
) f(x)
_
, then f
n
(x) converges to f
(x) at
each point where f is dierentiable.
B.8 Hint: Dene E
n
= x R
d
: [f(x)[ n and show that f
n
= f

En
f
in L
1
(R
d
).
B.9 Hint: To show L
p
(E) L
q
(E) when [E[ < and p < q < , apply
H olders Inequality to
_
E
[f[
p
1 using exponents q/p and (q/p)
.
To show that L
p
(R) is never contained in L
q
(R), consider functions of the
form x
(0,1)
(x) or x
(1,)
(x).
B.11 Hint: The set of all intervals [a, b] with rational endpoints a, b is count-
able. Show that
S =
_
n
k=1
r
k

[a
k
,b
k
]
: n N, Re(r
k
), Im(r
k
) Q, a
k
< b
k
Q
_
is countable and dense in L
p
(R) when 1 p < .
B.14 Hint: To show the inclusion, dene s =
qp
qr
and t =
qp
rp
, and observe
that
1
s
+
1
t
= 1 and
p
s
+
q
t
= r.
B.15 Hint: To show completeness, suppose that
f
n
is an absolutely con-
vergent series in the norm | |. Then there exist g
n
L
p
(R) and h
n
L
q
(R)
such that f
n
= g
n
+ h
n
and |g
n
|
p
+ |h
n
|
q
< 2
n
. Use the completeness of
L
p
(R) to show that g =
g
n
converges in L
p
-norm, and similarly h =
h
n
converges in L
q
-norm.
For the inclusion, given f L
r
(R), consider g = f

{|f|>1}
and h =
f
{|f|1}
.
B.16 Hint: Use the fact that
d
dy
y
x
2
+y
2
=
x
2
y
2
(x
2
+y
2
)
2
to show that for each x > 1
the function f
x
(y) =
x
2
y
2
(x
2
+y
2
)
2
is integrable on (1, ), and to nd the value of
_
1
x
2
y
2
(x
2
+y
2
)
2
dy.
B.17 Hint: Compare Exercise A.17.
B.19 Hints: (a) For the case < 1, x p + < < p 1, and use H olders
Inequality to show that
__
x
0
f(t) dt
_
p
C
_
x
0
f(t)
p
t
dt x
p1
.
Solutions 215
(b) To show C(p, p)
_
p
_
p
, set = p, and minimize over . To show
that
_
p
_
p
is best constant, consider f(t) = t
1/p
[0,R)
(t).
(c) To show that equality holds only for the zero function, note that there
is only one step where inequality can occur in the proof of part (a), namely,
when H olders Inequality is applied.
B.21 Hint: Show that f is dierentiable everywhere and f
is bounded, so f
is Lipschitz. To show g does not have bounded variation, choose a particular
appropriate partition, similar to the hint for Exercise B.70.
B.22 Hint: Given > 0, there exists an open set U A such that [U[ <
[A[
e
+. Every open set in R can be written as a union of at most countably
many disjoint open intervals (a
n
, b
n
). Use the fact that f is Lipschitz to show
that if E
n
= E (a
n
, b
n
), then [f(E
n
)[ C (b
n
a
n
).
B.24 Hints: (a) Let f be the CantorLebesgue function. Show that f maps the
complement of the Cantor set into the set of rationals in [0, 1] Then consider
the inverse image of a nonmeasurable subset of the irrationals in [0, 1] under f.
(b) Use Exercise B.14 to write E = F Z where F is an F
set and Z has

measure zero. Then F can be written as a countable union of compact sets,
and continuous functions map compact sets to compact sets.
B.26 Hints: Set E = x [a, b] : f
(x) = 0 and Z = [a, b]E and use

the BanachZarecki Theorem and Lemma B.95 to show that [f([a, b])[ = 0.
Combine this with the fact that f is continuous to show that f([a, b]) is a
single point.
B.29 Hints: Orthonormality follows from an examination of the supports
of the functions

(x k) and 2
n/2
(2
n
x k). The main issue is complete-
ness. Suppose that f L
2
(R) is orthogonal to each function

(x k)
and 2
n/2
(2
n
x k) for k Z and n 0. Show that this implies that
_
I
nk
f(t) dt = 0 for every dyadic interval I
nk
=
_
k
2
n
,
k+1
2
n
. Use the Lebesgue

Dierentiation Theorem to show that f = 0 a.e. Alternatively, use one of the
characterizations of Lebesgue measurability to show that
_
E
f(t) dt = 0 for
every measurable set E, and conclude from this that f = 0 a.e.
Remark: This direct approach settles the completeness of the Haar system,
but does not shed much light on the question of why the Daubechies wavelets
generate orthonormal bases for L
2
(R). Instead, the wavelet insight is the
role of multiresolution analysis in the wavelet construction. For this, we refer
to wavelet texts such as [Dau92].
Exercises from Appendix C
C.10 Hint: Expand |f + g|
2
= |Lf + Lg|
2
and |f + ig|
2
= |Lf + iLg|
2
using the Polar Identity.
C.14 Hint: (b) If /
, then there exists a subsequence (

n
k
)
kN
such that
[
n
k
[ k for each k. Let c
n
k
= 1/k and dene all other c
n
to be zero. Show
that f =
c
n
e
n
converges but Lf =
n
c
n
e
n
does not converge.
C.15 Hints: (a) For the case p < , to show the inequality |M
| ||
,
choose any > 0. Then E = [[ > ||
has positive measure, so some

set E
k
= E [k, k + 1] must have positive (and nite) measure. Consider
f = [E
k
[
1/p
E
k
.
(b) Suppose p < , and assume / L
(R). The sets E

k
= k [[ <
k + 1. are measurable and disjoint, and since is not in L
(X) there must

be innitely many E
k
with positive measure, say E
n
k
for k N. Choose
F
k
E
n
k
with 0 < [F
k
[ < , and consider f(x) = n
k
[F
k
[
1/p
for x F
k
,
and f(x) = 0 otherwise.
C.26 Hints: Fix f X. Since Y is dense in X, there exist g
n
Y such
that g
n
f. Show that Lg
n
nN
is Cauchy in Z, so there exists an h Z
such that Lg
n
h. Show that

Lf = h is well-dened and has the required
properties.
C.24 Hints: (b) Let A
n
nN
be a sequence in B(X, Y ) that is Cauchy in
operator norm. Given f X, show that A
n
f
nN
is a Cauchy sequence in Y
and hence converges, say to g. Dene Af = g. Use the fact that A
n
f Af for
each f together with the fact that A
n
nN
is Cauchy to show that A
n
A
in operator norm.
C.31 Hint: (b) Consider a function g C
0
(R) that is nonzero everywhere.
C.34 Hints: (b) If ,= 0, choose g / ker(). Let p be the orthogonal projection
of g onto ker(), and set e = g p. Show that (e) ,= 0, and set u = e/(e).
Show that f (f)u ker() for every f H. Use the fact that u ker()
to show that =
h
where h = u/|u|
2
.
C.37 Hints: H olders Inequality implies that |
g
| |g|
p
. For the case 1 <
p < , to show that equality holds let [(x)[ = 1 satisfy g(x) = (x) [g(x)[.
Then show that the function
f(x) =
(x) [g(x)[
p
1
|g|
p
1
p
satises |f|
p
= 1 and [f,
g
)[ = |g|
p
.
C.40 Hints: (b) Show that L
is bounded and linear with |L
| |L|. The
same reasoning applied to L
instead of L yields |(L
| |L
|.
Solutions 217
C.58 Hint: By Problem A.62, any closed and bounded subset of a nite-
dimensional vector space is compact.
C.62 Hint: (b) If V f = f with ,= 0, then f =
1
V f. Since V f is continu-
ous, we can redene f on a set of measure zero so that f is continuous. On the
other hand, V f is absolutely continuous and (V f)
= f a.e. Therefore (V f)
is equal almost everywhere to the continuous function f, so Problem B.27

implies that V f is dierentiable everywhere and (V f)
(x) = f(x) for all x.

C.66 Hint: Since T is self-adjoint, we have ker(T)
= range(T).
C.70 Hint: (a) Suppose that c = e
n
nN
and T = f
n
nN
are each or-
thonormal bases for H. Apply the Plancherel equality to expand |Af
m
|
2
in
terms of the basis c.
C.72 Hints: (f), (g) Let e
n
nN
be an orthonormal basis for H. Let P
N
be the
orthogonal projection of H onto H
N
= spane
1
, . . . , e
N
. Given T B
2
(H),
show that T
N
= TP
N
is bounded and nite-rank and converges to T in
HilbertSchmidt norm.
C.77 Hint: Since T
T 0, we can combine an orthonormal basis c for

range(T
T) with an orthonormal basis T for ker(T
T) to obtain an orthonor-
mal basis for all of H.
C.81 Hint: The method of Exercise C.70 can be adapted, or that exercise
can be applied to the compact positive operator [T[
1/2
.
C.92 Hint: Let T
p
: L
p
(E) L
p
(E)
and T
p
: L
p
(E) L
p
(E)
be the
antilinear isometric isomorphisms given by Theorem C.38. Suppose that F
L
p
(E)
. Then F T
p
L
p
(E)
, and therefore F T
p
= T
p
g =
g
for some
g L
p
(E). Show that F = g.
C.99 Hint: If f
n
nN
is a countable Hamel basis for a Banach space X,
consider F
N
= spanf
1
, . . . , f
N
and apply the Baire Category Theorem.
C.108 Hints: (c) |h
M
h
N
| |h
M
S
M
g
n
|+|S
M
g
n
S
N
g
n
|+|S
N
g
n
h
N
|.
(d) Use the fact that S
N+1
g
n
h
N+1
as n with respect to | | to
show that S
N
h
N+1
= h
N
. Note that S
N
restricted to the nite-dimensional
space spanf
1
, . . . , f
N+1
is continuous because S
N
is linear.
(g) Consider S
k
S
k1
.
C.114 Hint: (d) (b). Prove this rst for f span
_
f
k
kN
_
, and then
extend to general f X.
C.118 Hint: Suppose that f
n
w
f. Let

f
n
be the image of f
n
in X
under
the canonical embedding of X into X
, and show that sup [,

f
n
)[ < for
each X
. Apply the Uniform Boundedness Principle.

218 Solutions
Additional Problems from Appendix C
C.1 Hint: By Theorem A.57, all norms on X are equivalent, so it suces to
consider a norm such as the norm | |
1
constructed in Exercise A.56.
C.6 Hint: Choose an orthonormal basis e
n
nN
for H and an orthonormal
basis f
n
nN
for K.
C.7 Hint: The operator norms are
|A|
1 = max
j=1,...,n
_
m
i=1
[a
ij
[
_
and
|A|
= max
i=1,...,m
_
n
j=1
[a
ij
[
_
.
C.8 Solution. By the Axiom of Choice in the form of Zorns Lemma, Axiom
of Choice there exists a Hamel basis (vector space basis) f
i
iI
for X. That
is, span
_
f
i
iI
_
= X (i.e., the nite linear span is all of X), and every nite
subset of f
i
iI
is linearly independent. By dividing each vector by its length,
we can assume that |f
i
| = 1 for every i I. Let J
0
= j
1
, j
2
, . . . be any
countable subsequence of I. Dene f
jn
, ) = n for n N and f
i
, ) = 0 for
i IJ
0
. Then extend linearly to all of X: Each nonzero vector f X has a
unique representation as f =
N
k=1
c
k
f
i
k
for some i
1
, . . . , i
N
I and nonzero
scalars c
1
, . . . , c
N
, so we dene f, ) =
N
k=1
c
k
f
i
k
, ). This is a linear
functional on X, but since |f
jn
| = 1 yet [f
jn
, )[ = n, we have || = .
C.24 Hint: If c
2
, then T
c =
c
n
e
n
.
C.28 Hint: (b) Apply Problem C.17 to the normal operator L I.
C.31 Hint: Show that any subsequence f
n
nN
of e
n
nN
has a subsequence
g
n
nN
such that Tg
n
nN
converges in K, say Tg
n
h. Consider Tg
n
, h)
and Bessels inequality to show that h = 0.
C.33 Hint: Consider the nite-rank operators T
N
f =
N
n=1
n
f, e
n
) e
n
.
C.34 Hint: Consider any orthonormal sequence e
n
nN
in L
2
(R) such that
supp(e
n
) E, where E is such that [[ a.e. on E.
C.35 Hints: Since T(Ball
H
) is compact, it is totally bounded. Hence for
each n N we can nd nitely many balls of radius 1/n with centers in
T(Ball
H
) that cover T(Ball
H
). If we consider all these balls for every n, we
have countably many balls that cover T(Ball
H
). Show that this implies that
T(Ball
H
) contains a countable, dense subset. Then do the same for each ball
of radius k N instead of just k = 1. Combine all of these together to get a
countable dense subset of range(T).
Solutions 219
C.39 Hint: T
T is a multiple of the orthogonal projection onto the line

through h, hence has only one nonzero eigenvalue.
C.40 Hints: If X
is separable, then by renormalizing a countable dense

subset of X
we can nd a countable set

n
nN
that is dense in the closed
unit sphere D
=
_
X
: || = 1
_
in X
. Since |
n
| = 1, there must
exist some x
n
X with [x
n
,
n
)[ 1/2. Then M = span
_
x
n
nN
_
is a
closed subspace of X, and it is separable by Exercise A.84.
If M ,= X, choose x
0
XM with d = dist(x
0
, M) = 1. Then by Hahn
Banach, there exists some X
such that [
M
= 0, x
0
, ) = 1, and
|| = 1. Show that |
n
| 1/2 for every n, contradicting the fact that
nN
is dense in D
.
C.42 Hint: Write range(A) =
k
A(B
X
k
(0)). If range(A) is nonmeager, then
some set A(B
X
k
(0)) must contain an open ball. Apply Lemma C.103 to con-
clude that range(A) contains an open ball.
C.43 Hint: (b) Let f
n
(x) = xe
2inx
. If f
n
nZ
was a Schauder ba-
sis, then there would be a nite constant C (its basis constant ) such that
1 |f
n
|
2
|g
n
|
2
2C for all n (see Exercise C.108).
Exercises from Appendix D
D.17 Hint: Let R = P N be a Hahn decomposition for . Show that
[[(R) =
+
(P) +
(N).
D.18 Solution sketch. Let R = P N be a Hahn decomposition for . Set
(E) = sup
_
(A) : A B
, A E
_
, and x E B
. If A E, then
(A) =
+
(A)
(A)
+
(A)
+
(E), so it follows that (E)
+
(E).
On the other hand,
+
(E) = (EP) (E), so it follows that
+
= , and
the characterization of
is similar.
Now let
(E) = sup
_
n
k=1
[(E
k
)[ : n N, E
k
B
, E =
n
k=1
E
k
disjointly
_
.
If E B
and E =
n
k=1
E
k
disjointly, then
n
k=1
[(E
k
)[
n
k=1
[[(E
k
) =
[[(E), so (E) [[(E). On the other hand, E = (E P) (E N), so
[[(E) =
+
(E) +
(E) = [(E P)[ + [(E N)[ (E), and therefore

[[ = .
D.20 Hint: Consider =
1
|x|
dx.
D.45 Hint: If = f d, then (E) =
_
f
E
d. Now apply Problem D.6.
D.47 Hint: Consider the case 0, and suppose that the given equation
fails. Then there exists an > 0 such that for each = 2
n
there exists
E
n
B
with (E
n
) < 2
n
but (E
n
) . Consider F =
k=1
n=k
E
n
.
Show (F) = 0 but (F) . For a general signed measure , consider
Problem D.13
D.58 Hints: (a) By Exercise D.55, there exists a positive measure and
f L
1
() such that d = f d. By denition, we then have d[[ = [f[ d.
Hence if E B
, then
[(E)[ =
_
E
f d

_
E
[f[ d = [[(E).
(b) The fact that [[ follows from part (a). Therefore d = g d[[ for
some g L
1
([[). The denition of total variation implies that d[[ = [g[ d[[.
Show that
_
E
([g[ 1) d[[ = 0 for every Borel set E, and that this implies
that [g[ = 1 [[-a.e.
D.59 Hint: Let
1
(E),
2
(E),
3
(E) denote the quantities dened on the
right-hand side of the equality in each of the statements of parts (a), (b),
and (c), respectively. Show that
1

2

3
= [[ and
3

1
, using the
hints below.
2

3
: Suppose that E =
k=1
E
k
disjointly. Let c
k
be the scalars of
unit modulus such that
_
E
k
d
= c
k
_
E
k
d, and set f =
c
k
E
k
.
Solutions 221
[[
3
: By Exercise D.58(c), there exists f with [f[ = 1 -a.e. such that
d = f d[[. Then

f d =

ff d[[ = [f[
2
d[[ = d[[.
3

1
: Suppose [f[ 1. Since f is bounded, it follows from Exercise B.37
that we can nd a sequence of simple functions that converge to f uniformly.
Hence, if > 0 then there exists a simple function =
N
j=1
c
j
Ej
such that
sup
xR
[f(x) (x)[ . Consequently, [c
j
[ 1 + since c
j
= (x) for some
x, and
f d
_
d
+ ||.
D.62 Hints: To show the Triangle Inequality, use the equivalent form for the
total variation given in Exercise D.59(a).
Once it is shown that | | is a norm on M
b
(R), to show that it is complete,
use Exercise A.12, i.e., show that every absolutely convergent series converges.
That is, suppose that
k
M
b
(R) satisfy
|
k
| < . Show that the series
(E) =
k=1
k
(E) converges absolutely for each E B
. Show that
dened in this way is a complex Borel measure.
It still remains to show that =
k=1
k
, i.e., that the partial sums of
this series converge to in the norm of M
b
(R). To do this, again use the
equivalent form for the total variation given in Exercise D.59(a) to show that
_
_

N
k=1
k
_
_

k=N+1
|
k
| 0 as N . Note that you cannot simply
claim that
_
_

N
k=1
k
_
_
=
_
_
k=N+1
k
_
_
and apply the Triangle Inequality,
since you dont yet know that the series
k=1
k
converges in M
b
(R).
D.63 Hint: If d = g d with g L
1
(R), then by denition of the total
variation we have d[[ = [g[ d. Use the equivalent characterization of [[
given in Exercise D.59(c), and apply Problem D.15.
D.82 Solution sketch. (c) Part (a) implies that || ||. By Exercise D.58,
there exists a g with [g[ = 1 [[-a.e. such that d = g d[[. Since is bounded,
if > 0 then we have by Lusins Theorem (Theorem D.70) that there exists
a C
c
(R) such that = g except on a set E with [[(E) < , and
||
|g|
= 1. Hence
|| =
_
[g[
2
d[[ =
_
g g d[[
=
_
g d
_
d
_
( g ) d
[, )[ +
_
[ g [ d[[
[, )[ + || +.
Since this is true for every > 0, it follows that || ||.
Note that if 0, then we can give a more direct proof. In this case, let
f
n
C
c
(R) be such that 0 f
n
1 everywhere and f
n
= 1 on [n, n]. Then
222 Solutions
apply the Dominated Convergence Theorem to show that f
n
, ) ||. Since
|f
n
|
= 1, this implies that || ||.

Additional Problems from Appendix D
D.3 Hint: Show that
(E
k
) converges regardless of order. That is, show
that
k=1
(E
(k)
) converges for every bijection : N N. The series is
therefore an unconditionally convergent sequence of scalars, and hence must
converge absolutely (see Section A.10).
D.5 The elements of L
1
(
a
) are the equivalence classes
[f] = Borel measurable g : g(a) = f(a).
D.6 Solution sketch. First show that is a positive Borel measure. Suppose
that =
n
j=1
a
j
Ej
is a simple function with disjoint E
j
B
. Then
_
g d =
_
n
j=1
a
j
Ej
g d =
n
j=1
a
j
_

Ej
g d =
n
j=1
a
j
(E
j
) =
_
d.
Extend this to general functions f.
D.13 Hint: To show that implies [[ , use the equivalent charac-
terization of the total variation measure given in Exercise D.18.
D.24 Hints: Since Radon measures are both inner and outer regular, the
proof is similar to that of Theorem B.60 (consider also Problem D.7).
D.25 Hint: Show that since every subset of N is open and every nite subset is
compact, every positive measure on N is both inner regular and outer regular.
D.26 Hints: (a) The convergence criterion for the inductive limit topology
on c
00
is: f
n
f in c
00
if there exists a K > 0 such that for each n N we
have f
n
(k) = 0 for all k > K, and |f f
n
|
0 as n .
D.27 Hints: Show there exist innitely many integers n
k
with n
k
+1 < n
k+1
such that 0 < [n
k
, n
k
+ 1] < for all k but
k=1
[n
k
, n
k
+1] = . Given
N > 0, dene f
N
(x) = 1 for x [n
k
, n
k
+1] and k = 1, . . . , N. Extend f
N
to a
compactly supported continuous function on R, and show that f
N
, ) .
D.28 Hint: The Uniform Boundedness Principle implies that weakly conver-
gent sequences are bounded (see Exercise C.117).
Exercises from Appendix E
E.42 Hint: Given X
, r > 0, and x X, the set

U = z C : [x, ) z[ < r
is an open ball in the complex plane. Hence
1
(U) T .
Additional Problems from Appendix E
E.1 Hints: (a) A net f
i
iI
in T(R) converges to f with respect to this
topology if
x
(f f
i
) 0 for each individual x R.
(b) Find f
n
such that c
n
f
n
0 pointwise no matter what scalars c
n
we
choose. If | | is a norm on T(R), consider c
n
= 1/|f
n
|.
E.3 Hint: (c) This topology is not translation-invariant.
E.6 Hint: (a) (b). Suppose that S is strongly closed and choose x / S. By
HahnBanach, there exists a X
such that [
S
= 0 and x, ) = 1. Show
that U =
1
(C0) is open in the weak topology, and use this to show that
XS is weakly open.
E.7 Hints: (a) Plancherels Equality.
(b) Consider y =
c
n
e
n
where c
n
2 = 1/n and all other c
n
= 0.
(c) If U is any open neighborhood of 0 in the weak topology, then there
exist r > 0 and y
1
, . . . , y
N
such that the base element B =
N
j=1
B
yj
r
(0)
satises 0 B U. If n
1/2
e
n
does not belong to B for any n, apply the
Plancherel Equality to |y
1
|
2
+ +|y
N
|
2
to obtain a contradiction.
E.8 Hints: (b) By part (a),
n
nN
is contained in the closed unit ball in
(
, which by Alaoglus Theorem is weak* compact.

(c) Let be a weak* accumulation point of
n
nN
. Fix j and let e
j
denote the jth standard basis vector. The open strip
B
ej
r
() =
_
(
: [e
j
, )[ < r
_
is a weak* open neighborhood of . Show that we cannot have =
n
for
any n. It then follows from Problem A.18 that there must be innitely many
distinct
n
k
in B
ej
r
(). Use this to show that if = x for some x
, then
x = 0 and hence = 0. Finally, to show that ,= 0, consider x = (1, 1, 1, . . . ).
References
[Bab48] K. I. Babenko, On conjugate functions (Russian), Doklady Akad. Nauk
SSSR (N. S.), 62 (1948), pp. 157160.
[Bag92] L. W. Baggett, Functional analysis, Marcel Dekker, New York, 1992.
[Ben76] J. J. Benedetto, Real Variable and Integration, B. G. Teubner, Stuttgart,
1976.
[Ben97] J. J. Benedetto, Harmonic Analysis and Applications, CRC Press, Boca
Raton, FL, 1997.
[BeL76] J. Bergh and J. Lofstrom, Interpolation Spaces, An Introduction, Springer
Verlag, New York, 1976.
[BBT97] A. M. Bruckner, J. B. Bruckner, and B. S. Thomson, Real Analysis,
Prentice-Hall, Upper Saddle River, NJ, 1997.
[Con78] J. B. Conway, Functions of One Complex Variable, Second Edition,
SpringerVerlag, New YorkBerlin, 1978.
[Con90] J. B. Conway, A Course in Functional Analysis, Second Edition, Springer
[Con00] J. B. Conway, A Course in Operator Theory, American Mathematical
Society, Providence, RI, 2000.
[Dau92] I. Daubechies, Ten Lectures on Wavelets, SIAM, Philadelphia, 1992.
[DM72] H. Dym and H. P. McKean, Fourier Series and Integrals, Academic Press,
New YorkLondon, 1972.
[Enf73] P. Eno, A counterexample to the approximation problem in Banach
spaces, Acta Math., 130 (1973), pp. 309317.
[Fol92] G. B. Folland, Fourier Analysis and its Applications, Wadsworth and
Brooks/Cole, Pacic Grove, CA, 1992.
[Fol95] G. B. Folland, A Course in Abstract Harmonic Analysis, CRC Press, Boca
Raton, FL, 1995.
[Fol99] G. B. Folland, Real Analysis, Second Edition, Wiley, New York, 1999.
[GG01] I. Gohberg and S. Goldberg, Basic Operator Theory, Birkhauser, Boston,
2001 (reprint of the 1981 original).
[Gro01] K. Grochenig, Foundations of Time-Frequency Analysis, Birkhauser,
Boston, 2001.
[Haa10] A. Haar, Z ur Theorie der Orthogonalen Funktionensystemen, Math. Ann.,
69 (1910), pp. 331371; English translation in [HW06], pp. 155188.
[Hei10] C. Heil, A Basis Theory Primer, Birkhauser, Boston, 2010.
226 References
[HW06] C. Heil and D. F. Walnut, eds., Fundamental Papers in Wavelet Theory,
Princeton University Press, Princeton, NJ, 2006.
[Jam51] R. C. James, A non-reexive Banach space isometric with its second con-
jugate space, Proc. Nat. Acad. Sci. U. S. A., 37 (1951), pp. 174177.
[Kat04] Y. Katznelson, An Introduction to Harmonic Analysis, Third Edition,
Cambridge University Press, Cambridge, 2004.
[Kel50] J. L. Kelley, The Tychono product theorem implies the axiom of choice,
Fund. Math., 37 (1950), pp. 7576.
[LT77] J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces, I, Springer-
[LT79] J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces, II, Springer-
[Mal98] S. Mallat, A Wavelet Tour of Signal Processing, Academic Press, San
Diego, CA, 1998.
[MH87] J. E. Marsden and M. J. Homan, Basic Complex Analysis, Second Edi-
tion, W. H. Freeman, New York, 1987.
[Mar69] J. Marti, Introduction to the Theory of Bases, Springer-Verlag, New York,
1969.
[Mun75] J. R. Munkres, Topology: A First Course, Prentice-Hall, Englewood Clis,
NJ, 1975.
[Rud91] W. Rudin, Functional analysis, Second Edition. McGraw-Hill, New York,
1991.
[Sch11] I. Schur, Bemerkungen zur Theorie der beschrankten Bilinearformen mit
unendlichvielen Veranderlichen [Remarks on the theory of bounded bilin-
ear forms with innitely many variables], J. f ur Reine und Angew. Math.,
140 (1911), pp. 128.
[Sin70] I. Singer, Bases in Banach Spaces I, Springer-Verlag, New York, 1970.
[ST76] I. M. Singer and J. A. Thorpe, Lecture Notes on Elementary Topology
and Geometry, SpringerVerlag, New YorkHeidelberg, 1976 (reprint of
the 1967 edition).
[SS03a] E. M. Stein and R. Shakarchi, Fourier Analysis, An Introduction, Prince-
ton University Press, Princeton, NJ, 2003.
[SS03b] E. M. Stein and R. Shakarchi, Complex Analysis, Princeton University
Press, Princeton, NJ, 2003.
[SW71] E. M. Stein and G. Weiss, Introduction to Fourier Analysis on Euclidean
Spaces, Princeton University Press, Princeton, NJ, 1971.
[Str94] R. S. Strichartz, A Guide to Distribution Theory and Fourier Transforms,
CRC Press, Goca Raton, FL, 1994.
[SN96] G. Strang and T. Nguyen, Wavelets and Filter Banks, Wellesley
Cambridge Press, Wellesley, MA, 1996.
[Wal02] D. F. Walnut, An Introduction to Wavelet Analysis, Birkh user Boston,
Inc., Boston, MA, 2002.
[WZ77] R. L. Wheeden and A. Zygmund, Measure and Integral, Marcel Dekker,
New York-Basel, 1977.
[You01] R. Young An Introduction to Nonharmonic Fourier Series, Revised First
Edition, Academic Press, San Diego, 2001.
Index of Symbols
Miscellaneous Symbols
Symbol Description Reference
a.e. Almost everywhere Notation B.17
B
r
(f) Open ball of radius r centered at f Denition A.5
B
Borel -algebra on R Denition D.3

Ball
H
Closed unit ball in a Hilbert space Notation C.52
C Complex plane General Notation
d Metric Denition A.1
D Dierentiation operator Notation (C.3)
dx Lebesgue measure Example D.6
ess sup Essential supremum Denition B.18
f g Orthogonal vectors Denition A.91
L -algebra of Lebesgue measurable sets Denition B.6
M
Orthogonal complement Denition A.99

M N Orthogonal direct sum of subspaces Denition A.93
N Natural numbers, 1, 2, 3, . . . General Notation
Q Set of rational numbers General Notation
R Real line General Notation
span(E) Finite linear span of E Denition A.73
span(E) Closed linear span of E Denition A.76
T (c) Topology generated by c Exercise A.38
X
Dual space of X Denition C.2

x
n
w
x Weak convergence Example E.7
x
n
w*
x Weak* convergence Example E.8
Z Integers, . . . , 1, 0, 1, . . . General Notation
,
a
distribution Example E.33
(X, X
) Weak topology on X Section E.6

228 References
(X
, X) Weak* topology on X
Section E.6
(c) Sigma algebra generated by c Exercise D.2
Spaces
AC[a, b] Absolutely continuous functions on [a, b] Denition B.84
AC
loc
(R) Locally absolutely continuous functions on R Denition B.84
B(X, Y ) Bounded linear operators from X to Y Denition C.2
B
0
(H, K) Compact operators from H to K Denition C.53
B
00
(H, K) Finite-rank operators from H to K Section C.7
B
1
(H) Trace-class operators on H Denition C.80
B
2
(H) HilbertSchmidt operators on H Denition C.69
B
p
(H) Schatten class Denition C.78
BV[a, b] Functions of bounded variation on [a, b] Denition B.66
BV(R) Functions of bounded variation on R Denition B.68
c
00
Finite sequences Exercise A.63
c
0
Sequences vanishing at innity Exercise A.63
C
b
(R) Bounded continuous functions Exercise A.20
C
0
(R) Continuous functions vanishing at innity Example A.20
C
c
(R) Continuous, compactly supported functions Exercise A.20
C
m
(R) m-times dierentiable functions Section A.3.2
C
m
b
(R) m-times dierentiable functions with
bounded derivatives Exercise A.21
C
m
0
(R) m-times dierentiable functions decaying
at Section A.3.2
C
m
c
(R) m-times dierentiable functions with
compact support Section A.3.2
C
(R) Innitely dierentiable functions Section A.3.2

C
b
(R) Innitely dierentiable functions with
bounded derivatives Section A.3.2
C
0
(R) Innitely dierentiable functions decaying
at Section A.3.2
C
c
(R) Innitely dierentiable functions with
compact support Section A.3.2
C[a, b] Continuous functions on [a, b] Equation (A.7)
C(K) Continuous functions supported in K Section D.8
C
(K) C
functions supported in K Denition E.37

T
(R) Distributions Denition E.40

c
(R) Compactly supported distributions Denition E.34

References 229
p
(I) p-summable sequences Denition A.15
L
p
(E) Lebesgue space of p-integrable functions Denition B.48
L
1
() -integrable functions Denition D.32
L
1
loc
(R) Space of locally integrable functions Denition B.62
Lip[a, b] Lipschitz functions on [a, b] Denition B.71
M
b
(R) Complex Borel measures Denition D.61
o(R) Schwartz space Example E.4
o
(R) Tempered distributions Denition E.31

Operations on Sets
E Closure of a set E Denition A.58
[E[
e
Exterior Lebesgue measure of E R
d
Denition B.2
[E[ Lebesgue measure of E R
d
Denition B.6
vol(Q) Volume of a cube Denition B.1
Metrics, Norms, Seminorms, Inner Products
d Metric Denition A.1
Equivalent norms Denition A.54
| | Norm or seminorm Denition A.6
| |
p

p
or L
p
norm Denition A.15
| |
B1
Trace-class operator norm Denition C.80
| |
B2
HilbertSchmidt operator norm Denition C.71
| |
HS
HilbertSchmidt operator norm Denition C.71
| |
Bp
Schatten-class operator norm Denition C.78
, ) Inner product or semi-inner product Denition A.22
, ) Sesquilinear form Notation C.36
, ) Notation for functionals Notation C.36
Functions and Operations on Functions
f > a = x : f(x) > a Notation B.21
f g Convolution of f and g Exercise C.23
f
Negative part of f Denition B.41

230 References
f
+
Positive part of f Denition B.41
f Complex conjugate of f
f Involution of f Exercise C.33

f
(x) Classical pointwise derivative of f at x

f
n
f Monotone increasing sequence Section B.5
f
n
f a.e. Pointwise a.e. convergence
f
n
m
f Convergence in measure Denition B.26
g h Tensor product function Example C.17
supp(f) Support of f Notation A.1
V [f; a, b] Variation of f on [a, b] Denition B.66
V [f; R] Variation of f on R Denition B.68
E
Characteristic function of a set E General Notation
Riemann zeta function Equation (F.5)
Operators
A 0 A is a positive operator Denition C.50
A > 0 A is a positive denite operator Denition C.50
A
Adjoint of A Section C.6

g h Rank one tensor product operator Notation C.18
ker(T) Kernel (nullspace) of an operator T Denition C.1
L
k
Integral operator with kernel k Denition C.16
range(T) Range of an operator T Denition C.1
s
n
Singular numbers of an operator Denition C.73
Measures and Operations on Measures
,
a
measure Exercise D.8
Mutually singular measures Denition D.12
Negative variation of a signed measure Denition D.16
+
Positive variation of a signed measure Denition D.16
r
Real part of a complex measure Denition D.16
i
Imaginary part of a complex measure Denition D.16
Involution of a measure Problem D.20
[[ Total variation of a measure Denition D.57
|| Norm of a bounded measure Denition D.61
Index
C
-algebra, 104
F-set, 44
G
-set, 44
measure, 142
-eigenspace, 85
-eigenvector, 85
-almost everywhere, 147
-independent sequence, 38, 134
-algebra, 43, 140
generated by E, 140
-nite measure, 144
absolute value operator, 117
absolutely continuous function, 70
absolutely continuous measure, 150, 153
absolutely convergent series, 6
accumulation point, 18
adjoint, 100, 127
Alaoglus Theorem, 193
algebra, 93
C
, 104
, 43, 140
Banach, 93
Borel , 140
commutative, 93
complex, 93
normed, 93
real, 93
with identity, 93
almost everywhere, 46, 147
alternating harmonic series, 33
analysis operator, 105
analytic function, 197
angle between vectors, 14
antilinear isometric isomorphism, 84
antilinear operator, 81
associated family of coecient
functionals, 132
Axiom of Choice, 39, 43, 86, 139, 192,
193, 205
Baire Category Theorem, 128
Banach algebra, 93
Banach space, 5
BanachAlaoglu Theorem, 193
BanachSteinhaus Theorem, 129
BanachZarecki Theorem, 75
base for a topology, 174
basis
Hamel, 30, 86, 218
orthonormal, 37
Schauder, 31, 37, 132
standard basis, 29
vector space, 30
basis constant, 133
Beppo Levi Theorem, 51
Bessels inequality, 36
bijection, 81
biorthogonal sequence, 39, 134
Borel -algebra, 140
Borel measurable function, 146, 156
Borel measure, 141
Borel sets, 140
bounded measure, 141
bounded operator, 82
bounded variation, 64
232 Index
canonical embedding, 126
canonical projection, 192
Cantor set, 46
CantorLebesgue function, 69
Caratheodorys Criterion, 45
Cauchy sequence, 4
CauchyBunyakowskiSchwarz Inequal-
ity, 13
chain, 205
closed nite span, 29
Closed Graph Theorem, 131
closed set, 15
closure of a set, 23
compact operator, 106
compact set, 25
compactly supported distribution, 185
complete inner product space, 14
complete metric space, 4
complete normed space, 5
complete set, 25, 29
complex Borel measure, 152
complex conjugate measure, 156
complex derivative, 197
componentwise convergence, 11
continuity
at a point, 20
uniform, 27
continuity at a point, 83
continuity equals boundedness, 181
continuous function, 20
continuously invertible operator, 84
convergence
absolute, 6
in a metric space, 4
in a topological vector space, 179
in measure, 48
of a net, 18
unconditional, 32
weak, 136, 173
weak*, 136, 173
convergent sequence, 4
convergent series, 5
convex set, 6
coordinate vector, 22
counting measure, 142
cube, 41
volume of, 41
Daubechies
scaling function, 78
wavelet, 79
delta
measure, 142
dense subset, 23
densely dened operator, 82
derivative
complex, 197
RadonNikodym, 151
Devils staircase, 69
Dirac
measure, 142
direct sum, 35
orthogonal, 35
directed set, 17, 179
Dirichlet
series, 199
distance from a point to a set, 35
distribution, 188
compactly supported, 185
order, 189
tempered, 184
Dominated Convergence Theorem, 149
double integral, 61
dual index, 1
dual space, 82, 179, 184
DvoretzkyRogers Theorem, 34
Eberlein-
Smulian Theorem, 196

eigenspace, 85
eigenvalue, 85
eigenvector, 85
entire function, 197
equivalent norms, 21
essential supremum, 46
essentially bounded function, 46, 55
Euclidean norm, 7
everywhere dierentiable function, 65
exact sequence, 134
exponential function, 199
extended -integrable function, 148
exterior Lebesgue measure, 42
F. Rieszs Lemma, 28
Fatous Lemma, 52, 148
nite linear span, 28
nite measure, 141
nite sequences, 24
nite-rank operator, 81, 107
Index 233
nitely linearly independent, 30
rst category, 127
Frechet space, 183
Fubinis Theorem, 62, 157
function
absolutely continuous, 70
analytic, 197
Borel measurable, 146
CantorLebesgue, 69, 78
conjugate linear, 12
continuous, 20
Daubechies scaling, 78
Devils staircase, 69
entire, 197
essentially bounded, 46, 55
everywhere dierentiable on [a, b], 65
exponential, 199
extended -integrable, 148
Holder continuous, 78
holomorphic, 197
integrable, 55, 148
Lebesgue measurable, 47
Lipschitz, 65, 78
locally integrable, 59
monotone increasing, 64, 66
open mapping, 130
renable, 69
Riemann zeta, 200
scaling, 79
simple, 49, 146
singular, 70
sublinear, 123
uniformly continuous, 27
wavelet, 79
functional, 82
linear, 82
positive, 162
Fundamental Theorem of Calculus, 72
generated topology, 16
Holder continuous function, 78
Haar
wavelet, 79
Haar system, 80, 136
Haar, Alfred, 79
Hahn decomposition, 143
HahnBanach Theorem, 122124
Hamel basis, 30, 86, 206, 218
Hardys Inequalities, 63
harmonic series, 33
Hausdor topological space, 16
Hermitian
matrix, 100
operator, 100
Hilbert space, 14
HilbertSchmidt integral operator, 89,
109
HilbertSchmidt norm, 115
HilbertSchmidt operator, 115
holomorphic function, 197
homeomorphism, 20
ideal, 94
generated by, 94
left, 93
right, 94
two-sided, 94
idempotent operator, 86
independence
-independence, 38, 134
induced
metric, 5
norm, 12
seminorm, 12
topology, 15, 176
inductive limit topology
on Cc(R), 160
on C
c
(R), 187
injection, 81
inner Lebesgue measure, 46
inner product, 12
inner product space, 12
complete, 14
inner regular measure, 158
integrable function, 55, 148
integral
Lebesgue, 50
with respect to a positive measure,
147
integration by parts, 73
invariant subspace, 100
Inverse Mapping Theorem, 131
involution, 94, 156
isometric isomorphism, 84
antilinear, 84
isometry, 84
isomorphism
234 Index
antilinear isometric, 84
isometric, 84
topological, 20, 84
iterated integral, 61
Jordan decomposition
of functions of bounded variation, 66
of measures, 143
Jordan Decomposition Theorem, 143
kernel, 81
Kronecker delta, 1
Lebesgue decomposition, 151
Lebesgue Dierentiation Theorem, 71
Lebesgue Dominated Convergence
Theorem, 53
Lebesgue integral, 50
Lebesgue measurable function, 47
Lebesgue measurable set, 43
Lebesgue measure, 43
exterior, 42
inner, 46
Lebesgue point, 71
Lebesgue set, 71
Lebesgue space, 55
LebesgueRadonNikodym Theorem,
151, 153
left ideal, 93
left-shift operator, 85
linear functional, 82
notation for, 96
linear operator, 81
linear ordering, 205
linearly independent, 30
Liouvilles Theorem, 198
Lipschitz constant, 65
Lipschitz function, 65, 78
locally absolutely continuous function,
70
locally nite measure, 141, 158
locally integrable function, 59
Lusins Theorem, 159
matrix
Hermitian, 100
operator norm, 86
Maximum Modulus Principle, 198
meager, 127
measurable function, 47
measurable set, 43
measure
, 142
-nite, 144
Borel, 141
bounded, 141
complex conjugate, 156
counting, 142
Dirac, 142
nite, 141
inner regular, 158
locally nite, 141, 158
negative variation, 143
outer regular, 158
positive, 141
positive variation, 143
Radon, 158
regular, 158
signed, 141
total variation, 144, 154
metric, 3
corresponding to a family of
seminorms, 183
induced, 5
translation-invariant, 23
metric space, 3
complete, 4
metrizable topology, 16
minimal sequence, 38, 134
Minkowskis Inequality, 8, 56
Minkowskis Integral Inequality, 62
Monotone Convergence Theorem, 51,
147
monotone increasing
function, 64
sequence, 51
multiplicity, 114
multiresolution analysis, 215
mutually singular, 143
natural embedding, 126
negative variation, 66
negative variation measure, 143
neighborhood, 15
nonharmonic trigonometric polynomial,
200
nonmeager, 127
nonoverlapping intervals, 70
Index 235
norm, 5
equivalent, 21
Euclidean, 7
HilbertSchmidt, 115
induced, 12
operator, 82
Schatten class, 118
trace-class, 121
norm topology, 15, 189
norm-preserving operator, 84
normable topology, 16
normal operator, 100
normed algebra, 93
normed linear space, 5
normed space
complete, 5
nowhere dense, 127
null set, 143
nullspace, 81
open ball, 4
open mapping, 130
Open Mapping Theorem, 131
open set, 15
open strip, 176
operator, 81
absolute value, 117
adjoint, 100
analysis, 105
antilinear, 81
bijective, 81
bounded, 82
compact, 106
continuously invertible, 84
densely dened, 82
nite-rank, 81, 107
Hermitian, 100
HilbertSchmidt, 115
HilbertSchmidt integral, 89, 109
idempotent, 86
injective, 81
left-shift, 85
linear, 81
norm, 82
normal, 100
positive, 103
positive denite, 103
right-shift, 85
Schatten class, 118
self-adjoint, 100, 101
surjective, 81
synthesis, 105
trace-class, 121
unbounded, 101
unitary, 85
Volterra, 111
operator norm, 82
order of a distribution, 189
ordering
linear, 205
total, 205
orthogonal complement, 36, 125
orthogonal direct sum, 35
orthogonal projection, 36, 86
orthogonal subspaces, 35
orthogonal vectors, 34
orthonormal basis, 37
orthonormal vectors, 34
outer Lebesgue measure, 42
outer regular measure, 158
Parallelogram Law, 13
Parseval equality, 37
partial isometry, 117
partial order, 205
partial sum operator, 132
partial sums, 5
Plancherel equality, 37
point mass, 142
polar decomposition, 117, 155
Polar Identity, 13
polynomial
nonharmonic trigonometric, 200
trigonometric, 200
positive denite operator, 103
positive functional, 162
positive measure, 141
positive operator, 103
positive variation, 66
positive variation measure, 143
power series, 198, 209
pre-Hilbert space, 12
precompact set, 27
product topology, 17, 192
Pythagorean Theorem, 35
radius of convergence, 198
Radon measure, 158
236 Index
RadonNikodym Theorem, 151
range, 81
rank, 81
rare, 127
renable function, 69
renement equation, 69, 78, 80
reexive, 126
regular measure, 158
reverse inclusion, 18
Riemann hypothesis, 200
Riemann zeta function, 200
Riesz Representation Theorem, 95, 163,
167
RieszThorin Interpolation Theorem,
201
right ideal, 94
right-shift operator, 85
scaling function, 79
Daubechies, 78
Schatten class, 118
Schauder basis, 31, 132
associated family of coecient
functionals, 132
basis constant, 133
partial sum operators, 133
unconditional, 136
Schurs Test, 89
Schwartz space, 171
second category, 127
self-adjoint operator, 100, 101
semi-inner product, 12
seminorm, 4, 176
induced, 12
separable space, 24
sequence
-independent, 38, 134
biorthogonal, 39, 134
Cauchy, 4
complete, 29
convergent, 4
exact, 134
fundamental, 29
minimal, 38, 134
total, 29
series
absolutely convergent, 6
convergent, 5
Dirichlet, 199
power, 198
Taylor, 199
unconditionally convergent, 32
sesquilinear form, 12, 95, 160
set
complete, 29
rst category, 127
linearly independent, 30
meager, 127
nonmeager, 127
nowhere dense, 127
rare, 127
second category, 127
signed Borel measure, 141
simple function, 49, 146
standard representation, 49, 146
singular function, 70
singular numbers, 117
singular vectors, 117
span
closed, 29
nite, 28
Spectral Theorem, 112
standard basis, 29
standard representation of a simple
function, 49, 146
strong topology, 15, 189
subbase for a topology, 16, 174
sublinear function, 123
submultiplicative norm, 92
subspace
invariant, 100
support, 9
surjection, 81
synthesis operator, 105
Taylor series, 199
Tchebyshevs Inequality, 51
tempered distribution, 184
tensor product, 88, 110
Tonellis Theorem, 61, 157
topological isomorphism, 20, 84
topological space, 15
Hausdor, 16
topological vector space, 175
topology, 15
base for, 174
generated by E, 16, 174
Index 237
induced from a family of seminorms,
176
induced from a metric, 15
inductive limit topology on Cc(R),
161
inductive limit topology on C
c
(R),
187
metrizable, 16
norm, 15, 189
normable, 16
product, 17, 192
relative, 16
strong, 15, 189
subbase for, 16, 174
translation-invariant, 175
uniform topology on Cc(R), 160
weak, 173
weak*, 173
total ordering, 205
total variation measure, 144, 154
totally bounded set, 25
trace-class operator, 121
translation-invariant metric, 23
translation-invariant topology, 175
Triangle Inequality, 3, 4
trigonometric polynomial, 200
Tychonos Theorem, 193
unbounded operator, 101
unconditional basis, 136
unconditionally convergent series, 32
Uniform Boundedness Principle, 129
uniform continuity, 27
uniform norm, 9
uniform topology on Cc(R), 160
unitary operator, 85
upper bound, 205
Urysohns Lemma, 40
variation of a function, 64
vector space basis, 30
Volterra operator, 111
volume of a cube, 41
wavelet, 79
Daubechies, 79
Haar, 79
weak convergence, 136
weak topology, 172, 189
weak* convergence, 136
weak* topology, 173, 190
Weierstrass Approximation Theorem,
31
weighted
p
space, 11
weighted L
p
space, 99
Youngs Inequality, 91
Zorns Lemma, 39, 218

An Analysis Companion - Heil

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

An Analysis Companion - Heil

Uploaded by

Copyright:

Available Formats

Christopher Heil

and Reexivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

(I) consists of all sequences of scalars x = (x

(see the proof by picture in Figure A.1, or Problem A.10).

(I) satisfy |x|

t + (1 ) for t > 0, with

and = 1/p to show that ab a

, with equality if and only if

(R) are important examples of

(I) is not separable.

for some J, and

is open there must exist some r > 0 such that B

, which is a contradiction. Hence the

, so E is covered by nitely many of

is a closed subspace of H (even if A is not), and

, the orthogonal direct sum of M and M

= 0, i.e., the only vector orthogonal to every

-set if there exist nitely or

-set if there exist nitely or countably many closed

-set and [Z[ = 0.

-set and [Z[ = 0.

sets, zero measure sets, and nally arbitrary measurable sets).

-set H A such that [H[ = [A[

-set such that [H[ = [E[

= , so the Lebesgue integral of f on [0, ) does not exist. In essence,

(E) consists of all those measurable functions

(R) as the Lebesgue space of essentially bounded functions.

denote the dual index to p. If f L

a.e. (Problem B.12).

(R). However, in doing so we are really identifying C

(R) under the equivalence relation , i.e., if f C

(R) (really an equivalence class

-norm dened above.

(E) then it follows that f

(R) is not sepa-

(x) exists and is nite for all x [a, b].

is bounded on [a, b], then

are both continuous

] [a, b], then V [f; a

is the sum of the positive terms of S

is the sum of the

= f(b)f(a). The next exercise

[f; a, b] = V [f; a, b].

[f; a, b] is nite, then they

[f; a, b] = f(b) f(a).

[f; a, x] = f(x) f(a).

[f; a, x], and

f(b) f(a). (B.11)

(x) a.e. as k . In particular, f

(x) exists for a.e. x, and f

and the total variation function.

(x)[ for almost every x [a, b].

(t) dt, x [a, b].

is well-dened for each x [a, b]. Moreover, by the Lebesgue Dierentiation

a.e. Hence (Gg)

= 0 a.e., so the function Gg is singular

(t) dt = 0, conrming the fact that is not absolutely

, x [a, b]. (B.14)

exists a.e. and

= 0 a.e. Thus h is singular.

(x) dx = f(b)g(b) f(a)g(a)

(x) g(x) dx.

is bounded, then f is Lipschitz and hence is absolutely