Beruflich Dokumente
Kultur Dokumente
John Toland
Series Editors
Nicola Bellomo, Torino, Italy
Michele Benzi, Pisa, Italy
Palle Jorgensen, Iowa, USA
Tatsien Li, Shanghai, China
Roderick Melnik, Waterloo, Canada
Otmar Scherzer, Linz, Austria
Benjamin Steinberg, New York, USA
Lothar Reichel, Kent, USA
Yuri Tschinkel, New York, USA
George Yin, Detroit, USA
Ping Zhang, Kalamazoo, USA
Titles from this series are indexed by Web of Science, Mathematical Reviews, and
zbMATH.
John Toland
123
John Toland
Department of Mathematical Sciences
University of Bath
Bath, UK
Mathematics Subject Classification (2010): 46E30, 28C15, 46T99, 26A39, 28A25, 46B04
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
Assuming some familiarity with Lebesgue measure, integration and related func-
tional analysis summarised in Chap. 2, this is an exposition of topics that arise
when identifying elements of the dual space of L1 ðX; L; ‚Þ with finitely additive
measures on a r-algebra L when the measure ‚ is complete and r-finite. Such a
representation has its origins in the independent work of Fichtenholz and
Kantorovitch [14] and Hildebrandt [20] and culminated in a general abstract theory
due to Yosida and Hewitt [35] in 1952. However, even now it is not unusual ([16] is
an exception) for books on measure theory to give a detailed account of the dual
space of Lp ðX; L; ‚Þ, 1 6 p\1, while relegating the case p ¼ 1 to references, e.g.
[12, p. 296] which invokes the theory of finitely additive measures on algebras such
as is developed in [35].
An explanation may be that a study of finitely additive measures on algebras
necessitates a possibly unwelcome diversion from the mainstream theory of
countably additive measures that suffices for p 2 ½1; 1Þ. Whatever the reason, a
consequence is that L1 ðX; L; ‚Þ has acquired an aura of mystery, to the extent that
it is often not very clear beyond the mere definition what is meant by saying that a
bounded sequence in L1 ðX; L; ‚Þ is weakly convergent.
The aim here is to take Yosida and Hewitt theory on r-algebras beyond the
representation theorem for L1 ðX; L; ‚Þ , pointing out some of its consequences for
measurable functions generally and in particular for weak convergence of
sequences in L1 ðX; L; ‚Þ. The target audience is anyone who feels nervous about
representing the dual of L1 ðX; L; ‚Þ by finitely additive measures in the knowledge
that there exist uncountably many, linearly independent, finitely additive measures
m > 0 defined on the Lebesgue r-algebra of ð0; 1Þ with the property that
Z 11k Z 1
u dm ¼ 0 for all u 2 L1 ð0; 1Þ and k 2 N; but 1 dm ¼ 1: ð†Þ
0 0
An essential goal will be to come to terms with observations such as this one.
v
vi Preface
In their seminal work, Yosida and Hewitt [35] studied general Banach spaces
L1 ðX; M; N Þ of essentially bounded measurable functions, where measurability is
determined by an algebra M (closed under complementation and finite unions) and
essential boundedness is defined in terms of a family N M (closed under
countable unions with the added property that A B 2 N implies A 2 N ) that
mimics null sets. Obviously, L1 ðX; L; ‚Þ is a special case of L1 ðX; M; N Þ but in
general no measure of any kind is involved in the definition of L1 ðX; M; N Þ.
However, although [35] shows that the dual of L1 ðX; M; N Þ can be expressed in
terms of finitely additive measures, the exposition here is restricted to L1 ðX; L; ‚Þ
because
For a r-finite measure space the ultimate aim is to develop theory sufficient to
characterise weakly convergent sequences in L1 ðX; L; ‚Þ in terms of their ‚-
almost-everywhere pointwise behaviour. However, in the process, when ðX; ‰Þ is a
locally compact Hausdorff topological space and ðX; B; ‚Þ is a corresponding Borel
measure space, there emerges a natural way to localise weak convergence.
A sequence is weakly convergent in L1 ðX; B; ‚Þ if and only if it is weakly con-
vergent at every point x0 in the one-point compactification of ðX; ‰Þ. Here, weak
convergence at x0 is defined in terms of functionals which are zero outside every
neighbourhood of x0 ; for an example of such, see (†).
The essential range RðuÞðx0 Þ1 of a Borel measurable function u at x0 is similarly
defined in terms of those elements of L1 ðX; B; ‚Þ which are localised at x0 . Since
it need not be a singleton, RðuÞðx0 Þ can be interpreted as a multivalued repre-
sentation of the fine structure at x0 of u 2 L1 ðX; B; ‚Þ which is intimately related to
weak convergence.
The Literature
In her Foreword to the monograph by Bhaskara Rao and Bhaskara Rao [6], Dorothy
Maharam Stone cites Salomon Bochner as having said that “contrary to popular
mathematical opinion finitely additive measures were more interesting, more dif-
ficult to handle, and perhaps more important than countably additive measures”.
Oxtoby [25] described [6] as a comprehensive account of finitely additive measures
which effectively organises a large body of material that is widely scattered in the
literature and deserves to be better known, and in their preface the authors them-
selves described it as a reference book as well as a textbook.
The origins of this theory are to be found in the early days of modern integration
theory when there were many contributors: see [12, §III.15, p. 233 and §IV.16,
p. 3882] and the comprehensive bibliography with notes in [6]. However, pre-
sumably because they could not match the versatility of Lebesgue’s theory of
integration and the power of its convergence theorems, finitely additive measures
seem to have fallen out of fashion. Nevertheless, they continue to have significant
roles in, for example, mathematical economics, probability, statistics, optimization,
control theory and analysis [7, 9, 25, 35].
In a series of three papers on additive set functions on abstract topological
spaces, A. D. Alexandroff [2] studied bounded regular finitely additive measures
that represent linear functionals on spaces of continuous functions. On a similar
theme, but in a more general setting, a much-cited reference for the dual of
L1 ðX; L; ‚Þ is Dunford and Schwartz [12, p. 296] which covers the theory of
finitely additive set functions on algebras and includes extensive historical notes.
For a recent account, see Fonseca and Leoni [16, Theorem 2.24], and Aliprantis and
Border [3] for the abstract theory in which it is embedded.
It will soon be apparent that key results, including (†), rely on the axiom of
choice. For a discussion of the role of the axiom of choice, geometrical and
paradoxical aspects of finitely additive measures, and their invariance under group
actions on X, see Tao [32]. Oxtoby’s commentary [25] is of independent interest.
A key role is played throughout by the set G of finitely additive measures that
take only the values 0 and 1 and the observation that L1 ðX; L; ‚Þ is isometrically
isomorphic to a space of real-valued continuous functions on ðG; ¿Þ with the
maximum norm, where ¿ is a compact Hausdorff topology. Further analysis of G in
a Borel setting then leads to localization, and to other developments mentioned
above and outlined in the Introduction.
What follows is in large part an extension of a simplified version of Yosida and
Hewitt [35], set out in the notation and terminology of Chap. 2.
Acknowledgements I am indebted to Charles Stuart (Lausanne) for many things including his
encouragement of this project. I am grateful to Anthony Wickstead (Belfast) who obtained for me
a copy of [33] and drew my attention to [34], and to Geoffrey Burton (Bath) and Eugene
Shargorodsky (King’s London) for their interest and many comments. In addition, I would like to
thank Eugene Shargorodsky who contributed Sect. 9.4 and Mauricio Fernández (Stuttgart) who on
a visit to Cambridge asked a question that the account that follows attempts to answer.
2
The reference to Theorem 8.15 on p. 388 is a misprint; 8.15 is a Definition and obviously
Theorem 8.16 was intended.
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Notation and Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 Partial Ordering and Zorn’s Lemma . . . . . . . . . . . . . . . . . . . . . 7
2.2 Algebras and r-Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Measurable Sets and Measurable Functions . . . . . . . . . . . . . . . 9
2.4 Measures and Real Measures . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.5 Splitting Families of Measurable Sets . . . . . . . . . . . . . . . . . . . . 14
2.6 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.7 Function Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.8 Dual Spaces and Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.9 Functional Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.10 Point Set Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3 L1 and Its Dual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4 Finitely Additive Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.1 Definition, Notation and Basic Properties . . . . . . . . . . . . . . . . . 31
4.2 Purely Finitely Additive Measures . . . . . . . . . . . . . . . . . . . . . . 36
4.3 Canonical Decomposition: baðLÞ ¼ RðLÞ PðLÞ . . . . . . . . . . 38
4.4 L1 ðX; L; ‚Þ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5 G: 0–1 Finitely Additive Measures . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.1 G and Ultrafilters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.2 G and the ‚-Finite Intersection Property . . . . . . . . . . . . . . . . . 44
ix
x Contents
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Chapter 1
Introduction
Overview
In a normed linear space V , a sequence {vk } converges weakly to v (vk v) if
v ∗ (vk ) → v ∗ (v) for all v ∗ ∈ V ∗ , the dual space of V and, from the uniform bound-
edness principle, weakly convergent sequences are bounded in norm. However, it
has been known since the work of Banach that when V is a complete normed linear
space it may not be necessary to use all elements of V ∗ when testing for weak con-
vergence. Indeed, when C(Z ) denotes the space of real-valued continuous functions
on a compact metric space Z with the maximum norm, he showed that vk v in
C(Z ) if and only if vk (z) → v(z) for all z ∈ Z and {vk } is bounded. To do so he
observed [5, Annexe, Thm. 7] that Dirac δ-functions satisfy conditions for a set W ∗
in the dual space of a Banach space to have the property that
When (X, ) is a locally compact Hausdorff space and (C0 (X, ), · ∞ ) is the
Banach space of real-valued continuous functions on X that vanish at infinity (see
(2.9)), weakly convergent sequences are pointwise convergent because δ-functions
belongs to the dual space of C0 (X, ), and bounded by the uniform boundedness
principle. Conversely, from Theorem 2.37 (Riesz) and Lebesgue’s Dominated Con-
vergence Theorem [15, Thm. 2.24], sequences that are norm-bounded and pointwise
convergent on X are weakly convergent.
In particular, when Z is a compact Hausdorff space, for {vk } ⊂ C(Z) (the space
of real-valued continuous functions on Z with the maximum norm)
vk v0 in C(Z) ⇔ sup vk < ∞ and vk (z) → v0 (z) for all z ∈ Z. (V)
k
it is shown that
R(u) = u dω : ω ∈ G
X
and its relation to the isometric isomorphism is given by (7.2) and (7.3). If the setting
is a Borel measure space (X, B, λ) corresponding to a locally compact Hausdorff
spaces (X, ), these considerations can be localised to points x0 in the one-point
compactification X ∞ of X . This in turn leads to the definition of the essential range
of u ∈ L ∞ (X, B, λ) localised at x0 ∈ X ∞ and the possibility of regarding u as multi-
valued at every point while at the same time being single-valued almost everywhere.
These ideas are closely related to the dual of L ∞ (X, B, λ) and weak convergence.
The fact that L ∞ (X, L, λ) and C(G, τ ) are isometrically isomorphic appears to
be at variance with L ∞ (X, L, λ)∗ being represented by finitely additive measures
while C(G, τ )∗ is represented by σ-additive measures. Moreover, when (X, ) is
a locally compact Hausdorff space, elements of L ∞ (X, B, λ)∗ are represented by
finitely additive measures but their restrictions to C0 (X, ) are also represented by
Borel measures. Both these issues will be addressed.
As explained in the Preface, finitely additive measures will be considered only on
σ-algebras. The classical text [12] and the exhaustive monograph [6] both consider
more general situations from the outset, as does Yosida and Hewitt [35] which is the
motivation and main source for this account of the simplified theory. The approach
here is referred to as Yosida–Hewitt theory.
Layout
Chapter 2 is a brief survey of background material, collected with references from
different sources and organised in a consistent notation that will be used in the chap-
ters that follow. Although most of the material is entirely standard and does not need
to be digested unless a need arises in later chapters, some of it may be less famil-
iar. For example, Sect. 2.5 is a slight extension of Kirk’s [22] application of Baire’s
1 Introduction 3
category theorem to show that if (X, L, λ) has no atoms and G is a denumerable fam-
ily of sets of positive
measure, there exists an infinite family of disjoint sets A ∈ L
such that λ(G A) > 0 and λ(G \ A) > 0 for all G ∈ G. This construction is used
in Example 9.11 when considering possibilities for the localised essential range of
u ∈ L ∞ (X, L, λ) at x0 defined in (9.4).
Chapter 3 begins by stating Theorem 3.1, which is the Yosida–Hewitt representa-
tion of the dual space L ∞ (X, L, λ)∗ as the space of finitely additive measures on L
which are zero on sets of zero λ-measure. These finitely additive measures will be
denoted by L ∗∞ (X, L, λ), which should therefore be identified with, but not confused
with, L ∞ (X, L, λ)∗ , the bounded, real-valued, linear functionals on L ∞ (X, L, λ).
From Theorem 3.1, it follows intuitively that if {Ak } ⊂ L is any sequence of mutu-
ally disjoint sets, then χ Ak 0 in L ∞ (X, L, λ) as k → ∞, where χ Ak denotes the
characteristic function of Ak . Moreover, there are finitely additive measures that rep-
resent non-zero bounded linear functionals f on L ∞ (0, 1) with f (v) = 0 when v
is continuous. Since no such countably additive measure can exist because of the
dominated convergence theorem and Theorem 2.35 (Lusin), observations such as
these reflect the delicacy of Theorem 3.1 and the differences between σ-additive and
finitely additive measures.
Chapter 4 introduces basic notation and definitions, for example, of partial order-
ing, lattice operations, absolute continuity and singularity, for finitely additive mea-
sures. The important notion of pure finite additivity of measures follows, and it is
shown that every finitely additive measure is uniquely the sum of a σ-additive and a
purely finitely additive measure.
Chapter 5 introduces the set G of elements in L ∗∞ (X, L, λ) which take only the
values 0 or 1 on L and explains the sense in which every u ∈ L ∞ (X, L, λ) is constant
ω-almost everywhere when ω ∈ G. Elements ω ∈ G give rise to families of sets
Uω = {E ∈ L : ω(E) = 1} which are ultrafilters (Definition 5.3). The collection of
ultrafilters is denoted by U (Gothic U), and there is a one-to-one correspondence
(Theorem 5.4) between G and U. The existence of ultrafilters, and consequently of
elements of G with certain properties, follow from Zorn’s lemma. G will dominate
subsequent developments.
Chapter 6 defines the integral of essentially bounded measurable functions with
respect to the finitely additive measures that featured in Theorem 3.1 (the Yosida–
Hewitt representation theorem).
In Theorem 6.2, it is noted that for all ω ∈ G and u ∈
L ∞ (X, L, λ), u(x) = X u dω, ω-almost everywhere in the sense of finitely additive
measures (Remark 5.2). Observation (†) in the Preface is justified in Remark 6.1.
The chapter ends with an account of the Valadier–Hensgen example [19, 33] of
purely finitely additive measures on [0, 1] that are definitely not σ-additive but which
coincide with σ-additive Lebesgue measure when integrating continuous functions.
Chapter 7 introduces a compact Hausdorff topology τ on G and from theory
already developed, derives the existence, Theorem 7.4, of an isometric isomorphism
between the Banach algebra L ∞ (X, L, λ) and the space of real-valued continuous
functions C(G, τ ). It is immediate that the functionals corresponding to G have
property (W) in the Introduction.
4 1 Introduction
R(u) = R(u)(x0 ) where R(u)(x0 ) = u dω, ω ∈ G(x0 )
x0 ∈X ∞ X
The set of natural numbers {1, 2, · · · , } is denoted by N, N0 = N {0}, R is the real
numbers, C is the complex numbers, Q is the rational numbers, and extended real
numbers, R {+∞, −∞}, are denoted by R. The empty set is denoted by ∅, and a
set with only one element is called a singleton.
For an arbitrary set S, ℘ (S) denotes the collection of all its subsets, including
the empty set ∅ and S itself. A set S is said to be denumerable or countable if there
is an injection from S into N, and uncountable otherwise. If Sk is denumerable for
all k ∈ N, then k Sk is denumerable. If there is an injection from S into the set
{1, 2, · · · , K } for some K ∈ N, S is said to be finite, and infinite otherwise. If S is
an infinite denumerable set, there is a bijection from S onto N.
Example 2.2 The usual ordering ≤ on R is a partial ordering which is a total ordering
on every subset S of R.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 7
J. Toland, The Dual of L ∞ (X, L, λ), Finitely Additive Measures
and Weak Convergence, SpringerBriefs in Mathematics,
https://doi.org/10.1007/978-3-030-34732-1_2
8 2 Notation and Preliminaries
Lemma 2.4 (Zorn) Suppose is a partial ordering on S and every totally ordered
subset A ⊂ S has an upper bound. The S has a maximal element.
The proof depends on an assumption that for any collection T of subsets of S there
is a set which contains a point from each of the sets in T . More precisely:
The Axiom of Choice
Let S be non-empty and T a non-empty subset of ℘ (S) \ {∅}. Then there exists a
function f : T → S such that f (T ) ∈ T for all T ∈ T .
Conversely, the Axiom of Choice can be proved using Zorn’s lemma. So these two
seemingly different statements are in fact equivalent formulations of a fundamental
axiom of set theory.
Definition 2.5 (Equivalence Relation) A relation ∼ on a set S is called an equiv-
alence relation if for all x, y, z ∈ S, (i) x ∼ x; (ii)x ∼ y ⇔ y ∼ x; (iii) x ∼ y
and y ∼ z ⇒ x ∼ z. The set [x] = y ∈ S : x ∼ y is the equivalence class which
contains x.
K
ϕ(x) = ak χ Ak , ak ∈ R, Ak ∈ L, K ∈ N, (2.2)
k=1
are called simple functions and are measurable. The functions ϕn in (v) above are
non-negative simple functions.
In the measure theory literature, subtly different meanings are often assigned by
different authors to the same terminology. For example, in [15, Chap. 3, p. 85] a
signed measure may have infinite values, whereas in [28, Sect. 6.6, p. 119] it may
not. See also Remark 2.38. For this reason, the terminology chosen for subsequent
chapters is described in some detail below.
Definition 2.6 (Measures) In a measurable space (X, L), a measure λ is an extended
real-valued function on L with
λ(X ) > 0, λ(∅) = 0, λ(A) ≥ 0 for all A ∈ L, and
when E k ∈ L, k ∈ N, and E i E j = ∅, i = j,
λ Ek = λ(E k ). (2.3)
k∈N k∈N
A Borel measure which is both outer and inner regular on every B ∈ B is called
regular.
Example 2.13 (Dirac Measures) For any measurable space (X, L) and x ∈ X , let
δx be defined on L by δx (E) = 1 if x ∈ E and δx (E) = 0 otherwise. Then δx is a
measure, called a Dirac measure at x, and (X, L, δx ) is a finite, complete measure
space in which {x} is an atom.
Example 2.14 (Counting Measure on N) Let X = N, L = ℘ (N) and define λ(E) as
the number of elements in E ⊂ N if E is finite, and λ(E) = +∞ if E is infinite. Then
(N, ℘ (N), λ) is a σ-finite, complete measure space, λ is called counting measure and
every singleton {n}, n ∈ N, is an atom.
12 2 Notation and Preliminaries
Definition 2.16 (Real Measures) In a measurable space (X, L), a real measure [28,
Sects. 1.8 and 6.6] is a real-valued function μ (not necessarily one-signed) on L with
μ(∅) = 0 and
μ Ei = μ(E i ) when E i ∈ L and E i E j = ∅, i = j ∈ N.
i∈N i∈N
Since the left side is independent of the ordering of i ∈ N, the sum of the series on
the right is the same if i is replaced by σ(i), where σ is a permutation of N. Hence,
the series is absolutely convergent.
Then |μ(E)| ≤ |μ|(E), E ∈ L, and |μ| is a measure in the sense of Definition 2.6
[28, Thm. 6.2] with |μ|(X ) < ∞ [28, Thm. 6.4]. Hence μ± are real measures where
2.4 Measures and Real Measures 13
whence
|μ| is called the total variation of μ, μ± are the positive and negative parts of μ, and
μ = μ+ − μ− is the Jordan decomposition of μ.
Theorem 2.17 (Hahn Decomposition [28, Thm 6.14]) Let μ be a real
measure on a
σ-algebra L. Then there exist A± ∈ L such that A+ A− = X , A+ A− = ∅ and
for E ∈ L,
μ+ (E) = μ(A+ E), μ− (E) = −μ(A− E)
This notation will be generalised in Definition 4.6 and Remark 4.7 to accommodate
finitely additive measures that are not σ-additive. In Chap. 4, real measures on L are
seen as examples of finitely additive measures, in which context they are denoted by
Σ(L).
Remark 2.19 Note that in [15, Chap. 7] the term “signed measure” allows μ to take
one, but not both, of the values ±∞. Thus in that terminology a real measure would
be a signed measure but a signed measure might not be a real measure.
Example 2.20 Any finite linear combination of finite measures is a real measure. A
class of real measures that are absolutely continuous with respect to λ will be defined
in Remark 2.27.
Definition 2.21 (Regular Real Borel Measures) Let B denote the Borel σ-algebra
of a locally compact Hausdorff space (X, ). Then a real Borel measure μ is said to
be regular if both the measures μ± are regular in the sense of Definition 2.12: for all
B∈B
μ± (B) = inf μ± (U ) : B ⊂ U ∈ = sup μ± (K ) : K ⊂ B, K compact ,
This section concerns measurable sets which in an arbitrary measure space (X, L, λ)
split families of measurable sets in the following sense.
Definition 2.22 (Splitting Measurable Sets) A ∈ L splits G ⊂ L \ N if λ(G A) >
0 and λ(G \ A) > 0 for all G ∈ G.
Remark 2.23 If A splits G, then X \ A also splits G. Moreover, if G1 , G2 ⊂ L \ N
are such that for each G 1 ∈ G1 there exists G 2 ∈ G2 with G 2 ⊂ G 1 , then A splits G1
if A splits G2 .
The main result, Theorem 2.25, depends on Baire’s category theorem. To set the
scene define an equivalence relation ∼ on L by writing
E ∼ F if an only if EΔF is
null, i.e. the symmetric difference EΔF = (E \ F) (F \ E) ∈ N . For E, F ∈ L
let
−1
−1
∂(E, F) = tan λ(EΔF) = tan |χ E − χ F | dλ , (2.5)
X
where χ A is the characteristic function of A ∈ L and tan−1 maps R onto [−π/2, π/2].
Since L 1 (X, L, λ) is a Banach space and since a convergent sequence in L 1 (X, L, λ)
has a subsequence that converges pointwise λ-almost everywhere,
it is immediate
that the space (L, ∂) is a complete metric space. Let L̂ = G ∈ L : λ(G) ∈ (0, ∞)
and for G ∈ L̂ let
F(G) = E ∈ L : λ(E G) ∈ {0, λ(G)} .
The following simple adaption of Kirk’s argument [22] yields Theorem 2.25, Lemma
7.9 and Example 9.11.
Lemma 2.24 For any G ∈ L̂ the set F(G) is closed with empty interior in (L, ∂).
In other words F(G) is nowhere dense in (L, ∂).
Proof For G ∈ L̂ let g : L → R be given by g(E) = λ(G E). Then g is continuous
because for E 1 , E 2 ∈ L,
|g(E 1 ) − g(E 2 )| = χG E1 − χG E2 dλ
X
≤ |χ E1 − χ E2 | dλ = tan ∂(E 1 , E 2 ) .
X
So F(G) = g −1 {0, λ(G)} is closed in (L, ∂). To see that F(G) is nowhere dense let
E ∈ F(G) and > 0 be arbitrary. By Lemma 2.10 there exists E ⊂ G with E ∈ L
) < min{tan , λ(G)}.
and 0 < λ(E
If λ(E G) = 0 let
E + = E E , so that
0 < λ(E + G) = λ(E ) < λ(G)
and ∂(E, E + ) = tan−1 λ(E \ E) ≤ tan−1 λ(E ) < .
2.5 Splitting Families of Measurable Sets 15
In the other case, when λ(E G) = λ(G), let E − = E \ E , so that
λ(E − G) = λ((E G) \ E )) = λ(G \ E ) = λ(G) − λ(E ) ∈ (0, λ(G))
and ∂(E, E − ) = tan−1 λ(E \ E − ) = tan−1 λ(E E ) < .
In both cases E ± ∈ L with ∂(E ± , E) < but E ± ∈
/ F(G). Hence, F(G) is closed
with empty interior when G ∈ L̂, as required.
Proof Let G = {G j : j ∈ N}. Since there are no atoms, each G j has a subset G̃ j
with λ(G̃ j ) ∈ (0, ∞). So, by Remark 2.23, it suffices to assume that G j ∈ L̂ for
all j. Then, by Baire’s category theorem, L = j∈N F(G j ) ⊂ L since, by Lemma
2.24,each F(G j ) is nowhere dense in the complete metric space (L, ∂). If A1 ∈
L \ j∈N F(G j ) then, from the definition of F(G j ), A1 splits G.
Now let X 1 = X \ A1 , L1 = {E \ A1 : E ∈ L) and G1 = {G \ A1 : G ∈ G). Then
G1 ⊂ L1 \ N satisfy the hypotheses of the theorem in the measure space (X 1 , L1 , λ).
So there exists A2 ⊂ X \ A1 , A2 ∈ L1 , which splits G1 . By construction A2 A1 = ∅
and A2 splits G by Remark 2.23. Iteration of this argument yields {Ai : i ∈ N} ⊂ L
as in the statement of the theorem.
2.6 Integration
where ak ≥ 0 and γ(A) ≥ 0 for all A ∈ L. Since the functions ϕn in Sect. 2.3 (v) are
non-negative simple functions, the integral of a non-negative measurable function u
with respect to non-negative γ may be defined by
u dγ = lim ϕn dγ. (2.6)
X n→∞ X
Because γ is σ-additive it is easily seen that the limit on the right is independent of
the sequence {ϕn } which satisfies (v)(a)–(c).
16 2 Notation and Preliminaries
Remark 2.26 Note that for u ∈ L ∞ (X, L, λ) this definition coincides with the
definition in Sect. 6.1 when γ is σ-additive, but not when γ is only finitely addi-
tive. This is because
χ(0,1−1/k) (x) 1 for all x ∈ (0, 1),
but (†) in the Preface (established in Remark 6.1) contradicts (2.6). Indeed (†) shows
that the limit in (2.6) may not be independent of the sequence {ϕn } when γ is not
σ-additive.
in which case
u dγ := u + dγ + − u − dγ + − u + dγ − + u − dγ − .
X X X X X
It follows that u is integrable with respect to γ if and only if |u| is integrable with
respect to |γ| and
u dγ ≤ |u| d|γ| < ∞.
X X
Moreover
|u| dλ = 0 if and only if u is zero λ-almost everywhere. (2.7)
X
In fact part (b) of the next result indicates that all real measures μ that are absolutely
continuous with respect to λ arise in this way.
Theorem 2.28 (Lebesgue–Radon–Nikodym [15, Thm. 3.8], [28, Sect. 6.10]) For a
σ-finite measure λ and a real measure μ on L there is
(a) a unique pair μa , μs of real measures with μ = μa + μs , μa λ and μs ⊥ λ
(Definition 2.18), and if μ is positive then so are μa and μs , and
2.6 Integration 17
For spaces of pth-power integrable functions, the notation is that of [6, 12, 35],
namely, L p rather than L p , because for dual spaces L ∗p is easier to manage than
L p ∗ . With the exception of Sect. 9.4 only spaces of real-valued functions will be
considered.
L p -Spaces. When (X, L, λ) is a measure space an equivalence relation ∼ is defined
on measurable functions by u ∼ v if and only if u = v λ-almost everywhere. For
convenience with notation, it is usual to denote the equivalence class [u] (Definition
2.5) simply as u and to refer to it as a function rather than an equivalence class of
functions.
Let L p (X, L,λ), p ∈ [1, ∞), denote the (equivalence classes of) measurable
functions u with |u| p dλ < ∞. Then it is well known that with norm
1/ p
u p := |u| p dλ , p ∈ [1, ∞),
Example 2.29 Let X = (0, 1) with the usual metric and Lebesgue measure and con-
sider the intervals
I1 = 0, 1 ;
1
1
I2 = 0, ; I3 = ,1 ;
2 2
1
1 2
2
I4 = 0, ; I5 = , ; I6 = ,1 ;
3 3 3 3
1
1 1
1 3
3
I7 = 0, ; I8 = , ; I9 = , ; I10 = ,1 ;
4 4 2 2 4 4
and so on. Note that λ(Ik ) ≤ 1/(n + 1) if k > n(n + 1)/2 (the sum of the first n
natural numbers). Let u k = χ Ik . Then u k → 0 in L p (0, 1) for all p ∈ [1, ∞), from
which it follows that u k → 0 in measure, and hence there exists a subsequence with
u k j (x) → 0 for λ-almost all x ∈ (0, 1). However, it is obvious from the construction
that
18 2 Notation and Preliminaries
Definition 2.30 (Local Compactness [28, Defn. 2.3]) A topological space (X, ) is
locally compact if for every x ∈ X there is an open set G with x ∈ G and the closure
G of G is compact.
C0 (X, ) is a real Banach space. When X is compact C0 (X, ) consists of all the real-
valued continuous functions on X . When (X, ) is not compact C0 (X, ) consists
of the restrictions to X of real-valued functions that are continuous on the compact
space (X ∞ , ∞ ) and zero at x∞ .
2.7 Function Spaces 19
Remark 2.34 In a Hausdorff topological space (X, ) with Borel σ-algebra B every
singleton {a} is closed, and hence χ{a} , a ∈ X , is Borel measurable.
Now suppose
that all Borel measurable functions are continuous. Then {a} = χ−1 {a} (0, 2) is open,
because χ{a} is continuous. Therefore, all subsets of X are open, and hence = ℘ (X )
( is the discrete topology) when all measurable functions are continuous. Conversely,
if is the discrete topology every function is continuous. An important case is given
by Example 2.14.
Theorem 2.35 ([28, Cor. 2.24]) Suppose (X, ) is a locally compact Hausdorff
space, λ(X ) < ∞ and λ is a complete regular real Borel measure. Then for u ∈
L ∞ (X, B, λ) there exists a sequence {gk } of continuous functions on (X, ) such that
each gk has compact support (the closure of {x : gk (x) = 0} is compact), gk ∞ ≤
u∞ and u(x) = limk→∞ gk (x) for λ-almost all x ∈ X .
is the following.
Theorem 2.36 ([15, Thm. 6.15]) For 1 ≤ p < ∞ let f ∈ L p (X, L, λ)∗ . Then for
p > 1 there exists a unique g ∈ L q (X, L, λ) such that for all u ∈ L p (X, L, λ)
p
f (u) = ug dλ where q = and f L p (X,L,λ)∗ = gq .
X p−1
20 2 Notation and Preliminaries
Theorem 2.37 (Riesz [28, Thm. 6.19]) For f ∈ C0 (X, )∗ there exists a unique,
regular, real Borel measure μ on X (Definition 2.21) with
f (v) = v dμ for all v ∈ C0 (X, ). (2.12)
X
Moreover, |μ|(X ) = f C0 (X,)∗ < ∞ and f (v) ≥ 0 when 0 ≤ v ∈ C0 (X, ) implies
that μ ≥ 0 on Borel sets.
The space of all regular, real Borel measures on B is denoted by C0∗ (X, τ ) and
for every μ ∈ C0∗ (X, ), (2.12) defines f ∈ C0 (X, )∗ .
Remark 2.38 Although the statements are different because the nomenclature is
different, the Riesz representation theorems in [15, Thm. 7.17] and [28, Thm. 6.19]
are equivalent to Theorem 2.37.
Theorem 2.39 (Hahn–Banach [12, p. 62, II.3.10]) On a real linear space V let
p : V → R be sublinear:
Corollary 2.40 Let Y be a closed linear subspace of a normed linear space V and
let v0 ∈ V \ Y . Then there exists f ∈ V ∗ such that f (v0 ) = 1 and f (y) = 0 for all
y ∈ Y.
Definition 2.41 (Banach Limits [5, p. 34, Chap. II]) As in Example 2.14 and Remark
2.34, let ∞ (N) denote the space of bounded, real sequences indexed by N. Now
let c(N) be the subspace of convergent sequences and let l(u) denote the limit of
{u(k)} ∈ c(N). Since l : c(N) → R is positive, bounded and linear, there exists [11,
Thm. III.7.1] a positive bounded linear functional L on ∞ (N) such that for u ∈
∞ (N),
and L∞ (N) = lc(N) = 1. L : ∞ (N) → R is called a Banach limit; see [31] for
development of the theory. For an account based on finitely additive measures, see
Remark 6.5.
Let C(Z) with the maximum norm denote the Banach space of real-valued continuous
functions on a compact Hausdorff space (Z, ).
Theorem 2.42 (Stone–Weierstrass [11, Thm. V.8.1], [15, Thm. 4.45]) Suppose A ⊂
C(Z) is a linear space in which (i) f, g ∈ A implies the product f g ∈ A; (ii) the
constant function 1 ∈ A; and (iii) for z 1 = z 1 ∈ Z there exists f ∈ A such that
f (z 1 ) = f (z 2 ). Then A = C(Z), where A is the closure of A in C(Z).
Theorem 2.45 (Mazur [23, Chap. 10, Thm. 6]) If {yk } ⊂ K , where K is closed and
convex in a normed linear space, and yk y0 , then y0 ∈ K .
22 2 Notation and Preliminaries
mi
mi
yi = γ ij yk j , γ ij ∈ [0, 1] and γ ij = 1, for some m i ∈ N.
j=1 j=1
Proof For the sequence {k j } ⊂ N let K be the closure in V of the set of convex
combinations of elements of {yk j : j ∈ N}. Since yk j y0 and yk j ∈ K , which is
closed and convex, y0 ∈ K follows from Theorem 2.45.
Definition 2.47 (Weak* Topology on V ∗ [11, Chap. IV, Example 1.8]) For a normed
linear space V a set U ⊂ V ∗ is open in the weak* topology if and only if for every
u ∗ ∈ U there exists v1 , · · · , vn ∈ V and > 0 such that
n
v ∗ ∈ V ∗ : |(v ∗ − u ∗ )(v j )| < ⊂ U.
j=1
The collection
V ,v,u ∗ = v ∗ ∈ V ∗ : |(v ∗ − u ∗ )(v)| < , > 0, v ∈ V, u ∗ ∈ V ∗
Theorem 2.48 (Alaoglu [11, Thm. V.3.1]) For a normed linear space V , B ∗ =
{v ∗ ∈ V ∗ : v ∗ V ∗ ≤ 1}, the closed unit ball in V ∗ , is weak* compact.
Note that when V is a Banach space, weak* convergent sequences in V ∗ are bounded,
but not necessarily if V is not a Banach space.
Remark 2.51 If V is separable (has a denumerable dense subset) the closed unit
ball in V ∗ with the weak* topology is metrizable [11, Thm. V.5.1], and hence [23,
Chap. 10, Thm. 12] weak* sequentially compact.
For example, the four vertices of a closed square in the plane are its extreme points,
and there are no others. Every point of the unit circle is an extreme point of the closed
unit disc in the plane.
Theorem 2.53 (Krein–Milman [11, Thm. V.7.4]) If K is non-empty, compact and
convex in a locally convex space, K is the closed convex hull of its extreme points.
In other words, K is the closure of the intersection of all convex sets which contain
the extreme points of K . (Equivalently, K is the intersection of all closed convex sets
which contain the extreme points of K .)
The next result is a consequence of Theorem 2.48 (Alaoglu) and Theorem 2.53
(Krein–Milman), since V ∗ with the weak* topology is a locally convex topological
space when V is a normed linear space.
Corollary 2.54 When V is a normed linear space the closed unit ball B ∗ in V ∗ is
the weak* closed convex hull of its extreme points (i.e. B ∗ is the intersection of all
weak* closed convex set which contains the extreme points of B ∗ ). In particular, the
set of extreme points of B ∗ is non-empty.
The final theorem of this section says that when V is a Banach space the extreme
points of B ∗ satisfy property (W) in the Introduction.
Theorem 2.55 (Rainwater1 [26]) In a Banach space, xk x if and only if {xk } is
bounded and x ∗ (xk ) → x ∗ (x) for all extreme points x ∗ in the closed unit ball of the
dual space.
Definition 2.56 (Base and Sub-base for a Topology) In a topological space (X, )
a collection G of open sets such that every non-empty open set is a union of sets in
G is called a base for the topology . A sub-base for is a collection G of open sets
with the property that the finite intersections of elements of G form a base for .
Remark
2.57 Note that for any collection T ⊂ ℘ (X ) with the property that X =
G∈T G, there is a unique topology τ for which T is a sub-base.
Definition 2.58 (Separable) A topological space is separable if it has a denumerable
dense subset.
Definition 2.59 (First Axiom of Countability) A topological space (X, ) satisfies
the first axiom of countability if for every x ∈ X there is a denumerable collection
Gx of open sets G with x ∈ G such that x ∈ U ∈ implies that x ∈ G ⊂ U for some
G ∈ Gx . The collection Gx is called a local base at x for the topology .
1 Many internationally renowned mathematicians have published under the name John Rainwater,
see https://en.wikipedia.org/wiki/John_Rainwater.
24 2 Notation and Preliminaries
A completely separable space satisfies the first axiom of countability and is separable.
In metric spaces separability and completely separability are equivalent.
If, for any pair of disjoint closed sets Fi , i = 1, 2, in a topological space there
exists a pair of disjoint open sets G i with Fi ⊂ G i , the space is normal; if this
holds for any closed F1 when F2 is a singleton the space is regular, and if it holds
when both Fi are singletons the space is Hausdorff. Thus normal implies regular
implies Hausdorff and a compact Hausdorff space is normal. Metric spaces are normal
because, with {i, j} = {1, 2}, Fi ⊂ G i = {x : dist(x, Fi ) < dist(x, F j )}. A famous
characterisation of normality is the following.
Lemma 2.61 (Urysohn [21, Chap. 4, Lemma 4]) A topological space is normal if
and only if for every pair of disjoint closed sets F1 and F2 there is a continuous
function from the space into [0, 1] with f (F1 ) = 0 and f (F2 ) = 1.
This is not to say that f −1 {0} = F1 and f −1 {1} = F2 . To obtain that level of precision,
see Lemma 2.62(b).
A set is a G δ -set if it is the intersection of a denumerable family of open sets.
Closed G δ -sets feature in Corollary 7.11, Theorems 9.4, 10.2, 10.4, and the following
lemma.
Lemma 2.62 (a) For A = ∅ in a normal topological space X there exists a contin-
uous function f : X → [0, 1/2] with f (x) = 0 for x ∈ A and f (x) > 0 otherwise,
if and only if A is a closed G δ -set.
(b) For every disjoint pair, F1 and F2 , of closed G δ -sets in a normal space X there
is a continuous function f : X → [0, 1] such that F1 = f −1 {0} and F2 = f −1 {1}.
Proof (a) For the “only if” part suppose such an f exists. Then A = f −1 {0} is closed.
Also
A = f −1 {0} = f −1 [0, 1/k),
k∈N
and so A is a G δ -set. Conversely, suppose A = ∅ is closed and A = k∈N Uk where
Uk is open. Since X \ Uk and A are disjoint closed sets, by Lemma 2.61 (Urysohn),
there exists a continuous function f k : X → [0, 1] suchthat f k (A) = {0} and f k (X \
Uk ) = {1}. Now define f : X → [0, 1/2] by f (x) = k∈N 2−(k+1) f k (x). Since the
series is uniformly convergent f is continuous, f (x) = 0 on A and f (x) ∈ (0, 1/2]
elsewhere.
(b) By Urysohn’s lemma there exists a continuous h 0 : X → [0, 1] with
h 0 (F1 ) = 0 and h 0 (F2 ) = 1. Since X is a G δ -space, by part (a) there exists con-
tinuous h 1 : [0, 1] → [0, 1/2] with h −1 1 {0} = F1 and h 2 : [0, 1] → [1/2, 1] with
h −1
2 {1} = F2 . Now f : X → [0, 1] defined by
2.10 Point Set Topology 25
f (x) = inf h 2 (x), sup h 0 (x), h 1 (x)
X X
Since there is no loss in assuming that G x ∈ G, the denumerable base for the topology,
the normality of X follows from Theorem 2.63.
To see that X is a G δ -space, let F be closed. Then for every x ∈ / F there exists
open
sets G x and G x with F ⊂ G x , x ∈ G x , G x G x = ∅ and G x ∈ G. Since F =
/ G x and at most countably many G x are needed, F is a G δ -set and so X is a
x ∈F
normal G δ -space.
26 2 Notation and Preliminaries
Theorem 3.1 For every bounded linear functional on L ∞ (X, L, λ) there exists a
finitely additive measure (Definition 4.1) ν on L such that
Once the integral with respect to finitely additive measures in Theorem 3.1 has been
defined, the proof in Sect. 6.2 of the theorem is straightforward (indeed, an interested
reader might like to have a quick look now). However, there are differences from
standard theory, see Remark 2.26, and the development needs to be handled carefully.
In the next few chapters, notation and terminology for finitely additive measures,
their existence and basic properties on σ-algebras will be established. However, the
following implications of Theorem 3.1 are more or less immediate.
Corollary 3.2 If {Ak } ⊂ L is a sequence in L with λ Ai A j = 0, i = j, then
χ Ak 0 as k → ∞ in L ∞ (X, L, λ), where χ Ak denotes the characteristic function
of Ak .
Proof By Theorem 3.1, for any f ∈ L ∞ (X, L, λ)∗ there is a finitely additive measure
ν with
f (χ Ak ) = χ Ak dν = ν(Ak ) = ν + (Ak ) − ν − (Ak ),
X
where ν ± (the positive and negative parts of ν, see (4.1c)) are non-negative finitely
additive measures. Since, by finite additivity and (6.2d),
K
K
0 ν ± (Ak ) = ν ± Ak ν ± (X ) < ∞, for all K ∈ N,
k=1 k=1
When combined with the Hahn–Banach theorem, Theorem 3.1 yields a wide variety
of finitely additive measures.
Corollary 3.3 If (X, ) is a Hausdorff topological space and is not the discrete
topology,
Proof From the hypothesis and Remark 2.34, elements of L ∞ (X, B, λ) which are
continuous on X form a proper closed linear subspace of L ∞ (X, B, λ). The result
is then immediate from Theorem 3.1 and Corollary 2.40 of the Hahn–Banach
theorem.
The existence of finitely additive measures with particular properties will be consid-
ered again in Chap. 5.
Proof For ν ∈ L ∗∞ (X, B, λ), define f ∈ L ∞ (X, B, λ)∗ by (3.1b). The restriction
f˜ of f to C0 (X, ) has f˜C0 (X,)∗ f L ∞ (X,B,λ)∗ = |ν|(X ) < ∞, and the result
follows from Theorem 2.37 (Riesz).
3 L ∞ and Its Dual 29
It follows from Corollary 3.3 that the inequality in (3.2) may be strict and, from the
Valadier–Hensgen example in Sect. 6.6, that for a given regular real Borel measure
ν̃ there may be uncountably many finitely additive measures ν which satisfy (3.2).
Chapter 4
Finitely Additive Measures
Finitely additive measures are naturally defined on algebras (collections of sets which
are closed under complementation and finite unions), but here they are considered
on σ-algebras (closed under complementation and countable unions) because L in
Theorem 3.1 is a σ-algebra. Although the terminology “finitely additive measures”
may suggest only a slight modification of familiar measure theory, intuition based
on experience with σ-additivity can seriously mislead, see (†) in the Preface which
is justified in Remark 6.1.
Definition 4.1 (Finitely Additive Measures [35, Sects. 1.2–1.7]) A finitely additive
measure ν on L is a mapping from L into R with
ν(A B) = ν(A) + ν(B) for all A, B ∈ L with A B = ∅.
In the literature, it is usual for ba(L) (bounded and additive on L) to denote the
set of finitely additive measures on L. Let Σ(L) ⊂ ba(L) denote the σ-additive
measures.
Obviously, the set Σ(L) coincides with the set of real measures in Sect. 2.4. However
λ ∈ ba(L) if and only if λ(X ) < ∞. Hence λ ∈ / Σ(L) when λ(X ) = ∞ although it
is σ-additive on L. However, because by hypothesis λ is σ-finite on L, there exists
[28, Lemma
6.9] an integrable function f which is positive everywhere on X . Hence
γ(E) := E f dλ defines γ ∈ Σ(L) which is positive on L \ N and can be used as
a surrogate for λ in arguments involving finitely additive measures, for example, in
the proof of Theorem 10.4.
There follows a review of standard theory sufficient for our purposes. For a com-
prehensive account see [35], [12, Chap. III] or [6, Chap. 4].
Lemma 4.2 Suppose ν ∈ ba(L).
(a) Then ν ∈ Σ(L) if and only if ν(Ak ) → 0 as k → ∞ for any nested sequence
Ak+1 ⊂ Ak , Ak ∈ L with k Ak = ∅.
(b) If 0 ν γ for some γ ∈ Σ(L), then ν ∈ Σ(L).
Proof (a) Suppose ν ∈ Σ(L) and, for {Ak } ⊂ L nested with k Ak =∅, let E k =
Ak \ Ak+1 ∈ L. Then {E k } is a sequence of mutually disjoint sets with k E k = A1 .
Since ν(A1 ) is finite and ν is σ-additive
∞
∞
ν(A1 ) = ν(E k ) = ν(Ak \ Ak+1 ) = lim ν(A1 ) − ν(Ak ) ,
k→∞
k=1 k=1
k−1
∞
∞
ν(E j ) = ν E j − ν(Ak ) → ν E j as k → ∞.
j=1 j=1 j=1
Hence ν ∈ Σ(L).
Part (b) is immediate because its hypothesis implies that ν satisfies the criterion
for σ-additivity in (a).
The next result, which is a simple case of Nikodym’s convergence theorem [8, Thm.
4.6.3(i)], depends on the following trivial observation. Suppose
a jk ∈ R, j, k ∈ N, 0 a jk a j k M when j j , k k .
Then
∞
∞
j
γ E = lim γk E = lim lim γk (E )
k→∞ k→∞ j→∞
1 1 =1
j
j
= sup sup γk (E ) = sup sup γk (E )
k j =1 j k =1
j
j ∞
= lim lim γk (E ) = lim γ(E ) = γ(E ),
j→∞ k→∞ j→∞
=1 =1 =1
ν1 , ν2 ∈ ba(L) ⇒ ν1 ∧ ν2 , ν1 ∨ ν2 ∈ ba(L),
ν1 , ν2 ∈ Σ(L) ⇒ ν1 ∧ ν2 , ν1 ∨ ν2 ∈ Σ(L).
34 4 Finitely Additive Measures
ν ± 0, ν = ν + − ν − and ν + ∧ ν − = 0. (4.1c)
(a) Let E 1 , E 2 ∈ L where E 1 E 2 = ∅ and for F ⊂ E 1 E 2 let Fi =
Proof
E i F, i = 1, 2. Then by finite additivity,
(ν1 ∨ ν2 )(E 1 E2 ) = sup
ν1 (F) + ν2 (E 1 E2 ) \ F
{F∈L:F⊂E 1 E2 }
= sup
ν1 (F1 ) + ν1 (F2 ) + ν2 (E 1 \ F1 ) + ν2 (E 2 \ F2 )
{F∈L:F⊂E 1 E2 }
= sup ν1 (F1 ) + ν2 (E 1 \ F1 )
{F1 ∈L:F1 ⊂E 1 }
+ sup ν1 (F2 ) + ν2 (E 2 \ F2 ) = (ν1 ∨ ν2 )(E 1 ) + (ν1 ∨ ν2 )(E 2 ).
{F2 ∈L:F2 ⊂E 2 }
So ν1 ∨ ν2 ∈ ba(L), and hence ν1 ∧ ν2 = − (−ν1 ) ∨ (−ν2 ) ∈ ba(L) by (4.1b).
Now suppose ν1 , ν2 ∈ Σ(L) and {E k } ⊂ L is a sequence of mutually disjoint sets.
Then, for ∈ (0, 1) and k ∈ N, there exists Fk ⊂ E k , Fk ∈ L, such that
∞
(ν1 ∧ ν2 )(E) ν1 (F) + ν2 (E \ F) = ν1 (Fk ) + ν2 (E k \ Fk )
k=1
∞
∞
(ν1 ∧ ν2 )(E k ) + k = (ν1 ∧ ν2 )(E k ) + ,
k=1 k=1
1−
∞ ∞
(ν1 ∧ ν2 ) E k = (ν1 ∧ ν2 )(E) (ν1 ∧ ν2 )(E k ).
k=1 k=1
4.1 Definition, Notation and Basic Properties 35
For the opposite inequality note that, with F ∈ L and F ⊂ E = k Ek ,
∞
∞
ν1 (F) + ν2 (E \ F) = ν1 (F E k ) + ν2 (E k \ F) (ν1 ∧ ν2 )(E k ),
k=1 k=1
whence
∞ ∞
(ν1 ∧ ν2 ) E k = (ν1 ∧ ν2 )(E) (ν1 ∧ ν2 )(E k ).
k=1 k=1
and ν = ν + − ν − follows by taking the supremum on the left and the infimum on
the right of (4.3), over F ∈ L with F ⊂ E, using (4.2).
Finally suppose ν + ∧ ν − = 0. Then, since ν ± 0 are finitely additive by part (a)
there exists E ∈ L and α > 0 such that for any F ⊂ E, F ∈ L,
Hence for all F ⊂ E, F ∈ L, ν(F) ν + (E) − α. Since α > 0 this contradicts (4.2)
and it follows that ν + ∧ ν − = 0.
(c) Since 0 ν − γ1 γ2 − γ1 and γ2 − γ1 ∈ Σ(L), Lemma 4.2 implies ν −
γ1 ∈ Σ(L), and hence ν ∈ Σ(L) as required.
(d) The result follows from the hypothesis and (4.2) since F ∈ N when F ∈ L
and F ⊂ E ∈ N , because (X, L, λ) is complete.
Definition 4.6 (Absolute Continuity and Singularity [6, Chap. 6]) For ν1 , ν2 ∈
ba(L) write ν1 ν2 (ν1 is absolutely continuous with respect to ν2 ), if for all
> 0 there exists δ such |ν1 (E)| < when |ν2 |(E) < δ, and write ν1 ⊥ ν2 (ν1
and ν2 are mutually singular) if for every > 0 there exists E ∈ L such that
|ν1 |(E) + |ν2 |(X \ E) < .
Remark 4.7 In the special case when ν1 , ν2 ∈ Σ(L) the above definitions are equiv-
alent to simpler statements [6, Thms. 6.1.6 & 6.1.17]:
36 4 Finitely Additive Measures
Definition 4.8 (Purely Finitely Additive Measures) μ ∈ ba(L) is purely finitely addi-
tive if
γ ∈ Σ(L) : 0 γ μ+ = 0 = γ ∈ Σ(L) : 0 γ μ− .
+ −
Clearly, μ is purelyfinitely additive if and only if μ and μ are purely finitely
additive and Π (L) Σ(L) = {0}, where Π (L) denotes the set of purely finitely
additive measures.
Remark 4.9 When μ1 , μ2 ∈ Π (L) and μ1 ν μ2 for some ν ∈ ba(L), it is imme-
diate that ν ∈ Π (L) since ν + μ+ − −
2 , ν μ1 . Similarly, |ν| μ ∈ Π (L) implies
ν ∈ Π (L).
Theorem 4.10 (a) ν ∈ ba(L) is purely finitely additive if and only if
μi (Fi ) + γ(E \ Fi ) , i = 1, 2.
Therefore since μi 0,
μ1 (F1 F2 ) + μ2 (F1 F2 ) + γ(E \ (F1 F2 ))
μ1 (F1 ) + μ2 (F2 ) + γ(E \ F1 ) + γ(E \ F2 ) 2. (4.4)
|μ1 + μ2 | μ+ + − −
1 + μ2 + μ1 + μ2 ∈ Π (L) when μi ∈ Π (L), i = 1, 2,
Lemma 4.11 For 0 γ ∈ Σ(L), 0 μ ∈ Π (L) and > 0 there exists F ∈ L with
γ(F) and μ(F) = μ(X ).
K
K K
0 μ E0 \ Fk e K and γ Fk ek < .
k=1 k=1 k=1
38 4 Finitely Additive Measures
Now since L is a σ-algebra, F = ∞ k=1 Fk ∈ L with γ(F) . Moreover, μ(X ) =
μ(F) since μ(X \ F) K for all K and μ is finitely additive. This completes the
proof.
The sense in which a purely finitely additive measure on a σ-algebra is singular with
respect to any σ-additive measure is captured by the following.
Theorem 4.12 Let 0 ν ∈ ba(L). Then ν ∈ Π (L) if and only if for every non-
negative γ ∈ Σ(L) there exists a sequence {E k } ⊂ L with
4.4 L ∗∞ (X, L, λ)
This chapter closes by specialising briefly to the set L ∗∞ (X, L, λ) of finitely additive
measures that feature in Theorem 3.1 (Yosida–Hewitt).
Proof Suppose ν ∈ L ∗∞ (X, L, λ) and ν ± ∧ γ = 0 for all non-negative γ ∈ Σ(L)
L ∗∞ (X, L, λ). Let γ̂ ∈ Σ(L) be arbitrary with
0 γ̂ ν ± . Since, by Theorem 4.4(d),
± ∗
ν = 0 on N it follows that γ̂ ∈ Σ(L) L ∞ (X, L, λ). Hence, by hypothesis,
ν ± ∧ γ̂ = 0 and, by Definition 4.8, ν ∈ Π (L).
Conversely, if ν ∈ Π (L), by Definition 4.8 ν ± ∧ γ = 0 for all 0 γ ∈ Σ(L)
L ∗∞ (X, L, λ).
of finitely additive measures, each element of which would, by Theorem 3.1, corre-
spond to an element f ∈ L ∞ (X, L, λ)∗ with the property that
f (χ X ) = 1, f (χ E ) ∈ {0, 1}, E ∈ L; f (χ E ) = 0, E ∈ N .
Much of the importance of G in the sequel can be traced back to the next theorem and
subsequent remark which suggest that the action of elements of G on L ∞ (X, L, λ)
is reminiscent of Dirac measures acting on continuous functions.
Theorem
5.1 For u ∈ L ∞ (X, L, λ) and ω ∈ G there is a unique α ∈ I := [u]− ∞,
[u]+∞ (see (2.8)) such that
ω x ∈ X : |u(x) − α| < = 1 for all > 0. (5.2)
This means (5.2) holds for a unique α ∈ R, independent of the function chosen to
represent the equivalence class u ∈ L ∞ (X, L, λ).
K
ω(X ) = ω x : u(x) ∈ (αk − αk , αk + αk )
k=1
K
ω x : u(x) ∈ (αk − αk , αk + αk ) = 0.
k=1
Remark 5.2 If ω ∈ G but ω ∈ / Σ(L) (which often happens, see Theorem 5.9 below)
it does not follow from (5.2) that ω ({x ∈ X : u(x) = α}) = 1. Nevertheless (5.2)
is what is meant by saying that in the sense of finitely additive measures u ∈
L ∞ (X, L, λ) is a constant ω-almost everywhere.
A filter U which is maximal with respect to set inclusion, i.e. for a filter F
(iv) U ⊂ F ⇒ F = U,
5.1 G and Ultrafilters 43
Proof (a) To show that ωU , U ∈ U, is well defined on L it suffices to show that exactly
one of E and E
= X \ E is in U for every E ∈ L. To begin note that at most one of
E and E
is in U, for otherwise E, E
∈ U implies ∅ = E E
∈ U, a contradiction
since ∅ ∈ N .
To see that one of E, E
is in U, suppose E ∈/ U. Then suppose that E F ∈ /N
for all F ∈ U and let
F =U G∈L:E F ⊂ G for some F ∈ U .
implies that E ∈ / N , equivalently N ∈ N implies N ∈ U. In particular, E F̂ ∈ N
implies that E
F̂
∈ U, and hence by (ii) E
F̂ = (E
F̂
) F̂ ∈ U. Finally
E
F̂ ⊂ E
implies that E
∈ U by (iii). This shows that E ∈ / U implies X \ E ∈ U,
and hence U ω is well defined by (5.3a). It remains to check that ω ∈ ba(L).
If A B = ∅, A, B ∈ L, it follows from (i) and (ii) that at most one of ωU (A)
andωU (B) is 1, and if one of them is 1, then ωU (A B) = 1 by (iii). If both are zero,
A
B
∈ U by (ii), and hence ωU (A B) = 0. Hence ωU ∈ ba(L) and it follows
that ωU ∈ G.
(b) Now suppose Uω is given by (5.3b) where ω ∈ G. Then Uω satisfies (i) and
(iii), by (5.1) and finite additivity. Also, for E 1 , E 2 ∈ Uω ,
ω(E 1 \ E 2 ) + ω(E 2 ) = ω(E 1 \ E 2 ) + 1
1 = ω(E 1 E2 ) = .
ω(E 2 \ E 1 ) + ω(E 1 ) = ω(E 2 \ E 1 ) + 1
Hence ω(E 1 \ E 2 ) = 0 = ω(E 2 \ E 1 ) which implies ω(E 1 E 2 ) = 1. So E 1 E 2
∈ Uω and Uω satisfies (ii).
To see that U ω satisfies (iv), suppose it does not and let
F denote thecollection
of filters which contain Uω . If F ⊂ F is totally ordered by inclusion, F ∈F F ∈
F is an upper bound for F. Hence, by Zorn’s lemma, F has a maximal element,
44 5 G: 0–1 Finitely Additive Measures
E ⊂
Definition 5.5 (λ-finite intersection property) L has the λ-finite intersection
K
property if E k ∈ E, 1 k K , implies λ k=1 k > 0.
E
Theorem 5.6 If E has the λ-finite intersection property, there exists ω ∈ G with
ω(E) = 1 for all E ∈ E.
K
A∈L: A⊃ E k , E k ∈ E, K ∈ N ∈ F.
k=1
Now suppose that T ⊂ F is a family of filters which is totally ordered by set inclusion
and let F ∗ = F ∈T F. Then F ∗ ∈ F is an upper bound for T. So by Zorn’s lemma,
F has a maximal element U which is an ultrafilter. By Theorem 5.4 (a), ωU ∈ G
satisfies the theorem.
The next result shows that for given E there may be uncountably many ω ∈ G that
satisfy Theorem 5.6.
In particular,
ω x ∈ X : g(x) n = g dλ → 0 as n → ∞.
{x∈X :g(x)n}
Remark 5.10 This shows that all elements of G are purely finitely additive when
L has no atoms. In any case if X has infinitely many mutually disjoint subsets
with positive λ-measure, by Corollary 5.7(b) there are uncountably many distinct
ω ∈ G which are purely finitely additive. For example, in ∞ (N) G has countably
many σ-additive elements given by the atoms {n}, n ∈ N, and uncountably many
ω ∈ G which are purely finitely additive are given by Corollary 5.7 and Theorem 5.9.
Further statements can be made about G, e.g. Lemmas 7.7, or 9.1 in a locally compact
Hausdorff setting.
Chapter 6
Integration and Finitely Additive
Measures
To begin note that if u ∈ L ∞ (X, L, λ) and w(x) = u(x) for λ-almost all x ∈ X , it
follows that |ν|({x : w(x) = u(x)} = 0 for all ν ∈ L ∗∞ (X, L, λ) since |ν|(N ) = 0
when λ(N ) = 0. Therefore, by finite additivity the following construction is inde-
pendent of whether u or w is used to represent an equivalence class in L ∞ (X, L, λ).
Since, by (4.1c), ν = ν + − ν − and by Theorem 4.4(d) ν ± ∈ L ∗∞ (X, L, λ) when
ν ∈ L ∗∞ (X, L, λ), it suffices first to define integration with respect to non-negative
ν ∈ L ∗∞ (X, L, λ) and to extend by linearity.
For u ∈ L ∞ (X, L, λ) and p > u∞ , let
K
K
S P (u) := yk ν(E k (u)) yk+1 ν(E k (u)) =: S P (u), (6.1b)
k=0 k=0
N
Moreover, when E = n=1 E n , a finite union of sets in L with E i Ej ∈ N
when i = j,
6.1 The Integral 49
N
u dν = u dν. (6.5)
E n=1 En
K
(u + w) dν = K
(u + w) dν = (u + w) dν
X k=0 E k (u) k=0 E k (u)
K
K
= S P (u) +
X w dν
X u dν +
X w dν −
.
= S P (u) + X w dν X u dν + X w dν +
Since > 0 was arbitrary, (6.7) follows when 0 ν ∈ L ∗∞ (X, L, λ). The general
case is immediate from (6.2c).
Remark 6.1 When γ ∈ Σ(L) L ∗∞ (X, L, λ), it is obvious that integration of u ∈
L ∞ (X, L, λ) with respect to γ coincides with Lebesgue integration. However, if ν is
not σ-additive, familiar results such as the Monotone convergence theorem do not
hold for this integral. For example, by Corollary 5.7(b), on the interval (0, 1) there
are uncountably many finitely additive measures ν ∈ G such that ν(0, 1 − 1/k] = 0
and ν(1 − 1/k, 1) = 1 for all k > 1, k ∈ N. It follows from (6.2a) that for all such ν
1 1− k1
1 dν = 1 but u dν = 0 for all k ∈ N and all u ∈ L ∞ (X, L, λ).
0 0
Having prepared the ground in Sects. 4.1 and 6.1, the proof of Theorem 3.1, which
identifies elements of L ∞ (X, L, λ)∗ with elements of L ∗∞ (X, L, λ), and vice versa,
is straightforward.
In Sect. 6.1, it was shown that when ν ∈ L ∗∞ (X, L, λ) a bounded linear functional
f ∈ L ∞ (X, L, λ)∗ is well defined on L ∞ (X, L, λ) by
f (u) = u dν and | f (u)| u∞ |ν|(X ).
X
Hence f ∞ = sup{| f (u)| : u∞ = 1} |ν|(X ). To see that equality holds note
that, since ν + ∧ ν − = 0, by (4.1c) for any > 0 there exists a set A ∈ L with
ν + (A) + ν − (A ) < , where A = X \ A. Let u = χ A − χ A ∈ L ∞ (X, L, λ). Then
u∞ = 1 and
f (u) = u dν = u dν + + u dν + − u dν − − u dν −
X A A A A
= −ν (A) + ν (A ) + ν (A) − ν (A ) ν (A ) + ν − (A) −
+ + − − +
K
K
f (u) = lim f (u K ) = yk f χ Ek (u) = yk ν(E k (u)) → u dν
K →∞ X
k=0 k=0
by (6.2a) and (6.2c). Thus (3.1b) holds and the proof is complete.
Proof (a) With α given by Theorem 5.1, the first part of (6.10a) follows because
u dω − α = (u − α) dω = (u − α) dω <
X X {x∈X :|u(x)−α|<}
for all > 0. The second part follows because, for all > 0
ω x ∈ X : |u(x)| − |α| < ω x ∈ X : |u(x) − α| < = 1.
(b) If
u dω = α and v dω = β,
X X
In a general measure space (X, L, λ), the essential range [28, Chap.3, Ex.19] of
u ∈ L ∞ (X, L, λ) is defined as
R(u) = α ∈ R : λ {x : |u(x) − α| < } > 0 for all > 0 . (6.11a)
Proof From Theorems 5.6 and 6.2, the right side contains the left. Since ω(E) = 1
implies λ(E) > 0, by Theorem 6.2 the left contains the right.
For another viewpoint on the essential range, see (7.3). In a Borel setting the essential
range can be localised, see Sect. 9.3.
Recall from Example 5.8 that bounded real sequences {u j } ∈ ∞ can be identified
with the essentially bounded functions u ∈ ∞ (N) defined by u( j) = u j , j ∈ N, and
there are uncountably many ω̂ ∈ G, each with the property that ω̂(E k ) = 1 for all k,
where E k = k + N.
Theorem 6.4 A sequence {u j } ∈ ∞ has a subsequence that converges
to α ∈ R if
and only if there exists ω ∈ G with ω(E k ) = 1 for all k ∈ N and N u dω = α.
Proof Let ω̂ ∈ G have the property that ω̂(E k ) = 1 for all k and suppose α = N u d ω̂
where u ∈ ∞ (N) is defined by u( j) = u j . Then by (5.4) and Theorem 6.2(a)
ω̂ j ∈ E k : |u j − α| < = 1 for any k ∈ N and any > 0.
6.5 Integrating u ∈ ∞ (N) with Respect to G 53
Hence, for any k ∈ N and > 0, there exists j k such that |u j − α| < . In particu-
lar, there exists j1 ∈ N such that |u j1 − α| < 1 and j2 > j1 such that |u j2 − α| < 1/2.
Proceeding by induction,
for each n ∈ N there exists jn > jn−1 such that |u jn − α| <
1/n. This shows that N u d ω̂ = α, ω̂ ∈ G, implies a subsequence u jn → α as
n → ∞.
Now suppose {u j } has a subsequence u jn → α, let E = { Ẽ k : k ∈ N} where Ẽ k =
{ jn : n k} and let ω̃ ∈ G be a finitely additive measure with ω̃( Ẽ k ) = 1 for all k ∈ N
given by Theorem 5.6. Then for > 0 there exists k ∈ N such that
Ẽ k ⊂ { j : |u j − α| < }, and hence 1 = ω̃( Ẽ k ) ω̃ { j : |u j − α| < } .
In Remark 9.13, the preceding observation is related to the essential range at infinity
of u ∈ ∞ (N).
for all ω ∈ G with ω(E k ) = 1 for all k. Let S : ∞ (N) → ∞ (N) be the shift operator
defined by (Su)( j) = u( j + 1). Then by (6.10b), for ω ∈ G,
u S(u) dω = u dω S(u) dω , u ∈ ∞ (N).
N N N
In other words, u → N u dω, u ∈ ∞ (N), is not shift invariant, and therefore is not
a Banach limit (Definition 2.41 and [31]) for any ω ∈ G.
However, a Banach limit can be constructed as follows. For u ∈ ∞ (N) let T u ∈
∞ (N) be defined by
1
j
(T u)( j) = u(i), i, j ∈ N.
j i=1
T u( j) − T (Su)( j)| = 1 |u( j + 1) − u(1)| → 0 as j → ∞.
j
54 6 Integration and Finitely Additive Measures
It follows that if ω ∈ G and ω(E k )=1 for all k then ω j : |T u( j) − α| < =1
for all > 0 if and only if ω j : |T (Su)( j) − α| < = 1 for all > 0. Hence, by
Theorem 6.2(a),
T u dω = T (Su) dω if ω ∈ G and ω(E k ) = 1 for all k.
N N
Independently, Valadier [33] and Hensgen [19] made the following observation
which contradicts a claim in [35].
For x ∈ [0,
Proof 1] let μx ∈ G be given by Theorem 5.6 with the property that
μ (x − , x + ) [0, 1]) = 1 for all > 0, and note from Theorem 6.2 that
x
1
v dμx = v(x) for all v ∈ C[0, 1].
0
1 1 x nj
2n 1 2n
snv = n v(x nj ) = v dμn , where μn = n μ . (6.13)
2 + 1 j=0 0 2 + 1 j=0
n
For all n ∈ N, μn ∈ L ∗∞ (X, L, λ) and μn (E) ∈ [0, 1], E ∈ L, since each μx j ∈ G.
Moreover, snv is an n th Riemann sum of v ∈ C[0, 1] corresponding to partitioning
[0, 1] with 2n + 1 equally spaced points. Hence s v := {snv } ∈ ∞ is convergent and
1
lim s v = v dλ, v ∈ C[0, 1],
n→∞ n 0
Now define μ on L by
μ(E) = L {μn (E)} , E ∈ L, (6.14)
n 1 1
E k = 0, 1 x j − 2n , x nj + 2n .
{ j,n:
2 k 2 k
0 j2n
n∈N}
Note that since μn (E k ) = 1 for all n ∈ N, it follows that μ(E k ) = 1 = μ(X ) for all k.
Moreover, since
{Ek } is a nested sequence of open subsets
of [0, 1] with λ(E k ) 2/k,
it follows that λ k E k = 0. Hence γ(E k ) → γ k E k =
0 if 0 γ ∈ Σ(L). S-
ince μ 0 this shows that μ ∧ γ = 0 for any 0 γ ∈ Σ(L) L ∗∞ (X, L, λ) whence,
by Theorem 4.15, μ is purely finitely additive. It remains
M to prove (6.12).
When g is a simple function on [0, 1] (i.e. g = m=1 gm χG m , gm ∈ R, G m ∈ L)
it follows from (6.2d) and (6.14) that
1 1
L g dμn = g dμ. (6.15)
0 0
Now the first and third terms on the right converge to 0 as k → ∞ and by (6.15) the
middle term is zero for all k. Hence,
1 1
L u dμn = u dμ for all u ∈ L ∞ ([0, 1], L, λ).
0 0
Remark 6.7 Yosida and Hewitt [35, Thm. 3.4] claimed that
1
v dν = 0 for all v ∈ C[0, 1] (6.16)
0
implies ν is purely finitely additive but omitted the proof because it was consid-
ered straightforward. However, Valadier [33], and later Hensgen [19], independently
showed that this claim is wrong. In Theorem 6.6 μ ∈ Π (L), λ ∈ Σ(L) and so, by the
uniqueness statement in Theorem 4.13, ν := μ − λ ∈ L ∗∞ (X, L, λ) is neither purely
finitely additive nor σ-additive yet satisfies (6.16).
Different choices of the points x nj lead to different μ. So if {y nj : j ∈ N} {x nj :
j ∈ N} = ∅, there are two measures μ1 , μ2 ∈ Π (L) which satisfy (6.12). Hence
0 = μ1 − μ2 = ν ∈ Π (L) also satisfies (6.16). For further discussion of these issues,
see Sect. 10.2.
Chapter 7
Topology on G
K
K
Δ Ak = Δ1K Ak and Δ Ak = Δ1K Ak , Ak ∈ L, (7.1a)
1 1
Proof Let τ be the topology defined in Definition 7.2. To see that (G, τ ) is Hausdorff
suppose that ω1 = ω2 . Then there exists A ∈L such that ω1 (A) = 1 and ω2 (A) = 0.
Hence ω1 ∈ Δ A and ω2 ∈ Δ X \A where Δ A Δ X \A = ∅. Thus (G, τ ) is Hausdorff.
To show
that (G, τ ) is compact suppose that G = A∈A Δ A for some A ⊂ L and
G = A∈A Δ A for any finite subset A
that of A. Since Δ X \A = G \ Δ A , it follows
that A∈A Δ X \A = ∅ but, by (7.1a),
Δ( = ⊂ A is finite.
Δ X \A = ∅ when A
X \A)
A∈A
A∈A
By Lemma 7.1 this implies that A∈A(X \ A) ∈ / N for any finite A ⊂ A.
Since E := {X \ A : A ∈ A} satisfies the hypotheses of Theorem 5.6, there exists
ω ∈ G such thatω(X \ A) = 1, and consequently ω(A) = 0, for all A ∈ A. Since
ω ∈ G and ω ∈ / A∈A Δ A , this is a contradiction. It follows that G ⊂ A∈A Δ A for
⊂ A and hence, since T is a base for τ , (G, τ ) is compact.
some finite A
The next observation is that L ∞ (X, L, λ) and C(G, τ ) are isometrically isomorphic
with the consequence that their duals are isometrically isomorphic, even though one
is represented by finitely additive measures, Theorem 3.1, and the other by σ-additive
measures, Theorem 2.37 (Riesz).
For u ∈ L ∞ (X, L, λ) let L[u] : G → R be defined by
L[u](ω) = u dω for all ω ∈ G. (7.2)
X
Note from Lemma 6.3 that the essential range of u ∈ L ∞ (X, L, λ) is the range of
L[u] on G:
R(u) = L[u](G). (7.3)
and the mapping u → L[u], from L ∞ (X, L, λ) to C(G, τ ) is linear. Moreover, for
u, v ∈ L ∞ (X, L, λ),
L[u 0 ](ω) − α 0
=
u 0 dω − α0
< 0 if ω ∈ Δ A0 .
X
K
K
H= Δ Ak = Δ A where A = Ak ∈ L,
k=1 k=1
Theorem 7.6 (G, τ ) is totally disconnected, meaning all its connected sets are sin-
gletons.
Corollary 7.8 When (X, L, λ) has no atoms every element of G is purely finitely
additive and no element of G is isolated.
Proof Since by definition ω is isolated in (G, τ ) if and only if {ω} = Δ A for some
A ∈ L, by Lemma 7.7 no element of G is isolated if there are no atoms. The pure
finite additivity of ω ∈ G follows by Theorem 5.9.
In a Borel measure space (X, B, λ), a further implication of the absence of atoms is
the following. Recall from Sect. 2.10 that a topological space (X, ) is completely
separable and it has a denumerable basis (a denumerable collection T = {Ti } of
non-empty open sets such that every open set is a union of elements of T ).
This discussion of (G, τ ) is continued in Sect. 9.1, but for that the following
refinement of (7.1b) is needed.
Theorem 7.10 (i) If Γ = k∈N Δ Ak , Ak ∈ L:
(a) Γ ◦ = Δ A0 , where Γ ◦ is the interior of Γ in (G, τ ) and A0 = k∈N Ak .
(b) Both Γ and Γ ◦ are closed.
(c) The boundary ∂Γ = Γ \ Γ ◦ of Γ is characterised by
∂Γ = ω ∈ G : ω(Ak ) = 1 for all k ∈ N but ω(A0 ) = 0 .
7.3 Properties of G and τ 61
whence
1 = ω(E) = ω(E \ A0 ) + ω(E A0 ) = ω E \ Ak + ω(E A0 ) = 0,
k
a contradiction. Since E \ Ak0 ∈ / N , it follows from Theorem 5.6 that there exists
ω0 ∈ G with ω0 (E \ Ak0 ) = 1. This implies that ω0 (Ak0 ) = 0 and hence that ω0 ∈/ Γ.
But ω0 (E \ Ak0 ) = 1 also implies that ω0 ∈ Δ E ⊂ Γ . This contradiction implies that
Γ ◦ \ Δk Ak is empty and hence that Δ A0 = Γ ◦ .
(i)(b, c) Since Γ is an intersection of closed sets, Γ = Γ is closed, Γ ◦ is closed
by Lemma 7.5 and the formula for the boundary ∂Γ = Γ \ Γ ◦ follows.
(i)(d) Now suppose Ak+1 ⊂ Ak , k ∈ N, and Ak \ Ak+1 ∈ / N for infinitely many
k. Then there is no loss since, by Lemma 7.1, Ak \ Ak+1 ∈ N implies Δ Ak = Δ Ak+1 ,
in assuming that F k := Ak \ Ak+1 ∈ / N for all k. Since the Fk are non-null and
disjoint, and since jk F j ⊂ Ak , Corollary 5.7 implies the existence of a family
{ωa : a ∈ (0, 1)} ⊂ G of distinct finitely additive measures with ωa jk j =
F
ωa (Ak ) = 1, k ∈ N. Hence ωa ∈ Γ for all a ∈ (0, 1). However
A0 F j = A0 Ak \ A0 = ∅.
jk
62 7 Topology on G
Therefore, ωa (A0 ) = 0 for all a ∈ R and hence ωa ∈ Γ \ Γ0 = ∂Γ for all a ∈ (0, 1),
as required.
(ii) For the statements about Λ, let Ak = X \ Bk and Γ = G \ Λ in part (i). This
completes the proof.
The next result generalises the first part of Theorem 7.10 since Γ is a closed G δ -set
(a closed intersection of countably many open sets).
Proof Since (G, τ ) is compact and Hausdorff, it is normal (see Sect. 2.10) and since
H is a closed G δ -set, by Lemma 2.62 there exists a non-negative continuous function
f : G → [0, 1/2] with H = f −1 ({0}). By Theorem 7.4, for some non-negative u ∈
L ∞ (X, L, λ),
f (ω) = u dω for all ω ∈ G.
X
Since H = k∈N Δ Ak , by Theorem 7.10(i) H◦ = Δk∈N Ak .
Finally, note that since H is closed its boundary ∂H = H \ H◦ is non-empty if
and only if H is not open. If H is not open it follows that Ak \ Ak+1 ∈
/ N for infinitely
many k, for otherwise, by Lemma 7.1, the collection of sets {Δ Ak } is finite and so
H = k Δ Ak is open. Thus when ∂H is non-empty {Ak } satisfies the hypothesis of
Theorem 7.10(i)(d) and ∂H is uncountable.
7.4 G and the Weak* Topology on L ∗∞ (X, L, λ) 63
Recall from Definition 2.47 that a set U ⊂ L ∞ (X, L, λ)∗ is open in the weak* topol-
ogy if and only if for every f 0 ∈ U there exists u 1 , · · · , u n ∈ L ∞ (X, L, λ) and > 0
such that
n
f ∈ L ∞ (X, L, λ)∗ : |( f − f 0 )(u j )| < ⊂ U.
j=1
where
u 0 dν − u 0 dν0
< 0 , (7.7b)
X X
Lemma 7.12 The topology τ on G coincides with the restriction to G of the weak*
topology on L ∗∞ (X, L, λ).
χ A0 dω0 − χ A0 dν
< ∈ W.
X X 2
Then ω ∈ W0 G implies that |ω 0 (A0 ) − ω(A0 )| < 2 . Hence ω(A0 ) = 1 since
1
W0 = ν ∈ L ∗∞ (X, L, λ) :
u 0 dν − u 0 dω0
< 0 ∈ W.
X X
Let α0 = X u 0 dω0 and put A0 = {x ∈ X : |u 0 (x) − α0 | < 0 /2}. Then ω0 ∈
Δ A0 by Theorem 6.2 and {x ∈ X : |u 0 (x) − α̂| < 0 /2} A0 ∈ N if |α̂ − α0 | 0 .
Therefore ω({x ∈ X : |u 0 (x) − α̂| < 0 /2}) =
0, and consequently
X u 0 dω =
α0 | 0 . It follows that
X u 0 dω − α0
< 0 if ω ∈ Δ A0 .
α̂, if ω ∈ Δ A0 and |α̂ −
Hence ω0 ∈ Δ A0 ⊂ W0 G, as required.
64 7 Topology on G
V = ν ∈ L ∞ (X, L, λ) :
χ A dν −
χ A dν0
< .
X X
Then (see (6.2d)) ν ∈ V implies that |ν(A) − ν0 (A)| < , which means ν(A) ∈ /
/ G. Therefore L ∗∞ (X, L, λ) \ G is weak* open, and G is weak*
{0, 1}, and hence ν ∈
closed.
It follows from Theorem 2.48 (Alaoglu) that (G, τ ) is compact. For a direct proof,
see Theorem 7.3.
By Definition 2.52 f is an extreme point of B ∗ , the closed unit ball in L ∞ (X, L, λ)∗ ,
if and only if for f 1 , f 2 ∈ B ∗ and α ∈ (0, 1)
ν, ν1 , ν2 ∈ U ∗ = {ν ∈ L ∗∞ (X, L, λ) : |ν|(X ) 1}
ν − (X ) = 1 − ν + (X ) = ν − (A) + ν + (A) − ν + (X )
= ν − (A) − ν + (X \ A) ν − (A) + ν + (X \ A) = < 0
Since α ∈ (0, 1), ν1 (A) = ν(A)/|ν|(A) = 0 and ν2 (A) = 0, this shows that f is not
an extreme element of B ∗ if ν is not one-signed.
So suppose ν is one-signed, say 0 ν ∈ U ∗ (for ν 0 replace ν with −ν), but
ν∈/ G. Then there exists A ∈ L with ν(A) ∈ (0, 1). Let
ν(A E) ν((X \ A) E)
ν1 (E) = , ν2 (E) = for all E ∈ L.
ν(A) ν(X \ A)
Remark 7.15 Because of Theorem 2.55 (Rainwater), Lemma 7.14 implies that G has
property (W) in the Introduction. In Sect. 8, this conclusion is derived independently,
from the isometric isomorphism in Theorem 7.4.
Let (X, ) be a Hausdorff space, B its Borel σ-algebra and let D denote the set of
regular Borel measures (Definition 2.12) δ that take only values 0 and 1 on B with
δ(X ) = 1.
Lemma 7.16 For δ ∈ D there exists a unique x0 ∈ X such that δ = δx0 , the Dirac
measure introduced in Sect. 2.4.
In this chapter, theory so far developed yields a necessary and sufficient condition in
terms of behaviour λ-almost everywhere for a bounded sequence that is pointwise
convergent to be weakly convergent in L ∞ (X, L, λ). The resulting test is illustrated
by examples.
Since, by Sect. 7.2, L ∞ (X, L, λ) and C(G, τ ) are isometrically isomorphic, their
duals are isometrically isomorphic and u k u 0 in L ∞ (X, L, λ) if and only if
L[u k ] L[u 0 ] in C(G, τ ). Since (G, τ ) is a compact Hausdorff space, by (V) in the
Introduction L[u k ] L[u 0 ] in C(G, τ ) if and only if {L[u k ]C(G,τ ) } is bounded
and L[u k ] → L[u 0 ] pointwise on G. In other words, u k u 0 in L ∞ (X, L, λ) if and
only if for some M
u k ∞ M and u k dω → u 0 dω as k → ∞ for all ω ∈ G. (8.1)
X X
Thus the collection G of finitely additive measures plays a rôle for weak convergence
in L ∞ (X, L, λ) analogous to that of σ-additive Dirac measures for weak convergence
in spaces of continuous functions. The sequential weak continuity of composition
operators is an immediate consequence, in contrast to what happens for weak conver-
gence in L p (X, L, λ), 1 p < ∞. (When u k (x) = sin(kx) in L p (0, 2π), u k 0
but |u k | 0.)
Theorem 8.1 If u k u 0 in L ∞ (X, L, λ) as k → ∞ and F : R → R is continuous,
then F(u k ) F(u 0 ) in L ∞ (X, L, λ).
and
u k dμ → u dμ for all purely finitely additive μ ∈ G. (8.2b)
X X
Of course many familiar measure spaces have no atoms (Example 2.15), but when
they do (8.2a) is equivalent to
χ A u k dλ → χ A u dλ for all atoms A.
X X
∗
Hence for bounded sequences (8.2a) holds if u k u, and in particular if u k (x) →
u(x) for λ-almost all x ∈ X .
Since (X, L, λ) is σ-finite, L ∞ (X, L, λ) is the dual of L 1 (X, L, λ) [8, Thm.
4.4.1] and hence by Theorem 2.48 (Alaoglu) the closed unit ball in L ∞ (X, L, λ)
is compact in the weak* topology of L 1 (X, L, λ)∗ (Definition 2.47). If in addition
L 1 (X, L, λ) is a separable metric space, by Remark 2.51 the closed unit ball in
L ∞ (X, L, λ) with the weak* topology is metrisable, and hence weak* sequentially
compact (Definition 2.49). Thus a bounded sequence in L ∞ (X, L, λ) has a weak*
convergent subsequence which therefore satisfies (8.2a). (For example (X, L, λ) is
σ-finite and L 1 (X, L, λ) is separable when X ⊂ Rn and λ is any Borel measure on
Rn which is finite on bounded sets [8, Cor. 4.2.2].)
In circumstances such as these, deciding whether a bounded sequence is weakly
convergent or has a weakly convergent subsequence is mainly a question of deciding
whether (8.2b) is satisfied. Theorem 8.7 settles this issue by characterising sequences
that converge weakly to 0 in terms of their λ-almost everywhere pointwise behaviour.
To set the scene, recall some necessary conditions for the weak convergence of
sequences in L ∞ (X, L, λ).
In a metric space, let B(x, r ) denote the ball with centre x and radius r .
Lemma 8.4 Suppose (X, ρ) is a metric space and for u ∈ L ∞ (X, B, λ), there exists
a set E(u) ∈ L such that λ(X \ E(u)) = 0,
1
u(x) := lim u dλ exists for all x ∈ E(u), (8.3)
0<r →0 λ(B(x, r )) B(x,r )
Remark 8.5 Hypothesis (8.3) means that λ-almost all x ∈ X is a Lebesgue point of
u [28, 7.6 & Th. 7.7]. According to Heinonen [18, Thm. 1.8 & Remark 1.13] this is
not very restrictive. For example, it holds when λ is a regular Borel measure on B
which is doubling on (X, ρ) (i.e. balls have finite positive measure and there exists a
constant C such that λ(B(x, 2r )) Cλ(B(x, r )), or on Rn with the standard metric
when λ is a Radon measure, finite on compact sets and positive on balls.
Proof Let E = ∞ 0 E(u k ), which has full measure, and let V denote the linear span
in L ∞ (X, B, λ) of {u k : k ∈ N0 }. Now for fixed x ∈ E define a linear functional x
on V by
x (v) = v(x), v ∈ V.
Since |x (v)| v∞ , by the Hahn–Banach theorem there exists L x ∈ L ∞ (X, B, λ)∗
with L x (v) = x (v) for all v ∈ V . Therefore u k u 0 implies that u k (x) = x (u k ) =
L x (u k ) → L x (u 0 ) = u 0 (x) for all x ∈ E, as required.
Example 8.6 To see that the converse is false let X = (0, 1) and, for k ∈ N, define
u k on (0, 1) by u k (x) = 0, x ∈ [1/2k, 1); u k (x) = 1, x ∈ (0, 1/4k]; u k linear on
[1/4k, 1/2k]. Then {u k ∞ } is bounded, u k is continuous on X and u k (x) → 0
monotonically for all x ∈ X as k → ∞. But by Theorem 5.6 there exists ω ∈ G with
ω(E ) = 1, where E = (0, 1/4) for each . Therefore, by Theorem 6.2, X u k dω =
1 for all k and hence u k 0 in L ∞ (X, L, λ). Hence u k is not weakly convergent to
0 (indeed not weakly convergent to any limit) in L ∞ (X, L, λ) .
mi
mi
u i ∞ → 0 as i → ∞, u i = γ ij |u k j |, γ ij ∈ [0, 1] and γ ij = 1, m i ∈ N.
j=1 j=1
70 8 Weak Convergence in L ∞ (X, L, λ)
λ Aα (u k j ) = 0. (8.5)
j=1
This criterion is equivalent to saying that for all strictly increasing sequences {k j },
the sequence {v J } in (8.4) converges strongly to zero in L ∞ (X, L, λ).
Remark 8.8 In the Borel setting of the next chapter, this criterion for weak conver-
gence in L ∞ (X, B, λ) will be localised to arbitrary open neighbourhoods of points
in the one-point compactification of X .
Proof Suppose, for a strictly increasing sequence {k j } and α > 0, that (8.5) is false
for all J ∈ N. Then E = {Aα (u k j ) : j ∈ N} satisfies the hypothesis of Theorem 5.6
and hence there exists ω ∈ G such that ω(Aα (u k j )) = 1 for all j. It follows that
|u k j | dω |u k j | dω α > 0 for all j.
X Aα (u k j )
J
and so ω(Aα (u k j )) = 1 for all j. Hence ω j=1 Aα (u k j ) = 1, for all J , since ω ∈
G. Since ω is zero on N because ω ∈ G, it follows that (8.5) is false for all J . Finally,
note that for a strictly increasing sequence {k j } and α > 0,
=λ Aα (u k j ) .
j=1
Dini’s theorem [27, Thm. 7.13] says that on compact spaces monotone pointwise
convergence of a sequence of continuous functions to a continuous function is uni-
form; equivalently, by (V) in the Introduction, weak and strong convergence coin-
cide for bounded monotone sequences of continuous functions on compact spaces.
An analogue in L ∞ (X, L, λ) now follows either from Dini’s theorem in the light of
Theorem 7.4, or directly from Theorem 8.7. The sequence in Example 8.6 converges,
pointwise everywhere and monotonically, to zero, but not weakly in L ∞ (X, L, λ).
With ∞ (N) as in Example 5.8 and Sect. 6.5, for every j ∈ N a bounded linear func-
tional δ j is defined on ∞ (N) by δ j (u) = u( j), u ∈ ∞ (N). Hence u k 0 in ∞ (N)
implies that u k ( j) → 0 as k → ∞ for each j ∈ N. However, pointwise convergence
does not imply weak convergence.
72 8 Weak Convergence in L ∞ (X, L, λ)
Proof Suppose u k 0. Then by Theorem 8.7 there exists α > 0 and a strictly
increasing sequence {k j } such that for all J ∈ N
v J ∞ > α where v J (i) = min |u k j (i)| : j ∈ {1, · · · , J } for i ∈ N. (8.8)
Let J1 = 1 and let i 1 denote the smallest i such that |u k1 (i)| > α. Since u k j converges
pointwise to 0 as j → ∞, there exists a smallest J2 > J1 such that |u k j (i 1 )| α for
all j J2 . Therefore by (8.8) there exists a smallest i > i 1 , denoted by i 2 , such that
min |u k j (i 2 )| : j ∈ {1, · · · , J2 } > α.
Again there exists a smallest J3 > J2 such that |u k j (i 2 )| α for all j J3 . Let i 3
denote the smallest i > i 2 such that
min |u k j (i 3 )| : j ∈ {1, · · · , J3 } > α.
By induction, for the strictly increasing sequence {k j } satisfying (8.8), this process
yields strictly increasing sequences {i n } and {Jn } with the required properties.
Conversely, if (8.7) holds it is immediate that v J ∞ → 0 as J → ∞ and hence
u k 0 as k → ∞ by Theorem 8.7. This completes the proof.
Example 8.12 A sequence {u j } in ∞ (N) with
u i j → 0 as j → ∞, i < j
u j (i) = , i ∈ N.
α > 0, i j
Example 8.13 In Example 2.29 u k → 0 strongly in L p (0, 1) for all finite p, and
therefore {u k } has a subsequence (see Sect. 2.27) which converges almost everywhere
to 0, yet lim supk→∞ u k (x) = 1 for all x ∈ (0, 1). To see that u k 0 in L ∞ (0, 1) let
j ( j + 1) 1
kj = 1 + and put Ik j = 0, .
2 j +1
8.3 Applications of Theorem 8.7 73
Then 1
u k j = χ Ik where Ik j = , ,
j 2j 2 j+1
defines a subsequence for which u k j 0 by Remark 8.10, since {Ik j } is a sequence
of disjoint intervals.
Example 8.14 (Step Functions) In this example X = (−1, 1) with Lebesgue mea-
sure and 1 1
u k = χ Ak where Ak = k+1 , k .
2 2
Let
1
1
u+ (x) = u k x + = χ A+
, where A +
= 0, ,
k
2(k+1) k
k
2(k+1)
1
1 3
u−
k (x) = u k x − (k+1) = χ A− , where Ak =
−
k
, (k+1) .
2 k 2 2
So u k = u ± −
k = 1, and u k 0 and u k 0 by Remark 8.10 since {Ak } and {Ak }
−
I
|u k (x)| |αi |χ Aik (x) +
= |αi | +
.
i=1 i∈{1,··· ,I
}
κ(x,i)=k
where lcm denotes the least common multiple. Then, for j ∈ {1, · · · , J }, since p m
k j and p is prime,
1 lcm {k1 , · · · , k J }
= π where
k j xJ pm k j
lcm {k1 , · · · , k J }
= r mod p m , r ∈ {1, · · · , p m − 1}.
kj
j
2k j divides k j+1 and k j+1 2 ki for all j ∈ N.
i=1
+ k j +
j−1
1
J
i=1 ki π
J
i= j+1 ki i=1 ki π
= =
k j xJ k j 2 k j 2
π
1
= 2N + 1 + z where N ∈ N and |z| < .
2 2
Hence π
π
1
|u k j (x J )| ∈ sin , sin = √ , 1 , j ∈ {1, · · · , J }.
4 2 2
This chapter considers what more can be said when (X, ) is a locally compact
Hausdorff space, with B the corresponding Borel σ-algebra and λ a measure on B as
described in Chap. 3. In addition, here λ is assumed regular, finite on compact sets
and positive on open sets. Recall from Lemma 7.16 that a regular Borel measure that
takes only values 0 or 1 is a Dirac measure concentrated at a point x0 ∈ X . Now it
will be shown that G can be partitioned into a union of disjoint closed subsets G(x0 ),
x0 ∈ X ∞ , where X ∞ is the one-point compactification of X . Because of (8.1), this
yields a notion of weak convergence at points of X ∞ and leads to a related idea, the
essential range at x0 of u ∈ L ∞ (X, B, λ) (sometimes called the set of cluster values
of u at x0 ) which is parameterized by elements of G(x0 ). The pointwise criterion for
weak convergence in the previous chapter can then be localised and related to the
pointwise essential range.
The distinction between the essential range at x0 of u ∈ L ∞ (X, B, λ) and its
actual range can be rather striking. When Ω ⊂ Rn with positive Lebesgue measure,
exists u ∈ L ∞ (Ω, B, λ) for which the range is denumerable, {u(x) : x ∈ Ω} =
there
Q [0, 1], while at every x0 ∈ Ω the essential range of u is the continuum [0, 1].
The essential range of u ∈ ∞ (N) at the point at infinity in N∞ is the set of limits of
convergent subsequences {u(k j )} of {u(k)}.
9.1 Localising G
For all ω ∈ G there is at most one x0 ∈ X satisfying (9.1a) and when (X, ) is
compact there is exactly one such x0 .
(b) For x0 ∈ X at least one ω ∈ G satisfies (9.1a). If (X, ) is not compact, at
least one ω ∈ G satisfies,
Proof (a) Suppose that ω(K ) = 1 for some compact K and (9.1a) is false. Then for
each x∈ K there is an open set G x with x ∈ G x and ω(G x ) = 0. Since K is compact,
K
K ⊂ i=1 G xi . Since ω(G xi ) = 0, 1 i K and ω is finitely additive, ω(K ) = 0
which is false. Hence if ω(K ) = 1 and K is compact, there exists x0 ∈ K satisfying
(9.1a). Suppose there is another x1 ∈ X satisfying (9.1a). Since X is Hausdorff, there
exist open sets with x0 ∈ G x0 , x1 ∈ G x1 and G x0 G x1 = ∅. But
this is impossible
because, by finite additivity, it would imply that ω G x0 G x1 = 2.
Now suppose ω ∈ G and ω(K ) = 0 for all compact sets K . By local compactness,
for any x ∈ X there is an open set G x with x ∈ G x and its closure G x is compact.
Hence, since ω(G x ) ω(G x ) = 0, there is no x ∈ X with the required property.
Finally, the existence of x0 when X is compact follows because ω(X ) = 1.
(b) For x0 ∈ X , let E(x0 ) = {G ∈ : x0 ∈ G}. Since by hypothesis λ(G) > 0
when G is non-empty and open, by Theorem 5.6 there exists ω ∈ G with ω(G) = 1
for all G ∈ E(x0 ).
Finally if (X, ) is not compact let E∞ = {X \ K : K compact}. Since sets in E∞
are non-empty and open (because compact sets in Hausdorff spaces are closed), the
existence of ω ∈ G follows from Theorem 5.6 as before. This completes the proof.
Corollary 9.2 (a) For ω ∈ G there is a unique x0 ∈ X ∞ such that ω∞ (G) = 1 when
x0 ∈ G and G is open in X ∞ . Moreover x0 = x∞ if and only if ω(K ) = 0 for all
compact K ⊂ X .
(b) For x0 ∈ X ∞ there exists ω ∈ G such that ω(G) = 1 when x0 ∈ G and G is
open in X ∞ .
Definition 9.3 For x0 ∈ X ∞ , let G(x0 ) ⊂ G denote the set of ω ∈ G for which the
conclusion of Corollary 9.2(b) holds, and let U(x0 ) ⊂ U denote the corresponding
family of ultrafilters (Definition 5.3).
Properties of G(x0 ), x0 ∈ X ∞
(a) G(x0 )
= ∅ by Corollary 9.2(b);
(b) G(x1 ) G(x2 ) = ∅ if x1 = x2 ∈ X ∞ , because (X ∞ , ∞ ) is Hausdorff;
9.1 Localising G 79
(b) Suppose (X, ) is not compact and {x∞ } (X \ K k ) k∈N is a nested, denu-
merable local base for ∞ at x∞ (i.e. K k is compact, K k ⊂ K k+1 and for every
compact K there exists k ∈ N with K ⊂ K k ). Then
(i) G(x∞ ) is a closed G δ -set in (G, τ ),
(ii) G(x∞ )◦ = ∅ and G(x∞ ) is uncountable.
Proof (a) (i) For x0 ∈ X it is immediate that G(x0 ) = k ΔG k and since ΔG k is both
open and closed, G(x0 ) is a closed G δ -set in (G, τ ), and (G, τ ) is compact.
(ii) Since (X, ) is Hausdorff, ∩k∈N G k = {x0 } and by Theorem 7.10 the interior
G(x0 )◦ = Δ{x0 } . Hence G(x0 )◦ = ∅ if and only if λ({x0 }) > 0, i.e. {x0 } is an atom
and by Lemma 7.7 G(x0 )◦ is the singleton Δ{x0 } .
(iii) Since x0 ∈ G k , by hypothesis λ(G k ) > 0 because G k is non-empty and open,
and λ(G k ) → 0, since λ is σ-additive and λ({x0 }) = 0. That G(x0 ) is uncountable
and G(x0 )◦ = ∅ now follows from Theorem 7.10.
(b) Since a singleton
is compact, the hypothesis implies that X = k∈N K k , and
since G(x∞ ) = k Δ(X \Kk ) the results follow from Theorem 7.10 as in part (a).
The hypothesis of part (b) implies that X = k K k ; in other words, (X, ) is
σ-compact (Definition 2.31). However, (X, ) being σ-compact does not imply that
the hypotheses of part (b) are satisfied.
Remark 9.5 Given the one-to-one correspondence between the set G of 0–1 finitely
additive measures and ultrafilters U in (X, B, λ) in Theorem 5.4, there is an obvious
similarity with Lemma 7.16, but there are important differences. By Lemma 7.16
there is a one-to-one correspondence between points x0 ∈ X and Dirac measures
δx0 ∈ D, and no Dirac measure is concentrated at infinity. By contrast, there is a one-
to-one correspondence between points x0 ∈ X ∞ and the disjoint sets G(x0 ) ⊂ G
which by Theorem 9.4(iii) may be uncountably infinite.
80 9 L ∗∞ When X is a Topological Space
J
λ Aα0 (u k j |G ) > 0; (9.3)
j=1
(c) there exists a strictly increasing {k j } such that, for all open G ⊂ X ∞ with x0 ∈ G,
the non-increasing sequence {v J } defined by (8.4) satisfies
v J |G X → 0 in L ∞ (G X, B, λ) as J → ∞.
9.2 Localising Weak Convergence 81
Proof That (b) and (c) are equivalent is immediate from the definitions. If (a) holds
there exists α0 > 0, ω0 ∈ G(x0 ) and a strictly increasing sequence {k j } ⊂ N, such
that for any open G with x0 ∈ G
0 < α0 |u k j | dω0 = |u k j | dω0 for all j.
X G
Now (b) follow as in the proof of Theorem 8.7. If (b) holds then (9.3) implies that
E0 := Aα (u k j |G ) : j ∈ N, G open, x0 ∈ G
satisfies the hypothesis of Theorem 5.6. Hence, there exists ω0 ∈ G with ω0 (A) = 1
for all A ∈ E0 . It follows that ω0 ∈ G(x0 ) and from (9.3),
|u k j | dω0 = |u k j | dω0 α0 > 0 for all j,
G X
Example 9.9 Suppose (X, B, λ) has no atoms and (X, ) is completely separable
(i.e. has a denumerable base, G say, see Sect. 2.10). By Theorem 2.22 there exists
{Ak }, a sequence of mutually disjoints sets in B, each of which splits G.
Hence, for all k ∈ N and for every open set G ⊂ X the function χ Ak takes the value
1 on a subset of G with positive measure. However, since the sets Ak are mutually
disjoint it follows from Example 8.14 that χ Ak 0 in L ∞ (X, B, λ), and therefore
χ Ak 0 at every point of X . Note that {χ Ak } does not satisfy (9.3) for any {k j } ⊂ N
j
and x0 ∈ X .
By analogy with (6.11), in a Borel measure space (X, B, λ) the essential range of u
localised at x0 ∈ X ∞ is defined as
R(u)(x0 ) = α:λ x ∈G X : |α − u(x)| <
> 0 (9.4)
{
>0, G: x0 ∈G∈}
= u dω : ω ∈ G(x0 ) .
X
From the first line of (9.4), the multivalued mapping x0 → R(u)(x0 ) reflects the
measure-theoretic fine structure at x0 of u ∈ L ∞ (X, B, λ) and, since
R(u)(x0 ) = L[u] G(x0 ) , (9.5)
82 9 L ∗∞ When X is a Topological Space
the second line relates that fine structure to the isometric isomorphism L :
L ∞ (X, B, λ) → C(G, τ ) in Sect. 7.2. R(u)(x0 ) is closed because G(x0 ) is compact
and L[u] is continuous.
A point α ∈ R(u)(x0 ) given by α = X u dω, ω ∈ G(x0 ), may be thought of as
the directional limit of u at x0 , the “direction” being determined by the ultrafilter
Uω ∈ U(x0 ) (see (5.3b). Thus, by Theorem 9.7, weak convergence in L ∞ (X, B, λ) is
equivalent to convergence, for each U ∈ U(x0 ), of the directional limits of u k at x0 to
corresponding directional limits of u at x0 , for each x0 ∈ X ∞ . In other words, u k u
implies that for every x0 ∈ X ∞ and every α ∈ R(u)(x0 ) there exist αk ∈ R(u k )(x0 )
such that αk → α as k → ∞. This is not equivalent to αk → α when αk ∈ R(u k )(x0 )
and α ∈ R(u)(x0 ) because of the possibility that
αk = u k dωk and α = u dω, but ωk = ω.
X X
Thus the multivalued R(u)(x0 ) may be interpreted as representing the fine structure at
x0 of u ∈ L ∞ (X, B, λ) which is intimately related to weak convergence. A sufficient
condition for u k u is that for every x0 ∈ X ∞
sup |α| : α ∈ R(u k − u)(x0 ) → 0 as k → ∞.
As already noted, the necessary condition is not sufficient. The following example
shows that the sufficient condition is not necessary.
Example 9.10 With {Ak } in Example 9.9, u k = χ Ak has the property that R(u k )(x) =
{0, 1} for all x ∈ X while u k 0 in L ∞ (X, B, λ) by Corollary 3.2. For a simpler
example, let u k = χ Ak where {Ak } is a sequence of disjoint open segments cen-
in R . Then R(u k )(0) = {0, 1} but u k 0 by
2
tred on the origin 0 of the unit disc
Corollary 3.2. In these examples X u k dω → 0, but not uniformly for ω ∈ G(0).
The next example illustrates the strong distinction between the pointwise value u(x)
and the essential range R(u)(x) at a point x.
Example 9.11 Suppose (X, B, λ) has no atoms and (X, ) is completely separable
with denumerable base G. Let {Ai : i ∈ N} be the disjoint family of measurable
sets,
each of which splits G according to Theorem 2.25. Let q : N → Q (0, 1) be a
bijection and define a measurable function by
∞
q(i) if x ∈ Ai , i ∈ N
u(x) = q(i)χ Ai = .
0 otherwise
i=1
Since R(u)(x0 ) is closed, R(u)(x0 ) = [0, 1], even though u(x0 ) ∈ Q, for all
x0 ∈ X .
Lemma 9.12 Suppose (X, ) is completely separable and u ∈ L ∞ (X, B, λ). Then
u(x) ∈ R(u)(x) for λ-almost all x ∈ X .
K
Proof Consider first the case when u is simple, say u = k=1 αk χ Ak where the αk
are distinct and Ak = {x : u(x) = αk } ∈ B. Then let
Bk = x ∈ X : u(x) = αk ∈
/ R(u)(x) , k ∈ {1, · · · , K },
K
so that x ∈ X : u(x) ∈
/ R(u)(x) = Bk .
k=1
Let y ∈ Bk . Then u(y) = αk and there exists an open set G with y ∈ G and λ {x ∈
G : u(x) = αk } = 0. Since there is no loss in assuming that G ∈ G where G is
the denumerable base for , Bk is covered bya denumerable family of sets of zero
measure. Hence λ(Bk ) = 0 for all k, and so λ {x : u(x) ∈ / R(u)(x)} = 0.
For the general case, let u ∈ L ∞ (X, B, λ) be arbitrary, let {u k } be a sequence
of simple functions with u k − u∞ → 0 in L ∞ (X, B, λ) and for fixed k let
u k (x)∈ R(u k )(x) for all x ∈ E k where λ(X \ E k ) = 0. Then λ(X \ E) = 0 where
E = k∈N E k ,
u k (x) ∈ R(u k )(x) for all k ∈ N and x ∈ E,
and without loss of generality suppose u k (x) → u(x) for all x ∈ E.
Now for x ∈ E and k ∈ N let ωk ∈ G(x) be such that u k (x) = X u k dωk . Then
for x ∈ E,
u(x) = lim u k (x) = lim u k dωk
k→∞ k→∞ X
= lim (u k − u) dωk + u dωk = lim u dωk ∈ R(u)(x),
k→∞ X X k→∞ X
since X u dωk ∈ R(u)(x) and R(u)(x) is closed. The proof is complete.
1
n
yn → y0 in ∞ (N), where yn ( j) = xi+ j and y0 ( j) = s for all j.
n i=1
Theorem 9.14 Let f ∈ H ∞ (D) be an inner function and let eiϑ0 ∈ T be its singu-
larity. Then R( f )(ϑ0 ) = T.
Let α±ε be the two points where the circle {ζ ∈ C : |ζ− α| = ε} intersects T and let
T(α, ε) and D(α, ε) be the intersections of T and D T with the closed half-plane
9.4 A Localised Range from Complex Function Theory 85
Let > 0 be the distance from α to D(α, ε). (It is not difficult to see that is the
distance from α to the midpoint of the chord [α−ε , αε ] and = 1 − cos(2 arcsin 2ε ).)
For the Poisson kernel
1 − r2
Pr (ϑ) := , 0 ≤ r < 1, ϑ ∈ R,
1 − 2r cos ϑ + r 2
It follows from (9.6) that f eiϑ ∈ T(α, ε) for almost all ϑ with |ϑ − ϑ0 | < δ. Since
Pr ≥ 0,
Pr (θ − ϑ) dϑ < Pr (θ − ϑ) dϑ = 1,
|ϑ−ϑ0 |<δ |ϑ−ϑ0 |≤π
T(α, ε) ⊂ D(α, ε) and the latter is a closed convex set containing 0, it follows that
Pr (θ − ϑ) f (eiϑ ) dϑ ∈ D(α, ε). (9.8)
|ϑ−ϑ0 |<δ
for all r eiθ ∈ Bδ0 (eiϑ0 ). Then it follows from the definition of and from (9.8), (9.9)
that for z = r eiθ ∈ Bδ0 (eiϑ0 ),
f (z) = Pr (θ − ϑ) f (eiϑ ) dϑ
|ϑ−ϑ0 |≤π
= Pr (θ − ϑ) f (eiϑ ) dϑ + Pr (θ − ϑ) f (eiϑ ) dϑ,
|ϑ−ϑ0 |<δ δ≤|ϑ−ϑ0 |≤π
cannot belong to B/2 (α). Since this contradicts [17, Chap. II, Sect. 6, Theorem
6.6], every α ∈ T must belong to R( f )(ϑ0 ). So, T ⊆ R( f )(ϑ0 ) ⊆ T, i.e. R( f )
(ϑ0 ) = T.
86 9 L ∗∞ When X is a Topological Space
If f is an infinite Blaschke product, such that the closure of its zero set contains T, or
the singular inner function defined by a singular measure with closed support equal
to T, then the set of its singularities equals T (see [17, Chap. II, Sect. 6, the paragraph
after Theorem 6.6]).
Theorem 3.1 identifies elements of L ∞ (X, L, λ)∗ with finitely additive measures
in L ∗∞ (X, L, λ) and Theorem 2.37 (Riesz) identifies bounded linear functionals on
C(G, τ ) with regular real Borel measures, denoted by C ∗ (G, τ ), on (G, τ ). However,
the isometric isomorphism in Theorem 7.4 means that these representations must be
related to one another and the finite additivity of one should be reconciled with the
σ-additivity of the other. This issue will be addressed first.
Then, in a topological space setting, for a functional f ∈ L ∞ (X, B, λ)∗ let ν 0
be the finitely additive measure in the Yosida–Hewitt representation of f and ν̂ 0
the regular real Borel measure in the classical Riesz representation of f restricted
to the space C0 (X, ) of continuous functions that vanish at infinity. A minimax
formula for ν̂ in terms ν is obtained in Sect. 10.2 when ν is non-negative.
In this section, (X, L, λ) is the general measure space introduced in Chap. 3. For
a bounded linear functional f on L ∞ (X, L, λ), let f˜ ∈ C(G, τ )∗ be defined by
f˜ ◦ L = f , where L is as in Theorem 7.4. Then with ν ∈ L ∗∞ (X, L, λ) representing
f ∈ L ∞ (X, L, λ)∗ let ν̃ ∈ C ∗ (G, τ ) denote the regular real Borel measure on G that
represents f˜ ∈ C(G, τ )∗ . Since, for any A ∈ L and ω ∈ G,
L[χ A ](ω) = χ A dω = ω(A) = χΔ A (ω),
X
∞ ∞
∞
ν̃ Δ Ai = ν̃ Δ Ai = ν Ai
i=1 i=1 i=1
∞
∞
ν Ai = ν̃ Δ
∞ = ν̃ Δ Ai .
i=1 Ai
i=1 i=1
∞
γ̃(H) = lim γ̃(Δ Ak ) = lim γ(Ak ) = γ Ak = 0.
k→∞ k→∞
k=1
Conversely, suppose that γ̃(H) = 0 for every closed, nowhere dense G δ -set H in
(G, τ ). Let {E k } ⊂ L
with E k+1 ⊂ E k and k∈N E k = ∅. Since the nested sets Δ Ek
are open and closed, k Δ Ek is a closed G δ -set which is nowhere dense since, by
Corollary 7.11, its interior is empty. Therefore
∞
lim γ(E k ) = lim γ̃(Δ Ek ) = γ̃ Δ Ek = 0,
k→∞ k→∞
k=1
When X is compact it follows that ν(X ) = ν̂(X ), because in that case non-zero
constant functions belong to C0 (X, ). However, when (X, ) is not compact ν̂(X )
may be zero when ν is non-negative and ν(X ) = 1.
Since ν = ν + − ν − , without loss of generality attention will henceforth be
restricted to non-negative ν ∈ L ∗∞ (X, B, λ) with the goal of understanding how
ν̂ 0 depends on ν 0. Note:
(i) from the Yosida–Hewitt decomposition (6.9), ν = μ + gλ where μ ∈
L ∗∞ (X, B, λ) is purely finitely additive and gλ, g ∈ L 1 (X, B, λ), is σ-additive.
90 10 Reconciling Representations
As Valadier and Hensgen noted (see Sect. 6.6), the relations between μ and ρ, and
g and k are not straightforward: for Lebesgue measure λ on [0, 1], there is a non-
negative ν ∈ Π (B) with
1 1
v dν = v dλ for all v ∈ C[0, 1], (10.3b)
0 0
In particular, ν̂(X ) = ν(X ) implies ν(F) ν̂(F) for all closed F ⊂ X . (ν(X ) =
ν̂(X ) when (X, ) is compact was noted following (10.2).)
Proof For a given Borel set B and K ⊂ B ⊂ G as in the statement, let f be the
continuous function determined in Lemma 10.5 by K and G. Then
ν(K ) f dν ν(G) and ν̂(K ) f d ν̂ ν̂(G).
X X
10.2 Restriction to C0 (X, ) of Elements of L ∗∞ (X, B, λ) 91
From (10.2), it follows that ν(K ) ν̂(G) and ν̂(K ) ν(G) whence, since ν̂ is
a regular real Borel measure (Corollary 3.4), ν(K ) ν̂(B) ν(G). In particular,
if B = K is compact, ν(K ) ν̂(K ), and if B = G is open, ν̂(G) ν(G). That
ν(F) ν̂(F) + ν(X ) − ν̂(X ) when F is closed follows by finite additivity since
X \ F is open and 0 ν(X ), ν̂(X ) < ∞.
K ⊂ G n ⊂ K n ⊂ G, G n ⊂ G n−1 , K n ⊂ K n−1 ,
ν̂(K ) ν(K n ), ν̂(G) ν(G n ) and λ(K n ) < λ(K ) + 1/n.
Proof Since λ is a regular Borel measure that is finite on compact sets there exist
open sets G k with K ⊂ G k ⊂ G and λ(G k ) < λ(K ) + 1/k for k ∈ N.
Since K ⊂ G k , for k ∈ N there exists, by Lemma 10.5, a continuous function
f k : X → [0, 1] such that f k (K ) = 1 and {x : f k (x) > 0} is a compact subset of G k .
For x ∈ X , let gn (x) = min{ f k (x) : k n} so that gn gn−1 , gn is continuous on
X , gn (K ) = 1 and {x : gn (x) > 0} ⊂ G n is compact.
Let G n = {x : gn (x) > 0} and K n = {x : gn (x) > 0}. Then K ⊂ G n ⊂ K n ⊂
G n ⊂ G and, by Lemma 10.6,
and λ(K n ) < λ(K ) + 1/n because K n ⊂ G n . Now {G n } and {K n } are nested
sequences of open and compact sets, respectively, because gn (x) is decreasing in
n, with the required properties. This completes the proof.
Proof From Lemma 10.6, for any open set G and all compact K ⊂ G, ν(K )
ν̂(K ) ν̂(G). Since ν̂ is a regular real Borel measure, for any > 0 there exists
92 10 Reconciling Representations
compact K ⊂ G with ν̂(K ) > ν̂(G) − . Now by Theorem 10.8 there exists compact
K 1 with K ⊂ K 1 ⊂ G and ν(K 1 ) ν̂(K ) > ν̂(G) − . This establishes the first
identity. For a given compact set K and any open set G with K ⊂ G, ν̂(K ) ν̂(G)
ν(G) and for > 0, there exists an open G with K ⊂ G and ν̂(G) < ν̂(K ) + . By
Theorem 10.8, there exists an open set G 1 with K ⊂ G 1 ⊂ G with ν̂(K ) + >
ν̂(G) ν(G 1 ), and the result follows.
for all B ∈ B.
Proof (a) By Lemma 9.1, either ω(K ) = 0 for all compact K , in which case ω̂ = 0
by the first formula for ω̂(B), or ω(K ) = 1 for some compact K . In the latter case,
10.2 Restriction to C0 (X, ) of Elements of L ∗∞ (X, B, λ) 93
1. Y.A. Abramovich, A.W. Wickstead, Singular extensions and restrictions of order continuous
functionals. Hokkaido Math. J. 21, 475–482 (1992)
2. A.D. Alexandroff, Additive set-functions in abstract spaces I, II, III. Rec. Math. (Mat. Sbornik)
N.S. 8(50), 307–348 (1940); 9(51), 563–628 (1941); 13(55), 169–238 (1943)
3. C.D. Aliprantis, K.C. Border, Infinite Dimensional Analysis—A Hitchiker’s Guide (Springer,
Berlin, 2006)
4. J.M. Ball, Weak continuity properties of mappings and semigroups. Proc. Roy. Soc. Edin.
Sect. A 72(4), 275–280 (1975)
5. S. Banach, Théorie des Opérations Linéaires. Monografje Matematyczne, Tom I, Z Subwe-
hcji Fundusku Kultryy Narodowej, Warszawa, 1932; reprinted with corrections and added
material, Chelsea Publishing Co. New York, 1955; Translated from the French by F. Jellett,
with comments by A. Pelczyński and Cz. Bessaga. North-Holland Mathematical Library, vol.
38. North-Holland Publishing Co., Amsterdam, 1987 & Dover Books
6. K.P.S. Bhaskara Rao, M. Bhaskara Rao, Theory of Charges - A Study of Finitely Additive
Measures, vol. 109, Pure and Applied Mathematics (Academic Press, London, 1983)
7. N.H. Bingham, Finite additivity vs countable additivity: De Finetti and Savage. J. Électron.
Hist. Probab. Stat. 6(1), 35 (2010)
8. V.I. Bogachev, Measure Theory, vol. 1 (Springer, Berlin, 2007)
9. A.G. Chentsov, Finitely Additive Measures and Relaxations of Extremal Problems. Mono-
graphs in Contemporary Mathematics (Plenum, New York, 1996)
10. E.F. Collingwood, A.J. Lohwater, The Theory of Cluster Sets (Cambridge University Press,
Cambridge, 1966)
11. J.B. Conway, A Course in Functional Analysis, 2nd edn. (Springer, New York, 2007)
12. N. Dunford, J.T. Schwartz, Linear Operators, vol. I (Wiley Interscience, New York, 1958)
13. G. Fichtenholz, Sur les fonctions d’ensemble additives et continues. Fund. Math. 7(1), 296–
301 (1925)
14. G. Fichtenholz, L. Kantorovitch, Sur les opérations lináires dans lespace des fonctions bornés.
Stud. Math. 5, 69–98 (1934)
15. G.B. Folland, Real Analysis, 2nd edn. (Wiley, New York, 1999)
16. I. Fonseca, G. Leoni, Modern Methods in the Calculus of Variations: L p Spaces, Springer
Monographs in Mathematics (Springer, New York, 2007)
17. J.B. Garnett, Bounded Analytic Functions (Revised First Edition, Springer, New York, 2007)
18. J. Heinonen, Lectures on Analysis on Metric Spaces (Springer, New York, 2001)
A F
Absolute continuity, 35 Filter, 42
Algebra, 8 Fine structure at x0 of u ∈ L ∞ (X, B, λ), 81
Algebras and σ -algebras of sets, 8 Finitely additive measures, 31
Atom, 11 First axiom of countability, 23
Axiom of choice, 8 Functionals, bounded, 19
Function spaces, 17
B
H
Banach limit, 21, 53, 84
Hahn decomposition, 13
Banach space, 19
Hardy spaces, complex, 84
Base, 23
Hausdorff topology, 24
Blaschke product, 86
Borel
σ -algebra, 9, 77
I
measure, 77
Inner function, 84
sets, 9 Integration, 47
with respect to ν ∈ G, 51
Isolated, 60
C Isometric isomorphism, 58
Canonical decomposition, 38
Characteristic function, 10, 14
Closed G δ -set, 62, 79 J
Cluster set, 84 Jordan decomposition, 13
Cluster value, 77
Completely separable, 24, 60, 81
Complex function theory, 84 L
Convergence Lebesgue
in measure, 10 σ -algebra, 12
λ-almost everywhere, 10 decomposition, 17
Countable, 7 integrable, 16
Counting measure on N, 11 integral, 15
measure, 12
point, 69
Lemma
D
Urysohn, 24, 90
Darboux property, 11
Zorn, 3, 7, 8, 42–44
Denumerable, 7
Local base, 23
Dirac measure, 11
Localising
Directional limit of u at x0 , 82
G, 77
Discrete topology, 19
complex function theory, 84
Dual space, 1, 19
essential range, 81
weak convergence, 80
Locally compact, 18, 77
E
Equality almost everywhere, 42
Equivalence M
class, 8 Maximal, 8, 42
relation, 8 Maximal element, 8
Essential range, 52, 77 Maximum norm, 18
localised, 81 Measurable
Extended real numbers, 7 functions, 9
Extreme point, 22, 64 sets, 9
Index 99
space, 9 T
Measure, 10 Theorem
σ -finite, 10 Alaoglu, 22, 64, 68
Borel outer, inner regular, 11 Alexandroff, 91
finite, 10 Baire’s category, 3, 14, 15
Lebesgue, 12 Banach–Steinhaus, 21
space, 10 Dini, 71
Measure space Dominated convergence, 1
complete, 10 Hahn–Banach, 5, 20, 28, 44, 69
Krein–Milman, 23
Lebesgue–Radon–Nikodym, 16, 90
N Lusin, 19
Normal topology, 24 Mazur, 21, 69
Null sets, 10
Monotone convergence, 49
Nikodym’s convergence, 32
Radon–Nikodym, 17, 50
O
One-point compactification, 18 Rainwater, 23
Oscillatory functions, 74 Riesz representation, 1, 20, 27, 87
Stone–Weierstrass, 21, 59
Valadier–Hensgen, 3, 5, 54, 56
P Weierstrass approximation, 21
Partial ordering, 33 Yosida–Hewitt representation, v, 3, 27,
Poisson kernel, 85 50, 87
Positive and negative parts, 13 Totally disconnected, 60
Purely finitely additive measures, 36 Total ordering, 7
Total variation, 13
Translations, 74
R
Real measure, 12
Reconciling representations, 87
U
Regular
Ultrafilter, 42, 43
Borel measure, 11
Upper bound, 8
inner, 11
outer, 11
real Borel measure, 13
Regular topology, 24 W
Weak*
closed convex hull, 23
S convergence, 22
Second axiom of countability, 24 sequential compactness, 22
Separable, 22, 23 topology, 22
Sequential weak continuity, 67 topology on L ∗∞ (X, L, λ), 63
Shargorodsky, 5, 84 Weak convergence
Shift operator, 53 at a point, 77, 80
Simple function, 10, 73 criterion, 70
Singleton, 7
of sequences, 21, 67
Singular, 35
Splitting measurable sets, 14
Step functions, 73
Sub-base, 22, 23, 57, 63 Y
Symmetric difference, 14 Yosida–Hewitt decomposition, 89