
Citation 
 Permanent Link:
 https://ufdc.ufl.edu/UF00097651/00001
Material Information
 Title:
 Vector measures and stochastic integration
 Creator:
 Witte, Franklin Pierce, 1942 ( Dissertant )
Brooks, James K. ( Thesis advisor )
Varma, Arun K. ( Reviewer )
Dufty, James W. ( Reviewer )
Bedmarck, A. R. ( Degree grantor )
 Place of Publication:
 Gainesville, Fla.
 Publisher:
 University of Florida
 Publication Date:
 1972
 Copyright Date:
 1972
 Language:
 English
 Physical Description:
 v, 108 leaves. : 28 cm.
Subjects
 Subjects / Keywords:
 Algebra ( jstor )
Banach space ( jstor ) Indefinite integrals ( jstor ) Martingales ( jstor ) Mathematical integrals ( jstor ) Mathematical vectors ( jstor ) Mathematics ( jstor ) Measure theory ( jstor ) Perceptron convergence procedure ( jstor ) Scalars ( jstor ) Dissertations, Academic  Mathematics  UF Mathematics thesis Ph. D Measure theory ( lcsh ) Stochastic integrals ( lcsh )
 Genre:
 bibliography ( marcgt )
nonfiction ( marcgt )
Notes
 Abstract:
 This dissertation investigates the stochastic integration of scalarvalued functions from the point of view of vector measure and integration theory. We make a detailed study of abstract integration theory in Chapters I and II, and then apply our results in Chapter III to show that certain kinds of stochastic integrals, previously defined by other means, are special cases of the general theory. In carrying out this program we prove extended forms of the classical convergence theorems for integrals. We also establish a generalization of the standard extension theorem for scalar measures generated by a left continuous function of bounded variation. The special case of measures in Hilbert space is discussed, and a corrected form of a theorem of Cramer is proved. In Chapter III we show that certain sample path integrals, the WienerDoob integral, and a general martingale integral are included in the abstract integration theory. We establish a general existence theorem for stochastic integrals with respect to a martingale in L , 1 < p < infinity>.
 Thesis:
 Thesis  University of Florida.
 Bibliography:
 Bibliography: leaves 105107.
 Original Version:
 Typescript.
 General Note:
 Vita.
Record Information
 Source Institution:
 University of Florida
 Holding Location:
 University of Florida
 Rights Management:
 Copyright [name of dissertation author]. Permission granted to the University of Florida to digitize, archive and distribute this item for nonprofit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
 Resource Identifier:
 022670731 ( AlephBibNum )
13989845 ( OCLC ) ADA5262 ( NOTIS )

Downloads 
This item has the following downloads:

Full Text 
tegr*r
des
L
F
'TO TH E
UNIVE ITY OF FLORI
THE ENTS FOR HE op
0 roRJIXMILOSOPHY
FDW
1 m
Lk
UNIVERSITY OF FLORIDA
SIII lll1 llltl II 11
3 1262 08552 4642
.. ... ll ......... ....IlilllllliiC iii" " i;
..... . ......... ...l~
......... .... ... ...""'";;iiii iiiII
.... ... ....i
iii
S
To Do
Whose patience and encouragement made this possible.
ACKNOWLEDGEMENTS
I would like to thank my wife Dot for her under
standing and encouragement throughout the past four years.
I am indebted to each member of my conunittee; special
thanks are due to Dr. James K. Brooks, who directed my
research and guided my study of abstract analysis, and
to Dr. Zoran R. PopStojanovic, who guided my study of
probability theory. Finally, I would like to thank
Jean Sheffield for her excellent typing.
iii
TABLE OF CONTENTS
Acknowledgements
Abstract
Introduction
Chapter
I. Abstract Integration
II. Stieltjes Measures and Integrals
III. Stochastic Integration
References
Biographical Sketch
Page
iii
105
108
Abstract of Dissertation Presented to the Griaduate
Council of the University of Florida in Partial Fulfillment
of the Requirements for the Degree of Doctor of Philosophy
VECTOR MEASURES AND STOCHASTIC INTEGRATION
By
Franklin P. Witte
December, 1972
Chairman: J. K. Brooks
Major Department: Department of Mathematics
This dissertation investigates the stochastic integration
of scalarvalued functions from the point of viev. of vector
measure and integration theory. We make a detailed study of
abstract integration theory in Chapters I and II, and then
apply our results in Chapter III to show that certain kinds
of stochastic integrals, previously defined by other means,
are special cases of the general theory. In carrying out
this program we prove extended forms of the classical con
vergence theorems for integrals. We also establish a gener
alization of the standard extension theorem for scalar
measures generated by a left continuous function of bounded
variation. The special case of measures in Hilbert space is
discussed, and a corrected form of a theorem of Cramer is
proved. In Chapter III we show that certain sample path in
tegrals, the WienerDoob integral, and a general martingale
integral are included in the abstract integration theory.
We establish a general existence theorem for stochastic in
tegrals with respect to a martingale in Lp, 1 < p < m.
v
INTRODUCTION
This dissertation concerns stochastic integration, a
topic in the theory of stochastic processes. A stochastic
process (random function) is a function from a linear in
terval T into an L space over a probability measure space.
In its broadest terms, stochastic integration deals with
linear transformations from a class of functions x to a class
of stochastic process fxdz, which depend on a fixed process z.
The functions x which are transformed may be sure (scalar
valued) functions, or more generally, random functions. Since
random functions take their values in an L space they are
vectorvalued functions. Hence we are in the framework of the
general theory of bilinear integration developed by Bartle [1].
The vectorvalued function z generates a finitely additive
measure on an appropriate ring or algebra of sets. When the
integrand x is a sure function we are concerned with the in
tegration of scalar functions by vector measures. When the
integrand is another random function we have the full gener
ality of the Bartle Theory, where both the integrand and the
measure take values in Banach spaces X and Z respectively,
and there is a continuous bilinear multiplication from XxZ
into another Banach space Y, in which the integral takes its
values.
2
Until recently the field of stochastic integration
appeared to be in much the same state as the area of real
function theory was before the systematic introduction of
measure theory. Several relatively unrelated kinds of
stochastic integrals can be found throughout the literature
on stochastic processes and their applications. Most of
these integrals are defined as limits of approximating sums
of the RiemannStieltjestype. This approach, however, is
not wellsuited to stochastic integration in general, and
certain difficulties may arise. With the exception of some
work by Cramer [10], the measure properties inherent in the
stochastic integration scheme have not been exploited.
There is currently a great deal of interest in unifying
the field of stochastic integration using vector measure and
integration theory. E. J. McShane, who has been investigating
stochastic integration for several years, is presently writing
a book which should contribute to this unification. The
recent Symposium on Vector and Operatorvalued Measures and
Applications (Salt Lake City, August, 1972) was specifically
concerned with stochastic integration, and attracted many of
the known experts in both the fields of measure and integration
theory (Brooks, Dinculeanu, Ionescu Tulcea, Kelley, and
Robertson) and stochastic processes and stochastic integration
(Chatterji, Ito, Masani, and McShane). A very recent un
published paper by Metivier [23], presented at the Symposium,
deals with the problem of incorporating a rather general kind
of stochastic integral within the framework of the Bartle in
tegration theory.
3
The purpose of this dissertation is more modest. We
confine our attention to the case of scalarvalued inte
grands, anid present a detailed investigation and synthesis
of the appropriate vector measure and integration techniques,
which are then used to exhibit certain kinds of stochastic
integrals as special cases. In carrying out this program we
make improvements in known results concerning the convergence
of sequences of integrals, measures generated by vector
functions, and existence of stochastic integrals.
Chapter I develops the abstract integration theory of
Bartle [1] for the special case at hand, namely, the integra
tion of scalar functions by finitely additive vector measures.
In this setting more is true than in the general Bartle theory.
For example, Theorem 1.5.8 states roughly that the integral
commutes with any bounded linear operator. This fact enables
us to prove extended forms of the classical convergence
theorems, and sketch a theory of L spaces. We discuss
briefly the case of integration by countably additive vector
measures, and introduce another integral developed by Gould
[16]. This integral is defined by a net of RiemannStieltjes
type sums, as opposed to the Lebesgue approach used in the
development of the Bartle integral by means of simple functions.
We show that these two integrals are equivalent a fact not
discussed by Gould.
Since a stochastic process is a function from an interval
T of the line into an Lp space, as will be discussed in Chap
ter III, the stochastic integral of a scalar function f on T
4
with res':ect to z can be thought of as a Stieltjestype
integral Ff(t)z(dt) in Lp. With this in mind, we inves
tigate in Chapter II the properties of a Stieltjes measure
m generated by a vector function z on T. We discuss the
boundednzcs of m, and the question of the existence of a
countably additive extension to the Borel sets of T. For
a rather general class of Banach spaces we show in Theorem
11.3.9 that the classical extension theorem for scalar
measures on an algebra is valid. We also give a counter
example to show that the theorem fails to hold in the Banach
space c Finally we discuss the special case when z takes
its values in a Hilbert space. We consider a theorem of
Cramer [10], which is incorrect as stated, and apply our
results to establish a corrected version. Some implica
tions of this theorem are then discussed.
In Chapter III we first present some background infor
mation on probability and stochastic processes, and then
motivate the study of stochastic integration. We show that
the general integration theory of Chapters I and II includes
certain sample path stochastic integrals, the WienerDoob
integral, and a martingale stochastic integral. Several
related results are also discussed.
CHAPTER I
Abstract Integration
In this chapter we discuss the integration of scalar
functions by finitely additive vector measures. Section 1
establishes the notation we will use, while Sections 2 and 3
list some basic properties of vector measures. In Section 4
we discuss the space of measurable functions and convergence
in measure. Section 5 is concerned with integration theory
proper, and in Section 6 we apply this theory to prove con
vergence theorems for integrals. Section 7 outlines part
of the theory of Lp spaces in this setting. Countably addi
tive measures are discussed briefly in Section 8. Finally,
in Section 9,we compare a Riemanntype integral with the
Lebesguetype integral of Section 5, and establish the equiva
lence of these two approaches.
1. Notation. Throughout this dissertation the following
terminology and notation will be used. N is the set of
natural numbers (1,2,3,...) and Nt = (l,2,...,n). The lower
case letters i,j,k,d, and n will always denote elements of N.
A is a finite subset of N. R denotes the real number field;
we use the usual notation for intervals in R, so that (a,b)
.is an open interval, (a,b] a halfopen interval, and so on.
6
X and Y are Banach spaces over the scalar field 0,
which may be either the real or the complex field unless
otherwise specified. Ix denotes the norm of an element
x ( X. X1 is the unit ball of X, that is, X1= [x EX: xl s 11.
B(X,Y) is the Banach space of bounded linear transformations
from X to Y, with the usual topology of uniform convergence.
X* = B(X,f), and XI* is the unit ball of X*. Recall that by
definition,
Ix*l = sup Ix*xl ,
X
x1
while by the HahnBanach Theorem,
Ix = sup jx*xl
X *
Suppose that R and E are families of subsets of a non
empty set S. R is a ring if R is closed under the formation
of relative complements and finite unions. A ring R is a
aring if R is closed under countable unions. E is an
algebra if Z is a ring and S E E. An algebra E is a o
algebra if E is closed undercountable unions. If 8 is any
family of subsets of S, then a(g) denotes the smallest o
algebra containing g. If E E R, then n(E,R) denotes the
family of all partitions P = (Ei:1 i Sn} of E whose sets
belong to P. When R is understood we write n(E). If (Ai)
is a sequence of subsets of S which are pairwise disjoint,
we call (Ai) a disjoint sequence. Whenever we write A = EA.
instead of A = LIAi for some set A and sequence (A.), we mean
that (Ai) is a disjoint sequence, and say that A is the sum
of the A.'s. Similarly we write A + B instead of A U B, if
A n B = 0.
7
Suppose that R is a ring and jn: X is a set function.
m is finitely additive if m(A + B) = m(A) + m(B). m is
countably additive if m(FIA.) = Zm(A.) whenever (A.) is a
sequence in R such that EAi E R. We make the convention
that whenever we write m(A) it is assumed that A belongs to
the domain of m. A finitely additive set function defined
on a ring or algebra will be called a measure.
If m:R [o,w] is a set function, then m is monotone if
m(A) : m(B) whenever A c B and A,B E R. m is subadditive if
m(A U B) m(A) + m(B) for every A,B E R. mi 1 m2 means
ml(E) s m2(E) for every E E R.
Standard references for results in measure and integra
tion theory are Dinculeanu [12], Dunford and Schwartz [15],
and Halmos [17]. We adhere to the notation of Dunford and
Schwartz unless otherwise noted.
2. Properties of Set Functions. In this section we discuss
some elementary properties of measures and their associated
nonnegative set functions. R is a ring and m:R X is a
measure unless otherwise noted. m is bounded if the set
m(R) is a bounded subset of X, that is, sup EE m(E)j < The
total variation v(m,E) of the measure m on a set E E R is de
fined by
n
v(m,E) = sup E Im(Ei) .
T(E) i=l
It is well known (see Dinculeanu [12]) that the set function
v(m) is finitely additive but may not be finite even though m
is a bounded countably additive measure. As Theorem 2.2 (ii)
P8
infra shows, v(m) is bounded whenever m is a bounded scalar
measure. For this reason v(m) plays an important role in
the theory of integration with respect to a scalar measure m,
since it is a nonnegative measure .'.hich dominates m. When m
is a countably additive vector measure and v(m,E) < c for
each E R, Dinculeanu [12] has shown that most of the results
from the scalar integration theory carry over to the vector
case. In general, however, the total variation may be in
finite and hence of little value. Thus we must use a more
delicate device known as a control measure for m (see Theorem
3.3 infra). We now introduce a set function i called the
variation of m. For every E 6 R we define
if(E) = sup Im(F) .
F=E
FE6
It is obvious that m is bounded if and only if m is bounded.
The following Lemma will be used several times in this
dissertation.
2.1 Lemma. If aI,...,a are scalars, then
n
E Ia I 4 4 sup I ai .
i=l AGNn iEA
n
In particular, if Xl,...,xn E X and x* C X*, then
n
E Ix*xi. 4 sup I x*x.I S 41x* sup I xil,
i=l A=Nn iEA A=Nn iA
n
and I a.xi. 4 sup a.il sup I xil.
i=l lsien A=Nn iEC
9
Proof. LetAl, ...,L4 denote the sets of integers i such
that Re a. > 0, Re a. 0, Im a. > 0, and Im a. 0 respec
1 1 1 1
tively. Then
n n n
S a ail I Re aji + I IIma.i
= E Re a. E Re a. + I Ima.E Ima.
= Re a. E Re a. + Z Ima. Z Im a
A1 A2 A3 4
=Re E a Re E ai + Im E a. Im E a.
1 A2 A3 A4
4 sup ail
AcN iEA
where we have used the elementary inequalities a I 5 I Re a +
IImal and IRe a limal 5 lal for any a E '. The inequa
lities concerning the x*x.'s are then immediate. Since
1
n n n
I ai.il = Ix* E a.xil s sup Jai. E Ix*xil
i= 1 i=l igsin i=l
for some x* E XI*, the final inequality follows from the
preceding one. O
As a first application of this lemma we establish the
properties of the variation of a vector measure.
2.2 Theorem. (i) if is nonnegative, monotone, and subadditive.
(ii) If m is scalarvalued, then v(m) s 4fi.
(iii) Im(E) s m(E) s sup v(x*m,E) s 4ff(E), for EE R.
Xl*
Proof. (i) fi is clearly nonnegative and monotone by defini
tion. If E,F, and G belong to 9 and G Q E U F, then
10
G = (G n E) + (G n (F\E)), so that
m(G)j = Im(G n E) + m(G n (F\E))
Im(G n E) I + Im(G n (F\E))j
Sfm(E) + mi(F).
Therefore f(E U F) e mi(E) + m(F) for every E,F E R.
(ii) If m is scalarvalued and (E.) E r(E), then
n
E Im(Ei ) 4 sup I m(Ei)  4 sup Im(F)
i=l A=Nn iEA FeE
FER
by Lemma 2.1. Thus v(m,E) s 4im(E) for every EE R.
(iii) Suppose that E E R. It is obvious that
Im(E) fii (E). Using (ii) we have
fi(E) = sup Im(F)I
FcE
FER
= sup sup Ix*m(F)I
X FcE
FER
sup v(x*m,E)
X1*
4 sup sup Ix*m(F)l
X FCE
FER
= 4i(E). 0
2.3 Remark. By the principle of uniform boundedness (see
Dunford and Schwartz [15]) any function with values in X is
bounded if and only if it is weakly bounded. In particular,
m is bounded if and only if x*m is bounded for every x* E X*.
If R is a cring and m is countably additive, then each x*m
is also countably additive. It is known (see Halmos [17])
11
that a countably additive scalar measure on a aring is
bounded. Hence m is also bounded. The OrliczPettis Theorem
(see Dunford and Schwartz [15]) states that, conversely, if
x*m is countably additive for each x* E X* and R is a oring,
then m is countably additive.
The hereditary ring 3 generated by a ring R is the
family of all subsets of S which are contained in some element
of R. If p:R [o,m] is a monotone set function, then we
can extend p to A as follows. Define
i(E) = inf m (F),
F=E
FER
for every E E M. It is immediate that p is a monotone
extension of p, and p 2 0. If s is subadditive then so is
1
u To see this, suppose that G,H E M and c > 0 is fixed.
Choose E,F E R such that G c E, H z F, and p(E) < p (G) + c,
u(F) < p (H) + e. Then G U H c E U F and so
p (G U H) r p(E U F) p(E) +p(F) < p (G) +p (H) + 2e.
1 1 1
We conclude that L (G U H) p (G) + u (H). In this way we
can extend M to N. A set A E N is said to be an mnull set,
or simply a null set when m is understood, if f(A) = 0. We
shall always identify the extension of ff to by fi unless there
is a possibility of confusion. Note that A is a null set if
and only if for every e > 0, there is an E E 9 such that A c E
and m(E) < c.
3. Strongly Bounded Measures. We now introduce a fundamental
property of certain vector measures, and discuss some of its
implications. R is a ring and m:9 X is a measure unless
otherwise noted.
12
3.1 Definitions. (i) m is said to be strongly bounded
(sbounded) if for every disjoint sequence (E.) in 2,
lim m(Ei) = 0.
This property was introduced by Ricknar [25], and later
used by Brooks and Jewett [6] to establish convergence
theorems for vector measures. Brooks [4] showed that s
boundedness is equivalent to the existence of a control
measure. This in turn provides necessary and sufficient
conditions for the existence of an extension of a countably
additive measure from E to o(E) when is an algebra. These
results are presented in Theorems 3.3 and 3.4 infra.
(ii) A series Fx in X is unconditionally convergent
if for every e > 0 there is a L such that A r = 0d implies
E 6
thatI x. < c.
iEti
The following result states some basic properties of an
sbounded measure. The proof is omitted.
3.2 Theorem. (Rickart [25]) (i) If m is sbounded then m
is bounded.
(ii) The following statements are equivalent:
(a) m is sbounded.
(b) For every disjoint sequence (Ei) in R we have
lim m(Ei) = 0.
(c) For every disjoint sequence (E.) in R, Zm(Ei)
is unconditionally convergent.
13
Suppose that r, is a family of measures on R. We say
that the measures in i are uniformly mcontinuous, or uni
formly continuous with respect to m, if for every e > 0
there is a 6 > 0 such that Iu(A) < e for every U in if
Fm(A) < 6. When l contains only one measure p we say that p
is mcontinuous. p and m are mutually continuous if each
is continuous with respect to the other.
The most important characterization of sbounded measures
is the following result due to Brooks [4] which we state with
out proof.
3.3 Theorem. (Brooks) m:R X is sbounded if and only if
there is a bounded nonnegative measure m on R such that
(i) m is mcontinuous.
(ii) m a m on R.
Moreover, m is countably additive if and only if m is count
ably additive.
The measure m is called a control measure for m. (i) and
(ii) show that m and m are mutually continuous; it follows
that m and m have the same class of null sets.
A fundamental problem in measure theory is to find con
ditions under which a countably additive measure on a ring R
has a countably additive extension to the aring generated
by R. Using the previous result of Brooks we can prove
an extension theorem for countably additive sbounded measures
on an algebra. This subject has been investigated by many
authors. Takahashi [28] introduced a kind of boundedness
14
condition less natural than, but equivalent to, the concept
of sboundedness. His result in the case of an algebra is
therefore included in the following theorem. A measure
m:R X is said to be Tbounded if for every e > 0 there
is an n E N such that if El,...,En are disjoint sets in R,
then Im(Ei)I < E for some i E Nn
To see that Tboundedness is equivalent to sboundedness,
we remark that if m is not sbounded then there is a disjoint
sequence (Ei) and an e > 0 such that Im(Ei)I > e, i E N, so
m cannot be Tbounded. This shows that if m is Tbounded
then m is sbounded. Conversely, suppose that m is sbounded
with control measure m as in Theorem 3.3. Since m is a
bounded nonnegative measure there is a constant K such that
n
E m(E.) s K for every finite disjoint family (Ei). If m
i=l
is not Tbounded, then there is an e > 0 such that for every
n E N there are disjoint sets Eln,...,Enn in R with m(Ein ,
1 s i s n. This clearly contradicts the boundedness of m, so
m is Tbounded. Since m and m are mutually continuous, m is
also Tbounded.
We now state the extension theorem for measures on an
algebra.
3.4 Theorem. (Brooks [5]) Suppose E is an algebra and
m:E X is countably additive and sbounded. Then m has a
unique countably additive extension to ao().
Remark. Conversely, if m has a countably additive extension
to a (E), then m is sbounded on E.
15
Proof. Let m be a control measure for m as in Theorem 3.3.
Then m is countably additive and bounded on E. It is well
known (see Halmos [17]) that a nonnegative, bounded, count
ably additive measure on an algebra has a unique extension
to a bounded, countably additive measure on a(E) = E1. We
identify this extension by m.
The symmetric difference of two sets A and B, denoted
by A A B, is defined by A A B = (A\B) + (B\A). We also have
A A B = (A U B)\(A n B). The elementary set relations
(i) A C AA B U B A C,
(ii) A U B A C U D A A C U B A D,
(iii) A n B A C n D C A A C U B A D,
(iv) A\B A C\D C A C U B A D,
are straightforward to verify. Define d:E1 x Z1 [O,m) by
d(A,B) = m(A A B). Since m is subadditive, (i) shows that d
satisfies the triangle inequality, and hence is a semimetric
for E1. Moreover, from the mcontinuity of m on E and from
the inequalities m(A\B) s m(A A B) and Im(A) m(B)I Im(A\B) +
Im(B\A) it follows that m is a uniformly continuous mapping
from E into X.
Since Z1 is the smallest oalgebra containing E, it is
known that for every A E E1 there is a sequence (A ) in Z such
that lim m(A A A) = 0 (see Halmos [17]). This fact shows
that E is a dense subset of El. If we identify sets A and B
in EI whenever d(A,B) = 0, then m is welldefined on E, since
m(A) = m(B) whenever d(A,B) = 0 for A,B E E. Since X is com
plete, m has a unique extension to a continuous mapping from
E1 into X. We denote this extension by m.
16
To see that m is finitely additive on Z1, we remark
that (ii) (iii), and (iv) show that the mappings (A,B) Ai B,
A n B, and A\B respectively from 771 E 1 are continuous.
Suppose A,B E 21 and A n B = ,. Let (A ) and (B ) be sequences
1 n n
in Z converging to A and B respectively. (ii) and (iv) imply
that the mapping (A,B) A B is also a continuous function,
so lim d(A B A A B) = 0. By the continuity of m we have
lim m(An A B ) = m(A A B) = m(A + B) since A A B = A + B.
But m(An A Bn) =m(A\Bn ) + m(B \An), and by the continuity
of (A,B) A\B we have lim m(A \B ) = m(A\B) = m(A) and
n n
lim m(Bn\A ) = m(B). Thus m(A + B) = m(A) + m(B).
n n
Finally, to see that m is countably additive on El it
suffices to show that if Ai.' then lim m(A.) = 0. Since m
is countably additive, lim m(Ai) = 0. This in turn implies
that lim m(Ai) = 0 since m is continuous on the space (Zl,d). O
We quote without proof a final result from the theory
of vector measures. The following theorem is a generalization
of the wellknown VitaliHahnSaks Theorem.
3.5 Theorem. (Brooks and Jewett [6]). Suppose R is a oring
and v: [0,I ) is a bounded measure. If p :R X is a v
continuous measure for each n, and if lim un(E) exists for
each E E R, then the pn are uniformly vcontinuous.
4. Convergence in Measure and the Space M(m). This section
introduces the concept of convergence in measure for sequences
of functions. Using this concept we can define the class of
17
functions with which the integration theory of Section 5
will be concerned. From now on Z will denote an algebra
and m:Z X is a bounded measure. Then i is a nonnegative,
monotone, subadditive, bounded set function, and, in fact,
these are the only properties that are needed to obtain
the results of this section. F(S) denotes the family of
all scalar functions on S. If f E F(S) we use the abbre
viation [Ifl > c] for the set (s E S:jf(s)j > ) .
4.1 Definition. A sequence (fk) in F(S) is Cauchy in measure
if for every c > 0 we have
lim [ f.i f. > e] = 0.
i,jc 1 J
A sequence (fk) in F(S) converges to f E F(S) in measure if
for every e > 0 we have
lim m[I fk fl > ] = 0.
kc
If (fk) converges to f in measure we write m lim fk = f.
The following lemma gives a useful equivalent formulation of
convergence in measure.
4.2 Lemma. Suppose f,fk E F(S), k 6 N. Then m lim fk = f
if and only if there are sets Ak E 2 and a sequence of positive
numbers ek converging to zero such that lim M(Ak) = 0 and
Sfk(s) f(s)I E k if s Ak.
Proof. Let (Ak) and (Ck) satisfy the stated properties, and
suppose e,6 > 0 are given. Choose n such that ek < e and
m(Ak) < 6 if k > n. Since [Ifk fl > e] c [fk fl >Ck] Ak
if k n, we conclude that m[I fk f > e]< 6 ifk n. Thus mlim fk=f.
] ?!
Conversely, suppose rn lim f. = f. For f E F(S),
Dunford and Schwartz [15] define the quantity
lfll = inf ( + fm[i f > e]),
e>0
which is finite since in is bounded. To see that liml[fk fli= 0,
suppose, to the contrary, that liml[fk fl > 0. Without loss
of generality we may suppose that lifk fl > c for k E N and
some c > 0. There is an n such that a[I fk f > e/2] < ,/2
if k 2 n. Hence
E < lfk fll ,/'2 +F[ fk fl > e/2] < e,
which is a contradiction. Therefore limllfk fil = 0. For
each k E N choose Ek > 0 such that ek + m[Ifk fl > Ek] <
< Ifk fl + 1/k. Then choose A, F such that [Ifk fl > k] C Ak
and m(Ak) < m[ fk fl > Ck] + 1/'k. The sequences (Ak) and (ek)
evidently satisfy the properties stated in the lemma. [
If E c S, then IE denotes the indicator function of E,
that is, the function whose value is 1 for s F E and 0 for
s f E. We now define an important class of functions.
4.3 Definition. A Esimple function is any function f E F(S)
which can be represented in the form
n
f = a I ,
i=l i
where a. E 4, and the sets E. E E are disjoint and satisfy
1 1
S = ZE.. When Z is understood, f is called a simple function.
Note that every simple function f has a canonical repre
sentation with E. = [f = a.].
4.4 Definition. f E F(S) is measurable if there is a sequence
of simple functions converging to f in measure.
19
For f E F(S) and E c S we define the oscillation of
f on E, written 0(f,E), by
0(f,E) = sup If(s) f(t) .
s,t6E
We have the following characterization of measurability.
4.5 Theorem. f is measurable if and only if for every c > 0
there is a partition (E ) E n(S) such that m(E ) < e and
0(f,E.) < e for 2 i s n.
Proof. Suppose (fk) is a sequence of simple functions con
verging to f in measure. Let (Ak) and (ek) satisfy the
conditions of Lemma 4.2 for f and (fk). Fix e > 0, and
choose k such that i(Ak) < e and Ek < c/3. Suppose that
n
fk = ai IE.; define a partition of S by setting F1 = "k
i=l 1
and F. = Ei_ \Ak for 2 i i s n+ 1. If s,t E F. for some
i > 1, then If(s) f(t) I I f(s) fk(s)I + I fk(t) f(t) l 2c/3.
Hence 0(f,F.) < e for 2 s i s n+l, so (Fi) is the desired
partition.
Conversely, if the conditions of the theorem hold then
choose a sequence of partitions Pk for e = 1/k. If Pk = (Eik)'
set
nk
f = f(sik )I
i=l ik
where sik E Eik for 1 r i s nk. If s X Elk then
If(s) fk(s) r max 0(f,Ei) 5 1/k. Since mi(Elk) < 1/k ,
2sisnk
the sequences (Elk) and (1/k) satisfy the conditions of
Lemma 4.2, and so mlim fk = f. We conclude that f is
measurable. ]
20
Let M(m) = M denote the space of all measurable functions
in F(S). Since we have introduced a concept of convergence
in F(S) we can discuss the topological properties of M. The
following result from Dunford and Schwartz [15] lists some
of the useful properties of the space M. We shall omit the
proof.
4.6 Theorem. (Dunford and Schwartz) M is a closed linear
algebra in F(S). If g:' R is continuous, then the mapping
f gof is a continuous function from M into M.
The following result will be very useful in connection with
the Dominated Convergence Theorem.
4.7 Theorem. If (fk) is a sequence of simple functions and
mlim fk = f, then there is a sequence (gk) of simple functions
such that mlim gk = f, and Igkl 2fl on S, for each k E N.
Proof. Let (Ak) and (ek) be sequences as in Lemma 4.2 for f
and (fk). Define gk(s) = fk(s) if s E Ak and Ifk(s)I > 2ek'
and gk(s) = 0 otherwise, for each k. Then if s Ak and
Ifk(s)I > 2Ck, we have gk(s) f() f( < ek. If s Ak and
Ifk(s)l 2ek then
If(s)l If(s) fk(s) + Ifk(s) I g 3Ek'
so If(s) gk(s)I g 3ek if s A Ak. By Lemma 4.2 we conclude
that mlim gk = f. If s E Ak or if Ifk(s) s 2ek, then
gk(s) = 0 so certainly Igk(s)I s 21f(s)I. If s e Ak and
SIfk(s)I > 2ek, then
If(s) I fk(s)I fk(s) f(s)
Sfk(s) 
1 /2 fk(s)i = 1/21gk(s) .
Thus Igk(s)I < 21f(s)l for s E S and k E N. ]
5. A Lebesguetype Integral. We now present the standard
results of integration theory. Many of these theorems were
stated by Bartle [1] in his fundamental paper; they are
included for completeness. Throughout this section Z is
an algebra and m:E X is a bounded measure.
If f is a simple function with representation
n
2 a. IE we define the integral of f over E E 2 by
i=l i
n
SEfdm = a aim(E n E.).
i=l
The following result lists the basic properties of the inte
gral for simple functions. The set function X(E) = fEfdm is
called the indefinite integral of f.
5.1 Theorem. (Bartle) (i) The integral of a simple function
is independent of the function's representation.
(ii) For a fixed set E E 2 the integral over E is a
linear map from the linear space of simple functions into X.
(iii) For a fixed simple function f the indefinite
integral of f is a measure on E.
(iv) If f is a simple function bounded by a constant K
on a set E, then
JEfdmJI 4Kmf(E).
22
(v) If U 4 B(X,Y) and f is a simple function, then for
every E CE ,
SEfdUm = U(SEfdm).
Note that by (iv) the indefinite integral of a simple
function is a bounded and mcontinuous measure. If m is
sbounded, then so is the indefinite integral of a simple
function.
The Bartle integral is defined using sequences of simple
functions. The next two theorems discuss the properties of
these sequences which enable us to construct the general
integral. Bartle proved only that (i) implies (ii) in
Theorem 5.2.
5.2 Theorem. Suppose (fk) is a sequence of simple functions
that is Cauchy in measure. Then the following statements
are equivalent:
(i) The indefinite integrals of the fk s are uniformly
mcontinuous.
(ii) lim rEfkdm exists uniformly for E E E.
Proof. Suppose (i) holds and e > 0 is given. Choose 6 > 0
such that E E E and ii(E) < 6 implies that for each k E N,
IJEfdml < e.
There is an n E N and sets Ai. E Z such that mf(A ) < and
If.(s) f.(s)I < e/4f(S), if s Aij and i,j > n. Then for
E E and i,j > n, we have
IfEfidm SEfjdml S ('E\A ij(f f)dml +
+ ISEnAijf'dml + IErAijf f.dm
s [4e,/4fmi(S)]f(E\\Aij) + 2e < 3e,
23
using (iii) and (iv) of Theorem 5.1. Since E was arbitrary,
we conclude that the sequences (JEfkdm) are Cauchy uni
formly for E E 2. Since X is complete, (ii) holds.
Conversely, if (ii) holds and e > 0 is given, then
there is an n E N such that for every E E 7,
IE(fk fn)dml < C,
provided k 2 n. Since each indefinite integral is mcontinuous,
there is a 6 > 0 such that if A E E and m(A) < 6 then
IJAfkdmI < e,
for 1 s k s n. If fi(A) < 6 and k : n, we also have
SSAfkdm) 1 A (fk fn)dml + (SAfndmI
S2e.
Therefore the indefinite integrals of the fk 's are uniformly
mcontinuous. r]
5.3 Theorem. Suppose (fk) and (gk) are sequences of simple
functions such that
(i) mlim(fkgk) = 0.
(ii) The indefinite integrals of the fk's and the gk's
are uniformly mcontinuous. Then
lim SEfkdm = lim JEgkdm
uniformly for E E S.
Proof. Since hk = fk gk is a simple function for each k,
and since
Ehkdml < ISEfkdmI + ISEgkdml
for every E E 2, Theorem 5.2 implies that lim r Ehkdm exists
for every E E E. We need only show that this limit is zero
uniformly in E. Fix > 0. There is a 6 > 0 such that if
A E and mf(A) < 6, then for each k G N,
I Zh dm I < e.
Since inlim h = 0, there is an n E N and sets Ak E E such
that mf(Ak) < 6 and hk(Z l < e .'4(S) if s Ak, provided
k n. Thus for E E E and k n,
I Ehidml I fE ,khkddmI + IJEgAkdm
S[4c/''i(S) ]m(E\Ak) +
< 2e.
That is, lim Ehkdm = 0 uniformly for E E E. 0
Following Bartle [1] we now define the general integral.
5.4 Definition. f E F(S) is integrable if there is a sequence
(fk) of simple functions such that
(i) mlim fk = f
(ii) The indefinite integrals of the fk 's are uniformly
mcontinuous.
If f is integrable then any sequence (fk) of simple
functions satisfying (i) and (ii) of Definition 5.4 is said
to determine f. Theorem 5.2 shows that if (fk) determines f,
then lim YEfkdm exists uniformly in E. We denote this limit
by the usual symbols
YEfdm or SEf(s)m(ds).
Theorem 5.3 shows that the integral of f is independent of
the determining sequence. Let L(m) denote the family of all
functions that are integrable with respect to m. The following
25
theorem lists some standard properties of the integral. We
omit the proof.
5.5 Theorem. (i) For a fixed set E E E, the integral over
E is a linear mapping from the linear space L(m) into X.
(ii) For a fixed function f E L(m), the indefinite
integral of f is an mcontinuous measure on E.
Suppose f E M and there is a null set A such that
supsAI f(s)I < a. Then we say that f is essentially bounded,
and define the essential supremum lfll[ of f by
llfll = inf sup If(s) .
m1 (A)=0 s'A
The following standard result, whose proof refines that of
Bartle, shows that every measurable and essentially bounded
function is integrable.
5.6 Theorem. If f E M is essentially bounded, then f E L(m)
and for every E E E we have
IJEfdml 4l1flim(E)
Proof. Let K = Ilfll, and suppose (fk) is a sequence of simple
functions converging to f in measure. Define
fk(s) Ifk(s) I K,
gk(s) =
(K/i fk(S) )fk(s) I fk(s) > K,
for each s E S and k E N. Then (gk) is a sequence of simple
functions, and
Ifk(s)( K Ifk(s) > K
Igk(s) fk(s) I =
0 {fk(s) I K,
26
since Igk(s) fk(s) I = fk(s)(1K/Ifk(s) ) = fk (s) K
if Ifk(s)I > K. Suppose e > 0. Since f is essentially
bounded by K, there is a null set A such that If(s)I
if s e' A. Now if s ( A and Ifk(s)j > K + e/2, then since
If(s)I : K + e/4, we have
Ifk(s) f(s) I fk(s)1 If(s)
> K + e/2 K e/4 = c/4.
Therefore [Igk fk > e/2] = [Ifk > K + e/2]
A U [Ifk fl > e/4].
Finally,
[I9k fl > e] c [19k fkl > 3e/4] u [Ifk fl > /4]
c A U [Ifk fl > e/4].
Since A is null and mlim fk = f, we conclude that
mlim gk = f. Since the gk's are uniformly bounded by K,
Theorem 5.1 (iv) implies that for every k,
fEgkdml < 4Km(E)
for any E E E. This shows that the indefinite integrals
of the gk's are uniformly mcontinuous. It follows that
f E L(m), and that for every E E S,
jEfdm = limIEfgkdml < 4Km(E). .
Using Theorem 5.6 and the mcontinuity of the indefinite
integral of a function f E L(m) we have the following Corollary.
5.7 Corollary. If f E L(m), then the indefinite integral of
f is a bounded measure on E.
27
Proof. Since f is measurable, f is bounded except possibly
on sets of arbitrarily small variation. Choose 6 > 0 such
that if m(E) < 6, then
I(Efdml < 1.
Choose E E E such that m(E) < 6 and f is bounded on S\E, say
by K. For any A E E, therefore
'LAfdml IsA\Efdm + IfAfEfdml
: 4Kf(S) + 1. 0
When E is a oalgebra and m is a countably additive
vector measure, Bartle, Dunford and Schwartz [2] have shown
that the integral satisfies the following property. If
f E L(m) and U E B(X,Y), then f E L(Um) and for each E E,
SEfdUm = U(JEfdm).
We now extend this result to the case where m is a bounded
measure on an algebra E.
5.8 Theorem. Suppose that U E B(X,Y). If f E L(m) then
f E L(Um), and for every E E E we have
SEfdUm = U('Efdm).
Proof. Since IUm(E)I r IUIm(E) l.it follows at once that
Um(E)I 5 IUIm(E) for every E E E. From this inequality we
see that if mlim fk = f, then Um lim fk = f as well. If
f E L(m) and (fk) is a sequence of simple functions deter
mining f, then we need only show that the indefinite integrals
of the fk's with respect to Um are uniformly Umcontinuous.
Since we have
28
iJ(f ; fj)dUmI < IU1SE(fk fj)dml,
using Theorem 5.1 (v), this follows immediately from
Theorem 5.2 and the fact that lim 3Efkdm exists uniformly
for E E C:. ]
When m is a scalar measure, Dunford and Schwartz [15]
have introduced the following definition of integrability.
f E F(S) is integrable if there is a sequence (fk) of
simple functions converging to f in measure, such that
lim Slfi fjIdv(m) = 0.
1,j JCO
We shall show that the Bartle concept of integration and that
of Dunford and Schwartz coincide when m is a bounded scalar
measure. We state the following result without proof.
5.9 Theorem. (Dunford and Schwartz [15]) If m is a scalar
measure and f is integrable (in the sense of Dunford and
Schwartz), then the total variation of the indefinite inte
gral X(E) = Efdm is given by
v(X,E) = SEIf(s)Iv(m,ds).
Using Theorem 5.9 we now prove the equivalence of the
Bartle and the Dunford and Schwartz integration theories in
the case of a scalar measure.
5.10 Theorem. Suppose m is a bounded scalar measure. Then
f E L(m) if and only if f is integrable in the sense of
Dunford and Schwartz. In this case the same sequence of
simple functions determines f for both definitions; conse
quently the two integrals coincide.
29
Proof. Suppose (fk) is a sequence of simple functions.
By Theorems 5.9 and 2.2 (ii) we have
ISEfidm SEfjdml a ~Efi fjdv(m)
s 4 sup IF(f(i f.)dml
Fq.E
FEZ
for every E E E. Suppose f E L(m) and (fk) is a sequence
determining f. Then by Theorem 5.2 the righthand term
in the inequality goes to zero uniformly for E C E as
i,j C. Hence
(*) lim SIfi f. dv(m) = 0,
ijo J'
and so f is integrable in the sense of Dunford and Schw.artz.
Conversely, if m lim fk = f and (*) holds, then the left
hand term of the inequality converges to zero uniformly for
E E E as i,j . This shows that the sequences ( Efkdm) are
Cauchy uniformly in E, and hence by the completeness of X,
they converge uniformly in E. By Theorem 5.2 we conclude that
the indefinite integrals of the fk's are uniformly mcontinuous,
so f E L(m). In either case, the same sequence of simple
functions determines f for both definitions, so the two inte
grals coincide. [
We now prove some results concerning the domination of
functions by integrable functions.
5.11 Lemma. Suppose (fk) is a sequence of integrable functions,
and g E L(m). If Ifk . IgI on S for each k, then
SEfkdml 4 sup Ilfgdml
FE
FEE
30
for each E E Z; consequently the indefinite integrals of
the fk's are uniformly mcontinuous.
Proof. We use Theorem 5.8 ..ith U = x* E X*, and also
Theorems 5.10, 5.9, and 2.2 (ii) to compute
ISEfkdm = sup SEfkdx*m
X *
Ssup E I fk dv(x*m)
XI *
< sup SElgldv(x*m)
X *
4 sup sup IFgdx*ml
FcE X *
FEZC
= 4 sup fLFgdmI,
FCE
FEZ
for any k E N. Since the indefinite integral of g is m
continuous, it follows that the indefinite integrals of the
fk s are uniformly mcontinuous. 0
5.12 Theorem. If f E M, g E L(m), and Ifi < Igi on S, then
f E L(m).
Proof. Since f E M there is a sequence (fk) of simple functions
converging to f in measure. By Theorem 4.7 we may assume that
Ifkl 21fI on S for every k. Then Ifk < 21gI on S, and
2g E L(m). By Lemma 5.11 we conclude that the indefinite
integrals of the fk's are uniformly mcontinuous. Therefore
f E L(m). r
5.13 Corollary. Suppose f c M. Then f E L(m) if and only if
I fl L(m). If (fk) is a sequence of simple functions deter
31
mining f, then ( fkI) is a sequence of simple functions
determining If .
Proof. The first statement follows from Theorem 5.12. If
fk is a simple function, then so is Ifkl and by Lemma 5.11,
ISEIfkidmI < 4 sup IfFfkdmI.
FEE
FE
Thus the indefinite integrals of the Ifkl's are uniformly
mcontinuous, if the indefinite integrals of the fk's are.
Finally, since I fkI I fl I f fl, it follows that
[I Ifk I fl I > ] c [I fk f > ] If (fk) determines f,
we conclude that (Ifk ) determines If O
6. Convergence Theorems. In this section we prove stronger
forms of the convergence theorems for sequences of integrals
stated by Bartle [1]. Suppose that f and (fk) are integrable
functions. We say that f and (fk) satisfy property B if
lim SEfkdm = SEfdm
uniformly for E E E.
The main convergence theorem resembles the classical
Vitali convergence theorem.
6.1 Theorem. Suppose f E F(S) and (fk) c L(m). Then f E L(m)
and f, (fk) satisfy property B if and only if
(i) m lim fk = f.
(ii) The indefinite integrals of the fk's are uniformly
mcontinuous.
32
Proof. Suppose f E L(m) and f,(fk) satisfy property B.
To prove that m lim fk = f we proceed as follows. By
Theorem 5.13, gk = If fk E L(m) for each k E N. Let
(hkj) be a sequence of simple functions determining gk as
in Definition 5.4. By Theorem 5.13, (ihkjl) is a sequence
of nonnegative simple functions determining Igk = gk, so
without loss of generality we may assume that hkj 2 0 for
each j E N.
By Theorem 5.8 we have gk E L(x*m) for every x* E X*,
hence by Theorem 5.10 gk is integrable in the sense of
Dunford and Schwartz. By Theorems 5.9 and 2.2 (ii) we have
EgkdV' (x*m) = El fk f dv(x*m)
< 4 sup ISF(fk f)dx*ml
FSE
FEZ
S 4 sup I F(fk f)dm ,
FCE
FEE
if x* E X1*. Since f and (fk) satisfy property B, it
follows that given e > 0, there is an n such that if k 2 n
and x* E X1*, then for every E E E,
Egkdv(x*m) < e.
Since (hkj) determines gk, it follows in the same manner
that for a fixed k 2 n, there is an t such that j 2 C implies
that for every E E Z,
SElhkj gkldv(x*m) < ,
provided x* E X *. By omitting the first L1 of the functions
hkj if necessary, we therefore conclude that for x* E X *
E E E, and k 2 n,
SEh.jdv(x*m) < 2 .
Since hkj is a nonnegative simple function the sets
Ekj = [hkj > y] belong to Z for every y > 0. Moreover,
vv(x*imij) SShkjdv(x*m) < 2e,
so
v(x*m,E.j) < 2e/y,
if x* E X,*. By Theorem 2.2 (iii) we conclude that
m(Ekj) sup v(x*m,Ekj) < 2e/y.
1
Since mlim hkj = gk, there is an i E N and a set Fk E C
such that fi(Fk) < e/y and Ihki(s) gk(s) 5 y if s Fk.
Therefore, if s % Eki U Fk, we have
Ifk(s) f(s) = gk(s)
Igk(s) hki(s) + Ihki(s)l
S2y ,
and m(Eki U Fk) < 3e/y. Now if 61 and 62 > 0 are given,
then choose y > 0 such that 2y < 61. Then choose e > 0 so
that 3e/y < 62. It follows from the arguments above that
there is an n E N and sets Gk such that if k 2 n, then
fi(Gk) < 62 and Ifk(s) f(s) 61 if s Gk. We conclude
that mlim f = f.
To see that (ii) holds we note that the argument in the
proof of (ii) implies (i) for Theorem 5.2 applies.
Conversely, if (i) and (ii) hold, then for each k there
is a simple function gk such that T[Igk fkl > 1/k] < 1/k
and for every E E E,
IEgk fk)dml < 1/k.
34
As in the proof of Theorem 4.6, it follows that m lim gk= f.
Suppose e > 0. Choose n such that 1/n < e. Then choose
6 > 0 such that i(E) < 6 implies
IJEgkdmI < e and SEfidmi < E
for 1 s k 5 n and i E N. Then if m(E) < 6,
Egkdm I I Egk fkdm + ISE kdm
s 2e
for every k, so the indefinite integrals of the gk's are
uniformly mcontinuous. It follows that f E L(m). Since
ISEfkdm JEfdml I'SE(fk gk)dm + IJ(gk f)dml
S1/k + IjE(gk f)dml,
we conclude that f and (fk) satisfy property B. 0
We now state two important corollaries to Theorem 6.1
which are stronger forms of the results stated by Bartle [1].
6.2 Corollary. (Dominated Convergence) Suppose (fk) c L(m)
and g E L(m). If Ifkl < Ig\ on S for every k E N, then
mlim fk = f if and only if f E L(m) and f, (fk) satisfy
property B.
Proof. Suppose mlim fk = f. By Lemma 5.11 the indefinite
integrals of the fk's are uniformly mcontinuous, so the
conclusion follows from Theorem 6.1. The converse is immediate
by this same theorem. ]
6.3 Corollary. (Bounded Convergence) Suppose (fk) c M is a
sequence of functions uniformly bounded on S. Then m lim fk= f
if and only if f E L(m) and f,(fk) satisfy property B.
35
Proof. Since constant functions are integrable, the con
clusion follows from Corollary 6.2. n
7. Lp Spaces. As an application cf the previous theory, we
present a portion of L space theory. In this section Z is
an algebra and m:E X is a bounded measure.
Recall the following set of inequalities, which were
used in the proof of Lemma 5.11. If f E L(m) and E E E,
then
 EfdmI = sup jEfdx*mI
X *
< sup SE fl dv (x*m)
X *
1
s 4 sup sup I~Ffdx*ml
FCE X *
= 4 sup IJFfdmi.
FcE
These terms are finite by Corollary 5.7. Let f E M. If
1 ~ p < = and IfP E L(m), define
lfllp = sup [Jf IflPdv(x*m)]l /p
1
Let L denote the space of all functions f E M such that
p
IflP E L(m). Since F c E implies that
ISFfdmI 5 sup El fldv(x*m),
X *
the inequalities above show that the two expressions
sup [JEI flPdm]1/p l/p
sEup E1 fI Pdm] 1/p and sup [FS IffPdv(x*m)] /p
EEEuI
X *
1!
36
0
are equivalent "norms" for the spaces L We prefer to use
p
the second of these expressions, since the properties of the
scalar norms in L (x*m) are then available.
By Corollary 5.13, it follows that L(m) = LI. Moreover,
since 1 p q < m implies that IflP 1 + Ifl we see
that f E L implies that f E L, by Theorem 5.12. Hence
q p
O LO = LO
L D Lp q
1 p q
From the elementary inequality la+blI 2P(IalI+ IblP),
it follows that If+glP < 2P(IfIp + Ig1P); if f and g are
in L, then so is f + g, by Theorem 5.12. L is clearly
p' p
closed under scalar multiplication, and hence is a linear
space. It is easy to see that llafl = lallfip for any
scalar a. Moreover,
f +gip = sup[ I f +gPdv(x*m) ]l/p
P X*
< sup[(IfI flPdv(x*m))1/p + (Slg 9v(x*m))l1p]
X *
1 lfllp + llgllp
Since lflp = 0 if f = 0, we conclude that I1l lp is a semi
norm on L.
P
Suppose 1 < p < and 1/p + 1/q = 1. The HBlder
inequality shows that
IfgI Iflip Igl
S +
Ilf jj4 11gj PllfIIp q jgjjq
fllpllgllq p qlgll
if f E LO and g E L By Theorem 5.12 we conclude that
p q
fg 4 Lo If x* E X1 we have
fg 1L
37
fS fgjdv(x*m) < l\fllpIg. q(1/p + 1/q).
Hence Ilfgll < I flpllgllq
A sequence of functions (fk) in LO is said to converge
k p
in L to a function f E L if we have limlifk flp = 0. The
p p k
following convergence theorem is a consequence of Theorem 6.1.
7.1 Theorem. Suppose 1 s p < m and (fk) is a sequence in L .
Then f E LO and limllfk flp = 0 if and only if
(i) m lim fk = f.
(ii) The indefinite integrals of the Ifk P's are uni
formly mcontinuous.
Proof. Suppose f E L0 and limlifk fp = 0. If we set
p p
gk = Ik flp E L(m), then the first part of the proof of
Theorem 6.1 is easily adapted to show that m lim fk f
Since the indefinite integral of IflI is mcontinuous, and
llfkIElp < llfk fllp + lfEllp, (ii) follows as in the proof
of Lemma 5.2.
Conversely, suppose that (i) and (ii) hold. By Theorem
4.6, IfkiP converges to IflP in measure, so by Theorem 6.1,
f E Lp. Moreover, Theorem 4.6 implies that mlimlfk flP=0.
p
Since Ifk flp 2P(IfkIp + IflP), we have
EJfk flPdml 5 2p+2 sup iF(IfklP+ IflP)dml,
FcE
FE Z
by Lemma 5.11. Since the indefinite integrals of the ifklP's
are uniformly mcontinuous, we conclude that the indefinite
integrals of the Ifk flP's are also uniformly mcontinuous.
By Theorem 6.1, then,
lim JEIfk flpdm = 0
uniformly for E E E. Finally, since
lfk flp < 4 sup fEI fk fl dml,
P EE
we have limllfk fil 0. O
The inequalities at the beginning of this section show
that convergence in L1 is equivalent to property B. For if
f,(fk) belong to Lo = L(m), then
sup IE(fk f)dml < lIfk f1l1 4 supE (fk f)dm .
EE E EE E
Just as Corollary 6.2 followed from Theorem 6.1, we have:
7.2 Corollary. Suppose 1 < p < , (fk) Lp, g E Lp, and
Ifkl I Ig on S for each k. Then mlim fk = f if and only
if f E LO and limllfk f = 0.
p kp
Proof. Since IfklP < Igjl the indefinite integrals of the
Ifk P's are uniformly mcontinuous. The conclusions follow
from Theorem 7.1. ]
7.3 Corollary. Suppose 1 s p < c. The subspace of simple
functions is dense in L.
p
Proof. Suppose f E L Since f E M there is a sequence (fk)
of simple functions converging to f in measure, and by Theorem
4.7 we may assume that Ifkl  21fl on S for each k. By
Corollary 7.2 it follows that limllfk fil = 0. O
39
8. Integration by Countablv Additive Measures. Integration
of scalar functions by a countably additive measure m:Z X,
where S is a aalgebra, has been studied extensively by Bartle,
Dunford and Schwartz [2], and also by Lewis [20]. The Bartle,
Dunford and Schwartz Theory uses a result concerning the
equivalence of weak compactness in the space of scalar
measures with the existence of a control measure (as in
Theorem 3.3), and the VitaliHahnSaks Theorem. By means
of these powerful results the following definition of the
integral yields a theory which coincides with that in Section
5 for countably additive measures.
A sequence of functions (fk) c F(S) is said to converge
(pointwise) almost everywhere to a function f if there is a
null set A such that lim fk(s) = f(s) for every s A. In
this case we write ae lim fk = f.
8.1 Definition. f E F(S) is integrable if there is a sequence
(fk) of simple functions such that ae lim fk = f, and such
that for every E E E, the sequence (fEfkdm) converges in X.
Pointwise almost everywhere convergence replaces con
vergence in measure in the countably additive situation by
virtue of the nonnegative control measure m. Since m and m
are mutually continuous and countably additive, the standard
theorems on the relation between convergence in measure and
almost everywhere convergence can be shown to hold. In
particular, ae lim f = f implies m lim fk = f. Moreover,
K k
since the indefinite integrals of the function fk in Defini
tion 8.1 are mcontinuous meo'ures, the *quirement that
40
( Efkdm) converges for each E E E, together with the Vitali
HahnSaks Theorem, show that the indefinite integrals of the
fk 's are uniformly mcontinuous, as in Definition 5.4.
8.2 Remark. If f is finitely additive, E is a calgebra, and
m is sbounded, so that there is a control measure m by
Theorem 3.3, then we may apply Theorem 3.5. Thus if
(fEfkdm) converges for every E, we conclude that the in
definite integrals of the fk's are uniformly mcontinuous.
Convergence in measure must be retained however.
Bartle, Dunford and Schwartz establish a onesided
form of Theorem 6.1: If (fk) = L(m), ae lim fk = f, and
the indefinite integrals of the fk's are uniformly mcon
tinuous, then f E L(m) and lim JEfkdm = 1Efdm for E E 7.
In view of Theorem 6.1 and the VitaliHahnSaks Theorem we
have the following extension:
8.3 Theorem. Suppose (fk) c L(m) and ae lim fk = f. Then
f E L(m) and f,(fk) satisfy property B if and only if
lim rEfkdm exists for each E E E.
In the case of convergence in L we have by Theorem 7.1:
p
8.4 Theorem. Suppose 1 s p < , (fk) C Lp, and ae lim fk f
Then f E L0 and limllfk fl = 0 if and only if lim iE fkiPdm
exists for every E E E.
For future use we state the following useful approxi
mation theorem for the countably additive case.
41
8.5 Theorem. (Dunford and Schwart.z [15]) Suppose 1 < p < m,
o is an algebra, and E = o(E ). Then the space of 
simple functions is dense in L .
p
Lewis [20] has developed an integration theory for count
ably additive measures with values in a locally convex space
X using the Pettis approach of employing linear functionals.
This integration theory coincides with that presented above
when X is a Banach space.
9. A Riemanntype Integral. Gould [16] introduced an
integral for scalar functions, which in the case of bounded
functions is defined as a limit of RiemannStieltjestype
sums. Although he calls his integral an improper integral,
it is in fact equivalent to the "proper" Bartle integral
when m is a bounded measure. This alternative approach to
integration has been extended by McShane [22] to obtain a
very general theory. We identify the Gould and the Bartle
integrals as the (G) and the (B) integrals respectively.
In this section E is an algebra and m:E X is a bounded
measure. If P and P' are two partitions in n(S), we write
P' > P to mean that each set in P' is a subset of a set in P;
we say that P' is a refinement of P. If P E n(S) and
P = (E1,...,En), then for a function f E F(S) we introduce
the symbol S(f,P) to indicate any RiemannStieltjestype
sum
n
E f(si)m(Ei),
i=l
where s. E E., 1 s i < n. Let B(S,E) = B(S) denote the family
of all bounded measurable functions on S.
42
9.1 Definition. Suppose f E B(S).. f is integrable (G) if
for any e > 0 there is a P E n(S) such that if P' E n(S)
and P' 2 P, then IS(f,P) S(f,P')j < e.
We omit the proof of the following Theorem.
9.2 Theorem. Suppose f E B(S). f is integrable (G) if and
only if there is an x E X with the following property: For
every c > 0 there is a P E n(S) such that if P' E n(S) and
P' 2 P, then IS(f,P') xj < c.
The vector x is called the integral of f over S. We
now show that for bounded measurable functions the (G). and
(B) integrals coincide.
9.3 Theorem. If f E B(S), then f is integrable (G) and the
(G) and (B) integrals coincide.
Proof. By Theorem 5.6, f E L(m); let x = jSfdm, and let
(fk) be a sequence of simple functions converging to f in
measure. By Theorem 4.7 we may assume that ifkl 21fl on
S for each k. Suppose that If[ s K on S, and fix s > 0.
Choose k E N and A E Z such that (fA) < e/4K and
If(s) fk(s)I < c/4f(S) if s A. Suppose fk = ai I.
Let F = A and F. = E._\A, 2 < i + +1. Then P = (Fi) is
a partition of S and fk is constant over F2,...,Ft+1. If
P' = (B ) > P and if fk(s) = bi for s E Bi, then
n n
IS(fP') SfkdmI = I f(si)m(Bi) JB fkdm
i=1 i=l i
S E (f(s.) b.)m(B.) I+I (f(si)b.)m(B.i)
B. A=,O BicA
4 4[e/4ff(S)]m(S\A) + 4Kff(A) + 8Ki(A)
4c,
43
w.,here w',e have used Lemma 2.1.
Since Ifk < 21fj, Corollary 6.3 shows that f and (fk)
satisfy property B. There is then a k E N such that
S f3djm xI < e. The argument above shows that there is
a partition P E n(S) such that if P' > P then IS(f,P') xI 5c.
We conclude that f is integrable (G) and that the (G) and (B)
integrals coincide. 0.
The vector x will now be denoted by ,Sfdm. If E E E
and f Q B(S), then flE is bounded and measurable by Theorem 4.6,
and we write
Efdm = YSf Edm
Suppose f E M(m). The definition of measurability implies
that f is bounded except perhaps on sets of arbitrarily small
variation. Let B(f) denote the family of all sets in Z on
which m is bounded. It is easy to see that B(f) is a ring of
sets. Since f is integrable on each set in B(f), the class
I(f) = (SEfdm:E E B(f)) is nonvoid. Moreover, I(f) is a net
in X over the directed set B(f), where E s F means E C F.
With this motivation we can define the integral for unbounded
measurable functions.
9.4 Definition. f E M is integrable (G) if for every c > 0
there is an E E B(f) such that if F E B(f) and E n F = 0,
then I Ffdmi < e.
In view of the fact that I(f) is a net we have the
following result due to Gould [16].
44
9.5 Theorem. (Gould) f E M is integrable (G) if and only
if there exists an x E X having the following property:
For every e > 0 there is an E E B(f) such that if F E B(f)
and E c F, then ISFfdm xi < c.
Proof. If such an x exists, then since B(f) is a ring and
the indefinite integral of f is a measure on B(f), we have
I Ffdml ISF+Efdm xj + IEfdm xl,
for every E,F e B(f) such that E n F = 0. Fix > 0 and
choose E E B(f) such that if G E B(f) and E C G, then
IJGfdm xc < ,/2. If F E B(f) and E l F = O, then
E c E+F E B(f), and so
IJFfdml < e.
Conversely, suppose f is integrable (G). Definition 9.4
implies that I(f) is a Cauchy net in X. For suppose e > 0
is given. Choose E E B(f) such that if F E B(f) and E F = 9,
then IfFfdmI < e. Now if F,G E B(f) and F,G B E, we have
IJFfdm fGfdml < IF\Efdml + ISG\Efdm
s 2e,
where we have again used the additivity of the integral. Let
x denote the limit of I(f). It is then immediate that x
satisfies the requirement of the theorem. 0
From this result we see that when f is integrable (G),
then
fSfdm = lim fEfdm
B(f)
45
Suppose now that f E L(m). To see that f is integrable
(G), suppose e > 0, and choose 6 > 0 such that if fi(A) < 6,
then  Afdmj < e. Since f is measurable, there is a set
E E E such that f is bounded on E and m(S\E) < 6. It
follows that for any A E Z such that A n E = d, M(A) < 6,
and so IfAfdmI < e. By Definition 9.4, f is integrable (G).
To see that integrability (G) implies integrability (B)
we need the following result established by Gould [16].
9.6 Theorem. (Gould) If f is integrable (G), then the
indefinite integral of f is a bounded mcontinuous measure
on E.
Proof. Since I(f) is a Cauchy net it is a bounded set, so
the indefinite integral is bounded. Additivity follows from
the additivity of the indefinite integral for bounded func
tions. To see that the indefinite integral is mcontinuous,
suppose c > 0. There is an E E B(f) such that if F E B(f)
and E n F = 0, then ISFfdmI < c. Hence for A E 2, A n E = 0
implies that IfAfdmI  e. On E, however, the indefinite
integral is already mcontinuous by Theorem 5.6, so there is
a 6 > 0 such that if M(A) < 6 then IF AEfdml < c. Thus if
m(A) < 6, then
AfdmI I fdm + IfAEfdmI 2e. O
We now prove a result that is the counterpart of
Lemma 5.11.
46
9.7 Lemma. If f,g E M, Ifl < Igl on S, and g is integrable
(G), then for every A E E,
(*) LAfdmIl 4 sup Ifgdm
B=A
BEE
Thus f is integrable (G).
Proof. Suppose E E B(f) and e > 0. The indefinite integral
of f over E is an mcontinuous measure by Theorem 5.6, so
there is a 6 > 0 such that if ff(A) < 6, then 'A~EfdmI < e.
Since g is bounded except on sets of arbitrarily small
variation, there is a set F E B(g) such that mf(S\F) < 6.
Since f and g are both bounded on E n F, we have
IJEnFfdml 4 sup Bgadmi
4 sup IBgdml
BcE
by Lemma 5.11. Therefore,
IJEfdm ISE\Ffdm + 'IEnF fdm
C e + 4 sup I[Bgdm
BGE
Since c was arbitrary, we see that (*) holds for every A E B(f),
and hence for every A E E.
To see that (*) implies f is integrable (G), suppose e > 0.
Choose E E B(g) such that if F E B(g) and E n F = 0, then
IfFgdml < e. For any B E E with B n E = 0, we have
IfBgdmiI e. Since B(g) c B(f), we conclude by (*) that if
A E B(f) and A n E = ,, then
IfAfdm s 4 sup I Fgdml 4e. r
B A
BEZ
47
Suppose now that f is integrable (G). Since f E M,
there is a sequence (fk) of simple functions converging
to f in measure, and by Theorem 4.7 we may assume that
Ifkl < 21fl on S. Since 2f is integrable (G), Lemma 9.7
shows that the indefinite integrals of the fk's are uni
formly mcontinuous, since by Theorem 9.6 the integral of
f is mcontinuous. By Definition 5.4, f E L(m).
To see that the (G) and (B) integrals are equal, note
that the definition of the (G) integral as the limit of I(f)
shows that there is an increasing sequence of sets Ek E B(f)
satisfying
(G) Sffdm = lim fSf Ekdm.
Since the functions fTEk are measurable by Theorem 4.6, and
since we can choose the Ek's so that F(S\Ek) < 1/k, it
follows that mlim fk = f. Since f E L(m) and
If EkI 5 If on S for each k, Corollary 6.2 implies that
(B) Sffdm = lim JSfEkdm.
The integrals therefore coincide.
CHAPTER II
Stieltjes Measures and Integrals
In this chapter we specialize the integration theory
of Chapter I to the case where the measure is generated by
a vectorvalued function on an interval of the real line.
We shall call such measures Stieltjes measures, and the
resulting integrals will be called Stieltjes integrals.
By restricting the Banach space in which the vector function
takes its values to be an L space over a probability
measure space, we shall study stochastic integration.
(Chapter III). In the present chapter we shall investi
gate the question of boundedness of a Stieltjes measure,
and the question of the existence of an extension to a
countably additive measure on the Borel sets. Finally we
discuss the important special case of Stieltjes measures
with values in a Hilbert space.
1. Notation. Let T denote a closed, bounded interval in R,
which for convenience we shall assume is the unit interval
[0,1]. Let S denote the family of all subintervals of T
of the form [a,b) or [a,l], together with 0, where we shall
always assume that 0 s a < b < 1. Finally, let E denote
the family of all subsets of T which are of the form A= EAi,
where (Ai) is a finite disjoint family in 9. Note that Z
49
is an algebra of subsets of T, and that the aalgebra
generated by 7 is the family of Borel sets in T, which
is denoted by E1.
Suppose that X is a Banach space, and that z:T X
is a function. Define a set function m:S X by setting
m[a,b) = z(b) z(a) and m[a,l] = z(l) z(a). If A E "
and A = 7Ai, where each A. E 8, then it follows at once
by the definition of m on S that m(A) = Em(Ai). Suppose
n k
now that A E E, and that E Ai and E B. are any two re
i=l j=l 1
presentations of A by finite sums of sets in o. Then the
sets C.. = A. n B. belong to 8 for each i and j, and we
have
A = E C.., 1 i g n, and B. = T Ci., 1 r j k.
j i
By the preceding remark, it follows that
m(Ai) = E m(C ), and m(B) = E m(Cij),
j J i j
for each i and j. We therefore conclude that
( m(A) = m(C ) = E m(B.).
i i j j
We have just shown that if A E E and A = EAi is any repre
sentation of A by a finite sum of sets in 8, then we can
define m(A) = E m(A.), and this definition in independent
of the representation of A. It is now immediate that m
is finitely additive on E. In the sequel, we shall associate
with each function z:T X the unique measure m defined as
above on E. When we wish to emphasize that the measure is
generated by z, we write m We call such a measure m the
Stieltjes measure generated by z.
50
If m is a bounded Stieltjes measure on E, then m deter
mines a class L(m) of functions integrable in the sense of
Chapter I. The Zsimple functions will now be called step
functions; the class of step functions coincides with the
class of all Ssimple functions. Every function that is the
uniform limit of a sequence of step functions is obviously
bounded and measurable with respect to any Stieltjes measure
on Z. Hence by Theorem 1.5.6, such functions are integrable,
and the sequence of integrals of the step functions converges
to the integral of the limit function. If m has a countably
additive extension to E1, then Theorem 1.8.5 shows that the
class of step functions is still dense in L(m).
2. Bounded Stieltjes Measures. In this section z:T X is
a function generating the Stieltjes measure m:Z X. We
investigate conditions on z which insure that m is a bounded
measure. By Remark 1.2.3, m is bounded if and only if x*m
is bounded for each x* E X*. Now since
x*mz[a,b) = x*z(b) x*z(a) = mx*z[a,b), for every interval
[a,b) (or [a,l]) in 8, it follows that x*m = m for every
x* E X*. By Theorem 1.2.2 (ii) we know that the scalar
measure m*z is bounded if and only if it has bounded total
variation. Recalling that each set in S is a finite sum of
sets in 8, it follows that
n n
sup E Im(A.) = sup S Im(A) I,
(Ai)cS i=l (A i)C i=l
where each (A.) is a finite disjoint family. That is, the
total variation of any measure m on E is the same as the
51
total variation of m restricted to g. Therefore the measure
m*z is bounded if and only if
n n
sup Z Im z (A )I = sup jx*z(b.) x*z(ai.) <,
(Ai) i=l (Ai)s1 i=l
where we write each A. as [ai,b.) (or [ai,b.] if b. = 1).
But the righthand term is just the total variation of the
scalar function x*z. Hence mx*z is bounded if and only if
x*z is a function of bounded variation for each x* E X*.
This establishes the following theorem.
2.1 Theorem. m is a bounded measure on E if and only if the
function z is of weak bounded variation, in the sense that
x*z is a function of bounded variation for x* E X*.
2.2 Remark. Since
n n
sup Im(A) = sup I m(A.) = sup z(b) z(a.) ,
AES (Ai)cg i=l i=l 1 1
where the supremum on the right is taken over all finite
sets of points a. and b., with 0 < a1 b b .. an b n 1,
we have as a corollary to Theorem 2.1 that z is of weak
n
bounded variation if and only if supl S z(b.) z(ai)I < .
i=1
This result can also be proved directly using Lemma 1.2.1
and the principle of uniform boundedness.
Let C(T) denote the family of all scalarvalued contin
uous functions on T. If f E C(T), then since f is bounded
and uniformly continuous, it follows from Theorems 1.4.5
and 1.5.6 that f is integrable with respect to any bounded
measure on Z. In fact, if m is a bounded measure and f E C(T),
then in defining the integral according to Definition 1.9.1
52
we may restrict the partitions P to be from n (T,8). To
see this we introduce the concept of the norm of a parti
tion P E Tr(T,5). If P = (A.:1 < i n), where each A. is
1 1
an interval [ai,b.) (and A = [a ,l]), then the norm of P,
denoted by jPI, is defined as follows: jPI = max (b. a.).
l~i~n 1 1
We now state the above mentioned result.
2.3 Theorem. If m is a bounded measure and f E C(T), then
for every C > 0, there is a 6 > 0 such that if P and P'
belong to n(T,8) and IP', IPI < 6, then IS(f,P') S(f,P) <~.
In particular, this holds if IPI < 6 and P' 2 P.
Proof. By the continuity of f, there is a 6 > 0 such that if
js t < 6 and s,t E T, then If(s) f(t)j < c/4fi(T). Now
if P,P' E n(T,S) and if PI, IP' < 6/2, then
n k
IS(f,P') S(f,P) = I f(s )m(Ai) S f(t.)m(Bj)I
i=l j=l
n k k n
=  Z f(s.) m(A. n B.) E f(t.) E m(A. nBj)
i=l j=l j=l i=l
n k
= I E (f(si) f(t.))m(A. n Bj) .
i=1 j=l
The only nonzero terms in this double sum occur when
m(Ai n Bj) ) 0; in this case A. n B. 0, and hence
Is. tj. < 6. By Lemma 1.2.1 we conclude that
S(f,P') S(f,P) s 4[e/4f (T)]m(T) = e.
The final statement of the theorem is immediate. O
53
Motivated by this result and the classical definition
of the RiemannStieltjes integral, we make the following
definition. Recall that B(S) is the set of all bounded
measurable functions on S.
2.4 Definition. Suppose m:T X is a measure and f E B(S).
Then f E RS(m) if for every e > 0 there is a 6 > 0 such that
whenever P,P' E n(T,") and Pj, P'j < 6, then
S(f, P) S(f,P') < C.
Using this definition we could prove as in Theorem 1.9.2
that if f 6 RS(m), then there is a vector x E X to which the
sums S(f,P) converge as Pl converges to zero. When m is
bounded, Theorem 2.2 shows that C(T) a RS(m), and Theorem
1.9.3 and Definition 1.9.1 show that RS(m) c L(m) and that
the integrals coincide. In fact, if m is any measure from
E to X, and if C(T) c RS(m), then m is bounded. This follows
from Remark 1.2.3 and the following result.
2.5 Theorem. (Dunford [14]) If m is a scalar measure and
C(T) C RS(m), then m is bounded.
Proof. First we remark that if (P ) is any sequence of
partitions such that P  converges to zero, then for every
f E C(T), the sums S(f,Pn) converge to the integral of f
with respect to m. Suppose that m is not bounded. Since m
is a scalar measure we have only to suppose that the total
variation of m is infinite. It follows that there is a
54
sequence (P ) in n(T,g) such that PnI < 1.'n and
E Im(Ai)I > n
P
n
for each n, where P = (Ai); we may also assume that m(A) f0.
If for each n we choose a set of points si E A., then by the
preceding remark, lim S(f,P ) = STfdm for every f C(T),
where the sums S(f,P ) are formed using the points (si) chosen
for each P
n
Consider the mapping Un:C(T) X defined by U (f) =S(f,Pn),
n E N. Each U is linear and bounded, and lim U (f) exists
n n
for each f E C(T). Since C(T) with the norm llfl = suPT f(s)
is a Banach space, it follows by the principle of uniform
boundedness that there is a constant K > 0 such that
(*) IS(f,Pn) I i Kllfll,
for every f E C(T). To obtain a contradiction we shall con
struct a function f E C(T) such that liflj < 1, but for some
k
n E N, IS(f,P )I > K holds. Choose n such that Im(Ai)( >K,
n j i=l
where P = (A.). Let s., 1 5 i < k, be the sequence chosen
for Pn above. Define f(si) = m(Ai)/Im(Ai) 1 i < k;
f(t) = f(sl) for t in the interval [0,sl); f(t) = f(sk) for
t in the interval [sk,l]; and on each interval [si'si+)
define f to be linear between f(si) and f(si+1). Then
f E C(T), and we have l\fl = 1; however,
k
S(f,Pn) = Z f(s.)m(A.)
i=l
k
= Im(Ai) I > K.
i=l
Since this contradicts (*), m is bounded. ]
55
The following theorem summarizes the results concerning
the boundedness of Stieltjes measures.
2.6 Theorem. Let m:Z X be generated by z. The following
statements are equivalent:
(i) z is of weak bounded variation.
(ii) m is bounded.
(iii) C(T) c RS (m).
(iv) C(T) RS(x*m) for every x* E X*.
Proof.(i) = (ii) follows from Theorem 2.1. Theorem 2.3
shows that (ii) = (iii). We deduce from Theorems 1.5.8 and
2.3 that (iii) a (iv). Suppose that (iv) holds. Theorem 2.5
implies that x*m is bounded; hence, by Remark 1.2.3, m is
bounded. r]
3. Extensions of Countably Additive Measures. If z is a
scalar function on T, then the classical extension theorem
(see Halmos [17]) for the measure m states that m has a
z z
unique countably additive extension to El, the Borel sets
of T, if and only if z is leftcontinuous and of bounded
(total) variation. When z is a vector function, we shall
show that a similar theorem holds, provided that the Banach
space X does not contain a copy of c We need to introduce
several preliminary concepts and results.
A semiring S is a family of subsets of S such that (i)
A,B E c implies that A B E 8; (ii) A,B E S and A B implies
that there is a finite family (C.) in g such that
3
56
A =C1 c C ... C = B and C..C. 9 for 2 < is n
2 n 1 11
(see Halmos [17] or Dinculeanu [12] for information on
semirings and measures on semirings). If S is a semi
ring, then the smallest ring R containing ", which h we call
the ring generated by 8, is the family of all sums of finite
disjoint subsets of 8. An example of a semiring is just
the class 8 of halfopen intervals discussed in Section 1.
In this case the ring generated by 8 is the algebra Z.
3.1 Remark. Suppose that 8 is a semiring and that m:8 X
is a set function. It is known (see Dinculeanu [12]) that
if m is finitely additive on 8, then m has a unique finitely
additive extension to R, the ring generated by 8. Moreover,
if m is countably additive on 8, then its extension to R
is countably additive on R as well. From now on, given a
finitely additive set function on 8, we shall assume that m
has been extended to R, and shall denote the extension by m.
When we refer to the variation if we always mean the variation
of the extension, defined as in Chapter I, Section 2, by
rf(A) = sup Im(B) .
BcA
BER
Note that if R is the ring generated by the semiring 8,
then we also have
n
ff(A) = sup E m(Bi)I,
(B i)s i=l
where (Bi) denotes a finite disjoint family in 8.
57
If S is a topological space and A s S, then we denote
the interior of A by int A and the closure of A by cl A.
If S is a semiring in S and m:8 X is a finitely additive
set function, we make the following definition, with
Remark 3.1 in mind.
3.2 Definition. Suppose that S is a compact space. Then
m is regular if for every E E 8 and c > O, there are sets
A,B E g such that cl A c E c int B and m(B\A) < c.
We now prove a result due to Huneycutt [18] which
generalizes the classical Alexandroff Theorem for regular
scalar measures (see Dunford and Schwartz [15]).
3.3 Theorem. (Huneycutt) If S is a compact space, S is a
semiring in S, and m:S X is finitely additive and regular,
then m is countably additive on 8.
Proof. Let (Ei) be a disjoint sequence in S such that
E = ZEi E 8, and fix e > 0. By regularity there are sets
A,B E S with cl A Q E int B and f(B\A) < e. There are
also sequences of sets (A.) and (Bi) in S such that for
each i, cl Ai E. c int B. and f(B.\A.) < e/2'. It follows
that for each n,
n
m[( U Bi)\A] < 2e.
i=l
For
n n
( U B.)\A s [( U B.)\E] U [E\A]
i=l i=1
n
c [ U (Bi\Ei)] U [E\A],
i=l
n n
m[( U Bi)\A] f m[ U (Bi\Ei)] + T[E\A]
i=l i=l 1 1
n
Z E/2 + c < 2e.
i=l
Since cl A is compact and cl A c
k
is a k E N
such that A c U B..
i=l I
n n
Im(A) E m(Ei) < m( U
i=l i=l
n
+ Im( U
i=l
Finally, Im(E)
U Ei = U
i=l i=l
int B., there
1
For n 2 k, therefore,
n
B.) m( U E.) +
S mi=l
Bi) m(A)I
1
n n n
= m[ U Bi\( U E.)]l + Im[( U Bi)\A]
i=l i=l i=l
n n
< M[ U (B.\Ei)] + f[( U B.)\A]
i= 1 i=1
n
S e/2 + 2e < 3e.
i=l
n
 E m(E) j i m(E) m(A)1 +
i=l
n
+ Im(A) S m(Ei)
i=l
S f(E\A) + 3e 4,
w3
if n k. We conclude that m(E) = E m(Ei); hence m is
i=l
countably additive on 8. .
As an immediate corollary of this result we have:
3.4 Corollary.
(Huneycutt [19]) If m:Z X is a finitely
additive measure such that lim m[a h,a) = 0 for every
h0
a E (0,1], then m is countably additive on E.
59
Proof. It suffices to show that m is countably additive on
g, and in view of Theorem 3.3 we need only show that m is
regular. Suppose that [a,b) E 9. We may suppose that
0 < a < b < 1, since the cases where a = 0 or b = 1 are
proved in a similar fashion. Choose h > 0 such that
0 < ah < a < bh < b. Then we have [a,bh) E [a,bh] c
[a,b) r (ah,b) c [ah,b), and [ [ah,b)\,[a,bh)] !
in[ah,a) + m[bh,b). By hypothesis, given e > 0, we can
choose h > 0 such that n~[ah,a) + m[bh,b) < e. Therefore
m is regular. r
We now discuss an important class of Banach spaces
characterized by Bessaga and Pelczynski [3]. The signifi
cance of this class for our purposes lies in the fact that
it is precisely the class of range spaces for vector measures
in which boundedness is equivalent to sboundedness.
Let co denote the Banach space of all scalar sequences
x = (xn) such that lim xn = with norm defined by
jlx = sunlx nl. If X is a Banach space and (xn) is a
sequence in X, we say_that (x ) is weakly Cauchy if the
scalar sequences (x*x ) are Cauchy sequences for each
x* E X*. We say that (xn) converges weakly to x E X if
lim x*xn = x*x for each x*. Recall that a Banach space is
said to be weakly sequentially complete if every weak
Cauchy sequence converges weakly to some element in X. A
series Exn in X is said to be weakly unordered bounded if
for each x* E X* we have
sup E x*xil < .
A iEA
60
Note that by Lemma 1.2.1, Fxn is weakly unordered bounded
if and only if TIx*xil < c for each x* E X*.
If a Banach space X contains no subspace isometrically
isomorphic to co, then we write X co; otherwise '..e write
X 2 co. We state the following important result without
proof:
3.5 Theorem. (Bessaga and Pelczynski [3]) The following
statements are equivalent:
(i) X co.
(ii) Every weakly unordered bounded series in X con
verges unconditionally.
The following result was originally stated by Brooks
and Walker [7] when X is weakly sequentially complete, and
slightly extended to the present situation by Diestel [11].
3.6 Corollary. If X A co and m:R X is any measure, then
m is bounded if and only if m is sbounded.
Proof. By Theorem 1.3.2 an sbounded measure is always
bounded. Conversely, if m is bounded and (Ei) is a disjoint
sequence in R, then by Lemma 1.2.1 we have
n
Ix*m(E )I < 41x*I sup I m(Ei)I
i=l ANn iEA
< 4 x*[ sup Im(E) < w,
EER
for every n E N and x* E X*. It follows that the series
61
Zm(Ei) is weakly unordered bounded. By Theorem 3.5, there
fore, Em(E.) converges unconditionally. By Theorem 1.3.2
we conclude that m is sbounded. ri
3.7 Example. To see that Theorem 3.5 fails to hold when
X C co, consider the sequence of unit vectors en = (6 i) in
c where 6 ni is the Kronecker delta. The series Ze is
weakly unordered bounded, since for any 6 c N, I ei = 1,
iEA
but this series does not converge in c .
We now prove the final result that will be needed for
the main theorem of this section.
3.8 Lemma. Suppose m:Z X is a bounded measure such that
lim m[ah,a) = 0 for every a E (0,1]. If X c cO, then for
h0
every a E (0,1] we have
lim m[ah,a) = 0.
h0
Proof. Since i is monotone the limit above exists for each
a E (0,1]. Suppose there is an a E (0,1] and a 6 > 0 such
that lim fm[ah,a) > 6. Let n = 1. In view of Remark 3.1
h0
there is an n2 > nl and a finite disjoint family of sets
2 1 n2
A2....,A in 8, such that I m(Ai)I > 6. We may assume
2 i=n+l1
for convenience that A. = [a ,b.), where a
I i. 1 n n2
If b = a, then choose b such that a < b < b and
"2 "2 "2
Im[b,a)l < 6/2. This is possible by the continuity hypothesis
on m. Then
n21 n2
I m(A.) +m[a ,b) I I Z m(A ) m[b,a) > 6/2.
i=n+1 2i=nl+1
62
Redefining b = b, so that A = [a ,b), we have
2 2 2
n2
"2
+ m(A.) > 6/2.
i=nl+1
Now since M[b ,a) > 6, we can repeat this argument over
the interval [b ,a) to obtain an n3 > n and a finite
n 3 2
disjoint family of sets An ,...,An in S, such that
n3
"3
 E m(Ai) I > 6/2.
i=n2+1
By induction, therefore, we obtain an increasing sequence
(nk) and a disjoint sequence (Ai) in 8 such that for k E N,
nk+l
E m(Ai)I > 6/2.
i=nk+l
nk+1
Define Ek = E Ai; then (Ek) is a disjoint sequence in E,
i=nk+l
and for every k E N,
(*) Jm(Ek)J > 6/2.
Since X 6 co and m is bounded, m is sbounded by Corollary
3.6. This contradicts (*), so we conclude
lim +[ah,a) = 0. O
h40
The following theorem is our main result on the extension
of Stieltjes measures to countably additive Borel measures.
3.9 Theorem. Suppose X 6 c and let E1 be the aalgebra
generated by 2. The measure m :E X has a unique countably
additive extension to Z1 if and only if
(i) z is left continuous on (0,1].
(ii) z is of weak bounded variation.
63
If X 2 c then there is a function z satisfying (i) and
(ii) such that m has no countably additive extension to Ei
Proof. If m has a unique countably additive extension
m' to ZE, then by Remark 1.2.3, m' is bounded, so m is
z z z
also bounded. By Theorem 2.1 z is of weak bounded variation.
Since m is countably additive on 7, lim m (A.) = 0 whenever
Z Z 1
(A ) is a sequence in Z such that Ai' 0. For any sequence
an b where b E (0,1], the sequence [an,b) decreases to 0.
Hence we have 0 = lim m [an,b) = lim [z(b) z(a )], so z is
left continuous on (0,1].
Conversely, if (i) and (ii) hold, then by Theorem 2.1,
m is bounded. (i) implies lim m [ah,a) = 0 for every
h0
a E (0,1]. By Lemma 3.8 we have lim+ f[ah,a) = 0 for
h0
a E (0,1]. Consequently, by Corollary 3.4, m is countably
additive on Z. Since X c c and m is bounded, m is s
bounded by Corollary 3.6. We conclude from Theorem 1.3.4
that mz has a unique countably additive extension to Z1.
Now suppose that X 2 co. Let (e ) be the sequence of
unit vectors in co defined in Example 3.7. Let
el n = 1
an=
e Z e. n > 2
i=2
Consider the function z:T c defined by
0 l
z(t) = a I(,/n+l, 1/n] (t)
n= 1
64
for every t E T. One can show that z is left continuous
on (0,1]. Suppose that 0 < s < t < 1. If s = 0, then
since z(0) = 0, we have z(t) z(s) = a provided that
t E (1/n+1,l/n]. If s > 0, then for t E (1/n+1,1/n] and
k
s E (1/n+k+1,l/n+k], z(t) z(s) = an an+k = en+i'
i=1
provided k 2 1. If k = 0 then z(t) z(s) = 0. Now if
[albl),..., [ak,bk) are disjoint intervals in T, with
al < b1 g ... ak < bk, and if a1 > 0, then there is a
finite set nl < n2 ... s n2k of integers such that
n2i
z(b.) z(a.) = E e., 1 < i < k.
J="2i1
If al = 0, then z(bl) z(al) = a In either case, since
k n
E [z(bi) z(ai)] is a finite sum of the form E ciei for
i=l i=l
some n, where each ei is 0 or + 1, we conclude that
k
17 (z(b.) z(a ) ] <1.
i=l
By Remark 2.2, z is of weak bounded variation. Now if m
z
has an extension to a countably additive measure on El,
then m is sbounded. Let A = [1/n+l,1/n) for each n. Then
z n
(A ) is a disjoint sequence in 8, and m (An) =z(1/n) z(l/n+l) =
en, n E N. Therefore Imz(An) = 1 for every n, so mz is not
sbounded. Hence m cannot have a countably additive exten
sion. f]
The requirement that z be left continuous is in a sense
superfluous; if z(t) exists for every t E (0,1], then we
can define m[a,b) = z(b) z(a), (or m[0,b) =m(b) m(O)).
65
Essentially this is just normalizing the function z to be
left continuous. Moreover, as the following result shows,
if z(t ) exists at each point of (0,1] and z is of weak
bounded variation, then z determines a unique countably
additive measure on ZE.
3.10 Theorem. If z(t ) exists for every t E (0,1] and if z
is of weak bounded variation, then the function z' defined
by z'(t) = z(t), t E (0,1], and z'(0) = z(0) is left con
tinuous on (0,1] and of weak bounded variation.
Proof. Since by definition z' is left continuous on (0,1],
it suffices to show that z' is of weak bounded variation.
If [alb),..., [a ,b ) are disjoint intervals in T with
a1 < b ... s a < b then we can choose points a' < a.
1 1 n n 1 1
(if a = 0 let a' = 0) and b' < bi, such that for each i,
11 1
Iz(ai ) z(a')I < 1/2n and Jz(b ) z(b.') < 1/2n. Then
n n n
S z'(b) z'(a.)I E z(b.) z(b') + E Iz(ai) z(a.')
i=l 1i=l i=l
n
+ 1 Z z(b') z(a')
i=l
< 1 + iz (T).
z
By Remark 2.2 it follows that z' is of weak bounded variation.]
If z:T X is a function such that mz has a countably
additive extension to E1, then as we have already seen, z
is left continuous. It follows that righthand limits also
.exist at each point in [0,1). For suppose a E [0,1) and
S a Then the intervals a,b) decrease to a) and
bn a Then the intervals [a,bn) decrease to (a] E Z1, and
66
so by the countable additivity of.m mrn(a = lim m [a,b) =
z z z n
lim[z(b ) z(a)]. Thus z(a ) exists. By Theorems 3.10
and 3.9, every function z with right and lefthand limits
that is of weak bounded variation determines a unique
countably additive measure on Zi. Conversely, it is easy
to see that if m is a countably additive measure on l
with values in X, then there is a function z with right
and lefthand limits that is of weak bounded variation and
satisfies z(a) = m[0,a) if a E (0,1], and z(a ) = m[0,a]
if a E [0,1). If we normalize this function to be left
continuous and satisfy z(0) = 0, then we have established
a one to one correspondence between the class of all countably
additive measures on E1 and the family of all normalized
functions z that are of weak bounded variation. This is a
generalization of the classical theorem relating countably
additive Borel measures and functions of bounded variation
in the scalar case.
4. Measures in Hilbert Space. In this section we apply
our previous results to sharpen a theorem of Cramer [10],
and discuss its consequences in integration theory. H is
a Hilbert space with inner product (<,*), and z:T H is
function which generates a measure m = m:Z H.
z
: Recall from Section 1 that the family 8 of all half
open intervals in T is a semiring which generates E. If
8is any semiring, let Sx8 denote the class of all sets
AxB, where A,B E g. It is not difficult to see that 9x8 is
67
also a semiring. Elein.ents of gxg are called rectangles.
Consider the set ft:nction ml:gxB 1 defined by
I1 (AxB) = (m(A),m(B)).
By the next lemma, it will follow that mi is finitely addi
tive on x'g, and therefore by Remark 3.1, mi has a unique
extension to a finitely additive measure on the algebra 3
generated by S8x. 5 is just the family of finite sums of
rectangles in g,..
4.1 Lemma. If A:.B E Sx and AxB = AlxB1 + A2xB2, then
either A = A = \A2 and B = B. + B2, or B = B = B2 and
A = A1 + A2.
Proof. If A1 = A2, then clearly A = A1 = A2. Suppose that
Bl 0 B2 / ,. If b B1 n B2, then for any a E A,
(a,b) E A XB1 f A2xB2, which is a contradiction. Hence
B1 n B2 = Since B C B1 U B2 c B, E = B1 + B2.
If A1 / A2, then there are al'a2 E A such that
al A \A2 and a2 A2\A Since (a,,bl) and (a2,b2) E AxB
for any bi E Bi, i = 1,2, (al,b2) and (a2',b) E AxB. Since
(al,b2) E A1xB1, B2 c B1. Similarly, B1 z B2, so B= B1 = B2
Then as above we must have A1 n A2 = d and A = A + A2. 5
Now suppose that AxB = A xB1 + A2xB2, and that A= A =A2
and B = B1 + B2. Then
ml(AxB) = m (Ax(B1 + B2))
= (m(A), m(B1 + B2)
= (m(A), m(Bl)) +
= ml(AIxB1) + m (A2XB2)
68
It follows that mi is finitely additive on gxg, and so by
Remark 3.1, we may assume that mI is a scalar measure on J.
4.2 Remark. Suppose that (A.) is a finite disjoint family
in 8. Then we have
n n n
I m(A.) I ( m(A.), Z m(A.))
i=l i=l i=l
n n
= z E
i=l j=l
n n
< 7 E Iml(AixA) .
i=1 j=
We conclude that if mI is a bounded measure on 3, then m is
a bounded measure on Z, or equivalently z is of weak bounded
variation.
Since mI is a scalar measure, Theorem 1.2.2 (ii) implies
that ml v(ml) 5 4m1 on 3. It follows that in the defini
tion of regularity, Definition 3.2, we may replace m1 by
v(ml). In what follows we shall use v(ml). The next lemma
is the counterpart of Corollary 3.4.
4.3 Lemma. Suppose mI is bounded. Then mi is regular if
(i) lim v(ml, [s,so)x[t,t)) = 0, so E (0,1], tt0 ET.
ss
0
(ii) lim v(ml, [s,so)x[t,to)) = 0, t E (0,1], s,s ET.
tt
o
Proof. Suppose [a,b)x[c,d) E 8x8, and e > 0. By (i) there
is a 6 > 0 such that v(ml,[a 6,a)x[O,d)) < e and
v(ml,,[b 6,b)x[O,d)) < e. By (ii) there is an a > 0 such
69
that v(ml, [0,b)x[d a,d)) < e and v(ml, [0,b)x[c a,c)) < e.
Now let A = [a,b 6)x[c,da) and B = [a 6,b)x[c u,d).
It follows that cl A c [a,b)x[c,d) z int B, and that
B\A c [a 6,a)x[0,d) U [b 6,b)x[0,d) U [0,b)x[c a,c) U
U [0,b)x[d a,d), so v(ml,B\A) 5 4e. We conclude that mi
is regular, since the same type of argument can be applied
to any set in 8x8. O
We now apply Lemma 4.3 to obtain a connection between
the countable additivity of mi and the continuity properties
of the generating function z.
4.4 Theorem. Suppose that mi is bounded. Then z(t ) and
z(t ) exist for every t E (0,1) (with onesided limits at
0 and 1). mi is countably additive on J if and only if z
is left continuous.
Proof. Suppose that lir sup jz(t) z(s)I > 6 > 0. Then
Ss ,tt
there is a sequence s1 < t 1< s2 < t < ... < t such that
1 1 2 2 o
lz(ti) z(si) 2 6, i E N. Since Jz(ti) z(si) 2
2
Im[si,ti) 2 = ml [si,ti)x[si,ti), and since the sequence
([si,ti)x[si,t.)) is disjoint, this contradicts the fact
that m is a bounded scalar measure, and hence has finite
total variation. Therefore we have lim _iz(t) z(s)J = 0.
S : s, tt
O
Since H is complete there is then a z(to) E H such that
0
lim_ z(t) z(to)j = 0. The case for limits from the right
tt
is handled similarly.
70
Suppose that mI is countably additive on U; then m1
is continuous from above at 0. Since [t,to)x[t,t ) decreases
to 5 as t t we conclude that
0
0 = lim ml[t,to)x[t,to)
tt
0
o2
= lim z(t) z(t ) 2
tt
0
Therefore z(to) = z(to)
Conversely, suppose that z is left continuous. In view
of Lemma 4.3 and Theorem 3.3 we need only show
(*) lim_ v(m, [s,so)x[t,to)) =0, so E (0,l],t,to E T,
ss
0
and similarly for t t as in Lemma 4.3. By Schwartz's in
equality we have
ml[s,so)x[t,to) = ((m[s,so), m[t,to))l
S Iz(so) z(s) z (to) z(t) I ,
and so be the assumption of left continuity of z we see that
(**) lim_ ml[s,so)x[t,t ) = 0
ss
0
for s E (0,1] and t,to E T. Since v(ml) is monotone, the
limit (*) exists for every s E (0,1] and t,t E T. Suppose
that for some choice of so,t, and t we have
lim_ v(ml,[s,s )x[t,tO)) > 6,
ss
0o
for some 6 > 0. Then there is a family of disjoint rectangles
[ai,b.)x[ci,d.), 2 i n2, contained in [0,s )x[t,to), such
that
n2
"2
SIml[ai,bi)x[c ,di) > 6.
i=2
71
We shall now show that we may assume b. < s for each i.
Suppose that [a,b)x[c,d) is any rectangle in [O,so)x[t,to)
such that b = so, and
Iml[a,b)x[c,d) = a > c > 0,
for some c > 0. By the continuity condition (**) we can
select a b' such that a < b' < s and
0
Iml[b', so)x[c,d) < a c.
Then Iml[a,b')x[c,d) I > Iml[a,b)x[c,d) Iml[b',so)x[c,d)
> a (a c) = c.
Now if any of the b. = s then the argument above shows
that we can find new b.'s such that b. < s and
1 1 O
n2
"2
E Iml[ai,b.)x[ci,di.) > 6.
i=2
Let 6 = min (s b.); then v(m, [s 61 s )x[t,t )) > 6.
2i 0 1 o 0 o o
so we can repeat the process above. By induction we obtain
a disjoint sequence (A xBi) of rectangles and an increasing
sequence (nk) of integers (set nl = 1) such that
nk+1
S ml(AixBi) > 6,
i=nk+l
contradicting the boundedness of mi. We conclude that (*)
holds. The proof for tt is similar to the above. O
0
The theorem we are about to prove was essentially
stated by Cramer [10], but his statement and proof contain
a gap which we now rectify. In our terminology, the statement
of Cramer's result is as follows. Let z:T H be a function
generating m and mi as above. If mi is a bounded measure,
72
then m and mi have countably additive extensions to Z1 and
.1 respectively, where U1 denotes the aalgebra generated
by U. Since even for scalar functions the measure m is not
z
countably additive unless z is left continuous, the theorem
as stated is not correct. We now present the corrected
version.
4.5 Theorem. If mI is bounded, then the measures m and m1
have unique countably additive extensions to Z1 and 51
respectively if and only if z is left continuous. In this
case we have
(*) (m(A), m(B)) = m (AxB)
for every A and B in El, where we identify the extensions by
m and mi.
Proof. If mi has a countably additive extension then z is
left continuous by Theorem 4.4. Conversely, if z is left
continuous, then by this same theorem mi is countably addi
tive on J. By the assumption of boundedness, mi has a unique
extension to 31 (by the classical extension theorem or Theorem
1.3.4), and z is of weak bounded variation by Remark 4.2. By
Theorem 3.9, m has a unique countably additive extension to
1'
To see that (*) holds, fix B E E, and define set functions
p and X on E1 as follows:
p(A) = ml(AxB), X(A) = (m(A), m(B)),
for A E Z1. Since m and mi are countably additive, it follows
that p and X are finitely additive and continuous from above
at 0; thus p and X are countably additive scalar measures.
73
Since X(A) = u(A) for A E Z, it follows by the uniqueness of
extensions that X(A) = p(A) for A E ,El or m (AxB)= (m(A),m(B))
for A E and B E E fixed and B E E1 fixed. Repeating this
argument for A E 71 fixed and B E gives the desired result. M
The author is indebted to Professor Brooks for dis
cussions on the previous material; in particular the proof
of (*) in Theorem 4.5 is due to him.
As an application of Theorem 4.5, we discuss the most
important case, namely functions z with orthogonal increments.
A function z:T H has orthogonal increments if
(z(tI) z(sl), z(t2) z(s2)) = 0 whenever s < tl < s2 < t2.
The differences z(t) z(s) are called the increments of the
function z. This property simply states that for disjoint
sets A,B E 8 we have m (AXB) = (m(A), m(B)) = 0; mI therefore
reduces to a nonnegative measure on S if we define
ml(A) = (m(A), m(A)) = im(A) 2. It follows from orthogonality
that mI is finitely additive, since
ml(A + B) = (m(A+B), m(A+B))
= (m(A), m(A)) +(m(A), m(B)) +
+ (m(B), m(A))+ (m(B), m(B))
= m1(A) + m (B).
Since mI is nonnegative and finitely additive it is monotone,
and so m (A) 5 m (T) for every A E E. By Theorem 4.5 we
conclude that m and mi have unique extensions to El if and
only if z is left continuous. Moreover, since mI is bounded,
Theorem 4.4 implies that z(t+) and z(t) (z(0 ) and z(l))
always exist. It follows that any function z with orthogonal
74
increments determines countably additive measures m and ml
if we normalize z to be left continuous. In fact, z is
continuous except possibly at countably many points, since
F(t) = ml[O,t) is a nondecreasing function and Iz(t) z(s)12=
F(t) F(s).
Suppose now that z is any function with orthogonal in
crements, not necessarily left continuous, and z generates
the measures m:E H and m:E [0,) as above. Since
[m(A)]2 = sup Jm(B) 2 = sup ml(B) = m (A),
BcA BQA
BEE BEE
it follows that convergence in measure with respect to m is
equivalent to convergence in measure with respect to mi.
Moreover, if (fk) is a sequence of step functions on T, then
the indefinite integrals of the fk's are uniformly mcontinuous
if and only if they are uniformly mlcontinuous. Suppose
n
f = E a.IE is a step function; then
i=l E.
n n
I Efamd2 = (< a.m(E n Ei), Z a.m(E n Ei))
i=l i=l
n 1 
= ai 2ml(E n Ei) = Elfl2dm
i=l
using the orthogonality of the measure m. By the remarks
above, and by Definition 1.5.4 and Theorem 1.7.1, we conclude
that f E L(m) if and only if f E Lo(ml). Moreover, for
step functions f and g it can be shown that
(SEfdm, IFgdm) = EnFfgdml'
and so by continuity this equality holds for every f and g in
75
L(m) = Lo (m1). If z is left continuous, so that m and m
have countably additive extensions to Borel measures, then
by Theorem 1.8.5 the class of step functions is dense in
L(m) = L (m) and the results above still hold by con
tinuity.
CHAPTER III
Stochastic Integration
In this chapter we shall restrict the integration
theory of the preceding chapters to the case where the
Banach space X is an L space over a probability measure
space. Stieltjes integrals with respect to a function
z:T Lp are stochastic integrals in the sense that the
integral is an element of L or in probabilistic terms,
a random variable. We show that the general integration
theory of Chapters I and II includes certain stochastic
integrals previously defined by other methods.
In Section 1 we introduce and discuss the probabilistic
concepts which are used. Section 2 consists mainly of back
ground material on the theory of stochastic processes. In
Section 3 we attempt to motivate the study of stochastic
integration, and then discuss the general integration theory
in the present context. The sample path stochastic integral
is studied in Section 4; we compute some statistical pro
perties of the integral in a simple special case, and discuss
an integration by parts formula. The WienerDoob stochastic
integral in L2 is discussed briefly in Section 5. In Section
6 we deal with martingale stochastic integrals, and prove a
general existence theorem.
77
1. Probability Conceots and notation. We say that (1,F)
is a measurable space if 0 is a nonvoid set (w is a generic
element in in D) and F is a aalgebra of subsets of n. If
there is a countably additive measure P defined on F with
values in [0,1], and if P(n) = 1, then we say that (G,F,P)
is a probability measure space, and P is a probability
measure. A probability measure space, or probability space,
is therefore just a finite nonnegative measure space with
the property that 0 has measure 1.
Suppose that (Q,F) and (Q',F') are two measurable spaces,
and x:O n' is a function. x is said to be measurable
(relative to F and F') if the set [x E A'] x (A') belongs
to F for every A' E F The family of all sets [x E A'],
A' E F', is a aalgebra contained in F, and we denote it by
F(x). Note that F(x) is the smallest aalgebra F" contained
in F such that x is measurable relative to F" and F'. If P
is a measure on F, then a measurable function x determines a
measure P' on F' by the relationship P'(A') = P[x E A'] for
each A' E F'. If P is a probability measure on F then P'
is a probability measure on F' as well. In this case P' is
called the distribution of x. When n' = 1 we shall always
take F' to be the aalgebra of Borel sets in 0.
A measurable scalarvalued function on a probability
space (0,F,P) is called a random variable. As motivation
for this term, suppose we have a system whose state can be
described by a scalar variable x, and whose behavior is
subject to some sort of statistical variation. We imagine
78
the numbers x corresponding to the (random) states of the
system to be the values of a function x defined on some
probability space. The assumption that x is a measurable
function insures that the probability of each event x E B
is defined, where B is any Borel set. This probability is
just P[x E B].
Lp L (Q,F,P), where 1 s p < , will denote the Banach
space of all equivalence classes of random variables x such
that
E xlPx n Mx(U) IP(d,) < m.
L, denotes the space of equivalence classes of essentially
bounded random variables. It is known (see Dunford and
Schwartz [15]) that for 1 p < , L is weakly sequentially
complete. Since co is not weakly sequentially complete, it
follows that Lp j co; hence Theorem 11.3.9 applies to these
spaces. The number ElxlI is called the pth absolute moment
of x. If k E N, then Ex is called the kth moment of x.
In particular, the first moment Ex is called the mean or
expectation of x. It is well known from the theory of in
tegration that the expectation operator E is a continuous
linear functional on L1. If x E L2, then the number
2 2#
(2(x) = EIxEx2 = ElIx2 Ex2
is called the variance of x, and represents in some sense the
dispersion of x about its mean value Ex.
Standard references for results in probability theory are
Love [21] and Chung [9]; most of the material in this first
section can be found in either of these works.
79
2. Stochastic Processes. In this section we discuss the
definition of a stochastic process (random function) as a
mapping from a subset T of R into an L space over a pro
bability space. In order to illustrate some of the important
properties that a process may have, we discuss two classical
examples, the Poisson process and the Wiener process.
Let T be a subset of R whose elements are interpreted
as points in time, and suppose we have a system whose
observable state at time t E T is some scalar quantity x(t).
If the system is subject to influences of a random or
statistical nature, then we are led to suppose that x(t) is
a random variable for each t E T. Our mathematical model
of the system is then a family of random variables x(t), t E T,
which are defined on some underlying probability space (0,F,P).
/
If we assume that each x(t) belongs to Lp, so that the con
vergence properties of this space are available, then we have
defined a mapping x from T into L x is called a stochastic
process or random function since it represents the behavior
of a system subject to stochastic (random) influences.
Since Lp is a space of (equivalence classes of) functions
from Q into , we may also think of a random function x as a
mapping from TxO . Then each w E n determines a function
x(*,w):T 4; this function is called the sample function or
sample path corresponding to w. Every set A E F determines
a set of sample functions (x(.,wu):w E A), and we say that the
probability of this set of sample functions is p if P(A) = p.
80
A fundamental problem in the theory of random functions con
cerns the fact that we would like to discuss the probabilities
of sets of sample functions which may not correspond to sets
in F as above. For example, given any finite set of times
tl'...tn and Borel sets B1,...,B n the set
n
[x(t.) E B.:l
i=l 1
is in F. Even for countably many t.'s and B.'s such sets are
1 1
in F. However, the set
[x(t) E B:t E (a,b)] = n [x(t) E B]
tE(a,b)
clearly need not be measurable. Doob [13] discovered a
reasonable way to skirt this issue. He was able to show that
every random function x is equivalent to a random function y
which behaves nicely in this regard, where equivalence means
that for every t E T, P[x(t) = y(t)] = 1. We shall always
assume for the sake of simplicity that we are dealing with
such wellbehaved processes.
In order to discuss two important examples of stochastic
processes we need to introduce a few additional concepts.
Suppose that x(t), t E T,is a random function, and tl,...,tn
is a finite subset of T. Assume for simplicity that x(t) is
realvalued for each t. Let Rn denote real Euclidean nspace,
andlet Bn denote the aalgebra of Borel sets in Rn. It can
be shown that the family of all product sets Bx...xB n, where
B. C B for 1 s i < n, is a semiring which generates B.
Consider the mapping w [x(tlW),...,x(tn,)],which is a
81
function from 0 into Rn. Since each x(ti) is measurable,
it follows that the set
n
[[x(t ),...,x(t )] 6EB >.... ,xB ] = [x(ti) E Bi]
i=1
belongs to F, for every product set Blx...xB n As in Section
1, the mapping w [x(tlw),...,x(tnw)] determines a measure
on the semiring of product sets, and by a standard extension
theorem (see Halmos [17]), this measure extends to a probability
measure P on Bn. The measure P is called the
tmeasut tl..t1 .t
tl'''tn 1'*,'' n
joint distribution of the random variables x(tl),...,x(t ).
It seems reasonable, and in fact can be verified (see Loeve
[21]), that a random function is characterized by the family
of all its joint distributions Ptl
Suppose that x(t), t E T, is a random function. As in
Section 11.4, we say that x(t2) x(tl) is an increment of the
process. x(t) is said to have stationary increments if for
every choice of t > 0 and s < t s < t2 ... sn < t in
1 1 2 2 n n
T, the joint distribution of the increments x(ti + t) x(s + t),
1 s i s n, coincides with the joint distribution of the incre
ments x(ti) x(si), 1 s i < n. In other words, the prob
abilistic properties of the increments are invariant under a
shift in time.
Recall that every random variable x(t) x(s) determines
a aalgebra F(x(t) x(s)) as in Section 1. We say that
x(t), t E T, has independent increments if for each choice
of sI < t 1 ... sn < tn in T and Ai F(x(ti) x(si)),
n
1 s i n, P[ n Ai] = P(AI) ... P(A ). In this case it is
i=l
82
n
known (see Loeve [21]) that if E TT x(ti) x(si) < m,
i=l
then we have
n n
E T [x(t.) x(s )] = TT E[x(t.) x(si)].
i=l i=l
Now we consider a classical random function called the
Poisson process. Suppose we have a general situation where
events occur in a random fashion that is in some sense uni
form over T = [0,1]. Let x(t) denote the number of events
which have occurred in the interval [O,t). It seems reason
able to assume that the random function x has the following
properties:
(i) Independent increments. The increment x(t) x(s),
t > s, represents the number of events occurring in [s,t),
and we would expect that over disjoint intervals of time,
the numbers of events occurring in those intervals are in
dependent of each other.
(ii) Stationary increments. The uniformity assumption
means that over disjoint intervals of the same length, we
expect that the distribution of occurrences is the same.
(iii) x(0) = 0.
If in addition we assume that there is a X > 0 such that for
small values of h > 0 we have
P[x(h) = 1] = Xh + o(h),
and
P[x(h) = 0] = 1 Xh + o(h),
then
P[x(h) > 1] = o(h),
83
and it can be shown (see Doob [13]) that for every s < t
in T, we have
,(ts) n (t s)n
P[x(t) x(s) =n] = e( n(s)
n:
for n = 0,1,2,... This discrete probability distribution
is called the Poisson distribution, and the relation above
states that the random variable x(t) x(s) is distributed
according to a Poisson distribution, with parameter .. The
mean and variance of the increments are given by
E[x(t) x(s)] = X(ts),
and
2 (x(t) x(s))= E[x(t) x(s)] [E(x(t) x s))]2
= (t s).
The sample functions of a Poisson random function
(except for a set of sample functions of measure zero) are
monotone, nondecreasing, nonnegative, integervalued functions
of t, with at most finitely many jumps of unit magnitude in
T. This might be expected in view of the eventcounting
interpretation of x(t).
A second important example of a random function is the
Wiener process, which has been used extensively as a mathe
matical model for Brownian motion. Suppose we observe a
particle undergoing Brownian motion, and restrict our atten
tion to, say, the x coordinate of its position. Due to the
random bombardment of the particle by the fluid's molecules,
this x coordinate changes in an erratic fashion as time
progresses. If we denote the value of x at time t > 0 by
84
x(t), then it seems reasonable to assume that x(t) is a
random function which has the following properties:
(i) Independent increments. The increment x(t) x(s)
represents a change in position along the xaxis over the
interval [s,t). Over disjoint intervals of time these
changes should be independent of each other due to the ran
dom character of the molecular bombardments.
(ii) Stationary increments. Since we would assume that
the fluid in which the particle is suspended is homogeneous,
the distributions of change over two intervals of the same
length should be the same.
(iii) x(0) 0, assuming we use the position of our
initial observation as the center of the coordinate system.
(iv) Normally (Gaussian) distributed increments, with
mean zero and variance proportional to the length of the
time interval. This follows from the assumption that the
motion is random and continuous, and means that if a < b,
then
P[a < x(t) x(s)
/27c2 (ts) a
where s < t, and o > 0 is some constant, namely the variance
over an interval of unit length (we have assumed t 0
so T = [0,))
It can be shown that except for a set of sample functions
of measure 0, the sample functions of a Wiener process are
continuous, nowhere differentiable, and of infinite variation
over any finite interval. Although this last property runs
85
counter to what intuition says the sample paths should look
like, the Wiener process has played a very significant role
in the application of probability theory to physics and
engineering. In fact, the first stochastic integrals were
defined by Wiener for scalar functions with respect to a
Wiener process.
A standard reference for results on stochastic processes
is Doob [13]. Much of the material in this section can be
found in his classic work.
3. Stochastic Integrals. We now discuss as example which
motivates the theory of stochastic integration. Suppose we
have a system consisting of a signal processing unit with
an input signal z(t) and an output signal x(t). Suppose
that the effect of the processor can be described by a
scalar function of two variables g(t,T) in the following
way. A small change z(dT) = z(T + dr) z(T) in the input
at time T produces a change in the output of the system at
a time t 2 7 given by g(t,T)z(dr). Moreover, suppose that
the response of the system is linear, in the sense that
small disturbances z(d 1),...,z(diTn produce a change in
n
the output given by E g(t,T.)z(dT.), for t 2 max T.. If
1 1 1
i=l
we assume that g(t,T) = 0 if t < T, then for a given input
function z(t), we have formally that
x(t) = Jf g(t,T)z(dT).
As long as the functions g and z are sufficiently well
behaved scalar functions, the definition of this integral
poses no problem.
86
Suppose, however, that the input z(t) is a random
function, for example, a signal contaminated by random
noise. Then the output will also be random, and we are
faced with the problem of how to define the integral above
so that it is a random variable for each t. If the function
g is continuous in T for every t, and if almost every sample
function of z is a function of bounded variation, then we
could define the random output in a pointwise fashion as
fo 1 low0s:
(*) x(t,w) = SS_ g(t,T)z(dT,u).
Stochastic integrals defined in this way are called sample
path integrals; although they are the most restrictive kind
of integral, their study does provide motivation for the more
powerful kinds of stochastic integrals. In general, this
simple solution to the problem is not feasible, since even
processes as wellbehaved as the Wiener process have almost
every sample function of unbounded variation.
Using the integration theory of Chapters I and II when
the Banach space X is an Lp space over a probability space,
we obtain a very general definition of the stochastic inte
gral. Suppose that z:T Lp is a random function, where
1 < p < m. As in Chapter II we can define a measure m:7 Lp
by setting m[a,b) = z(b) z(a) on S. For each A E E, m(A)
is thus a random variable in Lp. The measure m is bounded
if and only if z is of weak bounded variation, and by the
form of Lp* this is equivalent, using Theorem 11.2.1, to
87
requiring that the scalar functions t E[z(t)x] are of
bounded variation for every x E L where 1/p + 1/q = 1
if 1 < p < , and q = if p = 1.
Suppose that z is of weak bounded variation. Then z
determines a class L(z) L(m) of integrable scalar functions,
and for f E L(z) we write
Sf(t)z(dt) = STf(t)m(dt).
We shall call such integrals stochastic Stieltjes integrals,
or norm integrals for short, since they are defined as the
limit of a sequence of integrals of step functions in the L
norm. If Xllp denotes the L norm (ElxlP) /p of an element
x E Lp, then it is known (see Loeve [21]) that 1 r s p and
x E L implies that llxllr s lXll This shows that Lr 7 L ,
and that the canonical embedding Up of Lp into L is continuous
pr p r
By Theorem 1.5.8 it follows that if f E L(z) then f E L(Uprz)
and
Upr(SEfdz) = SEfdUprz
That is, if f is integrable with respect to z as a mapping into
Lp for 1 p < , then f is integrable with respect to z as a
mapping into Lr, 1 r : p, and the integrals coincide. More
over, it is clear that the same sequence of step functions
determines the integrals of f in Lp and in Lr.
By Theorem 1.5.8 we have
E(jTf(t)z(dt))x = STf(t)E[z(dt)x],
for every x E Lq, since as is well known, this is the form
of linear functionals on Lp (see Dunford and Schwartz [15]).
Tr.us
E(STf(t)(d)) = Tf(t)E[z(dt)]
Also
E Tf(t)z(dt) 2 = E[Tf(t) z(dt) ] [Tf(T)z(d ) ]
= Tf(t)E[z(dt)JTf(T)z(dT) ].
In general
E[fjf(t)z(t))]k = Tf(t)E[z(dt) (STf(T)z(dT))k1,
since [ Tf(T )z(dT)]k1 belongs to Lk/k_ if z:T Lk. If
we evaluate the scalar functions
t E[z(t) (STf(T)z(dT))k] ,
then we can compute the kth moment of the integral by per
forming a scalar integration. In practice this may be of some
value.
4. Sample Path Integrals. Let T = [0,1] and g be a continuous
function on T. Let z(t). be a Poisson process on T with para
meter .. As described in Section 2, almost every sample path
z(.,w) is a monotone nondecreasing function, with z(0,w) = 0,
and finitely many jumps of unit magnitude. We can therefore
define the RiemannStieltjes integral
S. g(t)z(dt,uw)
for almost every w. By Theorem 11.2.3 it follows that
1 g(t)z(dt,w) = lim Z g(i1/2n) [z(i/2n, ) z(i1/2n u1)]
n i=l
for almost every w. Since the summands g(i1/2n) [z(i/2) 
z(i/2n)] are random variables it follows that the sample
89
path integral 1 g(t)z(dt) is also a random variable.
Moreover, we can evaluate the integral directly, due to the
simple behavior of the sample functions. Fix w and let
n = z(l,u1). Let tl,...,t denote the n points of jump of
o 1 n o
o
the sample function z( ,w). For every n and i, the difference
z(i/2n,w) z(il/2,uw) is either 1 or O, depending on whether
one of the t.'s belongs to (i1/2n,i/2n] or not. By the con
tinuity of g we see that
n
o
Si g(t)z(dt,w) = E g(ti).
i=l
In general, denote by ti () the point in T of the ith jump
of the sample function z(*.,) ; that is, for t s ti.(),
z(t,w) i 1, while for t > ti(w), z(t,w) 2 i. Since
[w:ti (u) < a] = [z(a) 2 i] E F for each a E T, t. is a random
variable for each i. Thus
1 z(l,w)
JO g(t)z(dt,w) = g(t (w)).
i=l
That is, the integral of g with respect to the Poisson process
z is the sum of a random number (z(l,w)) of random variables
of the form g(ti(w)).
S We can formally compute the mean of the integral as
follows: we have
2n 2n
E[ E g(il/2) [z(i/2) /2Z(i1,n)]= g(i1/2n) (1/2n
i=1 i=l
2n
Since lim E g(i1/2"n)(1,/2) = 1g(t)dt, and since the
i=l
2n
functions E g(i1/2n )[i,,2nI/2n) converge uniformly to
i=l [i/2nl/2
g on T, we would expect that
E[j g(t)z(dt)] = X1 g(t)dt.
90
This will follow once it is established that the stochastic
Stieltjes integral exists. Proceeding to the formal com
putation of the variance of the integral, we compute
2n
El Z g(i1/2n) z(i/2n) z(i1/2)]2 =
i=l
2n 2n
= E E g(i1/2n)g(jl/2n)E[z(i/2n) z(il/2n) i x
i=l j=l
x [z(j/2n)z (jl/2n)
n
2
= g(i1/2n) 2 E[z(i/2n)z(il/2 )] +
i=l
nn
+ E g(i1/2n)g(jl/2n) E[z(i/2n)z(i1/2n)] i
ifj=1
x E[z(j/2n) z(j1/2n)],
since z has independent increments. To evaluate the second
moments, we suppose that x is a Poisson random variable with
parameter a. Then
 n C ni
2 2 a a a a
Ex = n = ae L n n
Sn=0 n= 1
C n1 m n1
= aea[ (n1) (nl) + E (nl),
a a a
= ae [ae + e ] = a(a + 1).
Since z(i/2n) z(i1/2n) is a Poisson random variable with
parameter \2n, we conclude
E[z(i/2n) z(i1/2n) = 2 X(1/2n 2 + ) (1/2n).
Therefore,
n
2
El E g(i1,'/2) [z(i/2) z(i1/2"n ] 12
i=l
2n 2n 2n
1n.21/
= g(i1/2) g (j1/2 ").2(/2n) + 7 g(i1/" 2) (1/2n)
i=1 j=1 i=l
The double sum converges to
X 20 0 g(t)g(s) dtds = X (Jg(t)dt) (J g(s)ds)
= x2 0 g(t)dt 2
10
The sum on the right converges to Xj g(t) dt. We conclude
That if the integral J1 g(t)z(dt) exists as a norm integral
in L2, then
2 1 2 1 d2 1 2
S(0 g(t)z(dt)) = X[Ifo g(t)dt 2] +).Olg(t) dt 
1 2
l0 g(t)gdt.
= XfOIg(t)I2dt.
Now since g is continuous we need only show that z is of weak
bounded variation as a mapping into L2; then the norm integral
.. O g(t)z(dt) belongs to L2, hence to L as well. By compu
tations as above,
n n n
El I z(t ) Z(si) 12 S 2(t.i ) (tj s j)
i=l i=l j= 1
n
+ E X(t. s.)
i=l 1 1
< X2 + X,
92
since 0 s. s t. 1 for each i. .We conclude that z is of
1 1
weak bounded variation by Remark II.2.2. Since the simple
2n
functions E g (il/'2 n)
functions Ei g()iln n converge uniformly to
g on T, we conclude that
2n
o g(t)z(dt) = lim E g(i1l/2n) [z(if2n) z(il2n)
i=l
in L2 and hence in Ll. The expressions for the mean and
variance now follow.
Motivated by this example, we can obviously define
sample path integrals under the following conditions:
(i) f:T is a continuous function and z(t) is a
process with almost every sample function of bounded varia
tion.
(ii) f:T d is a function of bounded variation and z(t)
is a process with almost every sample function a continuous
function.
If either (i) or (ii) hold then by the formula for
integration by parts for RiemannStieltjes integrals (see
Rudin [26]) we have
1 1 1
(*1 .S f(t)z(dt, w) + 0 z(t,w)f(dt)= [z(*,w))f(.)]0
where both integrals are RiemannStieltjes integrals as in
Definition 11.2.4, and (*) holds for almost every w.
In case (i), random variables of the form
n
Z f(il/n) [z(i/n) z (i1/n)]
i=l
converge to f(t)z(dt) for almost every w. In case (ii),
convege to0 (i
sums of the form
n
E z(il/n) [f(i/n) f(il/'n)]
i=l
converge to [0 z(t)f(dt) for almost every w. In both cases,
then, the integral exists as a measurable function on (P,F,P);
that is, the integral is a random variable.
In general, the existence and computation of the moments
of a sample path integral depend on the properties of the
functions Ez(t), Ez(t)z(s), and so on, as we saw in the case
of the integral with respect to a Poisson process.
A case of particular interest concerns random functions
whose sample functions are nonnegative and nondecreasing
functions. Since s < t implies that 0 < z(s) s z(t) almost
everywhere, we see that 0 s Ez(s) Ez(t) Ez(l). If we
suppose that z(1) belongs to LI, then it follows at once
that z is of weak bounded variation, since the mapping
t Ez(t) is nondecreasing. If f E L(z), then we know from
Theorem 1.5.8 that Esf (t)z(dt) = f(t)Ez(dt). The Poisson
integral discussed previously is a special case of this, since
Ez(t) = Xt.
We now discuss an interesting connection between norm
integrals and certain sample path integrals. Suppose that
z:T L is a random function of weak bounded variation, and
suppose that z is also measurable with respect to the product
measurable space (TxC., 1ExF). (E1xF denotes the smallest
aalgebra containing all rectangles AxB with A E E1 and B E F).
Shachtman [27] asserts that the following integration by parts
theorem holds, when p = 1.
94
4.1 Theorcm. If f:T 4 t is a function of bounded variation
and z:T L1 is product measurable and of weak bounded
variation, then f E L(z), and
SAf(t)z(dt) = jAz(t)mf(dt),
where the integral on the right is the sample path integral
with respect to the negative of the LebesgueStieltjes
measure generated by f. A is taken to be an interval if
m is not countably additive, and a Borel set if m is count
z z
ably additive.
The proof of this theorem (Shachtman [27]) discusses
only the case when A = T. That the theorem is false as
stated is trivial, since it fails even if z is a nonrandom
function, for example z(t) = t. For if f is constant then
mf 0. The question remains whether the formula
1 f(t)z(dt) = S z(t)mf(dt) + [zf]0
can be true under these hypotheses. We establish the follow
ing result.
4.2 Theorem. Suppose f:T i is a continuous function of
bounded variation, and z:T L is product measurable and of
weak bounded variation, where 1 s p < m. Then
1 1 1
f0 f(t)z(dt) + J0 z(t)m(dt)= [zf],0
where the integral in the center is the sample path integral
of z with respect to the BorelStieltjes m on Z1 generated
by f.

Full Text 
xml version 1.0 encoding UTF8
REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchemainstance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd
INGEST IEID E2LMNFUPG_ZJB0FI INGEST_TIME 20170714T23:33:10Z PACKAGE UF00097651_00001
AGREEMENT_INFO ACCOUNT UF PROJECT UFDC
FILES
PAGE 1
Vector Measures and Stochastic Integration By Franklin P. Witte A DISSERTATION PRESENTED TO THE GRADUATE COUNCIL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 1972
PAGE 2
UNIVERSITY OF FLORIDA illilllilllilllil 3 1262 08552 4642
PAGE 3
To Dot Whose patience and encouragement made this possible,
PAGE 4
ACKNOWLEDGEMENTS I would like to thank ray wife Dot for her understanding and encouragement throughout the past four years. I am indebted to each member of my committee; special thanks are due to Dr. James K. Brooks, who directed my research and guided my study of abstract analysis, and to Dr. Zoran R. PopStojanovic, who guided my study of probability theory. Finally, I would like to thank Jean Sheffield for her excellent typing. Ill
PAGE 5
TABLE OF CONTENTS Page Acknowledgements iii Abstract v Introduction 1 Chapter I. Abstract Integration 5 II. Stieltjes Measures and Integrals 48 III. Stochastic Integration 76 References 105 Biographical Sketch 108 XV
PAGE 6
Abstract of Dissertation Presented to the Graduate Council of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy VECTOR MEASURES AND STOCHASTIC INTEGRATION By Franklin P. Witte December, 1972 Chairman: J. K. Brooks Major Department: Department of Mathematics This dissertation investigates the stochastic integration of scalarvalued functions from the point of view of vector measure and integration theory. We make a detailed study of abstract integration theory in Chapters I and II, and then apply our results in Chapter III to show that certain kinds of stochastic integrals, previously defined by other means, are special cases of the general theory. In carrying out this program we prove extended forms of the classical convergence theorems for integrals. We also establish a generalization of the standard extension theorem for scalar measures generated by a left continuous function of bounded variation. The special case of measures in Hilbert space is discussed, and a corrected form of a theorem of Cramer is proved. In Chapter III we show that certain sample path integrals, the WienerDoob integral, and a general martingale integral are included in the abstract integration theory. We establish a general existence theorem for stochastic integrals with respect to a martingale in L , 1 < p < Â«>.
PAGE 7
INTRODUCT ION This dissertation concerns stochastic integration, a topic in the theory of stochastic processes. A stochastic process (random function) is a function from a linear interval T into an L space over a probability measure space. In its broadest terms, stochastic integration deals with linear transformations from a class of functions x to a class of stochastic process fxdz, which depend on a fixed process z. The functions x which are transformed may be sure (scalarvalued) functions, or more generally, random functions. Since randoia functions take their values in an L space they are vectorvalued functions . Hence we are in the framework of the general theory of bilinear integration developed by Bartle [1] . The vectorvalued function z generates a finitely additive measure on an appropriate ring or algebra of sets. When the integrand x is a sure function we are concerned with the integration of scalar functions by vector measures. When the integrand is another random function we have the full generality of the Bartle Theory, where both the integrand and the measure take values in Banach spaces X and Z respectively, and there is a continuous bilinear multiplication from XxZ into another Banach space Y, in which the integral takes its values .
PAGE 8
2 Until recently the field of stochastic integration appeared to be in much the same state as the area of real function theory was before the systematic introduction of measure theory. Several relatively unrelated kinds of stochastic integrals can be found throughout the literature on stochastic processes and their applications. Most of these integrals are defined as limits of approximating sums of the RiemannStieltjestype. This approach, however, is not wellsuited to stochastic integration in general, and certain difficulties may arise. With the exception of some work by Cramer [10] , the measure properties inherent in the stochastic integration scheme have not been exploited. There is currently a great deal of interest in unifying the field of stochastic integration using vector measure and integration theoiry. E. J. McShane, who has been investigating stochastic integration for several years, is presently writing a book which should contribute to this unification. The recent Symposium on Vector and Operatorvalued Measures and Applications (Salt Lake City, August, 1972) was specifically concerned with stochastic integration, and attracted many of the known experts in both the fields of measure and integration theory (Brooks, Dinculeanu, lonescu Tulcea, Kelley, and Robertson) and stochastic processes and stochastic integration (Chatterji, Ito, Masani, and McShane) . A very recent unpublished paper by Metivier [23] , presented at the Symposium, deals with the problem of incorporating a rather general kind of stochastic integral within the framework of the Bartle integration theory.
PAGE 9
3 The purpose of this dissertation is more modest. We confine our attention to the case of scalarvalued integrands, and present a detailed investigation and synthesis of the appropriate vector measure and integration techniques, which are then used to exhibit certain kinds of stochastic integrals as special cases. In carrying out this program we make improvements in known results concerning the convergence of sequences of integrals, measures generated by vector functions, and existence of stochastic integrals. Chapter I develops the abstract integration theory of Bartle [1] for the special case at hand, namely, the integration of scalar functions by finitely additive vector measures. In this setting more is true than in the general Bartle theory. For example. Theorem 1.5.8 states roughly that the integral commutes with any bounded linear operator. This fact enables us to prove extended forms of the classical convergence theorems, and sketch a theory of L spaces. We discuss briefly the case of integration by countably additive vector measures, and introduce another integral developed by Gould [16] . This integral is defined by a net of RiemannStieltjestype sums, as opposed to the Lebesgue approach used in the development of the Bartle integral by means of simple functions. We show that these two integrals are equivalent a fact not discussed by Gould. Since a stochastic process is a function from an interval T of the line into an L space, as will be discussed in Chapter III, the stochastic integral of a scalar function f on T
PAGE 10
4 with respect to z can be thought of as a Stielt jestype integral ["^^f (t) z (dt) in L_j. With this in mind, we investigate in Chapter II the properties of a Stieltjes measure m generated by a vector function z on T. We discuss the boundedness of m, and the question of the existence of a countably additive extension to the Borel sets of T. For a rather general class of Banach spaces we show in Theorem II. 3. 9 that the classical extension theorem for scalar measures on an algebra is valid. We also give a counterexample to show that the theorem fails to hold in the Banach space c . Finally we discuss the special case when z takes its valvies in a Hilbert space. We consider a theorem of Cramer [10] , which is incorrect as stated, and apply our results to establish a corrected version. Some implications of this theorem are then discussed. In Chapter III we first present some background information on probability and stochastic processes, and then motivate the study of stochastic integration. We show that the general integration theory of Chapters I and II includes certain sample path stochastic integrals, the WienerDoob integral, and a martingale stochastic integral. Several related results are also discussed.
PAGE 11
CHAPTER I Abstract Integration In this chapter we discuss the integration of scalar functions by finitely additive vector measures. Section 1 establishes the notation we will use, while Sections 2 and 3 list some basic properties of vector measures. In Section 4 we discuss the space of measurable functions and convergence in measure. Section 5 is concerned with integration theory proper, and in Section 5 we apply this theory to prove convergence theorems for integrals. Section 7 outlines part of the theory of L spaces in this setting. Countably additive measures are discussed briefly in Section 8. Finally, in Section 9, we compare a Riemanntype integral with the Lebesguetype integral of Section 5, and establish the equivalence of these two approaches. 1. Notation . Throughout this dissertation the following terminology and notation will be used. N is the set of natural numbers {1,2,3,...} and N ={l,2,...,n}. The lowercase letters i,j,k,t, and n will always denote elements of N. A is a finite subset of N. R denotes the real number field; we use the usual notation for intervals in R, so that (a,b) is an open interval, (a,b] a halfopen interval, and so on.
PAGE 12
6 X and Y are Banach spaces over the scalar field Â§, which may be either the real or the complex field unless otherwise specified. x denotes the norm of an element X f X. X, is the unit ball of X, that is, X, = {x Â€ X: x ^ 1} . B(X,Y) is the Banach space of bounded linear transformations from X to Y, with the usual topology of uniform convergence. X* = B(X, Â§), and X_* is the unit ball of X*. Recall that by definition, x* = sup lx*x , ^1 while by the HahnBanach Theorem, lx = sup lx*x . ^1 Suppose that R and Z are families of subsets of a nonempty set S. R is a ring if R is closed under the formation of relative complements and finite unions. A ring R is a aring if R is closed un.der countable unions. S is an algebra if Z is a ring and Sen. An algebra Z is a aalgebra if Z is closed under countable unions. If S is any family of subsets of S, then a(S) denotes the smallest aalgebra containing g. If E Â€ R, then tt(E,R) denotes the family of all partitions P = {E.:l^i^n} of E whose sets belong to R. When R is understood we write tt(E) . If (A^) is a sequence of subsets of S which are pairwise disjoint, we call (A.) a disjoint sequence. Whenever we write A = ZA^ instead of A = UA. for some set A and sequence (A. ) , we mean that (A.) is a disjoint sequence, and say that A is the sum of the A. 's. Similarly we write A + B instead of A U B, if A B = 0.
PAGE 13
7 Suppose that R is a ring and jn:P * X is a set function, m is finitely additive if m(A + B) = m(A) + m(B) . m is countably additive if m(I'A.) = Em(A.) whenever (A.) is a sequence in R such that EA. Â€ R. We make the convention that whenever we write m(A) it is assumed that A "belongs to the domain of m. A finitely additive set function defined on a ring or algebra will be called a measure. If m:R Â• [0,00] is a set function, then m is monotone if m(A) ^ m(B) whenever A c B and A,B Â€ R. m is subadditive if m(A U B) s: m(A) + m{B) for every A,B Â€ R. m, ^ m means m, (E) < m(E) for every E e R. Standard references for results in measure and integration theory are Dinculeanu [12], Dunford and Schwartz [15], and Halmos [17] . We adhere to the notation of Dunford and Schwartz unless otherwise noted. 2. Properties of Set Functions . In this section we discuss some elementary properties of measures and their associated nonnegative set functions. R is a ring and m:R X is a measure unless otherwise noted. m is bounded if the set m(R) is a bounded subset of X, that is, sup_^m(E) I < 0= . The EfcR total variation v(m,E) of the measure m on a set E G R is defined by n v(m, E) = sup S lm(E. ) 1 . n(E) i=l ^ It is well known (see Dinculeanu [12]) that the set function v(m) is finitely additive but may not be finite even though m is a bounded countably additive measure. As Theorem 2.2 (ii)
PAGE 14
8 infra shov/s , v (m) is bounded whenever m is a bounded scalar measure. For this reason v (in) plays an important role in the theory of integration with respect to a scalar measure m, since it is a nonnegative measure which dominates m. When m is a countably additive vector measure and v (m, E) < Â«> for each E 6 ft, Dinculeanu [12] has shown that most of the results from the scalar integration theory carxy over to the vector case. In general, however, the total variation may be infinite and hence of little value. Thus we must use a more delicate device known as a control measure for m (see Theorem 33 infra ) . We now introduce a set function m called the variation of m. For every E 6 R we define m(E) = sup lm(F) I . FcE F6R It is obvious that m is bounded if and only if m is bounded. The following Lemma will be used several times in this dissertation. 2.1 Lemma . If a,,..., a are scalars, then n E la. 1 ^ 4 sup I Z a. I . i=l ACN^ iÂ€A In particular, if x^,...,x^ Â€ X and x* Â€ X*, then n E x*x. 1 ^ 4 sup I Z x*x. I ^ 4x* sup  E x^  , 1=1 ^ AcN^ iÂ€A AcN^ iÂ€A n and i S a.x. I ^ 4 sup  a .  sup  S x.]. 1=1 ^ ^ lii^n AcN iÂ€A
PAGE 15
Proof . Let A, , . . . , A^ denote the sets of integers i such that Re a, > 0, Re a. s; 0, Im a . > 0, and Im a . s respectively. Then n n n S a.i i: I iRea^l + Z llma^ i=l i=l i=l = E Re a . E Re a . + E Im a . E Im a . = Re E a^ Re E a^ + Im E a^ Im E a . ^1 ^2 ^3 ^4 ^ <: 4 sup I E a. I , AcN^ iÂ€A where we have used the elementary inequalities a i JRea] + lima] and JRea , j Im a  s  a for any a Â€ Â§. The inequalities concerning the x*x.'s are then immediate. Since n n n I E a.x I = Ix* E a.x. I ^ sup  a .  E  x*x .  i=l ^ ^ i=l ^ ^ isii^n ^ i=l ^ for some x* 6 X,*, the final inequality follows from the preceding one. D As a first application of this lemma we establish the properties of the variation of a vector measure. 2.2 Theorem . (i) in is nonnegative, monotone, and subadditive. (ii) If m is scalarvalued, then v (m) 5: 4m. (iii) m(E) I Â€. m(E) s sup v(x*m,E) i. 4m(E), for E 6 R. ^1 Proof . (i) fii is clearly nonnegative and monotone by definition. If E,F, and G belong to R, and G c E U F, then
PAGE 16
10 G = (G n E) + (G n (F\E) ) , so that I m (G) I = I m (G E) + m (G n (F\E) ) 1 s: iin(G n E) I + m(G n (f\E) ) ] s: fn(E) + ra(F) . Therefore m(E U F) s: ffi(E) + m(F) for every E,F Â€ R . (ii) If m is scalarvalued and (E.) Â€ tt(E), then n E lm(E.)l i 4 sup I T m(E.) n 4 sup m(F)I , i=l ACNÂ„ iÂ€A FcE Fea by Lemma 2.1. Thus v(m,E) s 4m (E) for every E e R. (iii) Suppose that E Â€ R. It is obvious that m(E)l i in(E) . Using (ii) we have m(E) = sup m(F) 1 FcE F6R = sup sup x*m(F) I X * FcE FÂ€R i sup v(x*m,E) ^1 ^ 4 sup sup x*m(F)  X * FcE FÂ€ft = 4m(E) . D 2.3 Remark . By the principle of uniform boundedness (see Dunford and Schwartz [15]) any function with values in X is bounded if and only if it is weakly bounded. In particular, m is bounded if and only if x*m is bounded for every x* G X* . If R is a aring and m is countably additive, then each x*m is also countably additive. It is known (see Halmos [17])
PAGE 17
11 that a countably additive scalar measure on a gring is bounded. Hence m is also bounded. The OrliczPettis Theorem (see Dunford and Schwartz [15]) states that, conversely, if x*m is countably additive for each x* ^ X* and R. is a aring, then m is countably additive. The hereditary ring )i generated by a ring R, is the family of all subsets of S which are contained in some element of R. If n:R [o,Â°Â»] is a monotone set function, then we can extend i to Ji as follows. Define \x^(E) = inf ^(F), F2E FeR for every E 6 W. It is immediate that \x is a monotone extension of ^, and yx s 0. If )a is subadditive then so is li . To see this, suppose that G,H ^ M and e > is fixed. Choose E,F Â€ R such that G c E, H c F, and n(E) < ji (G) + e, \l{F) < [1 (H) + e. Then G U H c E U F and so U^{G U H) S ^(E U F) i ^i(E) + ^ (F) < la"*" (G) + a"^ (H) +2e. We conclude that n (G U H) ^ ji (G) + u (H) . In this way we can extend fn to li. A set A Â€ )i is said to be an mnull set, or simply a null set when m is understood, if m(A) =0. We shall always identify the extension of m to !a by ffi unless there is a possibility of confusion. Note that A is a null set if and only if for every e > 0, there is an E ^ R, such that A c E and in(E) < e. 3. Strongly Bounded Measures . We now introduce a fundamental property of certain vector measures, and discuss some of its implications. R is a ring and m:R X is a measure unless otherwise noted.
PAGE 18
12 3.1 Definitions . (i) m is said to be strongly bounded (sbounded) if for every disjoint sequence (E.) in ii, liiii in(E. ) = 0. This property was introduced by Rickart [25], and later used by Brooks and Jewett [6] to establish convergence theorems for vector measures. Brooks [4] showed that sboundedness is equivalent to the existence of a control measure. This in turn provides necessary and sufficient conditions for the existence of an extension of a countably additive measure from S to o {Y.) when r. is an algebra. These results are presented in Theorems 3.3 and 3.4 infra. (ii) A series Tx in X is unconditionally convergent if for every e > there is a Â£i such that AHA = implies e e that Is X . 1 < e . i6A The following result states some basic properties of an sbounded measure. The proof is omitted. 3.2 Theorem . (Rickart [25]) (i) If m is sbounded then m is bounded. (ii) The following statements are equivalent: (a) m is sbounded. (b) For every disjoint sequence (E.) in R we have lim fn(E. ) = . (c) For every disjoint sequence (E.) in 9,, Em(E^) is unconditionally convergent.
PAGE 19
13 Suppose that IT* is a fsunily of measures on Ji. We say that the measures in tT\ are uniformly mcontinuous, or uniformly continuous with respect to m, if for every e > there is a 6 > such that u(A) j < e for every u f tTl if fh(A) < 6. When fU contains only one measure ^ we say that p is mcontinuous. [j, and m are mutually continuous if each is continuous with respect to the other. The most important characterization of sbounded measures is the following result due to Brooks [4] which we state without proof. 3.3 Theorem . (Brooks) m:R X is sbounded if and only if there is a bounded nonnegative measure m on R such that (i) m is mcontinuous . (ii) m s m on R. Moreover, m is countably additive if and only if m is countably additive. The measure m is called a control measure for m. (i) and (ii) show that m and m are mutually continuous; it follows that m and m have the same class of null sets. A fundamental problem in measure theory is to find conditions under which a countably additive measure on a ring R has a countably additive extension to the aring generated by R . Using the previous result of Brooks we can prove an extension theorem for countably additive sbounded measures on an algebra. This subject has been investigated by many authors. Takahashi [28] introduced a kind of boundedness
PAGE 20
14 condition less natural than, but equivalent to, the concept of sboundedness . His result in the case of an algebra is therefore included in the following theorem. A measure m:R Â— X is said to be Tbounded if for every e > there is an n ? N such that if E, , . . . , E are disjoint sets in R, then m(E.) < e for some i 6 N . To see that Tboundedness is equivalent to sboundedness, we remark that if m is not sbounded then there is a disjoint sequence (E.) and an e > such that m{E.) > Â£# i Â€ N, so m cannot be Tbounded. This shows that if m is Tbounded then m is sbounded. Conversely, suppose that m is sbounded with control measure m as in Theorem 3.3. Since m is a bounded nonnegative measure there is a constant K such that n E m(E.) ^ K for every finite disjoint family (E.) Â• If m 1=1 ^ is not Tbounded, then there is an e > such that for every n Â€ N there are disjoint sets E, Â„,..., E _ in R with m(E.^) >e, ^ in nn Â— m 1 s i S n. This clearly contradicts the boundedness of m, so m is Tbounded. Since m and m are mutually continuous, m is also Tbounded. We now state the extension theorem for measures on an algebra. 3.4 Theorem . (Brooks [5]) Suppose T. is an algebra and m:E X is countably additive and sbounded. Then m has a unique countably additive extension to a (D . Remark. Conversely, if m has a countably additive extension to a (Z) , then m is sbounded on Z.
PAGE 21
15 Proof . Let m be a control measure for m as in Theorem 3.3. Then m is countably additive and bounded on E . It is well known (see Halmos [17]) that a nonnegative, bounded, countably additive measure on an algebra has a unique extension to a bounded, countably additive measure on a (D = ^n Â• ^^ identify this extension by m. The symmetric difference of two sets A and B, denoted by A A B, is defined by A A B = (a\B) + (B\A) . We also have A A B = (A U B)\(A n B) . The elementary set relations (i) AACcAABUBAC, (ii) AUBACUDCAACUBAD, (iii) AnBACDDcAACUBAD, (iv) a\b a c\d caacubad, are straightforward to verify. Define d:i:, x Zi Â• [O,Â®) by d(A,B) = m(A A B) . Since m is sxibadditive, (i) shows that d satisfies the triangle inequality, and hence is a semimetric for E, . Moreover, from the mcontinuity of m on Z and from the inequalities m(A\B) s m(A A B) and m(A) m(B)  S m(A\B)  + m(B\A) 1 , it follows that m is a uniformly continuous mapping from Z into X . Since Z, is the smallest aalgebra containing Z, it is known that for every A G Z, there is a sequence (A ) in Z such that lim ni(A A A) = (see Halmos [17] ) . This fact shows that Z is a dense subset of Z, . If we identify sets A and B in Z, whenever d(A, B) = 0, then m is welldefined on Z, since m(A) = m(B) whenever d(A, B) = for A,B Â€ Z. Since X is complete, m has a unique extension to a continuous mapping from Z, into X. We denote this extension by m.
PAGE 22
16 To see that m is finitely additive on E, , we remark that (ii) , (iii) , and (iv) show that the mappings (A,B) AUB, A n B, and a\b respectively from T^^xy^, E, are continuous. Suppose A,B Â€ T.^ and A Ci B = 6. Let (A ) and (B ) be sequences i n n in E converging to A and B respectively. (ii) and (iv) imply that the mapping (A, B) A A B is also a continuous function, so lim d(A AB , A A B) = 0. By the continuity of m we have lim ni(A^ A B^) = m(A A B) = m(A + B) since A A B = A + B. But m(A A B ) =Â• m(A \b ) + m(B \a ) , and by the continuity . n n n n n n ^ ^ of (A, B) a\b we have lim m(A \B ) = m(A\B) = m(A) and ^' n n lim m(B \a ) = m(B) . Thus m(A + B) = m(A) + m(B) . Fina^ly, to see that m is countably additive on T. it suffices to show that if A."^ then lim m(A.) = 0. Since m is countably additive, lim in(A.) = 0. This in turn implies that lim m(A.) = since m is continuous on the space (Si/d) . Q We quote without proof a final result from the theory of vector measures. The following theorem is a generalization of the wellknown VitaliHahnSaks Theorem. 3.5 Theorem. (Brooks and Jewett [6]). Suppose R is a aring and V:fi? Â— [0,oo) is a bounded measure. If n :R, * X is a vcontinuous measure for each n, and if lim u (E) exists for ^n each E Â€ R/ then the n are uniformly vcontinuous . 4. Convergence in Measure and the Space M(m) . This section introduces the concept of convergence in measure for sequences of functions. Using this concept we can define the class of
PAGE 23
17 functions v/ith which the integration theory of Section 5 v/ill be concerned. From now on E will denote an algebra and m:S Â• X is a bounded measure. Then m is a nonnegative, monotone, subadditive, bounded set function, and, in fact, these are the only properties that are needed to obtain the results of this section. F(S) denotes the family of all scalar functions on S . If f Â€ F(S) we use the abbreviation [f > e] for the set {s e S:f(s) > e}. 4.1 Definition . A sequence (f, ) in F(S) is Cauchy in measure if for every e > we have lim m[lf.f.>e]=0. 1,300 Â» A sequence (f, ) in F(S) converges to f 6 F(S) in measure if for every e > we have lim m[f, fl > e] = 0. If (fK.) converges to f in measure we write m lim f, = f. The following lemma gives a useful equivalent formulation of convergence in measure. 4.2 Lemma . Suppose f Â» fi^ Â€ F(S), k e N. Then m lim f, = f if and only if there are sets A. 6 Z and a sequence of positive numbers e, converging to zero such that lim m(A, ) = and \\{s) f (s) 1 s Cj^ if s ;^ Aj^. Proof. Let (A. ) and (e, ) satisfy the stated properties, and suppose e,6 > are given. Choose n such that ei^ < Â€ and m (A^) < 6 if k s n. Since [j f^ fj > e] c [  f^ f  > e^] = A^ if k 5 n, we conclude that m[f, f >e]<6ifk2rn. Thus mlim fn^=f
PAGE 24
Conversely, suppose ra lim f,, f. For f 6 F(S), Dunford and Schwartz [15] define the quantity Ijfll = inf (e + m[f > e]), e>0 which is finite since in is bounded. To see that limH f, f jj = 0, suppose, to the contreiry, that limjf, fl > 0. Without loss of generality we may suppose that j! f, f > e for k Â€ N and some e > 0. There is an n such that ra[jf, f > e/2] < e/2 if k ^ n. Hence e < llfj^fll ^ e/2+m[lfj^f] > e/2] < e, which is a contradiction. Therefore limlf, fjl = 0. For each k C N choose e, > such that e, + fn[f f > e, ] < < f, f + 1/k, Then choose A, f S such that [f, f > e^] C A, and m(A, ) < fn[f, f > e, ] + 1/k. The sequences (A, ) and (e, ) evidently satisfy the properties stated in the lemma. Q If E c S, then XÂ„ denotes the indicator function of E, that is, the function whose value is 1 for s f E and for Â• S J^ E. We now define an important class of functions. 4.3 Definition . A Esimple function is any function f Â€ F(S) which can be represented in the form f = .^ ^i ^E,' 1=1 1 where a. Â€ $, and the sets E. 6 Z are disjoint and satisfy S = ZE. . When S is understood, f is called a simple function. Note that every simple function f has a canonical representation with E. = [f = a.]. 4.4 Definition . f Â€ F(S) is measurable if there is a sequence of simple functions converging to f in measure.
PAGE 25
19 For f 6 F(S) and E c S we define the oscillation of f on E, written 0(f,E), by 0(f,E) = sup lf(s) f(t) 1. s,tgE We have the following characterization of measurability . 4.5 Theorem . f is measurable if and only if for every e > there is a partition (E.) e tt(S) such that fti(E^) < e and 0(f,E.) < e for 2 s: i ^ n. Proof . Suppose (f, ) is a sequence of simple functions converging to f in measure. Let (A, ) and (e, ) satisfy the conditions of Lemma 4.2 for f and (f,^) Â• Fix e > 0, and choose k such that m(A, ) < e and e, < e/3. Suppose that n ^ Â• ^ f , = "Z a. Xp ; define a partition of S by setting F, = A, i=l i and F. = E._Aa, for 2:si^n + l. Ifs,tÂ€F. for some i > 1, then f(s) f(t) I i f(s) f3^(s)l + f3^(t) f(t)  s2e/3. Hence 0(f,F.) < e for 2 ^ i s: n + 1, so (F.) is the desired 1 Â• 1 partition. Conversely, if the conditions of the theorem hold then choose a sequence of partitions P, for e = 1/k. If P,^ = (E. ) / set f , = S f (s., )X , X ^^^ Ik E^^ where s^^ Â€ E., for ls:isn,. Ifs^^ E,, then f(s) f, (s) ^ max 0(f,E.) ^ 1/k. Since mCE^,) < 1/k / ^Jiisin, the sequences (E^) and (1/k) satisfy the conditions of Lemma 4.2, and so mlim f, = f. We conclude that f is measurable. G
PAGE 26
20 Let M(m) = M denote the space of all measurable functions in F(S). Since we have introduced a concept of convergence in F(S) we can discuss the topological properties of M. The following result from Dunford and Schwartz [15] lists some of the useful properties of the space M. We shall omit the proof. 4.6 Theorem . (Dunford and Schwartz) M is a closed linear algebra in F(S). If g:# Â— Â§ is continuous, then the mapping f * go f is a' continuous function from M into M. The following result will be very useful in connection with the Dominated Convergence Theorem. 4.7 Theorem . If (f, ) is a sequence of simple functions and m lim f, = f, then there is a sequence (g,) of simple functions such that mlim g, = f, and g,  i 2f on S, for each k Â€ N. Proof . Let (A. ) and (e, ) be sequences as in Lemma 4.2 for f and (f^) . Define <3^{s) = fj^(s) if s ;^ A^ and If^^Cs) I > 2e^, and gT^(s) = otherwise, for each k. Then if s ^ A, and f^(s) > 2e^, we have Igj^(s) f (s) j < e^^. If s ^ A^ and fj^(s) I i 2e^ then f(s) i If(s) fj^(s) 1 + \f^is) 1 ^ 3e^. so f(s) gT,(s) I ^ 3e^ if s ^ A, . By Lemma 4 . 2 we conclude that m lim g^ = f. If s Â€ A^ or if f^(s) ^ 2e^, then g, (s) = so certainly lg^(s) s; 2f(s). If s j^ A^ and I fj^(s) 1 > 2e^, then
PAGE 27
21 If(s) I ^ !fj^(s) If^(s) f(s) ^ lfk(s)l ^k ^ l/2lf3^(s)l = l/2lg^(s). Thus [g (s) I ^ 2] f (s) I for s Â€ S and k 6 N. D 5. A Lebesguetype Integral . We now present the standard results of integration theory. Many of these theorems were stated by Bartle [1] in his fundamental paper; they are included for completeness. Throughout this section E is an algebra and mrS * X is a bounded measure. If f is a simple function with representation n Z a. X , we define the integral of f over E Â€ E by i=l ^ ^i n r fdm = E a.m(E n E.) . The following result lists the basic properties of the integral for simple functions. The set function X (E) = J fdm is called the indefinite integral of f. 5.1 Theorem. (Bartle) (i) The integral of a simple function is independent of the function's representation. (ii) For a fixed set E Â£ T. the integral over E is a linear map from the linear space of simple functions into X. (iii) For a fixed simple function f the indefinite integral of f is a measure on Z. (iv) If f is a simple function bounded by a constant K on a set E, then IJgfdm] ^ 4Kffi(E) .
PAGE 28
22 (v) If U Â€ B(X,Y) and f is a simple function, then for every E Â€ Z, Jgfdum = U(Jgfdm) . Note that by (iv) the indefinite integral of a simple function is a bounded and mcontinuous measure. If m is sbounded, then so is the indefinite integral of a simple function. The Bartle integral is defined using sequences of simple functions. The next two theorems discuss the properties of these sequences which enable us to construct the general integral. Bartle proved only that (i) implies (ii) in Theorem 5.2. 5.2 Theorem . Suppose (f, ) is a sequence of simple functions that is Cauchy in measure. Then the following statements are equivalent: (i) The indefinite integrals of the f^^'s are uniformly mcontinuous . (ii) lim r f, dm exists uniformly for E Â€ S. Proof . Suppose (i) holds and e > is given. Choose 6 > such that E e S and rfi(E) < 6 implies that for each k Â€ N, IJgfdmj < e. There is an n Â€ N and sets A^. Â€ I such that m(A^.) < 5 and f^(s) f.(s) ^ e/4m(S), if s ;^ A^. and i,j 2 n. Then for E Â€ L and i,j ^ n, we have ^ [4e/4m(s)]m(E\A^.) + 2e < 3e,
PAGE 29
23 using (iii) and (iv) of Theorem 5..1. Since E was arbitrary, we conclude that the sequences (J f dm) are Cauchy uniformly for E Â€ S. Since X is complete, (ii) holds. Conversely, if (ii) holds and e > is given, then there is an n Â€ N such that for every E Â€ Z, provided k s n. Since each indefinite integral is mcontinuous, there is a 6 > such that if A Â€ S and m(A) < 6 then IJa^K^I < e, for 1 i k ^ n. If m(A) < 6 and k ^ n, we also have ^ 2e. Therefore the indefinite integrals of the f i^ ' s are uniformly mcontinuous . Q 5.3 Theorem . Suppose {f^^) and (g,^) are sequences of simple functions such that (i) mlim(f^ " "^k^ ^ Â°' (ii) The indefinite integrals of the f t^ ' s and the 9,^ ' s are uniformly mcontinuous. Then lim Jgf^^dm = lim J^g^dm uniformly for E Â€ Z. Proof . Since h, = f , g, is a simple function for each k, and since for every E 6 S, Theorem 5.2 implies that lim J hL dm exists for every E e Z. We need only show that this limit is zero
PAGE 30
24 uniformly in E. Fix e > 0. There is a 6 > such that if A e T. and fn(A) < 6, then for each k e N, IJaV^I < e. Since mlim h^, = 0, there is an n 6 N and sets A, 6 S such that m(A^) < 6 and hj^(s) j < e/4m{S) if s ^ A^, provided k ^ n. Thus for E 6 S and k s n, hdml Je\^^1 ^ IJe\A3^\^"^1 ^ IJeOA^^ S [4e/4m(S) ]m(E\Aj^) + e ^ 2e . That is, lim J h dm = uniformly for E Â€ S. G Following Bartle [1] we now define the general integral. 5.4 Definition . f G F(S) is integrable if there is a sequence (f, ) of simple functions such that (i) mlim f, = f. (ii) The indefinite^ integrals of the f .^ ' s are uniformly mcontinuous . If f is integrable then any sequence (f, ) of simple functions satisfying (i) and (ii) of Definition 5.4 is said to determine f. Theorem 5.2 shows that if (fn^) determines f, then lim Jpfi^dn^ exists uniformly in E. We denote this limit by the usual symbols Jgfdm or Jgf (s)m(ds) . Theorem 5.3 shows that the integral of f is independent of the determining sequence. Let L(m) denote the family of all functions that are integrable with respect to m. The following
PAGE 31
25 theorem lists some standard properties of the integral. We omit theproof. 5.5 Theorem . (i) For a fixed set E Â€ 2, the integral over E is a linear mapping from the linear space L(m) into X. (ii) For a fixed function f Â£ L (m) , the indefinite integral of f is an mcontinuous measure on E. Suppose f G M and there is a null set A such that sup y f(s) I < CO. Then we say that f is essentially bounded, and define the essential supremum f of f by llfll = inf sup f(s) 1 . fli(A)=0 S;^A The following standard result, whose proof refines that of Bartle, shows that every measurable and essentially bounded function is integrable. 5.6 Theorem . If f G M is essentially bounded, then f 6 L(m) and for every E Â€ E we have . IJjjfdml ^ 4lfllfn(E). Proof . Let K = Hfjl, and suppose (f, ) is a sequence of simple functions converging to f in measure. Define gv(s) = " f^(s) fj^(s) I ^ K, ^ (K/lf3^(s))f^(s) f3^(s)! > K, for each s Â€ S and k Â€ N. Then (g, ) is a sequence of simple functions , and 93^(3) f^(s) I =< fj^(s)l K Ifj^Cs)! > K fv(s) I ^ K,
PAGE 32
26 since \g^(s) fj^(s) = j f^^ (s) (lK/j f^ (s)  )  = f^(s)lK if f, (s) > K. Suppose e > 0. Since f is essentially bounded by K, there is a null set A such that f(s)l ^ K + e/4 if s ^ A. Now if s ^ A and Ifj^(s) j > K + e/2, then since f (s) I ^ K + e/4, we have fj^(s) f{s) 1 ^ !fj^(s)l f(s)l > K + e/2 K e/4 = e/4 . Therefore [k^ " f^l > e/2] = [f^ > K + e/2] c A U [fkf > e/4]. Finally, [gj^f > e] c [Igj^fj^l >3e/4] u [fj^f > e/4] c A U [f^ f > e/4]. Since A is null and mlim f, = f , we conclude that mlim g, = f. Since the g,^ ' s are uniformly bounded by K, Theorem 5.1 (iv) implies that for every k, IJ^g^dml ^ 4Kifi(E) for any E Â£ T,. This shows that the indefinite integrals of the g^^ ' s are uniformly mcontinuous . It follows that f Â€ L(m), and that for every E Â€ S, IJgfdml = limlj^gj^dml ^ 4Km(E) . n Using Theorem 5.6 and the mcontinuity of the indefinite integral of a function f 6 L (m) we have the following Corollary 5.7 Corollary . If f Â€ L (m) , then the indefinite integral of f is a bounded measure on E.
PAGE 33
27 Proof . Since f is measurable, f is bounded except possibly on sets of arbitrarily small variation. Choose 6 > such that if m(E) < 6, then IJgfdml < 1. Choose E Â€ E such that m(E) < 6 and f is bounded on S\E, say by K. For any A Â€ Z, therefore i 4Km(S) + 1. G When E is a aalgebra and m is a countably additive vector measure, Bartle, Dunford and Schwartz [2] have shown that the integral satisfies the following property. If f 6 L(m) and U Â€ B{X,Y), then f Â€ L (Um) and for each E 6 E, JgfdUm = U(Jgfdm). We now extend this result to the case where m is a bounded measure on an algebra E. 5.8 Theorem . Suppose that U Â€ B(X,Y) . If f Â€ L(m) then f Â€ L(Um) , and for every E Â€ Z we have JgfdUm = U(Jgfdm) . Proof . Since Um(E)  ^ lu  m(E)  ^. it follows at once that Um(E) <: um(E) for every E Â€ Z. From this inequality we see that if mlim f, = f, then Um lim f, = f as well. If f Â€ L(m) and (f, ) is a sequence of simple functions determining f, then we need only show that the indefinite integrals of the fi^'s with respect to Um are uniformly Umcontinuous . Since we have
PAGE 34
28 using Theorem 5.1 (v) , this follows immediately from Theorem 5.2 and the fact that lim J* f,dm exists uniformly for E e >:. a When m is a scalar measure, Dunford and Schwartz [15] have introduced the following definition of integr ability . f Â£ F(S) is integrable if there is a sequence (f, ) of simple functions converging to f in measure, such that lim Jgf^ f. Idv(m) = 0. i,j*co We shall show that the Bartle concept of integration and that of Dunford and Schwartz coincide when m is a bounded scalar measure. We state the following result without proof. 5.9 Theorem . (Dunford and Schwartz [15]) If m is a scalar measure and f is integrable (in the sense of Dunford and Schwartz) , then the total variation of the indefinite integral X(E) = Jgfdni is given by v(X,E) = Jgf(s)lv(m,ds) . Using Theorem 5 . 9 we now prove the equivalence of the Bartle and the Dunford and Schwartz integration theories in the case of a scalar measure. 5.10 Theorem . Suppose m is a bounded scalar measure. Then f e L(m) if and only if f is integrable in the sense of Dunford and Schwartz. In this case the same sequence of simple functions determines f for both definitions; consequently the two integrals coincide.
PAGE 35
29 Proof . Suppose (f^) is a sequence of simple functions. By Theorems 5.9 and 2.2 (ii) v/e have ^ 4 sup j J (f. f.)dm FcE ^ for every E Â€ Z. Suppose f f L(m) and (f, ) is a sequence determining f. Then by Theorem 5.2 the righthand term in the inequality goes to zero uniformly for E C E as i/j Â• 00. Hence (*) lim J*gf^ f.ldv(m) = 0, i,joo and so f is integrable in the sense of Dunford and Schwartz. Conversely, if m lim f , = f and (*) holds, then the lefthand term of the inequality converges to zero uniformly for EÂ€Easi,jÂ»oo. This shows that the sequences (J f,dm) are Cauchy uniformly in E, and hence by the completeness of X, they converge uniformly in E. By Theorem 5 . 2 we conclude that the indefinite integrals of the fi^' s are uniformly mcontinuous, so f Â€ L(m) . In either case, the same sequence of simple functions determines f for both definitions, so the two integrals coincide. Q We now prove some results concerning the domination of functions by integrable functions. 5.11 Lemma . Suppose (f, ) is a sequence of integrable functions, and g Â€ L(m) . If f  ^ g on S for each k, then J*j,fj^dml ^ 4 sup IJpgdm] FcE FÂ€S
PAGE 36
30 for each E 6 Z; consequently the indefinite integrals of the f, ' s are uniformly mcontinuous . Proof . We use Theorem 5.8 witli U = x* Â£ X*, and also Theorems 5.10, 5.9, and 2.2 (ii) to compute ^1 ^ sup r I f dv(x*m) ^ sup rÂ„ I g I dv (x*m) X * ^ ^1 i 4 sup sup I J gdx*m FGE X * = 4 sup I r gdrnj, FCE ^ FGE for any k Â€ N. Since the indefinite integral of g is mcontinuous, it follows that the indefinite integrals of the f. 's are uniformly mcontinuous. G 5.12 Theorem . If f Â€ M, g Â€ L (m) , and ] f 1 ^ [g] on S , then f Â€ L (m) . Proof . Since f G M there is a sequence (f, ) of simple functions converging to f in measure. By Theorem 4.7 we may assume that f, I ^ 2fl on S for every k. Then f^ ^ 2g on S, and 2g Â€ L(m) . By Lemma 5.11 we conclude that the indefinite integrals of the fi^'s are uniformly mcontinuous. Therefore f Â€ L(m) . a 5.13 Corollary . Suppose f e M. Then f Â€ L(m) if and only if I f I Â€ L(m) . If (f,) is a sequence of simple functions deter
PAGE 37
31 mining f, then (] f ,  ) is a sequence of simple functions determining 1 f j . Proof . The first statement follows from Theorem 5.12. If f, is a simple function, then so is  f > I Â» ^^'^ ^Y Lemma 5.11, Jgf^dm ^ 4 sup IJpfj^dmj. FcE FÂ€Z Thus the indefinite integrals of the ] f ,  ' s are uniformly mcontinuous , if the indefinite integrals of the f i^ ' s are. Finally, since   f,   f   ^  f, f  , it follows that [1 Ifj^l fl I > e] c [If^fl > e] . If (f^^) determines f, we conclude that (f, ) determines f. n 6. Convergence Theorems . In this section we prove stronger forms of the convergence theorems for sequences of integrals stated by Bartle [1] . Suppose that f and (f ) are integrable functions. We say that f and (f, ) satisfy property B if lim Jgf^dm = Jgfdm uniformly for E Â€ S. The main convergence theorem resembles the classical Vitali convergence theorem. 6.1 Theorem . Suppose f Â€ F(S) and (f,) c L(m). Then f e L (m) and f, (f, ) satisfy property B if and only if (i) m lim f, = f . (ii) The indefinite integrals of the fi^'s are uniformly mcontinuous .
PAGE 38
32 Proof . Suppose f Â€ L (m) and f/Cfi^) satisfy property B. To prove that m lim f , = f we proceed as follows. By Theorem 5.13, g^ =  f f ^^  ^ L (m) for each k 6 N. Let (h^.) be a sequence of simple functions determining g as in Definition 5.4. By Theorem 5.13, (  h, .  ) is a sequence of nonnegative simple functions determining g  = g, , so without loss of generality we may assume that h, . ^ for each j 6 N. By Theorem 5 . 8 we have g, 6 L(x*m) for every x* Â€ X*, hence by Theorem 5.10 g, is integrable in the sense of Dunford and Schwartz. By Theorems 5.9 and 2.2 (ii) we have S^g^dv{K*m) = ;gf3^fdv(x*m) ^ 4 sup lj'p(f3^f)dx*ml FCE F6E ^ 4 sup Jp(fj^f)dm, Fes if X* Â€ X^*Â« Since f and (f, ) satisfy property B, it follows that given e > 0, there is an n such that if k ^ n and X* Â€ X,*, then for every E Â€ S, jÂ£gj^dv(x*m) < e. Since (h, . ) determines g, , it follows in the same manner that for a fixed k s n, there is an t such that j ^ t implies that for every E Â€ E, Je1\j " gkldv(x*m) < Â€, provided x* Â€ X,*. By omitting the first ll of the functions h, . if necessary, we therefore conclude that for x* 6 X,*, E Â€ Z, and k k n.
PAGE 39
33 .dv(x*m) < 2e . Since h, . is a nonnegative simple function the sets E, . = [h, . > y] belong to E for every y > 0. Moreover, Yv(x*m,F^.) 5 Jgh^.dv(x*m) < 2e, so v(x*m,Ej^j) ^ 2e/y, if X* C X * . By Theorem 2.2 (iii) we conclude that m (Ej^.) ^ sup v(x*m,Ej^.) ^ 2e/Y X^* Since mlim h, . = g, , there is an i Â€ N and a set F, Â€ S such that fn(F^) < e/y and l^j^^^^ ~ ^k^^^ ' s: Y if s ;^ F^. Therefore, if s j^ R . U F, , we have f,^(s) f(s) I = g^(s) ^ gj^(s) hj^^Cs)! + \i(s)l ^ 2y , and in(E, . U F, ) ^ 3e/Y. Now if 6, and ft > are given, then choose y > such that 2y < 6, Â• Then choose e > so that 3e/Y < bjIt follows from the arguments above that there is an n 6 N and sets G, such that if k s n, then m(G,) < 5~ and f, (s) f(s)l 5 6, if s j^ G, . We conclude that m lim f, = f . k To see that (ii) holds we note that the argument in the proof of (ii) implies (i) for Theorem 5.2 applies. Conversely, if (i) and (ii) hold, then for each k there is a simple function g, such that m[lg, f, 1 > 1/k] < 1/k and for every E Â€ E,
PAGE 40
34 As in the proof of Theorem 4.6, it follows that m lim g,^ = f. Suppose e > 0. Choose n such that 1/n < s. Then choose 6 > such that fn(E) < 6 implies ljj.^3^dnil < e and IJ^f^drnj < e, for 1 ^ k ^ n and i Â€ N. Then if m(E) < 6, i 2e for every k, so the indefinite integrals of the g,^ ' s are uniformly mcontinuous . It follows that f Â€ L(m). Since llE^k^lE^^"^! ^ llE^^k gk)din +llE(gkf)dm ^ lA + IJ^^^k ^)^l' we conclude that f and (f, ) satisfy property B. O We now state two important corollaries to Theorem 6.1 which are stronger forms of the results stated by Bartle [1] . 6.2 Corollary . (Dominated Convergence) Suppose (fi^) c L (m) and g Â€ L(m) . If ] f,  ^ jg] on S for every k Â€ N, then m lim f , = f if and only if f Â€ L(m) and f, (f, ) satisfy property B. Proof. Suppose m lim f = f. By Lemma 5.11 the indefinite integrals of the fn^'s are uniformly mcontinuous, so the conclusion follows from Theorem 6.1. The converse is immediate by this same theorem. Q 6.3 Corollary . (Bounded Convergence) Suppose (f, ) c M is a sequence of functions uniformly bounded on S. Then mlimf, = f if and only if f Â€ L(m) and f, (f, ) satisfy property B.
PAGE 41
Proof . Since constant functions are integrable, the conclusion follows from Corollary 6.2. G 7, Lp Spaces . As an application cf the previous theory, we present a portion of L space theory. In this section Z is an algebra and m:I X is a bounded measure. Recall the following set of inequalities, which were used in the proof of Lemma 5.11. IffÂ€L(m) and E ^ T., then IJgfdml = sup IJgfdx ^m ^1* ^ sup rÂ„ f i dv (x*m) Â£. 4 sup sup IJ fdx*m FCE X^* = 4 sup I Jpfdml . FCE These terms are finite by Corollary 5.7. Let f Â€ M. If 1 s; p < CO and jfj^ Â€ L(m), define llfH = sup [f Jf!Pdv(x*m)]l/'P . Let L denote the space of all functions f 6 M such that lf^ Â€ L(m). Since F c E implies that IJpfdml s sup Jgfldv(x*m), the inequalities above show that the two expressions s E6 up [;glfPdm]l/P ^^^ ^^p fj^,fP^^(^*^)jl/p ^1
PAGE 42
36 are equivalent "norms" for the spaces L . We prefer to use the second of these expressions, since the propertJ.es of the scalar norms in L (x*m) are then available. It By Corollary 5.13, it follows that L (m) = L, . Moreover, since 1 ^ p ^ q < <Â» implies that f ^ 1 + [fj / we see that f G LÂ° implies that f ^ L , by Theorem 5.12. Hence 1 p q From the elementary inequality [a + b]^ < 2^(a^+b^), it follows that jf + gj^ < 2^(1 fP + Igl^); if f and g are in if, then so is f + g, by Theorem 5.12. L is clearly closed under scalar multiplication, and hence is a linear space. It is easy to see that llafH = allfll for any ir ir scalar a. Moreover, llf + gl! = sup[Jgf + glPdv(x*m)]^'^P ^1 ^ s up[(J lflPdv(x*m))^/P + (Jggl%v(x*m))^/P] ^1* ^ llfllp + llgllp. since llf = if f = 0, we conclude that 1 Â• ! is a semip P tO norm on L . P Suppose 1 < p < o= and 1/p + 1/q = 1. The Holder inequality shows that fg Ifl^ Igl"^ Ifllpllgllg pIUII^ . qllgll^ if f e L and q f L . By Theorem 5.12 we conclude that P q fg f LÂ° . If x* 6 X, we have
PAGE 43
37 Jgfgldv(x*m) < llfllpllgllqd/p + 1/q). Hence Hfgll^^ < llf llp!!gllq. A sequence of functions (f, ) in L is said to converge K p in LÂ° to a function f Â€ LÂ° if we have limllf, f j] =0. The p p k "p following convergence theorem is a consequence of Theorem 6.1 7.1 Th eorem . Suppose 1 < p < <Â» and (f, ) is a sequence in L . K p Then f g LÂ° and limllf, f 1! =0 if and only if p " K "p (i) m lim f, = f . (ii) The indefinite integrals of the f, ^'s are uniformly mcontinuous . Proof . Suppose f ^ LÂ° and liraf, f\\ =0. If we set g = I f 1^ f]^ Â€ L(m) , then the first part of the proof of Theorem 6.1 is easily adapted to show that m lim f. = f. Since the indefinite integral of f is mcontinuous, and 11Ve"p ^ l^^k " ^"p "^ II^eI'p' ^"^^ follows as in the proof of Lemma 5.2. Conversely, suppose that (i) and (ii) hold. By Theorem 4.6, 1 f 1. 1 converges to lf" in measure, so by Theorem 6.1, f Â€ L . Moreover, Theorem 4.6 implies that m lim f, f  ^ = Since fj^fP ^ 2^(1 fj^l^ + lf^), we have IJ^Ifj^flPdml ^ 2P"2 ^^p J^(lf^lP+flP)dml, FCK Fsr by Lemma 5.11. Since the indefinite integrals of the f, "'s are uniformly mcontinuous, we conclude that the indefinite integrals of the jf, fj^'s are also uniformly mcontinuous.
PAGE 44
38 By Theorem 6.1, then. lim ;^f^ flPdm = E' k uniformly for E Â€ E. Finally, since llf^^ fljP^ 4 sup \^^\f^ f^dml, EÂ€ S we have limjlf, f  = . D The inequalities at the beginning of this section show that convergence in L, is equivalent to property B. For if f, (f, ) belong to L? = L(m), then suplJgCf^^ f)dm ^ llf^ f^ ^ 4 supJ^(f^f)dml . E$ Zj Ec S Just as Corollary 6.2 followed from Theorem 6.1, we have; 7.2 Corollary . Suppose 1 ^ p < Â», (f, ) c L , g Â€ L , and K p p jf I <. g on S for each k. Then mlim f, = f if and only if f Â€ LÂ° and limHf^ f\\^ =0. Proof . Since jfi.!^ ^ 1 9,1 ^' thÂ® indefinite integrals of the f, 1^'s are uniformly mcontinuous . The conclusions follow from Theorem 7.1. G 7.3 Corollary . Suppose 1 ^ p < Â». The subspace of simple functions is dense in L . P Proof . Suppose f Â€ LÂ° . Since f Â€ M there is a sequence (f^) IT of simple functions converging to f in measure, and by Theorem 4.7 we may assume that j f ,  ^ 2fl on S for each k. By Corollary 7 . 2 it follows that limjjf f]! = 0. Q
PAGE 45
39 8. Integration by Countably Additive Measures . Integration of scalar functions by a countably additive measure m: Z Â» X, where r is a aalgebra, has been studied extensively by Bartle, Dunford and Schwartz [2], and also by Lewis [20], The Bartle, Dunford and Schwartz Theory uses a result concerning the equivalence of weak compactness in the space of scalar measures with the existence of a control measure (as in Theorem 3.3), and the VitaliHahnSaks Theorem. By means of these powerful results the following definition of the integral yields a theory which coincides with that in Section 5 for countably additive measures . A sequence of functions (f, ) c F(S) is said to converge (pointwise) almost everywhere to a function f if there is a null set A such that lim fi^(s) = f(s) for every s j^ A. In this case we write ae lim f, = f . 8.1 Definition . f Â€ F (S) is integrable if there is a sequence (f, ) of simple functions such that ae lim f, = f, and such that for every EST, the sequence (pEfv^"') converges in X. Pointwise almost everywhere convergence replaces convergence in measure in the countably additive situation by virtue of the nonnegative control measure m. Since m and m are mutually continuous and countably additive, the standard theorems on the relation between convergence in measure and almost everywhere convergence can be shown to hold. In particular, ae lim f. = f implies m lim f, = f. Moreover, since the indefinite integrals of the function f, in Definition 8.1 are mcontinuous me. ures , the quirement that
PAGE 46
40 (f f, dm) converges for each EST, together with the VitaliHahnSaks Theorem, show that the indefinite integrals of the f , ' s are uniformly mcontinuous , as in Definition 5.4. 8.2 Remark . If f is finitely additive, T is a aalgebra, and m is sbounded, so that there is a control measure m by Theorem 3.3, then we may apply Theorem 3.5. Thus if ( r f, dm) converges for every E, we conclude that the indefinite integrals of the f^'s are uniformly mcontinuous . Convergence in measure must be retained however. Bartle, Dunford and Schwartz establish a onesided form of Theorem 6.1: If (f,) C L(m), ae lim f^ = f, and the indefinite integrals of the f j, ' s slto uniformly mcontinuous, then f 6 L(m) and lim J^^k^"^ ~ Ie'^^"^ ^'^^ E Â€ FIn view of Theorem 6.1 and the VitaliHahnSaks Theorem we have the following extension: 8.3 Theorem. Suppose (f^) c L(m) and ae lim f^ = f. Then f 6 L(m) and f, (f, ) satisfy property B if and only if lim r f, dm exists for each E Â€ T. In the case of convergence in L we have by Theorem 7.1: 8.4 Theorem . Suppose 1 ^ p < <=, (f,) C L , and ae lim \= '^^ Then f Â€ LÂ° and limllf^^ fH = if and only if lim JEif]^.^Â«^ exists for every E ^ T,. For future use we state the following useful approximation theorem for the countably additive case.
PAGE 47
41 8.5 Theorem . (Dunford and Schwartz [15]) Suppose 1 < p < Â», r is an algebra, and Z = o {Z ) . Then the space of E^o o o simple functions is dense in L . Lewis [20] has developed an integration theory for countably additive measures with values in a locally convex space X using the Pettis approach of employing linear functionals . This integration theory coincides with that presented above when X is a Banach space. 9. A Riemanntype Integral . Gould [16] introduced an integral for scalar functions, which in the case of bounded functions is defined as a limit of RiemannStieltjestype sums. Although he calls his integral an improper integral, it is in fact equivalent to the "proper" Bartle integral when m is a bounded measure. This alternative approach to integration has been extended by McShane [22] to obtain a very general theory. We identify the Gould and the Bartle integrals as the (G) and the (B) integrals respectively. In this section E is an algebra and m:i: X is a bounded measure. If P and P' are two partitions in tt(S), we write P' ^ P to mean that each set in P' is a subset of a set in P; we say that P' is a refinement of P. If P 6 tt(S) and P = [E, ,...,E ], then for a function f Â€ F(S) we introduce the symbol S(f,P) to indicate any RiemannStieltjestype sum n Z f(s.)m(E.), where s. Â€E., l^i^n. Let B(S,E) = B(S) denote the family of all bounded measurable functions on S.
PAGE 48
42 ^Â•1 Definition . Suppose f Â€ B(S).. f is integrable (G) if for any e > there is a P 6 Tr(S) such that if p' $ tt(S) and P' s P, then s(f,P) S(f,P') < e. We omit the proof of the following Theorem. 9.2 Theorem . Suppose f Â€ B(S) . f is integrable (G) if and only if there is an x Â€ X with the following property: For every c > there is a P Â€ tt(S) such that if P' ^ tt(S) and P' 2: P, then ls(f,P')x 0. Choose k 6 N and A 6 E such that m(A) < e/4K and I f(s) f^(s) I < e/4m(S) if s j^ A. Suppose f^ = ^g.^ a^ X^ . i Let F, = A and F. = E. Aa, 2 ^ i ^ I + 1. Then P = (F.) is 1 1 11 1 a partition of S and f, is constant over FÂ„,...,F , . If P' = (B.) ^ P and if f, (s) = b. for s 6 B., then S(f,P') Jgf^dml = _E f(s^)m(B^) _E Jg fj^dm s I I (f(s.) b.)m(B.) I + I Z (f(s.)b )m(B.) B^nA=0 111 B_^g;^ 1 1 X s: 4[c/4m(S) ]m(s\A) + 4Kfn(A) + 8Km(A) ^ 4e,
PAGE 49
43 where we have used Lemma 2.1. ,; Since f,  < 2fl, Corollary 6.3 shows that f and (f,) satisfy property B. There is then a k Â€ N such that j r f , dm xl < e. The argument above shows that there is a partition P 6 tt(S) such that if P' ^ P then js(f,p') x s5e. We conclude that f is integrable (G) and that the (G) and (B) integrals coincide. Q. The vector x will now be denoted by f fdm. If E Â€ S and f ^ B(S) , then fXÂ„ is bounded and measurable by Theorem 4.6, and we write Je^^"^ = Js^^E^ Â• Suppose f 6 M(m) . The definition of measurability implies that f is bounded except perhaps on sets of arbitrarily small variation. Let B(f) denote the family of all sets in H on which m is bounded. It is easy to see that B(f) is a ring of sets. Since f is integrable on each set in B(f), the class 1(f) = [J fdm:E 6 B(f)} is nonvoid. Moreover, 1(f) is a net in X over the directed set B(f), where E ^ F means E c F. With this motivation we can define the integral for unbounded measurable functions. 9.4 Definition . f 6 M is integrable (G) if for every s > there is an E 6 B(f) such that if F Â€ B(f) and E n F = 0, then jj fdm < e. In view of the fact that 1(f) is a net we have the following result due to Gould [16] .
PAGE 50
44 9.5 Theorem . (Gould) f Â€ M is integrable (G) if and only if there exists an x $ X having the follov/ing property: For every e > there is an E 6 B(f) such that if F e B(f) and E c F, then  J f dm xj < cProof . If such an x exists, then since B(f) is a ring and the indefinite integral of f is a measure on B(f), we have IJpfdnil ^ IJp+E^dni x + jj^fdm xj , for every E,F Â€ B(f) such that E fl F = 0. Fix . > and choose E Â€ B(f) such that if G Â€ B(f) and E c G, then I r fdm x < e/2. If F Â€ B(f) and E F = 0, then E c E + F 6 B(f), and so IJpfdmj < e. Conversely, suppose f is integrable (G) . Definition 9.4 implies that 1(f) is a Cauchy net in X. For suppose e > is given. Choose E Â€ B(f) such that if F Â€ B(f) and EH F = 0, then I r fdm < e. Now if F,G 6 B(f) and F,G ^ E, we have IJf^^ Jg^^I ^ UfXe^^"^! " IJgXe^'^I s 2e, where we have again used the additivity of the integral. Let X denote the limit of 1(f) . It is then immediate that x satisfies the requirement of the theorem. D From this result we see that when f is integrable (G) , then r_fdm = lim f^fdm . ^2 B(f) '^^
PAGE 51
45 Suppose now that f Â€ L (m) . To see that f is integrable (G) , suppose e > 0, and choose 6 > such that if in (A) < 6, then 1 r fdm] < e. Since f is measurable, there is a set E Â€ E such that f is bounded on E and fn(S\E) < 6. It follows that for any A ^ Y. such that A n E = 0, m(A) < 6, and so ] f fdm] < e. By Definition 9.4, f is integrable (G) . To see that integrability (G) implies integrability (B) we need the following result established by Gould [16] . 9.6 Theorem . (Gould) If f is integrable (G) , then the indefinite integral of f is a bounded mcontinuous measure on Z. Proof . Since 1(f) is a Cauchy net it is a bounded set, so the indefinite integral is bounded. Additivity follows from the additivity of the indefinite integral for bounded functions. To see that the indefinite integral is mcontinuous, suppose e > 0. There is an E Â€ B(f) such that if F Â€ B(f) and E n F = 0, then IJ fdm < e. Hence for A6SÂ»AnE=0 implies that jf fdm ^ e. On E, however, the indefinite integral is already mcontinuous by Theorem 5.6, so there is a 6 > such that if ffi(A) < 6 then l/AnE"^^"^' ^ Â®* '^^^ ^^ fft(A) < 6 , then We now prove a result that is the counterpart of Lemma 5.11.
PAGE 52
46 9.7 Lemma . If f,g g M, jfj < g on S, and g is integrable (G) , then for every A Â€ S/ (*) !J;^fdml ^ 4 sup IJ^gdml . BCA B6S Thus f is integrable (G) . Proof. Suppose E Â£ B(f) and e > 0. The indefinite integral of f over E is an mcontinuous measure by Theorem 5.5, so there is a 6 > such that if m{A) < 5, then I f fdmj < e. Â•J ATI E Since g is bounded except on sets of arbitrarily small variation, there is a set F ^ B(g) such that m(S\F) < 6. Since f and g are both bounded on E H F, we have IJ^^pfdml ^ 4 ^sup^ IJ^gdml i 4 sup I f^gdn^l Â» BcE Â•'^ by Lemma 5.11. Therefore, IJe^^I ^ IJeXf^^^I ^ IJehf^^"^! ^ e + 4 sup I rÂ„gdm . BcE "^^ Since e was arbitrary, we see that (*) holds for every A 6 B(f), and hence for every A Â€ S. To see that (*) implies f is integrable (G) , suppose e > 0. Choose E Â€ B(g) such that if F Â€ B(g) and E n F = jZ5, then I J* gdml < e. For any B ^ T. with B fl E = 0, we have I r gdm s e. Since B(g) c B(f), we conclude by (*) that if A Â€ B(f) and A n E = 0, then I r,fdml ^ 4 sup I rÂ„gdm rs 4e. G iJa g^^ Ub B6S
PAGE 53
47 Suppose now that f is integrable (G) . Since f C M, there is a sequence (f, ) of simple functions converging to f in measure, and by Theorem 4 . 7 we may assume that f, 1 < 2f on S. Since 2f is integrable (G) , Lemma 9.7 shows that the indefinite integrals of the fi^'s are uniformly mcontinuous, since by Theorem 9.6 the integral of f is mcontinuous . By Definition 5.4, f g L(m). To see that the (G) and (B) integrals are equal, note that the definition of the (G) integral as the limit of I (f ) shows that there is an increasing sequence of sets E, Â€ B(f) satisfying (G) Jgfdni = lim J^n^dm. Since the functions fTÂ„ are measurable by Theorem 4.6, and since we can choose the E, 's so that m(S\E. ) < 1/k, it follows that m lim fXÂ„ = f. Since f f L(m) and I fX_ I ^ I f ! on S for each k. Corollary 6.2 implies that (B) Jgfdm = lim Jg^^E. '^Â• The integrals therefore coincide.
PAGE 54
CHAPTER II Stieltjes Measures and Integrals In this chapter we specialize the integration theory of Chapter I to the case where the measure is generated by a vectorvalued function on an interval of the real line. We shall call such measures Stieltjes measures, and the resulting integrals will be called Stieltjes integrals. By restricting the Banach space in which the vector function takes its values to be an L space over a probability measure space, we shall study stochastic integration. (Chapter III) . In the present chapter we shall investigate the question of boundedness of a Stieltjes measure, and the question of the existence of an extension to a countably additive measure on the Borel sets. Finally we discuss the important special case of Stieltjes measures with values in a Hilbert space. 1. Notation . Let T denote a closed, bounded interval in R, which for convenience we shall assiome is the unit interval [0,1]. Let S denote the family of all subintervals of T of the form [a,b) or [a,l], together with 0, where we shall always assvime that s: a ^ b ^ 1 . Finally, let E denote the family of all subsets of T which are of the form A=i:a., where (A.) is a finite disjoint family in g. Note that Z 48
PAGE 55
49 is an algebra of subsets of T, and that the aalgebra generated by T is the family of Borel sets in T, which is denoted by E, Â• Suppose that X is a Banach space, and that z:T . X is a function. Define a set function m:g X by setting m[a,b) = z(b) z(a) and in[a,l] = z(l) z(a). If A 6 Â§ and A = y'h, where each A^ Â€ Â§/ then it follows at once by the definition of m on Â§ that m(A) = i;m(A. ). Suppose n k ^ now that A e T., and that Z A. and S B. are any two rei=l ^ j=l ^ presentations of A by finite svims of sets in g. Then the sets C. . = A. n B. belong to S for each i and j, and we have A. = Z C. ., 1 :S i s n, and B. = Z C^., 1 s: j :Â£ k . By the preceding remark, it follows that m(A^) = T. m(C^.), and m(B.) = E m(C^ .) , for each i and j . We therefore conclude that r m(A^) = Y. T. m(C^.) = Y. m(B.) . i i j j ^ We have just shown that if A Â€ S and A = SA. is any representation of A by a finite sum of sets in Â§, then we can define m(A) = S m(A. ) / and this definition in independent of the representation of A. It is now immediate that m is finitely additive on S. In the sequel, we shall associate with each function z:T Â« X the unique measure m defined as above on S. When we wish to emphasize that the measure is generated by z, we write m . We call such a measure m the Stieltjes measure generated by z.
PAGE 56
50 If m is a bounded Stieltjes measure on Z, then m determines a class L(m) of functions integrable in the sense of Chapter I. The Ssimple functions will now be called step functions; the class of step functions coincides with the class of all Ssimple functions. Every function that is the uniform limit of a sequence of step functions is obviously bounded and measurable with respect to any Stieltjes measure on S. Hence by Theorem 1.5.6, such functions are integrable, and the sequence of integrals of the step functions converges to the integral of the limit function. If m has a countably additive extension to S, , then Theorem 1.8.5 shows that the class of step functions is still dense in L(m) . 2. Bounded Stieltjes Measures . In this section z:T Â» X is a function generating the Stieltjes measure m:i; X. We investigate conditions on z which insure that m is a bounded measure. By Remark 1.2.3, m is bounded if and only if x*m is bounded for each x* Â€ X*. Now since x*m [a,b) = x*z(b) x*z(a) = m . [a,b) , for every interval z x*z [a,b) (or [a,l]) in Â§, it follows that x*m = m for every X* 6 X*. By Theorem 1.2.2 (ii) we know that the scalar measure m . is bounded if and only if it has bounded total x*z ^ variation. Recalling that each set in Z is a finite sum of sets in Â§, it follows that n n sup S m(A.) = sup S lm(A.), (A^)cS i=l ^ (Ai)cE i=l where each (A.) is a finite disjoint family. That is, the total variation of any measure m on E is the same as the
PAGE 57
51 total variation of m restricted to Â§. Therefore the measure m ^ is bounded if and only if x*z ^ n n sup Z !m^^(A.)j = sup Z x*z(b.) x*z(a.) I
PAGE 58
52 we may restrict the partitions P to be from n (T, Â§) . To see this we introduce the concept of the norm of a partition P Â€ Tf(T,S). If P = { A. : 1 ^ i ^ n] , where each A. is an interval [a.,t)) (and A = [a ,1]), then the norm of P, denoted by I PJ , is defined as follows: [pj = max (b . a.) Â• l^i^n ^ ^ We now state the above mentioned result. 2.3 Theorem . If m is a bounded measure and f ^ C (T) , then for every e > 0, there is a 6 > such that if P and P' belong to TT(T,g) and p' , p < 5, then JS (f,P') S(f,P)  ^ c. In particular, this holds if ] p < 6 and p' s p. Proof . By the continuity of f, there is a 6 > such that if ist < 6 and s,t 6 T, then f(s) f(t) < c/4in(T). Now if P,P' 6 tt(T, S) and if p1,p'1 < 6/2, then n k S(f,P') S(f,P)l = I Z f(s.)m(A.) E f(t.)m(B.)l i=l ^ ^ j = l J ' = 1 E f (s.) Z m(A. n B.) Z f(t.) S m(A. HB.)  i=l ^ j = l ^ 3 j = i D i=i 13 n k = 1 E Z (f(s.) f (t.))m(A. n B.) I . i=l j=l J 1 J The only nonzero terras in this double sum occur when m(A. n B.) ^ 0; in this case A^ n B. 7^ 0, and hence Is. t . 1 < 6 . By Lemma 1.2.1 we conclude th at S(f,P') S(f,P) < 4[c/4m(T)]m(T) = e. The final statement of the theorem is immediate, n
PAGE 59
53 Motivated by this result and the classical definition of the Rieip.annStielt jes integral, we make the following definition. Recall that B(S) is the set of all bounded measurable functions on S. 2.4 Definition . Suppose miE ^ X is a measure and f f B(S) . Then f Â£ RS (m) if for every e > there is a 6 > such that whenever P,P' Â€ tt(T,Â§) and p,p'l < 6, then s(f,P) S{f,P') I < Â€. Using this definition we could prove as in Theorem 1.9.2 that if f f RS (m) , then there is a vector x Â€ X to which the sums S(f,P) converge as  p converges to zero. When m is bounded. Theorem 2.2 shows that C (T) c RS (m) , and Theorem 1.9.3 and Definition 1.9.1 show that RS (m) c L(m) and that the integrals coincide. In fact, if m is any measure from "Z to X, and if C(T) c RS (m) , then m is bounded. This follows from Remark 1.2.3 and the following result. 2.5 Theorem. (Dunford [14]) If m is a scalar measure and C (T) c RS (m) , then m is bounded. Proof . First we remark that if (P ) is any sequence of partitions such that  P  converges to zero, then for every f 6 C(T), the sums S(f,P ) converge to the integral of f with respect to m. Suppose that m is not bounded. Since m is a scalar measure we have only to suppose that the total variation of m is infinite. It follows that there is a
PAGE 60
54 sequence (P ) in tt(T,S) such that JP j < 1/n and S m{A. ) j > n for each n, where P = (A.); we may also assume that m(A.) ^0. If for each n we choose a set of pxDints s. 6 A., then by the preceding remark, lim S(f,P ) = f fdm for every f f_ C (T) , where the sums S(f,P ) are formed using the points (s . ) chosen for each P . n Consider the mapping U :C(T) * X defined by U (f) =S(f,P ), n e N. Each U is linear and bounded, and lim U (f) exists n n' for each f Â€ C (T) . Since C(T) with the normllfll= sup f(s)l is a Banach space, it follows by the principle of uniform boundedness that there is a constant K > such that (*) lS(f,P^) < Kllfll, for every f Â€ C (T) . To obtain a contradiction we shall construct a function f Â€ C (T) such that Hf]! < 1, but for some k n Â€ N, S(f,P ) > K holds. Choose n such that Y. m(A.) >K, ^ i=l ^ where P = (A.) . Let s., 1 < i ^ k, be the sequence chosen for P above. Define f(s^) = m(A^)/ m(A^)  , 1 ^ i ^ k; f(t) = f(s,) for t in the interval [0,s^); f(t) = f(s^) for t in the interval [s, ,1]; and on each interval [s.,s._j_,) define f to be linear between f (s.) and f (Sj^. n) Â• Then f 6 C(T) , and we have llf = 1; however, k S(f,PÂ„) = E f(s.)m(A.) k = E lm(A.) 1 > K. i=l Since this contradicts (*) , m is bounded. Q
PAGE 61
55 The following theorem summarizes the results concerning the boundedness of Stieltjes measures. 2.6 Theorem . Let mri: X be generated by z. The following statements are equivalent: (i) z is of weak bounded variation, (ii) m is bounded. (iii) C(T) C RS (m) . (iv) C (T) c RS (x*m) for every x* Â€ X*. Proof . (i) Â» (ii) follows from Theorem 2.1. Theorem 2.3 shows that (ii) ^ (iii). We deduce from Theorems 1.5.8 and 2.3 that (iii) =* (iv) . Suppose that (iv) holds. Theorem 2.5 implies that x*m is bounded; hence, by Remark 1.2.3, m is bounded. G 3. Extensions of Countably Additive Measures . If z is a scalar function on T, then the classical extension theorem (see Halmos [17]) for the measure m states that m has a unique countably additive extension to I,, the Borel sets of T, if and only if z is leftcontinuous and of bounded (total) variation. When z is a vector function, we shall show that a similar theorem holds, provided that the Banach space X does not contain a copy of c . We need to introduce several preliminary concepts and results . A semiring Â§ is a family of subsets of S such that (i) A,B Â€ S implies that A n B Â€ S; (ii) A,B Â€ S and A = B implies that there is a finite family (C.) in g such that
PAGE 62
56 A=C, cc C...CC =B and C.\c. , f Â§ for 2 f^ i ^ n 1 2 ^ 1 i"'(see Halmos [17] or Dinculeanu [12] for information on semirings and measures on serairings) . If S is a semiring, then the smallest ring R containing Â§, which we call the ring generated by S, is the family of all sums of finite disjoint subsets of Â§. An example of a semiring is just the class 8 of halfopen intervals discussed in Section 1. In this case the ring generated by S is the algebra S. 3.1 Remark . Suppose that Â§ is a semiring and that m:S > X is a set function. It is known (see Dinculeanu [12]) that if m is finitely additive on Â§, then m has a unique finitely additive extension to R, the ring generated by S. Moreover, if m is countably additive on Â§, then its extension to R is countably additive on R as well. From now on, given a finitely additive set function on S, we shall assum.e that m has been extended to R, and shall denote the extension by m. When we refer to the variation in we always mean the variation of the extension, defined as in Chapter I, Section 2, by ffi(A) = sup lm(B) I . BCA BÂ€R Note that if R is the ring generated by the semiring S, then we also have n ffi(A) = sup 1 E m(B.) I , (B^)=S i=l where (B.) denotes a finite disjoint family in Â§.
PAGE 63
57 If S is a topological space and A c S, then we denote the interior of A by int A and the closure of A by cl A. If S is a semiring in S and m: Â§ Â« X is a finitely additive set function, we make the following definition, with Remark 3 . 1 in mind . 3.2 Definition . Suppose that S is a compact space. Then m is regular if for every E Â€ S and e > 0, there are sets A,B 6 S such that cl A c E c int B and m(B\A) < c. We nov/ prove a result due to Huneycutt [18] which generalizes the classical Alexandroff Theorem for regular scalar measures (see Dunford and Schwartz [15]). 3.3 Theorem . (Huneycutt) If S is a compact space, Â§ is a semiring in S, and m:Â§ Â— X is finitely additive and regular, then m is countably additive on S. Proof . Let (E.) be a disjoint sequence in Â§ such that E = HE. Â€ S, and fix e > 0. By regularity there are sets A,B Â€ S with cl A c E s int B and m(B\A) < e. There are also sequences of sets (A.) and (B . ) in S such that for each i, cl A. c E. c int B. and fn(B.\A.) < e/2'^. It follows XI 1 11 that for each n, n m[( U B.)\a] < 2e. For n n ( U B.)\A c [( U B.)\E] U [E\a] i=l ^ 1=1 ^ n = [ U (bAe. )] U [E\A], i=l ^ ^
PAGE 64
58 so n n in[( U B.)\a] s m[ U (B.\E )] + fn[E\Al i=l ^ i=l ^ ^ n i t: c/2^ + c < 2e. i=l 00 oo since cl A is compact and cl A c U E. c U int B., there is a k ? N such that Ac U B. . For n ^ k, therefore, i=l ^ n n n lm(A) Z in(E^) I ^ lin( U B^) m( U E^) 1 + i=l 1=1 i=l n + lm( U B.) m(A) I 1=1 n n n = lm[ U B.\( U E.)] + in[( U B.)\A] 1=1 ^ i=l ^ 1=1 n n ^ ffi[ U (B.\E.)] + m[( U B.)\A] 1=1 ^ ^ 1=1 ^ " i ^ T. e/2 + 2e < 3e. 1=1 n Finally, m(E) Zm(E.)l ^ m(E) m(A) 1 + i=l n + m(A) S m(E.) I 1=1 !S ffi(E\A) + 3e ^ 4e, 00 If n ^ k. We conclude that m(E) = Z m(E.); hence m is 1=1 countably additive on S. G As an immediate corollary of this result we have: 3.4 Corollary . (Huneycutt [19]) If m:Z X is a finitely additive measure such that lim m[ah,a) = for every h0+ a 6 (0,1], then m is countably additive on E.
PAGE 65
59 Proof . It: suffices to show that m is countably additive on S, and in viev/ of Theorem 3.3 we need only show that m is regular. Suppose that [a,b) g Â§. We may suppose that < a < b < 1, since the cases where a = or b = 1 are proved in a similar fashion. Choose h > such that < ah < a 0, we can choose h > such that m[ah,a) + m[bh,b) < e. Therefore m is regular. r\ Vie now discuss an important class of Banach spaces characterized by Bessaga and Pelczynski [3] . The significance of this class for our purposes lies in the fact that it is precisely the class of range spaces for vector measures in which boundedness is equivalent to sboundedness . Let c denote the Banach space of all scalar sequences X = (x ) such that lim X =0, with norm defined by ^ n n ^ [x = sup x I . If X is a Banach space and (x ) is a sequence in X, we say that (x ) is weakly Cauchy if the scalar sequences (x*x ) are Cauchy sequences for each X* 6 X* . We say that (x ) converges weakly to x Â€ X if lim x*x = x*x for each x* . Recall that a Banach space is said to be weakly sequentially complete if every weak Cauchy sequence converges weakly to some element in X. A series Ex in X is said to be weakly unordered bounded if for each x* Â€ X* we have sup I Z x*x. I < <=. A i6A ^
PAGE 66
60 Note that by Lemma 1.2.1, Zx is weakly unordered bounded n if and only if ex*x.  < 0=, for each x* g X*. If a Banach space X contains no subspace isometrically isomorphic to c , then we write X ^ c ; otherwise v/e write X a c . We state the following important result without proof: 3.5 Theorem . (Bessaga and Pelczynski [3]) The following statements are equivalent: (i) X ;2^ c^. (ii) Every weakly unordered bounded series in X converges unconditionally. The following result was originally stated by Brooks and Walker [7] when X is weakly sequentially complete, and slightly extended to the present situation by Diestel [11] . 3.6 Corollary . If X ^ c and m:R X is any measure, then m is bounded if and only if m is s bounded. Proof . By Theorem 1.3.2 an sbounded measure is always bounded. Conversely, if m is bounded and (E.) is a disjoint sequence in R, then by Lemma 1.2.1 we have n T, x*m(E.) I ^ 4x* sup I S m(E.) i=l ACN^ iÂ€A ^ ^ 4[x* sup m(E) I < ", EÂ€R for every n e N and x* Â€ X*. It follows that the series
PAGE 67
61 Zm(E.) is weakly unordered bounded. By Theorem 3.5, therefore, Zin(E.) converges unconditionally. By Theorem 1.3.2 we conclude that m is sbounded. n 3.7 Example . To see that Theorem 3.5 fails to hold when X 2 c , consider the sequence of unit vectors e = (6 Â•) in o ^ n ni c , where 6 . is the Kronecker delta. The series He is o ni n weakly unordered bounded, since for any A c N,  E e. [ = 1, but this series does not converge in c . We now prove the final result that will be needed for the main theorem of this section. 3.8 Lemma . Suppose m:E X is a bounded measure such that lim m[ah,a) = for every aÂ€ (0,1]. IfX;^c, then for hO"^ Â° every a Â€ (0,1] we have lim lft[ah,a) = 0. hO"^ Proof . Since m is monotone the limit above exists for each a 6 (0,1]. Suppose there is an a Â€ (0,1] and a 6 > such that lim m[ah,a) > 6. Let n, = 1. In view of Remark 3.1 h^O ^ there is an n_ > n, and a finite disjoint family of sets 2 1 n2 A^,...,A in g, such that  i: m(A.)l>6. We may assume 2 i=n,+l for convenience that A. = [a.,b.)# where a_ 6/2 . i=n^+l ^ ^2 i=n^+l
PAGE 68
62 Redefining b = b, so that A = [a ,b) , we have "2 "2 ^2 "2 I Z m(A^) 1 > 6/2. i=n,+l Now since ^[t) , a) > 5, we can repeat this argument over 2 the interval [b , a) to obtain an n^ > n_ and a finite ^2 J 2 disjoint family of sets A ,,..., A in S, such that 2 3 ^3 I E m(A.) I > 6/2. i=n2+l ^ By induction, therefore, we obtain an increasing sequence (n. ) and a disjoint sequence (A. ) in S such that for k Â€ N, ^+1 I Z m(A.) ! > 6/2. i=n3^+l Define E, = Z A; then (E, ) is a disjoint sequence in Z, i=n3^+l and for every k 6 N, (*) Jm(Ej^) I > 6/2. Since X ^ c and m is bounded, m is sbounded by Corollary 3.6. This contradicts (*) , so we conclude lim m[ah,a)=0.n h0 The following theorem is our main result on the extension of Stieltjes measures to countably additive Borel measures. 3.9 Theorem . Suppose X ^ c , and let Z^ be the aalgebra generated by S. The measure m :Z X has a unique countably additive extension to Z, if and only if (i) z is left continuous on (0,1]. (ii) z is of weak bounded variation.
PAGE 69
63 If X 2 c , then there is a function z satisfying (i) and (ii) such that m has no countably additive extension to Z, . Proof . If m has a unique countably additive extension m z ' to Zt , then by Remark 1.2.3, m' is bounded, so m is z 1 z z also bounded. By Theorem 2.1 z is of weak bounded variation. Since m is countably additive on S, lim m (A.) = whenever (A.) is a sequence in E such that A.**^ 0. For any sequence a . b , where b 6 (0,1], the sequence [a ,b) decreases to 0. Hence we have = lim m [a ,b) = lim [z(b) z(a )], so z is z Â• n L \ ' ' n left continuous on (0,1]. Conversely, if (i) and (ii) hold, then by Theorem 2.1, m is bounded. (i) implies lim m [ah,a) = for every ^ h.O'^ ^ a e (0,1]. By Lemma 3.8 we have lim. fn [ah,a) = for h.0 ^ a Â€ (0,1]. Consequently, by Corollary 3.4, m is countably additive on S. Since X ^ c and m is bounded, m is sbounded by Corollary 3.6*. We conclude from Theorem 1.3.4 that m has a unique countably additive extension to E, . Now suppose that X 2 c . Let (e ) be the sequence of unit vectors in c defined in Example 3.7. Let e, n = 1 a = n n e, Z e. n ^ 2 i=2 Consider the function z:T Â— c defined by o ^ ^^^^ = J^ ^n^l/n.l,l/n](^)
PAGE 70
64 for every t Â£ T. One can show that z is left continuous on (0,1]. Suppose that ^ s < t ^ 1. If s = 0, then since z(0) = 0, we have z(t) z(s) = a , provided that t e (l/n+l,l/n]. If s > 0, then for t g (l/n+l,l/n] and k s Â€ (l/n+k+l,l/n+k], z(t) z(s) = a^ a^^^ = E e ^^, i=l provided k ^ 1. If k = then z(t) z(s) =0. Now if [a, ,b, ) , . . . , [a, ,b, ) are disjoint intervals in T, with a, < bi ^ Â•Â•Â• ^ ^ < ^vÂ» ^^*^ if ^1 ^ ''' then there is a finite set n, ^ n_ :S ... ^ n , of integers such that ^2i z(b^) z(a^) = E e., 1 :s i ^ k. ^=^2il If a, = 0, then z (b, ) z(a, ) = a . In either case, since k n E [z(b.) z(a.)] is a finite sum of the form Z c^e. for i=l ^ ^ i=l " ^ some n, where each eis or +1, we conclude that k I Z [2(b.) z(a.)] ^ 1. 1=1 ^ ^ By Remark 2.2, z is of weak bounded variation. Now if m has an extension to a countably additive measure on Z, # then m is s bounded. Let A = ri/n+l,l/n) for each n. Then z n (A ) is a disjoint sequence in S, and m (A ) =z(l/n) z(l/n+l) n z n e , n Â€ N. Therefore Im (AÂ„)  = 1 for every n, so m is not n z n z sbounded. Hence m cannot have a countably additive extenz sion. G The requirement that z be left continuous is in a sense superfluous; if z(t~) exists for every t g (0,1], then we can define m[a,b) = 2(b") z(a~), (or m[0,b) =m(b~) m(0)).
PAGE 71
65 Essentially this is just normalizing the function z to be left continuous. Moreover, as the following result shows, if z(t ) exists at each point of (0,1] and z is of weak bounded variation, then z determines a unique countably additive measure on E, . 3.10 Theorem . If z(t~) exists for every t $ (0,1] and if z is of weak bounded variation, then the function z' defined by z'(t) = z(t~), t Â€ (0,1], and z'(0) = z(0) is left continuous on (0,1] and of weak bounded variation. Proof . Since by definition z' is left continuous on (0,1], it suffices to show that z' is of weak bounded variation. If [a, ,b, ) , . . . , [a ,b ) are disjoint intervals in T with a, < b, s . . . ^ a < b , then we can choose points af < a. 11 n n 11 (if a, = let a,' = 0) and bf < b., such that for each i, 1 111 ' 2(a.~) z(aM < l/2n and z(b.") z(b.') < l/2n. Then n n _ n Is 2'(b.) z'(a.)l ^ E z(b.) z(b.') I + s z(aT) z(a;)l i=l ^ ^ i=l ^ ^ i=l ^ ^ n + I S z(b^) z(a() i=l ^ 1 + ffi (T) . z By Remark 2.2 it follows that z' is of weak bounded variation. G Ifz:TÂ»Xis a function such that m has a countably additive extension to S, , then as we have already seen, z is left continuous. It follows that righthand limits also exist at each point in [0,1). For suppose a ^ [0,1) and b^ a . Then the intervals [a,b ) decrease to [ a} 6 St / and
PAGE 72
66 so by the countable additivity of m , m f a] = lira m fa.b ) = lim[z(b ) 2(a)]. Thus z(a ) exists. By Theorems 3.10 and 3.9, every function z with right and lefthand limits that is of weak bounded variation determines a unique countably additive measure on S, . Conversely, it is easy to see that if m is a countably additive measure on S, with values in X, then there is a function z with right and lefthand limits that is of weak bounded variation and satisfies 2(a~) = m[0,a) if a Â€ (0,1], and z(a ) = ra[0,a] if a C [0,1) . If we normalize this function to be left continuous and satisfy z (0) 0, then we have established a one to one correspondence between the class of all countably additive measures on S, and the family of all normalized functions z that are of weak bounded variation. This is a generalization of the classical theorem relating countably additive Borel measures and functions of bounded variation in the scalar case. 4. Measures in Hilbert Space . In this section we apply our previous results to sharpen a theorem of Cramer [10] , and discuss its consequences in integration theory. H is a Hilbert space with inner product {Â•,Â•), and z:T Â• H is a function which generates a measure m = m:I Â• H. 5:Recall from Section 1 that the family g of all halfopen intervals in T is a semiring which generates S. If g :is any semiring, let gxS denote the class of all sets AxB, where A, B Â€ g. It is not difficult to see that gxg is
PAGE 73
67 also a semiring. Elements of Â§xS are called rectangles. Consider the set function m, :Â§xS Â• $ defined by F.j^(AxB) = . By the next lemma it will follow that m, is finitely additive on SxS/ and therefore by Remark 3.1, m, has a unique extension to a finitely additive measure on the algebra H generated by gxS. 3" is just the family of finite sums of rectangles in gxS. 4.1 Lemma. If AxB Â€ SxS and AxB = AXB, + AÂ„xB , then either A = A, = A^ and B = B, + B^ . or B = B, = Band Proof . If A, = Ap, then clearly A = A, = A. Suppose that B^ n B^ y^ ^. If b Â€ B, n BÂ„, then for any a Â€ A, (a,b) Â€ AjXB, n 7iÂ„xBÂ„, which is a contradiction. Hence ^1 1^ ^2 = 0. Since B C Bj^ U B^ c B, E = B, + BÂ• If A, 7^ Ap, then there are a,,aÂ„ Â€ A such that ^1 ^ ^1^^2 ^^^ ^2 ^ ^2^^!' ^^^^^ (a,,b,) and (a^/b) Â£ AxB for any b^ Â€ B^, i = 1,2, (a,,b2) and (a^/b,) Â€ AxB. Since (a.j^,b2) Â€ AjXB,, B2 = B,. Similarly, B, c B2/ so B = B, = B2 . Then as above we must have A, D Aj = and A = A, + A2 . D Now suppose that AxB = A,xB, + A2XB2, and that A=A, =A2 and B = B, + B2 . Then mj^(AxB) = m^(Ax(B^ + B2)) = + = m^(Aj^xBj^) + m^(A2XB2) .
PAGE 74
68 It follows that m, is finitely additive on SxS* and so by Remark 3.1, we may assume that m, is a scalar measure on 3". 4.2 Remark . Suppose that (A^^) is a finite disjoint family in g. Then we have n n n I E in(A. ) 1^ = < Z m(A.), Z m(A. )> i=l " i=l " i=l " n n = 1:1: i=l j=l ^ 3 n n ^ E Sim (A.XA.) I . i=l j=l ^ ^ 3 We conclude that if m, is a bounded measure on 3", then m is a bounded measure on H, or equivalently z is of weak bounded variation. Since m, is a scalar measure. Theorem 1.2.2 (ii) implies that in^ i v(mj^) i 4m, on 3". It follows that in the definition of regularity. Definition 3.2, we may replace m^ by v(m,) . In what follows we shall use v (m, ) . The next lemma is the counterpart of Corollary 3.4. 4.3 Lemma . Suppose m, is bounded. Then m, is regular if (i) lim_ v(m^, [s,s^)x[t,t^)) = 0, s^ Â€ (0,1], t,t^Â€T. sÂ»s o (ii) lim v(m^ , [s,s )x[t,t )) = 0, t Â€ (0,1], s,s Â€T. tÂ»t~ X o o o o o Proof . Suppose [a,b)x[c,d) Â€ SxS, and e > 0. By (i) there is a 6 > such that v (m, , [a 6,a) x [0,d) ) < e and v(m, , [b 6/b) X [0,d) ) < e. By (ii) there is an a > such
PAGE 75
69 that v(in, [0,b) X [d a,d) ) < e and v (m^, [0,b) x [c a, c) ) < c. Now let A = [a,b 6) x [c,d a) and B = [a 6,b) x [c u,d) . It follows that cl A c [a/b)x[c,d) c int B, and that b\a c [a 6,a)x[0,d) U [b 6/b) x [0, d) U [0,b) x [c a, c) U U [0,b) x [d a,d) , so v(m, ,b\a) ^ 4e. We conclude that m, is regular, since the same type of argument can be applied to any set in gxS. D We nov/ apply Lemma 4.3 to obtain a connection between the countable additivity of m, and the continuity properties of the generating function z. 4.4 Theorem . Suppose that m, is bounded. Then z(t ) and z (t ) exist for every t Â€ (0,1) (with onesided limits at and 1) . m, is countably additive on 0" if and only if z is left continuous . Proof . Suppose that lim sup z(t) z(s)  > 6 > 0. Thai s , tÂ»t o there is a sequence s,
PAGE 76
70 Suppose that m, is countably additive on S"; then m, is continuous from above at 0. Since [t,t )x[t,t ) decreases to as t t~ , we conclude that ^ o = lim m. rt,t ) X [t,t ) . . 1 Â•Â• ' o ^ o t*t o = lim_ lz(t) z(t^) 1 2 . tt " o Therefore z(t~) = z (t ) . Conversely, suppose that z is left continuous . In view of Lemma 4.3 and Theorem 3.3 we need only show (*) lim_ v(m^, [s,s^)x[t,t^)) = 0, s^ Â€ (0,1], t^t^ 6 T, sÂ»s o and similarly for t t~ as in Lemma 4.3, By Schwartz's inequality we have lm^[s,SQ)x[t,t^) I =   ^ z(s^) z(s) I z(t^) z(t), and so be the ass\imption of left continuity of z we see that (**) lim_ m^[s,s )x[t,t^) =0 sÂ»s o for s Â€ (0,1] and t,t Â€ T. Since v(m^) is monotone, the limit (*) exists for every s Â€ (0,1] and t,t Â€ T. Suppose that for some choice of s ,t, and t we have o o lim_ v(mj_, [s,s^) X [t,t^) ) > 5, s*s o for some Â£ > 0. Then there is a family of disjoint rectangles [a^,b^) X [Cj^,dj^) , 2 s. X S, n^, contained in [0,s^) x [t,t^) , such that S m, [a.,b.)x[c. ,d.)  > 6. i=2 ^ ^ ^ X i
PAGE 77
71 Vie shall now show that we may assume b. < s for each i. Suppose that [a,b)x[c,d) is any rectangle in [0,s )x[t,t ) such that b = s , and linj^[a,b)x[c,d)  = a > e > 0, for some e > 0. By the continuity condition (**) we can select a b' such that a < b' < s and o m^[b',s^) X [c,d) 1 a (a e) = e. Now if any of the b. = s , then the argument above shows that we can find new b. 's such that b. < s and 1 X o E m, [a.,b.)x[c. ,d.)  > 6. i=2 Let 6, = min (s b . ) ; then v (m, , [s 6, ,s ) x [t/t ) ) > 6, X o Â• ^' ,Â« ox X O X O O ' Z^x^n^ so we can repeat the process above. By induction we obtain a disjoint sequence (AxB.) of rectangles and an increasing sequence (n, ) of integers (set n, = 1) such that ^+1 S Im^ (A. xB.) I > 6, x=n3^+l contradicting the boundedness of m, . We conclude that (*) holds. The proof for t't" is similar to the above. Q o The theorem we are about to prove was essentially stated by Cramer [10] , but his statement and proof contain a gap which we now rectify. In our terminology, the statement of Cramer's result is as follows. Let z:T Â» H be a function generating m and m, as above. If m, is a bounded measure.
PAGE 78
72 then m and m, have countably additive extensions to Z, and 3"^ respectively, where ST, denotes the aalgebra generated by ST. Since even for scalar functions the measure m is not z countably additive unless z is left continuous, the theorem as stated is not correct. We now present the corrected version. 4.5 Theorem . If m. is bounded, then the measures m and m, have unique countably additive extensions to Z, and 7, respectively if and only if z is left continuous. In this case we have (*) = m^(AxB) for every A and B in E, , where we identify the extensions by m and m, . Proof . If m, has a countably additive extension then z is left continuous by Theorem 4.4. Conversely, if z is left continuous, then by this same theorem m, is countably additive on JT. By the assumption of boundedness, m, has a unique extension to IT, (by the classical extension theorem or Theorem 1.3.4), and z is of weak bounded variation by Remark 4.2. By Theorem 3.9, m has a unique countably additive extension to To see that (*) holds, fix B Â€ S, and define set functions H and X on Z, as follows: H(A) = mj_(AxB), X(A) = , for A Â€ Ei . Since m and m, are countably additive, it follows that ^L and X are finitely additive and continuous from above at 0; thus \i and X are countably additive scalar measures.
PAGE 79
73 Since X (A) = u (A) for A Â€ E, it follows by the uniqueness of extensions that X (A) = u (A) for A Â€ Y.^, or m^ (AXB) = for A Â€ S, and B Â€ E fixed and B Â€ S, fixed. Repeating this argument for A Â€ Ei fixed and B Â€ H, gives the desired result, n The author is indebted to Professor Brooks for discussions on the previous material; in particular the proof of (*) in Theorem 4.5 is due to him. As an application of Theorem 4.5, we discuss the most important case, namely functions z with orthogonal increments. A function z:T H has orthogonal increments if = 0; m, therefore reduces to a nonnegative measure on S if we define m, (A) = = m(A) . It follows from orthogonality that m, is finitely additive, since m^{A + B) = = + + + = m(A) + m (B) . Since m, is nonnegative and finitely additive it is monotone, and so m, (A) ^ m, (T) for every A Â€ E. By Theorem 4.5 we conclude that m and m, have unique extensions to E, if and only if z is left continuous. Moreover, since m, is bounded. Theorem 4.4 implies that zCt"*") and z(t~) (z(0"*") and z(l~)) always exist. It follows that any function z with orthogonal
PAGE 80
74 increments determines countably additive measures m and m. if we normalize z to be left continuous. In fact, z is continuous except possibly at countably many points, since F(t) = m, [0,t) is a nondecreasing function and lz(t) z(s) = F(t) F(s) . Suppose now that z is any function with orthogonal increments, not necessarily left continuous, and z generates the measures m:i; H and miZ * [0,co) as above. Since [m(A)] = sup m(B) I = sup m, (B) = m^ (A) , BCA BCA BeS BÂ€S it follows that convergence in measure with respect to m is equivalent to convergence in measure with respect to m, . Moreover, if (f, ) is a sequence of step functions on T, then the indefinite integrals of the fi^'s are uniformly mcontinuous if and only if they are uniformly m, continuous . Suppose n f = Z a. I is a step function; then 1=1 "Â• ^i ~ n n I Lfdmr = < S a.m(E HE.), S a.m(E HE.)) , ^ i=l ^ ^1=1"^ = _Z a.2m^(E n E.) = Jgf2dm3_, " using the orthogonality of the measure m. By the remarks above, and by Definition 1.5.4 and Theorem 1.7.1, we conclude that f Â€ L(m) if and only if f Â€ 'LÂ°{m^) . Moreover, for step functions f and g it can be shown that = Iehf^^^i' and so by continuity this equality holds for every f and g in
PAGE 81
75 L(m) = Lp (m, ) . If z is left contimaous, so that rti and m, have countably additive extensions to Borel measures, then by Theorem 1.8.5 the class of step functions is dense in L (m) = LÂ° (m, ) , and the results above still hold by continuity.
PAGE 82
CHAPTER III Stochastic Integration In this chapter we shall restrict the integration theory of the preceding chapters to the case where t?ie Banach space X is an L space over a probability measure space. Stieltjes integrals with respect to a function 2:T Â• L are stochastic integrals in the sense that the integral is an element of L , or in probabilistic terms, a random variable. We show that the general integration theory of Chapters I and II includes certain stochastic integrals previously defined by other methods. In Section 1 we introduce and discuss the probabilistic concepts which are used. Section 2 consists mainly of background material on the theory of stochastic processes. In Section 3 we attempt to motivate the study of stochastic integration, and then discuss the general integration theory in the present context. The sample path stochastic integral is studied in Section 4; we compute some statistical properties of the integral in a simple special case, and discuss an integration by parts formula. The WienerDoob stochastic integral in L, is discussed briefly in Section 5. In Section 6 we deal with m.artingale stochastic integrals, and prove a general existence theorem. 76
PAGE 83
77 1, Probability Concepts and Notation . We say that (n,F) is a measurable space if is a nonvoid set (w is a generic element in in n) and F is a aalgebra of s\absets of Q. If there is a countably additive measure P defined on F with values in [0,1], and if P(fi) = 1, then we say that (n,F,P) is a probability measure space, and P is a probability measure. A probability measure space, or probability space, is therefore just a finite nonnegative measure space with the property that Cl has measure 1. Suppose that (Q,F) and (n',F') are two measurable spaces, and x:Q q' is a function. x is said to be measurable (relative to F and f') if the set [x Â€ a'] = x~ (a') belongs to F for every a' Â€ f'. The family of all sets [x Â€ a'], a' Â€ f'# is a aalgebra contained in F, and we denote it by F(x). Note that F(x) is the smallest aalgebra f" contained in F such that x is measurable relative to f" and f'. If P is a measure on F, then a measurable function x determines a measure p' on F ' by the relationship p'(a') = P[x Â€ a'] for each a'Â€F'. IfPisa probability measure on F then p' is a probability measure on f' as well. In this case p' is called the distribution of x. When fi' = $ we shall always take f' to be the aalgebra of Borel sets in $. A measurable scalarvalued function on a probability space (Q,F,P) is called a random variable. As motivation for this term, suppose we have a system whose state can be described by a scalar variable x, and whose behavior is sxjbject to some sort of statistical variation. We imagine
PAGE 84
78 the numbers x corresponding to the (random) states of the system to be the values of a function x defined on some probability space. The assumption that x is a measurable function insures that the probability of each event x 6 B is defined, where B is any Borel set. This probability is just P[x Â€ B] . Lp L (n,F,P), where 1 ^ p < oo, will denote the Banach space of all equivalence classes of random variables x such that ex^ h J^x(uj)Pp(dtJu) < Â». L^ denotes the space of equivalence classes of essentially bounded random variables. It is known (see Dunford and Schwartz [15]) that for 1 ^ p < oo, Ly, is weakly sequentially complete. Since c is not weakly sequentially complete, it follows that L ^ c', hence Theorem II. 3. 9 applies to these p o ^^ spaces. The number exP is called the pth absolute moment "k of X. If k Â€ N, then Ex is called the kth moment of x. In particular, the first moment Ex is called the mean or expectation of x. It is well known from the theory of integration that the expectation operator E is a continuous linear functional on L, . If x Â€ L, then the number a^(x) = exEx1^ = ex1^ ex^ is called the variance of x, and represents in some sense the dispersion of x about its mean value Ex. Standard references for results in probability theory are Loeve [21] and Chung [9] ; most of the material in this first section can be found in either of these works.
PAGE 85
79 2. Stochastic Processes . In this section we discuss the definition of a stochastic process (random function) as a mapping from a subset T of R into an L space over a probability space. In order to illustrate some of the important properties that a process may have, we discuss two classical examples, the Poisson process and the Wiener process. Let T be a subset of R whose elements are interpreted as points in time, and suppose we have a system whose observable state at time t 6 T is some scalar quantity x(t) . If the system is subject to influences of a random or statistical nature, then we are led to suppose that x(t) is a random variable for each t Â€ T. Our mathematical model of the system is then a family of random variables x(t), t Â€ T, which are defined on some underlying probability space (Q,F, P) . If we assxime that each x(t) belongs to L , so that the convergence properties of this space are available, then we have defined a mapping x from T into L . x is called a stochastic IT process or random function since it represents the behavior of a system subject to stochastic (random) influences. Since L is a space of (equivalence classes of) functions from n into f, we may also think of a random function x as a mapping from Txfi Â• i . Then each uu Â€ Q determines a function x(Â»,uj):T $; this function is called the sample function or sample path corresponding to uu. Every set A Â€ F determines a set of sample functions [x(Â«,uj) :u) Â€ A] , and we say that the probability of this set of sample functions is p if P(A) = pÂ»
PAGE 86
80 A fundamental problem in the theory of random functions concerns the fact that we would like to discuss the probabilities of sets of sample functions which may not correspond to sets in F as above. For example, given any finite set of times t,,...,t and Borel sets B,,...,B , the set n [x(t^) Â€ B^:l^i<:n] = n [x(t.) Â€ B . ] i=l is in F. Even for countably many t.'s and B.'s such sets are in F. However, the set [x(t) Â€ B:t 6 (a,b)] = n [x(t) Â€ B] tÂ€(a,b) clearly need not be measurable. Doob [13] discovered a reasonable way to skirt this issue. He was able to show that every random function x is equivalent to a random function y which behaves nicely in this regard, where equivalence means that for every t Â€ T, P[x (t) = y (t) ] = 1. We shall always assvime for the sake of simplicity that we are dealing with such wellbehaved processes. In order to discuss two important examples of stochastic processes we need to introduce a few additional concepts. Suppose that x(t), t 6 T, is a random function, and t,,...,t is a finite subset of T. Assume for simplicity that x(t) is realvalued for each t. Let R denote real Euclidean nspace, and let B denote the aalgebra of Borel sets in R . It can be shown that the family of all product sets B,x...xB , where B. Â€ B for 1 s: i ^ n, is a semiring which generates B . Consider the mapping uu Â• [x(t, ,0)) , . . . ,x(t , u)) ] , which is a
PAGE 87
81 function from Q into R . Since each x(t.) is measurable, it follows that the set n [[x(tj_),...,x(t^)] Â€B^x...xB^] _n [x(t^) e B^] belongs to F, for every product set B,x...xB . As in Section 1, the mapping uu [x(t, ,uj) , . . . /X (t ,(Â«)] determines a measure on the semiring of product sets, and by a standard extension theorem (see Halmos [17] ) , this measure extends to a probability measure P. . on B . The measure P. . is called the joint distribution of the random variables x (t, ) , . . . ,x(t ) , It seems reasonable, and in fact can be verified (see Loeve [21] ) , that a random function is characterized by the family of all its joint distributions P. . . 1 n Suppose that x(t) , t Â€ T, is a random function. As in Section II. 4, we say that x(tÂ„) x(t,) is an increment of the process. x(t) is said to have stationary increments if for every choice of t > and s,
PAGE 88
82 n known (see Loeve [21]) that if E rr x(t.) x(s.)! < Â°' , i=l ^ ^ then we have n n E TT [X(t ) X(S )] = TT E[x(t.) X(S )]. i=l ^ ^ i=l ^ ^ Now we consider a classical random function called the Poisson process. Suppose we have a general situation where events occur in a random fashion that is in some sense uniform over T = [0,1] . Let x(t) denote the number of events which have occurred in the interval [0,t). It seems reasonable to assume that the random function x has the following properties : (i) Independent increments. The increment x(t) x{s), t > s, represents the number of events occurring in [s,t), and we would expect that over disjoint intervals of time, the numbers of events occurring in those intervals are independent of each other. (ii) Stationary increments. The uniformity assumption means that over disjoint intervals of the same length, we expect that the distribution of occurrences is the same. (ill) x(0) ^ 0. If in addition we assume that there is a X > such that for small values of h > we have P[x(h) = 1] = Xh + o(h), and P[x(h) = 0] = 1 Xh + o(h) , then P[x(h) > 1] = o(h).
PAGE 89
83 and it can be shown (see Doob [13)) that for every s < t in T, we have P[x(t) x(s) = n] = e ^X(ts) ^ \"(ts)" nl for n = 0,1,2,... . This discrete probability distribution is called the Poisson distribution, and the relation above states that the random variable x(t) x(s) is distributed according to a Poisson distribution, with parameter X. The mean and variance of the increments are given by E[x(t) x(s)] = X(t s) , and a^{x{t) x(s)) =E[x(t) ^x(s)]2[E(x(t) x(s))]^ = = \(t s) . The sample functions of a Poisson random function (except for a set of sample functions of measure zero) are monotones nondecr easing, nonnegativer integervalued functions of t, with at most finitely many jumps of unit magnitude in T. This might be expected in view of the eventcounting interpretation of x(t) . ; . A second important example of a random function is the Wiener process, which has been used extensively as a mathematical model for Brownian motion. Suppose we observe a particle undergoing Brownian motion, and restrict our attention to, say, the x coordinate of its position. Due to the random bombardment of the particle by the fluid's molecules, this X coordinate changes in an erratic fashion as time progresses. If we denote the value of x at time t s by
PAGE 90
84 x(t), then it seems reasonable to assume that x(t) is a random function which has the follov;ing properties: (i) Independent increments. The increment x(t) x(s) represents a change in position along the xaxis over the interval [s,t). Over disjoint intervals of time these changes should be independent of each other due to the random character of the molecular bombardments. (ii) Stationary increments. Since we would assume that the fluid in which the particle is suspended is homogeneous, the distributions of change over two intervals of the same length should be the same. (iii) x(0) = 0, assuming we use the position of our initial observation as the center of the coordinate system. (iv) Normally (Gaussian) distributed increments, with mean zero and variance proportional to the length of the time interval. This follows from the assumption that the motion is random and continuous, and means that if a < b, then fb , 2 .^ 2 P[a a x(t) x(s) is some constant, namely the variance over an interval of unit length (we have assumed t ^ so T = [0,co) ) . It can be shown that except for a set of sample functions of measure 0, the sample functions of a Wiener process are continuous, nowhere dif ferentiable, and of infinite variation over any finite interval. Although this last property runs
PAGE 91
85 counter to what intuition says the sample paths should look like, the Wiener process has played a very significant role in the application of probability theory to physics and engineering. In fact, the first stochastic integrals were defined by Wiener for scalar functions with respect to a Wiener process. A standard reference for results on stochastic processes is Doob [13] . Much of the material in this section can be found in his classic work. 3. Stochastic Integrals . We now discuss as example which motivates the theory of stochastic integration. Suppose we have a system consisting of a signal processing unit with an input signal z(t) and an output signal x(t) . Suppose that the effect of the processor can be described by a scalar function of two variables g(t,T) in the following way. A small change z(dT) = z(t + dr) z(t) in the input at time t produces a change in the output of the system at a time t ^ t given by g(t,T)z(dT). Moreover, suppose that the response of the system is linear, in the sense that small disturbances z (dx, ),..., z (dr ) produce a change in n the output given by Z g (t, x . ) z (dx . ) , for t ^ max x . . If i=l ^ ^ ^ we assume that g(t,x) = if t < x, then for a given input function z(t), we have formally that x(t) = J^^ g(t,x)z(dx) . As long as the functions g and z are sufficiently wellbehaved scalar functions, the definition of this integral poses no problem.
PAGE 92
86 Suppose, however, that the input z(t) is a random function, for example, a signal contaminated by random noise. Then the output will also be random, and we are faced with the problem of how to define the integral above so that it is a random variable for each t. If the function g is continuous in t for every t, and if almost every sample function of z is a function of bounded variation, then we could define the random output in a pointwise fashion as follows: (*) x(t,uj) = J"^ g(t,T)z(dT,uu) . Stochastic integrals defined in this way are called sample path integrals; although they are the most restrictive kind of integral, their study does provide motivation for the more powerful kinds of stochastic integrals. In general, this simple solution to the problem is not feasible, since even processes as wellbehaved as the Wiener process have almost every sample function of unbounded variation. Using the integration theory of Chapters I and II when the Banach space X is an Lspace over a probability space, we obtain a very general definition of the stochastic integral. Suppose that z:T * Lp is a random function, where 1 ^ p < OS. As in Chapter II we can define a measure m: Z Â» Lp by setting m[a,b) = z (b) z (a) on g. For each A Â€ I, m(A) is thus a random variable in Lp. The measure m is bounded if and only if z is of weak bounded variation, and by the form of Lp* this is equivalent, using Theorem I I. 2.1, to
PAGE 93
87 requiring that the scalar functions t Â• E[z(t)x] are of bounded variation for every x Â€ L , where 1/p + 1/q = 1 if 1 < p < <=, and q = Â» ifp= 1. Suppose that z is of weak bounded variation. Then z determines a class L(z) = L (m) of integrable scalar functions, and for f Â€ L(z) we write J^f(t)z(dt) = J^f(t)m(dt) . We shall call such integrals stochastic Stieltjes integrals, or norm integrals for short, since they are defined as the limit of a sequence of integrals of step functions in the L norm. If JixH denotes the L norm (ex^) ^^ of an element X Â€ L , then it is known (see Loeve [21] ) that 1 s; r :S p and It X Â€ L implies that jjxll s: IJxH . This shows that L a L , cind that the canonical embedding U of L into L is continuous ^ pr p r By Theorem 1.5.8 it follows that if f Â€ L{z) then f Â€ L (U z) and That is, if f is integrable with respect to z as a mapping into L for 1 i p < 00, then f is integrable with respect to z as a mapping into L , 1 ^ r ^ p, and the integrals coincide. Moreover, it is clear that the same sequence of step functions determines the integrals of f in L and in L . By Theorem 1.5.8 we have E(J^f(t)z(dt))x = J^f(t)E[z(dt)x], for every x Â€ L_, since as is well known, this is the form of linear functionals on Lp (see Dunford and Schwartz [15]).
PAGE 94
88 Thus E(J^f(t)z(dt)) = J^f(t)E[z(dt)] . Also El f^f (t) z (dt) I 2 = E [ J^f (t) z (dt) ] [J^f(T)z(d7) ] = J^f ( t) E [Z (dt) J^f(T)z(dT) ] . In general E[J,j,f(t)z(dt)]^ = J^f(t)E[z(dt) (J^f(T)z{dT))^~^], kÂ— 1 since [J f(T)z(dT)] belongs to W/t,_i if z:T L, . If we evaluate the scalar functions t E[z(t) (J^f(T)z(dT))^"^], then we can compute the kth moment of the integral by performing a scalar integration. In practice this may be of some value. 4. Sample Path Integrals . Let T = [0,1] and g be a continuous function on T. Let z (t). be a Poisson process on T with parameter X. As described in Section 2, almost eveiry sample path z(',uu) is a monotone nondecreasing function, with z(0,tju) = 0, and finitely many jumps of unit magnitude. We can therefore define the RiemannStieltjes integral . jj g(t)z(dt,uj) for almost every m. By Theorem II. 2.3 it follows that jj g(t)z(dt,uj) = lim Z g (i1/2'') [z (i/2'',aj) z (i1/2'', uO ] X1.00 i=l for almost every u). Since the summands g(il/2 ) [z(i/2 ) 'z(il/2")] are random variables it follows that the sample
PAGE 95
89 path integral J g(t)z(clt) is also a random variable. Moreover, we can evaluate the integral directly, due to the simple behavior of the sample functions. Fix o) and let n = z(l,o)). Let t,,...,t denote the n points of jump of o the sample function z (Â•,(Â«). For every n and i, the difference z(i/2 ,(d) z(il/2 ,(ju) is either 1 or 0, depending on whether one of the t. 's belongs to (i1/2 ,i/2 ] or not. By the continuity of g we see that n 1 Â° n g(t)z(dt,(ju) = E g(t.) . Â•^0 i=l ^ In general, denote by t . (tju) the point in T of the ith jump of the sample function z(Â«,uj)7 that is, for t ^ t. (uj), z(t,a!) i i 1, while for t > t.(iD), z(t,uj) ^ i. Since [ujrt. (uj) < a] = [z(a) 2: i] Â£ F for each a Â€ T, t. is a random variable for each i. Thus 1 z(l,u)) Jq g(t)z(dt,(ju) = r g(t^((ju)) . 1=1 That is, the integral of g with respect to the Poisson process z is the sum of a random number (z(l,uu)) of random variables of the form g (t . (uu) ) . ie: We can formally compute the mean of the integral as follows: we have . _ .. 2n 2^ : E[ S g (i1/2") [z(i/2")z(il/2'')] = E g (i1/2'') \ (1/2^) . i=l i=l 2^ , Since lim E g (il/2") X (1/2^) = X fi g(t)dt, and since the i=l JO 2^ functions Z g (i1/2 )^r;_i/2n 1/2^) converge uniformly to g on T, we would expect that E[Jq g(t)z(dt)] = XjJ g(t)dt.
PAGE 96
90 This will follow once it is established that the stochastic Stieltjes integral exists. Proceeding to the formal computation of the variance of the integral, we compute Ej S g(il/2") [z(i/2") z(il/2'')]r = i=l 2^ 2^^ = S S g(il/2'')g(jl/2'')E[z(i/2")z(il/2")] x i=l j=l X [z(j/2")z(jl/2'')] = E lg(il/2'')l^ E[z(i/2'')z(il/2'')]^ + i=l 2^ + Z g(il/2'')g(jl/2'')E[z(i/2'')z(il/2'')] x X E[z(j/2'') z(jl/2'')], since z has independent increments. To evaluate the second moments, we suppose that x is a Poisson random variable with parameter a. Then ^ OD _ n <Â» nÂ— 1 ^2 ^ 2 a a a v^ Ex = Zne Â— r=ae Sn , , x , n=0 ^n=l ^'^^^' ' '00 n1 03 n1 = ae [ae + e ] = a(a + 1) . Since z(i/2'^) z(il/2^) is a Poisson random variable with parameter X2~ , we conclude E[z(i/2'') z(il/2'')]2 = X^(l/2'')^ + X(l/2"). Therefore,
PAGE 97
91 2 Â• , e Z g(il/2'') [z(i/2'') z(il/2'')]r = i=l 2^ 2^ 2^ = I I g(il/2'')g(jl/2'')X^(l/2")^+ 7; {g{ii/2^) 2x(l/2"). i=l j=l i=l The double sum converges to \'^l\ Jo g(t)iTiT dtds = x2(jjg{t)dt) (jj ^TiTdF) = X^ljJ g(t)dtl2. 1 2 The sum on the right converges to XjQlg(t) dt. We conclude that if the integral J^ g(t)z(dt) exists as a norm integral in L_, then a^(jj g(t)z(dt)) = X^[!jJ g(t)dt2] +XjJg(t)  ^dt IXjJ g(t)dtl2 = XjJg(t)2dt. Now since g is continuous we need only show that z is of weak bounded variation as a mapping into L^; then the norm integral r_ g(t)z(dt) belongs to L_, hence to L, as well. By computations as above, n _ n n Â„ e! S z(t.) z(s.) ^= S Z X^(t. s ) (t s ) i=l ^ ^ i=l j=l 1 1 J J n + 2 X(t. s ) i=l ^ ^ i X^ + X,
PAGE 98
92 since ^ s. ^ t. ^1 for each i. We conclude that z is of weak bounded variation by Remark II. 2. 2. Since the simple 2^ functions H g(il/2 )1ri /," Â• /on< converge uniformly to g on T, we conclude that 2^ Jq g(t)z(dt) = lim I g(il/2") [z(i/2") z(il/2'')] i=l in Lp and hence in L, . The expressions for the mean and variance now follow. Motivated by this example, we can obviously define sample path integrals under the following conditions: (i) f:T $ is a continuous function and z(t) is a process with almost every sample function of bounded variation. (ii) f:T Â» $ is a function of bounded variation and z (t) is a process with almost every sample function a continuous function. If either (i) or (ii) hold then by the formula for integration by parts for RiemannStielt jes integrals (see Rudin [26]) we have (*1 Jq f(t)z(dt,u)) + jj z(t,uj)f(dt) = [z(.,(ju)f(.)]J. where both integrals are RiemannStielt jes integrals as in Definition II. 2.4, and (*) holds for almost every uu. In case (i) , random variables of the form n S f (il/n) [z (i/n) z (il/n) ] 1=1 converge to ["_ f(t)z(dt) for almost every uu. In case (ii) ,
PAGE 99
93 sums of the form n Z z(il/n) [f(i/n) f(il/n)] i=l converge to rÂ„ z(t)f(dt) for almost every uj. In both cases, then, the integral exists as a measurable function on (Q,F, P); that is, the integral is a random variable. In general, the existence and computation of the moments of a sample path integral depend on the properties of the functions Ez (t) , Ez(t)z(s), and so on, as we saw in the case of the integral with respect to a Poisson process. A case of particular interest concerns random functions whose sample functions are nonnegative and nondecreasing functions. Since s < t implies that i z(s) ^ z (t) almost everywhere, we see that ^ Ez(s) ^ Ez(t) ^ Ez(l) . If we suppose that z(l) belongs to L, , then it follows at once that z is of weak bounded variation, since the mapping t Ez(t) is nondecreasing. If f Â€ L(z) , then we know from Theorem 1.5.8 that Ejjf (t) z (dt) = jjf (t) Ez (dt) . The Poisson integral discussed previously is a special case of this, since Ez(t) = Xt. We now discuss an interesting connection between norm integrals and certain sample path integrals. Suppose that z:T Â• L is a random function of weak bounded variation, and XT suppose that z is also measurable with respect to the product measurable space (Txfi, SiXF) . (Z,xF denotes the smallest aalgebra containing all rectangles AxB with A 6 S, and B 6 F) . Shachtman [27] asserts that the following integration by parts theorem holds, when p = 1.
PAGE 100
94 4.1 Theorem, If f:T 4 $ is a function of bounded variation and z:T 4 L, is product measurable and of weak bounded variation, then f Â€ L(z) , and J^^f{t)z(dt) = J^z(t)m^(dt), where the integral on the right is the sample path integral with respect to the negative of the LebesgueStielt jes measure generated by f . A is taken to be an interval if m is not countably additive, and a Borel set if m is countz ^ z ably additive. The proof of this theorem (Shachtman [27] ) discusses only the case when A = T. Tliat the theorem is false as stated is trivial, since it fails even if z is a nonrandom function, for example z(t) = t. For if f is constant then m^ = 0. The question remains whether the formula jj f(t)z(dt) = jj z(t)m^(dt) + [zf]J can be true under these hypotheses. We establish the following result. 4.2 Theorem . Suppose f:T $ is a continuous function of bounded variation, and z:T Â» L is product measurable and of weak bounded variation, where 1 ^ p < <Â». Then jj f(t)z(dt) + jj z(t)m(dt) = [zf]J, where the integral in the center is the sample path integral of z with respect to the BorelStieltjes m on S, generated by f.
PAGE 101
95 Proof . Since z is of weak bounded variation and f is continuous, f Â€ L(z) . z is a bounded function into L_, so sup Elz(t)P < o=. Since the function t Ez(t)lP is measurable by the Fubini Theorem (see Dunford and Schwartz [15]), we have by this same theorem, E(jJlz(t)Pv(m,dt)) = jj Ez(t)lPv(m,dt) < =o. Therefore the function uj J  z (t, U))  ^v (m,dt) is Fmeasurable and finite almost everywhere. Since the function t Â• t^ is convex, the Jess en inequality for integrals implies that for almost every tju, we have \vd^ jjlz(t,a;) lv(m,dt) jP ^ ^^^ lll^it,^) Pv(m,dt) . Therefore EjJz(t,uu) lv{m,dt) jP <: v(m,T)P~^EjJz(t,a)) Pv(m,dt) < Â». We conclude that the mapping ^ * [q z(t,U))m{dt) is a random variable in Lp. Now since z is of weak bounded variation, each function t Â• Ez(t)x is bounded, for x Â€ Lg. By an application of Fubini 's Theorem we conclude that jj E[z(t)x]m(dt) = E(Jq z(t)m(dt))x, where Jz(t)m(dt) denotes the sample path integral whose existence was proved above. Since f 6 L(z), we know by Theorem 1.5.8 that E(J*J f(t)z(dt))x= jj f(t)E[z(dt)x], for each x Â€ Lq. Using the integration by parts theorem (see Rudin [26]) we have Jq f(t)E[z(dt)x] = [f(.)Ez(.)x]Jjj E[z(t)x]m(dt),
PAGE 102
96 for each x Â€ L_. Therefore E(Jq f(t)z(dt))x = E[[fz]J]x E(jJ z(t)in(dt))x = E[[fz]J jj z(t)m(dt)]x for each x G L . We conclude that jj f(t)z(dt) + jj z(t)in(dt) = [fz]J as desired, where the second integral is the sample path integral of z in Lp. D 4, Processes with Orthogonal Increments . In Section II. 4 we defined the concept of orthogonal increments for a function z:T Â» H, where H is a Hilbert space. In particular, then, when H = L2(n#F,P), the results of Section II. 4 apply. Recall that the inner product in LÂ» is defined by
PAGE 103
97 continuous functions n[g(t+l/n) g(t)]). Using the orthogonality of the increments of the Wiener process, he then showed by somewhat lengthy computation that jj g(t)z(dt) 1^ = jjg(t)!2dt. Using this identity and the fact that functions v;ith a E!j bounded derivative are dense in L (m^ ) , Wiener then extends the definition of the integr^ll to all of L (m, ) . This process of defining stochastic integrals was simultaneously simplified and generalized by Doob [13] . He replaced the functions with bounded derivative by step functions, and defined the integral with respect to any jorocess z:T L_ with orthogonal increments. The relation ejJ g(t)z(dt)2 = jlg(t)!2m^(dt), where m, is a nonnegative measure generated by z as in Section II. 4, and g is a step function, is central to the whole development. By the results of Section I I. 4 it follows that the WienerDoob integral in 1, is contained in our general stochastic integration theory. Using the relation (*) E(J^f(t)z(dt)) (Jgg(t)z(dt)) = J^^gf(t)iTt)m^(dt), which is valid for f,g Â€ L(m) = L (m, ) and A,B G S, it follows that if we define a new processes x:T > LÂ„ by X(t) = Jq f(T)z(dT), where f Â€ L(m), then x is a process with orthogonal increments. For if s, < t, ^ s < t , then we have E(x(t2) X(S2)) (x(t^)x(s^)) =E(Jg2f(t)z(dt)) ( J^^f (t) z (dt) ) = ;[s,,t,)n[s2,t2)f(^)l\(^^) =Â°
PAGE 104
98 More generally, if z is left .continuous, then m and z m, have countably additive extensions to Z, . Since the class of step functions is dense in L(z) by Theorem 1.8.3, relation (*) is still valid. It follov/s by this identity that the measure ij:E, L defined by M(A) = J^f(t)z(dt), A Â€ r,, is a measure with orthogonal values. Stochastic integrals with respect to processes with orthogonal increments have been used extensively in applications of probability theory to engineering. They are of particular value when the associated scalar measure m, can be determined, as is the case when z is a Wiener process. The measure m, allows us to compute the second moment of the stochastic integral of any f in L(z), by simply computing 1 ? the scalar integral fnl^l dm,. 6, Martingale Integrals . In this section we shall introduce the concept of conditional expectation, and use it to define an important class of random functions, the martingales. Using a deep result of Burkholder [8] , we show that every martingale z:T L is of weak bounded variation, provided 1 < p < 00. Consequently every such random function determines a class of integrable functions containing all uniform limits of step functions. Since unbounded measurable functions may also be integrable with respect to a particular random function, our integral is somewhat more general in this respect than the martingale integral of Millar [24] . The Millar integral is not restricted to scalarvalued integrands, however.
PAGE 105
99 Suppose X is a random variable in L, , and F, C F is a aalgebra. The indefinite integral of x is a Pcontinuous, count ably additive measure on F, ; by the RadonNikodym Theorem there is an integrable random variable z which is measurable with respect to F, , unique up to a Pnull set in F, , and satisfies r zdP = r xdP Ja J a for every A Â€ F, . The random variable z is denoted by E[xf, ], and is called the conditional expectation of x with respect to F, . The operator E['f, ] maps L, into L, , and can be shown to be a continuous, positive, linear contraction operator. Suppose that A is any set in F, . Then we have, for every B 6 Fj^, JgE[xX^lF^]dP= J^xI^dP = Jgl^EtxiF^ldP. It follows that E[xXj.If,] = X E[xJF,] almost everywhere. By linearity and continuity of the operator E[Â»F,], we conclude that for any function y that is measurable with respect to F, and such that E[xyF,] exists, we have E[xyF,] =yE[xF] almost everywhere. In particular, E[ylF, ] = yE[lF,] = y almost everywhere, so E[Â»f,] is a projection operator.
PAGE 106
100 Another useful property of the conditional expectation operator will now be established. Suppose that FÂ„ is another aalgebra, and F c F, g F. Then . J^E(E[xF2] iF^)dP = J^E[xF2]dP for every A Â€ F^, so we conclude that E[E[xF2] 1f^] = E[xF2] = E[E[xf^] IF2], up to some null set in FÂ„. Finally, since Ex = jQ^dP = J^E[xlF^]dP, we conclude that E[E[xF,]] = Ex. Let T = [0,1]. We now define the martingale property. 6.1 Definition . Suppose 1 i p ^ <= and zrT L is a random function. Suppose that (F(t):t Â€ T) is a family of aalgebras contained in F such that F(s) c F(t) if s ^ t. Then 2 (t) is a martingale with respect to (F(t):t e T) if z(t) is measurable with respect to F(t) for each t Â€ T, and if s < t, then (*) E[z(t)F(s)] = z(s) almost everywhere. We say for short that (z(t),F(t)) is a martingale. If (z(t),F(t)) is a martingale and t, < t2 < . . . < t^ is any finite set of points in T, then (z (t^) , F (t^) : 1 s: i 5: n) is a martingale in the sense that (*) holds for s and t restricted to the set { t, , ,t }. In this section we shall consider only realvalued random functions and scalar functions. An important result concerning conditional expectations is the conditional Jess en inequality. Ifg:RRisa convex
PAGE 107
101 function, if x Â€ L,, and if g (x) 6 L,, then for every aalgebra F, contained in F, we have g(E[xFJ) ^E[g(x)F,] almost everywhere. For a proof of this result see Chung [9].. Suppose that z:T Â» Lp is a martingale with respect to (F(t):t Â€ T) , Since t Itj^ is a convex function for every p s 0, it follows that for s ^ t, i2(s) P= lE[z(t)F(s)]P ^ E[!z(t)lPlF(s)], almost eveirywhere. Hence ez(s)^ ^ Ez(t)^, using the property E[E[xF]] = Ex of conditional expectations. We conclude that the family of pth absolute moments of a martingale in L is a nondecreasing family. Since EJzCt) 1^ ^ e1z(1) I^ for every t, we see that z is a bounded function from T to L_. P To show that any martingale z is actually of weak bounded variation when 1 < p < Â», we need to introduce the concept of a martingale transform. . 6.2 Definition . Suppose that (z ,F ) is a martingale in L , and (a ) is a sequence of real numbers. We say that a sequence (x ) of random variables is the martingale transform of (z ) by (a ) if for every n, n x^= i; a.[z. z. ,], where we set zÂ„ = 0. If (x ) is the transform of (z ) , then it follows that (x , F ) is a martingale. For x is measurable with respect
PAGE 108
102 to F , since each z., 1 ^ i ^ n, is measurable with respect to F . Moreover, ^ n+k E[x ., F ] = r a.E[z. z. , F ] ^ n+k' n" . , 1 *Â• 1 x1 ' n 1=1 n+k n = T. a.E[z.z. Jf]+ T. a.E[z.z. Jf] ... 1^1 11' n . , 1 1 11' n i=n+l 1=1 n+k n = E a. [z z ] + Z a. [z. z. ,] i=n+l ^^ ^ i=l ^^ ^^ = X almost everyv/here. The following theorem is due to Burkholder [8] . 6.3 Theorem . (Burkholder) For each p, 1 < p < <=, there is a constant KÂ„ such that if (x ) is the transform of (z_,F_) p n n n by a sequence (a^.) Â» with a  i. 1 for all n, then e1x^P ^ KpEz^P, n = 1,2,... . The proof of this result, which depends on showing, for each p, the equicontinuity of the sequence of bounded linear operators : n T^() = a3_E[.lFj_] + E a^{E[.lF^] E[.Fj^_^]}, will not be given. We can now state our final result. 6.4 Theorem. Suppose 1 < p < <Â» and z:T L is a martingale with respect to (F(t):t Â€ T) . Then (i) z is of weak bounded variation. (ii) If f Â€ L(z) then E(jJ f(t)z(dt)) = 0.
PAGE 109
103 (iii) If f Â€ L(z) and we define x(t) = Jq f(T)z(dT), for t e T, then (x(t), F(t):t Â€ T) is a martingale. Proo f. (i) Suppose that t^ < t< ... < t^^ is any finite family in T. Then (z(t.)# F(t.):l ^ i ^ n) is a martingale. Let aÂ„ . = 1 and aÂ„ . , =0 for 1 ^ i 5 n. Then 2i 2il El S z(t2^) z(t2i_l) P=El S a^[z(t^) z(t^_^)]P i=l i=l 5 KpEz(t2jP ^ KpElz{l)P < Â«, by Theorem 6.3. By Remark II. 2. 2, it follows that z is of weak bounded variation. (ii) Since z (t) = E[z(l)F(t)] for every t ^ T, we have Ez(t) = E[E[z(l) I F(t) ] ] = Ez(l) . Hence Em [a,b) = E [z (b) z (a) ] = z for every interval in T. It follows that for every step function f on T, eJÂ„ f(t)z(dt) = 0. Since the class of step functions is dense in L(z), (ii) holds. (iii) Since f Â€ L(z), there is a sequence of step functions f, such that the indefinite integrals of the f i^ ' s converge uniformly for A Â€ S to the indefinite integral of f. Define Xj^(t) = Jq f^(T)z(dT) for each k and every t Â€ T. Then lim x, (t) = x(t) for every t, the limit taking place in L . It suffices to show that each x^ is a martingale with respect to {F{t) , t Â€ T) . For if this is the case, then by continuity of the conditional expec
PAGE 110
104 tation operator, \ve have E[x(t)p(s)] =limE[x,(t)F(s)] k ^ = lim X, (s) = x(s) . k ^ Suppose then that f is a step function of the form n1 '^ ill """itirV * VttÂ„_3^,tj' where ^ t^ < t^^ < ... < t^ = 1 . If t 6 T, then x{t) = Jq f(T)z(dT) k1 = S a. [z (t^) 2 (t^_^)] + a^ [z (t) z (t^_^) ] , where t ^ [t , t^) . Nov/ suppose that s < t. Then k E[x(t)F(s)] = E a.E[z(t.) z (t . t)f(s)] i=l ^ ^ ^"^ k S a. [z(s) z(s)] +a.E[z(t.)z(t. ,) !f(s)] i=j+l ^ 3 J JL J1 + r a^[z(t^) 2(t^_^)] i=l = aj[E[z(tj) z(s) f(s)] +E[z(s)z(tj_j^) f(s)]] j1 + E a^[z(t^) z(t^_^)] i=l j1 = a. [z(s) z(t. ,)]+ S a. [z(t )z(t )] J '~i=l = x(s), where s Â€ [t._,,t.). We conclude that (x(t),F(t)) is a martingale. Q
PAGE 111
REFERENCES [1] R. G. Bartle, "A general bilinear vector integral," Studia Math. 15 (1956), 337352. [2] R. G. Bartle, N. Dunford and J. T. Schwartz, "Weak compactness and vector measures," Canadian J. Math. 7 (1955), 289305. [3] C. Bessaga and A. Pelczynski, "On bases and unconditional convergence of series in Banach spaces," Studia Math. 17 (1958), 151164. [4] J. K. Brooks, "On the existence of a control measure for strongly bounded vector measures," Bull. Amer. Math. Soc. 77 (1971), 9991001. [5] , "Contributions to the theory of finitely additive measures," (to appear). [6] J. K. Brooks and R. S. Jewett, "On finitely additive vector measures," Proc. Nat. Acad. Sci. 67 (1970), 12941298. [7] J. K. Brooks and H. S. Walker, "On strongly bounded vector measures and applications to extensions, decompositions and weak compactness," (to appear). [8] D. L. Burkholder, "Martingale Transforms, " Ann. Math. Stat. 37 (1966), 14941504. [9] K. L. Chung, A Course in Probability Theory , Harcourt, Brace and World, New York, 1968. [10] H. Cramer, "A contribution to the theory of stochastic processes," Proc. Second Berkeley Symp. (1951), 329339. [11] J. Diestel, "Applications of weak compactness and bases to vector measures and vectorial integration, " Rev. Roum. Math. Pures Appl., (to appear). 105
PAGE 112
106 N. Dinculeanu, Vector Measures ^ Pergamon Press, Oxford, 1967. J. L. Doob, Stochastic Processes , John Wiley, New York, 1953. N. Dunford, "Uniformity in linear spaces," Trans. Amer. Math. Soc. 44 (1938), 305356. N. Dunford and J. T. Schwartz, Linear Operators , Part I_, Interscience, New York, 1958. G. G. Gould, "Integration over vectorvalued measures," Proc. London. Math. Soc. 15 (1965), 193225. P. R. Halmos, Measure Theory , Van Nostrand, Princeton, 1956. J. E. Huneycutt, Jr., "Extensions of abstract valued set functions," Trans. Amer. Math. Soc. 141 (1969), 505513. , "Regularity of set functions and functions of bounded variation on the line," Rev. Roum. Math. Pures Appl. 14 (1969), 11131119. D. R. Lewis, "Integration with respect to vector measures," Pacific J. Math. 33 (1970), 157165, M. Loeve, Probability Theory , Third Edition, Van Nostrand, Princeton, 1963. E. J. McShane, "A Riemanntype integral that includes the LebesgueStieltjes, Bochner, and stochastic integrals," Mem. Amer. Math. Soc. 88 (1969). M. Metivier, "Stochastic integral and vectorvalued measure, " (to appear) . P. W. Millar, "Martingale integrals," Trans. Amer. Math. Soc. 133 (1968), 145166. C. E. Rickart, "Decompositions of additive set functions," Dioke Math. J. 10 (1943), 653665. W. Rudin, Principles of Mathematical Analysis , Second Edition, McGrawHill, New York, 1964. R. H. Shachtman, "The PettisStielt jes (stochastic) integral," Rev. Roum. Math. Pures Appl. 15 (1970), 10391060.
PAGE 113
107 [28] M. Takahashi, "On topologicaladditivegroupvalued measures," Japan Acad. Proc. 42 (1966), 330334. [29] N. Wiener, Nonlinear Problems in Random Analysis , M. T. Press, Cambridge, 1958 .
PAGE 114
BIOGRAPHICAL SKETCH Franklin P. Witte was born on May 30, 1942, in Davenport, Iowa. He attended parochial schools in Peoria, Illinois, and Conception, Missouri. In 1960 he graduated from high school at St. John Vianney Seminary, Elkhorn, Nebraska. From 1960 to 1964 he served in the U.S. Navy. He received his B.S. in Mathematics from the University of Missouri, Kansas City, Missouri, in 1957. As an undergraduate, he met his future wife Dorothy Silveirman; they were married in January, 1968. He worked at Midwest Research Institute in Kansas City for one year before enrolling as a gr.aduate student at the University of Florida in 1968. 108
PAGE 115
I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. / / .' Vi^'l^ , /j^CA. 'C _ Jazties K. Brook?^' 'chairman P^rofessor of Mathematics I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. 4: fC ^ Kim Assistant Professor of Mathematics I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. Zoran R. Pop$tojanovic Professor of Mathenyktics I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. a Stephen A. Saxon Assistant Professor of Mathematics
PAGE 117
I certify that I have read this study and that in my opinion it confomis to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. A k J APH ^ Arun K. Varma Associate Professor of Mathematics I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. l/aAU [P (v J^es W. Dufty ^sistant Professor o:^ Ph'ysics This dissertation was submitted to the Department of Mathematics in the College of Arts and Sciences, and to the Graduate Council, and was accepted as partial fulfillment of the requirements for the degree of Doctor of Philosophy. December, 1972 S^Vkfir^i Chairman, Department of Mathematics Dean, Graduate School
PAGE 118
e> V9. s.Â« ,y'J.

