Professional Documents
Culture Documents
S. J. Taylor - Introduction To Measure and Integration-University Press (1973)
S. J. Taylor - Introduction To Measure and Integration-University Press (1973)
MEASURE
AND INTEGRATION
BY
S. J. TAYLOR
Professor of Mathematics at Westfield College,
University of London
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521098045
A catalogue record for this publication is available from the British Library
CONTENTS
Preface page v
1 Theory of sets
1.1 Sets 1
1.2 Mappings 3
1.3 Cardinal numbers 5
1.4 Operations on subsets 9
1.5 Classes of subsets 14
1.6 Axiom of choice 19
3 Set functions
3.1 Types of set function 51
3.2 Hahn-Jordan decompositions 61
3.3 Additive set functions on a ring 65
3.4 Length, area and volume of elementary figures 69
PREFACE
There are many ways of developing the theory of measure and inte-
gration. In the present book measure is studied first as the primary
concept and the integral is obtained later by extending its definition
from the special case of `simple' functions using monotone limits. The
theory is presented for general measure spaces though at each stage
Lebesgue measure and the Lebesgue integral in Rn are considered as
the most important example, and the detailed properties are estab-
lished for the Lebesgue case.
The book is designed for use either in the final undergraduate year
at British universities or as a basic text in measure theory at the post-
graduate level. Though the subject is developed as a branch of pure
mathematics, it is presented in such a way that it has immediate
application to any branch of applied mathematics which requires the
basic theory of measure and integration as a foundation for its
mathematical apparatus. In particular, our development of the
subject is a suitable basis for modern probability theory - in fact this
book first appeared as the initial section of the book Introduction to
measure and probability (Cambridge University Press, 1966) written
jointly with J. F. C. Kingman.
The book is largely self-contained. The first two chapters contain
the essential parts of set theory and point set topology; these could
well be omitted by a reader already familiar with these subjects.
Chapters 3 and 4 develop the theory of measure by the usual process
of extension from `simple sets' to those of a larger class, and the
properties of Lebesgue measure are obtained. The integral is defined
in Chapter 5, again by extending its definition stage by stage, using
monotone sequences. Chapter 6 includes a discussion of product
measures and a definition of measure in function space. Convergence
in function space is considered in Chapter 7, and Chapter 8 includes
a treatment of complete orthonormal sets in Hilbert space. Chapter 9
deals with special spaces; differentiation theory for real functions of
a real variable is developed and related to Lebesgue measure theory,
and the Haar measure on a locally compact group is defined.
Starred sections contain more advanced material and can be
omitted at a first reading.
It will be clear to any reader familiar with the standard treatises
that this book owes much to what has gone before. I do not claim any
particular originality for the treatment, but the form of presentation
owes much to my experience of teaching this subject - at Birmingham
Vi PREFACE
University, Cornell University and the University of London - and I
readily acknowledge the stimulus received from this source. I am
grateful to Dr B. Fishel and Professor G. E. H. Reuter who made
helpful criticisms of an early draft, and to a great number of students
and colleagues who pointed out misprints and errors in the first
edition. However my main debt of gratitude is to Professor J. F. C.
Kingman who was co-author of the first edition of this book, and who
was much involved in every detail of it.
S.J.T.
London
December 1972
1
THEORY OF SETS
1.1 Sets
We do not want to become involved in the logical foundations of
mathematics. In order to avoid these we will adopt a rather naive
attitude to set theory. This will not lead us into difficulties because in
any given situation we will be considering sets which are all contained
in (are subsets of) a fixed set or space or suitable collections of such sets.
The logical difficulties which can arise in set theory only appear when
one considers sets which are `too big'-like the set of all sets, for
instance. We assume the basic algebraic properties of the positive
integers, the real numbers, and Euclidean spaces and make no attempt
to obtain these from more primitive set theoretic notions. However,
we will give an outline development (in Chapter 2) of the topological
properties of these sets.
In a space X a set E is well defined if there is a rule which determines,
for each element (or point) x in X, whether or not it is in E. We write
x r: E (read `x belongs to E') whenever x is an element of E, and the
negation of this statement is written x 0 E. Given two sets E, F we
say that E is contained in F, or E is a subset of F, or F contains E
and write E c F if every element x in E also belongs to F. If E C F
and there is at least one element in F but not in E, we say that E is a
proper subset of F.
Two sets E, F are equal if and only if they contain the same ele-
ments; i.e. if and only if E c F and F E. In this case we write
E = F. This means that if we want to prove that E = F we must prove
both x E E x E F and x E F x r: E (the symbol should be read
`implies').
Since a set is determined by its elements, one of the commonest
methods of describing a set is by means of a defining sentence: thus
E is the set of all elements (of X) which have the property P (usually
delineated). The notation of `braces' is often used in this situation
E = {x: x has property P}
but when we use this notation we will always assume that only
elements x in some fixed set X are being considered-as otherwise
logical paradoxes can arise. When a set has only a finite number of
2 THEORY OF SETS [1.1
elements we can write them down between braces E = {x, y, z, a, b}.
In particular {x} stands for the set containing the single element x.
One must always distinguish between the element x and the set {x},
for example, the empty set 0 defined below is not the same as the class
{ 0 } containing the empty set.
Cartesian product
Given two sets E, F we define the Cartesian (or direct) product E x F
to be the set of all ordered pairs (x; y) whose first element x E E and
whose second element y e F. This clearly extends immediately to the
product El x E2 x ... x E. of any finite number of sets. In particular
it is immediate that Rn, Euclidean n-space, is the Cartesian product
1.11 SETS 3
of n copies of R. For an infinite indexed class {Ej, i E I} of sets, the
product II El is the set of elements of the form {as, i E I} with aj E Es
iEI
for each i E I.
Exercises 1.1
1. Describe in words the following sets:
(i) {t a R: 0 5 t S 1};
(u) {(x, y) E R2: x2+ y2 S 1};
(iii) {k E Z: k = n2 for some n r: Z};
(iv) {keZ:nj k=> n= 1 or k};
(v)
(vi) {B: B c E}.
2. Show that the relation c is reflexive and transitive, but not in gen-
eral symmetric.
3. The sets X x (Y x Z) and (X x Y) x Z are different but there is a
natural correspondence between them.
4. Suppose x is an element of X and A = {x}. Which of the following
statements are correct : x e A, x e %, x e A, x c I, A E %, A c %, A e x?
5. Suppose P(a) and Q(a) are two propositions about the element such
that P(a) . Q(a). Show that {a: P(a)} c {a: Q(a)}.
1.2 Mappings
Suppose A and B are any two sets: a function from A to B is a
rule which, for each element in A, determines a unique element in B.
We talk of the function f and use the notation f : A -+ B to denote a
function f defined on A and taking values in B. For any x E A, f (x)
means the value of the function f at the point x and is therefore an
element of the set B: we therefore avoid the terminology (common
in older text books) `the function f(x)'. The words mapping and
transformation are often used as a synonym for function.
For a given function f : A B, we call A the domain of f and the
subset of B consisting of the set of values f (x) for x in A is called the
range of f and may be denoted f (A). When f (A) = B we say that f
is a function from A onto B. Given a function f : A -> B, by definition
f (x) is a uniquely determined element of B for each x e A; if in addition
for each y in f (A) there is a unique x e A (we know there is at least
one) with y = f(x) we say that the function f is (1, 1). Another shorter
way of saying this is that f : A -> B is (1, 1) if and only if for x1, x2 E A,
x1 4 x2=f(x1) 4f(x2)
4 THEORY OF SETS 11.2
Given f : A -> B there is an associated f : sad -* -4, where .sad is the class
of all subsets of A and .4 is the class of all subsets of B, defined by
f(E) = with y = f(x)}
for each E c A. (the symbol 3 should be read, `there exists': i.e. the
set described by {x E E: y = f (x)) is not empty). There is also a function
f-1: -4 -> &I defined by
f-1(F) = {xEA:f(x)EF},
for each F - B. The set f-1(F) is called the inverse image of F under f.
Note that if yEB-f(A), then the inverse image f-1({y}) of the one
point set {y} is the empty set. If f : A -> B is (1,1) and Y E f (A), then
it is clear that f -1 ({y}) is a one point subset of A, so that in this case
(only) we can think off-' as a function from f (A) to A. In particular,
if f: A -* B is (1, 1) and onto there is a function f-1: B -- A called the
inverse function off such that f-1(y) = x if and only if y = f(x).
Now suppose f: Al B, g: A2 -- B are functions such that A, ' A2
and f (x) = g(x) for all x in A 2: under these conditions we say that f
is an extension of g (from A2 to A1) and g is the restriction of f (to A2).
For example, if
g(x) = cos x (x E R);
f(x + iy) = cos x cosh y + i sin x sinh y (x +iyEC);
then f: C --> C is an extension of g: R --> C from R to C, and the usual
convention of designating both f and g by 'cos' obscures the differences
in their domains.
If we have two functions f: A -* B, g: B -a C the result of applying
the rule for g to the element f (x) defines an element in C for all x E A.
Thus we have defined a function h: A -+ C which is called the composi-
tion off and g and denoted g of or g(f). Thus, for x E A
h(x) = (g of) x = g(f(x)) E C.
Note that, if f : A -> B is (1, 1) and onto we could define the inverse
function p l: B --> A as the unique function from B to A such that
(fof-') (y) = y for all yEB,
(f-1 of) (x) = x for all x E A.
Sequence
Given any set X a finite sequence of n points of X is a function from
{1, 2,..., n} to X. This is usually denoted by xl, x2, ..., xn where
xi c X is the value of the function at the integer i. Similarly, an infinite
1.21 MAPPINGS 5
sequence in X is a function from Z to X (where Z is the set of positive
integers). This is denoted x1, x2, ..., or {xi} (i = 1, 2, ...), or just {xi}
where xi is the value of the function at i, and is called the ith element
of the sequence. Given a sequence {ni} of positive integers (that is, a
function f : Z -+ Z where f(i) = ni) such that ni > nn for i > j, and a
sequence {xi} of elements of X (a function g: Z --* X) it is clear that the
composite function g of: Z X is again a sequence. Such a sequence
is called a subsequence of {xi} and is denoted {xn,} (i = 1, 2,...). Thus
{x.,} is a subsequence of {xi} if ni E Z for all i E Z, and i > j = ni > n p
We can think of a sequence as a point in the product space Ij Xi
i=i
where Xi = X for all i. More generally a point in the product space
11 Xi with X i = X for i E I can be identified as a function f : I -+ X.
iel
Exercises 1.2
1. Suppose f : R R is defined by f (x) = sin x. Describe each of the
following sets:
f-1{0}, f l{1}, f-1{2}, f-1{y:0 <, y <
2. Suppose f : A - . B is any function. Prove
(i) E c f-1(f(E)), for each E c A;
(ii) F f(f-1(F)), for each F e B;
and give examples in which there is not equality in (i), (ii).
3. Suppose f : A -* B, g : B -+ C are functions and h = g of: show that
h-1(E) = f-1[g-1(E)] for each E e C.
4. If AcBCC,f:A-*X,g:B-+X,h:C-+X are such that his an
extension of g and g is an extension off, prove that f is the restriction of h
toA.
5. Show that the restriction of a (1,1) mapping is (1, 1).
6. Suppose m, n E Z, A is a set with m distinct elements and B is a set
with n distinct elements. How many distinct functions are there from A
to B?
map Z onto a proper subset of Z and are (1, 1). Instead we say that the
cardinal of a set A is less than the cardinal of the set B if there is a
subset B1 cz B such that A - B1 but no subset Al c A such that
Al - B.
From this definition of ordering we consider the following state-
ments, where m, n, p denote cardinals
(i) m<n,n<p=>.m<p;
at most one of the relations m < n, m = n, n < m holds so
(ii)
that m < n,n < m=> m=n.
(iii) at least one of the relations m < n, m = n, n < m holds.
Now (i) follows easily from the definition, for let M, N, P be sets with
cardinals m, n, p and suppose N1 c N, P1 c P with M - N1, N - P1.
The mapping f: N -. P1 when restricted to Nl gives an equivalence
N1-P2cP1 sothat M,P2cP. Further if P-1111cMthe map-
ping g: M -> Nl when restricted to M1 shows P -M1 - N2 c N which
contradicts n < p. (ii) can also be deduced from the definition (see
exercise 1.3 (5)), though this requires quite a complicated argument:
(ii) is known as the Schroder-Bernstein theorem. However, the truth
of (iii)-that all cardinals are comparable-cannot be proved. without
1.31 CARDINAL NUMBERS 7
the use of an additional axiom (known as the axiom of choice) which
we will discuss briefly in § 1.6. If we assume the axiom of choice or
something equivalent, then (iii) is also true.
A set of cardinal X. is said to be enumerable. Thus such a set
A - Z so that the elements of A can be `enumerated' as a sequence
a1, a2, ... in which each element of A occurs once and only once. A set
which has a cardinal m 5 No is said to be countable. Thus E is countable
if there is a subset A c Z such that E - A, and a set is countable if it
is either finite or enumerable.
Given any infinite set B we can choose, by induction, a sequence
{bi} of distinct elements in B and if B1 is the set of elements in {bi}
the cardinal of B1 is No. Hence if m is an infinite cardinal we always
have m > No. By using the equivalence
bi+-+b2i
between B1 and the proper subset B2 B1 where B2 contains the even
elements of {bi} and the identity mapping
b<-+b for
we have an equivalence between B = B1 v (B - B1) and B2 V (B - B1),
a proper subset of B. This shows that any infinite set B contains a
proper subset of the same cardinal.
In order to see that some infinite sets have cardinal > No it is
sufficient to recall that the set {x E R: 0 < x < 1} cannot be arranged
as a sequence.4 Now it tan-1 x + I = f (X), x E R defines a mapping f :
R -a (0, 1) which is (1, 1) and onto so that R has the same cardinal as
the interval (0, 1) and we have c > No. It is worth remarking that a
famous unsolved problem of mathematics concerns the existence or
otherwise of cardinals m such that c > m > No. The axiom that no
such exist, that is that m > No = m >, c is known as the continuum
hypothesis.
The fact that there are infinitely many different infinite cardinals
follows from the next theorem, which ccmpares the cardinal of a set
E with the cardinal of the class of subsets of E.
Theorem 1.1. For any set E, the class (f = (E) of all subsets of E
has a cardinal greater than that of E.
Proof. For sets E of finite cardinal n, one can prove directly that
the cardinal of '(E) is 2n, and an induction argument easily yields
n < 2n for n E Z. However, the case of finite sets E is included in the
general proof, so there is nothing gained by this special argument.
t See, for example, J. C. Burkill, A First Course in Mathematical Analysis (Cam-
bridge, 1962).
8 THEORY OF SETS 1.3
Suppose 2 is the class of one points sets {x} with x e E. Then
2 c ' and E - 2 because of the mapping x H {x}. Therefore it is
sufficient to prove by (ii) above, that ' is equivalent to no subset
El c E. Suppose then that g ' -* El is (1, 1) and onto and let
x: El -> W denote the inverse function. Let A be the subset of El
defined by
A = {x e El, x x(x)}.
Then A E 6 so that c(A) = xc E El. Now if x0 a A, x(xc) = A does not
contain x0 which is impossible, while if x0 0 A, then x0 is not in x(xo)
so that x0 E A. In either case we have a contradiction.
It is possible to build up systematically an arithmetic of cardinals.
This will only be needed for finite cardinals and No in this book, so
we restrict the results to these cases and discuss them in the next
section.
Exercises 1.3
1. Show that (0,1] .. (0,1) by considering, defined by
f(x)=I-x, for lj<x.1;
=I-x, for J < x < ,J;
=I-x, for }<x<,j;
2 -x, for 2n<x.ri
Deduce that all intervals (a, b), (a, b], [a, b] or [a, b) with a < b have the same
cardinal c.
2. Every function f : [a, b] --> R which is monotonic, i.e.
a <xl <x2 <b=t-f(xi) 4f(x2),
is discontinuous at the points of a countable subset of [a, b].
Hint. Consider the sets of points x where the size of the discontinuity
d(x) = f(x+0)-f(x-0) satisfies 1/(n+l) < d(x) < 1/n and prove this is
finite for all n in Z.
3. Show that R2 - R.
Hint.
Fig. 1
U Ai,
i=1
fl Ai,
i=1
U Ai,
i=1
fl Ai
i=1
UEi=E'n, in
i=n
flEi=flEE
i=1
for all n,
Exercises 1.4
1. Prove each of the following set identities:
A-B=A-(AnB)=AvB-B,
A n (B-C) = AnB-A n C,
(A-B)-C = A-(BvC),
A-(B-C) = (A-B)v (An C),
(A -B) n (C-D) =AnC-BvD,
Et (FAG) = (EAF)AG,
En(FAG) _ (EnF)A(EnG),
EA 0 =E,EAX =X-E,
EAE= s,EO(X-E)=X,
EA F = (Eu F)-(En F).
2. With respect to which of the operations L, v, n does the class of all
subsets of X form a group?
3. Show that E c F if and only if X - F c X - E.
4. Prove that A A B= C A D if and only if A A C= BL D, by showing
that either equality is equivalent to the statement that every point of X
is in 0, 2 or 4 of the sets A, B, C, D.
5. Show that if Il e I2, then
nEaDnEa, UE.CUEa.
aEIt aEI2 aEIt aEll
6. A real number is said to be algebraic if it is a zero of a polynomial
an xn +an-1x°`-1 + ... -- ao where the coefficients ai are integers. Defining
the `height' of a polynomial to be the integer
h = n+lanl+Ian-1I +...+laol,
show that there are only finitely many polynomials of height h, and deduce
that the set of all algebraic real numbers is enumerable. Deduce that the
14 THEORY OF SETS 11.4
set of transcendental numbers (real numbers which are not algebraic) has
cardinal > Mo.
7. Show that in Rn, the set Qn of points (x1, x2, ..., xn), where each co-
ordinate x= is rational, is an enumerable set.
Further, the class of all spheres with centres at points of Qn and rational
radii is an enumerable class.
8. Show that any sequence of disjoint sets converges to 0. Show that
{En} is a convergent sequence if and only if there is no point x of X such that
each of x e En, x e X - En holds for infinitely many n.
Suppose
0 < x 5 1- (1/n)} n odd,
E. {x:
{X: (1/n) 5 x < 1} n even;
show that {En} converges but is not monotone.
9. For any sequence {En} of sets prove
(i) lim sup En, lim inf E. are unaltered by the omission or alteration of
any finite number of sets in the sequence.
(ii) for any set F,
F-lim sup R. = liminf(F-En),
F - lim inf En = lim sup (F - En).
10. If En = A for n even, En = B for n odd, show that lim sup En = A v B,
liminfEn = An B.
11. Can an uncountable union of distinct sets be countable?
12. If {En} is a sequence of sets and
D1 = E1, D .n = Dn_10 En for n=2,3,...,
show that the sequence {DJ converges to a limit if and only if lim En = o.
13. Show that XE(x) < xF(x) for all x in X if and only if E c F. Suppose
A=EvF,B=EnF,C=E/ F: show that
xB=XE*XF' xC= IXE-XFI'
XA=xE+XF-XB,
Generalise the first two of these identities to finite unions and inter-
sections.
14. If xn is the indicator function of E. (n = 1, 2, ...) and A = lim sup En,
B = lim inf En, show that, for all x in X,
XA(x) = lim sup xn(x), XB(x) = lim inf xn(x)
n-4- OD n-4. OD
2. Ring
This is any non-empty class . of subsets such that
A,BER=>AnBEP and ALBEPP.
Since 0 = A A A, A v B= (A A B) A (A n B), and A- B= A A (A n B)
we see that a ring is a class of sets closed under the operations of union,
intersection, and difference and QS E R. Thus a ring is certainly also a
semi-ring. As examples the system { o, X} is a ring as is the class of all
subsets of X. However, the class 9 of half-open intervals in R is not
a ring, for it is not closed under the operation of difference.
4. Sigma ring
A ring . is called a if it is closed under countable unions, i.e.
if 00
ALB= i=1
U UDik
k=1
Um (UEk;)
j=1 k=1
so that the system 2 is also closed under the operation of taking the
symmetric difference.
Example. We have already seen that J', the class of intervals
(a, b] in R, is a semi-ring. The generated ring is the class off of finite
unions of disjoint half-open intervals. off is called the class of elemen-
18 THEORY OF SETS 11.5
tary figures in R. Similarly, the elementary figures in Rn form the
class Pn of finite disjoint unions of half-open rectangles from 6pn.
The next theorem is often important in proving that a given class is
a o--ring.
Theorem 1.5. If 9 is any ring, the monotone class _W(M) generated by
. is the same as the o--ring.9'(M) generated by M.
Corollary. Any monotone class .4' which contains a ring q contains
the .c"(.) generated by M.
Proof. Since a a-ring is always a monotone class and .9'(.) M
we must have Y (M) (9), denoted by .4'.
Hence it is sufficient to show that .,K is a o--ring, and this will follow
if we can prove that .4' is a ring. For any set F, let 2(F) be the class
of sets E for which E - F, F - E, E v F are all in .4'. Then if 2(F)
is not empty it is easy to check that it is a monotone class. It is clear
that 2(F) R for any F e 9 so that 2(F) .4'. Hence, EE.%',
FE. E e 2(F) . F E 2(E) by the symmetry of the definition of the
class 2, and it follows that ..Ill c 2(E) since 2 is a monotone class.
But the truth of this for every E E .4' implies that .4' is a ring. I
In § 1.2 we discussed mappings f : X -* Y and saw that any such
mapping induced a set mapping f-1 on the class of all subsets of Y.
If f-1 is restricted to a special class IF of subsets in Y, then the image of
' under f-I will be a class of subsets in X. The interesting thing is that
the structure of the class 'C is often preserved by such a mapping f-1.
Theorem 1.6. Suppose s' is a z-class of subsets of Y, f: X Y is any
mapping and f-1 (io) denotes the class of subsets of X of the form f-1(E),
EE'. Then f-1(co) is a z-class of subsets of X.
Proof. It is easy to check that the mapping!-': 'f --> f -'(W) commutes
with each of the set operations union, symmetric difference, countable
union and monotone limit. The closure of ' with respect to any of
these operations therefore implies the closure off-'(W) with respect to
the same operation. I
Exercises 1.5
Partial ordering
Suppose V is a set with elements a, b, ... and -< is a relation defined
between some but not necessarily all pairs a, b E V such that
(i) -< is transitive, i.e. a -< b, b -< c a -< c;
(ii) -< is reflexive, i.e. a -< a for all a in V;
(iii) a-<b,b-< a=> a=b;
then V is said to be partially ordered by the relation -<. V is said to be
simply (or totally) ordered if,
(iv) for each pair a, b E V at least one of a -< b, b -< a is valid.
Any partial ordering in a set V induces automatically a partial
ordering in every subset of V. If W V and the induced ordering in
W is a simple ordering, then W is said to be a chain in V.
For example, in R the usual S relation defines a total ordering of
R. However, in R2, if we say (xi, Yi) -< (x2, Y2) if and only if yi < y2
and xi 5 x2 we have an example of a partial ordering which is not
simple. A more useful example is the class ' of all subsets of a fixed
set X with A -< B meaning A c B.
1.6] AXIOM OF CHOICE 21
A chain W in a partially ordered set V is called a maximal chain if
it is not possible to obtain a larger chain by the addition of an element
in (V - W). We can now state
Kuratowski's lemma. Every partially ordered Bet V contains a maxi-
mal chain.
This means that there is a totally ordered subset W c V such that
for every x e V - W, there is some element yEW such that neither of
x -{ y, y < x is true.
For a partially ordered set V, the element a is said to be an upper
bound for the subset C c V if c -< a for every c E C. The element a
is said to be the least upper bound or supremum of the subset C if
(i) a is an upper bound for C;
(ii) if b is an upper bound for C, then a -< b.
It is easy to check that, in any partially ordered set V, it is impossible
for two distinct elements al, a2 to satisfy the above conditions (i),
(ii) so that the supremum of a set C is unique when it exists. However,
even when a set V is totally ordered, not all its subsets need have a
supremum. With the usual ordering R has the property that any
non-empty subset C which is bounded above has a supremum (this
is known as the least upper bound axiom), but Q does not have this
property.
Finally, we say that the element m e V is a maximal element of V
if m -< a . m = a. We can now state
Zorn's lemma. If V is partially ordered and each chain W in V has a
supremum, then V has a maximal element.
Both Zorn's lemma and Kuratowski's lemma can be deduced from
the axiom of choice, fi but we will not give the details as these are
complicated and outside the mainstream of our argument. However,
,the next theorem shows that, if we assume Zorn's lemma as an
additional axiom, then both the axiom of choice and Kuratowski's
lemma will be valid. This means that, in our subsequent work, we will
assume whichever of these three results happens to be most convenient.
Theorem 1.6. The statements of (A) Kuratowski's lemma and (B) Zorn's
lemma are equivalent. Either of them implies (C) the axiom of choice.
Proof. (A) . (B). Suppose W is a maximal chain in V, then by the
hypothesis of (B) there is a supremum m for W so that a -< m for all
a E W. If m is not a maximal element of V, then there is a be V such
that b 4 m and m -< b. Then b is not in Was this would imply b -< m,
t For a discussion of these and other axioms equivalent to axiom of choice see, for
example, J. L. Kelley, General Topology (Van Nostrand, 1955).
22 THEORY OF SETS [1.6
and b = m. Hence we may add b to the chain W and the new set ob-
tained is still a chain. This would contradict the fact that W is a
maximal chain.
(B) . (A). The chains in V form a class f which is partially ordered
by inclusion. If now VYis a chain in ' with elements W (each of which
is a chain in V), then the union U {W: W E *Y} is a chain in V so that
it is an element of ' which can only be the supremum of 0. Hence
by hypothesis ' contains a maximal element, i.e. V contains a maxi-
mal chain.
(B) (C). We now suppose given a class .NV of sets E. There are
clearly some subsets (in fact any finite subset) . c .'V on which
it is possible to define a choice function g: . --> U {E,: Ea E .} such
that g(Ea) E E. The set V of all such functions g is therefore non-
empty and it is partially ordered if we say g1-< g2 if gl is defined on
., 92 is defined on X., . c X. and g1(Ea) = g2(Ea) for Ea E
(i.e. g2 is an extension of g1). If now W is a chain in V containing func-
tions gi defined on M, the supremum of W is the function defined
on U which has the value gi(Ea) on any set E. E .. If we now
assume (B) it follows that the set V has a maximal element f. Then this
function f must be defined on all the sets Ea, for otherwise if f is not
defined on E1 we could choose an element x1 E E1, put f (El) = x1 and
this would be a proper extension off and therefore contradict the fact
that f is maximal.'
Exercises 1.6
Open sphere
In a metric space (X, p), if x c X, r > 0, then
S(x,r) = {y:p(x,y) < r};
Open set
A subset E of a metric space X is said to be open if, for each point x
in E there is an r > 0 such that the open sphere S(x, r) c E. Note
that the open spheres defined above are examples of open sets since
y E S(x, r) = p(x, y) = r1 < r,
so that, for 0 < r2 <, r - r1, S(y, r2) c S(x, r).
Theorem 2.1. In a metric space x, the class 9 of open sets satisfies
(1) 0,XETJ;
(ii) A1,A2,...,A.ET=> nAEV;
i=1
(iii) A. E V for a in I U A. c!?.
aEI
Proof. (i) Since any statement about the elements of 0 is true, 0 E 9,
and it is clear that S(x, r) c X for any x E X, r > 0 so certainly X E 9.
n
(ii) If x E fl Ai, then x E Ai for i = 1, ..., n and each Ai is open
i-1
so there are real numbers ri > 0 for which S(x, ri) c A. If we
put r = min ri, then 0 < r < ri so that S(x, r) S(x, rj) c Ai for
1-<i<n
n
i = 1,...,n; and S(x,r) c (1 Ai.
i=1
Remark. The condition (ii) says that 9 is closed for finite inter-
sections, while (iii) says it is closed under arbitrary unions. One
2.11 METRIC SPACE 25
cannot extend (ii) to give closure for infinite intersections for, in R
the intervals (0, 1 + (1/n)) are open sets, but
00
fl (0,1+11
n=1\ n
={x:0<x51}=(0,1]
is not open as it contains no open sphere centre 1.
It is more general to start with a set X and a class V of subsets
of X satisfying (i), (ii), (iii) of theorem 2.1 and to call these `the
open sets' in X. Such a class 9 and set X are said to form a topological
space, and 9 is said to determine the topology in X. A topological
space (X, 9) is said to be metrisable if there is a distance function p
defined on it which determines the class 9 for its open sets. Most topo-
logical spaces (X, T) of interest satisfy the rather weak conditions
which are sufficient to ensure metrisability, so that little is lost by
assuming in the first place that we have a metric space (X,p). Of
course two different metrics p1, p2 on a set X may define the same class
V of open sets, so that even when a topological space is metrisable,
the metric p is not uniquely determined-see exercise 2.4 (1).
In this chapter we will define most of the further concepts. which
depend on the toplogy of X in terms of the class 9 of open sets in X:
this means that the definitions will make sense either in a metric space
(X, p) or in a topological space (X, V). However, when it simplifies the
proof, we will assume that X has a metric p determining ( and use
this metric, so that some theorems will be stated and proved for metric
spaces even though they are true more generally.
Closed set
Limit of a sequence
Given a sequence {xi} of points in a metric space (X, p) we say that
the sequence converges to the point x E X if each neighbourhood of
x contains all but a finite number of points of the sequence. Thus
{xi} converges to x if given e > 0, there is an integer N such that
i>N=> p(x,xi)<e.
We then write x = lim xi or x = lim xi and say that x is the limit of the
sequence {xi}. Note that, in a metric space, the limit of a sequence is
unique-see exercise 2.1 (7).
In a metric space X, given a point x and a set E, the distance from
x to E, denoted by d(x, E) is defined by
d(x, E) = inf{p(x, y): yeE}.
This is always defined since {p(x, y) : y e E} is a set of non-negative
real numbers. If E c S(x, r) for some open sphere, we say that E is
bounded and define the diameter of E, denoted diam (E), by
diam (E) = sup {p(x, y) : x, y e E}.
Remark
Many readers will be familiar with the concepts of this section for
R and R2. Usually the proofs given in these special cases can be
generalised to a general metric space, and often even to a topological
space. The reader who has difficulty in working in an abstract situa-
tion should visualise the argument in the plane R2, but not use any of
the special properties of R2.
2 TIT
28 POINT SET TOPOLOGY [2.1
Exercises 2.1
1. In any set X, the class '' of all subsets of X satisfies the conditions for
a class of open sets. Show that this topology can be generated by the metric
p(x, y) = 1 for x $ y, x, y E X.
(This is called the discrete topology in X.) At the opposite end of the scale,
the indiscrete topology in X is that for which 9 = { 0, X}: in this case the
space is non-metrisable if X contains at least two points.
2. Show that in a metric space X containing at least 2 points
(i) single point sets {x} are closed, but they are open only if
d(x, X -{x}) > 0;
(ii) finite sets are closed;
(iii) any open set G is the union of the class of open spheres contained
in 0;
(iv) any open set 0 is the union of the class of closed spheres contained
in G.
3. In the topological space X, given a set E -- X a point x is said to be
an interior point of E if there is a neighbourhood N of x with N c E. Prove
(i) the set E° consisting of the interior points of E is open;
(ii) E is open if and only if E = E°.
4. In a topological space X show that
A1 v A2v...v A = A1v A2v ...
but this does not extend to arbitrary unions. Give an example in which
EAF+EnF.
5. If X is a 2-point space {x1, x2}, p(xl, x2) = 1 show that (X, p) is a metric
space in which the closure S(x1,1) of the open sphere S(x1,1) is not the same
as the closed sphere S(x1,1). However, if X is a normed linear space (see
§2.6) then S(x, r) = S(x, r).
6. Suppose A c R and A is closed and bounded below. Show that the
infinum of A is an element of A.
7. In a metric space, suppose is a sequence converging to x and E
is the set of points in this sequence. Show
(i) every subsequence converges to x;
(ii) either x is the only limit point of E or there is an integer N such that
x = x for n > N. Deduce that a sequence {xn} cannot converge to two
different limit points.
8. Suppose E is the set of points in a sequence in a metric space and
x is a limit point of E. Show there is a subsequence {x t} which converges
to X.
2.11 METRIC SPACE 29
9. For any set E in a metric space, show that
E = {x: d(x, E) = 0}.
10. If E, F are subsets of a metric space X, x, y e X, show
(i) p(x,y) > Jd(x,E)-d(y,E)I;
(ii) p(x,y) <d(x,E)+diam(E), if ycE;
(f) Jp(x, y1) -p(x, y2) < diam (E), if y1, y2 E E;
(iv) d(E, F) = d(F, E).
Is d a metric on the space of subsets of X?
11. In R show that a bounded open set is uniquely expressible as a
countable union of disjoint open intervals.
Hint. For each x E E, put a = inf {y: (y, x) c E}, b = sup {y: (x, y) C E};
and show that the open interval It,, = (a, b) contains x, is contained in E and
is such that any open interval I satisfying x E I c E satisfies I C I,,. Deduce
that for x1, x2 E E, either Ixi = Ix9 or I xl n Ixs = o, so that E _ U I,, is a
xEE
disjoint union. Enumerate the intervals Ix by considering those of length
greater than 1/n (n = 1, 2, ...).
In Rn (n > 2) show that a bounded open set can be expressed as a dis-
joint union of a countable number of half-open rectangles in Yn (but that
this expression is never unique). Show that in general an open set in Rn
(n > 2) cannot be expressed as a disjoint union of open spheres, or of open
rectangles.
(xn, -xni_1)
Covering systems
A class ' of subsets of X is said to cover the set E c X or form a
covering for E, if E c U {S: S Eon }. If all the sets of ' are open, and
le covers E, then we say that' is an open covering of E.
Compact set
A subset E of X is said to be compact if, for each open covering c'
of E, there is a finite subclass W. c' such that' 1 covers E. For ex-
ample, the celebrated Heine-Borel theorem states that any finite
closed interval [a, b] is compact. Though this is proved in most
elementary text-books we include a proof which starts from the least
upper bound property.
Lemma. If a, b are real numbers, the closed interval
[a,b]={x:aex<b}
is compact.
t The fact that absolute convergence implies convergence is a consequence of the
least upper bound axiom for R, that is, it follows from the assumption that every non.
empty subset of R which is bounded above has a supremum. (See, for example,
Burkill, A First Course in Mathematical Analysis, Cambridge, 1962.)
2.21 COMPLETENESS AND COMPACTNESS 31
Proof. Let le be any open cover of [a, b] and let c be the supremum
of the set of x in [a, b] for which some finite subfamily W1 c W covers
[a, x]. (This set is non-void since it contains a.) Choose a set GEW
with c e G and choose a point d E (a, c) such that the closed interval
[d, c] c G. Then there is a finite subfamily covering [a, d] and the
addition of 0 to this family gives a finite subfamily covering [a, c].
But unless c = b, since G is open, we have covered by a finite subfamily
the interval [a, e] for some e > c which contradicts the choice of c. 3
It is also possible to prove directly that any closed rectangle
in Rn is compact, but we will be able to deduce this from theorem 2.6.
We can use this to show that, in Rn, every closed bounded set is com-
pact. This will follow from the following:
Lemma. If E is compact, and F is a closed subset of E, then F is
compact.
Proof. Suppose le is an open covering for F. Then', together with
(X-F), which is open, forms an open covering for E. This has a
finite subcovering (1 of E and 6 n W, must be a finite subclass of
which covers F. J
It is not true in a general metric space that closed bounded subsets
are compact-see exercise 2.2 (3). However, we can prove:
Lemma. In a metric space X, every compact subset is closed and bounded.
Proof. If E is not closed, there is a point x0 E X which is a limit point
of E but is not in E. For every x E E, put S,, = S(x, r) with r = Jp(x, x0).
Then the collection of all such open spheres covers E, but every finite
subclass S(xl, r1), ..., S(xn, rn) has a void intersection with S(xo, r),
where r = min ri and so cannot cover E, for S(xo, r). contains points
16i<n
of E. On the other hand, if E is not bounded, the class of open
spheres of radius 1 and centres in E covers E, but no finite subclass
can cover E. J
Whenever the whole space X is compact, we talk of a compact
space. The above lemma shows that Rn is not compact because it
is not bounded. A space X is said to be locally compact if every point
x in X has a neighbourhood N such that N is compact. It is clear that
Rn is locally compact.
There are various other properties in a topological space which are
equivalent to compactness under suitable conditions.
Weierstrass property
A set E is said to have property (W) if every infinite subset of E
has at least one limit point.
32 POINT SET TOPOLOGY [2.2
so that the class of sets Fa, a e I has not got the finite intersection
property. This proves that compactness implies that any class sad
of closed subsets with the finite intersection property has a non-void
intersection. Conversely suppose a closed set E is such that any class
.d of closed subsets with the finite intersection property has a non-
void intersection, and suppose Ga, a E I is an open covering of E,
so that (1(X - Ga) = Q1. If E is closed E n (X - Ga), a E I is a family of
aEI
closed subsets of E, so there must be a finite set al, a2, ..., an such that
n
n (X - Gai) n E _ 0 . This means that Gal, ... , Gan form a finite sub-
covering for E.
(ii) Suppose first that E has not got property (W). Let A be an
infinite subset of E with no limit point. If A is not enumerable, choose
an enumerable subset B of A. Then B is closed and (X - B) is open.
Enumerate B as a sequence of distinct points {xi}, and for each xi
choose a neighbourhood Ni which contains xi but no other point of B.
2.21 COMPLETENESS AND COMPACTNESS 33
Then the sequence {Ni} together with the set (X - B) form an open
covering of E, which has no finite subcovering as none of the open
sets Ni can be omitted without `uncovering' the corresponding xi.
Conversely suppose E has property (W). Then there is a finite class
W,, of spherical neighbourhoods of radius 1/n which covers E; for
otherwise we could find an infinite subset of E all of whose points were
distant more than 1/n apart and such a subset can have no limit
w
point. Let' = U Wo n; so that ' is countable. Now if G is any open set,
n-1
for each x in G n Ewe can find a sphere S E 'o containing x with S c G:
for we can first choose S > 0 so that S(x, S) c G and if n > 2/4, the
sphere of Wn which contains x will be contained in S(x, S). Given
any open covering -9 of E, carry out the above process for each set
D of .9 which intersects E, and each x in D n E, and let "' c ' be the
countable collection of open spheres obtained. For each SEW', choose
one set D E -9 with D S, and let .9' be the countable class of sets
so obtained. Then, since W' covers E, the class _9' is a countable
subcovering. This means that, if we assume property (W), open
coverings can be replaced by a countable subcovering.
00
xnEE- U Gt
i=1
and the sequence {xn} must form an infinite set, so that there is a
limit point x0 E E. But x0 E Gk for some k, and Gk is open and therefore
is a neighbourhood of x0; this means we can find an n > k such that
n
xn E Gk U Gi,
i=1
which is a contradiction.
Compactification
Many operations can be carried out more easily in compact spaces
than in non-compact spaces. Given a non-compact space X a useful
trick is to enlarge it to a topological space X* X which is compact
and such that the system G of open sets in X is obtained by taking the
intersection X n 0 with X of sets G which are open in X*. This device
is known as the compatification of X. For example, R is not compact,
but if we adjoin two points + oo, - oo to form the space R* of extended
34 POINT SET TOPOLOGY [2.2
real numbers we can show that R* is compact if we call a set E c R*
open if E n R is open and
if + oo E E, there is a neighbourhood {x: a < x < + oo} c E;
if - oo E E, there is a neighbourhood {x: - oo 5 x < b} c E,
where a, b E R. Note that the extended real number system R* is
simply ordered if we put - oo < x < + oo for all x e R.
In general, a non-compact topological space X may be compactified
in many different ways. The simplest method is to adjoin a single
point oo (which can be thought of as a point at infinity) to give the
space X* = X v {oo} and say that a subset G of X* is open if either
(i) G c X and G is open in X; or
(ii) oo E G and (X* - G) is a closed compact subset of X. It is not
difficult to verify that this collection of `open' sets defines a topology
in X* in which X* is compact. This process is called the one-point
compactification of X. It is familiar in the theory of the complex
plane, where it is usual to add a single point at infinity (with neigh-
bourhoods of the form In > R) to make the resulting `closed plane'
compact. Note that, if G* is the class of open sets in X*, G is the class
of sets of the form X n E, E e G*.
There are other, more sophisticated, methods of compactifying a
topological space X, but we will not require these in the present
book.
Exercises 2.2
1. If (X, p) is a compact metric space, show that
(i) X is complete;
(ii) for each e > 0, X can be covered by a finite class of open spheres of
radius E.
2. If (X, p) is a complete metric space which, for each e > 0, can be
covered by a finite number of spheres of radius e, show that X is compact.
3. The open interval X = (0, 1) c R with the usual metric
p(x1,x2) = Ixl-x21
is a metric space. In (X, p) the set X is closed and bounded. Show that X
is not compact (and therefore not complete by example 2).
4. Construct a covering of the closed interval [0, 1] by a family of closed
intervals such that there is no finite subcovering.
5. If A, B are compact subsets of a metric space X, show that there are
points xo E A, yo c B for which
p(xo, yo) = d(A, B).
2.2] COMPLETENESS AND COMPACTNESS 35
Hint. Take sequences {xi} in A, {yi} in B with p(xi, yi) < d(A, B) + 1/i,
and apply property (W) to find convergent subsequences.
6. Give the details of the proof that the process of adjoining a point co
used to give the one-point compactification does yield a compact set X*.
If this process is applied to a space X which is already compact, show that
the one point set {co} is then both open and closed in X*.
2.3 Functions
In Chapter 1 we defined the notion of a function f: X --> Y. When
X and Y are topological spaces it is natural to enquire how the func-
tion f is related to these topologies. In particular do points which are
`close' in X map into points which are close in Y? We make this
precise first for metric spaces.
Continuous function
If (X, px), (Y, pp) are metric spaces, a function f: X -* Y is said to
be continuous at x = a if, given e > 0 there is a S > 0 such that
px(x,a) < 6- py(f(x),f(a)) < e.
If E c X, we say f is continuous on E if it is continuous at each point
of E. In particular f: X -. Y is said to be continuous (or continuous
on X) if it is continuous at each point of X.
Lemma. If (X, px), (Y, py) are metric spaces a function f : X --> Y
is continuous if and only if f-'(G) is an open set in X for each open
set G in Y.
Proof. Suppose first that f is continuous and G is an open set in Y.
If f-'(G) is void, then it is open. Otherwise, let a e f-'(G), f (a) e G
so that there is an e > 0 for which the sphere S(f(a), e) c G. But then
we can find a 8 > 0 such that
px(x, a) < 8 = f (X) E S(f (a), e) e G
so that the sphere S(a, S) c f-'(G).
Conversely consider f at a point a of X. For each e > 0, S(f (a), e) = H
is an open set in Y, so that if f-'(H) is open, we can find a 8 > 0 for
which S(a, 8) c f-'(H), that is such that
px(x,a) < 8=>- pr(f(x),f(a)) < e.]
If (X, 9), JV') are topological spaces, the function f: X -> Y
is said to be continuous if f-'(H) e 9 for every H in .'. The lemma just
.
Exercises 2.3
1. Consider the function f: (0, oo) -+ R given by f(x) = min (1,1/x), for
x>0.
Show that it is continuous. Find the image f (E) of
(i) the set E = (1,1);
(ii) the set Z of positive integers;
(i) shows that E can be open, f (E) not open; (ii) shows that E can be closed,
f (E) not closed.
38 POINT SET TOPOLOGY [2.3
2. The function
1
f : (0,1) R given by f (x) _
x(1-x)
is continuous on (0,1), but not bounded and not uniformly continuous,
so theorems 2.3, 2.4 fail if the set is not closed. To see they also fail if the set
is closed but not compact, examine
g: R R given by g(x) = exp (x).
3. In the argument of the proof to theorem 2.4 why could we not have
put
g = inf {Sx: x A}
before first restricting to a finite subset?
4. Suppose A is compact and f f,} is a monotone sequence of continuous
functions f,: A -* R converging to a continuous f : A -+ R. Show that the
convergence must be uniform, and give an example to show that the
condition that A be compact is essential.
5. Prove Lebesgue's covering lemma, which states that if le is an open
cover of a compact set A in a metric space (X, p), then there is a 8 > 0,
such that the sphere S(x, S) is contained in a set of ' for each x e X.
Exercises 2.4
1. If p11 P2 are two metrics in X such that cp1(x, y) < p2(x, y) 5 kp1(x, y)
for all x, y E X where c > 0, k > 0; show that pl and P2 generate the same
topology in X.
2. Show that, if X x Y has the product topology, then the projection
function p: X x Y -* X defined by p(x, y) = x is continuous.
In a space X, if T, T2 are two collections of `open' sets defining topologies
and 61 g2 we say that the topology given by T2 is coarser than that given
by T,. Show that the product topology in X x Y is the coarsest topology
for which projections are continuous. (For an arbitrary Cartesian product
ji Xa of topological spaces (Xa, 9a) the projection p,B can be defined as
ad
p g(xa, a e I) = xe a X, for any 6 E I. One method of defining the product
2.4] CARTESIAN PRODUCTS 41
topology in the Cartesian product space is to say that it is the coarsest
topology for which each of the projections is continuous.)
3. Suppose X x Y has the product topology and A e X, B C Y. Show
that
AxB=AxB,
and prove that the product of closed sets is closed.
Now diam (flA) diam (An) for each n so that diam (fl A) = 0
1
By the last lemma, there is a unique point xo a fl S(xn, rn) and this
n=1
point x0 cannot be in A. for any integer n. This means that
OD
n=1
Corollary. Rn is of the second category. For a < b, the interval
[a, b] c R is of the second category.
Dense subset
If A, E are subsets of a topological space X, we say that A is dense
in E if A m E. This means that any open set G which contains a
point of E also contains a point of A. In particular A is dense in
X if A = X, that is, if every non-void open set contains a point of A.
Separable space
A topological space X is said to be separable if there is a countable
set E c X such that E is dense in X. This implies the existence of a
sequence {xn} in X such that every non-void open subset contains a
point of the sequence. It is immediate that R and Rn are separable
as the set of points with rational coordinates is dense and countable.
Further, every compact metric space X is separable, since X can be
2.5J FURTHER TYPES OF SUBSET 43
covered by a finite class len of open spheres of radius 1 /n for n = 1, 2, ...
and the (countable) set consisting of the centres of all these spheres is
clearly dense in X.
Borel sets and Borelian sets
In any topological space X, we will call the i generated by
the open sets the class of Borel sets, and the .'1'' generated by
the compact sets the class of Borelian sets. (One must remark that
some authors use the term Borel sets for .7E.)
In a metric space the compact sets are closed, and therefore in
OD
Exercises 2.5
1. Show that, in R", any countable set is of the first category. Give a
category argument for the existence of irrational numbers.
2. Show that the class .N' of nowhere dense subsets of X is a ring, and the
class ' containing all sets of the first category is the generated o--ring.
a{xi} = {axi}.
Since for x, y real we have
Ix+yl Ixl lyl
1+lx+yl 1+1x1 +1+Iyl
Exercises 2.6
1. Show that s is bounded but not compact. If x = {xi} c s, and
E _ {y: lyil , Ixil},
show that E is compact in s, but show that s is not locally compact.
2. Show that each of the spaces M, c, m, s is complete.
3. Show that each of the spaces C, c, s is separable, but that m is not
separable.
Hint for C. Consider the set of functions which take rational values at
each of a finite set of rationals in [0, 11 and are defined by linear interpola-
tion between these points.
4. C*(X) denotes the space of functions f: % -+ R which are continuous
and bounded. Show that C*(R) is not separable by considering continuous
functions which take the values + 1, -1 on disjoints subsets Z1, Z2 of the
set Z of positive integers and are defined elsewhere by linear interpolation.
(The distance between any two such functions is 2, and there are c of them.)
However, let .9 be the subset of C*(R) consisting of those functions f
for which lim ft x), lim &) both exist. Prove that -9 is separable.
ro
5. Let 12 be the subset of s such that xZ converges. In the linear struc-
i=Z
ture of s show that 12 is a linear space and that
IIxIl=fix?
defines a norm. In the topology of this norm show that 12 is separable.
The subset {x: 1xil < 1/i} of 12 is known as the Hilbert cube: prove it is
compact.
Hint. Starting with an infinite sequence in the Hilbert cube pick a sub-
sequence in which the first coordinate converges, then successive sub-
sequences in which the 2nd, 3rd, ..., nth coordinate converges. Show that
the sequence to which these coordinates converge is in the Hilbert cube
and is a limit point of the original set.
2.71 CANTOR SET 49
Gn=r=1
UE,,r G= UGn;
n=1
it is clear that G is an open subset of [0, 1]. Its complement
C=[0,1]-G
is called the Cantor set. From its definition C is closed.
Lemma. The Cantor set C is nowhere dense and perfect.
Proof. If we express points xE [0, 1] in the form
ai
x= (a1 = 0, 1, 2), (2.7.1)
i=1 3'
then the set G. above is the set of x for which an = 1. Hence, the set
C consists of precisely those real numbers which have a representation
in the form (2.7.1) with each ai = 0 or 2. Given a point x1 E C, altering
the nth term an (replacing 0, 2 by 2, 0 respectively) gives a new
point x2 in C such that x2l =
I xl - 2.3-n.
This shows that every point of C is a limit point of C, so that C is
perfect.
If H is any open set with H n [0, 1] not void then H n [0, 1] contains
an open interval I of length S > 0. If S > 31-n, then I must contain
an interval En,r so that H contains an open set disjoint from C.
This proves that C is nowhere dense. I
From the above lemma and example 2.5 (7) we can deduce that C
is not countable. However, one can prove that C must have the same
cardinal as the continuum [0, 1] by considering the following mapping:
00
if 1 ], put X= (bi = 0 or 1),
zZ 2i
where the sequence {bi} does not satisfy bi = 1 for i > N. Put
00 ai ai = 0 if bi = 0,
P X) = i? 13' (ai = 2 if bi = 1.
50 FOINT SET TOPOLOGY [2.7
Then f: [0, 1] -* C is (1, 1) and maps [0, 1] onto a (proper) subset of C.
Since C c [0, 1] the cardinal of C is c.
We can think of f as a function on [0, 1] to [0, 1]. It is easy to see
that f is monotonic, that is,
{
x1 < x2 e f (x1) < J (x2)
so that, for each y E [0, 1],
f-1[0, y] = [0, z]. (2.7.2)
If z is defined by (2.7.2), then we say that
z = g(y)-
This defines g: [0, 1] -> [0, 1] as a monotonic function which is clearly
constant on each of the sets En ,. In fact
3r-2 3r-1 2r- I
3n -< y 3n g(y) = 2n
The function g is continuous and monotonic increasing, for
0 S y1 - Y2 < 3-n-1 0 5 g(y1) - g(y2) < 2-n.
Since the function g is constant in each En,, it follows that it is dif-
ferentiable with zero derivative at each point of G. One can easily
see that g increases at each point of C-and in fact the `upper deriva-
tive' at points of C is + oo.
Note that there is nothing magical about the integer 3 used in the
construction of C. Similar constructions using expansions to a
different base will give sets with similar properties.
Exercises 2.7
1. If x =
c(c2 = 0,1, ..., 9) is a decimal expansion of real numbers
i=110'
in [0, 1] and T is the set of such x for which cE + 7, show that T is perfect
and nowhere dense.
2. Construct a set which is dense in [0,1] and yet the union of a countable
class of nowhere dense perfect sets.
3. Show that the function g : [0,1]-x[0,1] defined above satisfies a
Lipschitz condition of order a = log 2/log 3, but not of any order 8 > a.
(A function h: I-* R is said to satisfy a Lipschitz condition of order a at
xo E I if jh(x) - h(xo) I < K I x - xo) j a for x e I and some suitable K E R.)
51
SET FUNCTIONS
3.1 Types of set function
We consider only functions u: ' -> R*, where ' is a non-empty
class of sets. Thus p is a rule which determines, for each Ee', a
unique element u(E) which is either a real number or ± oo. We always
assume that' contains the empty set 0. R* denotes the compacti-
fication of the real number field R by the addition of two points
+ oo, - oo, while R+ will denote the set of non-negative real numbers
together with +oo. It is not possible to arrange for R* to be an alge-
braic field extending R, though we will preserve as many of the alge-
braic properties of R as possible by adopting the convention that, for
any a e R,
-oo<a< +oo, -oo±a=-oo, +oo±a= +oo;
if a > 0, a(+ oo) = + oo, a(- oo) oo;
and suppose P1, P2, ... is a sequence of real numbers such that E pi
i=1
either converges absolutely or is properly divergent to + oo or - oo
(the case E pi convergent, E lpi I divergent is not allowed: why?). Put
,u(E) = E pi,
where the sum extends over all integers i = 1, 2, ... for which xi E E.
Any set function which can be defined as in example 6 is called discrete.
Note that example (4) can be thought of as a special case of example (6).
Although it is not sufficient to restrict our attention to set functions
,u: W -. R which are finite valued, the condition of additivity which is
usually assumed prevents ,u from taking both the values + oo, - oo
at least when ' is a ring. This is one of the results in the next theorem.
Theorem 3.1. Suppose T: V R* is an additive set function defined
on a ring' and E, FEW. Then
(i) if E F and r(F) is finite
T(E - F) = T(E) -T(F);
(ii) if E F and T(F) is infinite
T(E) = T(F);
(iii) if T(E) _ + oo, then r(F) + - oo.
Proof. (i) Since '' is a ring, E - F E'o and additivity implies, since
Fn (E-F) = 0, T(E) = T(E-F)+T(F). (3.1.2)
Subtracting the finite real number T(F) gives the result.
54 SET FUNCTIONS [3.1
(ii) If T(F) = +oo, then (3.1.2) can only have a meaning if
T(E - F) + - oo, and this implies T(E) = + oo. The case r(F) oo
is similar.
(iii) Since E n F, E - F, F - E are disjoint sets of le
T(E) = r(EnF)+T(E-F) = +oo,
T(F) =T(EnF)+T(F-E) = -oo
could only have meaning if r(E n F) is finite. But this would imply
T(E - F) = + 00, T(F - E) = - oo, and then, since EL F E W,
T(E0F) = r(E - F) + T(F - E) =+oo+(-oo)
which is impossible. J
Our definition of additivity means that for ,u: ' ->. R* to be additive
any set E0E' which can be split into a finite number of disjoint sub-
sets in ' must be such that ,u(Eo) is the same as the sum of the values
of ,u on the `pieces'. We often want this to be true for a dissection of
E0 into a countably infinite collection of subsets in W.
E _ UEiE'f,
i=1
00
+ oo = ,u(E) + 1 = =1
\n -n+1 ) _ n=1
E ME.).
1
3.1] TYPES OF SET FUNCTION 55
Notice further that even when '' is a ring it does not follow that
Ei c9 (i = 1, 2, ...) =>E = UEE a 1;
so that in testing (3.1.3) we can only use those sets E E' which can be
split into a countable sequence of disjoint subsets in W. In particular
if' is a finite class of sets then additivity for ,u: (f ->. R* implies
additivity. We also interpret (3.1.3) to mean that the right-hand side
is uniquely defined and independent of the order of the sets Ei;
thus if p is a decomposition E = U Ei,
i=1
we cannot have,u(Ei) = +oo, p(E5) _ -oo, nor can the series in (3.1.3)
converge conditionally.
It is easy to check that each of the set functions in examples (1), (2),
(4), (6) on p. 52 is and the set function of example (3) is
also o--additive though the proof of this fact is non-trivial. This proof
will be given in detail in § 3.4, as it is an essential step in the definition
of Lebesgue measure in R.
Measure
Any non-negative set function p:' R+ which is o-additive is
called a measure on ', (R+ = {x E R*: x > 0}).
We should remark that there is not general agreement in the
literature as to which set functions ought to be called measures.
According to our definition the set functions in examples (1), (2)
and (3) are measures, those in (4) and (6) are not because /I can
take negative values while the set function in (5) is not because it is
not o--additive.
The natural domain of definition of a measure, or indeed of any
or-additive set function, is a since then
OD
Continuity
Suppose q is a ring and p:. R* is additive with ,u(E) > -oo
for all E E M. Then for any E E R we say that :
(i) u is continuous from below at E if
lim #(E.) = p(E) (3.1.4)
n-aw
for every monotone increasing sequence {En} of sets in gP which con-
verges to E;
(ii) It is continuous from above at E if (3.1.4) is satisfied for any
monotone decreasing sequence {En} in 9 with limit E which is such
that p(En) < oo for some n;
(iii) It is continuous at B if it is continuous at E from below and from
above (when E = 0 the first requirement is trivially satisfied).
Theorem 3.2. Suppose 9 is a ring and ,u:.9 --3- R* is additive with
,u(E) > - oo for all E E R.
(i) If p is o--additive, then p is continuous at B for all E E 9;
(ii) if It is continuous from below at every set E E 9?, then p is v-
additive;
(iii) if p is finite and continuous from above at o, then p is o--additive.
Proof. (i) If;a(En) = +oo for n = N and {En} is monotone increasing
then ,u(E) = +oo and #(E.) = +oo for n >, N by theorem 3.1 (ii),
where E = lim En. Thus in this case ,u(En) ->. p(E) as n -a oo. On the
other hand, if p(En) < oo for all n and {En} increases to E, then
00
E = El U U (En+1- En)
%=1
En = (0, n) (n = 1, 2, ... )
Outer measure
If IF is the class of all subsets of 0, then ,u:' R+ is called an outer
measure on t if
(i) u(O) = 0;
(ii) u is monotone in the sense that E c F . p(E) < ,u(F);
(iii) u is countably subadditive in the sense that for any sequence
{Ei} of sets, OD 00
Exercises 3.1
1. If S2 = [0,1) and '' consists of the 6 sets
0, Q, [0, '), [0, 1), [0, 1),
li(o)=0, u[0,1)=2, p[0,1)=2,
,a[0,1) = 4, fz[4,1) = 2,
lt(Q) = 4,
show that ,a is additive on W. Canp be extended to an additive set function
on the ring generated by le?
3 TIT
60 SET FUNCTIONS [3.1
2. Show that if 9 is any finite ring of subsets of 0 and p is additive on
R then ,u is a--additive on R.
3. A set function fi: ' -+ R* is said to be monotone if p(0) = 0 and
E e F, E, F E' . p(E) < ,u(F). Show that monotone set functions are non-
negative, and if % is a ring, show that an additive non-negative set function
is monotone. Of the set functions in examples (1)-(7), which are monotone?
exists. Show that T is finitely additive, but not v-additive on e, but that
' is not even a semi-ring.
6. Ifp is finitely additive on a ring 9; E, F, G E 9 show
p(E) +p(F) = #(E v F) +,u(E n F),
p(E) +,u(F) +, a (G) +,u(E n Fn 0)
= p(EvFuC;)+,u(EnF)+p(FnG)+lu(GnE).
State and prove a relationship of this kind for n subsets of R.
7. Suppose. is a v-ring of subsets of n, It is a measure on Y. Show that
the class of sets EE.9' with p(E) finite forms a ring, and the class with p(E)
v-finite forms a a--ring.
8. If E is a set in So of a-finite #-measure (where p is a measure on So)
and 9 c 01 where -9 is a class of disjoint subsets of E show that the subclass
of those D e -9 for which p(D) > 0 is countable.
9. State and prove a version of theorem 3.2 (i) for set functions, defined
on a semi-ring W.
10. To show that the finiteness condition in the definition of 'continu-
ous from above' in theorem 3.2 (i) cannot be relaxed, consider any infinite
space.Q and put
T(E) = number of points in E, if E is finite;
T(E) = + oo, if E is infinite.
3.11 TYPES OF SET FUNCTION 61
Then T is a measure on the class of all subsets of 0, but for any sequence
of infinite sets which decreases to 0, we do not have urn T(E) = 0.
11. Suppose 9 is the semi-ring of half-open intervals (a, b], Q is the set
of rationale in (0, 1] and 9Q is the semi-ring of sets of the form (a, b] n Q.
Put
,u{(a,b]nQ}=b-a if 0<a<b<1.
Show that ,u is additive on 9Q and is continuous above and below at every
set in eQ, but is not a--additive. This shows that theorem 3.2 (ii), (iii) is not
true for semi-rings.
12. Show that if It is an outer measure on S2, E0 any fixed subset then
,uo(E) = p(E n Eo) defines another outer measure on Q.
,t Ei = Z lu(Ei)
i =1 i=1
co
Since U Ei is independent of the order of the sets in the sequence,
i=1
it follows that the series on the right-hand side must be either
3-2
62 SET FUNCTIONS [3.2
absolutely convergent or properly divergent. In the case of example (6)
the set function
,u(E) = E pi
x;EE
r=n r=n
m
If we put Dn = fl Cr we have Dn EF and, by theorem 3.2 (i),
r=n
T(D,) < A+21-n.
But now {Dn} is a monotone increasing sequence of sets in F so that
N = lim Dn = lim inf Cn E .F,
n->oo n--> ao
and T(N) = A,
by theorem 3.2 (i).
64 SET FUNCTIONS [3.2
Finally, put P = fl - N. If E e.F, E e P we must have T(E) > 0
for, if T(E) < 0, then T(E v N) = T(E) +T(N) < A. Also if Be F,
E c N we must have T(E) < 0 for, if T(E) > 0, then
T(N - E) = T(N) - T(E) < A.
If we now put
T+(E) = T(E n P), T_(E) = T(E n N),
it is clear that all the conditions of the theorem are satisfied.
Remark. It is usual to call the decomposition T = T+-T_, of T
into the difference of two measures, the Jordan decomposition while
the decomposition of SZ into positive and negative sets P and N is
called the Hahn decomposition. It is not difficult to show that the
Jordan decomposition is unique while the sets P, N are not uniquely
determined by T unless T(E) + 0 for all EE,F such that 14(E) + 0 and
p(F) = 0 or µ(E) for every F c E with F E.F. It is further clear that
T_(E) = -A(E),
(3 . 2. 2)
T+(E ) = sup T(B )
BCE, BE.i f
under the conditions of theorem 3.3, where A(E) is given by (3.2.1).
If one is given a a or-ring Y
which is not a then it is not, in general, possible to obtain the
Hahn decomposition, but the Jordan decomposition is still possible,
using (3.2.1), (3.2.2) as the definition of T_, T+.
Exercises 3.2
1. If :Y -> R* is a'-additive on a Y, show that, for any Ee.9',
there are sets A c E, B e E with A, B E So such that
c(A) = inf 0(0), O(B) = sup (Cry).
CCE,CE.f CCE,CE`.'
2. Showthat, given a (finitely) additive ,u: 3P --> R* defined on a o--ring M
and taking finite values, there is a decomposition
.u(E) = µ+(E) -u-(E)
of p into the difference of two non-negative additive set functions on M.
3. The set E0 E' is said to be an atom of a set function 0:' R* if
g5(Eo) + 0 and for every E e B0, E ET; ¢(E) = 0 or ¢(E0). Write down the
atoms of the set functions of examples (4) and (6) on page 53.
4. A set function 0: ' -> R* is said to be non-atomic if it has no atoms.
Suppose q5:.F-* R* is on the non-atomic, and finite
valued. Show that for any A e.., 0 takes every real value between - 0-(A)
and 6+(A) for some subset E e A.
3.31 ADDITIVE SET FUNCTIONS 65
where F e' and are disjoint. Put Hkj = Ek n F1. Then -since W is a
semi-ring the sets Hkj a le, are disjoint and
m
Ek=UHkj (k n);
j=1
n
Fj = U Hkj (9 = 1, 2, ..., m);
k=1
E1.(Ek)=E(E/(Hkj))=E(Eu(Hkj))=Elt(Fj)
k=1 k=1 j=1 j=1 k=1 j=1
n m
Hence v(A1 v A2) = E,u(Ek) +iE,u(F )
= v(A1) + v(A2),
since v is uniquely determined by (3.3.1). Thus v is finitely additive
on R (since R is a ring). It is obvious that v is non-negative.
Now let r be any extension of p from ' to °.rP which is additive. If
n
A e 9P and A = IJ Ek is a decomposition into disjoint sets of '',
k=1
n
T(A) = E r(Ek), since r is additive;
k-1
n
= Ep(Ek), since r is an extension;
k-1
= v(A) by (3.3.1).
Thus v is the unique additive extension of p from $° to Q.
If we start with a measure p: % --> R+ on a semi-ring ', then It
is clearly a non-negative finitely additive set function, and so possesses
a unique additive extension to the generated ring R. What can we
say about this extension?
Theorem 3.5. If ,u:'-.R+ is a measure defined on a semi-ring W,
then the (unique) additive extension of µ to the generated ring 3P(() is also
a measure.
Proof. In the last theorem we discovered the form of the unique
additive extension v of ,u from ' to R. It is sufficient to show that
v is on R. Suppose E e R, Ek (k = 1, 2,...) e 9 and are
disjoint, and E = Uco Ek,
k=1
n
Put E = U A Ar disjoint sets of %;
r-1
nk
Ek = U Bkd, Bk4 disjoint sets off.
4x1
3.3] ADDITIVE SET FUNCTIONS 67
Put
Crki=ArnB
then {Crki} forms a disjoint collection of sets in ', and
co nk n
Ar = U U Crki, Bki = U Crki
k-1 i=1 r=1
are disjoint decompositions into sets of W. Since ,a is additive on ',
n
u(Bki) = Ela(Cki)i
r=1
and since it is a--additive on W
0o nk
p(Ar) = z z/J'(Crki)-
k=1i=1
Since the order of summation of double series of non-negative terms
makes no difference, we have
(00 nk
v(E) _ E f,(Ar) _ I E fp(Crki)
r=1 r=1 k=1 i=1
k
= E ( E E (Cki)
k=1 i=1 r=1
Co nk °o
= Fi Zi lu(Bki) = Fi v(Ek) J
k=1 i=1 k=1
The above theorem gives one method of obtaining a measure on a
ring-it is sufficient to define a measure on any semi-ring which
generates the given ring. The extension to the generated ring is easily
carried out, is unique, and gives a measure. There are circumstances
in which one can define a set function p directly on a ring so that it is
easy to see that p is non-negative and additive. Under these circum-
stances one can often use theorem 3.2 as a criterion for determining
whether or not It is a measure. Another useful criterion is given by
the following theorem.
Theorem 3.6. Suppose ,u: 9 --> R+ is non-negative and additive on a
ring R. Then
(i) if E E R, and {Ei} is a sequence of disjoint sets of R such that
Co
°°
E=) UEE
i=1 p(E) i Tlu(Ei)i
i=1
(ii) p is a measure if and only if for any sequence {Ei} of sets in R
00
such that U Ei E E PA, 00
IA(R)>,,A
[3.3
n=1 n=1
p(E) % Eli(Ei);
i=1
p(E) 5 E,u(Ei)
i=1
Exercises 3.3
1. If n = {1, 2,3,4, 5,}, show that ' consisting of o, 0, {1}, {2, 3},
{1, 2, 3,}, (4,51 is a semi-ring and that 0, 3,1,1,2, 1 defines a set of values
for an additive set function ,u on W. What is the ring ? generated by 6?
Find the additive extension of p to M, and show that it is a measure.
2. Suppose . is any ring of subsets, 0: G -+ R+ is non-negative, finitely
additive on 9P, and p:. -->. R+ is a measure on 6. such that, for any sequence
of sets in R
0= 0 as n oo;
show that 0 is completely additive.
3.3] ADDITIVE SET FUNCTIONS 69
3. If ,u: R -+ R+ is finitely additive on a ring . and E, F E R are such
that #(E L F) = 0, we say that E - F. Show that - is an equivalence re-
lation in R and that
E- F-#(E) =,u(F) _,u(EvF) =,u(EnF).
Is the class of all sets E E R for which E -' 0 a ring?
4. In the notation of question 3, put p(E, F) = #(E A F) and show that
p(E, F) > 0, p(E, F) = p(F, E), p(E, F) 5 p(E, 0)+p(O, F). If E1,., E2,
F1- F2 are all sets in .', show that p(E1, F1) = p(E2, F2). Does p define a
metric in A?
set function in offk. Then one can either make use of the continuity
theorem 3.2 to show that ,u: fk -> R+ is a measure on 01, or one can
prove directly that ,u is a measure on 9k and appeal to theorem 3.5
to deduce that its extension is also a measure. We illustrate by apply-
ing the first method to the case k = 1, and the second method to the
case k = 2.
k=1
For each (a, b] E 9 we put µ(a, b] = b - a. It follows that, is additive
on 9 for if (a, b] _ U (ai, bi] and the (ai, bi] are disjoint we may assume
i=1
that these intervals are ordered so that bi < ai+1(i = 1, 2, ..., n - 1). It
follows that we must have a1= a, bn = b and bi = ai+1(i =1, 2, ... , n - 1)
so that, if an+1 = bn,
n n n
E u(ai, bi] = E (bi - a1) = F (ai+1- a1)
1=1 i=1 i=1
Thus by induction we can establish (3.4.1) for all n. Since II(E.) >, 46
for all n, we have #(F.) > 16, for all n
so that all the sets F. are non-void. Hence {Fn} is a decreasing sequence
CD
nFnc n=nEn= o,
n=1
so we obtain a contradiction.
k=2
Suppose C = {(x, y) : a < x < b, c < y < d} is a set of g2, and
p(C) = (b - a) (d - c). In order to prove that u is additive on g2,
C n
suppose that U Ci is a decomposition of C into disjoint rectangles
i=1
in each of which one of the sides (say (c, d]) remains the same. Then
the other sides (ai, biJ must be disjoint and satisfy
n
(a, b] = U (ai, bi]
i=1
so that by the corresponding result in 91, ,u is additive in this case.
More generally if
n
C = U Ci, Ci = {(x, y): ai < x < bi, ci < y < di}
i=1
is a decomposition of C into a finite number of disjoint rectangles,
use the infinite lines x = ai, x = bi, (i = 1, 2,..., n) to decompose
each Ci into a finite number of pieces Cik each with the same bounds
for the y-coordinate. Hence
n
E p(Ci) = E Eclu\cik),
/
i=1 i k
and we can sum the right-hand side by first summing over the rect-
angles whose x-coordinate is bounded by a pair of contiguous ai,
by and then summing over these intervals in x. Thus by repeated
application of additivity in 91 we get
n
A(C) = E lu(Ci),
i=1
Since Fo is compact and the sets FOi are open it follows that, for some
integer n, we have
C n n
PO U F° so that Fo c U Fi.
i=1 i=1
,u*(E) = inf
i=1
the infimum being taken over all sequences of sets {Fi} in 9 such that
OD
is immediate. If a*(E1) < oo for all i; for any e > 0, choose sets
Fik (k = 1, 2,...) in .ri'P such that
00 00
6
Ei c U Fik and E /(Fik) < ,u*(Ei) +
24
(i = 1, 2, ... ).
k=1 k=1
The countable collection {Fik} will now cover E, and
00 00
Exercises 4.1
1. Suppose p* is an outer measure on S2 = lim Ek where {Ek} is a mono-
tone increasing sequence of sets. Show that if a set E is such that E n Ek
is measurable (p*) for all sufficiently large k, then E is measurable (p*).
2. Show that if p* is a regular outer measure on S2 and p*(Q) < oo,
then a necessary and sufficient condition for E to be measurable (,a*)
is that
p*(S2) = p*(E)+p*(Q-E).
3. In each of the following cases, show that p* is an outer measure, and
determine the class of measurable sets
(i) p*(o) . 0, p*(E) = 1 for all E + 0.
(ii) p*(Q) = 0, p*(E) = 1 for E + 0 or S2, p*(SZ) = 2.
(iii) S2 is not countable; p*(E) = 0 if E is countable, p*(E) = 1 if E is not
countable.
4. Show that any outer measure which is (finitely) additive is o--additive.
5. Suppose p* is an outer measure on 0 and E, F are two subsets, at
least one of which is measurable (p*). Show that
p*(E) +p*(F) = p*(Eu F) +,a*(E n F).
6. Suppose is a sequence of sets in a o--ring .97, and # is a measure
on 9. Show that
(i)
then UEi=UXiuUN7=XvN,
i=1 i=1 i=1
00
Exercises 4.2
1. Suppose It is a measure on a a-ring .2 and ;u on .2 is its completion.
Show that if A, Bet with A c E c B, ,u(B - A) = 0 then E E , and
Z(E) =,u(A) =#(B).
2. Given a a-finite measure,u on a ring. the extension given by theorem
4.2 yields a complete measure on the class .4' of #*-measurable sets which
is the completion of .5o the generated a-ring. The following example shows
that this is not true if the hypothesis of a-finiteness is omitted: Let S2 be
non-countable, .9' the ring (also a a-ring) of all sets which are countable
or have countable complements, ju(E) = number of points in E for E EY.
Then .5o is complete with respect to a, but applying theorem 4.2 yields a
complete measure on the class of all subsets (as every subset is measurable).
Regular measure
Suppose .So is a of subsets of a topological space S2 which
includes the open and the closed subsets of S2, and p:$-->- R+ is a
measure. Then the measure ,u is said to be regular if, for each e > 0,
(i) given E E.Y, there is an open G E with ,u(G - E) < e;
(ii) given E E ., there is a closed F E with p(E - F) < e.
Since the class . of Borel sets in S is the generated by the
open sets, the condition that 3 includes the open sets implies .9' .4.
If p is regular on ., then . .9, where . denotes the completion of
. with respect to ,u; for if Sn is a sequence of positive numbers de-
creasing to zero one can find for any E in . an open set G. and a
closed set F. such that
µ(G. - Fn) < Sn and G. E = Fn,
and G=nGn,
n=1
F=IJFF
n=1
will then be Borel sets with G E F and µ(G - F) = 0.
Metric outer measure
An outer measure µ* defined on a metric space S2 and such that
p* is additive on separated sets, i.e.
d(E, F) > 0 . p*(E v F) = ,u*(E) +,u*(F),
is said to be a metric outer measure. It can be proved that, for any
metric outer measure, the class ,t of measurable sets contains the
open sets (and therefore contains -4), and that, if u* is also o--finite,
the restriction of µ* to . ' is regular. Since Lebesgue measure is
generated by a metric outer measure, this general theory would allow
us to deduce that Lebesgue measure is regular. However, we prefer
instead to prove the result only for the special.case of Lebesgue
measure.
Theorem 4.5. Lebesgue k-dimensional measure, defined on the class
2k of Lebesgue measurable sets in Rk, is a regular measure.
Proof. We give the details of the proof for k = 1; only obvious
alterations are needed for general k. Suppose E e 2 = 21; then
B n [n, n + 1) = E. e 2 for every integer n, and p(En) < 1 < oo. By
the construction of theorem 4.2, there is a countable covering {Cz}
of E. by f-open intervals of 9 such that
1e °°
lu(En) 4 F(Cni)
iE
4.3] APPROXIMATION THEOREMS 87
Enlarge each of these intervals Cni to an open interval Gni such that
1 e
u(Gni - Cni) < 4 21n1+i
00
Then Q. = U Gni is an open set which contains En and satisfies
i=1
u(Qn-En) <
221n1.
00
If we now put Q = U Qn, then Q is open, Q C E, and u(Q - E) < e.
n=-ao
This proves condition (i) for regularity.
For any E E 2', 1) - E E 2, and we can apply the above argument
to obtain an open R 12 - E such that ,u (R - (S1- E)) < e. Then
F = SZ - R is closed, F C E and #(E - F) = u(R n E) < e, so that the
second condition for regularity is also satisfied. I
Corollary. Given any set E e 2k, there is a Va-set Q and an .°F, set R
such that
Q=)E=)R and µ(Q-R)=0.
Proof. Note that 9ra and .F,,. sets were defined in § 2.5. For each in-
teger n, take an open set G. E and a closed set F. C E such that
In,
#(Gn-E) < n #(E-Fn) <' n1.
OD 00
The sets Q = n Gn and R = (,J F.
n=1 n=1
Exercises 4.3
1. Suppose 2 is the o--ring generated by a ring 9 and ,u, v are two a--
finite measures on R. Show that if E e 2 is such that both #(E), v(E) are
finite then, for any e > 0, there is a set E. e R for which
p(E A Eo) < e, v(E A E0) < e.
2. Suppose 0 is a metric space and p* is an outer measure on S2 such
that every Borel set is #*-measurable. Show that µ* is a metric outer
measure, i.e. that for E1, E2 C S1,
d(E1,Es) > 0 ,a*(E1vB2) = p*(E1)+,u*(E2)
88 PROPERTIES OF MEASURE [4.3
C=(1Fn,
n=0
defined in § 2.7 where F. = [0, 1] and F. is obtained from Fn_1 by replac-
ing each closed interval of Fn_1 by two closed intervals obtained by
removing an open interval of one third its length from the centre.
We proved that C was perfect and therefore non-countable. But
IFnl = I I Fn-lI = (J)'IF0l = (J)n,
so that ICI = lim I Fns = 0.
n--)- co
for all p > 0, x E Rk. This means that, for a given x and I, the effect
on the measure is the same for all intervals of the form (p1) (x).
Now any open set G can be expressed as a countable union of
disjoint sets of the form (pI) (x). In particular the unit open sphere
S centre the origin, can be expressed this way
OD
S = U (piI) (xi),
i=1
ro
and ISI = EI(piI)(xi)I.
i=1
But xS = 8, so that
OD
Ix(piI)xiI =
II co
EcUQicJ, FcUR1cK,
i=1 f=1
00 OD
A=UAncE,
n=1
B=UBmCF,
m=1
IE-AI = 0, IF-BI = 0.
4.41 LEBESGUE MEASURE: PROPERTIES 93
Since A x B is an F, set in Rk+a it is measurable and
1*(A x B) = IA x BI = IAI . IBI = IEI. IFI.
But A x B c E x F, and Lebesgue measure is complete so that we
must have E x F measurable and
IE x FI = ,u*(E x F) = IEI. IFI.
In order to remove the restriction of boundedness, apply the above
to E n S, , F n Sn, where S. S,', are spheres of radius n centre the origin
in k-space, l-space respectively. This shows that, for each n,
(E n Sn) x (Fn S.) E 2k+l,
I(EnS.)x(FnSn)I = IEnS.IIFnS;,,I
and the result follows from the continuity of measures on letting
n-.oo.1
Non-measurable sets
We have now seen that Lebesgue measure can be defined on 2k,
a large class of subsets of Rk, in such a way as to preserve the intuitive
geometrical ideas of volume. We also remarked earlier that it is
impossible to define such a measure on all subsets of Rk, so we now
demonstrate the existence of at least one subset which is not in 2'k.
Again we carry out the construction for k = 1. Consider subsets
E c (0,1] and for xE (0, 1] let E(x) be the set of real numbers z such
that z = x + y, yEE and x + y < 1,
or z = x+y-1, yEE and x+y > 1;
that is, E(x) is the result of translating E a distance x and then taking
the non-integer part. From property (i), it follows immediately that
EE2=> E(x)E2, IEI = IE(x)I.
Now let Z be the set of rationals in (0, 1]. Two sets Z(x1), Z(x2) will
be disjoint if (x1- x2) is irrational and identical if (x1- x2) is rational.
Let f be the class of disjoint sets of the form Z(x). By the axiom of
choice (see § 1.6) there is a set T containing precisely one point from
each of the sets in W. If Z is the set (r1, r2, ... ), we put
Qn = T (rn) (n = 1, 2, ... ).
Then
00
UQn=(0,1],
n-1
4-2
94 PROPERTIES OF MEASURE [4.4
since every point x E (0, 1] is in Z(xl) for some x1 and if q E Z(xl) n T,
we have q - x1 = rn so that x E Q. Also the sets Q. are disjoint as T
contains only one point from each set in and therefore cannot
contain two points differing by a rational. If T E 2, then Q. E 2
(n = 1, 2,...) and I TI = IQ.I (n = 1, 2, ...).
n=1
Exercises 4.4
1. Show that the set of points in [0, 1] whose binary expansion has zero
in all the even places is a Lebesgue measurable set of zero measure. Is it
a Borel set?
2. By changing the lengths of the extracted intervals in the construc-
tion of the Cantor set, show how to obtain a nowhere dense perfect set of
measure J.
3. Generalise (2) to show that for any e > 0 there is a nowhere dense,
perfect subset of [0, 1] with measure greater than 1-e.
4. Consider a union of sets of (3) to obtain a subset of [0, 1] of full measure
which is of the first category, and another subset of [0,1] of zero measure
which is of the second category.
5 Show that any bounded set in Euclidean space Rk has finite Lebesgue
outer measure. Is the converse of this statement true?
6. Suppose X is the circumference of a unit circle in R2. Show that there
is a unique measure aC defined on Borel subsets of X such that ,u(X) = 1
andp is invariant under all rotations of X into itself.
4.4] LEBESGUE MEASURE: PROPERTIES 95
7. By considering suitable approximating polygons (finite unions of
rectangles will do), show that the area of the plane region bounded by x = 1,
y = 0, y = x3 is J. Generalise to the case y = xk, where k > 0 but need not
be an integer.
8. Show that a subset E of a bounded interval I c Rk is measurable if,
for any e > 0 there are elementary figures Q1, Q2 a ek such that Q1 E,
Q2 I-E and IQ1I+IQ2I < III +e-
9. Suppose X is the unit square {(x, y): 0 < x < 1, 0 < y < 1}. If
E c [0,1] put 2 = {(x, y): 0 < y < 1] and let .' be the class of sets
P such that E is 21-measurable. Put µ(E) = IEI, and show that the subset
M = {(x, y):0 < x < 1, y = J} is not measurable with respect to the outer
measure p* generated by a on the class of all subsets of X. Show that
µ*(M) = 1, p*(X-M) = 1.
where V are the 2k vertices of the set I E .k and yy = + 1 for the vertex
in which each co-ordinate is largest and y, = (-1)f if the vertex Y
is such that r of its coordinates are at the lower bound (and (k - r)
at the upper bound). Any such function F is called a k-dimensional
Stieltjes measure function. With a little care. it is not difficult to show
that, under these conditions, ,uF is a non-negative additive set function
on 5k and that it therefore has a unique extension to to k. Either of
the arguments given in § 3.4 can now be modified to show that ,aF
is a measure on 1i k.
We can now apply theorem 4.2 to this measure It,, to extend it
to the o--ring 1k of Borel sets in Rk. As in the case of Lebesgue measure,
this extension automatically defines ,uF on the completion TF
of Oak with respect to It,. The class 2F is called the class of sets which
are Lebesgue-Stieltjes measurable for the function F. The class clearly
depends on the function F-for in the particular case F - c, TF
contains all subsets of Rk as 1t (Rk) = 0 and pF is complete; while
if F(x1, x2, ...xk) = x1x2 ... xk, then luF is the length function and
-T,k, is the Lebesgue class 2k.
Each of these measures ,up:YF --)- R+ is regular. The proof given
in theorem 4.5 can easily be modified to show this (we again do the
case k = 1) by using the fact that, for any e > 0, if (a, b] E 9, there is
a y > 0 such that
(a,b+y] (a,b+y) (a,b]
and #,(a, b + y) S #p (a, b + y] < ,aF, (a, b] + e,
to obtain economical coverings by open intervals.
Probability measure
Given a o'-field Fof subsets of 0, any measure P: F R+ such that
P(S2) = 1 is called a probability measure on F. If in addition F is
complete with respect to P we will say that the triple (S2, .F, P) form a
probability space.
Distribution function
A function F: R -* R is called a distribution function if
(i) F is monotonic increasing, continuous on the right;
(ii) F(x) -+ 0 as x -> - co, F(x) -+ 1 as x -* +oo.
4.5] LEBESGUE-STIELTJES MEASURE 97
A function F: Rk -- R is called a (k-dimensional) distribution function
if
(i) F is continuous on the right in each variable;
(ii) /AF.(I) > 0 for all I F_ 9k, where pp is defined by (4.5.1),
(iii) F(xl, x2, ..., xk) -+ 0 as any one of xl, x2, ..., xk-* - 00,
F(xl, x2, ..., xk) -* 1 as xl, x2, ..., xk all -+ +oo.
It is immediate from our definitions that any distribution function
F can be used to define a Lebesgue-Stieltjes measure OF on the
o--field 2'F. Further #,,(Rk) = 1 and ,aF is complete, so that every
distribution function determines a probability measure and (Rk,
°F, #F), is a probability space. There, is a sense in which these are the
only interesting probability measures on Rk.
Theorem 4.8. Suppose So is a o- field of sets in R,. contains the open
sets and ,a:.5o -* R+ is a complete measure which is finite on bounded sets
in Y. Then there is a Stieltjes measure function F: R --> R such that
. n .5°F and ,u coincides with It, on YF. If is a probability
space, then F can be chosen to be a distribution function.
Proof. Since contains the open sets and is a o--field, it must contain
-4, the Borel sets and in particular . 9, the class of half-open
intervals. Define F by
for x 0,
F(x) - {,u(0, x]
_p(x, 0] for x < 0.
Then F: R -> R is clearly defined and is monotonic increasing for all
real x (note that F(0) = 0). By theorem 3.2 (i), if {xn} is any monotonic
sequence decreasing to x, lim F(x,,) = F(x); since
n->- oo
Exercises 4.5
1. To see that condition (4.5.1) for k-dimensional Stieltjes measure
functions is not implied by the condition that F be monotonic increasing
in each variable separately consider
max (0,x1 + x2 -I-1) for x1 + x2 < 0,
F(x x,) =
t I for x1+x2 0.
Does this condition (4.5.1) imply that F is monotonic in each variable?
4.5] LEBESGUE-STIELTJES MEASURE 99
2. If F: R -> R is a Stieltjes measure function, show that
,aF(a,b) =F(b-0)-F(a), 1t_,[a,b] =F(b)-F(a-0)
and determine ,u, for intervals of the form
[a, b), (- c , a), (a, co)
3. If F is a Stieltjes measure function in R which generates the Stieltjes
measure ,u,, show that F(x) is continuous if and only if IuF{x} = 0 for all
single point sets {x}. What is the corresponding continuity condition in Rk?
4. Consider Lebesgue measure on 21-subsets of [0,1] and let E0 be a
subset of [0,1] which is non-measurable, such that the Lebesgue outer
measure of Eo and ([0,1]-E0) are both 1. Let .2 be the smallest 0--field
of subsets of [0,1] containing Eo and Y. Show that .2 consists of sets of
the form
E = A n Eo+B n ([0,1] - E0)
for A, B F2" and that #(E) = IA n [0,1]l defines a probability measure on
the a--field .2. By applying theorem 4.8 to this probability measure show
that, in general it is not possible to deduce in theorem 4.8 that .So = 2F.
5. Suppose F(x) - r0 for x < 0,
1 for x > 0.
Show that ,up(-1, 0) < F(0) - F(-1).
6. Give an example of a right-continuous monotone F such that
,up(a, b) < F(b) - F(a) < ,uF[a, b].
7. Show that, if F, G are distribution functions in Rk, then aF+bG is
a distribution function for any a > 0, b >, 0, a+b = 1.
8. In R2' 1 for xl >, 0, x2 3 0,
F(xi, x2)
fO for all other points.
Show that this F is a distribution function describing a unit mass at 0.
9. State and prove an n-dimensional form of theorem 4.8.
10. We can obtain completely additive set functions in RI which are not
necessarily non-negative by the following method. Suppose F: R -> R
is continuous on the right everywhere and of bounded variation in each
finite interval and F(b)-F(a) is bounded below for all a < b and define
TF(a, b] = F(b) -F(a).
Show that TF is additive on and can be extended to S By an extension
of theorem 4.2, TF can then be extended to a o--additive set function on a.
Now apply theorem 3.3 to express TF as the difference of two measures.
Finally, the argument of theorem 4.8 shows that TF is the difference of two
Stieltjes measures.
100
Dissection
0 n
If = U Ei and the sets Ei are disjoint, then El, E2, ..., E. are said
i=1
to form a (finite) dissection of Q. They are said to form an.'-dissec-
tion if, in addition, Ei E .F (i = 1, 2, ..., n).
Simple function
A function f: S2 -> R is called F-simple if it can be expressed as
n
f(x) = i1
=
cxhi(x)'
1
F'n+1= a-i=1
UEi$ 0
we can always put cn+1 0 and write
n+1
f = E ci xEi
i1
to see that the function is .f"-simple. If there is only one o--field .°F
under consideration we will talk of simple functions rather than F-
simple functions.
Lemma. The sum, difference and product of two simple functions is
a simple function.
Proof. Suppose we have the representations
n m
f =2Y- eixEi, 9 =jE djxdj;
=1 =J
then the sets Hi j = Ei n A j (i = 1, 2,..., n; j = 1, 2,..., m) are in .F and
form a dissection of Q. Further
f (x) = ci and g(x) = d, for xeHij, XH,j = XEi xd j
so that (f ± g) (x) = ci ± dj, (fg) (x) = ci dj for xeHij
n in n m
and f ± g =iZ (ci ± dj) XH, f9 =iE cidjxHij I
jZ jZ
5.2] SIMPLE AND MEASURABLE FUNCTIONS 103
Note that the constant functions
f(x) = c all xE S1
is the finite union of those sets Ei(e.F) for which ci < c, and is there-
fore in F. By condition (i) of the theorem, this implies that f is
measurable. 3
The next theorem examines further the relationship between simple
functions and measurable functions. It is both important and some-
what surprising.
Theorem 5.2. Any non-negative measurable function f : S2 - R+ is
the limit of a monotone increasing sequence of non-negative simple
functions.
Proof. For each positive integers, let
Q8 = (x: p-_28'
<f(x) < 2 } (p = 1,2,...,22,1);
228
so that {x:f(x) < c} differs from {x: g(x) < c} by a subset of a set of
zero measure. If .F is complete with respect to It, all such sets are in
.F so that
{x: f (x) < c} E .F if and only if {x: g(x) < c} E .F. J
Exercises 5.2
1. In theorem 5.1, show that the condition {x: f (x) S c} E.V for all rational
c is sufficient to imply that f : ) R* is 3---measurable.
2. Suppose {f,,) is a sequence of functions: S2 -> R* each of which is
finite a.e. Show that, for almost all x in S2, is finite for all n.
3. Suppose G is an open set in R and { is a convergent sequence of
functions: S2 R. Show that
{x: lim E G} = U U fl
00 {x: d(f (x), R- 0) > m),
n-.co m=1k=1n=k
where d(y, E) denotes the distance from y to E (defined in §2.1).
4. Show that, in theorem 5.2, the condition f >, 0 can be deleted provided
we do not require monotonicity for the sequence {f,,} of simple functions
converging to f. Show that if f is unbounded above and below it is impossible
to arrange for the sequence {f,,,} to be monotone.
5. An elementary function is one which assumes a countable set of values
each on a measurable subset of 0. Show that, if f: f R is measurable
then it is the uniform limit of a monotone sequence of elementary functions,
but that if f is not bounded it is not the uniform limit of simple functions.
6. If.'V is a finite field of subsets of S2, show that f : S2 -* R is 35-`-measurable
if and only if it is F.simple.
7. If S2 is a topological space, give examples to show that, for f : Cl -+ R,
the condition
`f is continuous a.e. in Cl'
110 PROPERTIES OF THE INTEGRAL [5.2
neither implies nor is implied by the condition
`there is a continuous g: S2 - R for which f = g a.e.'
8. Suppose S2 is a topological space,. .4 and (0, F, u) is such that
.F is complete with respect to It. Show that any function f which is continu-
ous a.e. is .F-measurable. Give an example of a measurable function which
cannot be made continuous by altering its values on any set of zero measure.
9. If 2 is the class of Lebesgue measurable sets in R, give an example
of an 2-measurable function f : R --> R and an 2-measurable set E c R
for which f-1(E) is not 2-measurable.
Hint. Use a suitable subset of the Cantor set (see §2.7).
ffczu E ci,u(Ei)
= i=1
This sum is always defined since each of the terms is non-negative.
It is called the integral off with respect to p. (Note that if ci = 0,
,u(Ei) = -boo our convention is that cip(Ei) = 0.) Since the repre-
sentation of a simple function in the form (5.3.1.) is not unique we
must first see that our definition of the integral does not depend on the
particular representation used. Suppose
f= n m
cj xE. = Fi dj XFj,
i=1 j=1
=Jfda+Jgdu.
It is now immediate from the definition that if a > 0, 8 > 0 and
q are non-negative simple functions then
r
f (af+ffg)du = a ffdu+ffJ 9du
so that our operator is linear on the class of non-negative simple
functions. It is also clear that it is order preserving; that is, if f, g
are simple functions and f > g then f fdu > f gdu.
These properties allow us to extend our definition to:
(2) Non-negative measurable functions
Given a measurable f : SZ --> R+, by theorem 5.2 there is a monotone
increasing sequence fn of simple functions such that fn ->.f. Since
112 PROPERTIES OF THE INTEGRAL [5.3
ff. d# is defined for all n, and is monotone increasing it has a limit in
R+ (which may be +oo). We define
ffd =lm f
n-oo J
fnd %(5.3.3)
Since there are many possible monotone sequences of simple functions
which converge to a given non-negative measurable f, we must show
that the integral f f du defined in this way is independent of the par-
ticular sequence used.
Suppose {fn} is an increasing sequence of non-negative simple
functions and f = lim fn > g, where g is non-negative simple. The
n-->.
first (and main) step in
00 showing that our definition (5.3.3) is proper
limu> J gdu.
n->oo
(5.3.4)
k
Put 9= ciXEi,
i=1
then if f g du = + oo, there must be an integer i, 1 < i S k such that
ci > 0, p(Ei) _ + oo. Then for any fixed e such that 0 < e < ci, the
sequence of sets {Ann Ej (n = 1, 2, ...) is monotone increasing to
Ej where
An = {x:fn+e > g}. (5.3.5)
Hence p(An n Ei) --,- + oo as n -* oo, by theorem 3.2. But
n-> oo
f
lim fnd1t = lim
mom J
gmdu.
fdu = ffxAda
fA
fa fdu = ffdlt
and we will usually continue to omit the set f when we are integrating
over the whole space.
Note that, if E e .°F and ,u(E) = 0, then any function f : 0 -> R*
is integrable over E with
fE1d = 0.
Exercises 5.3
1. Show that a simple function
f(x) = E C XE;(x)
is integrable if and only if c; = 0 for each integer i such that 4a(E) = +ao.
2. Let Sl be a finite set, u(E) the number of points in E. Show that all
functions on S2 are simple functions and that the theory of integration
reduces to the theory of finite sums.
3. If f: S2 -> R* is integrable (a) show that, for any e > 0
#rx: I f (x) I 1 e} Goo.
5.3] DEFINITION OF THE INTEGRAL 115
4. Suppose ,ul and 1a2 are two measures defined on f7 and v =111+,u2.
Show that if f is integrable with respect top and,u2 over a set E, then it is
integrable with respect to v and
f
= f+xAda- f f- xAda+ f f+xBdu- ff-xad1v;
ffdi = f/ld/L_ff2du.
f + g = (f++g+) - (f-+g-),
so that
Pf + g) du = f (f++g+) d4 - (f-+g-)dw f
f f
= f+dµ- f- da+ f g+d#- g-d# f
= ffdu+fgdu.
I ffdlt
I- ff+4 -ff-du s ff+d+f/_d.
(v) Ifc=0,cf=0and fcfdp=0=cffdu. Ifc> Othen
(cf)+ = of+, (cf)- = cf-
118 PROPERTIES OF THE INTEGRAL [5.4
and the result follows since it has already been proved for non-
negative functions (p. 113). Similarly, if c < 0
(cf ) + = - cf-, (cf ) - = - cf+,
f (cf)+da-f (cf)_du = (- c) ff- d1t + ff+du = cffdu.
C
then f = g a.e. For, suppose not, so that ,u{x: f(x) + g(x)} > 0. Then
5.41 PROPERTIES OF THE INTEGRAL 119
at least one of {x:f(x) > g(x)}, {x:f(x) < g(x)} has positive measure.
must be an integer n such that, for
By theorem 3.2 there{x:f(x)
Since the two extremes of the inequality are the same, we must have
as n -+ oo.
ffdu fIfIdP=ffNId#
+f,
IfI-IfNIdU<ie+ff2
(IfI-IfNI)da<e.3
Remark. The notion of absolute continuity for a set function v:
.F -+ R will be considered more fully in § 6.4.
Theorem 5.7 (Fatou). If {fn} is a sequence of measurable functions
which is bounded below by an integrable function, then
f liminffndp 5 liminfffndu
n--» n-aao
Remark. The operation lim inf picks out the small values of a
sequence. This theorem says that if this is done point-wise and the
result is then integrated the answer will be not greater than the result
of first integrating and then applying the operation.
5.4] PROPERTIES OF THE INTEGRAL 121
Proof. Since {fn} is bounded below by an integrable function g
we may assume without loss of generality that fn > 0 for all n. For
h,,.=fn-g> 0 a.e.t and
fhnd/ u = f/n du -J gdu, lim inf hn = lim inf fn - g a.e.
fd -*ffdiu.
Proof. (i) We first prove the special case of the theorem where
t Since g is integrable the set (x: jg(x) I = + oo } has zero measure, so that the opera-
tion f (x) -g(x) can be carried out at least outside the set (x: lg(x) l _ + co }. We put in
a.e. to cover the possible exceptional set of zero measure where (ff - g) is not defined.
By theorem 5.5 (viii) such exceptional sets do not effect the value of the integrals,
and we could arbitrarily define to be zero at the points where fa = g = ± oo.
122 PROPERTIES OF THE INTEGRAL [5.4
fn > 0 and fn - 0 as n -+ oo. In this case we can apply theorem 5.7
and corollary to give
lim sup flimsupfdP = fodfc = 0
= J lim inf ff,d# < lim sup fndµ.
J
Hence all the inequalities must be equalities, lim f fn du exists, and
has the value zero.
In the general case, put h,, = I fn -f I ; then 0 < hn < 2g, 2g is
integrable and hn is measurable with hn -* 0 as n -+ oo. But then
ffd1u_ffd/fIf_fJd1u -, 0 as n --> co,
00 fflfn!d4u< o.
n=1
w
Show that the series E fn(x) converges absolutely a.e. to an integrable
n=1
function f and that
ffd1i = E ffndlu
n=1
8. Suppose {Er,} is a sequence of sets in .°F, m is a fixed positive integer,
and G is the set of points which are in E. for at least m integers n. Then
G is measurable and 00
ff JgJ d# = y f I9Id1z.
= -n.
This shows that theorem 5.7 is not valid without the restriction that {
be bounded below by an integrable g.
12. State and prove a version of Fatou's lemma (theorem 5.7) for a
family ft, t e (a, b), of non-negative measurable functions.
13. Is it true that, for measurable f, g:12 R*,
f2 and g2 integrable =>fg integrable?
b f (x) dF(x)
Ja
unless we know that F is continuous for all x. (This condition is
sufficient to imply that the ,uF measure of single point sets is zero, so
that the integrals over El and E2 are the same).
Because Sfk is complete with respect to Lebesgue measure (2F
is complete with respect to pp) we see that if f : Rk -+ R* is integrable
and f = g a.e., then g: Rk --> R* is also integrable.
The theorems of § 5.4 were proved for any measure space (92, F, u)
so they are true in particular for Lebesgue measure in Rk. Thus the
Lebesgue integral is an order preserving linear operation on the class
.Qt of Lebesgue integrable functions. It is also a continuous operator
in the following senses.
Theorem 5.6 A. If {fn} (n = 1, 2,...) is a monotone increasing sequence
of non-negative Lebesgue measurable functions on Rk -> R+ and
f = lim fn,
Exercises 5.5
1. From first principles calculate the Lebesgue integrals
ro J1
(iii) fsf du, where It is Lebesgue measure in R2, f (X, y) = xy and S is the
unit square 0 . x . 1, O < y < 1.
2. Suppose f: R R* is Lebesgue integrable and
f 0<y<,f(x)}j.
a
Show that this definition is equivalent to the one we have given.
lim aJ f(x) dx
6+O+ e
0 for X E In,
hn(x) =
0 for x 0 Io.
Then for each integer n, x E Io
gn(x) < f(x) < hn(x);
130 PROPERTIES OF THE INTEGRAL [5.6
{gn} is a monotone increasing sequence of simple functions, and {hn}
is a monotone decreasing sequence of simple functions. If we put
g = lim gn, h = lim hn,
n--). OD n-1 Go
= limb-a
2"
-27EE Mn, = lim Sn, say.
n-->oo i=1 n-.w
We say that f is .?-integrable over [a, b] if, only if
b
lim sn = lim S. and 9 f (x) dx
n->oo a
so that, by theorem 5.5 (vii) g = h a.e. Since the set E of points where
f is discontinuous is contained in D u {x: g(x) + h(x)} it follows that
JET = 0. Further, since Lebesgue measure is complete, f is P-measur-
able and, by theorem 5.5 (viii),
b g(x) dx = b h(x) dx
YEa Yfa
so that f is . -integrable. I
5.6] CONDITIONS FOR INTEGRABILITY 131
Theorem 5.9 shows that .P-integrable functions have to be con-
tinuous at most points. We have many examples of 2-integrable
functions which are continuous nowhere. However, there is a sense in
which even 2-integrable functions have to be approximable by con-
tinuous functions-in fact by functions which are arbitrarily smooth,
that is, functions that can be differentiated arbitrarily often.
Theorem 5.10. Given any 2'-integrable function f: R -* R* and any
e> 0 there is a finite interval (a, b), and a bounded function g: R --> R
such that g(x) vanishes outside (a, b), is infinitely differentiable for all real
xand
Yflfx_uxiIdx < e.
Proof. We carry out the approximation in 4 stages.
(i) First, find a finite interval [a, b] and a bounded measurable
function fi which vanishes outside [a, b] and is such that
I/(x)-fi(x)I dx < Je.
Yf
This can be done by considering the sequence of functions
f (x) if x e [ - n, n] and If(x) I < n,
if x E f - n. n1 and f(x) > n.
a..(Xi
-n It xE[-n,nj ana j(x) < -n,
0 if xo[-n,n].
Then gn(x) -+ f(x) for all x and I gnI < If 1. By theorem 5.8 it follows
that
-If(x)-gn(x)Idx-*0 as n->oo
f
so that we can fix a sufficiently large N and put fi(x) = gN(x).
(ii) The next step is to approximate fl by an'-simple function/2
which vanishes outside [a, b] and satisfies
Exercises 5.6
1. In theorem 5.10 it was shown that any integrable function f could be
appproximated by a step function g in the sense that
fIfx)_(xi dx < e.
Show that in general it is not possible to arrange at the same time that
g <f-
2. Show if f : R -> R* is integrable, then
Rectangle
Any set in X x Y of the form E x F with E c X, F c Y is called a
rectangle (set).
Product of classes
If '', 3 denote classes of subsets in X, Y respectively, then ' x -9
denotes the class of all rectangles E x F with E E ', F E.9.
Sections of functions
Given any function f: X x Y R*; for each fixed x E X,
fx(y) = f (x, y) for y E Y
defines a function Y -* R* and for each fixed yE Y, ff(x) = f(x, y)
for x E X defines a function on X to R*. These functions fx, fv are
called the sections of f at x, y, respectively.
Corollary. Under the conditions of theorem 6.1, given any .-measur-
able function f: X x Y --> R*, each of the sections fx(y) is 9-measurable
and each of the sections ff(x) is .°F-measurable.
Proof. Suppose xa is a fixed point in X and M is a Borel set in R*;
then
{y: ,,.(Y) E.411 = {y: f (x0, y) E M} = {(x, Y): f (X, y) E M}x0
so that the test set is the section at x0 of a set in .'.
The results of theorem 6.1 and corollary can be extended in an
obvious way to finite Cartesian products X. X X2 x ... X X,1; there is
no difficulty in making the required modifications to the definitions
and proofs. It is not quite so immediate that they can also be extended
to arbitrary Cartesian products jj Xi.
iel
Let us recall that a point in jj Xi can be thought of as a function
iEl
f : I U X j such that f(i) E Xi for each i E I. Suppose then that we
iEI
have a collection {Xi, i E I} of spaces and o--fields .i of subsets of Xi.
Cylinder set
If i1, i2, ..., in is any finite subset of I and EjkE rzk, k = 1, ..., n; the
set of points f E IIXi such that
f(ik) E Eik (k = 1, 2, ..., n),
is said to be a (finite dimensional) cylinder set in 11X1. When we say
that f is in a cylinder set C, the values off are restricted only on a
finite set of indices. The class of all such cylinder sets in IIXi will be
denoted i E I).
Lemma. The class i E I) of cylinder sets is a semi-ring of subsets
ofiEI
IIXi.
Proof. We can think of .j as a semi-ring in Xi which contains the
whole space X. Then if two sets
A = {f: f(i)EEi,iEJ},
B = {f: f(i)EFF,iEK},
6.11 SUBSETS IN A PRODUCT SPACE 137
are in '(.Fi, i e 1), J and K must be finite subsets of I, and each of the
sets Ei, Fi must be in the relevant .j. The set J v K = L
is also a finite subset of I and, if we put
Ei =Xi for iEK-J, Fi = Xi for iEJ-K,
then A = {f : f (i) E Ei, i E L}, B ={ f: f (i) E Fi, i E L},
are now cylinder sets in which the same finite subset L of indices are
restricted. Since we know that any finite Cartesian product of semi-
rings is a semi-ring, we can deduce that A - B is a finite disjoint union
of sets of this type and A n B is a set of this type. Hence W(Aj, i E I) is
a semi-ring.
Note. The case I = Z is important. The cylinder sets in II Xi
i-1
00
then reduce to sets of the form E1 x E2 x ... x E. x jj Xi with
i=n+1
Ei E.Fj (i = 1, 2, ..., n).
00
Exercises 6.1
set E in rj Xi the (finite dimensional) section of E at x1, x2, ..., xk is the set
i=1
00
(in n Xi) of points (xk+i, xk+2, ...) such that (x1, x2, ...) r :E. Then if EE Y1
\\ i=k+1
the product o--field in rj Xi, all its k-dimensional sections belong to .Sok+1
i=1
ffdi = Ia1(E)uu2(F)
Now for each fixed x in X1 we have
U(EixFi)=ExF
i=1
because all the functions fi(x) are non-negative measurable, so that
the monotone convergence theorem 5.6 justifies the inversion of
integration and summation. Thus It is a measure on the semi-ring
.F1 x .F2. It can be extended uniquely by theorem 3.5 to the generated
ring, and then, by theorem 4.2, to the generated o--ring which is the
product o--field F1 * .5F2. The result is called the product measure
on <F1*c.F2. We have thus proved
Theorem 6.2. Given two measure spaces (X1, #i,1u1), (X21 'F21,02) such
that,u17u2 are o --finite, there is a unique measure a defined on the product
o field Al * ffl72 in X1 x X2 such that
It is clear that 0 < ,u(E) < 1 for all E in V. To see that It is finitely
additive on ' it is sufficient to see that, in any finite collection of
cylinder sets, only a finite number of coordinates are involved so that,
m
if C = U Cj is a dissection of C E' into disjoint sets of le, there is an
j=1
integer N such that C and Cj (j = 1, ..., m) can all be expressed in
the form
E1xE2x...xENx jj Xi. 00
i=N+1
We can then apply theorem 6.2 to the finite products to see that
m
fi(C) = E,u(Cj).
j=1
µ(E1x...xEnx
)
= lu1(E1) fi2(E2) . . . run(E.)
\\
ft
i=n+1
Exercises 6.2
f hdA=Jfdu=Jgdv,
where f(x) = fhdv when hx is v-integrable,
f+(x) = fh:dv
6.3] FUBINI'3 THEOREM 145
will always be defined, though it may take the value +oo. Since
ff+(x) dp exists, we must have f+ finite except for a set of zero u-
measure. Similarly, f- is finite almost everywhere. If we put
f(x) = f+(x) -.f-(x)
when both f+, f- are finite and f(x) = 0 otherwise, we see that
f hdA = fh+dA_fh_dA
= ff+du- rf-dµ = ffda.
(iii) Again split f into positive and negative parts. Since
0 <f+ < If 1,
we can apply (i) to each of the positive and negative parts to deduce
that f f+dA and ff-dA are both finite. The result now follows by (ii).
We should remark that theorem 6.5 is one of the most useful tools
in the theory of integration as we have developed it. This result again
exhibits the power and neatness of the absolute integral.
We have been careful to define the product measure A on the
smallest .V which contains F x T. Some authors define pro-
duct measure to be the completion of this A obtained by the process
of theorem 4.3. If one uses this definition then some of our statements
have to be modified to exclude possible subsets of zero measure,
though the essential content of the results remain valid. In particular,
given a function f(x, y) which is measurable with respect to the com-
pleted ar-field . f°, one can only say that the section fx is T-measurable
for almost all x. However, provided F and T are complete with respect
to their respective measures, theorem 6.5 remains valid as stated.
We can use our definition of product measure to give an alternative
definition of the integral of a non-negative measurable function.
Theorem 6.6. Suppose (S2, -,µ) is a o --finite measure space, (R, ., v)
denotes the real line with Lebesgue measure on it and z is the product
measure ,a x v defined on the product crfield dY in t x R. Then if
E E F and f : E -;,- R+ is non-negative, f is F-measurable over E if and
only if Q(E,f) E .ye, and in this case,
fE fd1i = r(Q(E',f));
where Q(E,f) is the ordinate set defined by
{(x,y):xEE,yR,0<y<f(x)}.
146 RELATED SPACES AND MEASURES [6.3
Proof. Suppose first that Q(E, f) e.. Then by theorem 6.1 all
its sections are in F. But the section of Q(E, f) at y = a is the set
{x:f(x)>a},
so that by definition, f is .F-measurable. Conversely, if f is .F-
measurable then there is a sequence {fti} of .F-simple non-negative
functions which increases to f. Now for any F-simple function
Q(E, fn) is a finite union of measurable rectangles and is therefore
in .-Y. Also Q(E, fn) increases monotonely to Q(E,f) so we must have
Q(E,f) E .-Y. Further if r
A = i-1 Cn, 4 xEn {
Exercises 6.3
1. Suppose S2 is any set of cardinal greater than X0, and F is the o-field
of sets in fI which are either countable or have a countable complement.
For EeJF, put p(E) = 0 if E is 1 if (S2-E) is countable.
Consider the Cartesian product of two copies of S2 and let E be a set in
S2 x SZ which has countable x-sections for every x and y-sections whose
complement is countable for every y. If is the indicator function of E, then
h
fhu(x)au(dx) = 1, fh(Y)(dY) = 0.
Why does this not contradict theorem 6.4?
2. Suppose (X, .F,#) (Y, OF, v) are o -finite measure spaces and A is the
product measure on the product a-field A. Show that
(i) If E, G c: A' are such that v(E.,) = v(G,,) for almost all x e X, then
A(E) = A(G).
(ii) If f, g are integrable functions on X, Y then f (x) g(y) is integrable on
XxYand
ff(x) g(y) dA = ffdufYdv.
3. X = Y = [0,1] an&F, 9 are the Borel subsets. Let p(E) be the Lebes-
gue measure of E, v(E) the number of points in E. Form the product mea-
sure It x v on Borel subsets of the unit square. Then if D is the diagonal
{(x, y); x = y}, D is measurable and
f{ffx,YdY}dx = 4 ,
0 0
-1 (f 1 f(x, ) = = f(ff(x,Y)dY)dx
but the integral over the unit square in R2 does not exist.
6. Given a countable collection of probability spaces (X2, .j u;) and
the product measure ,u on the product v-field, we can form the finite product
measures T.. = µ1 X P2 X ... x pn and the product measure A on the product
space rj X j. Then, if f (x1, x2, ...) is any p-integrable function on rj Xj we
{=n+1 i 1
have
fdu= jftx1, x, ...) d n dTn.
v(E) = fE fdu
J
defines a finite valued absolutely continuous set function v. In fact,
in § 5.4 we proved that v was and that (corollary to theorem
5.6) given e > 0, there is a 8 > 0 such that for E e.F,
p(E) < S' Iv(E)I < e. (6.4.1)
It is immediate that any set function v which satisfies (6.4.1) is
absolutely continuous with respect to It. The conditions are equiva-
lent for finite measures, but not in general (see exercise 6.4 (4)).
There is a partial converse given by:
Lemma. If (S2, F, p) is a measure space and v: F ->. R is finite valued,
ar-additive and absolutely continuous with respect to ,u, then v satisfies
condition (6.4.1).
6.4] RADON-NIKODYM THEOREM 149
Proof. By the decomposition of § 3.2, any such v is the difference of
two finite measures, so it is sufficient to prove the result for a measure v.
Then if (6.4.1) is false, there is an e > 0 and a sequence {En} of
sets of F such that v(En) > e and ,u(En) < 2-n. Put E = lim sup En.
Then
#(E) 5 p U Er) <, j U(Er) < 2-r`,
r=n+1 n+1
so that ,u(E) = 0 while
v(E) = lim v 1 u Er) > lim sup v(Er)
r=n+1
so that v(E) > e. This contradicts v << p. I
Thus we see that the indefinite integral of an integrable function
defines an absolutely continuous set function. Our object in the present
section is to obtain the converse of this statement under suitable
conditions. It is convenient at the same time to consider a more
general o--additive set function and to decompose it into a maximal
absolutely continuous component and a remainder which has to be
concentrated on a p-null set. It is convenient to give a further
definition.
Singular set function
Given a measure space a set function v: F - R* is said
to be singular with respect to p if there is a set E0E.F for which
p(E0) = 0 and v(E) = v(E n Eo), all EE.F. (6.4.2)
This condition clearly means that the parts of SZ outside the null set
E0 make no contribution to P. In fact if v is also a measure we see that
S2 can be dissected into two sets Eo, El E. such that
Eo n El = 0, Eo v E1 = S2, ,u(Eo) = 0, v(E1) = 0.
The symmetry of the relationship in this case is sometimes stressed
by saying that It and v are mutually singular.
Theorem 6.7. Given a afinite measure space (S2, . It) and a o--additive,
o--finite set function v, then there is a unique decomposition
V = V1+V2
into set functions vi which are ofinite and such that v1 is
singular with respect to ,u and v2 << p. Further there is a finite valued
measurable f: S2 --> R such that
v2(E) = gda
fE
for all E in .F, then f(x) = g(x) except in a set of zero,-measure.
Corollary. Under the conditions of the theorem if v < It then there is a
finite valued f : S2 -. R such that
v1- V3 = V4 - V21
Taking the union of support sets of v1, v3 gives a set Eo such that
(v1- v3) (E) = (v1- v3) (E n E0), ,a(Eo) = 0.
But (v4 - v2) is absolutely continuous and therefore zero on any null
set so that, for any E E .F,
(v4-v2)(E') = (v1-v3)(E) = (v1-v3)(EnEo)
= (v4 - v2) (E n Eo) = 0.
Thus vl(E) = v3(E), v2(E) = v4(E) for all E. The uniqueness of the
integral representation of v2 was proved in § 5.4. Thus it is sufficient
to find any decomposition and integral representation.
By theorem 3.3 we can decompose v into the difference of two mea-
sures. It is therefore sufficient to prove the theorem when v is a
measure. Now let .-° be the class of non-negative measurable
f: U -* R+
Exercises 6.4
1. Show that if µ, v are any two measures on a a-ring SP, then v < µ -{- v.
2. Suppose F(x) is the Cantor function defined in §2.7 and v is the
Lebesgue-Stieltjes measure with respect to F. Show that v is singular with
respect to Lebesgue measure.
3. Suppose (S2, .F, µ) is a measure space with µ(S2) < ac and visa measure,
v << It. Show there is a set & such that (S2 - E) has v-finite v measure and for
every measurable F c E, v(F) is either 0 or oo.
4. Let t be the set of positive integers,
µ(E) = E 2-n, v(E) 2n
nEE nEE
then v < µ, but (6.4.1) is not satisfied. This shows that (6.4.1) is a stronger
condition than absolute continuity when v is not finite.
5. Suppose Q is an uncountable set,." is the class of sets which are either
countable or have countable complements. For E e.ri", put µ(E) = the
number of points in E, v(E) = 0 or 1 according as E is countable or not. Then
clearly v <#, but no integral representation is possible. This shows that
in the Radon-Nikodym theorem we cannot do without the condition that
µ be ar-finite.
6.41 RADON-NIKODYM THEOREM 153
6. If A, ,u, v are or-finite measures on IF and A << It, ,u << v; show that
A<<vand
dA_dAdu
dv du dv
except on a set of zero A-measure.
7. A, a are v-finite measures on F with y < A. Then if f is ,u-integrable
ffdu = J f d dA.
8. If A,# are a--finite measures on F such that ,u <A and A <# then
d1t dA)
TA -(
except for a set of zero A-measure.
9. If ,u, v are o--finite measures on F such that v << ,u, show that the set
of points x at which dv/du is zero has zero v-measure.
10. Suppose {,u1} is a countable family of finite measures on a o,-field F.
Show that there exists a finite a on such that each of the pi is absolutely
continuous with respect to It.
11. Suppose
lun = r Ilk - lu,
n n
vn = E vk - v,
k=1 k=1
where all the u, v with suffices are finite measures on a or-field 317 and vn
is din-continuous for all n. Show that
(i) du1/dun - du1ldc almost everywhere (f1).
(ii) If each,un is v-continuous then d7n/dv --* a.e. (v).
(iii) v is 71-continuous and dvn/dun -* dv/du a.e. (F1).
Measurable transformation
Suppose f is a mapping from X into Y, F is a in X and 9r
is a o--field in Y, then we say that f is a measurable transformation
from (X, F) into (Y, 9) if f-1(E) EJF for every E in K This condition
can also be written f-1(g) c ,F.
In Chapter 5 we discussed `measurable functions'. In our new
terminology these are measurable transformations from (X, .F)
to (R*, 9) in which is the oS-field of Borel sets in R*. Given mappings
f:X Y, g: Y -± Z we can consider the composition g(f) : X -* Z
defined by g(f) (x) = g(f(x)). In particular if g: Y -* R* is an extended
real-valued function on Y, then g(f) defines an extended real function
on X.
Lemma. If f: X -a Y is a measurable transformation from (X,F)
into (Y, 9) and g: Y --> R* is EQ-measurable as a function with extended
real values, then the composition g(f) is .F-measurable.
Proof. For any Borel set B in R* we have
{x:g(f)(x)EB} = f-1 {y: g(y) E BI
= f-1(E) for some E E 9,
and is therefore in F. ]
Remark. We obtained a special case of this lemma when we proved
that a Borel measurable function of a measurable function is measur-
able (see §5.2).
If we start with a measure space (X, and f is a measurable
transformation from (X, F) into (Y, 9) it is natural to use f to define
a measure v on 9 by putting
v(E) _ ,a(f-1(E)) for E E 9. (6.5.1)
With this definition of v it is immediate that (Y, 9, v) is a measure
space. If (6.5.1) holds we will write v = ,uf-1. This allows us to carry
out a `change of variable' in an integral.
Theorem 6.8. Suppose f is a measurable transformation from a measure
space (X,.F,It) to (Y, 9) and g: Y -* R* is T-measurable: then
fd(f_1) = fg(f)du
in the sense that if either integral exists so does the other and the two are
equal
Proof. It is clearly sufficient to consider non-negative functions
6.5] MAPPINGS OF MEASURE SPACES 155
g: Y -). R+. Suppose first that g = XF, the indicator function of a set
E in 9. Then
g(f) (x) = 1 if x E f-' (E),
= 0 if x of'-'(E);
so that g(f) is the indicator function off-1(E), a set in F. Thus, in this
case, by (6.5.1)
Theorem 6.9. Given o-finite measure spaces (X, .F, ,u) and (Y, T, v)
and a measurable transformation f from (X, F) into (Y, T) such that
µf -' is absolutely continuous with respect to v
f
f g(f)du = g.Odv,
where 0 is the Radon-Nikodym derivative d(,af-')/dv, for every measur-
able g: Y -* R* in the sense that, if either integral exists, so does the
other and the two are equal.
f g(.f) du = f gd(1uf-1).
6 TIT
156 RELATED SPACES AND MEASURES [6.5
Since # -1 is absolutely continuous with respect to v, there exists a
measurable 0 such that, for every E E 9
/F(E) = fE' F
for intervals of 9 and therefore for all sets E E Y. Hence the measure
pp is absolutely continuous with respect to Lebesgue measure ,u
and q is a possible definition of the Radon-Nikodym derivative
daF/da. The second equality now follows from theorem 6.9.
Remark. It is clear from the above that the function y5 (or q in the
corollary) plays the part of the Jacobian (or rather the absolute
value of the Jacobian) in the theory of transformations of multiple
integrals. In general it is not easy to obtain an explicit value for the
Radon-Nikodym derivative d(,af-1)/dv, but in important special
cases this can be done. In particular, if both spaces are (Rk, 21k, ,u)
with ,a Lebesgue measure, and f: Rk -> Rk is a linear transformation
given by a non-singular matrix A so that y = Ax one can prove that
If(E)I = IIA[I.IEI,
Exercises 6.5
1. Show that the composition of two measurable transformations is
measurable.
2. If f is a measurable transformation from (X, -5F) into (Y, .P) and u, v
are two measures on such that u << v, show that ,uf-1 << vf-1.
3. (Integration by parts.) If F(x), G(x) are non-negative continuous
functions satisfying the conditions of §4.5 for a Stieltjes measure function
and E is any Borel set, then
If we put Vn = Dl n D2 n ... n D.
n
then T(Vn) = T(En)-T(En-Vn) i T(En)- ET(Ei-Di) > J8
i=1
JbEexpL-(62-61)21d62
2(t2-t1) J
f blexp(-*) d 1.
ai
162 RELATED SPACES AND MEASURES [6.6
The fact that this defines a consistent family of measures on .So which
can be extended to all sets of the form (6.6.2) can be proved directly
(it will follow from the discussion of the multinormal distribution in
Chapter 14). Hence, we can apply theorem 6.10 to give a measure ,a
on IT the class of Borel sets in RT = Q. This is called Wiener measure
in the space of functions f: T --> R, and is an example of a stochastic
process which will be discussed more fully in Chapter 15.
However, let us use the example of Wiener measure to illustrate
the inadequacy of theorem 6.10. This follows from the fact that the
o--field _IT is too small to contain interesting sets-for we have seen
that it contains no set in which a non-countable set of time coordinates
is restricted. Even if PT is completed with respect to a to give a
probability measure, the completed o-field is still too small. For
if A c RT is a set in which f is restricted at a non-countable set, the only
set of .IT which is contained in A is the empty set. This means that
A can only be measurable, if it has measure zero. But the same
argument applies to the complement of A so that if both A and its
complement involve restrictions on f : T R at a non-countable set
of indices, then the outer measure of A must be 1, and the inner mea-
sure must be 0. In particular the set
f f E RT : a < f (t) < b for all t E [tl, t2]} (6.6.4)
is not measurable, and if C is the set of functions f: T - R which are
continuous for all t E T, C also has outer measure 1 (and inner measure
0). Various methods can be used to extend ,a from .VT to a larger
v-field which includes C and (6.6.4) and other sets of interest. These
have been studied in detail and the interested reader is advised to
look in J. L. Doob, Stochastic Processes (Wiley, 1953).
6.7 Applications
In the second part of this book random variables will be defined as
.p-measurable functions f: S2 -+ R* where (S2, $ a) is a probability
space. Although it is usual to work with general carrier spaces S2,
there is a sense in which the real line R has a structure sufficiently
complicated to reproduce all the probability properties of the function f.
In fact, in many treatises on probability theory, the carrier space S2
is barely mentioned. This attitude is partially justified by the follow-
ing considerations.
Suppose (52, .° ', p) is any finite measure space and f : S2 -+ R is
.F-measurable. For all real x, define
F(x) = p{y: f(y) e x}. (6.7.1)
6.71 APPLICATIONS 163
Then F(x) -* 0 as x ->. -oo, F(x) --> ,u(S2) as x -->- +oo, and F: R ->- R
is continuous on the right. Thus we can define a Stieltjes measure
#p using this particular F.
Theorem 6.11. Suppose (02, .F, ft) is a finite measure space and!: c2 R
is .F-measurable, ,uF is the Lebesgue-Stieltjes measure in R given by
(6.7.1) and g: R R is Borel measurable, then g(f) is $ -measurable and
,u{x: g(f) (x) E B) is determined by uF for every Borel set B. Further
= /tF(B) = fdPF.
By linearity our result follows for non-negative simple functions and
the monotone convergence theorem then gives it for non-negative
Borel measurable g and then for all integrable g.
Corollary. In the notation of the theorem
ffdu = fxdF(x).
Remark. There is an n-dimensional form of theorem 6.11 and
corollary which links the behaviour of n measurable functions with a
Lebesgue-Stieltjes distribution in Rn-see Chapter 14.
Marginal distributions
Not all measures in product spaces are product measures. Suppose
X, Y are spaces, then the projection X x Y --* X given by p(x, y) = x
defines a mapping. This will be a measurable transformation on
(X x Y, .$) into (X,.9') provided E x Y E .F for every E E Y. In this
case, if It is a finite measure on J F, pp-1 defines a measure on Y. In
164 RELATED SPACES AND MEASURES [6.7
general it may not be a very interesting measure as there may be no
sets of finite positive measure. However, if (X x Y, . ,1u) is a finite
measure space, then the measure pp-1 on .9' is called the marginal
measure on X. The marginal measure on Y is similarly defined using
a projection on Y. If It is a probability measure these marginal mea-
sures are called marginal (probability) distributions.
If F(x, y) is a distribution function in R2 (see § 4.5) then
lim F(x, y) and lim F(x, y)
7H+M
will again define 1-dimensional distribution functions, and it is im-
mediate from theorem 4.8 that the corresponding Lebesgue-Stieltjes
measures will be the marginal distributions of,UF. If
F(x, y) = F1(x) F2(y)
is the product of two 1-dimensional distributions, then ,uF will be
the completion of the product measure 1aF, x,uF2 and F1, F2 will be
the marginal distributions for F. Conversely, if ,uF is a product mea-
sure, then it must be the product of its marginal distributions so that
F(x, y) = F1(x) E2(y) is a necessary as well as a sufficient condition for
FiF to be a product of two probability measures.
Thick subsets
For any finite measure space we can generate the outer
measure 00
Exercises 6.7
1. Formulate and prove a theorem of the form of theorem 6.11 for n
F-measurable functions fi: 52 -. R (i = 1, 2, ..., n).
2. Find the 2-dimensional distribution function F(x, y) which generates
the measure µF such that uF(R) is 11V2 (length of diagonal D in B) for any
rectangle R, where D is the segment joining (0, 0) to (1, 1). Calculate the
marginal distributions of µF, and show that µF is not a product measure.
3. If is a complete o--finite measure space, and the outer mea-
sure,a* is defined by (6.7.2) show that a set E is ,u*-measurable if and only
if it is in F.
4. Suppose (52, is a finite measure space and E. is a subset of 4
such that, for A1,
A1nE0=A2rE'o=µ(A1)=µ(A2)
Prove that E. is thick in Q.
166
so that ,a(E) = 0 and I f(x)I < C outside E. Thus I f(x)! < C a.e., and
if we define f(x) if I f(x)I <' C,
0 if If(x)I >C,
168 SPACE OF MEASURABLE FUNCTIONS 17.1
then I f * (x) I < C for all x and f * =f a.e. Further {x: I f *(x) l > C - e)
has positive measure for all e > 0, so that it is non-empty and we must
have sup If* I = C. It is clear that, if f = g a.e., then ess sup f = ess sup g,
so that we can think of ess sup as a functional on the subset Y,'C' of
the essentially bounded functions of -9. If we define (af +/3g) by
(af +/ig) (x) = af(x) +/3g(x) when f(x), g(x) E R,
=0 otherwise;
it is clear that (af +,6g) E _ if f, g E _W,0 for any a,,8 E R so that Y.,
is a linear subspace of .mil (over the reals) Further
P.(f, g) = ess sup if - gi
defines a metric in Y., for
(i) Pc(f,g) = P.(g,f);
(ii) p. (f, g) = 0 if only if f = g a.e.;
(iii) ess sup I f + gl < ess sup I f I +ess sup I g I so that
p. (f, g) < p, (f, h) +P.(h, g)
Now it is clear that, if {fn} and f are functions in - such that
fn -->.f uniformly a.e., then pro(fn, f) -* 0 as n -> oo. Conversely sup-
pose po(f f) --> 0, and let E. be a set of .F with,a(En) = 0 and
ess sup l fn -f I = sup I fn(x) -f(x) I
EQ-En
Put E = U En, then for x E SZ - E
I fn(x) -f(x) 15 sup I fn.(x) -f(x) I = ess sup If,, -fl
x e n -En
so thatfn --> f uniformly on SZ - E and #(E) = 0. A similar, but slightly
more complicated argument shows that, in 2 , a Cauchy sequence
uniformly a.e. is the same as a Cauchy sequence in p. norm.
Almost uniform convergence
Given functions fn: E -+ R* (n = 1, 2, ...) and f: E R* each of
which is finite a.e. on E we say that fn converges almost uniformly to
f on E if, for each e > 0, there is a set Fe ' E, FEE .F, ,u(F6) < e such that
fn -->f uniformly on (E-FE). The example E = [0,1] c R, fn(x) = xn
It Lebesgue measure shows that it is possible for a sequence to con-
verge almost uniformly on E while it does not converge uniformly a.e.
on E. However, it is immediate from the definitions that convergence
uniformly a.e. implies almost uniform convergence. What is more
surprising is that, under suitable conditions, convergence a.e. implies
almost uniform convergence.
7.11 POINT-WISE CONVERGENCE 169
Theorem 7.1. (Egoroff). Suppose E oo, and { fn} is a sequence
of measurable functions on E -> R* which are finite a.e. and converge
a.e. to a function f: E -> J2* which is also finite a.e. Then fn -+ f almost
uniformly in E.
Proof. By omitting a subset of E of zero measure, we may assume
that all the functions fn and f are finite and that
fn(x)-*f(x) for all xEE.
For positive integers, m, n put
ill
Then, for fixed m, Ai', A,-,..., An, ... is an increasing sequence of
measurable sets converging to E. Since #(E) is finite, by theorem 3.2
there is a positive integer N. = Nm(m) such that
,u(E - A') < e/2'n for i > Nm.
a)
If we put FE U (E-AmNm
M=1
then ,u(FE) < e. Further given S > 0 we can choose m so that 1/m < 8
and then
fi(x)- f(x) I < S for all i > Nm, xE (E-F,),
so that fn --> f uniformly on (E - FE).
Remark. The converse to theorem 7.1 is true and almost trivial.
For if {fn}, f are finite a.e. on E, measurable, and fn -->. f almost uni-
formly, this means we can find sets F. with ,a(F.) < 1/n such that
fn -> f uniformly on (E - Fn) and so fn f point-wise on (E - Fn).
Put a)
F = ll Fn, then ,u(F) = 0
n=1
andfn -,,-f point-wise on (E-F) so thatf,, - f a.e. on E.
Exercises 7.1
1. Let X be the space of positive integers, class of all subsets of X,
and,u(E) the number of integers in E c X. If fn(x) is the indicator function
of {1, 2,..., n}, then f,,,(x) -- 1 for all x. However, fn does not converge almost
uniformly to 1, showing that theorem 7.1 is false without µ(E) < oo.
2. Suppose the conditions of theorem 7.1 are satisfied except that
,u(E) = eo, show that given P > 0, there is a subset Fp c E with ,u(Fp) > P
such that fn f uniformly on Fp but that there need not be a subset F
with µ(F) _ +co with fn -* f uniformly on F.
170 SPACE OF MEASURABLE FUNCTIONS [7.1
3. Suppose EE,"-", ,u(E) < oo, f,,: E -* R* (n = 1, 2,...) is a Cauchy
sequence a.e. of measurable funtions each finite a.e. Prove there is a finite
c and a measurable F c E with ,u(F) > 0 such that, for every integer n,
all x E F, If (x) < c.
4. Suppose E E.°, E has v-finite measure, f (n = 1, 2,...) and f are finite
a.e. on E and f -* f a.e. on E. Show that there exists a sequence {E;} of
sets in .° such that
,u(E- U Ezl = 0 and d, -* f
\\ i=1
uniformly on each Ej. By considering the measure of example 2, §3.1,
and a suitable sequence of functions show that the condition that E has
v-finite measure is essential.
5. In §4.4 we produced a sequence of sets each of which was not
Lebesgue measurable. If we put f (x) = indicator function of
n
[0, 1) - LJ Q2, then f (x) -* 0 for all x in [0,1].
i=1
Show that f does not converge almost uniformly so that theorem 7.1 fails
if the functions are not measurable.
6. Suppose fa: E -* R, h > 0 is a continuous family of measurable func-
tions, each finite valued, #(E) < oo and for each x E E, f,,(x) --* f (x) as h 0
where f is finite valued. Then if a continuous parameter version of Egoroff's
theorem were valid we would have given e > 0, there exists Fs
FE e E, ,u(FE) < e such that f h(x) -* f (x) ash -* 0 uniformly on (E -FE). The
following example shows that this extension is false. In Chapter 4, we
saw that there is a non-measurable set E e [0, 1) such that every point
x e [0, 1) has a unique representation x = y + q (mod 1), y e E, q rational.
Prove that, if M is a measurable subset of [0, 1] such that M n E(r) is
non-void for finitely many rationale r, then I M I = 0.
Arrange the rationals Q as a sequence For x E [0, 1) let n(x) be the
integer such that x = y + y E E. If x/n(x) _ al a2 ... (decimal repre-
sentation not ending in 9 recurring), put O(x) _ /31/32... where Ak = ak
(k = 1, 2, ...); and N2k-1 = 1 for k = n(x), 0 otherwise. Put fh(x) = 1,
for x = ¢(h), fh(x) = 0 otherwise. Prove fh(x) -* 0 as h 0 for each x.
Show that if M any measurable set, I M > 0, then f h(x) +i 0 uniformly
on M.
7. Suppose f f,,} is a sequence of functions in .2, In f a.e. and there is
an integrable function g such that < g a.e. for all n. Show that f,, -a f
almost uniformly.
8. Define what is meant by saying that a sequence {fn} of a.e. finite
valued functions is a Cauchy sequence almost uniformly, and show that this
implies the existence of a limit function f such that fn -* f almost uniformly.
7.2] CONVERGENCE IN MEASURE 171
Note that the definition only makes sense for functions in W which
are finite a.e. We first see that the limit in measure is unique in .alt'.
For suppose fn --> f in measure, fn g in measure; then if 8 > 0,
{x: I AX) - g(x) I > 8} C {x: I fn(x) -Ax) I > 18} v {x: I fn(x) - g(x) I > Zs}
and both sets on the right can be made of arbitrarily small measure
by choosing n large. This means that
,u{x: I&) - g(x) I > S} = 0 for each S > 0,
and it follows that f = g a.e. (by taking a sequence 8n decreasing to
zero).
We say that the sequence { fn} of functions in. ' is a Cauchy sequence
in measure if, given e > 0, 8 > 0 there is an integer N such that
n > N, m > N - ,u{x: I fn(x) - fm(x) I > e} < 8.
The argument used to prove uniqueness of the limit also shows that
fn ->.f in measure {fn} is a Cauchy sequence in measure.
The converse is included in the following theorem.
Fk= i=k
UEi.
00
so that the sequence fmi converges uniformly on (S2 - Fk); that is,
since ,u(Fk) -> 0 as k -> oo, it is a Cauchy sequence almost uniformly.
(iv) By (iii) we can obtain a subsequencefmk of the given sequence
which is a Cauchy sequence almost uniformly. This means we can
find a function g E.% such that fmk - g almost uniformly as k -> oo.
Now, for e > 0,
{x:lfn(x) - g(x)I > e} -- {x:Ifn(X) -fmk(x)I >- e}
Er,k = k2kr
r[__ 1
(r = 1, 2,..., 2k; k = 1, 2, ...),
so that, for any e > 0, ,a{x: I fn(x) I > e} 5 ,u(F.) ->. 0. This means that
fn 0 in measure in [0, 1]. However, at no point x in [0, 1] does
f,,(x)--> 0; in fact, since every x is in infinitely many of the sets F. and
infinitely many of the sets (S2 - F,,,) we have
lim inf f,,, (x) = 0, lim sup fn(x) = 1 for all x E [0, 1].
Exercises 7.2
1. Suppose { fn} is a Cauchy sequence in measure, and fnt, f,n, are two sub-
sequences which converge to f, g, respectively. Prove that f = g a.e.
2. Show that if {f} is a Cauchy sequence in measure then every subse-
quence of { fn} is also a Cauchy sequence in measure.
3. If S2 is the set of positive integers and, a is the counting measure on the
class 0T of all subsets, show that convergence in measure is equivalent to
uniform convergence.
4. If #(S2) = co can we say that convergence a.e. implies convergence
in measure?
5. Suppose {An} is a sequence of sets in,'Z;', xn is the indicator function
of A,,, and d(A, B) = ,u(A A B) for A, Be ". Show that is a Cauchy
sequence in measure if and only if d(A,,, A.) -* 0 as n, m -> oo.
6. Suppose {fn} is a sequence of functions of M which are finite a.e.
and fn -* f a.e. with f finite a.e. Show that, if (i) ,u(S2) < cc, or <, go
for all n where go is integrable; then f,, -> f in measure.
7. Suppose (S2, is a finite measure space and { fn}, are finite
valued F-measurable functions which converge in measure to f, g respec-
tively. Show
(i) If,,l converges in measure to If ;
(ii) for all real a, 6 the sequence converges in measure to
(af+fig);
174 SPACE OF MEASURABLE FUNCTIONS [7.2
(iii) if f = 0 a.e., then fn converges in measure to f2;
(iv) the sequence {f,,g} converges in measure to fg;
(v) the sequence {f,,2,} converges in measure to f 2;
(vi) the sequence converges in measure to fg;
(vii) if f + 0 a.e. all n, f + 0 a.e., the sequence {1/fn} converges in measure
to 1/f.
Is the condition µ(S2) < oo essential for all these results?
The class .p
For p > 1, a function f in M is said to be of class Lp if If I p is ,u-
integrable. Since
21f(x)I, if If(x)I > Ig(x)I,
l.f(x)+g(x)I <
2lg(x)l, if Ig(x)I > If(x)I;
we have, for all x,
If(x)+g(x)Ip S 2p{lf(x)Ip+Ig(x)Ip}. (7.3.1)
Thus, if f, g E Lp we must have (f ± g) e Lp. With the usual convention
about the set of zero measure where (af +,6g) may not be defined, it
follows that Lp is a linear space. For f, g e LP we define
isp
Pp(f,g) _
Ifif-gipdµJ
and notice again that pp(f, g) = 0 if and only if f = g,a.e. so that in the
space Yp c .ill of equivalence classes we have
Pp(f, g) = Pp(g,f),
pp(f, g) = 0 if and only if f =gin gyp.
We will prove in the next section that pp satisfies the triangle in-
equality, which shows that it is a metric in rp. However, we can now
define:
Fk = U Ei,
i=k
we have u(Ek) < 2-k e, µ(Fk) < 21-k e, and if x is not in Fk,
fNi+i(x) - fNi(x)I < e2-i for all i >, k.
fIf_frIh/<6 if r>N(e)
so that fr -+ f in pth mean.
It is worth remarking at this stage that the theorem corresponding
to theorem 7.3 for Riemann integrals over a finite interval is false.
It is not difficult to construct an example of a sequence of functions
whose pth powers are Riemann integrable and which Cauchy con-
verges in pth mean, but for which the limit is necessarily discontinuous
on a set of positive measure and so cannot be Riemann integrable
by theorem 5.9 (see exercise 7.3 (10)). Thus theorem 7.3 exhibits
another way in which the Lebesgue integral is a big improvement on
the Riemann integral.
We now relate convergence in pth mean to convergence in measure.
Theorem 7.4. If {f.) is a sequence of functions of Lp (p 3 1) which is a
Cauchy sequence in pth mean then { is a Cauchy sequence in measure.
If fn -+ f in pth mean, then fn -+ f in measure.
Proof. For any h in LP, r/ > 0
rli{x: I h(x) I > V1/2P} > V1 fI hI pdu i r.
7.3] CONVERGENCE IN pTH MEAN 177
If {fn} is not a Cauchy sequence in measure, then there is an e > 0,
8> 0 for which
P'{x: I fn(x) - fm(x) I '> e} > 8
fifn(X)_fm(X)I4a i I> 0
for infinitely many n, m so that {fn} is not a Cauchy sequence in pth
mean. This proves the first statement: the second part of the theorem
is proved similarly.
Remark. The example after theorem 7.2 shows that {fn} may con-
verge in pth mean but not converge a.e., though theorems 7.2, 7.3
together show that there must be a subsequence {fnti} which converges
a.e. If we consider Lebesgue measure in R and put
n-1/P for x in [0, n],
f -W =
{0 otherwise,
ni/P for x in [0, 1/n],
{0 otherwise.
we see thatfn --> 0 uniformly (and therefore almost uniformly, a.e.,
and in measure) but not in pth mean. If t = [0, 1], then gn 0
almost uniformly, a.e. and in measure, but not in pth mean so that
even in a finite measure space we cannot deduce convergence in mean
from other types of convergence without some additional condition,
even if the functions concerned are all in Y p. The next definition
turns out to be appropriate:
vn(B) = fBIP
I fn P d< 2p+3' (7.3.3)
For each in, n put
(4Ae 11P
Cm.n = {x: I fm(x) -fn(x) I > ) )
ffIVd#< 21
Cm,fn Cm, n Cm, n
Uniform
* pth mean
Pointwise *
I T t
a.e. * Almos t a.e. * * Almost
uniform uniform
*
In measure In measure
Fig. 3. y(i) < oo. Fig. 4. f JP < h a.e.,
Exercises 7.3
1. Check Figures 2, 3, 4, stating in each case the theorem or theorems
which justifies A --* B, or the example which justifies the exclusion of
ABB.
2. Show that, if lc(f) < oo, then the condition in theorem 7.5 that {vn}
be equicontinuous at 0 can be replaced by a condition that {vn} is uni-
formly absolutely continuous.
3. Show that if {fn} Cauchy converges uniformly a.e. and each fn is
integrable over E with ,u(E) finite, then f (x) = lim f ,(x) a.e. is integrable
over E and
fE Ifn-.fI d1z-0 as n -> oo.
4. Suppose 0 is set of positive integers and ,u is the counting measure.
Then
(i) If 1/n for 1 < k <, n,
f .(k)
{0- for k > n,
show that fn(x) -->- 0 uniformly on Q, each f, is integrable but
ffnd/2-» f
This shows ,u(E) < oo is essential in question (3).
(ii) For the same sequence {f} show that
vn(E) = fE f. d#
is uniformly absolutely continuous, but not equicontinuous at 0. This
shows that the condition ,a(0) < co is essential in question (2).
(iii) If
1/k for 1 < k < n,
fn(k)
- 10 for k > n,
show that { fn} is uniformly convergent on 01, each fn is integrable, but the
limit is not.
182 SPACE OF MEASURABLE FUNCTIONS [7.3
5. Show that, if {f,,,} converges in pth mean to f, and g is essentially
bounded, then f f,, g} converges in pth mean to fg.
6. Show that if
vn(E) = fndu (n=1,2,.-.),
fIf.d =x
is a Cauchy sequence of real numbers for every E E .F, and { f,j is a Cauchy
sequence in measure. Give an example of a sequence which does not
converge in measure, for which
lim fE
for all E.
8. Suppose 1u(L) < oo, and for f, gE.,', and a.e. finite;
AM) = f l+If-gl
If-gI dµ
Show that p defines a metric in the space of a.e. finite functions of .4', and
that convergence in this metric is equivalent to convergence in measure.
9. Suppose #(Q) < co (1 < q < p). Show that Yi zD _a Yv Y,,, and
that p.(f, 0) = lim p9(f, 0) for f E 2,,. Show that the finite measure con-
dition is essential.
By considering a suitable function on [0, 1] show that 2' + n Y,, but
v>1
that if f E l Yp -then p,(f, 0) --* oo asp -* oo.
T>1
10. Suppose S2 = [0,1], ,u is Lebesgue measure. Let K be a nowhere
dense perfect set with positive measure and let {Ek} be the set of disjoint
open intervals such that (0, 1) - K = (J00 Ei. Let fn be the indicator function
i=1
n
of F,, = U Ei. Prove that fn (n = 1, 2,...) is Riemann integrable and con-
verges in mean to the indicator function f of (0, 1) - K. By considering the
construction of K, show that f is discontinuous a.e. on the set K of positive
measure, and so cannot be Riemann integrable. This shows that the class
of Riemann integrable functions is not complete with respect to convergence
in mean.
7.41 INEQUALITIES 183
7.4 Inequalities
We now obtain some inequalities which turn out to be important in
several branches of analysis. We need to use the algebraic inequality
xayfi < ax +/3y (7.4.1)
for x > O, y > 0, a > 0, fl > 0, a+ fl = 1, which is strict unless x = y.
This is most easily proved by taking logarithms to give
alogx+/3logy ( log (ax+fly)
and using the fact that log: R+ -> R is strictly concave since it has a
negative second derivative.
Conjugate indices
If p > 1, q > 1 and 1/p+ 1/q = 1, we say that p and q are conjugate
indices.
Theorem 7.7 (Holder). If p, q are conjugate indices, and f e Lp, g r: LQ
then fg is integrable and, for each E in F,
d)1/p
!glQd/A)
fE d1u< (fE IfI (fE .
The inequality is strict unless there exist real numbers a, b such that
alfIP=blgl4a.e.onE.
Proof. If
f IfI du = 0,
then the loose inequality is certainly satisfied and if the right-hand
side is also zero then either f = 0 a.e. on E or g = 0 a.e. on E. In either
case the condition alf Ip = bIgIQ a.e. is satisfied with b = 0 or a = 0,
respectively. Hence we may assume that
fIfgId>0.
Put El = {x: Ig(x)l < If(x)Ip-1},
E2 = S2 - El.
Then for xeEl, If(x)g(x)I < If(x)IP,
for x e E2, I f(x) g(x) I S I g(x) I Ql
so that I f (x) g(x) 15 I f (x) I P + l g(x) I Q, for all x,
this implies that fg is integrable.
184 SPACE OF MEASURABLE FUNCTIONS [7.4
Given E e . , put E0 = E n {x : f (x) g(x) = 0}. Then by our assump-
tion ,a(E - E0) > 0. For x e E - E0, we can apply (7.4.1) replacing
This gives
E-E,,
Iflpd#
fE-E0 Iglgdu
If < If(x)IP + Ig(x)Ig
\p f Iflpdi q fE-Eo IglgduE-Eo
(7.4.2)
If we now integrate over (E - E0) and note that the right-hand side
gives 1/p+ 1/q = 1, we obtain the desired inequality for the integral
over (E - E0). We can only obtain equality for the integrals if we have
equality in (7.4.2) for almost all x E (E - E0). The condition for
equality in (7.4.1) now shows that we must have aI f I P = blglq a.e. on
(E - E0) where
it
a= E-E IfIpdu and
0
b= fE- 0
IgIgdu.
The inequality for the integrals over E now follows, and we can
again only have equality if f = g = 0 a.e. on E0, since otherwise the
right-hand side is increased while the left-hand side remains the same
on replacing (E - E0) by E. Thus equality can only occur if all I P = b I g I g
a.e. on E. 1
Remark 1. The special case p = q = 2 of theorem 7.7 is called
Schwartz's inequality. A simpler proof of this case is possible. See
exercise 5.4 (13).
Remark 2. In the sense of theorem 7.7 the index conjugate top =1
is q = oo. It is easy to prove directly that, if f E L g E L. then
f1f1dt< (flflit)esssuPlxEI)
Theorem 7.8 (Minkowski). For p >, 1, if f, g E Lp then (f + g) E LP and
for any EEC,
(fi)l/p+(fEIglpd1a)l/p
(fEIf+gIPdl-t)l/p
\
For p > 1, equality is strict unless there are real numbers a > 0, b > 0
such that of = bg a.e. on E.
7.4] INEQUALITIES 185
Proof. We already proved in §7.3 that Lp is a linear space. For
p = 1, the result is immediate. For p > 1,
fElf+glpdu=fElf+gIIf+gIp-ld1t
f EIfIIf+gIp-1da+ f II If+ gI p-1du,
with equality if and only if f and g have the same sign a.e. in E.
If we now apply theorem 7.7 to each of these integrals we obtain
!'du)114
(fE If+gIpd1)
to obtain the desired result. If
fE If + gIP dy = 0,
in the sense that the convergence of both series on the right implies the
convergence of the series on the left and the inequality. Further, equality
is only possible if there is a constant k such that l ail p = kl bil4 for all i.
Minkowski.
co
lai+bilp
llp
(i=1
co
lailP)llp + (Z'
with equality if and only if there is a k > 0 such that ai = kbi for all i.
=1
lbilP)llp
Exercises 7.4
f E lfghl dp
\ (IE lflVI d1,)a
(fE IgIllad)ie (fE l
hlllyd1u)'.
5/k dp =Jfdu.
For a fixed m, let ak be the first m-leader, and let (ak + ... + ak+p_1)
be the shortest non-negative sum that it leads. Then for every integer
h with k < h < k + p -1, we must have ah + ah+1 + ... + ak+p-1 > 0,
so that ah is an m-leader. Now continue with the first m-leader in
ak+p, ..., an and repeat the argument until all the m-leaders have been
found. It follows that the sum of all the m-leaders of the original
sequence must be non-negative, as it is the same as the sum of the
non-negative shortest sums obtained by the above procedure.
We can now turn to the proof of the lemma and notice that, since f
is integrable, we may assume that it is everywhere finite valued. If
E. denotes the set of x such that
n
E fi(x) i 0
i= O
For a positive integer n, let s(x) be the sum of the m-leaders of the
finite sequence fo(x), f1(x), ...,fn+m_1(x). Let Ak be the set of x for which
fk(x) is an m-leader and let xk be the indicator function of Ak. Since
n+m-1
Ak is measurable, and s(x) = Z xk(x) fk(x), s is measurable and in-
k=0
tegrable and s(x) > 0 so that
n+m-1
,I
k=0
fdkfkda 0. (7.5.1)
Now notice that, for k = 1, 2, ..., n-1, T(x) E Ak_1 if and only if
fk-1(Tx) + ... +fk_l+p-1(Tx) > 0 for some p < m, which is equivalent
to fk(x) + ... + f k+p-1(x) > 0 for some p < m which in turn is the
condition for X E Ak. Thus Ak = T-'Ak-1 = T-kA0 for k = 1, ..., n -1.
Hence by theorem 6.8,
f akfk(x)dit=JT-"f(Tkx)d#= f df(x)da
7-2
190 SPACE OF MEASURABLE FUNCTIONS [7.5
so that the first n terms of (7.5.1) are all equal. Now A0 = Em, so that
(7.5.1) implies r
nfEmfdu+mJJfjda> 0.
fEf_sxd j_
/ 0,
fB (f - 8) d# > 0.
511*1 du s f IfI du
(iii) For fixed n, put
D(k, n) _ {x:
k k+1
< f* (X) <
and apply the lemma to the transformation T on the set D(k, n) which
can be assumed to be invariant. Then f*(x) 3 k/2n in D(k, n), so that
at least one of the sums
n-1 k
E (fi(x)-2n+ > 0
i=0
for each e > 0. Hence
k
L fdu >
k, n)
(_e)u(D(k,n)),
and so we must have
fdu % 2k ,u(D(k,n))
D(k,n)
192 SPACE OF MEASURABL E FUNCTIONS [7.5
Similarly (' fd \ k+1 (/D(k, n))
D(k,n) 2n
and
k
µ (D(k, n)) c f f dµ 5
k2 1 µ (D( k , n)).
J D(k,n)
For each integer k, it follows that
*d d 2nµ(D(k,n));
J D(k, n)f JD(k, n)f
and if we sum over k
Ifaf * d1i -fnfdu I
< I #(Q)
Since n is arbitrary we must have f f * dµ = f f du. ]
For applications to statistical mechanics one would expect the
equilibrium value f*(x) to be independent of the point x, so that the
limit function f* of theorem 7.9 is a constant. Unfortunately this is
not true without imposing an additional condition.
Ergodic transformation
A measure preserving transformation T is said to be ergodic (or
metrically transitive or metrically invariant) if for all invariant sets
E (sets for which T-1(E) = (E)), µ(E) = 0 or,u(S2-E) = 0.
Lemma. T is ergodic if and only if every measurable invariant function
is constant a.e.
Proof. Suppose g is measurable and invariant. Then {x: g(x) > a}
is invariant for all real a, and must either have zero measure or be
the complement of a set of zero measure. Hence g = constant a.e., if
T is ergodic. Conversely, if every measurable invariant function is
constant a.e., since the indicator function of an invariant set is an
invariant function, there cannot be any invariant sets other than
null sets and complements of null sets. ]
We can now apply this lemma to theorem 7.9 when T is ergodic.
There are two cases:
(i) µ(S2) = + oo. Since the only constant which is integrable over a
space of infinite measure is zero we deduce that
l n-1
Efi(x) 0 a.e.
-ni=0
(ii) µ(S2) < oo. We can integrate f * = c a.e. by (iii) to obtain
l n-1 1
LINEAR FUNCTIONALS
In this chapter all measure spaces (S2, F, p) will be o-finite, and F
will be complete with respect top, unless stated otherwise. In Chapter 7
we saw that Y. (1 < p < oo) with the metric
llp
Pp(f,g) _ (f If-glpd,a}
and Y. with the metric
P.(f,g) = esssupIf-gl,
were complete metric spaces. We also proved they were linear spaces
(over the reals); and it is immediate that the metric defines a norm
Ilflip = P1(.ff 0) (1 < p < co),
for which the spaces are normed linear spaces. Thus
Ilfllp > 0 if f + 0, I1011 = 01
Using #(E) < co, this means that such a function g can be found with
= f f-EIfI2dp+f if-gl2dp
< 1e2.
so that IIf - gll < 1e. If all the rk in the representation of g are zero we
are finished, so there is no loss in generality in assuming they are all
non-zero. Since ' is a basis for ,u and p(Ek) < oo, we can find sets
Ck of such that
2
,u(EkLCk) < Gr n) (k= 1,2,...,n).
k
E 2
Then II rkXEk-rkXC, II2 = (2n
so that, if
n n
h = F+rkXCk, we have IIg - hII , = II rkX k-rkXC.II < JE
k=1 k=1
Inner product
For any normed linear space K over the reals a function (f, g) on
K x K ->. R is called an inner product (or scalar product) if
(i) V, g) = (g,f);
(11) (fl+f2,g) = (f1,g)+(f2,g);
(iii) (Af, g) = A(f, g), for A E R;
(iv) (ff) = 11f J12.
For the normed linear space 'T2 we can define
(f, g) = P9 dl t, f, g E '42,
since, by theorem 7.7, the productfg is integrable. It is a simple matter
to check that, with this definition, (f, g) satisfies all the conditions
(i)-(iv) for an inner product.
Exercises 8.1
1. For any normed linear space with an inner product, prove that
V 'O '< IlfiIIIgII.
Hint. Consider (f + 8g, f + Og) >, 0 for all real 6.
2. If (f, g) = 0 in a normed linear space, show that
Ilf+gll2 = ilfll2+Ilgll2.
3. Suppose (92, .v u) is a discrete measure space, i.e. there is a sequence
{pi} of reals with E 1 pi I < oo and a sequence {xi} in f2 such that p(E) = F, pi.
x=eE
Prove that 22(f2 u) is separable.
4. If (Y, 9, v) are v-finite measure spaces each with countable
bases, show that the product measure A = It x v on X x Y also has a count-
able basis.
Hint. Consider finite unions of rectangles which are products of basic
sets.
5. Generalise the above to countable products of spaces
with ,ui(Xi) = 1. The example (8) below shows that it does not extend to
arbitrary products.
6. Let Q. be any set and define the counting measure ,u(E) = number of
points in E when E is finite; #(E) = + oo otherwise. Show (i) if f2 is countable,
the finite subsets of 0 form a countable basis; (ii) if f2 is uncountable, there
is no countable basis for ,u.
7. Show that any Lebesgue-Stieltjes measure (Rk, has a count-
able basis.
8.1] DEPENDENCE OF 22 ON (S2, . µ) 199
8. Suppose I is a non-countable index set and for a e I, Xa denotes
a 2-point space {0, 1} with µa{0} = ,aa{1} = 1. Form the product measure
p on the Cartesian product j-j {0, 1} = S2. Show that there is no countable
ael
basis for p.
9. Show that, if p has a countable basis, then 2(S2,µ), I < p < co
is a separable space.
Complete set
A family {ca} (a c: I) in a normed linear space K is said to form a com-
plete set if its closed linear span is the whole space.
Suppose now that a normed linear space K is separable, so that
there is a sequence x1, x2, ..., xn, ... of points dense in K. By omitting,
in turn, any point in the sequence which can be expressed as a linear
200 LINEAR FUNCTIONALS [8.2
combination of the previous ones we obtain a sequence 9.1 g2, ...
which is linearly independent, and has the same linear span as {x,z}.
Since {xn} is dense, the closed linear span of {gn} must therefore be the
whole space. Thus in any separable normed linear space there is a
complete set of independent points which is either finite or enumerable.
If there is a finite complete set (g1, g2, ..., gi.) of independent points and
K has an inner product, then we will see that K is isomorphic to
Euclidean k-space. For K = Y2(S2,1a), it is easy to see that K is
finite-dimensional if ,a is a discrete measure concentrated on a finite
set of points, for then the indicator functions of these individual
points will form a finite complete set. However, the interesting Y2-
spaces are infinite-dimensional. Any (S2, .° u) for which contains
infinitely many disjoint sets, each of finite positive measure, clearly
generates an infinite-dimensional since the indicator func-
tions of this sequence form an independent set.
Orthogonal system
Two points x, y in a normed linear space K with an inner product
are said to be orthogonal if (x, y) = 0. Any class %} (i E I) of points
of K which are different from zero and pairwise orthogonal is called an
orthogonal system. A non-zero element x E K is said to be normalized
if Jlxii = 1, i.e. (x, x) = 1. An orthogonal system of normalized points
is said to be an orthonormal system in K. Thus {O2} (i E I) is an ortho-
normal system if
1 for i = j E I,
10 for i + j.
Now any orthogonal system of points is certainly linearly indepen-
dent for, if we take the inner product of (8.2.1) with O; we obtain
cf(c,, 0f) = 0, so that cf = 0. Further, if K is separable, any ortho-
gonal system in K is countable. For any such system can be normal-
ised to give an orthonormal system {oi} (i E I), and then
JJ 0z - Y'7II = V2 for i 4 j ;
and, if {x,,} is a countable dense set, we can find for each i E I an
integer nz such that 11 x'i - 5zlj < J and this gives
11xni-xnjJJ > 4J2-1 > 0 for i + j, so that n z + nj.
In the study of finite-dimensional normed linear spaces it is helpful
to use a (finite) orthogonal normalized basis. In the general case, at
least for K separable, it is also possible to find a complete orthonormal
8.2] ORTHOGONAL SYSTEMS OF FUNCTIONS 201
sequence for K. This can be done by first obtaining a complete in-
dependent sequence and then orthogonalising it by the process of the
next theorem.
Theorem 8.2 (Gram-Schmidt orthogonalisation process). If K is a
normed linear space with an inner product and x1, x2, ..., xn, ... is a
linearly independent sequence in K, then there is an orthonormal sequence
y11 y21... Yn, ... such that
(1) yn = an1x14an2x24... +annxn, ann 4 0;
(ii) xn = bnlyl+bn2y2+...+bnnyn, bnn + 0;
where a11, b11 are real numbers. Further each y1 is uniquely determined
(up to the sign) by these conditions.
Proof. If yl = ax1, then
(yl, yl) = a2(x1, x1) = 1
if a is suitably chosen. The conditions are therefore satisfied with
n = 1 if b11 = 1/a,1 = , J(x1, x2). (Note that the linear independence
condition ensures that Il xlll + 0.) For n > 1, suppose y1, y2, ... , yn_1
have been found to satisfy all the conditions. Then
xn = bn1y1+... +bn,n-lyn-l+zn,
where bnj = (xn, yi) (i = 1, 2, ..., n - 1) so that (zn, yz) = 0 for i < n.
We must have (zn, zn) > 0, since otherwise zn = 0 and x1, x2, ..., xn
would be linearly dependent. If we put
Zn II
yn = // bnn = N(zn,zn),
(z zn))
then (yn, yn) = 1, (yn, y2) = 0 for i < n and (ii) is satisfied. We can
then deduce (i) since bnn 4 0. By induction the method of ortho-
gonalisation is established.
The uniqueness of the process (apart from sign) follows since the
values of the constants are all determined except for the ± sign in
the square root which occurs at each stage. J
Corollary. If is separable, then there is a complete ortho-
normal sequence in Y2.
Proof. Start with a sequence {fn} which is dense in 22, and replace
it first by a sequence {gn} of linearly independent functions with the
same closed linear span. If this is an infinite sequence, use the process
of theorem 8.2 to obtain the orthonormal sequence {hn}. It is clear that
this sequence has the same closed linear span as {gn} so that it is a com-
plete set. On the other hand, if 22 is finite dimensional, we will obtain
202 LINEAR FUNCTIONALS [8.2
a finite set {g1, 92, ... , 9n} whose linear span is 2'2. This finite set
can be replaced by a finite orthonormal set using the process of
theorem 8.2.]
In practice it is not always easy to prove that a given orthonormal
sequence {01, 452, ...} is complete. Various methods for proving com-
pleteness will be given in the next section.
Exercises 8.2
1. Suppose L1 = [0,1), ,u is Lebesgue measure, fo(x) - 1,
+1 if 2n-1x - y e [0, J) (mod 1),
fn(x) -{-1
if 2n-ix=ye[j,1) (mod1).
The functions fn: [0, 1) -> R are called the Rademacher functions. Show
that they form an orthonormal sequence in
2. For Q = [ - rr, 7r], It Lebesgue measure, show that the trigonometric
functions
1 1 1 1 1,
V2 rr
,
- cos x,
n
sin x, ...,
TV
cos nx, - sin nx, ...
Fourier coefficients
Given an orthonormal family (ei), j e J in a Hilbert space H, and
any point x E H, the real numbers
C, = (x, ei) (j e J),
are called the Fourier coefficients of x on the orthonormal family,
and the series
E ci ei
JEJ
is called the Fourier series of x. (Note we have not yet said in what
sense (if any) this series converges.)
The choice of the Fourier coefficients c, can be justified as follows.
If I is a finite subset of the index set J, re-label the indices 1, 2, ..., n
and consider the partial sum
n
sn = ai ei (n = 1, 2, ... ).
i=1
n n
Then 11sn-x112 = x- Zaiei, x- i=1
Zaiei
i=1
x n n
= JJxJJ1-2E(x,aiei)+E1 E1(aiei, a,ef),
n n
JJxJJ2-2aici+ ai,
i=1 i=1
n n
so that Ilsn-xll2 = 11x112- Ci (ai-Ci)2. (8.3.2)
i=1 i=1
Thus Ilsn - xll will be a minimum when all the terms of the last series
in (8.3.2) are zero, and aiei is the best approximation (in norm)
i=1
204 LINEAR FUNCTIONALS [8.3
to x when all the ai are Fourier coefficients. This generalises the well-
known geometrical theorem (for Rn) which states that the length of
the perpendicular from a point to a plane is smaller than the length of
k
any other line joining the point to the plane: for (x - I c i e,) is ortho-
\ i=1
k
gonal to all linear combinations of the form fi ei.
i=1
Bessel's inequality
We can make another deduction from (8.3.2) by noting that
I1sn-x112 > 0.
If we put ai = ci we obtain
n
k=1
=1
(!Ixll2.
If we now define c to be the supremum of c, for all finite subsets
jEJ jEI
I c J we find that
e IIxI12, (8.3.3)
jEJ
and this is known as Bessel's inequality. It follows as an immediate
corollary that at most countably many Fourier coefficients of a point
in H can be non-zero.
Theorem 8.3. If {ej} (j E J) is an orthonormalfamily in a Hilbert space,
each of the following conditions is equivalent to {ej} being a complete
family
(i) Z e; = JJx112 for every x E H (Parseval),
jEJ
where {c,} are the Fourier coefficients of x;
(ii) The finite partial sums sI = Eckek of the Fourier series of x
converge to x in norm for all x E H.
Note. For any arbitrary index set J we say that Z xj converges in
jEI
norm to x if, given e > 0, there is a finite set K such that if I is finite
and K c I c J then
11 j xj - xII < e.
It is easy to check that, when J is countable and the xj are real so
that we have a real series this notion reduces to the usual definition
of absolute convergence.
Proof. The conditions (i) and (ii) are clearly equivalent by (8.3.2).
Now suppose that (ii) is satisfied. Then any x can be approximated in
norm by a finite sum sn which is a linear combination of e1, e2, ..., en.
8.31 RIESZ-FISCHER THEOREM 205
Hence, each x is in the closed linear span of {ej} and the sequence must
form a complete system.
Conversely suppose {ej} is a complete system. Then given e > 0,
N
x E H, there is a finite sum y = E ai ei for which l l x - y II < e. But,
i=1
if SN is the corresponding partial Fourier sum, we know
ilx-y112 i Ilx-sNII2,
N
so that, by (8.3.2) Ec2 >' 11x112-e.
i=1
i=n+1
Since E f2 converges, it follows that {sn} is a Cauchy sequence in norm.
Since H is complete, there must be an x e H such that 8n -+ x in norm.
Further
(x, ei) = (sn, ei) + (x - sn, ei)
= Ni+(x-8n,e1) for n >, i
and, by exercise 8.1 (1)
I (x-sn, ei)I <- lleill Ilx-snll = IIx-snll --> 0 as n -9 oo.
Since (x, ei) is independent of n, we have & = (x, ei) for all i. J
206 LINEAR FUNCTIONALS [8.3
Corollary. An orthonormal system {ee} (j E J) in a Hilbert space H
forms a complete system if and only if the only point x E H which is
orthogonal to all the {e;} is the point x = 0.
Proof. Suppose {e1} is a complete orthonormal set and (x, ef) = 0
for all j. Then all the Fourier coefficients of x are zero and so
IIxl12=Zc1=0.
IEJ
for which Z ck converges is called the space 12. By using the discrete
k=1
version of theorem 7.7 one can check that if {ci}, {di} E l2,
00
llx(1)+x(2)1!2 = Ilx(1)II2+2(x(1)
x(2))+Ilx(2)II2
F..(Cz1))2+2}rc'c2)+ (c2))2,
n
.
sin x, ... ,
Exercises 8.3
m
1. Prove that a series E an of real terms converges absolutely to s if and
n=1
only if, for each e > 0 there is a finite set I e Z such that for every finite K
with I C K c Z we have
(s - Eanl < e.
nEK
2. If S2 = [ - it, 7f],µ is Lebesgue measure, f E Y2 (S2,µ)
am = -
1f f (x) cos mx dx (m = 0,1, 2, ... );
7f
IT
1
bm = f (x) sin mx dx (m = 1, 2, ... ),
_n
then the a,,,, bm are the classical Fourier coefficients off. Prove:
OD
0"
jao+ E (a2 +b2,,) < 00,
then there is a function f E 2'2 for which these are the Fourier coefficients
and such that
sn(x)
= Jao + En (am cos mx + bm sin mx) -± f
m=1
in second mean;
(iii) if {an}, {bn} are the Fourier coefficients of f in the above sense
n
8.(X) = Ja0+ E (am cos mx+bm sin mx),
m=1
o'n(x) [so(x)+sl(x)+...+sn(x)],
= n+1
then vn(x) -> f (x) in Y. norm (and in fact on -* f a.e.);
1 _nncr n
Eat+E(ak+bk)(___)a
(iv) (x)d/2= n-F1
n o0
< ao+(ak+bk)
1
<' ao+E(ak+bk);
1
E.fIT r
(v) since v(x) dµ-* f 2(x) dwe have for f
J Ir
show that lp is separable. I., is the space of bounded real sequences with
jjxli = sup lxil. By considering sequences of 0's and 1's show that l., is not
separable. Deduce that Y., is not separable if S2 has infinite but v-finite
measure.
Remark. In the above theorem, the only property of the norm which
we used was that
llx+yll < llxll+llyll for all x,yEH.
It is possible to state and prove the extension theorem in terms of any
subadditive bounding functional. This gives
Theorem 8.6A (Hahn-Banach extension). Suppose K is a linear
subspace of a linear space H, p is a subadditive functional on H such
that p(ax) = ap(x) for a >, 0, xEH; and f is a linear functional on K such
that f (x) < p(x) for all x E K. Then there is a linear functional f : H - R
such that
f (x) = f (x) for x E K, j (x) < p(x) for x E H.
214 LINEAR FUNCTIONALS [8.4
Exercises 8.4
1. If (0, .F, u) is such that there are two sets El, E2 .. with ,u(E1),
u(E2) positive and finite, show by considering the indicator functions of
El, E. that does not satisfy (8.4.1) if p + 2; and therefore is not a
Hilbert space.
2. Prove that m, the set of bounded real sequences {xi}, is a Banach space
with IIxII = sup I xi1
i
3. Suppose K is a Banach space, K* is its dual, and K** is the dual of
K*. Prove:
(i) if x is a fixed element in K, X (f) = f (x) for f E K* defines a linear
functional on K*;
(ii) for the above function IIXII = 11xii, so that T(x) = X is a norm
preserving map from K to K**;
(iii) this map T preserves the linear structure;
(iv) The set of elements X of K** such that X (f) =&) for some x K
forms a closed linear subspace of K**.
Ilf1I =
Let 1/p+ 1/q = 1 (if p = 1, q = co). Then
(i) for each
F(f) = ffdu
defines a linear funtional on Yp;
(ii) given any bounded linear functional F on 27 , there is a g c .T.
such that
F(f) = ffgdiu,
1/q
and in this case IIFII = (flglQdp} if p > 1,
= esssup IgI if P=1-
Proof. (i) This follows immediately from theorem 7.7 and the linear-
ity properties of the integral; for
{flfvd4u)"PJigIQdu}11q
IF(f)I <
216 LINEAR FUNCTIONALS [8.5
(ii) Suppose first that ,u(S2) < oo and F is a linear functional on
.T,. For any measurable E c S2 put
cr(E) = F(XE),
where xE is the indicator function of E. The linearity of F implies
immediately that 0' is finitely additive. Now suppose E _ (J Ei,
i=1
N
Ei disjoint. Then ,u( U E1) ->,u(E) as N oo, so that in -rp,
\\\i=1 1
N
where QN = U Ei.
II xQr, - XE II -->0,
i=1
Since F is continuous, we must have
00 N
E cr(Ei) = lim E o (Ei) = tr(E)
i=1 N-00 i=1
so that o- is completely additive on -5F. Further 1u(E) = 0 - v(E) = 0,
so that a- is absolutely continuous with respect top. By theorem 6.7
it now follows that there exists a function g which is integrable, such
that
F(,XE) = v(E) = I gdti for all E E z,
r
f
= yjpgdlc.
This gives the required representation for F on the class of indicator
functions of measurable sets. We must prove that g E .q, and that the
representation is valid on the whole of p.
It is clear from linearity that the representation is valid for F -
simple functions. If fo E Yp, fo > 0, we can find a sequence fn of
simple functions which increases monotonely to fo. Then by theorem
7.6, f n - f 0 in pth mean, and by the continuity of F
f
F(ffo) = limF(fn) = lim fngdu = f fogdu
n->00
and 11 F11.
(flgdIt)1/4S
so that, by theorem 5.5 IIFII. (8.5.1)
We can apply the above argument to each of the spaces (Qn, Win, a)
where ffln = F n Qn. By the uniqueness of the derivative g in the
Radon-Nikodym theorem, if f e and vanishes outside
N
RN
zUUQi' F(f) = ffgdi&
But if f > 0, we can apply the monotone convergence theorem to each
of {fng+}, {fn g_} where fn = AR to obtain this representation by
using the continuity of F. The final step is to use f = f+ -f- so that
the representation is valid on all of gyp. Further (8.5.1) follows since
it is true for the integral over each R.
Now by Holder's inequality (theorem 7.7) we have
Exercises 8.5
1. Suppose 1 < p S + oo, 11p+ 11q = 1 and f,, ->f in .9 norm, g g
in £a norm. Deduce that
ff- g. du - JfgdU.
2. If f2 is the set of positive integers and u is counting measure, then
2D (1 < p < oo) reduces to the set of sequences {xi} of real numbers such
that I xi I9 < oo; 2,,, reduces to the set m of bounded sequences.
i=1
3. Let X = [-1, 1], ,u Lebesgue measure. Show that the collection '
of continuous functions f : X -* R is a closed linear subspace of Y., (pro-
8.5] THE SPACE CONJUGATE TO Y 219
vided any function f which is equal a.e. to a continuous function is identified
with it). Hence, by theorem 8.6, extend the bounded linear functional
F(f) = f (O) from 9 to Y. without changing its norm. If possible, suppose
there is an f0 which is integrable and such that
fff0d - 0.
4. Extend example (3) to show that if S2 contains a disjoint sequence
of measurable sets of finite positive measure, then 21(Q,,u) is a proper
subspace of . . Deduce that ll is not reflexive.
= Ilk-yIl2-27lF(y)+y2IIFII2
= Ilk-yII2-V2 < e,
1 E 1IIf(Tix)-fk(Tix)II EfkT1(x)
+ 1n i=o
n i=o
IIf-All +nllhkll;
so that we can make IIgnII < e by first choosing fk with If - fkll < je and
then making n large.
The class of f E -T2 which satisfy either (a) or (b) is clearly a closed
linear subspace K of 22. By the corollary to theorem 8.8, any f E '2
can be written uniquely as
f=/1+f2 where f1 E K,
and (f2,Tf-f) = 0 for all fE22.
Now 0 = (f2,Tf-f) = (f2,Tf)-(f2,f)
(T-1f2,f) - (f2,f) = (T-1f2 -f2,f) for all f E Y2,
and in particular when f is the indicator function of a measurable set E
of finite measure. Hence T--1f2 = f2 a.e. so that f2 is invariant under T.
Hence l n-1
- Z f2(Tix) = f2(x) a.e. for all n,
ni=o
8-2
222 LINEAR FUNCTIONALS 18.6
so that f * = f2 is the limit in second mean of {gn}. Thus (i) and (ii) are
proved. To prove (iii), suppose g is invariant under T; then
(T if, g) = (.f, T -1g) = (f, g)
so that (gn, g) = (f, g) for each n and the result follows on letting
n --* oo since the inner product is continuous in the norm topology.
Corollary. Under the conditions of theorem 8.9, if T is ergodic, then
the limit (in second mean) f * = c a.e. Also
(i) if ,u(S2) = oo, then c = 0,
(ii) if #(Q) < oo, then f f*du = f fdu.
Proof. The only invariant functions are constants so (ii) of the
theorem implies that f * = c a.e. Now if µ(S2) = oo, we have 11 f *1I finite,
so c = 0. If µ(S2) < oo, then the function g(x) = 1 is in 2'2 and is in-
variant so that
(1,f*) = ff*diu = (1,f) = f du. f
Exercises 8.6
STRUCTURE OF MEASURES IN
SPECIAL SPACES
In the present book most of the theory of measure and integration
has been developed in abstract spaces, and we have used the properties
of special spaces only to illustrate the general theory. The present
chapter, apart from § 9.4, is devoted to a discussion of properties which
depend essentially on the structure of the space.
The first question considered is that of point-wise differentiation.
In the Radon-Nikodym theorem 6.7 we defined the derivative du/dv
of one measure with respect to another for suitable measures ,u, v:
but the point function du/dv obtained is only determined in the sense
that the equivalence class of functions equal almost everywhere is
uniquely defined. This means that at no single point (except for those
points which form sets of positive measure) is the derivative defined by
the Radon-Nikodym theorem. In order to define du/dv at a point x,
the local topological structure of the space near x has to be considered.
It is possible to develop this local differentiation theory in fairly
general spaces, but only at the cost of complicated and rather un-
natural additional conditions: we have decided instead to give the
detailed theory only in the space R of real numbers where the term
derivative has a clear elementary meaning.
There are several ways of defining an integral with properties similar
to those obtained in Chapter 5. So far in this book we have con-
sidered definitions which start from a given measure defined on
a suitable class of sets. In § 9.4 we describe the Daniell integral and
show that, under suitable conditions this can be obtained in terms of
a measure. Then, for locally compact spaces, we discuss positive
linear functionals on the space Cg of real-valued continuous functions
which vanish outside a compact set, and show that these also corre-
spond to integrals with respect to a suitable measure.
The final section of the chapter is devoted to the definition of Haar
measure in topological spaces which have the algebraic structure
of a group and in which the group operation is continuous. The
details are given only for locally compact metric groups.
224 MEASURES IN SPECIAL SPACES [9.1
Vitali covering
For a subset E c R, a class f of intervals is said to cover E in the
Vitali sense if, given x E E, e > 0 there is an interval J E / with
xEJandO < IJI < e.
Theorem 9.1. Suppose E c R has finite Lebesgue outer measure and
is covered in the Vitali sense by a class / of intervals. Then there is a
countable disjoint subclass f1 e / such that
IE-U{J:JE /1}I = 0.
Proof. We use JAI to denote the Lebesgue outer measure of A
whether or not A is measurable. There is no harm in assuming that
all the intervals J in / are closed since III = III for any interval I.
We may further assume without loss of generality that there is an
open set 0 D E with I G I < oo, and that all the intervals of f are
contained in G.
We choose f1 by induction as follows. Let J1 be any interval of /.
Suppose we have already chosen disjoint intervals J1,J2, ...,Jm and let
sm be the supremum of the lengths of the intervals in If which do not
intersect any of J1,J2, ..., Jm. Now sm < IGI < oo, and if E is not con-
m
tained in U Ji, we must have sm > 0. Thus if E is not already covered,
i=1
m
we can choose Jm+l disjoint from U Ji with IJm+1I > lism. Now the
i=1
M
theorem is proved if E c U Ji for any finite m. Otherwise we obtain
i=1
a sequence {J,,} of disjoint sets so that
00
Theorem 9.2. Suppose f:I --> R is monotone increasing. Then the set
E of points x in I for which f is differentiable at x satisfies II - El = 0.
The derivative f' is Lebesgue measurable, and if [a, b] c I,
where we re-define f(x) = f(b) for x >, b, then gn(x) is defined and
measurable and gn(x) --> g(x) for almost all x in [a, b] as n -* oo so that
g: I - R* is Lebesgue measurable if we define it arbitrarily to be zero
22$ MEASURES IN SPECIAL SPACES [9.1
on the exceptional set where Df(x) is not defined. By Fatou (theorem
5.7)
b g(x) dx 5 lim inf f b gn(x) dx
fa n-oo Ja
Jim inf n fa( f (x + n) - f (x) } dx
rb+(1/n) a+(1/n)
= liminf[nJb f(x)dx-n fa f(x)dx]
5 f(b) -f(a). J
This shows that the function g is integrable and so finite almost
everywhere. Thus f is differentiable a.e. in [a, b]. Since [a, b] is an
arbitrary subinterval of I, f is differentiable almost everywhere in I.
Functions of bounded variation
A function f: I R is said to be of bounded variation on I if
n
I f (xi) -f (xi-1) I
i=1
is bounded above for all ordered finite sequences xu < x1 < ... < xn
in I. Clearly if f: is of bounded variation on I, it is also of
bounded variation on each interval J c I. For an ordered sequence
a = {xi}, i = 0, 1, ..., n put
n
p(a) _ max [O,f(xi) -f(xi-1)],
i=1
n
n(a) min [O, f(xi) -.f(xi-1)],
i=1
n
t(a) = p(a) + n(a) _ I f (xi) -.f (xi-1) I
i=1
where each of the suprema is taken over all ordered finite sequences
a in [a, b]. It is easy to check that, in this case
Ta = Pa + Na, f(b) -f(a) = Pa - Na.
Now if f: [a, b] R is of bounded variation on [a, b] we can put
g(x) = Na, h(x) = PQ for all x e [a, b]
so that f (x) can be expressed as the difference of two non-decreasing
functions of bounded variation.
9.11 DIFFERENTIATING A FUNCTION 229
Corollary (Lebesgue). A function f: I -a R which is of bounded
variation on each finite interval [a, b] c I must be differentiable at x
for almost all x in I.
Proof. In each finite [a, b] we can express f as the difference of
two monotone increasing functions g and h. Each of these is differen-
tiable almost everywhere in [a, b] by theorem 9.2. Hence the difference
f is differentiable almost everywhere in [a, b]. I
Exercises 9.1
1. Show that, if g: I ->- R, h: I - R are each monotone increasing, then
f = g - h is of bounded variation on each [a, b] c I.
2. If f: I -* R is of bounded variation on each [a, b] c I, show that the
limits f(x + 0), f(x-0) exist at each interior point of I.
3. If c is an interior point of I and f : I -* R has a (local) maximum at c,
show that D+f (c) < 0, D_ f (c) > 0.
4. If f: [a, b] --> R is continuous and D+f(x) > 0 for all x in [a, b), show
that f (b) >, f (a).
5. Define f (o) = 0,
f(x) = x2 sin x 2 for x + 0;
g(0) = 0,
g(x) = x2 sin x-1 for x + 0.
Which of the functions f, g is of bounded variation on [-1, 1]?
6. Give an example of a function for which all the four derivates are
different at x = 0.
7. For any Lebesgue measurable f : I--> R, prove that D+f(x) is Lebesgue
measurable.
8. Show that theorem 9.1 as stated in R is false in R" for n > 2.
Hint. Take a Vitali covering of [0, 1] and for each J of covering consider
J x [0, 1] and J x [ 3,1 J. This will give a covering in the sense of our definition
of the unit square [0, 1] x [0, 1]. Show theorem 9.1 is not satisfied.
(In fact a more complicated construction shows that theorem 9.1 fails
even if we require each point of the set to be covered by an interval J of
arbitrarily small diameter.)
9. Show that theorem 9.1 is true in R" for all n if we restrict the covering
to cubes. (In fact it can be shown that it is true if there is a constant K
such that the ratio of the lengths of longest and shortest sides is bounded
for the intervals in f)
10. For the Cantor ternary function g: [0, 1] -> [0, 1] show that g'(x) = 0
for all x e [0, 1] - C.
230 MEASURES IN SPECIAL SPACES [9.1
(This shows that we cannot hope, in general, for equality in
ff'(x) dx
b
F(x) = 1:1(t)
then F: [a, b] --> R is differentiable in (a, b) with F'(x) = f(x). The
object of this section is to obtain the analogous theorem for the
Lebesgue integral, where it is not appropriate to assume that f is
continuous. (Of course, if f : [a, b] -* R is continuous on [a, b], we
know that F(x) = f(x) for all x in (a, b) since the Lebesgue integral
coincides with the Riemann integral in this case.) The first thing to
note is that, even for a monotonic function F, we cannot claim that,
in general, b
JF'(x) dx = F(b) - F(a), (9.2.1)
a
see exercise 9.1 (10). We will, however, obtain necessary and sufficient
conditions for the truth of (9.2.1).
Lemma. If f : [a, b] ->. R* is Lebesgue integrable on [a, b] and
Then F is differentiable with F'(x) = f (x) for almost all x in [a, b].
Proof. Assume first that f is bounded on [a, b] so that for a suitable
M in R, If(x) I < M, for all x in [a, b]. Now we know that F is absolutely
continuous and therefore differentiable almost everywhere. Put
n[F(x+n\\l
fn(x) = -F(x)J.
Then I f.I < M and fn(x) - F'(x) almost everywhere; so, by theorem
5.8 fora<c<b,
fF'(x)dx = 1imfa f(x)dx = lmn f ) -F(x)]dx
+(1/ n) a+(1/n)
= lim [nf F(x) dx - n f F(x) dx
a
f c {F'(x) - f(x)} dx = 0
and {F'(x)-f(x)}dx = 0.
Ja
This with (9.2.3) implies that F'(x) = f(x) a.e.
Lemma. If F: [a, b] -> R is absolutely continuous on [a, b] and
F'(x) = 0 a.e.,
then F is constant.
Proof. Suppose a < c < b, and E _ {x E [a, c]; F'(x) = 0}. For a
fixed e > 0, there are arbitrarily small intervals [x, x + h] for each
xEE such that IF(x+h)-F(x)I < eh.
Choose 8 > 0 to satisfy (9.2.2) in the definition of absolute continuity
and use theorem 9.1 to obtain a finite collection [xk, yk] of intervals
with
IF(yk)-F(xk)I < e(yk-xk)
which cover all of E except for a subset of measure less than 8. Order
these intervals so that
yo = a < x1 < y1 -< x2 < ... < yn -< C = xn+i,
n
and I xi+1- yi l < 8.
i=0
n
By (9.2.2) this implies Z I F(xi+i) - F(yi) I < e
i=o
234 MEASURES IN SPECIAL SPACES [9.2
and, from the choice of the covering family
n
F(yi)-F(xi)I < e(c-a)
i=o
so that
n n
IF(c) - F(a) = {F(x2+1) - F(yz)} + {F(yz) - F(xi)}
2=o a=o
< e(c-a+1).
Since e is arbitrary, we have F(c) = F(a). ]
Theorem 9.4. A function F: I R is an indefinite integral, that is
there is a measurable f : I --> R* such that
F(b)-F(a) =J bf(x)dx
a
for all [a, b] c I, if and only if F is absolutely continuous on I.
Proof. We have already seen that any indefinite integral is abso-
lutely continuous. Conversely suppose F: I -+ R is absolutely con-
tinuous. Then F is differentiable almost everywhere in [a, b] and
IF'(x) I 5 Fi(x)+F2(x) a.e.,
where F = Fi - F2 expresses F as the difference of two monotone
functions. By theorem 9.2, F' in integrable on [a, b]. Put
G(x) = faF'(t) d t.
F(x) = f
Jaa
Corollary. Every absolutely continuous function F:I --> R is the
indefinite integral of its derivative.
Density
Given a set A C R, X E R consider the ratio
IInAj
III
for all intervals I containing x where JEJ denotes the Lebesgue outer
measure of E. If this ratio converges to a limit as III -> 0, then
9.21 DIFFERENTIATING THE INTEGRAL 235
this limit is called the density of A at x and denoted ?-(x, A). The
point x is called a point of density for A if T(x, A) = 1, and a point
of dispersion for A if T(x, A) = 0. We can obtain the following as a
corollary of theorem 9.4.
Lemma (Lebesgue). If A - R, A is Lebesgue measurable, then
T(x, A) = 1 for almost all x E A,
T(x, A) = 0 for almost all x E R - A.
Proof. Suppose a < x < b. Then the indicator function yd is
Lebesgue integrable over [a, b]. Hence
F(x) = f xx.dx
a
is differentiable almost everywhere and
F'(x) = 1 for almost all x in [a, b] n A,
F'(x) = 0 for almost all x in [a, b] n (R - A).
But if x is such that F'(x) = 1, there is for each e > 0 a E > 0 such that
(i)
1>I[x,x]nAl >1-e for 0<h<4,
Exercises 9.2
1. If F: I R is absolutely continuous, show that FD is absolutely
continuous for each p > 1, but not, in general, for p < 1.
2. If F: [a, b] ->. R is such that F' exists everywhere in (a, b) and is
bounded show that rb
F'(x) dx = F(b) - F(a).
For F(x) = x2 sin l/x2 (x + 0), F(0) = 0 show that F'(x) exists for all
x but is not Lebesgue integrable over [-1, 1]. (This shows that even the
Lebesgue integral is not strong enough to integrate all derivatives.)
3. Construct a subset A c R for which T(0, A) = J.
9 TIT
236 MEASURES IN SPECIAL SPACES [9.2
4. Extend the density result to non-measurable sets A by showing that
for any A c R, T(x, A) = 1 for all x in A except a subset of zero measure.
Hint. Assume A is contained in a finite interval, and take a measurable
set B A with JBI = CAI.
Deduce that a set A c R is measurable if and only if r(x, A) = 0 for
almost all x in (R-A).
5. Prove that the Cantor function g: [0, 1] -* [0, 1] defined in §2.7 is
monotone increasing and continuous but not absolutely continuous.
6. The function f: [0, 1] --> R is absolutely continuous on [e, 1] for each
e > 0. Can one deduce that f is absolutely continuous on [0,1]? Does the
additional condition that f is of bounded variation on [0, 1] help?
forE.
#' = f E dm .
9.31 POINT-WISE DIFFERENTIATION 237
T(E) = dx
fE
is not always zero and T < ,aF. by theorem 9.2 so ,aF is not singular
with respect to Lebesgue measure. Conversely, if µF is not singular a
9.3] POINT-WISE DIFFERENTIATION 239
non-null absolutely continuous measure T 5 µF can be found, and this
corresponds to a function G, that is
G(b) - G(a) = b G'(x) dx.
Ja
But F(x) > G'(x) when both are defined, so F'(x) > 0 on a set of
positive measure. )
Theorem 9.5 (Lebesgue). Given any function F: R --> R which is
monotone increasing and continuous on the right, there is a decomposition
ofF F=F1+F2+F3
where F1 is a jump function,
F2 is singular,
F. is absolutely continuous.
This decomposition is unique if we insist that F1(0) = F2(0) = 0.
Proof. Use the function F to define a Lebesgue-Stieltjes measure
,up on -4. Decompose µF with respect to Lebesgue measure m by
theorem 6.7 so that
,aF = vl + V3
with v3 < m and v1 singular with respect to m. Decompose v1 by
lemma 2,
v1=A1+A2,
where Al is discrete and A2 is non-atomic.
Let F1, F2 be the monotone functions (with F1(0) = F2(0) = 0)
obtained by theorem 4.8 for which Al = PF,' A2 = UF2 on A Then by
lemmas 3 and 5, F1 is a jump function, and F2 is a singular function.
If one applies theorem 4.8 to v3 one obtains an absolutely continuous
G3 for which v3 = µa,. Finally, put F3(x) = G3(x)+F(0) and we still
have F. absolutely continuous, and v3 = µF8. Now
F(x) - F(0) = Fi(x) - F1(0) + F2(x) - F2(0) + F3(x) - F3(0)
for all x so that F(x) = F1(x)+F2(x)+F3(x).
The uniqueness follows from the uniqueness of the decomposition
µF = Al + A2 + v3, and theorem 4.8.1
In R we can also use the connexion between µF and F to define
differentiation. Thus if F:I -+ R is differentiable at x0, this means
that F(xo + h) - F(xo - k)
-)- F'(xo) as h,k -> 0
h+k
240 MEASURES IN SPECIAL SPACES [9.3
with h > O, k > 0, and
17L - - F(xo)
pF(xo-k,xa+h]
I(xo- k ,xo+ ]1
This can be written
#F(J) -a F'(xo) as IJI-->0
A
for intervals J containing x0, and we can write d aF/dm (xo) for the value
of this limit. More generally, if p, v are two measures in R which are
finite for bounded sets then
lim L/t(J)
I.n-- o L v(J)J
xEJ
when it exists, is called the derivative of p with respect to v at the
point x.
In Rn we can consider the values of the ratio
,a(J)
(9.3.2)
V(J)
for rectangles J (in 911) containing a fixed point x and ask whether or
not this ratio approaches a limit as diam (J) 0. The existence of
this limit for all x except for a set of zero v measure can be proved
when v is Lebesgue measure: the limit in this case is called the strong
derivate of p at x. This result is harder to prove than the result in
§ 9.1 because theorem 9.1 is not valid without some restriction on the
ratio of the sides of the covering class /. Essentially similar methods
to those of § 9.1 will work if only cubes J are considered. On the other
hand if in (9.3.2) one considers rectangles with arbitrary orientation
an example can be given for which the limit exists nowhere.
Differentiation point-wise in abstract spaces can also be defined
in terms of suitable `nets', and the theorems of this chapter can be
obtained if sufficient conditions are imposed. Since the results are
not often used in practice, we will not state them in detail.
Exercises 9.3
1. Enumerate the rationals as a sequence {ri}. By considering the discrete
measure with mass 1/i2 at ri (i = 1, 2,...) define a jump function which is
constant in no interval.
2. Give an example of a singular function which is constant in no interval.
3. If F, G are two monotone real functions differentiable at xo with
G'(xo) + 0, show that
d#F
(xo ) = lim -F exists and e quals , ()
duG xoEJ #G(J) G (x0)
IJI-+o
9.4] THE DANIELL INTEGRAL 241
for each g E L satisfying g(x) 5 lim fn(x) for all x E X. (Note that
n-co
lim fn(x) will be +oo if the sequence {fn(x)} is unbounded, and even
n-*00
if lim fn exists as a function with finite values we do not assume
that it is in L.)
In particular, this implies that, if f is a Daniell functional, {fn}
a monotone sequence in L such that f (x) = lim fn(x), x E X defines
n- C0
242 MEASURES IN SPECIAL SPACES [9.4
a function in L then 5(f) = lim5(fn). For if {fz} is increasing then
n--*oo
f > fn for all n, so 5(f) > -f(fn) since .1 is positive, which with (9.4.1)
gives the required equality. Thus a Daniell functional is continuous in
the sense that for any sequence { fn} in L which decreases monotonically
to the zero function we must have . .f (f.) -). 0. Any Daniell functional
is therefore an `integral' in the sense discussed in § 5.1. However,
for the integral to be useful we want the domain L to be as large as
possible: if {fn} is an increasing sequence in L which is bounded above
by an element of L we would certainly want lim fn to be in L. The
Daniell integral is essentially the result of extending a Daniell func-
tional.f from L to a class Ll L: it turns out that this extension can
be carried out in two stages.
Suppose f is a Daniell functional on a vector lattice L. Denote
by L+ the set of functions f: X -> R* which are limits of monotone
increasing functions of L. L+ is not a linear space but
a,f > 0 f,gEL+= of+/3gEL+.
Then if {fn} is an increasing sequence in L, {,f (f.)} is an increasing
sequence in R which has a unique limit in R u { + oo}. We can define
5 in L+ by
_f(lim fn) = lim5(fn).
This definition is proper because if {fn},{gn} are two monotone
sequences each converging to h in L+, condition (9.4.1) gives, for
fixed n,
fn S h = lim gn 5(fn) 4 "M -f (9n)
so that lim5(fn) S lim.f (gn) and the opposite inequality can be
similarly obtained. It is clear that f is linear on L+ in the sense that
a > 0, f >, 0; f,gEL+=.f(af+fg) = a.f(f)+fS(g)
For an arbitrary function f : X -> R* we define the upper integral
.f *(f) by .'*(f) = inf of (g),
8>f
DEL+
where we adopt the (usual) convention that the infimum of the empty
set is + oo. Similarly, the lower integral 5* (f) is defined by
5*(f) = -.f *(-f),
and we say that a function f: X -a R* is integrable (with respect to 5)
if 5*(f) _ 5* (f) and is finite. The class of integrable functions will
be denoted by Ll = L1(5, L). For f E Ll we call the common value of
5*(f ), 5* (f) the integral off and denote it by /(f ). We now show that
9.4] THE DANIELL INTEGRAL 243
this functional/ on L1 is a Daniell functional which extends 5, and
that L1 has the closure properties desired. It is convenient to obtain
a number of preliminary results before stating the theorem.
Lemma 1. If {gn} is a sequence of non-negative functions in L+, then
00 OD
5(f) = E 5(fn)
n=1
= E (E (fn,v))
n=1 v=1
Lemma 2. For arbitrary functions f: X --> R*, g: X --> R*:
(i) 5*(f+g)
(ii) if c % 0, 5*(cf) = c5*(f);
(iii) if f 5 g, 5*(f) < 5*(g), J*(f) 5-f*(g);
(iv) 5*(f) 5 J*(f);
(v) if fEL+, 5*(f) =5*(f) _ 5(f).
Proof. (i), (ii) and (iii) follow immediately from the definitions.
It is worth noting in (i), that we can put (f + g) (x) = + oo at those
points x for which one of f (x) is +oo and the other is - co so that (i)
is true whatever the value in R* chosen for (f + g) (x) at such points x.
(iv) Since 0 = 5(0) = 5(f -f) < 5*(f) +5*(- f) by (i), it follows
that. *(f) _ -5*(-f) < 5*(f)
(v) If f E L+, then by definition ./*(f) = .1(f ). Now if g E L, then
244 MEASURES IN SPECIAL SPACES [9.4
- g c L c L+ so that -0'* (g) = .fi(g). But each f in L+ is the limit of an
increasing sequence {gn} in L. Thus f > gn so J*(f) 3 5*(gn) = 5(gn)
andJ* (f) >, lim.f(gn) = 5(f).]
Lemma 3. If {gn} is a sequence of functions on X to R+, and
g= n=1
Egn, then .O*(g) < E.-.O*(gn)
n=1
Proof. If5* (gn) = +oo for some n, or if the series I.f*(gn) diverges
there is nothing to prove. Otherwise, given e > 0, for each integer
n choose hn > gn, hn E L+ such that .f*(gn) > 5(hn) - e 2-n. Then
h= E hn E L+ by lemma 1, h >,g and
00
c%0-5*(cf)=c.f *(f)=c.f*(f)=5*(cf),
c<0 5*(cf) = c-f*(f) = c5*(f) _ .f*(cf).
Further, if f and g are both in L1, using lemma 2 (i),t
-5*(f+g) =5*(-f-g) /(f)-/(g)
so /(f)+/(g) % *(f+g);
and, by lemma 2 (iv), f + g E L1 and
/(f+g) = /(f)+/(g)
Thus L1 is a real linear space, and f is a linear functional on L1.
To prove that L1 is a lattice it is sufficient to prove that
fEL1= f+eL1.
t As pointed out in the proof of lemma 2(i) the inequality is valid, whatever value
in R* is chosen for (f+g) (x) at points x where f(x) = + oo, g(x) = - oo. The proof
given then shows that, for f, g e L1, (f+ g) E L1 whatever values are assumed at such
points.
9.41 THE DANIELL INTEGRAL 245
For a fixed f in L1 and each c > 0, choose functions g, h in L+ such that
-h<f<g and
5(g) < /(f)+e < oo, f (h) < -/(f)+e < oo.
Now g = (g v 0) + (g A 0) and g A 0 E L+; so .5(gv 0) < .fi(g) - 5(g A 0) < oo.
Thus g+ = g v 0 E L+ and .f(g+) < oo. Similarly, - h_ = h A 0 E L+ and
h- < f+ < 9+. But (g+h) > 0; and separate consideration of each
possible pair of signs for g, h shows that g+- h_ < g + h. Hence
.f (9+) +.f (- h_) < .fi(g) +.f (h) < 2e.
00
Exercises 9.4
1. Show that the condition (9.4.1) for a positive linear functional is
equivalent to saying that, if {un} is a sequence of non-negative functions in
L and 0 E L satisfies 0< E u,n, then -0(0) 5 E 5(un).
2. If (S2, ,u) is a a -finite measure space, L is the class of u-integrable
functions and 5(f) = f f d u, show that .0 is a Daniell functional on L.
3. Let J be the class of continuous functions on R to R which are zero
outside [ - K, K] for some K and put
f
F(f) = fdv+- ffdv_
F(f) = ff dv.
Ilfll IvI(X)
so that IIFII < I vI (X). Further
IvI(X) < 1t1(X)+,u2(X) = F+(1)+F-(1) = IIFII
so we have IIFII = I vl (X).
To prove that v is uniquely determined by F, suppose there are
two signed measures v1, v22 with
Exercises 9.5-
Show that 0 is a Radon measure, but that 0 does not correspond to any
signed Baire measure.
lim Y. = yo lim xn 1 = xo 1.
n- oo n-aco
UHi:D E,
i=1
Exercises 9.6
1. Suppose S2 is the set of positive real numbers with the usual metric
and multiplication for the group operation. If (1, e) is the reference set Ho
used in the definition of Haar measure It in f1, show that
,u(a, b) = log b/a for each interval (a, b) c Q.
(Here e is the base for Napierian logarithms.)
2. With X = R and addition for the group operation define Haar
measure ,u with (0,1) taken as the reference set Ho. Show that It is Lebesgue
measure in R.
3. Let X be the set of 2 x 2 matrices of the form
(0 x)
with x > 0 and multiplication for the group operation. Define a metric
in X by using the Euclidean metric in R2. Show that in the topology of
this metric, X is a locally compact metric group.
Define
F(x y=-y
0 X/ x'
INDEX OF NOTATION
A-B, 9 lp, 1w, 209
-4,-4-,43 2k, 79
R*, 103 F, 96
C0(X), 250 168
C(X), 251 Y2(S2, µ), 194
C, 2 M, .,f, 166
C, 46 M(ca), 199
C(a, b), 209 M, m, 46
C*(X), 48 p, gn, 15
c, 6,47t Q, 2
Wx-9, 134 R, Rn, 2
W*-9, 135 R*, 34, 51
((3; iEI), 136 R+, 51
D, D+, D_, D+, D-, 224 RI, 158
d(x, E), d(E, F), 27 S(x, r), 24
E, 26 S(x, r), 25
Ex, EY, 135 s, 46
ExF, 2 2, 82
e, 15 5P(M), 18
gn, 18 XI, 157
f-1, 4 {x1}, 5
fog, 4 {x; P},1
f: A-i-B, 3 Z, 2
fe(y), 136 a.e., 109
(f, g), 198 diam, 27
Via, 44 ess inf, ess sup, 167
F(A), 19 lim inf, lim sup, 12
19 11+, ,u_, 62
9a, 44 96
-or, 100 ,af-1, 154
K*, 211 per, 168
43 Pl, 174
L1, $1, 174 pp, 175
Lp, Yp, 175 (r--q, 77
12, 48 T(x, A), 235
t Note that the symbol c has two distinct uses, which should not be confused.
262 INDEX OF NOTATION
XE, 12 u, n, to
No,6 11.11,45
E, O, z , =>, 1 f , see chapter 5
o,2 <<, 148
H,3 V, A, 241
3,4 The symbol ] is used to signal
-,6 the end of a proof.
v, n,A,9
263
GENERAL INDEX
absolute continuity of functions, 231; of local, 31
measures, 120, 148, 236 complement, 9
integration, 128 complete measure, 81, 109, 166
additive set function, 51, 65, 214 metric space, 29, 175, 194
algebra of subsets, 15 set in a linear space, 199
algebraic numbers, 13 completion, 82
almost everywhere, 108 composition, 4, 154
uniform convergence, 168 conjugate indices, 183
approximation in measure, 84 space, 213, 215
to measurable functions, 131 consistency conditions, 158
area, 69, 79 continuous function, 35
under a curve, 146 set function, 56
atom, 64, 237 continuum hypothesis, 7, 58
axiom of choice, 7, 19, 93 convergence, 180-1
in mean, 174
Baire sets, 132, 250 in measure, 171
measure, 250 in norm, 204
Baire's category theorem, 42 in pth mean, 174
Banach space, 194, 209 of sets, 12
Bessel's inequality, 204 countable, 7
Birkhoff's theorem, 190 basis for measure, 195, 207
Borel field, 16 counting measure, 185
-measurable function, 107, 154 covering, 30, 225
sets, 43 cylinder set, 136, 140, 158
Borelian sets, 43
bounded, 27 Daniell extension, 242
convergence theorem, 126 functional, 241
linear functional, 210, 251 integral, 101, 223, 241, 246
variation, 228 -Kolmogorov theorem, 159
bounds, upper and lower, 21 decomposition, Hahn-Jordan, 61, 64
Brownian motion, 161 Lebesgue, 149, 239
de Morgan's laws, 10
Cantor set, function, 49, 110, 152, 229, dense, 42, 195
236, density, 234
Cantor's lemma, 41 derivate, 224
Caratheodory, 127 derivative, 224, 240
cardinal numbers, 5, 6 Radon-Nikodym, 152_223
Cartesian product, 2, 38, 134 derived set, 26
category, first and second, 41 diameter, 27
Cauchy integral, 127 differentiable, 224
-Riemann integral, 129 discrete distribution
sequence, 29, 167, 171 measure, 53, 80, 237
chain, 20 probability, 98
maximal, 21 topology, 28
change of variable, 155 disjoint, 11
class, 2 dissection, 102
closed linear span, 199 distance, 23, 27
set, 25 distribution function, 96-7
sphere, 26 domain, 3
closure, 26 dominated convergence theorem, 121,
coarser topology, 40 125-6, 180, 249
collection, 2 dual space, 21, 215
compactification, 33, 34
compactness, 29, 30, 39 Egoroff's theorem, 169
264 GENERAL INDEX
elementary event interior point, 28
figures, 17-18; length, area and volume intersection, 9, 10
of, 69 invariant function, 190
functions, 109 measure, 90, 255
empty set, 2 inverse function, 4
enumerable, 7 image, 4
equicontinuity, 177 invertible, 187
equivalence relation, 6, 166
ergodic theorems, maximal, 188; mean, Jacobian, 156
221; pointwise, 190 Jordan decomposition, 61, 64
transformation, 192, 222 jump function, ,237
essentially bounded, 167
extended real numbers, 33 Kolmogorov, 159
extension of functions, 4 Kuratowski's lemma, 21
of set functions, 58
theorems, 65-6, 77, 244 least upper bound axiom, 21, 30
Lebesgue convergence theorem, 121
Fatou's lemma, 120, 249 covering lemma, 38
field, 15 decomposition theorem, 149, 239
finite-dimensional distributions, inter- density theorem, 235
section property, 32 integral, 124
Fourier coefficients, 203 measurable, 108
series, 203 measure, 69, 79, 195
Fubini's theorem, 143-4 -Stieltjes integral, 125; measure, 95,
function, 3 198, 236
space, 157 Legendre polynomials, 202
length, 69, 79
generated ring, 17, 65 limit of a sequence, 27
o -ring, 18, 77 point, 26
z-class, 17 linear dependence, 199
Gram-Schmidt orthogonalisation, 201 functional, 201, 215, 241, 250
groups, 254 space, 45, 194
span, 199
Haar measure, 255, 257, 259 subspace, 212
Hahn-Banach theorem, 212-13 Lindelof space, 22
decomposition, 61, 64 Liouville's theorem, 188
Hausdorff space, 250 local compactness, 31, 250, 254
Heine-Borel theorem, 30
Hilbert cube, 48 majorised, aee dominated
space, 194, 202 mapping, 3, 153
Holder's inequality, 183, 186 marginal distribution, 163
maximal ergodic theorem, 188
indefinite integral, 127, 149, 230, 234 mean ergodic theorem, 221
indicator function, 12 measurable function, 103, 107, 166,
indiscrete topology, 28 246
inequalities, 183 set, 74, 79, 96, 246
inner measure, 75 space, 246
product, 198 transformation, 154
integrable function, 113-14,127, 129, 246 measure, 55
set, 246 Haar, 255, 257, 259
integral, 100 Lebesgue, 69, 79, 195
Cauchy, 127 Lebesgue-Stieltjes, 95, 198, 236
Cauchy-Riemann, 129 Radon, 254
Daniell, 101, 223, 241, 246 measure-preserving transformations, 187
Lebesgue, 124 metric, 23, 185
Lebesgue-Stieltjes, 125 group, 255
Riemann, 100, 128, 129, 176 outer measure, 86, 88, 257
integration by parts, 157 space, 23
GENERAL INDEX 265
metrisable, 25 regular measure, 86
Minkowski's inequality, 184, 186 outer measure, 75
monotone class, 16, 79 representation of linear functionals,
class theorem, 18 250
convergence theorem, 119 restriction, 4
sequence, 12 of a set function, 58, 75
set function, 60 Riemann integral, 100, 128, 129, 176
monotonic function, 8, 224, 226 Riesz-Fisher theorem, 205
mutually singular measures, 149 Riesz's lemma, 188
representation theorem, 250
neighbourhood, 26 ring, 15
non-atomic, 64, 238 (in algebraic sense), 18
non-measurable set, 93
norm, 45, 194, 211 scalar product, 198
normal numbers, 193 SchrSder-Bernstein theorem, 6
normed linear space, 44-5 Schwarz's inequality, 184
nowhere dense, 41, 49 section, 135-6
null set, 2 semi-ring, 15
separable, 48, 187, 194
open covering, 30 separating functional, 219
set, 24 sequence, 4
sphere, 24 monotone, 12, 18
ordered pairs, 2 set, 1
ordering, 20 function, 51
ordinate set, 145 shift, 193
orthogonal system, 199, 200 sigma additive (v-additive), 54
orthogonalisation, 201 algebra (v-algebra), 16
orthonormal system, 200 compact (v-compact), 43
outer measure, 59, 74 field (v-field), 16
metric, 86, 88, 257 finite (Q-finite), 59
ring (Q-ring), 16
parallelogram law, 209 simple function, 102
Parseval's identity, 204 singular, 149, 238
partial ordering, 20 statistical mechanics, 188
perfect set, 44, 49 Stieltjes measure, 95-6, 99, 163
phase space, 187 see also Lebesgue-Stieltjes
point of density, 235 Stone's theorem, 247
of dispersion, 235 strong derivative, 240
pointwise convergence, 166 subadditive, 59, 213
positive linear functional, 241, 250 support, 250
probability measure, 96, 98 supremum, 21, 29
space, 96
product field, 134 thick, 164
measure, 139, 141, 164 topological group, 254
ring, 134 space, 25
v-field, 134 transcendental numbers 14
space, 3, 5, 134 transformation, 3, 89, 154
z-class, 134 measure-preserving, 187
projection, 40, 135 transitive, 20
proper subset, 1 triangle inequality, 23
trigonometric functions, 202
Rademacher functions, 202 Tychonoff's theorem, 39
Radon measure, 254
-Nikodym derivative, 152, 223; theo- uniform absolute continuity, 178
rem, 149 continuity, 37
range of a function, 3 convergence, 167
rectangle, 134 union, 9, 10
reflexive, 20, 218 uniqueness of extension, 77
266 GENERAL INDEX
upper bound, 21 Weierstrass property, 31
well ordered, 2, 22
vector lattice, 241 Wiener measure, 162
Venn diagram, 9
Vitali covering, 225 z-class, 16, 134
volume, 69, 79 Zorn's lemma, 21
Von Neumann, 15