Professional Documents
Culture Documents
by
Edwin Hewitt
Professor of Mathematics
The University of Washington
and
Karl Stromberg
Associate Professor of Mathematics
The University of Oregon
With 8 Figures
This book is first of all designed as a text for the course usually called
"theory of functions of a real variable". This course is at present cus-
tomarily offered as a first or second year graduate course in United
States universities, although there are signs that this sort of analysis
will soon penetrate upper division undergraduate curricula. We have
included every topic that we think essential for the training of analysts,
and we have also gone down a number of interesting bypaths. We hope
too that the book will be useful as a reference for mature mathematicians
and other scientific workers. Hence we have presented very general and
complete versions of a number of important theorems and constructions.
Since these sophisticated versions may be difficult for the beginner, we
have given elementary avatars of all important theorems, with appro-
priate suggestions for skipping. We have given complete definitions, ex-
planations, and proofs throughout, so that the book should be usable
for individual study as well as for a course text.
Prerequisites for reading the book are the following. The reader is
assumed to know elementary analysis as the subject is set forth, for
example, in TOM M. ApOSTOL'S Mathematical Analysis [Addison-Wesley
Publ. Co., Reading, Mass., 1957], or WALTER RUDIN'S Principles of Mathe-
matical Analysis [2 nd Ed., McGraw-Hill Book Co., New York, 1964].
There are no other prerequisites for reading the book: we define practi-
cally everything else that we use. Some prior acquaintance with abstract
algebra may be helpful. The text A Survey of Modern Algebra, by GARRETT
BIRKHOFF and SAUNDERS MAC LANE [3 rd Ed., MacMillan Co., New York,
1965] contains far more than the reader of this book needs from the
field of algebra.
Modern analysis draws on at least five disciplines. First, to explore
measure theory, and even the structure of the real number system, one
must use powerful machinery from the abstract theory of sets. Second,
as hinted above, algebraic ideas and techniques are illuminating and
sometimes essential in studying problems in analysis. Third, set-theoretic
topology is needed in constructing and studying measures. Fourth, the
theory of topological linear spaces ["functional analysis"] can often be
applied to obtain fundamental results in analysis, with surprisingly little
effort. Finally, analysis really is analysis. We think that handling ine-
qualities, computing with actual functions, and obtaining actual num-
VI Preface
are deeply grateful to Mrs. SHANTI THAYIL, who has typed the entire
manuscript with real artistry.
Our thanks are also due to the UniversitieE. of Oregon and Washington
for exemption from other duties and for financial assistance in the pre-
paration of the manuscript. It is a pleasure to acknowledge the great
help given us by Springer-Verlag, in their rapid and meticulous publica-
tion of the work.
July 1965
CHAPTER ONE
Set Theory and Algebra
From the logician's point of view, mathematics is the theory of sets
and its consequence,>. For the analyst, sets and concepts immediately
definable from sets are essential tools, and manipulation of sets is an
operation he must carry out continually. Accordingly we begin with two
sections on sets and functions, containing few proofs, and intended largely
to fix notation and terminology and to form a review for the reader in
need of one. Sections 3 and 4, on the axiom of choice and infinite arith-
metic, are more serious: they contain detailed proofs and are recommended
for close study by readers unfamiliar with their contents.
Plainly one cannot study real- and complex-valued functions seriously
without knowing what the real and complex number fields are. Therefore,
in § 5, we give a short but complete construction of these objects. This
section may be read, recalled from previous work, or taken on faith.
This text is not rigorous in the sense of proceeding from the axioms
of set theory. We believe in sets, and we believe in the rational numbers.
Beyond that, we have tried to prove all we say.
0 A ... =
00
then .. 1 {O}.
§ 1. The algebra of sets 3
B C 8
as rings or algebras of sets [see (1.25)]. Let B be any set. Suppose that
to each a EB there is assigned a unique element a* EB and that to each
pair of elements a, b EB there is assigned a unique element a V b EB
such that these operations satisfy
(i) a V b = b Va,
(ii) a V (b V c) = (a V b) V c,
(iii) (a*V b*)*V (a*V b)*= a.
Sets B with operations V and * [or similar operations J and satisfying
axioms equivalent to (i) -(iii) were studied by many writers in the period
1890-1930. They bear the generic name Boolean algebras, after the
English mathematician GEORGE BOOLE [1815-1864]. The axioms
(i) -(iii) were given by the U. S. mathematician E. V. HUNTINGTON
[1874-1952J [Trans. Amer. Math. Soc. 5, 288-309 (1904)].
The reader will observe that if a, b are interpreted as sets and V
and * as union and complementation, then (i) -(iii) are simple identities.
Other operations can be defined in a Boolean algebra, e.g., /\ [the analogue
of n for setsJ, which is defined by a /\ b = (a* V b*)*. A great deal of
effort has been devoted to investigating abstract Boolean algebras. In the
1930's, the contemporary U. S. mathematician M. H. STONE showed that
any Boolean algebra can be interpreted as an algebra of sets in the follow-
ing very precise way [Trans. Amer. Math. Soc. 40, 37 -111 (1936)]. Given
any Boolean algebra B, there is a set X, an algebra ~ of subsets of X,
and a one-to-one mapping -r of B onto ~ such that -r(a*) = (-r(a))'
[* becomes 'J and -r(a V b) = -r(a) U -r(b) [V becomes U]. Thus from the
point of view of studying the operations in a Boolean algebra, one may
as well study only algebras of sets.
STONE'S treatment of the representation of Boolean algebras was
based on a slightly different entity, namely, a Boolean ring. A Boolean
ring is any ring 5 such that x 2 = x for each xES. [For the definition of
ring, see (5.3).J
STONE showed that Boolean algebras and Boolean rings having a mul-
tiplicative unit can be identified, and then based his treatment on
Boolean rings. More precisely: for every Boolean ring 5, there is a ring
of sets ~ and a one-to-one mapping -r of 5 onto ~ such that
-r(a + b) = -r(a) £:" -r(b)
and
-r(ab) = -r(a) n -r(b) .
That is, addition in a Boolean ring corresponds to the symmetric differ-
ence, and multiplication to intersection.
Proofs of the above results and a lengthy treatment of Boolean al-
gebras and rings and of algebras and rings of sets can be found in
G. BIRKHOFF, Lattice Theory [Amer. Math. Soc. Colloquium Publications,
Vol. XXV, 2nd edition; Amer. Math. Soc., New York, N. Y., 1948].
6 Chapter 1. Set theory and algebra
The equality (ii) is true because the following statements are pairwise
equivalent: x Edomh; (x, y) Eh for some y; (x, y) Ef for some I Etr;
x Edoml for some I Etr.
Let x Edomh n doml = doml where IE tr. Then (x, I (x)) EIe h
and h is single-valued so hex) = I(x). This proves (iii).
The equality (iv) follows from the previous conclusions and (2.15.i)
since rngh = h (domh) = h (U{dom/: IE tr}) = U {h (dom/): IE tr}
= U{/(dom/): I Etr} = U{rng/: I Etr}. 0
(2.20) Definition. Let X be any set and E any subset of X. The
function ~E with domain X and range contained in {O, I} such that
~E (x) =
{
°lifXEE'
if x EX n E' ,
a least upper bound, i.e., if x, y EX, then there exists z EX such that
x ~ z, y ~ z, and if x ~ w, y ~ w, then z ~ w.
(2.26) Exercise. Let 1be a function from X to Y. Suppose that there
is a function g from Y to X such that log (y) = y for all y E Y and
g 01 (x) = x for all x EX. Prove that I is a one-to-one function from X
onto Y and that g = 1-1.
the set {I, ... , n}, then we write A I = An. Some authors write
A oo for AN.
A typical member of An is, to be sure, a function and as such is a set
of n ordered pairs. We follow conventional notation, however, and list the
values of such a function as an ordered n-tuple. Thus An= {(a v ... , an)
: ak E A for k = 1, ... , n}. Similarly AN = {(a v a 2 , • • • ) : ak E A for
kEN}. The set Rn is called Euclidean n-space and Kn is called unitary
n-space.
°
(3.4) Example. Let A = {O, I}. Then AN is the set of all sequences
a = (a v a 2 , ••• , an' ... ) where each an is or 1. In many ways this set
resembles CANTOR'S ternary set P = [0, 1] n (] !, ! [ u ] ~, ~[
U] ~, : [ u .. .y [see (6.62) infra]. The mapping cp defined by
cp (a) = 2 t ;: is a one-to-one mapping from
n=!
AN onto P. Anticipating
future developments, we remark that AN can be made a metric space by
introducing the metric e, where e(a, b) =~ if tlt = bI , a2 = b2 ,
°
••• ,
To show that (iii) implies (iv), lei (P, ~) be any nonvoid partially
ordered set in which each chain has an upper bound. By (iii) there is a
maximal chain Me P. Let m be an upper bound for M. Then m is a
maximal element of P, for if there is an x E P such that m ~ x and
m =1= x, then M U {x} is a chain which properly includes M, contradicting
the maximality of M.
To prove that (iv) implies (v), let 5 be any nonvoid set and let :!l'
denote the family of all well-ordered sets (W, ~) such that We S. For
example, ({x}, {(x, x)}) E:!l' for each xES. We next introduce an ordering
on :!l' by defining (WI' ~I) ~ (W2' ~ 2) to mean that either WI = W 2 and
~I = ~2 or there exists a EW 2 such that WI = {x EW 2 : x ~2 a, x =1= a}
and ~2 agrees with ~I on WI> i.e., ~I C ~2' We say that (W2' ~2) is a
continuation 0/ (WI' ~I)' The reader should see without difficulty that
~ is a partial ordering on :!l'.
Let us show that ZORN'S lemma can be applied to the partially ordered
set (:!l', ~). Let Cfi = {(W" ~J}tEI be any nonvoid chain (relative to ~)
in :!l'. Set W = U W, and ~ = U ~ t [recall that each ~ t is a set of
tEl tEl
ordered pairs]. We leave it to the reader to prove that ~ is a linear
ordering on W. Let A be a nonvoid subset of W. There exists lEI such
that A n W, =1= 0. Since (w" ~ t) is a well-ordered set, there is an element
a EA n W, such that a ~ t X for each x EAnn;. Assume that there is an
element b EA such that b ~ a. Then b EW, and b ~ t a, so b = a. Thus A
has a smallest element a in (W, ~). We conclude that (W, ~) E:!l' and
is an upper bound for Cfi.
By ZORN'S lemma, :!l' has a maximal element (Wo' ~o). If Wo = 5,
then ~ 0 is a well-ordering for 5 and we are through. Assume that Wo =1= S.
Let zEsnwh. Define ~=~oU{(x,z):xEWoU{z}} on WoU{z},
i.e., we place z after everything in WOo Then (Wo U {z}, ~) E:!l'. This
contradicts the maximality of (Wo' ~o), and so we have proved that
Wo= S.
It remains only to show that (v) implies (i). Let {At}tEI be any non-
void family of nonvoid sets. Let 5 = U At. Let ~ be a well-ordering
tEl
for S. For each lEI, let / (l) be the smallest member of At relative to the
well-ordering ~. Then / is a choice function for the family {At}tEI' 0
It is frequently useful to make definitions or carry out constructions
by well ordering a certain set Wand making the definition or construction
at a EW depend upon what has been defined or done at all of the prede-
cessors of a in the well-ordering. The general form of this process is
described in (3.13) and (3.14) below.
(3.13) Definition. Let (W, ~) be a well-ordered set and let a EW.
The set I(a) = {x EW: x ~ a, x =1= a} is called the initial segment 0/ W
determined by a.
§ 3. The axiom of choice and some equivalents 17
= DC2 = ... = DC" = 0.1 A nonvoid linearly independent set B such that
DC1
B ~ E C X implies that E is not linearly independent is called a Hamel
basis [or merely basis] for X over F. Thus a Hamel basis is a maximal
linearly independent set.
(3.19) Theorem. Every vector space with at least two elements contains
a Hamel basis.
Proof. Let X be a vector space with at least two elements. Let x =l= 0
in X. Then (3.16) shows that {x} is a linearly independent set. Thus the
family:F of all linearly independent subsets of X is non void. The definition
of linear independence shows at once that:F is of finite character. TUKEY's
lemma proves that :F contains a maximal member, i. e., X contains a
basis. D
(3.20) Theorem. Let X be a vector space over a field F and let B be a
Hamel basis lor X over F. Then lor each x EX there exists a unique lunction
DC Irom B into F such that DC (b) = 0 except lor finitely many b E B and
x = L DC (b) b, i. e., x can be expressed in just one way as a finite linear
bEB
combination 01 members 01 B.
Proof. Let x EX. If x EB, define DC(X) = 1 and lX(b) = 0 for bE B,
b =l= x. Then LIX(b)b = Ix = x. Suppose x ~ B. Then B U {x} is not
bEB
linearly independent, so there is a finite set {x, Xl' x 2, ... , x,,} C B U {x}
and a finite sequence ((3, (31) ... , (3,,) cF not all 0, such that (3x + (31Xl
+ ... + (3" x" = O. Since B is independent, we see at once that (3 =l= O.
Therefore x = - (3-1(31Xl - ••• - (3-1(3"X". Now define IX(X;) = - (3-1(3;
(j = I, ... , n) and IX (b) = Oforb EB n {Xl' ... ' x,,}'. Then x = LIX(b)b.
This proves the existence statement. bE B
To prove uniqueness, suppose that LIXI(b)b = LIX2(b)b. Then
bEB bEB
L (IXI (b) - 1X2(b))b = 0, and this is a finite linear combination of elements
bEB
of B. By independence, IXI (b) - 1X2 (b) = 0 for each b EB and therefore
the two functions IXI and 1X2 are the same. D
(3.21) Exercise. Given a nonvoid set A and a field F, let ~ be the
subset of FA consisting ofthose functions I for which the set {a EA : I (a) =l= O}
is finite. Let the linear operations in ~ be as in (3.17.b). Prove that ~ is
a vector space over F. Prove that every vector space is isomorphic qua
vector space with some vector space ~. 2
(3.22) Exercise. Prove that if P is a set and ;;;;; is a partial ordering
on P, then there exists a linear ordering ;::;; 0 on P such that ;;;;; C ;;;;; o.
1Note that [25 is linearly independent.
S Let Xl and XI be linear spaces over F. An isomorphism T of Xl onto XI is a
one-to-one mapping of Xl onto Xs such that T(X + Y) = T(X) + T(y) and
T (ocx) = OCT (x) for all x, y E Xl and all oc EF. Isomorphic linear spaces cannot be
told apart by any linear space properties.
§ 4. Cardinal numbers and ordinal numbers 19
(3.23) Exercise. Let (L, ~) be a linearly ordered set. Prove that there
exists a set We L such that ~ well orders Wand such that for each
x EL there is ay EW for which x ~ y.
(3.24) Exercise. Let G be a group and let H be an Abelian subgroup
of G. Prove that there exists a maximal Abelian subgroup] of G such
that He]; i. e., ] is Abelian, but no subgroup J* such that] ]* is s;:
Abelian.
(3.25) Exercise. Prove that the following assertion is equivalent to
the axiom of choice: If A and Bare non void sets and I is a function from
A onto B, then there exists a function g from B into A such that
g (y) E1-1 (y) for each y EB.
(3.26) Exercise. Let X be a vector space over a field F. Let A be a
nonvoid linearly independent subset of X and let 5 be a subset of X
such that each element of X is a finite linear combination of elements
of S. [The set 5 is said to span X.] Suppose that A c S. Prove that X
has a Hamel basis B such that A C Be S.
one version of axiomatic set theory, the cardinal number of a set is taken
to be a very specific well-ordered set, viz., the smallest ordinal number
that is equivalent to the given set.
We next define an order relation for cardinal numbers.
(4.5) Definition. Let u and 1) be cardinal numbers and let U and V
be sets such that U = u and V = 1). We write u ~ 1) or 1) ~ U to mean
that U is equivalent to some subset of V. One sees by considering com-
posite functions that this definition is unambiguous. We write u < 1) or
1) > U to mean that u ~ 1) and u =1= 1).
Proof. Let U and V be sets such that U = u and V = 1). Let 5'
denote the family of all one-to-one functions I such that doml C U
and rngl C V. It is easily seen that 5' is a family of finite character so,
by TUKEY's lemma (3.8), 5' contains a maximal member h. We assert
that either domh = U or rngh = V. Assume that this is false. Then
there exist x E un (domh), andy E V n (rngh)'. But then h U {(x,y)}E5',
contradicting the maximality of h. Thus our assertion is true. If domh= U,
then h shows that u ;:::;; 1). If rngh = V, then h- 1 shows that 1) ;:::;; U. 0
(4.9) Theorem. The ordering;:::;; lor cardinal numbers makes any set
01 cardinal numbers a linearly ordered set.
Proof. This theorem is just a summary of Theorems (4.6), (4.7),
and (4.8). 0
Our next theorem shows that there is no largest cardinal number.
(4.10) Theorem [CANTOR]' Let U be any set. Then (J < [ljJ(U).
Proof. We suppose that U =1= 0, since [ljJ(0) = 1> 0 = '0. Let
U = (J and 1) = [ljJ(U). The function I defined on U by I (x) = {x} E [ljJ(U)
is one-to-one, so U ;:::;; 1). Assume that u = 1). Then there exists a [one-to-
oneJ function h such that domh = U and rngh = [ljJ(U). Define
5 = {x E U: x ~ h(x)} .
Since each positive integer is a power of 2 [possibly the ot" power] times
an odd integer, I is onto N. We see that I is one-to-one, for otherwise
there would be an integer which is both even and odd. 0
(4.19) Lemma. II A is any nonvoid countable set, then there exists a
mapping Irom N onto A.
Proof. Since A is countable, there exists a one-to-one mapping g
from A into N. Let a EA. Define I on N by
n _ { g-l (n) for n E rngg ,
I( ) - a for n ~ rngg. 0
(4.20) Lemma. II A and B are two nonvoid sets and il there is a
mapping Ilrom A onto B, then A ~ B.
Proof. Let g be a choice function for the family {I-I (b)hEB. Then g
is a one-to-one mapping from B into A. 0
(4.21) Theorem. The union 01 any countable lamily 01 countable sets
is a countable set, i. e., il {AihEI is a lamily 01 sets such that I is countable
and each Ai is countable, then A = U Ai is countable.
iEI
Proof. Let {AihEI be as in the theorem. We obviously may, and do,
suppose that I and each Ai are nonvoid. Apply Lemma (4.19) to obtain
mappings/iandgsuch that doml i = domg = N, rngg = I, and rngli= Ai
for all i E I. Now define h on Nx N by hem, n) = fg(m) (n). Then h
is onto A. It follows from (4.20) and (4.18) that
'" x"
x="",,-
1'1=1 2"
° °
where each x,.. is or 1 and x,.. = for infinitely many n EN: see (5.40).
Define g on [0, 1 [ into A by g (x) = q; where q; (n) = x,.. for each n EN.
Then g is a one-to-one mapping, so that C ~ 21to. Now apply the Schroder-
Bernstein theorem. D
We next point out a few curious arithmetical properties of infinite
cardinal numbers. First we need a lemma.
(4.28) Lemma. II D is any infinite set and F is any finite set such that
D n F = 0, then 15 = D U F.
Proof. Let F = {Yv Ys, ... , y,..} where Yi =1= Yi for i =1= j and let
C = {Xi:j EN} be a countably infinite subset of D where Xi =1= xi for
i =1= j (4.15). Define I from D onto D U F by
I(x) = !
Yi for x = Xi' 1 ~ j ~ n ,
xi_,.. for x = xi,j > n,
x for xED n c/.
Then I is one-to-one. D
(4.29) Theorem. Let a be any infinite cardinal number. Then a + a = a.
Proof. Let A be any set such that A = a. Let B = A x {O, I}.
Then B = {(a, 0): a E A} U {(a, 1): a E A} so, by Definition (4.23),
we have 11 = a + a. Let 5' denote the set of all one-to-one functions I
such that doml C A and rngl = (dom/) x {O, I}. Since A is infinite,
there exists a count ably infinite set C such that C C A (4.15). In view
of (4.21), we see that Cx {O, I} is also count ably infinite. Hence there
26 Chapter 1. Set theory and algebra
°
(4.32) Corollary. II a is an infinite cardinal number and b is a cardinal
number such that < b ~ a, then a b = a.
Proof. We have a ~ ab ~ aa = a. 0
(4.33) Exercise. Prove that our ordering and our arithmetical opera-
tions for cardinal numbers agree on the set N with the usual ordering
and arithmetical operations for positive integers.
§ 4. Cardinal numbers and ordinal numbers 27
and so by (ii)
aa- 1 = e.
Note that the last equality also implies that a is a left inverse of a-I.
(b) For any e as in (ii) and any a E G, we have
ae = a(a- 1 a) = (aa- 1 ) a = ea = a,
i. e., e is also a right identity. If e1 and e2 satisfy (ii), then they are both
right identities also, and so
e1 e2 = e2 [e 1 is a left identity] ,
e1 e2 = e1 [e 2 is a right identity] ,
so that e1 = e2 , i. e., there is a unique left and right identity in G.
(c) Similarly one sees that a-I is unique.
(d) For Abelian groups, we often use additive notation [+ denotes
the binary operation]; in this case we denote the identity by 0, the in-
verse of a by -a, and a + (-b) by a-b.
(5.3) Definition. Consider a set A with two binary operations + and·
[called addition and multiplication, respectively], which is an Abelian
group under +, with identity 0, and in which the equalities
a· (b + c) = (a· b) + (a· c) [left distributive law] ,
(a + b) . c = (a . c) + (b . c) [right distributive law] ,
and
(a· b) . c = a· (b· c) [associative law for multiplication]
hold for all a, b, c EA. Then A is called a ring. If a· b = b· a for all a, b
in a ring A, A is called commutative. An element 1 E A such that 1 . a
= a . 1 = a for all a E A is called a [two-sided] unit for A. A non void
subset I of a ring A is called a left [right] ideal if a - bEl for all a, bEl
andx·aEI [a·xEI] forallaElandxEA. A subset IofA that is a left
and a right ideal is called a two-sided ideal.
(5.4) Remarks. (a) The notation in the statements of the distributive
laws is correct but clumsy. From now on we will follow the universal
algebraic convention that ab means a· b and that ab + cd means
(a· b) + (c· d).
(b) For all x in a ring A, we have xx = x(x + 0) = xx + xO, so
that xO = 0. Similarly Ox = 0.
(c) Evidently two groups or rings can be distinct objects and still
be indistinguishable as groups or rings. Formally, we say that rings A
and A are isomorphic if there is a one-to-one mapping -r carrying A
I
onto A' such that -r(a + b) = -rea) + -reb) and -r(ab) = -rea) db) for all
a, bE A. The mapping -r is called an isomorphism or an isomorphic
mapping. An analogous definition is made for groups. An isomorphism
of a ring or group onto itself is called an automorphism.
Hewitt/Stromberg, Real and abstract analysis 3
34 Chapter 1. Set theory and algebra
1 1 0 1 I0
(d) If P is a prime, the set {O, 1,2, ... , p - I} is a field under addi-
tion and multiplication modulo p. All necessary verifications are easy;
we will give the least easy one, namely that each nonzero element has
a multiplicative inverse. If a E{I, 2, ... , p - I}, then we must show
that there exists x E{I, 2, ... , p - I} such that
ax == 1 (mod P).
Since p is prime, the greatest common divisor of a and p is 1. It follows
that there are integers x and y, x =l= 0, such that
1 = ax + py.
In particular, there are integers x' and y' such that
1 = ax' + py' , 1 ~ x' < p;
hence ax' == 1 (mod P) and x' E {I, 2, ... , p - I}.
(e) If F is any field, then the elements 0, 1, 1 + 1, ... , n 1, ...
°
[where n is any positive integer and n 1 has an obvious meaning] are all
members of F. If n 1 = for some positive integer n, then the smallest
°
positive integer p such that p 1 = is obviously a prime. In this case F
is said to have characteristic p; otherwise F is said to have characteristic 0.
Fields of characteristic p are of no interest at present in elementary
analysis. If F has characteristic 0, then F has a sub field [the definition
of a subfield is obvious] which is isomorphic to the rational number field.
We will always denote the rational number field by the symbol Q.
To see that F contains a [unique] isomorph of Q, consider a mapping
i of part of F onto Q, and see what properties it must have in order to
be an isomorphism. For notational convenience, we ignore the distinc-
§ 5. Construction of the real and complex number fields 35
tion between the zeros of F and of Q and between the units of F and of Q.
It is clear that 1'(0) = 0 and 1'(1) = 1. A trivial induction shows that
l' (n 1) = n for all n E N. If l' is an isomorphism, we must also have
1
1'«(n 1)-1) = nand 1'(- (n 1)) = - n for all n EN. It follows that we must
have
1'«(m 1) (n 1)-1) = :
and
1'((-(ml)) (nl)-I) = - :
for all m, n EN. It is a routine matter to show that the mapping l' so
constructed is an isomorphism of a subfield of F onto the field Q.
(f) Let G be any group. For A, E C G, we define
A E = {xy: x E A, Y E E} ,
A-I = {X-l: x E A} ,
A2 = AA .
lal =
{ -aaifa~O,
if a < 0.
§ 5. Construction of the real and complex number fields 37
Since ~n
°
> and F is Archimedean ordered, we have S =1= 0. Also,
again by the Archimedean order property of F, there is a positive integer p
such that
1
p-;;>-a,
- (p !) < a,
1
(-p)-;;<a.
It follows that -p, - (p + 1), - (p + 2), ... lie outside of S. Hence S
is a non void set of integers which is bounded below, and so it contains a
least element m. Since mE S, we have a < ~
n
.
m-l
Since m - 1 4 S, we have-n- ~ a; and so
m m-l 1 1
-=--+-:0;:
n n n - a+-<a+
n (b-a) =b·'
i. e.,
a<~<b.
n
0
Proof. Let us first show that sums and products of Cauchy sequences
are again Cauchy sequences. Let (an), (bn) E <r be given. For a positive
e E F, let N (e) and M (e) be the positive integers associated with the
Cauchy sequences (an) and (b n), respectively. Ifp, q;;;' max{N(~), M( ~)},
I I
then we have lap + bp - (aq + bq)1 ;;:;; lap - aql + Ibp- bql < Z-e+ Z-e=e.
It follows that (an) + (bn) is a Cauchy sequence.
Since <r c m,
there are positive elements c and d of F such that:
lapl ;;:;; c for p = 1, 2, 3, ... ;
Ibpl ;;:;; d for p = 1, 2, 3, .. .
and
1
xp=-
a p
if P~ m.
We have
and hence
(x"a" - 1(,,») = (al - 1, a 2 - 1, ... , am - l - 1,0,0, ... ) .
Thus it is obvious that (x"a" - 1(,,») E m. To complete the proof, we need
only show that (x,,) E <r. If p, q ~ N ( ~ e), then we have
Ix p - xql = I:p - ~q I= I:p 1·1 ~q I lap - aql < +. ~ lap - aql .
For arbitrary positive dE F, it is now obvious that
implies that
if for every positive e in F there exists a positive integer L (e) such that
la" - bl < e for all n ~ L (e). An ordered field is said to be complete if
every Cauchy sequence in F has a limit in F.
(5.26) Lemma. A sequence with a limit is a Cauchy sequence. It (a,,)
is a Cauchy sequence and (a"k)%"= 1 (1 ~ n l < n 2 < ... < nl< < ... ) is a
subsequence with limit b, then (a,,) has limit b.
Proof. The first assertion is trivial. To prove the second, choose
e> 0 in F. If L( ~ e) ~ k,
then we have la"k - bl < ~ e. Since (a,,)
is a Cauchy sequence, we have
1
lap - aql <Z-e
42 Chapter 1. Set theory and algebra
(5.27) Lemma. For oc > 0, oc EF, there exists e EF such that 0 < e < oc.
It F is Archimedean ordered, then F is also Archimedean ordered.
Proof. Since oc > 0, there exists (an) E oc such that an> 0 for
n = 1, 2, 3, . .. and (an) ~ m. Hence there is a d EF such that
as = lasl ~ d for arbitrarily large s. We have
1
lap - aql <zd
Proof. Choose anye > 0 in F and any e > 0 in F such that 0 < e < e;
this is possible by (5.27). We have
lap - aql < e if p, q ~ N (e) .
§ 5. Construction of the real and complex number fields 43
so that
2kp 2k p- 1 1
ap+ l = 2 p+1 = ap or ap+ l = 2 p+1 = ap - 2 p+1 '
and hence
(p = 1,2,3, ... ).
1 Clearly we can also define suprema and infima in arbitrary partially ordered
sets.
§ 5. Construction of the real and complex number fields 45
If q > P ~ 1, then
o~ ap - all = (ap - aH1 ) + (a HI - aH2 ) + ... + (a ll - 1- all)
1
~ 29+1 + 29+1 + ... + 20
2
1
= 1(
29+1 1 + "2 + ... +
1 1)
2.-9-1
= 21'1+1 ( 2 - 2.-1'-1
1) 1
< 2P .
1
We thus have lap - alll = ap - all < 2P whenever q > p ~ 1. From (5.31)
we infer that (a p ) is a Cauchy sequence, and so lim ap exists; call it c.
p...... oo
It is plain that ap ~ c.
We claim that supA = c. To prove it, assume first that c is not an
upper bound for A. Then there is an x E A such that x > c, and hence
there is a positive integer p such that ap - c = lap - cl < x - c; i. e.,
ap < x. Since ap is an upper bound for A, the last inequality cannot
obtain. Therefore c is an upper bound for A. Assume next that there
exists an upper bound c' for A such that c' < c, and choose a positive
integer p such that ;1' < c - c'. We then have ap - ;1' ~ c- ;1'
> c + c' - c = c', and so ap - ;1' is an upper bound for A. However,
-r (: 11) = : 12 ,
46 Chapter 1. Set theory and algebra
(5.55) Recapitulation.
In the accompanying Figure 3 we illustrate addition and multiplica-
tion of complex numbers. Addition is componentwise; graphically, one
applies the parallelogram law for addition of vectors. Multiplication is
a little more complicated. We have
Arg(zw) = Arg(z) + Arg(w) modulo 2n, and Izwl = Izl ' lwl .
iY
/
/
/
Fig. 3
Just as with real power series, one proves that exp(z) exists [i. e., the
limit exists] for all z EK. The identity
exp (z + w) = exp (z) exp (w)
4*
52 Chapter 1. Set theory and algebra
iY
\
\
\
I
/
I
I
The function z --* i '
defined implicitly above, is used frequently.
It is called the signum and is defined formally by
sgn (z) = (
-IZI if z EK
Z
n {O}',
o if z = O.
(5.57) Exercise. Use Hamel bases to prove that the additive groups R
and K are isomorphic.
(5.58) Exercise. Define addition in K as usual and define multipli-
cation "coordinatewise":
(x + iy) (u + iv) = xu + iyv .
Prove that, with these operations, K is a commutative ring with unit.
Prove also that K is not a field.
CHAPTER TWO
§ 6. Topological preliminaries
Set-theoretic topology is the study of abstract forms of the notions
of nearness, limit point, and convergence. Consider as a special but
extremely important case the line R. For x, y ER, we can define the
distance between x and y as the absolute value of x - y:
e(x,y) = Ix - yl·
54 Chapter II. Topology and continuous functions
We can also associate with a point x ER all of the points y E R such that
e (x, y) is less than a specified positive number ex, i. e., the open interval
]x - ex, x + ex[. This gives us a systematic notion of nearness in R
based on the distance-function e. Even a more general idea is needed.
A subset A of R is called open if it contains all points "sufficiently close"
to each of its members. The abstract and axiomatized notion of open
set is one common and convenient way to approach the study of set-
theoretic topology.
To make all of this precise, we begin with some definitions. We
first extend the real number system, and describe certain important
subsets of the extended real numbers.
(6.1) Definition. (a) Let 00 and - 00 be two distinct objects, neither
of which is a real number!. The set R# = R U {-oo} U {oo} is known
as the set of extended real numbers. We make R# a linearly ordered set
by taking the usual ordering in R and defining - 00 < 00 and - 00< x< 00
for each x E R. For a and b in R# such that a < b, the sets
]a,b[={xER#:a< x< b},
[a, b] = {x E R#: a ~ x ~ b},
[a, b[ = {x E R#: a ~ x < b},
and
]a, b] = {x E R#: a < x ~ b}
are called intervals, with endpoints a and b. The interval ]a, b[ is an
open interval, [a, b] is a closed interval, and [a, b[ and ]a, b] are hal/-
open intervals. If a and b are in R, these intervals are said to be bounded.
Otherwise they are said to be unbounded. Note that R itself is the open
interval ]- 00, oo[ and that all intervals under our definition have
cardinal number c.
(b) For future use, we define sums and products in R#, with a few
restrictions, by the following rules. For x, y ERe R#, x + y and xy
are defined as usual. For x ER, we define:
oo+x=x+oo=oo;
(-00) + x = x + (-00) = x -00 = -00;
we also define
00+00=00;
(-00) + (-00) = -00 - 00 = -00;
-(00)=-00 and -(--00)=00;
the expressions 00 + (- 00) and (- (0) + 00 are not defined. For x E R
1 Many writers use the symbol + 00 for what we write as 00. The + sign is a
mere nuisance and so we omit it.
§ 6. Topological preliminaries 55
and
00'00=00.
For x ERand x< 0, we define
OO'x=x'oo=-oo;
(-00)' x = X· (-00) = 00;
the expressions 00' (- 00), (- 00) . 00, and (- 00) . (- (0) are not defined.
Open subsets of R are defined in terms of open intervals, as follows.
(6.2) Definition. A set U C R is said to be open if for each x E U
there is a positive real number e such that ]x - e, x + e[ C U.
Thus we may say informally that a subset U of R is open if for each
of its points x, it contains all points y such that y is sufficiently close to x.
[The "sufficiently close" depends of course on x.] Plainly every open
interval is an open set. Some important properties of open subsets of
R are listed in the following theorem.
(6.3) Theorem. Let () denote the family of aU open subsets of R. Then:
(i) 0 E () and R E ();
(ii) if o/t is a subfamily of (), then U o/t E () l;
(iii) if {Ul , Us' ... , Un} C (), then Ul n U2 n ... n Un E ().
The reader can supply the proof of this very simple result.
The properties of open sets in R given in Theorem (6.3) form the
basis for the concept of a topological space, which we now define.
(6.4) Definition. Let X be a set and () a family of subsets of X with
the following properties:
(i) 0 E () and X E ();
(ii) if o/t is a subfamily of (), then U o/t E ();
(iii) if {Uv Us, ... , Un} C (), then Ul n Us n ... n Un E ().
[That is, () is closed under the formation of arbitrary unions and of
finite intersections.] Then () is called a topology for X and the pair (X, ())
is called a topological space. When confusion appears impossible, we will
call X itself a topological space. The members of () are called open
sets in X.
Definition (6.4) by itself is rather barren. No great number of excit-
ing theorems can be proved about arbitrary topological spaces. However,
certain entities definable in terms of open sets are of considerable interest
in showing the connections among various topological concepts and in
1 Recall that the union of a void family is 0.
56 Chapter II. Topology and continuous functions
showing the simple nature of such ideas as continuity. Also by the ad-
dition of only two more axioms, we obtain a class of topological spaces
in which many of the processes of analysis can be carried out with great
profit. We first give a few examples of topological spaces.
(6.5) Examples. (a) Naturally, the real line R with the topology
described in (6.2) and (6.3) is a topological space. This topology for R
is known as the usual topology for R. We will always suppose that R
is equipped with its usual topology unless the contrary is specifically
stated.
(b) Consider R* and the family l!J* of all sets having any of the
followingfourforms: U, U U ]t, 00], U U [-00, s[, U U [-00, s[U]t,oo],
where U is an open subset of Rand s, t E R. Then (R*, l!J*) is a topo-
logical space; l!J* is called the usual topology for R*.
(c) Let X be any set. The pair (X, .9'(X)) is a topological space,
obviously. The family .9'(X) is called the discrete topology for X. A set X
with the discrete topology will frequently be denoted by XII'
(d) Let X be any set and let l!J = {X, 0}. Then l!J is called the in-
discrete [or concrete] topology for X. This topology is of little interest to us.
We proceed to the definition of various topological concepts from
our basic notion of open sets.
(6.6) Definitions. Let X be a topological space. A neighborhood of a
point x EX is any open subset U of X such that x E U. The space X
is known as a Hausdorff space if each pair of distinct points of X have
disjoint neighborhoods. A set A C X is said to be closed if X n A' is
open. For x E X and A C X, we say that x is a limit point of A if
(U n {x}') n A =1= 0 for each neighborhood U of x. For A C X, the
closure of A is the set A - = n {F : F is closed, A C F c X}; the interior
of A is the set A °= U {U: U is open, U C A}; and the boundary of A
is the set oA = A- n (A')-.
(6.7) Theorem. Let X be a topological space.
(i) The union of any finite collection of closed subsets of X is a
closed set.
(ii) The intersection of any nonvoid family of closed subsets of X is a
closed set.
(iii) The closure A-of a subset A of X is the smallest closed set con-
taining A, and A is closed if and only if A = A-.
(iv) The interior A ° of A is the largest open set contained in A,
and A is open if and only if A = A 0.
(v) A subset A of X is closed if and only if it contains all of its limit
points.
For subsets A and B of X, we have:
(vi) AO = A'-';
(vii) oA=A-nAo,;
§ 6. Topological preliminaries 57
n n
in 91 such that x Elfs c U. ~ c u v. Let .!II = {lfs: x E U V}. n
Then.!ll C 91 [.!IImay be voidl] and un V = U.!II E~. Thus ~ is closed
under the formation of finite intersections. According to (i) X is in ~,
and, since {21 C 91, we have {21 = U {21 E~. This proves that ~ is a topol-
ogy for X. Clearly 91 is a base for~. 0
The function (x, y) - Ix - yl defined on R x R is an obvious dis-
tance-function. An important although special class of topological spaces
are those in which the topology can be defined from a reasonable dis-
tance-function. The axiomatic definition follows.
(6.12) Definition. Let X be a set and let e be a function from X x X
into R such that for all x, y, z E X we have:
(i) e(x, y) ~ 0;
(ii) e(x,y) = 0 if and only if x = y;
(iii) e(x, y) = e(y, x) ;
(iv) e(x, z) ~ e(x,y) + e(y, z) [the triangle inequality].
Then e is called a metric [or distance-function] for X; e(x, y) is called
the distance from x to y, and the pair (X, e) is called a metric space.
When no confusion seems possible, we will refer to X as a metric space.
(6.13) Examples. (a) Let n be a positive integer, let X = RR or KR,
and let p be a real number such that p ~ 1. For re = (Xl' ... , xR ) and
y = (Yl' ... , YR) in X, define
I
Properties (6. 12.i) - (6. 12.iii) are obvious for ep. The triangle ineqUality
(6.12.iv) is a special case of MINKOWSKI'S inequality, which we will
prove in (13.7) infra.
The metric e2 is known as the Euclidean metric on R!' or KR.
(b) For re, y E Rfl or KR define
e(re, y) = max{lxi - Yil : 1 ~ j ~ n}. It is easy to verify that (KR, e)
and (RR, e) are metric spaces.
(c) Let X be any set. For x, Y EX define e(x, y) = ~SY [~ is KRON-
ECKER'S ~-symbol as in (2.20)]. Plainly e is a metric. It is known as the
discrete metric for X.
(d) Consider the set NNo, which we realize in concrete form as the
set of all sequences (a,,)A= I of positive integers. For a = (a,,) and b= (b,,)
in NNo, define:
e(a, b) = 0 if a = b;
the existence of the set A" U {x} contradicts the maximality of A".
It follows that A- = X. 0
(6.37) Theorem. Let X be a metric space. Then the following three
assertions are pairwise equivalent:
(i) X is compact;
(ii) X is Frechet compact;
(iii) X is sequentially compact.
Proof. The fact that (i) implies (ii) follows from (6.35). Suppose that
(ii) holds and let (x,,) be a sequence with values in X. If (x,,) has only
finitely many distinct terms, it is clear that there exists an infinite set
{n,,: kEN} C N such that nl < ns< ... and X"j; = x", for each kEN.
In this case the subsequence (X"k) converges to x",. Therefore we suppose
that (x,,) has infinitely many distinct values. Then the set {x,,: n EN}
has a limit point x EX. Let x", = Xl. Suppose that x"" ... , x"l: have
been chosen. Since each neighborhood of x contains infinitely many
distinct x,,'s, we choose x,," + 1 EB_l_(X) such that nHl > ni (1;;;; j;;;; k).
IHI
64 Chapter II. Topology and continuous functions
Then the subsequence (xnl:) converges to x. Thus (ii) implies (iii). Next
suppose that (iii) holds. According to (6.36) X is separable so, by (6.23),
X has a countable base &I. Now let !fI be any open cover of X. Let
.91 = {B E &I: Be U for some U E !fI}. For each BE.9I, choose UB E !fI
such that B C UB and let "f/' = {UB : BE .91}. Clearly "f/' is a countable
family. If x E X, then x E U for some U E !fI, and since &I is a base, there
is aBE &I such that x E B c U. Then BE.9I and x E B CUB' We
conclude that "f/' is a countable subcover of !fl. Enumerate "f/' in a se-
"
quence "f/' = (v,.). For each kEN let w,. = U v,.. To prove (i), we
n=1
need only show that w,. = X for some kEN. Assume that this is false.
For each k choose x" E X n W~. Then (x,,) has a subsequence (x"J)
converging to some x EX. Since "f/' is a cover there exists a ko EN such
that x E lIk. C w,. •. Thus w,.. is a neighborhood of x which contains x"
for only finitely many k. This contradiction establishes the fact that (iii)
implies (i). D
(6.38) Theorem. Let X be a Hausdorff space and let A be a subspace
of X that is compact in its relative topology. Then A is a closed subset
of X.
Proof. We will show that A' is open. Let zEA'. For each x E A
choose disjoint open sets Uz and Vz such that x E Uz , z E Yx. Then
{Uz n A : x E A} is an open cover of A, in its relative topology, so there
n n
exists a finite set {Xl' ... , x n } C A such that A C.U UZJ' Let V =.n VZj '
1=1 1=1
Then V is a neighborhood of z and V n A = 0, i.e., V CA'. D
(6.39) Theorem. Let X be a compact space and let A be a closed subset
of X. Then A is a compact subspace of X.
Proof. Let ofF be any family of closed [in the relative topology]
subsets of A having the finite intersection property. Then each member
of ofF is closed in X, so (6.34) implies that n ofF =1= 0. Thus A is compact,
by (6.34). D
We next present a striking characterization of compactness which
shows that we may restrict our attention to very special open covers in
proving that a space is compact.
(6.40) Theorem [ALEXANDER]. Let X be a topological space and let f/
be any subbase for the topology of X [see (6.10)]. Then the/following two
assertions are equivalent.
(i) The space X is compact.
(ii) Every cover of X by a subfamily of f/ admits a finite subcover.
Proof. Obviously (i) implies (ii). To prove the converse, assume that
(ii) holds and (i) fails. Consider the family )I{ of all open covers of X
without finite subcovers. The family )I{ is partially ordered by inclusion,
and plainly the union of a nonvoid chain in )I{ is a cover in)l{. ZORN'S
§ 6. Topological preliminaries 65
Lemma (3.10) implies that )K contains a maximal cover "Y. That is,
"Y is an open cover of X, "Y has no finite subcover, and if U is any open
set not in "Y, then "Y U {U} admits a finite subcover. Let "If" = "Y n f/.
Then no finite subfamily of "If" covers X, and so (ii) implies that "If" is
not a cover of X. Let x be a point in X n (U"If")" and select a set V
in the cover "Y that contains x. Since f/ is a subbase, there are sets
.
51' ... ,5" in f/ such that x E,n 5 i c V. Since x ~ (U "If"), no 5 i is in
1=1
"Y. Since "Y is maximal, there exists for eachj a set Ai which is the union
of a finite number of sets in "Y such that 5 i U Ai = X. Hence
.U
V U 1=1 Ai::::> (/1 5 i ) U (.U Ai)
1=1 1=1
= X,
and therefore X is a union of finitely many sets from "Y. This contradicts
our choice of "Y. 0
Another important class of topological spaces are those obtained
by taking the Cartesian product of a given family of topological spaces.
We need a definition.
(6.41) Definition. Let {X,},E! be a nonvoid family of topological
spaces and let X = X X, [see (3.l)J. For each, E I, define n, on X
lEI
by n , (~) = x,, The function n , is known as the projection of X onto X,.
We define the product topology on the set X by using as a subbase the
family of all sets of the form n;-1 (U;), where , runs through I and U;
runs through the open sets of X,. Thus a base for the product topology
is the family of all finite intersections of inverse projections of open sets.
A base for the product topology is the family of all sets of the form
XU;, where U; is open in X, for each, E I and U; = X, for all but a finite
'EI
number of the ,'so Whenever we discuss the Cartesian product of a family
of topological spaces, it is to be understood that the product is endowed
with the product topology unless the contrary is specified.
(6.42) Exercise. Let I = {I, 2, ... , n} for some n E N, and for each
,E I let X, = R [or KJ with its usual topology. Clearly X = 'EI X X, = R"
[or K"J. Prove that the product topology on X is the usual topology on X.
(6.43) TIHONOV'S Theorem 1. Let {X,},E! be a nonvoid family of
compact topological spaces. Then the Cartesian product X of these spaces
is compact [in the product topology J.
Proof. According to ALEXANDER'S theorem (6.40) it suffices to con-
sider open covers of X by subbasic open sets as described in (6.41).
Let CfI be any cover of X by subbasic open sets. For each, EI, let CfI,
denote the family of all open sets U C X, such that n;-I(U) E CfI. We
1 This theorem was proved by A. TIHONOV for the case in which each X, is the
closed unit interval [0,1] [Math. Annalen 102, 544-561 (1930)]. The general case
was first proved by E. CECH [Ann. of Math. (2) 38, 823-844 (1937)].
Hewitt/Stromberg, Real and abstract analysis 5
66 Chapter II. Topology and continuous functions
assert that U <¥t. = XI for some tEl. If this were not the case, there would
be a point ~ E X such that for every tEl, n l (~) = XI E XI n (U <¥t.)';
hence ~ ~ n;-l (U) for all n;-l (U) E tfI. That is, tfI would not be a cover of X.
Hence we can [and do] choose an 1] E I such that U tfl7/ = X7/' Since X7/
is compact, there is a finite family {Ut> ... , Un} C tfl7/ such that
X7/ = U1 U U2 U ... U Un. Plainly {nq-1(Ui):j = 1, ... , n} is a finite
subcover of tfI for X. 0
We next characterize the compact subspaces of R.n and Kn.
(6.44) Theorem [HEINE-BoREL-BoLZANO-WEIERSTRASS]. Let n EN
and let A eRn [or Kn]. Then A is compact [in the relativized usual
topology] it and only it A is closed and bounded 1.
Proof. The mapping X + iy -+ (x, y) of K onto R2 preserves distance:
1
I(x + iy)- (u + iv)l = «(x - U)2 + (y - V)2)2" = e«(x, y), (u, v)). Thus
Kn and R2n are indistinguishable as topological spaces. We therefore
restrict our attention to the case in which A eRn.
We first take the case that n = 1 and A = [a, b], a bounded closed
interval in R. A subbase for the topology of [a, b] is the family [/ of
all intervals of the form [a, d[ or ]c, b] where c, dE [a, b]. Let tfI be any
cover of [a, b] by sets in [/. Since b is covered by tfI, there is a set of the
form ]c, b] in tfI. Let Co = inf{c: ]c, b] E tfI}. Since Co is covered by tfI,
there is an interval [a, d1 [ E tfI such that co< dl . By the definition of
infimum, there is an interval ]ct> b] EtfI such that Cl < dl • Thus {[a, ~[,
]ct> b]} C tfI and [a, b] = [a, ~[U] cl , b]. It follows from this and ALEX-
ANDER'S theorem (6.40) that [a, b] is compact.
.
Now let n be arbitrary and suppose that A is closed and bounded .
Since A is bounded, there exists a cube C =.X [ai' bi] such that A c C.
1=1
The preceding paragraph and TIHONOV'S theorem (6.43) show that C
is compact. Using the fact that A is closed and citing (6.39), we see that A
is compact.
Conversely, suppose that A is compact. By (6.38), A is closed.
00
I t is clear that A cUB ~ (0) = Rn, where B ~ (0) is the open ball of radius
"=1
k centered at 0 = (0, 0, ... , 0) in Rn. Since A is compact, there exists
ko EN such that A c B",(O), i.e., A is bounded. 0
(6.45) Exercise. Prove the following.
(a) Any compact subset of a metric space is bounded.
(b) Theorem (6.44) is not true for arbitrary metric spaces.
(c) Every bounded sequence in Rn [or Kn] admits a convergent sub-
sequence.
1 A subset A of a metric space X is said to be bounded if there exist P E X and
(J E R such that e(P. x) ~ {J for all x E A.
§ 6. Topological preliminaries 67
e (xm' X.,) < B/2 and k ~ ko implies e (X"k' x) < B/2. Choose kl ~ ko such
that n", ~ no' Then n ~ n", implies e (x." x) ~ e (x." X.,,,) + e (X.,,,,, x)
< B/2 + B/2 = B. Thus lim X., = x, and so X is complete. 0
.. -..00
n A., =
.. =1
{x} for some X EX .
Proof. Suppose that (An) is a decreasing sequence of nonvoid closed
subsets of X such that diam (A.,) -+ O. For each n EN let X., E A.,. Then
m ~ n implies that e(xm' X.,) ~ diam (A.,) -+ 0 so (X.,) is a Cauchy se-
quence. Let X = lim X.,. For each m, X., E Am for all large n, and Am
.. -..00
00 00
Cauchy sequence, diam (An) -)- 0. Let nC]1 An = {x}. If e> 0, then there
is an no E N such that diam (An,) < e. But x E An., so n ~ no implies
that 12 (xn' x) < e. 0
(6.53) Definition. Let X be a topological space. A set A C X is
said to be nowhere dense if A - 0 = 0. A set F C X is said to be of first
category if F is a countable union of nowhere dense sets. All other sub-
sets of X are said to be of second category.
(6.54) Baire Category Theorem. Let X be a complete metric space.
Suppose that A C X and that A is oj first category in X. Then X n A'
is dense in X. Thus X is oj second category [as a subset oj itself].
00
a" ~ t.. < x .. and ]t.. , x] cU. Then ]a", x] = U ]x.. , x] C fI.=1S
fI.=n.
U ]t.. , x] cU.
o
Likewise, we have [x, bA C U, and hence ]a", bA C U.
We next show that a" ~ U, b" ~ U. Assume that b" E U. Since U is
open, there is a 15 > 0 such that ]b" - 15, b" + !5[ C U. But then [x, b" +!5[
= [x, bA U [b", b" + !5[ C U and b" + 15 > b". This contradicts the def-
inition of b". Thus b" ~ U. Likewise a" ~ U.
Let J = {Ja", b,,[: x E U}. Since x E U implies x E ]a", b"L we have
U = UJ. We next show thatJ is a pairwise disjoint family. Let x,y E U
and suppose that there exists u E ]a", bA n ]a:v, b:v[. If a" < a:v < u,
then a:v E U and if a:v < a" < u, then a" E U. But neither a" nor a:v is
in U. Therefore a" = a:v. Likewise b" = b:v. Accordingly any two inter-
vals in J are either disjoint or identical, i.e., J is pairwise disjoint.
For each IE J there is a rational number r[ E I. Since J is pair-
wise disjoint, the mapping f: I ~ r[ of f into Q is one-to-one, and so
:7 ~ Q= Ro. Thus J is countable.
It remains only to prove that J is unique. Thus suppose that U = U"
where" is a pairwise disjoint family of open intervals. Let ]a, b[ E ".
Assume that a E U. Then there exists an interval ]c, d[ E " such that
70 Chapter II. Topology and continuous functions
a E Jc, de· Thus Ja, b[ =l= Jc, dL but Ja, b[ n Jc, d[ = Ja, min{b, d}[ =l= 0.
This contradiction shows that a ~ U. Likewise b ~ U. Let x E Ja, be.
Then Ja, xJ C U and [x, b[ C U so Ja, b[ C Ja x , bx [ C U. Since a ~ U
and b ~ U, we have Ja, b[ = Ja x , bA Ef. Therefore f C f. If there
exists Ja x , bA Ef nf', then x E U while x ~ Uf = U, a contradiction.
Therefore f = f. 0
(6.60) Remark. The simple structure of open sets in R has no analogue
in Euclidean spaces of dimension> 1. For example, in the plane R2
open disks play the role that open intervals play on the line as the
building blocks for open sets, i. e., the base for the topology. But it is
plain that the open square {(x,y): 0< x< 1,0< y< I} is not a union
of disjoint open disks, for if it were, the diagonal {(x, x): 0< x< I}
would be a union of [more than oneJ disjoint open intervals, contrary to
the uniqueness statement of (6.59).
Neither do the closed subsets of R have such a simple structure as
the open ones do. The next few paragraphs show this rather complicated
structure. We begin with a definition.
(6.61) Definition. Let X be a topological space and let A eX. A point
a E A is called an isolated point of A if it is not a limit point of A, i. e.,
if there exists a neighborhood U of a such that UnA = {a}. The
set A is said to be perfect if it is closed and has no isolated points, i. e.,
if A is equal to the set of its own limit points.
We will now construct a large class of nowhere dense perfect subsets
of [0, IJ.
(6.62) Definition. Remove any open interval 11 ,1 of length < 1
from the center of [0, IJ. This leaves two disjoint closed intervals
11,1 and 11,2 each having length < ~. This completes the first stage of
our construction. If the nth step of the construction has been completed,
leaving 2n disjoint closed intervals I n,l' In, 2' ••. , In, 2" [numbered
from left to rightJ, each of length < ;n ,
we perform the (n + 1)8t step
by removing any open interval I n +1,k from the center of In,k such that
the length of I n +1,k is less than the length of In,k (1 ~ k;:::;; 2n). This
leaves 2 n +l closed intervals In+1,l"'" In +1, 2n+ 1 each of length < 2,,1+1'
2"-1 2" 00
Let ~ = k=1
U In ' k and Pn = k=1 n Pn = [0, IJ
U In' k (n EN). Let P = n=1
1
is exactly 3 of the length of In,k for all k, n E N, 1 ~ k ~ 2n, the
resulting set P is known as the Cantor ternary set [or simply the Cantor
§ 6. Topological preliminaries 71
setJ. In this latter case J1,l = [0, !]. J1,2 = [~, 1]. 1 2 ,1 = ]!, : [,
12,2 = ] ~, : [, J2,l = [0, ~-]. etc.
(6.63) Theorem. Let P be any Cantor-like set. Then P is compact,
nowhere dense in R, and perfect.
Proof. We use the notation of (6.62). Obviously each Pn is closed,
so that P is closed and bounded and hence compact (6.44). Since no
Pn contains an interval of length ~ ;n
and P C Pn for each n E N, it
follows that P contains no interval. Thus P- o = po = 0; that is,
P is nowhere dense in R. Next let x E P. For each n E N we have x E Pm
so that there exists kn such that x E J ..,k". Thus, given B> 0, there is an
n EN such that ;n < B, and therefore the endpoints of J ..,k" are both in
Jx - B, X + B[. But these endpoints are in P. Hence x is a limit point
of P. We conclude that P is perfect. 0
,,=2 .. -I
00
= {x E A : e(x o, x) ;;;; el} and A(I) = {x E A : e(xv x) ;;;; ell. Then A(O)
and A(I) are disjoint infinite closed sets each of diameter;;;; 1. Suppose
that n is a positive integer and for each n-tuple (~, ... , an) E {O, I}n
we have an infinite closed subset A (aI' ... , an) of A having diameter
;;;; ~ and such that no two of these sets have a common point. For
(av ... ,an) E {o,l}n, choose x(~, ... ,an,O)=l=x(al, ... ,an,l) in
A(~, ... , an) and let en +1 = min{2(n 1+ 1)' ~ e(x(al ,···, an, 0),
x(al , ... , an, 1))}. Define A(al , ... , an, j) = {x E A(al , ... , an):
e(x(~, ... , an, j), x) ;;;; en+1} (j = 0, I). Then {A(a l , ... , an+!) :
(aI' ... , an +1) E {O, I}n+l} is a pairwise disjoint family of closed in-
finite sets each having diameter;;;; n ~ 1 . Thus for each a = (an) E{O, I}N
we have a decreasing sequence (A(a l , .. . , an))':=l of infinite closed
subsets of A with diameters tending to 0. Hence by CANTOR'S theorem
00
odd, 0< a < 2"}'Let D = "goD". Thus D is the set of all dyadic rational
numbers in [0, 1]. We shall define by induction on n a chain {UthED of
subsets of X. First let U1 = A and Uo = U. For n = 1 we have D1 = {!}
and we apply (6.79) to obtain an open set UJ:. such that U1- C UJ:. c Ui: c Uo.
9 2 ~
Next let n 6 2 and suppose that open sets Ut have been defined for all
_1 _1 a
tE "=1
U D" SO that s< tin U D" implies u,-C U8 • For t =
"=0
-2" ED.., we set
76 Chapter II. Topology and continuous functions
a-I a+ I .
t' = ~ and t" = ~ and notIce that lIe, and lIe" are already defined
[a - 1 and a + 1 are evenJ. We again use (6.79) to obtain an open set
Ut such that lIe;;C Ut C lIe- C lIe,. Thus we obtain the desired family
{Ut}tED, and we have lIe- C Us whenever s < tin D.
Now define I on X by f(x) = 0 for x E U' and I (x) = sup{t ED: x E Ut }
for x E U. Clearly I (x) = 1 for all x E A = U1 • It remains to show that f
is continuous. To this end, let 0 ~ ~< 1 and 0 < (J ~ 1. Clearly I (x) > ~
if and only if x E Ut for some t > ~ and therefore l-l(J~, 1]) =
U {Ut : tED, t > ~}, which is open. In like manner I(x) ~ {J if and only
if x E Us for every s< (J. Therefore l-l([{J, 1]) = n {Us: sED, s< {J}
= n {lIe-: tED, t < {J}, which is closed. Taking complements we see
that 1-1 ([0, (J[) is open. These facts together with (6.70) show that I
is continuous. 0
We now take up the notions of limit superior and limit inferior for
sequences of real numbers.
(6.81) Definition. A nondecreasing [nonincreasing] sequence in R# is a
sequence (xn) C R# such that m ~ n implies Xm ~ Xn [xm ~ xnJ. A se-
quence (xn) C R# is said to have limit 00 [ - 00] if to each ~ E R there
corresponds an n" E N such that n ~ n" implies x" ~ ~ [xn ~ ~], and
we write lim Xn = 00 [lim Xn = - 00] or Xn ~ 00 [xn ~ - 00 J. A sequence
n---+oo n---+oo
that is either nondecreasing or nonincreasing is called monotone.
(6.82) Theorem. Every monotone sequence in R# has a limit in R#.
Proof. Let (xn) be nondecreasing and let x = sup{xn: n EN}. Then
lim Xn = X. 0
"->-00
(6.83) Definition. Let (xn) be any sequence in R#. We define the
limit superior 01 (xn) to be the extended real number
lim Xn = inf (sup xn)
"->-00 kEN ,,~k
Obviously the sequences (sup xn).~'_l and (inf Xn)k'=l are monotone
n~k - n~k
sequences, so that lim Xn and lim Xn are just their respective limits.
n---+oo 110---+00
The alternative notations lim sup Xn = lim Xn and lim inf Xn = lim Xn
n->-oo
are often used.
(6.84) Theorem. Let (xn) be a sequence in R# and let L = {x E R#: x
is the limit 01 some subsequence 01 (xn)}. Then lim Xn and lim Xn are in
n-::;OO 11---+00
Proof. We prove only the assertions about the limit superior, the
others being obvious duals. Let x = lim x"' and for each kEN, let YII
....... 00
field in § 5J.J
(6.86) Exercise. Let (X, e) be a metric space. For each nonvoid
subset A of X and each x E X, define
e(x, A) = inf{e(x, a) : a E A}.
The number e(x, A) is called the distance Irom x to A. Prove each of the
following statements.
(a) If 10 =F A C X, then A- = {x EX: e(x, A) = O}.
(b) If 10 =F A C X and x, y EX, then le(x, A) - e(y, A)I ~ e(x, y).
78 Chapter II. Topology and continuous functions
is continuous on X. Also h(A) = {I} and h(B) = {o}. Notice that this
gives a simple proof of (6.80) in the case that X is a metric space.
(6.87) Exercise. Let (X, e) be a metric space and let A and B be
nonvoid subsets of X. Define the distance Irom A to B to be the number
e(A, B) = inf{e(a, b): a E A, bE B}.
Prove the following assertions.
(a) If A is compact, then there exists a point a E A such that e(a, B)
=e(A,B).
(b) If A and B are both compact, then there exist points a EA and
bE B such that e(a, b) = e(A, B).
(c) If A is compact and B is closed, then e(A, B) = °
if and only
ifAnB=!=0.
(d) If X is a noncompact metric space with no isolated points, then
X contains nonvoid, closed, disjoint sets A and B such that e(A, B) = 0.
(6.88) Exercise. Let X be a nonvoid complete metric space. Suppose
that I is a function from X into X such that for some constant c E JO, 1 [
we have
e(t(x), I(y) ~ ce(x,y)
for all x, y EX. Prove that there exists a unique point u E X such that
I(u) = u. [Let xE X and consider the sequence x, I(x), 1(t(x), ... J.
This result is known as BANACH'S fixed-point theorem. It implies several
existence theorems in the theory of differential and integral equations.
(6.89) Exercise. Prove that the closed interval [O,IJ cannot be ex-
pressed as the union of a pairwise disjoint family of closed [nondegenerate J
intervals each of length less than 1.
(6.90) Exercise. Suppose that: X is a topological space; Y is a metric
space; and I is a function from X into Y. For each x E X, define
w (x) = inf{diam(t(U): U is a neighborhood of x}.
The function w is called the oscillation lunction lor I. Prove the following
statements.
(a) The function I is continuous at x if and only if w (x) = 0.
(b) For each real number IX, the set {x EX: w (x) < IX} is open in X.
(c) The set {x EX: I is continuous at x} is a Gd set.
(d) There is no real-valued function I defined on R such that
{x E R : I is continuous at x} = Q.
(e) There exists a real-valued function I on R such that {x ER :
I is discontinuous at x} = Q.
§ 6. Topological preliminaries 79
00 OO[OO{
A = kOI m':2 1 n£1m xE X: e(tm(x), In (x}) ;;;; k
l}]O
Prove that:
(a) I is continuous at each point of A;
(b) X n A' is of first category in X; I
(c) if X is of second category [in itself], then {x EX: I is continuous
at x} is dense in X;
(d) l-l(V} is an Fa set for each open set VcY [prove that I-I (V)
= kgl mgl nOm {x EX: e(tn(x), V') ~ !}] ;
(e) the function ~Q is the pointwise limit of no sequence of contin-
uous real-valued functions on R.2
(f) Prove that ~Q(x) = lim [lim {cos (m!nx)}2n] for all x ER.
m~oo n-+oo
(g) Prove that sgn (x) = lim ~ arctan (nx) for all x E R.
n-+oo n
(h) Prove that 1 - ~Q(x) = lim sgn{sin2(m!nx)} for all x E R.
m->oo
(6.93) Exercise. Let l'oo(N) denote the set of all bounded sequences
x = (xn) of real numbers. For x, y E l'oo(N) , define d (x, y) = sup {Ixn - Ynl :
n EN}. Prove the following.
(a) The function d is a metric for l'oo(N}.
(b) The metric space l'oo (N) is not separable.
(c) If (X, e) is any separable metric space, then there exists an isom-
etry I from X into l'oo(N), i.e., d(t(x), I(y) = e(x,y} for all x,yE X.
[Let (Pn)':=l be dense in X and define I(x) = (e(x, Pn) - e(Pn' Pl)':=l]'
(6.94) Exercise. Prove that if X is a compact metric space and I is an
isometry from X into X, then I is onto X.
(6.95) Exercise. Let X be a locally compact Hausdorff space and let
D be a dense subset of X such that D is locally compact in its relative
topology. Prove that D is open in X.
(6.96) Exercise. Let X be a linearly ordered set. The order topology
lor X is the topology on X obtained by taking as a subbase the family
00
1 Show first that if (Em);;'=l is any sequence of subsets of X, then ~ E-;;
00
C ( n Em U
) (00U ~n
'\ .
E;"; m-l
.. =1 m-l m
a Recall that ;Q is the characteristic function of Q (2.20).
80 Chapter II. Topology and continuous functions
of all sets of the form {x EX: c < x} and {x EX: x < d} for c, dE X.
Prove that X, with its order topology, is compact if and only if every
nonvoid subset of X has both a supremum and an infimum in X. [Use
(6.40) as in the proof of (6.44).]
(6.97) Exercise. (a) Use (6.96) to prove that a well-ordered set is
compact [in its order topology] if and only if it contains a greatest
element.
(b) Use (6.96) to show that R* with the usual topology (6.5.b) is
compact.
(6.98) Exercise. Prove that the set P D of all countable ordinal
numbers [see (4.49)] with its order topology (6.96) is sequentially compact
but not compact.
(6.99) Exercise. Let P be CANTOR'S ternary set and let X = {O, l}N
have the product topology, where {O, I} has the discrete topology. For
re = (xv x 2, ••• ) E X, define 11' (re) = 1;
,,=1
23:" . According to (6.64), 11' is
a one-to-one mapping from X onto P. Prove that both 11' and 11'-1
are continuous.
(6.100) Exercise. Prove that if Y is a compact metric space and P is
CANTOR'S ternary set, then there exists a continuous function I from P
onto Y. [Let {v,,}:=1 be a countable base for the topology of Y. For each
nE N, set A",o= v,;- and An,l = Y n V~. For a point x =
00
1;
,,=1
2;" in P,
the set "~1 An,x" is either void or contains just one point. Let
(b) Let E = [a, b] n "~1 F~. Prove that I is one-to-one on E and that
I(E)=/([a, b]). In fact, each xEE is equal to sup{Y:YE [a,b], I(y)
= I(x)}. Note that E is a Ga set.
(6.102) Exercise. Let I be a real-valued function with domain R
having a relative minimum at each point of R, i. e., for each a E R, there
is a number ~(a) > 0 such that I(t) 6 I(a) if It - al < ~(a).
(a) Prove that I (R) is a countable set.
(b) Find a function as above that is unbounded and also monotone
on no interval containing O.
§ 7. Spaces of continuous functions 81
= 8.
then there is an IE era (X) such that all I - In are bounded and
lim III - Inll .. =
n ...... oo
°.
We return to a study of function spaces on locally compact Haus-
dorff spaces.
(7.12) Definition. Let X be a nonvoid locally compact Hausdorff
°
space. Let <roo (X) be the subset of <r (X) consisting of all IE <r (X) such
that for some compact subset F of X [depending on fJ, I (x) = for all
x EF' n X. Let <ro (X) be the subset of <r (X) consisting of all IE <r (X)
such that for every positive number e, there is a compact subset F of X
[depending on I and eJ such that II (x) I < e for all x EF' n X. Functions
in <roo (X) are said to vanish in a neighborhood 01 infinity and functions in
<ro (X) to vanish at infinity: both phrases are loose but expressive.
(7.13) Exercise. Prove the following.
(a) The inclusions <roo (X) C <ro(X) C <reX) obtain. If X is non-
compact, then <roo(X) ~ <reX) (6.80). For the space P Q of (7.11.b),
we have <roo (X) = <ro(X). If X is compact, then <roo (X) = <ro(X) = <r(X).
(b) Let X be an arbitrary locally compact Hausdorff space, and
consider <r (X) and SB (X) as metric spaces under the uniform metric
III - gil ... The space <ro(X) is closed in SB(X) [and hence also in <r(X)J:
if InE<ro(X), IESB(X), and limllin-/II .. =O, then IE <ro(X). Also
n ...... oo
<ro(X) is complete as a metric space. [Recall (6.47).J The space <roo (X)
§ 7. Spaces of continuous functions 87
is not in general closed: its closure is <ro(X). [A proof of the last state-
ment is indicated in (7.41) below.]
Several concepts closely related to, but not identical with, continuity
are needed in our treatment of integration theory. We now take them up.
(7.14) Definition. Let X and Y be metric spaces with metrics e and a
respectively. A mapping q; with domain X and range contained in Y
is said to be uniformly continuous if for every e > 0 there is a c5 > 0
such that e(x, x') < c5 implies that a(q;(x), q;(x')) < e.
(7.15) Remarks. (a) It is easy to see from (6.71) that a uniformly
continuous mapping is indeed continuous.
(b) There is a notion of uniform continuity more general than that
in (7.14), based on what are called uniform spaces. We do not need this
concept and hence omit it.
(7.16) Exercise. (a) Consider the function exp (5.56) defined on K.
Prove that exp is continuous but not uniformly continuous.
(b) Let X be a nonvoid set and let e be the discrete metric on X
(6.13.c). Prove that every mapping of (X, e) into a metric space is
uniformly continuous.
(c) Find reasonable necessary and sufficient conditions for a metric
space to admit a continuous real-valued function that is not uniformly
continuous.
(d) Let (fn)':=l be a sequence of complex-valued uniformly con-
tinuous functions on a metric space X such that all differences fn - fm
are bounded and
lim Ilfn - fmll .. = 0 .
m,n~().,)
satisfy (7.14) for the preassigned 8. If x and x' are in A; n X, than (1)
may be applied. If at least one of x and x' is in A., we may suppose that
x E A •. Then x is in some B'lYk(y,,). If e(x, x') < 15, then we have
e (y", x') ~ e (y", x) + e (x, x')
< 'YJYk + 15
~ 2'YJYk'
That is, x' is in B2'lYk(y,,) and so (2) shows that a(<p(x'), <p(y,,)) < ~.
Since x is also in B2'lYk (y,,), (2) implies that a(<p (x), <p (y,,)) < ~ ,and hence
(i)
(ii)
converges, and the series (i) converges uniformly and absolutely in [-1, 1].
Finally, we have
(iii)
for x E J-1, 1[ for all real IX. For IX> 0, (iii) holds for all x E [-1, 1J.
Proof. To prove (i), we use the ratio test. If IX is a nonnegative integer,
then all but finitely many of the numbers (:) are 0, and so (i) trivially
converges. Otherwise, for Ixl < 1 and n = 0, 1, 2, ... , we have
IX )xn+l
( n+l
§ 7. Spaces of continuous functions 91
Hence the series (i) converges absolutely for Ixl < 1. Next we prove (ii).
The case in which a is a nonnegative integer is again trivial. For a
not a nonnegative integer, let an = 1(:) I· Then we have
an+! _ la-nl _ n-a
---a:: - n +1 - -n-+T '
the last equality holding for n ~ [aJ + 1 [[aJ is the largest integer not
exceeding a]. Hence for n ~ [aJ + 1, we have (n + 1) an +1 = nan - aan,
and so
(1)
Thus for n ~ [aJ + 1, (nan) is a decreasing sequence, and so has a
00
limit; let lim nan = y ~ O. Now consider the series I.: (nan -(n+ 1) an +1).
n-+oo n=O
The pth partial sum of this series is -(p + l)ap+I' which converges to
00
- y. Hence the series I.: (nan - (n + 1) an+1) converges, and since by (1)
n~O
00
for sufficiently large n, the series l: an converges; i. e., the series (ii)
n~O
converges. Since 1(:) xnl ~ 1(:)1 for Ixl ~ 1, the series (i) converges
absolutely and uniformly in [-1, IJ and so defines a continuous func-
tion on [- 1, IJ.
We now prove (iii). For x E J-l, 1 [ and a E R, let fa.(x) = f (:) xn.
n~O
f~(x) =
n~l
i; (n+ 1) (n +a 1)xn.
fn(a)xn-l=
n n~O
we see that
t~ (x) = an~
00 (a -1)
n xn = at a.-I (x) . (2)
Next we have
(l+x)ta._I(x)=(I+x)I.:
00 (a-I)
n xn
n~O
i.e.,
(1 + x) Icc-l (x) = Icc (X) • (3)
Combining (2) and (3), we see that
(1 + x)f~(x) - IXfcc(x) = 0
for all x E ]- 1, 1 [. Also we have
and the fact that the series converges absolutely and uniformly for
Ixl ~ 1.
(b) The STONE-WEIERSTRASS theorem depends ultimately upon order
properties of the space er' (X) for a compact Hausdorff space X. It is
evident that max{f, g} and min{f, g} are continuous if I, g are in er'(X),
and this simple fact will be very useful. The following theorem connects
polynomials with maxima and minima.
(7.27) Theorem. Let X be any nonvoid set [no topology] and let 1jl
and cp be functions in ~' (X). For every e > 0, there is a polynomial
P = L:m L:" lXi k 1jli cpk with real coefficients such that
i=O "=0
(i) Ilmax{1jl, cp} - Pllu < e.
A like assertion holds lor min{1jl, cp}.
Proof. For every real number t, the identity
1
It I = (1 + t2 - 1)2
1
is obvious. For It2 - 11 ~ 1, that is, for It I ~ 22, we have
(1 + t2 - l)i = i; (+)
,,=0n
(t2-1)",
and the series converges absolutely and uniformly for all t such that
§ 7. Spaces of continuous functions 93
1
It I ~ 22 (7.26). For every positive integer p and for all real t such that
1
It I ~ 22 , we thus have
(I)
and the right side of (I) is arbitrarily small for p sufficiently large. Now
consider any nonzero function 1.p in Q3' (X), and for brevity write the
number 111.p11 .. as fJ.
For x EX and PEN, we apply (I) to write
Setting P(1.p, rp) = ~ (Q(1.p- rp) + 1.p + rp), we obtain (i). To prove (i)
with "max" replaced by "min", note that min{1.p, rp} = -max{- 1.p, -rp}. 0
(7.28) Definitions. Let X be a set and e a family of functions on X
with values in a set Y. Suppose that for all x, y E X such that x y '*'
'*'
there is an 1E e such that 1(x) I(Y). Then we say that e is a separating
lamily 01 lunctions on X. Next suppose that e is a family of real-valued
functions on X. A real polynomial in functions Irom e is any finite
sum of functions a./~'/;··· 'I?I, where the coefficient a. is a real number
and the exponents nj are positive integers. Equivalently, a real poly-
94 Chapter II. Topology and continuous functions
Hence the function 1p = max{ If"" If"" ... , If"J is in e and satisfies the
following inequalities: 1p (x) < - ! for all x EE and 1p (x) > ! for all
x EF. Now define WI by
wI=min{max{1p,- ~}, ~}.
It is clear that wIEe, WI (E) ={- !}, wI(F) ={!}, and WI(X) C
[- ~ , ~] . The definitions of E and F show that
2
Ilf - WIll .. = 3'
The function ~ (f - WI) is in ~r (X) and has minimum - 1 and maxi-
mum 1. The method used to construct WI can again be used to approxi-
mate ~ (f - WI)' Thus there exists W 2 E e such that
II ~ (f - WI) - w2 11u = ~ .
Multiplying by ~ , we have
is a sub algebra of <rr (X). Suppose that I and g are in $-, and that
lim III - Inll" = 0 ,
n~oo
lim Ilg - gnll" = 0 ,
n~oo
where In and gn are in $. Theorem (7.27) proves that max {tn' gn} is in $-.
It is easy to see from (5.36) that
Ilmax{tn' gn} - max{t, g}ll" ~ Il/n - III" + Ilgn - gil" ,
and so max{t, g} is in ('ir)- = $-. Thus $- satisfies all the hypotheses
imposed on 6 in (7.29), and (7.29) therefore implies that ($-)- = $-
is <rr(x). 0
(7.31) Corollary [WEIERSTRASS]. Let X be a compact subset 01 R
and let I E<rr (X). Then there is a real polynomial P = P (x) such that
III - PII" is arbitrarily small.
Proof. The corollary is simply (7.30) with 6 = {t, I} where t (x) = x
for all x EX. 0
(7.32) Remarks. The hypothesis in (7.29) and (7.30) that 6 be
a separating family of functions is obviously necessary. Let X be a
compact Hausdorff space containing at least two points. Suppose that
5' is a nonvoid subset of <rr (X) such that for some distinct a, b EX, we
have I (a) = I (b) for all I E5'. Then P (a) = P (b) for every polynomial P
in functions from 5'. It follows that polynomials in functions from 5'
cannot be dense in <rr(x). In fact, no function having different values
at a and b can be arbitrarily uniformly approximated by polynomials
in the functions of 5', and Theorem (6.80) shows that there is a CfJ E<rr (X)
such that CfJ (a) = 1 and CfJ (b) = O. The beauty of the STONE-WEIERSTRASS
theorem lies in the fact that its hypothesis, trivially necessary, is also
sulficient. We will use the STONE-WEIERSTRASS theorem very frequently.
It is an essential tool for the analyst.
(7.33) Examples. (a) Let X be CANTOR'S ternary set (6.62), which we
write as the set of all numbers 2 1; ~:, where eachYk is 0 or 1. For n EN,
k~l
CfJn (2 i
k~l
~:) = (_1)Yn.
n
I
For n l < n 2 < ... < n z, let CfJnl,n" ••• ,nz = CfJnr
i~l
Then each CfJn is continuous on X, the set {CfJl> CfJ2' ••• , CfJn' ••• } is
a separating family on X, and CfJ; = 1. Hence every function in <rr (X)
is arbitrarily uniformly approximable by a linear combination of the
functions 1 and CfJnl'n" •.. ,nz.
(b) The function exp defined as in (5.56) is real-valued on Rand
satisfies, as every schoolboy should know, the inequality exp (Xl) < exp (x 2 )
§ 7. Spaces of continuous functions 97
1
IIQ - Imgll .. < 2 6 ,
and so
IIP+iQ-gllu<6. D
Hewitt/Stromberg, Real and abstract analysis 7
98 Chapter II. Topology and continuous functions
(7.35) Examples. (a) Let X = {z EK: JzJ ~ I}, and let e = {t}
where t (z) = z. Unifonn limits of polynomials in t and 1 are analytic
on the open unit disk {z E K: JzJ < I} and continuous on the closed unit
disk {z EK : JzJ ~ I}. There are certainly nonanalytic continuous func-
tions on the unit disk, and so polynomials in functions from e are not
dense in (i (X). Thus some additional hypothesis is necessary in the com-
plex STONE-WEIERSTRASS theorem. Conditions onX under which the theo-
rem holds for (i (X) with no additional hypothesis have been found by
W. RUDIN [Proc. Amer. Math. Soc. 7, 825-830 (1956)].
(b) Let T = {z EK: JzJ = I}, and consider the function t on T
given by t(z)=z. For any zET, we have z=exp(ix) = cos (x) +i sin (x),
for exactly one x E [0, 2n[. Hence t(exp(ix) = exp(ix), and we see
that {t} is a separating family for T. We have t(z) = t(exp (ix) = cos (x)
- isin(x) = cos (-x) + isin(-x) = exp(-ix). Recall that exp(ix)fI
= exp(inx), n = 1,2, .... The family {t, t, I} satisfies the conditions
of (7.34). Now let 1 E (i(T) and let e > O. By the above remarks there is
a function
.
exp (ix) -+ 1: cx" exp (ikx)
k=- ..
such that
It (X nU') C {O}. Let oct = max{oc, O} and {Jt = min{{J, O}. Then define
the function q; on X as
q; = min{max{/t, {Jt}, oct}.
It is clear that q; is in err (X), and it is easy to see that: q; (x) = I (x) for
xEY; oct=max{q;(x):xEX}; {Jt=min{q;(x):xEX}; q;(x) =0 for
x EX nu'. Thus q; is an extension of I of the sort we want for which
11q;llu = Il/ilu· That is, if lEe, then I admits an extension It such that
!Wllu = 11/11u·
N ow let g be a function in e~:r (Y) that is the uniform limit on Y
of a sequence (fn)":;:=l of functions in e. Choose a subsequence (fnk)":=l
such that Il/nk - InkJu < 2- k and write gk = Ink - Ink+! (k = 1,2,3, ... ).
Then we have
on U' and Ilgtllu = Ilgkllu < 2- k • The infinite series 1: gk converges uni-
k=!
formly on X to a function in crr (X) that vanishes on U' . Therefore g - Inl
is in e and hence g itself is in e.
We have therefore shown that e satisfies the hypotheses of (7.30)
and is uniformly closed in crr (Y). Therefore e is all of crr (Y). 0
(7.41) Exercise. Let X be a locally compact Hausdorff space. Every
function I in cro (X) can be arbitrarily uniformly approximated by
functions in croo (X). [Hint: consider a compact set outside of which I is
small and use (7.40).J
(7.42) Exercise. Let E be a real or complex normed algebra with
multiplicative unit u. For x E E, let
Illxlll = sup{llyxll:y EE, IIYII ~ I}.
Prove the following:
(a) Illulll = 1;
1
(b) l1Ulf Ilxll ~ Illxlll ~ Ilxll for all x E E;
(c) the algebra E with the norm III III is a normed algebra in the sense
of (7.5);
(d) if E is complete in the metric e(x, y) = Ilx - yll, it is also complete
in the metric Illx - ylll·
(7.43) Exercise. Let X be a nonvoid compact Hausdorff space and
let .s: be a lattice of real-valued continuous functions on X, i. e., I, g E .s:
implies max{j, g} E.s: and min {I, g} E.s:. Suppose that q; is a real-valued
continuous function on X having the property that for each B > 0
and each pair of points x, y E X there exists a function IE.s: such that
§ 7. Spaces of continuous functions 101
l<p(x) - I(X)I < e and l<p(y) - t(y)1 < e. Prove that for each e >
there exists a function h E~ such that 11<p - hll u < e.
°
(7.44) Exercise [ROBERT I. JEWETTJ. Let I denote the interval
[0, 1J. A family fr of functions is said to have property V if
(i) fr C IX for some set X,
(ii) I Efr implies (1 - t) Efr,
and
(iii) I, g Efr implies Ig Efr·
Supply IX with the topology of uniform convergence [the topology
induced by the uniform metricJ. Prove assertions (a)-(k). [This exercise
is rather difficult.J
(a) If X is a set and Q( C IX, then Q( is contained in a smallest sub-
family [closed subfamilyJ of!X having property V. [Intersect a collection
of families. J
Suppose that X is a topological space, and let ~ (X) = IX n ~(X).
For n EN, let '.Pn be the smallest subfamily of ~ (In) that has property
V and contains the n projections
(Xl' ... , x n ) -J> X,. (k = 1, ... , n) .
(b) If fr has property V, if {Iv' .. , In} C fr, and if p E'.Pn' then the
function I defined on X by
I(x) = P(tI(X), ... , In (x))
is in fro [Consider the set of all P E~ (In) for which the conclusion holds.J
(c) For every e> 0, there exist functions P and q in '.PI such that
P < e and q > 1 - e on I. [Consider P(x) = xm(1 - x)m.J
° °
(d) If e > and < a < b < 1, then there is a function P E'.PI
such that
p> 1 - e on [0, aJ,
P< e on [b, IJ .
[Take P (x) = (1 - xm)n for appropriate m, n EN.J
(e) If {aI' ... , an} U {b v ... , bn} C I, then
I
ii b,. - fI a,.\;:;;; i: Ib,. - a,. I .
1<=1 1<=1 1<=1
k k j - i -1
Hence L(I, ac, Lt*) ~ L(I, ac, Lt). The proof by induction on the number
of points in Lt* and not in Lt is now clear. Similarly, suprema decrease
on subintervals of an interval, and so the last inequality in (i) also
holds. 0
(8.S) Lemma. II Lt, Lt* are in 9J([a, b]), then L(I, ac, Lt) ~ U(/, ac, Lt*).
Proof. By (8.4) we have L (I, ac, Lt) ~ L (I, ac, Lt U Lt *) ~ U (I, ac, Lt*). 0
(8.6) Theorem. Let I be Riemann-Stieltjes integrable with respect to ac
on [a, b]. Then there exists a unique real number I' such that
L (I, ac,Lt) ~ I' ~ U (I, ac, Lt)
lor every Lt E9J ([a, b]). In fact,
I' = sup{L (I, ac, Lt) : Lt E9J ([a, b])} = inf{U (I, ac, Lt) : Lt E9J([a, bJ)}.
This number I' is called the Riemann-Stieltjes integral 01 I with respect
b
to acover [a, b]. Historically it is denoted by J I (x) dac(x), but in this text we
..
§ 8. The Riemann-Stieltjes integral 107
write it as sa.(I; [a, b]). In case a(x) = x for each x E [a, b], we call y
the Riemann integral 01 lover [a, bJ and denote it by 5 (I; [a, bJ).
Proof. Let y = sup{L(f, a, .1) : .1 EE&([a, bJ)} and b = inf{U(I, a, .1):
.1 EE&([a, bJ)}. It follows from (8.5) that y and b are real numbers
and that y ~ <5. We need only show that y = <5. Assume that y < <5.
Since <5 - y > 0, the definition of integrability (8.3) shows that there is a
.1 EE&([a, bJ) such that U (I, a, .1) - L (j, a, .1) < <5 - y. Then
b ~ U (I, a, .1) = (U (f, a, .1) - L (I, a, .1)) + L (f, a, .1) < (<5 - y) +y= <5,
= ec(b)-:(a) +1 (a(b)-a(a))<s. D
r= r 1 U r 2 , we have
°
It follows that ~ U (lj' a, r j ) - L (/j' a, r j ) < ~ for j = 1, 2. Setting
Then
(U(I, IX, Ll I ) - L(I, IX, Ll I ) + (U(I, IX, Ll 2) - L(I, IX, Ll 2)
= U(I, IX, LI) - L(I, ex, LI) < e.
It follows that I is integrable over both [a, c] and [c, b]. Let 5,,(I;[a,c])
= YI and 5,,(j; [c, b]) = Y2. Clearly 0 ~ U(f, IX, Ll I ) - YI < e and
o ~ U (I, IX, Ll 2) - Y2 < e. Adding these two inequalities, we get
o ~ U(j, IX, LI) - (YI + Y2) < 2e. Since a similar statement is true
for L(I, IX, LI), we conclude that 5,,(1; [a, b]) = YI + Y2. 0
(8.12) Theorem. Let IX be a real-valued nondecreasing lunction de-
fined on R. For I E<r~o(R), define 5,,(/) = 5,,(/; [a, b]) where [a, bJ is any
closed interval in R such that I vanishes outside 01 [a, b]. Then 5,,(/) is
unambiguously defined lor each I E<r~o([a, b]). Moreover the lunction 5"
has the lalla wing properties:
(i) 5,,(j + g) = 5,,(/) + 5,,(g);
(ii) 5,,(c/) = c5,,(I);
(iii) 5,,(1) ~ 0 il IE <rto(R).
Proof. Let I E<r~o(R) and suppose that 1=0 outside of both [a, b]
and [a', b/]. Let a" = min{a, a /} and b" = max{b, b/}. By (8.11), we have
5,,(/; [a", b"]) = 5,,(/; [a", a]) + 5,,(/; [a, b]) + 5,,(1; [b, b"])
= 0 + 5,,(1; [a, b]) + 0
= 5,,(/; [a, b]) .
Similarly we see that 5,,(/; [a", b"J) = 5,,(f; [a', b/]). Therefore
5,,(1; [a, bJ) = 5,,(j; [a', b/]) and so 5,,(/) is well defined. The remaining
assertions of the theorem are now trivial consequences of previous
results. 0
(8.13) Theorem. Let IX be as in (8.12). For IE <roo (R), define 5,,(/)
as 5,,(Re/) + i5,,(Imf). Then lor I, g E<roo(R) and c EK, we have
(i) 5,,(j + g) = 5,,-(j) + 5,,-(g) ,
110 Chapter III. The Lebesgue integral
such that II (x) -/(y)1 < <x(b)-IX(a) + 1 whenever Ix- yl <C}. BecauseD
6
xn =
co
°or 1, then tp(x)
.. = 1 6
.!.
'I
'\' Xn 5
= ~ z,;-' "6
.. =1 I
The accompanying Z
J
Fig. 5, showing part ofthe 6
graph of tp may be help- ~
fu!. [Use the facttha t the ~
dyadic rationals in [0, 1]
178 I l.IUIlZ '-(l1l1 X
[on the y-axis J form a iboT Jl'7l'78 9l'7&l
(8.29) Exercise. Let 1jJ be as in (8.28) and let t(x) = x for all x E[0,1].
Prove that Sop(t; [0, IJ) = ~. [Consider the sequence of subdivisions
(L1n) where L1n consists of the endpoints of the closed intervals Jn, 11
making up P n .] For the computation of S",(t1l; [0, 1]) for k = 2, 3, ... ,
see (22.44) and (22.45) infra.
°
(ii) I(ocf) = ocI(f) [homogeneous],
(iii) I (I) ;?; if f E<rto (X) [nonnegative J.
(9.2) Examples. (a) Let X = R and I = Sa. for any real-valued non-
decreasing oc on R.
00 00
Proof. For finite <S), (9.13) and an easy induction yield the result.
For infinite <S), apply (9.13) and (9.11). 0
We now make our second, and last, extension of I.
(9.15) Definition. Let 'J+ be the set of all functions defined on X
with values in [0, ooJ. For h E'J+, let
l(h) = inf{l (g) : g E!m+, g ~ h}.
(9.16) Theorem.
(i) For g E!m+, we have I (g). I (g) =
(ii) I I hi ~ h2' hi E'J+, then we have I (hi) ~ I (h2)'
(iii) For h E'J+ and 0 ~ IX < 00, we have l(ock) = IXI(h).
(iv) For hi' h2 E'J+, we have l(h l + h2) ~ l(hl ) + l(h2).
The proof is left to the reader.
Our next theorem is a generalization of B. LEVI'S monotone conver-
gence theorem.
(9.17) Theorem [Generalized B. LEVI theorem). II hn E'J+ lor
n = 1, 2, 3, ... , and hi ~ h2 ~ ... ~ hn ~ " ' , then l( lim hn )
....... 00
= lim l(hn ).
"--->00
We must now prove that lim I (h..) ~ I (h); clearly we may suppose that
,,--->00
lim 1 (hn ) < 00. Choose e > O. For each positive integer n, select g.. E!m+
;n < l(h..). We wish to apply (9.12). To
....... 00
and so
l(gp+l) < l(g;) + I(hp+l) - I(h1) +;
= I(hp+l) + (l(gl) - I(h1)) + ~
< l(hp +1) + e .
Thus the inequality I (gP+l) < I (hp+l) + e is valid for p = 1,2,3, ....
The same inequality obtains with p = 0, and so we have
l(g~)<l(hn)+e (1)
for n = 1, 2, 3, . . .. The sequence (g~) is nondecreasing and g~ E ml+
for n = 1, 2, 3, ... ; (9.12) thus implies that
lim I (g~) = I ( lim g~) = I (supg~) .
n---+ooo n-+oo
that (9.13) and (9.11) fail in general if I is replaced by i: see (10041). The truth of
(9.17) is by itself fairly remarkable.
120 Chapter III. The Lebesgue integral
00
= 1: t(Uo). 0
OE@
(9.23) Corollary. II (Jan, bn[):=l is any pairwise disjoint sequence 01
open intervals 01 R, then A CQl Jan, bn[) = n~ (bn - an)·
Proof. This is an immediate consequence of (9.22) and the fact
[which is a trivial exercise] that A(]a, b[) = b - a. [Note that a or b
may be infinite.] See also Exercise (9.9.a). 0
(9.24) Theorem. For every A C X we have
t(A) = inf{t(U) : U is open, A C U} .
Proof. If t(A) = 00, then the result is trivial since A C X, X is open,
e:
and 00 = t (A) ~ t (X) = 00. Thus suppose that t (A) < 00 and let e > 0
be arbitrary. Choose a real number ~ such that 0 < ~ < 1 + t(A) < 1.
Next select gEcm+ such that g~ ~A and l(g)-~<I(~A)=t(A).
Let U = {x EX: g(x) > 1 - ~}. Clearly U is open and A C U. For
1 1
x E U, we have 1-<5 g(x) > 1, and so 1-<5 g;;?; ~u. Thus
- -( 1 ) 1 - 1
t(U)=I(~u)~I 1-<5 g = 1-<5 I (g) < 1_<5(t(A)+~)<t(A)+e. 0
(9.25) Remark. It follows from (9.24), (6.59), and (9.23) that for
ACRwehaVeA(A)=inf{1: (bn-an): nQl ]an,bn[::::>A and {Jan,bn[}:=l
Proof. Take any real number fJ such that fJ < £(U). Since U is open,
we have £(U) = l(~u) = sup {I (I) : I E~rio, I ~ ~u}. Thus we can choose
I E~rio such that fJ < I (I) ~ t{U). For n = 1,2,3, ... , define
F,,={xEX:/(x)~ !}, and vv,,={xEX:/(x»!}; F,. is compact,
vv" is open, and vv,,- is compact. Let W = {x EX: I (x) > a}. It is clear
that lim ~F" (x) = lim ~w (x) = ~w (x) for each x EX, and it is clear
n-+ 00 n-+ 00 n
too that the sequences are nondecreasing. Applying (9.17), we have
fJ < 1(1) ~ l(~w) = lim I(~F.. ) = lim l(~w) = lim l(Fn) = lim £ (Tv..) •
n~ 00 K-+ 00 n n-+ 00 11-+ 00
°
the prefix "t-".
(9.30) Theore~. For h Etr+, we have I(h) = if and only if h is an
t-null function. If I(h) < 00, then h is finite t-a.e.
Proof. Let A = {x EX: h (x) > a}. The functions nh, n = 1, 2, ... ,
are all in tr+, and it is obvious that lim nh ~ gAo Thus if l(h) = 0,
n->-oo
(9.17) shows that t(A)=i(gA);:;;; i(lim nh)= lim l(nh) = lim nl(h) = O.
n-+ 00 fl.-+- co "-+ 00
If h is an l-null function, then teA) = 0; using the inequality h ;:;;; lim n;A'
"_00
we have i(h) ;:;;; lim i(ngA) = lim nl(gA) = lim nt(A) = O.
n~ 00 n---7OO K-+ 00
Next suppose that l(h) < 00, and let B = {x EX: hex) = oo}. For
all e> 0 we have gB ~ eh, and so t(B) = l(gB) ~ l(eh) = el(h). Since
i(h) is finite, we infer that t(B) = O. 0
(9.31) Corollary. Let (h,,)::'=1 be a sequence of functions in tr+ and
suppose that lim i(h,,) = O. Then there is a subsequence (h,,~)k°=1 such that
"_00
00
); h".(x) < 00 l-a.e. on X and in particular, lim h".(x) = 0 t-a.e.
k=1 k_oo
00
Proof. We first select a subsequence (h"k) of (h,,) such that}; 1(h",,) < 00.
k=1
Using (9.18), we see that I ( J'I h"l:) < 00, and it follows from (9.30)
00
(110 ):'=1 of open intervals such that U In C [0, 1], A(In) = an for each
10=1
n EN, and [0, 1] n CQl
In)' is nowhere dense and perfect in R. [See
(8.26).]
(9.39) Exercise [FATOU'S lemma]. Let X and I again be arbitrary.
Suppose that (hn ) is a sequence of functions in tr+. Prove that
I( lim hn ) ~ lim
n~oo ft.---ioOO
1(hn ) .
Also find a sequence (hn ) C <rto for which the strict inequality holds,
where X = R and I = S, the Riemann integral.
(9.40) Exercise. Let X be a locally compact Hausdorff space. Let X*
be a nonvoid closed subset of X [with the relative topology]. Let 1*
§ 10. Measures and measurable sets 125
(i)
126 Chapter III. The Lebesgue integral
holds. [We have already proved (i) for functions in ml+.] The avenue
we choose toward this goal is through abstract measures and integrals
defined in terms of these abstract measures. The problem of finding
a largest family of functions on X for which (i) holds is unsolved and ap-
parently very difficult. Our approach to the problem is not the only
one possible, but it has the advantages of simplicity and also of introduc-
ing abstract integrals, which every reader should know about anyway.
The present section is mainly concerned with properties of the set
function , and in particular with its good behavior on a certain well-
defined family of subsets of X. The properties of £ that we need in defining
this family are set down in Theorem (9.21). It turns out that set functions
enjoying properties (9.2l.i)-(9.2l.iv) can be studied in abstracto, with
no reference to a topological space or to positive functionals. We make
a formal definition, as follows.
(10.2) Definition. Let X be a set [no topology]. A function /-' de-
fined on 9'(X) is called a [CARATHJ£ODORyl] outer measure if the follow-
ing relations hold:
(i) 0 ~ /-,(A) ~ 00 for all A ex;
(ii) /-,(0) = 0;
(iii) /-,(A) ~ /-,(B) if A C Be X;
(iv) /-' e.gl Aft) ~,,~ /-' (Aft) for all sequences (Aft):'=1 of subsets of X.
Outer measures are not by themselves of great use. Far more important
for integration theory are measures, which we next define.
(10.3) Definition. Let X be a set and .91 an algebra of subsets of X.
A set function /-' defined only on .91 is called a finitely additive measure if:
(i) 0 ~ /-,{A) ~ 00 for all A E.Iif;
(ii) /-,(0) = 0;
(iii) /-,(A U B) = /-,(A) + /-,(B) if A, B Ed and A n B = 0 2 •
A finitely additive measure /-' such that
(iv) /-' e.Ql Aft) =,,~ /-' (Aft) for all pairwise disjoint sequences (Aft):'_1
00
(X, .JiI, f-t) is called a finite measure space and f-t a finite measure. If X
is the union of a countable family of sets in.Jil each having finite meas-
ure, then (X,.JiI, f-t) and f-t are called (I-finite. If f-t = 0 or f-t assumes only
the values 0 and 00, then (X, .JiI, f-t) and f-t are called degenerate.
Outer measures offer a convenient method of obtaining measures,
if we suitably restrict the family of subsets on which the outer measure
is defined. Before entering on the technicalities of this construction,
we give a few examples.
(10.4) Examples. (a) Let X be any set. For A C X, let
,,(A) =
{ .A if
00 if
A is finite ,
A is infinite.
Then " is a measure and also an outer measure on .9' (X). [It is easy to
see that a measure on .9' (X) is also an outer measure.] The measure "
is usually called counting measure.
(b) Let X be any set. For A C X, let
f-t(A) =
{o00 ifA=.0,
if A =F .0 .
Then f-t is a [plainly degenerate] measure and also outer measure on .9' (X).
(c) The set function identically zero is a degenerate measure on
.9'(X), for any set X.
(d) The most important outer measures for our purposes are the
set functions t constructed as in § 9 for all subsets of a locally compact
Hausdorff space X. See Theorem (9.21) for the proof that t is an outer
measure. The reader should be aware, however, that not all measures
used in analysis are derived from set functions t.
We now begin our construction of measures.
(10.5) Definition [CARATHEODORY]. Let X be a set [no topology]
and f-t an outer measure on .9' (X). A subset A of X is said to be f-t-measur-
able if
f-t(T) = f-t(T n A) + f-t(T n A')
for all TeX. Let ~ denote the family of all f-t-measurable subsets of
X and';y': the family of all subsets A of X such that f-t (A) = O.
(10.6) Remarks. (a) In view of (1O.2.iv) and (1O.2.ii), a subset A
of X is p-measurable provided that f-t(T) ~ f-t(T n A) + f-t(T n A')
for all Tc X.
(b) Definition (10.5) has a somewhat artificial air. It singles out the
subsets A of X for which A splits aU subsets T of X into two pieces on
which f-t adds. How CARATHEODORY came to think of this definition seems
mysterious, since it is not in the least intuitive. CARATHEODORY'S def-
inition has many useful implications. It gives us a (I-algebra, although
128 Chapter III. The Lebesgue integral
(i) p,(T) =n~ p,(T nAn) + p, (T n CQ! AS) for all TeX.
00
Then
n=l
U An = Al U (Aan A;) U (A3 n A~ n A;) U ... U (An n A~_l
n ... n AD U .. '. By (10.10), each set of the form Bn = (An n A~_l
n ... n AD is p,-measurable. Furthermore, the sets Bn are pairwise
disjoint.
Hewitt/Stromberg, Real and abstract analysis 9
130 Chapter III. The Lebesgue integral
Upon setting T =
k=1 U Bli in (1O.9.i), we obtain
00 00
p(Bn) + p(0) =
= ~
..=1 . =1 p(Bn) .
~
p(
n=1UAn) = lim P (An) .
,,-..00
00 00
Proof. Write Ao =
Then clearly U An = U (An
0.
. =1 ..=1 n A~ _ I). By
(1O.3.iv), countable additivity, we have
00
p ( "~IAn
)
= .. f 00
00
In particular, il n An =
n=1
0, then lim I-'(An)
11"""",+00
= O.
(ii) I-' tOI e.gk An)) ~ ~col-' (A,,) provided that I-' tQI A,,) < 00 •
co
n An C n=2
n An C ... cn=k
n An C •... Theorem
00 00
I-'(A,,} ~ I-' CBkAn) for all k. This implies that ~ool-' CB . . An) ~
lim I-'(A,,), from which (i) follows. The inequality (ii) is proved in like
" ...... 00
manner. 0
132 Chapter III. The Lebesgue integral
of positive numbers such that I: IXn < 00. Let I(rp) = I: IXnrp(xn)
for all
n=1 n=1
rp E<roo (X). Prove that: I is a nonnegative linear functional on <roo (X) ; the
00
family of all sets of the form U An' where, for each n EN, either An
n=1
or A~ is an element of .F. Let Q denote the smallest uncountable ordinal
number (4.49). We use transfinite induction to define a family Cf% for
each ordinal number IX < Q. Define Co as the family C. Suppose that
o < IX < Q and that C{J has been defined for each {3 such that 0 ;;;; {3 < IX.
U C{J)*, and write d for the family 0:[,f%<!J
Define Cf% as (0:['{J<f% U Cf%' We
assert that d = f/.
It is clear that tffo = tff c f/. Suppose that C{J C f/ for every {3 < IX
00
1 This theorem is not needed in the sequel and may be omitted by any reader
who is pressed for time; similarly for (10.24) and (10.25).
134 Chapter III. The Lebesgue integral
in .!7. Since .91= U C~, we have proved that .91 C .!7. It is trivial
0;;:;", <.Q
that C C .91, and so to complete the proof we need only to show that
.91 is a a-algebra. Since 0 ECo' we have
X= (0 / U 0 U 0 U ... ) EClCd.
Now let A E d. There is an IX < D such that A EC",; hence A'
= (A U A U ... ) Eiff: C Cp for every {3 > IX, and therefore A 'Ed. Next
I I
the sets U An ECl can be formed [at most 2 e choices for each AnJ,
n=1
we see that Cl ~ (2e)N. = eN. [(4.32), (4.34), (4.24.vii), and (4.24.xi)J.
Now suppose that C{J ~ eN. for all {3 such that 1 ~ {3 < IX, where
1 < IX < D. Then {J<",
U C{J ~ eN. Ko = eN. (4.32), and so, arguing as above,
i. ~ (eN.)N. = eN. [(4.24.viii) and (4.31)J. It follows by transfinite in-
duction that C" ~ eN. for every IX such that 0 ~ IX < D. Therefore
y=;;= a.<!}
U C.t X:::::eN''"l=e N•
- "
The problem of finding non ,a-measurable sets for a given outer meas-
ure ,a can be terribly complicated. Even for outer measures constructed
§ 10. Measures and measurable sets 135
00
x +Ac U Jan
n=!
+ x, bn + x[ , (1)
00
-A c U ]-b nJ -an [
n=l
are mutually equivalent, and since the three unions of intervals in (1)
have the same Lebesgue measure, the lemma is proved. 0
(10.28) Theorem. Let T be a A-measurable subset of R such that
A(T) > O. Then T contains a subset E that is not A-measurable. In fact,
E can be chosen to have the following property. If .s1 is a a-algebra of sub-
sets of R such that .4). C .s1 and x + A E.s1 whenever A E.s1 and x E R
[for example d = .4).], and it fl is a countably additive measure on d
such that fl(A) = A(A) for all A E.4). and fl(x + A) = fleA) for all
A E.s1 and x E R, then E ~ .s1.
Proof. Since T = U (T
n=!
n [-n, n]), (9.2l.iv) shows that
00
Now let D be any countably infinite subset of T (4.15) and let H be the
smallest additive subgroup of R that contains D. That is, H consists
m
of all finite sums}; nkdk' where the nk's are integers and the dk's are
k=l
in D; from this it is clear that H is countable.
Consider the cosets {t + H: t ER}. Since H is a subgroup of R,
these cosets are pairwise disjoint; i. e., for any t1 and t2, the cosets t1 + H
and t2 + H are disjoint or identical. Let {ty: I' Er} be chosen in R so
r,
that the sets ty + H; I' E are all of the distinct cosets of H, i. e., 1'1 =1= 1'2
implies (tYl + H) n (ty, + H) = 0, and for each t ER, there is a I' Er
such that t+H=ty+H. Let To={YEr:(ty+H) n T=I=0}. For
each I' ETo, choose just one element xl' E(ty + H) n T, and let
E = {xl' : I' ETo}. [Since T is contained in U {ty + H: I' ETo} and
ty + H = No, we must have
ro= T> No.J
In finding E we have twice made an uncountable number of arbitrary
choices, and to do this we must invoke the axiom of choice 1.
To prove that E possesses the pathological properties ascribed to it,
define the set J to be H n (T - T). Since D - D c J c H, it is obvious
that J = No· We claim that (Y1 + E) n (Y2 + E) = 0 for distinct
Y1> Y2 EJ. If not, then there are distinct x1> X 2 EE such that Y1 + Xl
= Y2 + x 2· Since Y1> Y2 EH, this implies
x2 = Xl + (Y1 - Y2) EXl + H,
a contradiction to the definition of E, as Xl and X 2 lie in disjoint cosets
of H. Hence the family {y + E : Y EJ} is pairwise disjoint. Now assume
that E is in the a-algebra d. If fl (E) = 0, our hypotheses on fl imply
that
fl (] + E) = fl (U (y + E) = }; fl (y + E)
yEJ yEJ
=};fl(E)=O. (1)
yEJ
If fl (E) > 0, a similar reckoning gives
fl (] + E) = 00 • (2)
Both (1) and (2) are impossible. To see this, we first prove that
TcJ+E. (3)
In fact, if vET, then v Ety +H = xl' + H for some y ETo, and so
1 Every example of a non A-measurable set has been constructed by using the
axiom of choice. A recent announcement by R. SOLOVAY [Notices Amer. Math.
Soc. 12, 217 (1965)] indicates that without the axiom of choice, non A-measurable
sets cannot be obtained at all.
§ 10. Measures and measurable sets 137
Using part (I), choose for each n EN a compact set Fn such that Fn C An
and t(Fn) ~ ! t(An). It is plain from (I) that
lim t(Fn) = lim t(An) = 00 = t(A). 0
n-+-oo n~oo
A n B' = CQIA. ) n Cgl BRr = CQI An) n COl m) c"gl (An n B~) ,
and so
00
Now choose 8> O. There is a function I Em1+ such that I ~ ft~.A + (J~B
and
1(/) -
-
a8
1 -
< I(ft~.A + (J~B) .
Choose 15 > 0 such that
0<!5<min{3'~Uo)' 3'~Vo)
,ft,{J}.
For all x EA, we have I (x) ~ IX. By the lower semicontinuity of I, there
is an open set U such that A cUe Uo and I (x) > ft - 15 for all x E U.
Similarly, there is an open set V such that B eVe Vo and I (x) > {J - 15
for all x E V. Thus we have I ~ (IX - 15) ~u + ({J - 15) ~v; therefore
I (I) ~ I «(IX-!5) ~u+ ({J -15) ~v)= (ft-!5) I (~u) + ({J -15) I (~v)
= (ft- !5)t(U) + ({J- !5)t(V) ~ ftt(A) + (Jt(B) - !5(t(U) + t(V))
2
~ IXt(A) + (Jt(B) - !5(t(Uo) + t(Vo)) > ftt(A) + (Jt(B) - a8 .
Summarizing, we have shown that
-
ftI (~.A) + {JI- (~B) - 8 < I- (I) - a1 8 < I- (IX~.A + (J ~B) .
Since 8 is arbitrary, it follows that (1) holds if A and B are compact.
(II) We suppose now that A and B are arbitrary t-measurable sets
of finite positive measure. Choose 8> O. Using (10.30), choose compact
1
sets E and F such that E C A and FeB, IXt(E) > IXt(A) - 2 8, and
(Jt(F) > (Jt(B) - ~ 8. Using part (I) for the compact sets E and F,
we have
I(IX~.A + (J~B) ~ l(ft~E + (J~F) = ftt(E) + (Jt(F) > ftt(A) + (Jt(B) - 8
= IXI(~.AJ + fJI(~B) - 8. 0
We close this section with a large collection of exercises. A few of
these [for example (1O.37)J are actually needed for subsequent theorems
in the main text. The rest of the exercises illustrate and extend the theory
in various directions, and we trust that all serious readers will work
through most of them.
(10.36) Exercise. Let X be an arbitrary set and d an algebra of
subsets of X [d need not be a a-algebraJ. Let fA be a countably additive
measure on d in the sense of (10.3). Define a set function p, on 9fJ(X)
as follows: for T C X, let
P,(T) = inf{..g fA (An) : T C nQIAn and AI' A!!, ... , An"" Ed} .
(a) Prove that p, is an outer measure on 9fJ(X).
(b) Prove that p, is equal to fA on the algebra d.
(c) Prove that all elements of d are measurable with respect to p,.
142 Chapter III. The Lebesgue integral
(iii) if Av A 2, ... Ed, if Al ::::> A2 ::::> ••• ::::> An ::::> •• " and n An =
.. =1
0,
then lim y (An) = O•
........ 00
Define y just as in (to.36). Prove that (a), (b), (c), and (d) of (10.36)
hold for the set functions y and y. [This is another version of E. HOPF'S
extension theorem.]
(10.38) Exercise. Let (X, d, p) be a measure space. Prove that p
can be extended to a measure p, on a a-algebra d such that every subset
of every set of p-measure 0 is p,-measurable and has p,-measure O.
The following exercise will be needed in the sequel to prove two im-
portant theorems of the main text [(20.56) and (20.57)], and so we
spell out the proof in some detail.
• (10.39) Exercise. Let X, d, p, and p, be as in (to.36). Let .9 = .9 (d)
be the smallest a-algebra of subsets of X that contains the algebra d.
Suppose that (X,.9,,,) is a measure space such that ,,(A) = peA)
for all A Ed. Prove the following.
(a) If BE.9, then P,(B) ~ ,,(B). [Hint. Let .s(. denote the family
00
n and X U F , then ,,(E) = P, (E) for all E E .9, i. e., the extension
n=1 n
=
(d) If the a-finiteness hypothesis in (c) fails, then fl may have more
than one extension to .:7. [Hint. Let X = [0, 1[ and let d be the algebra
°
of all finite unions of intervals of the form [a, b[ C [0, 1 [. Define fl
on d by fl (0) = and fl (A) = 00 if A =l= 0. Show that there are exactly
2c countably additive measures on the Borel sets of [0, 1[ that agree
with fl on d.]
(10.40) Exercise. An outer measure fl on gJ (X) is said to be regular
if for each E C X, there exists a fl-measurable set A C X such that
E C A and fleA) = fleE).
(a) Prove that any outer measure obtained from a measure on an
algebra of sets as in (10.36) is a regular outer measure.
(b) Prove that if X is a locally compact Hausdorff space and l is
as in § 9, then l is a regular outer measure.
(c) Let X = {O, I}. Construct an irregular outer measure on gJ(X).
(d) Let fl be a regular outer measure on gJ(X), let E C X, and let d
be the smallest a-algebra containing {E} U vII,.. Prove that if fl is finitely
additive on d, then E Evil,..
(e) Notice that, since A is a regular outer measure on R, none of the
extensions of A [as a measure on ~] mentioned in (1O.29.b) can agree
with the outer measure A except at sets in ~.
(10.41) Exercise. Let A be a subset of [0, 1] that is not A-measurable,
and let B = [0, 1] n A'. Prove that S (~A + ~B) < S (~A) + S (~B), where 5
is the Riemann integral.
(10.42) Exercise. Let X and Y be topological spaces. Prove the
following.
(a) If t is a continuous function from X into Y and if B E~(Y),
then 1-1 (B) E~ (X). [Consider the family of all sets B for which the asser-
tion holds.]
(b) If A E~(X) and B E~(Y), then A x B E~(X x Y). [Recall
the definition of the product topology (6,41) and use (a).]
(c) Generalize (b) to products of countably many topological spaces.
(10.43) Exercise [H. STEINHAUS]. Let T be a A-measurable set in R
such that A(T) > 0. Prove that the set T - T contains an interval
[-IX, IX] (IX> 0). The following steps may be useful.
(a) If U and V are open in R and have finite A-measure, the
function x -+ A«(X + U) n V) is continuous on R. [Begin with intervals
and use (6.59) for general U and V.]
(b) If A and B are A-measurable of finite A-measure, then
x -+ A«(X + A) n B) is continuous. [For U:::> A and V:::> B, prove first
that
JA«(X + U) n V) - A«(X + A) n B)J ;;i A(U n A') + A(V n B').]
144 Chapter III. The Lebesgue integral
· n (E, t)
extl/> (E) =
11m --(t)- .
ItO f{J
For all E c X, define
where the infimum is taken over all countable, pairwise disjoint families
00
+.
n(B, t), which is valid for t < e(A, B).]
(b) Compute "'1' for X = R with the usual metric for Rand tp (t) =
(c) Compare the outer measures "'1' [with suitable tp!] with Haus-
dorff p-dimensional measures.
(d) Prove that lI!p(E) = 0 if E is countable.
(e) Prove that "'1'(E) = lI.p(F) if there is an isometry of E onto F.
[An isometry is a mapping 1p of one metric space onto another such that
e(x,y) = e'(1p(x), 1p(y)), e and e' being the metrics on the two spaces.]
(10.52) Exercise. Let DC be any real-valued nondecreasing function
on R and let A" be the Lebesgue-Stieltjes measure on R induced by the
Riemann-Stieltjes integral as in § 9. Prove that A" ({x}) = 0 for x ER
if and only if DC is continuous at x.
(10.53) Exercise. Let DC and A" be as in (10.52) and suppose that DC
is continuous. Prove the following assertions.
(a) For each e > 0 there exists a nowhere dense perfect set A C [0,1]
such that A,,(A) > A,,([O, 1]) - e.
(b) There exists an Fa set Be [0, 1] such that B is of first category
and A,,(B) = A,,([O, 1]).
(c) There exists a Gd set of second category contained in [0, 1]
having A,,-measure zero.
(10.54) Exercise. In this exercise, we first sketch the construction of
a subset B of R measurable for no measure A" with continuous DC.
(a) Prove that every uncountable closed subset F of R has cardinal
number c. [Use (6.65) and (6.66).]
(b) [F. BERNSTEIN]. Prove that there is a subset B of R such that
B n F =1= {1) and B' n F =1= {1) for every uncountable closed subset F
of R. [Hints. There are just c open subsets of R and hence just c uncount-
able closed subsets. Let We be the smallest ordinal number with correspond-
ing cardinal number c [use (4.47) to show that We exists]. Let {I;: 1j< we}
be a well ordering of all uncountable closed subsets of R. Define B by
transfinite recursion and the axiom of choice, as follows. Let Xo and Yo
be any two distinct points in Fo. Suppose that Xy and yy have been
defined for all 'Y < 1j, where 1j < We' The set A'1= {Xy: 'Y < rJ} U {yy: 'Y < rJ}
has cardinal number < c, because We is the smallest ordinal number of
cardinal c. Hence the set I; n A~ has cardinal number c. Let x'1 and Y'1
be any two distinct points in the set I; n A~. Finally let B={x'1: rJ < we}'
It is clear that B n I; =1= {1) and that B' n I; =1= {1) for allrJ < we.]
(c) Prove that B is non A,,-measurable if DC is continuous and A" =1= O.
[Hints. Assume that B is A,,-measurable. Then by (10.30), we have
A,,(B) = sup{A,,(F) : F is compact, FeB}. The only compact subsets
of B are countable, and since A,,({X}) = 0 for all x ER, it follows that
§ 10. Measures and measurable sets 147
an -+ a. Then we have 1- 1([a, 00]) = n 1-1 (Jan, 00]). To see that (iii)
n=1
implies (iv), observe that 1- 1([-00, aD = (t-1([a, 00]))'. The proof that
(iv} implies (v) is similar to the proof that (ii) implies (iii). It remains
only to show that (v) implies (i). For a E R, choose a strictly decreasing
sequence (bn ) in D such that bn -+ a. Then
!
(iii) Let
II (x) IP il I (x) is finite,
h (x) = p_ il I (x) = - 00 ,
p+ il I(x) = 00 ,
where p_ and p+ are arbitrary extended real numbers and p is any positive
real number. Then h is .!ii-measurable.
!
(iv) Let
r
[f (x) il I (x) is finite,
hex) = p_ il I(x) = -00,
p+ il I (x) = 00 ,
where m is a positive integer and p_ and p+ are arbitrary extended real
(v) Let h = +
numbers. Then h is .!ii-measurable.
where I is finite and not zero, and let h assume constant
but arbitrary values p+, p_, and Po on the sets {x EX: I (x) = oo},
152 Chapter III. The Lebesgue integral
~ 1::00
if t = - 0 0 ,
W(t} if t ER,
if t = 00,
if IX> 0, and
if t=-oo,
p(t}~l~oo if t ER,
if t = 00,
if < 0; if IX = 0, the assertion is trivial.
IX
For (iii), let ep(±oo) = {3± and ep(t) = ItIP for real t. Since ep is con-
tinuous on R, it is clear that ep-l ([a, 00 J) n R is a Borel set for all real a.
The proof of (iv) is similar, with ep (t) = tm for realt and ep (± 00) = {3±.
To prove (v), let ep (t) = t1 for t =l= 0,00, - 00, and let ep (0) = {3o,
ep(±oo)={3±. D
(11.9) Lemma. Let I and g be d-measurable. Then the sets
(i) {x EX: I (x) > g(x)} ,
(ii) {x EX: I (x) ~ g(x)} ,
and
(iii) {x EX: I (x) = g(x)}
are in d.
Proof. We have
{xEX: I (x) >g(x)} = u~Q({x EX: I (x) > u} n {x EX: g(x) < u}),
and from this identity the measurability of (i) follows. The set (ii) is the
complement of the set (i) with the roles of I and g interchanged, and so
it too is measurable. The set (iii) is the intersection of two measurable
sets of type (ii) , and so is measurable. D
(11.10) Theorem. Let I, g be d-measurable. Let h (x) = I (x) + g (x)
lor all x EX such that I (x) + g (x) is defined and let h have any fixed value {3
[an extended real numberJ elsewhere. Then h is d-measurable.
Proof. For any real number a, we have
h- 1 (Ja, ooJ) = {x EX: I (x) + g(x) > a} U AfJ
= {x EX: I (x) > a - g (x)} U A fJ '
§ 11. Measurable functions 153
!
where
({X: I (X) = oo} n {X: g(X) = - oo})
AII= u ({x: I (X) =-oo}n{X:g(X) =oo}) if a<{J,
o if a~{J.
The set {x EX: I (x) > a - g (x)} is in .91 by (11.9); therefore, since All
is also in .91, the set h- 1 (]a, 00]) is in d. 0
(11.11) Theorem. Let I and g be d-measurable lunctions. Let h be
defined on X by
_ {/(X) g(x) il x ~ A
h (x) - {J il x EA ,
where {J is an arbitrary extended real number, and A = {x EX: I (x) = 00
and g(x) =-oo}U{xEX:/(x) =-00 and g(x)=oo}. Then h is an
d-measurable lunction.
Proof. Consider a ER such that a > 0, and let
A ={A if a<{J,
II 0 if a~{J.
We have h- 1 (]a, 00]) = {x EX: h(x) > a} = All U {x EX: I(x) = 00 and
g(x) > O} U {x EX: I(x) > 0 and g(x) = oo} U {x EX: I(x) < 0 and
g(x) = -oo} U {x EX: I(x) = -00 and g(x) < O} U {x EX: I(x) and
g(x) are finite and ~ [(f(x) + g(x))11 - (f(x) - g(X)2] > a}. Applying
(11.4), (11.8) and (11.10), we see thath- 1 (]a, 00]) Ed. Similar expressions
hold for a < 0 and a = 0, and it follows that h is d-measurable. 0
We next study limits of sequences of measurable functions.
(11.12) Theorem. Let (In) be a sequence old-measurable lunctions
defined on X. Then the lour lunctions inf In, sup In' lim In' and lim In
n n "---+-00 11---+-00
.
[recall that - ( 00) = - 00 and -
.
ability of inf In follows at once from the identity inf In (x) = -
..
sup(-In (x))
(- 00) = 00]' The rest follows from the
first two results and the identities
lim In (x) = sup ( inf In (x))
"->-00 k .. Oi:;k
and
lim I.. (x) = inf (sup I,,(x)). 0
"-+00 k .. Oi:;k
154 Chapter III. The Lebesgue integral
+,
all 01 the lollowing lunctions are d-measurable on X: I + IX,' IX/,' ItIP;
1m; il I(x) =!= 0 lor all x EX,' I + g,' Ig.
Proof. These results all follow immediately from Definition (11.15)
by applying (11.8), (11.10), and (11.11). 0
§ 11. Measurable functions 155
() _{/,,(X) if xEA,
g" x - 0 if x EA',
and define
if x EA,
g(x) = {~(X)
if X EA'.
Theorem (11.23) implies that g" is d-measurable for all n EN. Clearly
g,,(x)-+g(x) for all xEX. Applying (11.14) or (11.18), we see that g
is d-measurable. Again by (11.23), I is d-measurable. 0
Mathematical analysis is heavily concerned with convergence of
sequences and series of functions. Indeed, a main goal of analysis is the
approximation of complicated functions by means of simple functions.
[The terms "approximation", "complicated", and "simple" have dif-
ferent meanings in different situations.] Up to this point we have met
in this book two kinds of convergence: pointwise [almost everywhere]
and uniform. We now introduce a third kind of convergence and prove
some theorems that show a number of relationships among these three
kinds of convergence.
(11.25) Definition. Let (X, d, #) be a measure space and let I and
(1,,)':=1 be d-measurable functions on X. They may be either extended
real- or complex-valued. Suppose that for every «5 > 0, we have
lim #({x EX: I/(x) -I,,{x)l ~ «5}) =
,,->-00
o.
Then (I,,) is said to converge in measure [or in probability] to I. We write:
I" -+ I in measure.
(11.26) Theorem [F. RIEsz]. Let (X, d, #) be a measure space and let I
and (I,,) be d-measurable lunctions such that I" -+ I in measure. Then there
exists a subsequence (I..~) such that I..~ -+ I #-a.e.
§ 11. Measurable functions 157
1
Proof. Choose nl EN such that ",({x EX: I/(x) - In, (x)l ~ I}) < 2'
Suppose that nl , n a, •.• , n,. have been chosen. Then choose nk+1 such
that nk+l> n,. and
15 and e, there exist a set lEd and an integer no E N such that p (]/) < e
and II (x) - I" (x)l < 15 lor aU x EI and n ~ no.
Proof. Let E = {x EX: I (x) is finite, I" (x) is finite for all n EN,
I" (x) --+-1 (x)}. By hypothesis, p (E') = O. For each mEN, let Em = {x EE:
ex)
I/(x) - 1,,(x)1 < 15 for alln ~ m}. WehaveE1 C E2 C," and m=!
U Em=E.
00
Therefore E; :> E~ :> ... and n E:" = E'. Since p(E;) ~ p(X) < 00, it
m=!
follows that lim p (E:") = p (E') = O. Thus choose no EN such that
~oo
2: = e.
Also n ~ nm implies that
sup lI(x) - 1,,(x)1 ~ sup I/(x) - 1,,(x)1
xEA xE./...
:5:~
- m
for every mEN. Thus I" --+- I uniformly on A. D
(11.33) Caution. Theorems (11.31) and (11.32) depend heavily on
the hypothesis that p(X) < 00. For example, suppose that X = R,
d = u8J., P = A, I" = ~[".n+!l' and 1=0. Then I,,(x) --+- I (x) = 0 for all
x ER. But A({X E R: I/(x) - 1,,(x)1 ~ I}) = A([n, n + 1]) = 1-++0, and
so I" -++ I in measure. Also, if A Eu8J. and A(A/) < 1, then for each nEN
there exists x" E An en, n + 1], and so II (x,,) - I,,(x,,) I = 1, i.e., I" -++ I
uniformly on A.
§ 11. Measurable functions 159
such that t({x EX: I (x) =l= g(x)}) < B. Moreover il 11/11 .. < 00, then g can
be selected so that Ilgll .. ~ 11/11 ...
Proof. Let B > 0 be given. Use (9.24) to select an open set U such that
E e U and t(U) < teE) + : . The set U is fixed throughout the proof.
(I) Suppose that I = ~.A' Then AcE and A is t-measurable. Since
teA) < 00, we may apply (10.30) to find a compact set F such that
Fe A and teA n F') < ; . Next use (9.24) and (6.79) to produce an
open set V with compact closure such that Fe Ve U andt(VnF') < ~ .
Use (6.80) to obtain a continuous function g from X into [0, 1] such that
g(x) = 1 for all x EF and g (x) = 0 for all x E V'. Then we have: gE<roo(X),
g = 0 on U' , and {x EX: I (x) =l= g(x)} e (V n F') U (A n F'). It follows
that
t({x EX: I (x) =l= g(x)}) < ; + ; = B,
Clearly s" (x) = g" (x) for all x EA I and all n EN. Thus g" (x) -+ I (x)
for all x EA I. Since t (U n A') < 00, EGOROV'S theorem (11.32) shows
that there is an t-measurable set B e UnA I such that t (U n A 'n B') < :
and g" -+ I uniformly on B. Next use (10.30) to select a compact set
§ 11. Measurable functions 161
and Jl (Y) = inf {Jl (U) : U is open in the topology of D, U ~ Y} for all
Y Ef!l(D). That is, CD, f!l(D), Jl) is irregular, although "outer" regular.
n
[Hints. Construct a subset D of [0, 1] such that D F =1= 0 =1= D' F n
for every uncountable closed set Fe [0, 1]: see (1O.54.b). Then D is
°
non A-measurable, and so < A(D) ~ 1. Let X be a G6 set such that
Dc X c [0, 1] and A(X) = A(D). Then every A-measurable subset
of X n D' has A-measure 0, and we may take Jl to be ().(~) Af as in
part (a). All of the claims made for Jl are easy to verify. Compare this
result with (10.58).]
(11.40) Exercise: Extending a measure. Let (X, .91, Jl) be a measure
space. Let Y be a subfamily of f!P (X) such that:
(i) P E Y n .91 implies Jl(P) = 0;
00
which L (I) is the supremum. [Actually the equality L (I) = S (I; [0, 1])
holds: see (12.51.£) inlra.]
n
(12.4) Theorem. Let IE 16+, say I= };a.k~Ek' where the Ek'S are
k=1
n
pairwise disjoint and in .!if. Then L (I) exists and L (I) = } ; a.k '" (E k) .
k=1
Proof. We may suppose that the pairwise disjoint family {E k }%=1
n
covers X. Since inf{/(x): x EE k} = a.k, we have L(I) ~ };a.k",(Ek).
k=1
Now let {Bi}~1 be an arbitrary measurable dissection of X. Then
m m n
}; inf{/(x) : x EB i } ",(Bi ) = }; };inf{/(x): x EB i } ",(E k n B i )
;=1 i=1 k=1
m n
~}; };inf{/(x):xEEknBi}",(EknBi )
i=1 k=1
m n
= } ; } ; a.k ",(E k n B i)
i=1 k=1
n
= }; a." '" (E k ) •
k=1
1=1
ft
... "
L (I + g) = I: I: (ex; + Pk) ft (Ai n B lI )
;=1 k=1
m ,.. II m
= I: I: ex; ft (Ai n Bk) + I: I: Pk ft (Ai n B lI )
~1~1 ~1~1
= L (I) + L (g). 0
(12.8) Theorem. II I;;;; 0 and t ER, then we have L(tl) = tL(I).
The proof is easy and is omitted.
Our immediate aim is to establish the extremely important identity
Next suppose that I' < LE(f), and let {B k }Z'=l be any measurable
dissection of E such that
m
I' < J; inf{t(x): x EB k } p.(B k ) •
k=l
We have
m
I' < J; inf{t(x): x EB k } p.(B k ) + o· p.(E')
~ L(t) ,
k=l
and so LE (I) ~ L (I). The assertion for arbitrary functions follows im-
mediately. 0
The following result, which looks harmless enough, is the key to
the proof that L is countably additive.
(12.10) Theorem. Let (g,,):=l be any nondecreasing sequence in e+.
5uppose that h E e+ and that lim g" ~ h. Then we have lim L (g,,) ~ L (h).
~oo n~oo
It thus suffices to prove the theorem under the assumption that 1'1> o.
Case (I): p. (X) and I'm are finite. For any () > 0, choose e > 0 satisfy-
ing the inequality
e ~ min {2.u~X) , 1'1} .
For every positive integer n, let 5" = {x EX: g,,(x) > h(x) - e}. Since
00
lim g.. ~ h, we have X = U 5... The sequence (g,,) is nondecreasing,
~oo "=1
and so the sequence 51> 52' ... is nondecreasing. From these facts and
the countable additivity of p., we find (10.13) that
lim p. (5..) = p. (X)
,,-+00
and that
lim p.(5~) = 0. 1
n-+OO
Case '(III): I-' (Em) is positive and I'm = 00. Here we have L (h)
?;, I'm I-' (Em) = 00. Choose any real number I' > I'm-I, and let
h" = I'l~El + ... + I'm-I~Em-l + I'~Em' By Cases (I) and (II), we have
lim L (gn) ?;, L (h,,) ?;, 1'1-' (Em). Since I' can be arbitrarily large, it follows
n-->-oo
that lim L (gn) = 00 = L (h).
n-->-oo
m-I
Case (IV): I'm = 00 and I-'(Em) = O. Here we have L (h) = L: I'jl-' (E j).
i=1
Let B = E1 U E2 U ... U Em_I' Then
m-I
gn ?;, gn~B and lim gn~B ?;, h~B =
n-)-oo
L:
1=1
I'j~Ej'
Since I'm-I < 00, Case (I) or Case (II) applies to (gn~B) and h~B' so that
lim L(gn) ?;, lim L(gn~B) ?;, L(h~B) = L(h). D
n-+oo n-?OO
Proof. Let lim gn = cp, and let I' be any real number such that
n->oo
I' < L (cp). There exists a measurable dissection {AI' A 2 , ••• , Am} of X
such that
and
00 • pCB) + inf{/-(x) : x E B'} pCB') ~ L(n < 00.
explanatory.
(12.19) Theorem. Let I E ~~, and let 1=11 - 12' where 11 ~ 0, I. ~ 0
and II' I. E ~~. Then
J I dp. = J 11 dp. - fl. dp. .
x x x
§ 12. The abstract Lebesgue integral 171
x x x
r
J 11 dp + J dp = J 1+ dp + J 12 dp. 0
x
(12.20) Theorem. For I, g E ~~ and IX, fJ E R, we have
J (IXI + fJg) IX J I dp + fJ J g dp .
dp =
x x x
That is, the mapping I--+- J I dp is a linear lunctional on ~~.
x
The proof is easy and is omitted.
(12.21) Theorem [LEBESGUE]. Let (In) be a sequence 01 nonnegative,
extended real-valued, d-measurable lunctions on X. Then
1: Ii ~ 1: Ii;
00 '"
;=1 ;=1
therefore
f(£li)d PG f(#/i)dP=;~llidP'
x x
and consequently
x
f ( i Ii) dp il
~ j Ii dp .
11 + 12 + ... + 1m ~ lim gt ~
/1-+00
1: Ii
;=1
for each m. Taking the limit with respect to m, we obtain
172 Chapter III. The Lebesgue integral
f( f i= I
Ii) dp = lim J gk dp;;;;
k--->oo X
lim
k--->oo X
J (11 + 12 + ... + Ik) dp
X
k 00
= lim I: J Ii dp = I: J Ii dp. 0
k--->ooi=IX i=IX
(12.22) B. LEVI'S Theorem. Let (lk)k=1 be a nondecreasing sequence 01
extended real-valued, d-measurable lunctions on X such that J fk dp < 00
x
lor some k. Then
gk
(x) = {/Hdotherwise.
00
Ik(x)
X) - if Ik(x) < 00,
II lim In(x) exists lor ft-almost all x E X, then lim fin dft exists and
~oo fl.---+OO X
Proof. It is obvious that all In are in ~'i. Hence all In' and s, are finite
a.e. on X. Let A = {x EX: In (x) is ± 00 or lin (x) I > s(x) for some n EN},
let B = {x EX: In (x) is undefined for some n EN}, and let C = {x EX:
s(x) is infinite or is undefined}. Let In (x) = s(x) = 0 on AU B U C.
Since ft (A U B U C) = 0, (12.14) shows that none of the integrals appear-
ing in the statement of the theorem has been changed by this definition.
Furthermore we have lin (x) I ~ s (x) < 00 for all n EN and all x EX.
The sequence (s + Inr:~l consists of nonnegative functions. Applying
FATou's lemma (12.23), we find
Thus (i) holds. [The reader will note that the function lim In occurring
n-->oo
in (i) is defined only a. e., but is equal a. e. to the function lim In with
n-->oo
the In defined everywhere as in the proof. ] The inequality (ii) is proved
in like manner, starting with the sequence (s - Inr:~l and using the
equality lim IXn = - lim (- IXn).
~oo ~oo
Then fin dA = n' ~ = 1 for all n EN while lim In(x) = 0 for all
R n n-+oo
Proof. Let I (x) = lim In (x) whenever this limit exists. Clearly I
..-..00
J I dl-' =}; J I dp .
A 11=1 A ..
= lim }; J I dl-'
..-..00 k=1 A)
00
=}; J Idl-" 0
k=1 Ai
lunctions on X such that}; 11101 E~l [or, equivalently, }; J 1/101 dl-' < 00].
Then =
X X
Proof. Exercise.
176 Chapter III. The Lebesgue integral
lim I
n-+oo X
1pn dp = I lim 1pn dp
X n-+oo
= I
X
III dp .
where A". k = {x EX: k 2" 1 ~ I (x) < : .. } and B" = {x EX: I (x) ~ n}.
By (10.35) and (12.4), we have
It follows from (10.20), (9.27), (9.24), and (9.26) that every measure t
defined as in § 9 is a regular measure on Jt.. [Cf. the different but
related definition of regular outer measure in (10.40) supra.]
(12.40) Theorem. Let fl be a regular measure defined on a a-algebra d
01 subsets 01 a locally compact Hausdorff space X. Then
fl(A) = sup{fl(F) : F is compact, Fe A}
lor every A E d that is a-finite with respect to fl.
Proof. Repeat verbatim the proof of (10.30) with t replaced by fl. 0
(12.41) Theorem. Let X be a locally compact Hausdorff space and
let fl and v be regular measures defined on a-algebras Jt,. and Jt" respectively.
Suppose that f I dfl = f I dv lor all IE <rto(X). Then fl(E) = v (E)
x x
lor all E EJt,. n Jt".
Proof. Let F be any nonvoid compact subset of X. Use (12.39.ii)
to find sequences (Un) and (v,,) of open sets containing F such that
fl(UI) < 00, v(v;.) < 00, fl(Un) -7 fl(F), and v (v,,) -7 v(F). For each n EN,
n
set Wn = n
k~1
(Uk n Vk). Then each Wn is open, WI:::> W 2 :::> ... :::> F,
V(Wn) -7 v(F), and fl(Wn) -7 fl(F). For each n EN, use (6.80) to obtain
a function In E <rto(X) such that In (X) C [0, 1], In (F) = {I}, and In(W~)
C {o}. Next let gn = min {II' ... , In}. It is clear that gn E <rto (X), that
gl ~g2 ~"', and that !~gn(x) = ~F(X) for all x E COl Wn nF')',i.e.,
f gl dfl ~ fl(WI) <
fl-a.e. and v-a.e. Since 00 and f gl dv ~ v(WI) < 00,
x x
Theorem (12.24) applies to yield
Proof. Let E be in .91 and suppose that ft (E) < 00. Since ft is regular,
there exist sequences of sets (Fn) and (Un) such that Fn C E C Un'
each Fn is compact, each Un is open, ft(Fn) ->- ft(E), and ft(Un) ->-ft(E).
00 00
Let A = U Fn and B =
n=!
n Un" Then A and B are Borel sets, AcE CB,
n=!
ft(A) = ft(E) = ft(B), and ft(B n A') = O. It follows from (12.35)
and (12.41) that leA) = l(B) and l(B n A') = O. Therefore, since l
is complete, E = AU [E n (B n A')] E 1.. This argument proves:
for all IE <roo (Y). The measure'll is called the image 01 the measure fl
under the continuous mapping cpo
(12.46) Theorem. Notation is as in (12.45). For all a-finite 'II-measur-
able subsets B 01 Y, we have
(i) v(B) = fl(cp-l(B)) = f ~B 0 cp(x) dfl(x).
x
For every lunction IE -S:l (Y, JI", 'II), 10 cp is in -S:l (X, ~, fl) and we have
(ii) f I (y) dv (y) = flo cp (x) dfl (x).
y x
Proof. We proceed in steps. Suppose first that U is a nonvoid open
subset of Y. Let ~ be the set of all functions I in <rto (Y) such that
I ;£; ~u· URYSOHN'S theorem (6.S0) implies that sup {I: I E ~} = ~u.
§ 12. The abstract Lebesgue integral 181
B=CQ1Fn)UP, (5)
where Fl C F2 C ... C Fn C . ", each Fn is compact, P is disjoint
00
f
y
sn(Y) dv(y) ~ f
y
I(Y) dv(y) < 00 • (7)
Write Sn = I'" (Xk~Bk' where 0 < (Xl < ... < (Xm' Then V(Bk) IS finite
k~l
for all k, as (7) proves, and (12.4) and (i) imply that
m m
f sn(Y) dv(y) = I (X"v(B,J = I (X" ,u(q;-l(B,,))
Y k~l k~J
and let
An = {x EX: IX + (n - l)n(f3 - oc) ~ I(x) ~ p} .
The Lebesgue sums lor I are defined as the numbers
Next suppose that fl{X) < 00. Let I be any bounded, real-valued,
d-measurable function on X. Define Sn as above. Prove that lim Sn = JI dfl.
n--+oo X
(12.49) Exercise. Let (X, .91, fl) be any finite measure space and let
(tn):~1 and I be complex-valued, d-measurable functions on X such that
In -+ I a. e. Suppose that there exists PER such that I/nl ~ Pa. e. for all
n EN. Use EGOROV'S theorem, not (12.21)-(12.24) or (12.30), to prove
that
lim J In dfl = f I dfl .
n~oo X X
(12.50) Exercise. Let (X, .91, fl) be a a-finite measure space. Let g
be an d-measurable function on X such that I g E ~1 for every I E~1. Prove
that there is a number IX E R such that fl({x EX: Ig(x)l > IX}) = o.
Give an example to show that this conclusion may fail if the hypothesis
of a-finiteness is dropped.
(12.51) Exercise. Let - 00 < a < b < 00 and let I be any bounded
real-valued function on [a, bJ. For each 15 > 0 and x E [a, b], define
m.,(x) = inf{/(t) : t E [a, b] n ]x - 15, x + !5[}
and
M,,(x) = sup{/(t) : t E [a, b] n ]x - 15, x + !5[} ,
and define
m(x) =limm6(x), M(x) = lim Mb(X) .
6-}0 "to
(a) Prove that I is continuous at x if and only if m(x) = M(x).
184 Chapter III. The Lebesgue integral
Prove that:
(b) if x E [a, b] and x is distinct from all x~), then ~im CPi(x) = m(x)
1--+00
and Fm "Pi (x) = M (x) ;
1--+00
(c) m and M are Lebesgue measurable on [a, b];
(d) if L (I, Llj) and U (I, Ll i ) are the lower and upper Darboux sums
respectively [defined with IX (x) = x] for the function I corresponding to the
subdivision LI j, then
b
~im
1-'J-00
L(I, Ll i ) = f
a
m(x) dx
and
b
lim U(/, Ll i )
1~OO
=
a
f M (x) dx;
where 5 denotes the Riemann integral (8.6). [The reader should note
that the foregoing applies only to bounded functions on finite intervals.]
(12.52) Exercise. Find a real-valued function I on [0, 1] such that I
is continuous on ]0, 1], lim f I dA is finite, and the Lebesgue integral
d.j.O [d, I]
of lover [0, 1] is not defined.
(12.53) Exercise. Find a bounded, real-valued, Lebesgue measur-
able function Ion [0, 1] such that f 1/- gl dA > for every Riemann
[0, I]
°
integrable function g on [0, 1].
(12.54) Exercise. Let - 00 < a < b < 00 and let I be a Lebesgue
.,
measurable function on [a, b] such that f
a
I(t) dt = °for every x E [a,b].
Prove that I = 0 A-a. e. [This exercise is needed for the proof of Theorem
§ 12. The abstract Lebesgue integral 185
b
(16.34) inlra. Hints are as follows. Since J I (t) dt exists and is finite,
a
III is in ~l([a, bJ, vA)., A). It is evident from (6.59), (12.32), and our hy-
pothesis that J I dA = 0 for all open subsets U of [a, b]. Use (9.24) to infer
u
that J I dA = 0 for all A-measurable sets A C [a, b]. From this the identity
A
1= 0 A-a.e. is immediate.J
(12.55) Exercise. Let X be a locally compact Hausdorff space such
that every open subset of X is a-compact. Suppose that I-' is a measure
defined on 1!4 (X) such that I-' (F) < 00 for all compact sets F C X.
Prove that I-' is a regular measure. [First consider the case that X is
compact. Consider the family &f = {E E I!4(X) : I-'(E) = inf{I-'(U) : U
is open, U:::> E} and I-'(E) = sup {I-' (F) : F is compact, Fe En.J
(12.56) Exercise. Let I-' be a measure defined on I!4(R) such that
1-'([0, IJ) = 1 and I-'(E + x) = I-'(E) for every E E I!4(R) and x E R.
Prove that I-'(E) = A(E) for all E EI!4(R). [First prove that I-'({x}) = 0
for all x ER. Next show that I-'(Ja, b[) = b-a for all a < bin R. Use
(12.55)J.
(12.57) Exercise. Let (X, .91, 1-') be an arbitrary measure space and
let I and (In)':=1 be complex-valued measurable functions on X. Suppose
that In --+ I in measure and that there exists a function g E ~~ such that
1/.. 1 ~ g a.e. for all n EN. Prove that lim J II - Inl dl-' = o. [Assume
n~oox
that lim J II - Inl dl-' = > 0 and choose a subsequence (Ink) such that
n~oox
Q(,
Prove that for each Q(, E [0, I-'(B)J there exists a a-compact set A C B
such that I-' (A) = Q(,.
(12.60) Exercise. Let (X, .91, 1-') be a measure space as described in
(10.56.a). Let I be an d-measurable function on X. Prove that I is con-
186 Chapter III. The Lebesgue integral
stant on each E" except for a set of ,u-measure zero, and accordingly
write the integral J I d,u as a certain finite sum.
x
(12.61) Exercise. Let (X, .91, ,u) be a measure space that is degenerate
in the sense of (10.3): ,u (A) = 0 or ,u (A) = 00 for all A Ed. Show that
every function I E ~l (X, d,,u) vanishes except on a set of ,u-measure
zero and that Jill d,u = O. [This unpleasant property justifies the term
x
"degenerate" for the measure spaces under consideration.]
(12.62) Exercise. Let X be a locally compact, u-compact Hausdorff
space and let ,u be a regular measure defined on au-algebra .91 of subsets
of X [.91:::> gj(X)J. Let I be an d-measurable function on X such that
I(X) C [0, ooJ and such that I~F E ~l(X, d,,u) for all compact subsetsF
of X. [Such an I is called locally ,u-integrable.J Define the set-function"
ondby
,,(A) = J I (x) d,u(x).
A
(b) Let ibe a function in ~l (X, d, P). Prove that there is a function I
in ~l (X, .91, ,u) such that
J 1/- iI dp = 0
x
and
fld,u= fidp.
x x
[It suffices to consider nonnegative functions l By adding sets of ,u-
measure 0, it is easy to make the simple functions s.. of part (a) into a
nondecreasing sequence. By (12.22) we then find
x
J i dp = .......
lim J s.. dp = lim f s.. d,u = J I d,u.
00 x .......00 x x
The other equality is obvious.]
(c) Let X be a locally compact Hausdorff space and (X,.,(, t)
as in §§ 9 and 10. Suppose that X is u-finite with respect to t. Let gj (X)
§ 12. The abstract Lebesgue integral 187
y y
For any positive real number p such that p =l= 1, define p' = P~l
[thus ; + ;, = 1] .
(13.3) Corollary. For p > 1 and a and b any nonnegative real numbers,
we have
(i)
Ilfllp ~Igll,,' f [Ig[ d,u ~ p II~II~ f [tiP d,u + P' II~II~: f [g[P' d,u = ~ + ;, = 1,
and this proves (ii). The inequality (iii) is immediate. D
For p = p' = 2, the inequality (ii) is called CAUCHY'S inequality,
or SCHWARZ'S inequality, or BUNYAKOVSKiJ's inequality; sometimes the
three names are listed together.
(13.5) Conditions for equality in (13.4). To get equality in (I3.4.ii),
it is clearly necessary and sufficient that we have
II (u) I Ig(u)1 = ~ I/(u)lp +~ Ig(u)I'"
~/II.. IIglil" P 11/11: p' IIgll::
§ 13. The spaces ~p(1 ~ P< 00) 191
for almost all uEX. By (13.3), this happens if and only if 111;:1; = I\!II~;
almost everywhere. Thus equality obtains in (13.4.ii) if and only if there
are nonnegative real numbers A and B, not both zero, such that
A ItIP = B IgIP'
almost everywhere.
The reader can easily formulate from this and (12.29) a necessary
and sufficient condition that equality hold in (13.4.iii).
(13.6) Theorem [HOLDER'S inequality for 0 <p < 1]. Let 0 < p < 1
and let I and g be lunctions in ~t and ~t" respectively. Then we have
1 1
(i) fig dp. ~ (J IP dp.)p (J gP' dp.)P'"
unless f gP' dp. = 0 [note that P' < OJ.
Proof. In the case we are concerned with, we have 0 < f gP' dp. < 00,
and since P' < 0, this implies that g (x) > 0 for almost all x EX. Let
1 1 1
1 -- - - , ,
q = -, and define q; = g q and 1j! = g q I q • It is easy to see that q;q = gP •
P
and so q; E~q" If fig dp. = 00, then (i) holds trivially. Otherwise Ig
is in ~l' and so 1j! is in ~(l' Applying (13.4) with p replaced by q. we have
1 1 1
f IP dp. = f q;1j! dp. ~ (J 1j!q dp.F (J q;q'dp.)7 = (f Ig dp.)P (f gP' dp.)7.
It follows immediately that
1 1
fig dp. ~ (J IP dp.)p (J gP'dp.f7P •
and since - q~p = ;, • the theorem follows. 0
(13.7) Theorem [MINKOWSKI'S inequality]. For 1 ~ P< 00 and
I, g E~p. we have
(i) III + gllp ~ 1I/IIp + Ilgllp .
Proof. Suppose first that p > 1. We have
II + glP ~ (It I + IgJ)P ~ [2 max {III. Igl}JP
= 2Pmax{lfIP. IgIP} ~ 2P(lfIP + IgIP) .
This crude estimate shows that II + glP E~l' i.e .• 1+ g E~p, Thus (13.4)
implies that
11/+ gll~ = f II + glP dp. ~ f II + glP-l III dp. + f II + glP-1lgl df'
1 1 1 1
~ (J IflPdp.)P (J II + gl(P-l)P' dp.)P'" + (J IgIPdp.)p (J 1/+ gl(P-l)P' dp.)P'"
.P..
= (llllIp + IIgllp) 11/+ gill' •
192 Chapter IV. Function spaces and Banach spaces
The inequality
p_..P...
III + gllp P' ~ 1I/IIp + Ilgllp
thus holds. Observing that p - :' = 1, we obtain MINKOWSKI'S inequality
for p > 1. Since f II + gl dfl ~ fill dfl + f Igl dfl, the inequality is
trivial for p = 1. 0
We now give conditions for equality in MINKOWSKI'S inequality.
(13.8) Theorem. For p = 1, we obtain equality in (13.7.i) il and only il
there is a positive measurable lunction e such that
< n" < ... and I; 11I"A:+1 - I",tllp = at: < 00. This is possible: e.g., we can
k=1
select increasing n,,'s such that 111m - l"A:llp < 2-" for all m ~ n". Now
define
g" = 1/... 1+ 1/...- 1...1+ ... + 1/"A:+1- I..,tl, for k = 1,2,3, ....
It is clear that
r
Ilgtlll = Ilg"lI~ = (111/...1 + 1/..,- 1...1 + ... + I/n,t+1- In,tlllp)P
~ (IItn.llp +;~ 11/ +1- In;llp
1IJ ~ (1I/n.llp + at:)P < 00 •
Let g = lim g". By B. LEVI'S theorem (12.22), and the above estimate,
k~oo
we have
J gP dp. = J lim k~oo
g~ dp. = lim
k ..... oo
J g~ dp. < 00 •
The nonnegative integrand above must be finite p.-a.e., and so the series
00
also converges p.-a. e. The ktl' partial sum of this series is I"l:+1 (x), and
so the sequence (lnA:(X))'k=1 converges to a complex number I(x) for all
x EA, where A Ed and p.(A') = O. Define I(x) as 0 for all x E A'. It
is easy to see that I is d-measurable, and obviously I is complex-valued
onX.
We will show that I is the limit in S!.p of the sequence (In), and this
will of course prove that S!.p is complete in the metric induced by the ~p
norm. Given e > 0, let l be so large that
III. - Itllp < e for s, t ~ n, .
Then for k ~ land m > n l , we have
111m - InA:IlP < e.
By FATOU'S lemma (12.23), we have
J II - Iml Pdp. = J klim
.....
I/"A: - Iml Pdp. oo
is in ~p; and
lim
,,-+00
III - Inllp = O. 0
( ,f Ix; - Y;IP)P.
1=1
For p= 2, we
obtain the classical Euclidean
metric. The topologies induced by
the lp metrics on Kn and Rn are
all the same [d. (6.17)].
The first quadrant of the unit
balls in R2 for various values of p
are sketched in Fig. 7.
o (13.14) Examples. The spaces
~p([O, 1]) and ~p(R), where 0< p
Fi&. 7. < 00 [it is understood that I' = A.
§ 13. The spaces ~p(1 ~ P < 00) 195
and d = ..-HAJ, are very important function spaces in both pure and ap-
plied analysis.
(13.15) Example. For p = 2, we obtain the famous function space
~2(X, d, fl). In this case p = P' = 2, and so HOLDER'S inequality takes
the form
f [fg[ dfl ~ [[f[[2 [[g[[2
for f, g E~2· Consider the mapping that takes ~2 x ~2 into K by the rule
(I, g) --+ f f g dfl = (f, g) ,
where the equality defines (f, g). This mapping has the following prop-
erties:
(fl + f2' g) = (fl' g) + (f2' g);
(rxf, g) = rx(f, g) for rx EK;
(f, g) = (g, f);
(f, f) > 0 for f"4= 0 .
We infer from these identities [or directly] that
(f, gl + g2) = (f, gl) + (f, g2) ,
(f, rxg) = Ii (f, g) ,
and
(0, f) = (f, 0) = 0 .
The spaces ~2 can be described in abstract terms, as follows.
(13.16) Definition. Let H be a linear space over K having an inner
product
(x, y) --+ (x, y) EK
mapping H x H into K such that
(x + y, z) = (x, z) + (y, z) ,
(rxx, y) = rx(x, y) for rx EK ,
(x,y) = (y, x),
(x, x) > 0 if x "4= O.
[The other properties of ( , ) listed for ~2 in (13.15) can be proved for H
from the above relations.] Then H is called an inner product space or
a pre-Hilbert space. For x EH, define
1
[[xli = (x, x)-Z.
The inequalities
[(x,y)[ ~ [[xli . [[y[[
and
I[x + y[[ ~ [lx[1 + [[yl[
13*
196 Chapter IV. Function spaces and Banach spaces
where A is a constant such that II (x) I < A for all xED. The reader
should find it easy to construct an example illustrating that the inclu-
sion is proper if D is infinite. 0
(13.19) Theorem. II I E~p n ~q, where 0< P< q< 00, and if
p < r < q, then I E~,. Also, the function q; defined by
q; (r) = log (1Ifll~)
on [p, q] is convex, i. e., 0 < IX < 1 implies
+ (1 - IX)q) ~ IXq;(P) + (1 - IX) q;(q) .
q;(IXP
Proof. Let r = IXP + (1 - IX) q, 0 < IX < 1. Using HOLDER'S inequality
If a = 0, then a E <roo (X) and the proof is complete. Thus suppose that
Iiall .. = M > O.
Let E = {x EX: a (x) =1= O}. Since a E en S!,p, we have t(E) < 00.
We apply LUZIN'S theorem (11.36) to obtain a function f/J E <roo such
that Ilf/JII .. ~ M and if A = {x EX: f/J(x) =1= a (x)}, then t(A) < (4~ y.
We therefore have
1
=; + (!la-f/JIPdt)P~; + (!(2M)Pdt)P
=; +2M·t(A)P< e.
1
0
r
Let I E s!'p and e> 0 be given. Use (13.20) to find a simple function aEs!'p
m
suchthatllt-allp< ; ,saya= 1: .Bj~Bj,whereBjEJl.forallj.SinceaEs!'p,
1=1
we may suppose thaI< (B;) < 00 10' all j. Let b ~ ; (2 (I +:~ IP,lj
and fix j in {I, 2, ... , m}. Use (9.24) to obtain an open set U j such that
00
00
where the Ij,k's are pairwise disjoint open intervals. We have 1: t(Ij,k)
k=1
00
= t(Uj) < 00, and so we may choose ko such that 1: t(Ij,k) < ~.
k=k.+1
k.
Let V; = k':!l Ij,k and TV; = Uj n V;'. We see at once that £(TV;) < ~,
V; L. B j C ([1; n Bj) U TV;, and £(V; L. B;) ~ t ([1; n Bj) + £(l1j) < 2~.
Therefore
1
IlgB;- gvjllp= IIgv;6dlp< (2r5)P.
m
Next set s = 1: .BigVj' Since each V; is the union of a finite number of
;=1
§ 13. The spaces .s!p{l ~ P < 00) 199
Y
1
x
f
(t~1/~' ... r::) d,a ~ II/lll~' 11/211~' ... Il/nll~" .
(13.27) Exercise. Write out carefully and prove the conditions for
equality in MINKOWSKI'S inequality for p = 1 and also for 1 < p < 00.
(13.28) Exercise. Let X be a locally compact Hausdorff space and
tand Jt,. as usual. Suppose that t(X) > 0 and t({x}) = 0 for all x EX.
Let p be any positive number.
(a) Find a function I E5!.p such that I ~ 5!.PH for all <5 > O.
2""
[Hint. Use (12.59) and the fact that I; -;2 =
00
00 for all <5 > 0.]
n=!
§ 13. The spaces ~p(1 ~ P< 00) 201
(b) Find a function I on X such that I E~p_" for all r5 > 0 and I ~ ~p.
00
[Recall that 1.: n-rx converges if a > 1 and diverges if 0 < a ~ 1.]
n~l
(13.30) Exercise. Let (X, .91, p) be a finite measure space and let I
be any bounded measurable function on X. Prove that
lim
p-+oo
1I/IIp = inf{a ER: a> 0, p({x EX: I/(x)1 > a}) = o}.
(13.31) Exercise. Let p be a real number such that 1 ~ P< 00, and
let I be a function in ~p (R) such that I is uniformly continuous. Prove
that I E~o (R). Show by examples that each ~p (R) contains an unbounded
continuous function.
(13.32) Exercise. Let (X,d, p) be a measure space such that
p(X) = 1 and let I be a function in ~t(X,.9I, p). Define log (0) as - 00.
(a) Prove that
(i) flog I (x) dp (x) ~ log ( f I (x) dp (x) .
x x
[Hint. Check the inequality log (t) ~ t - 1 for 0 ~ t < 00. Replace t
by 11:111 I(x) and integrate.]
(b) Prove that equality holds in (i) if and only if I is a constant func-
tion a. e. [Hint. Check that log (t) < t - 1 if t =F 1. J
(c) Prove that
(ii) lim IIfll, = exp [f logf(x) dp(x)].
'to x
[Hints. Show that W - I)ly decreases to logl as y 1- 0, and apply the
dominated convergence theorem to prove that
lim y-l( flrdp - 1) = flog I dp.
'to x x
Using (i), show that
(13.33) Exercise. Let (X, .91, p) be a measure space and let p be any
positive real number. Prove that if I and (fn)':~l are in ~p (X, .91, p) and
III - Inllp --+ 0, then In --+ I in measure. Find a sequence (fn)':'~l C
~p([O, IJ, JI}., A) such that In --+ 0 in measure but II/nllr+-+ O.
202 Chapter IV. Function spaces and Banach spaces
Prove that:
(b) ~~ c ~(/) [use (13.2)J;
(c) ~(/) is a complex linear space;
(d) I II(/) is a norm on ~(/), where functions equal a.e. are identified;
(e) with the norm 1111(/), ~(/) is a Banach space. [First suppose that
ft(X) < 00. Prove that if Il/n - Imll(/) -+ 0, then Il/n - Imlll -+ O.J
The spaces ~(/) are called BIRNBAUM-ORLICZ spaces. For further infor-
mation about these spaces, the reader should consult A. C. ZAANEN,
Linear Analysis, Vol. I [New York: Interscience Publishers, 1953J.
(13.37) Exercise. Define if> on [0, 00 [ by if> (t) = 0 if 0 ~ t ~ 1 and
if> (t) = t ·logt if t ~ 1. The Birnbaum-Orlicz space ~(/) (X, d, ft) [see
(13.36)J is often denoted ~ log+~. Prove that for the measure space
([0, IJ,~, A) we have
~p C ~ log+ ~ C ~1
for all p > 1. The space ~ log+ ~ arises quite naturally in Fourier analy-
sis. See for example A. ZYGMUND, Trigonometric Series, 2nd Edition.
[2 Vols. Cambridge: Cambridge University Press, 1959J, and also Theorem
(21.80) inlra.
(13.38) Exercise: VITALI'S convergence theorem. Let (X, d, ft) be a
measure space and let 1 ~ P < 00. Let (fS:=l be a sequence in ~p(X,d, ft)
and let I be an d-measurable function such that I is finite ft-a.e. and
In -+ I ft-a.e. Then I E~p(X, d, ft) and III - Inllp -+ 0 if and only if:
(i) for each e > 0, there exists a set A. Ed such that ft (As) < 00
and J I/nl P d,u< e for all n EN;
A~
and
(ii) lim
p(E)-;'()
f I/nl P d,u = 0
E
uniformly in n, i. e., for each e > 0 there is a t5 > 0 such that E Ed
and ,u (E) < t5 imply J I/nl P dft < e for all n EN.
E
Prove this theorem. [To prove the necessity of (i), let e > 0 be given,
choose no EN such that 111- Inllp < e for all n ~ no, choose B., C. Ed
of finite measure such that J IfIP dft < e and J Ifnl P dft < e for n = 1, ... , no.
B~ c~
Then put A. = B. U C•. The necessity of (ii) is proved similarly by using
(12.34). Next suppose that (i) and (ii) hold. Use (i), FATou's lemma and
MINKOWSKI'S inequality to reduce the problem to the case that ft (X) < 00.
204 Chapter IV. Function spaces and Banach spaces
Next suppose that (ii) holds, and write E k = n~k Bn,k. Plainly
f IgdA = ~ f d: = 1
exp(n}
° R
Thus In --+ uniformly on R but In ++ weakly in S!.p(R). ° 1
Thus, combining (1). (2) [with E = A n B']. (3). and (4) and using
HOLDER'S and MINKOWSKI'S inequalities, we have
<~+~+~=e
333
for all n ~ no. 0
(13.45) Corollary. The hypothesis I.. ~ I p.-a.e. in (13.44) can be re-
placed by the hypothesis that In ~ I in measure.
208 Chapter IV. Function spaces and Banach spaces
lim j f
k...... oo X
(I - Ink) g dflj = rL • (1)
lim
n-+oo E
f Ilnl dfl ~ EJill dfl = X
Jill dfl - Jill dfl
E'
Hence lim J
n.....oo E
Ilnl dfl exists and (i) holds.
To prove (ii) , let B > 0 be given. Select A Ed such that fl (A) < (X)
and fill dfl < ; . Use (12.34) to obtain a 0> 0 such that if E Ed and
A'
fl (E) < 0, then fill dfl < ; . Next apply EGORov'S theorem (11.32)
E
to find BEd such that Be A, fl(A n B') < 0, and In ~ I uniformly
on B. Choose no EN such that
[sup I/(x) - In(x)I]' fl(B) < ;
xEB
§ 14. Abstract Banach spaces 209
(14.4) Theorem. Let A and B be normed linear spaces over F and let
S8 (A, B) denote the set of all bounded linear transformations from A into
B. Then, with pointwise linear operations and the operator norm, ~ (A, B)
is a normed linear space. Moreover ~(A, B) is a Banach space if B is a
Banach space.
Proof. We prove only that ~(A, B) is complete if B is complete;
the other verifications are routine and we omit them. Suppose that B
is complete and let (Tn) be a Cauchy sequence in S8(A, B). For x EA
we have IITn(x) - Tm(x)11 ~ IITn - Tmll'llxll, and so (Tn(x)) is a Cauchy
sequence in B. Thus for each x EA, there is a vector T(x) EB such that
IITn(x) - T(x)ll-+ O. This defines a mapping T from A into B. For
x,y EA, we have IIT(x + y) - [T(x) + T(y)]11 ~ I T (x+y) - Tn (x+y) I
+ IITn(x) - T(x)1I + IITn(y) - T(y)lI-+ 0, and thus T(x + y) = T(x)
+ T(y). We prove similarly that T(a:x) = a:T(x) for all x EA and a: EF.
Thus T is linear.
Since (Tn) is a Cauchy sequence, there is a positive constant fJ such
that IITnll ~ fJ for all n EN. For Ilxll ~ 1, we have I T (x) I ~ I T (x) -
Tn(x)1I + I Tn (x) I ~ 1 + fJ [for all large n], and hence T is bounded,
i.e., T EQ3(A, B). It remains only to show that liT - Tnll-+ O. Let
e > 0 be given. Choose an integer p so large that m, n ;;?; p implies
e
II Tm - Tnll < 2' Next let x EA be such that IIxll ~ 1 and choose mx EN
such that mx ;;?; p and I T (x) - Tmz (x) I < ; . Then n ;;?; p implies that
transfounation. Also
sup{lx(f)i: f EE*, 11111 ~ I} sup{l/(x)l: f EE*, 11111 ~ I}
=
~ sup{ll/llllxll : f
EE*, 11111 ~ I} ~ Ilxll
for each x EE. Thus the mapping x -+ x is a bounded linear transfouna-
tion from E into E** of noun ~ 1. Several questions arise. (1) Is this
mapping one-to-one? (2) Does it preserve nouns? (3) Is it onto E**?
(4) Indeed, are there any nonzero elements in E* ? In general none of these
questions have obvious answers; however we are able to answer (1),
(2), and (4) with the aid of the Hahn-Banach theorem, which is next
on our program. Question (3) will be answered in the exercises.
(14.8) Definition. Let E be a real linear space. A real-valued function
p defined on E is said to be a sublinear functional if
(i) P(x + y) ~ P(x) + P(y)
and
(ii) p (OI:x) = OI:p (x)
for all x, y EE and all positive real numbers 01:. Notice that a noun is
a sublinear functional.
(14.9) Hahn-Banach Theorem. Let E be a real linear space and let M
be a linear subspace 01 E. Suppose that p is a sublinear lunctional de-
fined on E and that I is a linear functional defined on M such that I (x)
~ p (x) lor every x EM. Then there exists a linear lunctional g defined on
E such that g is an extension 01 f [i.e., f c gJ and g(x) ~ P(x) lor every
xEE.
Proof. Let 8 be the set of all real functions h such that domh is a linear
subspace of E, h is linear, f c h, and h(x) ~ P(x) for all x Edomh.
Notice that I E8. Partially order 8 by c [recall that a function is a set
of ordered pairs]. Let ~ be any chain in 8 and let h = U ~. Then, with
the help of (2.19), we see that h E8. Applying ZORN'S lemma (3.10),
we see also that 8 has a maximal member, say g. To complete the proof
we need only show that domg = E. Assume that this is false, and let y
be any element in En (domg),. Let G = domg and define H = {x+ OI:y:
x E G, 01: E R}. Clearly H is a linear subspace of E and G H. Let c S
be a fixed, but arbitrary, real number and define h on H by
h(x + OI:y) = g(x) + OI:C.
Then h is well-defined since if Xl + OI:lY = X 2 + 0I:2Y, where xl> X 2 EG
and 01:1,01:2 ER, then (01:1 - 01:2) y = x 2 - Xl EG so that 01:1 = 01:2 and Xl =x2•
S
Clearly h is a linear functional and g h. If we can select c in such
a way that h (x) ~ P(x) for all x EH, then we will have h E8, which
contradicts the maximality of g and will complete our proof. The re-
mainder of our proof is therefore devoted to showing that c can be so
selected.
§ 14. Abstract Banach spaces 213
Our requirement is that g(x) + exc = h(x + exy) ~ P(x + exy) for
all x E G, ex ER. By the linearity of g and the sublinearity of p, this is
equivalent to the two requirements
= sup { 11% I:d~zll : x + oez EM, Ilx + oezll =!= O} = sup {II-/+ zll : YES}
= ~ = 1. Apply (14.11) if F = R or (14.12) if F = K to obtain the re-
quired functional g E E*. 0
We now return to the mapping x -c>- X discussed in (14.7).
(14.14) Theorem. Let E be a normed linear space and let :n; be the
natural mapping 01 E into E**: :n; (x) (f) = I (x). Then:n; is a norm-preserv-
ing linear translormation Irom E into E**. Consequently :n; is one-to-one.
Proof. We have already observed in (14.7) that :n; is a bounded linear
transformation from E into E** and that 11:n;11 ~ 1. Let x be any non-
zero element of E. According to (14.13), there is an element g E E*
such that Ilgll = 1 and g(x) = Ilxll. Thus
that is,
lin (x) I = Ilxll .
Clearly lin (0) I = 0 = 11011. We have thus proved that n preserves nonns.
Consequently x =l= y in E implies that Iln(x) - n(y) I = Iln(x - y)11
= Ilx - yll =l= 0, and so n(x) =l= n(y). 0
(14.15) Remark. In view of (14.14), a nonned linear space E is
indistinguishable qua nonned linear space from the subspace n(E)
of E**. The mapping n need not be onto E** [see (14.26)]. In case
n(E) = E**, the space E is called reflexive. Since E** is complete and n
is an isometry, every reflexive nonned linear space is a Banach space.
In § 15 we will show that every 2.p space (1 < p < 00) is reflexive.
We next present three theorems which, together with the Hahn-
Banach theorem, are often regarded as the cornerstones of functional
analysis. These are the open-mapping theorem, the closed-graph theorem,
and the uniform boundedness principle. Several applications of these
theorems will be given in the corollaries and the exercises. Unlike the
Hahn-Banach theorem, these three theorems require completeness.
(14.16) Open mapping theorem [BANACH]. Let A and B be Banach
spaces and let T be a bounded linear transformation from A onto B.
Then T(U) is open in B for each open subset U of A.
Proof. For each e > 0, define A. = {x EA : Ilxll < e} and B. = {y EB:
Ilyll < e}. Let e> 0 be given. We will show that there exists a 15 > 0
such that T (A.) ::::> BiJ. For each n EN, let en = ; ... It is clear that if n
is fixed, then
00
A =.U jA.,.
1=1
[we definejA.,. as in (5.6.£)], and so we also have
00
Since B is complete, the Baire category theorem (6.54) implies that not
every T(jA ...), j = 1,2, ... , is nowhere dense. Thus there is a jn EN
such that [T(jnA.,,)r has nonvoid interior. But
[T(A .../2 )]- = 2~.. [T(jn A .,.)]-,
and (X W is open in B if (X =l= 0 and W is open in B. Thus there exists a
v..
nonvoid open set C [T(A.nI 2)]-' It follows that
[T(A.,.)]-::::> [T(A.nI2) - T (A",/2)]-::::> [T(A",/2)]-
- [T(A",/2)]- ::::> v.. - v...l
1 For subsets C and D of B, we write C-D={x-y:xEC, yED}: see
(5.6.£).
216 Chapter IV. Function spaces and Banach spaces
(2)
which has limit 0 as m --+ 00. Thus (zn) is a Cauchy sequence in A ; since A
is complete, there is an x EA such that Ilx - znll --+ O. It is clear that
I T(x)11 = n--+co
lim liT.. (x) II ~ Mllxll '
for each IE E*. Prove that sup{llxll : xED} < 00. [Consider n(D) CE**.]
(14.30) Exercise. Let E be a Banach space and let A and B be
closed linear subspaces of E such that A n B = {o}. Prove the following.
(a) If A + B is closed in E, then the mapping x + Y --+ x for x E A,
Y E B is a continuous linear mapping of A + B onto A.
(b) For x E A, Y E B let Ilx + yll' = Ilxll + Ilyll· Then 1111' is a complete
norm for A + B.
(c) The set A + B is closed in E if and only if IIII and 1111' induce the
same topology on A + B.
(14.31) Exercise. (a) Let E be a normed linear space and let M be
a closed linear subspace of E. Suppose that z EE n M'. Let 5 = {x+az:
x EM, a EF}. Thus 5 is the smallest linear subspace of E containing
M and z. Prove that 5 is closed in E. [Define I on 5 by I(x + az) = a.
Show that IE 5* and IIIII ;;:;:; dist(~,M) . Then use the fact that F is
complete.]
(b) Prove that every finite-dimensional linear subspace of E is closed
in E. [Use (a) and induction].
(14.32) Exercise. Prove that there exists no Banach space of al-
gebraic dimension No. [Use (14.31) and the Baire category theorem.]
For an arbitrary nonzero cardinal number m [finite or infinite], construct
a normed linear space of algebraic dimension m.
§ 14. Abstract Banach spaces 221
(x + M) + (y + M) = (x + y) + M
and
ex (x + M) = (exx) +M .
Define the quotient norm on EjM by the rule
(i) Ilx + Mil = inf{llx + mil: m EM}.
222 Chapter IV. Function spaces and Banach spaces
Prove that
(a) e is a metric for B,
and
(b) with the metric e, B is a compact metric space.
Let P be CANTOR'S ternary set.
(c) Prove that there exists a norm-preserving linear transformation
T from E onto a closed subspace of <r(P), where <r(P) has the uniform
norm. [Use (b) and (6.100) to obtain a continuous mapping q; from P
onto B. Define T(x)(t) = q;(t)(x), for x EE and t E P.]
(d) Prove (c) with P replaced by [0, IJ.
(14.40) Exercise. Let E be a normed linear space and let S be a
linear subspace of E that is dense in E. Suppose that I is a bounded linear
functional on S. Without recourse to the Hahn-Banach theorem, prove
that there exists a unique g EE* such that g(x) = I (x) for all xES.
Prove also that Ilgll = 11/11·
ative is
(/)' (x) = - X;+1 [(1 + X)I>-1 + (1 - X)I>-1 - 21>-1] • (2)
Write IX = P- 1 [note that IX ~ 1], and consider the function lJI de-
fined by
We have
lJI'(x) = IX(1 + X)ot-l_ IX(I- X)ot-l ~ 0 for 0 < x < 1.
Thus lJI is a nondecreasing function on [0, 1], and since lJI(I) = 0, the
mean value theorem implies that lJI (x) ~ 0 for 0 ~ x ~ 1. Going
back to (2), we infer that (/)' (x) ~ 0 for 0 < x < 1, and since (/)(1) = 0,
(/)(x) is nonpositive for 0 < x < 1. The definition (1) shows that F(x) is
also nonpositive for 0 < x < 1. 0
(15.4) Lemma. Let z and w be complex numbers, and suppose that
p ~ 2. Then we have
(i) z+w II> +-2-
I-2-
Iz-wll> ~2+-2-·
\zIP Iwlt>
(1)
which we will now prove. The inequality (1) can be written in the form
r
11 + r~Xp(ilJ) + 11 -r~p(ilJ) II> ~ ~ (1 + rl» (2)
Proof. We may suppose that I and g assume complex values and are
defined p,-a.e. [(12.18) and (12.26)]. Then for all x EX such that I(x)
and g (x) are defined, (15.4.i) implies that
for °< u < 1. It is clear that the steps going from (i) to (3) are reversible,
Hewitt/Stromberg, Real and abstract analysis 15
226 Chapter IV. Function spaces and Banach spaces
As shown in (7.25), the last line of (4) converges absolutely and uni-
formly for U E [0, I]. We will show that each term [... J in this series
is nonnegative. Plainly this will prove (3). The kth term is
t' t'
I
The left side of (1) is the left side of (i), since IIIW' lip-I = Ilhll~' for
any h E 5!p. The right side is
[f 2P-l (I ~ IP + I; n I
I !Pnllp = 1; (2)
lim L(!Pn)
n-+oo
= IILII . (3)
We will show that (!Pn) is a Cauchy sequence in 5!-p. In the contrary case,
there are a positive number (X and subsequences (!pnk)'k~l and (!pmk)k~l
such that I !pnk - !pmkllp > (X for k = 1,2, .... For p ;;?; 2, we use CLARK-
SON'S inequality (15.5) to write
(4)
[~ I !pmkll~ + ~ I !pnJ~p-l
1
(6)
(7)
From (6) and (7) we can find, for each p> 1, a number {3 EJO, 1 [ that
is independent of k and such that
(8)
lim L (gk)
k--;.oo
~ 1 ~ f3 IILII .
Since Ilgkllp = 1, this is an evident contradiction. Therefore (gJn) is a
Cauchy sequence in S!p and so has a limit gJo in S!p (13.11). It is clear
from (3) that lim L (gJn) = L (gJo) = IILII. 0
n--;.oo
1 This lemma is due to E. J. MCSHANE [Proc. Amer. Math. Soc. 1 (1950), p. 402].
230 Chapter IV. Function spaces and Banach spaces
OJ(I) ~ OJ(O) = f
P It'l + g1P-2 [(t'/I + gI) 11 + (t'/2 + g2) 12] df-' (3)
x
where 0 < WI < It I and t' is a lunction 01 x EX. [If 1 < P < 2 and t'/(x)
+ g(x) = 0, then the integrand is zero.] Since (t'I, + g,) ~ It'l + gl
and I, ~ III, the absolute value of the integrand in (3) is less than or
equal to 2p It' 1+ glP-l III. If It I ~ 1, then we have 2p It' 1+ glP-l III
~ 2P(1/1 + Igl)P-1I/I· The functions III and Igi are both in S!,p, and so
([II + Igl)P-l is in S!,P" and (III + Igl)P-1I/I is in S!,l' by HOLDER'S inequality
(13.4). Thus for all It I ~ 1, the integrand in (3) is less than or equal to
the fixed function 2p ([II + Igl)P-1I/I, which is in S!,l' LEBESGUE'S theorem
§ 15. The conjugate space of 21>(1 <p< 00) 231
~~ f
Ig + ttl; -Igll> dp = f
P /g/P-2[gdl + g2f2] dp. (4)
x x
[If 1 < p < 2 and g(x) = 0, then the integrand in (4) and in the following
integrals is zero.] Combining (2) and (4), we see that w'CO) exists and
that
w' (0) = f P /g/P-2 [glfl + g2f2] dp . (5)
x
Consequently "PI (0) also exists. Using (5), we write
(!
1
= f
/g/P-2 [gdl + g2f2] dp . (6)
x
Lemma (15.10) and (6) imply that
L (I) = IILII ("PI (0) + i"P'--i/(O))
= IILII f /g/P-2((gdl+g2f2)+i(glf2-g2fl))dp=II L II' f /g/P-2gjdp.
x x
The theorem follows when we set h = IILII . /g/P-I sgn(g); i.e.,
f fit dp. 0
L (I) =
x
(15.12) Theorem. Let (X, .91, p) be an arbitrary measure space and let
p be a real number such that 1 < p < 00. Then the mapping T defined by
T(g) = Lg
[see (15.1)] is a norm-preserving linear transformation from !i!.p' onto !i!.t-
Thus, as Banach spaces, !i!.p' and !i!.; are isomorphic.
Proof. The fact that T is a norm-preserving mapping from !i!.p'
into !i!.: is (15.1). It follows from (15.11) that T is onto !i!.t- It is trivial
that T is linear. Since T is linear and norm-preserving, T is one-to-one. 0
(15.13) Exercise [J. A. CLARKSON]' Let (Y, .91, p) be a measure
space such that .91 contains two disjoint sets of finite positive measure.
There is a [unique] least positive number c such that
~ < lit + gil; + lit - gil; <
c = 2 (11tll; + Ilgll;) = c
for all f, g E !i!.p (1 < p < (0) such that IIfllp and IIgllp are not both zero.
12-PI 1
Prove that c exists and that c = 2 P . Also, the constants c and -
c
are attained.
232 Chapter IV. Function spaces and Banach spaces
be any element of E n 5'. Show that there is one and only one element
Yo E5 such that
Ilyo - xii = inf{lly - xii: Y E5} .
(15.17) Exercise: Weak convergence and the RAnoN-lb:Esz theorem.
In (13.41), we defined weak convergence for sequences of functions in ~p
and thereafter studied relations between weak convergence and other
sorts of convergence. Weak convergence can be defined for sequences
(xn) in any normed linear space E, as follows. The sequence (xn) converges
-+
weakly to x EE if I (xn) I (x) for all I in the conjugate space E* of E.
Theorem (15.11) shows that the present definition is consistent with
the definition offered in (13.41) for E = ~p (1 < P < (0): for E = ~l>
see (20.19) inlra.
(a) Let E be a normed linear space [over R or K] with the property
that if (xn) is a sequence in E, x EE, Ilxnll = 1, and Ilxll = 1, then the
relation
IIX
n
:xll-+ 1
implies
Ilxn - xii -+ 0 .
[Such spaces are called locally unilormly convex.]. Let (Yn) be a sequence
in E and y an element of E such that:
(i) IIYnl1-+
Jlyll ;
-+
(ii) Yn Y weakly.
Prove that llYn - yll -+
o. [If Jlyll = 0, the assertion is trivial. Otherwise,
write Xn = IIYnll-1Yn and x = IIYII- 1y. Then Xn x weakly and Ilxnll-+
= Ilxll = 1. Also if Ilxn - xii -+
0, it follows that llYn - yll O. Hence -+
we need only to show that Ilx n- -+
xii O. Assume the contrary. Then by
local uniform convexity, there is a subsequence (Xnk)k~l such that
x = (xn) such that 1: Ixn 2 <l 00, is a Hilbert space in which «(xn), (Yn)
n~l
00
n
= 1: IIZjI12. 0
i~l
§ 16. Abstract Hilbert spaces 237
(16.11) Definition. Let E be a nonned linear space and let (Xn ):'=1
00
and we write 1: Xn
n=!
= x, if there exists an xEE such that lim
P-+oo
[[x - !; Xn[[
n=!
=0.
(16.12) Theorem. Let H be a Hilbert space and let {Zn}:'=l be an
Zn -1 Zm tor n =1= m. E Zn converges it and
00
= <il IXkXk, Xi) = t!lXk(Xk , Xj) = IXj Ilxj l1 2. But Xj =1= 0, and so IIxjl12 =1= O.
Therefore alllX/s are zero. 0
(16.14) Definition. Let E be an arbitrary orthogonal set in an inner
product space H. For X EHand Z EE, we define the Fourier coefficient
ot X with respect to Z to be the number (x, z).
238 Chapter IV. Function spaces and Banach spaces
JI(t) exp(-int) dt
n
Proof. We have
.. \\2 .. ..
\\ x - k-?: (Xk Z,. = I xl1 2 - k~ (Xk(X, Zk) - k~ (Xk(Z,., x) + k~.. l(Xkl 2
n n
= IIxl1 2 + I; [l<x, Z,.)12 - OC,. (x, Z,.) - (Xk (x, Z,.) + l(XkI 2] - I; I(x, Z,.)12
k=l
. . k=l
For every x EH, the vector Y = I; (x, Z,,)Zk exists in H and x - Y is orthog-
k=l
onal to every z,..
Proof. The existence ofy follows from (16.12) and BESSEL'S inequality.
Let mEN. We must show that (x - y, zm) = O. For each n EN let
n
Yn = I; (x, Zk) Zk' Then for all n ~ m, we have
k=l
I(x - y, zm)1 ~ I(x - Yn' zm)1 + I(Yn - Y, zm)1
If {Yl> Y2' ... , Yn} is aU of the y's, stop the construction. If Yn+1 exists,
let
.
zn+1 = Yn+1 - E (Yn+1' UII) ull .
11=1
Now zn+1 is not the zero vector, because Yn+1 ~span{Yl'" .,Yn}
= span{ul , ... , un}. Next define
1
(un+1' u;) = II Zn+1II- (Yn+1 - 1;1 (Yn+1' UII) ulI' U;)
= IIZ"+111- 1 (Yn+1' u;) - (Yn+1' u;) • (u;, u;») = 0 .
Thus the set {Ul> ... , un+1} is orthonormal; it remains to show that
it spans the same subspace as {Yl> ... , Yn+1}' It is obvious from the def-
initions of zn+1 and Un +1 thatY"+1is a linear combination oful , ... , un+!'
Therefore
{Yn+!} U (span{Yl' ... , Yn}) C span{ul> ... , Un+l} •
and so
span{yl> ... , Yn+!} C span{ul> ... , un+l} .
Similarly un+1 is a linear combination of the vectors Yn+1 and Ul> ...• Un'
By the inductive hypothesis, it follows that un+! Espan{Yl> ... , Yn+l},
and therefore span{ul>"" Un+l} C span{Yl>" "Yn+l}' Thus these two
subspaces are the same. [Alternatively, we could have reversed the
first inclusion by a dimensionality argument.]
The Gram-Schmidt process yields an orthonormal set which is es-
sentially unique. More precisely, if {Ul> ... , un} is to span the same
subspace as {Yl> ...• y,,} for each n, then we must have u l = II~II Yl>
where Irl = 1. Hence the choice of u l is unique up to a multiplicative
constant of absolute value 1. Having defined {Ul> ...• un}. we must take
un+l Espan {Ul' ... , un.Yn+1}· Thus for some complex numbers Pl' ... , Pn'
ex, we have
and so also
and
lJ j = - (Yn+l' Uj) (1;;;' j ;;;, n) .
Thus we have
just as we defined zn+l' The number (X is now determined from this last
equality by taking norms and noting that IIUn+ll1 = 1. Clearly (x, and
hence U n + v is unique up to a multiplicative factor of absolute value 1.
(16.23) Theorem. Let H be an inner product space, not {O}, that con-
tains a countable dense subset D. Then H contains a countable complete
orthonormal set that is obtained from D by the Gram-Schmidt process.
Proof. Suppose, as we may, that 0 ~ D, and enumerate D as (Xn):'l'
Define Yl = x n , where n 1 = 1. Suppose that Yl = x n " ..• , y,. = xnkhave
been defined and are linearly independent, and that n 1 < n 2 < ... < n,..
If there is no j > n,. such that {yv ... , Y,., Xj} is linearly independent,
stop the process. Otherwise let n"+l be the smallest such j and define
Y"+l = xnk+1' We have thus defined a finite or countably infinite linearly
independent set {Yl' Y2' ... } C D. Let 5 be the smallest linear subspace
of H containing {Yv Y2' ... }. It is clear that DeS, since if Xj ED, then
Xj is a linear combination of Yl' ... , y", where k is chosen so that
n" ;;;, j < n1<+ 1 [or else y" is the last Y selected and n,. ;;;, j]. Hence 5 is dense
in H. Let {u v us' ... } be the orthonormal set obtained from the set
{Yl' Y2' ... } by the Gram-Schmidt process. We will show that {uv us' ...}
is complete. Suppose that x E H and (x, un) = 0 for all n. Then
(x, g /XnUn) = 0 for all finite linear combinations of the un's, and so
(x, y) = 0 for all yES. It follows from (16.21) that x = o. 0
(16.24) Corollary. Let n EN. T.hen an inner product space H is in-
distinguishable [as an inner product space] from Kn if and only if the
algebraic dimension of His n. [In Kn we define «(Zl' ... , zn), (wv ... , wn)
n
= 1: ZjWd
i=1
Proof. It is clear that if H has dimension n, then the process of
choosing the y's in the previous proof stops with {Yv ... , Yn}. Thus we
obtain a complete orthonormal set {ull •• '1 un} C H. The mapping of
H onto Kn given by
n
1: (Xi Ui -+ ((Xl' • • ., (Xn)
i=1
§ 16. Abstract Hilbert spaces 243
where the superscript (n) denotes the nth derivative of the function
x -+ exp [- x 2 J. The functions (Hn):~o are clearly all polynomials. They
are called the Hermite polynomials. The first three Hermite pol ynomials are
Ho(x) = 1,
HI (x) = 2x,
H 2 (x)=4x 2 -2.
One can go on computing them as long as patience will permit. Next let
these functions are called the Hermite lunctions. They are all in
S!.2(R, Jt)., A), and, as we will now show, they are an orthogonal set.
First we have
+ 2xexp[ ~2] exp(n+l) [_ X2] + exp[ ~2] exp(n+2) [-X 2]}. (1)
-00 -00
and therefore
H~(x) = (-1)n {2x exp [X2] exp(n) [_X2]
+ exp[x2] (-2x exp(n) [_X2] - 2n exp(n-l) [_X2J)}
= (_1)n-l 2n exp [X2] exp(n-l) [_X2] = 2nHn_1 •
00
00
-00
= lim {(-1)nHn(X)
A~oo
exp(n-1)[-x2]1~.A
+ (_1)n-1 /H~(x) exp(n-1) [-x2J dX}
-A
00
-00
00
= 2n
-00
J CP;-l (x) dx .
This establishes the recursive formula
00 00
n"22nn!
-00
= (n+ 2n n!t"2 CPn{x) = {n+ 2 n n!t"2 (-I)n exp [~2] exp(n) [_X2] form
an orthonormal set. The functions {1J!n}:=o are obtained from
{xn exp [- ~2 ]}::o
by the Gram-Schmidt process. This follows readily
from the fact that Hn has degree n for all n, and from the essential unique-
ness pointed out in (16.22).
(16.26) Theorem. Let H be a Hilbert space. The following properties
01 an orthonormal subset E at H are equivalent.
(i) The set E is complete.
(ii) For each x EH, we have x = I.; <x, z)z [Fourier seriesJ.I
zEE
(iii) For all x EH, we have II xII 2 = I.; I<x, z)12 [PARSEVAL'S identity].
zEE
1 The equality (ii) means that the right side has only a countable number of
nonzero terms and that for every enumeration of these terms the resulting series
converges to x as in (16.11). The equalities (iii) and (iv) have analogous meanings.
246 Chapter IV. Function spaces and Banach spaces
(iv) For all x,y EH, we have (x,y) = I; (x, z) (y, z) [PARSEVAL'S
lEE
identity].
(v) The smallest subspace 01 H containing E is dense in H.
Proof. Suppose that (i) holds and let x EH. According to BESSEL'S
inequality (16.17), there are only countably many z EE such that
(x, z) =l= 0; enumerate these as (z..). By (16.18), the vector y = I; (x,z.. )z..
= I; (x, z)z exists and x - y is orthogonal to E. Since E is complete,
.
JEE
it follows that x - y = 0; and so (ii) holds.
To show that (ii) implies (iv), let x, y EH be given and let (z,,) be
an enumeration of all z EE such that (x, z) =l= 0 or (y, z) =l= o. Let
.
x.. = I; (x, z,,) Z"
"=1
and
.
y .. = I; (y, z,,) z".
"=1
Then we have
because Ilx - x.. II -40, Ily.. - yll --+ 0, and II x.. II --+ Ilxll. Thus (x, y)
00
= I; (x, z,,) (y, z,,) = I; (x, z) (y, z); hence (iv) is established.
"=1 JEE
It is obvious that (iv) implies (iii). if (iii) holds and (x, z) = 0 for all
z EE, then it is plain that Ilxll = o. Thus (iii) implies (i). This completes
the proof that (i), (ii), (iii), and (iv) are pairwise equivalent.
Plainly (ii) implies (v). Finally we show that (v) implies (i). Suppose
that x EH and (x, z) = 0 for all z EE. Then it is clear that (x, y) = 0
for ally in the linear span ofE. It follows from (v) and (16.21) that X= o. 0
(16.27) Theorem. Any two complete orthonormal sets in a Hilbert
space H have the same cardinal number.
Proof. Ignoring a trivial case, we suppose that H =l= {O}. Let A and B
be any two complete orthonormal sets in H. If A is finite, it follows from
(16.26.ii) and (16.13) that A is a Hamel basis for Hover K. Since B
is linearly independent, (3.26) shows that B is contained in a Hamel basis
C, so that 11 ~ {j, and {j = .1 by (4.58). Thus B is also finite, and is
also a Hamel basis. Another reference to (4.58) shows that 11 = .1.
The case in which A and B are infinite remains to be treated. For
each a E A, let Ba = {b EB: (a, b) =l= O}. Then Ba is countable for all
§ 16. Abstract Hilbert spaces 247
exp (it) ~
"
}; IX" exp (ikt)
k=-"
where IX" EK. These functions are called, for an obvious reason, trig-
onometric polynomials. We have proved in (7.35.b) that <I' is uniformly
dense in <reT). It is plain that this implies that any function IE<r([-n, n])
such that I(-n) = I(n) can be uniformly approximated [arbitrarily
closely] by functions in <I' [regarded as functions of t E [- n, n]]. Now
let rp E~2([-n, n]) and let e be a positive real number. Use (13.21)
to select a function gE<r([-n,n]) such that Ilrp-gI12~;' Next,
by changing the values of g on an interval [n - ~, n] for an appropriate
~ > 0 [if necessaryJ. it is easy to find a function I E<r ([- n, n]) such that
1(- n) = I (n) and Ilg - 1112 < ;. Finally, choose a function P E<I'
such that III - PII" < ; . Then
Ilrp - PI12 ~ Ilrp - gl12 + Ilg - 1112 + III - PI12
I
= e.
§ 16. Abstract Hilbert spaces 249
The number I(n) is called the nth Fourier coelficient 011. The function I
on Z is called the Fourier translorm 011.
(16.34) Theorem. Let I E~l([-~' ~]). II the identity I = 0 holds, then
1= 0 in ~l'
Proof. Let ~ be the set of all trigonometric polynomials on T, defined
as in the proof of (16.32). If 1= 0, then it is obvious that
n
J I(t) pet) dt = 0
-n
for allp E~. Let x be a fixed number in ]-~, ~]. It is easy to construct a
nondecreasing sequence (g,,)::"=l of nonnegative continuous functions
such that:
g,,(-~) = g,,(~) = 0;
for each y ER. Once again the equality 1=0 implies f = 0 a.e. Also,
lim I (y) = O. These transforms are important in several parts of analysis.
Iyl-+oo
Using them, it is possible to prove that the Hermite functions are a
complete orthonormal set in ~2 (R); for a short proof, see (21.64.b)
infra.
(16.37) Theorem [PARSEVAL'S identity]. For f E~2([-n, n]), the
equality
and g EiY imply g I E921 [i. e., 921 is invariant under multiplication by
lunctions in iYJ. Let P be the projection of ~2 onto 921 as defined in (16.47).
Prove that peep) = P(I)· ep for all ep E~2 and that P(I) = ~E for some
set E Ed. Thus 921 is exactly the set of all functions in ~2 that vanish
on E'. Note that every such set is plainly a closed subspace of ~2 invariant
under multiplication by functions in iY. [Hints. Consider any hE 921..1.,
IE 921, and g EiY. We have
J I g h dfl = JUg) ii dfl = 0
x x
and so 921..1. too is invariant under multiplication by functions in iY.
Since 1- P(I) E921..1., it follows that P(g(1 - P(I))) = 0 and so
peg) = P(gP(I)) = gP(I) for all g EiY. As iY is dense in ~2' we infer
that P (ep) = ep P (1) for all ep E~2. Since P (I) = P (1)2, we haveP (I) = ~E.]
(b) Let (X, d, fl) be a a-finite measure space, and let 921 and iY
be subspaces of ~2 (X, d, fl) just as in part (a). Prove that the projection
operator P of ~2(X, d, fl) onto 921 has the form P(f) = ~EI for some
00
E Ed.! [Write X = }:11 F n, where the Fn's are pairwise disjoint sets in d
of finite measure. Apply (a) to each subspace ~Fn ~2(X, d, fl) and add.]
(16.49) Exercise. Let H be the real Hilbert space l~ = l~(N). Define
C= {I EH : [I (n) [ ~ : for all n EN}. Prove that C is a compact, no-
where dense subset of H. The metric space C is known as the Hilbert
parallelotope.
(16.50) Exercise. Let x, y, z be elements of an inner product space H
such that [[x[[ = llY[[ = [[z[[. Prove that
(i) [<x, x)<z, y) - <x, y) <z, X)[2 ~ [<x, X)2 - [<y, X)[2][ <x, X)2_[<Z, X)[2].
[Hint. You may obviously suppose that [[x[[ = llY[[ = [[z[[ = 1. Use the
Gram-Schmidt process to replace H by K3, X by (1,0,0), Y by (ex:, fl, 0),
and z by (y, b, s), where [ex:[2 + [fl[2 = [y[2 + [b[2 + [S[2 = 1. Then (i)
becomes nearly triviaL]
(16.51) Exercise. Let x, y, z and H be as in (16.50). Prove that
(i) <x, X)2([<y, Z)[2+ [<y, X)[2+ [<z, X)[2)
~ <X,X)4 + <x,x) <z,y) <y,x)<x,z) + <x,x)<y,z)<x,y)<z,x).
(16.52) Exercise. Prove that a Hilbert space H is separable if and
only if the orthogonal dimension of H is ~ toto [see (16.29)J.
(16.53) Exercise. Let (X, d, fl) be a measure space such that
0< fleX) < 00 and let (d, e) be the metric space defined in (10.45).
Let m be the smallest cardinal number of a dense subset of (d, e).
1 For a yet more general result, see (19.76).
§ 16. Abstract Hilbert spaces 255
Differentiation
This chapter contains first a brief but reasonably complete treatment
of the theory of differentiation for complex-valued functions defined
on intervals of the line. Section 17 is severely classical, containing
examples and LEBESGUE'S famous theorem on differentiation of functions
of finite variation. In § 18, we explore the conditions under which the
classical equality
b
t(b) - t(a) = J f'(t) dt
a
These two extended real numbers are called the lower right limit and the
upper right limit 01 rp at a respectively. If rp is a real-valued function
defined on ]a - ~, a[, define the lower lelt limit and the upper lelt limit 01 rp
at a to be the extended real numbers
lim rp(h) = sup{inf{rp(h): t < h < a}: a - ~ ~ t < a}
"t"
and
lim rp(h) = inf{sup{rp(h): t < h < a}: a - ~ ~ t < a}
"t"
respectively.
(17.2) Definition. Let a ER and ~ > o. If I is a real-valued function
defined on [a, a + ~ [, define
D I(a) = lim I(a + h) - I(a)
+ "to h
and
D+/(a) = lim I(a + h) - I(a) •
lito h
If I is a real-valued function defined on ]a - ~, a], define
D_/(a) = lim I(a + h) - I(a)
"to h
and
D-/(a) = lim I(a + h) - I(a) •
"to h
These four extended real numbers are known as the Dini derivates 01 I
at a; D+/(a) is the lower right derivate, D+/(a) is the upper right derivate,
D_/(a) is the lower lelt derivate, and D-/(a) is the upper lelt derivate.
(17.3) Remarks. The inequalities
(D+f)(a) ~ (D+/)(a)
and
(D_/)(a) ~ (D- I)(a)
obviously hold. Also it is easy to see that (D+t)(a) [(D+t)(a)] is the
largest [smallest] limit of a sequence (' (a + h~~ - I (a) ) , where h" > 0
and lim h" = o. Similar statements hold for (D- t)(a) and (D_t)(a).
n--->-oo
°
where < h' < h. The absolute value of the right side of (1) is less than
or equal to a"n, and so
n-l a"b" _ 1 na"b"
IS", I ~ 1: akbkn = n ab _ 1 < ab _ 1 • (2)
k=O
Now write
a"'x = IX", + f3", ,
·
Smce 3
2"" ~ 1-
f1", > 2""1 ,we h ave 3 1 - (J"
2a" ~ -a-"- > 1
2a"
d
an so
2a" 1
-~-<2a'"
3 - k" •
since cosn (In ~ 0. Combining this estimate for IRnl with (2), we obtain
the inequality
If(x + h)h - f(x) I <IX
holds for every h EJO, ()[. Select an integer n for which n > max {~ , IX}.
00
1 Many writers have made constructions of this sort. Our construction is taken
from S. BANACH, Studia Math. 3, 174-179 (1931). See also K. KURATOWSKI.
Topologie I. Deuxieme Edition. Monografie Matematyczne. Tom XX. Warszawa-
Wroclaw. 1948. pp. 326-328.
262 Chapter V: Differentiation
1 We give the ingenious proof of this theorem due to S. BANACH [Fund. Math.
6, 170-188 (1924)J.
264 Chapter V: Differentiation
It follows that
(3)
and, since Ie Uq _ v we have
1.(1) ~ <5 q_1 < 21. (Iq) . (4)
Since A.(}q} = 51. (Iq) , (3) and (4) show that
n A 'eU
00 00
so that x E U
n=p
In. Hence we have E
n=p
In' which implies that
A.(E n A'} = o.
Now let e > 0 be given and choose an integer p so large that
00
}; A.(In) < e .
n=p+l
Then
En Ap C (E n A') U (
n=p+l
U In),
and so
A(E U In) <
n Ap} ~ 0 + A ( n=p+l e.
Thus the proof is finished if A(E) < 00.
Case II: A(E) = 00. For each nEZ, let En = E n ]n, n + I[ and let
n.
i'; = {I Er : I C ]n, n + 1 Clearly i'; is a Vitali cover of En. Apply
Case I to find a countable pairwise disjoint family J" C i'; such that
ACEn n (U J,,),) = 0 for each n EZ. Let J
00
Choose an open set U:::> ElL." such that A(U) < 0: + 8. For each x EE".",
there exist arbitrarily small positive numbers h such that [x, x + h] C U
n [a, b] and
f(x+h)-f(x)<uh. (1)
The family "f/ of all such closed intervals is a Vitali cover of E u ,,,, and so,
by (17.11), there exists a finite, pairwise disjoint subfamily {[Xi' Xi+ hi J}i"=l
of "f/ such that
...
Let V = .U ]Xi' Xi
.=1
+ hie. Then we have
A(Eu•v n V') < 8. (2)
The inclusion V C U implies that
...
1: hi = A(V) ~ A(U) < 0: + 8,
i=1
have
n m
L (t(y; + k;) - 1(y;)) ~ L (f(Xi + hi) - 1(Xi)) . (7)
i=1 i=1
for all x EA and all n, and s' (x) exists and is finite for x EA. For any
x E]a, b[ and every h> 0 such that x + h E]a, b[, it follows from the
equality
Six + h) - six) sn(x + h) - sn(x) rn(x + h) - r,,(x)
h = h + h
that
S,,(x + h) - sn(x) S s(x + h) - s(x) .
h - h '
and this inequality implies that s~ (x) ;;;;; s' (x) for all x EA. The inequality
s~ (x) ;;;;; S~+1 (x) is clear, and so we have
1 This is not the theorem ordinarily called "FUBINI'S theorem", which deals
with product measures and integrals and will be taken up in Chapter Six.
268 Chapter V: Differentiation
lim s~ (x) =
11-->00 i= 1
exists a. e., and it remains to show that lim s~ (x) = s' (x) a. e. Since the
,,~oo
1:
[s(b) - s"k(b)J < 00.
k=l
For each nk and for every x EJa, be, we have
o~ s(x) - S"k(X) ~ s(b) - s"k(b) .
The terms on the left side of this inequality are bounded by the terms of
1: [s (x)
00
f{J(x) = Ix
1- x if
I(x) =
1
0 if x<an ,
In (x) = Un if X = an ,
Vn if x> an.
00
Prove that s (x) = 1: In (x) has a finite derivative a. e., and that s' (x) = 0
.. =1
a. e. [Hint. The function s has finite variation; find each V; by using
the numbers Iunl and Ivnl. Then apply (17.18).]
(17.22) Exercise. Find a real-valued strictly increasing function I
on R such that I' (x) = 0 a. e.
(17.23) Exercise. Let I E~r([a, b]). Suppose that there exist real
constants ex < f3 such that
each n, define
.....
Vn = I: ~Bn.k '
k=l
where Bn,k = 1([X~~l' x~")]). Prove that vn(y) -+ v(y) for ahnost all y ER
and apply B. LEVI'S monotone convergence theorem.] The function v is
known as the Banach indicatrix 01 I.
(17.35) Exercise. For an interval [a, b] c R, let ~ ([a, b]) denote the
set of all complex-valued functions I on [a, b] such that V; 1< 00 and
I(a) = O. For I E~ ([a, b]), define II1II = Yab I. Prove (a)-(c) and answer
(d) and (e).
(a) With pointwise operations ~ ([a, b]) is a complex linear algebra.
(b) For I E~ ([a, b]), the inequality 11/11 .. ~ IIIII holds.
(c) With the above norm, ~ ([a, b]) is a Banach space.
(d) Is it true that Il/gll ~ II1II . Ilgll for all I, g E~ ([a, b]) ?
(e) Find the least cardinal number of a dense subset of ~ ([a, b])
in the topology defined by the variation norm.
(17.36) Exercise. Let x be a real number and let I be a real-valued
function defined in a neighborhood of x. The upper [lower] first and
second symmetric derivatives of I at x are defined to be the limits superior
[inferior] of the expressions
f(x + h) - f(x- h)
(1)
2h
and
I(x + h) + f(x - h) - 2f(x)
(2)
hI
as h t 0, respectively. These derivatives are denoted by D11 (x) and
IJ21(x) Wd (x) and lI21 (x)] respectively. If Dd (x) = lI11 (x) [IJ 21(x)
= lI 2 /(x)], we call this common value the first [second] symmetric deriv-
ative 01 I at x and denote it by Dd(x) [D 2 /(x)]. Prove the following.
(a) If f' (x) exists, then so does D11 (x) and they are equal.
(b) The converse of (a) is false.
(c) If f' exists and is finite in a neighborhood of x and I" (x) is finite,
then D 2 /(x) exists and is equal to f" (x). [Use the mean value theorem
on (2) as a function of h.]
(d) D 2 /(x) may exist even when I is continuous only at x.
(17.37) Exercise: More on convex functions. Let I be an open interval
in R and let I be a convex function [see (13.34)] defined on I.
(a) Prove that 1'+ (x) and 1'- (x) exist and are finite for all x EI,
also that 1'+ and 1'- are nondecreasing functions and 1'- ~ 1'+ on I.
Thus f' exists and is finite a. e. on I. [Hints. For x < y < Z, we have
272 Chapter V: Differentiation
Proof. For x' > x, the equality IF(x') - F(x)1 = Iff(t) dtl holds;
therefore it is clear from (12.34) that F is uniformly continuous. If
a = Xo < Xl < ... < X" = b, then we have
are dense in ~l([a. b]) (13.23). Consider the function sgnf; for every
positive integer m, select a step function am of the form (1) such that
Since amj has the form (1), the absolute value of the last integral in (3)
has the form
..
~ 1: Irx-kl . IF(Xk) - F(Xk_l)1
.
k=l
5(x)=Js(t)dt= };IX"JEAk(t)dt.
a "=1 a
Theorem (18.2) implies that
5'(x)=s(x) a.e.in]a,b[. (1)
For a nonnegative function I in ~l' let (S,,):=1 be a nondecreasing se-
quence of simple measurable functions such that lim s" (x) = I (x) for all
........ 00
1 We have actually proved a little more than claimed in the theorem. We have
n U.. ;
00
9" (x) = 1 a.e. on the set if A is nonmeasurable, the nonmeasurable set
(n
..-1
.. -1
U.. ) n A' does not have measure O.
lS*
276 Chapter V: Differentiation
x
X E [a, b] (11.35). Write S .. (x) = J s.. (t) dt for n EN; B. LEVI'S theorem
II
It follows that
s+h %+11 s+h
!f I/(t) - (XI dt - !f I/(t) - fJnl dt ~ !f ; dt = ; ,
and in turn that
s+h
!f %
I/(t) - 0:1 dt - I/(x) - 0:1
%+h s+h
~ !f I/(t) - 0:1 dt - !f I/(t) - fJnl dt
lim~f
11+0 h
I/(t) - 0:1 dt = I/(x) - 0:1
s
lim~
litO h
f "I/(t) - 0:1 dt = I/(x) - 0:1
x-II
(i) lim
lito
hI f I/(x + t) + I(x - t) - 2/(x)1 dt =0
o
lor almost all x E]a, b[.
Proof. For fixed x E ]a, bL write
II
%+11 %
Applying (18.4) with 0: = I(x), we see that (i) holds for almost all
x E]a, b[. 0
278 Chapter V: Differentiation
(18.6) Definition. Suppose that I E~l ([a, bJ) and that x EJa,b[.
Then x is called a Lebesgue point lor I if (18.5.i) holds. The set of all
Lebesgue points for I is called the Lebesgue set lor I.
It is obvious that every point of Ja, b[ at which I is continuous is
a Lebesgue point for I. However I need not be continuous anywhere,
and yet almost every point is a Lebesgue point. The Lebesgue set plays
an important role in the theory of Fourier series and integrals, as we
shall see in (18.29) and (21.43).
We now inquire into the "reverse" [not really a converseJ of Theorem
(18.3). Given a continuous function cp that is differentiable a.e., is it
true that
x
cp(x) - cp(a) = J cp'(t) dt?
a
That is, does the fundamental theorem of the calculus hold with the
Riemann integral replaced by the Lebesgue integral and differentiability
replaced by differentiability a. e.? The following examples answer this
question with an emphatic "no".
(18.7) Example. As in (8.28), let 1jJ be LEBESGUE'S singular function.
Then 1jJ is continuous on [0, 1J and it is clear that 1jJ' (x) = for all x °
in [0, IJ that are not in P [CANTOR'S ternary set]. Thus 1jJ' (x) = a.e. °
It follows that
1jJ(I) -1jJ(0) = 1 =!= ° =
o
1
J 1jJ'(x) dx.
The next example seems not as geometrically obvious as (18.7),
but it is much more dramatic.
(18.8) Example [Adapted from RIEsz-NAGY]. We will exhibit a
real-valued function F on [0, IJ such that F(O) = 0, F(I) = 1, F is con-
tinuous and strictly increasing, and F' (x) = °
a.e. We define continuous
functions Fn, n = 0, 1, 2, ... , inductively as follows. First, let (tn);:'=l
be any sequence of numbers in JO, 1 [. Let Fo (x) = x, and define Pi (0) = 0,
Pi(I) = 1, and Pi (21) = I-t
-2-
1
• °+ l+t
-2-
1
,1; and define FI to be
linear on [0,
~] and [~, 1]. Suppose that Fo' FI> ... , Fn have been
defined. Then we define:
F.
n+l
+ 1)
( 2k2n+1 = 1 - tn+1 F. (~)
2 n 2n + 1 +2tn+1 F.n (~)
2n
strictly increasing. Indeed, for 0 < t < 1 and any oc, Psuch that oc < P,
the inequalities
R
{'- ( -12 t OC + -12+-t{R)
- - ' -_ -12 t (R
- - {,-OC) > °
and
I-t l+t l+t
- 2 - OC +-2- P - oc= -2-(P- OC) >0
(2k+l)
Fn~ <Fn+l ~
(2k+l)
for k = 0, 1, ... , 2n- 1. Hence, again by linearity, the inequality
Fn (x) ~ Fn+ 1 (x)
holds for all x E [0, 1]. Thus the sequence (Fn (X)):=1 converges for all
x E [0, IJ; let
F(x) = lim Fn(x) .
n-->oo
(2)
where k n E{O, 1, ... , 2n- I} (n = 0, 1,2, ... ). Thus we have OCn = OCn+1
1 1
and Pn+1 = Pn - 2n +l ' or Pn+1 = Pn and OCn+1 = OCn + 2n +l • In the first
case, we go to the left in proceeding from (oc n, Pn) to (OCn+I' Pn+l);
in the second, we go to the right.
280 Chapter V: Differentiation
r -;0+1
F(PHI) - F(aHI) = FHI (PHI) - FHI (aHI)
= F" (P,,) - F" (a,,) + 1 +;k+l F" (P,,)}
= 1 -;0+1 (F(P,,) _ F(a,,)) . (3)
If we go to the left in going from (a", P,,) to (aHI' PHI)' then a like
computation shows that
F(PHI) - F(aH1) = 1 +2t 0+ 1 (F(P,,) - F(a,,)) . (4)
"g e+tt
A simple induction shows that
F(P,,) - F(a,,) = k
) ,
and if all t" are less than a number less than 1, F is obviously continuous.
[We will not bother with exploring necessary conditions for the conti-
nuity of F.]
We now look at the derivates of F. Consider any dyadic rational
;" such that 0 ~ ;" < 1. Define a sequence (a,., P,,) satisfying (1) and (2)
for which ap = ap+l = ... = ;" : we care not what al> PI'··· ,ap-I,pp-I
are. Then we must have
I 1
PI>+s = 2P + 2"+1 for s = 0, 1, 2, ....
Applying (4), we see that
F{PH') -F(;,,)
/12 (1 + tp+;) (F(Pp) -
s 1
I = 2P+ 8 F(ap))
PH. - 2P
II (1 -
,,=1
tIS) = O. For these two results it is sufficient that lim tIS be
~oo
each value is assumed for infinitely many k's, i. e., x is not a dyadic
rational. For each n, there is a unique l such that ;" < x < 1 ~ I ; let
1 1 +1 .
cx" = 2" and {J" = ~. In fact these numbers are gIVen by
In the second case the factor 1 ~ t" is replaced by 1 ~ t" in the preceding
line, and so we obtain
F(fJ,,} - F(oc,,} =
1
2" (1 + (-W"t,,)
2
• (F. ({J ) _ F. (
,,-1 ,,-1 ,,-1 IX,,_I
)
2"
II
verges to 1 if and only if lim tIl = O. Thus if lim t" > 0, the product
~oo ~oa
lI(1 + (-I)"kt k) = 0
k=l
x
E 2= .
00
for almost all numbers x =
k=l
We now identify the class of functions that are indefinite integrals of
functions in ~l.
(18.10) Definition. Let I be a complex-valued function defined on a
subinterval J of R.l Suppose that for every e > 0, there is a () > 0 such
that
" k) - I(Ck)1 < e
(i) EI/(d
k=l
for every finite, pairwise disjoint, family {JCk, dk[}k=l of open subinter-
vals of J for which
"
E (dk - ck ) < () .
(ii)
k=l
Then I is said to be absolutely continuous on J.
(18.11) Examples. (a) Theorem (12.34) shows that the indefinite
integral of a function in ~l([a, b]) is absolutely continuous. Our next
project will be to prove that every absolutely continuous function is an
indefinite integral.
(b) LEBESGUE'S singular function 'IjJ is not absolutely continuous.
00
1 Recall that by (6.1) J can be open, closed, or half-open, and that J can be
bounded or unbounded.
§ 18. Absolutely continuous functions 283
1
z
00 n
see that 1: C"p(b,,) - tp(a,,)) = 1, and so 1: (tp(b,,) - tp(a,,)) ;;:;; for
k=1 k=1
n
sufficiently large n, while 1: (b" - a,,) is arbitrarily small.
k=1
(c) None of the functions F of (18.8) is absolutely continuous. This is
most easily seen from Theorem (18.15) inlra.
We first set down some elementary properties of absolutely continuous
functions.
(18.12) Theorem. Any complex-valued absolutely continuous lunction I
defined on [a, b] has finite variation on [a, bJ.
Proof. Let () > 0 satisfy the conditions of Definition (18.10) for e = 1.
Let n be any integer such that n > b ~ a ,andsubdivide[a,b]bypoints
b-a
a= Xo < Xl < ... < Xn = b such that x" - X"_l = -n- < () for k = 1, 2,
... , n. From our choice of () it follows that V~~.t;;:;; 1 for all k. Thus
n
V:I = 1: V~~.t ;;:;; n. 0
k=1
(18.13) Theorem. Any absolutely continuous lunction I on [a, b]
is continuous, and can be written as
(i) 1= 11 - 12 + i(f3 - 14) ,
where the I; are real, nondecreasing, and absolutely continuous on [a, b].
Proof. If I is absolutely continuous, then the continuity of I and the
absolute continuity of Iml and Rei are obvious. For a real-valued,
absolutely continuous function g, write gl (x) = V; g. Then g = gl - (gl - g),
and the proof will be complete upon showing that gl is absolutely con-
tinuous. [Note that gl and gl - g are nondecreasing (17.16).] For an
n
arbitrary e > 0, let () > 0 be so small that 1: Ig(d,,) - g(c,,) I < ;
k=1
whenever the pairwise disjoint intervals ]c", d,,[ are such that
.
1: (d" - c,,) < () . (1)
k=1
Let {Jc", d,,[}Z=1 be a fixed system of pairwise disjoint intervals satis-
fying (1). Since g has finite variation, there is for each k E{I, 2, ... , n}
a subdivision c" = a~k) < a\k) < ... < al:> = d" such that
Hence we have
.. n n 1~-1
f (x + ~) - f (x)
I.. (x) = 1
n
for n = 1,2, 3, ... and a ~ x ~ b. Then (I..) is a sequence of nonnegative
measurable functions and
lim I.. (x) = f' (x)
........ 00
for almost all x E]a, b[ [see (17.12)]; thus f' is measurable. By FATOU'S
lemma (12.23) and (12.44) we have
J
....
f' (x) dx =
.. ft-io.OO n-+oo
=
..
~ ~ [ntf (X)4X-n7~(X) dX]
'" ~ [ntf(b) dx - nJ'-f(4) 4%]
= I(b) - I (a) .
This proves our first assertion. The second assertion plainly follows from
the first and the fact that g can be expressed as a linear combination of
four nondecreasing functions. 0
(18.15) Theorem. Let I be an absolutely continuous complex-valued
lunction on [a, b] and suppose that f'(x) = 0 a.e. in ]a, be. Then I is a
constant.
Proof. We lose no generality by supposing that I is real-valued, for
otherwise we examine Rei and Iml separately. We will show that
I(c) = I(a) for all c E]a, b]. Thus let c E]a, b] and e > 0 be arbitrary.
Select a number ~ > 0 corresponding to the given e for which the condi-
tion in the definition of absolute continuity (18.10) is satisfied. Let
§ 18. Absolutely continuous functions 285
E = {x E ]a, c[: f' (x) = o}. Clearly A(E) = c - a. For each x EE there
exist arbitrarily small h > 0 such that [x, x + h] C ]a, c[ and
ell
I/(x+ h) - l(x)1 < c _ a . (1)
The family of all such intervals [x, x + hJ is a Vitali cover of E, and so by
VITALI'S theorem (17.11) there exists a finite pairwise disjoint family
{[Xk' Xk + hkJ}i:~1 of these intervals such that
We may [and do] suppose that Xl < X 2 < ... < x n . It follows from (2)
that the sum of the lengths of the open intervals
]a, xI [, ]XI + hI' x2 [, ••• , ]Xn + hn' c[
n
complementary to k~l U [Xk' Xk + hk] is less than 15, and so, in view of our
choice of <5, we have
n-l
I/(a) - I (Xl) I + I: I/(Xk + hk) - I (Xk+l) I + I/(xn + hn) - l(c)1 < e. (3)
k~l
Therefore
x x
I (x) = hex) + g(x) = heal + J f' (t) dt = I (a) + J f' (t) dt
a a
for all x E [a, b]. 0
(18.17) Theorem. A lunction I on [a, b] has the lorm
x
I(x) = I (a) + J rp(t) dt
a
lor some fjJ E5~\([a, b]) il and only il I is absolutely continuOZls on [a, b].
In this case we have fjJ (x) = f' (x) a. e. on ]a, b[.
Proof. This is just a summary of (18.3), (18.11.a), and (18.16). 0
Indefinite integrals on R can be characterized in much the same way.
(18.18) Theorem. A lunction I on R has the lorm
x
(i) I(x) = J rp(t) dt
-00
lor some fjJ E~1 (R) il and only il I is absolutely continuous on [- A, A] lor
all A > 0, V~ool is finite, and lim I(x) = 0.
x-+-oo
Proof. Suppose that I has the form (i). Then I is absolutely con-
tinuous on [-A,A] by (12.34), and by (18.1) we have
V~ool = J \fjJ(t) \ dt < 00.
R
for every finite, pairwise disjoint, family {Jc k, dk[}k~1 of open subintervals
of [a, b] for which
n
(ii) E (d k - ck ) < 15 .1
k~1
and so
00
1 Thus we can replace the condition (18.10.i) by the apparently stronger condi-
Plainly we have
epeE) c ep (kgl Jc", d,,[) kgl ep(Jc", d,,[) .
= (2)
It is also evident that
{ JC(n)
II
d(n)[
I
J
(n) d(n)[} _ 17),
, . . . ,GIn' In -;;zJnJ
with the following properties. First, the intervals comprising each f!)n
are pairwise disjoint. Second, the inequalities
In
'"' W (c(n) den») > ~ (3)
£..oJ '" k ' k = "0
k=1
hold for all n. Third,
(4)
n=1 k=l
For each n, and for y E [IX, fJJ, let Nn(Y) be the number of intervals
Mnl, d~n)[ in f!)n that have nonvoid intersection with ep-l({y}). As the
intervals Jc~n), 4n )[ are pairwise disjoint, it is evident that
(5)
here v is the Banach indicatrix of ep, defined in (17.34). It is obvious that
In
Nn = 1: ;"'(]C~n), dkn )[).
k=1
Since ep ([4n), din)J) is a closed interval [whose measure is w'" (c~n), 4n »)]
and since (ep (Jc\:'\ d~n) [))' n ep ([cin ), 4n )J) contains at most two points,
N n is Borel measurable [actually the pointwise limit of a sequence of
continuous functions] and
(6)
E = i Ogi t~l
l Mn), dkn )[) •
It is obvious that Xo EE. From (4) and (10.15) we infer that l (E) = o.
Since ep is an N-function, we have l(ep(E) = O. We have proved above
that ep(E) :J A n A;, so that
l(A) = l(A n A;) = O. (7)
The definition of A and (7) show that lim N n (y) = 0 for almost all y in
1<->00
Since (6) and (8) contradict each other, the proof is complete. 0
The hypothesis in (18.25) that ep have finite variation is essential,
as the following example shows.
(18.26) Example. Consider any interval [a, b] and a perfect nowhere
dense subset F of [a, b] that contains both a and b. The measure l(F)
may be zero or positive. Write the open set [a, b] n F' as U Jan, bn [, where
00
n~l
k~l
We leave both proofs to the reader. To see that g is an
N-function, consider any set E C [a, b] such that l(E) = O. Using
(2.15.i), we have
00
and so
+ E A(g(E n ]ak' bk[)) =
00
A(g(E)) ~ A({O}) O.
k=1
E tk =
00
If 00, then g certainly fails to be absolutely continuous, since it
k=1
has infinite variation.
(18.27) Discussion. We close this section by giving a famous applica-
tion of (18.5) [which in fact led LEBESGUE to the definition (18.6) of the
Lebesgue set]. Consider a function I E ~1([-:n;, :n;]) and its Fourier co-
efficients I(n) (16.33). The uniqueness theorem (16.34) tells us that I is
determined [as an element of ~l([-:n;, :n;]), of course] by the function I
defined on Z. This theorem leaves untouched the problem of reconstruct-
ing I from f. This problem is important not only for its own sake but also
for applications to physics, chemistry, and engineering, since many data
obtained by spectroscopy, X-ray analysis, and the like, are nothing other
than Fourier coefficients of functions which one wishes to determine.
The simplest way to try to recapture I from I is by means of the Fourier
series 01 I, the partial sums of which are defined as
n
(i) snl(x) = E I(k) exp(ikx) (n = 0, 1,2, ... ) .
k=-n
In order to rewrite (i) and other expressions to be defined shortly, we
define I on the entire line R by periodicity: I(x + 2k:n;) = I(x), for
x E [-:n;,:n;[ and k EZ [the number I (:n;) has no importance for
I E~l([-:n;, :n;])].
We then have:
n
(ii) s,d(x) = i;
k=-n
2~ f I(t) exp(-ikt) dt· exp(ikx)
-n
The function defined by (iii) is called the Dirichlet kernel and is denoted
by Dn(t). Thus we may write
n
-"
The expression [... J in (vi) is called the Fejer kernel; it is denoted by
Kn (t); and one easily proves that
1 [sin(f(n+l)t)]2 1
1
if sinh-t) =l= 0 ,
(vii) Kn(t)= n+l sin(ft)
n +1 if sin (ft) = O.
The reader can easily verify the following:
(viii) Kn (- t) = Kn (t) ;
(ix) 0 ~ Kn (t) ~ n + 1;
n
(x) 21n
-n
f Kn(t) dt = 1;
(18.28) Theorem. Let p be a real number such that 1 ;:;;;; p < 00, and
let I be a lunction in ~p([-~, ~]). Then
(i) lim III - O"nilip = 0.
n->-oo
21n J I/(x) - I(x - t)i Ig (x) I dx;:;;;; III - I-tllp' IIgll p'
-n
(3)
Going back to (1), we therefore have
;:; ; "' - '-tllp·
n n
=_1
2n
J+- J 1
2n
Itl;:;;~ Itl>"
;:;;;; sup{II1 - I-tllp: It I ;:; ; 15} 21n J Kn(t) dt
Itl;:;;;"
+ 2","p 2~ J Kn(t) dt. (4)
Itl>"
By (13.24), the supremum in (4) is arbitrarily small if 15 is sufficiently
small. By (18.27.xii), the limit of the last expression as n -+ 00 is zero, no
1 We will prove in § 21 that this iterated integral is well defined.
294 Chapter V. Differentiation
matter how small 6 may be. That is, the first expression in (4) is arbitrar-
ily small for n sufficiently large. This implies by (15.1) that
For p = 1, use (3) and repeat the argument with obvious changes. 0 1
Our point in going through (18.28) was to lead up to the much
subtler fact that ani converges to I not merely "in the mean" [i.e., in the
~p norm] but also pointwise almost everywhere.
(18.29) Theorem [LEBESGUE]. Let I be a lunction in ~1([-n, nJ).
Then il x is in the Lebesgue set 011, we have
(i) lim ani (x) = I (x) .
n->-oo
+~(x,
and we write the number ~ (x, n) as a. Theorem (18.5) shows that
t) ~ 0 as t ~ 0 [x is in the Lebesgue set of II]. Consider any e > 0
and choose IX>O so that j+~(X, t)j < e for ItI ~ IX. Next use (18.27.xi)
n
ex
~ no and IX ~ t ~ n imply IKn(t) I < a: I . (1)
Note that
n ~ no implies n~(x, !) < e. (2)
It is easy to see that
n
51 ~ j Icp(x, t)1 (n + I)dt = (n + I)q> (!) ~ 2nq> (!) < 2e. (5)
o
Using (18.27.xi), we have
" "
52 = j Icp(x,t)1 Kn(t) dt~ j Icp(x, t)1 n: I fdt.
lin lin
Hence
:n;2
52~ n+1 q>(x,ex)~
I
+
2:n;2
n+1
j "q>(x,t)r dt 3
lin
e 2:n;2 j "
:n;2
:s;: - - -
-n+lex
+-
n+1
- q>(X' t) r 3dt
lin
<n2 e+54 •
Finally we have
"
5 :s;: ~j er2 dt = 2:n;2e [n-~] < 2:n;2e • n < 2n2e
4 - n+1 n+1 ex n+1
lin
and hence
(6)
The relations (3), (4), (5), and (6) show that n ~ no implies
2nlQ'nl(x) - l(x)1 < 3e + 3n2 e < 33e. 0
(18.30) Exercise. Let E be a subset of R [not a priori measurable]
such that for some positive real number b, the inequality .A. (E n I) ~ b.A. (I)
holds for all intervals Ie R. Prove that E' is a set of measure zero.
[Use (18.2).]
296 Chapter V. Differentiation
tk = ~ (b k - ak) + ek'
Then g has a finite derivative at no point of F.
Thus if A(F) > 0, we have an N-function that fails on a set of positive
measure to have a derivative. [Hints. Show first that if g is differentiable
at x EF, then g' (x) = 0. Show next that a derivative of g is nonzero
if x = ak or x = bk • For other points x EF and for each index n, there is
an interval Jaj(n), bj(n) [ among the intervals Jal> blL ... , Jan, bn[ at mini-
mum distance from x [there are obviously at most two such intervalsJ.
A moment's thought shows that lim j (n) = 00, that
n.-,.oo
that ({J E~i([a, b]) and that J ((J(t) dt < 1. Let 1jJ(x) = J ((J(t) dt, and
a a
%
({J"
(t) dt =
n.
"
Thus 1jJ~ (x) = 00; similarly for 1jJ'- (x) .]
(b) [A. ZYGMUND]. For I E(t"([a, b]), define
E = {x: a ~ x < b, D+/(x) ~ O}.
Suppose that t (E) contains no interval. Then I is nondecreasing. [Hints.
If I(c) > I(d), where a ~ c < d ~ b, choose any Yo E]t(d), I(c)[ and
define Xo as sup {x : c ~ x < d, I (x) ~ Yo}. Show that t (xo) = Yo, and hence
that D+/(xo) ~ O. This implies that I(E) ::::> ]/(d), l(c)L a contradiction
to our hypothesis.]
(c) Consider any IE (t"([a, b]). Suppose that D+I is nonnegative
almost everywhere on [a, b[. Suppose also that the set B, defined by
B = {x: x E[a, bL D+/(x) = - oo}, is countable. Then I is nondecreasing.
[Hints. Let A = {x: x E]a, b[, D+/(x) or D+/(x) is negative and finite}.
Let1jJbe as in part (a) for the set A, let a (x) = 1jJ(x) +x, andletg= 1+ ea,
§ 18. Absolutely continuous functions 299
I (x) = j
/(C) +
I(c)-fq;(t)dt
j q;(t) dt
••
for
for
x
x<c.
~ c,
"
[The "if" part is a tiny extension of (13.35). For the "only if" part,
combine (17.37.b), (18.16), and (17.37.a).]
(18.44) Exercise. Let I be a complex-valued function of period 2:rt
defined on R such that I E~1 ([ -:rt, :rtJ). Suppose that I is continuous at
every point of [a, b]. Prove that (an l)':=1 converges to I uniformly on
[a, b]. [Use uniform continuity and the Fejer kernel as in (18.29).]
(18.45) Exercise. Use the uniform boundedness principle to prove
that there exists a real-valued continuous function of period 2:rt defined
on R whose Fourier series diverges at O. [Let
'-13 = {t E<rr([-:rt, :rt]): I(-:rt) = I(:rt)}.
Then '-13, with the uniform norm, is a real Banach space. For each n EN,
n
tion is as in (18.27). Prove that T..E'-13* and that IIT..II= 21n JID.,(t) Idt
-n
for each n. Next show that lim IIT..II = 00. Finally conclude from (14.23)
...... 00
[Compare this with the definition of unl in (18.27.v).] The function IXrl
1:
00
is called the rtl• Abel sum 01 the series I(k) exp (ikx). Using the uniform
k=-oo
convergence of the series in (i). show that
n
(ii) IXrf(X) = 2~
-n
f I(x - t) L=~oo rlkl eXP(ikt)] dt.
The expression [... ] in (ii) is called the Poisson kernel, and is denoted
per, t). A simple computation shows that
''') P()
(11l r, t = 1 + 1'21_- 1'2
2rcos(t) .
(vi) 2~ f P(r, t) dt = 1,
-n
.., (1 - r2) 2r sin(t)
(vn) P (r, t) = - (1 + r2 _ 2r cos (t))2
Consider the functionF (x) = 21;n; f ! (t) dt, and consider any xE]-n, n[
-n
at which the symmetric derivative
lim F(x+h) -F(x-h) = D F(x)
hj-O 2h 1
= 21;n; fF(x-t)P'(r,t)dt
-n
n
= - 2~ f F(x+ t) P'(r, t) dt
-n
= _1_
2;n;
f n
F (x + t) - F (x - t)
2 sint
M( )d
r, t t ,
-t<
so that
n
A(I (En) ~ }; A(I (En n In. k) < ({3 + 8) }; A(In, k) < ({3 + 8) (A (E) + 8)
k=l k=l
~ }; [J II'(x)1 dx + d(Bn)]
n=l B..
= J I!,(x) I dx + deB) .J
B
[a, b]. [Hints. For [c, d] C [a, b], let B = {x E ]c, d[: I' (x) exists and is
finite}, and let A = [c, d] n B'. Then It(d) - t(c)1 ~ A(t([C, dJ)) = A(t(B))
d
+ A(f(A)) = A(t(B)) ~ J If'(x)ldx. Recall that I' E~l([a, b]).p
c
for E Ed. Clearly '/I is complex-valued, '/1(0) = 0, and '/I is count ably
additive (12.32). Thus '/I enjoys two essential properties of a measure.
Since '/I can assume arbitrary complex values, it is not always a measure
in the sense of (10.3). This leads us to define and study signed measures
and complex measures.
(19.1) Definition. Let (X, d) be an arbitrary measurable space. An
extended real-valued function '/I defined on d is called a signed measure if
(i) '/1(0)=0
and
(ii) '/I ("Ql E..) = ~1 '/I (E ..)
Our first goal is to show that just as a function of finite variation can
be expressed as a linear combination of four monotone functions, so a
complex measure can be expressed as a linear combination of four
measures. It is obvious that any complex measure v can be expressed
uniquely in the form v = VI + iV2 where VI and V2 are real-valued signed
measures; simply let VI (E) = Re V(E) and V2(E) = 1m V(E) for all E Ed.
We therefore take up signed measures first.
(19.3) Theorem. Let V be a signed measure on a measurable space
(X, d). We have:
(i) if E, FEd, Iv(E)1 < 00, and FeE, then Iv(F)1 < 00;
(ii) if An Ed (n = 1, 2, 3, ... ) and Al C A2 C ... CAn C .. "
then
lim V(An)
n--)o-oo
= U An);
V (n=l
(iii) if An Ed (n = 1, 2, 3, ... ), if Al:J A 2 :J···:J An:J"', and
Iv (AI) I < 00, then
Proof. To prove (i), observe that v(E) = v (F) + v(E n P). In order
that V (E) be finite, it is necessary and sufficient that both summands on
the right side be finite. Conclusions (ii) and (iii) are proved by repeating
verbatim the proofs of (10.13) and (10.15), respectively. In the proof of
(iii) we use (i) to write V(AI n A~) = V(Al) - v(An). 0
(19.4) Definition. Let v be a signed measure on a measurable space
(X, d). A set P Ed is called a nonnegative set for v if v (P n E) ;:;:; 0 for
all E Ed. A set M Ed is called a nonpositive set for v if v (M n E) ~ 0
for all E Ed. Note that 0 is both a nonnegative set and a nonpositive
set for v. If P is a nonnegative set for v and P' [the complement of P in X]
is a nonpositive set for v, then the ordered pair (P, P') is called a Hahn
decomposition of X for v.
(19.5) Lemma. Let v be a signed measure on (X, d) and suppose that E
is a set in d such that 0 < v (E) < 00. Then there exists a set 5 Ed such
that 5 C E, 5 is a nonnegative set for v, and v(5) > O.
Proof. It follows from (19.3.i) that Iv(F)1 < 00 for allF Ed such that
FeE. Assume that no set 5 of the required sort exists. In particular,
E is not a nonnegative set for v. Let n l be the smallest positive integer for
which there exists a set F; E d such that FJ. C E and v (FI) < - ~.
n1
Then we have
v(E n F I ') = v(E) - v(FrJ > v(E) > 0,
and so, by our assumption, E nFl' is not a nonnegative set for v. As
before, let n2 be the smallest positive integer for which there exists a set
Hewitt/Stromberg, Real and abstract analysis 20
306 Chapter V. Differentiation
1
+ 1) - > 0 .
00 00
00 > v(E n F') = v(E) - v (F) = v (E) -1) v (F,,) > v (E)
k~1 k~1 n k
Thus 1; _1_
k~1 nk
< 00 and E n F' is not a nonnegative set for v. Choose
A Ed such that AcE n F' and v(A) < 0, and then choose k so large
1
that v(A) < - -nk and n" > 2. Now
that each Pn is a nonnegative set for v and that v (Pn) ~ v (An). Applying
(19.3.ii), we have v (P n E) = lim v (pn n E) ~ 0 for all E Ed. Thus P
n-+oo
is a nonnegative set for v and v(P) = IX. [Note that IX < 00.]
We next show that P' is a nonpositive set for v. Assuming the contrary,
choose a set E Ed such that E C P' and v (E) > O. Since v does not
assume the value 00, we have v (E) < 00. Apply (19.5) to obtain a set
5 C E such that 5 Ed, 5 is a nonnegative set for v, and v(5) > o. Then
5 U P is a nonnegative set for v and v(S U P) = v(5) + IX> a; this
violates the definition of IX. Therefore P' is a nonpositive set for v.
§ 19. Complex measures and the LEBESGUE-RADON-NIKODYM theorem 307
To prove our uniqueness assertion, suppose that (PI' pn and (P2 , P;)
are two Hahn decompositions of X for'll, and select E Ed. Since
E n PI n P; is a subset of both PI and P; , we have'll (E n PI n pn = 0;
similarly v(E n PI' n P2 ) = o. Thus we have v(E n Pd = 'liCE n (PI U P2 ))
= v(E np2) and v(E npn = 'liCE n (P{ UP;)) = v(E np;). 0
(19.7) Definition. Let 'I' be a signed measure on (X, d) and let (P, P')
be a Hahn decomposition of X for'll. Define '1'+, '1'-, and 1'1'1 on d by:
v+(E) = v(E n P) ;
v-(E) = -v(E n P');
and
1'1'1 (E) = v+(E) + v-(E)
for all E Ed. The set functions'll+' '1'-, and 1'1'1 are called the positive
variation 01 '1', the negative variation 01 '1', and the total variation 01 '1',
respectively.
(19.8) Theorem. Notation is as in (19.7). The set lunctions '1'+, '1'-,
and 1'1'1 are well-defined measures on (X, d). Also we have
(i) v(E) = '1'+ (E) - v-(E) lor all E Ed .1
The proof is very simple, and we omit it.
(19.9) Example. Let (X, d, f-l) be a measure space and let I be an
d-measurable, extended real-valued function on X for which J I df-l is
x
defined. Define 'I' on d by
'I' (E) = J I df-l .
E
Then we have
and
1'1'1 (E) = Jill df-l
E
for all E Ed. If P = {x EX: t (x) > O}, then (P, P') is a Hahn decom-
position of X for'll. Notice that the nonequality 1'1'1 (E) =1= 1'1' (E) I must
occur for some sets E Ed if '1'+ and '1'- are nondegenerate.
tf
(19.10) Theorem. Let 'I' be a signed measure on (X, d). Then
(i) 1'1'1 (E) = sup 1'1' (E k) I: {El> ... , En} is a measurable dissection 01 E}
lor every E Ed.
1 The expression v = v+ - v-is known as the]ordan decomposition of v. in
analogy with (17.16).
20·
308 Chapter V. Differentiation
Proof. Let E be any fixed set in d, and let (3 denote the right side
of (i). Then we have
n n
£ Iv(E,,)1 = £ Iv+(E,,) - v-(Ek)i
"=1 k=1
n
;2; £ ('11+ (E,,) + v-(E k )
k=1
n
= £
k=1
1'111 (E,,) = 1'111 (E)
for every measurable dissection {EI' ... , En} of E; hence {3 ;2; 1'111 (E).
Consider the dissection {E n P, En PI} where (P, PI) is a Hahn de-
composition of X for v. We get
(3 ~ Iv(E np)1 + Iv(E n PI)I = v+(E) + v-(E) = 1'111 (E) . 0
In view of (19.10), we make the following definition with no risk of
inconsistency.
(19.11) Definition. Let v be a complex measure on (X, d). The total
variation of v is the function 1'111 defined on d by the formula (19.1O.i).1
(19.12) Theorem. Notation is as in (19.11). The set-function 1'111 is a
measure on (X, d).
Proof. It is obvious that 1'111 (0) = O. Thus we need only show that 1'111
is count ably additive. Let (Ai)~1 be a pairwise disjoint sequence of sets
00
n 00
reverse of this inequality is obvious. Thus suppose that 1'111 (A) < 00. For
any jo EN and any measurable dissection {B1' ... , Bm} of Ai., we have
and so 1'111 (Ai) < 00 for all j EN. Let e > 0 be arbitrary. For each j,
choose a measurable dissection {Ei'l' ... , E i ,lIj} of Ai such that
~ e
1) Iv(Ei,k)1 > 1'111 (Ai) -V·
k=1
x
J Id 1'111 = ;=1
1: Pi 1'111 (EI )
~ £P; C~ 'II,,(E,))
4
= }; J Idv" , (I)
k=IX
and
fill d"" ~ fill d 1"1
x x
for all k. D
We may now make the following definition.
(19.17) Definition. Notation is as in (19.16). For I E~l(X, .91, 1"/),
define
4
°
(19.21) Lemma. Let fl and v be measures on (X, d) such that
veE) ~ fleE) lor all E Ed. II P > and I E~p(X, .91, fl), it lollows that
I E~p(X, .91, v) and J IfiPdv ~ J IfiPdfl·
x x
Proof. If I E~p (X, .91, fl), then there is a sequence (O'n)~l of .91-
measurable, nonnegative, simple functions increasing to IfiP, and so
The next lemma, which may at first glance appear rather strange, is
actually the crucial step in our proof of the LEBESGUE-RADON-NIKODYM
theorem.
(19.22) Lemma. Let fl and v be finite measures on (X, d) such that
v ~ fl. Then there is an d-measurable lunction g on X such that g (X) C [0, 1 [
and
(i) JI(I-g)dv= JIgdfl
x x
lor all I E~2(X, .91, fl + v).
Proof. For I E ~2(X, .91, fl + v), define
L(f) = J I dv. (1)
x
Since fl and v, and with them fl + v, are finite, I is also in ~1 (X, .91, fl + v);
and so by (19.21), I is in ~l(X, .91, v). Thus L is defined and finite on
~2(X, .91, fl + v). It is clear that L(rxl + (Jg) = rxL(f) + (JL(g) for all
rx, (J EK and I, g E~2 (X, .91, fl + v). Inequality (13.4.iii) shows that
1 1
IL(f)1 = I Jldvl ~(J IfI2dv)"2(v(X)"2
x X
1 1 1
~(J IfI2d(fl + v))2(v(X))"2 = II/Mv(X)"2.
X
314 Chapter V. Differentiation
[Here 11/112 denotes the norm in ~2(X, .91, I' + v).] Thus L is a bounded
linear functional on ~2(X, .91, I' + v), and so by (15.11), there is a func-
tion h E~2(X, .91, I' + v) such that
L (I) = fiJi d(1' + v) .1 (2)
X
satisfies (I' + v)(A) > 0. Then L (';A) = f Reh d(fJ, + v) - i f Imh d(1' + v)
A A
is not real. By (1), L is obviously real-valued on real-valued functions.
This is a contradiction. Similarly, if h were negative on a set B such that
(I' + v)(B) > 0, we would have L (';B) < 0, which again contradicts the
definition of L. Thus h is real and nonnegative (I' + v)-a. e.; and we
may suppose it to be so everywhere. The definition (1) of L and the rep-
resentation (2) show that
fl(l-h)dv= flhdl' (3)
x x
for all 1 E~2(X, .91, I' + v).
Next, let
E = {x EX: hex) ~ I} .
Since ';E is in ~2 (X, .91, I' + v), we may apply (3) with 1= ';E to obtain
° ~ I'(E) = f ';Edl'
x x
= f ';E(l - h)dv ~ 0.
x
~ f ';Eh dl'
°
Thus we have I' (E) = 0, and so v (E) = also. [This is our only use of the
hypothesis v ~I'.] Let g = h';E" Then g(X) C [0, 1[ and g = h almost
1 Theorem (15.11) is of course much more than we need to produce the represen-
tation (2): only the case p = 2 is needed. As is the case with many problems in-
volving 2p spaces, the case p = 2 is much the simplest, and there is in fact a proof
of (15.11) forp = 2 couched in terms of abstract Hilbert spaces. This proof is sketched
in (16.56). Using (16.56), we could then prove (19.22) and so also (19.24) and (19.27)
without recourse to (15.11). It would be then possible to prove (15.11) for all
2p (X, d, p,) such that (X, d, p,) satisfies the hypotheses of (19.27). We prefer the
proof given in (15.11), partly because it is completely general and partly because it is
constructive and classical in spirit. In § 20 we construct the conjugate space of
21 (X, d, p,), a process that apparently requires (19.27). Then the general case of
(15.11) could be proved from (19.24).
§ 19. Complex measures and the LEBESGUE-RADON-NIKODYM theorem 315
x
J(1 - gn) I d'll =
x
J
1 ~ g (1 - gn) I dft . (1)
x
J
Id'll=
x
J
1 ~g Idft· (2)
This proves (i) for nonnegative, d-measurable I; (i) for I E~1 (X, d, 'II)
and (ii) follow at once.
To prove the uniqueness of 10' let go be d-measurable and satisfy (ii).
Assume that there exists a set E Ed such that p (E) > 0 and 10 (x) > go (x)
for all x EE. For some n, we have p(E n En) > 0; if A = En En' then
we have 'II (A) < 00,0 < p(A) and 10 - go> 0 on A. Applying (12.6) and
(ii), we obtain
0< !(/o - go) dp = 'II (A) - 'II (A) = O.
A
sets (Bn):=1 in £0 such that lim fl (Bn) = ex; let D = U Bn· Since fl is
n~oo n=l
count ably additive, it is clear that fl (D) = ex (10.13). The set D is in £0,
for if C Ed and C CD, then
v(C) = v(C n B l ) + !; v(C n (Bn n B~_l» .
n=2
Since each set C n (Bn n B~_l) is in d and has v measure 0 or 00, the
same is true of C.
Now consider the set D'. We will show that for every set FeD' such
that FEd and v (F) > 0, there exists a set FI in d such that Fj c F and
(1)
If v(F) < 00, then (1) is trivial. Thus suppose that v(F) = 00 and assume
that v(G) = 0 or v(G) = 00 for every subset G of F such that G Ed.
Under this assumption, it is clear that F U D E£0. Since v ~ fl and
1 This possibly uncountable sum is defined as the supremum of the sums
E f.l (E n F), where p) runs through all finite subfamilies of ~.
FEfId
318 Chapter V. Differentiation
lim ,u (Bn) = sup{,u (B) : B E(g} = y. Let G = U Bn, and let E = D n G'.
n~oo n=l
Since '/I (G) = 0, '/I is a-finite on E' = D' U G; i. e., (iii) is satisfied. The
assertion (i) is clear since E cD. To prove (ii), assume that there is a set
BeE such that BEd, '/I (B) = 0, and ,u (B) > 0. Then we must have
G U B E(g. But this is impossible since
,u(G UB) = ,u(G) + ,u(B) > ,u(G) = y ~ ,u(G U B);
and so (ii) is proved. 0
We can now prove our final version of the LEBESGUE-RADoN-
NIKODYM theorem.
(19.27) LEBESGUE-RADON-NIKODYM Theorem. Let (X, d,,u) be
decomposable with decomposition .f7, and let '/I be any measure on (X, d)
such that '/I ~,u. There exists a nonnegative, extended real-valued, .91-
measurable lunction loon X [which can be chosen finite-valued on each
FE.f7 where '/I is a-finite] with the lollowing properties:
(i) '/I(A) = flo d,u
A
lor all A Ed that are a-finite with respect to ,u;
(ii) f I d'/l = f Ito d,u
x x
lor all nonnegative, extended real-valued, d-measurable lunctions I on X
such that {x EX: I (x) > O} is a-finite with respect to ,u;
§ 19. Complex measures and the LEBESGUE-RADON-NIKODYM theorem 319
(iii) il I E~1 (X, .91,,,) and {x EX: I (x) =1= O} is a-finite with respect
to ft, then Ito E~1 (X, .91, ft) and J I d" = J Ito dft·
x x
Also 10 is unique, in the sense that il go is any nonnegative, extended real-
valtted, .9I-measurable lunction on X for which
(iv) ,,(A) = J go dft
A
lor all A Ed such that ft(A) < 00, then 10~E and gO~E are equal ft-a.e. lor
all E Ed that are a-finite with respect to ft.
Proof. For eachF E ff, the restriction of" toF is absolutely continuous
with respect to the restriction of ft to F, and so by (19.26) there are sets
DF and EF in .91 such that: DF n EF = 0 ; DF U EF = F ; (19.26.i) and
(19.26.ii) hold for E F ; and" is a-finite on DF . If" is a-finite on F, then
we take DF = F. Since" is a-finite on Dp and ft is finite on DF, we can
apply (19.24) to assert that there is a nonnegative, finite-valued, .91-
measurable function f&F) defined on DF such that the conclusions of
(19.24) hold for ft and" restricted to DF . Now let fo be the function on X
such that for all F Eff,
x = { fbF) (x) if x EDF ,
foe) 00 ifxEEF • (1 )
Therefore
Combining (3), (4), and (5), and harking back to (12.21), we find that
v(A)=,Ev(AnF)=,E flodf-t=flodf-t. (6)
FE.'F, FE.'F, AnF A
Assertion (i) is now obvious, since both ends of (6) are countably additive.
The equality (ii) is proved by considering characteristic functions,
then simple functions, and finally passing to the limit using (11.35) and
(12.22). Equality (iii) follows upon writing I as a linear combination of
functions in ~t (X, .91, v).
It remains to prove the uniqueness of 10. Let go be as in liv). Then for
every FE.'F and every .9I-measurable subset A of DF , we have
v(A) = flodf-t= fgodf-t,
A A
and so by (19.24), 10 (x) = go (x) for f-t-almost all x EDF • Now assume that
there exists a set A Ed such that A C EF , f-t (A) > 0, and
go(x) < 00 = 10 (x) for all x EA.
From (19.26.ii), we see that v(A) > O. For each n EN, write
An = {x EA: go (x) < n}.
00
Then (A n):'=1 is a nondecreasing sequence and ,,=1
U An = A. Applying
(10.13), we infer that
0< v(A) = lim v (An) ;
"->-00
Proof. Let 'P = 1:" eX"'P,, be the Jordan decomposition of 'P [(19.13.ii)
,\=1
and (19.15.c)]. According to (19.13) and (19.20), "1' lIa, 'P3' 'P" and I'PI are
21*
324 Chapter V. Differentiation
[here we have used the fact that Ilk E ~l (X, .91, ft) for all k, which is
evident from (1)]. This proves (i). The identity (ii) follows from (i) upon
taking I = ~.A' since Ivi is a finite measure.
To prove the uniqueness of 10 in ~l (X, .91, ft), suppose that
ho E ~l (X, .91, ft) and that v (A) = J ho dft for all A Ed. Let v = Cf + iT,
A
where Cf and T are real-valued. Then
J Reho dft = Cf(A) = J Re/odft
A A
= lim
'n~oo
J 10Cfm dft .
X
(3)
§ 19. Complex measures and the LEBESGUE-RADON-NIKODYM theorem 325
II
Each (Jm has the form 1: Pi~AJ' where {AI> ... , An} is a measurable dis-
;=1
section of A and IPil ~ 1 for all j. Therefore
Now (2) and (5) together imply (iii). Setting A = X, we get (iv). 0
(19.37) Note. The reader should find it illuminating to compare the
statement and the proof of (19.36.iii) with the corresponding result about
absolutely continuous lunctions (18.1).
(19.38) Corollary. Let v be a complex measure on (X, d). Then there
exists an d-measurable lunction 10 on X such that:
(i) Ito I = 1;
(ii) v(A) = J lodlvl loraU A Ed;
A
and
(iii) J I dv = J!fo d Ivl lor aU IE 5:.1 (X, d, Ivl) .
x x
Also,
(iv) I J I dvl ~ Jill d Ivl lor aU I E5:.1 (X, d, Ivl) .
x x
Proof. Obviously we have v ~ Ivl. Define 10 as in (19.36); then (ii) and
(iii) are immediate. LetA = {x EX: 110 (x) I < l}andB = {x EX: 110 (x) I > I}.
Apply (19.36.iii) to get
J (1 - 1/01) dlvl = 0 = J (1/01- 1) dlvi .
A B
and so
J (1 - 11) dv = 0 . (3)
enB'
The function 1 - 11 is positive on B', and so the equality (3) implies that
'1'1 (C) = v(C n B') = 0, as desired. For an arbitrary C in d such that
00
Then
(i) /t2 ~ /to
1 There is a proof of (19.42) in S. SAKS, loco cit. (18.42), which does not use the
LEBESGUE-RADON-NIKODYM theorem.
§ 19. Complex measures and the LEBESGUE-RADON-NIKODYM theorem 329
and
(1'1') --=--.--
dps dpB dP1 p,-a e
dpo dp1 dpo 0 ••
(i) ac(t) = 0 j
t([O,t[) il t> 0,
il t = 0,
-t([t,OD if t < o.
Then ac is a nondecreasing, real-valued, left-continuous function on R. Also,
we have lim ac(t) > - 00 if and only if t(]- 00, O[) < 00 and lim ac(t) < 00
~-oo ~oo
£(A)=
{
°I ifif °O~A
EA ,
,
it is easy to see that the corresponding ex is not right continuous at 0. The
choice of the definition of ex was in this respect arbitrary ; it could as well
°
have been defined so as to be right continuous and nondecreasing. Also,
the choice that ex (0) = is an arbitrary normalization. For finite measures
£ [i. e., £(R) < 00], it is often more convenient to normalize ex so that
lim ex(t) = 0. As in (8.20), we shall use the term normalized nondecreasing
I~-(X)
_co
holds for n = 1, 2, . We complete the proof by showing that
lim Stz(ln) = ex (t). If L1n is a subdivision of [-e1> t] such that
{-e n,O,t-en}CL1,
then
Stz (In) ~ U (In' ex, L1n)
~ (ex(t) - ex(t - en)) + (ex(t - en) - ex (0)) + (ex(O) - ex(-en ))
= ex(t) - ex(- en) -+- ex(t) - ex(O) = ex(t)
and
Stz (/n) ~ L (In' ex, L1n) ~ ex (t - en) - ex (0) -+- ex (t -) = ex (t) .
[Recall that ex is left continuous.] Thus
ex (t) = lim Stz (In) = lim
n~oo ~OOR
J IndAtz = f3 (t) .
A similar argument shows that f3(t) = ex(t) for t < O. Relation (i) now
follows from the definition of f3. 0
(19.48) Theorem. The mapping
(i) t -+- ex
defined in (19.45) is a one-to-one mapping 01 the set 01 aU regular Borel
measures on R onto the set 01 aU nondecreasing, real-valued, lelt continuous
lunctions ex on R such that ex (0) = O. The inverse 01 this mapping is the
mapping
(ii) ex -+- Atz.
Thus every regular Borel measure on R is a Lebesgue-StieUjes measure.
Proof. Let ex be given. Theorem (19.47) shows that ex is the image of
Atz in (i) and so this mapping is onto. Suppose that ex is also the image of
the regular Borel measure t. Now (19.47.i) and (19.45.i) show that
Atz([a, b[) = ex(b) - ex(a) = t([a, b[) (1)
for all a < bin R. Each open subset U of R can be expressed as a countable
disjoint union of sets of the form [a, b[, e.g., ]0, 1[= ,,=1 n +1 1 ' ~[,
n
and U[
so (1) implies thatAtz(U)=t(U) for all open UCR. Since Atz and t are both
regular, this implies that Atz(E) = t(E) for all E EfJI(R). Thus A.. = t and
the mapping (i) is one-to-one. The rest is clear. 0
332 Chapter V. Differentiation
as (19.47.i) shows. 0
(19.53) Theorem. The lunction IX is absolutely continuous (18.10) in
every interval [- p, P], PEN, il and only il the corresponding measure t is
absolutely continuous with respect to Lebesgue measure k 2
1 Theorems (19.50) and (19.48), together with (12.56), prove that the Riemann
integral is the only invariant nonnegative linear functional on ~oo(R), up to a positive
constant. See (8.16) supra.
2 The a-algebra Jt). of course contains ~ (R). In applying the definition of
absolute continuity of measures we here consider the measure spaces (R, ~ (R), £)
and (R. ~ (R). J.).
§ 19. Complex measures and the LEBESGUE-RADON-NIKODYM theorem 333
such that A n ]- p, P[ C
U ]a k, bk[ C [- P, P] and}; (b k- ak) < tJ (P).
k=1 k=1
By the choice of tJ(P), we have
00
define
Since t (F) is finite for every compact subset F of R, both integrals on the
right side of (3) are finite. Thus (2) shows that I. is a nonnegative linear
functional on ctoo(R). Let t. be the regular Borel measure on R constructed
from I. as in § 9. Theorem (12.36) shows that
f I dt = f I dt. + f I dtel = f I d(tc + t4)
R R R R
for all I Ectoo(R). The equality (i) follows from (19.54) and (12.41). For
all x ER, it is clear that tc({x}) = t({x}) - tel ({x}) = 0, and a moment's
thought shows that the decomposition (i) of t is unique. 0
(19.58) Remark. Let t be a regular Borel measure on R and let t. and
be as in (19.57). Let the correspondences t -'>- a, t. -'>- a., and td -'>- ad
tel
be as in (19.51). It is easy to see from (19.45) that
a = ac + ael'
The function a c is continuous (19.52). The function ael is called a saltus
lunction; it has derivative zero a. e. (19.56) and a jump [or saltus] equal
to a(x+) - a(x) = t{{x}) at each of its [countably many] discontinuities.
Next we turn to the question of singularity as regards measures and
their corresponding functions a. First we need a lemma.
(19.59) Lemma. Let ft and v be a-finite measures on a measurable space
(X, d). Then we have v .1 ft il and only il there exists no a-finite measure ji
on (X, d) such that ji =i= 0, ji ~ v, and ji ft. <
°
Proof. If v .1 ft, then v = + v is the unique Lebesgue decomposition
of v given in (19.42). Suppose that ji is a a-finite measure on (X, d) such
<
that ji ~ v and ji ft. Then there is a a-finite measure n on (X, d) such
that v = ji +n [define n (E) = lim (v (E nAn) -
11->'00
ji (E n An)), where
336 Chapter V. Differentiation
the desired v. 0
(19.60) Theorem. Let l be any regular Borel measure on R and let oc
be the normalized nondecreasing function associated with l as in (19.45).
Then l and A are mutually singular if and only if 0c' vanishes almost every-
where.
Proof. Suppose first that l and Aare not mutually singular. By (19.42),
we have l = II + l2' where II =1= 0, II ~ A, and l2 1.. A. Then II is a regular
measure, as (19.49) shows. 1
Next suppose that II -> OCl> i.e., II = A"". Then OC1 is absolutely continuous
on every interval [-P, PJ (19.53), and (19.45) and (18.16) imply that
Since II =1= 0 and II is regular, oc~ is positive on a Borel set E such that
A(E) > o. If l2 -'>- oc2 , then oc = OC1 + oc2, and so oc' = oci + oc~ ~ oc~ a. e. ;
hence oc' cannot vanish a. e. We have shown that 0c' = 0 a. e. implies i 1.. A.
To prove the converse, suppose that 0c' > 0 on a Borel set A of positive
Lebesgue measure. By (18.14), 0c' is Lebesgue measurable on R and of
course oc' ~ 0 a. e. Define a on !!J (R) by
a(E) = J oc' (t) dt.
E
It is obvious that
(1)
and that
a(A»O. (2)
1 Here is an alternate proof. If E is a bounded Borel set, choose a decreasing
sequence (U")~=l of bounded open sets such that Ee n Uti =
n=l
00
B, and A(B n E')
= O. Then 11(B n E') = 0, 1,(U,) ~ I(U,) < 00, and 1,(E) = 1,(B) = lim 1,(Un ) •
...... 00
Unbounded Borel sets of finite I,-measure are now easily dealt with. If U is
open in R, choose an increasing sequence (Fn)~=l of compact sets such that
n A') = O. n A') = 0
00
A = U Fn e U and A(U Then 1,(U and 11(U) = 1,(A)
n=l
= lim 1,(Pn). Thus 1, is regular .
...... 00
§ 19. Complex measures and the LEBESGUE-RADON-NIKODYM theorem 337
for all a < bin R. Like tl in the preceding paragraph, 0' is a regular Borel
measure on R. It follows from (3), as in the proof of (19.48), that
O'(E) ~ t(E) (4)
for all E EffI(R). Combining (1), (2), (4), and (19.59), we see that t..L A
cannot obtain. Thus t ..1. A implies <x' = 0 a. e. 0
We now present our main decomposition theorem for regular Borel
measures on R and their corresponding nondecreasing functions.
(19.61) Theorem. Let t be any regular Borel measure on R. Then t can
be expressed in just one way in the form
(i) t = ta + t8 + td'
where ta' t8, and td are regular Borel measures on R, ta ~ A, ta ..L A, td is
purely discontinuous, and ta is continuous. If <X, <xa, <X., and <Xd are the
corresponding nondecreasing functions [see (19.45)], then:
(ii) <X = <Xa + <X, + <Xd;
<Xa is absolutely continuous on every compact interval; <X, is continuous;
<X~ = 0 a. e.; and <Xd is a saltus function. Furthermore, we have
"
Proof. We proved in (19.57) that t = 'c + 'd' where 'c is regular and
continuous and 'd
is purely discontinuous. We also proved that this
decomposition is unique. In (19.58) we produced the decomposition
<X = <Xc + <Xd' where <Xc is continuous and <Xd is a saltus function. In (19.56)
we showed that <X~ = 0 a. e. ; hence <X' = <X~ a. e. Now define by formula 'a
'a
(iii). Then is the regular measure 0' that appears in the proof of (19.60).
It is clear that 'a ~ A. Let 'a~ <Xa. Then <Xa is absolutely continuous on
every compact interval (19.53) and 'a(E) = J <x~(t)dt for all E EffI(R).
E
Applying (19.45) and (18.14), we have
b
<Xa(b) - <Xa(a) = 'a([a, b[) = J <x~(t)dt ~ <Xc(b) - <Xc (a) (1)
II
whenever a < b in R. Let <X, be the function <Xc - <Xa. Then <X, is continuous
Hewitt/Stromberg, Real and abstract analysis 22
338 Chapter V. Differentiation
and, by (1), IX. is nondecreasing. Next let t8 -+ IX" i.e., t, = All.' Then t, is
continuous (19.52) and regular and
(2)
Formula (iv) follows at once from (iii) and the definition of lXa' From this
fact and (18.3) we get IX~ = IX' = IX~ a. e.; hence a; = 0 a. e., and so t,.l A
(19.60). Thus (2) gives the unique Lebesgue decomposition of tc with
respect to A. 0
As usual, we close the section with a set of exercises. None of them is
essential for subsequent work, but they all illustrate the theory in one
way or another, and we recommend their study to the reader.
(19.62) Exercise. Let" be a signed measure on (X, d). Prove that
,,+(E) = sup{,,(F) : FEd, FeE}
and
,,-(E) = -inf{"(F) : FEd, FeE}
for all E Ed.
(19.63) Exercise. Let X be a nonvoid set and let d be an algebra of
subsets of X. Suppose that" is a real-valued set function defined on d
such that "(0) = 0 and " (A U B) = ,,(A) + " (B) if A and B are disjoint
in d. Let ,,+
and ,,- be defined as in (19.62).
(a) Prove that ,,+
and ,,- are finitely additive measures on (X, d).
(b) Suppose that sup{I,,(A)1 : A Ed} < 00. Prove that" =,,+ - ,,-.
[This is the Jordan decomposition of a finitely additive signed measure.]
(19.64) Exercise. Let X = [- 1, 1[ and let d be the algebra of all
finite disjoint unions of intervals of the form [a, b[ C [-1, 1[. Let
f (x) = x1 for x =l= 0 and f (0) = O. Define" on d by
y
J g(y) de(y) = yJg(y) 11(y) d'll(Y) (1)
for all Borel measurable functions g on Y for which the left side of (1) is
defined. Also, if g is in ~1(Y' 9i(Y), e), then gft is in ~l(Y' 9i(Y), 'II).
We now appeal to (12.46.ii) to write
y
J g(y) 11(y) d'll(Y) = xJ go ffJ(x) 11 0 ffJ(X) dp,(x) (2)
for allg E~1(Y' 9i(Y), e). Since ffJ is continuous and ft is Borel measurable,
§ 20. Applications of the LEBESGUE-RADON-NIKODYM theorem 343
Inspection of (1), (2), and (3) shows that equality must hold in both (2)
and (3). Since w E~l([a, bJ) by (20.4) and
x
cp(x) = f w(t) dt + cp(a) ,
II
cp(O) = O.
[Draw a sketch of the graph of cp.]
(a) Prove the following assertions. The variation Vol cp is equal to 1.
The function cp is absolutely continuous; in fact, cp is in ~ipd[O, 1]). The
derivative cp' assumes only the values ± 1.
(b) Consider cp as a mapping of [0, 1] onto [0, !]
to which Theorem
(2004) may be applied. Find all of the ..A).-measurable functions w on
[0, 1] for which (2004.i) holds.
(20.8) Exercise. Let g be the function constructed in (18040.a) supra.
It is an N-function mapping [0, 1] onto [0, {3], where
Thus Theorem (20.4) can be applied, with g = cp. Since g' does not exist
on F, no function w for which (20.4.i) holds can possibly be the derivative
of g. Find all ..A).-measurable functions on [0, 1] for which (20.4.i) holds
for cp = g.
(20.9) Note. There are obvious questions concerning classical trans-
formations of integrals other than those treated in (20.4) and (20.5). For
example, for a continuous mapping cp of Rn into Rn, the function w of
(20.3) is in certain cases the absolute value of the Jacobian of the trans-
formation cp. Lack of space and time compels us to omit this interesting
subject, although its main outlines are clear enough from (20.3) and the
n-dimensional Lebesgue integral to be defined in § 21.
In § 15 we computed the conjugate space for each of the Banach spaces
£p(X, .91, p,) for 1 < P < 00 and arbitrary measure spaces (X, .91, p,), but
omitted all mention of this problem for spaces £1 (X, .91, p,). We now
address ourselves to this computation. The results and the techniques
employed here are quite different from those of § 15: our main tool is the
LEBESGUE-RADON-NIKODYM theorem.
(20.10) Remarks. Let (X, .91, p,) be any measure space. Let g be a
bounded d-measurable function on X and let Lg be defined on
£1 (X, .91, p,) by the rule
(i) Lg (f) = J f g dp, .
x
Obviously Lg is linear on £1 and
ILg (f) I ;;:; Ilgll .. Ilflll
for all f E £1' Hence Lg is a bounded linear functional on £1' If the function
g is tampered with in any way whatever on a set of measure zero to obtain
a new d-measurable function h, then it is clear that Lh = L g • Thus g
need not be bounded to give rise to an element of £1. It turns out in many
important cases that all elements of £1 indeed have the form L g , where g
is "essentially" bounded. We begin by making this idea precise.
(20.11) Definition. Let (X, .91, p,) be a measure space. A set A E .91 is
said to be locally p,-null if p, (A n E) = 0 for all E E .91 such that p, (E) < 00
[d. (9.29)]. An d-measurable function g on X is said to be essentially
bounded if for some real number a;;;: 0 the set {x EX: Ig(x)1 > a} is
locally p,-null. Let Ilglloo be the infimum of the set of all such numbers a.
The number Ilglloo is called the essential supremum of Igl. This number is
sometimes denoted by the symbols "ess sup Igl" or "vrai max Igl".
(20.12) Theorem. If (X, .91, p,) is a a-finite measure space and A Ed,
then A is locally ",,-mtll if and only if "" (A) = O.
Proof. Trivial.
(20.13) Theorem. Let (X, .91, p,) be any measure space. Suppose that
t E£1 (X, .91, p,) and that g is essentially bounded. Then:
(i) the set {x EX: If (x) I > O} is a-finite;
§ 20. Applications of the LEBESGUE-RADON-NIKODYM theorem 347
(ii) Ig E~1;
and
(iii) I Jig d",1 ~ Ilgll eo 11/111'
x
Proof. Let E = {x EX: If(x) I > O} and En = {x EX: I/(x)1 ;;;; ~} for
eo
U En and", (En) <
n = 1, 2, . . . . Obviously E = 11=1 for all n. Thus (i) 00
holds. Let a be any nonnegative real number such that the set
A = {x EX: Ig(x)J > a}
is locally ",-null. By (12.22), we have
J I/gld", = J Ilgld", + lim J I/gld", +EnA'
J Itgld",
x E' ....... eo E"nA
~ 0 + 0 + aEnA'
Jill d",
{x EX: Ig (x) I > "" {x EX: {J,. < Ig(x)i ;:;;; {J"+i}
{J,.} = i~1
is not locally ,u-null, and so for each n there exists a j such that
{x EX: {J,. < Ig (x) I ;:;;; {J,.+i} has a subset of finite positive ,u-measure.
Define the subsequence (1Xn) of ({J,.) by letting IXI = {Jl' and, having
defined 1Xn, letting 1Xn+1 be the smallest value of {J such that
{x EX: IXn < Ig(x)i ;:;;; 1Xn+1}
has a subset A,. of finite positive ,u-measure. This establishes (1) -(3).
Let
"" 1
I = ..~ IXnI'(A,,) ~.A".
1
J Id,u =
CD
Then we have }; - < 00, and so I E~. But we also have
x .. =1 IXn
jllgld,u = ..~
"" 1
IXnI'(A,,) 1lgld,u
1
J IXnd,u =
CD
~ }; (A) 00. 0
.. =1 IXnI' .. A"
§ 20. Applications of the LEBESGUE-RADON-NIKODYM theorem 349
such that 0< e < Ilglloo' Then the set {x EX: Ig(x)1 > Ilglloo - e} is not
locally f-t-null, and so it has a subset E Ed such that 0 < f-t (E) < 00.
Define
1
1= p,(E) ;E sgn(g) .
It is obvious that I E~l and that 11/111 ~ 1. Also, we have
_ 1
Lg (I) = jig df-t = fL (E) j Igl df-t
~ Ilglloo - e . (2)
It follows from (1), (2), and the definition of IILgl1 that IILgl1 = Ilglloo' 0
It is tempting to conjecture that every bounded linear functional on
an ~l space has the form Lg for some g E~oo' This, however, is not the
case for all measure spaces, as the following example shows.
(20.17) Example. Let 1= [0, 1] and let X be the unit square I x I.
Let d be the a-algebra of Borel sets in X [usual topology] and define v
and f-t on d by
v(E) = L: A({y EI: (x, y) EE})
xEI
and
f-t(E) = v(E) + L: A({x EI: (x,y) EE}).
rEI
[This construction is very like the construction in (I9.71.d).] Define Lon
~l (X,d, f-t) by
L(I) = J Idv.
x
Since v ~ f-t, it is plain that L E ~T (X, d, f-t) and that IILII ~ 1. Assume
that there exists a function g E ~oo (X, d, f-t) such that L = Lg. For each
fixed y E I, set H)' = {(x, y) : x E I} and 1= ;Hysgn(g). Then I is in
350 Chapter V. Differentiation
(ii) f I dp = I f Idp .
x ..=IF..
Proof. For n EN, write
~ n~ I ~F. = 1ft-a. e. and II - 1;1 I ~F.I ~ III E ~l for all PEN. Thus
(i) and (ii) follow from LEBESGUE'S theorem on dominated convergence. 0
(20.19) Theorem. Let (X,.9I, ft) be a decomposable measure space
(19.25). Suppose that L is a bounded linear lunctional on ~l (X, .91, ft).
Then there exists a lunction g E~<XI (X, .91, ft) such that L = L, as in
(20.16).
Proof. (I) Suppose that ft (X) < 00. For E E .91, define
(1)
We claim that'll is a complex measure on (X, .91) such that 'II ~ ft. Let
<XI
.. =1
for all E E.9I. We assert that Igl ~ IILII ft-a.e. To see this, let
A = {x EX: Ig(x)1 > IILII}
and assume that ft (A) > O. From (3) and (12.6) we have
1'111 (A) =
A
J Igl dft > AJ IILII dft = ilL I ft(A) .
Thus there exists a measurable dissection {Al' ... , An} of A such that
n
E I'll (Ai)! > IILllft(A).
;=1
352 Chapter V. Differentiation
J sg af'
L(s) = (4)
x
for all complex-valued, d-measurable, simple functions s defined on X.
Now let I E~l(X,d,ft), and use (11.35) to select a sequence (Sn):'=1 of
d-measurable simple functions such that ISll ~ IS21 ~ ••• ~ III and
sn(x) ~ I (x) for all x EX. Plainly ISngl ~ IILII '1/1 E ~l and II - snl
~ 21/1 E ~l for all n EN; hence (12.30), (4), and the continuity of Limply
that
This proves the theorem for the case that ft (X) < 00.
for all I E ~l(X, d, ftF)' Since ~dX, d, ft) C ~(X, d, ftF), (5) plainly
holds for all I E ~l (X, d, ft).
Next consider the functions gF' All of them are defined everywhere
on X, but the values of gF on F' are of no consequence in determining L F.
In any event, we can [and do] define a function g on all of X by
ITI (A) = sup{t; IT (A;) I : {AI> ... , An} is a measurable dissection of A}.
= [[g[[oo[['i[[ .
Thus LT E~!, and
[[LT[[ ~ [['ill . (1)
Let e > 0 be given and select a measurable dissection {AI> ... , An} of X
such that
n
L ['i(Aj)[ > [['ill - e.
i~l
n
For each j, let l;.j = sgn('i(Aj)) and set g = L l;.j~Ar It is plain that
i~l
Thus (20.27.i) holds for -r. Also, if A and B are disjoint sets in .91, then
~A + ~B = ~AUB' and so
-r(A U B) = L(~AUB) = L(~A + ~B) = L(~A) + L(~B) = -r(A) + -r(B);
hence (20.27.ii) holds. Next, if A Ed and A is locally p-null, then ~A =0
in ~ 00 and therefore
-r(A) = L(~A) = L(O) = O.
Thus 120.27.iii) holds, and so -r EF (X, .91, p). Let g be any function in ~oo.
Choose a bounded t E ~oo such that Ilg - 11100= 0 and 11/11 .. = Ilglloo' Choose
a sequence (sn) of d-measurable simple functions such that III - snllu -+ O.
It is clear from (1), the linearity of L, and (20.29.i) that
L(sn)=fsnd-r (2)
x
for each n EN. Since 111- snll 00 -+ 0 and L is continuous on ~oo' (2) and
(20.29.ii) imply that
L(g) = L(f) = lim L(sn) = lim f Sn d-r = f I d-r
n->-OO n->-OO X X
= f g d -r = Lr (g). 0
x
(20.35) Theorem. Let (X, .91, p) be an arbitrary measure space. Then
the mapping T defined by
T(-r) = Lr
[see (20.33)] is a norm-preserving linear mapping 01 F(X, .91, p) onto the
conjugate space ~!. Thus F is a Banach space and, as Banach spaces,
F and ~! are isomorphic.
Proof. The fact that T is a norm-preserving mapping from F into~!
is (20.33). It is trivial that T is linear. Theorem (20.34) shows that Tis
onto ~!. Being both linear and norm-preserving, T is one-to-one and
preserves Cauchy sequences. Since ~t, is complete, so is F. 0
(20.36) Remark. Let X be an arbitrary nonvoid set. As in (7.3), let
~ (X) denote the space of all bounded, complex-valued functions on X.
This space has several other names: if X is regarded as a topological space
with the discrete topology, then ~ (X) = <r (X) (7.8); if P is the counting
measure defined on gJ(X) (lOA.a), then ~(X) = ~oo(X, gJ(X), p); and
in (14.26) this space was denoted by 100 (X) [the conjugate space of 11 (X)].
In all cases the norm used on ~ (X) has been the uniform norm. Thus
Theorem (20.35) shows that the conjugate space ~ (X)* of the Banach
space ~ (X) is isometrically isomorphic to the space of all bounded,
complex-valued, finitely additive measures defined on gJ(X).
(20.37) Exercise. Let X be a nonvoid set. Suppose that -r is a finitely
additive measure defined on gJ (X) such that -r (X) = 1 and -r (A) = 0 or 1
for all A eX. Let 0/1 = {A : A eX, -r(A) = I}.
358 Chapter V. Differentiation
[First consider the case that lim I(n) = O. In general, find functions
11->-00
prove that inf{h (x) : x ER} ~ 0. Assuming that this is false, choose
°
e > such that h (x) ~ e for all x E R. Let p be an arbitrary positive
integer, and let (/J denote the set of all functions cp from {I, 2, ... , n} into
{I, 2, ... , P}. Clearly ($ = pIt. For each cp ((/J, let x (cp) = cp (1) tl + cp (2) t2
+ ... + cp(n)t". Show that
1: [Mx(cp) + tk ) - Mx(cp)] ~ 2P"-11IM"
'PE<l>
1I(A)= 1: T((An[n,n+l[)-n).
1$=-00
Show that 11 is finitely additive and that 11 (A) = A(A) for all A E~.
For A C R, define IA on R by the rule IA(t) = 1I(A + t). Let M be as in
part (a) and define ft on fJJ(R) by the rule
ft(A) = lim M(min{IA' 1~}).
~oo
It is easy to prove (iii). To prove that I-' is additive, use the inequalities
min {fA + fB' n} ~ min {fA, n} + min {fB' n} ~ min {fA + fB' 2n}. To prove
(iv), use the equality !cHt) (x) = fA (x + t).]
Our third application of the LEBESGUE-RADON-NIKODYM theorem is
to the study of yet another conjugate space. In (12.36), we saw that if X
is a locally compact Hausdorff space, then every nonnegative linear
functional I on <E:oo(X) has the form I (I) = J f dt for some regular Borel
x
measure ton X. This fact is useful in identifying the conjugate space of the
complex Banach space <E:o(X).
(20.41) Definition. Let X be a locally compact Hausdorff space.
A complex measure I-' defined on the a-algebra !Ji (X) of Borel sets of X
is said to be a complex regular Borel measure on X if for each E E!Ji (X)
and each e > 0 there exist a compact set F and an open set U such that
FeE c U and II-' (A)I < e for all A E!Ji(X) such that A c F'. Let un
M (X) denote the set of all complex regular Borel measures on X.
(20.42) Note. If I-' EM (X) and I-' ~ 0, then, since all complex measures
are bounded (19.13.v), I-' is a finite measure. Theorem (12.40) thus shows
that the definitions of regularity of I-' given in (20.41) and in (12.39) are
equivalent.
(20.43) Theorem. Let X be a locally compact Hausdorff space, let
1-', 'II EM (X), and let IX (K. Then, operations being defined setwise, we have
(i) IXI-' EM(X)
and
(ii) I-' + 'II EM(X).
Thus M(X), with the norm defined by 1/1-'1/ = 11-'1 (X), is a complex normed
linear space.
Proof. Exercise.
(20.44) Theorem. Let X be a locally compact Hausdorff space, let I-' be
4
a complex measure defined on !Ji (X), and let I lXi I-'i be the Jordan decom-
i=l
position of I-' as in (19.15.c). Then the following three statements are pairwise
equivalent:
(i) I-' EM(X);
(ii) II-' I EM(X);
(iii) I-'i EM(X) for j = 1,2,3,4.
Proof. We make heavy use of Theorem (19.13). The fact that (i) implies
(ii) follows from the inequality
II-'I(U n F') ~ 4· sup {II-' (A)I : A E!Ji(X), A c un F'}.
The inequality l-'i(U n F') ~ II-'I(u n F') (j = 1, 2, 3, 4) shows that
§ 20. Applications of the LEBESGUE-RADON-NIKODYM theorem 361
4
(ii) implies (iii). From the inequality 1.u(A)1 ~ E .ui(A), we infer that
;=1
(iii) implies (i). 0
(20.45) Theorem. Let X be a locally compact Hausdorff space, let
.u EM (X), and define L,. on (£:o (X) by the rule
(i) J fd.u.
L,.(I) =
x
Then L,. is a bounded linear functional on the Banach space {£:o (X) and
(ii) IIL,.II = 11.u11 .
Proof. Each f E (£:o is a bounded continuous function and so, since l.u I
is a finite measure (19.13.v), f is in ~l(X, &6'(X), l.ul). Thus L,. is a linear
functional on {£:o' Also (19.38.iv) shows that if f E {£:o, then
for all f E ~l(X, &6'(X), l.ul). Let e > 0 be arbitrary. By (13.21), there
exists a function g E (£:oo such that Ilgllu ~ Illoliu = 1 and
x
J 110 - g I dl.ul < e . (3)
1(/) = I W) - I (n .
Notice that if gl and g2 are any two functions in <rt such that I = gl - g2'
then 1+ + g2 = gl + r.
and so (4) implies that 1(/) = I (gl) - I (g2)'
A simple computation shows that I is real linear on <r~. Next. for I E <roo
define
1(/) = I (Rei) + il (lm/) .
Another obvious computation shows that I is complex linear on <roo
It is now clear [using (i)J that I is a nonnegative linear functional
on <roo Thus. just as in (9.4). for all I E <ro we have
II (I) I ~ I (lfI) .
This fact together with (1) yields
11111 = sup{ll (I) I : I E <roo II/I! .. ~ I} ~ sup{I(I/D : I E <roo 11/11 .. ~ I} ~ IILII·
§ 20. Applications of the LEBESGUE-RADON-NIKODYM theorem 363
To prove the reversed inequality, let e> 0 be given and then select
g E<io such that Ilgllu ~ 1 and IL (g) I > IILII - e. From (i) we have
IILII - e < IL(g)1 ~ I(lgi) ~ 11 1 11.
Hence 11111 = IILII. 0
(20.47) Theorem. Let X be a locally compact H ausdortJ space and let
L E<io (X) *. Then there exists a complex measure p, EM (X) such that
(i) L (f) = J I dp, = Lp(f)
x
lor all I E<io (X) .
Proof. Let I be constructed from L as in (20.46). By F. RIEsz's
representation theorem (12.36), there exists a measure £ on X as in § 9
such that
I(f)=Jld£
X
for all I E<ioo (X). It follows that
£(X) = 1(1) = sup {I (f) : I E<ito, I ~ I}
~ sup{II(f)I: I E<io' 1I/IIu ~ I}
= 11111 = IILII·
Thus t is a finite measure. We know (13.21) that <ioo is a dense linear
subspace of ~l (X, J(, t). It follows from (20.46.i) that, relative to the ~l
norm, L is a bounded linear functional of norm ~ 1 on this dense sub-
space: in fact,
IL(f)1 ~ 1(1/1) = J I/ldt = 11/111
x
for all IE <ioo . Thus, by (14.40) [or the Hahn-Banach theorem], L can be
extended to a bounded linear functional L' on ~l (X, J(, t) such that
IIL'II ~ 1. Now apply (20.19) to obtain a function g E~oo (X, J(, t) such
that Ilglloo= IIL'II ~ 1 and
L'(f)=JIgdt
x
for all I E~l (X, J(, t). Next define p, on J( by the rule
p, (E) = J g d t.
E
in the ~1 (X, vII., t)-norm on ~o and agree on the ~l-dense subspace ~oo.
All that remains is to show that p, is regular. Let E Evii. and E> 0 be
given. Since t is regular and finite, there exist a compact set F and an open
set U such that FeE c U and t(U n F') < E. Thus if A Evii. and
A C U n F', then
1p,(A)I = If gdtl ~ teA) ~ t(U n F/) < E.
A
(d) If '/I EM(X) and '/I = '/11 + '/12 is the Lebesgue decomposition of '/I
with respect to t, then '/Ii E M(X) (j = 1,2).
(20.52) Exercise. More on the structure of 'ioo(X) and'io(X). (a) Let
X be a non void locally compact Hausdorff space, and let M be a multi-
plicative linear lunctional on <ro(X). By this we mean that M satisfies
(9.l.i) and (9.l.ii), that M =1= 0, and that M (I g) = M (I) M (g) for all
I, g E<ro(X). Prove that there is a point a EX such that M (I) = Ea(l) = I (a)
for all IE <ro(X). [Hints. It is convenient to prove first that M is con-
tinuous. In fact, we have 1M (1)1 ~ 1I/IIu for all I E <ro. To see this, assume that
00
M(I) = IX IIXI> 11/11u. Let g = - L IX-kfk. The series converges
where
k=l
uniformly and so g E<ro. Check that IX-I I + g - IX-I I g = O. Then one has
0= M(O) = IX-1M(f) + M(g) - IX-IM(I)M(g) = 1 + M(g) - M(g) = 1.
Now knowing that M is bounded, we apply (20.47) to write
M(f) = J Idl-' = J Igdt
x x
where I-'EM(X), tEM+(X), gE~oo(X,&#(X),t), and Ilglloo~ 1. If
A, B E &# (X) and t (A) and t (B) are positive, then A n B =1= 0. This
follows readily from the regularity of t and the identity M (lIi2)
=M(fl)M(f2)' Now it is easy to see that t assumes only one positive
value, since teA) = t(A n B') + t{A n B) and t(B) = teA' n B) +
teA n B). The family {F eX: F is compact, t(F) > O} is nonvoid and
has the finite intersection property and hence has non void total inter-
section, say Fa. It is not hard to see that t (Fa) > 0 and that Fo has no
proper nonvoid subsets, i.e., Fa = {a} for some a EX. It follows immediately
that M(f) = I(a).]
(b) Prove that every multiplicative linear functional M on <roo (X)
has the form Ea for some a EX. [Hints. Choose 10 E <roo such that
M (to) = 1. Let U = {x EX: lo(x) =1= O}. Let 5' = {I E<roo (X) : I(U') c to}}.
Then 5' can be identified in an obvious way with <r0(U) [note that U is
locally compact]. The functional M is multiplicative and nonzero on 5'
and so part (a) shows that M(f) = I (a) for all I E5', a being a point in U.
Given g E <roo (X), the function glo is in 5' and M (glo) = M (g). Since
lo(a) = 1, the proof is complete.]
(c) Generalize part (a) as follows. Let (X, .5#, 1-') be a finite measure
space such that I-' (X) > O. Prove that Jig d I-' = J I d I-' J g d I-' for all
I E~l (X, .5#, 1-') if and only if I-' assumes only the values 0 and 1. [Compare
this with (20 37.c).]
(d) Part (a) can be interpreted in the following way. A lelt ideal in an
algebra A is a linear subspace I that is closed under left multiplication
by arbitrary elements of A: x E I and yEA imply yx EI. Right ideals
are defined similarly. A set I that is both a left and a right ideal is called
a two-sided ideal. In commutative algebras, the distinction between right
366 Chapter V. Differentiation
and left ideals disappears, of course, and we use the term "ideal". An
ideal is called maximal if the only ideal containing it properly is the entire
algebra. In the commutative Banach algebra <ro(X), every set
{f E <ro (X) : I (a) = O} is a closed maximal ideal. Furthermore, every
maximal ideal [not assumed a priori to be closed] in <ro(X) has this form.
[All assertions are simple to verify except the last, for which the following
hints are offered. If S is an ideal in <ro (X) and if there is a point a E X such
that I (a) = 0 for all f E S, then S can be maximal only if S = {I E<ro (X):
I (b) = O} for some b EX. If S is an ideal such that for all a E X, I (a) =f= 0
for some IE S, then a simple compactness argument, and the fact that
11 E S if IE S, show that S ~ <roo (X). Consider the algebra = <ro(X)jS. e
It is a linear space over K and also an algebra with no proper ideals at all.
This implies that e is a field or is K as a linear space with all products
equal to o. Let -r denote the canonical mapping of <ro (X) onto e:
-r (I) = I + S for all I E <ro (X). If e is a field, let h be a function in <ro (X)
such that -r(h) is the multiplicative unit of e. Since S contains <roo (X) ,
00
there is an element q; E S such that Ilh - q;11 .. < 1. Let 1p = - 1: (h- q;)k.
k=1
Then as in part (a), one can show that
h = q; - 1p q; + 1ph - 1p .
Clearly q; - 1pq; is in S, and 1ph - 1p is in S because
-r(1ph - 1p) = -r(1p) -r(h) - -r(h) = -r(h) - -r(h) = 0.
Hence h is in S, and so -r(h) = 0, a contradiction. If e is not a field, then
e
all products in are 0, and in particular, if IE <rt, then 1= f2 . l2- is in S.
1 1
l
f-t [apply (19.24) to 'fJ! and 'fJ; to obtain It and I~]. Now define I by
fo (x) for x EB',
f(x) = ~ 00 for x EB n P,
for x EB n P'.
Since I = fo f-t-a.e., it is clear that f is a derivative.
The following lemma characterizing derivatives will be very useful.
It has the advantage that it makes no mention of Lebesgue or Hahn
decompositions.
(20.54) Lemma. Let g be an extended real-valued, d-measurable func-
tion on X. For each real number cx, let Gar. = {x EX: g (x) ~ cx} and
Lar. = {x EX: g (x) ~ cx}. The function g is a derivative of fJ with respect to f-t
il and only if the following conditions obtain:
(i) for every real number cx and every A Ed, we have
fJ(Gar. n A) ~ cxf-t(Gar. n A);
(ii) for every real number cx and every A Ed, we have
fJ(Lar. n A) ~ cxf-t(L« n A) .
Proof. Suppose that g is a derivative of fJ with respect to f-t and
let Band P be as in (20.53) [with I replaced by g of course]. Then for
all A Ed and cx ER we have
'fJ(Gar. n A) = 'fJa(Gar. n A) + 'fJ,(Gar. n A)
= J g df-t + 'fJ,(Gar. nAn B)
G",nA
or
(b) there exist real numbers IX and (3 and a set FEd such that
0< fl(F) < 00, I'I].I(F) = 0, and g(x) ~ IX < (3 ~ t(x) for all x EF.
Assume that (a) holds. By (i) we have
'I] (F) = 'I](G{J nF) ~ (3fl(G{J nF) = (3fl(F)· (2)
We also have
'I] (F) = 'l]a(F) = J t dfl ~ IXfl(F) . (3)
F
An obvious contradiction ensues from (2) and (3), and so (a) must fail.
Likewise (b) fails. Thus g = t fl-a.e., and so (20.53.i) holds for g.
Now make the following definitions: L={xEX:g(x)=-oo};
G = {x EX:g(x) = oo};Bo= B n (L U G); and Po = B n G. We wish to
show that (20.53.ii)-(20.53.iv) hold for B o, Po, and g. Since Bo C B,
it is obvious that fl (Bo) = O. It is also obvious that (20.53.iv) holds. Thus
it suffices to show that 1'1].1 (B~) = 0 and that (Po' Po') is a Hahn de-
composition for '1] •.
Since fl (B) = 0 and 'l]a < fl' we have l'I]al (B) = O. Thus condition (ii)
implies that for all A Ed and IX ER, we have
'I].(L cx nAn B) = 'I](Lcx nAn B) ~ IXfl(L cx nAn B) = O. (4)
It is obvious that
00
L 1 C L Z C··· and U Ln
n=! = G" ,
hence (19.3.ii) and (4) imply that
'fJ.(G' nAn B) = lim 'I].(L n
n-->oo
nAn B) ~ 0
for all A Ed. We conclude that G' nB is a nonpositive set for '1] •.
00
Likewise we use (i) and the facts G_1 C G_ z C ... and U G_n = L'
n=!
to prove that L' n B is a nonnegative set for '1] •. Thus the set L' n G' n B
is both nonpositive and nonnegative for '1]., and so I'I].I(L' n G' n B) =0.
Now B~ = B' U [(L' n G') n BJ and l'I]sl(B') = 0; hence I'I].I(B~) = O.
It remains only to prove that (Po, Po') is a Hahn decomposition for '1] •.
00 00
It is plain that G = n Gn
n=!
and that L =
n=!
n L_n . Thus we may argue as
above to prove that
'I].(G nAn B) ~ 0 and 'fJ.(L nAn B) ~ 0
for all A Ed [this time we use (19.3.iii) and the fact that '1]. is a-finite];
we leave the details to the reader. Therefore G n B is a nonnegative set
for 'fJ. and L n B is a nonpositive set for '1]8' Since Po is equal to B n G,
Po is a nonnegative set for 'fJ •. Since
P; = (B n L' n G') U(L n B) UB'
§ 20. Applications of the LEBESGUE-RADON-NIKODYM theorem 369
U J(,. C vIt. We suppose that f-l and 11']1 are a-finite on each of the a-alge-
k~l
bra J(,..
For each n EN, consider the measure spaces (X, J(,., f-l) and (X, J(,.,1']);
that is, we restrict the domains of f-l and 1'] to vltn. We write f-ln for f-l
on J(,. and 1']n for 1'] on vltn- Let fn be a derivative of 1']n with respect to f-ln'
We are concerned with pointwise limits of the sequence of functions
(fn) under two hypotheses concerning (J(,.). Our first theorem is the
following.
(20.56) Theorem. Suppose that
(i) vitI C vlt2 C vita C .. "
00
and write vltw for the smallest a-algebra containing U vltn· Let f-lw and 1']w
n~l
be f-l and 1'], respectively, restricted to vltw' Then both of the functions
(ii) I = lim In and f = lim fn
- n---+oo ~ 00
are derivatives of 1']w with respect to f-lw' Thus: lim fn (x) exists for f-l-almost
,.....,.00
n A) = n A) ,
00
the relations (3) holding for all n ~ no. If i'fJi(A) < 00 and peA) < 00,
Since i'fJi and p are a-finite on 1}> (4) holds for all A E U.A(. .
.. =1
Let {Fk}k'=1 be a pairwise disjoint family of sets in 11 such that
00
X = U Fk and p and i'fJi are finite on each Fk . Let 11k be the set function on
k=1
00
U .A(. defined by
.. =1
lIk(A) = OCp(Fk n L .. n A) - 'fJ(Fk n L .. n A) .
A routine computation and (4) show that 11k is a countably additive,
00
nonnegative, finite-valued measure on the algebra U.A(., in the sense of
.. =1
00 00
(10.3). Let 11 be the set function}; "k' also defined only on U.A(.. It is
k=1 ..=1
easy to see that 11 is a nonnegative, countably additive, a-finite measure
00
on U.A(.. By (1O.39.c), 11 admits a unique [count ably additive!] exten-
#=1
sion over the a-algebra 1w. All this implies that (4) holds not only for
00
A E U.A(. but for all A E1w. Thus condition (1) is established.
n=1
The proof of (2) is very like the proof of (1), and we leave it to the
reader. 0
The following special result is frequently useful.
§ 20. Applications of the LEBESGUE-RADON-NIKODYM theorem 371
(20.57) Corollary. Suppose that rJn ~ fln lor all n EN and that the
equality
(i) lim J In dfl = J I., dfl
1\-700 E E
00
holds for all E in the algebra U Jtk. Then rJ., is absolutely continuous with
k=1
respect to fl.,.
Proof. For each fixed kEN and every E E Jtk , we have
rJ (E) = rJk (E) = (1]k)a (E) = J Ik dflk = J Ik dfl . (1)
E E
For n > k, E is also in Jt.., and so (1) can be extended to
1] (E) = rJn (E) = (rJn)a (E) = J In dfl . (2)
E
=1
Jt... Let flo and rJo be fl and rJ, respectively,
restricted to the a-algebra Jto. Suppose that flo and IrJol are a-finite. Let f
and 1 be defined as in (20.56.ii). Then both t and 1 are derivatives 01 rJ~
with respect to flo.
24*
372 Chapter V. Differentiation
t 1
Proof. We note first that and are ~-measurable. To see this, write
hn = k->-oo
lim [min {!n' !n+v ... , !n+k}].
It is plain that hn is ~-measurable, and also that lim hn =
n->-oo
t [d. (6.83)].
-
By (i), all of the functions hm' hm+1' ... are Jim-measurable, and since
t=
-
lim hn +m , we infer that t is Jim-measurable for all nt, i.e., t is Jlo-
~oo - -
Then In,P is in JIp, {]n,p}p=1 is a pairwise disjoint family, and p'd,dn,p = In.
§ 20. Applications of the LEBESGUE-RADON-NIKODYM theorem 373
derivative of 'YJn with respect to I'n' Now apply (20.56) and the definition
of derivative in (20.53).J
(b) Let cp be a Lebesgue measurable function on R such that
a
f Icpl dA < 00 for all a> O. For every x ER, let ](n, x) be the interval
-a
[k . 2- n , (k + 1) . 2-n [ that contains x (n EN, k EZ). Prove that
lim 2n J cp d'A = cp(x) for A-almost all x ER. [Hint. Apply part (a)
........ 00 ](n,x)
with X = R, d = ~, ~ = {[k2- n, (k + I)2- n [}k'=_00, 'YJ (A) = J cp dJ....J
A
Compare this result with (18.3).
(c) Let 'YJ be a a-finite signed measure or a complex measure on
PA (R) such that I'YJI and 'A are mutually singular. Prove that
lim 2n 'YJ(](n, x)) = 0 for 'A-almost all x ER. For 1] a signed measure,
........00
prove that lim 2n 'YJ(](n, x))
........ 00
= ±oo on a set E such that I'YJI (E') = O.
[Apply part (b), (20.56), and (20.53.iv).J Compare this with the behavior
of (X's described in (19.60).
(20.62) Exercise: Densities. Notation is as in (20.61). Let E be any
set in d. Let 'YJ be the measure on d such that
I'(A) = I'(E n A) forall A Ed.
Plainly 'YJn ~ I'n; let Wn be a LEBESGUE-RADoN-NIKODYM derivative of
'YJn with respect to I'n [here n E {I, 2, ... , w}]. The function 'Ze'n is called
a density 0/ E with respect to JI...
(a) Prove that lim Wn (x) = Wro (x) for I'-almost all x EX. [This is
....... 00
a direct application of (20.56).J
(b) Suppose that each Wn is a constant real function I'-almost every-
where, in addition to being JI..-measurable. Suppose also that there is a
set D Evitro such that I' (D 6 E) = O. Then I' (E) = 0 or I' (E') = O.
[Plainly lim Wn is a constant I'-almost everywhere. A simple argument
........ 00
shows that Wro is equal to ~D I'-almost everywhere. Thus ~D is a constant
I'-a.e. J
(c) Consider the space X = [0, I [, and for each n EN, let
~ = {[k 2- n , (k + I)2-n[: k E{O, 1, ... , 2 n - In.
Let JI.. be the [obviously finite J algebra of sets generated by ~. Let
vIt = ~([O, I[). Let P be the Cantor-like set in [0, I[ obtained by re-
moving middle one-quarters at each step (6.62). Determine 'A(P). Let I'
be A and let 'YJ(A) = 'A(A n Pl. Compute Wn as defined above and find
lim W n •
>1---+00
°
E EnB,'
Since ,uo assumes only the values and 1, it follows from (12.60) that
there is a number IX for which ,uo({x EX: g(x) = IX}) = 1 and so lim In = IX
...-..00
,uo-almost everywhere on B~ and so ,uo-almost everywhere on X.]
(b) Prove that the IX of part (a) is equal to f 11 (x) d,ul (x) = 'YJla (X),
x
where 'YJla is the ,ul-absolutely continuous part of 'YJl'
(20.64) Exercise. Let X = [0, 1[. For each n EN, let ..II.. = {A eX:
2"-\
there is an .A).-measurable set Be [0,2- n [ such that A = U (B + k2-n)}.
k=O
Note that ..II.. is a a-algebra and that .AI ~ .A2 ~ ... ~ ..II.. ~ ....
Prove that if A E.Ao = ""
n..ll..,
,,=\
then A(A) = °
or A(A) = 1. [Hints. Apply
(20.62.b) to the set A, and use the increasing algebras of sets described in
(20.62.c). It is evident that each Wn is a constant, and so A(A) = or °
A([O, 1[ n A') = O.J
(20.65) Exercise [JESSENJ. Let I be any function in ~d[O, 1[,.A)., A)
and extend lover R by the definition I (x + k) = t (x) for all k (Z and
x E [0, 1[. Prove that
for A-almost all x E R. [Hints. Let ..II.. and .Ao be as in (20.64). Let In
be the function on the left side of (i). On the domain [0, 1 [, In is plainly
.An-measurable, and is the LEBESGUE-RADON-NIKODYM derivative of the
measure 'YJn' where 'YJ(A) = f IdA for A E.Al . By (20.59), In converges
A
A-almost everywhere to the LEBESGUE-RADON-NIKODYM derivative 10
\ \
of 'YJo with respect to A.a. Since 'YJo([O, 1[) = flo dAo = f t dA, and since
o 0
1
(20.64) and (20.63) hold, we have lim In (x) = f I dA A-a.e. on [0, 1[;
""-"00 0
and by periodicity, A-a.e. on R.]
(20.66) Exercise: A martingale theorem. Let X,.A,..II..,.AU) and ,u
be as in (20.56) and suppose that ,u(X) < 00. Suppose that (fn)~\ is a
sequence of functions on X each with values in [0, oo[ such that:
(i) In is .An-measurable for all n EN;
376 Chapter V. Differentiation
(ii) J In df-t = J In+1 df-t for all A E.,A;. and all n EN;
A A
(iii) J In df-t = 1 for all n EN.
x
00
U .,A;.
(a) Prove that there is a finitely additive measure 'YJ on ,,=1
such that
(iv) 'YJ (A) = J In df-t for all A E.,A;. and all n EN.
A
(b) Prove by giving an example that n need not be countably ad-
00
U .,A;..
ditive on ,,=1
(c) Prove that lim In exists and is finite f-t-almost everywhere on X.
"-.00 00
[Hints. Consider the set Q of all finitely additive measures w on "~1 .,A;.
that assume the values 0 and 1 and no other values and vanish
for f-t-null sets. Make Q into a topological space by neighborhoods
00
LI.A = {w EQ: w(A) = I}, for A E,,=1 U vIIn . Then Q is a compact Haus-
dorff space, and f-t and 'YJ can be transferred to Q. Apply (20.56) appro-
priately and go back to X.]
CHAPTER SIX
y~ 1 (flYl) .
To indicate that we "integrate" with respect to x, we denote this function
by 1%(1). Similarly we obtain a function J:v(t) on X. Next choose open
sets U C X and V C Y such that U- and V- are compact and U x V
:::> {(x, y) EX x Y : t (x, y) =1= O}-. Use (9.5) to find positive constants ac
378 Chapter VI. Integration on Product Spaces
measure spaces include most of those that arise in classical analysis. The
subject exhibits several technicalities, which can be an annoyance or a
source of fascination: depending upon one's point of view. We proceed to
some exact definitions.
(21.2) Definitions. Let (X, vii) and (Y,.AI') be any two measurable
spaces. For A Evii and BEvY, the set A x B c X x Y is called a
measurable rectangle. The smallest a-algebra of subsets of X x Y contain-
ing all of the measurable rectangles is denoted by vii x .AI' 1 and is
called the product a-algebra. For E C X x Y and x EX, let
Ex = {y EY : (x, y) EE} ;
similarly, for y EY let
£3' = {x EX: (x, y) EE} .
These sets are called X- and Y-sections 01 E, respectively. For a function I
on X x Y and a fixed x EX, let /rx] be the function defined on Y by
/rx](Y) = I (x, y) ;
similarly, for each y EY define Iry] on X by
I[Y] (x) = I (x, y) .
These functions are called X- and Y-sections all, respectively.
(21.3) Theorem. Let (X, vii) and (Y,.AI') be measurable spaces. Then
the lamily .91 01 all finite, pairwise disjoint unions 01 measurable rectangles
is an algebra 01 subsets 01 X x Y.
Proof. First we note that if Ax B and C x D are two measurable
rectangles, then
(AxB) n (CxD) = (A n C) x (Bn D),
m
as the reader can easily verify. Thus if E =.U (Ai x B i ) and F =
<=1
Thus, in particular, if the algebra .91 satisfies (i) and (ii), then .91 is a
a-algebra!.
1 Families that satisfy (i) and (ii) are called monotone families.
§ 21. The product of two measure spaces 381
Proof. Note first that the family ~ exists, for [!P(T) is a monotone
family and the intersection of all monotone families containing d is
again a monotone family containing d; this intersection is ~.
Since any a-algebra is a monotone family and g (d) :J d, it follows
that g (d) :J~:Jd. To finish the proof it therefore suffices to prove
that ~ is a a-algebra. For a monotone sequence (Hn):=I, we write limHn
00 00
U F,.
n=1
= U Gn
n=1
= lim Gn E ~. 0
Proof. Let F be the family of all sets E E.Lx.Al" such that (i), (ii),
and (iii) hold for E. We will show that F is a monotone family that
contains all finite disjoint unions of measurable rectangles. Then (21.7)
will show that F = .Lx.AI'.
Suppose that E = Ax B is a measurable rectangle. Since E~= B
or 0 according as x EA or x EA', and E'Y = A or 0 according as y E B or
y EB', we have '/I(E~) = v(B) ~.A (x) and ",(E'Y) = '" (A) ~B(Y); hence (i)
and (ii) hold for this E. We may also write
J'/I(E~) d",(x) = J'/I(B) ~.Ad", = '/I (B) • ",(A)
x x
=
y
J",(A) ~B d'/l = yJ ",(EY) d'/l(y) .
Thus (iii) holds for this E, and so E EF. Thus F contains all measurable
rectangles.
Let {E v "" E1>} be a finite, pairwise disjoint subfamily of F. Since
(
..~l
1>
..) x = 1> (1)))' 1>
"~I(EfI)~ and "~I Efl = ,,~,(EfI)'Y for all x EX andy EY, it is
clear that (..~lfl) E F. Thus, F contains all finite disjoint unions of
measurable rectangles.
00
Now let (EfI):=, be an increasing sequence in F and let E = U Efl
.. =1
= lim Efi' For all x EX, we have
as (10.13) shows, and so (11.14) implies that (i) holds for E. In like
manner, (ii) holds for E. Applying B. LEVI'S theorem (12.22), we have
J '/I(E~) d",(x) = lim J '/I «(EfI)~) d",(x)
x "-+cox
= J ",(E'Y) d'/l(y) .
y
AxB for which peA) < and '/I (B) < 00. Write F = ..C1IFfI' For each
00
x EX, we have (~)~ c B, and so '/I «(~)~) < 00. It follows from (10.15)
§ 21. The product of two measure spaces 383
that
for every x EX. Thus the function x -+ ,,(F%) is the pointwise limit of a
sequence of .A-measurable functions, and so it is .A-measurable (11.14);
hence (i) holds for F. Likewise (ii) holds for F. Since
J,,((~)%)dft(x);:;;; J,,(BHAdft< 00
x x
and
J ft ((~),,) d,,(y) ;:;;; J ft(A) ~Bd" < 00,
y y
= Jft(P) d,,(y) ,
y
00
= ~ f.txv(E ..) .
.. =1
Thus f.t x v is countably additive. Plainly f.t x v ;;S 0 and f.t x '11(0) = 0;
hence p x v is a measure on vIIx %. To show that p x v is a-finite, let
00
for every x EX. Thus (iii), (iv), and (v) for ~E follow at once from (21.8)
and the definition (21.9) of p, x 'II. For a simple -A'x ';v-measurable
function I, assertions (iii)-(v) are clear from the linearity of all of our
integrals.
Finally, let I be any nonnegative, extended real-valued -A'x';v-
measurable function on X x Y. Let (0'.,):=1 be an increasing sequence of
nonnegative, real-valued, -A'x .;V-measurable simple functions on
XxY such that O'.,(x,y)-+I(x,y) for all (x,y) EXxY (11.35). For
every y EY, (12.22) shows that
= lim J Jan(x,y)dft(x)dv(y)
~OOY x
= J [lim J an (x, y) dft (x)] dv(y)
Y ~oox
= J J f(x,y)dft(x)dv(y).
YX
1 If one of these integrals is finite, then by (21.12.v), all are finite [and equal].
It is to be understood that this function is defined only for those y E Y such
2
that x -+ t (x, y) is in ~l (X, JI, p). A similar remark applies to assertion (iv).
§ 21. The product of two measure spaces 387
x -+ J ~H(X, y) dv(y)
y
and write E = n~\ En. If P x v (E... ) = 00 for some no, then (i) is trivial for
E and, by (ii) , En. has compact subsets of arbitrarily large [finite!]
measure; hence E is in Bt. Thus suppose that p x v (En) < 00 for all
n EN. For arbitrary e > 0 and for each n EN, choose a compact set
Fn E..A,. x JI" and an open set Un E..A,. x JI" such that Fn C En C Un
e
and p X v(Un n Fn} < ~. Let U = n~\ Un and F
I 00 00
= n~\ Fn. Then we
have: FeE C U; U, FE ..A,. x.A,.; and also
00
~ 1; p X
n=1
V (Un n F~) < l' ;..
n=\
= e.
It follows that
p x v(U) = P xv(E) +Px v(U n E') ~ p x veE) +e (1)
and
p X v (E) = p x v (F) + P x v (E n F') ~ P X v (F) + e . (2)
Since U is open, (1) implies that (i) holds for E. Applying (2) and (10.13),
we obtain
p x v (E) ~ lim p x v (1\ U·· . UFp) + e,
p.-..oo
and since 1\ U ... U Fp is compact for all p, (ii) holds for E. Hence Bt
is closed under the formation of countable unions.
Next, let A x B be a measurable rectangle. Use our a-finiteness
hypothesis to select ascending sequences (An)~\ c..A,. and (Bn):'=\ C
00 00
.A,. such that X = n=\ U B n , and p(An) <
U An' Y = n=\ 00 and v(Bn) < 00
A x B = n=\
U [(A n An) X (B n Bn)] ,
and so if we can show that Bt contains each measurable rectangle having
sides of finite measure, it will follow from the preceding paragraph that
A x BE Bt. Thus we suppose that peA) < 00 and v(B) < 00. Use the
390 Chapter VI. Integration on Product Spaces
and
{' x v(A x B) = {'(A) . v(B)
= lim [{'(Cn) . v (Dn)J = lim {' x v(Cn x Dn) .
n---+oo n-+oo
and that
{' x v(W n F') - {' X v(] nU') = {' X v((Wn F') n (] nu')')
~ {' X v(W n F' n]') + {' x v(W n F' n U)
< B +B= 2e.
This proves that (An X Bn) n E' is in ~n' We conclude that ~n is a
a-algebra of subsets of An X Bn.
Now let tff = {E EJI", x Jt" : E n (An X Bn) E~n for every n EN}.
It is easy to see from our previous results that tff is a a-algebra of subsets
of X x Y that contains every measurable rectangle. Thus tff = ~ x Jt".
Then for any E E~ we have E' Etff; hence E' n (An X Bn) is in ~ for
§ 21. The product of two measure spaces 391
<XI
(21.21) Exercise. (a) Let (X, Jt, p,) and (Y,.AI, 11) be a-finite
measure spaces. Suppose that there exists a set A C X such that A ~ Jt
and suppose that there exists a nonvoid set BE.AI such that lI(B) = o.
Prove that (X x Y, Jt x.Al, p, X 11) is incomplete.
(b) Prove that (R2, ~ x~, Ax A) is an incomplete measure
space.
(21.22) Exercise. For each of the functions I on R x R defined by
1 1 1 1
the following formulas, compute f f I (x, y) dxdy, f f I (x, y) dydx,
o 0 0 0
1 1 1 1
ff I/(x,y)1 dxdy, and ff It(x,y)1 dydx:
o 0 0 0
x2_y2
(a) I(x, y) = (x2 + y2)2 ;
= 0 + f 0 dp,t (x) = 0.
A'
This shows that p,t x vt ~ I' x v. To prove (i), write d:: as 10 and ~vvt
as go [for brevity's sake]. Applying (21.13) to an arbitrary I belonging to
.~\(X x Y,..A x.Al, 1'+ x vt ), and taking note of (19.24), we obtain
f Idp,+ x vt = ff I(x,y)dvt(y)dp,t(x)
XXY x Y
Now combine (1) and (2) and let I= ~A' where A Evi( x % and
p.t x vt(A) < 00. This gives
p.txvt(A) = J
XXy
~A(x,y)/o(x)go(y)dp. x v(x,y). (3)
Since p.t x vt is a"-finite, (3) holds for all A E vi( x%, and the unique-
ness provision in (19.24) proves (i).
N ow suppose that p.t 1- 1'. Let B be a set in vi( such that I' (B) = 0
and p.t (B') = O. Then, as noted in (21.11), we have
p.xv(BxY) =p.(B)v(Y) =0
and
p.t x vt«(B x Y)') = p.tx 'lit (B'xY) = p.t(B') vt(y) = O.
This implies that p.tx 'lit 1- P. x v. For arbitrary 1', p.t, v, and vt , we have
p.txvt = (p.! + p.!) x ('II! + 'II!)
=Wx~+Wx~+Wx~+Wx~,
from which (ii) and (iii) follow easily. 0
FUBINI'S theorem and LEBESGUE'S dominated convergence theorem
are cornerstones of analysis. The theory of Fourier transforms, as well as
many other theories, depends in the last analysis on these two theorems.
We devote the remainder of the present section to a number of applica-
tions of these theorems. Our first result is a simple lemma that is useful
in establishing the p. xv-measurability of functions on spaces X x Y.
(21.30) Lemma. Let cp be a real-valued Borel measurable lunction
defined on R2 such that il MeR and A(M) = 0, then cp-I(M) E ~ x ~
and A x A (cp-I(M)) = O. Then 10 cp is ~ x ~-measurable lor every
complex-valued Lebesgue measurable lunction I defined A-a. e. on R.
Proof. First suppose that I = ~A' where A (~. By (10.34), we have
A = BUM where B EfJI (R) and A(M) = O. Then we may write ~A °cp
= ~Al' where
Since
and
cp-l (M) E ~ x ~ ,
the lemma follows for I = ~A' Since (I + g) 0 p = lop + gop, the
lemma also holds if I is a simple function.
Finally let I be a complex-valued, Lebesgue measurable function
defined on R n F', where FeR and A(F) = O. Use (11.35) to obtain a
sequence (sn) of complex-valued, Lebesgue measurable, simple functions
on R such that sn(x) -+ I(x) for all x ERn F'. Then SnO P is ~ x~-
396 Chapter VI. Integration on Product Spaces
Then I * g E~l (R) and III * gill ~ 11/111 . Ilglll' [The lunction 1* g is
called the convolution of and g.] t
Proof. Suppose for the moment that the function (x,y) ~/(x-y)g(y)
is ~ x ~-measurable on R2. We apply (21.12) to write
00 00 00 00
J J If(x-y)g(y)ldxdy=Jlg(y)I'JI/(x-y)ldxdy
-00 -00 -00 -00
00
= J Ig(y)1 .
-00
11/111 dy = 11/111' Ilglll
< 00.
III * gill = J
-00
IJ
-00
I(x - y) g(y) dYI dx
00 00
~ J J I/(x-y)g(y)ldydx
-00 -00
00 00
= J J If(x-y)g(y)ldxdy
-00 -00
Iyl ~ n}. We see at once that (B~) C 81 (R2) and Pn C n B~. Use (10.15),
k=l
(21.12), and (12.44) to write
n
A x A (k=l B~) = lim A. x
k ..... oo
A.(B~)
00 00
= lim
k-+oo
J J ~Bk(X, y) dx dy
-00 -00
.. 00
= lim
k.-..oo _
J J ~Uk(X -
-00
y) dxdy
.. 00
= lim
k.-..oo _ J J ~Uk(X) dx dy
-00
= lim 2n A(U~)
k.-..oo
= o.
00 _ __
[As in (8.14) and (12.44), g-t denotes the translate of g by -t.] Since h
can be taken to never vanish, e. g. h(x) = exp(-x2), (1) implies that the
integrals f I/(x - y) g(y)1 dy and f I/(t) g(x - t)1 dt are both finite for
R R
almost all x ER. This proves our first assertion.
It also follows from (1) and (21.13) that the mapping
h ->- f h (x) 1* g (x) dx
R
for almost all x, i. e., for all x such that these integrands are in 5!,1 (R). 0
(21.33) Theorem. Let 1 ~ P < 00, I E5!,p (R), g E5!,p' (R) [where
P' = p P 1 it P > 1 and I' = 00]' Define t * g on R by
1* g(x) = f I(x - y) g(y) dy.
R
lJ > 0 such that J lu (t)1 dt < 1. Let 1= ;[-.".,1. Then I E ~ (R), and so
-2~
Proof. Let I E~p(R) and let B > 0 be given. Apply (13.24) to obtain
a neighborhood V of 0 in R such that 21If-:v - flip < B whenever y EV.
Next use (21.36.iii) to choose an no EN such that 411fllp J Un (y) dy < B for
v'
all n ~ no' Now fix an arbitrary n ~ no' Then (21.31) or (21.32) shows
that (f * Un - f) E~p (R), and so for any h E~p' (R) [recall that l' = 00],
FUBINI'S theorem (21.13) and HOLDER'S inequality (13.4) show that
If (f * Un (x) -/(x)) hex) dxl = IJ J(f(x- y) un(y) - f(x) un(y)) dy hex) dxl
R RR
~ J IUn (y) I J If (x - y) - f (x) Ilh (x) I dx dy
R R
<
Ilhllp"
B
Thus the bounded linear functional h -'>- J (f * Un - f) hd)" on ~P' (R) has
R
norm not exceeding B, and so it follows from (15.1) that Ilf*un - fllp~ B
if p > 1. For the case p = 1, we simply take h = 1 in the above com-
putation. 0
(21.38) Remarks. (a) We now take up the Fourier transform for various
classes of functions on R. This transform is of great importance in
applications of analysis, and it is also very useful in describing the
structure of the Banach algebra ~l (R). There are close similarities, as
well as some important differences, between the theories of Fourier
series and Fourier transforms. We will point these out as they come up
in our exposition.
§ 21. The product of two measure spaces 401
Thus
1 1
21/(y)1 = 1(2n)-2" ft (X) exp(-ixy) dx - (2n)-2" f t(x- ;)exp(-ixy)dxl
R R
1
~ (2n) -2" f It (x) - t (x - ; )1· lexp(-ixy)1 dx
R
;)1 dx.
1
= (2n)-2" f It(X) - t (x -
R
~ f It(x)l·lexp(-itx) - 11 dx
R
Our next theorem shows that this transformation also preserves products
[convolution in ~l turns into pointwise multiplication in ~oJ. The
Fourier transform is also one-to-one (21.47) and its range is uniformly
dense in ~o (21.62.b). There seems to be no simple way to describe
intrinsically the functions in ~o that have the form I for some I E~l'
(21.41) Theorem. Let I, g be lunctions in ~l (R). Then
............
(i) I * g = I- g.
Proof. For all y ER we have
1
l--;g(y) = (2n)-2j I*g(x) exp(-iyx) dx
R
1 1
= (2n)-2j (2n)-2j I(x-t)g(t) dtexp(-iyx) dx
R R
=f(y)·g(y);
1 1
1* un (x) - I (x) = (2n)-"2I I(x - t)un(t) dt - (2n)-"2II(x) un(t) dt
R R
1 00 _.!.. 0
= (2nf"2I[t(x-t)-/(x)]u n (t)dt+(2n) 2 1[/(x-t)-/(x)]un(t) dt
o -00
1 00
1 00
"'"
j
00
2 e (5)
~--; u(y)dy<c'
"'"
+jlep(t)1 nu(nt) dt
'"
(6)
406 Chapter VI. Integration on Product Spaces
12 = j Irp(t)1 nu(nt) dt
lin
0:
~ :~(rx.)ntf(nrx.)+n~(:)u(O)- j~(t)n2u'(nt)dt
lin
j
0:
e'
< C. + Ce u (0) - e
t c n 2u' (nt) dt 1
lin
j
0:
e2 e e [ nrx.u(nrx.)-u(l) ] +-c
=c.+c-u(O)-c en u(nt)dt
lin
j
no:
< ce' e
+cu(O) + c + cu(O) + c
e' e e
u(y) dy
!
I
Ia = jlrp(t)1 nu(nt) dt
0:
00 00
Thus, combining (2), (6), (7), (8), and (9), we infer that
(b) Let kn (y) = (1- I~I) E[-n,nl (y). For every ex> 0, integration
by parts yields
!
« «
= 2! «
sin (ty)
ext
dy
o
2 (l--cos (ext))
- ext2
_ [Sin(t ext)]2
- ex text •
1 This is one of the few Fourier transforms that is computable by inspection.
408 Chapter VI. Integration on Product Spaces
Hence
" --.!. [Sin(tnt)]2
U.,(t) = k.,(t) = (2n) 2 n tnt .
In this case, we may take u (t) = 1 ~t • . This sequence (u.,):= 1 is known as
FEJER'S kernel. See also (21.55) inlra.
(c) Let k., (y) = exp (- ;;.). It is shown in (21.60) inlra that
u.,(t)=k.,(t)=nexp(- n~t.).
Here we take u = u1 • This sequence is called GAUSS'S kernel.
(21.46) Notes. (a) Theorem (21.43) and Examples (21.45) show why
1
the factor (2n)-"2 is used in the integral defining f. With this factor, we use
the same integral for integrating Fourier transforms that we use for inte-
grating the original functions. This is convenient, and it becomes useful
in some later developments (21.53).
(b) All of the kernels listed in (21.45) can plainly be taken to depend
upon an arbitrary positive real number a instead of a positive integer n.
The equality (21.43.iv) holds as a -+ 00 for all three kernels.
(21.47) Corollary [Uniqueness Theorem]. I I I and g are in ~l (R) and
1= g, then I = g a.e. Thus the mapping I -+ I is one-to-one.
Proof. Let h = I-g. Then h = 0, and so (21.43) and (21.45.a) imply
that
h(x) = lim
n .....oo
fO. kn(y) exp(ixy) dy = 0
R
for ally ER. Since III E~1 (R), it follows from (1), (2), (3), and LEBESGUE'S
dominated convergence theorem (12.30) that (i) holds. Theorem (21.39)
shows that the left side of (i) is a function in <ro (R); and so the second
assertion holds. Two continuous functions that are equal a.e. are the same
function, and so the last statement holds. D
We propose now to define the Fourier transform for all functions in
~2(R). We need two lemmas.
(21.50) Lemma. Suppose that I E~1 (R) n ~oo (R) and that I is real-
valued and nonnegative. Then I is in ~1 (R) and so the conclusions 01
(21.49) hold lor f.
1
=~lfl(x-~)-1
1£ n 1 + dsl S2
R
for all x ERand n EN. We set x = 0 in (21.42.i) and use (21.45.a) to infer
that
1
(2n) -2"f I(y) exp (- I~I) dy = I * un (0) ~ 11/1100 < 00 (1)
R
410 Chapter VI. Integration on Product Spaces
TI=/.
Since <roo C ~1 n ~2' it follows from (13.21) that ~1 n ~2 is dense in ~2'
For I E~2' let (In) be any .,equence in ~1 n ~2 such that 111- Inl12 -+ O.
Then (In) is a Cauchy sequence in ~2' and so (21.52) implies that
I Tills = lim
H-+OO
I Tlnlls = lim IIInl12 =
n~oo
lillis,
where (In) is as above, and so (i) holds. Since T preserves norms, it is
bounded and one-to-one. Conclusion (ii) follows from (i) and the polar
identity (16.5), which shows that the inner product is determined by the
norm. The uniqueness statement follows from the fact that if two con-
tinuous mappings [into a Hausdorff space] agree on a dense subset of
their common domain, then they agree throughout that domain.
It remains only to show that T carries ~s onto all of ~2' Since T
preserves norms and ~s is complete, the range of T is closed in ~s. Thus
it will suffice to show that the range of T is dense in ~2'
Let us compute the adjoint T* of the operator T (16.40). For
I, rp E~1 n ~2' write
f
1
g(x) = (211:)-2 rp(y) exp(ixy) dy. 2 (1)
R
1 The function Tf is called the Fouriet' transform of f. Some writers call it the
Planckerel transform, but the term "Fourier transform" is more common, and we
will retain it.
a That is. g is the inverse Fourier transform of rp. as defined in (21.49.i).
412 Chapter VI. Integration on Product Spaces
Then we have
1
(I, T*qJ) = (TI, qJ) = (2nr2 f /(y) qJ(y) dy
R
A
1
qJ * tp(y) = (2n)-2 I (x) exp(iyx) dx
R
1
= (2n) -2II(-x) g(-x) exp(iyx) dx
R
1
= (2n)-2/(lg)(x) exp(-iyx) dx
R
/'.
=Ig(y). (5)
Also, (4) and (21.39) show that I g E<ro' and so
IgE~ln<rOC~ln~2·
[We omit the easy verification of the above inclusion.] Hence (5) shows
that qJ * tp is in the range of T, for all qJ, tp E~l n ~2'
Now let (tpn) be an approximate unit (21.36) such that (tpn) C ~ n ~Il
and let qJ E~ n ~2' The preceding paragraph shows that qJ * tpn Erng T
§ 2l. The product of two measure spaces 413
hence, rng T being closed in ~2' we have <p Erng T. Thus ~1 n ~2 C rng T,
and so rng T is dense in ~2' 0
(21.54) Remarks. (a) The proof of (21.53) shows that if <p, VJ E~1 n ~2'
then there exists a function h E~1 n <ro such that Ii = <p * VJ. In fact,
h = I g where I = T* <p and g = T* '!jJ.
(b) Theorem (21.53) is the analogue for the line of the RIEsz-FISCHER
theorem (16.39). It is of course much harder to prove than (16.39). This
is accounted for by the fact that ~d[-n, n]) ::> ~2([-n, n]), while ~2(R)
neither contains nor is contained in ~dR).
(21.55) Example. PLANCHEREL'S theorem can be used to evaluate
certain integrals. We illustrate by integrating FEJER'S kernel (21.45.b).
For oc > 0, it is obvious that
1 1
II~[-cx,cxlll~ = (2n) -2 j ~[-cx,cxl(x) dx = (!)2 oc.
R
By (21.52.i), we have
1
_ (2)2
- -;; -y-.
1
sin(IXY)
(2)
r
Combining (1) and (2), we obtain
j [ Sin;IXY) dy = noc (3)
R
for all positive real numbers oc. For FEJER'S kernel, then, we have
I' q (1 1) (1 1)
(i/(x - y)la Ig (y) Iii) (i/(x - y)li> "i-a) (ig(Y)l q -q-a ).]
and that I' E~ (R). Prove that ?(y) = i Y I (y) for all y ER. [HintS.
b
Write I(b) -/(a) = J I' (x) dx and apply (12.30) to prove that lim I(b)
a HOO
'" 1 b
= lim I(a) =0. Then write I'(y) = lim (2n)-2
/1-+-00 HOO
J I'(x)
-b
exp (-iyx) dx
Integrate by parts to obtain cp' (y) = - y cp (y) for all y ER. Conclude
that d~ [cp (y) exp ( ~2 ) ] = 0 for all y ER. Show that cp (0) = 1 by
noticing that
= fob
f expO
0
O]
(- ~) r dr dO 2n. =
(21.61) Exercise. (a) Let cp E~1 (R). Suppose that cp is twice dif-
ferentiable on R, that cp', cp" are in ~1 (R), and that cp and cp' are absolutely
continuous on R. Prove that there exists a function f E~1 (R) such that
f = cpo [Use (21.59.b) to show that ?(y) = _y2 ¢(y) for ally ER. Thus
conclude that ¢ E ~l(R), and then use the inversion theorem (21.49).]
(b) Prove that cp satisfies the hypothesis of (a) if cp, cp', and cp" are
all in ~l(R) n cro(R).
(21.62) Exercise. (a) Let F be a compact subset of R and let U be an
open subset of R such that FeU. Prove that there exists a function
f E~1 (R) such that /(y) = 1 for ally EF and /(y) = 0 for ally ERn U ' .
[Use (21.61) or (21.54.a).]
(b) Let Ql(R) = {/: f E~1 (R)}. Prove that Ql(R) is dense in cro(R)
in the topology induced by the uniform norm. [Use part (a) and the
STONE-WEIERSTRASS theorem.]
r
(21.63) Exercise. (a) For a ~ 0, prove that
f [ Sin;a y) dy = 2~n .
R
(b) Compute
J
R
r
[Sin;a y ) dy.
J [sin;a r
(c) Evaluate
y) dy
R
Show that F (z) is defined for all z EK, and that F is analytic in the
entire z-plane. Show too that the nth derivative F(n) of F is given by
1
F(n) (z) = (2n) -2 (-i)n J xnexp (- izx - ~2_) t(x) dx) .
R
for all n E{O, 1,2, ... }, and so F itself must vanish identically. For real
z, this yields
=M(g*/)
= M(I).]
(f) Every multiplicative linear functional M on the Banach algebra
~l (R) has the form
1
M(/) = (2n)-"2 ft(x) exp(-iyx) dx
R
for some fixed y ER. That is, the Fourier transform f of I describes the
values at I of all multiplicative linear functionals. [Use part (e) and
(18.46.a).]
(21.66) Note. The closed ideals of <£o(X) are identified in (20.52.f).
For ~ (R) the closed ideals are far more complicated, and no complete
description of them has as yet been found. An obvious class of such
ideals is the following. For a closed set FeR, let SF = {I E~ (R) :
f (y) = 0 for all y EF}. It is known that there are closed ideals in ~ (R)
not of this form, but little more can be said at present. For a discussion,
§ 21. The product of two measure spaces 419
hence tii) holds. We obtain (iii) by adding equalities (i) and (ii).
To prove (iv) , let E = {(x, y) E [a, b] x [a, b] : y ~ x}. Since E is
compact, we have E EfJ(R2) = fJ(R) xfJ(R). Applying (2 1.8.iii) , we
obtain
J Ap(Ex) dA",(X) = J A",(E~) dAp{Y) . (I)
[a,b) [a,b)
Since Ex = [a, x] and E~ = [y, b] for all x,y E [a, b], (iii) applied to (1)
27*
420 Chapter VI. Integration on Product Spaces
Now replace y by x in (2) and rearrange the terms to obtain (iv). To get
(v), simply interchange the roles of ac and 13 in (iv), add this new equality
to (iv), and then divide by 2. 0
(21.68) Remarks. A theorem more general than (21.67) can be for-
mulated by allowing ac and 13 to be any functions that are of finite
variation on each bounded interval of R and considering the correspond-
ing signed or complex Lebesgue-Stieltjes measures. All that is involved
is to reduce the more general case to that of (21.67) by invoking the
Jordan decomposition theorem 117.16). We leave this to the reader. In
case ac and 13 are both absolutely continuous functions, i. e., A", ~ A and
Ap ~ A, (21.67) reduces to (18.19).
(21.69) Theorem [First mean value theorem for integrals]. Let p, be any
finite [nonnegative, countably additive} Borel measure on [a, b] and let
IE (,t'"([a, b]). Then there exists a real number ~ such that a < ~ < band
J I(x) dp,(x) = 1m p,([a, b]) .
[a,b]
Proof. This follows immediately from the fact that I([a, b]) is a
closed interval [perhaps a single point] and from the obvious in-
equalities
min{/(x): x E [a, b]} ~ I-'([!,b]) f I (x) dp,(x)
[a,b]
Proof. Let ac (x) = ac (a) and f3(x) = 13 (a) for all x < a, and let ac (x)
= ac (b) and 13 (x) = 13 (b) for all x > b. Let A", and Ap be the Lebesgue-
Stieltjes measures corresponding to ac and 13. Then ac(a-) = ac(a),
ac(b+) = ac(b), and f3(x+) = f3(x-) = 13 (x) for all x ER, and so (21.67.v)
§ 21. The product of two measure spaces 421
becomes
f {3(x) dA.",(x) + f IX{X+) ~ IX{X-) d)'{J(x) = ct(b) {3(b) - ct(a) {3(a) (1)
[a,b) [a,b)
=Jp,(EnGt)rp'(t)dt.D
o
(21.72) Corollary. For (X, Jt, p,), E, I, and Gt as in (21.71) and lor
p > 0, we have
(i) J IPdp, = jptP-1p,(E n Gt) dt.
E 0
Proof. Set rp(t) = tP in (21.71.i). 0
(21.73) Note. For p = 1, the equality (21.72.i) serves as a definition
of J I dp, if the Lebesgue integral on [0, is known. This technique (Xl [
E
was used by J. RADON [Theorie und Anwendungen der absolut additiven
M engenlunktionen, Sitzungsberichte Akad. Wissenschaften Wien 122,
1295-1438 (1913)J to define Lebesgue-Stieltjes integrals. Note also that
(21.72.i) holds for I E~p(X, Jt, p,) even if X is not a-finite, as the func-
tion I vanishes outside of a certain a-finite Jt-measurable set.
Our final application of FUBINI'S theorem is to the proof of a famous
theorem due to the English mathematicians G. H. HARDY (1877-1947)
and J. E. LITTLEWOOD (1885- ).
(21.74) Notation and Definitions. We will adhere to the following
notation and definitions throughout (21.74)-(21.83). First, I is a non-
negative, extended real-valued, Lebesgue measurable function on R
such that J I dA< for all compact sets F. Define functions 1,1(,), tA(l),
(Xl
F
and 1,1 by the following rules:
It is easy to see that the set Ml'> is open, since the function s ~ -1-fl
s-x
dA.
x
is continuous on Jx, 00[. Let {]Pk' Yk[}k'=l be the unique family of pair-
wise disjoint intervals such that
00
We will prove that s" = Yk' If Sx < Yk' then the equality J I dA = t(sx-x)
" set
holds; an obvious continuity argument proves this. The N.z is non-
:Y
void, so there is a real number y E Js", YkJ such that J IdA> t(y-s,,).
Sz
:Y
It follows that J I dA > t(y - x), a contradiction since y > s". Hence for
x
all x E JPk' YkL we have Sx = Yk' and so the inequality
l'k
J ldA.~ t(Yk- X )
x
~ ~ 1f d)'.
M,
0
/!
1 1
=1 ptP- )'(M\i») dt
00
IW(i) (x))P dx 1
R 0
00
= p!tP- 2!f(x) dx dt
o Mil)
= p
R 0
00
= 1
P I (x)
[fd(i) (X)]P-l
p_ 1 dx
R
1 1
°
the set Ja, ooJ under the map (x, t) -+ ~Mlj)(x) is 0 if a ~ 1, is R x [O,oo[
if a < 0, and is {(x, t) : /,1 (i) (x) > t} if ~ a < 1. Each of these sets is
product measurable. A like calculation, based on (21.75.ii), proves (ii). D
The preceding theorem is ordinarily stated with R replaced by an
interval and with smaller functions /,1('), /,1(1), and /,1. This case is con-
tained in the following corollary.
(21.77) Corollary. Let / be as in (21.74), and suppose lurther that E
is a Lebesgue measurable set such that /(E') = {O}. For p> 1, we have
(i) f W(i)(x)Pdx ~ (p pIt f /(x)Pdx
E E
lor j = r, 1; and
(ii) f
E
t
1,1 (x)P dx ~ (p 2p 1 f l(x)P dx.
E
Proof. Since f g d)' ~ f g d)' for nonnegative g, the result is an im-
E R
mediate consequence of (21.76). D
(21.78) Remark. If the set E of (21.77) is contained in an interval
!
[(X, {3], then it is clear that
and
(ii) A(M t );;:;; (1 2k)t IldA.
Gk.
Proof. Define a function g on R by
x = {I (x) if I(x»kt,
g () 0 otherwise .
We have
IA(r)(x)=supj_l_IYgdA+_l-
y-x y-x
IldA:Y>X)
x ]X,y[nGk'
Since g = 0 on {x: I (x) ;;:;; kt}, the integral on the right side of (1) is
J IdA;
N(l_k).nG~
it follows that
A(M~'»);;:;; (l~k)t I IdA.
G~
(ii) I IAdA;;:;;
E
! A(E) + 1 2 k I f(x) log+/(x) dx.
R
For 0 < P< 1, we have
(iii) I (fA(i»)Pd)';;:;; ~(~l~P (1IdAY (j=r,l)
E R
and
(iv) I (fA)P dA ;;:;; 2P ~(~; (I I dAy.
E R
----
1 Recall that log+t = max{logt, O} (0 < t < (0): see (13.37).
§ 21. The product of two measure spaces 427
J IA(j) dJ. =
E
r
0
).({y : $E IA(J) (y) > t}) dt
J ).(Mlj) n E) dt
00
=
o
=j +j ~ ~
o
11k
11k
00
)'(E) + 11k j
11k
00
+ j I (X) dx dt
GlcI
=-k-+
A(E) 1
l-k
JOO 1
t j $GlcI(X) I (X) dxdt
11k R
= -k-
A(E) + l-k
1 j I (X) { JOO$GlcI(X) tI}
dt dx. (1)
R 11k
j +dt = log/(x)
11k
if I (x) > 1 and is 0 if I (x) ~ 1 [use elementary calculus or, if you like,
(20.5)J. Thus the last line of (1) is equal to
A(E)
-k- + 1- k
1 j +
I(x) log I (x) dx . (2)
R
From (1) and (2), (i) is immediate. To prove (ii) , use (21.79.ii) instead
of (21.79.i) in (1).
To prove (iii), a slightly different argument is required. Let ex be
any positive real number. We may suppose that), (E) > OandJ Id)' < 00.
R
Then, using (21.72), we write
00
o a.lk
00
Regarding (S) as a function of 0;, we see that it has exactly one minimum
value, attained at 0; = 1 ~ h (A (E))-l f /dA. The value of (S) for this 0; is
R
1
(1- h)P
A(E)l-P
--r=p-
(f fdA,)P
R
f (tA(i))PdA~ 1 A(E)l-P
(1-k)P I-p
(f/dA)P.
E R
Letting k --+ 0, we obtain (iii). Obvious modifications in the above proof,
using (21.79.ii), yield (iv). 0
(21.81) Exercise. Let
/(xl j
~ ;(IO~X)' if
otherwise.
xE ]0, ~ [,
Prove that / ES!,l (R), but that tA(l) ;]O,-H ~ S!,l (R). [Hint. Show that
1
/ ,1(1) (x) ~ x jlogxj f or x E] 0'"2
1[ ]
.
(21.82) Exercise [T. M. FLETT]. For a function / ES!,t (R), p > 1, let
Af(P) denote the number such that II/A(I)llp = Af(P) II/lip. Define a se-
quence in s!'p (R) by
x(n-'-1)p- if x E]O, 1[, 1
{
/n(x) = 0 otherwise.
§ 22. Products of infinitely many measure spaces 429
fAt x) ~ sup ( ~
u , /f d.l : - = < t ,; x ,; u < =. t '" U) ,
r: Q' = r n Q'. For any setLl such that 0 =l= LI c r, we let ~ = X T"
yELl
[thus in particular Tr = T]. Note that ~ is not a subset of T if LI S; r.
r
For Q = {I'l> 1'2' ••. , I'm} C [the I'/s are distinct], let GD be the family
of all sets
A", x A". x ... X A"m (A"J EJ{"J) .
These are the analogues of measurable rectangles for the product of
two spaces. For an arbitrary subset LI of r,
let ~ be the smallest
algebra [not a-algebra] of subsets of ~ that contains all sets AD x TLI nD',
where Q runs through all finite subsets of LI and AD through all sets in
GD.l Let ~ be the a-algebra !/ (~). We write ,AI' for %r and vii for
vIIr. Note that for finite Q, the a-algebra vIID is just !/(GD).
(22.3) Discussion. Our aim is to construct a [countably additive]
measure p on vii such that
(i) p(T} = 1
and
IL((A
(ii) r "l
X ' " x A } x 1',
".. {"I.",,·· .•"..},)
= p".(A".) P,," (A,,") ••• p""(A,,.. )
if A"J EJ{"J (j = 1,2, ... , n). We will then prove two analogues of
FUBINI'S theorem for this product measure. We begin with a technical
lemma.
(22.4) Lemma. Let {LlQ}eEP be a pairwise disjoint family of nonvoid
r
subsets of such that U LlQ =
eEP
r.
Then the mapping
(i) (ty) -+ ((t,,)yELle)QEP = <P(t)
is a one-to-one mapping of T onto the product space
XTLI = T+.
I!EP Q
The mapping <P carries vii onto the smallest a-algebra vIIt of subsets of Tt
that contains all sets
XA LI ,
eEP e
where ALI EvilLI for all e EP and only a finite number of the sets ALI are
e Q Q
dilferent from TLI Q.
Proof. The first assertion of the lemma is obvious on a little re-
flection: the mapping <P merely "regroups" the terms t" in (t,,}yEr' To
prove the second assertion, consider first a set
A=XA"CT, (1)
yEr
1 We commit a slight inaccuracy in writing AD X TA nD' as a subset of T...
although the intended meaning is clear enough. See (22.4) for a justification.
§ 22 Products of infinitely many measure spaces 431
Since Q is finite, all but a finite number of the sets X Ay are equal to
YEA(/
TA , and it is plain that each X Ay is in ~ C vilA . Therefore W(A)
Q YEA~ II Q
and
I-'oa(A02 x TOanoia) = I-'02(A02 ) • 1 = I-'02(A02 ) .
Thus I-' is well defined by (i).
For every set A E,Ai, it is easy to see that there exists an D and a set
AD E.A"iJ such that
A = Ao x To' ; (3)
we omit the proof of this simple fact. If A1 and All are disjoint sets in,Ai,
write
(j = 1,2)
1 Again we commit a slight solecism in writing AOI x T 0i = A O2 X T 02' as
the elements of these two sets are really different entities. Again Lemma (22.4)
saves the day.
§ 22. Products of infinitely many measure spaces 433
pairwise disjoint families {Cn}:=1 c.AI such that n~ICn E.AI, but no
..AD contains all of the sets AD.. E ..AD" such that Cn = AD" x TD;.. For
example, let r = N, T = [0, I]N, and Bn = {t ET: 0 <t~ <! for
k = 1, 2, ... , n}. Then write
C1 = B~, Cn = B~ n (B1 n ... n B n - 1) for n> 1.
00
It is clear that n~ICn = T, that Cn = A.. x 1{1, ... ,nY for A.. E ..4{1, ... ,n},
and that Cn does not have the form Dn- 1 x 1{1, ... ,n-1Y for any D n- 1
E..4{1,... ,n-l}' We proceed to the theorem itself. The proof is perhaps
complicated, but its basic idea is simple enough.
(22.8) Theorem. The finitely additive measure f' on.AI admits a
unique extension over ..A that is countably additive.
Proof. The uniqueness of f"s extension over ..A [if it exists at all]
is proved by an obvious application of (21.6). To prove that f' has some
count ably additive extension over ..A, we need only show that
lim f' (Fn) = 0 (1)
..-..00
for every sequence (Fn)::"=1 such that F.. E.AI, 1;;. :::> F2 :::> ••• :::> Fn :::> ••• ,
00
and n Fn =
n=1
0. This follows from (10.37).1
1 The reader will recall that (10.37) is an exercise, for which he must supply
the proof. We believe that all readers who have worked through the text to this
point will now be able easily to prove (10.37), if they omitted to do so on first
reading § 10.
Hewitt/StlOmberg, Real and aLstract analysis 28
434 Chapter VI. Integration on Product Spaces
Let L1 = n':!l [In' By (22.7), there is a finitely additive measure ft.!1 on.;¥,1
such that ft.!1 (A.Q x T.!1n.Q') = ft.Q(A.Q) for all [I c L1 and A.Q E .A.Q. For
each n, let F~ = A.Q.. x T.!1n.Q~ C JA. It is then clear that:
each F~ belongs to .;¥,1;
ft.!1 (F~) = ft (F..) ;
F1 :::> F1 :::> ••• :::> F~ :::> ••• ;
n F.!1n =
00
and n=l
10.
n Em = n Fn =
00 00
must show that lim ft (Em) = O. Let em = {I, 2, ... , m} for each m.
tn--+oo
Note that each Em has the form Am X Te;.., where Am E.A-@",.
As noted in the proof of (22.5), we have fte"+1 = fte .. x ftm+l for all m.
It is also easy to see that a set in .A-@"'+1 is .A'e.. x .Am+l-measurable.
Now apply (22.7) and (21.12.v) to write
ft (Em) = fte .. (Am) = J ~A .. (t) dftem (t)
Te",
for m = 1, 2, 3, .... That is, we carry out all but the outermost integration
in (3), and leave the first variable unintegrated. It is plain that
# (Em) = J II,m (tl) d#l (tl)
T,
and that II,.,. (I;.) C [0, 1J. It is not the case that
lim II,m(tl ) = 0 everywhere on T I ; (4)
m--+oo
If it were the case that lim la,.,. (S2) = 0 for all Ss ET s' then by (12.24)
m--+oo
we would also have lim Jla,.,.(ta)d#s(t2) = 0; i.e., we would have
tn--+oo T"
lim II,.,. (a l ) = O. Hence there is a point as ETs such that 12,m (as) does
m--+oo
not go to zero as m ~ 00.
In this way we construct a sequence of points (at> a2, ... , an' ... ) = a
in T with the following property. For every n EN, the sequence of
numbers
defined for m > n, does not have limit zero as m ~ 00. Now the integrand
in (5) cannot vanish identically for all large m, and so
(at> aa, ... , an, Sn+t> ... , sm) EA.,. (6)
for appropriately chosen Sj E T;, and for arbitrarily large m. Since
Em = Am X T@;,., we can choose a point s(m,n) E T@~ such that
(ai' aa, ... , an) x s(m,n) EEm .
Since E.,. C En' we also have
(ai' aa, ... , an) x s(m,n) EEn . (7)
Since En =
.
An X T@', (7) implies that
(at> as, ... , an) x T@, C En. II
a E..n
00
=1
En .
n En =
00
This contradicts the equality 0 and completes the proof. 0
.. =1
28*
436 Chapter VI. Integration on Product Spaces
This proves (I). From (I) and (13.33), we infer that hn -+ 0 in measure.
Thus for every e > 0, the equality
The equality (2) is one form of the weak law ot large numbers. [See
also (22.32.b) intra.]
(22.10) Exercise. Consider a generalization of (22.9), as follows.
Let rbe arbitrary [but infinite of course]. For each y, let Ay be any set
in~and let Iy be the function on Tsuch that Iy(t) = ~A y (ty) - fly(A y}.
For Q = {Y1' Y2' ... , Yn} C r, let w(Q) be a positive number and let
hn = w(Q} I: Iy·
yEn
lor all S2 ETa· Then 51 is in ~p(T, vii, fl) and IISllip ~ II/IIp, so that 5 is a
norm nonincreasing linear translormation 01 ~p(T, vii, fl} into itsell.
Proof. Since fl(T} = 1, we have ~p(T, vii, fl} c ~l(T, vii, fl}. Thus
I is in ~1 (T, vii, fl), and it follows from (21.13.iv) that the function
liS fll~ = J I J f (SI' t2) dfl2 (t2) IP d (fll x fl2) (s1> S2)
T,XT, T,
Thus (i) is established for our special function I. To establish (ii) for I,
again let D ::::> Do and observe that the integrand in
1 Observe that (22.13) holds for any finite measure space (T. vii. t-t) and any
algebra,Al" C vii such that [/ (,AI") = vii.
2 By this limit we mean that for every e > 0, there is an Do C r
such that if
D::::> Do, then IltD' - f f dt-tllp < e. The limit in (ii) has a similar definition.
T
440 Chapter VI. Integration on Product Spaces
is equal to 1 for all UD' if tDID; EAD.; therefore in this case 1.0 = 1=1.
If tDID. ~ AD., the integrand vanishes and so 1.0 (tD' t D.) = O. Thus we have
(II) To establish (i) and (ii) for all I E~p(T, vii, p), let 5 be the
subset of ~p for which (i) and (ii) are true. We prove first that 5 is a
closed linear subspace of ~p. It is obvious that 5 is a linear subspace, and
so we have to show that it is closed. Suppose that lim II/(n) - Ilip = 0,
~oo
where I(n) E5 for n = 1, 2, 3, .... For every 8 > 0 there is a set Dn such
that "/~) - J I(n) dpllp < 8 (n = 1, 2, 3, ... ), and such that the same
n T
inequality obtains with Dn replaced by any larger finite set. By (22.11),
the inequality II/(n) - Ilip < () implies the inequalities IIf~) - ID·llp < ()
and II/~) - ID·llp < () (n = 1,2,3, ... ) for any () > 0 and any"D. Choose n
so large that 11/(n) - Ilip < 8, and let D ::::> Dn. Then we have
~ 28 + J lI(n) -II dp
T
Here we have used (13.17) and the fact that IIC11p = ICI for any con-
stant C. Thus the inclusion D::::> Dn implies that 11/.0'- J I dpll p < 38, and
T
as 8 is arbitrary, (i) follows for the function I. The relation (ii) for I is
proved in like manner, and so 5 is closed. By step (I), 5 contains all
functions of the form ~A for A E .AI', and since 5 is closed, Lemma (22.13)
1
and the trivial identity II~Ellp = p(E)P prove that 5 contains all ~B for
BEvil. Thus 5 contains all vii-measurable simple functions, and as these
are dense in ~p (13.20), the proof is complete. 0
(22.15) Note. Theorem (22.14), which is of course two theorems,
tells us all we could hope for about mean convergence of integrals over
partial products of X I'y, either to the integral (22. 14.i) or to the
yEr
integrand (22. 14.ii). For finite products, (22.14.i) becomes trivial, and
(22.14.ii) becomes meaningful [and immediately trivial] only if we agree
that integration for a void set of coordinates does nothing at all.
§ 22. Products of infinitely many measure spaces 441
where the ..It.. of (20.56) is our present ..II(n). The a-algebra ..1(0) is our
present ..II. As the measure ,u of (20.56) we take our product measure ,u,
and we define the measure 'YJ of (20.56) by
'YJ(A) = J Id,u
A
for all A E..II. [To satisfy the hypothesis that 'YJ be a signed measure,
we must consider first I E~~ (T, ..II, ,u). The complex case obviously
follows at once.J It is trivial that i'YJi < ,u.
Now look at the definition (22.16.i) of In' and use Lemma (22.12)
and (21. 12.iv). These assertions show that In is ..II(n}-measurable. We
must show that In is a LEBESGUE-RADON-NIKODYM derivative of 'YJ(n)
with respect to ,u(n) ['YJ(n) and ,u(n) are restrictions to ..II(n}, as in (20.55)J,
i. e., we must show that
'YJ(A) = J Itl d,u (1)
A
for all A E j(n). Write A = An" x Tn~, where An" E ..II!}.. , Applying
442 Chapter VI. Integration on Product Spaces
for p,-almost all t ET. [Hints. Let 17 be the product measure on T defined
from measures l7n on each {O, l}n as follows. If n is even, then l7n({O})
§ 22. Products of infinitely many measure spaces 443
= 'l}n ({1}) = ~ . If n is odd, then 'l}n ({O}) = 1 and 'l}n ({1}) = O. Show
that 'l} and '" are mutually singular. Let J{(n) be defined as in (22.16).
Show that 'l} on J{(n) is absolutely continuous with respect to '" on J{(n)
and that 2n ~s" is its LEBESGUE-RADON-NIKODYM derivative. Now
apply (20.56).]
(c) Find a set D of 'l}-measure 1 and ",-measure 0 such that _lim
00
2n~s (t)
..
(20.53.iv).
We now present an important corollary of Theorem (22.17).
(22.20) Definition. Let t = (tn ):'=1 and u = (Un ):'=1 be points in T.
We say that t und u are ultimately equal if there is some no EN such that
tn = Un for all n ~ no.
(22.21) Theorem: The Zero-One Law. Let U be a set in J{ such that lor
all t ET, t is in U il and only il all points u ultimately equal to t are also
in U. Then ",(U) is either 0 or 1.
Proof. Consider the function I = ~u and form the functions In as
in (22. 16.i). For all n and all (tv' .. , tn) and (t;, ... , t~) in TD", the points
(tv' .. , tn, un+!' un+2 , ••• ) and (t;, ... , t~, Un+!' un+2 , ••• ) are ultimately
equal, and so ~u has the same value at these two points. The definition
(22.16.i) shows that In(t) is actually a constant on all of T, and so ~u(t),
which by (22.17.i) is the limit ",-a.e. of In (t), must be a constant ",-a.e.
As ~u assumes only the values 0 and 1, we have", (U) = 0 or '" (U) = 1. 0
With Theorems (22.21) and (20.59), we can prove another pointwise
limit theorem.
(22.22) Theorem [JESSEN]' Notation is as in (22.16). The relation
_00
(i) lim t~ (t) = Tf Id",
holds tor ",-almost all t E T.
Proof. We wish to apply the limit theorem (20.59). To do this,
consider the a-algebra J{D~ of subsets of TD~' and now let J{(n) be the
family of all subsets of T having the form TDn x AD'n AD' EJ{D'. It
for i l"
is evident that J{::> J{(1) ::> J{(2) ::> •.• ::> J{(n) ::> • •• and that each
n
00
J{(n) is a a-algebra of subsets of T. A set U in .-A;, = J{(n) clearly
.. =1
satisfies the hypotheses of (22.21) and so has ",-measure 0 or 1. Define the
measure 'l} on J{ by
'l}(A) = f Id",.
A
)~ll (kQn Ak) where the A,,'s are measurable sets in any measure space.]
To prove the second statement, we need our special sets En. Observing
that F' = nQI U5n E;) and that E; = (E(k)') x 1{kY, we have
fl (F') = lim fl (
n---+oo
0 E;) = lim fl (C X(E("»,) x
k-n n---+oo k=n
TI x ... x Tn-I)
00 00
If £ flk (E(k»)
k=1
= 00, then it follows from (22.25) that li (1 -
k=n
fl" (E(k»))
= 0 for n = 1,2,3, ... , and hence that fl (F') = 0; thus fl (F) = 1. 0
Theorem (22.22) has many applications. As an example, we will use
it to prove a famous limit theorem called the strong law large numbers. at
We first present two elementary lemmas.
(22.27) Lemma. Let (fJn);:'=1 be any sequence at complex numbers
having a limit; say lim fJn = fJ. Then we also have
1>-->00
lim
n->oon
2- (fJI + fJ2 + ... + fJn) = fJ .
Proof. Given B> 0, choose no such that k > no implies that IfJ" - fJl
s
< 2. If n > no, then
§ 22. Products of infinitely many measure spaces 447
I n.
Thus for n so large that - E IP~ - PI < 26 , we have
nk=l
I~ k~ P~ - pi < E. 0
(ii) lim
n--+oo
[~f
n k=l
f~(t)] = 0 for almost all t ET.
448 Chapter VI. Integration on Product Spaces
=
H=k
i
.£ j~ fg; df-t; f tk df-tk + k=m+l ;2 flgkl2 df-tk .
7j Tk Tk
We have used (22.24.i) to write the last equality. The first term of the
last expression above is zero by hypothesis, and the second term goes
to zero as m and n go to 00, by (i). Thus we have
shows that h is also the ~1 sum f ~ .We claim that J hdf-t = O. Write
k=l T
If hdf-tl =
T
IJ: f! tk df-t +f (h - g ! tk) df-tl
T T
The first term on the right is zero and the second term goes to zero as
n goes to 00. It follows that J hdf-t = O. We now use Theorem (22.22)
l'
to write
(1)
for almost all t E T. [Recall that Qn= {I, 2, ... , n} and note that the
expression J hdf-tD" is a function of t which is independent of the first n
TD..
coordinates.] We have
The first integral on the right is zero, and the integrand in the second is
§ 22. Products of infinitely many measure spaces 449
f( 1; ~/,,) dftQn
+I,,;
k~n+1
TQn
lim
n-+oo k~n+
£ 1
~ I" (t) = 0 for almost all t ET. That is, the series £
k~1
~ I" (t)
converges a. e. in T. By (22.28), we have
lim ~(fl(t)
n--:;..oon
+ ... + In (t)) = 0
a. e. in T. 0
(22.30) Example. Let 7;" ftn and In be just as in (22.9). We have
11f"J[~ = ~ ,so that the hypotheses of (22.29) are satisfied. Hence we have
a. e.; i. e.,
lim (tt + t2 + ... + tn _~) = 0
~oo n 2
a. e. Interpreting the occurrence of obtaining heads or tails upon flipping
an unbiased coin as a 0 or 1, respectively, this result says that the portion
of heads [or tails] obtained in n flips goes to ~ as n goes to infinity for
almost all sequences of flips. This result is far stronger than that obtained
in (22.9).
The following is another version of the strong law of large numbers.
(22.31) Theorem. Let r = N. For each kEN, let g" be a junction in
~1 (T", vii", ft,,). Write g,,= If,,+ i1p", where If" and 1p" are real-valued,
and suppose that
(i) the numbers ft"({t,, ET" : If" (t,,) > ce}) and ft"({t,, ET" : 1p,,(t,,) > ce})
are independent 01 k lor every real number ce.
Let I" be the lunction on T such that I,,(t) = g,,(t,,). Then we have
We will show first that for almost all t E T, there is a positive integer mo
[depending on t] such that
Im(t) = I'",(t) if m ~ mo. (1)
Under the hypothesis (i), it is trivial to show that the numbers
f-lk({tk E T k : Igk(tk)1 > IX}) are independent of k for every IX ER. Using
this fact, we have
00 00
E f-l({t ET: Ik(t) =1= I~(t)}) = E f-l({t ET: IIk(t)1 > k})
k=1 k=l
00
J I~ df-lll~ =
E k- 2 [J 1;'2 df-l - (J I;' df-l)2]
"" 00
E k- 2 III;' -
k=1 T 11=1 T T
"" 2
~ E k- J 1;'2 df-l . (5)
11=1 T
From (i), from the primeval definition (12.2) of the integral, and from
(12.21), it is clear that
£ k- < f
00
2 dx =~
k=fJ+I 0 (x+ P)2 P
f
k=p
k- 2 P < ~ +1~ 2. (8)
lim
....... 00
[~n k=1;1 f~(t) - ~n k=E1 Tff~ dll-] = 0
for almost all t ET. In view of (3), the present proof will be completed
by showing that
(9)
T
J I~ dll- = T,J h" dll-1 ' (10)
(ii)
k=1
converges for almost all t ET.
(b) Prove the following analogue of (22.29), which is known as
the weak lawai large numbers. Again notation is as in (22.29). Replace
(22.29.i) by the hypothesis
+1'
in measure.
(22.33) Exercise. Let T..= {O, 1,2, ... , r - I} and let fl .. (A) =
(n = 1,2, ... ). For a fixed l E{O, 1, ... , r - I}, define
lift.. =l,
{
g.. (t) = 0 if t.. =1= l
for all t in the product space T.
Prove that
.
00
X =
~
'\' r-" x '
n=1
x.. E{0,1, ... , r - I}, and x.. =1= 0 for infinitely many n's. Prove that
lim ~ b" (x) = ~ for [Lebesgue] almost all x E]0,1[. [This follows
k--->oo k r
from (22.31).]
(22.35) We now present an application of the limit theorem (20.56)
somewhat different from those given above. As in (22.16) and (22.17),
§ 22. Products of infinitely many measure spaces 453
let F={1,2,3, ... }, let Qn={1,2" ... ,n}, and let vIt(n) be the
a-algebra of all sets Aa x Ta, for Aa Evita . Now consider measures II.n
n " r-
and 'YIn on Cr,.,.A;,) such that P,n Cr,.) = 'YIn Cr,.) = 1, and let p, und 'YI be
1& "
the product measures formed from the measures P,n and 'YIn' respectively.
Our first result deals with the case that 'YIn ~ P,n for all n, and establishes
a remarkable fact about 'YI and p,.
(22.36) Theorem [So KAKUTANI]. Notation is as in (22.35). Suppose
that 'YIn ~ P,n lor all n. Then we have either
(i) 'YI ~ P,
or
(ii) 'YI 1. p,.
Let In be a lunction in ~i Cr.., .A;" P,n) such that
(iii) J In dp,n= 'YIn(En)
E"
lor all En E.A;" i. e., let In be a LEBESGUE-RADON-NIKODYM derivative
dd'YJ
n in the sense 01 (19.43). Then (i) holds il and only it
/-In
00 1
(iv) II( J I! dp,,.) > 0,
k=l Tk
and (ii) holds il and only il
00 1
(v) II (J I! dp,k) = O.
k=l Tk
as (13.4) shows. Hence the infinite product in (iv) and (v) is a number
in [0,1]. For each n EN, consider the finite product Ta.. and the product
measures P,1 x . .. X p,n and 'YIl x ... x 'YIn on vita... It follows from
(21.29) by induction on n that
'YIl x . .. x 'YIn ~ P,1 X •• • X P,n
and that the function
(1)
is a LEBESGUE-RADON-NIKODYM derivative of 'YIl X· •• x 'YIn with
respect to fl-I x ... X p,n' Now consider the a-algebra vIt(n) of all sets
Aa11 x Ta, where Aa1E1
tI
vita" . Let I(n) be the function on T such that
and let p,(n) and 'YI(n) be the measures p, and 'YI, respectively, restricted
454 Chapter VI. Integration on Product Spaces
II(n) df-t =
T
l1 Tk
II,. df-t,. = kiJ 'fJ,.(T,.) = 1 (2)
exists for f-t-almost all t ET and is a derivative of 'fJ with respect to f-t
in the sense of (20.53).
Suppose that (v) holds. Then we apply (22.24.i), (12.23), and (4) to
write
(5)
From (5) it follows that 1= 0 f-t-a.e., and so (20.53) implies that 'fJ .1 "',
since the ",-absolutely continuous part of 'fJ is obtained by integrating I.
Regardless of the value of /100/ Ii1df-t,., we can compute as follows.
Tk
For m < n, we use (13.4) and (2) to write
= (6)
T T
Now taking note of (3) and (22.24.i), we write
=
k=m+l
Ii Tk
II: df-t,. . (7)
§ 22. Products of infinitely many measure spaces 455
/1: dft k) 2]
1
lim n
m, ....... OOk=m+1
(/I:dftk) = 1.
Tk
Hence (8) shows that (l<n»):=1 is a Cauchy sequence in ~l(T, JI, ft).
We now appeal to (20.58) and (20.57) to infer that 'YJ ~ ft. 0
(22.37) Remarks. Notation is as in (22.35). If not all 'YJn are absolutely
continuous with respect to ft", then 'YJ cannot be absolutely continuous
with respect to ft, but it may still have a large absolutely continuous
part. Suppose that for some lEN, we have
(i) 'YJ! = lX!e! + (1 - IX!) a! •
°
where e! and a! are measures on (~, JI!) such that e!(~) = a!(~) = 1.
e! ~ ft!, a! ..L ft!, and ~ IX! < 1. If IX! = 0, i.e., if 'YJ! ..L ft!, then (21.29)
shows that 'YJ ..L ft. Otherwise, let 'YJ' be the product of all 'YJk for k =1= l,
on the space 1{W. It is easy to see that 'YJ = IX! (e! x 'YJ') + (1 - IX!) (a! x 'YJ').
As (a! x 'YJ') ..L ft. 'YJ cannot be absolutely continuous with respect to ft,
but it is possible for e! x 'YJ' to be singular with respect to ft, absolutely
continuous with respect to ft, or to be "mixed". A precise description
of'YJ in terms of the decomposition (i) for all lEN could be given: we leave
the details to any interested readers.
(22.38) Exercise. Notation is as in (22.35). Let T., = {O, I} for all
n EN, and let /Z be a sequence (a.,):=1 with values in ]0, 1[. Let ft ...
be the measure on (T, JI) that is the product of the measures ft" on
{O, I} such that ft" ({O}) = IX", ft" ({I}) = 1 - IX". Suppose that /Z and p
are any two such sequences.
(a) Prove that exactly one of the two following assertions holds:
(i) ft ... ~ ftp and ftp ~ ft ... ;
or
(ii) ft ... ..L ftp·
Prove also that (i) holds if and only if
00 1 1 1 1
(iii) 1: (1- IX!{J: - (1 - IXn)9(I- (J,,)9) < 00,
n=1
and that (ii) holds if and only if
(iv) the series in (iii) diverges.
[Hints. Apply (22.36) to the measures ft ... and ftp. The factor measures
are evidently absolutely continuous with respect to each other, and the
integral over T., of I! is IX! {J! + (1 - a.,)9(1 - {J,,)9. Now apply (22.29).
1 1 1 1 1
can be applied. Otherwise (vi) shows that the terms of the series in (iii)
do not have limit O.J
(22.39) Exercise. Prove that there is a set 8 of measures on (R, gj (R))
such that: 0" (R) = 1 for all 0" E8; each 0" E8 is regular; each 0" E8
is continuous; each 0" has support the interval [0, 1]; 0" 1.. 0"' for distinct 0"
and 0"' in 8; and § = c. [Hints. Let T be as in (22.38), and let ffJ (t)
E 2-k tk
00
= for t ET. The mapping ffJ carries Tonto [0, 1]. For every
k=!
number I' E]0, 1[, let fly be the measure on T constructed from the con-
stant sequence (1', 1', 1', ...) as in (22.38). For I' =!= 1", fly and fly' are
obviously mutually singular. For I' E]0, 1[ let ay be the measure on [0, 1J
constructed as in (12.45) and (12.46) from the measure fly on T and the
continuous mapping ffJ. It is simple to verify that {ay : I' E ]0, 1 [} can be
taken for the set 8 of measures.]
(22.40) Exercise. A set 8 1 of measures on gj (R) with all of the pro-
perties of 8 [see (22.39)] except that supports be [0, 1] can be constructed
without recourse to KAKUTANI'S theorem (22.36). Fill in the details of
the following argument. Let T = {O, I}N and consider the measure fll
2
on (T, .-'I) as in (22.39). For each u = (u}> u 2, •••• 1t.. , ••• ) ET, Let ffJu be
the mapping of T into [0, 1] given by
q; of T into R by q; (t) = 2 E 3-"t". For I' E]0, 1[, let Ty be the image of
k=l
Ily under q; as in (12.45) and (12.46). Prove that the support of each Ty
is the Cantor ternary set. Prove too that T.! is the Lebesgue-Stieltjes
2
measure that corresponds to LEBESGUE'S singular function (8.28).
(b) Prove that f x dTy(X) = 1 - y. [Hint. The following steps are
[0,1]
easy to check:
f
~,I]
x dTy(X) =
T
f (2 g 3-"t,,) dlly (t) = 211 f t" dlly(t) 3-"
T
00
f
[0,1]
x 2dTy(X) = 4
T
f C~3-"t"rdlly(t)
= 4 [g f t"dlly(t) + 21~ C£1 f t"t, dlly(t))]
3-210 3-"-'
T T
(1.) 11 n = a1 _
n
(1 - y) ( ;,
an kJ
(n)j 11 n-)
.) .
1
1~1 ;
[Hints. For n EN, we have
in the form
note that
n!
(n - ill! (il - i.l! (i. - isl! ... (i h - l - ihl! i k ! =
(n)il (il)
i.'"
(ik-l)
ik '
the sum 1:" being taken over all k-tuples (jI' j2' ... , jk) of positive integers
such thatjI + j2 + . , . + jk = n. [Hint. Rewrite (i).]
Index of Symbols
dE 155 L [Lebesgue integral] 164
arg, Arg 50 lp, lp(D) 194
19 loo (D) 219
30 ~1 173
301 ~~ 170
~p 188
f!I(X) [Borel sets] 132
~oo 347
~(X) 83 ~4> 203
~(H) 251 ~ log+ ~ 203
B.(x) [E-neighborhood] 60 L (f, IX, .1) [Darboux sum] 106
~(A,B) 211
lim, lim 76, 256
c 19 ~iPa; 270
Co (D) 218 A [Lebesgue measure] 120
(£(X) 84 Aa; 120
(£0 (X), (£00 (X) 86
430
Dv D2 271 [,-measurable sets] 128
D" [Dirichlet kernel] 292 [Lebesgue measurable sets]
D+, D+, D-, D_ 257 128
.@([a, b]) 105 viifJ [p,-measurable sets] 127
diam 67 m(D) 219
domf [domain] 7 921 (X) 88
!5xy [KRONECKER'S delta] 11
M(X) 360
t,. [symmetric difference] 4 max {x, y}, min {x, y} 8
E" [E,,(f) = f(a)] 114 max{t, g}, min{f, g} 82
tin 430 161
exp [exponential] 51
Ea [Ea(A) = ~A (a)] 120 N [positive integers] 2
1)", 1)8 366 .t¥,j,% 430
Fa 68 ordA 28
F(X, d, p,) 354 w 28
Ga 68
[J 29
[J [Cr] 429
I [nonnegative linear function-
al on (£00] 114
I 116
P' [ = _P-]
P-l
190
HEWITT, E. 137, 250, 378, 445 infinite products of measure spaces 429
HEWITT, M. 32 infinite set 21
Hilbert parallelotope 254 - -, countably 22
Hilbert space 196, 235 initial segment 28
-, bounded linear functionals on a - - of a well ordered set 16
248,255 inner product space 195, 234
-, orthogonal dimension of a 247 integers 2
-, pre- 195 - , positive 2
-, projections in a 253 integrability, Riemann 105, 183
HIRSCHMAN JR., I. I. 250 integrable function 173
HOLDER'S inequality for 1 < p < 00 190 - -, locally 186
- - for 0 < p < 1 191 - -, Riemann-Stieltjes 106
- -, generalized 200 integral as a set function 175
HOLLADAY, J. C. 103 integral of a function, indefinite 272
HOPF's extension theorem 142 - relative to a complex measure 312
HUNTINGTON, E. V. 5 integral representation of convex func-
tions 300
ideals in a ring 33 integral, absolute continuity of the 176
- in (£0 (X) 365 -, Denjoy 299
- in £l(R) 417 -, derivative of an indefinite 275
identity, PARSEVAL'S 245,246,250 -, iterated 385
-, polar 235 -, Lebesgue 164
image sets, measures on 161 -, Lebesgue integral as an extension
images of measures, continuous 180 of the Riemann 184
imaginary part of a complex number 48 -, linearity of the 169, 171, 174
increasing function, strictly 105 -, Riemann 107
indefinite integral of a function 272 -, Riemann-Stieltjes 106
- -, derivative of an 275 integrals as areas 392
independence, linear 17 for finitely additive measures 355
indicatrix of a function, Banach 271 on infinite product spaces, iterated
indiscrete topology 56 439
induction, transfinite 17 on the completion of a measure space
inequalities, CLARKSON'S 225, 227 186
-, HOLDER'S 190, 191 -, change of variable in 342
-, JENSEN'S 202 -, first mean value theorem for 420
inequality, BESSEL'S 239 -, fundamental theorem of the integral
-, BUNYAKOVSKII'S 190 calculus for Lebesgue 285
-, CAUCHY-BuNYAKOVSKrI-SCHWARZ -, integration by parts for Lebesgue-
234 Stieltjes 419
-, CAUCHY'S 190 -, second mean value theorem for 420
-, generalized HOLDER'S 200 -, summability of Fourier 404
-, MINKOWSKI'S 191 integration by parts 287
-, SCHWARZ'S 190, 234 - - for Lebesgue-Stieltjes inte-
-, YOUNG'S 189 grals 419
infimum in an ordered field 44 by substitution 342
- inR#69 integration, term by term 171, 175
infinite product measures 432 interior of a set 56
-, absolute continuity of 453 intersections of sets 2
-, singular measures induced by interval, bounded 54
456 -, closed 54
-, singularity of 453 -, open 54
infinite product spaces, iterated integrals -, subdivision of an 105
on 439 -, unbounded 54
Hewitt/Stromberg, Real and abstract analysis 30·
468 Index of authors and terms