Documente Academic
Documente Profesional
Documente Cultură
OF MODERN
ANALYSIS
AVNER FRIEDMAN
Northwestern University
I ..
'
' '
'
!' '
1.1 Rings and Algebras, 1 3.3 Completion of Metric Spaces; H"' Spaces, 99
1.2 Definition of Measure, 4 3.4 Complete Metric Spaces, 105
I .3 Outer Measure, 7 3.5 Compact Metric Spaces, 107
1.4 Construction of Outer Measures, 11 3.6 Continuous Functions on Metric Spaces, 110
1.5 Completion of Measures, 12 3.7 The Stone-Weierstrass Theorem, 116
1.6 The Lebesgue and the Lebesgue-Stieltjes Measures, 13 3.8 A Fixed-Point Theorem and Applications, liB
1.7 Metric Spaces, 15 Chapter 4. ELEMENTS OF FUNCTIONAL ANALYSIS IN
BANACH SPACES, 123
1.8 Metric Outer Measure, 17
4.1 Linear Normed Spaces, 123
1.9 Construction of Metric Outer Measures, 20
4.2 Subspaces and Bases, 129
1.10 Signed Measures, 25
4.3 Finite-Dimensional Normed Linear Spaces, 132
Chapter 2. INTEGRATION, 29 4.4 Linear Transformations, 135
2.1 Definition of Measurable Fuctions, 29 4.5 The Principle of Uniform Boundedness, 139
2.2 Operations on Measurable Functions, 32 4.6 The Open-Mapping Theorem and the Closed-Graph Theorem, 141
2.3 Egoroff's Theorem, 35 4.7 Applications to Partial Differential Equations, 145
2.4 Convergence in Measure, 37 4.8 The Hahn-Banach Theorem, 150
2.5 Integrals of Simple Functions, 40 4.9 Applications to the Dirichlet Problem, 155
2.6 Definition of the Integral, 42 4.10 Conjugate Spaces and Reflexive Spaces, 159
2.7 Elementary Properties of Integrals, 48 4.1l Tychonoff's Theorem, 164
2.8 Sequences of Integrable Functions, 51 4.12 "'{eak Topology in Conjugate Spaces, 168
iv
Vi CONTENTS
Chapter 5.
5.1
COMPLETELY CONTINUOUS OPERATORS
186
Basic Properties, 186
,
MEASURE
5.2
5.3
The Fredholm-Riesz-Schauder Theory, 189
Elements of Spectral Theory, 194
THEORY
5.4 Application to the Dirichlet Problem, 197
of those elements which belong to all the sets E,. The following relations can
easily be verified:
(u ,)" n
~-e A
= ~.
~E:=.A
(1.1.1)
(n = u
~!:";:A.
~)<
. i%E A
~. (1.1.2)
2 MEASURE THEORY RINGS AND ALGEBRAS 3
Let {En} be a sequence of sets (n = 1,2, ... ). The superior limit of {En} is, (b) and (d) also hold. Indeed, (d) follows from (f) with A = 0. and (b)
is the set consisting of those points which belong to infinitely many En. follows from
We denote it by
A - B = A n B" = (Ac u B)".
E* =lim sup E.= flm E.= Tiii1 E.
I"'""""'GC n-+~
A disjoint union is a union of sets that are mutually disjoint. We intro-
duce a condition that is weaker than (c), namely:
Similarly, the inferior limit
(c') If A E 9t, Be 9t and An B = 0, then Au BE !Jt.
E* =lim inf E.= lim E.= lim E.
n-+oo n-+oo n ~~ We then have: A class 9t satisfying (a), (b), (c') is a ring. Indeed, (c)
is the set consisting of all those points which belong to all {uta finite number follows from A u B = A u (B - A).
of theE,. If E* = E*, then we say that the sequence {E.} has a limit, and we Next we introduce a condition weaker than (e):
denote the set '1' by lim E, or lim E.
tt-[1:) " (e') Any countable disjoint union of sets A. (n = 1,2, ... ) which belong
Note that if {E.} is monotone increasing (decreasing)-that is, if to !X also belongs to !X.
E. c ,+ 1 (. => E,+ 1) for all n 2 I, then limE, exists.
Theorem 1 .1.1 If !X satisfies (a), (b), (e') [and {d)], then 3 is a cr-ring
"
The following two definitions are fundamental in measure theory. (a-algebra).
Definition I .I .I A class fJt of sets is called a ring if it has the following Proof. Taking, in (e'), A. = 0 for n 2 3, we see that (c') holds. ~ence
properties : (c) follows, and 9f is a ring. To prove (e), let {A.} be a sequence of sets m 9f.
Then the sets B., defined by
(a) 0 e 3.
(b) If A E fJf and Be at, then A -Be 3.
(c) If A e .11 and Be fft, then A v Be Bf.
A ring 3 is called an algebra if it has the additional property: also belong to 91. Since B. n B,. = 0 if n #- m and UB. = UA., (e') implies
that U A. e 91. " "
(d) X efH.
f"\ "
Note that if !f/1. is a ring and if A, B e fft, then A --' B = A - (A - B) is in
!f/1.. Note also that if !f/1. is a ring, then any finite union and any finite intersection PROBLEMS
of sets in ffi are in ffi.
Definition 1.1.1 A class 3 of sets is called a u-ring if it has the properties
1.1.1. (tim E.)"= Jim~. (rm E.)"= tim~.
" n " . .,
(a), (b), and the additional property:
re5ult holds for u-algehras, and we speak of the rr-algebra generated by 2. whenever E Em,
,., FE (}l
'
En F = 0- r, is called finitely additive if, for any
1.1.7. If 2 i5 any class of sets, then every set in :Jf(ft) can be covered by positive integer m,
(that is, is contained in) a finite union of sets of 2. [ Hinl: The class .X' of sets
that can he covered by finite unions of sets of 9 forms a ring.]
1.1.8. If 9 is any class of sets, then every set in .:1'(9) can be covered
/leo\ Ej) j~l J.i(Ej)
=
(1.2.3)
mutually disjoint intervals on the real line, having the form {1; a< t:::; h}, "1 Let".. he a measure with domain Ci. Then
where a.b are any real numbers. Then (i is a ring but not an algebra. Tlreorem 1.....
1.1. I 1. Find W/.(01) when 1!1! consists of two distinct sets, each consisting (i) 1fE E fi, FE (i, E c F, then l!(E)::::;; l!(F).
of one point.
(ii) /fEE (i, FE (i, E c F, J.l(F) < oo, then
1.2 DEFINITION OF MEASURE (1.2.4)
tt(F - E) = p(f) - J.i(E).
We introduce two symbols, "+oo" (or, briet1y, oo) and "-oo"
(ca \led " pi us infinity " and " minus infinity "), and the following relations: (iii) If {En} is a monotone-increasing sequence of sets of Ci, then
- oo < x < oo if x is real. Any entity that is either a real number or one of
(1.2.5)
these two symbols will be called an extended real number. We also define
arithmetic relations by: x oo = :n. oo + (x) = oo, x(oo) =
u-J if x > 0, x( co)= + oc if x < 0, and so forth. The relations (iv) if {En} is a monotone-decreasing sequence of sets of Ci, and if
+ oo + (- oo), 0( oo) are not defined. J.l(Eno) < oo for some no, then (1.2.5) holds.
A srcrfunction pis a function defmcd on a class :JJ of sets. It is said to be
extended real-valued, if its values are extended real numbers. iis called the Proof. To prove (i), write F as a disjoint union F = E u (F - E). Then
domain (of definition) of tL
J.i(F) = J.l(E) + fJ(F- ). (1.2.6)
Let p be a set function defined on a ring :Jf. We say that p is completely
additiue if Since fJ(F _ E) ~ O, (i) follows. If /l(F) < oo, then /.1(E) < oo, and ( 1.2.4)
follows from (l.2.6). To prove (iii) we first note, by Problem 1.1.3, that
(1.2.1)
lim E. belongs to fi. Next, setting Eu = 0, we have
whenever the E" are mutually disjoint sets of-'* such that U E. is also in H.
Definition 1.2.1 A measure is an extended real-valued set function
p. having the following properties:
(i) The domain fi of /l is a u-algcbra.
(ii) J1 is nonnegative on fi.
(iii) Jl is completely additive on fi. which proves (iii).
(iv) J.i(0) = 0.
6 MEASURE THEORY
OUTER MEASURE 7
To prove (iv), we use (ii):
Corollary 1.2.3 1fli~ En exists and if llCQ E.) < oo, then
1
II [1i"m (E. E.)] = p[E." - Ii:n E.J = p(E.J - .u(li:n E.).
0 -
8 MEASURE THEORY
Proof. We shall give the proof in several steps.
whenever the sets 1 and their union belong to the domain of v. Finally, (i) If J.t*() = O, then E is measurable. Indeed, for any set A we have
it is called countahly subadditive if
J.l*(A n ) + J.l*(A -E) :S J.t*(E) + p*(A) = p*(A)
Taking 1 = 0 in (1.3.2), for j:?: n, and using (v), we see that an outer ,u*(A) ;::-: ,u*[A n ( 1 u EJ] + J.l*[A - (El u E2)].
measure is finitely subadditive. Since A is an arbitrary set. the ,u*-mcasurability of E 1 u 2 fol\ows.
It is obvious that every measure whose domain consists of all the
subsets of X is an outer measure. (v) Jf
1
,
2
belong to a, then E 1 - 2 belongs to a. Indeed, this
The usefulness of outer measures is primarily in the role they play in follows from the relation
the construction of measures. On one hand, it is fairly easy (see Section 1.4) to
construct outer measures. On the other hand, one can always construct a
measure from an outer measure. The latter assertion will be proved in this using {iii) and (iv).
section (Theorem 1.3.1). We have proved so far that ct is an algebra.
(vi) Let {E.} be a sequence of mutually disjoint sets of a and let
Definition 1.3.2 Given an outer measure J.l*, we say that a set E is
J.t*-measurable (or, briefly, measurable) if s
"
= lJ Em. Then, for any
m=l
n:?: 1,
n
p*(A) = p*(A n ) + J.l*(A - E) (1.3.3) p*(A n s.) = I J.l*(A n Em) (1.3.5)
m-=1
for any subset A of X. for all sets A. The proof of (1.3.5) is by induction on n. The case n = 1 ts
The (vague) motivation for this definition is that the sets we want to
obvious. The passage from n ton + 1 is given by
single out as measurable should be such that p* will become additive on them.
Since 11* is subadd iti ve, ( 1.3.3) is equivalent to the inequality: p*(A n s.+ tl = J.l*(A n s.+ I n S.) + {l*(A n s.,! - s.)
J.t"'(A):?: J.l*(A n E)+ p*(A - ). ( 1.3.4) = !t*(A n S.) + p*(A n E., 1)
n+ 1
Theorem 1.3.1 Let 1-1* be an outer measure and denote by a the class
= I
m-=1
J.l*(A n Em)+ ,u*(A n E.i-1) = J 1
J.l*(A n Eml
.
I
J.I.*(A) = t~"(A n S.) + J.I.*(A- S.) For example, the bounded open intervals on the real line form a sequential
covering class of the real line.
~ J.," tt*(A n E,) + t~*(A - S). Let A. be an extended real-valued, nonnegative set function, with domain
.i:, such that J.(0) = 0. For each set A (of X), let
Taking n __,. OC! and using (I
. .3.6) , v.e
. get
~he double sequence {E.k}. with indices nand k, then satisfies We call ji the completion of II
:(. ~
U A. c n,k=l
n=l
U E.k Proof. We first shov. that R is a a-algebra. Let E e Ci, l';' c A E Ci.
(A) = 0. Then " - A = 0 n A" c E" n ,\'" = (E u l'>i"Y c E". Hence,
Hence, by the definition ( 1.4.1) and by ( 1.4.3). r u NY= (Ec- A) u N' where N' c: A. Since [Ec- A] E (i, it follows that
(E u N)' Ea. Si!!ce 1 - 2 ': (] u 2y, it remains to show that a countable
Jl*CV! AnJ ~ n.~l }(.) ~ ntl l~~*(A.) +f.!~ ntl /I*(Arr) + r;. union of sets of (i is again in (i. Let these sets be En U Nm where Nn C An,
p(An) = 0. Then "' (" cc )
Since is arbitrary, ( 1.4.2) follows.
f. U (E.uN.)= U En uN,
"=1 n=l ,
Theorems 1.4.1 and 1.3. I enable us to construct a measure frolllany
given pair .If", ..1.. where N =
UN.. Since N UA.= A and p(A)
~1
c:
n~I _
= 0, the assertion follows .
Let .:>r consist of all the open intervals (a.b) of R 1 and defme A by ).(0) =0
and
A{(a.b)} f(b) -f(a) if a< b. (1.6.3)
which is impossible (since this would imply that ~, n E~, i=_ 0).
=
The outer measure pj determined by this pair .:.f-, ). is ca!led the Lebesgue- Suppose now that F is measurable. By ~robl.em 1.6.5, each set Fk
Stieltjes outer measure induced by f The measure 11 1 corresponding to this (k 2: 2) is then also measurable, and p(Fk) = p(f). Smce all the sets Fk are
outer measure is called the Lebesgue-Stieltjes measure induced by f disjoint and are contained in [0, 1], it fo!lows that
It will be proved later on that closed sets and open sets are Lebesgue-
mt-1
measurable. mJ-1-(F) = L.
k= 2
p(f 1,) :-::; I for any integer m.
PROBLEMS
Hence p(F) = 0. . .
1.6.1. The Lebesgue measure of a point is zero. Consider next the sets G, = {r + 1/; 1J e F}, where r ts any ratwnal
1.6.2. The Lebesgue measure of a countable set of points is zero. number. By Problem 1.6.5, G, is measurable and J-1-(G,) = 0. If r and s are
1.63. The outer Lebesgue measure of a closed bounded interval distinct rational numbers, then G, n G, = 0. Since, further
[a,b] on the rea! line is equal to h - a. [Hint: Usc the Heine-Bore\ theorem ry,
[a,b),(a,b] is equal to h- a.
it follows that J-1-{(- oc ,co)} = 0, which is impossible.
1.6.5. Consider the transformation Tx = :xx + fl from the real line
on1o itself, where :x.# are real numbers and 11. # 0. It maps sets E onto sets T(E).
Denote by J.1 (!t"') the Lebesgue measure (outer measure) on the real line.
Prove: 1.7 METRIC SPACES
(a) For any set , !<*(T(E)) = l:xllt*(t.:). Suppose that there exists a real-valued fun~tion p de~ned for every
(b) E is Lebesgue-measurable if and only if T(E) is Lebesgue-measur- ordered pair (x,_r) of points of X, having the followmg propertieS:
able.
(c) If E is Lebesgue-measurable, then J-1-(T(EJ) = l:;~lp(). (i) p(x,y) 2: 0, and p(x.y) = 0 if and only if x = y.
Additional information on the Lebesgue measure (in n dimensions) (ii) p(x.y) = p(y.x) (symmetry). . . .
wi l! he given in Problems 1. 9. 1-1 .9. I I . J\.1 ore information on the Lebcsgue- (iii) p(x,z) :-::; p(x.y) + p(y,z) (the tnangle mequahty).
Example. The function A set is said to be closed if its complement is an open set. Note that the
112
p(x,y) =lx - Yl ={E(x, - J= I
}';)
2
} (1.7.1) closed balls are closed sets.
The interior of a set E is the set of all points y in such that there is an
is a metric function in the Euclidean space W. We call it the Euclidean metric. ~:-neighborhood of y contained in E. We denote the interior of E by int E.
When we speak of R" as a metric space, the metric we usua11y have in mind is A sequence {xnl is called a Caurhy sequence if p(x",x,)-> 0 as n.m .... CXJ.
the one given by (1.7.1). Note that every convergent sequence is a Cauchy sequence. Also. a Cauchy
We define the distance between two sets A and B by sequence cannot have more than one point of accumulation. If every Cauchy
sequence is convergent, then we say that the metric space is complete. The
p(A,B) = inf p(x,y). Euclidean space (with the Euclidean metric) is complete, but the subset
X~A
y~B consisting of all points with rational components forms an incomplete metric
If A = {x}, then we also write p(x,B) for p(A,B). space (under the Euclidean metric).
A set A in a metric space (X,p) is said to be bounded if its diameter, 1n the next section we shall need the following definition:
d(A) =sup p(x,y), Definition Denote by g the tHing generated by the class of all the
'<$A
y~.t
open sets of X. The sets of /4 are called Borel sets.
is finite. For any x EX, e > 0, the set Note that dd coincides with the u-algebra generated by the class of all
B(y,e) = {y; p(x,y) < e} the open sets of X.
n-1
LemiiUJ 8.1.1 Let 11* be a metric outer measure and let A, B be any Using the relation A 2 ~ => U Gzk and the property (vi) of metric outer measures,
k~t
sets such that A c B, B open, 11--*(A) < "" For any positive integer n, let
we get
A"= {x; x E A, p(x,B<) ~ ~} ii"'(A) ~ Ji*(Az.) 2: Ji* cv: Gn) = :t: J1*(G2k)
Then Since p*(A) < oo, this shows that the series
. _Proof . Since A. c A.+ 1 and p* is monotone, the sequence {Ji*(A.)} is convergent. Similarly one shows that the series
ISan mcreasmg sequence. Also, .u*(A") :::;: ii"'(A). It therefore suffices to show
that
lim ~t*(A 2 ") ~ I!*( A). (1.8.1)
n~oo is convergent. Taking n--> ro in (1.8.3), we then obtain
Each point y of A lies in the open set B. Hence there is some ~:-neighbor
hood of Y that is contained in B. It follows that p(y,B") > e. This shows that
(1.8.2) Proof. Let F be a closed set. For any set A with p*(A) < oo, A - F is
contained in the open set F. Hence, by Lemma 1. 8.1, there is a sequence E. of
Let G" =An+ I -An for n ~ l. Then (1.8.2) implies that, for any subsets of A - F such that
n ~ 1,
(n = 1,2, ... ), (1.8.4)
Proof Denote by a the a-algebra of all the p*-measurable sets. Taking n ....... ro, we get
Every open set E has the form E = X - F, where X E a and FE a (since F
is closed). Hence E E a. It follows that a contains the a-ring generated by
the class of all the open sets-that is, a ::::> !Jd.
p~col E.)= !~n; ~~:col El) ~ J~ ll~(Ed.
Jt remains to prove the condition (vi) in Definition 1.8.1. I~ p(A,B) > 0,
PROBLEMS then p(A,B) > lin for all n 2: n 0 For any e > 0 and n :?: n 0 , there ts a sequence
1.8 .I. Prove the converse of Corollary I .8. 3-that is, if ll * is an outer {E.k} in f . such that
measure and if every open set is measurable, then ll* is a metric outer measure.
1.8.2. Let fJ be a measure with domain a. For any two sets E and A v B c:
"' .,,
U
Fin a, let p(E,F) = 11[(- F) u (F-E)]. Prove that p(E,F) = p(F,E) and I<= I
+ p(F,G).
p(E,G) .:-::; p(E,F)
1.8.3. Let (X,p) be a metric space and let {x.} be a sequence of points f
k=l
A.( E)::::::; p:(A v B) + B.
in X. Deft ne p*( E) to be the n urn her of points x. that belong to E. Prove that
JJ* is a metric outer measure. . d(E ) < 1/n no set E k can intersect both A and B. Hence the sequence
Stnce nk ' " {E } d {E } the
1.8.4. Let (X,p) be a metric space. Define JJ*() = I if E i= 0 and {E.d can be written as a disjoint union of two sequences . an nk"'
!< *(0) = 0. Is ll * a rnctric outer measure'! first being a covering of A and the second being a covenng B. lt follows that
MEASURES
Taking n _, ro, and then c-> 0, we get
Let f be a sequential covering class with 0 E f, and let p~(A) + p~(B) .:-::; JI~(A v B).
f.={AEf;d(A):::::;~} [we define d(0) =OJ Since the reverse inequality is obvious, the proof is complete.
for any positive integer n. Then .ff.+ 1 c .ff. Assume that each f . is a
Theorem 1.9.2 Let the assumption on .ff,"A made in Theorem 1.9.1
sequential covering dass. Let A. be a nonnegative, extended real-valued set
hold. Assume further that for each set A E f , for any E > 0, and for any
function defined on f , with }.(0) = 0, and denote by J-1: the outer measure
positive integer n, there exists a sequence {Ex} in ..ff n such that
defined by each pair fn,A. (according to Theorem 1.4.1). Clearly
J-l:(A) s J-1:+ 1 (A)
for any set A. Let
,--+.:o
(1.9.1) L"' ..l(11 )::::::; i.(A) +e.
11=1
Theorem 1.9.1 Let the foregoing assumptions on f,). hold. Then the
Then the outer measure ~* defined by ( 1.4.1) coincides with the metric outer
set function )1~, defined by ( 1.9.1 ), is a metric outer measure.
measure constructed by (1.9.1 ).
Proof The conditions (il, (ii), (iv), and (v) in Definition 1.3.1 hold for
each p:.Taking n-> w, we sec that they hold also for fJ~. To prove (iii), Proof. Since f => f n we have
we note that p:(A) s ~~~(A). Hence, \1.9.2)
for any set A and n = I ,2, ....
P:(Ql .) .:-: ; ltl f.I:(E.)::::::; Jl .uUE.). We shaH now prove the reverse inequality. From the
d fi . . f * it
e mtwn o P
22 MEASURE THEORY CONSTRUCTION OF METRIC OUTER MEASURES 23
follows that for each set A and r. > 0, there exists a sequence {Ei} in :;(" E. with diameter < 1fm, such that L, ..1.( Ei) < e. J can be covered by a finite
such that n~mber of (n - 1)-dimensional intervals Ej lying on xi = a; + r ,b; and having
diameter < 1/2m. Let C =I
I.(Ej). Take '1 < t/2C, '1 < lf4m, and let
Ei = {x; (x1 , ... ,x,_ 1 , a,+ y/J;, x,+ 1 , .. ,x.) E Ej, a,+ y,b,- '1 < x, < a1
f
k~l
A.( Bid :o;; ..i.(Ej) + .: I.
~
c
Since UBil< covers A, we have 1.( I~ ~.) $ ..1.(1 Ck J + -;:- (C independent of r).
j.k
[Hint: Let c" = (cuc""), ek = (e, 1, ... ,e... ). Let o:ki be the largest number
of the form a,+ y;b, that is less than c.,. Let {l.; be the smallest number of the
form a~., +ri'b; that is larger than eki. Then lck ... c 1,.. ~.-Also
Taking e-> 0, we get tt:(A):::;: tt(A). Combining this with (1.9.2), we get
= tt*(A). Hence J.i~(A) = 1-1*(A).
tt:(A) (C' constant).
PROBLEMS
Now use Problem 1.9.2.]
1.9.1. Let a,< a1 + .5,. < ai + 23i < <a,+ r,o, = b;for i = I,2, ... ,n. I .9.5. Let i . b be the closure of a bounded interval la,b in R". Show
Denote by I . b.y the open interval that
a,+(!';- l)b; < x; <a, + y,b, (I :::;: i:::;: n)
when the l'; vary from 1 to 'i Let ..1. be defmed by ( 1.6.2). Prove that
[Hint: ttU .):::;: p.*(l.-,,0+.) for any > 0. Use Problem 1.9.2 to deduce
..1.(/d) = I A.U.,b,)-
&
y, integers] either lies entirely in I~ . p, or does not intersect I.,,p, By Problem (compare Problem 1.5.2) the Lebesgue measure is the completion of its
1.9.1, ),(!~ !1.) = I(J:J I
..1.(/.,b.y), A.U. b) = ..1.(/.,b,y)- Since restriction to the Borel sets.
We conclude this section with a few results that we shall not prove.
(They will not be used in the sequeL) .
It is known that not every Lebesgue set 1s a Borel set. In fact, the two
it follows that every term J.Ua,b.,) that occurs in the last sum must also occur
in at least one of the sums I (kl_ This gives classes of sets have different cardinal numbers. _
Denote by f-t* the Lebesgue outer measure o_n R". Then for_any bounde~
,. m m m
interval Ia.b the following is true: a sel E c Ia,b IS measurable 1f and only 1f
TI (b;- a;):=::; I A.U. ,.> :=::; I A.Un.~.) + t: = I l(Ed + e.J
i~J J:-1 J:~J k~l
/--1-(la,b) = /--1-*(E) + /--1-*(la,b - ),
1.9.6. Let Ia,b be an open bounded interval in R". Show that
n The inner measure Jl* is defined by
p(l.,b) = p*(l.,b) = !--1-0a,b) = ..i.(I.,b) = D (b;- a;).
i=l 11 *(E) =sup (Jl(F); F c E, F Lebesgue-measurable}.
1.9.7. If a set F in R" is Lebesgue-measurable and if p(F) < oo, then It is known that
for any t: > 0 there exists an open set E such that E ::;, F and Jl(E) < Jl(F) + e.
1.9 .8. If a set F in R" is Lebesgue-measurable, then there exists a Borel Jl,(E) = ftUa.b) - ;t*Ua,o- E)
set E such that E => F and Jl(E- F)= 0. [Hint: Consider first the case where for any subset E of Ia,b Consequently, E is measurable iF and only if
p(F) < 00.1 Jl*(E) = II*(E).
1.9.9. A straight line in R" (n ~ 2) has Lebesgue measure zero. More
generally, a k-dimensional plane in R" (n 2 k + 1) has Lebesgue measure zero.
1.9.10. The boundary of a ball in R" has Lebesgue measure zero.
1.9.11. Consider the transfonnation Tx = Ax+ kin R", where A is a 1.10 SIGNED MEASURES
nonsingular n x n matrix and x, k are column n-vectors. T maps sets E onto
sets T(E). Assume that ).[T(l.,b)] = ]det A ]A.(l.,b)- (This will be proved in Definition An ex.tended real-valued set function 11 is ca\led a si~ne_d
Problems 2. 16.7 and 2.16.8.) Prove that T satisfies the properties (a)-{ c) measure if it assumes at most one of the value~ + oo and - oo, and tf 1t
of Problem I. 6. 5, with I:t:] replaced by Idet A 1- satisfies the conditions (i}, (iii), (iv) of Definition 1.2. I.
1.9.12. Extract from the interval (0,1) the middle third-that is. the The motivation for this notion comes by looking at the difference
interval I 1 = (-i,t). Next extract from the two remaining intervals, (0,-j-) and ll = p 1 - p 2 of two measures p 1 ,p 2 , one of which is finite, defined by
(1,1), the middle thirds I 1 and 13 [12 = {i-,-t), I 3 = (-i-.t)J, and so on. Let (E E <i), (1.10.1)
I= I 1 u I 2 u 13 u ... , The set C = [0,1]- I is called Cantor's set. Prove
that the Lebesgue measure of Cis zero. (It can be shown that Cis not count- where fit , 112 have the same domain a.
It is clear that J1 is a signed measure.
able; in fact, it has the cardinal number of the continuum.) More generally, any linear combination (with real coefficients) of measures
1.9.13. The Lebcsgue-Stieltjes outer measure 1-11 , induced by a mono- having the same domain, defined in the obvious way (analogously to 1.10.1)
tone function f (continuous on the right), satisfies is a signed measure, provided all but one of the measures are lin ite. .
fl/a,h] = f(b)- f(a). 1n this sect ion we shall prove that every signed measure is neccssanly
of the form (1. 10.1 }.
1. 9 .14. The Lebcsgue-Stiel tjes outer measure is a metric outer measure. Let fl be a signed measure with domain a.
A measurable set E is called
1.9.15. Let f(x) = 0 if x < 0, f(x) = 1 if x 2 0. Prove that positive (negatire) with respect top if f-t(E n A) 2 0 ( :S 0) for any measurable
J11 {(- I ,0)} </(0) -/(-I), set A.
As we have seen, every Borel set in R" is a Lebesgue (measurable) set. Theorem 1.10.1 Ij)J is a signed measure, then there exist two measurable
From Problem I .9.8 it also follows that every Lebesgue set has the form sets A and B such that A is positire and B is negative with respect to 1-1-. and
E - N, where E is a Borel set and N is a null set [that is, Jl(N) = 0]. Thus X= A u B, A n B = 0-
SIGNED MEASURES 27
26 MEASURE THEORY
To prove Theorem 1.10.1 we shall assume that By Lemma 1.10.2, J.I.CQ. ,) < oo. Since the sets Ek are mutually disjoint,
-oo < p(E)::::; oo for all E E ct. (1.10.1) it follows that JJ.(E,)-> 0 as k -> oo. Hence
The case where - oo ::::; I'( E) < oo for all E E (i can be treated by looking . 1
at -J.!. l!m- = 0.
Observe that the difference of two negative sets is again a negative set
-"" m,
00
and that the union of two negative sets is a negative set. Since also the count- It follows that every measurable subset F of Fo = Eo - kl),E satisfies
able disjoint union of negative sets is a negative set, we conclude (by Theorem
1.1.1) that any countable union of negative sets is a negative set. 1
Define J.l.(F)::::; - - - + 0 if k-+ 00
mk- I
P= inf J.!(B 0 ), B 0 varies over all the negative sets.
-that is, J.!(F) = 0. Thus F 0 i~ a negative set. Since
Then there is a sequence {Bj} of negative sets such that J.!(Bi)-+ p. The set 00
B
00
= UBj is then a negative set and, therefore, J.!(B) .:-:;; p(Bi). It follows that J.I.(F0 ) = JJ.(E 0 ) - L
k; I
p(Ek) < p( 0 ) < 0,
j;\
we get II(B u F 0 ) = p(B) + p(F0 ) < 11(B). Since 11(8 u F0 ) ~ p (for B U Fo is
p = p(B). (1.10.2)
a negative set), we get a contradiction to ( 1. 10.2).
Note, by (1.10.1), that p > - oo. Set
We shall show that the set A =X - B is positive; this will complete (l.\0.4)
the proof of the theorem. p +(E)= 11( n A),
If A is not positive, then there is a measurable set E0 c:: A such that Then p + and p- are measures, and at least one. ~f them is finite. They are
ll(E 0 ) < 0. E 0 cannot be a negative set, for otherwise B u E 0 is a negative independent of the particular Hahn decompositiOn (see .Problem 1.10.4).
set and thus Also, p() = p +(E)- 11 -(E). We can now state the followmg result, known
p::::; J.!(B u E 0) = J.!(B) + p(E0 ) = p + ,u(Eo) < fi. as the Jordan decomposition of a signed measure.
Therefore E0 contains a set 1 with J.!(1 ) > 0. Let m 1 be the smallest positive Theorem 1.10.3 Let ].1 be a signed measure. Then at least one of the
integer for which such a set E 1 exists with JJ.(E 1) ~ 1 jm 1 By Lemma 1.10.2, measures ll + ,l..l-, given by ( 1.1 0.4), is finite, and
J.!(Ed < oo. Since
p = J.1. + - J.!- . ( 1.1 0.5)
. !he measures 1-l + ~- given by (1.10.4) are called the upper and lower
CHAPTER 2
vanotwns of J.l (re~p~ctJvely). The measure Lui given by II' I = J.l + + J.l- is
c~lled the total vanatwn of 1-l Note that II' I( E) = J.l +(E) + J.l-(E) is genera!!
different from Lu(E)I = Lu +(E) - ~~-(E) I.
A signed measure is said to be .finite (a-finite) if 11-11 is finite (o--fmite).
y
INTEGRATION
PROBLEMS
.lJ/ and
00
r- 1
{(a,b)} =/- 1 {(-oo,b)} -r 1
{(-co,a]}, Proof Consider the class Q! of all the sets Eon the real line for which
f - 1(E) is measurable. It is easy to verify that !1! is a a-ring. Si nee 9 contains
we see that /- 1 {(a,h)} is measurable. Since every open set M is a countable all the open sets, it also contains all the Borel sets.
union of open intcrvals,/- 1 (M) is also a measurable set. This, together with
(ii), shows that f is measurable. PROBLEMS
2.1.6. The characteristic function of a set E is the function XE defined
PROBLEMS
by
2.1.1. Let (X,CL,Jl) be a measure space and let Ybe a measurable set. 1 ifx EE,
XE(X) = { ' if X j: E.
Denote by eLy the class of all the measurable sets that are subsets of Y and 0,
denote by Pr the restriction of p. to eLy. Prove that (Y,CLy,Jiy) is a me;sure
Prove that the set E is measurable if and only if the function XE is measurable.
space. It is called a measure subspace of(X,CL,p).
2.1. 7. Let {EJ be a sequence of sets and let E* = fiffi E. Prove that
2.1.2. Let Z be a subset of a space X, and let {Z,al,v) be a measure n
space. Denote by CL the class of all sets E in X such that En Z is in fM and
define a set function 11 by p(E) = v(E n Z). Then (X,CL,11) is a measure ~pace
and (Z,.:.l1,r) coincides with (Z,dz,Jlz). 2 .1.8. The positive part f + of a function f is defined by
2.1. 3. A function I defined on a measurable set Y of a measure space
f+(x) = {.f(x), if f(x) > 0,
(X,CL,p) is measurable if and only if it is measurable as a function on Y if f(x)::::;; 0.
0,
considered as a set in the measure space (Y,dy,l/r).
Similarly, the negathe part f- off is defined by
Let f! be a Lebesgue-measurable set in R". The measure subspace of
the Lebesgue space R" corresponding to f! is 1alled the Lebesgue measure rex)= {0,- f(x), ~f f(x) > 0,
If f(x) ::::;; 0.
space of f!. By Problem 2.1.3, f is measurable on the measure space f! if
and only iff is Lebesgue-measurable on f! (as a subset of the Lebesgue Prove thatfis measurable if and only iff+ and/- are measurable.
space R"). 2.1.9. lfjis measurable, then 1/1 and IJY arc measurable.
2.1.10. A monotone function defined on the real line is Lebesgue-
2.1.4. Prove that a function/is measurable if and only if: (i) the sets measurable .
.r 1{(- co ,c]} are me as ura ble for all real numbers c, and (ii) the set sf- 1( + GO), 2.1.11. A function f(X), defined in a metric space X, is said to he
f- 1 ( - cr;) are measurable. upper-semicontinuous (lower-semicontinuous) if for any point x EX,
2.1.5. Prove that a function f is measurable if and only if (i) the sets
1 1
/ - {(c,oo)} (or the setsf- {[c,GO)}) are measurable for any real number c, rrm f(y)::::;; f(x) [ lim f(y) :2= f(x)].
r
and (ii) the set sf - 1( +X ).. 1( - GO) are measurable. ,... " ,... "
OPERA T!ONS ON MEASURABLE FUNCTIONS 33
32 INTEGRATION
In this section and in the following ones, X is a fixed measure space. Proof The assertion for sup/~ follows from Problem 2.1.4 and the
Furthermore, unless the contrary is explicitly stated, all measurable functions
are defined on the whole space and are extended real-valued functions. relations
E = {x; f(x) < g(x)} = U [{x; f(x) < r.} n {x; g(x) > r.}]. The assertion for inff. follows from inf f~ = -sup (-/.). Now use the
'" relations
Since each set on the right is measurable, E is also measurable.
When we consider the sum f + g of two measurable functions f and g, lim fn = inf sup fn,
k P1:::.k
we always tacitly assume that at no point x, f(x) = oo, g(x) = - oo or
f(x) = - oo, g(x) = oo. Similarly, when we consider the productfg, we always lim fn = sup inf f .
l il1 ~k
assume that at no point x, f(x) = 0, g(x) = oo or f(x) = oo, g(x) = 0.
n
Since the set on the right is measurable (by Lemma 2.2.1), also E is measur- Definition A sequence {/") of functions is said to be (pointwise)
able. Next, the sets (f + g)- 1(oo) = f- 1 (oo) v g- 1(oo) and(/+ g)- 1 ( - 00 ) = convergent almost everywhere (or ererywhere) if there exists a function g such
f -J ( - oo) u g - 1 ( - oo) are also measurable. Now apply Theorem 2.1.1 to that
conclude that/+ g is measurable. Similarly one proves that f- g is measur-
able. lim f.(x) = g(x) a.e. (or everywhere).
The measurability of fg follows from the measurability of If+ gj 2
and If- gj 2 (compare Problem 2.1.9), and the relation We then say that {fn} conrerges (pointlrise) almost ererywhere (or everywhere)
tog, and write limfn = g a.e.
fg =i {(f + g)2 - (f- g)2}. n
EGOROFF'S THEOREM 35
34 INTEGRATION
f(x) = (CI:;,
if x e E, (i = 1,2, ... ,rn), g(x) = {f~x)' if f(x) #- 0,
0, if f(x) = 0.
u
m
0, if X t 1,
j;::l
Prove that g is measurable.
We can write 2.2.4. The nth Baire class :11. is a class of functions/ defined as follows:
fe :11. if and only iff is the pointwise limit everywhere of a sequence of
f(x) = L'"
j::::o:J
Cl.;X,(x). functions belonging to :11. _ 1 , and 81 0 is the class of all continuous functions.
Prove that all the functions of :11. are measurable.
2.2.5. Show that the characteristic function of the set of rational
. Theorem 2,2,5 Let f be a nonnegative measurable function. Then there
numbers in R1 belongs to J4 2 .
ex1sts a monotone-increasing sequence {f"} of simple nonnegatit'e functions,
2.2.6. Letfbe a measurable function, and define
such that {fn} comerges to f everywhere.
{~:
if f(x) is rational,
Proof. For each integer n ;:::: 1 and for each x eX, let y(x) = if f(x) is irrationaL
At a point x where f(x) = oo, /,(x) = n. This shows that, for any point
if n > f(x).
A measurable function f is said to be a.e. real-ralued if the set
{x; lf(x)l = oo} has measure zero.
I
i
L
x, j~(x)--> f(x) as n--> oo. (
Dejinitian A sequence {/"} of a.e. real-valued, measurable functions,
PROBLEMS in a measure space (X ,<l,J.I), is said to conrerge almost uniformly to a measur-
able function f if for any t: > 0 there exists a measurable set E such that
2.2.1. If g is an extended real-valued Bord-measurable function on the
J.I(E) < e and {f"} converges to f uniformly on X - E. Note that f is neces-
real line and iff !s a real-.valued measurable function on a measure space x,
sarily a.e. real-valued.
then the composite functwn h(x) = g[f(x)] is a measurable function on X.
CONVERGENCE IN MEASURE 37
,,
36 INTEGRATION
1
I
Theorem 2.3.1 If a sequence {f"} of a.e. real-valued, measurable PROBLEMS 'i
:I
functions conrerges almost uniformly to a measurable function f, then {f"}
23.1. Let X be the Lebesgue measure space on the real line and let
comerges to f a.e.
f. be the characteristic function of (n,oo ). Show that Un} is convergent to
Proof For every positive integer m there is a measurable set ,such zero a.e., but not almost uniformly.
that p(,) < 1/m and {/,} converges uniformly to f on ':..=X-,. The 23.2. Let {f.} be a sequence of measurable fum:tions in a finite
sequence {/,} is then convergent to f on the set measure space X. Suppose that for almost every x, {f,(x)} is a bounded set.
Then for any e > 0 there exist a positive number c and a measurable set E
F = mQl E~= U]l E,r with p(X - E) < 6, such that 1/,(x)l :<:;; c for all x E E, n = 1,2, ....
Since p(fc) = I' CO/) :<:;; p(Em) < I /m for any positive integer m, f.l.(f<) = 0.
2.4 CONVERGENCE IN MEASURE
Hence {f.} converges to f a.e.
If X is a finite measure space, then the following converse of Theorem Definition A sequence {f.} of a.e. real-valued, measurable functions
2.3.1 is also true, and is known as Egoroff's theorem. is said to be convergent in measure if there is a measurable function f such
that, for any e > 0,
Theorem 2.3.2 Let X be a finite measure space. If a sequence {f.}
lim f.l.[{x; lfnCx)- f(x)l ~ e}] = 0. (2.4.1)
of a. e. real-mlued, measurable functions converges a.e. to an a.e. real-l'alued,
measurable function f, then {f.} converges to f almost uniformly.
We then say that {f.} converges in measure tof
Proof We may assume that f and the f, are real-valued everywhere.
For any two positive integers k and n, let It is easily seen that if {f.} converges in measure to two functions,
f and g, then f = g a.e. Also, if {f.} converges to fin measure, then f is a.e.
E~ = ,{J. {x; lfm(x)- f(x)l < ~} real-valued.
Hence, for any E > 0 there is an integer nk such that Proof For any e > 0, b > 0, there is a set E with p(E) < b such that
\f.(x) - f(x)l < 6 for all x EX - E and n sufficiently large (independently
f.I.(X -E.)<
k
2iie of x). But this immediately implies convergence in. measure.
Let F =
k~
n"' E~. This is a measurable set, and
I
Corollary 2.4.2 If Jl(X) < oo, then any sequence {fn} of a.e. real-
valued, measurable functions that converges a.e. to an a.e. real-valued, measure-
The sequence {fn} is uniformly convergent, on F, to f In fact, this follows This fo!lows from Theorem 2.4. t and Egoroff's theorem.
from the inequalities
Definition A sequence {fn} of a.c. real-valued, measurable functions
1
lf.(x) - f(x)l < k if x E F c ~.. n ~ n,. is called a Cauchy sequence in measure if, for any 6 > 0,
f.l.[ {x; lf.(x)- };,(x)l ~ e}] ..... 0 as m,n _, oo.
The proof is thereby completed.
38 INTEGRATION CONVERGENCE IN MEASURE
if X EX - Fm )~ h ~ m.
I If [Hint: Use Problem 2.4.1.] ~
2.4.3. Prove the following result (which immediately yields another 1
1
proof of Corollary 2.4.2): Let f.. (n = 1,2, ... ) andfbc a.e. real-valued measur- I
Thus, on each set X - F"" {/,,} satisfies tL: Cauchy criterion for uniform
convergence. It follows that there exists a function f defined on able functions in a finite measure space. For any e > 0, n ~ l, let ''
E,(e) = {x; lf,(x)- f(x)l ~ e}.
"'
A= U (X- F'")
m=.:.l Then {f..} converges a.e. to f if and only if
such that {f.;} con verges to f uniformly on every set X - F.,. By the first
part of Corollary 2.2.4, fis measurable on A. Definef(x) = 0 for x EX- A. !~~ ~L.Q. Em(e)] = 0 for any e > 0. (2.4.2)
Xt!.[is also an integrable simple function. We define the integral off over E by
2.4.5. Definefunetionsf~(x) on [0,1) byf~(x) = 1 ifxE [(m- 1)/n, m/n)
and j:,(x) = 0 otherwise. Enumerate the/:, in any way, and call the sequence (2.5.2)
thus obtained {op1}. For instance, op 1 =fi, op 2 = fi, op 3 =Jf, op4 =f~, ....
Prove that {op1} converges in measure to zero, but lim op /x) docs not exist
In particular we have
for any XE [0,1).
2.4.6. If {fn} is a sequence of a.e. real-valued measurable functions
in X, and if t~(X) < oo, then there exist positive constants }." such that
{f,/A,} converges a.e. to zero. [Hint: lf..(x)l s const. on a set En with if Eisa measurable set with finite measure.
p.(X- E") < rM.J The following theorem contains some basic results on integrals. The
proofs are all rather obvious and are left for the reader.
2.5 INTEGRALS OF SIMPLE FUNCTIONS Theorem 2.5.1 Let f and g be integrable simple functions, and let a, {1
be real numbers. Then:
M
A simple function!= L a, XE, on a measure space (X,<i,p) is said to (a) crf + ~g is an integrable simple function, and
!=I
be integrable if ,u(E;) < w for all the indices i for which
M
a; f= 0. The integral
J(af + f1g) dp. =a Jf dt~ + fJ J9 dJ-1.
ofjis the number L a;!-l( 1), where we agree to take et. 1 J1(E 1) = 0 for all the
i=I
(b) lf f:::::. 0 a. e., then f f d!l :::::. 0.
indices i for which Ct.; = 0, Jl(E;) = oo. We denote this sum also by f(x) dp(x) J J J
(c) Jf f : : :_ g a.e., then f dj.l :::::. g dj.l.
J
or f d!J. Thus,
(d) If I is an integrable simple function, and
~ "' M
= L a, i-1
i-l
L Jl(E; n F,) = L Ct.;J1(E;).
i-1
Definition A sequence {f.} of integrable simple functions is said to be
a Cauchy sequence in the mean if
Consequently, the definition (2.5.1) of the integral is independent of the
representation of j, and is thus unambiguously defined.
If E is any measurable set andjis an integrable simple function, then
f IJn - /,.1 dt~-> Q as n,m ...... oo. (2.5.3)
42 INTEGRATION
DEFINITION OF THE INTEGRAL 43
Lemma 2.5.2 If {f"} is a sequence of integrable simple functions that function g. Theorems 2.4.3 and 2.3.1 imply that there is a subsequence
is Cauchy in the mean, then there is an a.e. real-ralued, measurable function f U~J of UnJ that converges a.e ..to g. Since (~) holds, g =fa. e. Hence {f"}
such that {fn} converges in measure to f.
converges in measure to f-that IS, (a) and (b ) hold. . ..
Suppose conversely that f is integrable accordmg to the defim~wn
Proof For any c > 0, let (a), (b'). That means that there is a sequence {g"} ofintegrable_simplc functiOns
E"m = {x; lf,(x)- /,.(x)l :2:: t}. such that {g"} is Cauchy in the mean and {g"} converges 1n measure to f
By Theorems 2.4.3, 2.3.1 there is a subsequence {.qn.} of {g"} .that is convergent
Since.f" -/,.,is integrable, ""' has finite measure. From Theorem 2.5.1 we to f a.e. Denote this subsequence by {fd. ~~en {f~} satisfies (a) and (b).
then get
Thusfis integrable according to the first defimtwn [w1th (a), (b)].
From (b') we conclude: Iff is integrable, thenfis a. e. real-valued.
Definition 2.6.2 Let f be an integrable function and let _(a), (b) hold.
It follows that J.I.(Enm)....,. 0 if n,m ..... oo. Thus UnJ is a Cauchy sequence in The integral off is defined to be the number ~J f" dJ-1., and IS denoted by
measure. The assertion of the lemma now follows from Corollary 2.4.4.
Jf(x) d/l(x) or Jf djJ. Thus
PROBLEMS
2.5.1.Iff and g are integrable simple functions, then fg is also an Jf(x) d{J(X) = Jf dJ-1. = ~~~ Jfn d{J. (2.6.1)
integrable simple function.
2.5.2. An integrable simple functionfis equal a.e. to zero if and only Note that
if JEf d{J = 0 for any measurable set E.
2.5.3. If/is a nonnegative integrable simple function and if ffdf.l = 0,
thenf = 0 a. e. ifn,m....,. oo,
2.6 DEFINITION OF THE INTEGRAL so that the limit on the right-hand side of (2.6.1) indeed exists. .,
,;
Theorem 2.6.1 A function f is integrable according to the first definition for any measurable set E. Then !.(E) Lv a complete/}' additire set function.
[~tith (a), (b)] if and only if it is inte_qrable according to the second definition
[with (a), (b')]. Proof The limit defining ).(E) exists uniformly with respect to E.
Indeed, this follows from
Proof Suppose f is integrable according to the first definition. By
Lemma 2.5.2, (f.} converges in measure to an a.e. real-valued, measurable It fn dJ-1. - t fm d{JI =:;; t lfn - f.,l djt =:;; JIf. - fml d{J (2.6.2)
I
,;
;I
-i"
44 INTEGRA TI0:-.1
DEFINITION OF THE INTEGRAL 45
nd (a). Note also that J.(El is finitely additive. Now let E be a disjoint union
The first integral on the right is less than sp(). The second integral is
"' E, of measurable sets E;.
U Then estimated by
i:-= 1
+ IJ ~ i = 1
E; f. dp- ;{01 E;)l
.
if n 0 is sufficiently large and n ;::; n 0 . With n 0 fixed, we have lfnol :s; c a.e.,
For any e > 0, the first and third terms on the right are less than c,/2 for all where c is a finite number. Since p(.)--> 0 as n--> oo, we can choose n 1
n ;::; n0 , where 11 0 is sufficiently large, depending on c:, but not on m. Hence, sufficiently large so that
I itl I + It
i.()- J.(E;) S e ho dp - itl t; fno dp I J EnEJ1.
lfnol dj.l :s; cp(E n E.) < e
Taking m ....... cJ::; and using Theorem 2.5.1 (h), we get We conclude that
Since e is arbitrary, the assertion follows. The third integral on the right-hand side of (2.6.3) is estimated in the same
way.
Lemma 2.6.4 Let {fn} and {gn} be two Cauchy sequences in the mean Combining the estimates for the integrals on the right-hand side of
of integrable simple functions, such that (2.6.3), we get
I.( E) = lim I' f. dp, v(E) = lim r g. dp. II.() - v()1 s lmi I If. - g.l dp s t:J.l(E) + 4e.
~t---+oc... .f .
n--'~- '::1." L' E it.l-::t:! ... E
Since e is arbitrary, .l.(E) = v() provided p(E) < oo.
!fE is a-finite, then ).(E)= v(E).
Consider now the case where E is cr-finite. Then we can write E as
Proof Consider first the case where p(E) < w. By the proof of a countable union "" Ei
U of sets with finite measure. Taking F 1 = E 1 ,
Theorem 2.6.1 we have that Unl and {g.} converge in measure to f For
any e > 0, F1 = Ei- [v:E,J, j~J
E.= {x; lf,,(x)- g.(x)l;::; c} c: {x; lf"(x)- f(x)j;::; ~} are mutually disjoint and have a finite measure. By what we have already
proved, ).(F)= I'(F) for all j. Since }. and v are completely additive (by
Lemma 2.6.3), it follows that .l.() = v(E).
U{x; lg.(x)- f(x)l ~ ~). We shall now complete the proof of Theorem 2.6.2. We introduce the
sets
Hence p(E.)--> 0 if n--> oc-. Next,
N(f.) = {x; fn(x) #- 0}, N(g.) = {x; g.(x) #- 0},
r If. -
.. E
9.1 dp s 1
"'E- 11
If.- g"l dp + r
.... Er.Et\
If. I dp + r
.. E,.En
IY"I dp. (2.6.3)
N = [.QI N(fn)J v LQI N(g")J .
DEFINITION OF THE INTEGRAL 47
46 INTEGRATION
Iff and g are integrable functions, then they are a.e. real-valued. Hence
Since each of the sets N(f"), N(q") has a finite measure, N is a-finite. Lemma
2.6.4 then gives
f +g and f" g are defined a.e. In what follows we shall think off+ g and
f" g as everywhere defined measurable functions. All we have to do is define
lim
,.._a."::!
J In dJ1
N
= lim
n--+a."::!
J Bn dJ1.
!V
(2.6.4) f(x) + g(x) [or f(x) g(x)] on the set
E= [/- 1 (::D) n g- 1 ( - co)] u [/- 1 ( -oo) n g- 1 (co)]
But, clearly.fn = x"f.., gn = XNgn everywhere on X. Hence,
(or E = [f- 1(0) n g- 1 {co,-CQ}] n [/- 1 {oo,-co} n g- 1(0)]) such that
the resulting function f + g (or f"g) is measurable. This can be done, for
instance, by setting/+ g = 0 (or fg = 0) on E. Theorems concerning con-
Combining this with (2.6.4), we get vergence a.c., almost uniformly, and in measure are obviously unaffected by
the particular definition off+ g (or f"g) on E. Also, by Problem 2.6.1, the
lim
n__,.~
JIn dJ1 = lim JBn dp.
,._,.oo
integrals J (f + g) dp., J (/g) dp. are independent of the definition off+ g,
f"g on E. In fact, none of the results given in this book will depend on the
This completes the proof of Theorem 2. 6. 2. particular definition used for f + g, f" g on the sets E.
Jf
E
dg or Jl(x) dg(x).
E
and lim (()m =fa. e. [Hint: Prove it first when f is a simple function using
"' 1.9.7.]
Problem
We finally give the following useful definition. 2.6.6. letfbe a Lebesgue-integrable function over a bounded interval
(a,b). Prove that for any e > 0, i5 > 0, there exists a continuous function g
Definition 2.6.6 If f is a nonnegative measurable function (on a in [a,b] such that lg(x) - f(x)l < e for ull x in a subset E of (a,b), and
measurable set E) and if/ is not integrable (on E), then we say that the J.l[(a,b) - E] < <l.
integral off (over E) is equal to infinily, and write 2.6.7. A complex-valued function f is said to be simple if it has the
n
form I c; XE , where E, are measurable sets and ci are complex numbers.
J~l '
48 INTEGRATION ELEMENTARY PROPERTIES OF INTEGRALS 49
f is then said to be integrable if 11( 1} < oo for any i for which c; -# 0. The in the mean and that satisfies: lim f. = f a.e. Then {II. I} is also a Cauchy
n
integral off is defined by .L C; J.l(;). Now let I
~~~
be any measurable complex- sequence in the mean, and lim lf.l = fa.e. Hence
n
integrable function f), then it is also Cauchy in the mean. The converse is
Ce) I If+ gl ctli .:=::; I If I ct11 + I lgl d11. also true and will be proved in Section 2.8.
(f) lfm $ f $ M a.e. on a measurable set E, and 1/l-t(E) < oo, then
Theorem 2.7.2 II {f"} is a sequence of integrable functions that con-
mp.(E) .:=::; JI
.
d11 $ M p.(E). verges in the mean to an integrable function f, then {f.,} conrerges in measure
to f.
(g) Iff:::-: 0 a.e. and if E and F are two measurable sets with E c:: F,
then Proof For any c > 0, let
Proof. Let En = { x E E ;f(x) C. 1./n}. Then {E.} is an increasing We begin with a lemma that is concerned with Definition 2.6.1. ~
'< :J
seq lienee of sets and E - U En has measure zero. Hence It( E) = lim p(E.).
Lemma 2.8.I Jff is an integrable function and if {fn} is a sequence of
i
n=1
0= _I
.
f dp ;::: t" j
dJt ;:::
1
ni-!(En);::: 0, lim
ii'J--Ii'~;
!11- J.l dfl =0.
...
;
j
~
Theorem 2.8.2 If {f"} is a sequence of integrable functions satisfying Theorem 2.8.4 Let f be an integrable function and let l be the set
the conditions (a),(b) [or (a),(b')] in Definition 2.6.1, then f is integrable and Junction defined by
J J
f d).t = lim f" d1-1.
J.(E) = f dp J (2.8.1)
E
Proof Suppose that (a) and (b') hold. By Lemma 2.8.1, for each n for all the measurable sets E. Then '"A is completely additive and absolutely
there is an integrable simple function!, such that continuous.
lim ( lfn - /I dp = 0.
We conclude this section with a few definitions that will be needed I
"_,.a:: ..J
in Section 2.14.
A function g(x) defined on a real interval (a,b) is said to be absolutely
continuous if for any e > 0 there is a b > 0 such that for any sequence of
Definition 2.8.1 A real-valued set function l is said to be absolutely
continuous if for any <. > 0 there exists a number b > 0 such that, for any
mutually disjoint intervals (a 1,b;) in (a,b), with (b 1 - a;) < b, L
measurable set E with JI(E) < b, II.(E)I < ~- I lg(bJ - g(a1)1 < e.
54 INTEGRATION
LEBESGUE'S BOUNDED CONVERGEXCE THEOREM 55
A functionf(x) on a real interval [a,b] is said to be of bounded mriation converges either in measure or a.e. to a measurable function f. If lfn(x)l :'S: g(x)
if there is a constant K such that for any partition a = x 0 < x 1 < < x._ 1 < a.e.for all n, where g is an integrable function, then f is integrable and
x. = b of (a,b),
L lf(x,)- f(x,_,)J :<S: K. lim
n-~
fIf. - fl d11 = 0.
...
(2.9.1)
The g. I. b. of all the constants K is called the total rariation off in (a,b).
Proof Consider first the case where {/.} converges to fin measure.
PROBLEMS We shall prove that {/.} is a Cauchy sequence in the mean. Let
2.8.1. A measurable function f is called a null jimction iff= 0 a.e. 00
We shall say thatfis equivalent tog (and write/ ..... g) iff- g is a null function. E = U {x; f.(x) # 0}.
Denote by J the class of all measurable functions that arc equivalent to f
n~ I
00
We denote by L1(X,Ci,J1), or, more briefly, by L1(X,).l), the set of a!! classes J
for which f is integrable, and define on it the function
Then, by Problem 2.6.2, E is a-finite. We can therefore write E UEk,
=k=!
where E~ c EH 1 and 11(Ek) < oo, for all k. Let Fk E- Ek. Then
p(J,g) = p(f,g) = JIf- ol dJ1. =
[Note that if fo Ef, g 0 E iJ, then p([0 ,g 0 ) = p(.{,g).] Prove that L1(X,J1) is a f
F,
lfm- !.1 d11 ~
Fo
I (1!... 1+ lf.D d11 :<S: 2 J h
g dJ.l.
complete metric space with the metric p.
2.8.2. For any Lebesgue-integrable function g(y) on a bounded Noting that Fk ~ Fk+ 1 , lim Fk = 0, we can apply Corollary 2.8.5 and con-
interval (a,b), the function clude that, for any '1 > 0,
f(x) = J
(o,x)
g(J') dy (2.9.2)
is absolutely continuous.
where k 0 is a large positive integer that depends on '1 but not on m,n.
2.8.3. A function f(x) is of bounded variation in a bounded interval
For any e > 0, introduce the sets
(a,b) if and only if it is the difference g 1 (x) - gix) of monotone-increasing
functions g 1 ,g2 If f(x) is also continuous, then g 1(x) and gix) can be taken G... ~= {x; lf...(x)- f"(x)l ~ e}.
to be continuous monotone-increasing functions. [Hint: Take g 1 (x) to be the
total variation off in the interval (a,x).] We have
2.8.4. If f(x) is absolutely continuous, then it is of bounded variation.
2.8.5. Prove that the function .f(x) = x sin (1 ;'x), for 0 < x <I, is fE~t
If.., - f.j d/1 = f Ek- Gm,.
lfm - f"j d)1 + f EknG;r11,"
If,. - fnl dJ.l
uniformly continuous but not of bounded variation. (2.9.3)
2.8.6. Iff is not integrable, then the absolute continuity of ..1.() (in
Theorem 2.8.4) can still be proved, provided fis integrable on sets E of finite
:<S: eJ.l(Et) +2 f EhnG'"'J'l:
g d11.
measure (see Corollary 2.12.3). However, the same extension does not hold Since {f,.} converges to fin measure, /1(Gm.)-> 0 if m,n-> oo. Theorem 2.8.4
for the property of complete additivity. Give an example.
then gives
THEOREM
where n0 is a large positive integer depending only on e and g. Combining
One of the most important theorems in the theory of integration is the (2.9.2)-(2.9.4), we get
following one, known as Lebesgue's bounded convergence theorem.
if k > n.
Urn
m~"-to-ro
f lfm - fnl dp. $ '1 Then {f..} is a sequence of integrable functions that converges uniformly to
f f f
=:= 0, but lim f.. dp. =1- f dp.. This example again shows that the bounded-
Since 11 is arbitrary, we find that Um} is indeed a Cauchy sequence in the mean. ness condition If. I ::;; gin Theorem 2.9.1 is essentiaL
We now use Theorem 2.8.3. We conclude that there exists an integrable 2.9.3. Let (X,<i,JJ.) be the measure space introduced in Problem 2.4.4.
function h such that Let
lim
n-ro
Jlfn- hi dJl = 0. (2.9.5) f..(k) = r~. if 1 :-;:,; k::;; n,
0, ifk > n,
Theorem 2.7.2 shows that {f.} converges in measure to h. Sincefis also the
limit in measure of {f.}, we get f = h a.e. Thus, f is integrable. Finally, and f(k) = 1,ik for all k ;::= I. Then {f.} is a sequence of integrable functions
(2.9.1) follows from (2.9.5). that converges uniformly to f, but f is not integrable.
Suppose next that {f..} converges a.e. to f (instead of in measure). If 2.9.4. If J.J(X) < oo and if {f.} is a sequence of measurable func-
we prove that {f.} converges to fin measure, then the proof of the theorem tions that converges uniformly to a function f, then f is integrable and
is completed, by the previous part. Denote by N the set of points for which lim fl. dJ.l = ff dp..
either 1/(x)l > g(x) or lf,(x)l> g(x) for some n. Then, for any e > 0,
Since g is integrable, p.(E.) < oo (by Problem 2. 7.2). The hypothesis that Theorem 2.10.1 Let f and g be measurable functions. If If I ::0:: g a. e. and
g is integrable, then f is integrable.
Ii:;nf.. =f a.e. implies that JJ.(j~\E) = 0. Hence, by Theorem 1.2.1(iv),
lim p.(E.) = 0. Since Proof It suffices to prove that 1/1 is integrable (sec Problem 2.6.3).
By Theorem 2.2.5 there is an increasing sequence {h"f of simple nonnegative
"
functions such that lim h. = 1/1 a.e. The relations h. :-;:,; g show that the h.
{x; lfix)- f(x)l;:::: e} c E.,
are integrable simple functions (by the result of Problem 2.7.2). Now use
it follows that {f..} converges to fin measure. the Lebesgue bounded convergence theorem to conclude that Ill is integrable.
This follows from Theorem 2.10.1, noting that 1/91 .:5 cl/1 a.e. and
cl/1 is integrable. (2.10.4)
Corollary 1.10.3 If a measurable function f is essentially bounded on a If, in particular, lim Jfn dj.t < oo, then f is integrable.
n
measurable set E with finite measure, then f is integrable over E.
Indeed, the function IXE/1 is bounded a.e. by CXE where cis -a (finite) Proof. If lim J/.
d;t = oo, then (2.10.4) is obviously true. It thus
real number, and eX is an integrable function. Now apply Theorem 2.10.1. remains to consider the case where lim Jf~ d).l < c.o. Let Un(x) = inf f/x).
j~n
The next theorem is often referred to as the Lebesgue monotone con- Then {g.} is a monotone-increasing sequence of nonnegative integrable
rergence theorem. functions, and Un .::s;J It follows that
Recall that iff is not integra hie then the integral on the right of (2.1 0. 1)
If d;t = !~~ JOn djl.
is taken in the sense of Definition 2.6.6. Combining the last equality with (2.10.5}, we obtain the inequality (2.10.4).
Themem 1.10.5 Let {f"} be a sequence of nonnegatire integrable Prove that Z is a complete metric space with the metric p.
functions, and let f(x) = lim f,(x). Then 2.10.5. Let J.l(X) < oo and let .f(x,t) be a function of x EX, l E (a,b).
Assume that for each fixed t,f(x,t) is integrable, and that the partial derivative
THE RIEMANN INTEGRAL 61
60 INTEGRATION
i:f(x,t)/ct exists and is uniformly bounded for x EX, t E (a,b). Then, for each
2.10.14. Give an example where Fatou's lemma holds with strict
t E (a,h), rif(x,l),.-'a is integrable, Jf(x,t) d;1(x) is differentiable, and inequality.
2.10.15. Let/be a real-valued measurable function in a finite measure
a I f(x,t) dji(x) = I(:Ft f(x,t) dji(x).
dt space. Let
2.10.8. Let {f.} be a sequence of nonnegative integrable functions. If f"";?,O and one series, say that of s., is convergent, then use the inequality
Prove that if the series .f(xl = :[f.(x) is an integrable function, then f(x) ~[,(x) + 1/2".]
2.10.16. With the notation of Problem 2.8.1, we denote by L"'(X,6.,,J1)
L"' Jfn dJl < OC. or, more briefly, L"'(X ,f.l), the space of all classesJof measurable and essentially
n=l
2.10.9. Let f and .f. (n = 1,2, ... ) be integrable functions such that bounded functions[ Prove that L""(X,p) is a complete metric space with the
0 5,f.(x) -5,/(x) a.e. Then metric
p(f~g) = p(f,g) = ess sup lf(x)- g(x)l.
I (li~ t.) dp "2.: 11~ Jt. dp "2.: 11!11 It. a,..2 I (1i~ t") a,... ~Ex
S",s" arc called the upper and lower Darboux sums associated with the par- Ek = {x; lim sup lf(y)- f(z)l > ~}
tition IT. Note that sll 2 Sn. ~-o ly-xl <~
1::-...T <J
00
Definition f is said to be Darhoux-integrable if lim Sn and lim sn The set E where f is not continuous coincides with
o
U Ek.
k~l
Iff is not con-
exist as ITI I --+ 0, and if
tinuous a.e., then p*( 0 ) > 0. Therefore also
lim Sn = lim sn. (2.1l.l) p*(Ek) = rx > 0 for some k.
llll->0 Iili-O
The common limit is then denoted by k is now fixed. .. . < x = b and let
Take any parttt1on II. a= Xo < X1 < < Xn-1 ". _
1
b 1 = {x~" ... ,x"}. The intervals (x 0 ,x 1), ,(x._ 1 ,x.) form a covenn~ ~f Ek
a
r f(x) dx and p*(Ek _ J) = p*(Ek) = rx. Denote by J' the subset of J consistmg of all
the points x, for which
and is called the integral off from a to b, in the sense of Darboux. (x;-t>X;) n Ek "# 0 -that is,
r
"[d,b]
f(x) dx. On the closed set f = [a,b] - E, f is continuous and, therefore, uniformly
continuous. Hence there exists a J > 0 such that
We shall prove the following two theorems. (2.11.4)
lf(x)- f(y)l < 2(b -a)
if X E f, lx- J!l < J.
Theorem 2./1.1
is Riemann-integrable
A bounded function f on the bounded intenal [a,b]
if and only if it is continuous a. e.
Let II, (k = 1,2) be partitions wit_h. me~h <J, given b_:: sets
{ x1, ... ,x~i:" 1 ,x~.= b}, and let II be the partltlon given by the set 1 - 11 u 12 -
Jk:
{xl' ... ,xn-l'x" = b}. We shall compare sums
Themem 2.11.2 If a bounded function f over a bounded interval [a,b]
is Riemann-integrable, then it is also Lebesgue-integrable and the two integrals
S, = L f(y~)(x7 - X~-1) (k = l,l),
Xik e: Jk
agree-that is,
S= L J(zi)(x,-xi_ 1),
ta
f(x) dx = r
[a,b]
f(x) dx. (2.11.2)
)[j~J
64 INTEGRATION THE RIEMANN INTEGRAL 65
where x~ = x 0 = a, y~ e [x7 _ 1 ,x';], and z1 E F n [x1 -I>x1] if the latter set is For any x e E 1 with f(x) e I there is an open interval lx about x with
not empty. If f(lJ c I. (Here we needed the fact that xi= a, xi= b.) Denote by J the set
F n [x 1 _ 1 ,x1] = 0. (2.11.5)
U J x when x varies over the set 1 n f- 1(I). J is an open set and /1 1 (1) =
1 n J. Since 1 and J are measurable sets, also/1- 1(I) is a measurable set.
we require only that z1 e [x1 _ux1]. As for the sct/2- 1 (1), it is a subset of the set E 2 of measure zero. Hence it is a
We can clearly write a measurable set. From (2.11.7) it then follows that f- 1 (1) is a measurable
set. Hencefis a measurable function.
Since f is a bounded, measurable function, it is Lebesgue-integrable
where y1 =
1
y ; if [x1 _ 1 ,x1 ] c [xl_ 1 x}]. Then (by Corollary 2. I0. 3). It remains to show that (2. 11. 2) is valid.
We shaH associate with each Darboux lower sum Sn a step function
IS- S1l :=:: L
Xj ~ j
lf(z1) - f(Yj)l(x 1 - x 1- 1 ). u0 defined by
un(x) = m 1 if x,_ 1 < x:::; x,.
Denote by J' the subset of J consisting of those points x 1 for which (2.11.5)
holds, and let J" be the set J- J'. Recalling that jz1 - y 1 j < b and using Since f(x);?. un(x), J1a.b1f(x) dx;?. b.bl un(X) dx. However,
(2.11.3), (2.11.4), we get
IS- S1l :=:: L lf(z)- f(y)j(x 1 - x 1 _ 1 ) j' u 0 (x) dx = Lm 1 tt{[x,_ Hx;]} = L m;(x; - x,_ 1) = sn.
"' [a~b] i i
Xj ~J
Hence
+ L lf(zj)-f(y1)1(x1 -x1 _ 1)
.;t"jE'J"
I!
It follows that lim Sn exists as ITil --> 0. Similarly lim s0 exists as ITil --> 0. A function f may then be Riemann-integrable without its absolute value
Next take I1 1 = Il 2 and choose the y,J such that jS1 - S[J,J < e and the yi2 1/1 being Riemann-integrable. This situation cannot occur for Lebesgue-
such that IS 2 - s[J 1 <e. We then obtain
1 integrable functions.
PROBLEMS
This proves (2.11.1 ). Hence f is Riemann-integrable.
2.11.1. Prove the following theorem of Darboux: Let f(x) be a
bounded function in a bounded interval [a,b]. Then there is a number S
Proof of Theorem 2 .11.2. We shall first prove that f is measurable.
such that for any sequence n .. of partitions of [a,h] such that ITiml ..... 0,
Denote by l 1 the set of points where/is continuous and let E 1 == 1 - {a,b}.
Sn, _,. S.
The set E 2 == [a,b] - 1 has measure zero. Denote by/; (i = 1,2) the restric-
2.11.2. Prove that a bounded monotone function can have at most a
tion of /toE,. For any open interval I on the real line,
countable number of points of discontinuity. (Hence, it is Riemann-
(2.11. 7) integrable on any bounded interval.)
66 INTEGRATION THE RADON-NIKODYM THEOREM 67
2.11.3. The function f(x) =(sin x)/x is Riemann-integrable (but 2.11.8. Analogously to the Darboux integral f(x) dx, one defines J:
not Lebesgue-integrable, by Problem 2.7.5) over the interval (l,c.c ). the Riemann-Stieltjes integral as follows: Consider upper Darboux-Stieltjes
2.ll.4. Construct a sequence {f"} of Riemann-integrable functions
sums
on the interval [0, 1], such that lfn(x)l ::<::; 1 for all 0 ::<::; x ::::; 1, n = 1,2, ... , and n
limj~(x) =f(x) exists everywhere, but f(x) is not Riemann-integrable on 50 = I M;[g(x;) - g(X;-t )]
[0,1]. i=l
2.11.5. Let{, fn (n = 1,2, ... ) be bounded Riemann-integrable functions and lower Darboux-Stieltjes sums
on the interval [0, 1]. Assume that there is a constant K such that Ifn(x) I ::::; K
s 0 =I m;[g(x,)- g(x,_t)].
for 0 ::::; x ::::; 1, n = 1,2, .... Show that if Iimfn(x) =fix) a.e., then
1 1 If Jim S and lim s exist, as IIII ---> 0, and if they are equal, then we say that
lim J fn(X) dx = ( f(x) dx. f is Darboux-Stiel:;es integrable (with respect to g) ~nd _the_ limit is denot_ed
,._,.00 0 ... 0
by J: fix) dg(x). This limit is called the f!a~bo~x-Stlelt!es mtegral off :'It?
2.11.6. Let f be a bounded Riemann integrable function on bounded respect tog, from a to b. The Riemann-SIIelt;es mtegralis defined as the_lm_ut
intervals of the real line. Assume that lfl is Riemann-integrable on (- co,oc)-
that is, lim
"~"'
r"1/(x)l dx < w. Prove that f is Lebesgue-integrable on
of D~ 1 f(y,)[g(x;)- g(x,)] when IDI--+ 0; it is equal to the Darboux-StJeltJe.s
integral. In the above sum, they, are any points in [x,_ 1 ,x;~ Prove t~at_ 1f
1 is continuous and g of bounded variation, then the Rtemann-StteltJeS
(- w,w) and
I (~oo,oo)
J(x) dx = r
~a:,
f(x) dx.
integral exists. .
2.11.9. Prove that iff is continuous in [a,b] and g Js monotone
increasing and continuous from the right, then
2.11. 7. Let R be a rectangle given by a ::<::; x 5 b, c ::<::; y ::<::; d, and let f
be a bounded function on R. Let II 1 : a= x 0 < x 1 < ... < x" = b be a
partition of (a,b) and let II 2 : c = Yo < y 1 < ... < Ym = d be a partition of
J f(x) dg(x) = J
b
a (a,b]
f(x) dg(x),
su = L L
;~ l j~ l
mii(x; - x,_ 1)(yi - Yi-1 ).
If lim Sn and lim sn exist as IIII __,. 0, and if they are equal, then we say that 2.12 THE RADON-NIKODYM THEOREM
f is Darboux-illlegrable on R. We then define the integral off orer R to be
lim Sn. Similarly one defines the Riemann integral off over R as the limit In this section we consider measure spaces (X ,C1,J1), where fl. ts a
JnJ~o signed measure.
of the Riemann sums If('., 1"/i)(x, - x,_ 1)(yi- Yi- 1 ), when III I---> 0. It is
well known that the Riemann integral exists if and only if the Darboux Definition 2.12.1 Let (X,C1,J.I) be a measure. space and let v. be a
integral exists, and then the two integrals coincide. signed measure with the domain a.
We say that v IS absolutely contmuous
Prove: Iff is measura hie, then the Lebesgue integral off over R is equal with respect to , and write v 4 11 , if v(E) = 0 for any measurable set E for
11
to the Riemann integral off m'er R. which lp.I(E) = 0.
68 INTEGRA T!ON THE RADON-NJKODYM THfOREM 69
Notice that if Jt is a measure and if r is absolutely wntinuous with function /..{E) defined by (2.8.1) is absolutely continuous in the sense of Definition
respect to J1 according to Definition 2. 8.1, then I' is also absolutely continuous 2.8.1.
with respect to p according to Definition 2.12.1. We shall prove (Theorem We now state the theorem of Radon-Nikodym.
2. 12.2) that if IvI is finite on every set where p is finite, then the two concepts
of absolute continuity are equivalent. Theorem 2.12.4 Let (X,ct,11) be a a-finite measure space with 11 a
measure, and let v be a cr-finite signed measure on a, absolutely continuous
Lemma 2.12./ For any two signed measures 11 and v, the following with respect to 11 Then there exists a real-ralued measurable function f on X
three conditions are equiralenl: such that
{a) r4p;
(b) v+ 4 11 and r- 4 11; v(E) =
'E
rf d11 (2.12.1)
(c) il'l4p.
for every measurable set Efor which lvi(E) < w.lfg is another function such
Proof. Let X =A u B be a Hahn decomposition with respect to v. that v(E) = h g dl! for any measurable set E for which lvi(E) < w, then
Suppose (a) holds and let ltti(E) = 0. Then ltti(E n A) = 0, l;ti(E n B)= 0. f = g a.e. with respect to 11
Hencev+(E) = v(EnA) = O,v-(E) = v(EnB) = O.Thisproves(b).Next,(b)
implies (c), since lrl = v+ + v-. Finally, (c) implies (a), since 0 .s;lv(E)I.s; Note that f is not asserted to be integrable. In fact, it is clear that f ~s
lvi(E). . integrable if and only if Ivi( X) < w. Thus if~~ 11 and lvi(X) < w, then v 1s
the indefinite integral of some ,u-integrable function f
Theorem 2.12.2 Let )..l be a signed measure and let v be a s(qned measure Note also that iff is any measurable function, then, by Theorem 2.7.4,
such that lvi(E) < wfor any measurable set Efor which IJ..li(E) < w. Suppose the set function v(E), defined for all measurable sets E for which JE lfld;t < w,
o
that v 4 )..l. Then for any 1: > 0 there is a > 0 such that lvi(E) < 1: for any satisfies lvi(E) = 0 if IJli(E) = 0. If, further,fis integrabl~ on any measura?le
measurable set Efor which 1111(E) < 8. set E with 1111() < w, then (by Corollary 2.12.3) v(E) 1s absolutely contm-
uous in the sense of Definition 2. 8.1.
Proof. If the assertion is false, then there is an e > 0 and a sequence
{"} for measurable sets such that 1111(") < 1/2" and lvi(E,);;:: e for all Definition The function f occurring in (2.12.1) is ca!led the Radon-
n :2': I. Let E = ITmE,. Then Nikodym derivative of I' with respect to ,u, and one writes
n
dv
f = d,u' ordv = f d,u. (2.12.2)
for any n;;:: I.
It follows (by Problem 2.1 0. 7 with f. = .fxEnz) that f is J.l-i ntegrable on E Consider the set function ..1.() = v(E) - ftf dJ.l. If we show that A 0, =
an? (by Theorem 2.8.4) that (2.12.1) holds. As for uniqueness, if g is a function then we have completed the proof of the theorem. Note that A<:>; J.l, and . 1.
as m the last part of Theorem 2.12.4, then JF g d}' = JF f dJ.l for any measurable is a measure.
subset of each set z. Hence, by Theorem 2.7.6, g = f a.e. in Z". Therefore Suppose A" 0 and let X= Am v Bm be a Hahn decomposition for
also g =fa .e. in X. We have thus shown that it suffices to prove the existence the signed measure A- (1/m)J.l (m = 1,2, ... ). Write
part of the theorem in each subspace z., where I' and v are both finite. Thus,
without loss of generality we may assume that J.l(X) < oo, [v[(X) < oo. A0 =
"'
U
11'1=1
A,, B0 = n"' B,..
m.= I
Let X= A v B be a Hahn decomposition for v. The argument above
shows that it suffices to prove the existence part of the theorem in each of the Since B 0 c: B,.,
subspaces A and B separately. Since vis a measure on A and -vis a measure
on B, we may assume, without loss of generality, that vis already a measure 0 5, i.(B 0 ) .:-::; ..!_ J.l(B 0 )-0 if m-oo.
m
on X.
We now consider the case where J.l and v are finite measures, and prove Hence A..(B0 ) = 0. Since i. " 0 and A 0 n B0 = 0, A 0 u B0 =X, we must
the existence off Denote by !2 the class of all nonnegative functions j have A(A 0 ) > 0. By absolute continuity of . 1. with respect to J.', we also have
integrable with respect to J.l, such that ' p(A 0 ) > 0. Hence J.l(Am) > 0 for at least one m. Denoting l/m by ;; and
Am by A, we then have [by the positivity of Am with respect to A- (1/m)J.l]:
J J dp .:-::; v(E) for all E Ea.
Let
E
ep(E n A).:-::; ..1.( n A)= v(E n A}- f
"EnA
f dJ.l
Then there is a sequence {f.} in 9 such that t g dp. = tf d}' + BJ.l(E n A).:-::; t-.-~ f dJ1 + v(E n A)
lim
-a::
Jf. dp. = a. .:-::; v(E - A) + v(E n A) = v()
Let g.= max (/1 , .. ,/.). We write each measurable set E as a disjoint union for any measurable set E. Hence g E !2. However, by (2.12.4),
i 1)"1 E1 of measura hie sets E1 defined as follows: x E 1 if J;_ (x) ;:.: fix) for f g d}' = f f dJ.l + CJ.l(A) > !X,
2 =::;j =::; n; x E 2 if x f. 1 and if / 2 (x) zfj(x) for 3 =:;;j .:-::; n; and so on. It
follows that g,(x) = ~(x) for x E 1. Hence thereby contradicting (2.12.3). We have thus proved that . 1. = 0.
let (X,a,J.l) be a measure space, with p a signed measure, and let
n " J.l = p. + - J.l- be the Jordan decomposition of J.l. A measurable function f is
f g. dJ.l
E
= .L r J, dJ.l .:-::; I
J=l 1 j=!
v(Ej) = v(E).
said to be integrable (on E) if it is integrable (on E) with respect to the measure
Let fo(x) = sup f.(x). Then / 0(x) = lim g.(x), and {g.} is a monotone- [p[. The integral of/(over E) is then defined by
"
increasing sequence of nonnegative integrable functions. The Lebesgue
monotone convergence theorem then implies thatf0 (x) is integrable, fo dJ.l = J (f f dJ.l f f dJ.l+ - f f dJ.l-)
=
f. E
(2.12.5)
I
~ _and r:lo dp. .:-::; v(E) for any E E a (so that foE.@). Since fo is integrable,
PROBLEMS
II IS real-valued a.e. Let f be a real-valued function that is equal a.e. to j .
0
Thenfalso belongs to !2, and 2.12.1. Prove that iff is integrable with respect to J.l then it is also
integrable with respect to J1 + and p.-. [Therefore the integrals in (2.12. 5)
(2.12.4) are well defined.)
THE LEBESGUE DECOMPOSITION 73
72 INTEGRATION
[Hint: Consider first the case where 11 and v are measures and rp is a simple The decomposition of v, given by (2 .13 .I), is called the Lebesgue
function.]
decomposition of v relative to 11-
2.12.5. Let J,J.I. be u-finite measures and let v be a u-finite signed
measure. Assume that v 4: J.l., f1 4 J. Then Proof As in the proof of Theorem 2.12.4, we first reduce the general
dv = dv dJ.I. situation to the special one where 11 and v are finite. We may further assume
a. e. with respect to J. that J.l. is a measure, since v0 .l 1111 and v1 4 1111 imply that v0 .l JJ. and v1 4: J.l..
d). dl1 d).
Finally, we may assume that v is a measure, for otherwise we can treat v+
2.12.6. Let J.1. be a a-finite measure and let v1 , v2 be a-finite signed and 1'- separately.
measures such that v1 4: J.l., v2 4 p.. Then v1 + vl 4 /J. and Since p. and v are measures, v is absolutely continuous with respect to
d(v 1 + v2 ) dv 1 dv /). + v. By the Radon-Nikodym theorem, there exists an integrable function/
dp.
-2 .
=dj.l.- +dp. such that
2.12. 7. Let J.1. and v be a-finite measures and assume that 11 - v is v(E) = j"E f d!< + rf d1 (2.13.2)
a measure. If v 4p.- v, then the set of points where dvfdfl = 1 has zero
p.-measure. for any measurable set E. Since 0 ~ v(E):::; JJ.(E) + v(E), we have (by Problem
2. 7.3) 0 ~ f:::; 1 a.e. with respect to the mea~ ure p. + I'. Hence 0 :::; f:::; I a .e.
with respect to the measure v. Let
2.13 THE LEBESGUE DECOMPOSITION B= X-A.
A= {x:j(x) = 1},
Let fl and v be two signed measures defined in the measure space Then
(X,G.). We say that fl and v are mutually singular, and write /J. .L v, if there
exist measurable sets A and B such that v(A) = j"
A
dtt +
"A
I' dv = J.I.(A) + v(A).
X= A uB, A nB=0,
Since v(A) < oc, it follows that JJ.(A) = 0. Therefore, if we define
and (E E Ci),
v0 (E) = v(E n A),
l.ui(A) = 0, lvi(B) = 0.
then v0 .l p. It is also clear that v0 + 1' 1 = v.
We also say that /). is singular relatire to v (or that v is singular relative to JJ.). It remains to prove that v1 4 p.. Suppose p(E) = 0. Then (2.13.2) gives
PROBLEMS (
~ f.'nB
dl' = r
~ EnB
f dv
2.13.1. For each signed measure JJ., J.l.+ .L ~~-, p.+ 4: IJJ.I, /J.- <:t; 11'1-
2.13.2. If v1 and vl are singular relative to /)., then v1 + v2 is also
-that is, )E,.,B (I -f) d1 = 0. Since, on En B, l - f> 0 a. e. with respect
singular relative to p.. [Hint: If X = A 1 u B; is a decomposition relative to to v, we must have 1'(E n B)= 0-that is, v1(E) = 0.
74 INTEGRATION
THE I:EBESGlJfi INTEGRAL ON THE REA LINE 75
T~ prove uniqueness, suppose v = v0 + i'J is another decomposition The assertion g'(x) = f(x) a.e. can also be stated as follows:
of v, With v0 J. p., v1 ~ p.. Then, the signed measure ..1. = v0 - v0 = v - v
is both singular (by Problem 2.13.2) and absolutely continuous with r~spec~ 1 rx+h f(t) dt =
lim- f(x) a.e.
to fJ. Hence, by Problem 2.13.3, A= 0. Thus Vo = Vo and VI = vl" -o
h x
2.13.4. Let A be a sequence {x.,} in the Euclidean space R~. and let Lemma 2.14.2 Let f(x) be a Lebesgue-integrable function on (a,b).
{p,} be a sequence of positive numbers. Denote by v the u-finite measure If J: f(t) dt = 0 a.e., then f(x) = 0 a.e.
(cf. Problem 1.2.5) given by
Proof. The assumptions on/imply that ftJ(x) dx = 0 for any interval
v(E) = L
XmE'E
Pm (E any subset of R"). E. Since the set function J,J(x) dx is completely additive, it vanishes on any
open set E. Hence it vanishes also on every closed set E of (a,b). Suppose now
Find the Lebesgue decomposition of I' with respect to the Lebesgue measure that the assertion is false. Then there is a set F of positive measure such that
p. of R~. either f> 0 on For f < 0 on F. It is enough to consider the first case. Since
p.[(a,b)- F] < b- a, there is (by Problem 1.9.7) an open set E containing
(a,b) - F such that p(E) < b- a. The set G = (a,b)- E is a closed subset
2.14 THE LEBESGUE INTEGRAL of F and p.(G) > 0. Since j> 0 on G, we get (by Theorem 2.7.5)
ON THE REAL LINE JGf(x) dx > 0, a contradiction.
In this section we restrict our interest to the real line, and find a con- Lemma 2.14.3 ljf(x) is bounded and Lebesgue-integrable on (a, b), then
nection between integrals and derivatives. This connection, stated in Theorem the assertion of Theorem 2.14.1 is mlid.
2.14.1, generalizes the Fundamental Theorem of Calculus. It will have
implications for some of the abstract theorems we have previously proved, Proof. Write
such as the Radon-Nikodym theorem.
We shall denote the Lebesgue integral of a functionf(t) over an interval
(c,d) by Jed f(t) dt. We also define f.{ f(t) dt = - J~ f(t) dt, Jccf(t) dt = 0. We
g(x) = r
"'a
r(t) dt- r r<tl
"'"
dt. (2.14.2)
shall take (a,b) to be a fixed bounded interval. Each of the indefinite integrals is a monotone-increasing function. But it is
known (the proof, which is rather involved, is omitted) that monotone
Definition Iff is Lebesgue-integrable on (a,b), then the function functions are differentiable a.e. Hence g'(x) exists a.e. Let 1/(x)l .:<:;; M for all
g(x) = r ~
f(t) dt +c (c any constant) (2. 14.1)
a< x < b. Then, if 0 < h < b - x,
a
[g'(x) - f(x)] dx = 0.
a
Since, however, g'(x) Z.j(x) a.e., we get g'(x) = f(x) a.e.
-that is, Recall (see Problem 2.8.2) that the indefinite integral is an absolutely
coni inuous function.
I' [g'(t)- f(l)] dt = 0 The converse is also true, namely, if g(x) is an absolutely continuous
function in (a, b) then there exists a Lebesgue-integrable function f(x) such that
for all x. Lemma 2.14.2 then gives g'(xl = f(x) a.c.
then <:p'(xl 1s integrable, and We shall briefly indicate the proof. From Problems 2.8.3 and 2.8.4
,.h
we know that g(x) can be written as a difference oft wo monotone-increasing,
j <;J'(x) dx::;; 09(6)- <p(a). (2.14.4) continuous functions, say g(x) = g 1(x)- _qlx). Lemma 2.14.4 shows that
a g'1 (x) and g1(x) arc integrable. Hence g'(x) is integrable. Consider the function
Proof Define ~.p(x) = ip(b) if x > b and ~.p(x) = ~.p(a) if x <a. Since <P(x) = g(x)- r~ g'(t) dt.
[~~(x + hl - ip(xl}'h Z. 0 and [ip(x + h) - rp(x)].ih ..... ip'(x) a.e. as h __. O, a
Fatou's kmma gives
'P is absolutely continuous and ~.p'(x) = 0 a.e. But it can be shown (we omit
. .o rp(x +h)- rp(x) -" . the proof, which is rather involved) that such a function is necessarily a
hm 1
h 0 a h
dx ::0: I
a
rp(x)dx; constant.
if the left-hand sidt is fmitt. then <p'(xi is integrable. The integral on the kft
PROBLEMS
can be eva I uatcd in the same way as the integral on the left in (2.14.3). We
fmd that, ash....,. 0, this integral converges to <p(h)- ~.p(a). This proves (2.14.4). 2.14.1. To every function /, monotone-increasing and continuous
We can now compkte the proof of Theorem 2.14.1. We may suppose from the right on (a,b), there corresponds a Lebcsgue-Stieltjes measure,
~hat /(x) Z. 0 a.e., for otherwise we can treat each of the indefmite integrals which we denote by 111 . Every fundionfof bounded variation and continuous
m (2.14.2) separately. Sinc:c g(x) i~ monotone. g'(x) exists a.e. Deline from the right can be written in the form f = J; -/2 , where / 1 and / 2 arc
/,(x) =min {f(xl,n}. Then/(x)- fl'(l Z. 0. It follov.s that monotone-increasing and continuous from the right. We write /lf = p. 1 ,
- p. h' Prove thaI the function f(x) is absolutely continuous if and only if the
'
I [f(tJ-J..<nJdt signed measure )lf is absolutely continuous with respect to the Lebesgue
measure (which we shall denote by) Jlx
is nondccrcasing. Henn~ its derivative (whi<.:h exists a.c.) is ;:::0. This gives 2.14.2. Let /(x) be absolutely continuous on (a,b). Prove that
dJ1 1 /dp. .. =f'(x) a.e., and that for any Printegrable function 'f,
d d
-,. I
i .\ a
,x
f(t) dt ;::: -I
dx 0
X
f"(l) dt.
f ~ df =
0
r <Pf' dx.
"(a,b) a
By Lemma 2.14.3, the expression on the right is equal to f"(x) a. e. Hence Here we use the notation: df = dJ1 1 .
78 INTEGRATION
PRODUCT OF MEASURES 79
(This is actually true iff and g are just assumed to be continuous and of
Lemma 2.15.2 Let (X,a,)l) and (Y,M,v) be measure spaces. Denote
bounded variation.)
by ff the class of afl finite disjoint unions of rectangles of (i X ff4 that are
2.14.5. A function f(x) is called singular if f'(x) = 0 a.e. Show that
contained in a fixed subset of X x Y. Then :F is a ring.
every function g of bounded variation in (a,b) can be written as a sum
9 = 9 1 + 9 2 , where g 1 is absolutely continuous and 9 2 is singular. Show also
that this decomposition is unique, except for an additive constant. Proof :F is clearly dosed under the formation of finite disjoint
2.14.6. Let g, 9 1 , g 2 be as in Problem 2.14.5. Show that in the de- unions. One easily verifies the relations
composition f/.g = f/.g, + f/.g,, p. 9 , is absolutely continuous with respect to
fl.x and f/. 92 is singular relative to /J,..
2.14.7. If f(x) is a continuous function in (a,b) and if f'(x) exists
/)
1
(Aj x B;) =col A~) x Co~ Bj) (2.15.1)
in a bounded interval [a,.BJ and let g(x) = J;_, xit) dt. Then 9'(x) = x.dx)
m
F = UF1. Then
a.e. Note now that 9(x +h) - 9(x- h) = p.[E(x, h)].] j~l
m
n (E,-F).
~
E-F= UI
i~ j~ I
2.15 PRODUCT OF MEASURES
By (2.15.2), for each i and j, E,- F, =D11 v D 12 , where Di! and DJ2 are
Let X and Y be two spaces. The Cartesian product X x Y is the set disjoint rectangles of a x rM (depending on i). We shall use the relation
of ali ordered pairs (x,y), where x E X, y e Y. Any subset of X x Y of the
form A x B, where A c X, B c Y, is called a generalized rectangle (or, n (D
m
J~ I
11 v D12 ) = U [Du, n
I.
... n Dmd where k, = 1,2; s = l, ... ,m.
briefly, a rectangle) with sides A and B.
Let <i be a a-algebra of subsets of X and let I'J be a a-algebra of subsets Noting that the sets in the union, on the right, are mutually disjoint, and
of Y. We shall denote by <i x I'J the a-algebra generated by all the rectangles using (2.15.1 ), we see that each set G, = n
(E, - F;) is in 9'. Since E - F
j
A X B with A E <i, BE 14. We call a X 14 the Cartesian product of a and f!l. is a disjoint union of the G 1, it also belongs to :F.
-~~-
80 INTEGRATION
'C
PRODUCT OF MEASURES 81
Lemma 2.15.3 If f.! and v are cr-fi'nite measures, then every set in X x Y
definition of Jt 0 it follows that .1( 10 ::::> Jt 0 . Since this is true for all E E J! 0 ,
can be covered by a countable disjoint union of rectangles of Ci x f!4 having
sides offinite measure.
.it 0 is a ring.
If !?); is a class of sets and A is any set, we denote the class of all sets
E n A, where E E 2, by :2 n A.
Proof It is sufficient to cover X x Y. Let X = UXn be a countable
disjoint union of sets in Ci with fi(Xn) < co. Similarly: let Y = U Ym be a Lemma 2.15.6 Let P be any class of sets and A any subset of X. Then
countable disjoint union of sets in f!4 with v(Y.,) <co. Th~n X x y 9(2) n A = .9'('2 n A).
= U (X" x Ym) gives the desired covering of X x Y.
n,m
Proof Since Y(P) n A is a a-ring containing 2 n A, it contains also
Y"(f? n A). To prove the converse, denote by % the class of all sets of the
Definition A class .it of sets is called monotone, if for any monotone form B v (C- A), where BE Y'(Y n A) and C E 9(f?). Writing E =
sequence {"} of sets in J! we have: lim E" is in J!. (E n A) v (E -A) for any E E 9, we sec that 2 c %. Since, as easily
verified,% is a a-ring, we find that .9'(9) c %. Hence Y(.') n A c ,1{ n A.
Lemma 2.15.4 A monotone ring & is a a-ring. Noting, however, that.'>( n A = Y'(EZ n A), we get Y'(.Q') n A c 9(0: n A).
Proof We have to show that if is closed under the formation of Theorem 2.15,7 Let (X,ct,f.l) and (Y,.;il',v) be a-finite measure spaces
countable unions. Thus, let {E.. } be a sequence in &. Since rJt is a ring, the and let E he any set in a x JQ. Define functions f and g by f(x) = v(E.) jor all
finite unions F" = lJ Em belong to&. Since :1 is monotone, UEm= lim F"
m=l m=l ,.
x EX and g(y) = IJ-(P).for ally E Y. Then f and g are measurable functions, and
lim" - F =lim (" - F) E .4/ 0 , (c) Let 0 = A 0 x 8 0 be a rectangle of ct x .!8 with Jl(A 0 ) <co,
" v(B 0 ) < co. Let {E.) be a monotone sequence of sets in ct x :$ such that
F -limE"= lim (F-E") E J/ 0 , E. c 0 if n ~ 1. We claim that iff,. E 9 for all n::::: I, then also E =lim E.
n n
is in . Since a X :Ja is a a-algebra, E belongs to ct X .@, The sequence of
X -sections E".x of E. (for each fixed x) clearly converges monotonically to the
corresponding X-section Ex of E. Similarly, {t~} wnverges toP. Denote by
/,and g,. the functions f and g corresponding to E. Then {.f.l and {g"} con-
Jf FE if, then E E.? implies E E Jf' F Hence :Jf c % r Therefore .If0 c% F verge monotonil:ally tofand g, respectively, and
This implies that if E E J! 0 , FE:!!, then FE% E Thus .ffE => Jf. From the 0 s; J~ s; v(B 0 1,
-~
82 INTEGRATION
PRODUCT OF MEASURES 83
The Lebesgue-bounded convergence theorem implies that .f and Suppose A: is another measure satisfying (2.15.5). By Lemma 2.15.3,
integrable and g are
we can write X x Y as a countable disjoint union of rectangles E" of x [JIJ a
with sides having finite measure. Denote by Pl. the class of all sets in
Jf dt~ = li:U Jfn dt~ = li:n I g" dv = Jg dv. {ct x .14) n E. on which ..l and X agree. Denote by X. the class of all finite
disjoint unions of rectangles of <l x [j contained in E,.. Clearly X,. c '21:,. and 21:.
Hence E belongs to !0. is a monotone class (by Theorem 1.2.1 (iii)). Since, by Lemma 2.15.2, X,. is a
(d) ~t E be a rectangle in a
x .14 with sides having finite measure. ring, Lemma 2.15.5 implies that~. contains the class X n E,., where X is the
Denote by 9 the class of all subsets of 9 that are contained in E. Denote class of all the rectangles of a x !4. It follows (using Lemma 2.15.6) that '21:. =
by !1' E the class of all finite disjoint unions of rectangles of a
x rJ6 that (a )( .14) n E". Writing any set fin a x_M in the form f= UCF n .)and using
are subsets of E. By Lemma 2.15.2, !J' E is a ring. By (c) and (a), (b), q E is the last result, it follows that ..I.( F)= J.(F). Thus l =l
a n:onotone class and qE ::J JE. Hence, by Lemma 2.15.5, 9iE contains the The rectangles of a x J6 are also called measurable rectangles. The
a-nng genemted by jT f." Denote by ,-;J;' the class of all rectangles of X ::Jd.a reason for this name is that if a rectangle A x B is a set in a
x M, then (by
Then .:f n 1:.: = {FE .X; F c }. ~on~equently, .9"(. ...) ::J 9"(.)( n ). By Lemma 2.15.1), A E a, BE !4.
Lemma 2.15.6, the latter class comc1des with SI'(Jf") n -that is, with
(ax Jl) n E. Thus, 9 => PE ::J (a x 86) n E. PROBLEMS
a
(e) Let F be any set of x iJI. By Lemma 2.15.3, F is wntained in a 2.15.0. Prove that 2lim X 00k = OOm+k
countable disjoint union nY/n""
of rectangles En of aX J having sides of 2.15 .] . A Borel set E in R" is called a G ~set if it has the form E = n
k~ I
Ak,
fin!te measure. By ~d), ea~h set F n E" is in !2. Hence. by (a), the disjoint where the A 1 are open sets. From Problems 1.9.7, 1.9.8 we have that every
unJOn V
(F n E") IS also m 9-that is, FE 2. This completes the proof. bounded Lebesgue set Fin R" has the form F =E-N, where Eisa G~ set
and t~(N) = 0. Verify (2.15.1) for m = co, thereby concluding that if A and B
are GJ sets in R" and R', respectively, then A x B is a GJ set in R"H.
Theorem 2.15.8 Let (X,a,J.l) and (Y,b.?,v) be cr-finile measure spaces.
Then Ihe set function ), defined for erery set E in X :3d by a 2.15.2. Let N be a set in R" having Lebesgue measure 0. Then for any
set B in R\ N x B is a Lebesgue set in R"+k having Lebesgue measure 0.
).(E)= I v(Ex) dt~(x) = J,u{P) dv(y) (2.15.4)
[llinl: Consider first the case where B is bounded.]
2.1 5.3. Denote by 2'" and !t'" the classes of the Lebesgue sets in R"
and R", respectively. Then any bounded rectangle of !!'" x !t'" has the form
is a cr-finite measure on ax 36, haring the property that, for any rectangle
E-N, where E = 1 x E 1 is a G 10 set in R"+~<, N is a set of Lebesgue measure
A X 8 oj'a X :.H
0, and 1 , E 1 are GJ sets of R" and R', respectively. It follows that !t'" x !t'"
..I.( A X B) = ,u(A). I'( B). (2.15.5) is contained in the a-algebra a
of all sets E- N, where E is any Borel set in
This last properly determines ), uniquely. R"+" and N is any set in R"u having Lebesgue measure 0. Note (compare
Problem 2.15.1) that a coincides with the class !t'"H of all Lebesgue sets
in R"+k.
Definition ~he. measure ..l is called the product of the measures 11 2.15.4. Every open set in R"H is in fl!" x fl!k. Hence fl!" x !'"
~nd 1,and we wrJte ..t = 1-1 x v. The measure space (X x Y, a x .JA, /.1 x v) contains all the Borel sets of R"u.
JS called the Cartesian product of the measure spaces (X,a,/.1) and ( Y,ai', v). 2.15.5. Prove that!!'" x !'" #- ff!"+k.
2.15.6. Denote by [Rm,fl!m,(dx)m] the Lebesgue measure space in
Pro_of. Since v(E,) is a nonnegative measurable function, its integral Rm. Prove that the measure (dx)"+k and the product (dx)" x (dx)k coincide on
...l(E) IS e1ther a nonnegative real number or +co. The complete additivity rectangles A x B with bounded sides, when (i) A and Bare intervals; (ii) A
of A follows from the complete additivity of l' and from the Lebesgue mono- and Bare open sets; (iii) A and Bare G ~sets, and, finally, (iv) A and B are any
tone con vergence theorem (compare Problem 2. 10.7 J. Si nee X x y can be Lebesgue sets.
covered by a countable class of rectangles of a x J whose sides have finite 2.15.7. Prove that the measure space (R"+k,fl!n+~<,(dx)"+") is the
measure, ...l is a-finite. It thus remains to prove uniqueness. completion of the Cartesian product [(R",!i'",(dx)") x (R\!t'k,(dxl]. Prove
_\
84 INTEGRATION FlJBIN!'S THEOREM 85
also that the latter product is an extension of the restriction of For any measurable function h on X x Y, we write the integral Jh dJ..
(R"+k,!l'H,(dx)"+k) to the Borel u-algebra of R"H. (l = J1 x v) also in the form
2.15.8. If fJ 1 ~ f.J 2 , v, ""!1 v2 , then Jl1 x v1 ""!1 f.J2 X Vz; the p,;, v; are finite
measures.
2.15.9. A function tp in a subset Q of R" is called a special simple
Jh(x,y) d..\(x,y) or Jh(x,y) d(!-l x v)(x,y).
function if there is a finite number of intervals I "bk and real numbers ck We call it the double integral of h. Note that here h is assumed either to be
such that tp = L
ck x.. where Xk is the characteristic function of !"'" Prove integrable or, if it is not integrable, to be nonnegative (and then h d).. = co J
that iff is Lebesgue-integrable in a bounded open set Q of R", then there
by definition). .
exists a sequence {cp,} of special simple functions such that Jn
If- cp,l dx....,. 0 By Lemma 2.16.1, the X-sections h"' are measurable functions. We set
as m-+ co, and lim tp, = f a.e. in n. [Hint: Compare Problem 2.6.5, and use
Problem l.9.7.]
2.15.10. Iff is Lebesgue-integrable on a bounded open set Q of R",
f(x) = Jh,(y) dv(y) (2.16.1)
Proof Let/be a measurable function. For any open interval I on the Proof If h is a characteristic function of a set E, then
real line,
J; 1(J) = {yJ,(y) E J} = {y;f(x,y) E J} Jh(x,y) dv(y) = v(EJ, Jh(x,y) d!-l(x) = Jl(P),
1 1
= {y; (x,y) E/- (I)} = [f - (1) Jx arid the assertion follows from Theorem 2.15. 7. By linearity, the assertion
Since f- 1(1) is in a X @, the set on the right is in @ (by Lemma 2. I 5.1). follows also when h is a nonnegative simple function. By Theorem 2.2.5,
Thus f, is measurable. The proof for fY is similar. there exists an increasing sequence {hn} of nonnegative simple functions that
\
..
1-"J?;-
~~.
86 INTEGRATION
FUBIN!'S THEOREM 87
converges to h everywhere. By the Lebesgue monotone convergence theorem
we then have are indeed defined. Furthermore,
lim Jh. d/ = j' h dl.. (2.16.4) JIhi d(JJ. x v) = fJ lhl dp dv = fJ Ihi dv di!.
"-+"' J
Hence lhl d(p x v) < oo. We conclude that lhl is integrable on X x Y.
J
Letf,.(x) = hJx,y) dv(y). Then {/.}is a nonnegative increasing sequence It follows that his also integrable. Now apply Theorem 2.16.3.
of measurable functions and, again by the Lebesgue monotone convergence Theorem 2.16.2 for h a characteristic function of a measurable set E,
theorem,
or Theorem 2.15.8, gives
lim f"(x)
jl'l-+:))
= rh(x,y) dv(y) = f(x).
...
/() = Jv(EJ dJJ.(X) = Jfl(Y) dv(y). (2.16.5)
Proof. We may suppose that h ;;:-: 0, for otherwise we write h = h + - h- The Lebesgue measure space of R" coincides with the completion of the pro-
1
and consider each of the functions h+,h- separately. The integrability off duct of n measure spaces, each being the Lebesgue measure space of R .
J
and g follows from (2.16.3) and the fact that h d1: < oo. Since f and g are Fubini's theorem also extends to the case of a product of n measure spaces.
then finitevalued a.e., it follows that almost all sections of h are integrable.
We give another version, perhaps the most useful one, of Fubini's PROBLEMS
theorem. 2.16.1. If E and F are measurable sets of X x Y such that v(E,J =
v(F.,) for almost all x E X, then .l.() = J..(F) (). = J.1 x v).
Theorem 2.16.4 If h is a measurable function on X x Y and if eilher 2.16.2. Let f(x) and g(y) be integrable functions on X and Y, re~
JJIhi UJ.l dv < w or JJ
Ihi dv d!-! < XJ, then h is integrable on X x Y and the spectively. Then the function h(x,y) = f(x)g(y) is integrable on X x Y, and
assertions of Theorem 2.16.3 are ralid.
2.16.4. If j(x,y) is a nonnegative Borel-measurable function in R"+k has continuous derivatives of a!! orders in 0:, and
J
(x E R", y E R'), then ff(x,y) dx, f(x,y) dy are Borcl-mea~urable and
D;h(x) = ~ D':f(x,_v) g(y) dy.
Jf(x,y) dx dy = J J
dy f(x,y) dx = J J
dx f(x,y) dy.
c
Here IY; denote-; any partial derivative of order m. Note that the assertion
Here dx dy is the product of the Lebesgue measures dx and dy. form = I, n = I is similar lo the assertion in Problem 2.1 0. 5.
2.16.5. Let E be a planar domain bounded by two continuous curves 2.16.Il. Let n be a bounded open -;ct in R". Let f(xl and g(x) be
Y = <P 1(x), Y = tp 2 (x) for a::;;; x::;;; b, where <P 1 (x) < tpAx) if a< x <b. functions having continuous derivatives in X En, of all orders s; m. A~sume
Prove that iff is a Lebesgue-integrable, Bord-measurable function on, then further that g(x) = 0 outside a compact sttbset of n. Prove that for any
partial derivative n:
of order m
I f(x,y) dp = I
. b ('PI(X)
f(x,y) dy dx, (2.16.6)
... .. a ~ I".P'tL~)
f f(x)D';g(x) dx
Q
= (-I)'" J D;J(x) g(x) dx.
'!l
where Jl is the Lebesgue measure in R 2 .
2.16.6. The formula (2.16.6) remains true (with the same proof) if x [Hint: Suppose m = I, n; = iJi1~x 1 . Replace 0: by an interval containing it
1
varies in a bounded domain of R'" instead of the interval [a,b] of R 1 Using and use Fubini's theorem with X = R 1, Y = R"- , and Problem 2.14.3.]
this formula, compute the volume of the unit ball in R".
2.16.7. Consider the transformation Tx = Ax+ k in R", where A is a
nonsingular n x n matrix and x,k arc column n-vectors. Denote by ). the
Lebesgue measure in R". Prove the relation
(2.16.7)
for n = 2. [Hint: use (2.16.6).]
2.16.8. Prove the relation (2.16.7) for any n;;:: 2. [Hint: Proceed by
induction, and assume that the assertion is true for n - l. Write A = B 1
Bk A 0 , where the Bi arc elementary matrices and A 0 is the unit matrix. Usc
Problem 2.16.1 and the inductive assumption to prove that A.[B/E)] =
ldet Bil..i.(E) for any interval ; hence, by Problem l.9.11, also for any
Lebesgue set.]
2.16.9. Let f.! be a bounded open set in R''. Let Tx = Ax+ k
be a linear map of R" into itself with det A =1- 0. Denote by 0' the image of 0
under the map T. Prove: If f(y) is Lebesgue-integrable for y E 0', then
f(Tx) is Lebesgue-integrable as a function of x in 0, and
Q'
r f(y) dy =
Q
r f(Tx) ldet AI dx.
[Hint: Use Problem 1.9.11.]
2.16.10. Let G be a bounded open set in R" and let n be an open set in
R". Let f(x,y) be a function of (x,y) E n X G, and a~sume that all its deriva-
tives with respect to x exist and arc bounded measurable functions in n x G.
Let g(y) be a Lebesgue-integrable function in G. Prove that the function
h(x) = J f(x,y)g(y) dy
G
I. TOPOLOGICAL AND METRIC SPACES 91
It is easily seen that A and Bare open sets, and, obviously (sec Problem 1. 7.6), If y = (f1 1 ,1h, ... ,rf", ... ) is another point of l\ we define
E cA,F c B, An B = 0.
A subset Y of a metric space X is called bounded if sup p(x,y) < a:_. p(x,y) = L"'
n=l
I~"- l"lnl
:x.reY
N ole that a set is bounded if and only if i I is con I ai ned in some ball. 1
Note that/'"' and / 1 are special cases of the spaces L""(X,J.t) and L (X,J.i.)
If Y is a subset of a metric space X, then we can make Y into a metri~: introduced in Problems 2.10.16 and 2.8.1, respectively. Indeed, they are
space (called a metric subspace of X) by taking its metric to he the restriction obtained by taking X to consist of all the positive integers and J.I(E), for any
of the metric p of X to Y. We call this metric the one imluced bv the metric subset E of X, to be the number of positive integers contained in E.
of X.
A subset Y of a metric space X is called sequentiafly compact if cwry Example 4. The space c is the metric subspace of I"' consisting of all se-
sequence {rnl in Y has a subsequence {y".} that converges to a point in quences ( ( 1 , ~ 2 , ... , ~"' .. ) for which lim ~" exists. The space c0 is the metric
Y--that is, p(y"k'y)--> 0 as n,--> w, for some y E Y.
"
subspace of c consisting of all the sequences for which lim ~" = 0,
A subset Y of X is called dense if f = X. Y is called nowhere dense if
n
f has no interior points.
A space X is called separable if it contains a countable dew,e set ExampleS. The spaces consists of all the sequences x = (~1 .~ 2 , ... ~"''".) with
We finally recall that a metric space is called complete if any Caucl1y real componcn ts. Let y = ( f1 1 ,11 2 , .. , 1J" ... ) be another sequence. Then we define
sequence {x"} [that is, p(x",x,.,) __, 0 as m,n--> w] is convergent.
We shall derive later on various relations between some of the. concepts "" 1 1~. - '1.1 (3.1.2)
defined ahove. In particular, it viill be shown that a set Y is comp:H.:t if and p(x,y) = ~ 2" 1 + I"<::;" -
n -I fin
I
only if it is sequentially compact.
To prove the triangle inequality, it suffices to show that for any real numbers
Example I. R", the n-dimensional Euclidean space, is a metric space with a,b,
__:[a__
+hi:...._ < - [a
-I- Ihi
+ ---. (3.1,3)
p(x,y) = {.f (x;- J)<}l/2,
[....:.1
(3.1.1) l + Ia + bl- I +lui I + lhl
Suppose a and b have the same sign, say both are positive. Then
where x = (x1 , ... ,x"), y = (y 1 , ... ,y"). R" is not compact. The Heine-Bore! , ' a+b
----=
a
+
h
<--+--.
a h
theorem asserts that bounded closed subsets of R" arc compact sets. R" is 1+a+b l+a+b l+a+b l+a. l+h
complete. It is also separable, since the points x = (x1 , ... ,x") with rational
componcn ts x, (I ::;; i ::;; n) form a dense sequence. Suppose next that a and h have different signs. We may assume that lal C. lbl,
so that Ia + bl ::::; Ia[. Since the function x/(1 + x) is monotone increasing
Example 2. The real (complex) space !'". Its clements are sequences x = for x > 0, we get
(~1.~2~", ... ) with real (complex) components~"' such that sup1("1 < oc-. [a+ bl
----'--- < - - - < - - -
lui lal lhl
+ ---
n
If y = (ll!J/ 2.... ,1/ 0 , . ) is another point ofF', we define 1 +[a+ hi - 1 + lal - I + lal 1 + lbl'
Example 6. Let - w < a < b < w. C[a,b] is the space of all continuous
p(x,y) = sup I(, - 1/,J functions on [a,b] (real-valued for the real space C[a,b], and complex-
valued for the complex space C[a,b].) If f,g arc two continuous functions on
A sequence {x"'} in I, with x, = (:;'.,. 1 ,. .. ,~,.,"'" .. ), is convergent to y = [a,b], then we define
(r 1 , .. .. y., .. .) if and only if lim x.,." = y, uniformly with respect to n. p(f,g) = max lf(t)- g(t)l. (3.1.4)
"' Q~t:5b
Example 3. The real (complex) space 11 . Its elements are sequences x = We call this metric the maximum metric. Convergence of a sequence in
(~ 1 .( 2 , ... ,(", . . ) with real (complex) wmponcnts .;., such that L
l~nl < CXJ. C[a,b] (with the maximum metric) means that the sequence of functions is
94 METRIC SPACES LP SPACES 95
uniformly convergent on [a.b]. For this reason we call the metric (3.1.4) 3.1.2. Let X,p,p be as in Problem 3.1.1. Prove that p(xmx)-0
also the uniform metric. if and only if j)(x.,,x) ...... 0. Give an example shmving that p and p arc not
equivalent in general.
Definition T"'o mdrics p and p on a space X are called equiralent 3.1.3. Prove that the Cartesian product (X,p) dc!ined in Example 7
if there exist positive constants :x and {3 such that is a complete (separable) space if and only if each space (X,,p;) is complete
:x/i(x,r) ::::; p(x,r) ::::; {3p(x.y) (separable).
for all x,y in X. 3.1.4. Prove that a sequence {x.,}, with x., = (x'" 1 , ... ,Xmn> .) in the
Cartesian prod net defmcd in Example 8, is convergent if and only if, for each
For example, the metrics
n, {x., .. } is a convergent sequence in (X".p"). Prove aho that (X,p) is complete
fi(x,y) =
I
sup
~i~n
lx, - Yd. p(x,y) = J" 1
lx, - .\";I if and only if each space (X",p") is complete.
3.1.5. Prove that the spaces / 1 ,f"",s,c,c0 , and C[a,h] are complete
arc both equivalent to the Euclidean metric (3.1.1) of W. metric spaces.
3.1.6. Prove that the spaces f1 ,s,c,c 0 arc separable metric spaces.
Example 7. Let (X 1 ,p 1 ), .. . ,(Xm,Pml be metric spaces. Their Cartesian product
3.1.7. Prove 1l1at /" is not separahle. [Hint: Consider the points
x~ = (~t.~ 2 .... ~n), where each~ .. is either 0 or I. This is a nondcnumerable
(X, p) is defined as follows: X is the Cartesian product xl X ... X xm
set, and p(x~,x, ) = 1 if ~ #- ('. The open halls B~ of radius j and center ,;
-that is, its clements arc n-tuples x ~, (x 1 , .. . ,xml with x 1 E X., and
satisfy B~ n B:; = 0 if ~ #- ~. If {x111 } is any sequence, then there exist
m
indices ~ such that xm ~ B; for all 111. Another proof: Suppose A = {xm}
p(x,y) = L p,(x; ,y;) (3.1.5)
i.= 1 is a dense sequence, and xm = (<:';' ..... ~:, ... ).Construct a point~ = (.::-1 , ... ,.::-", .. )
not in .-1.]
if y = (y 1 .... ,y.,J Note that if c 1 , .. .,em are any positive numbers, then the
3.1.8. Construct. in each of the spaces l 1 ,fa,s,c,c0 ,C[a,b], bounded
metric
m closed sets that are not seq ucn tia!ly compact.
Pc(x,y) = L
j-J
c, p,(x; ,y;) 3.1.9. Let X be a metric space, and let Ybe a metric subspace of X.
Prove that Y is also a topological subspace or X.
is equivalent to p.
3.1.10. If p(x".x)-+ 0, p(y ... y)--> 0, then p(xn,Yn) --> p(x,y).
Example H. Let {(X,.p;)} be an infinite sequence of metric spaces. Let X 3.1.11. For what values of pis p(x.r) = lx- !Y a metric on the real
be the Cartesian produtt X 1 x X 2 x .,. x X., x. -that is, it consists of !inc?
points x = (x1 ,x 2 , ... ,x.. , .. .), with X; EX i We cannot define p simply by 3.1.12. If K is a dosed subset of a topological space X, and if Lis a
(3. 1. 5) with m = X, since the series may not converge. Instead we define closed subset or the topological subspace K, then Lis a dosed subset of X.
(3. 1.6)
3.2 U SPACES
where .r = (v 1 .y 2 , ... ,y.,, ... ). We call (X,p) the Cartesian product of the Let (X .ct,ft) be a measurable space, and let p he any positive number.
sequence {(X;.p;)} . .!'\ole that the spaces is obtained as a special case, with We denote by ..'f'P(X,p) the class of all measurable functions f such that
all the X; coinciding with R 1 . J.fiP is integrable. We write
PROBLEMS ( j'IJIp 1 'P
I PI1
IIJ.. p = (. t . (3.2.1)
3.1.1. Prove that if(X,p) is a metric space, and if
Similarly, ::t"(X,p) is the class of all measurable and essentially bounded
A( ) _ p(x,y)
p X,J - ' functions. We write
1 + p(x,y)
'f!l., =esssupl.fl. (3.2.2)
then also (X,fi) is a metric space. [Hint: Cf. the proof of l3.!.3).] X
/
96 METRIC SPACES LP SPACES 97
We shall derive an important inequality known as Holder's inequality. Proof. The assertions for p = I and for p = oo are rather obvious.
We shallthcrefore assume that I < p < oo. From
Theorem 3.2.1 Let p and q be extended real numbers, I p :;; co,
P {2P lf(x)]P, if lf(x)] ~ ]g(x)],
$
I :;; q :;; oo, 1/p + 1/q = l. Iff E ,!l'P(X,j.l.), g E .!l'q(X,11), then fg E 2' 1(X,11) lf(x) + g(x)l 5 2P lg(x)IP, if ]g(x)l ~ lf(x)l,
and
it follows that 1/ + giP s 2P]fiP + 2P]g]P. Hence f + g belongs to ZP(X,p).
11/gill :;; I /II Pile I q (3.2.3) Next, using Holder's inequality, we can write
To prove it, consider the function h(t) = tPfp + t-qjq. It (strictly) decreases
from co to I in the interval (0, 1], and it (strictly) increases from I to oo in the
I = (!1/llp + !1g!!p) II!+ gll~ 1 q.
interval [l,oo), and h(l) = I. Hence From this (3.2.6) immediately follows.
We say that/is equivalent tog iff= g a. e. For any measurable function
for all t > 0. f, we denote hy J the class of all the measurable functions equivalent to f.
Note that iff and g belong to ,PP(X,p), then g E}if and only if II/- BliP =0.
We denote by I!( X ,p) the space of all classes J of functions/ of !f'P(X,p)
Suppose a> 0, h > 0. Taking t = a 11 qWiP and multiplying the resulting and define
inequality by ah, (3.2.4) follows for a> 0, h > 0. The general case a~ 0,
h ~ 0 follows immediately by continuity. p(j,) = p(f,g) = 11/- ellp (3.2.7)
We may assume that II/IlP i= 0, l]g]]q i= 0, for otherwise the assertion
of the theorem is trivially true. We then take Theorem 3.2.3 If 1 s p S oo, then LP(X,).l) is a complete metric space.
Since each of the two terms on the right is in 2' 1 (X,tJ), the same is true of 1/g]. The proof of Lemma 2.5.2 shows that Um} is Cauchy in measure. Henc~,
Further, by integrating both sides of (3.2.5) we obtain (3.2.3). by Theorem 2.4.3, there is a subsequence {fmJ that converges almost um-
We shall use Holder's inequality to derive Minkowski's inequality: formly to a measurable function f. {fm,} then converges to f also a.e. .
Note now that (by (3.2.8)] {J lf~IP dt~} is a bounded sequence. Smce
Theorem 3.2.2 Let I :!>: p s oo. Iff and g belong to .,2"P(X,Jl), then lim 1/...]P = lfiP a.e., Fatou's lemma gives
also f + g belongs to .,2"P(X,Jl), and
(3.2.6)
JlfiP djl :;; li~ J1/,,lr dJI < oo.
98 METRIC SPACES COMPLETION OF METRIC SPACES; H"'P SPAC[S 99
[Hint: Introduce a measure space (A,v), where A= {1,2, ... ,m}, v{i} = 1,
Corollary 3.2.4 Jf ~ = (1<':, 1~~~' 2 , ... ,"-:,n'
1< ... ) E JP and
11 = (TJ 1' 11 2, ... , 11 n~++ ) and apply Minkowski's inequality in the product of the spaces (X,J.!) and
El\ where I :s;p::;; co, 1 < q <co, 1/p + 1/q = 1, then {A,v).]
3.2.9. If j~-'" f in IJ'(X,!<) (that is, if [If. -/II" ...... 0) and if 9n ...... g
in U(X,J.!), where 1 ::;; p::;; oo, 1 ::0:: q::::; oo, 1/p + 1/q = I, then f. g.__,. fg
(3.2.10)
in J.!(X,I').
(3.2.11) Let (X,p) and {X,,O) be two metric spaces. A map a 'om X into X is
called isometric if ;!(O"x,ay) = p(x,y) for all x,y in X. We call a also an
. Holde:'s inequality for p = q = 2 is also known as Schwarz's inequality. imbedding of(X,p) into (X,fi). If a is onto. then we say that O" is an isomorphic
The mequahty (3.2.10) for p = q = 2 is also known as Cauchy'!_i&quality. map and the two metric spaces are said to be isomorphic to each other.
.,......
COMPLETION OF METRIC SPACES; H"'P SPACES 101
100 METRIC SPACES 1
that
Proof. We divide the proof into several steps.
. (a) Two Cauchy sequences {x.} and Cr.} in X will be called equirafent 1
1f p(x,,y,)-> 0 as n-> w. This relation is reflexive, symmetric, and transitive.
p(x,,., ,Xn,kJ $-;;
We can therefore form classes x of Cauchy sequences, each class containing
only mutually equivalent Cauchy sequences, such that sequences of different Consider the sequence
classes are not equivalent to each other. We take the space X to consist of ( 3.3.3)
these classes x, and we define
We claim that this is a Cauchy sequence in X. Indeed, if x:,k. is the class
P(xS) = lim p(x, ,y.) (3.3.1) of the constant sequence of x .. then
v.-e get
by (3.3.4). Since also, for any e > 0,
lim p(x, ,y.) ::-; lim p(x. ,Y.).
n p(x,x:. J = lim p(x j.k;x.,,) < e
j
The reverse inequality is obtained in a similar way.
102 MHRIC SPACES COMPLETION OF METRIC SPACES; H"'P SPACES 103
.\i E X, let {x"} be a seq ucncc in X such that a(x.) --. ~. Since ii is an isometry, jX- }')
{xn} is a Cauchy sequence. Denote by
a correspundcm:e /' by ;(~) =x.
x a class containing it.
Note that if y EX
We then define
and a'(y.) _,.p,
{y"} E y,
h(x) = e- f p\-e- G dy,
then, by Problem 3.LIO, where G is a bounded open set, A c: G c: G c: B, and e > 0 sufficiently small.
Use Problems 2.16.9, 2.16.10.] P( ) h
p(ty) =lim Niixn,GY") =lim p(x.,y.) = (J(x,S'). 3 3 2 We denote by ll'(O) (I 5 P < oo) the space ~ X,fl' :' er.e
n d" t pen set n m R
(X,fl) is the Lebesgue measure space correspon mg o an o .
Thus }' is an isometry. r is also an isomorphy. Indeed, every E 1? con- x Let u e ll'(0:), and define
tains a Cauchy sequence {x"}. The sequence {ax.} is a Cauchy sequence
in X. It we denote by~ its limit in X, then dearly;.'(.~)= This completes x. (J,u)(x) = e- fa r(x ~ y)u(y) dy (e > 0). (3.3.7)
the proof of the theorem.
Theorem 3.3.1 shows that metric spaces can be completed. However, We call J.u a mollifier of u. Prove that
the e Iemen ts of the completed space are more complicated entities than those
(i) J,u is in C'''(R"). hes outside
of the original space. In many instances when this completion needs to be (ii) If u vanishes outside a subset A of n, then J u vanJ s
carried out, the problem arises of giving a simple characterization of the -neighborhood of A (that is, outside the set {x; p(x,A) < e}}.
clements of the completed space. We shall give one important example. an e h (K R" - !1) > () > 0 then
(iii) If K is a closed subset of Q wtt P - '
Let 0: be a bounded open set in R". We denote by cm(n) the set of
functions u(x) that are continuous in n together with all their first m deriva-
tives. The subset consisting of the functions that have a compact support
(J u)(x) = e-
J
]y-xl<<
p(x - Y)u(y) dy =
e
f11< 1
p(z)u(x - ez) dz
3 .3.5. Cij(Q) is dense in I!'(!l). a Cauchy sequence {urn} in the metric (3.3.8) of functions that belong to cm(O),
- 1"1
By Theorem 3.2.3 there exist functions u (0 < "'" -< m ) m
. '"(,....)
u such that J..; where D"(uv), D~u are taken as strong derivatives, and fi :.-;:; tt means that
fi, ,:.-; :t:; for all i.
L Jn JLPu, -
11 "'"' u"']Pdx __,. 0 as m __,. oo. (3. 3.10)
Any two equivalent Cauchy se u
Thus, each class of equivale~t e~C:~ ~ve the same u ~as elements of I!'(Q)].
i:
{u~; 0:::::; jrtj :::;; m} with components bc~):uences grves a unique vector
3.4 COMPLETE METRIC SPACES
Theorem 3.4.1 Let {Fn} be a monotone-decreasing sequence of non-
Lemma 3.3.2 For any rp e C"' (0), empty closed sets, in a complete metric space (X,p). If the sequence of the
diameters d(Fn) converges to zero, then there exists one and only one point
t u({J dx = ( -1)11 I uolP({J dx. (3.3.11) that belongs to
n~t
nF
~
Proof Integration by parts gives This theorem is an extension to metric spaces of the nested sequence
t D"u,. ({J dx = ( -l)ll I u, D"([J dx. theorem on the realli ne, with F n being closed intervals.
Indeed, if uo = 0, then, by (3.3.11) J d (Fm)---+0 if n :?: m----><>C. Thus {x"} is a Cauchy sequence. Denote by x its limit.
Now use Problem 3.3.6. ' n u 'P dx = 0 for all ({J e CQ'(O).
Since each set f" is a closed set that contains all the xn, with m :?: n, we have
by uo.The corollary shows that the u for 0 < jrtj <
- m are de termtne
. d umquely
.
X E F.. Hence X E nFn.
We denote the completion of C'"(Q) . "
by H'""(Q) W h with respect to the metric (3 3 8)
e can t en state :
i ,.r
. ~-
_-
.' COMPACT METRIC SPACES
x, inK such that p(x0,x1);.;:: d0/2. Proceeding by induction, we choose xn+l Let B 1 ,B 2 , ... be the sequence of all the balls B(Ym 1/n') that occur in
in K such that min f!(X1,x. + 1 ) ;.::: d.i2, where d, = sup min p(xi,x). the above way when G varies over the sets E~- Each BJ is contained in some
set E~ The sets {~;} then form an open covering on K We can now apply
1
1 :::;_j.:::;;P1 xEK 1 ::;.j~n
Clearly d0 ;.::: d 1 ;.::: .. . If d. ;.::: .5 > 0 for all n, then no subsequence of {x,} is a
the result of the last paragraph.
Cauchy sequence. This, however, contradicts the assumption that K i~
Combining Theorems 3.5.1-3.5.4, we get:
sequentially compact. Hence d._,. 0 as n ___,.co. But this implies that {x.}
is dense in K-that is, K is separable.
Corollary 3.5.5 A compact subset of a metric space is closed, bounded,
Theorem 3.5.4 A subset of a metric space is compact if and only if' it is
sequentially compact.
r and separable.
I"' '
lx.IP :::; e for all I f(y) = M. Thus/has a maximum on Y. The proof thatfhas a (fimte) mml-
mum is similar.
The following result is usually called the Tietze extension theorem.
XE K. I
3.5.10. The Cartesian product of a countable number of compact I Theorem 3.6.3 Let f be a hounded, real-valued, continuous function on.
metric spaces is a compact metric space. a closed set y of a metric space X. Then there is a continuous function F
defined on X with F(x) = f(x)for x E Y, and
is equivalent to the one given in 2.1. A function f is said to be uniformly For each x EX, Mx(r) is bounded and monotone increasi~g in r. Hence it
o
continuous on Y if for llny E > 0 there is a > 0 such that lf(y) - f(x)l < e is integrable on finite r-intcrvals. We define F by F(x) = flx) 1f x E Y, and
whenever y and x belong to Yllnd p(y,x) <b.
One easily verifies that f is continuous at y 0 if and only if for any 1 H(x)
sequence {x.} in Y such that p(xn,Yu)--+ 0 as n ..... co, /(x") --+f(y 0 ) as n--+ 'X!.
f(x) = ~
o(x)
J O{x)
MAr) dr (3.6.3)
Theorem 3.6.1 Tl f is a continuous function on a compact suhset Y of a if x 1- Y. where (i(x) = p(x, Y). N ole that b(x! > 0 if x ~ Y, since Y is closed.
metric space X, then f is uniformly continuous on Y . We first prove that F is continuous at any pomt y E Y. Smce, for x 1- Y,
Proof. if the assertion is false, then there is an e > 0 and sequences min f ~ F(x):::; max f, where o= p(x,y),
{y.}, {x"} in Y such that Yntl{y.Jol) Y,-,!l(y,Jol)
(3.6.1) and since 3.5 __, 0 as x -+ y (x f$ Y). it follows that, for any e > 0, IF(x)- f(y)l
< if x 1- y and b(x) :::; () 0 , where ,) 0 is sufficiently small. On the_ othe~ ?and,
and p(y.,x,)--+ 0. There exist subsequences {y"J and {x.J that converge to IF(x) - f(y)l = lf(x) - f(y)l < e ifx E Yandp(x,y) :s b 1 , whereo1 tssuffiCJently
points y and x, respectively, belonging to Y. Since small. Thus F(x) is continuous at y. . .
Consider next the continuity of F at a pomt z f Y. Let X be any P?lllt
p(y,x) :::; p(y,y"J + p(.r,,,xnJ + p(x,,,x)--+ 0 as k --+ a:;, in X with p(x,z) < ~p(z,Y). Let h = p(x,z). Then 2h < p(x, Y). Smce
we have p(y,x) = 0--that is, y = x. Hence
112 METRIC SPACES
CO~TINUOUS FUNCTIONS ON METRIC SPACES 113
lb(x) - o(z)j :5" h, Mz(r) :<::: M:.;(r- h), and M.(r);:::: 0, we have
Theorem 3.6.4 Let % be a family of functions, uniformly bounded and
equicontinuous on a compact metric space X. Then any sequence {fn} offunctions
F(x)- F(z) of % has a subsequence that is uniformly convergent in X to a continuous
function.
1 .2~( x) 1 2~(<)
= b(
X
) J6(x) M,Jr) dr- -J
o(z) ~(z)
Mz(r) dr
Proof Let {xm} be a dense sequence in X. Since the sequence {f,(x 1 )}
is bounded, there is a subsequence {f.. 1 (x1)} that is convergent. Next, the
:5" b()
1 (26(x)
A..J_.(r) dr-
1
(
2~(;<)- 2h
M (r _h) dr
sequence {f,,,(Xz)} ~s bounded. Hence it has a subsequence u..
2(Xz)} that
X ~(-') O(X) + h '&(x)+h x is convergent. We proceed in this way step by step. In the kth step, we extract
.
a convergent subsequence u~ (x._)} of the bounded sequence Un.k-I(xd}.
1 26(x)- 3h 1 2~(><) Consider now the diagonal sequence {/,,,} of the double sequence {f...~},
= .<:(x)
-
U
J~(x)
M x (r) dr + -J:( )
u X
J l~(x)- Jh
M "( r) d r and write g, =!.. .. Then {gnCx.)} is convergent for every k, since, except for
the first k terms, it is a subsequence of {f.,1 (x1)}.
1 2~(x)- Jh We shall prove that {g.} is uniformly convergent in X. Since the family
- b(x) + h J.,.,) M_.(s) ds {g,} is equicontinuous, for any e > 0 there is a b > 0 such that
C(X) is then a metric space. It is easily seen that C(X) is also complete. "' ~axIf- g[
1
(or conditionally compact) if its closure K is compact. Prove also that p(f.,,f)--> 0 if and only if Um} converges to f uniformly on
We can now state a criterion for relative compactness of subsets of C(X). compact subsets of Q.
3.6.2. Let Q be an open set in the complex plane. Denote ~y !fCQ)
the space of all the complex analytic functions that arc holom~rph1c m Q.
Theorem 3.6.5 Let X be a compact metric space. A family Jf" of/unctions
Let {Om} be as in Problem 3.6.1, and defi~e p(f,g), for f and g m H(Q), by
ofC(X) is relatively compact in C(X) if and only if Jf" is both uniformly hounded
and equicontfnuous. (3.6.8). Prove that H(Q) is a complete metnc _space. . . .
3.6.3. A family sF of complex analytiC functwns holo~orph1c 1_n Q
that are uniformly bounded on every compact subset of Q 1s a relatively
Proof The lemma of Ascoli-Arzela shows that if Jf" is uniformly compact subset of H(Q). [Hint: Use Cauchy's formula and the lemma of
bounded and equicontinuous, then any sequence of functions {/.} in Jf" has Ascoli-Arzela.] ..
a subsequence that converges in the metric of C(X) to an element of C(X). 3.6.4. A function f that is upper semicontinuous (for defimtton, see
From this it easily follows that the closure :it of JY" is sequentially compact Problem 2.1.11) on a compact metric space is bounded from above and has
and, hence, compact. Thus X is relatively compact. a maximum. .
Suppose conversely that Jf" is relatively compact. By Theorem 3.5.2, 3.6.5. Let z be a set of functions u in I!'(Q) (1 $ p < co), where Q
JY" is a bounded set in C(X). But this is equivalent to Jf" being a uniformly is a bounded open set in W, and extend each u into R" - Q by 0. Prov~ that
bounded family of functions. To prove that the family .Jf" is equicontinuous, z is relatively com pact in I!'( Q) if the following conditions are sat! sfied:
take an e-covering of Jf. by a finite number of balls whose centers arc func- (i) p(J,u,u) =. [[J,u - u[[ l.P(R"l--> 0 as e ..... 0, u~iformly w.ith res?ect ~o u E Z,
tions / 1, .. . ,fh Then, for any /E Jf' there is an jj, with l .,;, j.,;, h, such that an d ('")
u
11 u 11 < C for all u E Z Here J ' u Js . the funct
LP(n) - .
wn defined m (3.3. 7).
k
p(f,jj) < e. That means that [Hint: It suffices to show that Z has a finite b-covenng for any o > 0..Ta c
eo>O such that [[J, 0 u-u[ILP(R"J<M2 for all UEZ. It s~fficcs to find a
lf(x) - //x)l < e finite (b/2)-coveringof W = {1, 0 u; u E Z}. Show that W is uniformly bounded
for all x EX. (3.6.6) and equicontinuous and usc Theorems 3.5.4, 3.5.6, and 3.6.4.]
3.6.6. Let Z be a set of functions u in l!(Q), 1 $ p < oo, where Q
Since each of the functions / 1 , .. . ,fh is uniformly continuous (by Theorem
3.6.1), there is a b > 0 such that is a bounded open set in R", and extend each u into R"- ~ by 0. Prove
that if z is relatively compact, then the conditions (i) and (u) of Problem
3. 6. 5 arc satisfied. .
lfj(y) -.f/z)l < e if p(y,z) <,) (3.6.7) 3.6.7. In every infinite metric space there is an infimte sequence such
that no point of accumulation of 1he sequence is a point of the sequence.
for all y,z in X and for alii 5} _.,;,h. Combining (3.6.6), (3.6.7), we get 3.6.8. If a compact metric space X is such that bounded cl~sed sub~ets
of C(X) are compact, then X consists of a finite number of pomts. [Hmt:
IJ(y) - j{z-)[ < 3e if p(y,z) < b. Use Problem 3.6.7 and the Tietze extension theorem.]
3.6.9. IfF is a Lebesgue-measurable bounded set in R", then .for any
This proves the equicontinuity of the family Jf". E > 0 there exists a closed set G c F such that p( G) > p( f) - c. [ Hm t: Use
Problem 1.9.8.]
'
THE STONE-WEIERSTRASS THEOREM 117
]]6 ~HRTC SPACES
\
2 112
3.6.10. Prove Lusin's theorem: lctf(x) be an a.e. real-valued Lebesgue- Proof One can easily check that the function (te + t) has a
measurable function on a Lebesgue-measurable and bounded set E c R". uniformly convergent Taylor series (about t = t) for 0 :<::; t ~ I. Hence there is
Then, for any h > 0, there is a continuous function 'P(X) in R" such that a polynomial q(t) such that
;t{x;f(x) #- 'P(x)} < 3. Furthermore, if lf(x) I :<::; C a.e., then I'P(x)l:<::; C. E
[Hint: Use Prohlem 2.15.11, Egoroff's theorem, Problem 3.6.9, and the ](}cl + t)112 _ q(t)i < 2: if 0 :<::; t :<::; l.
Tietze extension theorem.]
Now, if -1 :<::;A.:<::; I,
IIA.I - q(A.z)l :-s; IIA.I- ({-62 +A.2)ll21 + 1(!62 + A.2)112- q(A.2)1 <e.
3.7 THE STONE-WEIERSTRASS THEOREM
2
Thus, the assertion follows with p(A.) = q(A. ).
Let X be a compact metric space. In the real metric space C(X) ,,.e
introduce the following operations for any ;;_q in C(X) and}. a real number: Proof of Theorem 3.7.1. We first prove tha_t if fEll, then 1.(1 Ea.
The proof depends upon the follov.ing lemma, which is a special case fiz) <f(z) + e for all z EX.
There is a ball D, with center x such that.f,(z) > f(z)- e for all z E Dx
of the V.'eierstrass theorem. Since X is compact, we can cover it by a finite number of such balls; call them
Lemma 3.7.1 Let }.. be a real rariahfe. For an~y <: > 0 !here exists a Dx 1 , ... ,Dx. The function
polynomial p(f...) such that ii/...1- p().)l <<:for -I $ ). :<::; I.
118 METRIC SPACES A flXED-PO!NT THEOREM AND APPLICATIONS 119
Let
.
(X,p)
.
and (Y,a) be metric spaces. A function .(from a subset X 0 c X :-::;: p(z,x.) + Op(xn-I ,z)-> 0 if n -> CD.
mto Y IS satd to be cant inuous at a point y E X 0 if I~H any e > 0 there is a
b > 0 such. that f~r all x E X 0 with p(x,y) < J, the inequality a[((x),f(y)] < 10 It remains to show that z is the unique fixed point of T. Suppose then
holds. lffts contlnuous at each pointy of X 0 then we say thatfh continuous that y is also a fixed point of T. Then
on X o and if. furthermore, b can be taken to be independent of y, then we say p(y,z) = p(TJ,Tz) :-::;; Op(y,z).
thatfis uniformly continuous on X 0 .
The following result is very useful. Since 8 .P 1, we conclude that p(y,z) = 0-that is, y = z.
120 METRIC SPACES A FIXDPOINT THEOREM AND APPLICATIONS 121
We restate Theorem 3.8.2 in a form that is often more convenient: and jq>(x) - Yol ::5: M/5. (3.8.6)
y is a closed subset of C(1 6). We define a transformation T from Y into
Theorem 3.8.3 Let Y be a closed subset of a complete metric space X C(I 6 ) as follows :
IjT is a map from Y into itself, and ifT is a contraction (on Y), then there exists
a unique point z in Y such that Tz = z. (Trp)(x) =Yo+ r
xo
f(t,rp(t)) dt for x E JJ. (3.8.7)
In applying this theorem one has to verify two facts about T: (i) T Tis well defined, since the points (t,q>(t)) lie in R.
maps Yinto itself, and (ii) Tis a contraction on Y. If y(x) is a fixed point of T, then y(x) f>atisfies (3.8.5) and is thus a
Letf(x,y) be a real-valued function defined on an open set 0 of the solution of (3.8.3). (3.8.4). Conversely, every solution of (3.8.3), (3.8.4) in
Euclidean plane R 2 We say thatf(x,y) is uniformly Lipschitz continuous with I~ is a solution of (3.8.5) and it thus satisfies (3.8.6)-that is, it bel.ongs to Y.
respect toy in !l if there is a constant K such that It is therefore a fixed point of T. Thus what we have to show JS that the
1/(x,y,)- j(x,y2)l :::;; KIYI - Y21 transformation T has a unique fixed point in Y.
(3.8.2)
T maps Yinto itself, since (Tq>)(x 0 ) = Yo and since
for all (x,yd and (x,y 2 ) in !l.
Consider the ordinary differential equation
j(Ttp)(x)- Yo I S: I( f(t,tp(t)) dt IS: M lx- Xol :;<; M/5.
dy
dx = f(x,y) (3.8.3) Next, T is a contraction since
In the space C(I 6) we consider the subset Y consisting of those continuous where A is a complex number. Prove that there exists a small positive ~um
functions q>(t) for which ber }.0 such that, for any complex number J., with ]).1 < }.o, there ex1sts a
122 METRIC SPACES
unique solution u(x) of (3.8.8). [Hint: Consider the operator (Tu)(x) = f(x)
CHAPTER 4
J
+A K(x,y)u(y) dy on L2 (0:).]
3.8.4. Let 0 be the closed unit ball B(O,l) in the Euclidean metric space
(R~,~). Let T be a map ?f n into itself such that p(Tx,Ty)::;; p(x,y) for all
x,y m n. Then there exists at least one fixed point ofT. [Hint: Consider
(1 - 1/k)T.]
ELEMENTS OF
3.8.5. Give an example where n (in R~) is not a unit ball {but is a
. and the assertion of the last problem is false-that is , T has no
closed set) FUNCTIONAL
fi xed pomts.
ANALYSIS IN
BANACH SPACES
We denote by IR the field of the real numbers, and by C the field of the
complex numbers. Let " denote either IR or C. The elements of" are often
called scalars.
(i 1 ) x + (y + z) = (x + y) + z;
(iJ x +y = y + x;
(i 3 ) there is an element 0 (the zero element) such that x + 0 = x for
all x EX; and
(i 4 ) for every x EX there is an element ( -x) (the inverse of x) such
that x + ( -x) = 0.
We shall denote by 0 both the scalar 0 and the zero element of the group
X; this should cause no confusion.
123
124 ELEME~TS OF FUNCTIONAL ANALYSIS I~ BANACH SPACES LINEAR NORMED SPACES J25
From the rdations p.x = p.(x + 0) = p.x + p.O we get p.O = 0. Hence if The vector sum A + B of two sets A and B is the set {a + b; a E A, bE B} .
.-lx = 0 and .-1. "# 0, then We also define A.A = {.-lx; x E A}.
x= lx = (~ })x = ! (Ax) = ! 0 =
). ..1. .-1.
0 Definition 4.1.3 A normed linear space is a linear vector space X on
which there is defined a function x ...... llxll, called a norm, having the fol!owing
-that is, x = 0. The last property is equivalent to any one of the following properties:
properties:
(a) i:xll ~ 0, and llxll = 0 if and only if x = 0;
if }.x = 0 and x # 0, then .-1. = 0; (b) IIJ.xll = 1;.1 !lx!i;
if}# 0, X# 0, then A.x # 0; (c) llx + y'll ::<::,; r.x!l + IIYII-
if J.x = JlX and ;_ # Jl, then x = 0.
Definition 4.1.4 A metric linear spaee is a linear vector space X that is
The reader may easily verify that, in a linear vector space, also a metric space, with metric p, having the following additional property:
Ox= 0, (-l)x= -x. Tf )."-+ ..1., Jln--+ )1, x,--+ X, y"--+ y, then A,Xn + finYn-+ AX + JlY
Definition 4.1.2 A real (complex) linear vector space is a linear vector The last property is equivalent to the statement that the maps (x,y)--+
space with sF = :R (,9" =C). x +yand (A.,x) ...... }.x are continuous maps from the metric spaces X x X
and.' x X, respectively, into X.
Elements x 1 , ... ,x" of X arc called linearly independent if there exist no
numbers } 1 , ... ,}." in. such that Definition 4./.5 A Frechet space is a metric linear space X having the
additional properties:
and I" IJ-j12 > o.
j~l
(4.1.1) (a) p(x,y) = p(x- y,O);
(/3) X is complete.
If there exist numbers ..1. 1 , ... ,}." such that (4.1.1) holds, then we say that
Given a normed linear space we introduce on it a function p by
x 1 , ... ,x" are linearly dependent. We can then express any one of the x,, with
}.J # 0, as a Iinear combination of the others: p(x,y) = llx - Yll- (4.1.2)
Proof Since
Conversely, if p satisfies (4. 1.4), ( 4.1.5), then I x I = p(x,O) is a norm and the
m
space is a normcd linear space. The metric space corresponding to it by
Theorem 4.1.1 is the originally given metric linear spao::e.
llsm- sk\1 :::;; I
n~k+l
llx.ll--+ 0 ifm > k---> oo,
Since (by Theorem 4.1.1) a normed linear space is also a metric space, the sequence {s... ) is a Cauchy ~equence in X. Since X is complete, this
the theory of Chapter 3 applies to it. We recall that x.---> x if llx. - x I ---> 0;
sequence has a limit. . . .
{x.} is a Cauchy sequence if llx. - xm il ---> 0 as n,m--+ oo; if every Cauchy We note that a set A in a normed linear space X ts bounded 1f and only
sequence is convergent, then X is called complete. Note that if x.---> x, then if there is a positive number R sue h that !i x I :::;; R for all x E A
"x.ll---> 11x11. and if {x.} is a Cauchy sequence, then {1\x,J} is a Cauchy
sequence. Definition 4.1.8 Two norms 11 11 1 and i\ 11 2 on a linear vector space
x are said to be equivalent if there exist positive numbers a and f1 such that
Definition 4.1.6 A real (complex) normcd linear space that is complete
is called a real (complex) Banach space. all x lit :::;; I x!lz :::;; fll';x'!lt for all x EX.
This is clearly the case if and only if the metrics corresponding to them by
Note that Banao::h spaces are Frechet spaces. ( 4.1.2) are equivalent in the sense defined in Section 3.1.
Theorem 3.3.1 implies that every normed linear spao::c X can be com-
pleted as a metric spao::c. But actually the completed space X can be made Example 1. The spaces LP( X ,Jl) (I ::s; p :::;; oo) _are linear vector spao::es with
into a linear vector space by defining f + g and If defined in the obvious way-that Is,
(f + g)(x) = f(x) + g(x),
x+J=z, J3 =u (4.1.6)
(A.f)(x) = ..lf(x).
as follov.s: if {x.} Ex and {y.} E J\ then z is the class containing the sequence
{x. + r.} and il is the class containing the sequeno::e {ix"}. :? is a normcd Furthermore, the function/---> ';i/11 given by
space with the norm given by
if 1:::;; p < oo, ( 4.l.7)
I!_\' II= lim l:x.ll
and (4.1.8)
Thus, for every normed linear space X there is a Banach space X such that X is 11!11 = ess sup lfl if p = oo,
isomorphic to a dense subset u(X) of X; the map a from X into :? preserces X
the algebraic operarions of' addition and of' multiplication hy scalars. is a norm. Unless the contrary is explicitly ~tatcd, we shall ~!ways co~slid7er
U(X,Jl) to be the normcd hncar space w1t~ the norm given by ( ),
Definition 4.1.7 Let {x"} be a sequence in a normed linear space X ( 4.1.8). U( X ,Jl) is a Banach space. If the functwns fare ta_ken to be a.e. real-
"' valued, then F(X,p) is a real Banach space. If the functw~s fare taken to
and let s., = I x . If the sequence {sm} has a limit s, then we say that the
be wmplex-valued, then LP(X ,Jl) can be con~idered 10 be etther a real or a
.,, l"i:-1
series I x. is convergent and its sum iss. We then write complex Banach space.
h-1
."!;_.
Example 2. The spaces 1r (I :::;; p ::s; oo) are Banach spaces. Here addition
.Iii= LXII' and multiplio::ation by scalars are defined in the obvious way, and the norm
n= l
is given by
If I
m
'j'i=: 1
I x" '' < oo, then we say that the series
m
I
n= 1
x. is ahsolut e~r converyent. WI = Ltl l~.lp} l/P if 1 :::;; p < oo,
Finally, the condition (c) in Definition 4.1.3 is proved as follows: number of elements of J4 are linearly independent Denote such elements by
x 1 , ... ,xm and let x; E a~, for I ::::; is n. Since the set {<l~J is totally orde~ed,
II Y1 + Y2ll = :q~:tJ
inf llxt + x 2 ll s inf llx 1 ll
x1~Y1
+ x2.:=Y.1
inf llx2JI = IIY1 II + II Y2 ll. one of the subsets Cia, contaim all the others. Denote that set by a"". Smce
X;2.E':t"2
{x 1 , ... ,X"} c: a"", the elements XI, .. . ,Xn are linearly indcpend~nt.
Definition A partially ordered set is a nonempty set S together with a
Applying Zorn's lemma, we conclude that S has a maxtmal element a.
We shall show that every y EX is a finite linear combination of elements
relation " 2:: " subject to the following conditions:
of a. If this is not the case, then the set fJ = {a,y} belongs to S, a : : ; J4,
(i) x s x; but not ::;w sa. This contradicts the maximality of a.
(ii) if x s y and y s z, then x s z.
If for every two elements x,y in Sat least one of two relations x s y, y s x PROBLEMS
holds, then we say that S is a totally ordered set. 4.2.1. Let X be a normed linear space and let Y0 be a closed linear
subspace. Let { Y"} and Y belong to X I Y0 Prove: :: Y" - Y II -> 0 as n -> OCJ if
Definition Let T he a subset of a partially ordered set S. An element
and only if there exists a sequence {y.} with Yn E Y. and a point x_o E Y such
x of S is called an upper bound of T if y s x for all y E T.
that llYn - Xo il ...... 0 as n .... OC!. [Hint: If II Y. - Yl! = em there extst. x. E Y",
x~ E Y such that llxn- x~i! S 2e". Fix x 0 E Y and define Yn = Xo- x.- x . ]
Definition Let S be a partially ordered set. An element x of Sis called
maximal if for any element yES, the relation x s y implies that y s x. 4.2.2. Let X be a Banach space and let Y0 be a dosed linear subspace.
Prove that X 1Y0 is a Banach space. .
The following result is known as Zorn's lemma. 4.23. Let X be a normed linear space and let Y0 be a closed hnear
subspace that is complete. If X/Y0 is a Banach space, then X is a Banach
Theorem 4.2.2 If S is a partially ordered set in which et>ery totally space.
ordered subset has an upper bound, then S has a maximal element. 4.2.4. Show that c is a closed linear subspace of/"" and that it is no-
where dense in /~.
This theorem is equivalent to both the well-ordering theorem and the 4.2.5. Is C[a,h] a dosed linear subspace of U(a,b) (I ::::; p S oo)? .
axiom of choice. We shall not give here the proof of the theorem. 4.2.6. If a linear vector space is infinite-dimensional, then there cxtst
We shall use Zorn's lemma in the proof of the next theorem. on it norms that are not equivalent. [Hint: Let {y.} be a Hamel hasis and d~fi.nc
norms by llx11 2 =I ca1X.1 2 , where x has the form (4.2.2) and Ca are postttve
Theorem 4.2.3 Every linear vector space X contains a set Ci of linearly
independent elements such that the linear subspace spanned by coincides a numbers.]
with X. 4.2.7. If a normed linear space has a countable Hamel hasis, then it is
separable.
132 ELEMENTS OF FUNCTIONAL ANALYSIS IN BANACH SPACES
FINlTE-DIMENSIONAL NORMED LINEAR SPACES 133
that tends to oo. For simplicity we denote this sequence again by {Am}. Since
4.3 FINITE-DIMENSIONAL NORMED {y,} is a convergent sequence, it is also a bounded set. Hence
LINEAR SPACES
IIY,I1_. 0 as m oo. (4.3.3)
. V.:e shall characterize in this section the normcd linear spaces of finite --+
(;.~m)
Lemma 4.3.1 Let Y be a closed proper linear subspace of a normed
linear space X. Then, /or any t > 0, there exists a point z EX, ;lz'' = I, such
z, = ,... , ..1,~"'))
A
A, "m
that liz- yll > I - efor alt y E Y.
on the unit sphere in n dimensions has a convergent subsequence {Zm}.
Thus,
Proof Take a point x 0 EX, x 0 Y and define n
Proof. Let {y'"}_ c: Y, Ym--+ y as m--> oo. We have to show that y E Y. Proof. Suppose first that X is finite-dimensional. We have to show
We fix a basts e 1 , .. . ,e" m Y and write every clement y'" in the form that every bounded sequence has a convergent.subsequence. Let {y.,} be a
bounded sequence. Choose a ba~is e1 , ,e. in X and represent the Ym in the
(4.3.1) form (4.3.1). The arguments used in the proof of Theorem 4.3.2 show that
there is a constant C such that (4.3.2) holds. But, again by the arguments used
We claim that there is a constant C such that in the proof of Theorem 4.3.2, there is a subsequence {m'} of the sequence
n of natural numbers such that
:L
i= 1
14~'" 1 1
2
sc for all m. ( 4.3.2) (..1.\"'"\ ... ,).~'"'1) ...... (v 1 , ... ,v") as m'--+ 00.
Indeed, otherwise, the sequence of numbers Am= I IJ.)'"ll 2 has a subsequence Hence, y., ....... I v;e; as m' ...... oo. This completes the proof of one part of the
i theorem.
] 34 ELEMENTS OF FUNCTIONAL !I."''AL YSJS II\ BA"''ACH SPACES
LINEAR TRANSFORMATIONS 135
Suppose now that X has the property that bounded subsets are rela-
tively compact, and suppose that X is infinite-dimensional. We shall derive a 4.4 LINEAR TRANSFORMATIONS
contradiction. First we take a point x 1 such that llx 1:: = l. We now apply
Lemma 4.3.1 with Y being the linear space spanned by x 1 . We conclude that Let X and y be two linear vector spaces, and let T be a function from a
j T
there exists a point Xz linearly independent of x 1 such that llx 2 1! = 1, subset Dr of X into Y. We call T also an operator, a mapping. or a transforma-
<ix 2 -x1 :: > J. tion. The set Dr where Tis defined is called the domain ofT. If Y = Tx. then
Next we apply Lemma 4.3.1 with Y the linear space spanned by x 1 and we call y the image of x and we call x an inverse image of y. For any subset A
x 2 We conclude that there exists a point x 3 linearly independent of x 1 ,x 2 of Dr we write T(A) = {Tx; x E A}. The set T(D 1 ) is called the range _of T.
such that 'f We shall often use the notation T(x) = Tx. Tis called a linear operator If Dr
~xJII = 1, 11x3 -xJ > {-. ! is a linear subspace of X and if
Proceeding in this way by induction, we get a sequence {xn} such that T(A. 1 x 1 + }. 2 x 2 ) = J.. 1 Tx 1 + A. 2 Tx 2
llx.ll = 1, ifn>j. for all x ,x in Dr and }.1 .}.. 2 scalars. It follows that T(O) = T(2 0)
1 2
= 2T(O),
so that T(O) = 0.
Note that at each step, the linear space Y that is used is finite-dimensional Unless the contrary is explicitly stated, we shall always assume that
and hence (by Theorem 4.3.2) closed. Thus Lemma 4.3.1 can in fact be Dr= X -that is, Tis defined on the whole space X. If T(X) = Y, then we
employed. Since the sequence {x.} is bounded, and since it obviously has no say that T maps X onto Y. If Tx 1 = Tx 2 implies x 1 = x::, then we s~y that T
convergent subsequences, we ha vc arrived at a contradiction. is a one-to-one map. When Tis linear, Tis one-to-one 1f and only If Tx = 0
implies x = 0.
PROBLEMS Let X and y be metric linear spaces. A continuous linear transfor-
4.3.1. Let X be a finite-dimebsionallinear space. Then any two norms mation T from X into Y is a linear transformation that is also continuous-
on X are equivalent. (According to Problem 4.2.6, the assertion is false if X that is, if x.---> x, then Tx.-+ Tx.
is infinite-dimensional.)
4.3.2. Let Y be a finite-dimensional linear subspace of a normed Theorem 4.4.1 Let X and Y be normed linear spaces. A linear trans-
linear space X, and let x 0 EX, x 0 ~ Y Then there exists a point y 0 E Y such formation T from X into Y is continuous if and only if it is continuous at one
that point.
inf ilx 0 - Yll = llxo- Yol:.
yEl' Proof. Let The continuous at a point z. If x.-+ x, then x.- x + z---> z.
4.3.3. A norm II II is said to be strictly convex if !! x !' = 1, llyll = 1, Hence T(x"- x + z)-+ Tz, so that Tx.- Tx + Tz---> Tz. Thus Tx.-+ Tx.
II x + y II = 2 imply that x = y. Prove that if the norm of X is strictly convex, Let X and Y be normed linear spaces, and let T be a linear map from
then the point Yo occurring in the assertion of Problem 4.3.2 is unique. X into Y. If there is a constant K such that, for all x E X,
4.3.4. Prove that the norm of LP(X,JJ.) is strictly convex if 1 < p < oo, IITxii-::;KII~II,
( 4.4.1)
and is not strictly convex if p = 1 or if p = oo.
4.3.5. Prove that in C[a,b] the uniform norm is not equivalent to then we say that Tis bounded and we call T a bounded linear map: No_te that
the LP norm (for 1 -::; p < oo ). the norms on the different, sides of the last inequality are taken m different
4.3.6. Let 0 be a bounded open set in R" and let I -::; q < r-::; p < oo. ' ' _j .
spaces. , -v>"- ; . .
(
} fU I0 1,I~ i
_ .
'
',
.
Prove that in the linear vector space U(O) the norms of U(O.) and of The g.l.b. of all the constants K for wh1ch (4.4.1) Is satisfied IS denot_ed
L'(O) are not equivalent. by IITII and is called the norm of T. It is obvious that (4.4.1) holds with
4.3.7. Let n be a positive integer, I -::; p < cc, and let f(x) be a con- K =II Til-that is,
(4.4.2)
tinuous function on 0 -::; x -::; I. Then there exists a unique polynomial Q" of IITxll -::; IITII llxll-
degree n such that for any other polynomial P" of degree n We also have {.(~ h \
) IITxll ( 4.4.3)
1 .I
II Til= l.u.b.- = l.u.b. IITxll-
( lf(x)- P"(x)IP dx > j lf(x)- Q"(x)IP dx. "*o llx11 - llxll ~ 1
o o
LINEAR TRANSFORMATIONS 137
136 ELEMENTS OF FUNCTIONAL ANALYSIS IN BANACH SPACES
ll.tTII = sup ll(..lT)xll = 1..1.1 sup IITxll = IA.IIITIJ, Example t. Let X be a compact metric space. Then C(X) is a Banach algebra
llx!l~l llxll-1
with (fg)(x) =f(x)g(x).
liT+ Sll = sup II(T + S)xll s: sup iiTxll + sup IISxll =II Til+ IISII.
llxll=l llxll-1 llxii=I Example 2. If X is a Banach space, then &l(X) is a Banach algebra with
(ST)(x) = S(Tx). Indeed,
Definition 4.4.1 A sequence {T"} of bounded linear operators, from
a normed linear space X into a normed linear space Y, is said to be uniformly IISTII = sup II(ST)xll = sup IIS(Tx)ll:::; !!S'.'. su~ !ITxl!, = iiSii liT'.!
convergent if there exists a bounded linear operator T from X into Y such ll:xll ~ 1 l;x'.'. ~ 1 llx'l-1
that I TM - Til ---> 0 as n ___,. oo. We then say that {T.} is uniformly convergent
YI(X) has an identity/, given by /x. = x. . .
toT. If x = R", ::ll'(X) can be identified wrth the space of 11 x n matnces.
Note that uniform convergence means convergence in the norm of
.'B(X, Y).
PROBLEMS
Theorem 4.4.4 If X is a normed linear space and Y is a Banach space, 4.4.1. Let X be a normed linear space and let Y be a Ban~ch space.
then 1/(X,Y) is a Banach space. Let T be a linear operator from X into Y with domain Dr dense Ill X. If T
THE PRINCIPL OF UNIFOR~ BOUNDED NESS 139
138 ELEMENTS OF FUNCTIONAL ANALYSIS IN BANACH SPACES
is bounde~-that is, if (4.4.1) holds for all x E D:r-then there exists a unique
bounded hnear operator Tfrom the whole space X into Y such that fx = 4.5 THE PRINCIPLE OF UNIFORM
Tx for all x E Dr. We call T the continuous extension of T. BOUNDEDNESS
4.4.2. let The an additive operator [that is, T(x 1 + x 2 ) = Tx 1 + Tx 2 J The following important theorem is known as the Banach-Steinhaus
from a real normed linear space X into a normed linear space Y. If T is con- I -~
theorem and also as the principle of uniform boundedness.
tinuous, then T is homogeneous [that is, T().x) = A.Tx]. [Hint: Prove that
T[(m/n)x] = (mfn)Tx, where m,n are integers.] Theorem 4.5.1 Let X be a Banach space and let Y be a normed linear
. ~ 4.~.3. Let A be a symmetric n x n matrix. Consider A as an operator
space. Lei {T.} be a family of bounded linear operators from X into Y. If for
m R g1ven by x ..... Ax. Prove that IIAII =max IA.il, where }.j are the eigen-
j each x EX the set {T.x} is bounded, then the set {liT. II} is bounded.
values of A.
Proof If there exists a ball B(x 0 ,e) on which {T.x} are uniformly
4.4.4. let f(z) = I"' anz" be an entire complex analytic function.
boundr:;d-that is, for some constant K,
n=O
IS absolutely convergent in "'(X). One definesf(T) by -then we can easily complete the proof of the theorem. Indeed, for any
y - 0 define
s
f(T) = I"" a. P. z= lfYjf Y + X 0 .
n=O
4.4.5. Let f(z), g(z) be entire complex analytic functions, and let Since liz- x 0 ll $ e, (4.5.1) implies that IIToz!! $ K. Hence
h(z) = f(z)g(z). Prove that for any TE @(X), X a Banach space, h(T) =
f(T)g(T). In particular we have: euel'r = e(J.+pJr_ _c_IIT v\l- liT x 0 ll
!! y II "~ '
$11-/:--
li Y II
T.y + T,x 0 ll
I
= IIT.z',i $ K.
4.4.6. Find the norm of the operator A E "'(X) given by (Af)(t) = tf(t)
(0 :$ t :$ 1), whe~e (a) X= C[O,I], (b) X= U(O, I) and (I ::-; p :$ cx:;). . We thus ge:t
4.4.7. A lmear operator from a normed linear space X into a normed K + IIT.xoll K + K' ...
linear space Y is bounded if and only if it maps bounded sets onto bounded IIT.rll =:; ll_r!l :S - - tr:l,
s c
sets.
4.4.8. A linear operator from a normed linear space X into a normed where K' =sup .T.x 0 1l < oc. We conclude that IIT.!I $ (K + K')/e, and the
a
linear space Y is continuous if and only if it maps sequences wnverging to 0
assertion of the theorem fol!ows.
into bounded sequences. To complete the proof of the theorem we only have to show that there
4.4.9. Let (X,p) and (Y,v) be measure spaces and consider the operator exists a baH on which the {T,x} are uniformly bounded. We shall wppose
A defined by that no such ba11 exists, and derive: a contradiction.
Fix a ball B 0 . Then there exists a point x 1 in B 0 such that
(Af)(x) = JK(x,y)f(y) dv(y) IIT.x 1 11 > I
for some index 'l 1 . By continuity we have II r.,x" > 1 in some ball B(x~>el)
from L'(Y,r) into U(X,p), where 1/p + 1/r =I, I <p < CJJ. Prove that if
lying in 8 . We take c1 < l. ln this new ball the family {T,x) is not uniformly
K(x,y) E U(X x Y,p x v), then A is a bounded operator. 0
4.4.10. Prove that L 1(R") is a Banach algebra if multiplication is hounded. Hence there exists a point x 2 E B(x 1 ,e 1 ) such that
defined by convolution--that is, :;r.,x) > 2
for some index a 2 , a} 1= a 1. By continuity, 1.::r.,xl\ > 2 in some ball B(xl,c2)
(f g)(x) = JRn f(x- y)g(y) dy. lying in B(x ,e ). We take e2 < 1- We can proceed in this way to define
1 1
THE OPEN-MAPPING THEOREM AND THE CL.OSED-GRAPH THEOREM 141
140 ELEMENTS OF FCNCTIOi'OAL ANALYSIS IN BA'\ACH SPACES
ame'"'x, where a, = -2
1! 0
r f(~)e-''"~ d~.
m=-::t::
Prooj: For any x EX the sequence {T.x} is hounded, sim:e it is con- from X into Y is hounded, and
vergent. The principle of uniform houndcdness implies that II T" !I :::::; K for all 1 .2n
n ~ 1, where K is a constant. Hence, II T.ll = - J
2n o
IDn(x)l dx.
forallxEX. (4.5.2)
4.5.5. There exists a continuotts function f(x) on 0 ::0:: x $ 211, with
Now define Tx = li~n T.x. [t is clear that Tis a linear operator. Fro in (4.5.2) f(O) = f( 2n:), such that its Fourier series diverges at x = 0. [Hint: Use
Problems 4.5.2, 4.5.4, and the Banach-Steinhaus theorem.]
it_ also follows that ;iTx, ::0:: Kl!xl!. Thus Tis a bounded linear operator.
Fmally, by the definition of T it follows that {T"} is strongly convergent
toT. THE OPEN-MAPPING THEOREM
4.6
AND THE CLOSED-GRAPH THEOREM
PROBLEMS
4.5. I. Let X, Y, T, he as in Theorem 4.5.1. Let The following theorem is called the open-mappin_q theorem.
.
X.= {x; sup IIT,xll ::0:: n!lxii}. Theorem 4.6.1 Let X and Y be Banach spaces and let T be a bounded
linear map from X onto Y. Then T maps open sets of X onto open sets ofY.
Then X = l) X n L' se 1he Baire category theorem to give a (somewhat)
,.2n I sin (n + nx I with center 0 and radius. We shall prove that for any e > 0 there 1s a J such
I I
o I
-,
SlllzX
ldx----> X asn---->oo.
that
( 4.6.1)
[flint: lsin lxl ::0:: lx.]
142 ELEMENTS OF FUNCTIONAL ANALYSIS!~ BA"'ACH SPACES
THE OPEN-MAPPING THEOREM AND THE CLOSED-GRAPH THEOREM 143
Corollary 4.6.3 Let X be a Banach space with any one of two norms
(4.6.4)
1 1 , ilb Assume that there is a constant K such that
The series L x. is absolutely convergent
"
~incc , x I <
I " _ e. an
d \
L.. c. < oo. By for all x EX. (4.6.5)
!!xll s; L\ ,',.-x.ll s; L\ ~,." < 2oc-o Proof. Denote by X 1 and X 2 the Banach spaces obtained by providing
n=O n=O X with the norms ','1 1 1 and I 1i 2 , respectively. Consider the map Tx = x
Since T is continuous, and since ti ----.0 as n----. w (4 61 4) - . from X 1 onto X 1 . By (4.6.5), Tis bounded. Hence, by Theorem 4.6.2, T- 1 is
h h " - - gl\es y = Tx We also bounded. This gives (4.6.6).
ave t us shown that for any y E }Jo there exists a point x EX such. that
TX= Y Th1s proves (4.6.2). 2
'"
Definition 4.6.1 Let T be a linear operator from a linear vector space
(c) We shall prove that for any open ~el c 1 r X' d r
_,_T_ - l an tOr anv pomt X into a linear vector space Y, having a domain Dr- The graph GrofT is the
J-- x, x E G, there IS a ball Y, (IJ > 0) ~uch that v + y c T(C) Th- "II
show that T(u) is open, and thus establish the assertion of"the th' . ISS_wl set or all points (x, Tx) in X x Y, where x varies in Dr- If Gr is a dosed set
eorem. mce in the Cartesian product X x Y. then we say that Tis a closed operator.
144 ELEMENTS OF FUNCTIONAL ANALYSIS IN BANACH SPACES
APPLICATIONS TO PARTIAL DIFFERENTIAL EQUATIONS 145
Note that Gr is a linear subspace in X x Y. Thus T is closed if and 4.6. 6. Let X and Y be Banach spaces and let T be a linear map from
only if Gr is a closed linear subspace. Note also that Tis closed if and only if a linear subspace Dr of X into Y. If Dr (in X) and the graph ofT (in X X Y)
Xn E Dr, Xn---> X, Txn---> y imply that x E Dr and Tx = y.
are closed, then T is bounded-that is, IITxll ~ Kllxll for all x e Dr (K
The following theorem is known as the closed-graph theorem. constant).
4.6.7. Let X be a normed linear space with any one of two norms
Theorem 4.6.4 Let X and Y he Banach spaces and Jet T be a linear 11 1 , II II~ If II Xn11 2 ---> 0 implies II x.";i 1 ---> 0, then (4.6.5) holds.
operator from X into Y (with Dr =X). IfT is closed, then it is continuous.
ll(x,Tx)ll = llr 1 x[[ ~ Kl!xll (for all x E X) j Definition 4.7.1 Let T be a linear operator from a linear vector space
X into a linear vector space Y, having domain Dr. A linear operator S from
for some constant K. Consequently, llxll
bounded.
+ IITxil ~ Kllxll, and Tis therefore
I X into Y is called an extension ofT if Ds :::1 Dr and Tx = Sx for all x E Dr.
PROBLEMS
Definition 4.7.2 LetT be a linear operator from a normed linear space
If T,S,T- \S-
1
4.6.1. belong to !?B(X), then (TSr 1 E !?B(X) and X into a normed linear space Y with domain Dr. Suppose there exists an
crs)- 1 = s-~r- 1 . operator S with the following properties: (i) S is a closed linear operator;
4.6.2. Let X be a Banach space and let A e .Jd(X), II A II < I. Prove (ii) S is an extension of T, and (iii) if S' is any operator with the properties
that (/ + A) -J exists and is given by (i), (ii), then S' is an extension of S. Then we call S the closure of T. When
the closure of T exists, we often denote it by f.
(/+A)-1 =I "' (~ltA", The following theorem gives a necessary and sufficient condition forT
,. .. o to have a closure.
where the series is absolutely convergent [in .'I( X)]. Show also that
IIU+A)- 1 11~ 1/{1-IIAII). Theorem 4.7.1 Let T be a linear operator from a linear subspace DT
4.6.3. Let X be a Banach space and let T and r- belong to .:B(X). 1 of a Banach space X into a Banach space Y. T has a closure T if and only if the
Prove that if S E Ji'(X) and II S- Til < 1/11 r- 1 11, then S -J exists and is a following condition is satisfied:
bounded operator, and imply (4.7.1)
Let z. E Dr, z"---> z, Tz.---> u for some u E Y. Then, for each n we can Ask___,. XJ, the integrals on the right converge to zero, whereas the integrals
find a point x. in D-r such that on the left converge to Jn v r.p dx. Hence
- 1
IITz.- Tx.!l < -.
n
J v <p dx = 0
ll
We then have: x. E Dr, x" ..... z, Tx.---> u. Hence z E Dr and Tz = u. Thus Tis for all r.p E C~(O). By the result of Problem 3.3.6 we conclude that u = 0.
indeed a closed operator. In the remaining parl of this section we shall consider the differential
Let P(z) = P(z 1 , ,z.) be a polynomial inn variables z1 , .. ,z. of degree
operator P(iD) (where i = \ ../~-1) rather than P(D).
m, say
P(z) -- '\
L. aa1o:P1 z"' z~"
1 1'1
E.:.:J5m Definition The polynomial P(z) [or the operator l'(iD)] is hypoelliptic
We shal! use th_e notation: ::x = (o:1 , .. ,o:"), lo:l = o: 1 + ... + o:., z~ = z~' ... z~". if every function u E C"'(O:) that satisfies P(iD)u = 0 inn is in C''(O:).
Then we can wnte There are necessarv and sufficient conditions for P(z) to be hypoelliptic.
P(z) = L a~z~. These conditions are independent of 0:. Denote by .-\'(P} the set of complex
l~l.:<;m zeros of the polynomial P(z). We can then state:
We write Di = c;'i!xi, D~ = D~' D:", and define
P(D) = L a~D~. (4.7.2) Theorem 4.7.3 P{z} is hypoelliptic if and only if any one of the following
1~1sm conditions holds:
We call P( D) a parti~l differential operator with constant coefficients of degree (i) For any constant A > 0 there exists a constant B > 0 such that all
~- P(D) can be considered as an operator from C"'(O:) into itself, where 0:
the points s =a + ir of N(P) satisfy the inequality
IS any open subset of R". It maps a function u(x) into the function
L a~D~u(x).
[P(D)u](x) = {4.7.3) lrl 2 A log lal - B.
'~',;'" (ii) If Ia I ---> oo, a + ir E N(P}, then lrl ..... w.
We can also consider it as an operator from Ck{O:) into ck-"'(0:) for an
k ~ m. However, if k < m or if Ck(O:) is replaced by the space U(O), the~ (iii) For any 9 ERn,
:(D)u cannot be defined on the whole space. Tn some important problems it P(O" + 0)-->
IS necessary to consider P( D) in spaces LP(Q) (I :-::;: p < w ). What we then do
1 as a E R", lal --> x.
P(a)
is de~ne P_(D) !sin (4.7.3) on the linear subspace cm{Q) consisting of all the
functwns m C (.0.) that have a finite norm (iv) For any il., lctl 2 0,
D"P(a) ___,. O
(4.7.4) ds O" E R", IO"I ---> oo.
P(rr)
Tt is often desirabl~ t_o exte~d P(D) to a closed linear operator in L~'(O:). We
shall prove that thrs rs possible. The proof of this theorem is beyond the scope of this book. However,
one part of it, namely, condition (ii) as a necessary condition for hypoellip-
Theorem 4.7.2 The operator P(D) from P(O) (1 :-::;: p <co) into itself, ticity, will be derived here as an application of the dosed-graph theorem. In
with domain (m(n), has a closure. fact, we shall prove the following stronger theorem.
Proof ln view of,Theorem 4.7.1 it suffices to show that ifuk E Cm(O:), Theorem 4.7.4 Assume that there exists a domain Q in Rn such that
- Vii o.p--> 0, then v = 0. Let r.p E cg (0:). r ntcgration by
I uk I O.p ____,. 0, I P( D)uk all the solutions of P(iD)u = 0 in Q that belong to C 11 (0:) (for some h ~ m)
parts gives belong also to C 11 + 1(0:). Then the condition (ii) of Theorem 4. 7.3 holds.
J P(D)u< r.p dx
Q
= J'Q
u, P( -D)<p dx. Proof For simplicity we may assume that 0 En. Denote by X the
148 F I.E MEI'TS OF FU ..... CTIONA L A1\ALYSJS IN BAN ACJJ SPACES APPLICATIONS TO PARTIAL DIFFERENTIAL EQUATIONS 149
space of all the functions of Ch(f!) that arc solutions in fl of P(i D)u = 0 and Tf u and v are in U(O) (I -::;, p < oo) and if for any compact subset K of n
that have a finite norm there exists exists a sequence {tpi} in ciI(Q) such that
(4.7.5) f ID"tp
'K
1 -vi~' dx----> 0 asj ...... oo, (4.7.8)
X is a Banach space. Let fl 0 be a bounded domain containing 0 and whose then we say that r is the ath LP strong derivatil'e of u and write Du = 1' (s.d.).
closure lieS in fl, and denote by Y the Space of functiOnS in Ch+ 1(Q 0 ), having For example, the functions u(x) of the space H'"-P(Q) introduced in 3.3
a finite norm have LP strong derivatives of any order -::;, m, and all these derivatives belong
llull' = sup L ID'u(x)j. to U(O).
xE:n(ll'l:sn~l
Y is abo a Banach space. We shall denote by u' the rcstril:tion of u (in Q) to PROBLEMS
fl 0 . By assumption, if u EX, then u' E Y. Consider now the map Tu = u'
4.7.1. If D'u = v (s.d.), then D"u = v (w.d.).
from X into Y. Tis closed. Indeed, if
4.7.2. If D'u = v (w.d.) and D 11 v = w (w.d.), then DP+au = w (w.d.).
4. 7.3. If II and v belong to u locally in n (I :::; p < 00 ), and if D"u = v
(w.d.), then also D"u = v (s.d.). [Hint: Take tp 1 = J,;(u), e.i----> 0, and use
then clearly r = u' _, Tu. The closed-graph theorem l:an now be used. It Problems 3.3.2, 3.3.4 and the relation D"J.u = J,v.]
implies that T is a bounded map-that is. 4.7.4. Denote the closure of P(D) in U(O:) again by P(D) and denote
its domain by Dp. Prove that if the polynomial P(z) has degree m, then
ilu!l' = 'u' ':0: K u (K constant). (4.7.6) H"'P(Q) c Dp-
4.7.5. Prove that H'"~'(O:) = (] DIY>-
If s E N(P), then u(x} = e-j~- belongs to X. Hence, by (4.7.6), 11 ,lm
4.7.6. Prove that P(D) cannot be extended into a bounded linear
lsi sup ex :0: K' sup exp [ -lxl 2 +x r], operator from U( fl) into itself.
_"'( f:ilo xeO
4.7.7. Introduce in C''(f!), n a bounded open set in R", the norms
where K' is a constant. Since 0 E fl 0 , the term on the left is ;::.: lsi- Since lui'"= sup L ]fful. Denote by C*(Q) the subset of C''(!l) consisting of
n o :;;11 ,;'"
, lrl 2 all functions with lulm < oo for all m :2: 1. Introduce in C*(O) a metric
-lxl 2 + x r :0: -lxl- + lxllrl :0:
4 , p(u,v) = p(u - v,O) by
we get ~ I lui .. I
p(u,O) = ,1;1 2'" 1 + lulm.
[I~
2
I 11D'rp dx = (- 1)11 J-
vrp dx, (4.7.7)
(b) (z~ + zi)(z~ + 1);
-n k
(c) z"+ L aj(z!+ .. -+z;_y (a i constants).
then we say that r is the lth weak deriratire of u and write D'u = r (..d.). j~O
150 ELEMENTS OF FUt-;CTIONAL ANALYSIS !!'>l BA'SACII SPACES
THE HAHN-BANACH THEOREM 151
To choose c that satisfies (4. 8. 1), we note that for any two points
4.8 THE HAHN-BANACH THEOREM
x,y in Y0 ,
Let X be a linear vector space. An operator (a linear operator) from X go(Y)- go(X) = go(Y- x) :$ p(y- X):$ p(y + Yr) + p(- Yt - x).
into the space :R is called a real (linear) functional on X. Next let X be a
normed linear spa<:c. A bounded linear operator from X into iRis called a real Hence
continuous linear functional (on X). Similarly we define a complex (linear) -p(- y 1 - x)- go(x) :$ p(y + Yl)- go(y).
functional (on X) and a complex continuous linear functional (on X) when X This implies that
is a complex space and iRis replaced by C. We denote by X* either one of the
spaces M(X,:R), ::2(X,C) of continuous linear fun<:tionals on X, and call it the A= sup {- p(- y 1 - x)- g 0 (x)} :$ inf {p(y + Yt)- Bo(Y)} =B.
xeYo y~Yo
conjugate (or dual) space. By Theorem 4.4.4, X* is a Banach spa<:e.
The following theorem is one of the most important results in functional Let c be any number satisfying: A:$ c :$B. Then
analysis. It will enable us to show that the space X* is ' sufficiently rich.'' It is c :$ p(y + y 1) - g 0 (y) for all y E Y0 , ( 4.8.2)
called the Hahn-Banadt lemma.
- p(- y 1 - y) - g 0 (y} :<:;; c for all y E Y0 . (4.8.3)
Theorem 4.8.1 Let X he a real linear vector .;pace and let p be a real
Multiplying both sides of(4.8.2) by A.,}> 0, and replacingy by y/}, we obtain
functional on X, satisfying
.tc :$ p(y + }.y 1) - g 0 (y). (4.8.4)
p(x + y) :$ p(x) + p(y), p(..lx) = }p(x) for afl / :?: 0, x E X, y E X.
Multiplying both sides of (4.8.3) by .A., .t < 0, and replacingy by y/1., we again
Let f he a real linear functional on a linear subspace Y of X, satisfying
obtain (4.8.4), after using the homogeneity of p. Note that (4.8.4) holds also
f(x) $ p(x) }Or all x E Y. if).= 0. Since (4.8.4) implies (4.8.1), the proof of the theorem is complete.
The following theorem, whose proof is based on Theorem 4.8.1, is
Then there exists a real linear functional F on X such that
called the Hahn-Banach theorem.
F(x) = f(x) for all x E Y, F(x) :$ p(x) for afl x EX.
Proof Denote by .;(" the set of all pairs ( Y~,g~), where Ya is a linear Theorem 4.8.2 Let X be a normed linear space and let Y be a linear
subspace of X that contains Y, and 9a is a real linear functional on Y~ satisfying subspace. Then to every ~* E Y * there corresponds an ,x,* EX* such that
We define a relation"$" by (Y~,gJ $ (}p.gp) if Y, c Yp and ga = gp on Y. Proof. If X is a real space, then the assertion follows from the Hahn-
Then .%' becomes a partially ordered set. Every totally ordered subset =
Banach lemma withp(x) = i\y*ll iixll..f(x) y*(x) and x* =F. To prove that
{( Y11 ,gP)} clearly has an u ppcr bound ( Y' ,g') given by: Y' = U Yp, g' = g fi lix*ll $ lly*ll, write for any x EX, x*(x) = 81x*(x)i, where 0 = I. Then
on Yp. Hence, by Zorn's lemma (Theorem 4.2.2), there is a maximal clement
( Y0 ,g 0 ). If we can show that Y0 = X, then the proof of the theorem is complete
lx*(x)l = tlx*(x) = x*(8x) :<:;; p(8x) = lly*ll 118x = lly*ll llxll-
11
,
(with F = g 0 ). This shows that II x* II :<:;; IIY* 11. The reverse inequality is obvious.
We shall assume that Y0 #- X and derive a contradiction. Let Y1 EX, Suppose now that X is a complex normed linear space and define real
Y1 Y0 , and consider the linear manifold Y1 spanned by Y0 and y 1 Each point functlonalsf1 ,f} over Y by
X in Yr has the form X= y + AYr. where y E Yo and ;, E IR are uniquely
determined (for otherwise we get J\ E Y0). Define a linear functional g 1 on
y*(y) =ft(y) + !MYl
Y1 by g 1(y + ;_Y1) = g 0 (y) + ),c. If we can choose the constant c such that From the relations
Bo(Y) + ;_c :$ p(y + }.yl) (4.8.1) y*(Xy + A'y') = ft(}.y + }'y') + if2(J..y + A'y'),
for all }. E IR, y E Y0 , then ( Y1,g 1) E ,;f" and (Y0 ,g0 ) :$ ( Y1,g 1), Y0 #- Y1 . This y*(.ty + }'y') = }y*(y) + }'y*(y') = }[.f,(y) + i.fl(y)] + A'[.f,(y') + ifiy')]
contradicts the maximality of ( Y0 ,g 0 ).
it follows that if .l.-, )_' E IR, thenf1(.l.-y + A'y') = A/1(y) + A'f1(y'). We a[so have
THE HAHN-BA~ACH THEOREM 153
152 ELEMENTS OF FUNCTIONAL ANALYSIS IN BANACH SPACES
1/l(y)l ~ ly*{y)l ~ lly*ll llyll. Thus, regarding X as a real normed linear Thus \',z'!! = Ijd. Applying Theorem 4.8.2. withy* replaced by z*, we get the
space, w_e can appl~ the result of Theorem 4.8.2 for X real and conclude that assertion of the theorem.
there eXIsts a real hnear functional F 1 on X such that
Corollary 4.8.4 Let X be a normed linear space. For any x #- 0 there
for ally e Y. (4.8.5) exists an x* eX* with llx':i =I, x*(x) = 11 x.',.
Now define x* on the complex linear space X by
Indeed, by Theorem 4.8.3 with Y = {0}, there exists a z* r= X* such that
x*(x) = F1 (x) - iF1(ix). liz* I = 1/llxll, z'"(x) = I. Take x* = llxllz*.
x* .is clearly a~ditive, and x*().::) ~ .h*(x) if . 1. is real. Also, x*(ix) = F 1 (ix)
Corollary 4.8 .5 If y and z are distinc 1 points in a normed linear space X,
- '.Ft( -x) = 1x*(x). Hence x* IS lmear. To prove that x'" is an extension of
y* 1t suffices to show that then there exists an x* E X* such that x*(y) #- x*{z).
-that is, thatfz{y) = - /1 (iy). But this follows from Corollary 4.8.6 For any x in a normed linear space X,
ft(iy) + if2(iy) = y*(iy) = iy*(y) = if!(y) - fiy). lx*(x)l
ilxll = sup--,..,-= sup ]x*(x)l. (4.8.7)
It remains to prove that 11 x* I ~ 11 y*jl. For any x e x, write x'"(x) = re;e llx il
"';<0 !'x,, = 1
where r ~ 0, (J real. Then, by (4.8.5), ' lo
The right-hand side is clearly less than or equal to the left-hand side.
lx*(x)l = lx*(e-; x)l = F1(e-ex) ~ lly*ll lle-' xll = liy*llll x 11.
6 6
On the other hand, by Corollary 4.8.4, there exists an x~ such that x~(x) =
Thus llx*ll ~ lly*ll. I xi!, 11x:;11 = 1. Hence the left-hand side of (4.8.7) is less than or equal to the
We shall use the Hahn-Banach theorem to prove the following useful right-hand side.
result.
Corollary 4.8.7 Let Y be a linear subspace of a normed linear space
Theorem 4.8.3 Let Y be a linear subspace of a normed linear space X X. If Y is not dense in X, then there exists a functional x* #- 0 such that
and let x 0 e X be such that ' x*(y) = Ofor ally E Y.
inf IIY- x 0 ll = d > 0. (4.8.6)
1eY Indeed, if Y =f. X, then there is a point x 0 satisfying (4.8.6). Now we
Then there exists a point x'" eX* such that can apply Theorem 4.8.3.
Let X be a normed linear space and let x* EX*. The null space of x*
is the set N". = {x; x*(x) = 0}. If x* #- 0, then there exists an element
for ally e Y.
x 0 =f. 0 such that x*(x 0 ) = I. Each element x eX can then be written in the
I form x = z + .h 0 , where l = x*(x) and z = x - J.x 0 E ;\,. This gives a
Proof Denote by Y1 the linear space spanned by Y and x . Since
x 0 Y, every p_oint x in Y1 ~as the form x = y + ..1.x 0 , where y e y ~nd the
decomposition of X into a direct sum of N., and the one-dimensional space
scalar A are umquely determmed. Consider the linear function z* defined by spanned by x 0 .
Let X be a normcd linear space and let x* E X*, x* #- 0. For any real
z*(y + ..1.x 0 ) = A. If }. #- 0, then
number c, the set {x; Re x*(x) = c} is called a hyperplane. It divides the
space X into two half-spaces defined by the inequalities Re {x*(x)} ~ c
IIY + holl = lAIII I+ II~ Ill d.
Xo and Re {x*(x)} ::;;:; c. If X is a real space, then this hyperplane wineides with
the vector sum Nx + cx 0 , where x 0 is any point for which x*(x 0 ) = 1.
Hence iz*(x)l ~ llxll/d for all x e Y1 -that is, liz* II~ Ijd. Now let {y} c. y
I Xo - Y.ll ..... d. Then n '
Definition 4.8.1 Let K be a subset of a normed linear space X and let
1 = z*(xo- y.) ~ llz*ll llxo- Y.ll-+ llz*lld. x E K. If an x* eX* satisfies Rc {x*(x)} ::;;:; Re {x*{x 0 )} for all x E K, then
0
I,
154 ELEMENTS OF FUNCTIONAL ANALYSIS IN BANACH SPACES
APPLICATION TO THE DIRICHLET PROBLEM !55
we say that x* supports Kat x 0 (or that it is tangent to Kat x 0 ). The hyperplane topology), then we say that A (I) is uniformly differentiable. Prove that
Re {x*(x)} =II x 0 ll is called a supporting (or a tangent) hyperplane to K at x 0 . eA [A E .':f(X)] is uniformly differentiable and de'Afdt = Ae'A. .
Corollary 4.8.4 can now be rephrased in a geometric language as 4.8.10. Let X be a real normed linear space, and let u(t) be contmuous
follows: and strongly differentiable in (a,b). Then for any a< cc < f3 < b,
Corollary 4.8.8 Let X be a normed linear space and let x0 belong to the du(t) II
llu(,B)- u(o:)ll $ (,8 ~ ct) ~:~~P dt
boundary of the unit ball (that is, llxoll = 1). Then there exists a tangent 1\
hyperplane to the unit ball at x0
[Hint: Apply x* to u(,B) ~ u(o:).]
PROBLEMS 4.8.11. Let X be a Banach space and let A E &l(X). Prove that for
any x 0 EX there exists a unique solution u(t) of
4.8.1. let X be a normed linear space and let {x.} c X. A point y 0
n
is the limit of linear combinations L cixi if and only if x*(y 0 ) = 0 for all x*
-du + Au = 0 (0..;, t $ 1), u(O) = Xo
j= 1
for which x*(x) = 0 for l ..;,} < oo. dt
4.8.2. Any two Hamel bases in a linear vector space X have the same
[Hint: To prove uniqueness, show that e'Au(t) is indcp~ndent of 1.]
cardinal num her. This n urn her is called the dimension of X.
4.8. 12. For every n ormed linear space X there 1s a set A such that. X
4.8.3. Two linear vector spaces X and Y are called isomorphic if there
is isomorphic to a subspace of the Banach space .of functions f on. A w1th
is a one-to-one linear map T from X onto Y. Prove that X and Y are iso-
norm 11111 = sup lf(t)l. If X is separable, A 1s countable. [Hmt: Let
morphic if and only if they have the same dimension.
4.8.4. Prove that dim 1~ = c, where c is the cardinal number of the {x,; o: E A} be d;nse in X. Let f(x,cc) be the bounded linear f~nctional.(in
continuum. [Hint: The elements ().,2 2 , .. . ,..1." ... ), 0 <A.< 1, are linearly x EX) satisfying llf(.,:dll = J,f(x~,:x) = llx~ll Define the 1somorph1sm
independent.] x _,. g,(o:), where g,(cc) = f(x,ct). Prove: llf(x,:xJI.~ .llx Ill ..;, 211.x~ ~ xll.]
4.8.5. Let X be an infinite-dimensional Banach space. Prove that 4 .8. 13. A separable n ormed li ncar space Is IsomorphiC to a subspace
there exists an infinite, strictly decreasing sequence { r.} of infinite-dimensional of/"". f . I
closed linear subspaces of X. [Hint: Take Y1 to be the null space of some 4.8.14. Let x, y be Banach spaces and letf(x,y) be a unctwna o.n
xi # 0 in X*. Take Y2 to be the null space of some x! -,t. 0 in Y~, and so on.] X x Y, continuous and linear in each variable separatc.ly .. Then f(x,y) IS
4.8.6. Let X be an infinite-dimensional Banach space. Prove that the continuous in the variable (x,y). [Hint: To prove contmmty at (0,0), l~t
linear space I"' is isomorphic to a Iinear subspace of X. [Hint: Take x" E r. _1 , ym = {y; liYII $ 1/m}, c > 0, A"= {XE X; lf(x,y)l'"'{, e for. all yE Yn}. A" IS
x. f: Y. with Y" as in Problem 4.8.5, Y0 = X, such that II x":J = 1/2", and con- closed [since f(x,yl is continuous in x] and UA.
=X [smce f(x,y) IS con-
sider the map {(.} -> L (.x".] tinuous in y]. Use the Baire category theorem to deduce that .lf(x,y)l ..;, 2e
4.8. 7. Prove that the dimension of an infinite-dimensional Banach if x E u, y E y'" for some m and for some neighborhood U of 0 m X.]
space is 2 c.
4.8.8. A Banach space X is finite-dimensional if and only if every
linear subspace is closed. [Hint: Consider the linear space spanned by the
4.9 APPLICATION TO THE DIRICHLET PROBLEM
x" in Problem 4.8.6.]
4.8.9. Let u(t) be a function defined on a< 1 < b with values in a
Rn d 1 t A - ~ ' 1 !'~ 1 The equation
Banach space X. We say that u(t) is strongly differentiable at t [on (a.b)] if Let n be a hounded domam m an e '-' ~ ;~t c r.~, 1
lim {[u(t +h) ~ u(t)]ih} exists [for all t E (a,b)]. The limit is denoted by 2
L\u = 0 is called the Laplace equation. Any function in C (Q) that satisfies the
h~O
du(t)/dt and is called the derimtive of u( 1). For functions A(l) with values in Laplace equation in Q is called a harmonic jimction in f.!_ We shall denote hy
~(X), if lim {[A(t + h)x- A(t)x]/h} exists for any x EX, then we say that an the boundary of n. We say that an is in C 1 if an can be covered by _a
h~O
finite number of open subsets G;, with each G,ncn having a paramctnc
A(t) has a strong derimtive. If lim {[A(t + h) - A(t)](h} exists (in the uniform 2
h~O represe n ta ti on in terms of functions in C .
!56 ELEME~TS OF FU>JCTlOl'iAL A~IILYSIS IN ll!INACH SPliCES APPLICATION TO THE DIRICHLET PROBLEM 157
The Dirichlet problem is the follov.ing one: given a continuous function Definition 4.9.1 A function G(x,y), defined for each y E 0: and
f on Dn, find a function u in C 1 (Ql, continuous in O, and satisfying X E n- {y}, is called Green's function (for the Laplace operator in Q) if:
1w = 0 inn, (4.9.1) {i) G(x,y) = k(x,y) + h(x,y), where h{x,y) is harmonic in x when
U=f on en. (4.9,2)
X En.
(ii) G(x,y) is continous in x when x E 0: - {y}.
Theorem 4.9.1 The Dirichlet problem has at most one solution. (iii) G(x,y) = 0 for x E DO:.
This follows immediately from the following re,ult, known as the One can usc Green's function in order to represent the solution u of
(ll"eak) maximum principle. the Dirichlet pro hlem (4. 9.1 ), ( 4. 9. 2) in terms of an integral i nvo! ving the
boundary values f (compare Problem 4.9.2). What we shall now do is con-
Theorem 4.9.2 Let u be a h armomc'
f'unction in struct Green's function hy using, as a tool, the Hahn-Banach theorem. We
n, continuous in n.
Then max u = max u. shall wnsider only the case n = 2, which is somewhat ~mpler than the case
1~ tn where n ~ 3. We shall assume that an consists of a finite number of con-
tinuously differentiable closed curves. The following condition then holds (the
Proof. If the assertion is not true, then there is a pointy E Q such that proof is omitted):
u(y) > ';: ~X u. Consider the function r(x) = u(x) + jx _ yj2. If > 0 is
1 ( P) For any z near an,
denote by T" the tangent plane to at the en
point on (;Q nearest to z. Denote by z' the reflection of z with respect to Tz-
sufficiently small, then v(y) = u(y) >max v. Hence the maximum of v on n Then
. . en
JS attamed at a point z in n. At that point, jz- xi
max------1
xen I='- xi
(i = l, .... n). Theorem 4.9.4 lfn = 2 and en is in C 1
, then Green's(unction exists.
Hence t..r :"0: 0 at z. However, Proof Denote by X the Banach space of all continuous functions on
an with the maximum norm, and denote by X' the linear subspace consisting
of those functions f for which the Dirichlet problem (4.9.1), (4.9.2) ha' a
-a contradiction. solution. For any y E n, consider the linear functional Ly on X' defined by
As for the existence of a solution to the Dirichlet problem. we quote the Ly(fl = u(y), where u is the solution of (4.9.1 ), ( 4. 9.2). From Theorem 4.9.2
following result: . it follows that LY i' bounded and that its norm is 1.
By the Hahn-Banach theorem L, can be extended into a bounded linear
Theorem 4.9.3 lf Zn is in C 2 then for any continuous function f on en functional on X, having norm i. We denote such an extension again hy Ly.
there exists a solution to the Dirichlet problem (4.9.1 ), (4. 9.2). For each z an consider the element(, of X given by
h(x) =log jx- zj (x E an),
Theorem 4.9.3 is much deeper than Theorem 4.9. I. There are various
methods of proving it, hut each of them require' lengthy developments. We and define k,(z) = Ly(f,). We claim that k,(z) is a harmonic function. To
shall present one of these methods in Section 5.4. o,
prove it let z' = (z 1 + z 2 ). Then
Consider now the function
ky(z')- ky(z) = L {f,- f.).
2 0 b .
k(x. r) = /I
X - l'j
-"
' if II> 3.
y \
In the same way one shows that k,(z) has any number of derivatives
and these arc equal to the application of Ly to the corresponding derivatives 4.10 CONJUGATE SPACES
o~ f.. In particular, L>.k,(z) = Ly(L>.f.). Sim:c, however, L>.f, is the function AND REFLEXIVE SPACES
ky(z) = Ly(f.J =log [y- z[ if z n. (4.9.3) The converse is not true in general (see Section 4.14).
Now take a point z in 0: and near 20:. Denote by z' its reflection with
respect to T" [see the condition (P) above]. Then Proof. Let {x~} be a dense sequence in X*, and choose x. EX such
. 1:- x[ that 11x.11 = 1. fx!(x.)l > Jlx!::/2. Denote by A the set of all linear combina-
!If,- j,:! =max log-,-----. 0 if Z-+ Zo, z0 E i30:, tions of the x. with rational coefficients. A is countable, and it suffices to
XEcn [.::: - xl
show that it is dense in X. If A -1= X, then, by Corollary 4.8.7, there exists an
Tt follows that L).(f, - j~.)---> 0-that is, x"' -1= 0 such that x*(x) = 0 for all x EA. Let {x:} be a subsequence of
{x~} that converges to x*. Then
ky(z)- ky(z')--+ 0 if z---> z0 .
But, by ( 4. 9.3), lim kJ(z') exists and equals log [y - z 0 [. Hence k,.(z) (z E 0:) * C. l(x * - x.)(x.,)l
llx * ""'x.;ll * * II x.~
) I C. - II
~an be~extended,into _a continuous fum:tio~n ky(z) inn, and k,(z) =log IY- z[
= lx._(x., ~-
2
~f z E cO:. The function -log [x - y[ + k).(x) satisfies the condition (i)-{iii)
111 the definition 4. 9. I of Green's function. Hence x:,---> 0. It follows that x* = 0-a contradiction.
PROBLEMS Definition 4.10.1 Let X and Y be normed linear spaces. Suppose there
2
exists a one-to-one linear map T from X onto Y. Then we say that X and Yare
4.9. 1. If cO: is in C and I' denotes the out ward normal to cO:, then linearly isomorphic (or, briefly, isomorphic) and we call T a linear isomorphism
Proof We can apply the principle of uniform boundedness to the ,; .. Now let y~"' e Y*"' and define x 0 = K- 1(qri*). For any y* e Y*, let
family {i,.}. We conclude that sup lli~ll < oo. From this the assertion follows. x be any extension of y to an element of X*. Then y* = ax*, and
~
y6*(y*) = (r.r6*)(x*) = (Kx 0 )(x*) = x"'(x 0 ) = y*(x 0 ),
Definition A Banach space X is called reflexive if K(X) =X""". since x 0 e Y. Thus y~* is the image of x 0 under the natural imbedding of
Y into Y **. This proves that Y is reflexive.
We shall prove later on (in Section 4.14) that LP spaces are reflexive if
1 <p < oo, but are not reflexive if p = I, p = oo. Theorem 4.10.6 A Banach space is reflexive if and only if its conjugate
is rejlexire.
Lemma 4.10.4 Let X and Y be isometrically isomorphic normed linear
spaces. If Y is reflexive, then X is reflexive. Proof. If X is reflexive, then the same is true also of X*"', which is
then iso~etrically isomorphic to X (with the map K). It is obvious that
(X"')*"'= [(X"')*]*= (X*"')*. Now let x~** e (X*)*"'. Then x~** E (X**)*.
Proof Let a be the isometric isomorphism from X onto Y. Define a
Thus the functional x~ = x~**K defined by x~(x) = x&*"'(KX) for x EX is
map -r from X"' into Y"' by
(n"')(y) = x"'(a- 1y).
in X*, and ,
x**(x~) = (Kx)(x~) = x~(x) = x6*"'(Kx) = x~*"'(x**)
It is easily seen that r is an isometric isomorphism from X* onto Y*. Similarly
define an isometric isomorphism p from X*"' onto Y"'* by for any x"'* EX**. Hence x~** is the image of x~ under the natural imbedding.
Since x&"'* is an arbitrary element of (X"')**, we conclude that X* is reflexive.
(px**)(y"') = x**(-r- 1y*). Suppose conversely that X* is reflexive. Then, by what we have just proved,
Denote by K1 and K2 the natural imbeddings of X into X** and of Y onto x is reflexive. Since K(X) is a closed linear subspace of X**, Theorem
Y**, respectively. Now take any element x;i* in X*"' and define x 0 = 4.10.5 shows that it is reflexive. Hence, by Lemma 4.10.4, also X is reflexive.
a- 1Kz 1 px~*. The point x 0 is in X and, for any x* EX*, We now introduce an important concept of convergence.
x*(xo) = x"'[a- 1(K2 1 px6*)] = (tx*)(K2 1 px~*) = px6*(n*) = x6*(x*) Definition 4.10.2 Let X be a normed linear space. A sequence {x.}
Thus K 1 x 0 = x~*. We have thus shown that K1(X) = X"'*-that is, X 1s in X is said to be weakly convergent if there exists an element x EX such that
reflexive. Jim x*(x.) = x*(x) for any x* EX"'. We call x the weak limit of {x.} and we
s;y that {x.} com,erges weakly to x. A set K c: X is called weakly sequentially
Theorem 4.10.5 A closed linear subspace of a reflexive Banach space is compact if every sequence {x.} in K contains a subsequence that converges
reflexive. weakly to a point in K. A sequence {x.} in X is ca1led a weak Cauchy sequence
if {x"'(x.)} is a Cauchy sequence for any x* eX*. The space X is called
Proof Let Y be a closed linear subspace of a reflexive Banach space weakly complete if every weak Cauchy sequence has a weak limit. A set
X. Consider the map a: x .....
Y* defined by (ax*)(y) = x"'(y) for y E Y. Note K c: X is weakly closed if the weak limit of any w~akly convergent sequence
that !lax* II :S: ilx*ll Next define for y** E Y**, (-ry**)(x"') = y**(ax*) for {x.} in K is also in K.
x EX"'. Then l(ry"'*)(x*)l :S: lly"'*ll !lux* II :S: lly"'*ll llx*ll. Hence ry*"' EX*"'.
Since X is reflexive, the natural imbedding K maps X onto X**. Hence PROBLEMS
K _,( Ty"'*) is a well-defined element of X. We shall prove that it belongs to
4.10.1. A sequence {x.} t:annot have two distinct weak limits.
Y. If x = K- 1(Ty*"') 1/: Y, then there exists an x* eX* with x*(y) = 0 for all
4.10.2. Let X be a normed linear space and let B be a dense subset of
y e Y and x*(x) # 0. Since x"'(y) = 0 for all y e Y, ax"' = 0. Hence
X*. If a sequence {x.} in X is bounded, and if lim x*(x.) exists for each
0 = y**(ax*) = (-ry**)(x*) = (Kx)(x*) = x*(x) n
! :
Theorem 4.10.7 Let {X 0 } be a weakly convergent sequence in a normed exists a y** e Z** such that lim (Ky.)(z*) = y**(z*) for all z* e Z*. Since Z
k
linear space X, then: is reflexive there is then a pointy E z such that its image in Z* * (by the natural
(a) {x 0 } is bounded; ,..
imbedding) is y**. Hence,
(b) its weak limit x belongs to the closed linear subspace spanned by
{x 1 ,x 2 ~. ~ .,xnP .. }; Jim z*(yx) = z*(y) for any z* E Z".
k
(c) llxll ~lim llx.ll.
n Now take any x* e X*. It determines an element z* E Z* by z*(z) = x"(z)
for a11 z e z. Since {y.} c: Z, y e Z, we conclude that
Proof. The assertion (a) follows from Theorem 4.10.3. Next, if (b) is
not true, then by Theorem 4.8.3 there is an x* eX* such that x*(x.) = 0 for lim x*(.rd = x*(y).
k
all n ;::: I, but x*(x) =1- 0. This is impossible, since {x.} converges weakly to
Hence {y } is weakly convergent toy. Since K is weakly closed, y E K. Thus
x. Finally, from x*(x) =lim x*(x.) we get 1
J K is weakly sequentially compact.
"
lx"'(x)l =lim lx*(x.)l ~lim {llx"'llllx.ll} = llx*ll fim llx.ll- Theorem 4.10.9 A reflexil;e normed linear space X is weakly complete.
n
Now use Corol:lary 4.8.6. Proof. Let {x.} be a weak Cauchy sequence in X. Then li~ x*(x.)
The next two results are concerned with important properties of
exists for all x* ex. By Theorem....4.10.7, {x.} is bounded. From the proof of
reflexive spaces.
Theorem 4.10.8 it then follows that there is a subsequence {x.;} that converges
Theorem 4.10.8 Let X be a reflexive Banach space. A set K c: X is weakly to some element x E X. But then
weakly sequentially compact if and only if it is both bounded and weakly closed. lim x*(x.) = lim x*(x.) = x"'(x) for any x* E X
n;
Proof. Suppose K is weakly sequentially compact. Then K is weakly Thus {x.} is weakly convergent to x.
closed. Indeed, let {x,} c: K, {x,} weakly convergent to x. There is asubsequence
{x . } that is weakly convergent to a point y in K. Since x = y, x e K. Next, PROBLEMS
K is bounded. Indeed, otherwise there is a sequence {x.} c K such that 4.10.3. In a finite-dimensional normed linear space, the concepts of
llx.ll > n. By our assumption on K, there is a subsequence {x~.} that is weakly
convergence and weak convergence coincide. . .
convergent. But then, by Theorem 4.10.7 (a), {x.,} is bounded, which is 4.10.4. A weakly closed set is closed. [The convers~ 1s not true, Problem
impossible since llx.; I > n;. 4.14.6 shows that the closed set {x; ll_x\'1 = I~ in L (0, \)IS not ~eakly closed.]
2
4.10.10. Let X be a normed linear space and Lemma 4.Jl.1 A topological space X is compact if and only if every
class {F.} of closed sets in X with empty intersection contains a finite subclass
converges weakly to y, then there exists a sequence
{F.,, ... ,F.J with empty intersection.
scalars) that converges toy.
4.10.11. Let T be a bounded linear operator from a normed linear The proof is obtained by noting that the sets F~ = X - F~ are open,
space X into a normed linear space Y If {x") is a sequence in X that is weakly and using the relations:
convergent to x 0 , then {Tx"} (in Y) is weakly convergent to Tx 0 .
4.10.12. A sequence {x"} in a normed linear space X is weakly con-
vergent_ to x 0 if and only if the following conditions hold: (i) the sequence
{llx.ll} IS bounded, and (ii) x*(x")---> x*(x 0 ) as n .... cc, for any x* in a set B
dense in X*. Lemma 4.11.2 A dosed subset ofa compact topological space is compact.
xEK Si nee (by Problem 3.1.12) a closed subset of B is a closed subset of X, Lemma
tionf(x) = xll takes a minimum on K.
1
4.11.1 yields the assertion.
4.10. 14. Prove that the unit ball K of a normed linear space X is A class {F.} of sets of X is called centralized (or is said to have the
weakly closed. Prove also that if the norm is strictly convex (for definition. finite intersection property) if any finite number of sets F, have nonempty
see Problem 4.3.3), then the minimum of f(x) = !lxli on K is attained at one intersection.
point only. Lemma 4. 11.1 im mcd iatel y yields:
Conversely, if for any x EX there is assigned a class .:f 0 (x) of sets of Theorem 4.11.5 Let X be a topological space satisfying the first count-
X such that (I )-(3) hold, then a topology can be defined on X by taking the ability axiom. If X is compact, then it is also sequentially compact.
open sets to be all the possible unions of sets rV(x) in Jf'0 (x) with x EX.
The proof of this fad is left to the reader. Nok that each set Jf0 (x) is then a Proof Let {x"} be a sequence in X. For every positive integer n let
neighborhood basis at x. M. = {x.,x"+i>-l- The class {M.} is centralized. Hence, by Lemma 4.11.4,
TYCHONOFt''S THEOREM 167
166 ELEME!\TS OF FUNCTIONAL ANALYSIS !'i BA'iACI! SPACES
i=l
0
).
be the space whose elements are the ordered sets x = {x,}. We define on X
a topology by giving a neighborhood basis .Jf 0 (x 0 ) at each point x 0 = {x 0 }. Since x~ is a point of contact of the class {M~J. U(x~) intersects each 1\.1L
A set U(x 0 ) is in f 0 (x 0 ) if there exists a finite number of indices a 1 , .. . ~ct.n Hence U;(x 0 ) intersects each set M'". By (a), U;(x ) intersects any fin~te
0
0
from A and neighborhoods U(x~.) of x~. in Xa, such that U(x 0 ) is the set of intersection of sets of {M'}. But then (b) implies that U;(x ) is in {M"}.
0
all points {x~} with x.; E U(x~) for i = 1, ... ,n and x. arbitrary in X~ if Hence, by (a), U(x 0 ) is in {M'-}. It follows that U(x ) intersects each set M'".
a of= :r: 1 , ... ,:r: of= a". 1t is easy to verify that the conditions (1 )-{3) for neighbor-
hood bases are satisfied. Thus X is a topological space with the classes PROBLEMS
f 0 (x 0 ) as neighborhood bases. We call X the Cartesian product of the
4.11.1. Let X andY be topological spaces. A function/from X into
topological spaces x .
Y is called continuous if the inverse image of open sets are open sets. Prove
Consider the function fo(x) = x, from X onto X,. The image of a set
that f is continuous if and only if the inverse images of closed sets are closed
M c X under this map is called the projection of M on X .
The next result is called the Tychonoff theorem. sets.
4.1/.2. A continuous function maps compact sets onto compact sets.
Theorem 4.12.1 The neighborhood bases % 0 (y*) satisfy the conditions M = {A; \A.- A. 0 1< e'}, N" = N(y*; x 1 , .. . ,x"; e').
( 1)--{3) of Section 4.11 and, therefore, determine a topology on X*. With this
topology, X* is a topological linear space. Then, if A. E M and x* EN",
~ lh*(x;)- ).0 y*(x;)l = lh*(x;)- /.y*(x;)l+ 1).- ).olly*(x;)\
. Definition The topology Introduced by the neighborhood bases.)(" (y*)
IS called the weak topology. 0 ~ \i.\e' + Ke' ~ (IA.ol + r.')B' + Ke',
where K =max Jy*(x,)l- Taking c' < 1, CIA.ol + I + K)e' < r., we find that
0
- A. y~(x )1 < e for 1 ~ i ~ n. Hence A.x* E N'. Thi~ completes the
1
Proof. The condition (1) is obvious. The condition (2) follows by
noting that the set IA.x*(x 1)
proof of (iv) and of the theorem. _
{x*; lx*(x 1) - y*(x;)l <c for I ~ i ~ n} n {x*; lx*(x;) - y*(x;)l < e' Whereas in the norm topology bounded sets in X* are not relatiVely
compact in general, they arc relatively compact in the weak topology. This
for 1 ~ f ~ m} will be proved in the next theorem, called the theorem of Alaog{u. This result
contains the set is one of the most important illustrations of the importance of the concept of
{x*; lx*(z;)- y*(z;)l < e" weak topology.
for I ~ i ~ n + m},
where e" =min (e,e'), z,- = x- if J < i < n z . = x--' if 1 < < m Theorem 4.12.2 Let X be a normed linear space. Then the closed unit
l ~ - ~ t~+~ - 1- .
To prove (3) let y! E N(y*: x 1 , ... ,xn; e). Then ball in X* is compact in the weak topology.
IY!(x;) - y*(x;)l < e' if I ~ i:::;; n, l'roof. To each x EX we correspond the closed real interval /"' =
for some 0 < e' <e. It follows that II, I x II] if X is a real space, and the closed disc in the complex plane
[ - 'H x
Jx = {z; lzl ~ l,',x 11} if X is a complex space. Denote by I the Cartesian product
f1 (,. By Tychonoff's theorem, I is compact.
;<eX
170 ELEMENTS OF FUNCTIONAL ANALYSIS IN BANACH SPACES
WEAK TOPOLOGY IN CONJUGA"rE SPACES 171
Consider now the set r = {f} of all the elements f of x with I /!1 :::;; 1. Proof Denote by X~ (R > 0) the topological subspace {x* EX*;
We have to prove that r is compact. For each fEr, 1/(x)l :$ llxll. Thus we llx*ll ::o;; R} (with the weak topology). Its topology can be defined by the
can correspond to fa point J in /, having coordinates f(x). We shall denote neighborhood bases consisting of thersets x;
n N, where N varies over the
this correspondence by a. Thus, af = ]. Jt is clear that r:r is one-to-one. Let sets N(x*; x 1 , .. ,x,; e). Let {y"} be a dense sequence in X. For any z* E XZ,
I' = u(r). From the definition of neighborhoods in X* and in I it follows the sets
that u and its inverse are both continuous. Hence r is compact if and only
if I' is compact.
To prove that l' is compact it suffices (by Lemma 4.11.2) to show that x; n N ( z*; Yt, .. ,ykn; ~) (n = 1,2, ... ; m = 1,2, ... )
I' is closed. Let]0 be a point in 1'. We first show that the functionalf0 defined
by fo(x) =fo(x) ( = the x-coordinate of / 0 ) is linear. Take any points x ,x form a countable neighborhood basis (for the weak topology of X!) at z*.
1 2
in X and any scalars A.1 ,A.2 . Let U0 = U(]0 ; x 1 x 2,..1. 1 x 1 + ..1. 2 x 2 ; e) be a neigh- Thus, the topological linear space x:
satisfies the first countability axiom.
borhood of ] 0 consisting of the points {yJ with Now let {x:} be a sequence of continuous linear functionals with llx!ll :::;; K,
K constant. By the Alaoglu theorem, the set Xi is compact in the weak
topology. By Theorem 4. 11 .S, this set is then sequentially com pact in the
IY .. ,x, +!.,x, -loC..tixl + ).2 x2)1 < e weak topology. It follows that there is a subsequence {x:,} that is weakly
convergent.
and Y.. arbitrary if x 'I- x 1 , x -1= x 2 , x -1= A. 1 x 1 + ..1. 2 x 2 . Since ] 0 E 1', there is an
element J in I' n U 0 . It has the form J = uf, where f is a continuous linear PROBLEMS
functional. Thus
4.1 2.1. Let X be a normed linear space. A sequence {/"} in X* is
weakly convergent to fr=. X*if and only if the following conditions hold:
f
lf(ltxl + ..l.2x2) -/o(AtXI + ..\ 2x 2)1 < e, (i) the sequence {II/,;;} is bounded, and (ii) limf,(x) = f(x) for all x in a
n
and](..t,x1 + A. 2 x2) = + A.zl(x 2 ).
A.j(x1 ) Consequently, dense subset of X.
l..l.do(x,) + A.2foCx2) -/o(A.tx 1 + ..l. 2x2)l 4 .12.2. If {f..} is weakly con vergen I to f, then I Iii ::o;; lim II/,}.
4.12.3. Give a direct proof of Theorem 4.12.3. [Hint: Choose a sub-
"
:::;; IA.d 1/o(xl)- f(xdl + l..\211/o(Xl) -/(x2)1 + l]o(..tlxl + ..\ 2x 2)
sequence {x:, 1 } such _that lim x:. 1 (~d exists. Choose a subse~ue~ce {x:,z}
- fVtxl + ..t2x2)l 5 (1..1.11 + P-21 + lk of {x:.l} such that lmi. x:.ix 1 ) exists, and so on. Then {xn,n} JS weakly
convergent.]
Since e is arbitrary, we get]0 (A. 1 x 1 + ..1. 2 x 2 ) = A.j0 (x 1 ) + A.zf0(x 2 ). Thus the
4.12.4. Weak topology can be introduced in the normed linear space
functional / 0 defined by / 0 (x) = ] 0 (x) is linear. Since ] 0 (x) E !,, lf0 (x)j :::;; llxll.
X by taking as a neighborhood basis of any pointy the class of sets
It follows that fo E X* and II fo I :$ I . Hence /o E r. But then fo = ufo E !'.
We have thus proved that]' = !'.
N(y; xi, .. . ,x:; e) = {x; lxj(x) - xj(y)l < forl::o;;j::o;;n}.
Definition 4.12.2 Let X be a normed linear space. A sequence {xn in Prove that this turns X into a topological linear sequence. Prove also that a
X* is said to be weakly convergent if there exists an element x* EX* such sequence {x,} is convergent to x in the weak topology if and only if {x,} is
that lim x!(x) = x*(x) for all x EX. We then call x* the weak limit of {x!} weakly convergent to x according to Definition 4.10.2.
~
4.12.5. If a Banach space X is reflexive, then its closed unit hall is
and we say that {x!} is weakly convergent to x*.
compact in the weak topology. (The converse is also true.)
Note that a sequence {x:} cannot have two distinct weak limits. 4.12.6. Any Banach space X is isometrically isomorphic to a closed
linear subspace of the space C(r) of continuous functions on a compact
Theorem 4.12.3 Let X be a separable normed linear space. Then etwy Hausdorff space with the uniform topology. [Hint: Let r be the closed unit
bounded sequence of continuous linear functiona!s in X* has a weakly convergent ball in X* provided with the weak topology. To each x EX corresponds a
subsequence.
x
function E C(r) by x(x*) = x*(x).]
172 ELEMENTS OF FUNCTIONAL ANALYSIS l'l BANACH SPACES ADJOINT OPERATORS 173
(T~ where J* is the identity operator in dJ( Y*, Y*). Similarly T*(T~ )*
1 )*T*, 1
Hence T* y* E X*. The map T--> T* is obviously linear. Thus it remains to
show that IIT*II = IITI! By Corollary 4.8.6 we have is the identity operator in M(X*,X*). It follows that (T*) ~ exists and is
1
Proof We first prove that R1- =:J N~ . What we have to show is that exists a constant K such that for any y ERr there is an x with Tx = y, llxll :$;
if y 0 rj Rr, then y 0 rj N~ By Theorem 4.8.3 there exists a y* E Y* such that Kl!yjj. Hence
y*(y 0 ) of= 0 and y*(Tx) = 0 for all x EX. Hence T*y* = 0. Since
lg(y)l = ]x*(x)] $ Kllx*ll llyll.
N~. = {y E Y; y*(.r) =0 for any y* such that T*y* = 0}, Thus g is a bounded linear functional on Rr. By the Hahn-Banach theorem
we can extend g into a continuous linear functional y"' E Y"'. Since, for any
it follows that Yo N~-
XEX,
We next prove that Rr c N~ . Let y E R.,. Then there exists a sequence
{y.} ERr such that lim Y. = y. If T*y* = 0, then y*(y.) = y*(Tx.) = (T*y*)(x) = y*(Tx) = g(Tx) = x*(x),
(T*y*)(x.) = 0, where Tx. = y . Hence also y*(y) = 0. We conclude that we have T*y* = x*. Thus x* E Rr
yEN~ . Suppose conversely that x* ER-r We shall prove that x* EN~. Let
y* E Y* be such that T*y* = x*. If x E Nr, then
Remark. Theorem 4.13.5 can be interpreted as an existence theorem
x*(x) = (T*y*)(x) = y*(Tx) = y"'(O) = 0.
for the equation Tx = y. It then implies the following result: if T* is one-to-
!'
one and Rr is closed, then Rr = Y -that is, for any given y E Y there exists a This shows that x* E Ny.
solution x in X of the equation Tx = y. In Section 5.2 we shall solve such The assertion that Rr is closed follows from (4.13.2) and the fact
equations more systematically. that Ny
is dosed.
The dual of (4.13.1) is R,. = ,v}. But this relation is not true in general.
\
We shall prove, however, the following weaker form of it.
PROBLEMS
Theorem 4.13.6 Let X and Y he Banach spaces and let T E da(X,Y). 4.13.1. Choose a basis {e1 , ... ,e.} in R" and choose a basis {ef, ... ,e:}
If Rr is closed, then Rr is closed and in R"*. To every linear operator Tin da(R",R") corresponds a matrix (al),
The proof depends upon the following lemma. sponds a metrix (af;), where Sej = t
l= 1
a7/t. Find the relation between (a; 1)
and (a;j) if S = T*. . .
Lemma 4.13.7 Let X, Y be Banach spaces and letT E ::iii'(X,Y). IfRr is 4. 13.2. Letf(z) be an entire complex analytiC functwn and letT E .>id(X).
closed, then there exists a con stall/ K such that _fi1r any y E Rr there is a point Prove that (f(T)]* = f(T*).
x EX 11'ith Tx = y and :x:l ::<; KIIY!!. 4.13 3. Let X and Y be Banach spaces and let T E .>id(X, Y). Prove :
(a) If T has a continuous inverse, then Rr is closed.
Proof By the open-mapping theorem (Theorem 4.6.1) T maps the (b) T* is one-to-one if and only if Rr is dense in Y.
(c) If T maps X onto Y, then T* has a bounded inverse with domain
unit ball 8 1 of X onto a set containing a ball in Rr with center 0 -tlmt is,
R,..
for some b > 0. 4.13.4. Let X, Y be normed linear spaces. Denote by d(x,L) the distance
(in X) from x to a set L. Let T E 9.J(X, Y). Prove that
Now let 0 of= y E R,. Then by/21lyll is in TB 1-that is, Tz = by/211YII for some
d(x,Nr) =max {x*(x); llx*! 1, ::<; 1 and x* Ny}. (4.13.3)
z, 11=11 < I. Let x = 211YIIz/b. Then Tx = y and llxll < (2/J).'y!l. E
Also,
4.14 THE CONJUGATES OF LP AND C[0,1]
x(f) = Jfg dp. for all f e l!(X,p.). (4.14.1) sup max I 11" yj dp =
r1/gl dp ~ llgll.=l lltnllp ___,_ oo
llgll.=l.
sup
ll9llo= 1
I lfgl dt~ = oo. v obviously is finitely additive. We shall prove that it is completely additive.
Let {Ed be a sequence of mutually disjoint measurable sets and let S" =
n cr..:
Proof To prove (a) we may clearly assume that 11/llr =1- 0. By Holder's U ~,, S = U 1
By the Lebesgue bounded convergence theorem [here
-~ l k~ l
inequality, we use the assumption that t~(X) < oo] we get
m=l
v('") =
a::
L
m=l
v(E'") .
THE CONJUGATES OF f.P AND C[O,I] 179
178 ELEMENTS OF FUNCTIONAL ANALYSIS IN BANACH SPACES
We have thus proved that v is a complex measure. If 11() = 0, then XE is Sincef.g?: 0 and limf.g = lfgl a.e., Fatou's lemma gives
zero as an element in U(X,J1). Hence v(E) = x"'(XE) = x*(O) = 0. Thus the
complex measure v is absolutely continuous with respect to the measure p. I lfgl dJ1 ~ lim f!. g dp
"
= lim x*(f,)
n
~ llx* 1111/ II p
Finally, v is a finite complex measure, since any function XE is in LP(X,p.),
and thus lv(E)I= lx*(XEll < oo. Here we have used (4.14.5) with/replaced by/, (f, is bounded) and (4.14.6).
By the Radon-Nikodym theorem there exists an integrable function g We can now apply Lemma 4.14.2(b) (with the roles of p and q interchanged)
such that and conclude that g e U(X,p.).
Now let fe U(X,J1) and define J.(x) = f(x) if l.f(x)l ~ n,f.(x) = 0 if
~(E)= I
E
g dJ1 lf(x)l > n. Then {f,} converges to f in U(X,p.). (This fo!lows using, for
instance, Coro!lary 2.8.5 with f replaced by 1/1 P.) It fallows that
for any measurable set E. Hence
x*(f) =lim x"'(f.).
n
By Holder's inequality we also find that IIJ..g - fgll 1 --> 0 as n--> oo. Since
By linearity of both sides with respect to XE we get x*(f.) = J_r.,g dJ.l, we conclude that (4.14.5) holds. Having proved (4.14.1),
with g e U(X,p.), the assertion (4.14.2) follows from Lemma 4.14.2(a). We
x"'(f) = I fg dp (4.14.5)
have thus completed the proof in case p.(X) < oo.
ln the general case where X is a-finite, write X as a countable union of
for any simple function f We shall extend this relation to any bounded
monotone-increasing measurable sets X 1, where J1(X 1) < oo. Denote by x:
the restriction of x* to X.-that is, for any .fin U(X.,p.), x:(fl = x*cl.l.
measurable functionj. Writing/in the formj, -j2 + lf3 - i]4 , where tfle]; where i.
is the extension off to X by 0. By what we have already proved,
are bounded, nonnegative measurable functions, and applying Theorem there exists a functional g. in Lq(Xn,Jl.) such that, for any fin U(Xn>p),
2.2.5 to each i;,
we conclude that there exists a sequence {f.} of uniformly
bounded simple functions that converges to f everywhere on X. By the
and
Lebesgue bounded convergence theorem it follows that
lim I fn g dp. = Jfg dp.. For every fe LP(X,,p.) and for every n > m, denote by F the extension of
fto X, by 0. Then j, = F,. It follows that
"
Since (by the Lebesgue bounded convergence theorem) II/,- fliP ...... 0 as
n--> oo, we also have
Letfbe any function in U(X,p.) and let(,= x.f, where Xn is the charac- Proof The proof is the same as that of Theorem 4.14.1, except for
teristic function of X . By the Lebesgue monotone convergence theorem, one difference, which occurs in part (a) of Lemma 4.14.2, when p =co (but
lim Jx" lfiP dp. = lfiP dJ1. Hence
J not when p = 1). The assertion now is that
We conclude that In proving (4.14.8) we may clearly assume that llf,i"' i= 0. It is clear that the
lim x*(f") = x*(f). term on the right is not larger than the term on the left. Next, from the
definition of the essential supremum it follows that for any e > 0 there is a
Also, by HOlder's inequality, llf,g- full 1 _.. 0 as n _..co. Thus, in particular, set E of positive measure such that lf(x)l > IIP,Ioc - E for all x E .Take
Corollary 4.14.5 Let (X,J.l) be a a-finite measure space and let 1 < p < oo. Corollary 4.14.7 The space L 1 (Q) is not reflexive.
Then P(X,J.l) is a reflexiFe space.
We shall now find the dual of the real space C[O, 1]. We denote by BV
Proof Take any x** in [U(X,JJ)]**. For any x* in [U(X,/1)]*, let the space of all functions g(x) on [0,1] of bounded variations, satisfying the
TX* = g and define a functional x** on Lq(X,p.) by x**(g) = x**(x*). This is a condition g(O) = 0. Denote by V(g) the total variation of g.
bounded linear functional. Hence by Theorem 4.14.1, there exists a function Let x* be a continuous linear functional over C[O,l]. Since C[O,l] is a
fin U(X,p.) such that x"'*(g) = Jfg dp.. Since the last integral is equal to closed linear subspace of L "'(0, 1), there exists a continuous linear functional
x*(_f) (by the definition of r), we get x**(x) = x*(f). This shows that x** 4l over L 00 (0, 1) that satisfies:
is the image off under the natura I imbedding. <b(f) = x*(f) if jE C[O,l] and lltPii = llx*ll.
Theorem 4.14.6 Let (X,J.l) he a a-finite measure space. Then for every For each x E [0, I] we define a function 1/1 x as follows:
x* E [U (X,J.l)]* there corresponds a unique function g in L "'(X,J.l) such that
1/J(t)={l, ifO:;;;t-5x,
x(f) = Jfq d11 X
Furt hemwre.
.g(x) = IP(ifr,). (4.14.9)
llx*ll = IIBII"''
and the map r: x*-" g is an isometric isomorphism from [L 1(X,J.l)]* onto
Since ifro = 0, g(O) = <b(O) = 0. We shall prove that g E BVand that
L"'(X,J.l). V(g) :;;; I x*ll (4.14.10)
182 ELEMENTS OF FUNCTIONAL ANALYSIS IN BANACH SPACES
THE CONJUGATES OF LP AND C[O,l} 183
Let 0 = x 0 < x 1 < ... < x._ 1 = x. =I, and write;= sgn [g(x;)-
g(x;_ 1 )]. Then By Problem 4.14.3, if i1 e BV, then
J f dg =I
I I
Elg(x;) -
i=l
g(x;- 1 )1 = f
i=l
e;[g(x;) - g(x; _ 1)]
0 0
f d{j for all f e C[0,1],
if and only if g = g at all points where one of the functions is continuous. Let
= L"
i~l
e;[<b(if!,) - <b(if/,,_ ,)] = <b(if/), us define equivalence " ,.., " in the space BV by saying that g ,.., {j if and only
if g and {J differ only at the points of discontinuity. Note that this set of points
where if!= L"
i-1
e;[o/,,- iflx,_ ,]. Since llo/11,., = I, we get is countable. We shall denote by BV0 the space of equivalent classes and by
[g] the class containing g.
J" I
1
g(xt) - (g; _,)1 ::::; ll<blllll/111"" = ll<bll = llx*l!.
Given any class [{J], consider the bounded linear functional x* on
C[O,I] defined by
I
This proves (4.14.10).
Now letf(t) e C[0,1], and define
x*(f) = Jf 0
d.
f.( I)= J, f (~)[iflkfn(t)- o/(k-1)/n(t)]. By the proof of Theorem 4.14.8, there exists a function g in BV such that
(4.14.11) holds and llx*ll = V(g). Since
It is easily seen that II/. - f II"' _,. 0 as n _,. co. Hence _ 1
llx*ll = sup [J f dl::::; V(g)
lim <b(f.) = <b(f) = x*(f). llfll=l 0
The right-hand side of (4. 14. 13) defines a norm on BV0 We shall designate
[91
the norm of [g] by II[] 11. Combining (4.14.13) with (4. 14.12), we can now
as n _,.co. We conclude that
state:
x*(f) =I 1
0
f(t) dg(t). Corollary 4.14.9 The map -r: x* ..... [g] defined in Theorem 4.14.8 is an
isometric isomorphism of the dual of C[0,1] onto the space BV 0 with norm
It follows (by Problem 2.1 1.10) that
ll[g] II= g.l.b. V().
~e[g]
lx*(f)l = 1.( f(t) dg(t) I::::; 0~~~\ 1/(t)l V(g). In every class [.ii] there is a function g as in (4.14.13). There is also a
Recalling (4.14.10), we get llx*ll = V(g). We sum up: unique g 0 having the following properties: (i) g 0 (t) is right continuous for
all 0 ;<=;; t < 1 ; (ii) g 0 (0) = 0, and (iii) g 0 (t) is left continuous at t = 1. In fact,
Theorem 4.14.8 For every continuous linear functional x* over C[O,l], g 0 (t) = g( t + 0) - g(O +) if 0 ;<=;; t < 1. We shall call g 0 the normalization of
there corresponds afunction g in BY such that [9], or of g. We shall denote by NBV the space of all these functions g 0
I This space is clearly a normed linear space with norm ll9oll = V(g 0 ).
x*(f) = Jf 0
dg for every f e C[O, 1]. (4.14.11) We claim that if g 0 is a normalization of g, then
V(g 0 ) = V(g). (4.14.14)
Further,
In view of ( 4.14.13), we only have to show that
llx*ll = V(g). (4.14.12)
V(g 0 ) ::::; V(g). (4.14.15)
\84 ELEMENTS OF FUNCTIONAL ANALYSIS IN BANACH SPACES THE CONJCGA TES OF U AND C(O, J] 185
To prove (4.14.15), take any .e > 0 and any partition 4.14.3. Let g and rJ belong to BV. Prove that
,I .1
= t 0 < 11 < ... < t. _ 1 < t. = 1
0
I 1 dg = o I
o
1 drJ for all f E C[O, 1]
of [0,1]. For each 11 (0:::;; i:::;; n- 1) choose T 1 > 11 and T;- t; so small that
if and only if g(l) = rJ(t) at each point I, v.here either g or {J is continuous.
e 4.14.4. Every continuous linear functional x"' over fP (1 < p < oo)
lo(t, + O)- o(-r;)l < -.
2n can be represented in the form
Choose also Tn such that r._ 1 < -r. < 1 and 1 - T" so small that -,:_.
x*(x) = I ry 1 x 1
e i=1
lg(l - 0)- g(-r.)l <-.
2n where
Then, if 1 s i :::;; n - I,
fl = (fll>fll> .. ) E lq (~ + ~ = 1) and llx*ll = {_f lr, 1lq}t/q.
Oo(tJ- 9o(t;-l) p q = 1
= [g(t, + 0) - g(O + )] - [g(t1_ 1 + 0)- g(O+ )] 4./4.5. Every linear functional/on R" has the form
= g(t; + 0)- g(t;-1 + 0)
f(x) = L" ry 1 x;
= (g(/ 1 + 0)- g(r;)J - (g(t1 _ 1 + 0)- g(T;_ 1 )] + (g(T;)- g(T1 _ 1 )J. i-=-l
Hence V(g 0 ) :::;; V(g) + e. Since e is arbitrary, (4.14.15) follows. /14.8. A sequence {.f~(rl} in P(O, I) (1 < p < OC) is weakly convergent
Combining Corollary 4.14. 9 with (4.14.13) and ( 4.14.14 ), we get : tof0(t)EP(O,l) if and only if (i) {"f,!lp} is a bounded sequence, and (ii)
J
lim 0J.(t) dt = s~ j~(t) dt for any T E (0, 1}.
Corollary 4.14.1 0 Consider the map cr: x* .... g 0 where g 0 is the normaliza- " 4.14.9. Prove Helfy's principle: if {g.) is a sequem:e of monotone and
tion of g and-g is defined by Theorem 4.14.8. This map cr is an isometric iso- uniformly bounded functions, then there exists a subsequem:e {g.,} that con-
morphism/rom the conjugate of C[O, 1] onto the space NBV. verges evervwhere to a bounded monotone function. [Hint: By Alaoglu's
Recall that g 0 is related to x"' by theorem and Theorem4.14.8 we may assume that f.jj(r) dg.(r) ____,. Hf(r) dg 0(T)
I
for any continuous function f, where g 0 is monotone. Let I be a point of
x"'(f) = J f(t)dg (t). continuity of g 0 . Take f to run over a sequence tp.,, (Pm( -r) = I if 0 s r :::;; t,
0
0
IPm(T) = 0 if t + ( 1/m) .:-::; T < oo and (Pm( T) linear in [t,l + ( l /m)]. Deduce
PROBLEMS lim [g,(tl- g,(O)] :::;; g 0(t)- g 0(0). Next take .f(-r) = tp.,[T + (l/m)], and
deduce lim [.q.(t) - g.(O)] ~ Bo(t) - Bo(O).]
4.14.1. If fU(X,p) (I <p < oo), then there is a function g m 4.14.!0. Find the adjoint of the operator given in Problem 4.4.9.
U(X,p) (!jp + 1jq =
1) such thatfg ~ 0 and Jfg dp = oo.
4.14.2. Let Q be an open set in R". Prove that L "'(Q) is not separable.
[Hint: Compare with Problem 3.1.7.}
BASIC PROPERTIES 187
CHAPTER 5 Proof Let A be a bounded set in X. For any e > 0 there exists an
operator T, such that I T, x - Txll < e/3 for every x e A. Since T,(A) is a
compact set, it is also totally bounded. It follows that there is a finite number
OPERATORS inf
1 s..is.m
I Tx - Tx;]! <E.
186
188 COMPLETELY CONTINUOUS OPERA TORS THE FREDHOLM-RIESZ-SCHAUDER THEORY 189
Using (5.l.2), we then get 5 .1.6. Let f(z) be an entire complex analytic function with f(O) = 0,
and let X be a Banach space. If T E .c.i'(X) is completely continuous, then
IY!,(Txm) -lim y,.:',(Tx,)l ;;:=: 1'1 , (5. 1.3)
f(T) is also completely continuous.
" J
Since ]]x.,)l = I and Tis compact, there exists a subsequence {mj} of {m } 5.1.7. Let T be a compact linear operator from a Banach space X
onto itself. If r- 1 is a bounded operator, then X is finite-dimensional.
such that lim Tx,./ = Yo exists. Hence, for any c > 0 there is an n0 such thit,
if mj ;;:=: n0 , 5.1.8. Let The a bounded linear operator in a reflexive Banach space.
If T maps weakly convergent sequences onto convergent sequences, then Tis
IIYo- Tx,/11 < e, IY!,-Cro) -limY!. (y 0 )] <e. completely continuous.
k
[Here we have used the facts that y 0 E T(X) and Lv!,(Y)} is convergent if
y e T(X).] Therefore, 5.2 THE FREDHOLM-RIESZ-SCHAUDER THEORY
IY!/(Txm/) -lim Y!.-(Tx,/)1 Throughout this section we shall denote by Ta completely continuous
" linear operator in a Banach space X, and by A a nonzero scalar. From the
s IY!;-(Tx,./ - Yo) I + IY!,-(Yo) - lim Y!.-(Yo)l + ]lim Y!.(Yo - Tx,/)1 proof of Theorem 5.1.4 we conclude that T* is completely continuous (X
s Me+ e +Me, " " need not be complete). We shall consider the bounded linear operator
AI - T and write
where lly!ll sM. Since to> 0 is arbitrary, we get a contradiction to (5.1.3).
Suppose conversely that T* is compact. By what we have just proved Nl = {x; Ax- Tx =0}, N: = {y*; ...ly*- T*y* = 0},
it follows that T** is compact. Rl = {y; y = ...lx- Tx, x EX}, R: = {x*; x* = ...ly*- T*y*,y*eX*},
Now let {x.} be any bounded sequence in X. Then the sequence {.X,}, where T* is the adjoint of T.
where .R, = KX, (K the natural imbedding of X into X**), is also bounded. It We recall the following fact proved in Section 4.3.
follows that I here exists a subsequence {i,.} such that { T** i".} is a conver-
gent sequence in X**. Since. by Theorem 4.13.3, K(Tx,) = T**i,, (here K Lemma 5.2.1 A closed linear subspace X 0 of X is finite-dimensional
is the natural imbedding of Yinto Y**), and since K is an isometry, it follows if and only if every bounded sequence in X 0 has a conLergent subsequence.
that {Tx,J is a Cauchy sequence in Y. Since Y is complete, this sequence ~~
convergent. Hence T is compact. We shall now prove:
, _ (A.I-T)x. y
().1 - T )w., =(AI - T)x., = n, = -"-'- _,. 0.
d(x.) d(x.)
Thus the sequence {Tx,} has no convergent subsequences. Since, however,
Since T is compact, it follows that a subsequence { Tw.) of {T~~".} is {x.} is a bounded sequence, we have derived a contradiction to the assumption
convergent. Hence w. , __,. w as n; _,. w. But then (a - T) w = 0 'and
that Tis compact.
d(w) =lim d(w.) =lim d(x.) = 1. This is impossible. We have thus proved
that {d(x.)} ts a bounded sequence, say d(x.) ~ C for all n. We can there-
Lemma 5.2.6 lf R_. = X, then N _. = 0.
fore choose z. c
in N;. such that l'x.- z.ll ~ + 1. Now take zn =Xn - n z
(b) From the compactness of T it follows that there exists a sub- Proof If N ~ #- 0, then there exists a point x 0 #- 0 in Nk Since R;. =X,
sequence {z.} of {z.} such that {Tz.,} is convergent. Hence also { z } we can construct a sequence {x.} in X such that Lx 1 = x 0 , Lx 2 = x 1 , Lx 3 =
!S convergent. Denoting its limit by x 0 , we get y 0 = ..1.x0 - T.-.;: 0 . This sho~s x2, .... But then r:x. = Xo I= 0, v+ l x, = Lxo = 0. Thus N).+ I #- N~ for all n.
that y 0 E R_.. This contradicts Lemma 5.2.5.
Combining lemma 5.2.3 with Theorems 4.13.5, 4.13.6, we get: Theorem 5.2.7 R~ =X 1/and only ifN~ = 0.
Theo~~m 5.2.4 (a) RJ. = N:1_--that is, the equation }...x- Tx = y has Proof Suppose N. = 0. By Theorem 5.2.4, R1 =X*. Since T* is
a solutwn if and only ify lies in the orthogonal complement ofN*. compact, we can apply Lemma 5.2.6 to T* (instead ofT) and conclude that
( b) RJ."' = NJ:-that
' . the equation A.y*- T*y* = x*. has
1s, J.
a solution if NT = 0. Theorem 5.2.4 then gives R. =X.
and on(v if x * lies in the orthogonal complement of 1'\ i.
Lemnw 5.2.8 Let xi .... ,x~ be linearly independent elements of X*.
Let L = ..1. I - T. Then
Then there exist points Xp- .,x" in X such that x:(xJ) = 8;; (8;J = 0 if i I= j,
i'i;i = 1 ifi = j) for all 1 ~i, j ~n.
It follows that L" has the form J.l.[ - T0 , where J.l. = A." is a scalar and T0 is
Proof Denote by NJ the null space of x7 and let Mi = N 1 n n
Ni_ 1 n Ni+l n n N . We shall prove that
a c~mpact operator. Hence the null space N~ of L" is finite-dimensional.
Notlce that N~ c: N~+ 1 . MJ is not contained in Ni, 1 ~} ~ n. (5.2.1)
Lemma 5.2.5 There exists a posilire illfegrr k such that N~ I= N~+J It suffices to prove (5.2.1) for j = L lf M 1 c: N 1 , then xt(x) = 0 whenever
1
x1(x) = = x:(x) = 0. Consider the linear mllp A: X ...... 9'"- (!F the
for all n < k and N~ = N~for all n> k. field of scalars of X) given by
Proof We first show that if N~ = N~+ 1 , then N~+ 2 = N~. let Ax= [x!(x), ... ,x;(x)],
xeN~+z. Then 0=L"+ 2 x=L!'-'- 1 (Lx). Hence LxeNZ+ 1 =N;. This
and define on A{X) a linear functional ijl by
means that L"(Lx) = 0-that is, x eN~+ =N1. Continuing in this way
1
~tep by step, we deduce that NZ-'-" = N; for any positive integer h. Thus, iji(Ax) = ijl[x!(x), .. .,x!(x)] = x~(x).
Jt remams to show that N~ I= N~+ 1 cannot occur for all n. If this does
ijl is defined unambiguously. Indeed, if Ax= Ai, then x!(x- i) = =
occur, then, by Lemma 4.3.1, there exists a sequence {x } in X such that
P1+1 li x:(x- x) = 0. Hence xj(x- i) =0-that is, iji(Ax) = iji(Ai).
x. E N J. , !lx.ll = 1, and llx.- xl' > l for all x EN~.
192 COMPLETELY CONTINUOUS OPERATORS THE FREDHOLM-RIESZ-SCHAUDER THEORY 193
By the Hahn-Banach theorem, we can extend the linear functiona1 !jJ From (5.2.2) and Theorem 5.2.7 it follows that the range of}.[- S is
into a linear functional !/1 1 in yn- 1 But then !jl 1 has the form X. Thus, in particular, there is an x eX such that }.x- Sx = Yn+I But then
n
!/lt(Y2>,Yn) I a,yl,
= i=2
where a 1 are constants. It follows that y: .r:+
since + 1 (y 1) = 0 (I :::;; i ::;;; n) and since 1 e IV~. The contradiction obtained
n
proves that n ;::= v.
xi(x) = I a,xj(x),
1=2
Applying this result to T*, we conclude that the dimension m of the
null space of A.l - T** is not larger than the dimension v of the null space of
thus contradicting the linear independence of the xj. A.J- T*-that is, v;::.: m. Since obviously m ;::= n, we get v 2': n. This completes
the proof.
From (5.2.1) it follows that there exist points xj E Mj such that Theorems 5.2.4, 5.2.7, and 5.2.9 constitute the Fredlwlm-Riesz-Schauder
x 1 ~ N 1 . This means that x{(xj) = 0 if i # j, and xj(x1) ,p 0. Now theorJ. We summarize these theorems, using the notation N 8 =null set of
take Xj = Xjjxj(x). S, R 8 =range of S:
Theorem 5.2.9 N,_ and N! have the same finite dimension. Theorem 5.2.10 Let X be a Banach space and let T be a completely
continuous linear operator in X. For any A. i= 0,
Proof In view of Lemma 5.2.2, n =dim N 1 and v =dim Ni are
finite numbers. Let x 1 , ,x. be a basis in N 1 and Jet yt, ... ,y: be a basis in dim Nu-r =dim N;.t-T' < oo, (5.2.3)
N!. By the Hahn-Banach theorem there exist elements in X"'xt, ... ,x: Ru-T= Nj,_p, Ru-T*= Nr- T (5.2.4)
such that x{(xj) = biJ for a!! I :::;; i, j:::;; n. In view of Lemma 5.2.8 there exist
elements y 1 , . . ,y. in X such that yj(y) = bu for all 1 :::;; i,j:::;; v.
Theorem 5.2.1 0 can also be stated in the following form (called the
Suppose n < v. Let
Fredholm alternatire).
[ither (a) for any y E X there exists a unique solution x of (U- T)
Sx = Tx + I" x{(x)y1 x = y, or (b) there is an x #- 0 such that (:U- T)x = 0. If (a) holds, then for
i= I
any y* eX* there exists a unique solution x* of (}.I- T*)x* = y*. If (b)
S is a finite sum of compact operators and therefore it is also a compact holds. then dim Na- r = dim ,'..,'a- r is finite and, furthermore, the equation
operator. We shall prove that (i.I ~ T)x = y [(}.!- T*)x* = y*] has a solution if and only if y (y*) is in the
orthogonal complement of Nu-J"O (NAI-T).
Nu-s= 0. (5.2.2)
Suppose x eN u-s Then Ax= Sx-that is, PROBLEMS
0 = (...lyj- T"'yj)(x) = yj(...lx- Tx) = xj(x). f(x) = g(x) + J. JK(x,y)f(y) d{l(y). (5.2.5)
Thus xj(x) = 0 for l :::;; j :::;; n. Hence ..tx - Tx = 0-that is, x e N 1 . Con-
" We ca!l this equation an integral equation of Fredholm's type. Prove that if
sequently, x = I
I= I
...l,x,. But then 0 = xj(x) = ).1. Thus x = 0. This proves
g = 0 implies f = 0, then there exists a unique solution of (5.2.5) for any
(5.2.2). g e L2 (X,p).
194 COMPLETELY CONTINUOUS OPERATORS ELEMU...;TS OF SPECTRAL THEORY I95
5.2.2. Let X be a compact metric space and let 1-i be a finite measure (b) (..1.I - T) _,exists, but its domain is not dense in X. We then say that
on X. Let K(x,y) be continuous on X x X. Assume that the only continuous I . 1. belongs to the residual spectrum of T.
solution of
~ (c) a- Tdoes not have an inverse---that is, there is an x #- 0 satisfy-
ing lx- Tx = 0. We then say that . 1. is an eigenwlue ofT.
I
f(x) = A K(x,y)f(y) d!-i(Y) I If . 1. is an eigenvalue, then any x #- 0 satisfying ..1.x - T x = 0 is called
an eigenvector (or eigenelement) of T (corresponding to ..1.).
is f = 0. Prove that for every continuous function g(x) on X there exists a
unique continuous solution f(x) of (5.2.5).
5.2.3. Let X,p,K be as in Problem 5.2.2. Denote by <:p 1 , .. 'Pm a basis Theorem 5.3.1 Let T E al'(X), X a Banach space. Then p(T) is an open
for the space of solutions <:p of the equation set and
R(11;T)- R(J.;T) = (..1.- Jt)R(Jc;T)R(p;T) (5.3.I)
I
<:p(x) = A K(y,x)<:p(y} dp(y).
if A.,11 E p(T). Furthermore, R(A ;T) is analytic in A E p(T).
Then the equation (5.2.5) has a solution if and only if Jg(x)op;(x) d11(x) =
0 for I ::; i ::; n. Proof If Ao E p(T) and lA- ..lol < I! R(}.o;T) r 1, then . 1. E p(T).
5.2.4. Consider the Volterra integral equation Indeed,
+ (). -
f(s) = g(s) + J' K(s,t)f(t) dt
0
(0::; t ::; l), (5.2.6)
..1.1 - T = ().
0
I - T)
1
Jc 0 )I = ( ..lo I - T)[I - (J.o - ..l)R(J.o; T)].
op(t) = . 1. r 0
(sin st)op(y) dy
Next,
(..1.1- T)- 1 - (!-!I- T)- 1 = (..1.1- T)- 1 [(!11- T)- VI- nJ(pi- n- 1
a> 0. It remains to prove that R(Jc;T) is analytic in . 1. E p(T). This follows from
(5.3.l). Indeed, if fl.--> ..1.;
5.3 ELEMENTS OF SPECTRAL THEORY R(!<; T) - R(.-1.; T) = - R(Jl ;T)R(I.; n--> - [R(..l; T)]l
p-Jc
Let T be a bounded linear operator in a normed linear space X.
in the uniform topology.
Definition The resolvent set p( T) of T consists of all complex numbers
. 1. for which (..1.1 - T) _, is a bounded operator <with domain X). The spectrum Theorem 5.3.2 Let X be a Banach space ami let T he a completely
a(T) of T consists of the complement of p(T) in C. continuous linear operator in X. Then the equation x - }.Tx = 0 has nontririal
The operator R(.-1.; T) = ( ..1.1 - T)- 1 if bounded (and with domain X) solutions only for a countable set (~(complex numbers haring no finite points of
is called the resolrent of T. accumulation.
Let . 1. E a(T). Then there are three possibilities:
Proof Suppose there exi:;ts a bounded sequence {)..} of mutually
(a) The range of J.l- Tis dense in X and (AI- T)- 1 exists but is di:;tinct numbers such that x - }" Tx = 0 has a solution x = X" i= 0 for each
unbounded. We then say that . 1. belongs to the continuous spectrum ofT. n. We first claim that, for each n, x 1 ,. ,x. are linearly independent. I ndecd, if
'.
...~
] 96 COMPLETELY CO:-IT!NUOUS OPERA TORS
I APPLICATION TO THE DIRICHLET PROBLEM I97
this is false, then there exists an integer k such that x 1 ,---,x,_ 1 are linearly PROBLEMS
independent and x 1 ,xk are linearly dependent, say 5.3.1. u(T) is nonempty.
k 5.3.2. p(T)= p(T*) and R(A. ;T*) = R(A.;T) *.
I
j= I
c,.x 1 =0 (5.3.2) 5.3.3. If A.0 E u(T) and p(z) is a polynomial, then p(A. 0) E u[p(T)].
Then [Hint: p(A. 0 )1 - p(T) = (A. 0 l - T)q(T), q(z) a polynomial. If p{A 0 )1 - p(T)
has a bounded inverse, then also A. 0 1 - Thas a bounded inverse.]
0 = )< ~
L
k
c1 Tx 1 = Ik '
/.<cJ
-,- x1. 5.3.4. If A.' E u[p(T)], p a polynomial, then A.'= p(A. 1 ) for some
j=l j~J ).j
Comparing this with (5.3.2), we get A. 1 E u(T). [Hint: p(T)- A'l = f1" (T- A.J), the l; being the zeros of p.]
i-1
Theorem 5.4.1 The following alternatire holds: either (a) for erery Using Theorem 5.4.1 (or rather a variant of it, with G, dx replaced by
g E C(G) there exists a unique solution fin C(G) of S, dS), we conclude:
f(x) = g(x) +A. Jo:; K(x,y)f(y) dy, (5.4.3) Theorem 5.4.2 If A = I is not an eigenvalue of the operator 11(x) _,.
J K(x,y)ll(Y) dSr, then there exists a unique solution of the Dirichlet problem.
or, (b) there L> a nontrhial solution in C(G) of(5.4.3) when g = 0. J/(b) holds,
then the spaces of rontinuous solutions of the equations It can be shown (cf. Problem 5.4.5) that A = I is indeed not an eigenvalue.
Consider now the Neumann problem-that is, the problem of finding a
harmonic function u in n satisfying the boundary condition
tp(x) = A ~ K(x,y)tp(y) dy, (5.4.4)
''G
I' au= f on S.
l}!(x) = ,t
'G
K(y,x)l}!(y) d:y, (5.4.5) av
We now employ the simple layer potential V. It satisfies
hare the same finite dimension. Furthermore, there exists a solution of (5.4.3)
So
if and only if f(x)ljl(x) dx = 0 for any solution \j1 of (5.4.5).
(5.4.7)
A proof of Theorem 5.4.1 can be given on the basis of the results of
Problems 5.2.2 and 5.2.3 and (5.4.2). The details will not be given here. Instead, This leads to an integral equation for a with the kernel K(y ,x):
we shall show how Theorem 5.4.1 (or rather a result similar to it) can be
used to solve the Dirichlet problem. For simplicity we take n = 3. a(x 0 ) = f(x 0 ) - Js K(y,x)u(y) dSY.
Let n be a bounded domain in R 3 with the boundary Sin C2 . Consider
the functions, in 0, The number }. = -1 is an eigenvalue. But it can be shown that the only
J
nontrivial solutions of the homogeneous equation !f(x) = - K(x,y)!ft(y) dSy
V (x) -_ -I j" - u(y)
- - dS,,
are the nonzero constants. Hence the Neumann problem has a solution if
21t sx-y
1 1 and only if JsfdS = 0.
1 . j) 1
W(x) =-I
2TI s
tt(y)-A -~- dS.,
ovY lx - yl '
PROBLEMS
5.4 .I. Let K be an integrable kernel. Prove that the operator T deli. ned
where vY is the outward normal to Sat y. We call V(x) a simple layer potential by
and we call W(x) a double layer potential. If ;t is continuous, then it can be
proved that, for any x 0 ES,
1 j) 1
(Tf)(x) = t K(x,y)f(y) dy
where v is the outward normal. Use this relation to prove that if 0 V/ OiJ.y =0
CHAPTER 6
for all y e S, then V = const.
5.4.5. It is known that
. oV(x)
hm -~~ -
x->y O>'y
. oV(z)
lim~-=
z->y OiJ.y
2u(y) (yeS). (5.4.8) HILBERT SPACES
J
Prove that ..t = I is not an eigenvalue of the map u(x)-+ K(y,x)u(y) dSr It
J
foll.ows that ). = I is not an eigenvalue of the map 11(x)-+ K(x,y)!l(Y) dSr
AND SPECTRAL
[Hmt: Use (5.4.7), (5.4.8).]
THEORY
Theorem 6.1.2 A Hilbert space H is a Banach space with the norm dition (ii) in Definition 6. 1.1 of the scalar product, we use the parallelogram
I = ( , }112.
law to obtain:
2
Proof The only nontrivial property of the norm that has to be verified llu + v + w';l 2 + llu + v- wll 2 = 21iu + Vil 2 + 2';',wll ,
is the inequality I x + y I ~ II x I + I y 11. Since llu- v + wl!12 + llu- v- wl1 2 = 211u- tll 2 + 211wll 2 .
(x,y) + (y,x) = 2 Re {(x,y)}:::;; 211xii11YII, Hence
then x = -...ly. From the proof of Theorem 6.1.2 it then follows that if Taking w = u, we get (2u,r) = 2(u,!). Taking u + w = x, u- w = y, v = z,
llxll + llyll = llx +.vii and y i= 0, then x = -..i.y. Hence if llxll = llyll = I, we then obtain
llx + y II = 2, then l..\1 = I, II -J.I = 2. Therefore ..1. = -!-that is, x = y.
We conclude (see Problem 4.3.3):
(x,z) + (y,z) = 2 ( -X+-y, :: ) = (x + y,z).
2
Corollary 6.1.3 The norm of a Hilbert space is strictly convex.
To prove the condition (iii) in Definition 6.1.1 we first note that, by (ii),
Theorem 6.1.4 For any two elements x,y in a Hilbert space H, (mx,y) = ((m - !)x + x,y) = ((m - I )x,y) + (x.y)
llx + 2
Yll + llx - y!i
2
= 2 ilxll + 211 YV
2
(6.1.2) = ((m- 2)x,y) + 2(x,y) = = m(x,y).
204 HILBERT SPACeS AND SPECTRAL THEORY THE PROJECTION THEOREM 205
Hence also n(x/n,y) = (n(x/n),y) = (x,y). It follows that (x/n,y) = (1/n)(x,y). M .l N. We denote that Mj_ the set of all elements that are orthogonal toM.
If r =min, where m,n are any positive integers, then We call MJ. the orthogonal complement of M. It is easily seen that M~~ _a__ _
closed lineaf_S_P!I-Ce.If N .l M, then, clearly, N c MJ. .
......__.----
r(x,y) = -m(x,y) = m (x
-, y ) = ('m
- x,y ) = (rx,y). Lemma 6.2.1 Let M be a closed convex set in a Hilbert space H. For
n .n , .n
euery point x0 E H there exists a unique point Yo in M such that
Since (x,y) is a continuous functional in x, we obtain ),(x,y) = (J.x.y) for
any positive real number. If;_ < 0, then llx 0 - Yoll = inf IIXo- YII- (6.2.1)
}'E M
l.(x,y)- (i.x,_r) = /(x,y)- (1)1( -x)._r) = l.(x,y)- li.l( -x,y) Proof Denote the right-hand side of (6.2.1) by d. Then there exists
= /.(x,r) + /.(- x,y) = i.(O,y) = 0. a sequence {_r.} in M such that llx 0 - y.ll-+ d as n-+ oo. By the parallelo-
Hence (iii) holds for any real 1.. It is easily seen that (iii) also holds for ). = i. gram.law,
2
Finally, for any complex number i. = 11- + i v, 4\IXo- !(y,. + y.)ll 1 + IIYm- r.ll 1 = 2ilxo- Ymll 1 + 211Xo- Yn11
l.(x,y) = 11-(x,y) + iv(x,y) = (px,y) + i(vx.y) if m,n-+ oo.
= (!n,y) + (il'x,y) = ((fl + il)x,J) = (}.x,y). Since -1-(y,. + y.) EM, we also have
The condition (iv) is easily verified, and (vl follows from the completeness 411xo- -1-(y,. + Y.)ll 2 ~ 4d 2 -
assumption on the space H. Thus the proof is complete.
Hence IIYm- Ynll-+ 0 if m,n-+ oo. Since H is complete and M is closed,
lim _r. = y 0 exists and belongs to M. Since llx 0 - Yoil =lim llxo :- Y.ll = d,
PROBLEMS
6.1.1. Let f E L2 (0,1 ). Show that for any positive integer n there exists (6.2.1) follows.
Suppose now that there exists another point y 1 satisfying
a unique polynomial Pn of degree :s;; n such that l!f- p I ::0: III- Pn ,, for all
,.-
polynomials p of degree :s;; n. , ,._ llx 0 - Y1ll = inf llxo- YII-
j ~~M
6.1. 2. Prove that the norm of U(O, 1) does not satisfy the pa ralle lo-
gram law if p =1- 2. Then
6.1.3. If {x"} is weakly convergent to x, in a Hilbert space X, and if
lim llxn = llxll, then lim x" = x.
n
211x 0 - Yo; Yl II:<::: llx 0 - Yoll + llxo- Y1ll .
6.1.4. Is the assertion of Problem 6.1.3 true in LP(O,l),p =1- 2?
s 2:~~
6.!.5. If {x.} is weakly convergent to x, in a Hilbert space fl, then 1
6.2 THE PROJECTION THEOREM Si nee the Hilbert norm is strictly convex (hy Corollary 6.1.3 ), we get:
x 0 - Yo= x 0 - y 1 . Thus, y 1 =Yo
Definition Let x.y be two points in a Hilbert space H. If (x,_r) = 0, We shall now prove the projection theorem.
then we say that x is orlhogotmf to y and write x 1 _r. Let .H be any suhset
of H. If x is orthogonal to all the elements of ;\1, then we say that x-is ortlwg- Theorem 6.2.2 Let M be a closed linear subspace of a IIi/bert space H.
i- Any x 0 E H can be written in the form x 0 =Yo+ z 0 , where Yo EM, ZoE MJ..
ona/ to M, and write x .l i\-1. Let l\-' be a suhset of H. If x 1 )H for all x E N,
then we say that,\' is orllrogonafto A!, and write .N .l IH. We then also have The elemenls y 0 ,z 0 are uniquely determined hy x0 .
206 HILBERT SPACES AND SPECTRAL THEORY
THE PROJECTION THEOREM 207
Proof If Xo EM, then Yo= x 0 ,z0 = 0 give the asserted decomposition Consequently,
of Xo. If x 0 M, let Yo be the point in M satisfying
The existence of J.'o is asserted in Lemma 6.2.1. Take now any point y EM Taking z = [ii/(z 0 ,z 0 )]z 0 , we get x*(x) = (x,z) for all x E H.
and any scalar A. Then Yo + J.y E M. It follows that Suppose that there is another point z' such that x*(x) = (x,z') for all
2
llxo - Yoll :=::; llxo- Yo- 2.rll
2
= llxo- J'oll 2 - 2 Re J.(y,x 0 - y0 ) + f,WIIYII 2 x E H. Then (x,z- z') = 0 for all x E H. Taking x = z- z', we find that
liz- z'll = 0--that is, z = z'.
Hence Next,
-2 Re J.(y,xo - Yo) + jJ.[ 2 IIY 11 2 ~ 0.
Taking J. = t > 0, dividing by t, and letting c--> 0, we obtain the inequality
llx*ll = sup lx*(x)l = sup l(x,z)l s sup (llxllll=ll) = 11=11
11-"ll=l llxll=l llxll=l
Re (y,x 0 - y 0 ) ~ 0. On the other hand,
If we take J. = - ie, e > 0, then we obtain, after dividing bye and letting ll=il 2 = (z,z) = lx*(z)l :=::; ]lx*llllzl];
-->0, .
lm (Y,Xo - Yo) :=::; 0. hence llz ~ lix*]l. We have thus proved that llx*ll = 11=11. This completes
1
,',
Corollary 6.1.3 ~f M is a closed linear substance and M t= H, then Corollary 6.1.6 A Hilbert space H is reflexive. Thus, in particular,
there exists an element z0 #- 0 such that z0 .l M. H is weakly complete, and a set in H is weakly compact if and only if it is
bounded and weakly closed.
Indeed, there exists a point x 0 E H, x 0 M. Let x 0 =Yo+ z 0 , where
Yo EM, z0 E MJ.. Then z 0 .l M and z 0 #- o: .~
PrQof It is easily seen that the Banach space H* is a Hilbert space
with the scalar product (ux,ay) = (y,x). [In verifying the condition (iii)
Theorem 6.2.4 For every bounded linear functional x*on a Hilbert space of Definition 6.1.1 we use the fact that a(J.x) = A: ax.] Denote by '! the iso-
H there exists a unique element z of H such that x*(x) = (x,z) for all x E H, metric imbedding from the Hilbert space H* onto its conjugate H**. Thus
and llx*l: = !\zl:. (Tay)(ax) = (ax,ay) for all x,y in H. Now letfhe any element of H**. Then
f = '!O"J! for some y E Hand, for all x E H,
This theorem is cal!ed the Riesz theorem.
f( <1x) = ( r<1y)(ux) = (<1x,ay) = (y,x) = (ax)(y).
Proof Denote by N the null space of x*. N is dearly a closed linear It follows that f is the image of y under the natural in bedding K of H into
subspace of H. If A'= H. then x* = 0 and x*(x) = (x,O). If N =1- H, then by H**. Thus" maps H onto H**; His therefore reflexive.
Corollary 6.2.3 there exists a point z0 E N.1, z0 #- 0. Clearly o: = x*(z0 ) t= 0. The following extension of the Riesz theorem is called the Lax-Milgram
For any x E H, the point x- x*(x)z 0 jx is inN. Hence lemma.
(x - x*~)zo, z 0 ) = 0. Theorem 6.1.7 Let B(x,y) be a bilinear functional [that is, B(x,y) is
linear in x and B(x,y) is linear in y] in H x H, where H is a Hilbert space.
208 HILBERT SPACES AND SPECTRAL THEORY PROJECTION OPERATORS 209
Assume that Notice that the projection theorem states in effect that if M is a dosed
IB(x,y)l 5 CllxiiiiYII (C > 0), (6.2.2) linear subspace, then H = M $ M1.. Thus MJ. (M) is the orthogonal com-
plement of M (MJ.) in H.
IB(x,x)l ~ cllxll 2
(c > 0), (6.2.3)
for all x,y in H. Then for any bounded linear functional x* in H there exists a PROBLEMS
unique point z in H such that x*(x) = B(x,z) for all x E H.
6.2.1. If M and N are closed linear spaces and M .l N, then M EB N
is a closed linear space.
Proof For fixed y E H, IB(x,y)l s C'llxll (C' = CIIYII). Hence, by
6.2.2. Let .ll,f be any subset of a Hilbert space H. Then (MJ.)J. is the
Theorem 6.2.4 there exists a unique element Ty such that
dosed linear space spanned by M.
B(x,y) = (x,Ty).
T is a linear map of H into itself. Since
6.3 PROJECTION OPERATORS
II Tyll 2 = (Ty, Ty) = B(Ty,y) 5 Cll Ty IIIIYII,
Definition Let Af be a closed linear subspace. By the projection
IITyll s CIIYII-that is, Tis bounded. Next,
theorem, every x E H can be written uniquely in the form x = y + z, where
cllxll 2 s IB(x,x)l = l(x, Tx)l 5": llxllll Txll, y E M, z E M J.. The point y is cal led the projection of x in M, and the operator
so that II Txll ~ cllxll. It follows that T maps H in a one-to-one way onto a P given by Px = y is called the projeclion on .H. We also denote this operator
closed linear subspace H 0 of H. We claim that HI)= H. Indeed, otherwise by PM. We call M the subspace ofrhe projection P.
there exists (by Corollary 6.2.3) an element z #- 0 orthogonal to all the points
Tx, x EX In particular Definition Let T be a bounded linear operator in a Hilbert space
H. The adjoint T* of Tis defined by the relation (Tx,y) = (x,T*y) for all
0 = l(z,Tz)l = IB(z,z)l ~ cllzll 2 x,y in H. lf T = T*, then we say that Tis self-adjoint. When we speak of
Thus z = 0, which is impossible. self-adjoint operators we always mean bounded, linear self-adjoint operators.
Now let x* E H*. By Theorem 6.2.4 there exists a point z 0 E H such The present definition of adjoint operators is not the same as the one
that x*(x) = (x,z 0 ). Since T(H) = H, there exists a point z such that Tz = zl). given by Definition 4.13.1, with X = Y, in the case of a normed linear space
Therefore X. In view of Corollary 6.2.5, the two definitions are related by an isometri;
x*(x) = (x,Tz) = B(x,z). map a; a is additive and u(l T) = J.o-T.
Note that if T is a self-adjoint operator in a Hilbert space 11, then
To prove uniqueness, suppose B(x,z) = B(x,z') for all x E H. Then also (Tx,x) is a real number for any xE H.
B(x,z- z') = 0 for all x E H. Taking x = z- z', we obtain
0 = IB(z - z', z - z')l ~ cllz - z'll 2 PROBLEMS
Hence z' = z. 6.3 .1. Let T and S be bounded linear operators in a Hilbert space H,
and let }, be a scalar. Prove: (T + S)* = T* + S*, (TS)* = S*T*, (}.T)* =
Definition Let M and N be linear subspaces of a Hilbert space H .l:T*, I*= I, T** = T, IIT*II =II Til If, further, r-l is a bounded linear
and assume that every element in the vector sum M + N has a unique map (with domain H), then also ( T*)- 1 is a bounded linear map (with do-
representation of the form x + y, where x EM, yEN. Then we call M + N main ll) and (T- 1 )* = (T*)- 1 .
the direct sum of M and N. lf M .l N, then we write M + N also in the form 6.3.2. State Theorem 5.2.10 in a Hilbert space, using the definition
M $ N. Note that M $ N is the direct sum of M and N. If Y = M $ N, then of '1: self-adjoint operator as given in the present section.
we write N = Y 8 M and call N the orthogonal complement of M in Y.
Similarly we write M = Y 8 N and call M the orthogonal complement of Theorem 6.3.I A projection P is a self-adjoint operator satisfying
N in Y. P2 = P, and IIPII = 1 ifP #- O.
PROJECTION OPERATORS 21 I
,' 210 HILBERT SPACES AND SPECTRAL THEORY
It follows that Multiplying on the right by PM, we get 2PMPNPM = 0. Hence PMPN = 0.
2
Conversely, if PM P.v = 0, then also PN PM= 0. Jt follows that P = P. Since
P(..l1x1 + }.2 X:z) = AtYt + ..l2 Y2 = At Px1 + ..1. 2 Px 2 P is also self-adjoint, it is a projection. Finally, Px = PMx + P.'lx varies
over M ffi N as x varies in H. Hence, P = PMffiN
Thus Pis a linear operator. Next, P 2 x 1 = P(Pxd = Py = y = Px so that
2 . 2 2 2 1 1 1
P = P. Smce llxtll = IIYtll + llztll ;::::; 11Ytl! 2 = 11Px1 11 2 it also follows that
Theorem 6.3.5 The product of two projections PM and PN is a pro-
liP II .:=::; I. If P #- 0, there is a point x #- 0 in the subspace of the projection P.
jection if and only if they commute- that is, PM PN = PN PM. In that case,
Thus, IIPxll = llxll. This gives IIPII = I. It remains to show that P is self-
adjoint. This is proved as follows; PMPN = PM,.,N
(Px!,x2) = (y,,xz) = (yt,Yz) = (xt,Y2) = (x 1 ,Px 2 ). Proof If P =PM PN is a projection, then P* = P. Hence PM P,. =
P~-P'!t = PNPM. Conversely, let PMPN = PNPM = P. Then P* = P, so that
The converse of Theorem 6.3.1 is'also true. In fact, we have: P is self-adjoint. Also, P 2 = PMPNPMPN = PiiP~ = PMPN = P. Thus P
is a projection. Further, Px = PM(PNx) = PN(PMx). Therefore, Px EM n N.
Theorem 6.1.1 A self-adjoint operator P with P 2 = P is a projection. On the other hand, if x EM n N, then Px = PM(PNx) = PMx = x. Thus
p = PM,.,N
Proof Let M = P(H). M is a linear subspace of H. It is also closed.
Indeed, if Yn = Pxn,Yn-> z, then Pyn = P 2 Xn = Px. = Yn Hence z =limy = Definition A projection PL is a part of a projection PM if L c: M.
lim Py. = Pz EM. Next, since Pis self-adjoint and P 2 = P, "
Theorem 6.3.6 A projection PL is a part of a projection PM if and only
(x- Px,Py) = (Px- P 2 x,y) = 0 for ally E H,
if one of rhe following conditions holds:
so that x_-:- Px lies in MJ.. Therefore x = Px + (x - Px) is the unique
decomposition of x as a sum y + z withy EM and z E MJ.. This shows that (i) PMPL = PL;
P is the projection on M. (ii) PLPM = PL;
(iii) IIPLx II ~ II PM x\1 for any x E H.
Definition Two projections P 1 and P 2 are orthogonal if P,P 2 = 0. Since Proof Let L c: M. Then PLx eM for any x E H. Hence PMPLx =
(PIPz)* = P; P! = P 2 P 1,P 1 P 2 = 0 if and only if P2 P 1 = 0.
PL x, and (i) follows. If (i) holds, then
Theorem 6.3.3 Two projections PM and PN are orthogonal if and only PL =Pi.= (PM Ptl* = Pi.P'!t = PLPM,
ifM .lN.
and (ii) follows. Next, if (ii) holds, then, for any x E H,
Proof Let PM PN = 0 and let x EM, yEN. Then IIPLxll = IIPLPMxll ~ IIPd IIPMxll .:=::; IIPMxll,
(x,y) = (PMx,P.v y) = (PHPMx,y) = 0. so that (iii) holds. It remains to show that if (iii) holds, then L c: M. If this is
not true, then there is a point x 0 e L, x 0 ~ M. Let x 0 = Yo + z 0 where Yo eM,
Thus M .lN. Suppose conversely that M .lN. For any x E H, P.vx EN
z 0 .l M, z 0 #- 0. Then
and therefore (PNx) .l M. Jt follows that PM(P~X) = 0. Thus f:'MPN = 0.
2 2 2
11PLxoll 2 = 11Yoll + iizoll > IIYof = IIPMxoll
Theorem 6.3.4 The sum of two projections PM and PN is a projection
This contradicts (iii).
if and only if PMPr< = 0. In that case, PM + PN = PMffiN
t
~
212 HfLBERT SPACES ANO SPECTRAL THEORY ORTHONORMAL SETS 213
Proof I - P is self-adjoint and (I- P) 2 =I - P- P + P 2 =I- P. A set K in a Hilbert space H is called orthonormal if each element of
K is a unit element-that is, it has norm l, and if any two distinct elements
of H are orthogonal. An orthonormal set K is called complete if there exist
We shall now prove a theorem that contains the last lemma. no nonzero elements that are orthogonal to K -that is, if K j_ = 0.
Theorem 6.3.8 The difference P = PM- PL of two projections P'>I,PL Theorem 6.4.1 Let {x.} be an orthonormal sequence in a Hilbert space
is a projection 1jand only i/PL is a part ofP'>I. If this is the case, then P = PMeL H. Then, for any x E H,
(1- PM)PL = 0-that is, PL =PM PL. Hence, by Theorem 6.3.6, PL is a part
of PM. Suppose conversely that Pr. is a part of PM. Then (I- PM)PL = 0. This inequality is called Bessel's inequality. The scalars (x,x") are called
Hence, by Theorem 6.3.6, (I- PM)+ Pr. is a projection. Thus I - P is a the Fourier coefficients of x with respect to the sequence {x"}.
projection. By Lemma 6.3.7, also I - (I- P) =Pis projection.
Finally, if PL is a part of PM, then PM- P 1 and PL are orthogonaL Proof We have
By Theorem 6.3.4, the subspace Y of the projection PM - P L satisfies
Y$L = M. Hence Y = M 8L This completes the proof.
PROBLEMS
6.3.3. If Pis self-adjoint and P 2 is a projection, is P a projection?
6.3A. Let (X,J.l) be a measure space and denote by XE the character-
istic function of a measurable set E. Then the operator QrJ = XEf defined Therefore,
in L 2 (X,p) is a projection. Under what e-ondition onE, F is QE + QF a pro-
jection? (6.4.2)
6.3.5. Consider the operator ( Q/)(1) = a(t)f(t) in L 2 (0,1), where a(t)
is a scalar function. Find necessary and sufficient conditions on a(t) for Q
to be a projection. It follows that
6.3.6. Let A be an 11 x 11 matrix with real elements and denote by Q
the operator Qx = Ax in Rn. Find necessary and sufficient conditions on
A for Q to be (i) self-adjoint, (ii) a projection.
L"' l l(x,x.W ~ llxll 2
n~
= llxll f lc,l + f
2
-
n~ 1
2
n= 1
lc,-..1.,1 2
converges independently of the order by which the terms are taken, and
(iii) E is the projection on the closed linear space spanned by K.
(6.4.3)
and the sum I lln x, is independent of the order in which its terms are arranged. Theorem 6.4.2 then gives
Proof If m > n,
(6.4.6)
(6.4.4) We may assume that y1 E K, for I sj s n [otherwise we strike out from the
last inequality those y i for which (x,y 1) = 0 ]. Arrange the set K ~ in a sequence
From this relation and from the completeness of H, we get the first assertion {y), where I sj s oo. From (6.4.2) we see that the left-hand side of (6.4.6)
of the theorem. Taking n = I, m---+ oo in (6.4.4), we obtain the assertion does not increase with n. Hence, taking n ..... oo, we get llx- Exll <e. Since
(6.4. 3). Now let z = I o: i. xi. be a rearrangement of the series x = I r1. . x _ e is arbitrary, Ex= x for any x E K. We have thus proved that E is the pro-
and assume that I jr1.1 j2 < oo. Then 1 1
jection on R.
Ux- zjj 2 = (x,x) + (z,z)- (x,z)- (z,x). (6.4.5)
Definition A set K is called an orthonormal basis of H if K is an
One easily verifies (x,x) = (z,z) =I la) 2 . Writing orthonormal set and if for any x E H
m
s,. = I
j-1
t:i.jXj, t., =
P1-1
I'" t:J.i.xi. x = I
JE=K;~e
(x,y)y. (6.4.7)
we also have
ro
Theorem 6.4.5 Let K be an orthonormal set in a Hilbert space H. Then
(x,z) =lim (s,.,t.. ) = jo:i1 2 I the following conditions are equivalen 1:
"' J-1
(i) K is complete;
Also (z,x) = (x,z) = (x,z). We conclude from (6.4.5) that llx- zl1 2 = 0- (ii) The closed linear subspace spanned by K isH;
that is, z = x. This completes the proof of the theorem. (iii) K is an orthonormal basis;
(iv) For any x E H,
Theorem 6.4.4 Let K be an orthonormal set in a Hilbert space H: Then, t
(i) for any x E H, (x,y) = 0 for all but a countable number of points llxll 2 = I
)'!f::K~
l(x,y)] 2 (6.4.8)
yofK;
The relation (6.4.8) is called Parseval's formula.
216 HILBERT SPACES AND SPECTRAL THEORY ORTHONORMAL SETS 217
Proof By Corollary 6.2.3, (i) implies (ii). Suppose (ii) holds. Then, by PROBLEMS
Theorem 6.4.4, Ex= x for any x e H-that is, K is an orthonormal basis.
6.4.1. If Tis an isometric isomorphism of a Hilbert space H 1 onto
Suppose next that (iii) holds and arrange the elements of K"' in a sequence
a Hilbert space H 2 , then also (Tx,Ty) = (x,y) for all x,y in H1.
{x.}. Taking in (6.4.2) n ..... oo, we then obtain (6.4.8). Finally, if (iv) holds and
6.4.2. Let H 0 be a linear subspace of a Hilbert space H, spanned by a
if x .l K, then llxll 2 = L l(x,y)l 2 = 0--!hat is, x = 0. Thus (iv) implies (i).
sequence {xm} of linearly independent elements. Show that there exist
Theorem 6.4.6 In every Hilbert space there exists an orthonormal basis. scalars .l.m such that the sequence {ym} given by
where x = (xu ... ,x.), y = (y 1 , ... ,y.). We call H the direct sum of
the balls B are mutually disjoint. If there exists a dense sequenoo {y.} in the Hilberr spaces H 1 , ,H. and write H = H 1 EB EB H . Show that
the space, then there is a ball B.o that does not contain any of the points y . the map x ..... (O, ... ,O,x;,O, ... ,O) is a projection in H.
Hence x'"" is not in the closure of {y.}-a contradiction. 6.4.4. Let f(z),g{z) be complex-valued functions in a closed disc
D: izl : : ; r, and assume thatfe C0 (D) and g is analytic in D. Prove that
Theorem 6.4.8 Any two infinite-dimensional separable Hilbert spaces
are isometrically isomorphic. JJ. f(z) g'(z) dx dy = 2t~ J f(z) g(z) dz.
D
i!D
Proof Denote two such spaces by H 1 and H 2 . By Lemma 6.4.7, there 6.4.5. Let D be the disc lzl < 1 in the complex plane. Denote by
exist sequences {x.} and {y.} that form orthonormal bases for H 1 and H 2 , H2(D) the linear subspace of L 2 (D) consisting of_ al_l funct_ions holomorphic
respectively. For any point x e H 1 andy e H 2 we have in D. Prove that H 2 (D) is a closed set, so that 1t 1s a Htlbert space. Next
prove that the sequence { ,j n:i~z" -t} is orthonormal in H 2 (D) and that it is
(6.4.9) also complete. [Hint: the nth Fourier coefficient is equal to
. 1 r2 "
Fn 21 Jlzl~r -z"d z .
where c. = (x,x.), d.= (y,y.). When {c.} and {d.} vary in / 2 , the points x f(z)
a =ltm---
andy vary in H 1 and H 2 , respectively. We define a map Tfrom H 1 into H 2 " r~l
as follows: y = Tx if, in (6.4.9), c.= d. for all n ~ L Tis clearly linear,
and it maps H 1 onto H 2 . Since On the other hand, if /(z) = L b" z",
11Txll 2 = f ld.l 2 = f !c.l 2 = lfxll 2 , b
n-l
= _1 f
2ni [zl= z"
f(z) dz-
'
n~ l n~ I
Tis also an isometry. This completes the proof. Now verify Parseval's formula.]
6.4.6. Show that an orthonormal sequence {If'.} IS complete in
Coroll4ry 6.4.9 Any infinite-dimensional separable Hilbert space is
isometrically isomorphic to L 2 (0,1). L 2 (a,b) if n~l (r !p.(t) dtr =X- a for all X E (a,b).
218 HILBERT SPACES AND SPECTRAL THEORY SELFADJOINT OPERATORS 219
6.4.7. If {x"} is a complete orthonormal sequence in a Hilbert space Proof If A. is not a real number, then, by Theorem 6.5.2, E.l. = {0} =
Hand if {y.} is an orthonormal sequence in H satisfying 1 . Hence the assertion of the lemma is equivalent to the assertion
L"'
n=1
llx.- Y.ll < I,
2 H = E1 fBR~.
If x e E;;, y e Rl, then (A - Xl)x = 0 and y =(A - ...ll)z for some z e H. It
then {y.} is also complete. follows that
(x,y) = (x,(A - ..U)z) =((A - XI)x,z) = 0.
6.5 SELFADJOINT OPERATORS If y e R.l, then y = lim Yn for some sequence {y"} in R.l.. Since (x,y.) = 0 for
any x e E;: and for any n ~ 1, we get (x,y) =lim (x,y.) = 0. We have thus
Throughout this section we designate by A a self-adjoint operator in proved that El .1 Rl' It remains to show that if x .l (Ex. fB R.l.), then x = 0.
a Hilbert space H. For any z, (x,(A - ..ll)z) = 0. Hence ((A - li)x,z) = 0. Since z is arbitrary,
(A - XI)x = 0. Thus x e E;;. But since x l. E;;, it follows that x = 0.
Theorem 6.5.1 I Ail = sup I(Ax,x)l.
llll =1 Corollary 6.5.4 If E). = 0, then R). = H. Thus, in particular, A has
no residual spectrum.
Proof Let ':1. = sup I(Ax,x)j. If iixll = 1, then
11<11 =I
Lemma 6.5.5 If, for some A,
I(Ax,x)l :=::; IIAxllllxll = I Axil:=::; IIAIIIIxll = IIA 11.
I Ax- ...lxll ~ cllxll (c > 0) (6.5.1)
Hence IX:=::; IIA 11. To prove the converse inequality, let z e H, Az i= 0 and
introduce u = (Az)/..1. where A.= (11Azll/llz11) 112 Then (compare Problem 6.3.7) for all x E H, then Abelongs ro the resolvent set p(A) of A.
jjAz11 2 = (A(A.z),u) = !{(A(...lz + u),Az + u)- (A(J.z- u),..lz- u)} Proof The assumption (6.5. i) implies that E;. = 0. Hence, by Corollary
:-: ; t1X(II..lz + ull + IIA.z- ull 2 ) = tC!(IIhll 2 + llull
2 2
)
6.5.4, R..l. =H. If we prove that R~ is closed, then, by Theorem 4.6.2, it
follows that (...ll - A)_, is a bounded operator-that is, ...l e p(A). To prove
that R~ is closed, let {y,} c: Rl, Yn _,. y. Then Y. =(A - ..li)x~ for some
= }C!(A 2 IIzll 2 + ; 2 11Az11 2) = C!llzll iiAzll.
Xn E H, and
Hence II Azll :=::; cillzll. It follows that IIA II :=::; !Z. lly,- Ymll = II(A- J.l)(x.- x,.)ll ~ cllx,.- x,.ll.
It follows that {x.} is a Cauchy sequence. Let x =lim x . Then y =
Theorem 6.5.2 The eigem;alues of a self~adjoint operator are real lim (A - .U)x. =(A - ...li)x-that is, y e R...
numbers. Eigenuectors corresponding to distinct eigenvalues are orthogonal.
Corollary 6.5.6 A point A belongs to the spectrum cr(A) of A if and
Proof Let Ax= ...lx, x #- 0. Then (Ax,x) = A.llxll 2 . Since (Ax,x) is real only if there exists a sequence {xn} in H such that
and llxll #- 0, we conclude that ..l is real. Next let Ay =p.y, where y #- 0
and J1 #- L Then J1 is a real number, and we have lim I Ax.- ..lx.ll = 0.
Hence Corollary 6.5.6 then implies that Me o-(A). The proof that me o-(A) is
(x,y) - (y,x) = (..1. - .l:)(x,x) = 2villxll 2
similar. In fact, one takes a sequence {y.} c: H with I!Y.II = I, (Ay..,y.. ) =
m + e., e. 2 0, e .... 0, and proves that
Therefore
IIAy. - my.ll 2 = II(A - yl)y, - (m - y)y.ll 2 -> 0 if n-> oo.
21vlllxll 2 s l(x,y)l + l(y,x)l = 21(x,y)l s 211xiiiiYII,
-that is, lvlllxll s IIYII = II(A- U)x]l. Now apply Lemma 6.5.5. PROBLEMS
Definition The upper bound of a self-adjoint operator A is the number 6.5.1. Let B(x,y) be a complex-valued functional defined on H x H.
M = sup (Ax,x). The lower bound of a self-adjoint operator A is the number We call it a bilinear form if it is linear in x and antilinear in y [that is, B(x,y)
llxll-1 is linear in y]. If IB(x,y)l s Cllxllllyll for all x,y in H, then we say that
m = inf (Ax,x). B(x,y) is a bounded form. If B(x,y) = B(y,x), then we say that B(x,y) is a
llxll-1 Hermitian form. Prove that if B(x,y) is a Hermitian form, then there exists
Note, by Theorem 6.5.1, that I All =max (lmi,IMI).
a self-adjoint operator A such that B(x,y) = (Ax,y).
6.5.2. If B(x,y) is a Hermitian bilinear form, then B(x) = B(x,x) is
Tlu!orem 6.5.8 The spectrum of a self-adjoint operator A lies in the
called a quadratic form. Prove that
closed interval (m,M], where m and M are the lower and upper bounds of
A, respectively. 4B(x,y) = B(x + y) + B(x- y) + iB(x + iy) - iB(x - iy).
Proof Suppose . 1. > M. Then 6.5.3. Consider the operator (Ax)(t) = tx(t) in L 2 (0,1). Prove that
a( A) consists of the closed interval [0, I]. [Hint: If 0 < ..t < I, take x.( t) = ...;-;;
-((A- ..l.I)x,x) = -(Ax,x) + ..l(x,x) 2 (..1.- M)(x,x).
for . 1. < t <A+ (1/n), x.(t) = 0 if t [..1.,..1. + (1/n)] and use Corollary 6.5.6.]
It follows that 6.5.4. If A is an operator in Hand L is a closed linear subspace of H
such that A(L) c: L, then Lis called an invariant subspace of A. Prove that if
(..t- M)llxll 2 s I((A- Al)x,x)l s II(A- ..ll)xllllxll. Lis an invariant subspace of a self-adjoint operator A, then also LJ. is an
Hence II(A- ..ll)xll 2 (...1.- M)llxll. Lemma 6.5.5 then implies that ). e p(A). invariant subspace of A.
Similarly one proves that if . 1. < m then A e p(A). 6.5.5. Let L be a closed linear subspace, invariant with respect to a
self~adjoint operator A. Denote by a 1 (A) and o- 2 (A) the spectra of the re-
Tlu!orem 6.5.9 The lower and upper bounds, m and M, of a self-adjoint striction of A to Land L \respectively. Prove that a(A) = u 1(A)v o-z(A).
operator A belong to the spectrum of A. [Hint: Prove o-(A) c: o- 1(A) u o- 2 (A) by using Lemma 6.5.5, and prove
o-(A);:::, o- 1(A)u aiA) by using Corollary 6.5.6.]
ProofBy definition of M, there exists a sequence {x.} in H such that
=M-o., o. 2 0, o.. -> 0. Let y >max (IMI,Iml). The
llx.ll = I, (Ax x.)
upper bound of A + yl is M + y and the lower bound of A+ yl is m + y. 6.6 POSITIVE OPERATORS
Since M + y 2m+ y > 0, it follows, by Theorem 6.5.1, that the norm of
A+ yl is equal to M + y. Hence II(A + yl)x.ll s (M + y)llx.!! = M + y. We A self-adjoint operator A is ca!led positite if (Ax,x) 2 0 for all x E H.
now have If A is positive, then we write A 2 0. If A and B are self-adjoint operators
I Ax.- Mx.ll 2 = II(A + yl)x.- (M + y)x.ll 2 and if A - B 2 0, then we say that A is larger than B and that B is smaller
2 2 than A, and write A 2 B and B ~A. Note that for any bounded linear
= II(A + yJ)x.ll 2 + (M + y) llx.:: - 2(M + y)
operator T, TT* and T*Tare positive operators.
x ((A + yl)x,.,x.)
s (M + y) 2 + (M + y) 2 - 2(M + y)(M - b. + y)-> 0 Lemma 6.6.1 Let A be a posit ire operator. Then there exists a sequence
if n-> oo. of operators An, which are polynomials in A with real coefficients, such that
POSITIVE OPERATORS 223
222 HILBERT SPACES AND SPECTRAL THEORY
Therefore,
n Theorem 6.6.4 Every positi11e operator A has a unique positive square
L
m~l
(B, x,B,x) s (B 1 x,x). root B. B commutes with any bounded linear operator that commutes with A.
It follows that liB, xll -+ 0 as m-+ oo. Consequently, (6.6.3) gives Proof Suppose first that A.:<:;; 1. Set B0 = 0 and
......
adjoint operator B such that lim B,.x = Bx for any x E H. Applying both To prove (iii), note that, for any x E H,
sides of (6.6.4) to x E H, and letting n--> oo, we obtain
(A - B)(A + B)x = (A 1 -
2
B )x = 0.
Bx = Bx + }(A 2
- B )x
1
Hence (A + B)x e L. It follows that P(A + B)x =(A + B)x-that is,
-that is, 8 =A. Since each B" is positive, also B is positive. Since each
P(A + B)= A + B. Since (compare Lemma 6.5.3) the null space of A - B is
B" commutes with every operator that commutes with A, the same is true
orthogonal to the range of A - B, P(A - B) = 0. Hence
of B.
We have assumed, so far, that A :::; I. In the general case we consider P(A +B)- P(A- B) =A+ B
the operator e1 A, where e > 0, e2 il A I < I. Then e2 A :::; I, and it therefore
has a positive square root B'. Take B = 8'/e. -that is, A = (2P - l)B.
To prove uniqueness, suppose D is another positive square root of A. We shall use Lemma 6.6.5 to prove the following result, which will be
Since A= D 2 ,D commutes with A. Hence B commutes with D. For any of fundamental importance for the development of the spectral theory given
x E H, let y = (B- D)x. Then in the next section.
(By,y) + (Dy,y) = ((B + D)y,y) = ((B + D)(B- D)x,y)
Lemma 6.6.6 For any self-adjoint operator A there exists a projection
= ((B 2 - D 2 )x,y) = 0. E+ having the following properties:
Since (By, y) 2 0, (Dy, y) 2 0, it follows that (By, y) = (Dy, y) = 0. Let (i) any bounded linear operator that commutes with A commutes
C be a positive square root of B. Then
also with E + ;
I C..vll 2 = ( C 2 y,y) = (By,y) = 0. (ii) AE+ 2 0 and A(\ - E+):::; 0;
(iii) If Ax = 0, then E+x = x.
Hence Cy = 0. We conclude that By= C(Cy) = 0. Similarly, Dy = 0.
Therefore
Proof. Let B be a positive square root of A 2 and let E + be the pro-
IIBx- Dxll 2 = ((B- D?x,x) = ((B- D)y,x) = 0 jection operator on the null space of A -B. Then (i) and (iii) follow from
-that is, Bx = Dx. Since x is a~bitrary, B = D. Lemma 6.6.5. This lemma also gives A= (2+ - I)B. Hence
be the partition whose points are obtained by combining the points that
6.7 SPECTRAL FAMILIES OF SELF-ADJOINT occur in both partitions II 1 and IT 2 We shall estimate IISrr- Srr,ll.
OPERATORS Let IT 1 be given by 20 < 21 < < ..1.. and Jet IT be given by
Let {E.d be a family of projections for -co < l < co satisfying the ..1.o < 2o,t < ... < 2o,tn = 2, < A.t.t < ... < A.,,k, = 12 < ... < 2.-t.k.-t = 1.
foJlowing conditions: In Srr, there occur the points J.lt :::;; J.1 2 :::;; :::;; 11. with J.ik E Pt-t>lJ. Anal
(a) E ... s;;EI'if.l.<J.i; ogously in Sn there occur the points
(b) EA is strongly continuous from the left-that is,
llo,J :::;; J.lo,2 :::;; .:::;; J.l.o,t 0 ~ J.I.Lt :::;; J.l.1,2 :::;; :::;; 11t,k 1 :::;; J.l2,1 : : : ; :::;; J.l.-1,kH- 1
for any x E H;
We have
(c) EA = 0 if 2 < m, EA =I if 1 > M. Srr =
n-1
L L
'
f(J.I.;)[(1;)- E(A.;,j- t}],
;-o i-t
We call such a family of projections a spectral family or a resolution
of the identity. .where . 1. 0 , 0 = 10 ..i.t.o = . 1. 1 , ... .1.-t.o = ..1.,._ 1. We can write Sn, in the form
Let/{..1.) be a continuous complex-valued function form:::;; . 1. ~ M. We n-1 ii
extend/(1) toM < ..i. < M + l so that the extended function, which we again Srr, = L L f(I1;+J)[E(l;,)- E(J.;.j-1)].
denote by /{..1.), is continuous form :::;; 1 < M + I. We shall define an integral i=O j-1
"'
Take IT to be any partition of [m,M + e]: m = . 1. 0 < 11 < <
Hence
1.-t < 1. = M +e. Let ak = (..1.,._ 1 ,1t], E(at) = E(.-1.")- E(}."_ 1) and take
points Jl.k in &,.. Form the sum n-1 '
Srr, :::;; Sn + fl L I
i-0 j= 1
[E(1;)- E(2;,j- 1 )] = Sn + fll.
Srr = I" f(J.I,.)E(a,.).
Similarly Sn:::;; Sn, + fll. Hence -fll s;; Srr- Sn, s;; fll. It follows (compare
t-1
Problem 6.6.4) that IISn- Sn,ll s;; fl. Similarly IISn- Srr,ll s;; fl. We conclude
Denote by III I the mesh of IT-that is III I = max (..1." - 1,._ Jl.
k that IISn,- Sn,ll :::;; 2fl. Thus S = lim Sn exists, in the uniform topol-
IIII~o
ummo 6.7.1 If IIII-+0, Sn converges, in the uniform topology, to a ogy. Furthermore, S is independent of the choice of the points llt Since
bounded linear operator S that is independent of (i) the continuous extension EA - E, = 0 if M < J1 < ..1., it is readily seen that S is independent of the
off; (ii) the choice of; (iii) the choice of the points Ilk choice of e and of the extension off This completes the proof.
Note that I= J~+ dEw
Note that the limit of Sn, as III I -+ 0, is taken in the following sense: We now state the main result of the present section.
o
for any f1 > 0 there exists a > 0 such that IISn - Sll < fl whenever !Ill < /5.
We write Theorem 6.7.2 To e!;ery self-adjoint operator A there corresponds a
unique family {Ed of projections, - oo < 1 < oo, satisfying the following
conditions;
and call it the integral off with respect to the spectral family {El}. (i) Any bounded linear operator that commutes with A commutes
also with the E~;
Proof It suffices to consider the case where f(..l.) is real-valued. For (ii) E~ s E, if A. < J..l;
o.
any f1 > 0 there exists a b > 0 such that lf(..l.) - f(..1.')1 < f1 if 1..1. - ..1.'1 < (iii) E~ is strongly continuous from the left-that is, for any x E H,
Let II1 ,II 2 be two partitions of [m,M + e] with lll 1 1< /5, III 21< o. Let II lim E~x = E~x. Furthermore, E~+oX =lim E~x exists for any x E H.
~I'~ A':,p
228 HILBERT SPACES AND SPECTRAL THEORY SPECTRAL FAMILIES OF SELF-ADJOINT OPERATORS 229
(iv) El = 0 if A. < m and E, = I if A. > M, where m and M are the decreasing function of i.. Therefore lim (E., x,x) exists. It follows that
lower and upper bounds of A; Ai'l'
(v) A is gi~en by I E, x - E;. x II~ = ((E, - E;.)(E,. - E;.)x,x) = ((E. - E.,)x,x)....,. 0
if . 1. < v < fl, }, ....,. 11 Thus
(0 < e < 1). (6.7.1)
lim E;.x = EroX exists.
.l;'p
The family {;.} is ca11ed the spectral family of A, or the resolution of
the identity corresponding to A. Similarly,
lim E;.x = Eu+oX exists.
Proof Denote by +(..1.) the projection operator constructed by '"'-~'
Lemma 6.6.6 for A - J.I. The operators E.,= 1- +(..1.) are projections. We 'fo complete the proof of (iii) it remains to show that the operator E(L\ 0 ) =
s~all prove that they satisfy the conditions (i)-(v). Since +(..1.) commutes E~- E~-o is zero.
w1th any operator that commutes with A, the same is true of .. Hence (i) For any x E H,
is satisfied. In particular it follows from (i) that E;. E,_. = ,.."for Aall )., 11 .
To prove (ii), let J. < 11 and set P = E~(I - ~'). We have as ), /' 11
From (6.7.4) we then conclude that the operator pE(8. 0 ) - A(8. 0 ) is both
(6. 7.2)
positive and negative. Hence (cf. Problem 6.6.4) 11(8. 0 ) - A(8. 0 ) = 0.
From the definition of E;. we also have Hence, for any x E H, and y = E(8. 0 )x, (A - pl)y = 0. By the assertion (iii)
of Lemma 6.6.6 we conclude that E~' y = 0-that is, E~E(8. 0 )x = 0. Hence
(A - J.I)E;.:::; 0, (6.7.3)
E(8. 0 )x =lim E(h..)x =lim E~ E(L\)x = E~ E(A0 )x = 0.
Take any x E Hand let y = Px. Then (6.7.2) implies that A/'~ .!i'l'
EJ.y = E.,Px = Px = y We have thus proved that E(A 0 ) = 0. This completes the proof of (iii).
To prove (iv), let A < m and suppose that E;. # 0. Then there exists a
and, similarly, {1 - E~')y = )'. Using (6. 7.3), we get point x E H such that E;.x - 0. Let y = E;.x. We may assume that IIYII = 1.
Then
((A - A.l)y,y) =((A - AI)E;, y,y).::::; 0,
(Ay,y)- l =((A- J.I).v,y) =((A- U)E;_y,y) .:$0
((A - pl)y,y) =((A - 111)(1 - E,.)y,y) 2 0.
It follows that (A- 11)(y,y) 2 0. Since ). < J.l, (.v,y) :::; O~that is, Px = .v = 0. by (6.7.3). Thus m s (Ay,y) s ..1.-a contradiction. Thus E;. = 0 if A< m.
We have thus proved that P = O~that is, E;. = E;. Ew But this implies Suppose next that ), > M. If E;. -1= !, then there exists a point x E II such that
(compare Problem 6.6.3) that E.,.::::; Ew Thus (ii) is satisfied. z =(I - E;.)x -1= 0. We may assume that llz',l = 1. Then
Let . 1. < J.l, 8. = [.l.,p],(8.) = E~'- E.,. Then (Az,z)- A= ((A - ..ll)z,z) =((A- ..1.1)(1- E;.)z,z) ~ 0
E~E(A) = (8.), (f- ;,)(8.) = (8.). by (6.7.3). Hence M ~ (Az,z} ;;::: J.~a contradiction.
Therefore, by (6.7.3) (with . 1. and 11 interchanged) and the fact that (8.) ~ 0, We shall now prove (v). Take a partition
(A - pi)(8.) =(A - J.l/)~' E(h..) ::<:;; 0, m = Ao < . 1. 1 < < /."_ 1 <A"= M +c (0 < r. < 1),
(A - ).I)E(d) =(A - Al)(J - ..)(8.) ~ 0. and set 8.x = (;x-t.Ak], E(Ak) = E(Ak)- E(..lk- 1 ). From (6.7.4) and the
Consequently, relation I= " (8.x), we get
L
k~ I
A(8.) .::::; A E(h..) ::--; p(8.). (6.7.4) n n
Now take a sequence of partitions with mesh going to zero and use Lemma To prove (6.7.6) we take sums S 0 = Lp(p.)(6l) that converge to
6.7.1. p(A) as ITII-+ 0, and notice that
It remains to prove uniqueness. Before proving it, we shall establish (Sn x,x)-+ (p(A)x,x),
the following lemma.
M+e
(S 0 x,x) = L p(tlk)(E(AA)x,x)-> ( p(A.) d(E.J.x,x),
m
Lemma 6.7.3 Let A be a self-adjoint operator and let {Ed satisfy
as ITII->0.
(i)-{v). Then ,for any polynomial p(l),
We can now prove the uniqueness assertion of Theorem 6.7.2. Suppose
,M+e {Fl} is another family satisfying (i)-{v). By Corollary 6. 7.4, for any poly-
p(A) = jm p(}.) dE;.- (6.7.5) nomial p(l)
M+
n
Proof For any '1 > 0 there is a o> 0 such that any finite sum S = J.. p(.l.) du(A) = 0, (6.7 .7)
L
kM\
Ak E(~,J with max V..
k
- ..1.k_ 1 ) s o satisfies where a(A.) = (Elx,x)- (Flx,x) is a function of bounded variation. Since
n any continuous function can be approximated uniformly by polynomials, we
IIA-2: .l.kE(~k)ll::s:'l, obtain (using Problem 2.11.10) the relation (6.7.7) for any continuous
k~l
function p(.l.). Since a(m) = 0 and since a(.l.) is continuous from the left, it
or, n
follows (by Problem 4.14.3) that u(.l.) = 0. Thus (.. x,x) = (F, x,x). The
II [A - 2: ).kE(8.k)f II ::s: '1 2 operator El - Fl is self-adjoint, and we have just proved that its lower and
k~l
upper bounds are equal to zero. Hence, by Theorem 6.5.1, E.~- Fl = 0.
Thus
This completes the proof.
IIA
2
+ Lt AkE(Ak)]
2
- 2ASII $ 2
1j
define
Not11.tion Let f(A.) be any continuous function for m sA sM. We
Using the relations
2
if k '# j and liAS- A 11 ::s:IIAIIIJ. we find that From Lemma 6.7.3 it follows that if f(.l.) is a polynomial I: a. A.", then
n
L
f(A) = a" A".
I: lfE(Ad II::;;
II A 2 - kMI Clj.
One easily gets, by the argument used in proving Corollary 6.7.4,
M+
=J
2
This gives (6.7.5) with p{J.) = ..1. . Similarly one can prove by induction that
(6.7.5) is valid for p(A) =A", n any positive integer. From this (6.7.5) follows
f(A)x
.. f(.l.) dE.J.x, (6.7.9)
M+<
for any polynomial. (f(A)x,x) =J f(.l.) d(E .. x,x). (6.7.10)
"'
Corollary 6.7.4 For any x e Hand for any polynomialp(J..), Theorem 6.7.5 Let f(J..), f1 (J..), fi11.) be continuous functions and let
M+e a1 ,a 2 be scalars.
(p(A)x,x) = J p(.l.) d(E.J.x,x). (6.7.6) (i) If f(l) = a 1f 1(J..) + a 2 f2 (l), then f(A) = a 1f1 (A) + al2(A).
"' (ii) Jff(J..) = f 1 (J..)f2(J..), then f(A) = f 1(A)fiA).
The integral on the right is defined analogously to the definition of (iii) [f(A)]* = i(A), where l'(J..) = f(l).
dE;.. Since (E~x,x) is monotone nondecreasing, this integral coin-
J::!+ p(J,) (iv) llf(A)II s max lf(J..)I iff is real-ralued.
,.,,;;l,;;M
cides with the Riemann-Stieltjes integral (defined in Problem 2.11.8).
232 HILBERT SPACES AND SPECTRAL THEORY THE RESOLVENT OF SELF-ADJOINT OPERA TORS 233
(v) If B is any bounded linear operator that commutes with A, then B 6.7.2. LetA be a self-adjointoperatorand let {Ey} be its spectral family.
also commutes with f(A). Then, for any continuous functions f,g,
M +
Proof
E with
Consider partitions IT: m = . 1. 0 < }. 1 < < ..1..- 1 < ..1..
IIII-+ 0, and let E(Ak) = E;.k- E;.,_,. Then
= ( t M+e
f(J.) dE~.x, L
M+<
g(..l.) dE;.x =
)
t .M+e
f(..1.)g(..1.) d(E, x,x).
= l:f(..t.)E(A.).
[Lf(..1..)E(A.)]"'
To prove (iv) note thatf(A) is self-adjoint. Also, by (6.7.10),
(Ax,x) = f
n=l
..1.. ](x,x.)l 2 = f
n=l
..1..( P. x,x).
1
Since the last integral is equal to I x II, we obtain, upon taking B -+ 0, R(/..;A)x = J:"' (x,x.). 1. x. = J:"'
1
..1.. _
1
;" _ . 1. P.x;
PROBLEMS
6.7.1. Let A be a self~adjoint operator and let {E;.} be its spectral 6.8 THE RESOLVENT OF SELF-ADJOINT
family. Let E(A) = E(JJ) - E(o:), where o: < p. Prove that for any continuous OPERATORS
function /(..1.),
Throughout this section A denotes a self-adjoint operator in a Hllbert
E(A) [ f.. M+<
j(..1.) dE.<
] p
= { j(..l.) dE;,. space H. The lower and upper bounds of A will be denoted by m and M,
respectively. We denote by o{A) the spectrum of A and by p(A) the resolvent
234 HILBERT SPACES AND SPECTRAL THEORY THE RESOLVENT OF SELF-ADJOINT OPERATORS 235
of A. Recall that A E p(A) if and only if the resolvent R(l;A) = (U- A)- 1 We shaJI now prove the converse of Theorem 6.8.1.
exists (as a bounded linear operator in H).
By Theorem 6.5.8 the resolvent exists for all complex A not in the reaJ Theorem 6.8.2 If /.. 0 E [m,M] and R(/.. 0 ; A) exists, then there exists
interval [m,M]. By Theorem 6.5.9, the resolvent does not exist at the points an interval (et,IJ) such that et < ). 0 < !i and ~ = const. for all et < A < fJ.
A= m and A= M-that is, m and M belong to a(A). We shaJI characterize,
in this section, the points A of [m,M] for which R(A;A) exists-that is, the Proof. Let A= [a:,,B] be an interval with . 1. 0 as its midpoint, and let
points A which belong to p(A). E(&) = E({J) - E(a:). Applying R().0 ;A)E(.6.) to both sides of the identity
Theorem 6.8.1 Let J... 0 E [m,M]. If there exists an interval (et,fJ) with
et < J... 0 < IJ, such that E~ = const. for et < ). < IJ, then ). 0 E p(A) and
(A- A0 I)x = f..
M+<
(A- A0 ) dE,.x (x E H),
R = I.. M+
1/1(1) dE,..
= (,8 - a:) IIE(Ii)xll 2
4
By Theorem 6. 7.S(ii), Hence
R(A - 10 I) = (A - lol)R = ., f.
M+<
1/l(A)(A - A0 )d,.. IIE(A)xll ::;: fJ; a IIR(..1.0 ;A)IIIIE(Ii)xll.
By the remark of (6.8.2), the last integral is equal to I. This proves the
If (,8 - a) < 211 R(A 0 ;A) 11- 1, then II E(A)xll = 0-that is, E~ =const. for
theorem.
a < . 1. <,B. This completes the proof.
The integral So far we have characterized the points of the spectrum as follows:
M+ dE;. ,1. E u(A) if and only if Ao E [m,M] and there is no open interval containing
f"' . 1. ~ Ao
0
A0 on which El is constant. We shall now characterize the eigenvalues of A.
for l 0 E [m,M] has not been defined, since 1/(A- A0 ) is not continuous. Theorem 6.8.3 A point ). 0 in [m,M] is an eigenwlue of A if and only
However, under the assumptions of Theorem 6.8.1 it can be defined as a ifE~ is discontinuous at/...= A0 -that is, EJ.o+O =F-E._,.
limit of sums corresponding to partitions II, III I-> 0, provided we agree to
take (p.k- A0 )- 1 E(Iik) = 0 whenever E(Ak) = 0 (even when Ilk= A. 0 ). With Proof. Suppose ). 0 is an eigenvalue and let Ax0 - ..1. 0 Xo = 0, Xo =F- 0.
this agreement one then easily verifies that Then ((A - } 0 !) 2 x 0 ,x 0 ) = 0. It follows that
.M+t dE;.
jm ,----;- JM + (A - }. 0 ) 2 d(E;. x0 ,x 0 ) 0.
R(A 0 ;A)=
A - Ao
(6.8.3)
.. =
236 HILBERT SPACES AND SPECTRAL THEORY EIGENVALUE PROBLEMS FOR DIFFERENTIAL EQUATIONS 237
Since the integrand is non ncgat ivc and (E). x 0 , x 0 ) is nondecrcasing, PROBLEMS
~.U+l:
6.8.1. Letf(l) be a continuous function of J., with values in a Banach
I (}.- ).
... lo +~
0)
2
d(E;. X 0 ,x 0 ) = 0 for any IJ > 0.
space X, where ). varies in an open set G of the complex plane. Let r be a
continuously differentiable curve in G. Show that the contour integral
Since ic - i. 0 ;;:,: IJ in the last integral, we get Jr f(l) d). can be defined in the same way as for complex-valued functions
...\f ...... i!: J
f(l). If r is oriented counterclockwise, then we write r f(l) d). also in the
I d(E;_ X 0 .XO) = 0, Ur
form r f().) d)., Prove: If A is a bounded operator in X, then A f(A) dl] =
)o~~
Jr [A/(..1.)] die.
-that is, (x 0 ,x 0 ) - (E;.o+~x 0 ,x 0 ) = 0. Thus 6.8.2. Let T be a bounded linear operator in a Banach space X and
let r be a continuously differentiable simple closed curve containing the
ii(J- E; 0 +~)xoll =((I- E; 0 +q)x 0 ,x 0 ) = 0.
2
spectrum t1(T) of Tin its interior. Let f(J.) be an entire complex analytic
It fol!ows that (I - E." +~)x 0 = 0-that is, EJo .._~ x 0 = x 0 Similarly we obtain function. Define
E;. 0 -~
x 0 = 0. Taking IJ-+ 0, we find that 1
f(T) = - . J, j(A)R(2;T) dA. (6.8.5)
(6.8.4) 2l!l j r
Hence, E~, 0 + 0 =!- E).o Use Cauchy's theorem to show that the definition is independent of r. Prove
Suppose, conversely, that E;_ 0 +o =!- E).o Since , 0 :::;; F).o+O there exists that iff(..l) = r, thcnf(T) = rn. [Hint: R(..l;T) =Iympm+l.]
a point x 0 =!- 0 that belongs to the subspace of the projection E,o+ 0 but 6.8.3. Let f,g be entire complex analytic functions and let a.,p be
that is orthogonal to the subspace of the projection E).o Thus sea Jars. Prove:
(i) (a.f + [Jg)(T) = a.j(T) + {Jg(T).
(ii) (fg)(T) = f(T)g( T). [Hint: Modify the contour r for g(T),
We then also have , x 0 = x 0 if J. > ). 0 . Hence replacing it by a contour r 2 that contains the contour r = r I for f( T), and
use (5.3.1 ). ]
E(!1)xo = -"o (iii) [f(T)]* = f(T*), where j(A) = j(.t).
where IJ > 0. We deduce that 6.8.4. If T is self-adjoint, then f(T) as defined in (6.8.5) coincides
;_o+lJ with.f(T) defined by J~+ f(A) dE,, where {E.,} is the spectral family ofT.
Ax 0 = AE(L1)x 0 = r
... ;.o
}. dE;,X 0 6.8.5. If A is self-adjoint with upper boun9 M, and if Re ic > M, then
Also.
..O.o+:tt
;,oXo = i,o E(Mxo = r
... ;_0
,:0 dE, Xo .
1 J. ~o
(/,- i. 0) dE;. x 0 I = J.. .<n+~
Au
2
( l - i- 0 ) d(E!. x 0 ,x 0 ) Let A be a compact self-adjoint operator in a separable Hilbert space
2 2
If. From Problems 6.7.4, 6.7.5 we know that there exists an orthonormal
:::::: '1 li(1'1)xll .
basis {xm}, where Xm are eigenvectors.
Since IJ is arbitrary, Ax 0 - / 0 x0 = 0. This completes the proof. We shall consider now the case where A is an integral operator in
L2(0:):
Corollary 6.8.4 The space of the eigemectors corresponding to ).0 (Af)(x) = j' K(x,y)f(y) dy. (6.9.1)
coincide.\ with the subspace of the projection operator El-o + 0 - E).o !l
238 HILBERT SPACES AND SPECTRAL THEORY
EIGENVALUE PROBLEMS FOR DIFFERENTIAL EQUATIONS 239
Here f.!: is a bounded open set in R~. The function K(x,y) is called a kernel.
If K(x,y) is measurable in f.!: x f.!: and if where g,. = (g,(jJ,.). By Bessel's inequality
II (! (!
2
IK(x,y)l dx dy < oo,
(6.9.5)
We shall assume from now on that K(x,y) is symmetric, that n is a Hence, by Bessel's inequality,
bounded open set, and that
sup r 2
IK(x,y)l dy < cc. (6.9.2)
,,t IA-..1 1(/J,.(YW I
2
:S jK(x,yW dx :S C < oo, (6.9.6)
n a
where Cis a constant. Using (6.9.5), (6.9.6) we get, for j > i,
The eigenvectors of A are functions in L2 (0). We therefore call them also
1 j
eigenfunctions. I Jcm(j),.(x)l = L l..1.,.g..,(/Jm(x)l
Since A is compact and self-adjoint, it has an orthonormal basis of m=.i
eigenfunctions {x,.} with eigenvalues {.u,.}. Some of the eigenvalues .u~ may j "'
be equal to zero. We shall denote by{(/!,.} the subsequence of {xm} consisting :S I
m.r:::i.
lgm1 2 L
m ... l
2
1..1... 1 1(/J,.(XW
of those eigenfunctions with non-zero eigenvalues. We arrange the (/J., in
such a way that the corresponding eigenvalues .A.,. satisfy: 1..1... I ~ 1..1., + 1 1.
The sequence {(!).,.} may be finite or infinite.
The following theorem is called the Hilbert-Schmidt theorem. if i .... co. This proves (6.9.4).
We shall give an important application in case n is a bounded open set in
Theorem 6.9.1 Let f E L 2 (f.!:) have the form f = Ag-thar is, R 1 . Consider the ordinary differential operator
= I IK(y,x)(/Jm(x)
g(y) dx dy where ..1. is a complex number. We call this problem an eigenvalue problem.
Lis called self-adjoint if
=A..,J =
g(y)(/J .. (y) dy A.mg,.,
(Lu,v) = (u,Lv) (6.9.9)
for all u,v in C(/(a,b). The pair (L, U) is called self-adjoint if (6.9.9) holds for
240 HILBERT SPACES AND SPECTRAL THEORY EIGENVALUE PROBLEMS FOR DIFFERENTIAL EQUATIONS 241
all u,v in C[a,b] for which Uu = 0, Uv = 0. We then say that (6.9.7), (6.9.8) Applying the Hilbert-Schmidt theorem we conclude:
form a self-adjoint eigenFalue problem. If the eigenvalue problem (6.9.7),
(6.9.8) has a solution x(t) 0, for some J., then we call). an eigenvalue and Theorem 6.9.2 Euery function u in C"[a,b] that satisfies Uu = 0
x(t) an eigenfunction. can be written in the form
(6.9.12)
PROBLEM
6.9.1. If (L, U) is self-adjoint, then all the eigenvalues are real numbers. where the series is con~ergent uniformly and absolutely.
Furthermore, eigenfunctions corresponding to distinct eigenvalues are
orthogonal in eca,b). Since the fum:tions in C"[a,b] that vanish in some neighborhoods of
t = a and 1 = b form a dense subset in L2 [a,b ], it follows that the closed
We shall assume from now on that (L,U) is self-adjoint and ). = 0 linear subspace spanned by {!J'm} coincides with L2 (a,b). Thus
is not an eigenvalue. It can then be shown that there exists a function G(t,T) Corollary 6.9.3 The sequence {q>m} of the orthonormal eigenfunctions
such that, for any continuous function f( I),
of the self-adjoin I system ( 6. 9. 7), ( 6. 9. 8) is complete in L 2 [a, b] .
x(t) =IG(t,r)f(r) dr
a
(6.9.10)
value.
Recall that it was assumed in this corollary that i. = 0 is not an eigen-
x(t) =).I a
b
G(l,r)x(r) dr. (6.9.11) x(O) = 0, x(l) = 0.
Verify that
It can be shown tli.it G(t,T) G(r,l). Hence we can apply the Hilbert-Schmidt
theorem.
=
G(
t;r
) = {I((1 I-- t)r,
T), ifO '5. t < r,
ifr < t :-::;I,
We arrange the sequence {J... } of all the eigenvalues of (L,U) such that
IJ.,l '5. IJ.,. + 1 1, and denote by {tp,.} an orthonormal sequence of eigenfunc- is Green's function. Prove also that for any continuous f then: is a unique
tions, with tp,. corresponding to J.... (An eigenvalue 1., 0 will occur k times in solution of Lu =fin (0,1), u(O) = u(l) = 0, and it is given by
the sequence {).,.} if there are k linearly independent solutions of (6. 9. 7), l
(6. 9. 8) for ). = 1..0 .) The notation here is slightly different from that used in u(t) = ( G(t,r)f(r) dr.
the Hilbert-Schmidt theorem, since the numbers J.,. that occur here are the o
reciprocals of the eigenvalues of the integral operator with kernel G(t,r). 6.9.3. In the eigenvalue problem of Problem 6.9.2, show that the
Jfu is in C"[a,b] and if Uu = 0, then setf= Lu. It follows that eigenvalues are given by ).,. = n 2m 2 (m = 1,2, ... ) and the orthonormalized
b eigenfunctions are given by rp ..(t) =,/2
sin nmt. Applying Corollary 6.9.3,
u= J G(t,r)f(r) dr.
a it fo!lows that {,/2 sin nmt} is an orthonormal basis in L2 (0,1).
EIGENVALUE PROBLEMS FOR DIFFERENTIAL EQUATIONS 243
242 HILBERT SPACES AND SPECTRAL THEORY
6.9.8. Let {rp.,}, {A,} be as in Problem 6.9.7. Then, for any fixed y,
6.9.4. Consider the eigenvalue problem
Lx =
dx
i - - 11.x = A.x (11. not an integer), tiK(x,y)- j~llj(/lj<x><PJY>r dx->o as m-> oo.
dt
x(O)- x(l) = 0. 6.9 .9. Prove Dini' s theorem: if a monotone-increasing sequence of
continuous functions on a compact metric space S converges to a continuous
This is a self-adjoint problem. Prove that the function function on S, then the convergence is uniform on S.
itt<-) 6.9.10. Assume that K(x,y) is continuous in (x,y) for X E n, J' E n...
ifO .$ t < T, X #- y. Prove that the qJ,(x) are continuous on 0 and that K 2 (x,y) is con-
l - ei'
tinuous in (x,y) for X En, yEn. By Problems 6.9.7, 6.9.8 it follows that
G(t;r) = iei( 1 + -tl k
[ if T < I .$ 1, L ..t,! I<P,(Y)I 2 -> Kiy,y)
t - e~ m=l
ask-> oo,
If p,q are real-valued functions and if r.x,{J,y,b are real numbers, then the system
is self-adjoint.
In the following problems, A and K are as in the Hilbert-Schmidt
theorem.
6.9.6. A kernel K(x,y) is called degenerate if there exists a finite
number of functions a;(x),b;(y) such that K(x,y) = I: a;(x)b;(y). Prove that
K(x,y) is a degenerate kernel if and only if A has a finite number of eigen-
values different from zero. [Hint: If A has a finite number of eigenvalues
At, .. . ,..t, different from zero, and if {~p 1 , ... ,~p.,} is an orthonormal set and
Ap; = A. 1 ~p 1 , then, by the Hilbert-Schmidt theorem,
J[K(x,y)- J . t1 ~p;(x)qJj(y)lh(x)
1
dx =0
for any function h in L2 (0). Hence
m
K(x,y) = L ..t 1 rp 1 (x)~p 1 (y).]
j=l
Bibliography [8]
Reading, Mass., Addison-Wesley Publishing Company, Inc., 1959.
Natanson, I. P., Theory of Functions of Real Variables. New York:
Fredrick Ungar Publishing Company, vol. I, 1955; vol. 2, 1961.
[9] Rudin, W., Principles of Mathematical Analysis, 2nd ed. New York:
McGraw-Hill, Inc., 1964.
[I 0] Saks, S., Theory 9f the Integral. New York: Hafner Publishing
Company, 1937.
[11] Titchmarsh, E. C., The Theory of Functions, 2nd ed. London: Oxford
University Press, 1950.
[12] Williamson, J. H., Lebesgue Integra/ion. New York: Holt, Rinehart
and Winston, Inc., 1962.
FUNCTIONAL ANALYSIS
[13] Banach, S., Theorie des operations /ineares, 2nd ed. New York: Chelsea
Publishing Company, 1963.
[14] Davis, M., First Course in Functional Analysis. New York: Gordon
and Breach, 1966.
We have made no attempt to provide a systematic and detailed biblio- [15] Day, M. M., Normed Linear Spaces.-Ergebnisse der Mathematik,
graphy. References [1]-[31] inc! ude most of the relatively recent books that vo I. 21. Berlin : Springer-Verlag, 19 58.
deal with the subject matter of the book in a substantial way. The reader will [16] Dunford, N., and Schwartz, J. T., Linear Operators. New York:
find a proof of Zorn's lemma in [33], [35]. More details on the application lnterscience Publishers, vol. I, 1958; vol. 2, 1963.
given in Section 5.4 can be found in [36]. A detailed theory on eigenvalue [17] Edwards, R. E., Functional Analysis, Theory and Applications. New
problems for ordinary differential equations is available in [32]; for further York: Holt, Rinehart and Winston, Inc., 1965.
applications to ordinary differential equations, see [34], [38]. The construc- [18] Goffman, C., and Pedrick, G., First Course in Functional Analysis.
tion of Green's function in Section 4.9 is taken from [37]. Englewood Cliffs, N.J.: Prentice-Hall, Tnc., 1965.
[19] Halmos, P. R., fntroduction to Hilbert Space. New York: Chelsea
MEASURE AND INTEGRATION Publishing Company, 1957.
[20] Kantrovich, L V., and Akilov, G. P., Fu,nctional Analysis in Normed
[1] Asplund, E., and Bungart, L., A First Course in Integration. New
Spaces. New York: Pergamon Press, Inc., 1964.
York: Holt, Rinehart and Winston, Inc., 1961.
[21] Kolmogorov, A. N., and Fomin, S. V., Elements of the Theory of
[2] Caratheodory, C.. Vorlesungen iiber Reele Funktionen, 2nd ed. New
Functions and Functional Analysis. Rochester, N.Y.: Graglock Press,
York: Chelsea Publishing Company, 1948.
vol. I, 1957.
[3] Halmos, P. R., Measure Theory. Princeton, N.J.: D. Van Nostrand
[22] Liusternik, L.A., and Sobolev, V. J., Elements of Functional Analysis.
Company, 1959.
New York: Fredrick Ungar Publishing Company, 1961.
[4] Hildebrant, T. H., Introduction to the Theory of Integration. New
[23] Sz.-Nagy, B., Spektraldarstel!ung Linearer Transformationen des
York: Academic Press, Inc., 1963.
Hilbertschen Raumes. Ergebnisse der Mathematik, vol. 39. Berlin:
[5] Graves, L. M., The Theory of Functions of Real Variables. New York:
Springer-Verlag, 1967.
McGraw-Hill, Inc., 1946.
[24] Naimark, M.A., Normed Rings. Gronigen, Netherlands: P. Noordhoff,
[6] McShane, E.J.; Tntegrarion. Princeton, N.J.: Princeton University
N.Y., 1959.
Press, 1944.
[25] Riesz, F., and Sz.-Nagy, B., Functional Analysis. New York: Fredrick
244 Ungar Publishing Company, 1955.
246 BIBLIOGRAPHY
[26] Royden, H. L., Real Analysis, 2nd ed. New York: The Macmillan
Company, 1968.
[27] Taylor, A. E., Introduction to Functional Analysis. New York: John INDEX
Wiley & Sons, Inc., 1958.
[28] Vulikh, B. Z., Introduction to Functional Analysis. New York: Perga-
mon Press, Inc., 1963.
[29] Wilansky, A., Functional Analysis. Waltham, Mass.: Blaisdell Pub-
lishing Company, 1964.
[30] Yosida, K., Functional Analysis. Berlin: Springer-Verlag, 1965.
[31] Zaanen, A. S., Linear Analysis. Amsterdam: North-Holland Publishing Absolute convergence, 126 Closed set, 17, 90
Company, 1953. Absolutely continuous, 52, 53, 67 Closure, 16, 91
Additive set function, 5 Compact liTiear operator, 186
MISCELLANEOUS Adjoint operator, 172, 209 Compact set, 91
Alaoglu's theorem, 169 Complete measure, 12
[32] Coddington, E. A. and Levinson, N., Theory of Ordinary Differential space, 29
Algebra, 2
Equations. New York: McGraw-Hill, Inc., 1955. of functions. I 16 Complete metric space, 17, 105
[33] Hausdorff, F., M engenlehre. New York: Dover Publications, Inc., 1944. J-, 2
Complete orthonormal system, 213
[34] Hellwig, G., Differential Operators of Mathematical Physics. Berlin: Almost everywhere, 33 Completely additive set function, 5
Springer-Verlag, 1964. (Addison-Wesley Publishing Company, Inc.) convergence, 33 Completely continuous operator, 186
[35] Kelley, J. L., General Topology. Princeton, N.J.: D. Van Nostrand Ascoli-A rzela's lemma, 112 Completion of measure, 12
Company, 1955. Completion of metric space, 99
Complex measure, 28
[36] Kellogg, 0. D., Foundation of Potential Theory. New York: Dover Haire class, 3 5
Banach algebra, 137 Conditio!lally compact space, 114
Publications, Inc., 1953.
Banach space, 126 Conjugate space, 150, 159
[37] Lax, P. D.," On the existence of Green's function," Proc. Amer. Math. Continuous function, 3 I, 11 0, 118
Soc., vol. 3 (I 952), pp. 526-531. BanachSteinhaus" theorem, 139
Bessel's inequality, 213 uniformly, 110, 118
(38] Naimark, M. A., Linear Differential Operators, part II. New York: Continuous linear operator, 135
Bilinear form, 221
Fredrick Ungar Publishing Company, 1968. Contractio!l, 119
Borel set, 17
Boundary operators, 239 Convergence, 16
Bounded linear operator, 135 absolute, 126
Bounded set, 16 almost uniformly, 35
Bounded variation, 54 in the mean, 49
in measure, 37
Cantor set, 24 weak, 161
Cartesian product, 78 Convex. set, 118
of measure spaces, 82, 87 Countahly additive set function, 8
of metric spaces, 94 Covering, <-, 109
Cauchy sequence, 17
iTI the mea!l. 41, 49 Darboux integral, 62, 66
in measure, 37 DarbouxStieltjes integral, 67
CeTit ralized class, 165 Darboux sums, 62
Characteristic functio!l, 31 Degenerate kernel, 242
Closed hall, 16 Dense, 92, I 07
Closed graph theorem, 144 Diameter of a set, 16
Closed linear operator, 143 Dime!lsion of a linear space, 124, 154
247
248 INDEX INDEX 249
Dini's theorem, 243 Graph of an operator, 143 Lebesgue's bounded convergence complete, 29
Direct product of normed linear spaces, Green's function, 15 7 theorem, 54 finite, 29
128 Lebesgue's decomposition, 72, 73 Lebesgue, 29, 30
Hahn-Banach's lemma, !50
Direct sum of normed linear spaces, I 28 Lebesgue's monotone convergence o--linite, 29
Hahn-Banach's theorem, 151
Direct sum of subspaces, 208 theorem, 58 totally finite, 29
Hahn decomposition, 26
Dirichlet problem, 156 Lebesgue's-Stieltjes, integral, 46 ~Iesh, 62
Hamel's basis, 13I
Distance function, 15 measure, 14 Metric function, 15
Harmonic function, 155
Domain of an operator, 13 5 outer measure, 14 Metric linear space, 125
Hclly's principle, 185
Double integral, 85 Limit, 16 Metric outer measure, I 7
Hermitian form, 221
Double layer, 198 Linear isomorphism, 159 construction of, 20
Hilbert space, 201
Linear operator, 135 Metric space, I 5
real, 201
Egorolf's theorem, 35, 36 bounded, 135 compact, 107
Hilbert-Schmidt theorem, 238
Eigenclemcnt, 19 5 closed, 143 complete, I 05
Holder's inequality, 96
Eigenfunction, 242 closure of, 145 J\fetric subspace, 92
Homeomorphism, 167
Eigenvalue, 195, 240 compact, 186 Minkowski's inequality, 96
Hyperplane, 153
Eigenvalue problem, 239 completely continuous, 1llf:i Mollifier, 103
supporting, !54
self-adjoint, 240 continuous, 13 5 Monotone class, 80
tangent, 154
Eigenvector, 195 continuous extension of, 138 \1onotone set function, 8
Hypoelliptic operator, 147
Equicontinuous, 112 extension of, 145
Equivalent, functions, 97 inverse of, 143 N cgative part of, a function, 31
Indefinite integral, 53, 74
mctrics, 94 norm oL 135 an operator, 225
Inferior limit, 2
norms, 127 Linear vector space, 123 Negative set, 25
Inner measure, 25
Essential supremum, 57 dimension of, 124, 154 1\cighborhood, 90
Inner product, 201
Essentially bounded, 57 1\eighhorhood basis, 164
space, 201 subspace of. I 29
Euclidean metric, 16 Neumann's problem, 199
Integrable, 40, 42, 46, 48, 71 Lower-semicontinuous, 31
Extended real number, 4 Lower variation of a measure, 28 Norm, 125
kernel, 197
Lusin's theorem, 116 Normed linear space, 125
IntegraL 40, 42, 43, 46
Factor space, 129 Nowhere dense, 92
with respect to a spectral family, 226
Fatou's lemma, 58 1\ull function, 54
Integral equation, 121 Mapping, 135
Finitely additive set function, 5 '\lull space of a functional, I 53
of Fredholm type, 193 Maximal element, 130
First category, I 06, 107 Integral operator, 237 Maximum metric, 93 Open ball, 16
First countability axiom, 165 Interior, 17, 91 Maximum principle, !56 Open covering, 91
Fixed point theorc m, 118, 119 Invariant subspace, 221 1\teasurable function, 29, 32 Open interval, 13
Fourier coefficients, 21 3 Isometric imbedding, 99 Borel, 29 Open mapping theorem, 141
Frechet space, 125 Tsomet ric isomorph ism, 159 Lebesgue, 29 Open set, 16. 90
Fredholm alternative, 143 Iterated integral, 8 5 Measurable, set, 8 Operator, 135
Fredholm-Riesz-Schaudcr's theory, rect<tn gle, 8 3 Orthogonal, 204
189, 193 Jordan decomposition, 27 subspace, 30 Orthogonal complement, J 73, 205, 208
Fubini's theorem, 84, 86 1\.leJsure, 4 Orthonormal, 21 3
Kernel, degenerate, 242
Functional, 150 complete, 12 basis, f-15
symmetric, 238
cominuous linear, 150 complex, 28 Outer measure, 7, 8
Fundamental solution, 156 Laplace'> equation, 15 5 tinite, 5 construction of, 11
Lax-Milgram's lemma, 207 o--finite, 5 metric, 17
Generated by, ring, 3 Lebesgue, integral, 46
, __
signed, 25
o--algcbra, 4 measure, 14 totally finite, 5 Parallelogram law. 203
o--ring, 4 outer measure, 14 Measure space, 29 Parseva1's formula, 215
Gram-Schmidt process, 217 sets, 14 Borel, 29 Partially ordered set, 130
250 INDEX