ISBN: 0-8176-3884-9

Text
                    Malcolm Adams
Victor Guillemin
Measure Theory
and Probability
BlRKHAUSER
Boston • Basel • Berlin


Malcolm Adams Department of Mathematics University of Georgia Athens, Georgia 30602 Victor Guillemin Department of Mathematics MIT Cambridge, MA 02139 To Jon Bucsela in memoriam Library of Congress Cataloging-in-Publication Data Adams, Malcolm Ritchie. Measure theory and probability / Malcom Adams, Victor Guillemin. p. cm. Originally published: Monterey Calif. : Wadsworth & Brooks/Cole Advanced Books and Software, c 1986. Includes bibliographical references and index. ISBN 0-8176-3884-9 (he : alk. paper). - ISBN 3-7643-3884-9 (he : alk. paper) 1. Measure theory. 2. Probabilities. I. Guillemin, V., 1937- II. Title. [QA273.A414 1995] 515\42~dc20 95-46511 CIP Printed on acid-free paper © 1996 Birkhauser Boston Reprinted with corrections from the 1986 Wadsworth edition. Birkhauser )® Copyright is not claimed for works of U.S. Government employees. All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, without prior permission of the copyright owner. Permission to photocopy for internal or personal use of specific clients is granted by Birkhauser Boston for libraries and other users registered with the Copyright Clearance Center (CCC), provided that the base fee of $6.00 per copy, plus $0.20 per page is paid directly to CCC, 222 Rosewood Drive, Danvers, MA 01923, U.S.A. Special requests should be addressed directly to Birkhauser Boston, 675 Massachusetts Avenue, Cambridge, MA 02139, U.S.A. ISBN 0-8176-3884-9 ISBN 3-7643-3884-9 Typesetting by ASCO Trade Typesetting, Hong Kong Printed and bound by Maple-Vail, York, PA Printed in the U.S.A. 987654321
Preface to the 1996 Edition We have used Measure Theory and Probability as our standard text in the basic measure theory courses at M.I.T. and the University of Georgia for over ten years and have been agreeably surprised at the enthusiasm with which students have reacted to the 5-3 mix of measure theory and probability. It is not clear that we've converted lots of aspiring mathematicians into probabilists, but we do seem to have left the non-mathematicians, our students from electrical engineering and computer science, feeling upbeat about the Lebesgue theory and its practical uses. On die down side, our students have been annoyed at the plethora of typos and silly mistakes in the first edition. (For instance the absence of a superscript bar on the right hand side of identity (v, w) = (w, v) made a travesty of our definition of Hilbert space!) This edition has been ridded of these errors thanks to the efforts of the editorial staff at Birkhauser, our former students, Leonard Shulman, Roy Yates, Gregg Wornell, and Mastafa Terab, and, in particular, thanks to Professor Bo Green and a diligent group of his students at Abilene Christian University (who went through the book with a fine tooth comb and assembled a pretty definitive list of errata). To them our warmest thanks. vii
Preface to the First Edition Probability theory became a respectable mathematical discipline only in the early 1930s. Prior to that time it was viewed with scepticism by some mathematicians because it dealt with concepts such as random variables and independence, which were not precisely and rigorously defined. This situation was remedied in the early 1930s largely thanks to the efforts of Andrei Kolmo- gorov and Norbert Wiener, who introduced into probability theory large infusions of measure theory. In retrospect, it was fortunate that the kind of measure theory they needed was already available; it had, in fact, been created some thirty years earlier by Henri Lebesgue, who had not been led to the invention of Lebesgue measure by problems in probability but by problems in harmonic analysis. It seems strange that it took more than 30 years for this fusion of probability and measure theory to occur. In fact, since that time, probability theory and measure theory have become so intertwined that they seem to many mathematicians of our generation to be two aspects of the same subject. It also seems strange that the basic concepts of the Lebesgue theory, to which one is naturally led by practical questions in probability, could have been arrived at without probability theory as their main source of inspiration. Saddled as we are with the fact that the theory of measure didn't develop along these lines, this doesn't mean we cannot teach the subject as if it had developed this way. Indeed, we believe (and this is the reason we wrote this book) that the only way to teach measure theory to undergraduates is from the perspective of probability theory. To teach measure theory and integration theory without at the same time dwelling on its applications is indefensible. It is unfair to ask undergraduates to learn a fairly technical subject for the sake of payoffs they may see in the distant future. On the other hand, the applications of measure theory to areas other than probability (e.g., harmonic analysis and dynamical systems) are fairly esoteric and not within the scope of undergraduate courses. Of course probability theory, taught in tandem with measure theory, is also not thought of as being within the scope of an undergraduate course, but we feel this ix
X Preface is a mistake. Discrete probability theory is taught at many institutions as a freshman course (and at some high schools as a senior elective). The kinds of problems we will be interested in here, i.e., the amorphous set of problems that go under the rubric of the law of large numbers, are lurking in the background in these discrete probability courses, and are often so bothersome to bright students that they arrive, unaided, at quite original ideas about them. By formulating these problems in measure theoretic language, one is often doing little more than vindicating for undergraduates their own intuitive ideas and, at the same time of course, convincing them that the measure theoretic methods are worth learning. By now we have probably given you the impression that this book is basically about probability. On the contrary, it is basically about measure theory. Sections 1.1 and 1.2 nominally discuss probability, but primarily discuss why measure theory is needed for the formulation of problems in probability. (What we hope to convey here is that had the Lebesgue theory of measure not existed, one would be forced to invent it to contend with the paradoxes of large numbers.) Section 1.3 deals with the construction of Lebesgue measure on Rn (following the metric space approach popularized by Rudin [see References, p. 199]. In §1.4 we briefly revert to probability theory to draw some inferences from the Borel- Cantelli lemmas, but §§2.1-2.5 are straight measure theory: the basic facts about the Lebesgue integral. Only to illustrate these facts do we return to probability at the end of the chapter and discuss expectation values, the law of large numbers, and potential theory. Sections 3.1-3.5 are also consecrated entirely to measure theory and integration: £l9$£2, abstract Fourier analysis, Fourier series, and the Fourier integral. Fortunately, the last two items have some beautiful probabilistic applications: Polya's theorem on random walks, Kac's proof of the classical Szego theorem, and the central limit theorem. With these we end the book. All told, taking into account the fact that we have packed quite a few applications to probability into the exercises, the ratio of measure theory to probability in the book is about 5 to 3. The notes on which this book is based have served for several years as material for a course on the Lebesgue integral at M.I.T. and for a similar course at Berkeley. They have been the basis for a leisurely semester course and an intensive quarter course and have proved satisfactory in both (though in using these notes in a quarter course, we had to delete most of the material in §§2.6-2.8 and §§3.6-3.8). We divided the book into the three chapters not just for aesthetic reasons, but because we found that in teaching from these notes we were devoting approximately the same amount of time to each of these three chapters, i.e., four weeks in a typical twelve-week semester course. Incidentally, we found it very effective, for motivational purposes, to devote the first three class periods of the course to the material in §§1.1 and 1.2, even though in principle this material could be covered in a much more cursory fashion. We discovered that with these ideas in mind, the students were much better able to endure the long arid trek through the basics of measure theory in §1.3.
Preface xi We would like to thank Marge Zabierek for typing the notes on which this book is based, and we would like to thank our students at Berkeley and M.I.T., in particular, Tomasz Mrowka, Mike Dawson, Harold Naparst, Mike Conwill, Christopher Silva, and Ken Ballou for suggestions about how to improve these notes and for weeding out what seemed to have been an almost endless number of errors from the problem sets. We have dedicated this book to Jon Bucsela, to whom we owe an exhaustive revision of the manuscript before we had the final version typed. His untimely death in the spring of 1984 was a source of acute grief to all who knew him. Malcolm Adams Victor Guillemin
Suggestions for Collateral Reading For background in probability theory, we recommend Feller, An Introduction to Probability Theory and Its Applications,* We feel that at the undergraduate level, this is the best book ever written on probability theory. Its charm resides in the fact that there are literally hundreds of illustrative examples. This makes it hard to read through from cover to cover, but it is a gold mine of ideas. Another beautiful book, though more advanced than Feller, is Kac's 96-page monograph in the Carus series, Statistical Independence in Probabilityy Analysis and Number Theory, Our treatment of Bernoulli sequences and the law of large numbers in §1.1 was largely borrowed from this book, and one can go there to find further ramifications of these topics. There are several treatments of measure theory in conjunction with probability written for graduate students. In our opinion, the best of these is Billingsley's book Probability and Measurey which a bright undergraduate will, with a little effort, find accessible if he or she ignores the more technical sections toward the end. Finally, for material on metric spaces and compactness, we have attempted to remedy the fact that we presuppose a nodding acquaintance with these topics by summarizing the main facts in the appendix. To learn this material, however, we recommend either Hoffman (Analysis in Euclidian Space) or Rudin (Principles of Mathematical Analysis), ♦For complete bibliographic information for the titles listed here, see the Reference section on page 202. xii
Contents Chapter 1 Measure Theory 1 §1.1 Introduction 1 §1.2 Randomness 14 §1.3 Measure Theory 24 §1.4 Measure Theoretic Modeling 42 Chapter 2 Integration 53 §2.1 Measurable Functions 53 §2.2 The Lebesgue Integral 60 §2.3 Further Properties of the Integral; Convergence Theorems 72 §2.4 Lebesgue Integration versus Riemann Integration 82 §2.5 Fubini Theorem 89 §2.6 Random Variables, Expectation Values, and Independence 102 §2.7 The Law of Large Numbers 110 §2.8 The Discrete Dirichlet Problem 115 Chapter 3 Fourier Analysis 118 §3.1 ^-Theory 118 §3.2 ^-Theory 124 §3.3 The Geometry of Hilbert Space 130 §3.4 Fourier Series 137 §3.5 The Fourier Integral 145 §3.6 Some Applications of Fourier Series to Probability Theory 156 §3.7 An Application of Probability Theory to Fourier Series 164 §3.8 The Central Limit Theorem 170 xiii
xiv Contents Appendix A Metric Spaces 178 Appendix B On $£p Matters 183 Appendix C A Non-Measurable Subset of the Interval (0,1] 199 References 202 Index 203
Measure Theory and Probability
Chapter 1 Measure Theory §1.1 Introduction In this section we will talk about some of the mathematical machinery that comes into play when one attempts to formulate precisely what probabilists call the law of large numbers. Consider a sequence of coin tosses. To represent such a sequence, let H symbolize the occurrence of a head and T the occurrence of a tail. Then a coin-tossing sequence is represented by a string of H's and T's, such as HHTHTTTHHT... Now, let Sff be the number of heads seen in the first N tosses. The law of large numbers asserts that for a "typical" sequence we should see, in the long run, about as many heads as tails. That is, we would like to say that (1) lim£-i for the "typical" sequence of coin tosses. We do not expect this assertion to be true for all sequences, because it is possible, for instance, for our sequence of coin tosses to be all heads. Experience tells us, however, that such a sequence is not typical. In what follows, we describe a mathematical model of coin tossing in which we precisely define what is meant by a "typical" sequence of coin tosses. With this model, the law of large numbers can be rigorously demonstrated. Because James Bernoulli first stated the law of large numbers, in the seventeenth century, we will call a sequence of coin tosses a Bernoulli sequence. 1
2 Chapter 1 Measure Theory Let 0$ represent the collection of all possible Bernoulli sequences. Notice that <% is an uncountable set. (See exercise 1.) This fact is also clear from the following proposition. Proposition 1. If we delete a countable subset from ^, we can index what is left by points on the real interval / = (0,1]. Proof. We construct a map / -> 0$ that is one to one and fails to be onto by a countable set. The map is constructed as follows. Every coel can be written in the form (2) *>=ZS «i = <U Because the a^s determine co, we introduce the notation co = . ax a2 #3... which is called the binary expansion of co. From this representation we produce a Bernoulli sequence by putting an H in the nth term of the sequence if an = 1 or a T if an = 0. Unfortunately, this does not give a well-defined map l-*@ because co does not necessarily have a unique binary expansion. For example, co = \ has the two binary expansions .1000... and .0111... To avoid this problem we prescribe that, if co has a terminating and a nonterminating expansion, we give it the nonterminating one. This convention gives a one-to-one map / -* & that is not onto because it misses out on those Bernoulli sequences that end in all tails. Let ^dcg denote the collection of Bernoulli sequences that, after a certain point, degenerate to all tails. We claim that ^dcg is countable. Proof. Let ^Jeg be the Bernoulli sequences that have only tails after the fcth toss. Then ^dcg *s finite and ^deg = Q ^Seg fc=l is a countable union of finite sets. Thus ^deg is countable. □ Because ^deg is a countable subset of the uncountable set ^, we consider it to be negligible in our consideration of "typical" elements of ^. Thus, for all intents and purposes, we can consider & to be identified with /. In order to describe other features of our model of ^, we need some familiarity with the idea of Lebesgue measure. We will not yet attempt to define Lebesgue measure precisely, but we will describe some of the properties it
§1.1 Introduction 3 should have. We ask the reader to believe that it exists until we examine it more rigorously. A measure \i on a space AT is a nonnegative function defined on a prescribed collection of subsets of X, the measurable sets. If A is a measurable set, the nonnegative number /i(A) is called the measure of A. Of course we will require that jx have certain properties so that it behaves as our intuition tells us a measure should behave. For example, we will require additivity: If A and B are measurable and disjoint, then A u B is measurable and fi(A u5) = fi(A) + /x(B). This and other properties of measures will be discussed in §1.3. The particular measure in which we are interested here is called Lebesgue measure (denoted fxL) and is defined on certain subsets of the real line R. For the intervals (a,b), (a,*], [a,6], [«,&) the Lebesgue measure is just the length, b — a. More generally, by the property of additivity, if i=l is a finite disjoint union of finite intervals A,-, then A is Lebesgue measurable and i=l Using the concept of Lebesgue measure, we can now formulate what we will call the Borel principle. Borel principle. Suppose £ is a probabilistic event occurring in certain Bernoulli sequences. Let 0&E denote the subset of & for which the event occurs. Let BE be the corresponding subset of /. Then the probability that E occurs, Prob(E), is equal to fJ.L{BE). Let us show that this principle works for some simple probabilistic events. 1. E is the event that H appears on the first toss. BE = {fl)6/;fl) = .l...}=(i,l] so fiL(BE) = \. 2. E is the event that the first N tosses are a prescribed sequence. BE = {co e /; co = ,a\ a2a3.. .an .. } where au...,aN are prescribed and everything else is arbitrary. Let s = . a1 a2... aN 0 0..., then BE = (s, s 4- (1/2N)] so that fiL(BE) = 1/2* as expected.
4 Chapter 1 Measure Theory 3. E is the event that H occurs in the Nth place. BE = {coel; co = .ax a2...aN^ 1 aN+1...} Fix a particular s = .ax ...a^^ 1000.... Then B£ contains the interval (s, s + (1/2N)]. We can choose the al9..., aN_t in 2N~l different ways, and each of these intervals is disjoint from the others; so MBE) = 2-(i,)4 13 15 3 4 8 2 8 4 The shaded region is BE for the event that H occurs on the third toss. 4. E is the event that, in the first N tosses, exactly k heads are seen. Be = {co € /; co = .a\di... aN ..., where exactly £ of the first N at9s are 1} Fix .£*!,...,%, k of which are 1. Let s = .<z1a2...0NOOO..., so that ££ contains (s,s + (1/2N)]. There are (£) such intervals, all mutually disjoint, so Pl(Bb) -m 5. Start with X dollars and bet on a sequence of coin tosses. At each toss you win $1.00 if a head shows up and you lose $1.00 if a tail shows up. What is the probability that you lose all your original stake? To discuss this event we introduce some notation. Rademacher Functions For coel we define the kth Rademacher function Rk by (3) Rk(co) = 2ak-l where co = . ax a2... is the binary expansion of co. Note that „ , , f+1 ifak= 1 (4) -„-, , , ,. Q so Rk(co) represents the amount won or lost at the kth toss. To familiarize ourselves with the Rademacher functions, we graph the first three, Riico), R2(co), and R3(co).
§1.1 Introduction 5 *» + 1- -1-*- R2 +1- i _/ l ^ [to) v- 1 1 I 4 ■ ■ — ~w 1 1 I 2 LJ-™ -™" I 1 3 4 — ™ "• 1 1 1 **(») + 1- H h -!-*■ Using the Rk\ we can describe BE for event 5. First consider the event Ek, representing loss of the original stake at the fcth toss. Let (5) Sk(co) = £ *,(«) Sk(co) gives the total amount won or lost at the fcth stage of the game. Then BEk ={coe /; Si(co) > -X for / < k, and Sk(co) = -X} and *E = U BEk
6 Chapter I Measure Theory We will postpone the computation of fxL(BE) to §1.4 because BE is not a finite union of intervals. Now we return to the law of large numbers. Our assertion is that "roughly as many heads as tails turn up for a typical Bernoulli sequence." We formulate this statement mathematically as follows. For (Del, with co = .a1...aN..., let sN((o) = at + a2 + •• • + aN. This sum gives the number of heads in the first N terms of the Bernoulli sequence corresponding to o. Now fix e > 0 and consider f T\sN(co) ll 1 MH^~ 2\>e\ i 2 This set represents the event that, after the first N trials, there are not "roughly as many heads as tails." We can restate our assertion as follows. Theorem 2. (Weak law of large numbers) ViX^n) -*" 0 as N -* °° Proof. We first describe BN using Rademacher functions. Recall that Rk(co) = 2ak-l where co = . ax a2.. • ak Thus N SN(co) = X Rk(o>) = 2(flx + a2 + ••■ + flj - N = 2sN(o)) - N Now N 2 > eo\2sN((o) - N\> 2eN which is equivalent to |5N(a))| > 2sN. So, by altering e slightly, we restate the theorem as follows. Let AN = {weI; \SN(w)\ > Ns} Then jUL(/4N) -> 0 as AT -> oo To prove this form of the theorem, we will need the following special case of Chebyshev's inequality. Lemma 3. Let / be a nonnegative, piecewise constant function on (0,1]. Let a > 0 be given. Then
§1.1 Introduction 7 fiL({(oeI;f((o)> *})<\[ fdx aJo (Here the integral Jo fdx is the usual Riemann integral.) Notice that we know how to compute fiL({co e I;f{co) > a}) because {co e I; f(co) > a} is a finite union of intervals. Proof of lemma. When / is piecewise constant, there exist xl9...,xk with 0 = xx < • • • < xk = 1 and / = ct on (xh xi+1) i = 1,..., k — 1. (This is what we mean by piecewise constant.) Then I i k-i fdx = X Ci(xi+1 - Xi) > X'cr(xl+1 - xt) 0 i=l where Z' means sum over the i's such that c{ > a. Therefore, Z,ci(*m - xt) > a£'(xm - xt) = ocfiL({coeI;f(co) > a}) so 1 f1 - fdx > fiL({(oeI;f(co) > a}) aJo Now we continue with the proof of our theorem. Notice that AN = {coeI;\SN(co)\>Ns} = {coeI; SN(co)2 > N2€2} An application of the preceding lemma gives hMh) = M{">e/; |S*(<d)|2 > N2e2}) < ^ P S^dx To exploit this inequality we need to compute jo S%dx- However f1S^x= [\YJRk)2dx= t ['rUx + Z ^^Rjdx JO JO *=lJo i+JJO Because Rk = 1, each of the first N terms is equal to one. What about I i RtRjdx i ^/? Suppose i <j. Let J be an interval of the form (//2\(J + l)/2'], 0 < / < 2'. Then Rt is constant on J and Rj oscillates 2(j — i) times so that
8 Chapter 1 Measure Theory i Rjdx = 0 Thus which proves that Thus I 1 RiRjdx = 0 1 i S*dx = N o ^(^)<(^iJN = ^t->0 as N^oo □ The astute reader has probably noticed that we haven't proved exactly what we said we intended to prove at the beginning of this section. Namely, we wanted to prove that, for a "typical" Bernoulli sequence, (6) l_£»£Uoas^oo By "typical" we should mean equation 6 fails on a set of zero probability. By the Borel principle, an event E has probability zero if the corresponding set BEcz I has Lebesgue measure zero. The only sets we know thus far with zero Lebesgue measure are finite collections of points. When we extend Lebqsgue measure to a collection of sets much bigger than the collection of intervals, we will find many more sets of measure zero. In fact we can describe these sets now without developing the general theory of Lebesgue measure. Given a subset A czR and a countable collection of sets {At}jiu we will say the >4£'s are a countable covering of A if A c [j^ A{. Definition 4. A set A c R has Lebesgue measure zero if, for every e > 0, there exists a countable covering {At} of A by intervals such that (7) f>t04,)<e Remarks. 1. In this definition we can allow the i4f's to be finite unions of intervals. 2. If A has Lebesgue measure zero and B c A, then B has Lebesgue measure zero.
§1.1 Introduction 9 3. A single point has Lebesgue measure zero. 4. If A1, A2,... is a countable collection of sets, each having Lebesgue measure zero, then \J flx At has Lebesgue measure zero. In particular, countable sets have Lebesgue measure zero. Proof. (Remarks 1, 2, and 3 are clear.) To prove remark 4, choose e > 0. Because A{ has Lebesgue measure zero, there exists a countable collection of intervals Aitl,Ait29... covering A{ such that j-i * The collection {AtJ} is countable, it covers (J?^ Ah and i,j i=l j=l i=l ^ Now let N = {co e I; (sn(co)/n) -► 1/2 as n -► cxd}. N is called the set of normal numbers. Let Nc denote the complement of N. Theorem 5. (Strong law of large numbers) Nc has Lebesgue measure zero. Remark. Nc is uncountable; in fact, Nc contains a "Cantor set." Consider the map a:I -► J defined by o-(co) = . ax 11 a2 11 a3 1 1... for co = ,a1a2a3 This map is one to one, so its image is uncountable. Notice also that the image is contained in Nc. In fact, if co' = a(co)9 then s3n(co') > 2n\ so 3n "3 Now we will prove theorem 5. Let An = {coel; \Sn(co)\ > en} = {coel; S*(©) >e4n4} Then, by Chebyshev's inequality, MlW.) < i I sndx where | #ix = | ( £ Kfc J dx e n Jo Jo Jo \*=i / Multiplying out the integrand, we obtain five kinds of terms: 1. Rt a=l,...,n 2. RlR] oc*P
IU Chapter 1 Measure Theory 3. R*RfiR7 OL^P^y 4. R^Rp ol^P 5. RaRpRyRd a^P^y^d Because R£ = 1 and R^Rj = 1, J£fl£<frc = JoRlRjdx = 1. We claim that the other terms all integrate to zero. In fact, and R%RpRydx= RpRydx = 0 Jo Jo RlRpdx = RaRpdx = 0 Jo Jo What about RaRpRyRdl Assume a < p <y <8 and consider an interval of the form (//2y, (/ + 1)/2V]. Ry is constant on J and, because a < /? < y, RaRfiRy is constant on J as well. Finally, Rd oscillates 2(8 — y) times on J, so 1 RaRfiRyRddx = 0 and RaRpRyRddx = 0 Because there are n terms of the form R* and 3n(n — 1) terms of the form ^a^fl> J Sw4dx = 3n2-2n<3n2 and 3 *&> Hl(A„)?Z \-r-7 3«2^ Lemma 6. Given 5 > 0, there exists a sequence £,, e2,... such that s„ -> 0 and (8) 2i<5 n=l " «>w Proof. Choose, for instance, e„ such that e4 = cn~1/2 for some constant c.
§1.1 Introduction 11 Then oo 3 3 co j If c is chosen large enough, this quantity is less than 5. V Finally, for each n, set Bn = {co; \Sn(co)\ > snn} Hd^n) < 3/e£n2, hence Yj?=i A*l(Ai) < <5. Notice that the B„'s are finite unions of intervals since Sn is piecewise constant. Thus, if we can show that Nc a (j£=i An ^e theorem will be proved. Now Nc <= (Jw°°=1 Bn if N s fln^i5«- Bu*> if coef|„% ££, then, for each n, \Sn(co)\ < e„n; that is, \Sn(co)/n\ < en. Because en-^0, we conclude that | Sn(co)/n| -► 0; that is, co e N. □ Remarks. 1. We have just proven theorem 5 by showing that (9) fiL(N<) = 0 Notice that we needed a relatively sophisticated definition of "measure zero" to make sense of this statement, because Nc is such a bad set. In particular Nc is uncountable. (The only intervals of length zero are points, and Nc is not even a countable union of such sets.) Later, when we discuss the connection between measure and integration, we will see that this example provides a good illustration of why Riemann integration is inadequate for probability theory. 2. Notice that the strong law of large numbers (theorem 5) does not indicate at what point we can expect about as many heads as tails. In §3.8 we will discuss the central limit theorem, which has some bearing on this question. Exercises for §1.1 1. Prove that the set $ of Bernoulli sequences is uncountable by the Cantor diagonal argument. 2. a. Let coel = (0,1]. Show that co can be written in the form Yj%i aJlK a{ = 0,1. Show that this expansion is unique when we restrict to nonterminating series. b. Show that, for any integer k,coeI can be written in the form J^ ajk\ where a{ = 0,1,..., k — 1. Show that the expansion is unique when we restrict to nonterminating series. 3. A gambler has an initial stake of one dollar. Calculate the probability of
Chapter 1 Measure Theory ruin at times 1, 3, and 5. Show that the chance of eventual ruin is at least 70%. Show that I Ry>Rv2''RyJx = 0 or 1 for any sequence yx <y2<"' <yn- When is the answer one? Define the Rademacher functions on the whole real line by requiring them to be periodic of period one—that is, by setting jRfc(x + 1) = Rk(x). With this definition, show that Rk+l(x) = Rk{2x) and, by induction, that Rk(x) = R^l^x). Show that Rn(x)= -sgn[sin(27c2n~1x)] except at a finite number of points. (Notation: For any number a, sgna is one if a is positive and minus one if a is zero or negative.) We will see later in the text that some interesting analogies exist between the Rademacher functions and the functions sin(27c2n~1x). Prove that 2t - 1 = £ Rk(t)2~k Every number ewe(0,1] has a ternary expansion °> = X) ai3~' with a{ = 0, 1, or 2 (see exercise 2). We can make this expansion unique by selecting, whenever ambiguity exists, the nonterminating expansion— that is, the expansion in which not all af's from a certain point on are equal to 0. With this convention, define Tk{a>) = ak - 1 Draw the graph of Tk for k = 1,2, 3. Can you discern a general pattern?
§1.1 Introduction 13 9. Obtain a recursion formula for the 7^'s similar to the recursion formula for the Rk's in exercise 5. 10. Let C be the set of all numbers on the unit interval [0,1], which can be written in the form 00 <o = Z a*3~* k=l with ak = 0 or 2. Show that C is uncountable. (C is called the Cantor set.) (Hint: Use the Cantor diagonal process.) 11. Prove that the Cantor set (see exercise 10) can be constructed by the following procedure: From [0,1] remove the middle third, (i,f); from the remainder—that is, the intervals [0,y] and [f, 1]—remove the middle thirds, and so on, ad infinitum. The remainder is the Cantor set. I 2 3 3 12 1 2 2 8 9 9 3 3 9 9 12 1 2 2 8 9 9 3 3 9 9 12. Show that the Cantor set is of measure zero. 13. Describe geometrically the set a(I) discussed on page 9. 14. Show that the nonnormal numbers are dense in the unit interval. 15. a. Show that a positive number c3 exists such that, for all JV, / [SN(x)]6dx<c3N3 Jo b. Let An be the set {co e /; |S„(co)| > en}. Show that the Lebesgue measure of An is less than c3e"6n"3. 16. More generally, show that a positive number cK exists such that, for all N, L [SN(x)]2Kdx < cKNK. 0 17. Prove a refinement of the strong law of large numbers, which says that -£-►0 as N-+oo N*
i«t Chapter 1 Measure Theory for any 3 > j. (Hint: Use exercise 16. We will see later that, for <5 = j9 the situation is much more interesting.) 18. Prove that (Hint: By induction. Write etSn(x)dx= etS»-iix)etR»ix) dx Jo Jo Break up the unit interval into 2""1 equal subintervals on each of which #„__! is constant. Show that f etS»™dx = (^yA f ets^x)dx if J is one of these intervals.) 19. From exercise 18, derive the formula i:—&rm 1=0 20. Let / be a nonnegative monotone function defined on the unit interval. Prove Chebyshev's inequality ■» «;/.'• fiL({(oeI;f(co)>a})<-\ fdx aJo with the integral on the right being the Riemann integral. 21. We have already defined Lebesgue measure for two kinds of sets: finite unions of intervals and sets of Lebesgue measure zero. Show that these two definitions are not contradictory; that is, show that the interval [a, fc], a < b, is not a set of measure zero. (Hint: Use the Heine-Borel property of compact sets.) §1.2 Randomness In §1.1 we saw how to identify the set 0& of Bernoulli sequences with the set of points on the unit interval J. In terms of this identification, a probabilistic event E, associated with Bernoulli sequences, gets identified with a subset BE of /. We saw that, at least for simple events, the Borel principle
§1.2 Randomness 15 applies; that is, (1) Prob(E) = fiL(BE) We will attempt in this section to describe some slightly more complicated probabilistic events in measure theoretic terms. Example 1. Gambler's Ruin A gambler has X dollars and bets at even odds on a coin flip. What is the probability of his ruin? We discussed this event already in §1.1. We showed that BE=\jBEk where BEk = {coeI; Si(co) > -X for / < k, and Sk(co) = -X}. After developing some measure theoretical tools, we will see that (2) fiL(BE) = £ KBBk) = 1 fc = l In other words, with probability one, if a gambler bets long enough, he will eventually lose all his money no matter how big his initial stake. Example 2. Random Patterns Pick a finite pattern of coin tosses, for example, T H H T. Let E be the event that T H H T occurs in a given Bernoulli sequence. Then BE = {cogI; there exists n0 with R„o((o) = -1, Rno+1((o) = 1, Rno+2((o) = 1, and R„0+3((o) = -1} We will prove in §1.4 that this set is of measure one. In fact we will prove that, if one fixes any finite pattern, this pattern appears infinitely often in a Bernoulli sequence with probability one. This result can be interpreted as follows. Put a monkey in front of a telegraph key and let him punch a series of dots and dashes as he pleases. With probability one, the monkey will eventually tap out in Morse code all the sonnets of Shakespeare infinitely often.
16 Chapter 1 Measure Theory Example 3. Random Variables In example 1 let jR„ be the amount of money won or lost at the nth toss. jRn can be thought of as a function on the set 38 or, via the identification 38 <-> J, as a function on the unit interval. It is, of course, just the nth Rademacher function, discussed in §1.1. Rn is a typical example of what probabilists call a random variable. It is a variable—that is, a quantity that one can measure each time one performs a sequence of Bernoulli trials—and it is random, because the values it assumes are a matter of hazard or chance. Another example of a random variable is the sum which is the total amount won or lost by the nth stage of the game. Notice that the set BEk in example 1 is completely described by the S„s. This is not surprising. Most interesting random events are describable by random variables. For instance, consider winning streaks. Suppose that, starting at time t = n, a gambler tosses an unbroken sequence of heads for a certain length of time. The relevant random variable connected with this phenomenon is the variable /„, which counts the number of times H occurs consecutively starting with the nth toss. Example 4. Expectation Values Let 38 be the set of Bernoulli sequences. A random variable associated with the Bernoulli process is, by definition, a function /: 38 -► R. Thanks to the identification of 38 with J, we can also think of / as a function on J. In Chapter 2 we will address the question of what kinds of functions correspond to the "physically interesting" random variables. For these functions we will be able to define the Lebesgue integral (3) jVdfe The probabilists call equation 3 the expectation value of the random variable /. Roughly speaking, it is the value that / is "most likely" to assume in a series of frequently repeated experiments. To use a simplistic example, for the Rademacher function jRn, the integral in equation 3 turns out to be just the usual Riemann integral which, as we saw in the previous section, is zero; that is, the "most likely" value
§1.2 Randomness 17 of Rn is zero (even though Rn takes only the values +1 and — 1). We will justify this somewhat paradoxical assertion in §2.6. Example 5. Random Walks A Bernoulli sequence, that is, a sequence of coin tosses, can be considered to describe a random walk on the real line. That is, a particle is placed at the origin; a flip of a head causes the particle to move one step forward, and a tail moves it one step backward. As one tosses the coin an infinite number of times, the particle moves erratically backward and forward along the real line. We will call the path traced out by such a particle a random path and the sequence of motions itself a random walk. Obviously each Bernoulli sequence gives rise to a random path and vice versa. If we denote by 0t the set of random paths, we can identify & with ^ and, by means of binary expansions, both 0t and S& with the unit interval /. Probabilistic events associated with 0t can be reinterpreted as events associated with & and vice versa. For example gambler's ruin <-> passing through —X for the first time In probability jargon the space of all possible outcomes of a probabilistic process is called the sample space. For Bernoulli sequences, the sample space is 3&\ for random walks the sample space is 0t. For all intents and purposes, 0t and & are identical, even though one thinks of 0t in connection with the motion of particles and & in connection with games of chance. Example 6. Random Walks with Pauses To perform a random walk with pauses, one needs a gadget of the type depicted in the figure below. Place a particle at the origin of the real line and spin the pointer. If it lands on +1, move the particle one unit to the right; if it lands on — 1, move the particle one unit to the left; and, if it lands on 0,
Chapter J Measure Theory leave the particle fixed. By repeating this operation infinitely often, we get a random walk with pauses. Let $bp be the sample space of this process. Identify <%p with / using ternary expansions of points, a> e J; that is, each coel can be written as (4) «=ZS at = 0,1,2 The ternary expansion of co is then co = . ax a2 a3... Notice that ^ = . 10 0 0... or . 0 2 2 2...; so, in order to make ternary expansions unique, we will always choose the nonterminating expansion in cases like the one above (see §1.1, exercise 2). Now make the identification + 1~1 0~0 -l<->2 to identify a random walk with pauses with the digits in such a ternary expansion. This identification gives a map / -+ <%r The Borel principle in this instance says that, if an event E associated with this random process corresponds to the subset BE of /, then, just as before, (5) Prob(£) = iiL(BE) We suggest you check equation 5 for a few simple events. (See exercise 4.) Example 7. Random Walks in the Plane For two-dimensional random walks, we need a gadget similar to the one we used on the previous page: N W X E s
§1.2 Randomness 19 Let Z2 = {(m, n); ra, n integers} denote the integer lattice in the plane. Place a particle at (0,0) e Z2 and spin the pointer. If it lands on N, move the particle to (0, l)eZ2; if it lands on E, move the particle to (l,0)eZ2, and so on. By repeating this operation ad infinitum, one produces a random walk, the successive stages of which are indexed by an infinite sequence such as (6) NSSEWE... Let ^piane be the sample space of this process—that is, the set of all sequences like the one in display 6. We can identify each sequence with a point coel using base-four expansions; that is, we I can be written as (7) *>=£S ** = 0,1,2,3 The base-four expansion of co is then. ax a2 a3..., which can be identified with a sequence like the one in display 6 by means of the correspondence 0 <-> East l^West 2<-»North 3<->South Of course we must deal with the problem of nonuniqueness in this identification as above, by selecting nonterminating rather than terminating expansions whenever ambiguity exists. Just as in example 6, to every event E associated with this process there corresponds a subset BE of/. We urge you to check that Prob(E) = iiL{BE) for a few simple, typical events. (See exercise 5.) Example 8. The Discrete Dirichlet Problem Let Q be a smooth, bounded region in the plane with boundary B. An important problem in electrostatics is the Dirichlet problem: Given a continuous function / on B, find a function u satisfying Am = 0 in Q (8) u = / on B where Am = (d2/d2x)u + (d2/d2y)u. This problem has a discrete analogue that is itself quite interesting. Let Q
Chapter 1 Measure Theory be a finite subset of Z2. A point p = (w, n) of CI is an interior point if its four next-door neighbors (m,n + 1), (m + l,w), (m,n — 1), and (m — l,n) are also in Q; otherwise, p is a boundary point. For instance, in the figure below, pi is an interior point and p2 a boundary point of the shaded region. For a function u on Z2, we define Adiscretew by the formula (Kiscrcuu)(m9n) (9) w(m, n + 1) 4- w(m, n — 1) + w(w + 1, n) + w(m — 1, n) = - w(m, n) [Notice that the first term on the right is just the average of u over the next-door neighbors of the point (m, n).] The discrete analogue of the Dirichlet problem is to find a function u: Z2 -» R such that (10) Adiscreteu = 0 at the interior points of Q, and (11) " = / on the boundary, 3Q, of Q, / being a given function on 3Q. One can solve this problem elegantly by using the random walk described in example 7: Given a point peQ and a random path co starting at p, let F(a>,p) be the value of/ at the first point at which co hits 3Q. [If co never hits the boundary, set F(co, p) = 0.] If we fix p and regard F as a function of the random path co alone, then F is a random variable in the sense of example 3. We will show in §2.8 that its expectation value is the value at p of the solution of the Dirichlet problem described in equations 10 and 11.
§1.2 Randomness 21 Example 9. Randomized Series Probabilistic considerations have another way of entering into classical analysis. Consider the series 00 1 00 (-W Z± and E^ «=i« «=i n The first of these series diverges, whereas the second converges. We can enrich this problem by adding a probabilistic component. Consider a general series 00+1 z— where the plus or minus is determined by the flip of a coin; that is, for each co e I we get a series (12) £ *!.(*) w=1 n with Rn being the nth Rademacher function. Now let E be the event that this series converges: BE = ico e I; ^ converges\ Whatis/iL(BBJ? In §3.3 we give a series of exercises in which we sketch a proof that l*L{BE) = 1. The intuition behind this result is that a typical Bernoulli sequence has roughly as many pluses (heads) as minuses (tails). Example 10 We end this section by considering a collection of sample spaces that includes all of those we have considered up to this point. Take n marbles of k various colors. Say the colors are labeled cl9c2,--.,ck and suppose that, of the N marbles, Nj of them have the color Cj, 1 <j<k. Now put all of the marbles into a cylindrical wire cage that can be spun on its axis to mix the marbles fairly well within the cage. After the marbles are mixed, a blindfolded assistant removes one marble from the cage. If the marble has the color cj9 one gets as a reward a preassigned number, rj9 of dollars. (Incidentally, we will allow r- to be positive or negative.) After the color of the marble is recorded, the marble is returned to the cage and the process is repeated. The probability that the color Cj will be chosen is
22 Chapter J Measure Theory and k k (13) I>j=1 because £ W), = AT Notice that this game can serve as a model for coin tossing (and random walks) by allowing only two colors of equal number, say cx = red and c2 = white, and setting the rewards at +$1.00 for a red marble and —$1.00 for a white marble. With three colors of equal number, say cx = blue, c2 = white, and c3 = red, and rewards rx = +$1.00, r2 = 0, and r3 = —$1.00,'we get a model for the random walk with pauses. If we alter our process slightly by allowing the r£'s to be vectors in R2, we can model random walks in the plane; namely, take four colors of equal numbers with rx = (0,1), r2 = (1,0), r3 = (—1,0), and r4 = (0, — 1). In §2.6 we will develop a measure theoretic model for this process based on a "Borel principle" similar to that in the preceding examples. Exercises for §1.2 1. Under the correspondence <% «-* /, describe the subset of / corresponding to the event that a run of 15 heads will occur before a run of 11 tails. 2. Describe the subset of / corresponding to the event that no run of heads longer than 15 occurs in a Bernoulli sequence. 3. Prove that the pattern H T has to occur infinitely often in a Bernoulli sequence (with probability one) using the Borel principle. 4. With the ternary numbers as a model for the random walk with pauses, test the Borel principle by using it to compute the probability of a. a pause at time t = 1. b. a pause at time t = n. c. forward motion at times t = 1,2,3,..., n. d. forward motion at times t = fe,/c+ l,...,fe + n. 5. With the quaternary numbers as a model for the random walk in the plane, test the Borel principle by using it to compute the probability that a. the first move is due east. b. the wth move is due east. c. the first n moves lie on a straight line. 6. With the ternary numbers as a model for the random walk with pauses, prove that with probability one an infinite number of pauses occur. (Hint: See §1.1, exercise 12.)
§1.2 Randomness 23 7. Sum the series l±? by the following procedure. For each Bernoulli sequence, put a (+) sign in the feth place if a head comes up and a (—) sign if a tail comes up. What is the sum? (Hint: See §1.1, exercise 7.) 8. Let Q be the subset {(0,0), (1,0), (0,1), (-1,0), (0,-1)} of Z2. (That is, Q consists of the origin and its four next-door neighbors.) Check directly that the recipe described in example 8 for solving the "discrete Dirichlet problem" on Q is correct. 9. For the process described in example 10, show that, if one uses an equal number of marbles of each color, the sample space of the process can be identified with the unit interval using expansions in base fe. 10. For the ordinary random walk starting at the origin, show that the probability of a particle's being in position k at time t = n is 0 if |fe| > nor if n + Zeis odd (*) m where r = —-— otherwise 11. (On Markov processes.) Let P = (Py), — oo < ij < oo, be an infinite matrix with the following properties: (i) ^>0 (**) («) £j Pu = 1 for all i (Hi) For fixed i, Pu = 0 for all but finitely many/s. For the "generalized random walk" associated with P, a particle moves along the line according to the following probabilistic rule: If the particle is at position i at time t = w, then at time t = n + 1 it can be at any position j for which Pi} # 0, and the probability of its being there is P0.. (For instance, if Pti = 1 and Ptj = 0 for i # ;, then the particle stays forever at its initial position.) The matrix P is called the matrix of transition probabilities associated with the process. a. Show that the process described in example 10 is a process of this kind. (Think of the position of the particle as being the total number of dollars won or lost by time t = k.) b. For the process described in example 10, show that the matrix of transition probabilities is of the form Ptj = P£_j. c Show, conversely, that if Ptj = P^j the corresponding process is a process of the kind described in example 10.
24 Chapter 1 Measure Theory 12. a. Show that, if P and Q are matrices of the form in equation (**), the usual matrix product PQ is well-defined and is of the same form. b. Show that, for the generalized random walk associated with P, if the position of a particle at time zero is i, the probability that its position at time t = nisj is just the i —jth entry of the matrix Pn. 13. Show that, for the matrix PM+1 = Pi+iti = i> Pij = 0> otherwise, the generalized random walk is the usual random walk. Derive the formula (*) of exercise 10 by computing directly the i —jth. entry of Pn. (Hint: Consider the vector space V consisting of all finite sums: Y,ckekt cksR On this vector space consider the linear mapping "multiplication by (e~l + e')/2." Show that, if we take for a basis of V e~kt e~l 1 el ekt ...c ,...e ,x,e,...,e ,... then, in terms of this basis, this linear mapping has P as its matrix.) 14. Can you construct a measure theoretic model for random walks in space similar to the measure theoretic model for random walks in the plane? (Hint: Expansions in base six.) §1.3 Measure Theory We mentioned earlier that Lebesgue measure assigns to each set A, belonging to a certain collection of subsets of R, a nonnegative number liL(A) called the Lebesgue measure of A. We also mentioned that \iL has certain additivity properties. We will now study these properties in more detail. We need to begin with a large number of technical definitions. Keep in mind the vague notion of Lebesgue measure we have already discussed so as to put these technicalities in perspective. Let AT be a fixed set. Suppose A and B are subsets of X. We recall the following notation: Notation Meaning 0 empty set AkjB = {xeX;xeA or xeB} union of A and B Ar\B = {xeX;xeA and x e B} intersection of A and B Ac = {x e X; x $ A) complement of A B — A = {xeX; xeB and x$ A} B minus A S(A9 B) = (A — B)<u(B — A) symmetric difference of A and B (see figure, page 25)
§1.3 Measure Theory 25 A ring of sets in X is a nonempty collection ® of subsets of X satisfying the following two properties 1. AuBe® whenever/!, BeM 2. A- Be0t whenever A, Be®, Remark. 0e@ since A - A = 0. Also if A, Be® then A n Be® since A H B is obtained from A U £ by deleting A - 5 and 5 - A. Two examples with which we will soon be very familiar follow. Example 1. Let 2X denote the set of all subsets of X; 2X is a ring. Example 2. Let X = Rn. Suppose {ax,...,an) and (bl9...9bn) are given, with each at <bhi= 1,2,..., n. Let A be the set of points xeR" such that (1) at <xt < bt i= l,...,n A is called a multi-interval. More generally, a multi-interval is a set of the form shown with perhaps some of the <'s replaced by a <. Define ®Leh by Ae®Leh<=>A = \J?=iAh where the A/s are a disjoint collection of multi-intervals. We let the reader check that ®Leh is a ring. Now, fix a ring 0t of subsets of X. Let \i be a nonnegative set function on ^; that is, to each AetX, p assigns a nonnegative number \i{A). Definition 3. \i is additive if /z(A u5) = /z(A) + /z(£) whenever A, Be® are disjoint. Example 4. 0t — ®Leh. Suppose Ae®Lch is a multi-interval described by the inequalities at <Xi<bi i = 1,...,m (Again, some <'s may be replaced by <'s.) We define (2) rtA) = (bi-a1)(b2-a2)-(bn-an)
26 Chapter 1 Measure Theory More generally, if A = (JfLi Ai is a disjoint union of multi-intervals, we define (3) A*M)*5>Wi) I t=i Then u is a well-defined additive set function on ^Leb* Proposition 5. Let 0t be a ring of subsets of AT and /z an additive, nonnegative set function on 0t. Then 1. M0) = O. 2. (monotonicity) HA9Be@ with A c 5, then /z(A) < n(B). 3. (finite additivity) If Al9A2,.-*,An€<% are mutually disjoint, then MU?=i^-) = Z?=iM^). 4. (lattice property) If A, B e 0t then fi(A uB) + /z(A n5) = /z(A) + fi{B). 5. (finite subadditivity) For any Al9..., Ane@, fi(\JUi A{) < JJ.j /*(>!;). Aw/. 1. A e ®9 ii(A) = ii(A u0) = /i(X) + ix{0\ so /i(0) = 0. 2. £ = (£ - A) u 4 is disjoint, so /x(£) = /*(£ - A) + /*(A) > M>0- 3. Induction on n. 4. A = {A-B)Kj(AnB) B = (B-A)Kj{AnB) AKjB = (A-B)Kj{B-A)Kj(AnB) so fx(A) = fi(A - B) + fi(A n 5) p(JB) = /i(B-i4) + /iC4nB) /*(4 u B) = |i(i4 - £) + /i(B - A) + ^(i4 n 5) 5. Induction on m; case n = 2 follows from item 4. □ So far we have done nothing very deep. We have just given an abstract setting for the situation in example 4. Our eventual purpose is to extend the definition of the set function in example 4 to a much larger ring of subsets of R. For instance, this ring should contain the sets of measure zero described in §1.1. In order to carry out this extension in a natural way, we will need the following refinement of additivity. As the proof of theorem 7 will suggest, this property is much more intricate than finite additivity. Definition 6. 1. Let 0t be a ring of subsets of X and \i an additive set function on 31. We say \i is countably additive on 0t if, given any countable collection {A£}f* x c 0t with the Afs mutually disjoint and such that A = \Jf^t A{ is also in 0t9 then
§1.3 Measure Theory 27 (4) M^)=ZA*Ml) i=l 2. A countably additive, nonnegative set function fi on a ring ^ in X is called a measure. Warning. Equation 4 makes no sense unless we assume Ae0t, Theorem 7. If X = Rw, ^ = ^Lcb5 and \i is the set function in example 4, then /* is a measure. Lemma 8. Let A e @Leh9 and let e > 0 be given. There exist F,Ge ^Leb such that F is closed, G is open, F ^ A^G, and /i(F) > /i(i4) - s li(G) < ii(A) + e Proof. Suppose A is a multi-interval given by the inequalities ax <xt <bt i= l,...,n where some of the <'s may be replaced by <'s. We can find a 8 such that Y\l(bi-d)-(ai + d)l = Y\(bi-ai-2d)>Y\(bi-ai)-s »=1 i=l i=l and II [ft + d) - (a, - *)] = n& - a. + 25) < FI& - «i) + « t=l 1=1 i=l Let F be given by the inequalities at + 5 < Xi < bt — 8 i = 1,..., n and G by the inequalities af — <5 < x, < 6f + <5 i=l,...,n We then have H(F) > ii(A) - s 11(G) <; 11(A) + 6 Now, if A = U}sl i4£ is a disjoint union of multi-intervals, find for each At an Ff and G£ such that
28 Chapter 1 Measure Theory fift) 2: p{At) -1 MGi)^K) + | Then, with F = (JJ=1 Ft and G = (J?=1 Gf, we have MO * E A*(G|) <S E U^) +1] = A*W) + « V Now, take {A^t to be a disjoint collection of sets in ^Leb, and suppose /I = (J.%! /!(is also in ^Leb. Notice that (Jf^ At e A, so (5) fi(A) £ // f (J X^ = £ ^(yl,) for every N Thus (6) li(A) :> £ M4) £=1 Choose a closed stt F ^ A such that /*(F) > fi(A) — e, and for each A{ choose an open set Gx containing A{ with /z(Gf) < ^(Ai) + 6/2*. Because F is closed and bounded, it is compact. Because it is covered by the Gf's, it must be covered by a finite number of them, say Gx, G2,..., GN. Then li{A) - a < ii(F) < ji(jj Gf) < t m) < t [a*Wi) + j] ^ Z MWi) + £ Being true for all e, this yields (7) l*(A) zttiAi) Putting inequality 7 together with inequality 6 shows that \i is a measure. □ We have now constructed a measure on a collection of subsets of R". The sets on which this measure is defined, MLeb, are very simple, however. As remarked above, the property of countable additivity will allow us to extend this measure to a much larger ring of sets. Let \i be a measure on a ring M in X. We attempt to extend \i to the ring 2X by mimicking the definition of measure zero in §1.1.
§1.3 Measure Theory 29 Definition 9. Let A be a subset of X. A number / > 0 will be called an approximate outer measure of A if there exists a covering of A by a countable collection of sets Al9A29A39... with each A{eM such that (8) i^At)<l Remark. I is allowed to be + oo. Definition 10. Let A be a subset of X. The outer measure of 4, /i*(>4), is the greatest lower bound of the set {/: / is an approximate outer measure of A). If this set is empty, then ix*{A) = + oo. We now have a set function, fj,*, on the ring 2X. Unfortunately, /x* is not generally a measure (see, for example, exercise 1, or, for a more rewarding example, Appendix C). We will show, however, that \i* is a measure on a large ring of subsets of X; this ring will be called the ring of measurable sets in X. Proposition 11. 1. If A g 0, then y*(A) = fx(A). 2. If A <= B9 then fi*(A) < fx*(B). 3. fi* is countably subadditive; that is, if Al9 A2, A39... are subsets of AT, then Proof. 1. Covering A by the sequence Ax = A9 A2 = 0, A3 = 0,..., we see that fi(A) is an approximate outer measure for A9 so (9) ti*(A) < ti(A) To prove the other inequality, let e > 0 be given. Because f**(A) is the greatest lower bound of all approximate outer measures of A9 a cover {A^%t c= 0t must exist such that (10) fi*(A) + e^t^Ai) i=l Let A\ = Al9 A'2 = A2 — Al9 A'3 = A3 — {At u A2)9 and so on. Then the A"s are mutually disjoint and (11) fi*(A) + e > t VW) i=l If we let A" = A\ n A9 we have that A[ ^ A for all i9 the /4J"s are mutually
Chapter 1 Measure Theory disjoint, and still (12) M*(^) + e>X>W) Now, since A" ^ a for all i and (j£x At => A, we must have {JfLx A" = A. So V(A) = L-i A*W) and thus (13) /i*(^) + £ > /*M) Because this inequality is true for all £, we have (14) »*(A) > n(A) 2. If / is an approximate outer measure for B, then surely it is for A. Thus fJL*{A) < fi*{B) 3. Given e > 0, for each i we can find a cover, {;4,j}j!Li <= ^, of ^ such that (15) ^%4,) + J>lM^) Then the countable collection {Aitj}™j=1 covers A = (J£x A{ so that *,j=i t=i i=i (16) < -J+f/1%4,) 1 = 1 This holds for all e > 0, so (17) /*%4) <: |>%4,) □ i=l Remark. This proof is essentially the same as that used in §1.1 to show that a countable union of sets of measure zero is itself of measure zero. Now our original ring, 0ty is a subset of 2X. We wish to find a larger ring, M, containing ^, that will be the measurable sets. Our strategy will be to think of 2X as a metric space and define a distance function on it, so that, roughly speaking, Jt will be the closure of 0t in 2X with respect to this distance function. (For a quick review of metric spaces, see Appendix A.)
§13 Measure Theory 31 For A, B c= X we define the distance from A to B by (18) d(A,B) = fi*lS(A,Bn where S(A, B) is the symmetric difference S(A9 B) = (A — B)v(B — A). If A and 5 are the unit squares pictured, then d(A, B) = 1^. Caution. 1. d(A9 B) may be + oo. 2. Although we are calling d a. distance function, d(A9 B) = 0 does not necessarily imply A = B. Proposition 12. Suppose A, B, Ce2x. Then 1. d(A9B) = d(B9A) 2. d(A9A) = 0 3. d(A9B) + d(B9C)>d(A9C) Proof. Lemma 13. 1. S(A9B) = S{B9A) 2. S(A9A) = 0 3. S(A9B)vS(B9C)=>S(A9Q Proof. Items 1 and 2 are obvious. To see item 3 we have S(A9B) = {A-B)kj{B-A) and S(B9C) = {B-C)u{C- B) so- S(A9B)uS(B9C) = {A-B)u{B-A)kj{B- C)u(C - B) But ^~Cc(A-B)u(5-C)
32 Chapter I Measure Theory and C - A s (B - A) u(C - B) so S(/4, C) = {A - C)u(C - A) £ S(X,5)u S(B, Q V The proposition follows from the lemma. □ Note. d(A, B) = 0 if n*{S(A, B)) = 0; that is, A and B symmetrically differ by a set of outer measure zero. Although the preceding note says that d is not quite a distance function in the standard sense, we can still use d to define the notion of convergence in 2X. That is, we say a sequence {/lj}j™i e2x converges to A e2x, written At -* A, ttd(A„A)-*0. Proposition 14. The Boolean operations in 2X are continuous with respect to d. That is, if An -► A and Bn -* B in 2X, then An<oBn-* AkjB A„nB„-* AnB An-Bn-+A-B and Acn -> Ac Proof. Lemma 15. If AltA2, B1,B2e2*, then 1. S(A<uB\) = S(AuBl) 2. S(^! uAj.B, uB2) s S^.BJuS^.Bj) 3. S(A, ni42,B, nB2) cS^LB^uSf^,^) 4. S(>l1-yl2,B1-B2)sS(yl1,B1)uS(i42,B2) Pro©/. 1. S{A,B) = (/4 - B)u(B - A) = (/InB°)u(Bn/lc) so S(AC, W) = (/4£ n B) u (B° n ,4) = S(/4, B) 2. S(/l1u4B1uB2) = [(>l1u^)-(B1uB2)]uPiuB2)-(/(1u/l2)] = [(^1u^2)n(B1uB2)c]u[(B1uB2)n(/41u/42)c] = [(^! u/42)n(Bc1 nB2)] u [(Bj uB2)n(^C! n/42)] s (/41 n B0!) u (/12 n B^) u (Bt n A',) u (B2 n 42) = S(>l1,Bl)uS(/l2,B2) 3. S^j n /42,B! n B2) = S(A\ u 4C2, B^ u B2) SSMl.BDuS^Bi) = S(X„B1)uS(/(2,B2) 4. S(/4! - 42, Bt - B2) = S(.At n 4C2, B, n Bc2) sS(i41(B,)uS(4fli) = S(>l1,B1)uS(>l2,B2) V
§1.3 Measure Theory 33 From the lemma we have the immediate corollary 1. d{A,B) = d(Ac9Bc) 2. d(AluA2,BiKjB2) < d(AuBx) + d(A2,B2) 3. d(Ai n A29B1 nB2) < d(AuB{) + d(A2iB2) 4. d{Ax - A2,BX - B2) < d(AuBl) + d(A2iB2) from which the proposition follows. □ Proposition 16. fx* is continuous in the following sense: Let A, Be2x and suppose either fi*(A) or fi*(B) is finite, then (19) \fi*(A) - fx*(B)\ < d(A,B) Proof. Suppose fi*(B) < oo; also assume fji*(B) ^ n*(A). Then li*(A) = d(A,0) <d(B,0) + d(B,A) = V*(B) + d(B9A) Thus \lx*{A) - ^(B)\ = ii*(A) - fi*(B) < d(B, A) □ Definition 17. Let MF be the closure of 0t in 2X. That is, A e JtF if and only if there exists a sequence of sets {i4£}," t c 0t such that d(Ah A) -► 0 as i -» oo. Theorem 18. 1. JtF is a ring. 2. For A e JtF, fi*(A) < oo. 3. fi* is a measure on ^F. 1. Assume A, Be JtF. We need to show that AuB and A — B are in *MF» Now, because A, BeJtF there are sequences {Aj-3^ and {5j£i in 0t such that y4f -► A and ££ -» B. By the continuity of the Boolean operations At kjB^AkjB Ai-B^A-B so JtF is a ring. 2. A eJF implies that there is a sequence {A^fLx <= ^ with 4£ -► A. For some m, then, d{An9 A) < 1.
j*» Chapter I Measure Theory Thus lx*(A) < fi*{An) + 1 < oo 3. We first show that fx* is additive, or—what amounts to the same thing—we will prove the lattice property; that is, if A, Be JtF then (20) ii+(A uB) + fx*(A n B) = fi*(A) + fi*{B) Choose An -» A and Bn -» B in 01. Because on 9t, /z* = \i, fi* is additive on M. Thus li*(An u Bn) + m*(^4„ n B.) = ji*(4,) + /**(£„) But An u Bw -*• X u B and y4„njB„->An5,so the continuity of y* implies li*(A u B) + /x*(/l n5) = /**(/*) + /**(£) We now prove countable additivity. Let {A^fLi be a mutually disjoint sequence in JtF with 4 = (j£t >4£ also in ^#f. By the subadditivity of ft* we know that (21) »*(A) < £ /i%4,) i=l Furthermore i=l H*(A) > n* ( y XjJ = £ m*(^i) for all N That is, (22) /*%4)>f>*K) D 1=1 Definition 19. A is a measurable set, AeJi,\ithere exist {AJ£*x cr ^F such thati4 = Ui-Mi- Theorem 20. \iAsM then Xe^F<=>/z*(i4) < oo. Pnra/. Part 2 of theorem 18 gives "=>", so to establish the theorem we must show that, if fi*(A) < oo and AeJi, then AeM¥. Because A e Jt, there exist A£ g M? such that A = \]%\ At. We can assume this union is disjoint for, if it isn't, we can replace the i4t's by ^'s as follows: so 1,=AX
§13 Measure Theory 35 A2 — A2 — A1 ^3 = A3-(A1uA2) and, because M¥ is a ring, we know A( e M¥. Thus we can assume A = [jflx At is a disjoint union. Now consider ix*(A). First, subadditivity gives fi*(A) < Yj^i /J*(^i)- We claim that, in fact, ix*(A) = Yfii l**^). To see this, notice that so [JA^A i=l Because this equality holds for any N, we have 1 = 1 and thus (23) £ ix*(Ai) = lx*(A) Now, fix e > 0 and let BN = \J"=1 i4£; then BNsJtF, and (24) W ' because J^ /i*(i4£) is convergent. Thus /I eJt¥ because BN-+ A and ^#F is closed. □ We now consider properties of the collection M. Definition 21. Let Sf be a collection of subsets of a set X. ¥ is called a a-ring if 1. it is a ring and 2. given {Aj^ in Sf, \)?=x A£ is also in Sf. Theorem 22. J( is a a-ring. /V00/. First we will show property 2. Suppose y41,y42>... are elements of Jt. Let /I = (J^ ^£. Because each y4£e^#, there must be {A^}^ in MF such that
36 Chapter 1 Measure Theory At = U *u i=i Then A = U*v/=i ^y' a countable union, so /I eJi. Now we will show that Ji is a ring. It suffices to show that if A9 BeM ihQTiA-BeJt. First, suppose A e M¥ and write B = U£ t J5£ with J5£ e */#F. Because JiF is a ring, A nBteJiF so i4n5 = U«^i^n^ *s a member of Jt. Moreover, li*(A n B) < ix*(A) < oo, so Ar\BeMF. Now ^-J5 = ^-(^nJ5) and, because c^F is a ring of which A and ^4 n J5 are members, we have A — Be JtF. Now let A be a general element of ^# and write ^4 = \j£x i4£ with A£ g ^#f. Then but from the discussion above, A,- — J3 e ^#F, so we are done. □ Theorem 23. If A1, A2,... is a countable collection of disjoint sets in ^#, then Proof. Let >4 = (j£ i /l^, y4 e ^#. We consider two cases separately. 1. ix*(A) < oo Because At c /I, /x*(y4f) < oo so A and all of the A?s are elements of JtF. Because ju* is a measure on ^#F, we have then (x*(A)=fifi*(Ai) 2. fj.*(A) = oo In this case subadditivity tells us that oo = fi*(A) ^ f /,%4i) t=l 00 SO £ A**(4) = °° □ Now that we have constructed measurable sets in the abstract case, let us return to the example of Lebesgue measure in R". Example 24. X = Rn, ^ = ^Leb = finite unions of multi-intervals, and fi is as given in example 4. Here we call J( the set of Lebesgue measurable sets in
§1.3 Measure Theory yj Rn, and the extension of \x to Jt (the restriction of /x* to Jt) is called Lebesgue measure \ih. What do the sets in Jt look like? First we remark that RneJt. Indeed let IN = {xeRn; -N <Xi<N9i= l,...,n} Then 00 R" = (J iN and each INeM c JtF N=l Proposition 25. Every open subset of Rn is in Jt. Proof. Let c = (al9a2,--.,an,blib2,...,bn)e R2n with the a£'s and fy's rational and a{ < b(. Let Ic = {xe Rn; ai<xi<bi,i= 1,2,..., n}. The collection { Jc} is countable. Now let 0 be any open subset of Rn; 0 is equal to the union of all sets Ic such that Ic c (9, (If xeO we can find a c such that xe/cc 6?.) Because such a union is countable, GeJt. Corollary 26. Every closed subset of Rn is in Jt. Proof. A is closed so Ac is open. A = Rn — Ac and, because Jt is a ring and Rn, Ac 6 ^, we have AeJt. Corollary 27. All countable unions and intersections of closed and open sets are measurable. We have shown that the measurable sets are a <x-ring containing the open subsets of Rn. Are they the smallest <x-ring with this property? That is, if one starts with the closed and open sets, forms countable unions and intersections, and then from these forms further countable unions and intersections, and so on, does one eventually end up with all measurable sets? The answer is no, unfortunately; so we are forced to make the following definition. Definition 28. The Borel sets are the smallest a-ring containing the open sets. Although not all measurable sets are Borel (for an example see Halmos, P. Measure Theory. [Van Nostrand: Princeton, NJ] p. 67), the following theorem says that any measurable set is close to being a Borel set. Theorem 29. HAeJt there exists a Borel set B ^ A such that [j*(A - B) = 0; that is, A can be written as A = (A — B)vB9 where B is Borel and tx*(A -B) = 0.
38 Chapter 1 Measure Theory Lemma 30. XiAeJl and if e > 0 is given, then there exists a Borel set G such that G Z5 A and n*(G - A) < e. Proof. First suppose ix*(A) < oo. Then by definition of /x* we can find a cover A c UJSjAj such that where each of the A/ 's is a multi-interval; so we take G = U£j Ar-, which is Borel. More generally, if AeM we can write A = (j£x A,- where each A{tJ(F. By the preceding argument we can find Borel sets G{ with At c Gt and /z*(G£ - A,) < e/2l. Then with G = (JSi Gi we have /**(G - /4) < a V Lemma 31. If AeJt there exists a Borel set F c A with li*(A -F)<e Proof. Choose a Borel set G such that Ac a G and /x*(G — Ac) < e by lemma 30. Let F = Gc. Then /4 - F = G - Ac and ^♦(,4 - F) = /x*(G - Ac) < e V Now we prove theorem 29. Take A e Jt. For every N choose a Borel set FN(=A such that /x*(A - FN) < l/N. Let F = (j£=1 F^; then F is Borel and /**(>! - F) < ix*(A -FN)<~- for every N N Thus ix*(A - F) = 0. □ We conclude this section with a few remarks about notation. Let X be a set, 0t a ring of subsets of X, and /x a measure on ^2. By theorem 18 \x extends to a measure on a much larger ring of sets, Ji¥. In fact \x can be regarded as a measure on the a-ring Jt, providing we define it to take the value + oo on sets A that are in M but not in JtF. Note that proposition 5 is then still true if one observes the usual addition conventions for + oo, namely ( + oo) + a=+oo foraeR and (-f-oo) + (-foo) = -hoo Moreover, \i is countably additive on Jl by theorem 23. Note that, if X itself is in Jif, these problems with infinity don't arise; that is, J(F = J(. For all examples of measures that we will encounter in this text, the set X is either in JtF or in Jl—that is, X satisfies the conditions of the following definition.
§1.3 Measure Theory 39 Definition32. X is cr-finite if there exist sets X^J^p, i= 1,2,..., with For example, Rw is a-finite because R"=LM i=l with Bi being the ball of radius i about the origin. Exercises for §1.3 1. Let X be an uncountable set. Let & be the collection of all finite subsets of X. Given Aeffllet fi(A) be the number of elements in A. Show that 0t is a ring and that \x is a measure on ^?. Identify /x*. What are «/# and Jt/i Is every subset of X measurable? 2. Let X be an infinite set and let ^ be the following collection of subsets: Ae0t\i and only if A is finite or AC is finite. Let \x be the following function on M\ fi(A) = 0 if A is finite, and fi(A) = 1 if >4C is finite. Is \i a measure? 3. a. Let X be an infinite set and St the collection of all countable subsets of*. Is^aa-ring? b. Let \i be a measure on 0t. Show that there exists a function f\X ^ [0, oo) such that (*) AiM) = Z /(*) foralM60. c. Show that the function / in part b has to have the following two properties: (1) The set {xeX;f(x) ^0} is countable and (2) d. Show that, if/ has the properties in part c, the formula (*) defines a measure on 0t. 4. Let X be the real line and & — t%Leh. (That is, finite unions of intervals.) Given AeSt\t\. fi(A) = 1 if, for some positive e, A contains the interval (0, e). Otherwise let fx(A) = 0. Show that \i is an additive set function but is not countably additive. 5. Let F be a continuous, monotone increasing function on the real line. If A is an interval with endpoints a and b, let VM) = F(b) ~ F(a) More generally, if A is a disjoint union of intervals N A=\J A, i=l
40 Chapter 1 Measure Theory let fiF(A) = Yj=i ^f(^i)- Show that fxF is a measure on the ring ^Leb; that is, prove it is countably additive. Remark. If one takes for F an antiderivative of the function (\/^j2n)e~x212, fxF is called the Gaussian measure. We will encounter it several times later on. 6. a. Let A be a measurable subset of R. One says that the density of A is well defined if the limit r-oo 2.1 exists. If the limit exists, this expression is called the density of A. Can you produce an example of a measurable set A whose density is not defined? b. Show that, if A1 and A2 have well-defined densities and are disjoint, then Ax u A2 has a well-defined density and D(AlvA2) = D(A1) + D(A2) c. Show that there exist sets A and Ai9 i= 1,2,..., with well-defined densities such that 00 ^ = (J At (disjoint unions) t=i but D(A)^ltD(Ai) 7. Let X be a set, 0t a a-ring of subsets of AT, and \ix and fi2 measures on 9t. Let ££ be the family of all those sets Aefflfor which ixx(A) = fi2(A). Assume Xe@ and fx1(X) = f*2{X) < oo. Show that S£ has the following properties: (i) XeS£. (**) (ii) If A, Be& and B s A9 A - BeS£. (iii) If A( e if, i = 1,2,... and /I = (J^ y4£ (disjoint union) thence if. Remark. A collection of sets if having the properties listed in (**) is called a A-system. 8. Let X be the three-element set {Px ,P2,P3 }, and let ^ be the ring of subsets of AT. Let ix t and \x2 be measures on 0t. When is the set ££ a ring? Show that S£ doesn't always have to be a ring. 9. Show that the example described in exercise 1 is not cr-finite. 10. Remember that a metric space is complete if every Cauchy sequence
§1.3 Measure Theory 41 has a limit. Show that, with respect to the distance function d(A9 B) = fi*(S(A, B))9 2X is complete. 11. a. Given any collection # of subsets of a set X, show that there is a smallest ring of sets 0t containing <€. (That is, M has the property that it contains #, and any ring that contains # contains M) Describe explicitly how to construct 0t from #. b. Show that the ring ^Leb is the smallest ring containing the multi- intervals. 12. Given any collection # of subsets of a set X, show that there exists a smallest a-ring of sets 0ta containing #. This justifies definition 28. 13. Show that for any 8 > 0 there exists an open dense subset U of R with iiL(U) < 5. 14. Let ceR". Given any subset A of Rn, let A + c = {weRn;w - ceA}. Prove that, if A is measurable, then A + c is measurable and (**) M^ + c) = ixL(A) (Hint: First, prove this for multi-intervals. Next, show that in equation (**) the outer measures are equal.) 15. Let /: R -► R be the linear mapping x -> ax + b9 a and b being constants with a > 0. Show that, if X is measurable, f(A) is measurable and 16. Let ^ be a Lebesgue measurable subset of R and let CA = {(x,y)eR2;xeA} Such sets are called cylinder sets. Show that the collection of these sets forms a ring 0tc, Show that the set function \ic defined by Pc(CA) = VM) is a measure on this ring. Show that, if S is a proper subset of R and fxL(A) # 0, the set /I x S = {(x9y); xeA9yeS} is not a measurable subset of R2 with respect to the measure \xc. (Hint: What is its outer measure, computed with respect to ficl) 17. Let /: Rw -> Rn be a continuous map. Show that, if A is a Borel subset of R", then f~\A) is a Borel subset of Rw. Define VM) = M/_1(>4)) Show that fif is a measure on the Borel subsets of Rn. 18. Let £% be a ring and /x a measure on ^. Prove that, if Al9A2,*.*,An are in 0t then
42 Chapter 1 Measure Theory fi(A1 u • • • u An) = £ /x(^) - £ /x^ n Ay) £</<* + (-irV(^in---n^J 19. Let X be a set, 0t a ring of subsets of X9 and /x a measure on ^. Let /x* be the corresponding outer measure. Show that if A eJi¥ then, for all E^X, (t) »*(A n £) + ju*(/1c n £) = ,u*(£) (Hint: First check this for /I e^.) 20. Prove the converse result; that is, suppose n*(A) is finite and A satisfies property (f) for every subset E of X. Prove that A e Jt¥. (Hint: Show that, for every set A and every e > 0, there exists a set EeJi such that E => A and ju*(£) < fi*(A) + e. Use property (f) to conclude that d(E9A) < e). Remark. In many textbooks the property (f) is used as the definition of a measurable set. §1.4 Measure Theoretic Modeling Now that we have developed the basic notions of measure theory we can examine a little more closely the ideas involved in what we have called the "Borel principle." First we provide some definitions. Definition 1. Let X be a set and 3* a ring of subsets of X. 1. ^isa/ieWifXe^. 2. !F is a a-field if X e !F and if ^ is a a-ring. Definition 2. Let X be a set and !F a field of subsets of X. Suppose \x is a measure defined on «^\ Then ju is a probability measure iffi(X) = 1. In this case the triple {X9 «^, /*} is called a probability space. Example 3. Let X be the unit interval /, and let 3F be the measurable subsets contained in /. Then !F is a cx-field and the Lebesgue measure is a probability measure. Now let X be the sample space of a probabilistic process. A measure theoretic model of the process is a <7-field &* of subsets of X and a probability measure /x defined on !F so that, for any "plausible" event E in X9 we have
§1.4 Measure Theoretic Modeling 43 BEe^ and Prob(E) = ix(BE\ where BE is the set of points in X for which E occurs. Of course this definition is not precise; the word plausible is left to be interpreted by the modeler. We need to include the plausibility qualification because the desired measure may not be defined on all subsets of X. For example, as we show in Appendix C, not every subset of the unit interval is Lebesgue measurable. Let us consider some examples of measure theoretic models. I. Discrete Probability Theory Suppose the sample space X is finite or countable, say X = {xl9x2,x39...}. Further, suppose that each point x£ has the probability ft of occurring and that Yj Pi = 1- The measure theoretic model for this process is given by letting ^ be the collection of all subsets of X and by defining \i as (1) fx(A)= £ Pl torA^X xteA It is left as an exercise for the reader to check that \x is a measure. (See exercise 3 in §1.3.) Notice that in this case we need not interpret the word plausible because SF contains all subsets of X; that is, all events are considered plausible. II. Bernoulli Sequences and Random Walks In this case the sample space can be identified with the unit interval /, and the measure theoretic model is given by the Borel principle. In §1.2 we saw that, for many "plausible" events E, BE is a finite union of intervals (and thus measurable) and that Prob(£) = fi(BE) in these cases. The events considered there were rather simple; let us now confirm that BEs^ for some more complicated, yet still "plausible," events. 1. Let E be the event that a prescribed finite pattern, for example, H T T H, occurs infinitely often. To describe BE we let En be the event that the pattern occurs beginning at the nth step. Because BEn is described by a finite number of conditions on the Rademacher functions, it is a finite union of intervals. (If the pattern is HTTH, then BEn = {(oeI; Rn{co)=l, Rn+1(co)= -1, /J„+2(fi)) = — 1, Rn+3(co) = 1}.) Thus, because (2) BE=f] [jBEn it is Borel.
44 Chapter 1 Measure Theory 2. (Law of large numbers) Let E be the event that a Bernoulli sequence obeys the law of large numbers. That is, with Sn(co) = ££=1 Rk((o)9 SJco) BE = \(dgI\ -^-^-+ 0 as n -> oo ■{• We know this is measurable because we showed its complement has measure zero. However, let us describe BE as a Borel set. Recall that the statement 5w(co) -> 0 as n -► oo means that, for every integer r > 0, there is a fc> 0 such that S>) If we let (3) we can write (4) < - whenever n > k r f r |S„(<b)| 11 r=lfc=in>k which is Borel, because each An%r is a finite union of intervals. 3. Let E be the event that £ (Rn(a>)/n) converges. We claim BE is Borel. Let (5) *=1 Then the Cauchy condition tells us that X(K„(co)/n) converges if, for every integer r > 0, there is a fc> 0 such that | Tn(co) - Tm(co)\ < - for all m9n>k If we let n,r = \« we see that (6) which is Borel, because Amnr is a finite union of intervals. bb - n u n ^.n,
§1.4 Measure Theoretic Modeling 45 We have now shown that these "plausible" events correspond to measurable sets. Thus, if we assume that the Borel principle holds, we can determine the probability of these events by finding the measure of these sets. We know already that, for the law of large numbers, the set described has measure one. We now develop the necessary tools to determine the measure of the set described in event 1. This example is a special case of the following general situation: Start with a countable collection of events {E1,E2,...} and define a new event E to be the event that infinitely many of the events Et occur. Can we determine Prob(£) if we know Prob(jEt) for all i? Two theorems address this problem; they are called the Borel-Cantelli lemmas. In order to formulate them, we first restate the problem in measure theoretic terms. Let X be the sample space of our process, equipped with a a-field !F and a measure \i. Let Bt denote the subset of X on which E{ occurs. We assume that Bte^ and that Prob(jEj) = /x(Bf). If we let BE denote the subset of X corresponding to the event E, then, in terms of the 5/s, (7) BE=(\{JB„ k=ln>k We give this a name. Definition 4. Given sets Bt, B2,B3,... in &r, then (8) {Bl.;i.o.}=limsupBn= f] [j Bn is called "Bi9 infinitely often" or the limes supremum of the Bf's. Theorem 5. (First Borel-Cantelli lemma) Given BUB2>... in ^, let B = {Bt; i.o.}. Then £>i fifa) < oo implies that fi(B) = 0. Proof. Let Ak = \J„>kBn so that B = [)?=! Ak\ in particular, B ^ Ak for all k. Now, by subadditivity Thus, because Y£=i M(Ai) < oo, for e > 0 there is a k > 0 such that M(4k) ^ I KBn) < e n^k Because B ^ Ak we have that fi(B) < e, and because e is arbitrary we must have fi(B) = 0. □
46 Chapter 1 Measure Theory Application. (Run lengths) For oel define the nth run-length function /„ by letting l„(co) be the number of consecutive l's in the binary expansion of co starting at the nth place. That is, ln(co) = k if Rn(co) = 1, Rn+1(co) = 1,..., Kn+k_i(a>) = 1, and Rn+k(co) = -1. Now take a sequence of non-negative integers, ri,r2,r3,..., and let En denote the event that ln(uj) > rn. Let E = {En ;i.o.}. Then BEn = {coeI; R„(co) = Rm+1(a>) = • • • = R,vl(a)) = 1} so ii{BEr) = (|)rn and we can use theorem 5 to conclude the following. Corollary. If £n°°=1 (1/2/" < oo then fi(BE) = 0. The second Borel-Cantelli lemma supplies a partial converse to the first. It is restricted by applying only to independent events. Definition 6. Two events Et and E2 are independent if the outcome of Ex tells us nothing about the outcome of E2. Let us try to make this definition more precise by restating it in measure theoretic terms. Knowing that the event Ex occurs means that the elements of the sample space in which we are interested are already in BEr Now, for what proportion of the elements in BEl does the event E2 occur? Clearly the answer is (9) KBElnBS2) 1 ' KBEl) This ratio is called the conditional probability of E2 given Et. Now, if E2 is independent of Eu this conditional probability is just the probability of E2 computed without prior knowledge of Ex—that is, n(BE2); hence KBe^ mow This leads us to the following measure theoretic definition. Definition 7. Let A" be a sample space with c-field !F and probability measure \i. Two sets Al9 A2e!F are independent if (10) VL(AlnA2)-11^)^2) Example 8. Given X = /, \i = Lebesgue measure, Al = {coel; R^co) = 1}, and A2 = {coel; R2(co) = 1}, then ^=(il] and ^2 = (ii]u(il]
§1.4 Measure Theoretic Modeling 47 AtnA2 = &ll Thus, ii(Ax n A2) = i = (i)2 = KA1)fi(A2) Definition*). More generally, Al9A2,...,A„ are independent if, for any sequence of integers 1 <i1 < i2 < • • • < ik < n, we have (11) MX£i n Ai2 n • • • n XJ = ji^i,)/z(Af2) • • • fi(Aik) Further, a countable collection of sets is independent if every finite subcollec- tion is independent. Example 10. Let At = {coel; R^co) = 1}. It is left as an exercise to prove that the collection Ax, A2,... is independent. Theorem 11. (Second Borel-Cantelli lemma) Let (AT, ^, //) be a probability space and let Al9A2,... be an independent collection of sets from J5". Suppose that Yj^i^t) = °°jtnen Ki^il i-°-}) = 1- Lemma 12. Let i4l9 A2,... be an independent collection of sets in «F. Then A\, Ac2, A%,... is an independent collection of sets in SF. The proof of the lemma is left to the reader. (See exercise 10. We suggest that the reader give this exercise a few moments of thought before continuing the chapter.) Proof of theorem. Let A = {At; i.o.}. Then A=(\{JA„ so A<={J()A<n k=ln^k k=l»>k To show that fi(A) = 1, it is enough to show that fi(Ac) = 0; and, to establish this fact it is enough to show (by subadditivity) that «(.q/:)-° Now, by independence, \n=k J n=k but fi(Ac„) = 1 — fi(An\ which in turn is less than or equal to e~^A^ because it is true in general that 1 — x < e~x. (Prove this inequality yourself!) Thus
48 Chapter 1 Measure Theory (12) \i I f] Acn ) < n e-^ = g-H-*"M"> \n=k J n=k But g-Il-*^") -» 0 as / -» oo, because £^=1 //(XJ = oo. Thus v(C)n>kAcn) = 0. D Example 13. Let Hn denote the event of a head at the nth toss of a Bernoulli sequence. The corresponding subset of / is An = {coeI;Rn(a>)=l} In exercise 11 the reader will show that these sets are independent. Furthermore, fi{An) = j for each n9 so £^04J = oo. Thus a head occurs infinitely often in a Bernoulli sequence with probability one. (This result can be proved much more trivially. What is the proof?) Example 14. Example 13 is an example of a finite pattern (the pattern H) occurring infinitely often in Bernoulli sequences. More generally we now show that any finite pattern occurs infinitely often in Bernoulli sequences with probability one. For simplicity of notation, consider the particular pattern HTTH. Proposition 15. The pattern HTTH occurs infinitely often in a Bernoulli sequence with probability one. Proof. Let En be the event that HTTH occurs starting at step n, and let Bn be the corresponding subset of /. Because the A„'s are independent and Bn = AnnAcn+1nAcn+2nAn+29 we have fi(B„) = (£)4 = ^; so ^=1 »(Bn) = 00. Unfortunately, B„ and Bn+l are not independent, so the second Borel-Cantelli lemma does not apply. However, the sets B„,Bn+4,Bn+B,... are independent; in particular, BuBS9B99...9B4k+l9...are independent and 00 fc=l so the second Borel-Cantelli lemma applies to give M{*4*+i;i.o.})=l But {B4k+1; i.o.} c {Bn;lo.} so 1 = /*({B4k+1; i.o.}) < fi({B„; i.o.}) < 1 Thus ix({Bn9 i.o.}) =1. □
§1.4 Measure Theoretic Modeling 49 Remark. This same proof works for any finite pattern of H's and T's—for example, Shakespeare's sonnets translated into Morse code, with the dots and dashes changed to H's and T's. Exercises for §1.4 1. We will say that an event E involving Bernoulli sequences is plausible if the subset BE of the unit interval corresponding to it is a Borel subset. Show that the following events are plausible: a. A gambler quadruples his initial stake. (Beware: he will not quadruple his initial stake if he gets wiped out beforehand.) b. In an infinite sequence of trials, a gambler breaks even an infinite number of times. c. In an infinite sequence of trials, arbitrarily long run lengths occur. d. In an infinite sequence of trials, H comes up "on the average" more often than T. (Incidentally, event d shows that "plausible" does not necessarily mean "probable.") 2. Show that, for random walks on the line, the following events are plausible: a. The origin is visited infinitely often. b. Every integer point on the real line is visited infinitely often. 3. Show that, for random walks in the plane, the following events are plausible: a. The origin is visited infinitely often. b. Every point (m, n) is visited infinitely often. 4. Let / be a function from the integers to the real numbers. Show that, for random walks on the line, the event 00 Z /(*») < °° i=l is plausible, nt being the position at time i. 5. Let S = ZS=i ± 2~l be the series obtained by flipping a coin to decide whether a plus sign or a minus sign goes into the ith place. Show that the event |S| < e is plausible, and compute its probability. {Hint: See §1.2, exercise 7.) 6. Let AT be a set, SF a a-field of subsets of AT, and \x a probability measureon SF. Let Al9A29A3,...bea. sequence of subsets of AT belonging to !F. a. Show that, if Ax 3 A2 3 A3 • • •, then \i=l / i- ftm n(Ai)
50 Chapter 1 Measure Theory b. Show that, if At ^ A2 e A3--, then \f=l / i-^oo 7. Let AT be a set and Al9A2,A3,...a. sequence of subsets of X. The set k=1 \n>k J is denoted as liminf A„ or as {An; a.a.} (abbreviations for limes infimum of A„ and "An, almost always," respectively). Show that {Acn; a.a.} is the complement of {An; i.o.}. If X is a probability space and the >ln's are measurable, conclude that the probability of {Acn; a.a.} is zero if and only if the probability of {An; i.o.} is one. 8. Let AT be a set, & a a-field of subsets of X, and \i a probability measure on <F. Show that, if Al9A2,A39... are in ^, then ^(liminf An) < liminffi(An) < limsup/x(i4„) < fi(\imsupAn) 9. a. Let X be a set and let A x and A2 be subsets of AT. Show that the smallest (7-field containing ^ and A2 consists of at most 16 sets. (Hint: Take unions of the four sets: Atr\ A2, A\ n A2,AX n >42, A\ n A2O b. Let Al9A29.-.9Ak be subsets of X. Let <Fk be the smallest <r-field containing the A^s. Show that !Fk has at most 22k members. c. Show that the upper bound in part b cannot be improved. (Hint: Let M be the Jc-element set {px,..., pk}, and let X = 2M be the set of subsets of M. Let A( be all subsets of M that contain the point p(. 10. Let X be a set, ^ a a-field of subsets of X, and /x a probability measure on #". Suppose that ^ ,...,>!„ are independent sets belonging to &. a. Show that Ac1,A2,...9A„2ire independent. b. Let A be any one of the sets Al n A2, A\ n Aj, ^ n ^2, A\ n 42. Show that A, 43, >44,..., A„ are independent. c. Let 3Fk be the smallest subfield of & containing A l9...9 Ak. Show that if Ae3Fk then A, Ak+l,..., An are independent. d. Let 3Fk be the smallest subfield of SF containing Al9...9Ak and ^,_k be the smallest subfield containing Ak+l,..., A„. Show that, if A e !Fk and A'e&n-k9 then 4 and >4' are independent. 11. a. Let A{ be the subset of the unit interval corresponding to the event "H at the ith trial" in a Bernoulli sequence. Show that the A?s are independent. b. Let Bt be the subset of the unit interval corresponding to the event "HTHattheith,* -I- lstandi + 2nd trial." Show that BuB4r,Bl9Bl0, ... are independent. 12. a. For the random walk with pauses, let A{ be the subset of the unit
§1.4 Measure Theoretic Modeling 51 interval corresponding to the event "a pause at time i." Show that the Ai's are independent. b. For the random walk in the plane, let A{ be the subset of the unit interval corresponding to the event "an eastward move at time i." Show that the A?s are independent. 13. Let N be a large positive integer. Prove that in a Bernoulli trial run lengths of length N occur infinitely often with probability one. 14. Prove that run lengths of arbitrary length occur infinitely often with probability one. (Hint: Consider the random pattern: THTHHTHHHT...TH...HT with the last term involving n H's. Let En be the event that this pattern occurs infinitely often and let E be the intersection of the events En. 15. For the random walk with pauses, prove that with probability one there are infinitely many pauses. Use the Borel-Cantelli lemma. (An alternative proof of this fact was suggested in §1.2, exercise 6.) 16. Let N be a large integer. Prove that the random walk on the line, starting at zero, passes either through the point N or the point — N with probability one. Conclude that it passes through N with probability at least ^. 17. Let Z2 be the integer points in the plane, and let Q be a finite subset of Z2 containing the origin. Prove that a random path starting at the origin hits dQ in a finite time with probability one. The path depicted hits dSl on the ninth step. 18. In proving the law of large numbers in §1.1, we used Chebyshev's inequality to prove > e j ) < 3>T V4 PI ]<*>; Sn(co)\
52 Chapter 1 Measure Theory Choose a sequence en -*> 0 such that ^=1 n 2en 4 is finite and let An be the set «f|>,} Use the first Borel-CanteUi lemma to prove that fi({An; i.o.}) = 0 and deduce from this the law of large numbers. 19. Let X = {x1,x2,...} be a countable set, Pl9P29... a sequence of non- negative numbers such that £P£ = 1, and [i the measure ri(A)= £ P( xteA Show that X cannot contain an infinite sequence of independent sets i41,i42,... such that, for all i, ^(Ai) = \. (Hint: Start by observing that every point xeX must lie in one of the four sets Ax r\A2, A\ nA2, At n Ac2, or A\ n Ac2. Thus the measure of the one-point set {x} is less than or equal to £. Notice, by the way, the moral of this exercise: A discrete measure theoretic model for the Bernoulli process does not exist. Just let A( be the subset of AT corresponding to the event "an H at the ith trial.") <co;
Chapter 2 Integration Now that we have the tools of measure theory, we are ready to discuss integration. The student of the Riemann integral is accustomed to considering only integrals over subsets of Rn. However, we will see that integrals can be defined whenever we have a triple (X, #", pi), where X is a set, SF is a a-field of subsets of X9 and \i is a measure defined on #\ Such a triple is called a measure space. Our basic example of a measure space is, of course, Lebesgue measure on the Borel sets of [0,1] or of the whole real line. The last section of Chapter 1 suggests that the theory of probability is rife with other examples. Notice that for the real line some sets have infinite measure— for instance, fiL(R) = oo. We will allow this to occur in general. (See the comments at the end of §1.3.) §2.1 Measurable Functions In the study of integration, it is convenient to allow functions to assume the values + oo and — oo. To make this notion concrete, we define the extended real number system to be the set R u { + oo} u { — oo}. The elements + oo and — oo in the extended reals have the special properties 1. — oo<a<+oo, aeR 2. a + ( + oo) = ±oo, aeR 3. a-(±oo)=+oo, aeR9a>0 4. — l*(±oo) = +oo Now let (AT, SF, fi) be a measure space. Let / be a function on X with values in the extended real numbers. 53
54 Chapter 2 Integratio Definition 1. The function / is measurable if, for all aeR, the set {xeX f{x) > a) is an element of #\ Measurable functions can be characterized in various ways. Proposition 2. The following are equivalent: 1. For all a e R, {x; f(x) >a}e& 2. For all aeR, {x; f(x) > a} e ^ 3. For all aeR, {x; f(x) <a}e^ 4. For all a e R, {x; f(x) <a}e^ Proof. lo4: The sets in items 1 and 4 are complementary and, because 3F is a c-field, we know that A e & o Ac e 3F. 2o3: Same as above. 1=>2: For all aeR {x;/(x)>a} = ni|X;/(x)>a-iJ By item 1 each set {x; /(x) > a — 1/n} e &. Because !F is a c-field, the countable intersection is in «^\ 2=>1: For aeR {x;/(x) > a} = Q |x;/(x) > a + U By item 2 each set {x; /(x) > a + 1/n} e^. Because 3? is a a-field, the countable union is in &. □ In keeping with our notion of extended real numbers, we define the extended Borel sets as the collection of subsets ofRu{ + oo}u{ — oo} having one of the following forms: A, i4u{ + oo}, Au{ — oo}, Av{ + co,—co} where A is a Borel set. One can easily see that the extended Borel sets form a o--field. Theorem 3. Conditions 1 through 4 of proposition 2 are equivalent to 5. For every extended Borel set B (1) {x;/(x)e£}e^ Proof It is obvious that 5 => 1, 2, 3, 4. We will show that 1, 2, 3,4 => 5. Let # be the collection of all subsets CofRu{ + oo}u{ — oo} with the property that
§2.1 Measurable Functions 55 (2) {x;f(x)eC}e& We need to show that the extended Borel sets are contained in c€. Note that, by items 1,2,3, and 4, the sets (a, -f oo], [a, + oo], [— oo, a), and [— oo, a] are members of #. Also notice that ^ is a <r-field. Indeed, if /4fe#, 1 < i < oo, then jx; f(x)e H a\ = H {*; fix) eAt}e^ so flJii A^V. Similarly, (J^ >i£e«. Now notice that the extended Borel sets are the smallest <r-field containing all the infinite intervals mentioned above. Thus the extended Borel sets must be contained in #. □ Example 4. Let X = R", and let & = Jt be the Lebesgue measurable sets. If/: R" -► R is continuous, then / is measurable. Indeed, if a e R, then {x e R"; /(x) > a} is open and thus is a Borel set. Example 5. Let X = S3, the sample space for the Bernoulli process. As usual, identify <% with /, the unit interval. Consider Rn(coi), the Rademacher functions. These are measurable because they are piecewise constant; namely, for any subset AofRu{ + oo}u{ — oo}, we have that {coel; Rn(co) e A} is a finite union of intervals. Similarly, if we let Sn{co) = ]T£=1 Rk(co), we see that Sn is also piecewise constant and hence measurable. Notation. In §1.2 we used the term random variable for a function /: X -> R when (X9^,fj) is a probability space. The interesting functions on X are always measurable. For this reason we use the name random variable interchangeably with measurable function when discussing probabilistic notions. Example 6. Let T((o) be the number of times the random walk corresponding to a) returns to 0, and let ln(co) be the number of consecutive /f's beginning at the nth toss of co. As an exercise, show that T and /„ are random variables. (See exercise 4.) Remark. The variables /„ and T are pathologically discontinuous. The possibility of integrating such functions, that is, computing their expectation values, vindicates the work we are about to put into the theory of integration. For example, we will show in Chapter 3 that T = + oo with probability one; that is, with probability one a random walk returns to 0 infinitely often. We continue now with the study of the properties of measurable functions. Let (X, &, fj) be a measure space.
56 Chapter 2 Integration Theorem 7. Let / and g be measurable functions on X. Let max(/,0) be the function on X defined by max(/,0)(x) = max[/(x),^(x)]. Then max(/,^) is measurable. Similarly, if min( f,g)(x) = min[/(x),^(x)], then min(/,0) is measurable. Proof, {x; max(/,^)(x) > a} = {x; f(x) > a} u {x; #(x) > a} and is thus an element of &. Similarly, {x; min(/,^)(x) > a) = {x; fix) > a) n {x; #(x) > a} is also in^. □ Corollary 8. Let / be a measurable function on X. Let (fix) if f(x) > 0 (-fix) if/(x) < 0 /+(X) = 10 if/(x)<0 aDd f-{x) = { 0 if/(x)>0 Then /+(x) and /_(x) are measurable. Proof. f+ix) = max(/, 0) and /_(x) = max(-/, 0). D Notice that fix) = /+(x) — /-(x), so we have just shown that every measurable function can be written as the difference of two nonnegative measurable functions. (See the figure below.) X± /X We now consider how measurable functions behave under limits. First, if fi> fi> • • • are functions on X, define functions sup f and inf f on X by (3) sup/w = sup{y;.(x); 1 ^ / < 00} i>l inf fi{x) = influx); 1 < i < 00} «>i Theorem 9. ISfufi* -. - is a sequence of measurable functions, then sup£>0/;. and infj>0yi are also measurable. Proof ForaeR, {x; supy;(x) >a}=\J {x; /f(x) >a}e&
§2.1 Measurable Functions 57 Similarly, {x; infflx) >a}=(\ {x; /,(x) >a}e& □ i=l Now, if/l5/2,... is a sequence of functions, let gk = s\xpn>kfn. Notice that gx > 02 ^ 03 ^ ' * * so that lim gk(x) = inf 0k(x) fc->oo fc>0 Recall that lim sup fi = inf gk k>0 is called the "lim sup (limes supremum) of the sequencefl9f2,...." Similarly, lim inf f( = suphk where fck = inf/n fc>0 n>k Theorem 10, If fl9f2,... is a sequence of measurable functions on X, then lim sup/w and lim inf fn are measurable. Proof. From the previous theorem, gk = supn^k/n and hk = infw>k/w are measurable. Applying theorem 9 once again, we have that lim sup /„ = inf gk and lim inf fn = sup hk fc>0 k>0 are also measurable. □ Corollary 11. Let /i,/2»--- ^e a sequence of measurable functions on X. Suppose that the /„'s converge pointwise to a function /; then / is also measurable. Proof f = lim sup /„ = lim inf fn □ Thus we see that we can generate new examples of measurable functions by taking pointwise limits. Another way of getting new measurable functions is by composition. Theorem 12. Let /: X -► R be a measurable function. Let g be a continuous function on R. Then g°f is measurable. Proof Note that we are not allowing / to take the values of + oo or — oo so that g °/is defined. Now, for all aeR, let Oa = {teR; g(t) > a}. Then {xeX; g°f(x) > a} = {xeX; fix) e &a}. Because g is continuous, (9a is a Borel set. So {x e X; f(x) e &a} e ^", because / is measurable. □
58 Chapter 2 Integration Example 13. If/: X -» R is measurable, then eixf, e~f2,2> sin/, |/|, and so on are measurable. In a similar vein we have the following theorem. Theorem 14. Let fx, /2, ...,/„ be measurable functions on X with ft: X -► R, 1 < i < n. Let G: Rw -* R be continuous. Then G(/l5. ..,/„) is measurable. We leave the proof to the reader as an exercise. Example 15. Let ft: X -> R, i = 1, 2, be measurable functions. Since x + j; and xy are continuous functions from R2 to R, fx + /2 and /^ are measurable. Exercises for §2.1 1. Given a set X and subsets Al9 A2,..., let X+ = limsupylB = {An; i.o.} = f| (J A„ *=1 «>k and 00 A. = liminf^ = {An; a.a.} = (J f) An Let /+, /_ and fn be the characteristic functions of the set A+, A. and An, respectively. Prove that /+ = limsup/, and /_ = liminf/,. 2. a. Let X be a set, <F a <r-field of subsets of X, and /* a measure on !F. A map / = (/i ,..,/„) of AT into R" is said to be measurable if each of the coordinate functions,/., is measurable. Show that/is measurable if and only if f~l{B)e& for every Borel set B c Rw. b. Let g be a real-valued function on R". Then g is 5ore/ measurable if, for all numbers a, the set {xeRw; g(x) > a} is a Borel set. Prove that, if /: X -► Rw is measurable and g: Rw -► R is Borel measurable, then 0 o/: AT -► R is measurable. 3. a. Let /„: X -> R, n = 1,2,..., be a sequence of measurable functions. Show that the set of points, x e X, where the sequence {/„(x); n = 1,2, ...} converges, is measurable. b. From part a deduce that the set of points, coel, for which the "randomized harmonic series" «=i « converges, is a measurable subset of J.
§2.1 Measurable Functions 59 4. The examples below describe some random variables that occur naturally in the theory of Bernoulli trials. Show that these random variables are measurable functions on the unit interval /. (As usual we make the identification 88 = Bernoulli sequences = /.) a. R = ruin time = the time it takes a gambler with an initial stake of N dollars to be reduced to penury. b. l„ = the number of successive heads that appear beginning with the nth toss. c. T = the number of times a random walk returns to the origin. d. limsupSw/n, where Sn is the sum of the first n Rademacher functions. This random variable measures "violation of the law of large numbers." 5. Let T be the random variable in part c of exercise 4. Show that the set {coeI;T(co)< oo} is uncountable and dense in /. (We will show in Chapter 3 that this set is of measure zero.) 6. a. Let S be the random variable in part d of exercise 4. Show that, for every subinterval of the unit interval and for every real number a with — 1 < a < 1, S((o) = a uncountably often. (Hint: Suppose that co = .a1a2a3...is2i sequence for which S(co) = a. Let co' = ax bxa2a3b2a4asa6b3ana^a^a^b^,.. Show that S(co') = a no matter what the fy's are.) b. On the other hand, show that the set {weI; S(co) ^ 0} is of measure zero. 7. a. Let Q be a finite subset of Z2 containing (0,0) as an interior point. (See §1.2, exercise 8.) Let H be the time at which a random walk starting at the origin hits 5Q. (H is called the "hitting time.") Show that H, regarded as a function on the unit interval, is measurable and is finite except on a set of measure zero. (Hint: See §1.4, exercise 17.) b. Assume that the points (1,0) and (0,1) are also interior points. Show that the set {(dgI;H(co)= +oo} is uncountable. 8. Let /: R -> R be monotone increasing. Show that / is measurable. 9. A function /: R -> R is upper-semicontinuous at x if for every e > 0 there exists a 5 > 0 such that f(y) < f(x) + e when \x — y\ < 5. Show that if/ is upper-semicontinuous at all points of R, it is measurable. 10. Two functions / and g are said to be equal almost everywhere if / = g except on a set of measure zero. Show that if / = g almost everywhere and if f is measurable, g is also measurable.
60 Chapter 2 Integration 11. Let (X, ^, n) be a measure space with fx(X) < oo. A sequence of measurable functions fn:X->R, n= 1,2,..., is said to converge to zero in measure if for all e > 0 lim^{xeX;|/„(x)|>e} = 0 n-*co a. Prove that, if/, converges to zero pointwise except on a set of measure zero, then /„ converges to zero in measure. b. Show that the converse of part a is not true. (Hint: Let X = J, !F be the Borel subsets of /, and \i be Lebesgue measure. Let Ax = [0,£], A2 = [i H ^3 = [0,i], ^4 = B,fl, A5 = [|,J], ^ = [|, 1], ^ = [0,i], A8 = [i,|], and so on. Let fn be the function that is 1 on An and 0 elsewhere.) §2.2 The Lebesgue Integral Let (X9^9fi) be a measure space and s: X -> R be a measurable function. The function s is called a simple function if it takes on only a finite number of values. Example 1. Let £e«f and define :eE fl ifxe '""[O ifx^ M*)= IA :,^E Then l£(x) is called the characteristic function of E; it is clearly simple. In general, if s is a simple function taking on the values cx,..., cN9 let £.• = s-1^) = {xeX; s(x) = c£} {1 < i < AT} We then have that N s(x) = X CilEt(x) It is easy to define the integral of nonnegative simple functions. Definition 2. Let s: X -► R be a nonnegative simple function and let £ e3F. Let c!,..., cN be the distinct nonzero values of s and let Ex? = s~* (c£), 1 < i < AT. Define the integral of 5 over £ with respect to \x as the sum (1) /£(5)=fciM(£n£l.) Remark. This value may be + oo because fi(E n jE£) can be + oo.
§2.2 The Lebesgue Integral 61 Proposition 3. Let st: X -► R, i = 1, 2, be nonnegative simple functions and letEeJ^. 1. (linearity) a. IE{csx) = cl^Si) for c e R, c > 0 b. ^(S! + s2) = J^) + IE(s2) 2. (monotonicity) If sx < s2 then /^Sx) < IE{s2) Proof, la is clear. lb: Let cx,..., cm be the distinct values of Sj and dx,..., dn be the distinct values of s2. Let E£ = s^fo), 1 < i < m, and i? = sj1^), 1 <; < n. The E^'s form a mutually disjoint cover of AT and so do the F/s. Thus E{ n jF}, 1 < i < m and 1 <j <n, also form a mutually disjoint cover of X, and 5j + s2 has the constant value ct + dj on ££ n jF}. Hence IE(Sl + 52) = £(c£ + dy)/i(£i n .F) n £) = £cil/«<*i n^n £) + E^ZMBinFjO £) * i j = X c,./i(£,. n £) + Y,dj»(Fjn E) i J = h(si) + IE(s2) 2: s2 — $i is a nonnegative simple function, and s2 = sx + s2 — sl9 so J(s2) = I(SX) + 7(52 - St). Q We now extend our notion of integration to nonnegative measurable functions by approximation with simple functions. Definition 4. Let / be a nonnegative measurable function from X into the (nonnegative) extended real numbers and let Ee&. Then the integral of/ on E with respect to \x is defined by (2) I fdfi = sup{/£(s);0 < s </, s simple} The following proposition shows that this definition of the integral is consistent with our previous definition when / is a simple function. Proposition 5. IE(s) = J£ s d\i if s is a nonnegative simple function. Proof. Clearly IE(s) <$Esdfx because s is a simple function with s < s. To show equality let 5' be any simple function with 0 < s' < s. By monotonicity J£(s') < J£(s). Hence sup {J£(s'); 0 < s' < s, s' simple} < IE(s). D At this point the reader is probably asking why the integral of nonnegative measurable functions should be approximated by the integrals of simple
62 Chapter 2 Integration functions. We justify this by showing that nonnegative measurable functions can themselves be approximated by nonnegative simple functions. Theorem 6. Let / be a nonnegative measurable function on X with values in the (nonnegative) extended real numbers. There exists a sequence of non- negative simple functions 0<5l <s2 <••</ such that s( -> /pointwise. Moreover, if/is bounded, then st -» /uniformly. Proof. We begin by defining sn. Consider the interval [0, n) on R. Divide this interval into nln subintervals of length 1/2"; namely, let (3) J, = IteR;1-^- < t < jV 1 < i < n2n Let Et = rl{fi) and Fn = /_1([n, + oo]). Together the £/s with Fn form a mutually disjoint cover of X (n is fixed). Define (4) sn(x) = f(~-yEi{x) + nlFn(x) Notice that on Et we have i-1 . i . i-\ — <f<y; and sn = -^- Thus, sn(x) < f(x) for xeEhi= 1,..., nln Similarly, on Fn n< f and sn = n so s„(x)</(x) forx6Fw Hence sn < f on all of X Notice also that sn < sn+1. Indeed, let / be one of the intervals Notice that / = /' u I" where I 2" '2"/'
§2.2 The Lebesgue Integral 63 _ p - 2 2i - l\ _ [2t-l 2/ ^ ~ 2n+1 ' 2n+1 y a _ 2n+1 92n+1) Let £ = /-l(/), E' = f~\I'\ and £" = /_1(/"). Then SnM = ^r forxeE whereas sm+i(x) = -^r forxeE' and 5B+1(x) = -^T1- forxeE" Then, because E = E'uE" and ^—^ < ^-J- we have shown that sn(x) <sn+1 (x) for all x e E This argument is clearly independent of which It we began with. It also works with minor changes on [n, + oo], so sn(x) < sn+ j (x) for all x e X We now show that sn-*f pointwise. Two separate cases are involved. Case 1. f(x) = H-oo In this case xeFn for all n, so sn(x) = n for all n. That is, sn(x) -> + oo. Case 2. /(x) is finite Say f(x) < n0. Then, for all n> n0, /(x) lies on one of the intervals /£; that is, But then s«W = -yr so |/(x) - sn(x)\ < — for n > n0, proving that sn(x) -> f(x). Finally, suppose that / is bounded, say f(x) < n0 for all xeX. The preceding argument shows that for n > n0 |/(x)-5„(x)|<i for all xe* That is, sH->f uniformly on X. □
64 Chapter 2 Integration Remark. Hidden in this proof is a crucial ingredient of the Lebesgue theory. In the Riemann theory one approximates a function / by simple functions by dividing the domain of the function into small intervals, as shown in part (a) of the figure. In the Lebesgue theory one approximates by dividing the range of the function into small intervals, as shown in part b of the figure. (b) Graph of/ (The dark lines represent the approximating function.) The second procedure has two conspicuous advantages. First, the x axis no longer has to be the real line; it can be any measure space X. Second, one gets a good approximation of / by simple functions without assuming / to be continuous. We now look at some properties of the integral. Proposition 7. Let £, Fe^ and let / and g be nonnegative measurable functions. 1. (monotonicity) If f <,g then
§2.2 The Lebesgue Integral 65 2. IfFcFthen 3. H>(F) = Othen fdfi< gdfi JE JE J/*** J/** l/d/i = ° Proof. 1. This is obvious because, if s is a simple function and s < f, then s < g. 2. We first verify this for f=\G where Gs^ Then J£(lG)d/i = /£(1G) = [i(E n G) and likewise JF(lG)d/x = j*(F n G). ButEnGcFnG, hence /x(F n G) < fi(F n G) Now, by the linearity of IE we know that J£ s rf/x < JF s <fyx when s is a simple function. But, by the definition oi\EfdpL and $Ffdft, the statement has to be true in general. 3. If / = s is a simple function, this assertion is clear; thus sup{/£(s); 0 < s < /} = 0 □ Remark. We will defer to §2.3 the proof that, for nonnegative measurable functions, I (/i+/2)^= I /i^+ I /2<fo Je Je Je Unfortunately, this fact, which looks as though it should be practically obvious, requires a somewhat delicate proof. We can also now prove Chebyshev's inequality in full generality. Theorem 8. (Chebyshev) Let / be a nonnegative measurable function. For Ee& and c> 0, let Ec = {xsE;f{x) > c}. Then (5) M£c)<- fdli »*\L
66 Chapter 2 Integration Proof. Because f >c on Ec, f dfi (by mono tonicity) I cdii<\ j JEe JEe But cdix = IEe{c) = cfi{Ec) JEe Thus cfi(Ec)< fdfi< fd\i □ JEC JE Corollary 9. If / is a nonnegative measurable function with jEfdfi< oo, then {x e E; f(x) = + oo} has measure zero. Notation. If a property holds on a set E e 8F except for a subset of measure zero, we say that the property holds almost everywhere on E (abbreviated as a.e.). Thus corollary 9 can be restated as I, (6) fdfi < oo =>/(x) < oo a.e. on E Proof. Let An = {xe£; f(x) > n}9 and let A = {xeE; f(x) = + oo}. By Chebyshev ii(An)<z\fdli .sif But A c An for all n, so ii(A) < fi{A„) < - \fdfi for all n njE Because \Ef d\i < oo we must have fi(A) = 0. □ Corollary 10. Let / be a nonnegative measurable function and let EslF. If lEfdfi = 0, then / = 0 a.e. on E. Proof. Let /I = {xgE; f(x) ^ 0}, and let An = {xe£; /(x) > 1/n}. Because y4 = (J"=i An it is enough to show that fi(A„) = 0. By Chebyshev ■*"J, Another property we can now prove is that the integral behaves nicely on disjoint unions of sets.
§2.2 The Lebesgue Integral 67 Theorem 11. Let / be a nonnegative measurable function on X. Let Al9A2, ... be a sequence of pairwise disjoint members of #\ Let A = \JfLx Ar Then f/«*#« = £ f - (7) /<*/•=! /<fc J A i-1 Ji4, /V00/ First we prove the theorem in the case that f(x) = lE(x) for some Ee#\ In this case 1Edfi = ^(y4 n £) and \Ed\i = fi(A( n £) By countable additivity of ^, we know that That is, r _?° r f Ma* = Z f : •M *=1 .M, By the linearity of IA and J^., the theorem is now automatically true for all simple functions. We prove the general case with / being an arbitrary non- negative measurable function as follows. For e > 0 pick a simple function s < f with [■ fdfj. < IA(s) + e Because the theorem is true for simple functions, we have that «-l 1=1 J A, [ fdn<± [ , J A « = 1 J At so fdfj. < X | fdfi + s JA i= Because £ is arbitrary, we must have (8) f fdfi < £ I fdfi \fdn<t \ . J A i=l jAt To prove the opposite inequality, first consider two disjoint sets Al9A2e SF. Let sx and s2 be simple functions with 0 < sf < / and Sidfi> , JAi jAt (9) | s(dfi> | fdii-1 for/=l,2
68 Chapter 2 Integration Let s = max(5x, s2); then 0 < s < f, and s is a simple function. Also sx < s and s2 < s, so we can replace s( by s in inequality 9; namely, Sdfi>\ JA{ JA{ /d\i — - fori =1,2 Adding these inequalities gives or Jsdfi+ sdfi> fdfi+ fdfi — i A\ JA2 JAi JA2 sdii>\ fdii+ I fdii-i JA JAx JA2 because the theorem is known to be true for simple functions. Now iAfdfi>jAsdfi9 so I fdfi> \ fdti+\ fdp-i JA JAV JA2 Because e is arbitrary, f fdti> \ fdgi+ f fdii JA JAi JA2 Thus we have the inequality for A1 and A2. An induction argument gives (10) f fdn>t \ fdfi JAlvA2yj-~vAn i=l JAi Returning to the general situation—that is, A = \jf=i At—we have [fd»>[ fdfL J A JAiKJA2KJ'~KJAn because Ax u A2 u • • • u An c A. Hence, by inequality 10, fd\i for all n [ fdn>t I , JA «=1 JAi That is, fdfi [ fdn>t [ - J A *=1 JAi Application. Theorem 11 tells us that we can use the integral to define new measures. For example, we define Gaussian measure, \iG, on the measurable
§2.2 The Lebesgue Integral 69 subsets of R by (11) /xG(A) = -Lf e~x2*diiL x/27r ja Theorem 11 says that this is countably additive and so is indeed a measure. In fact, \iG is a probability measure because 'in Jr ldfiL= 1 The theorem also has the following corollary. Corollary 12. Suppose / and g are nonnegative measurable functions and Ee#\ Then, if / = g a.e. on E, JE JE gdfi Proof. Let Ax = {x eE; f(x) = g(x)} and A2 = {xeE; f(x) / g(x)}. Clearly Ax and A2 are disjoint. Moreover, by assumption fi(A2) = 0, so JA2 JA2 gdfi = 0 Also $Al f dfi = JXi gd\x because / = g on Ax. Thus fdfi= fdfi+ fdfi= gdfi+ gd\i= gd\i □ JE .Mi Ji42 .Mi JA2 JE Exercises for §2.2 1. Let / be the unit interval. Show that 1 xdfiL = - i 4 using only properties of the Lebesgue integral discussed in this section^ (Hint: Approximate x from above and from below by simple functions.) 2. Let J be the interval 1 < x < oo. Show that IO- + oo using only properties of the Lebesgue integral discussed in this section.
70 Chapter 2 Integration 3. Let (X, &9 fj) be a probability space. Let /: X -+ [0, + oo) be a random variable (that is, measurable function). The integral E = \xfdfi is called the expectation value of / (or "most likely value" of f) and the integral K = £(/-JE)2** is called the variance off Show that, if the variance off is small,/deviates from its expectation value with very small probability. Explicitly, show that the probability that / deviates by s from E—that is, fi({xeX;\f(x)-E\>e}) is less than or equal to s~2V. 4. Let Hn be the number of heads occurring in the first n trials of a Bernoulli sequence. Compute its expectation value and variance. 5. Consider the "random" series < 1 1 1 ~2~4~8~ with the assignment of a plus or minus in the nth term being decided by the toss of a coin. Compute its expectation value and variance. 6. Let (X, #*, fi) be a measure space and/a nonnegative measurable function. Forallfle(0,oo),let Q>(a) = fi{{xeX;f(x)>a}) Suppose Jx/k<fy* < 00, k > 0. Show that there exists a constant C > 0 such that 0(a) < Ca'k That is, show that $ goes to zero at least as fast as a~k as a -> +00. 7. Let J be a finite subinterval of the real line and f:J-+R a simple function taking on values cx,..., cn. The function/is called a step function if y "~1(c£) is a finite union of intervals for each i. Given a simple function s: J -> R and a positive number e, show that there exists a step function/ such that (*) j\s-f\diJiL<e
§2.2 The Lebesgue Integral 71 (Hint: Show that, if A is a measurable subset of J, there exists a finite union of intervals B such that d(A, B) = fi(S(A, B)) < e. Now prove the inequality (*) for s = 1A. Proceed.) 8. Let X be a countable set; that is, X = {■X1,x2,x3,...) and let Pl9P2,P3,... be a sequence of nonnegative numbers such that £Pf= 1. For/I c* let x{eA We saw in §1.4 that \i is a probability measure on the cr-field of all subsets of X. Prove that every function /:AT->Ru{±oo} is measurable, and prove that, for / nonnegative, I X 9. Let /: R -> [0, oo) be measurable. Given aeR, let fa(x) = f(x — a). Show that fa is measurable and that {/.<*-/. (Hint: See §1.3, exercise 14.) 10. Let (X, J*\ //) be a measure space and A and £ be measurable subsets of X. Show that, if fx(S{A, B)) = 0, then, for every nonnegative measurable function /, 1/*-1. B 11. Let (X, #",/*) be a measure space and / and g positive measurable functions. Show that, if g is simple, then, for all £e^, \ (f + Q)dfx= I fdp+ I gdfi jE JE JE 12. Let (X,^^) be a measure space with fi(X) < oo. Let / be a bounded nonnegative measurable function and let {sn} be the sequence of simple functions constructed in theorem 6. Show that, for Ee^, sndfi^> JE JE sndfi^> fdfx (Hint: The sequence sn converges uniformly to /. Moreover,
72 Chapter 2 Integration \(f-sjdn+ I sKdn= I fdfi JE JE JE by exercise 11.) §2.3 Further Properties of the Integral; Convergence Theorems The reader has probably noticed that the definition of the integral does not make evaluation of integrals easy. The set of simple functions s with 0 < s < f is a formidably large set. In this section we will develop some effective techniques for computing integrals and for manipulating integrals and limits. The three key results are the monotone convergence theorem, Fatou's lemma, and the Lebesgue dominated convergence theorem. Let fx, f2,... be a sequence of measurable functions with 0</i</2<- - Note that / = lim^^ f( exists and is measurable. Theorem 1. (Monotone convergence theorem) Let / and fh i= 1,2,3,..., be as above. Then (1) I /4i = lim| Jfrfp forEe^ JE i~*oo JE To prove this theorem we need the following lemma. Lemma 2. Let / be a nonnegative measurable function on X and let El9E2, E3,... be a sequence of sets in & with Ex c E2 a E3 c • • •. Set E = (Jji1 Et; then f fdfi = Km( JE i-*x>jEi fdfi Proof. Let At = £t A2 = E2 - Et A3 = E3- E2 and so on. Then the ^4,'s are pairwise disjoint and 0 At = E and Q At = E„
§2.3 Further Properties of the Integral; Convergence Theorems 73 So by countable additivity fdn \fdn=t \ JE i=l JA{ >=i Ja, = lim X | fdn n-*oo £=1 , = lim fdfi We now prove the monotone convergence theorem. We have 0<fi<f2<-<f=limfn By monotonicity \ fidfi< I f2dfi<-'< I fdix JE JE JE so lim^oo J£/nd^ exists and must be less than or equal to §Efdti. Let a = lim,,.^ J£/nrf^. We must establish that (2) a > [fdfi Let s be any simple function, with 0 < s < /, and let c e R, with 0 < c < 1. Let En = {xeE;fn(x)>cs(x)} Notice that Ex <= E2 <= • * * because ft < f2 < * * *. Also notice that U"=1 En = E. Indeed, if x e 2s with s(x) = 0, then xe£„ for all n, and, if x6E with s(x) ^ 0, then because c < 1 fix) > s(x) > cs(x) So, for some n, fn(x) > cs(x) because that is, xeEn. Taking integrals, we get a=lim fnd\i> fnd\i> fnd\i> csd\i n^cojE JE JE„ JEn because fn> cs on E„. We apply the lemma to E = \J™=X En to get
74 Chapter 2 Integration a > lim cs(x)dfi= cs(x)dfx = c s{x)dfi n-*oojEn JE JE Because this is true for all c with 0 < c < 1, it must also be true that a > sdfi By now taking the supremum over all simple s with 0 < s < f, we get inequality 2. □ Remark. Thanks to the monotone convergence theorem, we can apply theorem 6 of §2.2 to the evaluation of integrals. If / is a nonnegative measurable function and sn is the sequence of simple functions constructed in theorem 6 of §2.2, then (3) j\**-> [ fdn JE JE We will use this formula to clear up some matters that we left dangling in §2.2. Theorem 3. Let / and g be nonnegative measurable functions and let c > 0 be a real number. For E e $F we have 1. 2. -I = { fdn+[ JE JE cfdfi = c fdfi E JE (f + g)dfi= I fdn+l gdfi E JE JE Proof. 1. The first part is clear because we know that, if s > 0 is a simple function, then IE{cs) = dE(s) and also that s < f if and only if cs < cf. 2. Again we know that, ifs1 and s2 are nonnegative simple functions, then h(si + s2) = /£(sx) + IE(s2) Now, choose a monotone sequence of simple functions 0 <s2 <s2 < *•• with sn-+ f pointwise. Similarly, choose simple functions 0 < s'i < s'2 < • • • with s'„ -► g pointwise.
§2.3 Further Properties of the Integral; Convergence Theorems 75 Note that 0 < s1 + s\ < s2 + y2 < and s„ + s'„^ f + g pointwise. By formula 3 (f + g)dfi= lim (sn + s'„)dfi n-*oo Je | (/ + «7)^=lim I JE n-*ooji = limf s„dfi+ s'ndn) n-oo \JE JE / = lim sndfi+ lim ^ d/i n-*aojE n-*aojE /<^ + ^ n Corollary 4. Let /i,/2,... be a sequence of nonnegative measurable functions. Then ^Jij ./j is a nonnegative measurable function and f fl/.)^= E f/„^ for£e^ j£\n=l / n=l j£ /V00/I Let F„ = ££=j fk. Then Fx < F2 < • • •, and the F„'s are all measurable. Now apply the monotone convergence theorem and theorem 3. □ So far we have integrated only nonnegative measurable functions. When extending our definition to more general measurable functions, we must beware of the problem of adding + oo to — oo. Let / be an arbitrary measurable function from X into the extended real numbers. Recall that /+ = max(/,0) and /_ = max(-/,0) are nonnegative measurable functions with / = /+-/- Lemma 5. The following two conditions are equivalent. ■•I' 1. |/| dfi < oo I. /+ d\i < oo and I /_ d\i < oo JE JE
76 Chapter 2 Integration Proof. Notice that \f\ = f+ + /_, so (4) [ \f\dfi= [ f+dii+ \ f.dti V J E J E J E Definition 6. A measurable function / is integrable over E if either of the equivalent conditions of lemma 5 holds. In this case we write fe «£?(/*,£) or / g S£(\x) on E. If E = X we write / e J^(/x). For / e S£(\i, E) we define (5) [ fdfi= \ f+dix- [ f.dti J E J E J E Theorem 7. (Linearity) Let /, g e £f(n, E) and c e R. Then a. cfe£f{fi,E) and c/d/i = c /d/x Je Je b. f + ge&(n,E) and (f + g)dfi = fdfi + gdfi J E J E J E Proof. a. If c > 0, then c{f+) = (c/)+ and (c/)_ = c(/_), so cfdfi= cf+dfi— cf-dfi JE JE JE = c f+dfi-c\ f.dfi JE JE = c fdfi A similar argument treats the case of c < 0. b. Let h = f + g. We first deal with the special case that none of /, g9 or h changes sign on £. The six subcases are: 1. />O,0>O,fc>Oon£ 2. /<O,0<O,fc<Oon£ 3. / > 0, g < 0, h > 0 on E 4. f>0,g<0,h<0onE 5. f <0,g>0,h>0onE 6. /<O,0>O,fc<Oon£ Case 1 has been dealt with previously. Case 2 can be reduced to case 1 because we can rewrite the formula as
§2.3 Further Properties of the Integral; Convergence Theorems 77 f (-*)«! = f (-/)*!+ \(-9)dil JE JE JE Case 3: rewriting h—f+g as f = h + (—g) reduces this to case 1. Similarly, cases 4, 5, and 6 can be reduced to case 1. Now to complete the proof we write E = E1u E2W"Kj E6 so that Et is the set for which case i holds, i= 1,2,...,6. Then, because jEfd^ = Zf=i Je,/^A*> anc* similarly for g and h, the theorem follows by applying it to each E( separately and summing. □ Corollary 8. (Monotonicity) Let /, g e S£(/x, £) with / < #. Then (6) [ fdli< [ JE JE gdfi Proof. Because / < g, g — f > 0, so 0< J (g-f)dfx= J gdii- \ fdfi JE JE JE by linearity. Corollary 9. life $£ (/x, £), then (7) \ fdix\< \ JE I JE \fW Proof. Because / < \f\, by monotonicity [ fdn* \\f\dp JE JE Similarly, -/ < |/|, so -f/<*/*< f |/| i|i Je Je That is, 1 fdp\<Z {\f\dft E | JE D We end this section with the other convergence theorems mentioned above.
78 Chapter 2 Integration Lemma 10. (Fatou's lemma) Let /i,/2,... be a sequence of nonnegative measurable functions, and let / = liminf/,. Then (8) I /^<liminf| f^ JE JE Proof. Let gk = infn^k /„ and ak = inf„^ JE fn d[i. Then gi<g2^'" and al^a2^-" By definition / = lim gk and lim inf f„dfx = lim ak k-*oo JE k->ao Notice that ak > \E9k dp because gk < fn for n > k. Hence, by the monotone convergence theorem, fdfx = lim gkdp < lim ak = liminf fkd[i D JE k->aojE k->ao JE Theorem 11. (Lebesgue dominated convergence theorem) Let /i,/2>/3»-- be a sequence of measurable functions, and let EefF. Assumptions: 1. limn_>00 f„(x) exists for all xeE. 2. There is a nonnegative measurable function ge^(fiyE) with g > \fn\ on £,n = 1,2,.... Conclusion: The function f(x) = limn_>00 fn(x) is integrable and lim fndfi = lim fndfi j£n->co n-+cojE Proof. By Fatou's lemma, liminf|/n|d/x \\f\dix= [ lir < lim inf \fn\dfi<\ gd\x Hence fe&(fi,E). Now we note that # -b /„ is nonnegative, so by Fatou's lemma I lim inf(g + Qd\x < liminf | (g + f„)d/i JE JE
§2.3 Further Properties of the Integral; Convergence Theorems 79 But liminf(0 + /„) = g + lim /„ = g + / and liminf {g + fn)dfj.= gd\x + liminf fnA\i JE JE JE so we have fd/i< liminf fnd\i JE JE Repeating this argument, with g + /„ replaced by g — /„, we get - fdfx < liminff - fndfi) = -lim sup ( fndfi) so fdfi > lim sup! f„dfi\ Combining these results limsupM fndfi\< /d/i<liminf( fndA But it is always true that lim inf < lim sup, thus we get equality: /<//* = liminfM /n^J = limsupM /nd/xJ = limM fndA D Corollary 12. Let fx ,/2,... be a sequence in S£(fi9 E) with \dfi < oo 2 f i/.i n=l JE Then !• Z^=i fn converges absolutely a.e. on E and is integrable on E. 2. j£Zr=i/„^ = Z„°°=ijE/„^. Frw/. Let g = £*=1 l/J. Corollary 4 tells us that \ Gdfi= £ \fn\dfi< oo JE «=1 JE so geJ?(fi,E). In particular, # is finite a.e. on E, so Z^=i/n converges absolutely a.e. on E. To prove part 2 let F„ = XZ=i /k- Then I^J ^ Z*=i l/*l ^
Chapter 2 Integration g so, by the dominated convergence theorem, Y*=ifn = ^mn-*oo^n *s *n j£f(A*,£)and { (tfn)dv= f limFn^ = lim f /*,<*/*= £ [ f.d/i D jE\n=l / j£n-»oo n-»oo J£ «=1 j£ Exercises for §2.3 1. Let (X, «^", /*) be a measure space and / be a bounded nonnegative measurable function. Show that This formula is Lebesgue's original definition of the Lebesgue integral. (Hint: See formula 3, page 74.) 2. Let/„: R -► R be 1/n times the characteristic function of the interval (0, n). Show that/, -► 0 uniformly but that ]"/„ d//L = 1. Why isn't this example a counterexample to the Lebesgue dominated convergence theorem? M Compare J lim inf /„ &\iL and lim inf \fn d[iL for the sequence /„ in exercise 2. Can the inequality in Fatou's lemma be replaced by an equality? For n = 1,3,5,... let fn be the characteristic function of the interval (0, \\ and for n = 2,4,6,... let /„ be the characteristic function of the interval (^, 1). Compare J lim inf/„ &\iL and lim inf \fn dfiL. Let (X,^,fi) be a measure space. A measurable function f:X->R is mean-square integrable if \f2dn < oo. Show that, if pi(X) < oo, every mean-square integrable function is integrable. (Hint: Consider separately the integral of |/| over the set where |/| > 1 and the integral over the set where |/| < 1.) Let X = /, ^ the Borel sets, and \i Lebesgue measure. Show that there exists an integrable function on X that is not mean-square integrable. Let X = R, 3F the Borel sets, and \i Lebesgue measure. Show that there exists a mean-square integrable function on X that is not integrable.
§2.3 Further Properties of the Integral; Convergence Theorems 81 8. Let (X9^,fi) be a measure space and /„, n = 1,2,..., a sequence of measurable functions. Then /„ is said to converge to zero in measure if, for all e > 0, fi({xeX;\fn(x)\>e})^0 as n -► oo. (Compare with exercise 11 of §2.1.) Show that, if j \fn\ &\x -► 0, then fn converges to zero in measure. Show that the converse is not true. (Hint: See exercise 2.) 9. Let /: R -► R be an integrable function. Show that, if J/*1 = ° for every subinterval J of the real line, then / = 0 a.e. 10. Let J be a finite subinterval of the real line and /: J -> R an integrable function. Show that for every e > 0 there exists a step function g such that \f-g\dn<e (Hint: See §2.2, exercise 7.) 11. Let (X,^,fi) be a measure space and / and g measurable functions. Show that, if/ is integrable and g is bounded and measurable, then fg is integrable. 12. Let (X^.fi) be a measure space and / a nonnegative measurable function. For Ae^let l*AA) = \ fdf* Show that fif is a measure on &. Moreover, show that, if g is a nonnegative measurable function, then for Ee^ 9dnf= gfdfi JE JE 13. a. In exercise 12, let A e SF and let / be the characteristic function of the set A. Describe the measure \if. b. Suppose, more generally, that / is a simple function; that is <*) f=iciiAl i=l Describe the measure \if. 14. a. In exercise 12, show that, if/is bounded, S£(E,\if) contains S£(E,\ji). Moreover, show that, for g e £f(E, /x)
82 Chapter 2 Integration 9dflf= gfdfJ, JE JE b. Show that S£(E,\xs) need not necessarily contain £f(E,fi) if / is not bounded. 15. a. Let (X,^,fj) be a measure space and f:X^>R a measurable function. For every Borel set ^^R, let vf(A) = tif-H*)) Show that this formula defines a measure vf on the Borel sets of R. Moreover, show that, if/us a probability measure, so is vf. b. If / is the function (*), describe the measure vf. 16. A function g: R -» R is Borel measurable if, for every Borel set A ^ R, g~x(A) is a Borel set. Let vf be the measure in exercise 15 and let g be a nonnegative Borel-measurable function on the real line. Show that (**) gdvf = \ g(f(x))dfi Jr jx (Hint: What does equation (**) say when g is a simple function?) §2.4 Lebesgue Integration versus Riemann Integration By the results of §2.3 we can now integrate complicated limits and sums of series. What about simple integrals? We will show that in the Lebesgue theory, just as in the Riemann theory, these integrals can be evaluated by the second fundamental theorem of calculus; that is, for the Riemann integral one has the following theorem. Theorem. Let g be a continuous function on an interval [a, fc] c R. Then g has an antiderivative G and (bgdx = G(b)-G(a) where $9dx denotes the Riemann integral of g on the interval [a,fc]. We will show that the same is true for the Lebesgue integral when our measure space (X, 2F, \x) is X = [a, b\ \i = \xL, and !F is the field of Lebesgue- measurable subsets of X. Rather than prove the fundamental theorem of calculus directly for the Lebesgue integral, we prove a much more general theorem, as follows.
§2.4 Lebesgue Integration versus Riemann Integration 83 Theorem 1. Let / be a bounded Riemann-integrable function on [a, ft] with Riemann integral \bafdx. Then fe «£?(/xL, [a, ft]) and (1) [bfdx=\ fdfiL Ja J[a,b] Before proving this theorem, let us recall how the Riemann integral is defined. Riemann Integral A partition P of [a, ft] is a finite, ordered sequence of points a = x0 < xx < x2 < * • * < xN = ft The maximum of the numbers xi+l — xh i = 0,..., N — 1 is denoted m(P) and is called the mesh width of the partition. Fix a partition P of [a, ft], and let / be a bounded function on [a, ft]. Let and Define Mi = sup{/(x); Xj-x < x < xt} mt = inf{/(*); xi^1 < x < xt} U(fP)=tMi(*i-Xi-i) and L(fP)=fjmi(xi-xi.1) i=l to be the upper and lower Riemann sums, respectively. Notice that L{fP) < U(fP). In fact, it is true that, if Px and P2 are any two partitions of [a, ft], then L{f Px) < U{f P2). To see this we introduce the notion of refinement. A partition F is called a refinement of P if the ordered sequence of points comprising F contains the points of P as well as some additional points; that is, F = P plus additional points. Clearly, if F is a refinement of P, then L(/, F) > L(f P) and U(f9 F) < U(f P). Now, if Px and P2 are two partitions of [a, ft], let P be a partition of [a, ft] that refines both Px and P2 simultaneously. Then (2) L(f, P,) < L(/, P) < U(f, P) < U(f, P2)
84 Chapter 2 Integration We define fdx = inf { [/(/, P); P is a partition of [a, fc] } (3) r t fdx = sup {L(/, P); P is a partition of [a, fc] } Note that by inequality 2 (4) [ fdx< [fdx Ja Ja Definition 2. The function / is Riemann integrable if JJ fdx = \baf dx. In this case we define (5) I fdx = I fdx = I fdx Ja Ja Ja To compare the Riemann integral with the Lebesgue integral, we will first show the following lemma. Lemma 3. There exists a sequence of partitions P1,P2,P3,... such that 1. Pk is a refinement of Pfc_ x. 2. The monotone sequence U(/,i\)£l/(/,P2)*"' converges to \bafdx. 3. The monotone sequence Uf,Pi)*L{f,PK' converges to \hafdx. 4. The mesh width m(Pk) -> 0 as k -> oo. Proof. We describe how to define Pk. Let P^ be a partition with i/(/,ifl<j'V*+£ and P£ a partition with
§2.4 Lebesgue Integration versus Riemann Integration 85 uf,pn*jl fdx-k We know that J^ and Fk' exist by the definition of \bafdx and \bafdx. Let Pk be a partition with m(Pk) < 1/fc refining Pk, Pk, and P*^ all at once. Then L(f,Pk)>L(f,Pu)>^fdx-1- U(f,Pk)<iV(f,Fk)<jbfdx + l- lim L(/,P,)= | Vdx and WmV(f,Pk)= [" fdx V k-+oo Ja k-*co Ja and So For a fixed partition P, define the simple functions f(a) at x = a m1 on x0 < x < xx Lp{x) = ^ m2 on xx < x < x2 and mN on Xx-i < x < xN Cf(a) at x = a Mx on x0 < x < xx t/P(x) = < M2 on xx < x < x2 MN on Xpf-i < x < xN Notice that L (x) < f(x) < UP(x) and I LP(x)dfxL = £ ^(x, - x,.!) = L(/,P) J[a,b] i=1 See figure on page 86.
Chapter 2 Integration Graph of/ C mn xQ = a x, *3 x=b f ™3 Graph of LP xQ = a xx x2 x3 xn-\ Xn — b Also, if F is a refinement of P, then £/P<(x) < UP(x) and LP,(x) > LP(x) Now choose a sequence of Pk's as in lemma 3 and let Lk = LPk and Uk = UPk Then L1 < L2 <-</<"< U2 < Ux Let L(x) = linv^ Ln(x) and ir(x) = lim,,.^ Un(x). By construction, L(x) < f(x) < U(x), and, by the dominated convergence theorem,
§2.4 Lebesgue Integration versus Riemann Integration 87 and lim LndfxL= LdfiL n->ao J[a,b] J[a,b) lim f UmdpL=[ UdfiL n->co J[a,b] J[a,b] Now we know that f LndfxL = L(f,Pn) and ! UndfiL= U(fPn) J[a,b] J[a,b] and we chose the P's so that lim L(fPn) = [fdx and lim !/(/,PJ = [ fdx Hence, we conclude LdnL=\ fdx and Ud/xL= \ fdx J[a,b] Ja_ J[a,b] J a To prove the theorem, we assume that / is Riemann integrable; that is, p r* fdx = fdx Ja Ja In other words, Ld\iL= UdfiL= fdx J[a,b] J[a,b] J a Thus i[a,b](U - V)d\xL = 0. But U > L so U - L > 0; hence, U - L = 0 a.e. Now L<f<U so f=U = L a.e. and / is Lebesgue integrable with f fdliL=[bfdx J[a,b] Ja Remark. The standard notation for the Riemann integral of a function / over an interval [a, b] is the notation we have been using—namely, (6) f(x)dx I
88 Chapter 2 Integration Unfortunately, the Lebesgue integral has no such standard notation. Heretofore we have been using the notation (7) f fdfiL J[a,b] or (8) jV*iL with / denoting the interval [a, ft]. Now that we've shown that the Lebesgue integral and the Riemann integral are identical for Riemann-integrable functions, we will be less methodical with our notation. We will sometimes use display 6 for Lebesgue integrals and will sometimes use displays 7 and 8 with the subscript deleted from \iL when it is clear from the context that \i is \iL. Exercises for §2.4 1. Compute the Lebesgue integral of the following functions: x2, x3, sinx, ex, xlogx. (You are encouraged to use the tools of elementary calculus in making these computations.) 2. In the proof of theorem 1, choose an x e [a, ft] that is not equal to any of the partition points of any of the Pk's. Show that / is continuous at x if and only if U(x) = L(x). 3. Conclude from exercise 2 that, if a bounded function / on the interval [a, ft] is Riemann integrable, then it is continuous almost everywhere. 4. a. Conversely, suppose that / is a bounded function that is continuous almost everywhere on the interval [a, ft]. Conclude from exercise 2 that ]imU(f9Pk) = limL(f9Pk) b. Deduce that, if / is bounded and continuous a.e. on [a, ft], then it is Riemann integrable. 5. (Improper integrals) a. Let / be a nonnegative measurable function on the interval J = (0,1]. Suppose that / is Riemann integrable on all of the intervals [a, 1], 0 < a < 1. Show that fdpL = lim f(x)dx J J a->0Ja with the integral on the right being the Riemann integral. b. Compute the Lebesgue integral over J of the function f(x) = 1/^/x.
§2.5 Fubini Theorem 89 6. a. Let / be a nonnegative measurable function on the interval J = [1, oo). Suppose that / is Riemann integrable on all of the intervals [1, a], a > 1. Show fdnL = lim f(x)dx b. Compute the Lebesgue integral over J of the function /(x) = 1/x2. 7. Let /(x) = (sin x)/x. Show that lim f(x)dx a-*°o J i exists. On the other hand, show that f(x) is not Lebesgue integrable over the interval [l,oo). 8. Let /: R -+ R be continuous and have a continuous first derivative /' that is positive everywhere. If \i is Lebesgue measure on R, show that the measure vf of §2.3, exercise 15, is of the form vf(A) = j" 0<fo where l/g(x) = /'(/"'(*)) §2.5 Fubini Theorem Section 2.4 and the convergence theorems of §2.3 allow us to compute a number of integrals on subsets of R. In order to compute integrals on subsets of R", we must justify the use of iterated integrals. This is the purpose of Fubini's theorem. The general situation is as follows: Let {X,Jt,fji) and (y,«yT,v) be two measure spaces. Let X x Y denote the space (1) *x r={(x,j/);xe*,j/ey} If A a X and Bc7, then A x B (= X x Y. On the other hand, you should notice that most subsets of X x Y are not of this form. Definition 1. A x B <= X x Yis called a product set \iAtJt and Be Jf. The smallest a-field in X x Y containing all product sets A x Bis denoted M x jV. In a moment we will define a measure on Jt x Jf. First we must describe sets in J( x Jf in terms of J( and JT, Definition 2. For E <= X x Y and xeX fixed, let Ex = {ye Y; (x,y)eE}. Then Ex is called the x-slice of E.
90 Chapter 2 Integration Notice that, if E and F are subsets of X x Y, then (EnF)x = ExnFx (2) (E-F)X = EX-FX K = (Ec)x and, if Ex, E2, E3,... are subsets of X x 7, then \n=l /* n=l Proposition 3. If E e ^# x Jf, then £x e Jf. Proof. Fix xeAT and let £fx be the collection of all sets E ^ X x Y with Ex e Jr. Note that 1. Sfx contains all product sets A x B. 2. Sfx is a (7-field. Thus ^. contains the smallest cr-field containing all product sets—namely, Jt xJf. □ Corollary4. Let f:Xx y-»Ru{±oo} be measurable with respect to Jt x Jf. For x0eX fixed, define fXo: 7-> R by /Xo(.y) = /(x0,.y). Then, for each x0eX, fXQ is a measurable function on Y Proof. Fix x0 e AT. If a e R we need to show that {yeY;fXo(y)<a}ejr Let £ = {(xj)el x Y;f(x,y) <a). EeJt x Jf because / is measurable and {yeY; fXo(y) < a} = EXoeJf by proposition 3. □ Remark. We could just as easily have studied y-slices as x-slices. The corresponding proposition and corollary are obviously true for j/-slices as well. Thus far we have made no assumptions about the measure spaces (X, Jt> fi) and (7, JT, v). We will now assume that both of these spaces are a-finite. (Recall that a measure space (X, Ji^ fi) is cr-finite if there exist X/ e ^#, i = 1,2,3,... with /i(Xt) < oo and 0 xi=x (See definition 32 in §1.3.)
§2.5 Fubini Theorem 91 We now use the measures \i and v to define a measure on Ji x Jf. First let's recall how one computes the areas of regions in the plane in elementary calculus. Consider the region in the figure below. For each point x on the interval [a, 6], we let <f>E(x) be the length of the interval Ex\ that is, <t>E[x) = y2 — y^. For reasonable-looking regions, such as the one we've drawn here, <f>E{x) is a continuous function of x, so the Riemann integral (3) r <f>E(x)dx is well defined. In elementary calculus one proves that display 3 gives the area of the region E (or uses the integral as the definition of area). To define a measure on the product cr-field, J( x Jf, we will mimic this process. Let £eJ x ./. For each xeX,ExeJf, so we can define a function </>E: X -► R by (4) fc(x) = v(Ex) We would like to define the measure of E to be the integral (5) J. <f>E(x)dfx To use this, we have to check that <f>E(x) is a measurable function on X. Proof of this fact requires a general argument about tx-fields called the n-k theorem. Definition 5. Let Z be a set and let Sf be a collection of subsets of Z. £f is called a X-system if the following three properties hold. 312. If £iC£2C£3c is an increasing sequence with each Ene£f, then X3. If £, Fe^ and £ c F, then F - Ee^.
92 Chapter 2 Integration Definition 6. A collection s# of subsets of Z is called a n-system if Til. ,4, Bes/^AnBes^. Theorem 7. (7r-A theorem) If $f is a A-system and ^ is a 7r-system with s0 <^¥, then the smallest <7-field containing j^, <r(«s/), is contained in £f. Proof.* First notice that, if a collection / of subsets of Z is both a A-system and a rc-system, then it must be a cr-field. Now let l(s/) be the smallest A-system containing s/. If we can show that l(srf) is also a 7r-system, then we will know it is a cr-field. Because <r(jtf) is the smallest a-field containing stf and l(s#) is the smallest A-system containing j/, we will have which proves the theorem. To see that l(s/) is a rc-system, we need to show that if A,Bel(sf) then AnBel(sf). For 4 c Z let ^ = {5cZ;in5Gi(i)} Notice that if ,4 e /(j</), then ^ is a A-system. Indeed, the three properties are satisfied as follows. JUL If A e /(j/), then Ze^, because /I n Z = /I. X2. If Ael(s#) and Ft <= F2 cr F3 c ••• are in ^, then (A n Fx) c (A n F2) c • • • is an increasing sequence in l(s/)9 so (J*=1 (A nEn) = A n (U*=1 F„) is also in J(j^); that is, (J~=1 £ne^. X3. If Ael(st) and F, Fe^ with E c F, then XnF, infGl(i), and (yln£)c(inF). So >4 n(F - F) = (>4 nF) - (A n£)e!(4 that is, F-Ee%. Furthermore, HA, Be srf, then AnBes/ a /(j</), soBe&A; that is, s/ a^A when y4ej^. Thus we see that J(j</) <= &A when y4Gj^, by the minimality of l{s/). In other words, we now know that, if A e st and B e l(st\ then An Be Thus, if B e l(sf) we have shown that st c <&B. Again using the minimality of Z(jaO, we get l(s/) c <§B for B e \(s4)\ that is, An Be l(s/) when A and 5 are elements of l(jtf). This is property n\ for /(.s/). D With the 7T-A theorem we can now prove the following proposition. ♦This proof is rather technical, and you might want to skip it when reading this material for the first time.
§2.5 Fubini Theorem 93 Proposition 8. If E e Ji x Jf and <j>E: X -* R is defined by ^£(x) = v(£x), then <j>E is measurable. Proof. First assume that v(7) < cxd. By the n-X theorem, the proposition will follow from the following three facts. a. Let Sf be the collection of sets E such that </>E is measurable. Then Sf is a A-system. b. All product sets, A x B, with Xeif and 56^T, are in ^. c. The collection j</ of product sets is a 7r-system. We prove these statements as follows. a. XI. <j>x x y(x) = v(y) for all x; hence, is clearly measurable. 12. Let Ex £ E2 e • • • be an increasing sequence with En e Sf\ that is, </>En(x) is measurable. Let £ = (j£=i £„. We need to show that <f>E(x) is measurable. Now +b(x) = v(EJ = v[ Q (£jj = lim v[(fiJJ = lim ^(x) Thus ^£ is the pointwise limit of measurable functions and so is measurable as well. X3. Let E,Fe$f with E cz F. Then F = E u (F - E) is a disjoint union, and Fx = Ex u (F - E)x is also disjoint. Thus v(Fx) = v(Ex) + v[(F - E) J; that is Because </>E and ^F are measurable, <f>F-E is too. b. Let £ = Ax5bea product set. Then _fJ5 ifx ~\<f> ifx Gi4 so (6) «x)-io ifxM That is, <f>E(x) = v(B)lA(x)9 which we know to be measurable. c. Let A x B and A' x 27 be two product sets. Then (A xB)n(A' x B') = {Ac\A')x(Br\B') so the collection of product sets is a 7r-system.
94 Chapter 2 Integration Now, what happens if we drop the assumption that v{Y) < oo? Because Y is c-finite we can find a sequence of subsets Yl9 Y2, Y39... of Y with Y{eJf, v{Y^ < oo, and 00 ^ = U ^ (disjoint union) i=l LetEe^T x Jf. Then 00 E = (J ££ (disjoint union) i=l where £< = E n (X x Yt) and (7) ^(x) = f ^,(x) i=i We proved above that ^£. is measurable, so by equation 7, ^£ is measurable. □ We can now use display 5 to define a measure on Jt x Jf. Definition 9. Let EeJtxJf and define (8) tt'(£) = j^£(x)d/i to be the product measure of E. Proposition 10. n' is a measure. JV00/I We must check that ri is countably additive. Let EuE2>... be a pairwise disjoint family of sets in Jl x ^T. Let £ = (J"=1 £„. We need to show that n'(E)=fjn'(En) »=i Now, Ex = (J^ (En)x is a disjoint union, so v(Ex)= Zv[(£„)J n=l 00 that is, <j>E(x) = £ <t>En{x) n=l We can now apply the monotone convergence theorem to get
§2.5 Fubini Theorem 95 JX n=ljx that is, n\E) = g *'(£„) □ n=l Notice that we could repeat this whole procedure using ^-slices instead of x-slices. Ostensibly this method gives a different measure n" on Jt x Jf. Our first version of Fubini's theorem is that these two methods yield the same measure. Theorem 11. (Fubini, version 1) n = n Proof. First assume that ix(X) and v(7) are finite. By the %-X theorem, it is enough to establish the following three assertions. a. Let Sf = {E e Jt x JT; n\E) = n"(E)}. Then Sf is a A-system. b. The product sets are in Sf. c. The product sets form a 7c-system. We proved assertion c above. Let us look at assertions a and b. a. U. n\X x Y) = ix(X)v(Y) =n"(X x Y)9 so X x Y e#>. \2. Let Ex s £2 c • • • be an increasing sequence with F„ e £f. We wish to show that E = (J*=1 En eSf. It is clear that lim^^ n'{En) = rc'(F) and \imn^Q07i"(En) = n"(E). But te'(F„) = n"(En\ because EneSf. Hence 7c,(£:) = 7c"(£:);thatis,£G^. 3l3- Let E.FeSf with F c F. Then F = E u (F - F) is a disjoint union, so n'(F) = 7c'(F - F) + te'(F) and n'\F) = te"(F - E) + tt"(F). Hence ^(F - E) = tc"(F - F), because tc'(F) = tc"(F) and n\E) = n"(E). b. Let E = y4 x B be a product set. Then ^£(x) = v(jB)1^(x), so n\E)=[<t,E(x)dn = v{B)n(A) Reversing the procedure gives tt"(F) = n(A)v(B) ThusFe^. Now, what happens if fi(X) and v(7) are not finite? Because X and 7 are <7-finite, we can find subsets XteJt,i = 1,2,3,..., and Yje Jf, j = 1,2,3,..., such that fi(Xi) and v(l^) are finite and
96 Chapter 2 Integration X = [j Xi (disjoint union) and Y = [j Yj (disjoint union) MEeJt x J^let Then 00 (9) E= (J £u (disjoint union) By what we proved above, so, by equation 9,7c'(JB) = n"(E). D Definition 12. The measure n' = %" is denoted p x v and is called the product measure on Jt x Jf. Example 13. Let X = Y = R and ^# = .yf = &l9 the Borel sets in R. Also let ju = v = pl9 Lebesgue measure on R. We claim that M x Jf = 08 2, the Borel sets in R2, and p x v = p29 Lebesgue measure on R2. Proof. Notice that, if J and J are intervals in R, then J x J is a product set so J x J g M x Jf. Now, $b2 is the smallest a-field containing sets of the form J x J, so $2 c Jl x tAr. We now show Jt x Jf c <%2\ Fix an interval / c R. Let ^ = {5cR;/x Be@2) Note the following. 1. ^7 contains all intervals JcR. 2. Jfj is a <7-field. Indeed, suppose Bl9B29... are elements of 0bl9 then /xfQlj/U Q(/xBn)e^2 \n=l / n=l Because Si1 is the smallest <7-field containing the intervals, items 1 and 2 imply that SlcSII\ that is, \{Be08l we have I x Be@2. Now fix a Borel set Be£x. Let ^B = {A c R; A x 5e^2}. Note the following.
§2.5 Fubini Theorem 97 1. @IB contains all intervals /cRby the above result. 2. @lB is a <7-field by an argument just like the one above. Because ^\ is the smallest a-field containing all of the intervals, we have ^i <= ®b\ that is, A x Be@2 for any A, Be^x. But ^ x ^ is the smallest t7-field containing all sets of the form A x B9 so we have shown that SSl x Stx c ^2 and thus J^ x ^ = ^2. We now show that Mi x Mi = ju2. Let KN = {(*,)OeR2;-Ar<x,);<N} It is enough to show that, for all N, Mi x Mi = M2 on Borel subsets of KN. (Why?) To establish this fact, let ^N be the Borel subsets B of KN for which Mi * Vi(B) = n2(B) c^v is a A-system: XI. MiXMi(^) = 4N2 = M2(^iv). XI. Let Bt c B2 cz • • • be an increasing sequence with each Bne£fN. Let B = U»=1Bn;then (Hi x Mi)(5) = I™ (Aii x A*i)(Bn) = lim n2{Btt) = ^2(5) n-*ao n-»oo X3. Let £ and F be elements of £fN with F c F. Then F = E u (F - £), so Mi x Mi(^ - E) = Mi x Mi(^) - Mi x Mi(£) = M2W - M2(£) = M2(^ " E) Next, let ja/jy = {I x J; I, J subintervals of [—N,N]}. «s/N is a 7c- system because (It x Jx) n (72 x J2) = (/x n J2) x (Jx n J2) Furthermore, jrfN <= ^ because M2(/ x J) = (length/) (lengthJ) whereas (mi x Mi)(J x J) = MiCO'MiO^) = (length/) (lengthJ) From the rc-A theorem we conclude that SfN contains the smallest a-field containing stN. Hence M2 = Mi x Mi on all Borel subsets of KN. □ Exercise. Write R3 = R2 x R1 and show that ^3 = @2 x ^1 andM3 = M2 x Mi- We return now to the general situation: (X9 Jty \x) and (Y, J^*, v) are <7-finite
98 Chapter 2 Integration measure spaces and (X x Y9 M x jV, fx x v) is the measure space constructed above. Theorem 14. (Fubini, version 2) Let/:X x 7-► R be a nonnegative measurable function. Then 1. a. For each xeX9 f(x9y) is a measurable function of y. b. For each yeY9 f(x9 y) is a measurable function of x. 2. a. jV /(x, j;) rfv is a measurable function of x. k I* /(x> JO <fy* *s a measurable function of y. 3. f T f /(*, y) dvl d/i = f [ f /(*, y) dfA dv = f /(x, y) d(/, * v) JxLJY J JrLJ* J J*xy Proof. Part 1 was proved at the beginning of the section. To prove parts 2 and 3, first note that they are true for f(x9y) = 1e(x,j;) if EeJt x Jf. Indeed I, l£(x,y)dv = v(Ex) = (f>E(x) Y which we have shown to be measurable, and similarly lE(x,y)dfi = i*(Ey) = \I/E(y) I )x is also measurable. The fact that m\E(x9y)dv \dfx= lE(x,y)dii \dv = \ lE(x9y)d(fi x v) Y J JyLJX J JXxY is version 1 of the Fubini theorem. Now, because parts 2 and 3 are true for characteristic functions, they must be true for simple functions by linearity. To prove the theorem in general, let f{x,y) be a general nonnegative measurable function and choose an increasing sequence of nonnegative simple functions 0 < sx < s2 < • - • with sn->f pointwise. Then, for x e X fixed, sn(x9y)dv^> f(x9y)dv by the monotone convergence theorem. Thus \yf(x9y)dv is a measurable function of x, because we know that Jy s„(x, y) dv is measurable. Similarly, \xf(x>y)dn is a measurable function of y.
§2.5 Fubini Theorem 99 Furthermore, part 3 holds for all sn; that is, msn(x,y)dv \dfi= sn(x9y)dn\dv= sn(x,y)d{n x v) r J JyLjx J Jxxy Applying the monotone convergence theorem to each of these terms separately yields m/(x,j,)dv~L = [ [ f f(x,y)dfi~\dv = [ f(x9y)d(n xv) D Theorem 15. (Fubini, version 3) Let/be integrable onXx Y. Then 1. a. For almost all x, /(x, y) is integrable as a function of y. b. For almost all y, f(x9 y) is integrable as a function of x. 2. a. j"y /(x, j;) dv is equal a.e. to an integrable function of x. b. ]x /(x, y) d\i is equal a.e. to an integrable function of y. 3. f [ [ f(x9y)dv~\dfz = f I" [ f(x,y)d»\dv = [ /(x, jOdfo x v). /V00/. Write / = /+—/_ where /+ and /_ are nonnegative. Because / is integrable with respect to /* x v, we know that jxxy/+(x,j;)d(ju x v) and ixxYf-(x>y)d(fi> x v) are finite. Version 2 of Fubini then gives f Mx,y)d(n x v) = [ [" f /+(x,^v~L JXxY Ja LJ< _J < 00 Thus jV/+(x,y)dv is finite a.e., and similarly $xf+(x9y)d/j. is finite a.e. This gives part 1. Parts 2 and 3 follow by applying version 2 of Fubini to/+ and/_ separately and then adding. □ A Final Remark. Instead of considering only products of two measure spaces, we could have considered products of three or more measure spaces. For instance, let (XhJKh jx^ i = 1,2,3, be measure spaces. One can define a product <7-field Jtx x M2 x Jtz by defining it as either {Jtx x Jli) x M-$ or M^ x (Jt2 x Jt3) or as the smallest <j-field containing the product sets iT.j X iT_2 X iT,g
100 Chapter 2 Integration with AteJii9 i = 1,2,3. It turns out that these three definitions give the same (7-field. (See exercise 3.) Moreover, we can define on Jtx x Jt2 x Jtz a product measure \xx x 1*2 x /*3 w^h ^e property that, on product sets, (10) (Mi x \i2 x fi3){A1 x A2 x A3) = MiO^M^W^a) That is, we can define this measure as Mi x (M2 x Ms) or (Mi x M2) x Ma You will be asked in exercise 3 to show that these definitions are the same. By simply using the Fubini theorem twice, we get analogous statements for this triple product. In particular, iff is integrable on A\ x X2 x X3 (with respect to/i! x ix2x ju3), then the various partial integrals make sense a.e. and (11) fd(ixx x fx2 x /x3)= ( /<*Mi W2 Ui3 JXlxX2xX3 J*3LJ*2\J*i / J What we have said about products of three measure spaces applies equally well to products of any finite number of measure spaces. (See exercise 4.) Here the Fubini theorem looks like (12) f fd(fx1x-"Xfzn)=! -\[ fdixX-d^ JXlX'"XXn JXn |_J*l J Exercises for §2.5 1. Let R be the region * = {(*, j>);-1<x<1,-1<j><1} in the plane. Compute the Lebesgue integrals xy2dfx2 (x2 + y2)dfj,2 and ye* JR JR JR d\i2 2. Let (X, M, fj) and (Y, Jf9 v) be a-finite measure spaces. Call a product set A x B finite if fi(A) < oo and v(B) < oo. Show that the product measure fx x v is the only measure satisfying (fi x v)(A x B) = fi(A)v(B) for all finite product sets A x B. (Hint: Use the n-X theorem.) 3. a. Let (Xi9 Mi9 fit), i= 1, 2, 3, be a-finite measure spaces. Show that (J(1 x M2) x Jt3 = Jt^ x (M2 x M3)
§2.5 Fubini Theorem 101 and that these equal the smallest a-field containing the product sets Ax x A2 x A39 where A^Jt^ i = 1, 2, 3. b. Show that (ix1 x fx2) x fx3 = fj,1 x (/z2 x fi3). c. Call a product set Ax x A2x A3 finite if ^(^f) < oo,i = 1,2,3. Show that the measure in part b is the only measure satisfying (j"i x /i2 x n3)(A1 x A2 x A3) = jM^JiM^W^a) for all finite product sets Ax x A2 x A3. 4. Generalize exercise 3 to n-fold products. 5. In part a of exercise 3, let Xx= X2 = X3 = R, JK1 = Ji2 = M3 = the Borel sets of R, and ixx = fx2 = fx3 = Lebesgue measure. Show that Jtx x Jt2 x J(3 is the Borel sets of R3 and ixx x \i2 x \i3 is Lebesgue measure. Can you prove an equivalent statement for R4? for Rw? 6. Let (X,Jt,fj) and (Y, Jf9v) be <7-finite measure spaces, and let E be in Jl x Jf. Show that E is of measure zero if and only if Ex is of measure zero for almost all xeX. 7. a. Let (X, Jl, \x) and (Y, Jf, v) be a-finite measure spaces, and let /: X -* R and g: Y -► R be measurable functions. Let h(x9 y) = f(x)g(y). Show that h is a measurable function on X x Y. b. If/ and g are integrable, show that h is integrable and that its integral is (L-XJ gdv 8. (The integral as "area under the curve.") Let {X9Jt9\x) be a a-finite measure space, and let /: X -* R be a nonnegative measurable function. Let Af = {{x,t)eX xR;0<t< f(x)} Show that Af is a measurable subset ofXxR (that is, belongs to M x (%x) and that the measure of Af with respect to the product measure A* x jULebisequaltoJx/d/i. 9. Show that property XI of a A-system can be replaced by the property 12'. If Al9 A2,... are disjoint subsets of Sf9 then (J^ i4£ is in y. 10. Let Ibea set and ^ a a-field of subsets of X. Let ixx and /z2 be finite measures on & with the property that \i\[X) = ^PO- Show that the collection of sets ' , A ^ , ^ , ,,, {Ae&;iix(A) = »2(A)} is a A-system. 11. Show that Lebesgue measure is the only measure on the Borel sets of the interval [0,1] with the property that, for all subintervals J, fi(J) = length of J. (Hint: Use exercise 10 and the n-X theorem.)
102 Chapter 2 Integration 12. a. Show that for the function xy f(*,y) = (x2 + y2)2 the iterated integrals r [ P /<* y) dAdx and j \ [ fl f(*> y) dx~\ dy exist and are equal. b. Show that / is not integrable over the square — 1 < x < 1 and — 1 < 13. Let X = Y = R and M = Jf = Borel sets. Let \i be Lebesgue measure, and let v be counting measure; that is, v(B) = number of elements in B Let EeM x Jf be the set E = {(x,y)eXx Y;x = y} Recall that </>E(x) = v(Ex) and \jfE{y) = /i(£y). Show that <f>E and ^£ are measurable but that j ^(x)rf/i^j <f>E(x)dfj.^ il/E(y)dv §2.6 Random Variables, Expectation Values, and Independence In the next two sections we will discuss some probabilistic applications of the material in §2.5. Let X be a set, 2F & a-field of subsets of X9 and [i a probability measure on X. A random variable f is, by definition, a measurable function f:X-*Rv{±oo}. For instance, let X = J = @ and let (1) / -K"+i;*) Interpreted probabilistically, / is the number of times H comes up in the first n stages of a Bernoulli sequence. It is a "random quantity" or "random variable" that can be measured each time we perform a sequence of Bernoulli trials. The expectation value of a random variable is its integral E(f) = [fdv
§2.6 Random Variables, Expectation Values, and Independence 103 Of course, the expectation value need not always be defined; that is, / need not always be integrable. However, all the random variables we will consider do have well-defined expectation values. For instance, if we integrate equation 1 over the unit interval we get J>-«/>-! representing the fact that n/2 is the number of heads "most likely" to turn up in a sequence of n Bernoulli trials. (This meaning of expectation value will become clearer in §2.7.) Given a random variable /,: X -► R and a Borel subset A ^ R let (2) f*f(A) = tilf-HAn The right-hand quantity is well defined because f~1(A)e^. (See §2.1.) We leave it for the reader to check that equation 2 defines a measure on the Borel subsets of R. (See §2.3, exercise 15.) We will call this measure the probability distribution associated with the random variable /. If, for two random variables / and g, [if = [ig9 we will say that / and g are identically distributed. The essential property for us of the measure fxf is the following. Theorem 1. Let ^ be a nonnegative Borel-measurable function on R. Then (3) f *(/)**- f ***/ Jx Jr Proof. First, suppose that ^ is the characteristic function of a Borel subset XgR. Then ^(/) is the characteristic function of f~1(A), so the left-hand side of equation 3 is ix\_f~l(A)~\ and the right-hand side is fif(A). By equation 2 these quantities are equal. Next observe that equation 3 holds for finite linear combinations of characteristic functions of sets—that is, for simple functions. Finally, by theorem 6 of §2.2 there exists an increasing sequence sn of non- negative simple functions with sn -► <j>. Then sn(f) -► </>(f); so equation 3 follows from the monotone convergence theorem. □ Corollary 2. Let ^ be a Borel-measurable function on R. Then <j> is integrable with respect to the measure \is if and only if <j>(f) is integrable with respect to \i. When such is the case, equation 3 holds. Proof. Let <j> = ^+ — ^_ and apply theorem 1 to ^+ and ^_ separately. □ Notice that if <j>(x) = x, then equation 3 becomes [ xdfif= \ fd/i = E(f) Jr J*
104 Chapter 2 Integration This shows that, if random variables f and g are identically distributed, they have the same expectation value. More generally, if / and g are identically distributed, then, for every Borel-measurable function <j> on R, <j>diif= L </>dtig so by equation 3 f Kf)dii- ( Jx Jx (4) | *(/)**- «*)«* For instance, taking ^(x) = x2 we get g2d\i //>*.-/. Given several random variables /i,...,/„, let F:X->Rn be the map F(x) = (/i (x),... ,/„(x)). If A is a Borel subset of Rn, set (5) nfl fn{A) = nlF-i(A)-] This formula defines a probability measure on the Borel subsets of Rw (check this!) called the joint probability distribution associated with /i,...,/„. The analogue of equation 3, (6) <f>(fi,--->fn)dn= I ^dfifi fn Jx Jr" holds for any nonnegative Borel-measurable function (j> on Rw and is proved in exactly the same way. A set of random variables fi9...9fn is said to be independent if, for any sequence of Borel subsets Ax, A2,..., An of R, the sets are independent as subsets of X. An infinite sequence of random variables fl9 f2,...is said to be independent if every finite subsequence is independent. A very simple criterion for independence in terms of the joint probability distributions of the/J's is the following. Theorem 3. The random variables fl9...,/, are independent if and only if the probability measure At/lt...f/n is equal to the product measure \iSx x fxf2 x •"X/i/n. Remark. The product is, of course, defined as in §2.5.
§2.6 Random Variables, Expectation Values, and Independence 105 Proof. To check that the measures agree, it is enough, by the n-X theorem, to check that they agree on sets of the form At x • • • x An. By definition and by independence the right-hand expression is equal to /*[/f H^i)] x • • • x li[fnl(An)] which equals fifl(Ai) x •. - x (An) which equals(/Aflx • • • x /X/b)(Ai X • • • X An). □ One consequence of theorem 3 is the identity (7) J/iX-x/.^ = £(/l)x-x£(/B) Indeed, if we apply equation 6 to the function faxl9..., x„) = xt • • • xn we get A x--x/.dfi= xix2--xndiifx /n Jx Jr» x1x2'-xndfxfl x ••- x^ Jr» =(Lxi^-)(1X2^)"(Lx"^) by Fubini's theorem. However, the ith term on the right is just E(fy. A remark about independence that will be useful below is the following. Let /i,...,/, be independent, and let fa,..., <j>n be Borel-measurable functions on R. Then (8) Mfl)>~.,<f>n(fn) are independent. In fact, let Al9...,AH be Borel subsets of R, and let A't = ^f1 (At). Then the A"s are also Borel subsets, and M(/i)rl(>4i) = /r1w) *-i,...,n By assumption, the sets on the right are independent; hence, so are the sets on the left. Example 4. If ft and f2 are independent, then /i3 and |/2| are independent. The notion of independence plays a central role in measure-theoretic models of probabilistic processes. For instance let's go back to the gambling process described at the end of §1.2. Recall that this process involves a cage filled with colored marbles. There are assumed to be k different colors, with N( marbles of each color i and JV = J]f=1 Nt marbles in all. The process consists of mixing the marbles, then drawing a marble out of the cage. If the color of the marble is i, the player receives a reward (or penalty) of rt dollars. The
106 Chapter 2 Integration marble is then replaced, the marbles are again thoroughly mixed, another marble is drawn, and the game continues. \ifn is the amount of the reward or penalty at the nth draw, then fn takes on the values of r!,...,rk with the probabilities of pl9...9pk, where pi = Ni/N. What is an adequate measure- theoretic description of this situation? We claim that the data needed to "model" this process are: (a) a set X9 a c-field 3F of subsets of X, and a probability measure /x on SF\ and (b) an infinite sequence of random variables /i>fiy - • • with the following properties: (9) The y|'s are independent and (10) /^)=Ep; rjeA for every Borel subset A of R. Indeed, the yj's give an identification of points xeX with infinite sequences of draws from the cage, Le.,fk(x) describes what happens in the sequence corresponding to x at the fcth draw. Property 9 just says that what happens at the nth draw is independent of what happens at any of the other draws. This is justified by the fact that the marbles are thoroughly mixed after each draw. If one sets A = {rm}, then, by equation 10, the probability that at stage n the reward or penalty incurred will be rm is just which is what we expect because of the number of marbles of color m in the cage. Notice that equation 10 implies that the //s are identically distributed. We will now show that a probabilistic model with all the above features does exist. In fact, we will show that we can even take for X the unit interval J, for & the Borel subsets of/, and for the probability measure on 3F ordinary Lebesgue measure. Theorem5. There exist bounded measurable functions fuf29... on I such that property 9 and equation 10 hold with fx = fxL = Lebesgue measure. Proof. Decompose the unit interval into k disjoint subintervals Ix,..., Ik such that Jj is of length ph and define /i by setting ft=rt on/, l = lf...f* . Next decompose each of the intervals It into k disjoint intervals Ilt w, m = 1,..., k, such that IUm is of length pfpw. (Because It is of length pt and J]pm = 1, such a choice of 7/tW's is clearly possible.) Define f2 by setting fi = rm on Ilttn
§2.6 Random Variables, Expectation Values, and Independence 107 Notice that, if At = {r,} and Am = {rm}, fi-i(Al)nf2-HAJ = Iltm so ^fi'i(Al)nf2-HAM=PlPm On the other hand, k fi'HAi) = // and f2x{AJ = (J IUm (disjoint union) k so M[/i_1Mi)] = Pi and m[/2_1WJ] = Z P/Pm = Pm From these computations we conclude that or, in other words, fC1(Al) and f2~1(AJ are independent. Now, suppose that Bt and 2*2 are arbitrary Borel subsets of R. Then B1=B'1v(\Ja) and B2 = B'2J (J Xm) with B[ and £2 containing none of the r/s. Then /J"1 (2*1) and f2~1(B2) are empty; so /i"1(51)= U /i_1(^) (disjoint union) r,eB, and /2"1(^2)= U /2_1(^m) (disjoint union) are independent. (Why?) Because Bx and 2?2 are arbitrary, ft and /2 are independent. We let the reader check that ft and f2 satisfy equation 10 and move on to the construction of/3. Decompose Ilm into k disjoint subintervals Ii,m.n>*=l>...9 K of length ViVmVn and define f3 by setting f3 = rn on Ilttnt„ One checks that fl9 f2, and f3 are independent in exactly the same way as above. It is also clear by now how to construct /4, fs, and so on. We leave details to the reader. □ Remark. Let fe = 2, rx = 1, r2 = — 1, px = i, and p2 = h Then the functions constructed above are exactly the Rademacher functions!
108 Chapter 2 Integration Stage 1 Stage 2 1 1 1 This figure indicates the first two stages in the construction of the/'s with the data k = 3,rl = \, r2 = 0, r3 = -1, and/>! = p2 =p3 =* £. Exercises for §2.6 1. Let v be a probability measure on the Borel sets of the real line. Then v is said to be supported on the Borel set A if v(Ac) = 0. Using theorem 1, show that, if A is a Borel set containing the image of/, then nf is supported on A. In particular, if / is a simple function taking on the values rl9...9rh, then fif is supported on {rt,..., rfc}. 2. Let X = / be the unit interval, and let \i be Lebesgue measure. Describe the measure \is for the function f(x) = x2. 3. If, for i = 1,2,/ = Ri is the i th Rademacher function, what is the measure fxfif Verify directly that ju/l>/2 = fxfl x ju/2. 4. (The unfair coin.) Using theorem 5, let fe = 2, rt = 1, r2 = —hPi = P> and p2 = 1 — p. Describe the first three of the functions /x, /2, /3,.... 5. In exercise 4, let Sn = ft H + /,. Compute the expectation value E of Sn and the variance V(Sn)=f(Sn-E)2dfiL 6. (The random walk with pauses.) Using theorem 5, let k = 3, rx = 1, r2 = 0, r3 = — 1, and px = p2 = p3 = ^. We have already drawn the graphs of/! and /2. Draw the graph of/3. Can you discern a pattern? 7. a. Let Rt be the ith Rademacher function, and let /*(*>= zQf)w Compute I etf*ix)dx (Hint: Use independence.)
§2.6 Random Variables, Expectation Values, and Independence 109 b. Using the formula 2x - 1 = lim fk(x) (proved in exercise 7 of §1.1) deduce Vieta's formula 8. a. Let ft, f2,... be as in theorem 5, and let Sn — f^ + • • ■ + /„. Prove that ets»Wdx = YJProb(Stt = r)etr b. Conclude from part a that (*) £P">MS„ = r)e» = (£ Pie»>Y (Hint: Write the integral on the left in part a as e*fie*f2 x ... x etfndx 1 I and use independence.) 9. Let /i, f2,/3,... be independent, identically distributed random variables taking on the value of 1 with probability p and the value of 0 with probability 1 — p, where 0 < p < 1. (That is, using theorem 5, take k = 2, ri = 1, r2 = 0, Pi = p, and p2 = 1 - p.) Let Sn = /x + • • • + /„. Show that (**) Prob(Sn = r) = (")p'(l-p)n-' if 0 < r < n and is zero otherwise. (Hint: Use exercise 8.) 10. Let {rx, r2,. •.} be a countable subset of R, and let pj, p2,... be a countable sequence of nonnegative numbers with Zp«-i For every subset 4 of R, let (t) v(A)=£p* i"| e A Show that there exists a sequence /i,/2,... of independent, identically distributed random variables on the unit interval such that fifl = fif2 = ... = v. 11. For i = 1,..., n let {X,-, ^£, /jj be a probability space, let M± x-xMn be the product of the Jtl^, and let \ix x • • • x \in be the product measure
110 Chapter 2 Integration on Jt± x ••• x Mn. For each i let ft be a measurable function on Xt. Consider /, as a function on Xl x • • x Xn by setting y;.(x1,...,xn) = /i(xl) Show that the /£'s, regarded as functions on Xx x ••• x Xn, are independent. §2.7 The Law of Large Numbers Let's now return to the question posed at the beginning of §2.6. What does the expectation value of a random variable really represent? Consider a simple probabilistic process (such as the toss of a coin) and a numerical quantity Q associated with the process. (For instance, for the toss of a coin let Q = 1 if an H occurs and Q = — 1 if a T occurs.) Now repeat the process and again measure the quantity Q; repeat it a third time and again measure Q, and so on ad infinitum. Let AVn(Q) be the average value of Q, averaged over the first n stages of this infinite sequence of experiments. Does AVn(Q) tend to a limit as n tends to infinity? The answer is yes, provided that, each time the experiment is repeated, the conditions under which it is performed are not biased by the results of the preceding trials. (For instance, if the experiment consists of drawing a marble from a cage, recording its color, and then replacing, it, the marbles must be thoroughly mixed each time.) We will show that, if these experimental requirements are met, then, not only does AVn(Q) tend to a limit as n -► oo, but in fact (1) lim AVn(Q) = E(Q) n-*oo is the expectation value of Q. (We mean, of course, that equation (1) holds with probability one.) To see this, let's first describe the experimental setup above in somewhat more precise terms. Let /„ be the measured value of the quantity Q at the nth stage of the sequence of experiments. Then, under the hypotheses above, the ^'s are independent, identically distributed random variables. The underlying space X on which they are defined is, technically speaking, the totality of "all infinite sequences of repetitions of the experiment." For instance, if the experiment consists of the toss of a coin, X is the set 0$ of all Bernoulli sequences as in §1.1. Actually it isn't terribly important to describe X this explicitly. What is important are the i.i.d. (independent, identically distributed) random variables fl9f2,... and their common probability distribution fxfl = fifl = • • •. For instance, for the experiment described in §2.6 (a colored marble drawn from a cage) we showed that X could be taken to be a very simple set: the unit interval. The important point was that on the
§2.7 The Law of Large Numbers 111 unit interval we could produce a sequence of independent random variables fiyfi*- • aH with the probability distribution of equation 10 in §2.6. The following result is what is traditionally called the law of large numbers. Fix a set X, a <7-field & of subsets of X, and a probability measure \i on «^\ Theorem 1. Let fx ,/2,... be a sequence of bounded random variables on X that are independent and identically distributed. Let E = E(f^) = E(f2) = • • • be the common expectation value of the /.'s. Let X0 be the set of points xeX for which (2) /t(*)+ -+/■(*) _,g ft as n -► oo. Then i*(X0) = 1. Remark. The assumption that the /£'s are bounded is not essential, but it simplifies some details of the proof. Proof. Replacing ft by f{ — E9 we can, without loss of generality, assume that E = 0. Let V = (\f*d»)2 and w = \f? dfi Because the /f's are identically distributed, these quantities are the same for all f s. The first step in the proof will be to establish the following inequality for all s > 0: /f y. |/i(x)+ "-+/.(*)! ^ /1\ ^ 3n(n-l)K + K^ ^reX'|—s—|>8jJ-—s?— (3) The left-hand side of inequality 3 is equal to //({xeX;(A + ---+/J4 >nV}) and, by Chebyshev's inequality, this is less than (^)!(/i I'8"/J so inequality 3 reduces to (4) [(/i + -"+ fnfdfi < 3n(n - \)V + nW If we multiply out the expression on the left, we get five sorts of terms— namely,
112 Chapter 2 Integration j flfjdn a^p J flMidix a^P^l J faffifyfsdn a^p^j^S J fiffidp The first integral is equal to W, and the second integral is equal to (Jj*d^(J$d^=V by equation 7 of §2.6. Similarly, the third integral is equal to (//2<fc)(//***)(/A**) and the fourth and fifth integrals are equal to (J fa d^j (j ffi dp\ (J /7dji) (J hdl^j and J fid/* J ffid/jL Because the expectation values are zero, these three terms are zero. Because there are exactly n integrals of the first type and 3n(n — 1) integrals of the second type (see §1.1), the sum of all these integrals is the right-hand side of inequality 4. Now choose a sequence of numbers st, e2» £3* • • • such that sn -> 0 and » 3n(n- l)V + nW L, 4-4 < °° (See lemma 6 in §1.1.) Let A, n = IxeX; /iM+ - + «*) >£n[ Then, by inequality 3, Yj?=i l*{AH) < 00. So, by the first Borel-Cantelli lemma, fi{An; i.o.) = 0. This result means that, if we exclude a set of measure zero from X9 then for x in the complement /i(*) + -+/.(x) <£„
§2.7 The Law of Large Numbers 113 for all but finitely many values of n. This clearly implies that /i(x)+ -+/.(*) -> 0 as n oo □ Exercises for §2.7 1. Show that, if the /„'s in theorem 1 are the Rademacher functions, then equation 2 is the strong law of large numbers formulated in §1.1. 2. Let (AT, «^, fi) be a probability space. Recall (§2.3, exercise 8) that a sequence of measurable functions fn9n= 1,2,... converges to zero in measure if for all e > 0 ti({xeX;\fn(x)\ >e})->0 as n-+co Show that, if {fn} converges pointwise to zero almost everywhere, then /„ converges to zero in measure. (Hint: Let An = {xeX; \fk(x)\ < eforfc > n}. Show that Al ^ A2 ^ A3 ^ • • • and that X — (J"=1 An is of measure zero.) 3. Deduce from exercise 2 and theorem 1 the weak law of large numbers. Show that, if/*i,/2>--- are a sequence of bounded independent, identically distributed random variables and if £ is their common expectation value, then for all e > 0 Prob /i + -+/. -£ >e)->0 as n- oo Show that, if/i,/2>... are as in exercise 3, then I/1 + -+/. Prob -E *-)'&> where Kis the common variance of the//s, i.e., V = ^(f — E)2 d\i. Use this to give another proof of the weak law of large numbers. Let /i,/2,.. be bounded, independent, identically distributed random variables with E = E(fx) = E(f2) = • • = 0. Let Sn = fx + • • • + /„. Show that if a > 0 then Sn(x)/na/2)+a -► 0 a.e. as n -> oo (Compare with exercise 17 in §1.1.) (Hint: Show that there exists a constant Ck such that \ S?dn<Cknk for every integer k > 0.)
114 Chapter 2 Integration 6. Let /i,/2» • • be independent, identically distributed random variables on the unit interval. Suppose that the common probability distribution of the fiS is given by equation 9 of §2.6, with k = 2, rx = 1, r2 = 0, px = p, and p2 = 1 — p (p being any number between zero and one). Let Sn = f\ + " • • + fn- Show that, if <j> is any bounded measurable function on the interval [0,1], <•> «Kr)]-MD*-K)(i)^-^ (Hint: See section 2.6, exercise 9.) Remark. We will denote the right-hand side of equation (*) by Bn{<f>9p). Notice that it is a polynomial of degree n in p. We will call it the nth Bernstein polynomial associated with </>. 7. Show that for n very large Sn(x)/n is very close to p for most values of x. Explicitly show that for all 5 > 0 (**) <{■ jiNxeflU]; S„(x) -p >*})*(£)'<,-'> (Hint; See exercise 4.) 8. Let <j> be a continuous function on the interval [0,1]. Show that (t) ■['©]■ ► ^(p) as n -► oo (Hint: Given 8 > 0 choose 5 so that |^(s) — <l>(t)\ < e when 0 < s, t < 1 and |s - t| < 5. Let /j be the subset of 0 < x < 1 on which |(Sn(x)/n) - p\ < 5, and let I2 be the complementary set. Show that im- m dx < e and estimate im- m dx using inequality (**) and the fact that <j> is bounded.) 9. Show that the convergence in display (t) is uniform in p. By equation (*) conclude that, as functions of p, the Bernstein polynomials Bn(<l>) converge uniformly to <f> on the interval [0,1]. (The result we have asked you to prove is a constructive form of the Weierstrass approximation theorem: Given a continuous function ^ on the interval [0,1], there exists a sequence of polynomials Bn converging uniformly to ^ as n -► oo.)
§2.8 The Discrete Dirichlet Problem 115 §2.8 The Discrete Dirichlet Problem Let 6 be an open set in R2. A twice-differentiable function /: 0 -* R is called harmonic on 0 if d2f d2f „ Now let Q be a compact subset of R2 with a continuous boundary dCl. Suppose that g : dCl -* R is continuous. The classical Dirichlet problem asks one to find /: Q -* R such that / is harmonic on Int Q and / = g on 3Q. Many solutions to this problem have been discovered, some of which are quite ingenious. In particular, in Two dimensional Brownian motion and harmonic functions (Tokyo: Proc. Imp. Acad., 20, 706-714 [1944]), S. Kakutani showed how to construct / using probabilistic methods. He used a kind of limit of the random walk in R2 called the Wiener process or Brownian motion. Although the theory of the Wiener process is beyond the scope of this book, we can understand the ideas behind Kakutani's construction by looking at a discrete version of the Dirichlet problem due to Courant (Courant, R., Friedrichs, K. O., and Lewy, H. Ueber die partiellen Differenzengleichungen der mathematischen Physik. Math. Ann. Vol. 100. pp. 32-74 [1928]). (In fact, Courant showed that the solution to the classical problem can be obtained as a limiting case of the solution of the discrete problem described below!) Before we describe this discrete version of the Dirichlet problem, we need to translate the definition of harmonic functions into a form that is easily dealt with measure theoretically. Theorem. (Mean value property) Let 0 c R2 be open and let /: (9 -> R be harmonic. Let x0 e 0 and assume that the circle of radius a around x0 lies entirely in 0. Then (1) /(*o) = (^j £ 7(*o + «") d9 Conversely, if /: 0 -* R is continuous and equation 1 holds for all x0 and a such that the circle of radius a around x0 lies entirely in 0, then / is twice- differentiable and harmonic in (9. For a proof of this theorem see, for example, L. Ahlfors, Complex Analysis (New York: McGraw-Hill [1953]). Using this characterization of harmonic functions, we can formulate a plausible discrete analogue of the Dirichlet problem. The space R2 is replaced by the integer lattice Z2 = {(m, n); m, n are integers} and the compact region O becomes a finite subset of Z2.
116 Chapter 2 Integration For x e Z2, there are four nearest neighbors, xN, xs, x£, and xw, as pictured below. xN xwm mx mxE • If x e Q we say x e Int CI if xN, xs, x£, and x^ are all in CI as well. We then define dCl = CI - IntCl. To define harmonic functions on Intfi, the integral in equation 1 is translated to be the average over the nearest neighbors. Namely, if /: CI -► R we say / is harmonic on Int CI if fix) = \U(XH) + /(*s) + f(xE) + /W)] for all xg IntQ. Now let's consider the following problem. Discrete Dirichlet Problem Given g: dCl -► R find /: fi -> R such that / is harmonic on Int fi and f = g on dCl. We ask you to solve this problem by yourself. The following three exercises should be of some help. 1. Let 0tXQ denote the set of all random walks on Z2 with x0 as the starting point. This set can be identified with the set of all sequences of APs, E% S's, and W's (for example, N W W E S N...). Assign to N, E, 5, and W the numerical values 0,1,2, and 3. Let J = (0,1] = the half-closed unit interval. If coel, the quaternary expansion of co gives rise to a sequence of 0's, l's, 2's and 3's and hence to a sequence such as that above. Therefore, we can identify J with 0tXo. (For the details of this identification, see §1.2.) Now suppose x0 g CI, Consider the random walk rw e 0tXQ indexed by co el. Two possibilities exist: Either rw stays inside Intfi forever, or it eventually gets to a boundary point xb(co). (For instance, if x0edCl, then xb(co) = x0.) a. Show that the first of these two possibilities occurs with probability zero. (See §1.4, exercise 17.) b. Let fXo(co) = g [x6(co)]. Show that fXo is a measurable function of co e I.
§2.8 The Discrete Dirichlet Problem 117 2. Let 0t^.Q be the set of all random walks starting at x0 that move directly to xN on the first step. Define ^f0,^0, and #J£ similarly. a. Show that 0tXQ = «Jo u ^o u ^o u ^ (disjoint union) and show that, under the correspondence MXQ ~ J, 0$Xo corresponds to the interval (0,£], fflXo10 the interval (i,£], and so on. b. There is an obvious bijective map p: fflXo -> fflXfl. Namely, take the random walk whose first position after x0 is xN and think of it as a random walk starting at xN. Show that, if we identify 0tXft with (0,1] as in exercise 1 and identify M*o with (0,£| as in part a above, the mapping p becomes the mapping a> -> 4co. c. Show that, with the identifications in parts a and b, w /*»=/*0(f) Obtain comparable identities for fXE,fXs, and/X|K. 3. Define /: Q -> R by setting /(*o)= fx0(o>)dfi for all x0eQ, with fi being Lebesgue measure. Prove that / is harmonic and equal to g on 3Q.
Chapter 3 Fourier Analysis §3.1 if ^Theory Let (X, #", fi) be a measure space and let /: X ->C be a complex- valued function. We can write f(x) = u(x) + it;(x), where u and t? are real- valued functions on X and i = */— 1. Definition 1. f = u + iv is measurable if m and i; are both measurable. Note. tff = u + iv is measurable, then |/| = y/u2 + u2 is measurable by Theorem 14 in §2.1. Proposition 2. Let / = u + iv be measurable. The following two statements are equivalent. 1 \u\dfi < oo and |i;| Jx Jx 1. |/|<fo<co Proo/. Notice that |u| + \v\ > (u2 + t>2)1/2 £ |u| (or |p|). But |/| = (m2 + v2)1*2, so integration yields [ \u\dfi+ [ \v\dfi> f l/lrf/x 118
§3.1 &l-Theory 119 and \x[ndli~\xMdlX (0rLW*1) The first of these inequalities shows that statement 2 implies statement 1; the second inequality shows that statement 1 implies statement 2. □ Definition 3. If / = u + iv is a complex-valued measurable function on X, we say / is integrable if I \f\dp< co X In this case we define the integral of / to be the complex number (1) fd\x — udfi + i\ vdpt JX JX JX We denote by S£ 1{X,fj) the set of all such functions: (2) ^{X.ii) = j/:X->C,measurable; | \f\dpt < ool Proposition 4. Let f,ge££?x(X,pi), ceC, then 1. f + ge<el{X,p!) and | (f + g)dfi = I fd/i + | ^d/i 2. cfe&\X,ix) and | (cf)dpi = c \ fdpL 3. | /£fo = /<fo Jx Jx If fdn\< [ \Jx I Jx 4. /d/i < \f\dfi Proof. We leave proofs of 1,2, and 3 as exercises. To prove 4 let a = $xfdl*. Then (5/|a|)\xfd\i = \a\ is a positive real number. Let # = (fl/|fl|)/ and write g = u + iv, where u and v are real valued. Then fdfi \ = \a\ =— /d/* = #4u IJx I Mjx Jx = u dfx + i i; d\x Jx Jx
120 Chapter 3 Fourier Analysis Thus $x v d\i = 0 because the left-hand side is real. Therefore If fdfi\= [ udix<[ \u\dpi< \ \g\dti = \ \f\dii \jx I Jx Jx Jx Jx since \f\ = \g\. Q Parts 1 and 2 of proposition 4 are a proof that the space S£ 1{X,ii) is a vector space over the complex numbers. In fact, S£x{X,\i) is a normed vector space. Definition 5. Let V be a vector space over C. A norm on V is a function || • ||: V -> R with the following properties: a. \\v\\>0,veV b. \\v\\ =0ot; = 0 c. \\cv\\ = \c\\\vlceC,veV d. ||t; + w||< ||t;|| + ||w|| Given a norm || • || on a vector space K, we can define a metric d(-9 •): K x F ^ R on F by d(u, w) = || i; — w ||. It is easy to check that d satisfies the properties of a metric; that is, 1. d(v9 w) = d(w9 v\ v9weV 2. d(v, w) -+- d(w, u) > d(v> m), u, v, w e K 3. d(u, w) = 0 if and only if v = w (See Appendix A for a review of metric spaces.) There is a natural candidate for a norm on the space S£ 1(X,fx). Definition^ Let feSel(X,ii). We define H/IU = j*l/l4u to be the if1- norm of/. Unfortunately, ||. Id does not quite satisfy property b. Instead it satisfies property b': ll/l!i=0~/ = 0a.e. This statement means that two functions / and g in ££?l(X,ii) have to be considered the same if they are equal a.e. With this convention, it is easy to show the following theorem. Theorem 7. || • || x is a norm on if1 (AT, /*). Proof. Properties a, b', and c are obvious. Property d follows by integrating the inequality |/(x) + g(x)\ <: |/(x)| + \g(x)\ Q
§3.1 ^-Theory 121 We will say that a sequence of functions fne££l(X,ii\ n= 1,2,..., converges to a function fe&1(X,ii) in the J^-norm (or simply converges in (3) U/^/II^O as *->ao Convergence in this sense is not the same as pointwise convergence almost everywhere. In the exercises, you will find examples of sequences that converge in one of these two senses but fail to converge in the other. (See exercises 1 and 2.) The best one can conclude about the relationship between these two notions of convergence is the following. Theorem8. Suppose /„, n=l,2,..., converges to / in the if^-norm. Then there exists a subsequence /„£, i = 1,2,..., that converges to / almost everywhere. We will, in fact, prove a somewhat stronger result. Recall that, for a metric space (K, d), a sequence vneV9n= 1,2,..., is said to be a Cauchy sequence if d{vm9vn) -> 0 as m, n -> oo. In particular, a sequence of functions/„g ££l(X,iji), n= 1,2,..., is a Cauchy sequence if (4) ll/m-/Jli->0 as m,w->oo Theorem 9. Let /n, n = 1,2,..., be a Cauchy sequence in if1. Then there exists a subsequence /„., i = 1,2,..., that converges almost everywhere to an if1 function /. In addition, the original sequence converges to / in the J^-norm. Proof. Choose nx such that, for m, n > nu ||/m — fn\\i <i- Next choose n2 > n1 such that, for m, n > n29 \\fm — fn\\i <i- Continuing inductively, choose nt > n-^ such that, for m, n > ni9 \\fm — fn\\x < 1/2I+1. We will show that the subsequence {f„t} converges pointwise almost everywhere. By construction (5) ll/.l+|-/»,lli<2HT Let ^ = fni and let gt = fHi - fni_x for i > 2. Then (6) /„, = t 9r r=l and || 3, ||! <l/2r.Thus r=lJX
122 Chapter 3 Fourier Analysis so, by the corollary of the Lebesgue dominated convergence theorem (corollary 12 of §2.3), the series GO r=l converges pointwise almost everywhere; and, in view of equation 6, the sequence {f„{} converges almost everywhere. Let / be the pointwise limit of the /,{'s. By assumption,/is defined for almost all xeX, and we can define it at the remaining points of X by setting it equal to zero at these points. It remains for us to show that fn -> f in 5£1. Given e > 0, there exists an n0 such that \fm-fn\dfJ.<S for m, n > n0. Fixing n > n0 and letting m -> oo, we get e > lim inf |/m - fn\ dp > lim inf |/m - fn\dfi >Jl/-/J^=ll/-/Jli by Fatou's lemma. Hence /„ converges to / in the <£ ^norm. □ Recall that a metric space (K, d) is complete if every Cauchy sequence v„ e V has a limit veV. (Intuitively speaking, there are no "holes" in V.) A normed vector space (K, || • ||) that is complete with respect to the metric d(v,w)= \\v- w\\ is called a Banach space. By theorem 9, <£x{X,\x) has this property, so we conclude the following. Theorem 10. S£ * (X, p) is a Banach space. Exercises for §3.1 1. Let / be the unit interval 0 < x < 1, and let lk%n be the subinterval k fc+! „ , -<x < 0<fc<n n n Let/x be the characteristic function of/0>1,/2 and/3 the characteristic functions of /0,2and 7i>2, /4,/s, and /c the characteristic functions of /0,3, /lf3, and /2,3, and so on. Show that the sequence {/„} converges to 0 in S£X{1) but does not converge pointwise anywhere.
§3.1 ^-Theory 123 2. Let /„ be the function on the interval (0,1] that is equal to zero for 1/n < x < 1 and is equal to n for 0 < x < 1/n. Show that /„ converges point wise to zero everywhere as n -► oo but does not converge in ifl. 3. In exercise 1 extract a subsequence of the sequence {/„} that converges pointwise almost everywhere. 4. Let AT be a finite interval and \i Lebesgue measure on X. Show that there exists a countable family of functions {/> i = 1,2,3,...} with the property that the /f's are dense in S£x(X,\ji). That is, given any function / e .3? * (AT, /*) and any number e > 0, then, for some fh \\ f{ — /1| t < e. (Hint: See §2.2, exercise 7.) 5. Let AT be a set, and let @(X) be the set of all bounded, complex-valued functions on X. Forfe@(X) let ||/||= sup |/(x)| xeX Show that || || is a norm, and show that @(X) is a Banach space with respect to this norm. 6. In exercise 5 suppose the set X is infinite. Show that, if /i,/2,.-. is a sequence of functions in @(X), there exists a function fe@(X) such that n/-y;ii>i for all i. (Compare with exercise 4.) 7. a. Let p and q be numbers greater than 1 with (1/p) + (1/q) = 1. Prove that , ap bq ab< — + — P Q for any pair of nonnegative numbers a and b. b. Let (X,^9fi) be a measure space, and let / and g be nonnegative measurable functions. Prove that (*) jfgdp < ( [rd^fAg^d^ (Hint: Let a = (J/p^)1/p and 0 = ($gqdfi)1/q. At each point xeX, apply the inequality in (a) with a = /(x)/a and b = g(x)/P, and integrate with respect to x.) 8. Let / and g be as in exercise 7. Show that a\ i/p / r \ i/p / r \ i/p (/+0)p^J ^(J/P^J +(K<H (Hint: Write (/ + ^ = /(/ + g)p~l + g(f + 0)'"1 and apply equation (*) to each of the two products /(/ + g)p~l and g(f + g)p~l)
124 Chapter 3 Fourier Analysis 9. a. Let 1 < p < oo. A complex-valued measurable function /: X -* C is said to be ifMntegrable if \\f\pdp < oo. Denote by £fp(X,ii) the set of all such functions. Show that S£P(X9 /*) is a vector space. That is, show that, if / and g are in S£P{X9 /*), so is / + g, and that, if / is in £fp(X,n), any constant multiple of/ is in £fp(X9fi) as well. b. If/ei?*(X,/41et / r V/p WfWp -(Jl/I'*)' Show that || • ||p is a norm on S£p(X9\i). §3.2 ^2-Theory Let (AT, !F9 fi) be a measure space. A measurable function/: X -► C is said to be S£2-\ntegrab\e or square-integrable if (1) f |/|2^<oo We denote by 5£2{X9\x) the set of all such functions; that is, (2) S£\X9fi) = if: X -* C,measurable; | |/|2 d\i < oo 1 Definition 1. The quantity (3) ii/ii2=(£i/i2^)i/2 is called the i?2-norm of fe$e2{X9\i). We will see in a moment that equation 3 does indeed define a norm and that ££2 is a Banach space with respect to this norm. First, however, we will establish a few elementary facts about S£2. Theorem 2. If/ and g are in S£2{X9/4 fg is in S£*{X9fi). Proof Let X1 = {xeX; |/(x)| > |<?(x)|} and let X2 = {xeX; \g(x)\ > |/(x)|}. Then, on Xl9 \fg\ < |/|2; and, on X2, \fg\ < \g\2. So f \f9\dfi< f |/|2^+ [ \g\2dfi JX JXi J AT 2 *mi + Mi n
§3.2 &2-Theory 125 Corollary 3. If \i(X) < oo, then S£2(X, p) is contained in J&?1 (AT, /*). Proof, li fi(X) < oo, the constant function 1 is in JSf2(AT,/x). □ Corollary 4. If/ and # are in $£2{X9fi), so is / + g. Proof. It is enough to show that \f + g\2 is in JSP1, but |/ + #|2 < |/|2 + 2|/||fl| + I0I2. D Corollary 4 says that J£?2(X, p) is a vector space over the complex numbers. We will soon show that it has some other nice properties as well. First, however, we need to discuss briefly the subject of inner product spaces. Definition 5. A vector space V over the complex numbers is an inner product space if it is equipped with a mapping <-,*>:Kx V-+C such that 1. <>! +V2iW} = <U1,W> + <>2,W> 2. <cu,w> = c(v,w} 3. <u,w> = (wpv) 4. <u, t;> > 0 and <u, t?> = 0 if and only if v = 0 An example of an inner product space with which you are already familiar is the finite dimensional space C". If v — (a!,..., an)eCn and w = (bl,...ib„)e C", the inner product of v and w is (4) tfl,5 We will show that the much more complicated space j£?2(AT,m) is also an inner product space. Indeed, by theorem 2, the quantity (5) </,<?>= f fgdfi is well-defined for/, #e j£?2(AT,/j); and it is obvious from proposition 4 of §3.1 that it satisfies properties 1-3. It doesn't quite satisfy property 4; in fact, if <//> = Jl/l2^ = 0 the most we can conclude is that / = 0 a.e. But, if we put in force the convention that an ££2 function is zero "in the S£2 sense" when it is zero a.e., then property 4 is reinstated and we have proved the following theorem.
126 Chapter 3 Fourier Analysis Theorem 6. £?2(X, pi), equipped with the inner product given by equation 5, is an inner product space. We now prove a few facts that are true for inner product spaces in general and, thus, for <£2{X, fi) in particular. Given an inner product space (7, < , » and veV, let (6) iM=y<^> By property 4 this is well-defined and is zero if and only if v = 0. We call ||i;|| the norm of the vector veV. Theorem 7. (Schwarz's inequality) If v, w e V, then (7) \<v,n>\<\\v\\\\w\\ Proof. If w = 0, the inequality is obvious. If w ^ 0, consider, for Ae C, 0 < <t; + Xw,v + Aw> = <v,v> + I<u,w> + l<v,w> + |A|2<w,w> Letting A = — <t;, w}/(w, w>, this inequality implies <w,w> or, equivalently c |<u,w>|2<<t;,i;><w,w> □ Corollary 8. (Triangle inequality) For v,weV (8) ll» + w||<H| + ||w|| Proof. Squaring equation 8 we get (9) l|t; + w||2<||t;||2 + 2||t;||||w|| + ||w||2 But ||i? + w||2 = <t; + w,v + w> = <i;,i;> + <i;,w> + <w,t;> + <w,w>=||y||2 + 2Re(v, w> + || w||2, so equation 9 reduces to the inequality 2K*<»,w><2||t;||||w|| which is an immediate consequence of inequality 7. □ From this corollary we conclude the following. Corollary 9. The norm || • || on V is a norm in the sense of definition 5 of §3.1; that is, (K, || • ||) is a normed vector space. In particular, we restate this result for the vector space if2(AT,/x).
§3.2 ^-Theory 127 Corollary 10. ££2(X, pi) is a normed vector space. Moreover, applying Schwarz's inequality to ££2(X,ii\ we deduce the following. Corollary 11. If/ and g are in if2, then (10) ll/fflli^ll/llillfllli An inner product space (V, < , >) that is complete with respect to the norm || • || (that is, one that is a Banach space with respect to this norm) is called a Hilbert space. For example, Cn is a Hilbert space. We will show that if2(X,/x) is a Hilbert space. To simplify the proof we will make an assumption about the underlying measure space (X, ^, pi). We recall from the last paragraph of §1.3 the following definition. Definition 12. A measure space (X, ^, pi) is a-finite if there exists a sequence Xne&r,n= 1,2,... with (Jn°°=1 Xn = X and pi(Xn) < oo. Without loss of generality, one can assume that Xx £ X2 ^ • * •. Theorem 13. if2(Ar,/x) is complete with respect to the norm || • ||2; that is, it is a Hilbert space. Proof. Assume that X is a-finite. (See exercise 9 for a way to get rid of this assumption.) Let {/„} be a Cauchy sequence in £f2(X9pi). Choose Xn9s as in definition 12. By corollary 3, {/„} is a Cauchy sequence in £f1(Xl9pi). So, by theorem 9 of §3.1, we can extract from {/„} a subsequence {/i,„} that converges a.e. on Xx. Repeating the process, extract from this subsequence a smaller subsequence {f2t„} that converges a.e. on X2. Continuing inductively, one obtains for each i a subsequence^!^,,} of {/-!,„} that converges a.e. on XL. Now apply the Cantor diagonal process: The subsequence flil9 f2t2>... converges a.e. on X and its pointwise limit is equal a.e. to a measurable function g. Let gx =/n, g2=fi.2> and so on. Because the 0„'s are a subsequence of the /„'s, they are also a Cauchy sequence in if2(Ar,/x). So, for any e > 0, there exists an n0 such that ||^m — gn\\2 < e when m,n> n0. By Fatou's lemma, with n > n0 fixed and m -► oo, J- liminf|0m - g„\ dpi < liminf But the term on the left is \g-gn\2d\i \0m ~ 9n\2 dp < e J' because gm converges to g pointwise a.e. Hence, we conclude that g is in
128 Chapter 3 Fourier Analysis if2(Ar,/z) and that gn converges to g in if2(Ar,/z). Finally, if a subsequence of a Cauchy sequence converges, the sequence itself converges as well, so /„ also converges to g. □ The following general fact about inner product spaces will be useful in §3.3. Theorem 14. Let V be an inner product space. Then the inner product < •, • > is continuous in both variables with respect to the norm given by equation 6. In other words, i(vn-+v and wn -► w, then (vn, w„> -> <u, w>. Proof. livn^v with respect to || • ||, then || vn — v \\ < 1 for n large, so \\vn\\<\\vn-v\\ + \\v\\<l + \\v\\ for n large. Then Kvniw„} - (v,w}\ < !<!;„,w„> - <X,w>| + !<!;„,w> - <t;,w>| ^ I<v„, w„ - w>| + !<!;„ - v, w>| ^ll^llllw^-wll + ll^-^IMIwll ^(l + ll»ll)l|wll-w|| + ||i;il-r||||iy|| Hence | <t;„, w„> — <u, w>| tends to zero as n -> oo. D Exercises for §3.2 1. Let V be an inner product space. Show that Schwarz's inequality K^w>|<|M|||w|| is an equality if and only if w = 0 or v = cw for some complex number c. 2. Let V be an an inner product space. Show that (*) ll^ + w||2+||i;-w||2 = 2(||t;||24-||w||2) (Geometrically, the sum of the squared lengths of the diagonals in the figure below is equal to the sum of the squared lengths of the sides.) w
§3.2 ^-Theory 129 3. Let V be a normed vector space whose norm satisfies the identity in equation (*) in exercise 2. Show that there exists an inner product on V such that \\v\\=y/<^> (Hint: Show that 2Re<v,w> = \\v + w\\2 - \\v\\2 - \\w\\2 if an inner product exists.) 4. Let X = (0,1], equipped with Lebesgue measure. Show that the function f(x) = x"3/4 is in se^X,!*) but not in S?2(X,fi). 5. Let X = [1, oo), equipped with Lebesgue measure. Show that/(x) = x~3/4 is in &2(X,ti) but not in &l{X,n). 6. Let aua2>... be a sequence of positive numbers with Y*=ian = °°- Let sn = £?=1 a(. Show that Z^=oo but that oo a (Hfwt: Let /„ = £?=1 *ilst and let gn = £"=i ajs2. Compare fn with logs„ and gn with l/s„.) 7. Let (X, &, fi) be a measure space that is cx-finite but not finite. Show that &2(X,jj) is not contained in if l(X,fi). (Hint: Use exercise 6.) 8. Let (X,&9fi) be a measure space and fl9f29-.. a sequence of if2 functions on X. Let AT' = {xeX\ fi(x) =£ 0 for some i} Show that X' is a-finite; that is, show that it is a countable union of measurable sets of finite measure. (Hint: Let Emt„ = {xeX; \fn(x)\ > 1/m}. Show that X' = LLnEm,n) 9. Using exercise 8, show that theorem 13 is still true without the hypothesis that X is (7-finite. 10. Let (X, #", fi) be a measure space. Prove that J§?P(X, fi) is a Banach space— that is, complete with respect to the norm || • ||p. (See §3.1, exercise 9.) 11. (Sobolev's inequality) Let / be a function on the interval [0,1] that is continuous and has a continuous first derivative /'. Show that sup l/(x)-/(y)|<||/'||2 0<x,y<l
130 Chapter 3 Fourier Analysis §3.3 The Geometry of Hilbert Space In this section we will discuss some of the geometric properties of a Hilbert space ££ with inner product < , >. When we apply this material in the later sections of this chapter, ££ will always be if2(Ar, pi) where (AT,^, pi) is a measure space. Definition 1. Iff^ge^^v/es^y f is orthogonal tog {written f Lg)ii{f9g} = 0. Theorems (Pythagoras) If /, ge <£ with fig, then ||/||2 + \\g\\2 = II/+0II2. Proof. \\f + 9\\2 = <f + 9,f + 9> = </,/> + </.*> + <*,/> + <g,g> ll2 + \\g\\2 □ More generally, suppose that /x ,/2,... ,/„ e if with /£ _L yj-, i ^./. Then by induction it is easy to prove that || f, + f2 + • • • + fn\\2 = £?=11|/,||2. One basic example of a Hilbert space you should always keep in mind is C" with the inner product given by equation 4 of §3.2. This Hilbert space has finite "dimension." We will see, however, that some Hilbert spaces are "infinite dimensional"; in fact, these are the spaces that are most interesting to us. Definition 3. A sequence (/>i9 <f>2, <f>$,... in S£ is called orthonormal if Example 4. Let ^£ be C" with the inner product given by equation 4 of §3.2. Leti;1=(l,0,...,0),i;2 = (0,l,0,...,0),...,i;ll = (0,...,0,l). Then t^,...,*;,, is an orthonormal sequence. Example 5. Let X = [—7r,7t], pi = Lebesgue measure, and 5£ = if2(AT,ju). Let <j)k = (\ls/hi)eikx, — oo < k < oo. It is easy to check that the ^k's form an orthonormal sequence. Indeed f M-^ = 2^f e^'^dii -{ 1 iffc=J 0 if k*j
§3.3 The Geometry of Hilbert Space 131 Recall now that an orthonormal basis in Cn is, by definition, an orthonormal sequence of vectors vl9v2,...9vn. This definition depends on the dimension of Cn. Another characterization of orthonormal basis in Cn allows us to turn the tables and determine the dimension from the length of the basis: Namely, if v1,v2i...ivn is a basis of C", then there are no nonzero vectors that are simultaneously orthogonal to all the t^'s. This motivates the following definition. Definition 6. An orthonormal sequence ^, ^2>•••is called complete if, for any fe&, the conditions f-Lfa i=l,2,... imply / = 0. Remark. If if = ££2(X, /*), we must interpret / = 0 as / = 0 a.e. Definition 7. Let if be a Hilbert space, and suppose that <f>l9<f>2,...,<f>nis a. complete orthonormal sequence in S£. Then ^£ is said to have dimension n. If 5£ contains an infinite orthonormal sequence <j>x, </>2,..., ^ is said to be infinite dimensional We leave it to the reader to check that the dimension of a Hilbert space is well-defined (see exercise 1). Remark. Example 5 shows that, if X = [ — n9iz] and \x = \iL9 then £P2{X9jx) is infinite dimensional. We will see in §3.4 that the ^k's described in example 5 are complete. In infinite dimensions the notion of completeness of an orthonormal sequence replaces the notion of an orthonormal basis for finite dimensions. We now study some of the properties of a complete orthonormal sequence in if. Let <j>x, <j>2,... be a complete orthonormal sequence in 5£. Given f e if, let ci = </>^»>5 this is called the ith Fourier coefficient of/ with respect to the sequence <j>x, </>2,.... The formal series Yj%i ci</>i is called the Fourier series of / with respect to the sequence (j>u </>2.... Theorem 8. Let <t>u<l)2,... be a complete orthonormal sequence in S^. Let f^5£ and let c, = </,^/>. Define Sn = £?=1 c^ to be the nth partial sum of the Fourier series of/. Then Sn -► / in £f. Proof. Write / = / - Sn + Sn. Notice that </ - Sn, ^> = 0 as long as i < n because <f,</>i> = ci = <Sn,(f>iy for i<n
132 Chapter 3 Fourier Analysis Thus (/ — Sn) J_ S„ because Sn is a linear combination of ^'s with i < n. Then, by theorem 2, imi2 = ii/-sji2 + usnii2 > iisji2 Now, ||S„||2 = <Sn,Sny = (t^t^ = i\ct\2 1=1 So, we have shown that (2) £k-|2<:|mi2 for all n 1=1 That is, Y,T=i \ct\2 converges. Now consider Sn — Sm for n > m. Sn — Sm = /j"=OT+i Ci</>i, so by theorem 2 l|S„-SJ|2= t \c,\2 i=m+l Because we have shown that Yj=i \ci\2 converges, we can conclude that the sequence of S„'s is Cauchy. Because «£f is complete, we know that there is a g e 5£ such that Sn -> g in if. To finish the proof we need to show that / = g. Notice that, by theorem 14 of §3.2, (g,(f>i> = lim,,.^ <SW,^> = ct. Thus </ - g,^} = c, - c£- = 0 for all i. Because the ^'s are complete, we conclude that / = g. □ Theorem 9. (Plancherel) Let <j>u <j>2,... be a complete orthonormal sequence in if. For/eif, let ct = </,<*£>. Then ||/||2 = £Si |c,|2. Proof. S„ -> / in if so, by the continuity of the inner product, <s„,s„> -*</,/> = imi2 But ^.S,,) = £?=1 IcJ2, so we conclude that I kil2 = ll/ll2 D 1=1 You may have noticed that the proof of theorem 8 did not use the completeness of the orthonormal sequence until the final line. When the sequence is not necessarily complete, we get the following.
§3.3 The Geometry ofHilbert Space 133 Theorem 10. Let <j>i, ^2> • • • be an orthonormal sequence in ££. For fe <£, let c,= </»^i> and S„ = £"=1 c,^. Then S„ is Cauchy in if and thus converges to a limit g e if. Moreover, <#, ^> = </, ^> for all i. Proof. This proof is the same as the proof of theorem 8 with the last sentence omitted. □ If the sequence {</>„} is not complete, we get the following in place of the Plancherel theorem. Theorem 11. (Bessel's inequality) Let / and c£ be as in theorem 10. Then ll/li2>ZN2 f=l Proof. This inequality is a direct consequence of equation 2. □ Corollary 12. The Fourier coefficients ct tend to zero as i: -► oo. An important example of a Hilbert space is the space whose elements are infinite sequences (3) s = (a1,a2ia3i...) of complex numbers satisfying (4) £ lfl'l2 < °° The set of all such sequences is denoted I2 (read as "little S£2"). It is easy to see that it is a Hilbert space. In fact, it is a Hilbert space of the form if2(X,/i). Take for X the set of positive integers—that is, X = {1,2,3,...}. Let 3F be the (7-field of all subsets of X9 and let \x be the counting measure: fi(A) = number of points in A A function on X is just a sequence such as in equation 3. In the exercises we will ask you to show that such a sequence is in &2(X, p) if and only if equation 4 holds. We will also ask you to show that for two such sequences s = (al9a2,a3,...) and *= (bl9b29b39...) their inner product is
134 Chapter 3 Fourier Analysis (5) I a,b, i=l (Compare with equation 4 of §3.2.) Notice that Schwarz's inequality for I2 says that (6) E«A /oo \l/2/oo \l/2 few) CH We will use this fact in §3.4. Notice that, if we are given any Hilbert space 5£ and a complete ortho- normal sequence <j>x, </>2,... in £f, then, by the Plancherel formula, the sequence of Fourier coefficients s = (c1,c2,...) is in I2. Conversely, we claim that, given a sequence (ci,c2,c3,...)e/2 the sequence of partial sums converges in if to a limiting element /. Indeed, for m > n m USm-S„||2= I \c,\2 »=«+! so Sn is a Cauchy sequence and the assertion follows from the completeness of ^£. Thus, if we have a complete (infinite) orthonormal sequence in ££, we get a bijective map of 5£ onto I2. This statement is rather surprising in view of the fact that S£ can be, in principle, a much more complicated space than I2—for example, the space of square-integrable functions on Rn. Exercises for §3.3 1. Show that the dimension of a Hilbert space is well-defined. 2. Let X be the set of positive integers and pi the counting measure on X. Show that S£2(X,\x) = I2. Moreover, show that the 5£2 inner product on S£2(X^\i) is the inner product given by equation 5. 3. Let / be an <£2 function on the interval [ — n, %]. Show that I f(x)e inx dx^O as n -> oo (Hint: See corollary 12.)
§3.3 The Geometry ofHilbert Space 135 4. Let (AT, «^, fi) be a measure space. A sequence fn e if 2(AT, fj),n = 1,2,..., is said to converge in the if2 sense to fe£f2(X,n) if ||jf — fn\\2->0 as n -> oo. Prove that, if/„ -^ /in the S£2 sense, there exists a subsequence/„., i = 1,2,3,..., that converges to / a.e. 5. Let X be the unit interval and \x Lebesgue measure. Show that S^2- convergence in S£2(X,\i) does not imply pointwise convergence a.e., and vice versa. 6. (if 2-convergence of the randomized harmonic series) Let \x be Lebesgue measure on the unit interval, and let R„ be the nth Rademacher function. Let s--!,(t>' Show that Sn converges in the S£2 sense to a function //Eif2(/,/z). 7. For fixed m and n with m> n, let A be the set {coeI\ \Sk{a>) — Sn{a>)\ > £ for some k between m and n) Prove that fx(A)<^[(Sm-Sn)2dfx Here are some hints: (i) With k fixed, let J <= / be a union of intervals of the form i/2k < t < (i + l)/2*, with i between zero and 2k — 1. Show that, if n < k < m, ]jRmRHdp = 0. (ii) If J is as in part i, show that f (Sm-Sk)(Sk-S„)dfx = 0 and also show that f (Sm-Sn)2dfi> ((Sk-Sn)2dn JJ JJ (Hi) For n < fc < m, let /4k be the set {coeI; \Sj{co) - Sn(co)\ < e for n <j < k and \Sk(co) - Sn(co)\ > e} Show that A = \Jk=n Ak (disjoint union) and H(Ak)<\[ (Sm-Sn)2dfi
136 Chapter 3 Fourier Analysis 8. Using exercise 7 prove that the randomized harmonic series I(t)« converges almost everywhere to the function H of exercise 6. 9. (The Gram-Schmidt process) a. Let S£ be an inner product space and/i,/2,...,/n elements of S£. Show that, if/i,...,/„ are linearly independent, then there exists an orthonormal sequence fa,.., fa such that, for all i, (*) <t>i = Ysaijfj j<i with au > 0. (Hint: Let fa = c^/i where c[l is the length of ft. Let ^2 = c2[/2 — (f29fa)fal where cj1 is the length of/2 — (/i^iWi- Continue.) b. Let fx ,/2»• • • be an infinite sequence of elements of S£. Suppose that, for all n, fx,... 9f„ are linearly independent. Show that there exists an orthonormal sequence fa, fa,... such that equation (*) holds for all i. 10. a. Let S£ = ££2([-1,1]). Let); = l,/2 = x,f3 = x2, and so on. Apply the Gram-Schmidt process to this sequence (see exercise 9). Show that the resulting ^-'s are polynomial functions of x. (These functions are called the Legendre polynomials) Compute the first few of these functions. Answer: fa = —j=. fa = -x 1 3 *-4(§H) *>-M*-r) b. Use the Weierstrass approximation theorem (see §2.7, exercise 9) to show that the Legendre polynomials are a complete orthonormal sequence in S£2([— 1,1]). 11. a. Show that the Haar functions, #00(x)=l for0<x<l and #..*(*) = < -2"2 for^-Ux<^-i 2" 2n 2nl2 [ox±zA<x<h_ 2" 2n 0 elsewhere
§3.4 Fourier Series 137 where n > 1 and 1 < k < 2W, are an orthonormal sequence in S£2([0, 1]). b. Suppose that fe &2([0,1]) and (/, Hn,k) = 0 for all n, fc. Show that, if A is an interval of the form [fc/2B, Z/2n], with 0 < k < I < 2W, then 1 (**) fdfi = 0 c. Show that equation (**) holds for every subinterval A of [0,1]. Conclude that {H„fk} is a complete orthonormal sequence. d. Let Rn+1 be the n -h 1st Rademacher function. Show that 1 2" *Ni+l = ^ Zj "».* §3.4 Fourier Series We pointed out in §3.3 that the functions VV^J einx — oo < n < oo form an orthonormal sequence in the space if2 [—tt, jt]. We will show in this section that this orthonormal sequence is complete. In the course of proving this result, we will also prove a number of classical results about convergence of Fourier series. To begin, note that, by corollary 12 of §3.3, (1) yjllt J -K tends to zero as |n| -► oo, provided that / is an S£2 function on the interval [-7C,7T]. Let / be a measurable function defined on the whole real line. We will say that / is periodic of period 2% if (2) f(x + 2n)=f(x) a.e. Given any measurable function defined on the interval (—n, 7t], one can extend it uniquely to a periodic function on the whole real line by requiring that equation 2 hold. Moreover, if/ is periodic of period 2n and integrable over the interval [—7c,7t], then by equation 2 it is integrable over every compact subinterval of the real line. In fact, if / is a subinterval of length 2n9 then (3) \ fdx = fix J>-Lj
138 Chapter 3 Fourier Analysis To see this, suppose that / is of the form [a — 7t, a + 7t]. Then f(x)dx = f(x)dx+ f{x)dx- f(x)dx JI Jn J —n J —n But by periodicity the first term and third term cancel. We will first study convergence of Fourier series for functions that are rather nicely behaved. Let / be a continuous function periodic of period 2n, and let x0 be a point on the interval [ —7c,7i]. Suppose that the right and left derivatives of/exist at x0; that is, the limits (4)+ lim /(*>-^o> and (4)_ lim MZ/M x-(x0)_ X — X0 exist. We will prove the following theorem. Theorem 1. The series 1 °° S cnelnx y/27tn= converges at x = x0 and its limit is f(x0). Proof. Let SN(f)(x0) be the ATth partial sum of this series. By equation 1 SN(f)(x0) )= £ ckeik*<> 4j"/M(.l/'"-"",)J* Setting (5) DN(x) = ±- t eik* znk=-N we get for SN(f)(x0) the formula S*(/)(*o) = j * f(x)DN(x0 - x)dx
§3.4 Fourier Series 139 Making the change of coordinates x0 — x -► x, this integral becomes fx0+n f{x0-x)DN(x)dx Jx0-n So, by equation 3 we get finally (6) SN(f)(xo) = P f(x0 - x)DN(x)dx To estimate the right-hand side of this formula, we note first the following properties of DN(x): I DN(x)dx = (7) | DN{x)dx=l and i ei(N+l)x _ -iNx <8> W°> e«-i Proof of properties 7 and 8. To obtain equation 7, just integrate equation 5 term by term and note that all terms except k = 0 have integral zero. To obtain equation 8, rewrite DN(x) as 1 /2N -iJVx/ y gikx 27C \fc=o and note that, with a = eIJC, the second factor is just 2N a2N+1 - 1 V^afc, which equals —. Next we note that the denominator in equation 8 has a zero of first order at x = 0. In fact, eix - 1 d and eix — 1 has no zeroes on the interval [—n,7t] except at x = 0; so the function eix - 1 is continuous on this interval providing we define it to be — i at x = 0.
140 Chapter 3 Fourier Analysis We now return to the proof of theorem 1. By equation 7 /(xoHj f(x0)DN(x)dx Subtracting this from equation 6, we get (9) SN(f)(x0) - f(x0) = f* [/(x0 - x) - /(x0)]D*(x)<lx Set , , /(x0 - x) - /(x0) /(x0 - x) - /(x0) f x ^ *W = ?TTi = -x [T^Ti) The second factor on the right is continuous on the interval [—7t,7t], as we just observed. The first factor is continuous except at x = 0, and at x = 0 it is continuous on the left and on the right by assumptions 4±. Hence g is piecewise continuous and a fortiori in «Sf2. By equations 8 and 9 SN(f)(*o)-f(xo) = ^^ gMeW+^dx-^j" g(x)e-iNxdx The first term on the right is the — (N + l)st Fourier coefficient of g9 and the second term is the iVth Fourier coefficient; so by equation 1 both these terms tend to zero as N -► oo. □ A continuous function / is called piecewise differentiable if its domain of definition is a finite union of closed intervals and if, on each of these intervals, df/dx exists and is continuous. It is clear that, if / is piecewise differentiable, it satisfies assumptions 4± at all points in its domain of definition. We will show that for such functions theorem 1 can be considerably improved. Theorem 2. Let / be a continuous function that is periodic of period 2n. If /is piecewise differentiable on the interval [—7c,7t], then SN(f) converges to / uniformly and absolutely on this interval. Proof. Let g be the derivative of/ By assumption, g is defined and continuous on the interval [—7c,7t] except at a finite number of points, and we will define it everywhere by defining it arbitrarily at these points. We will first show that, if cn(f) and cn(g) are the nth Fourier coefficients of/ and g, respectively, then (10) cn{g) = incn(f) Proof of equation 10. We can find a0 = — n < at < a2 < • • * <ar = n such that g is continuous on (af,ai+i). Then
§3.4 Fourier Series 141 T ge~inx dx = Y. \ 9e~inx dx = H\ (if) *~inx d* fe~inx\ +in\ fe~inxdx\ U, J a, J = fe-inx\ +in I fe~inxdx | —n J—n However, the first term vanishes because / is periodic. Remark. Integration by parts is justified on [af, aI+1] because this integral equals the Riemann integral. To prove theorem 2 it is enough to show that £|c„(/)| < oo. Indeed, because we already know by theorem 1 that Sn(f) -► / pointwise, this fact will imply that the convergence is absolute and uniform. Because g is in if2, Zlcw(0)l2 < °°> so by equation 10 E"2kC0i2<a> Therefore, by Schwarz's inequality for I2 (see equation 6 of §3.3), i^--i(whcj4^)\^M2) Because both terms on the right are finite, so is their product. □ Let's now return to the proof that the functions <t>n = (l/^/2n)einx form a complete orthonormal sequence. We have to show that, if/ is an if2 function with </, <f>n) = 0 for all n, then / = 0 a.e. We will first show that, if / has this property, then (id [b fdx = 0 for every subinterval [a,fc] of [—n,it]. Let e > 0 and let Xt be the function indicated in the following figure: y=\ y = 0
142 Chapter 3 Fourier Analysis This function is piecewise differentiate. So, by theorem 2, SN(xe) -> Xe uniformly and, hence, a fortiori in if2. Thus 0 = </,SNfc.)>-*</,*.> that is, </, /£> = 0. As e tends to zero, xe converges in if2 to the characteristic function of the interval [a,fc]. So, by a repetition of the preceding argument, lim fXedx= fdx = 0 e-0 J -* Ja Thus, we have established equation 11. Consider now the collection of all measurable subsets of the interval [—7r,7c] for which (12) fdx = 0 This collection contains all the subintervals of [—7r, n] and is a A-system; so, by the 7r — A theorem (theorem 7 of §2.5), it contains all Borel subsets of [—7T, 7r]. Because every measurable set is a disjoint union of a Borel set and a set of measure zero, equation 12 holds for all measurable sets A. Prom now on we will assume that / is real-valued. (If not replace / by its real and imaginary parts.) Let A+ be the set where / > 0 and A- be the set where / < 0. Then / \f\dx= f fdx- [ fdx = 0 J-* Ja+ Ja- so / = 0 a.e. The Plancherel theorem now gives us P \f\2dx= £ \cj J-n n=-oo (13) with (H) C„—i=r/Me-** dx y/2K J -K for feJ?2( — 7i, 7c). We will discuss some applications of this identity in the exercises. Exercises for §3.4 1. a. Let / be an integrable function on the interval [—n9 n\. Let (f(x) when |/(x)| < M Show that j" \fM - f\ dfx -> 0 as M -> oo. *Uy~' ]n when |/(x)| > M
§3.4 Fourier Series 143 b. Show that the nth Fourier coefficient of/ tends to zero as n -* oo. (Hint: What can you say about the nth Fourier coefficient of /M?) 2. Let / be a periodic function of period 2n possessing continuous derivatives up to order k. Show that kcn(f) \ax j 3. a. Let / be the function <0)=«»>" ,, s fO for -n < x < 0 11 for 0 < x < n What are its Fourier coefficients? b. Prove that 00 1 „5(2n-l)2 4. The zeta function 7C2 ~8 and n=\n o C(s) = I n"s s > 1 is of considerable importance in number theory. By judicious use of the Plancherel theorem, evaluate this function at s = 2 and s = 4. Can you devise a method for evaluating £(s) at all even integers? (//int: See exercise 2.) a. Let / = f(x, t) be a function that has continuous second derivatives in x and t and is periodic of period 2n in x. Let cn(t) be the nth Fourier coefficient of f(x, t), regarded as a function of x (that is, with t fixed). Show that, if / is a solution of the heat equation {) dt dx2 thencn(t) = e-"2tcn(0). b. Given a function f0{x) that is periodic of period 2rc and has continuous second derivative, show how to construct a solution of the heat equation (*) with initial data: /(x,0) = /0(*)- The Weierstrass approximation theorem says that, if / is a continuous function on the interval [a, fc] and 6 > 0, there exists a polynomial p with SUP |/(X) - p(x)| < 6 a^x<b We sketched a proof of this theorem in exercises 8 and 9 of §2.7. Deduce from theorem 2 a second proof. Here are some hints: (i) Show that one can assume — n < a < b <n.
144 Chapter 3 Fourier Analysis (ii) Show that every continuous function / on the interval [a, b] can be extended to a continuous function on [—7r,7c] that is periodic of period 2n. (iii) Show that, if / is a continuous function that is periodic of period In, there exists a piecewise differentiate function f0 that is periodic of period 2n and is e/4 close to /; that is, SUP |/-/ol<T -ic<x£n 4 (Hint: See figure.) (iv) Let Sn(x) = Y,-Ncneinx> the JVth partial sum of the Fourier series for f0. Show that, for N sufficiently large, sup \SN-f\<- (v) In the formula for SN9 replace einx by J]?=0 (l/r!)(mx)r with k large. 7. Show that in theorem 1 it is enough to assume that / is continuous and dijfferentiable from the left and from the right at x0 and is piecewise continuous elsewhere. 8. a. Compute the Fourier series of the sawtooth function s(x) {x — Tt \x -I- n — n for 0 < x < n for —7t < x < 0
§3.5 The Fourier Integral 145 b. Show that the series 1 £ -einx converges everywhere on the interval — n < x < n except at the origin. 9. Let / and g be if2 functions on the interval (—71,71]. Extend them to functions on R by requiring that they be periodic of period 2n. Show that the convolution «-±r. is in if 1( — n9n) and that its Fourier coefficients cn are just where a„ and bn are the Fourier coefficients of/ and #. §3.5 The Fourier Integral Let / be a complex-valued integrable function defined on the real line. Its Fourier transform is the function -I' (1) fiy)=\me-">dx Notice that this function is well-defined because the absolute value of the integrand is |/(x)|. Indeed \M\ = f{x)e-^dx\< \\f(x)\dx J' so f(y) is bounded by the JSP1-norm of/ (In exercise 2 you will be asked to show that f(y) is continuous and that f(y) -» 0 as y -> + oo.) On the interval (—n,n)9 i?2-integrable functions are automatically if1- integrable; however, for functions defined on the real line, this is no longer the case (see §3.2, exercise 5). Therefore, equation 1 does not make sense if the integrand is an arbitrary S£2 function. Nevertheless, we will show that equation 1 can be appropriately defined for S£2 functions and that, just as for Fourier series, the 5£2-theory of the Fourier integral is remarkably simple and elegant. We will start by studying the Fourier transform for a very well-behaved class of functions.
146 Chapter 3 Fourier Analysis Definition 1. Let / be a complex-valued function defined on the real line whose derivatives of all orders—that is, df/dx, d2f/dx2, and so on—exist and are continuous. Then / is called a Schwartz function if, for each pair of nonnegative integers m and n, there exists a constant C (depending on m and n) such that (2) <® <C We will denote by 5 the set of all Schwartz functions. Iff and g are in S, so is / 4- g; and, iff is in S, constant multiples off are in S and so are xf and df/dx. Also, equation 2 implies that, given N, there is a constant C, depending on AT, so that (3) i/mi < c(i + \x\2rN So Schwartz functions go to zero very rapidly as x -► ± oo. The basic example of a Schwartz function, about which we will have much to say in the next two sections, is the function e~(x2/2). By equation 3 Schwartz functions are S£ Mntegrable; so their Fourier transforms are defined. We will prove that the Fourier transform of a Schwartz function is again a Schwartz function. To see this fact we need a fundamental property of the Fourier transform. Lemma 2. 1. LetfeS and let g(x) = xf(x). Then g{y) = yf^\(d/dy)}(y). 2. Let/e S and let h = df/dx. Then k(y) = yf^itfiy). In other words, up to factors of ^f — 1, the Fourier transform interchanges the operations "differentiation by x" and "multiplication by x." Proof. By definition f(x)e~ixydx yw-Jj The integrand on the right is differentiable with respect to y, and the derivative is again integrable; so the left side is differentiable with respect to y, and = -i xf(x)e-ixydx = -ig(y)
§3.5 The Fourier Integral 147 which proves part 1. To prove part 2 we note that = -//oo<J--'">«* K)e'ixydx = iyj7(x), The integration by parts is justified by the fact that / is going to zero very rapidly as x -► + oo. V It follows from the lemma that, if / is a Schwartz function, then (d/dy)f and yf are the Fourier transforms of Schwartz functions, and by induction ym(dn/dyn)f is the Fourier transform of a Schwartz function for all m and n. In particular, ym(dn/dyn)f is bounded; so/is a Schwartz function. Example. Let / = e~{x2/2\ We will show that (4) /(y) = v^^(y2/2) That is, up to a constant, / is its own Fourier transform. Proof. Notice that / satisfies the differential equation (5) f + */=0 Indeed, up to a constant factor, / is the only solution of this equation, for, if dh , r, — + xh = 0 dx then (d/dx)ex2/2h = ex2/2[(dh/dx) 4- xh] = 0. So ex2/2h is equal to a constant C and ft = Ce~ix2,2). By lemma 2, / satisfies equation 5 if / does; so f(y) is a constant multiple of e~{y2,2\ All that remains to check is that this constant is JIjl. But, if f(y) = Ce~{y2/2\ then = /(0) = JV(': 72) rf), The integral on the right can be evaluated by elementary means and shown to be yfhi. □ We can now state the first main result of this section.
148 Chapter 3 Fourier Analysis Theorem 3. The mapping / -> / maps S bijectively onto itself. Moreover, if feS and g = /, then / = g where (6) g(x) = —\g(y)eixydy Remarks. 1. The function g in display 6 is called the inverse Fourier transform of g. Notice that it is very simply related to the Fourier transform of g—namely, -(s)«- (7) M-[^)«-y) From this identity it is clear that the inverse Fourier transform maps the Schwartz space into itself. 2. Equation 6 implies that the Fourier transform is injective as a map of S into S. (That is, iffeS and g = / = 0, then, by equation 6,/ = 0.) It also implies that the inverse Fourier transform is surjective as a map of 5 into S. (That is, iffeS and # = /, then/ = g) But, because of the simple relation (equation 7) between the Fourier transform and its inverse, we conclude that both the Fourier transform and the inverse Fourier transform are injective and surjective. So, if we can prove equation 6, we will have automatically proved the rest of theorem 1. Incidentally, equation 6 is usually referred to as the Fourier inversion formula. For the proof of equation 6 we will need some additional properties of the Fourier transform. Lemma 4. Let / and g be Schwartz functions. Then (8) jf(y)g(y)dy = jf(xMx)dx Proof. By Fubini's theorem jto)g(y)dy = ^f{x)e-^dx\g{y)dy = \(\g{y)e-ixydy\f{x)dx = J6(x)Xx)dx The interchange of integrations is justified by equation 3.
§3.5 The Fourier Integral 149 Lemma 5. Let feS and let a be a real number. Then, iffa(x) = /(x + a), (9) faiy) = e*f(y) Proof. By definition /.GO- |/(x + a)<T'*Mx So, if we make the change of variables x = s — a, this becomes Liy) = *iay IfWe-tods = eia*f(y) Lemma 6. Let fe S and let a be a positive number. Then, if fa(x) = /(x/a), (10) Liy) = af(ay) Proof. By definition fa(y) = ^ffye-lxydx So, if we make the change of variables x = as, this becomes h(y) = a^me-ia**ds = aKay) We will now prove equation 6. Let/ = /(x) be an arbitrary Schwartz function, and let g = e~{y2/2a2). Then, combining equations 4, 8, and 10, we get (11) J f(y)e-{y2/2a2) dy = Jhta | /(x)£T<fl2*2'2> dx When we make the substitution ax = s, the right side becomes (12) V&f/Qe-*2* Now let a -> + oo. Then e"(y2/2fl2) tends to e° = 1 uniformly on compact sets; so the left side of equation 11 tends to J/QO dy. On the other hand, by display 12, the right side of equation 11 tends to y/2Hf(0)[e-<* i2/2)ds = 2tc/(0) and we obtain
150 Chapter 3 Fourier Analysis (13) /(0) -spw* Now let g be the function g(x) = f{x + a). Replacing / by g in equation 13 and taking into account equation 9 we get M = 9(0) = ^ [ mdy = ^\f(y)eiaydy □ Next we will show that the Fourier transform preserves 5£2-norms up to a scalar factor. Theorem 7. Let / be in 5. Then (14) ll/lli = 2*||/||5 Proof. If we take complex conjugates of both sides of the identity we get = ±\e*f(y)dy f{*) = ^e-ixytfmdy That is, 2nf is the Fourier transform of / Let g = f, so § = 2nf. Then, by lemma 4 |l/l2^ = J//^ = J/ff^ = Jj»ix = \f(2nf)dx = 2n\\f\2dx D Remark. This identity is called the Plancherel formula for the Fourier transform. Using theorems 3 and 7 we can now define the Fourier transform of an arbitrary if2 function. The idea for this definition is based on a theorem about metric spaces: Suppose M and N are metric spaces and A is a dense subset of M. A map fiA^Nis called uniformly continuous if, for every £ > 0, there exists a S > 0 such that dN(f(x)9f(y)) < e whenever djuix, y) < 8. The theorem we will need is the following. Proposition 8. If/: A -> N is uniformly continuous and N is complete, there exists a unique continuous mapping g: M -+ N extending/
§3.5 The Fourier Integral 151 A proof of this fact is outlined in Appendix A. By the Plancherel theorem, \\f — 0\\% = 2n\\f- g\\l; so the Fourier transform is uniformly continuous as a map of S into J?2. Moreover, as we saw in §3.2, <£2 is complete. Therefore, if we can show that S is dense in J?2, then, by proposition 8, there is a unique extension of the Fourier transform from S to S£2. In other words, if we can show that S is dense in &2, we will have succeeded in our goal of extending the Fourier transform to S£2 functions. To show that S is dense in &29 we need to show that there is a large supply of Schwartz functions. The results that we describe next make this point. We say that a function /: R -► R is C00 if all of its derivatives—that is, df/dx, d2f/dx2, and so on—exist and are continuous. Lemma 9. There exists a C00 function f0 that is zero for x < 0 and positive for x > 0. Proof. The function "(1/x) forx>0 «-fc" fo(x' '" forx<0 has this property. Lemma 10. Given an interval (a, b)9 there exists a C00 function fx such that /i = 0 for x${a,b) and/t > 0 for xe(a9b). Proof. Let f0 be as in lemma 9 and let /iM = /o(x-fl)/0(fr-x) V Lemma 11. Given an interval (a, b), there exists a C00 function f2 such that f2(x) = 0 for x < a, /2(x) = 1 for x > b, and 0 < f2 < 1 on the interval (a, b). Proof. Let ^(sjds /2(*) = " J —oo (s)ds Lemma 12. Given £ > 0, there exists a C00 function / of the type depicted in the following figure.
152 Chapter 3 Fourier Analysis a a+ e b—e b Proof. By lemma 11 there exists a C00 function g such that g = 0 for x < a, g = 1 for x > a + e, and 0 < g < 1 on the interval (a, a + s). Similarly, there exists a C00 function ft such that ft = 0 for x < ft — e, ft = 1 for x > ft and 0 < ft <_1 on the interval (ft - £, b). Now let /(x)_= #(x) [1 - ft(x)]. V Let S be the closure of S in if2; that is, feS if and only if there exists a sequence f„eS,n= 1,2,..., such that || f„ — /1| 2 -» 0. It is clear that, if g and ft are in 5, then # + h is in S and constant multiples of g are in S. Proposition 13. Let A_ be a finite union of intervals. Then the characteristic function \A of /4 is in S. /V00/. It is enough to prove this proposition for an interval A = [a, ft]. Given e > 0, let /be as in lemma 12. Because/is C00 and is zero outside [a, ft], / is in S. Moreover, \f — 1A\ < 1 on (a9a + s) and (ft — e,ft), and f=lA elsewhere, so J|/~lyl|2dx<2e □ Proposition 14. Let >4 be a measurable set of finite measure. Then the characteristic function 1A of A is in S. Proa/ Choose £ > 0. By the definition of J(¥ (see §1.3), there exists a finite union of intervals B such that fi(S(A,B)) < s. Then [\lB-lA\2dx = ti(S(A,B))<e □ Now let / be a nonnegative J£?2 function. By theorem 6 of §2.2, there exists an increasing sequence of simple functions sn > 0 such that s„ -> / By proposition 14, s„eS. Moreover l/~s«l2^-^0 as n->oo by the monotone convergence theorem, so feS. Finally, let / be a complex-.
§3.5 The Fourier Integral 153 valued if2 function. Then / = Re(f)+ - Re(f). + >/=Ilm(/)+ - ,/=! Jm(/)- so / is in S. Thus we have proved the following theorem. Theorem 15. S is dense in if2. We can now prove the second main result of this section. Theorem 16. There is a unique linear mapping (15) ~:i?2->if2 and a unique linear mapping (16) ~:£e2^&2 such that, restricted to S, equations 15 and 16 are the usual Fourier transform and inverse Fourier transform. These mappings are bijective and satisfy the Fourier inversion formula /=(/r and the Plancherel formula ll/ll! = 2K||/||! Proof. We have already indicated how the Fourier transform can be extended to S£2. The inverse Fourier transform can be extended the same way. Moreover, the Fourier inversion formula and the Plancherel formula hold on 5; so, by continuity, they hold on S£2. □ Remark. For a general if2 function /, the integral in equation 1 doesn't make sense. How then do you evaluate/? Of course, you have to approximate /by functions for which equation 1 does make sense, and then take limits. See exercise 4 for an explicit way to carry out this manipulation. Exercises for §3.5 1. Show that, if/eif^R), then for every e > 0 there exists a Schwartz function g such that \\f — g\\i < e. 2. If/ is in ifx (R), show that / is continuous and /(<!;) -> 0 as ^ -> ± oo. (Hint: These assertions are true when feS. Now use exercise 1.) 3. Show that, if / is both in &l(R) and in «^2(R), the two definitions of /—that is, equation 1 and the definition by continuity—coincide.
154 Chapter 3 Fourier Analysis 4. a. For/e«£P2(R)andM>0,let r ( \ _ J/M when M ^ M when |x| > M Show that \\fM - f\\2 -► 0 as M -+ oo. Show that, if /eif2(R), then lim Af-*oo J -M (*) lim f(x)e~ixydx M-+ao J -M exists, in the sense of if2, and is equal to /. (Equation (*) is often used as the definition of the S£2 Fourier transform.) Show that, if/, ge&2{R\ then -(s)0.« (This identity is called Parseval's identity.) (Hint: The real part of </, #> is equal to i(||/ + 0ll5-!/»!-11*111).) 6. a. Compute the Fourier transform of xe *2/2 and of x2e *2/2. Can you devise a scheme for computing the Fourier transform of xme~x2/2 for any ml b. Show that for every integer m there exists a polynomial //m(x) of order m such that the Fourier transform of Hm(x)e~x2/2 is a constant multiple of itself. Moreover, show that one can choose the Hm's so that the sequence Hme~x2/2, m= 1,2,..., is orthonormal. (The polynomial Hm(x) is called the mth Hermite polynomial.) (Hint: Use exercise 5.) 7. a. Let c be a positive number and let fc be the function '-6 .. . fe"" forx>0 UX)= 0 forx<0 Show that its Fourier transform is l/(c -I- iy). Use the Plancherel formula to compute the integral I dy -„c2 + y2 8. a. Let / be the characteristic function of the interval [—1,1]. Show that its Fourier transform is (2 sin y)/y. b. Use the Plancherel formula to compute Jl(^ ) dx 9. a. If/ and g are in Sf1, the convolution off and g is the function
§3.5 The Fourier Integral 155 (f*9)(x)=\f(x-y)g(y)dy Show that / * g is in S£l (R) and that || / * g \\ x < \\ f \\ x \\ g || x. (Hint: Use the Fubini theorem.) b. Show that the Fourier transform of /* g is the product fg. c. Conclude from part b that the convolution operation is associative and commutative. 10. a. Show that the function satisfies the heat equation dg/dt = d2g/dx2 for 0 < t < oo and — oo < x < oo. b. Show that, iff e S, the function ix(x,t) = gt*f= g(t9x - y)f(y)dy satisfies the heat equation for 0 < t < oo and — oo < x < oo. c. Show that, as t -► 0 +, /x(x, t) -> f(x). (Hint: Using part b of the previous exercise, show that J \(lt(y) — f(y)\ dy -► 0 as t -► 0 +. Here ixt(x) = fx(x,t).) 11. a. Let AT be a set, 3F be a c-field of subsets of AT, and /x a probability measure on X. Given a random variable/: X -► R, the function XrW= f «v*i is called the characteristic function of /. Show that x7 is continuous and |x/(OI < 1. b. Suppose / is bounded. Show that all the derivatives—(d/dt)xf, (d2/dt2)xf, and so on—exist and are continuous. Show that 12. a. Let (AT, «^\ /x) be as in exercise 11. Show that, if two random variables / and g are identically distributed, then Xf = Xg- b. Conversely, show that if X/ = Xg then f and g are identically distributed. (Hint: Let h(x) be a Schwartz function and ft(t) be its Fourier transform. Using the Fourier inversion formula, show that 00 Xf(t)kt)dt = f h(f)dn -co J* 2*1
156 Chapter 3 Fourier Analysis and interpret the right-hand side as I-**' where vf is the probability distribution of/) 13. Let (X9^,fi) be as in exercise 11. Show that, if/ and g are independent random variables, then X/+g = XfXg- (Hint: See equation 6 of §2.6.) 14. a. Let X be the unit interval, & the Borel sets, and /x Lebesgue measure. Show that, if/ = R„ is the nth Rademacher function, then xf = cos t. b. Let S„ = Yj=i Rt> Show that the characteristic function of Sn is (cos t)n. c Using part b of exercise 11, show that fsn2fcd/x = (~l)k^[(cosOn]| (Compare with §1.1, exercise 19.) 15. Let H = X^=i(Vw)^n be the randomized harmonic series. Let % be its characteristic function. Show that *w=flc°s(0 §3.6 Some Applications of Fourier Series to Probability Theory Let pn, — oo < n < oo, be a sequence of nonnegative real numbers with the following three properties: (1) Pn = P-n (2) pn = 0 for all but finitely many n's and (3) Zp.= l We will consider in this section a generalized version of the random walk in which a point-mass moves randomly along the real line with transition probabilities #_,. To be more specific, suppose that, at time fc, the position of the point-mass is the integer point i. At time k + 1 the point-mass is allowed to move to any integer positionj for which p£_j is nonzero and the probability of its moving to this position is assumed to be p£_j. (For instance, if px = p^ = \
§3.6 Some Applications of Fourier Series to Probability Theory 157 and the other pk's are zero, the process we have just described is the usual random walk.) To normalize, we will assume that the position at time zero is the origin. The basic random variables associated with this process are for each k9 (4) the difference between the positions at time k and time k + 1 Notice that the probability distributions v associated with these random variables do not depend on k; in fact, they are all just the measure (5) v(A)=Y,Pr reA for every Borel subset A of R. Indeed, if A is the one-element set consisting of the integer r, then v(A) = pr9 the probability that the point-mass moves r units to the right (or left) at time k. Now let X = I be the unit interval, fF = Bj the Borel sets, and fx = fxL. In §2.6 we showed that there exist independent, identically distributed random variables ft: I -> R, i = 1,2,..., such that equation 5 is their common probability distribution. If we take (X9lF,fx) to model the sample space of the process described above and the /fc's to model the random variables described in display 4, it is clear that we get an adequate measure theoretic model of this process. In this model the sum (6) Sn=tft 1=1 is the position of the point-mass at time n. Let's consider the question of when and how often the point-mass returns to its initial position. In our model the probability that the point-mass returns to its initial position at time n is the measure of the set {coeI;Sn(co) = 0} In the remainder of this section, we will show that (7) »({Sn = 0;Lo.})=l That is, with probability one the point-mass returns infinitely often to its initial position. (Incidentally, we will assume from now on that the transition probability p0 is less than one, for otherwise equation 7 is trivially true: The point-mass stays at the origin forever with probability one.) The first step in the proof will be to get a simple description of the measure of the set where S„ = 0. This step will be done using Fourier series. Consider the sum (8) g(t)= £ pne int
158 Chapter 3 Fourier Analysis By equation 2 this sum is finite; so there are no problems of convergence. We will show the following proposition. Proposition 1. The measure of the set {coeI;Sn(co) = r} is the rth Fourier coefficient of the function gn. Proof. We claim that for all k (9) g(t) = E(eitf*) = I e^dpt Indeed, display 4 tells us that fk is an integer-valued function, taking on only a finite number of integer values. In addition, it tells us that the measure of the set Em = {(oeI;fk(a)) = m} is just pm. Thus the right-hand side of equation 9 is Z^(/k = m) = ^/- = # as claimed. Because /i,...,/, are independent, so are eitfl,...,eitfn; so E(eits») = E(eitf> x ••• x eitf») = E(e^) x ••• x E(eitf») by equation 6 of §2.6. By equation 9 the right-hand side is g(t)n. On the other hand, because S„ is also an integer-valued function, taking on only a finite number of integer values, E(e^) = (e^d/i = 2>(S„ = r)el» Comparing the Fourier coefficients of this series with those of g(t)n, we see that fi(Sn = r) is the rth Fourier coefficient o{g(t)n. □ Before continuing with the proof of equation 7, we point out a few properties of the function g that we will use in our proof: (10) g is real-valued, (11) <7(0) =1, <7'(0) = 0, and 0"(O)<O, (12) \g(t)\ < 1, and (13) \g(t)\ < 1 except at a finite number of points on the interval [—n, it]
§3.6 Some Applications of Fourier Series to Probability Theory 159 Proof of properties 10-13: By equation 1 so g is real-valued. Differentiating equation 8, 0(0) = EP„, g'(0) = i^npn, and s"(0) = -2>2p„ By equation 3 the first sum is 1, and by equation 1 the second sum is zero. Finally, the third sum is negative, because, by equation 3, pn ^ 0 for some n # 0 (because we are assuming that p0 ^ 1). To prove inequalities 12 and 13, note that g(t) = Reg(t) = YJPncosnt so l0tol<IpJcosnt|<Xp„=l with equality holding if and only if cos nt = ± 1 whenever pn # 0. v Using these facts we will prove the following proposition. Proposition 2. The sum (14) JM{^U» = 0}) n=0 is infinite. Proof. By proposition 1 this sum is identical to the sum The integrand in the integral on the right is nonnegative and monotone- increasing (why?) and converges to the limit 1 1-0(0 except at those points where g(t) = — 1. Because the points with this property are finite in number, by equation 13, we get from the monotone convergence theorem 00 1 Cn dt Because the integrand on the right is nonnegative, we can show that the
160 Chapter 3 Fourier Analysis left-hand side is infinite by showing that dt i: + 00 for some e > 0. Let C be a constant with g"(0)>-2C Because g(0) = 1 and #'(0) = 0, we get from Taylor's formula with remainder that g(t) > 1 - Ct2 > 0 on a small interval, — e < t < e. Hence f£ dt fE dt If the events Sn = 0 were independent, we would now be finished: We could deduce that fx({Sn = 0; i.o.}) = 1 from proposition 2 using the second Borel-Cantelli lemma. However, because these events are not independent, we have to resort to a slightly more complicated argument. Let k be a positive integer and, for every positive integer /, let fl = fk+l Let Bk be the set of co e / where (15) Z/iM#0f r<K and £ fta) = 0 »=i f=i Similarly, let B\ be the set of coel where (16) tyi'M^O, r</, and £//(a)) = 0 Because the ^s in display 15 and the /"s in display 16 are independent, Bk and B\ are independent. Let pk = n(Bk) and pz' = ii(B\). We claim that, for k = /, pfc = p/. Indeed, let ^ be the joint probability distribution associated with fi,...,fk, and let n' be the joint probability distribution associated with //,..., fk. By theorem 3 of §2.6, n = fxfi x fxf2 x • • • x fifk = v x • • • x v and a similar identity holds for n', so n = n'
§3.6 Some Applications of Fourier Series to Probability Theory 161 Now pk is the measure with respect to /x of the set in display 15. But, by the definition of n, this measure is the same as the measure with respect to n of the set <(xi,...,xn)eRn; |>f t* 0, r < k, and J>, = 0> Similarly p'k is the measure of this set with respect to n'. Because n = n\ Pk = Pk as claimed. By definition, Bk is the set of random paths that return for the first time to the origin at time k; so the sum oo is the probability that a random path returns at least once to the origin. Similarly Bk n B\ is the set of random paths that return to the origin for the first time at time k and for the second time at time k + /. Because Bk and B\ are independent, the probability of this event—that is, the measure of Bk n B\—is pkpt; and the probability that a random path returns at least twice to the origin is L PkPi = [ LPk)\LPi) = P JM=1 \k=l J \l=l J We leave for the reader to show, by a similar argument, the following proposition. Proposition 3. If p is the probability of a random path returning to its initial position at least once, the probability of its returning at least k times is pk. We will deduce from this proposition that p must equal 1. In fact, suppose that p < 1. Let Ak = {a>eI;Sk(co) = 0} and let (17) h = £ 1^ Notice that, for m < oo, the set (18) {(oeI;h(a)) = m} is the set of random paths that return to the origin exactly m times; so, by proposition 3, the measure of this set is pm — pm+l or pw(l — p). For m = oo, the set in display 18 is the set of paths that return to the origin infinitely often; so, by the proposition, its measure is less than pk for all k; in other words, it is zero (assuming that p < 1). Thus
162 Chapter 3 Fourier Analysis I hdfi = £ mpm(\ - p) < oo f m<oo On the other hand, by equation 17 oo I, and the quantity on the right is infinite, by proposition 2; so we get a contradiction and conclude that p = 1. Let Ck be the set of paths that return to the origin at least fc times. Because p = 1, p(Ck) = 1 by proposition 3. Because Q => C2 ^ C3 • • •, "(,QC') 1 Hence, we conclude that random paths return to the origin infinitely often with probability one. Exercises for §3,6 1. Show that, for the classical random walk (p_x = p1 = j), every integer point neZ is visited at least once with probability one. (Hint: Suppose that the random walk visits the integer point n with probability p < 1. If n > 0, consider the set of random walks having the following two properties: (i) The first n moves are to the left, (ii) The origin is never revisited. Prove that the probability that a random walk belongs to this set is (1 — p)(l/2") > 0, contradicting equation 7.) 2. Show that, for the classical random walk, every integer point is visited infinitely often. (Hint: Let q be the probability that the point n is visited at least once, and let p be the probability that the random walk returns at least once to the origin. Show that the probability that n is visited at least k times is qpk~l. (Compare with proposition 3.) Now use the fact that p = q=l.) 3. a. For the generalized random walk with transition probabilities satisfying equations 1 through 3, let nl9...,nk be those integers for which pn t* 0. Let A be the set of those integers that can be written in the form rxnx + h rknk) with integers rx,..., rfc, and let n be the greatest common divisor of nl9...9nk. Show that A consists of all integer multiples of n. b. We will call an integer point a on the real line accessible if a e A. Show that a is visited with probability greater than zero if and only if it is accessible. 4. Show that, if an integer point a is accessible, then with probability one it is visited infinitely often by the generalized random walk. (Compare with exercises 1 and 2.)
§3.6 Some Applications of Fourier Series to Probability Theory 163 5. a. Let g(t) be the function given by equation 8. Show that g(t) = 1 if and only if t is a multiple of 2n/n (the n here being the same n as in exercise 3, part a). b. Conclude that g(t) = 1 if and only if cos kt = 1 for all admissible fc's. c. Show that if g{t) = 1 then Conclude that g"(t) < 0 when g(t) = 1. 6. Let TN(0) = t Prob(Sfc = 0) and TN(r) = £ Prob(5fc = r) *=i fc=i Show that, if r is admissible, lim[7^(0) — TN(r)~] is finite as N -> oo and is equal to 1 f* 1 -cosrt (*) ^- "J rr* (Hint: Use part b of exercise 5 to show that the integrand in equation (*) is a bounded function of t) 7. a. Show that, for the classical random walk (p_x = px = \\ Prob(S2„ -•»-#) b. By Stirling's formula (see S. Lang, A Complete Course in Calculus. Reading Mass., Addison-Wesley, 1968), there exists a number 0, with 0 < 0 < 1, such that n\ = x/2nnnne-nee/12n Deduce from Stirling's formula that Prob(S2„ ~ 0) = -= for n large. c. Prove from part b that £ Prob(S2n = 0) = oo n=l [Notice that Prob(S2n+1 = 0) = 0. Why?] 8. a. For the "unfair coin" [the process described in theorem 5 of §2.6 with fe = 2, rt = 1, r2 = -1, Pi = p ±\, and p2 = (1 - p) # ±], let S„ = /i+ ••+/„. Prove that Prob(S2„ = 0) = (2nj p"(l - pf and Prob(S2n+1 = 0) = 0
164 Chapter 3 Fourier Analysis b. Using Stirling's formula show that X Prob(Sn = 0) < oo and conclude (by the first Borel-Cantelli lemma) that Prob(S„ = 0; i.o.) = 0. Why doesn't this contradict equation 7? 9. Suppose that the transition probabilities pn, — oo < n < oo, satisfy equations 1 and 3 but not equation 2. Show that if £ n2p„ < oo, proposition 2 is still valid. 10. a. For the generalized random walk with transition probabilities satisfying equations 1 through 3, let Nn(k) be the number of times the point k is visited during the interval of time 0 < t < n. Show that the expectation value of Nn(k) is £ Prob(S, = k) 1=0 (Hint: Let Br be the set of paths that visit the point k exactly r times during the period 0 < t < n. Show that AT„(fc) = LrlBr Let A{ = {o; S^o) = k}. Show that Nn(k) = t UO 1=0 b. Show that the expectation value of Nn(k) is -r 2tiJ_ cos kt— at 1 -g c Using exercise 6 show that, if k is admissible, the expectation value of Nn(k) differs from the expectation value of Nn(0) by a quantity that tends to a finite limit as n -► oo. §3.7 An Application of Probability Theory to Fourier Series In this section we are going to discuss a classical theorem about Fourier series due to G. Szego. This theorem not only is of considerable theoretical interest, but also has a number of practical real-life applications. For those who want to learn about these applications, we recommend the very readable book by Grenander and Szego (Toeplitz Forms and Their Applications. Berkeley, Calif.: University of California Press, 1958).
§3.7 An Application of Probability Theory to Fourier Series 165 Szego proved his theorem in 1916, and since then several other proofs of it have been discovered. The proof described below is due to Mark Kac and consists of reversing one of the key arguments of the previous section. We begin by making a few definitions: Let 88 be the Borel subsets of the real line. Suppose we are given a probability measure \i and a sequence of probability measures /*!, /x2> • • • on ^- We will say that \in converges weakly to \x if, for every bounded continuous function /, h-l> a) yd^^fdix The notion of weak convergence will play an important role, not only in the following discussion, but also in the formulation of the central limit theorem in the next section. The second notion we will need involves some elementary linear algebra. Let T = (atj), 1 < ij < N9 be an N x N matrix of complex numbers with (2) au = an It is a standard theorem in linear algebra that T has N real eigenvalues; that is, the equation det(A - T) = 0 has N real roots Xx,..., XN (potentially occurring with multiplicities). For every subset A e ^, let fi(A) be the number of A£'s contained in A9 counting multiplicities. The measure fi defined by this recipe is called the spectral measure of T. If / is a bounded continuous function then (3) jfdii = jtf(X,) Having the notions of weak convergence and spectral measure, we can state a provisional form of the Szego theorem. Let a„, — oo < n < oo, be a sequence of complex numbers satisfying (4) 0-n=^ and (5) t l«nl<<*> n=— oo Associated with these numbers is the infinite matrix T = (amn) with amn = am_„ - oo < m, n < oo Matrices of this form are called Toeplitz matrices. Notice that, by equation 4,
166 Chapter 3 Fourier Analysis T satisfies the symmetry condition in equation 2. Let (6) TN = (am_n) 0<m,n<N-l be the AT x AT principal minor of this matrix. The Szego theorem in its provisional form says that, if fiN is the spectral measure of TN9 then fiN/N has a weak limit as N -> oo. We will identify this limit shortly, but before doing so let's observe an interesting tie-in between Toeplitz matrices and Fourier series. Let (7) <Z(0)= f a„eM n=-co By equation 5, q is continuous. Moreover, by equation 4 m = lL*-nein0 = YJanein° = q(e) That is, q is real-valued. Let i?2 be the space of i?2-integrable functions on the interval [ —7c,7c] and let Tq: &2 -► J&?2 be the linear mapping that sends fe £f2 to qfeS£2. Then, by equation 7 (8) Tf«"= J) a„e<<«+^= J an.m ^fcrf That is, an_m is the matrix associated with Tq in terms of the basis einB9 — oo < n < oo, of i?2. The matrix 7^ has a similar description: Let VN be the vector subspace of «£?2 spanned by the functions eM9 0 < n < N — 1, and let P:J2f2->Fw be the orthogonal projection of i?2 onto J^. In other words, iff is in i?2 and its Fourier series is — oo then Let be the linear mapping (9) JV-1 pf= EV- 0 (VN:VN-*VN (Tq)Nf = PTqf
§3.7 An Application of Probability Theory to Fourier Series 167 Then, for 0 < m < N - 1, (Tq)Neime = Y an_mein° n=0 by equation 8; so the matrix associated with (Tq)N in terms of the basis ein09 0 < n < N - 1, is exactly TN. Consider the probability measure ® /*L on the interval [ — n97t\. If we think of the function q: [ — n,7t] ->R as a random variable, its probability distribution \x is defined by the formula in equation 3 of §2.6; that is, for every Borel function /, (10) [fdii = ^\*flq(0ndiiL We can now state the Szego theorem in its sharp form. Theorem 1. Let fiN be the spectral measure of 7^, and let \x be the measure in equation 10. Then (id ^ weakly as N -► oo. Let us see what equation 11 says in concrete terms: Let Aj"> i=l,...,AT be the eigenvalues of 7^, and let / be a bounded continuous function on the real line. Then by equation 3 so equation 11 is equivalent to the assertion that (12) jjt fMN)) ^^[ /M] de as N -> oo, for every bounded continuous function /. We will now give a heuristic justification of equation 12 in terms of probability theory. Suppose that an = Pn — 00 < M < 00
168 Chapter 3 Fourier Analysis with the p„'s being transition probabilities satisfying equations 1 through 3 of §3.6. As in §3.6 we will consider the random walk based on these transition probabilities; but, instead of assuming that the random path starts at the origin, we will assume that it starts at one of the points x = 0,..., N — 1 and that all these points are equally likely as starting points. We will confine ourselves to a fixed finite interval of time 0 < t < m. Thus a sample path can be described as a zigzag line consisting of m segments. An example of a path for m = 3 is shown in the following figure. The t coordinate indicates time, and the x coordinate indicates position. t (2, 3) f (2,2)i >(3,1) I 4- x (2,0) We will compute the return-time probabilities for this process as in §3.6 but with a "confinement" condition imposed: What is the probability that a random path returns to its initial position at time t = m and stays confined in the box 0 < x < N — 1 and 0 < t < m? (For instance, in the figure the confinement condition says that, at times t = 0, t = 1, t = 2, and t = 3, the x coordinate of the path has to lie in the interval 0<x<N — 1.) Let us denote the probabilities in question by p(m, N). We claim that (13) p(m,N) = i£(A<.">r JM £=1 Proof. We will prove the case of m = 3, the general case being essentially no more difficult. Let k be a point on the interval 0 < x < N — 1. Because our random path has to start at some point in this interval and because all N points are equally likely, the probability that it starts at k is 1/AT. What is the probability that at t = 0 its position is fe, at t = 1 its position is Z, at t = 2 its position is n, and at t = 3 its position is again k? Clearly this is ( "y jPl-kPn-lPk-n The probability p(3, N) is therefore the sum
§3.7 An Application of Probability Theory to Fourier Series 169 (14) (jjlZPl-kPn-lPk-n over 0 <k,l, n < N - 1. But the sum in display 14 is also just l/N times the trace of the matrix T$\ that is, (15) p(3,N) = (^) traced In terms of the eigenvalues X\N\ this trace is just EOT establishing equation 13 for m = 3. □ Let's now drop the constraint condition—that is, no longer require the random path to be constrained to lie on the interval 0 < x < N — 1 but only require that its initial position lie on this interval. What is the probability, for an unconstrained random path, that its positions at time t = 0 and at time t = m coincide? Clearly this is just the "return at time m" probability computed in §3.6; that is, it is just **-*£ q(0)mdO where q(6) = Y,Pneine by proposition 1 of §3.6. Now it is intuitively clear that as N-> oo, with m fixed, p(m, N)-> p(m). Indeed, if we make the interval [0, N — 1] extremely large relative to m, relatively few paths with an initial point in this interval leave it before t = m in view of property 2 of §3.6. Hence, we conclude <"> sJ^-sJI mmde for f(x) = xm. By taking linear combinations of xm's we see that equation 16 is true for any polynomial function /, and a simple application of the Weier- strass approximation theorem shows that it is true in general. The following series of exercises will give you a chance to strip the probabilistic scaffolding from this proof. Henceforth, an, — oo < n < oo, will be an arbitrary sequence of complex numbers satisfying equations 4 and 5; that is, the a„'s will not necessarily be the transition probabilities associated with a random walk. 1. Show that 271 J- q3dO= £ arasa, i r+s+t=0
170 Chapter 3 Fourier Analysis 2. Suppose that all but finitely many of the a„'s are zero. Show that 3. Let a'n, — oo < n < oo, be another sequence of numbers satisfying equations 4 and 5. Suppose that, for all n, \an — a'n\ < e. Let M be the larger of the two sums Efc.1 and £WI Show that Z ar<*s<tt- Z ( r+s+t=0 r+s+t=0 <3eM2 4. Similarly, show that (*) (— ) | trace Tj - trace (T^)31 < 3eM2 5. Show that equation (*) is true without the assumption that all but finitely many of the an's are zero. 6. Prove equation (*) with the power 3 replaced by the power m. 7. Let a = min q and b = max q. Show that the eigenvalues of 7^ lie on the interval a < X < b, (Hint: Use equation 9.) 8. Let / be a continuous function on the interval [a, b]. By the Weierstrass approximation theorem, there exists for every s > 0 a polynomial function p such that sup|/ — p\ < e. (See §2.7, exercise 9, or §3.4 exercise 6.) Use this theorem in conjunction with exercise 7 to prove that fiN converges weakly to \i. 9. (Szego's original version of the Szego theorem) Let DN be the determinant of TN. Show that, if q is bounded below by a positive number, then DN>0 and lim Dk,N = exp|^- \ log q(9) dO] N-oo L2nJ-n J §3.8 The Central Limit Theorem In §2.7 we proved the following version of the law of large numbers. Theorem 1. Let/i,/2,... be a sequence of bounded random variables on X that are independent and identically distributed. Let E = £(/j) be the common expectation value of the /£'s. Let X0 be the set of points xe X for which
§3.8 The Central Limit Theorem 111 n as n -> oo. Then /jl{X0) = 1. In other words, if S.(x) = /i(x) + "- + /.(x)-n£ then (2) ^U n as n -> oo with probability one. In many practical problems, one would like to know how fast this convergence is. A close look at the proof of theorem 1 gives the following theorem. Theorem 2. Let/t ,/2> • • • be as in theorem 1, and let a > 0 be given. Then as n -> oo with probability one; that is, Sn(x)/n -> 0 faster than n~il,2)+a with probability one. Proof. The first step of the proof is to verify the following lemma. Lemma 3. For each k> 0 there is a Ck > 0 such that (4) £tfk^<Ckn* This lemma can be proved by induction. The case of k = 2 is done in the proof of theorem 1. The induction step is fairly messy but straightforward, and we leave it to the reader. (See §2.7, exercise 5.) Now, given a > 0, choose k so that 1/fc < a; notice that equation 2 and Chebyshev's inequality give (W>«})-'({i^>'"}) - E2knk+2ka ]x (5) *tt*s)xSr*l* < < — Ob Olrrt — £2kn2ka - e2*„2+/> where J? = 2fca - 2 = 2fc(a - 1/k) > 0.
172 Chapter 3 Fourier Analysis Now, as in §1.1, choose a sequence el9e2,... with en -► 0 and CO C La 2k 2+P „=l tn n Let < oo _ f r y. l ^W J . 1 71 "~ J X e A' M(l/2)+a n ( Then, by equation 5, Y*=i A*(4„) < °°. So, by the first Borel-Cantelli lemma, fx({An; i.o.}) = 0. But, for x in the complement of {An\ i.o.}, we must have Sn(x) M(l/2)+« <£„ for all but finitely many n's. Hence, we conclude S„(jc) ^ __->0 as M^°° with probability one. □ A natural question to ask at this point is "What happens to SJn112 as n -> oo?" To formulate the answer to this question, we need a definition. Let / be a bounded random variable with expectation value E. The integral -J,"- V(f) = ^(f-E)2dfi is called the variance of /. It is regarded by probabilists as a good measure of "deviation of / from its expectation value," because by Chebyshev's inequality, (6) ll{xeXt\f-E\>M}^^(f-Efdn = ^- For instance, inequality 6 says that the set where / deviates by one unit from its expectation value is less than V(f). If V(f) is very small, so is this deviation. Notice that, if v is the probability distribution associated with /, then by equation 3 of §2.6 = f {x-Eft V(f)=\(x-E)2dv So, if two random variables are identically distributed, they have the same variance. Let's now return to the question posed earlier. Because/i,/2,... are identically distributed, V(fx) = V(f2) = •••.
§3.8 The Central Limit Theorem 173 Theorem 4. Let a = V(fl) = V(f2) = • •. Then, for every pair of numbers a and b, with a < b (7) JxeX;a<^-<b)^-^[be-t2f2adt as n tends to infinity. This theorem is called the central limit theorem. It is sometimes stated as saying that, if the deviations of the /£'s from their expectation value E, for 1 < i < n, are rescaled by the factor n1/2, then these deviations tend to be normally distributed for n large. Notice that, if we denote by \in the probability distribution of the random variable SJnl/2 and by J the interval (a, ft), the left-hand side of equation 7 is fi„(J). The right-hand side, on the other hand, is J2%G J J (8) V.(J)- 7=\e-'^'dliL The measure \ia defined by equation 8 is called the Gaussian or normal distribution with variance a. The central limit theorem says that, for every interval J, (9) \in(J)^\ia(J) as n->oo Proof of theorem 4. Replacing f{ by ft — E, we can assume that E = 0. We will first prove a statement similar to equation 9 for the Fourier transforms of/*„and/v Lemma 5. Let Xn(t)=\ e-ixtdfin Then, for fixed t, xn{t) -+ e~ff'2/2 as n -► oo. Proof By equation 3 of §2.6, = f f e-Wirfdi*} x ••• x M e-itf»'fndv\
174 Chapter 3 Fourier Analysis by independence. Because the yj's are identically distributed, equation 3 of §2.6 tells us that I e-itf^dfi= I e-'W"d/i = ---= I e-*fnisT"dii Jx Jx Jx In fact, each of these expressions is equal to 1 e-itx/s/"dv R where v is the common probability distribution of the /,'s. Therefore, letting / be any one of the ft% we get for xn the formula do) ^w=(jv^^y Now notice that, for t fixed, where rn is a bounded function that tends to zero uniformly as n -> oo. Integrating the right-hand side, taking into account the fact that E = \fd\i = 0, we get (11) £^//^^=1_^(1+fin) where <* Jx S« = T / rndfi The Lebesgue dominated convergence theorem implies that (12) e„ -► 0 as n -> oo If we substitute equation 11 into equation 10, we get '■'"-['-©"-H"
§3.8 The Central Limit Theorem 175 Therefore, the proof of lemma 5 reduces to showing that (13) [i_(£)(i+0" • e-<"2/2 as n -> oo. For this proof, set a = —at2/2 and take the log of both sides. The left-hand side becomes »log[l + Q(l+o] or, with s = (1 + £„)/n, (l+£rt) s In view of equation 12, the limit of this expression as n tends to infinity is log(l + as) d km = — log(l + as)\s=0 = a s-o s as which is exactly the log of the right-hand side of equation 13. V We now return to the proof of theorem 4. We begin by proving an essentially equivalent statement. We will show that, if/ is a Schwartz function, then (14) f /<te,-> [ fdixa Jr Jr as n -> oo. To see this, note that L^-sLd.'^*)*- !H/M* The last expression, however, limits to 4= f f(y)e-**l2'dy 12-kg Jr by lemma 4 of §3.5 and lemma 5 of this section. Hence, we have established equation 14. To show that equation* 14 is essentially equivalent to equation 9, consider a Schwartz function ft of the form indicated in the following figure. (We showed in §3.5 that Schwartz functions of this type do exist.)
176 Chapter 3 Fourier Analysis y=l a a+ t Let J be the interval (a, b). It is clear that \fsdfxn<fxn(J) so by equation 14 liminf/i„(J) > lim ftd\in = fzd\i, n-*oo J J rb-B > e-x2/2°dx Ja+E The last inequality holds for all e > 0, so we obtain liminf/*„(J)> e~x2/2(rdx = fia(J) A similar argument shows that limsupA*n(J)<A*ff(J) so we conclude that the limit exists and is equal to the expression on the right. □ Another formulation of the central limit theorem is that the sequence of measures \xn converges weakly to the measure fj,ffasn-+ oo. Recall from §3.7 that this statement means that (15) Jr Jr for every bounded continuous function /. By equation 14 we know this fact to be true when / is a Schwartz function; and, by approximating by Schwartz functions, one can easily show it to be true for any bounded continuous function. Example: coin tossing. Suppose that in theorem 4 the /£'s are the Rademacher functions Rt. The strong law of large numbers says that (SJn)(co) -> 0 as
§3.8 The Central Limit Theorem 177 n -> oo. A gambler might want to know how many trials it takes to be reasonably sure this quantity is near zero. For example he or she might want to know that \Sn\/n < .01 with a probability of 99%. Because a = ViRJ = V(R2) = ••• = !, one gets from equation 7 the estimate .99 = JLeX;{^<m\) (16) =/i(|x6X;^<.01>Aj) *-t= e-t2*dt V27r J-.01JH .Oly/n ly/n By numerical methods one can show that, if 2= I e-'2/2dt = .99 /2n j-a then a = 2.57.... Hence, by equation 16, .01 y/n « 2.57 or n » 66,000; that is, after 66,000 tosses one can be 99% sure that \Sn\/n < .01.
Appendix A Metric Spaces We collect here, in a minimal sense, the important facts about metric spaces used in the text. This exposition is in no way complete and is meant only as an easy reference for the reader who is already familiar with these concepts. For a more thorough treatment of this matter, see W. Rudin, Principles of Mathematical Analysis, 3rd Ed. (New York: McGraw-Hill, 1976). Let M be a set. Definition 1. A metric on a set M is a map d( •, •): M x M -> R satisfying the properties 1. d(x,y) = d(y,x) 2. d(x,y)>0 3. d(x9 y) = 0 if and only if x = y 4. d(x, j;) < d(x, z) + d(z, j;) If d is a metric on M, the pair (M, d) is called a metric space. Example 2. Let (K, || • ||) be a normed vector space (see §3.1). Then V is a metric space with metric d(x,3>)=l|x-)i Metric spaces are nice because they allow us to define the basic topological objects we are used to considering in Rn—for example, open and closed sets, compactness, convergence, and so on. We first discuss convergence. 178
Metric Spaces 179 Definition 3. Let (M, d) be a metric space. Let xx, x2, x3,... be a sequence in M. We say {xw} is a Cauchy sequence if (1) d(x„,xm)->0 as n,ro->oo We say x„ converges to x e M (written x„ -► x), if (2) d(x, x„) -> 0 as n -> oo Proposition 4. If xt, x2,... is a sequence in M with x„ -► x e M, then the x/s form a Cauchy sequence. Proof. d(xni xm) < d(xn9 x) + d(x, xm) -> 0 as n, m -> oo D Thus, just as in Rn, every convergent sequence in (M, d) is a Cauchy sequence. The converse, however, is not true in general. Example 5. Let M = R — {0} and let d(x9 y) = |x — y\ for x, yeM. Clearly the sequence x„ = 1/n is Cauchy, and yet there is no xeM such that x„ -> x. Experience in Rn tells us that it is nice to be able to use display 1 as a criterion for convergence. This motivates the following definition. Definition 6. (M, d) is called complete if every Cauchy sequence in M converges to an element of M. Example 7. Rn with the metric / n \l/2 (3) d(x9y) = ^Z(xi-yi)2) is a complete metric space. We leave this fact for the reader to check. It follows from the fact that R with the metric in example 5 is complete. When we consider convergence of sequences, it is sometimes useful to know if a subsequence converges. One check for this is that the sequence be contained in a compact set. To define compactness in a metric space, we need to study the open and closed sets. Let (M, d) be a metric space and let A c= M be a subset. We say x e M is a limit point of A if for every s > 0 there is an xe e A with xe ^ x and d(x, x£) < e. The set A is called_c/osed if it contains all its limit points. In general, if A c= M, the closure of A, A, is the smallest closed set containing A, It is easy to check that A consists of the points in A together with all the limit points of A. A set 1/cMis called open if its complement is closed. Proposition 8. If U a M is open, then for each x0 e U there is an r > 0 such that the ball
180 Appendix A (4) Br(x0) = {x e M; d{x9 x0) < r} is contained in U. Proof, By definition, U is open if and only if Uc is closed. Now x0 £ Uc9 so it is not a limit point of Uc. Hence there exists an r > 0 such that Br(xo)nUc = 0. a A set K <=. M is called compact if, whenever {Ua} is a collection of open sets covering K9 there is a finite subcollection [/ai,l/a2 l/^ covering K. (A collection {Ua} of sets covers K if K c (J [/a.) Proposition 9. Let (M,d) be a metric space and let xl9x2,... be an infinite sequence in M. Suppose there is a compact set X c M such that xneK for all n. Then there is a subsequence xni, x„2,... of the xw's that converges. Proof. It is enough to show that the set A = {xj, x2,...} has a limit point. Suppose this is not true. Then for each yeK there is an sy such that, if By = {x; d(x9y) < sy}9 then By contains at most one point of A. The collection {By}yeK is an open cover of K with no finite subcover. This fact contradicts the compactness of K. □ This proposition gives a nice criterion for existence of convergent subsequences. The rub is that compactness is generally hard to check. However, in the case that the metric space is Rw with the usual metric given by equation 3, the Heine-Borel theorem gives a simple criterion for compactness. Theorem 10. (Heine-Borel) Every closed, bounded subset of Rw is compact. Remark. A set A cz Rn is bounded if there is a number M > 0 such that d(x, y) <M for all x, y e A. Warning. The Heine-Borel theorem is not true in a general metric space. To prove theorem 10 we need the following lemma. Lemma 11. Let (al9a2,•••,«„) and (bi9..,bn) be elements in Rn with aj < bj for all j = 1,2,..., n. Then the closed multi-interval J = {x e Rn; at < x£ < fcj is compact. Proof. Suppose, on the contrary, that J is not compact. Then there must be a cover of J by open sets {Ua} that has no finite subcover. If w& let Jf c R be the interval It = [ai9 fcj, we note that J = Ix x I2 x • • • x J„. Let ct = j(a( + bt) and let ir = lai9Ci] It = [ci?bj i = l,...,n
Metric Spaces 181 Then there are 2" multi-intervals of the form /±x/± x-x/± all of which are covered by the Ua's. At least one of these multi-intervals must not allow a finite subcover because J doesn't. Choose one such multi-interval and call it Jx. Now repeat this process ad infinitum to get a sequence of multi-intervals J => Jx =) J2 => • • * none of which can be covered by finitely many of the l^'s. For each fc take xkeJk, Notice that the sequence {xfc}£L, is a Cauchy sequence because the size of Jk decreases as fc -> cxd. Because Jx is closed there is a limit point x0 e Jx. Also, for each / the tail of the sequence {xfc}£L, is in the closed set Jt. Hence x0eJt for each /; that is, x0e f]fit Jt. Choose <x0 so that x0 6 Uao. By proposition 8 there is an r > 0 such that jBr(x0) c Uao. We claim that, for some fc, Jk c Br(x0). This claim contradicts the construction of the Jks and thus proves the lemma. To prove this claim let / n \l/2 ^ = (Z(^-^)2J and note that for x, y e Jk, d(x, y) < A/2*. Now, x0 e Jk for all fc, so we have that, if xe Jk, then d(x9x0) < X/2k\ that is, Jk c Bx/2k(x0). Choosing fc large enough so that X/2k < r we are finished. V Proof of theorem 10. Let C be a closed, bounded subset of Rn. Because C is bounded, it is contained in some closed multi-interval J. Because C is closed, the set Rn — C = U0 is open. Now let {Ua} be an open cover of C; the collection {Ua} u {U0} is then an open cover of J. By the lemma, J is compact so there is a finite subcover. If U0 is among these, throw it out; what's left still covers C. □ Finally, we can combine proposition 9 with theorem 10 to get the Bolzano-Weierstrass property. Theorem 12. (Bolzano-Weierstrass) Every bounded infinite set in Rn has a limit point. Now suppose that (M,dM) and (N,dN) are two metric spaces. A function /: M -> N is continuous at x0eM if for every e > 0 there exists 5 > 0 such that dM(x0iy) < d implies that dN(f(x0),f(y)) < e. The function / is called continuous if it is continuous at each point of M. Notice that when checking continuity the 8 may vary depending on x0. The function / is called uniformly continuous if the 6 can be chosen independently of x0; namely, / is uniformly continuous if for every s > 0 there is a 8 > 0 such that dN(f(x),f(y)) < e whenever dM(x, y) < 5.
182 Appendix A Proposition 13. Let/: M -* N be uniformly continuous and let xl5 x2,... be a Cauchy sequence in M. Then/(x1),/(x2),... is a Cauchy sequence in N. Proof. Given s > 0 we need to find K such that dN(f(xi)9f(Xj)) < s whenever i, j > K. Because / is uniformly continuous, there is a 8 > 0 such that dM(x,y) < 8 implies that dN(f(x),f(y)) < e. Because the x,'s are a Cauchy sequence, we can find K such that dM(xh xf) > 8 whenever ij < K. This is the K we sought. □ Now let (M, dM) be a metric space and let A c M be a subset. A is automatically a metric space with the metric induced from dM. In §3.5 we encounter a continuous function f:A->N and we wish to extend it to a continuous function g:M ->N. When/is uniformly continuous, we can make this extension if we assume that N is complete and that points in M can be "approximated" by points in A in the following sense: A subset A cz Mis called dense if A = M. Theorem 14. Let (M, dM) and (AT, djy) be metric spaces. Let A c M be a dense subset of M, and let /: A -► JV be uniformly continuous. Assume that N is complete. Then there exists a unique continuous mapg:M -> AT such that 0(x) = /(*) f°r all x g ^ /Voq/I We begin by defining g. Let x e M. If x e X we set #(x) = /(x). If x £ A then, because 4 is dense in M, x must be a limit point of A, Choose a sequence xl5x2,... in A with Xj-^xinMas/-> oo. By proposition 13 the sequence /(x£) in N is Cauchy. Because N is assumed to be complete, we know that this sequence has a limit; define this limit to be #(x). Notice that this definition is independent of the choice of sequence xl9x2i... because, if x\9x'29... is another such sequence, then f(x[) -> g(x). In this fashion it is also easy to see that, given s > 0, there exists a 5 > 0 such that, if dM(x, y) < 8 for x e M and ye A, then dN[g(x\g(yy] < e. To prove continuity of g, take x, yeM with dM(x, j;) < 8/2. Because A is dense in M, we can find zeA with dM(x,z) < 8/2. Then dM(y,z) < dM(y,x) + dM(x>z) < ^ an(* so dw[0(x),0(y)] £ 4v[0(*)>0(*)] + dN[g{z\g(yy] < 2s whenever dM(x,y) < 8/2. The uniqueness of g is a direct consequence of its continuity. □
Appendix B On <£p Matters You recall that in §3.8 we proved the following improved version of the law of large numbers. Theorem. Let/i,/2>--- be bounded, independent, identically distributed random variables on the probability space (X9fi). Let E be the common expectation value of the /j's, and let Sn(x) = fx(x) + • • • + fn(x) — nE. Then, for any a > 0, as n -> co with probability one. To prove this theorem we assumed that the /j's were bounded so that we wouldn't have to worry about the integrability of the functions S*k. Actually, if a > 1/fc it is easy to see that equation 1 holds as long as Juir (2) I l/l* < oo Indeed, in order to get the estimate in inequality 4 of §3.8, all you need to know is that (3) [\fi\lf£ x - x #| < oo when lt + - + I, = 2k It turns out that inequality 3 follows from inequality 2 once we have some basic facts about if p-spaces, which you have already proven in the exercises of §3.1 (see corollary 5 of this appendix). Here we first review those basic facts 183
184 Appendix B (don't peek until you've looked at exercises 7, 8, and 9 of §3.1 and exercise 10 of §3.2) and then we develop some ifp analogues of some of the i?2-theory in Chapter 3. Basic Theory Let (X,n) be a measure space. Recall that, for p > 1, S£p{Xy\x) (or just S£p if X,fi is understood) is the set of complex-valued measurable functions f:X^C such that (4) f|/|'4i<oo The value (5) ii/iip=(Li/|pd/x)1/P /ejsfp is called the JS?p-norm of/. Theorem 1. S£P(X, fi) is a vector space and || • ||p is a norm on S£P(X, /i). To prove theorem 1 we will need the following lemma of calculus. Lemma 2. Let #(t) be a convex function on the interval (a, b); that is, <j>"(t) > 0 for all te(a,b). Let x, ye(a,b) and let a and P be nonnegative numbers with a + p = l. Then ^(ax + #y) < ou/>(x) + /ty(j;) /Vaq/I Suppose this lemma is not true. Then, for some a, /?, x, and y as above, we have </>(<zx + ft;) > <x^(x) + jS^O) or <xO(ax + ft;) ~ <*(*)] > PW(y) - <*(** + ft,)] because a + /? = 1. Dividing by a/?(j; — x), we get #(coc + fly) - </>(x) ^ fly) - ^(ax + ft;) (ax + ft;) - x y — (ax + ft;) By the mean value theorem there exist f and rj with x<£<ax + Py<ri<y such that ^'(f) > fiW- This contradicts f(0 > 0. V
On Sev Matters 185 Proof of theorem 1. First, to check that J?p(X,fi) is a vector space, we need to see that, if/,ge ££P(X, fi), then J|/ + g\p < oo. To do this, notice that for p > 1 the function <ft(t) = tp is convex; hence, with a = /? = ^, we conclude from lemma 2 that @m+5i*i)'4i/r+5i*r pointwise in X. Integrating this inequality gives J 1/ + 0|'4* < 2'j Q|/| + ^Id)'** < 2*"1 (j l/IM/z + | IflfK^ Hence,/ + 0 e JS?'(*f ji). To show that || • ||p is a norm on £fp(X9fi)9 we need to prove the triangle inequality; that is, \\f + g \\p < \\f\\p + II0II p. (The rest of the norm properties are obvious.) To prove this inequality, we need the following. Lemma 3. Let / and g be nonnegative measurable functions on X, and let p and q be numbers greater than 1 with (1/p) + (1/q) = 1. Then (6) ffgdp < [\fPd^JP(\ ^^J" Proof Let a and b be positive numbers. Define the numbers x and y by a = ex,p and b = ey,q. Then, because ex is convex and (1/p) + (1/q) = 1, we have from lemma 2 that ab = e(x/p)+(y/q) < igx + \ey = iap + ±fcf p q p q Now let a = -m n-ttt and b = (W (I9**)'" and integrate to get inequality 6. To prove the triangle inequality, consider = 11/1(1/1 + \9\y~1 dp + j \g\(\f\ + \g\r~1 dp
186 Appendix B Apply lemma 3 with q = p/(p — 1) to get | (i/i + M)p^ < (| mp^)1/P (ja/i+ i^dp^)(p 1>/P so (7) (J !**)"([ (jfl/l + \g\rdfij" : \(p-D/p (l/l + Wdftj f + gL*[ (l/l + MW <\\f\\p +\\g\\ Remark. The triangle inequality for || • ||p, inequality 7, is called Minkowski's inequality. We can now give an argument to show that inequality 2 implies inequality 3. First, we have the following proposition. Proposition 4. (Holder's inequality) Let p and q be numbers greater than 1 with (lip) + (1/q) = 1. Let fe&p(X9fi) and gtSe\X,\i\ Then fge&l(X,p) and I fgdfi ^ UUgh Proof. By lemma 3 we have j\fg\dii = j\f\\9\dii£m,M, sofge^X.^and Ufgdfi\^j\fg\dn^\\f\\p\\g\\q D Corollary 5. Let pt and p2 be greater than 1, and let /i e J?"'(X, /i) and f2 e 2 "*{X, n). Then ft f2 e #'•'*«'»+'*>(X, /t) and II/1/2L II /1ll„II/2 Proof. Consider l/l/2|PlP2/(Pl+P2>^= |/1|PlP2/(Pl+P2)|y2|PlP2/(Pl+P2)^
On Sev Matters 187 Notice that/x e&p*(X9/i) implies that//lP2/(p'+P2)e J^*'2^2, and, similarly, //ip2/(Pl+P2)eif(P1+P2)/Pl.Now[p2/(pi +pa)] + q,i/(Pi +pa)] = i, so we can apply proposition 4 to conclude that f /r \P2/(pi+P2) / r \pi/(pi+p2) j (i/iii/2Dpip2/(pi+p2)^ < y i/iipi^j m i/2ip2^j so 11/1/2 HplP2/(Pl+P2) <ii/iIU/2iu n Now take k functions in iffc,/l5/2,...,/fc. By corollary 5, fx &£&**, Applying corollary 5 again gives/i/^ = (fif2)he^m- Continuing in this fashion, we get/i/2 x • • • x fke S£1; hence inequality 2 implies inequality 3. Notice also that corollary 5 gives the following. Corollary 6. Suppose fi(X) < 00. Then &p(X,ii) <= Se\X,\i) for 1 < r < p and (8) 11/11, <c||/||p for fe&IX.p) where c = [/*(*)] (p~r)/pr- Proof. Because fi(X) < 00 the constant function 1 is in «£?5(X, fi) for all s > 1. Let 5 = pr/(p - r). Then, by corollary 5Je&p*p+s) = Ser and 11/11, <ll/llpll Ills □ To continue now with the basic properties of ifp-spaces, recall that a normed vector space is called a Banach space if it is complete in the metric topology. Theorem 7. JSf P(X, p) is a Banach space. Proof. We will prove this theorem in the case that X is a-finite; that is, X = (JjSj Xt with Xx < X2 < • • • and /*(AT£) < 00 for all i. We leave the more general case to the reader (see exercises 8 and 9 in §3.2). Let {fn}™=i be a Cauchy sequence in «S?p(Z,/z); that is, given e > 0 there is an N such that || /„ — fm \\p < e whenever n,m> N. We wish to show that there is a function feJ?p(X,n) such that /„ -> f in «£?p; that is, given e > 0 there is an N such that \\fn — /||p < e when n> N. We proceed-as in the proof of theorem 13 of §3.2. Because ii{Xx) < °° we get from corollary 5 that $gp{Xu\i) c S£\Xu\i)> and the /„'s form a Cauchy sequence in i?1^,/*). Thus, from theorem 9 of §3.1, we can extract a subsequence {/i,„}^=i that converges a.e. on Xx. Similarly, because n(X2) < 00
188 Appendix B we can extract a subsequence {/2fW}"=i from the sequence {/i,n}"=i that converges a.e. on X2. Continuing inductively we extract a subsequence {fiin}ZLi from {/•-!,„}"=i that converges a.e. on Xt. By the Cantor diagonal process, the subsequence fiti,f2,2> • • • converges a.e. on X to a measurable function / on X. Let gx = /ltl, g2 = /2,2, anc* so on. The sequence {gn}T=i is Cauchy in &p(X9ii); that is, for s > 0 there is an JV such that \\gm — gn\\p < e when m,n> N. By Fatou's lemma, with n > N fixed and m -> oo, liminf |fc - flj'dji < liminf |^m - gn\pd\i < e The term on the left is J I/— gn\pdfi9 so this inequality shows thatfeJ?p(X,fi) and that gn-+ f in J?p(X9n). Because {^„}^=i is a subsequence of {/„}£Li, it follows that fn -► f in J£?p as well. D Representation Theorems Recall that, for the Hilbert space £f2(X, n), we have Schwarz's inequality If/^Ull/IUMU fovf,ge<?2(X,ti) I V I The generalization of this inequality to if-spaces is Holder's inequality I fgdp\z \\f\\P\\g\\q, feX1X,n\ gB<e\X,ti\ - + -=1 This inequality can be interpreted in the following way: For geJ?q(X,ij) consider the linear map (9) /,: X"(X, rf^C where /,(/) = jfg dp It is an immediate consequence of Holder's inequality that this map is continuous. Indeed, to show that a map F: ££p(Xy pi) -> C is continuous at /0, we need to show that, given e > 0, there is a 3 > 0 such that \\f — f0 \\p < 3 implies that \F(f) — F(f0)\ < e. In this case Uf) - Wo)l = \IJJ - /o)l < 11/ -fo\\P\\9\\q so, if we take 3 = e/\\g \\q9 we are done. The purpose of this section is to convince you that, in fact, every continuous linear map /: ££P{X, n)-+Cis of the form lg for some g e £?q(Xy fx) where (1/p) + (l/q) = 1. We will not prove this fact in general; we will just prove some special cases of it. If you want to see the general proofs, we recommend the
On Sev Matters 189 treatment in Reed and Simon (Methods of Modern Mathematical Physics: Functional Analysis, vol. 1. New York: Academic Press, 1980). We first introduce some nomenclature. Let (K, || • ||) be a normed, complex vector space. A continuous linear map /: V -> C is called a continuous linear functional. The dual space V* is the space of all continuous linear functionals /: V -► C. Notice that V* is a vector space. Proposition 8. Let I: V -> C be a linear functional. / is continuous if and only if there is a constant c> 0 such that (10) l(v)<c\\v\\ fomllveV Proof. The argument that inequality 10 implies continuity is the same as the argument that lg is continuous. To show that continuity of / implies inequality 10, notice that, because / is continuous at zero, if we choose e > 0 there is a 5 > 0 such that ||u|| < 8 implies \l{v)\ < e. Let c = 2e/d. Then for ve V notice that \M<> - I'Gr)I=>>»<' Hence |/(v)| < (2fi/«)|N| = c||w||. D Now let H be a Hilbert space with inner product <•,•>. If v e H we define lv e H* by lv(w) = < w, t;>, weH. (lv is continuous by Schwarz's inequality.) This definition gives a map L:H-*H* by L(v) = lv It is easy to see that Lis one to one. Indeed, if L{v) = 0 then lv(w) = <», w> = 0 for all w; hence i; = 0. The surprising fact is that L is onto. Theorem 9. The map L:H -> H* defined by L(v) = lv is bijective. Remark. We call this theorem a representation theorem because it "represents" the abstract space H* in terms of the known space H. To prove theorem 9 we will need a geometric result on Hilbert spaces, which is interesting in its own right. Let V <= H be a vector subspace of H. If V is also a closed subset of H (in the norm topology), then it is automatically a Hilbert space itself; in this case V is called a Hilbert subspace of H. If weH we write w ± V if <w, »> = 0 for all veV. Let V1 be the set of all weH such that w J_ K. Notice that K1 is a vector subspace of /f. Proposition 10, Let V c: H be a Hilbert subspace of the Hilbert space ff. Assume that V ^ H. Then there exists weV1 such that w # 0.
190 Appendix B Proof. Because V # H there is an x e H such that x $ V. Let a = inf ||x — v\\ veV We claim that there is a y e V such that || x — y || = a. To see this fact choose vneV such that ||x — vn\\ -»a as n -► oo. Then ||o. - vm\\2 = 2b. - x||2 + 2\\vm - x||2 - 4||i(t,n + oj - x||2 <2||t,n-x||2 + 2||i,m-x||2-4a2 Hence the vn's form a Cauchy sequence in if. Because V is closed in if, the sequence of i;„'s converges to y e V and ||x-y|| = lim||x-i>B||=a n->oo Now let w = x — j;, w # 0, because j/eK and x £ V. We claim that w e K1. If this claim is not true, then there is a v0 e V such that <w, v0} ^ 0. Let boll2 y =y+ .... in ^o Then y'eKand I foil <WM>0> ,|2 !|x-/|r = ||x — y — lln 2 v0 ,2 , I<W^q>|2 <W,P0) vn , <W,Pq>, boll2 Kll2^ * °> Kll2 II .,2 , IN^^O/ N^j^O// v \YV9U0/ , v = l|x,y.|2^l<w?i;o>|2 " '" Kll2 <II*-J>II2 This inequality contradicts the fact that ||x — y|| = a. □ We now prove theorem 9. Let /eH*. We want to find veH such that / = /„. If / = 0 we can take v = 0, so we will assume / ^ 0 from now on. Let K cz H be the subspace of if defined by K = {wetf;/(w) = 0} Note that K # H because / ^ 0. If {vn} is a Cauchy sequence in K, then vn -> i; for some veH.By the continuity of /, we see that /(»„) -► l(v) so /(») = 0. Hence veK. Thus we have seen that K is closed. From proposition 9 there exists w # 0 in K1. Let
On $£* Matters 191 Z(w) V = - rrW Then l) IMI2 Let x be any element of H. Notice that (x-j$v) = lix)-l{x) = 0 so X-M?VeK ButveK1, so \\v\\: 0 = ( x - -^2V,v ) = <x,v> - l{x) That is, l(x) = <x, i?> for all xeH. □ Now let's return to ifp-spaces. In the beginning of this section, we showed that, if ge£?q with (1/p) + (1/q) = 1, then g defines a continuous linear functional lg e £fp* given by lg(f) = \fgi[i for / e ££p. This definition gives a map L :&-+&** byL(g) = lg This map is one to one because if lg = 0 then, in particular, ^(1^) = J^ g dfx = 0 for all measurable sets A with pt(A) < oo. This implies g = 0 a.e. In fact, when X is a-finite, it is also true that L is onto; that is, L is an isomorphism. We won't prove this in full generality; instead we'll prove the following special case. Theorem 11. Assume that fi(X) < oo and let 1 < p < 2 and q = p/{p — 1). Then L\S£q'->£fp* is an isomorphism. Proof. Let Ze«£fp*. We want to find ge&q such that / = lg. Recall that, because \i{X) < oo, corollary 6 implies that S£\X,\£) c J?p(X9fi) and that there is a constant a > 0 such that ll/llp<a||/ll2 for/Gif2 Now, by proposition 8, there is a constant c> 0 such that |/(/)l<c|mip torfeX'
192 Appendix B Thus |/(/)l<c||/||p<ca||/||2 for/eJS?2 Applying proposition 8 again we see that, if we restrict / to ££2, we get something in if2*. By theorem 9 there is a g e ££2 such that ■J- 1(f) = ] fid? for/ei?2 We claim that g is actually in if*. To see this, fix K > 0 and set *»-fi tO it\g(x)\>K Then ft(x) is bounded, so heSC2(X,n) and |/(A)|~ |rf4t£c||fc». "I \J\g(x)\^K J \J\g(x)\£K J So (f lffMlf4*Y This bound holds for all K, so, by the monotone convergence theorem, we conclude that g(x)e ££q. Finally, because £f2(X,fj) is dense in ££p(Xy\i) (e.g., the simple functions are dense in JSfp), it is clear by continuity that 4 1(f) = fgdfi for all / e <£\X, fi D Remark. It is an easy exercise to extend this result to the case when X is (7-finite. Convolution In the exercises of §3.5, we defined the convolution of / and g when / and g are in ^(R) by
On &p Matters 193 (11) f*0{x) = j^fix-y)g(y)dy In those exercises you proved the following proposition. Proposition 12. f*geif X(R) and ii/*0iii<imiiii0iii Proof. Consider the iterated integral \f(x-y)g(y)\dxdy=W\\dg\\i JrJr JRJR By Fubini's theorem (theorem 15 of §2.5), the integral 1 f{x-y)g(y)dy R makes sense for almost all x and is equal a.e. to an &l function. Thus f*geSCl(R) and a second application of Fubini's theorem gives \\f*g\\i=\\\f(x-y)g{y)dy\dx JrJr JrJr \f(x - y)g(y)\dydx \f(x-y)g(y)\dxdy JRJR = II/Illll0lll □ Corollary 13. Convolution is a continuous map from if1 x if1 ->• SC1; that is, if/, -»/in JS?1 and gn-*ginSe1, then/, * gn -> f * g in S£*. Proof. Choose n large enough so that ||/ — /Ji < 1. Then ||/,Hi< (l + ll/IIJ-Now ll/*0-/.*ftJiHI(/-/,)**+ /.*(*-ft,)111 *ll/-/Jlill*lli + ll/.llill0-ftJi £fl/-/JillffOi+(l + fl/lli)ll0-flJ->O as n^co a
194 Appendix B 2. (f*g)*h(x) = | |r/(x - z - y)g(y)dyh(z)dz f(x — s)g(s — z)dsh(z)dz, s = z + y R Proposition 14. Let f,g,he &\R). Then 1. f*g = g*f 2. (f*g)*h = f*(g*h) Proof. 1. f*g(x)=( f(x-y)g{y)dy = /(s)0(x-s)ds, s = x-y = g*f(x) *g)*k{x)sShlt JrJi = /(x-s) g(s-z)h{z)dzds Jr Jr = f*(9*h)(x) Q One of the main strengths of the convolution is that it tells us what corresponds to a product in Fourier transform land. Proposition 15. (f*g)* = f$ for /, g e JSf^R). JYw/. (/ * 0) A(£) = f e-ix* f /(x - ,)flf(y) rfj, dx Jr Jr = f f ^/(x-j^OOdxdy JrJr JrJr =M)^e-Mg(y)dy The convolution can be extended to S£p-spaces in various combinations. For instance, if/ e Sep and g e <£q for (1/p) + (1/q) = 1, then /(x - y)g(y) e S£1
On S£v Matters 195 for each fixed x; so equation 11 makes sense. In fact by Holder's inequality we have (12) l/*0<*)|£ 11/11,11*11, for all x Hence, / * g is a bounded function when fe S£p and g e «£?*, (1/p) + (1/q) = 1. Proposition 16. Let fe&p, geS£q, and (1/p) + {1/q) = 1. Then f*g is a bounded, uniformly continuous function on R. Proof. Let s > 0 be given. We need to find 3 > 0 such that if |x — y\ < 8 then 1/ * g(x) — f * g{y)\ < e. Notice that \f*g(x)-f*g(y)\ = I U(x-z)-f(y-z)Mz)dz *Ux-f,hM< where fx(z) = f(x — z). Now let ^ be a compactly supported continuous function with ll/-0llP<e (You can prove that such ^'s exist.) It is easy to see that there is a S > 0 such that if |x — y\ < S then Ux-<f>y\\P<e Then ||/, - fy\\p < \\fx - d>x\\P +HX- <M, + ||/, - 4IIF < 3e □ Another combination for which the convolution can be defined is feJ?p zndge&1. Proposition 17. Let / e S£p, 1 < p, and g e J?l. Then / * g is well-defined and is in S£p. Moreover ll/^llp<ll/llpll^lli Proof. Let q = p/(p — 1) and notice that | \f(x - y)g(y)\dy = | \g(y)\1/p\f(x - y)\ \g(y)\l">dy Now, by proposition 12, \g(y)\ \ f(x — y)\p is integrable for almost all x, because \g\ and \f]p are both in <£x. Thus, \g(y)\llp\f{x - y)\ is in <£p for almost all x. Hence, for almost all x we have by Holder's inequality [ \f(x - y)g(y)\dy < Q \g(y)\ |/(x - y)\'dy\" \\g\\\f*
196 Appendix B Thus we have that / * g is defined for almost all x; furthermore, we get \f*g(x)\p < (| \g(y)\ \f(x - y)|'dy) MP By Fubini's theorem U*gVP < (^ \g(y)\\f{x - y)\"dxdy^ MP < miMi*" so ii/*0iiP<imipii0iii n Another important use of the convolution is that it allows us to give explicit smooth approximations to «£?p functions. We describe this in the following. Proposition 18. Let feJ?1 and let ^ be a Schwartz function. Then <f>*f is C00. Proof, j^(4*/)(x) = jj J 4(x - y)f(y)dy = \d^x-y)Mdy where the differentiation under the integral can be justified by the dominated convergence theorem. □ Now let <f>0(x) be a smooth function with support in (— 1,1) and such that <f>0(x)dx = 1 Let <f>k(x) = k<j>0(kx\ fc=l,2,.... Then <l>k(x) is smooth, supported in (-l/fc,l//c),and 1 <f>k{x)dx= 1 R Theorem 19. Let/ e S£p. Then <j>k * /converges to / in ifp, as k -► oo. /V00/. Let 0 be a smooth, compactly supported function with || / — g \\p < e. Then 11^*/ - /||p < 11^*/ - 4*0llP + 11^*9 - 0II, + \\g - /lip
On &p Matters 197 By proposition 17, \\<f>k*(f - g)\\p < ||/- g\\p because \\<t>k\\x = 1. Thus \\</>k*f-f\\P<2s+\\<f>k*g-g\\p Now notice that \<f>k*g(x) - g(x)\ = lg(* -y)- 0(*)M(y)<ty because \<l>k(y)dy = 1. Because # is compactly supported, it is uniformly continuous; that is, given s > 0 there is a 5 > 0 such that if | j/| < 5 then \g(x — y) — 0(y)| < e. Now ^k is supported in (— 1/fc, 1/k), so, if we take fe large enough, we can make \<f>k*g(x) - g(*)\ < e when xesuppg Furthermore, if g is supported in (a, b), then <j>k * gf is supported in (a — l//c, b + 1/fe). Thus, for large enough fe, we can conclude Uk*g-g\\p<e D Fourier Transform in «Sf P(R) In §3.5 we defined the Fourier transform We then used the density of the Schwartz space S in if2, along with the Plancherel formula, to extend the Fourier transform to an isomorphism of S£2 onto if2. The same techniques can be used to extend the Fourier transform to S£p functions for 1 < p < 2. We need the following two results. Proposition 20. The Schwartz space S is dense in ifp(R), p > 1. We leave the proof of this proposition to the reader as an exercise. It is essentially the same as the proof for p = 2 (theorem 15 of §3.5). The other ingredient we need is an S£v replacement for the Plancherel formula. Theorem 21. (Hausdorff-Young inequality) Let 1 < p < 2 and q = p/(p — 1). Then, for Schwartz functions/ (13) ||/||€ < C||/||p for some constant C A good reference for the proof of this is the book by Reed and Simon (see the reference section, page 202).
198 Appendix B This result tells us that the Fourier transform is uniformly continuous as a map from the dense set S c i?p(R) into the space &q{R). By theorem 14 of Appendix A we can then extend it to a map of «£?P(R) into if9(R), and inequality 13 still holds for all/e&P(R).
Appendix C A Non-Measurable Subset of the Interval (0,1] Let Sl be the set of complex numbers of modulus one (complex numbers of the form, c = a + ib, a, b e R, with a2 + b2 = 1.) This set forms a group under complex multiplication: If c\, c-i e S{ then c\c2 and c\/ci e S1. As a set one can identify Sl with the interval, / = (0,1], by means of the map (1) f:I—>S\ f(t) = e27rit. (Notice that this map is both one-one and onto.) We will use the identification (1) to transport the Lebesgue outer measure, /x*, on subsets of / to subsets of Sl; in other words if A is a subset of Sl we will define /x*(A) to be fA*(f~l (A)). (Thus, to show that there exists a non-measurable subset, U\ of / it will suffice to show that there exists a non-measurable subset, £/, of Sl and then take V = /-1 (£/)•) Let QCM be the set of rational numbers in the interval, /, and let SlQ = f(QCM). We claim Lemma 1. Sq is a subgroup ofS1. In other words, ifc\ and c-i are in SlQ so are c\C2 and c\/C2. Proof. By assumption cx = e2nitx and c2 = e2nitl with tu t2 6 Q O /. Choose Ji and S2 in QHI so that (t\ +1\) — s\ and (t\ —12) — S2 are integers. It's clear that Clc2 = e2ltiSi and c{/c2 = e2***2. D Given a subset, A, of 51 and co G Sl let coA = {coaya e A}. We leave the following as an exercise 199
200 Appendix C Lemma 2. For all subsets, A.ofS1 and all co e Sl (2) tf(fi>A) = tf(A). Hint: §1.3 exercise 14. Since the rational numbers on the interval, /, are a countable set, so are the points of Sq. Let {a>i,i = 1,2,...; a* e SlQ] be a sequence to which every element of Sq belongs. Without loss of generality we can assume that o>,- ^ coj for i ^ j. We will say that a subset, B, of Sx has the coset property (with respect to SXQ) if the sets Bj =(OjBt j = 1,2,..., are mutually disjoint: Bj O Bt = 0 if j' ± k. It is clear that sets with this property exist. In fact let c be any point in Sl and let B be the one element set consisting of c. Then /?,- H Bj ^<j> implies ccoi = ca>j or, dividing by c, on = a>7, i.e. i = 7. Let's pick a set, 5wflJO with the coset property which is as "large as possible," i.e. which is so large that if we add one more point to it, it no longer has the coset property. l We claim that one of the sets (3) Bk=cokBmaX9 * = 1,2,..., is non-measurable. Indeed, by assumption, these sets are mutually disjoint; and the fact that Bmax is "as large as possible" implies that (4) \JBk = S\ (Proof: if this were not the case, there would exist a c 6 Sl not in the union of the fljk's; and by adding c to Bmax we would get a larger set with the coset property.) Suppose now that the Z?*'s are all measurable. Then, since the union (4) is disjoint: (5) EM*(ft) = M*(S,) = M*(/) = l. But, by (2), (6) M*(ft) = M*(*m«x). Our naive intuition tells us that such a set has to exist; however, the existence of such a set involves some delicate issues in set theory (such as the axiom of choice and Zorn's lemma). Suffice it to say that the standard axioms of set theory permit us to assert that such a set exists, confirming our naive intuition.
A Non-Measurable Subset of the Interval (0,1 J 201 so the left hand side of (5) is either zero or infinity depending on whether fi*(Bmax) is zero or greater than zero. This argument by contradiction proves that one of the Bk's has to be non-measurable. Exercise. In fact show that all the Bkys are non-measurable and that Bmax is also non-measurable. Hint: See the exercise in §1.3 cited in the previous hint.
References P. Billingsley, Probability and Measure. New York: Wiley, 1979. W. Feller, An Introduction to Probability Theory and Its Applications, 3d ed. New York: Wiley, 1967. U. Grenander and G. Szego, Toeplitz Forms and Their Applications. Berkeley, Calif.: University of California Press, 1958. K. Hoffman, Analysis in Euclidean Space. Englewood Cliffs, N.J.: Prentice-Hall, 1975. M. Kac, Statistical Independence in Probability Analysis and Number Theory. Mathematical Association of America (Cams Mathematical Monograph, no. 12) 1959. S. Lang, A Complete Course in Calculus. Reading, Mass.: Addison-Wesley, 1968. M. Reed and B. Simon, Methods of Modern Mathematical Physics: Functional Analysis, vol. 1, revised and enlarged edition. New York: Academic Press, 1980. W. Rudin, Principles of Mathematical Analysis, 3rded. New York: McGraw-Hill, 1976. 202
Index almost always (a.a.), 50 almost everywhere (a.e.), 66 approximate outer measure, 29 Banach space, 122, 187 Bernoulli sequence, 1,2, 43-44 Bernstein polynomial, 114 Bessel's inequality, 133 binary expansion, 2 Bolzano-Weierstrass property, 181 Borel-Cantelli lemmas, 45, 47 Borel measurable, 58 Borel principle, 3-4 Borel sets, 37 extended, 54-55 Cauchy sequence, 179 central limit theorem, 170-177 characteristic function of a random variable, 155 of a set, 60 Chebyshev's inequality, 6, 9, 65-66 complete metric space, 122, 179 complete orthonormal basis, 130, 131 conditional probability, 46 convergence in^1, 121 in #2, 124 in metric spaces, 179 in 2X, 32 weak, 165 convex function, 184-185 convolution, 145, 154, 192-197 dimension, 131 discrete Dirichlet problem, 19-20, 115-117 discrete probability theory, 43 dual space, 189 expectation value, 16-17, 70, 102, 103 extended Borel sets, 54-55 extended real numbers, 53 Fatou's lemma, 78 field, 42 Fourier coefficient, 131 Fourier series, 131 Fourier transform, 137-145 Fourier transform in £p, 197-198 Fubini's theorem version 1,95-96 version 2, 98-99 version 3, 99 gambler's ruin, 15 Gaussian measure, 68-69, 173 Gram-Schmidt process, 136 Haar functions, 136-137 Hausdorff-Young inequality, 197-198 203
204 Index Heine-Borel theorem, 180-181 Hermite polynomials, 154 Hilbert space, 127 Holder's inequality, 186 identically distributed, 104 independent events, 46 independent random variables, 104 independent sets, 46-47 infinitely often (i.o.), 45 inner product, 125 integrable, 76 integral of integrable functions, 76 of nonnegative measurable functions, 61-63 of nonnegative simple functions, 60-62 inverse Fourier transform, 148 joint probability distribution, 104 X-system, 91-93 law of large numbers, 1-2, 44 strong, 9-11 weak, 6-9,113 2Oj0, 2(|i, E), 76 2\ 118-124 £2, 124-129 /2, 133-134 £p, 124, 184-188 Lebesgue dominated convergence theorem, 78-80 Lebesgue measure, 3, 36-39 Lebesgue measure zero, 8-9 Legendre polynomials, 136 lim inf, 50, 56-57 lim sup, 45, 56-57 linear functional, 189 marbles, 21-22, 105-106 Markov process, 23 measurable function, 53-58 measurable set, 34 measure, 27 space, 53 theoretic model, 42-43 mesh width, 83 metric space, 178-182 Minkowski's inequality, 186 monotone convergence theorem, 72-74 multi-interval, 25 Parseval's identity, 154 partition, 83-84 piecewise constant, 6-8 piecewise differentiable, 140 TT-X theorem, 92-94 ir-system, 92 Plancherel formula, 132, 150 probability distribution, 103 probability measure, 42 product measure, 94, 96 product set, 89 Rademacher functions, 4-6 random variable, 16, 55, 102 random walk, 17 generalized, 156 with pauses, 17-18 in the plane, 18-19 refinement, 83 representation theorems, 188-192 Riemann integrable, 84 Riemann integral, 83-88 Riemann sums (upper and lower), 83 ring, 25 run lengths, 46 sample space, 17 Schwartz function, 146 Schwarz's inequality, 126 set function, 25 additive, 25-26 countably additive, 26-27 Shakespeare, 15, 49 <r-field, 42 a-finite, 39 a-ring, 35, 42 simple function, 60 slice, 89 spectral measure, 165 square-integrable functions, 124 symmetric difference, 25 Szego theorem, 165
Index 205 ternary expansion, 12, 17-18 Toeplitz matrix, 165 triangle inequality, 126 unfair coin, 108, 163-164 uniform continuity, 150, 182 variance, 70, 172 vector space norm, 120 Weierstrass approximation theorem, 114, 143-144 winning streaks, 16