lntroduetion to

TOPOLOGY AND MODERN

ANALYSIS

INTERNATIONAL SERIES IN
PURE AND APPLIED MATHEMATICS

William Ted Martin and E. H. Spanier

CONSULTING EDITORS

Ah/.f-ora Complex Ane.Jysis

Bellman Stability Theory of Differential Equatiom
Buck Adv&need Calculus
Coddington and Let.rinBon Theory of Ordinary Djfferentja] Equatione
DeUman Mathematical 1\fethods in Physic.s and Engineering
Ep1tein Pa.rti&l Differen Ua.1 Equationa
Golomb and Sha-n.ks Elem en ts of Ordiru.ry Ditieren tial Eq u& tions
Gratie8 The Theory of Functions of Real v aria. bl es
Greenspan Intro d uc tio n to PRrtia.l D ifferen ti&l Eq ua. tions
Griffen Eletn.en ta.ry Theory of Numbers
Hamming Numerica.l Methods for Scientists and Engineers
H ilde~and Introduction to Numerical Analysis
H 01J8ekald.er Principles of Numerical Analysis
l-JiUB Elem en ts ol Pu re e.nd Applied 1\{ a them.a. tics
La8s Vec tor and Tensor A..n a1ysis
Lepage Complex Va.riabEes and the La pl&ce Tra..nsl orm for Engineers
N ehari Co nf orma.l ri.:t a.pp.ing
Newell Veetor A ualysis
Rosaetr Logie for 1-f athem a tic ians
Rudin Princip]es of Mathematical Analysis
Saaty and Bram Nonlinear l-rf &them a ties
Simtnons Introduction to Topology and Modern Analysis
Sneddon Elements of Partial Differential Equations
Snedd!ni. F ourior Transforms
Stoll Linear Algebra and Matrix Theory
Si,-uble Non] in ear Differen ti.al Equ a. tiom
Weimtoek Cs]culus of Variations
W mg Algebraic Number Theory

11ttrod11ctio11 to
TOPOLOGY AND
MODERN ANALYSIS

GEORGE F. SIMMONS
Associate Professor of Mathematics
Colorado College

INTERNATIONAL STUDENT EDlT ION

McGRAW-HlLL BOOK COMPANY, INC.

New York

San Fra.ncisco

Toronto

London

Kc5GAKUSHA COMPANY( LTD.

Tokyo

INTRODCUTION TO TOPOLOGY AND MODERN ANALYSIS

f":Oi"TER/\lA "J'lO~\AI~ S TlIDF.J.VT

EDlTIO~\T

Exclusive rights by Kogakusha Co., LtdT for lnanufaclure and exporc from Japan.
This book cannot be re-exported front the country to which ic is consigned by
Kogaku.sha Co.~ Ltd. or by McGraw~Hill no~Jk Company, Inc. or any of its :subsldiaries.
I

Copyright !963 by the ~fcGrav{-lfill Book Cornpany, Inc~ .All Rights Reserved.
This bookt or pares lhereof, n1ay nol be reproduced in any for1n l\Tithout pern1isa
sion of the publishers.
l .. HlR ARY OF -GO.~CRESS CATALOG C,\RD ~U\IBER

62-15149

For Virgie May Hatcher

and Elizabeth B. Blossom
TO EACH OF WHOM

I OWE MORE
THAN I CAN POSSIBLY EXPRESS

Preface
For some time now, topology has been firmly established as one of
the basic disciplines of pure mathematics. Its ideas and methods have
transformed large parts of geometry and analysis almost beyond recogni~
tionr It has also greatly ~timulated the growth of .abstract algebra.. As
things stand today, much of modern pure mathematics must remain a
elosed book the person who does not acquire a working knovlledge of at
least the element~ of topology~
There are many domains in the broad field of topologyr of which the
following are only a few: the homology and cohomology theory of complexes~ and of more general spa.ces as well; dimension theory; the theory
of differentiable and Riemannian manifolds and of Lie groups; the theory
of continuous curves; the theory of Banach and Hilbert spaces and their
operators 1 and of Banach algebras; and abstract harmonic analysis on
locally compact groups. Each of these subjects starts from roughly the
same body of fundamental knowledge and develops its own methods of
dealing with its o-..vn characteristic problems. The purpose of Part 1 of
this book is to make available to the student this .:ihard core" of fundamental topology; specificn.lly, to make it available in a form which is
general enough to meet the needs of modern mathematics, and yet is
unburdened by excess baggage best left in the research journals.
A topological spar.e can be thought of as a .set from which has been
swept away all structure irrelevant to the continuity of functions defined
on it. ~art 1 therefore begins with an informal (but quite extensive)
treatment of sets and functions+ Some writers deal with the theory of
metric spaces as if it were merely a fragment of the general theory of
t.opological spaces. This practice is no doubt logically correctt but it
8eems to me to violate the natural relation between these topics, in which
metric spaces motivate the more general theory. Metric spaces arP
therefore discussed rather fully in Chapter 2, and topological spacPs are
intl"oduccd in Chapter 3. The remaining four chapters in Part 1 are
concerned with various kinds of topological spaces of special importance
in applications and 'vith the continuous functions carried by them .
It goes without saying that one aspect of this type of mathematics
i8 its logical precision~ Too many writers 1 however, are content 'vith
this, and make lit t.le effort to help the reader maintain his orientation in

to

Yii

viii

Preface

the midst of maze.s of detail. One of the main features of this book is the
attention given to motivating the ideas under discussion. On every
possible occasion I have tried to make clear the intuitive meaning of -.,vhat
is taking place, and diagrams are provided~ whenever itt seems feasiblei
to help the reader develop skill in using his imagination to visualize
abstract ideas. Also~ each chapter begins with a brief introduction
which describes its main theme in general terms~ Courses in topology
are being taught more and more widely on the undergraduate level in
our colleges and universities, and I hope that these features, which tend
to sof tcn the austere framework of de futltions, theorer ns t and proofs~
will make this hook readable and easy to use as a text.
Historically speaking, topology has follo,vcd t'vo principa~ lines of
development. In homology theory~ dimension theory, and the study of
manifolds~ the basic motivation appPars to have co1ne from geometry.
In these fields, topological spaces are looked upon as generalized geometric
configurations, and the emphasis is placed on the structure of the spaces
themse]ve~..
ln the other direction, the n1ai n stimulus has been analysis.
Continuous functions are the chief objects of in t.erest here, and topological
space.s are regarded primarily as carriers of snr~h functions and as domain8
over which they can be integrated. These idea~ lead naturally into the
theory of Banach and Ililbert spaces and Banach algebras, the modern
theory of integration~ and abstract harmonic analysis on lo cal1y compact
groups.
In Part 1 of this book, I have attempted an even balance bet\veen
these two points of view4 This part is suitable for a basic semester course
and most of the topics treated are indispensable for further study in
almost any direction. If the instructor wishes to devote a second
semester to some of the extensions and applications of the theory, many
possibilities are open. If he prefers .applications in modern analysist he
can continue 'With Part 2 of this book 1 supplemented, ]JerhapsJ with a
brief treatment of measure and integration. aimed at the general form of
the I{iesz re presen ta ti on theorem4 Or if his tastes incline him to ward
the geometric aspects of topology, he can switch over to one of the many
excellAnt books which deal with these matters.
The instructor who intends to continue with Part 2 must face a
question which only he can answer. Do his students know enough about
algebra? This question is forced to the surface by the fact that Chapters
9 to 11 are as much about algebra as they are about topology- and analysis.. If his students know little or nothing about modern algcbrat t.hen a
careful and detailed treatment of Chapter 8 should make it possible to
proceed without difficulty. And if they knov{ a good deal~ then a quick
survey of Chapter 8 should suffice. It is my o~,.n opinion that education
in abstract. mftthematics ought to begin on the junior level \Yith a course
in modecrn alg;ebra and that topology f::l1011]d he oJTr.rcd only to students
1

Preface

ix

who have acquired some familiarity~ through such a course, with abstract
methods.
Part 3 is intended for individual study by exceptionally well ... qualified
student~ with a reasonable knowledge of complex analysis. Its principal
purpose is to unify Parts 1 and 2 into a single body of thought, along the
lines mapped out in the last section of Chapter 114
Taken as a whole, the present work stands at the threshold of the
more advanced books by Rickart {34L Loomis [271., and Naimark [~2];
and much of its subject matter can be found (in one form or another and
with innumerable applications to analysis) in the encyclopedic treatises
of Dunford and Schwartz [81 and Hille and Phillips 1201~ l This book is
intended to be elementary, in the sense of being accessible to well-trained
undergraduates, while those just mentioned are not.. lts prerequisites
are almost negligible.. Several facts about determinants are used l\rithout
proof in Chapter 11, and Chapter 12 leans heavily on Liou ville 's thcorctn
and the Laurent expansion from complex analysis4 "W~ith these exceptions, the book is essentially self-contained.
It seems to me that a "\i\rorth while distinction can be drawn bet ween
two types of pure mathematits~ The first-which unfortunately is
somewhat out of style at present-centers attention on particular function8 and theorems which arc rich in meaning and historyi like the gamma
function and the prin1e number theorem, or on juicy individual fa{~t~, like
Euler's wonderful f ormu]a
1

+ 7~ + 79 +

..

1f ~ /6~

1'he Recond is concerned primarily \vith form and strnct.urP.. The present
book belongs to this ca.mp; for its dominant theme can be expressed in
just two -.,vords, continuity and linearity, u.nd iii~ purpose is to illuminate
t.he meanings of these words and their relations to each other. Mathematics of this kind hardly ever yields great and memorable results like
the prime number theorem and Ruler's formula~ On the contrary~ its
theorems arc generaUy small parts of a. much larger whole and derive
their main significance from the place they occupy in that whole. In
my opinion, if a body of mat.hem a tics like th is is to justify it.self~ it must
possess aesthetic qualities akin to those of a good piece of architecture+
It. should have a solid foundationi it.s v.r,.alls and beams f-lhould be firm1y
.and truly placed, each part shouJd hear a meaningful relation to every
other part, and its to,vers and pinnacles should exalt the mind. Tt is my
hope that this book can contribute to a wider appreciation of these mathen1a tica.l values.

George F'. t.;immans

1

The num hers in

bracket~

reier to works listed in the Bibliography.

A /Vote to tlte Xeader

Two matters call for special comment~ the problems and the proofs.
The majority of the problems are corollaries and extensions of
theorems proved in the textJ and are freely drawn upon at all later stages
of the bookT In generalJ they serve as a bridge between ideas just treated
and developments yet to come 1 and the reader is strongly urged to master
them as he goes along.
In the earlier chapters, proofs arc given in considerable detail~ in an
effort to smooth the way for the beginner. As our subject unfolds
through the successive chapters and the reader acquires experience in
foil O\vi ng a bst..ra ct mathematical argurncn ts~ the proofs bee omc briefer
and minor details are more and more left for the reader to fill in for
hilnsclf. The serious student will train himself to look for gaps in proofs,
and should regard them as tacit invitations to do a little thinking on his
o~Tn.
Phrases like '~it is easy to seeJ~J 'ione can easily sho,v,~' ~'evidently,''
uclearly,1' and so on~ are always to be taken as "Taming signals which
indicate the presence of gaps, and they should put the reader on his

guard.
It is a basic principle in the study of mathematics, and one too
seldom emphasized, that a proof is not really understood until the
stage is reached at which one can grasp it as a whole and see it as a single
idea. In achieving this end, much more is necessary than merely follo'\\ring the individual steps in the reasoning. 1'\his is onJy the beginning.
A proof should be che,ved, swallowed, and digested, and this proces8 of
assimilation should not be abandoned until it yields a full comprehension
of the overall pattern of thought .

CoJttents
vu

Preface

A Note to the Reader

Xt

PART
Chapter One

ONE~

SETS AND F(INCTIONS

1~ Sets and set inc1usio n

TOPOLOGY

2. The n.Jge bra of sets 7

3. Functions 14
4. Products of set:s 21
5. Partitions and equivaJence relation8 25
6. Countable sets 31
7~ Uncountable sets 36
8. Partially ordered sets .and lattices 43

Chapter Two
9.
10.
11.
12.
13.
14.
15.

49

METlllC SPACES

The definition and some e-,;: am plea 51

Open sets 59
Closed sets 65
Con vergencc, com pletenees, and Bai.re,. a: theorem
Continuous mappings 75
Spaces of continuous functions 80
Euclidean and unitary epace.s 85

Chapter Three

Chapter Four

91

TOPOLOGICAL SPACES

16. The definition snd some ex amp] es 92

17. Elementary concepts 95
184 Open bases and open subbases 99
19. Weak topologies 104
20. The function algebras e(X 1 R) and e(X.C)

CO:AfPACTN ESS

21. Comp.e.ct spaces lll

22. Prod uets of spaces 11 S

70

106

110

xiv
23~
24~
25~

Contents
Tychonoff's theorem and locally compacl epaees
C.Ompa.ctne.ss for metric spaces 120
A.soolfts theorem 124

Chapter Five

118

SEPARATION

129

26. T i-spa.ces and Hausdorff s paec.s 130

27. Completely regu1 ar sp.acP.s and normal s pac~s. l .'3 2
28~ Urysohn~s lemma and the Tietze ext~nsion theorem
29. The Urysohn imbedding theorem 1:.17
30+ The Stone-Cech compe.ctification 139

Chapter Six

135

142

co--1NECTEDNE"SS
.

:,,..

3 l. C.Onnc~ted spaces 143

32. The components of a. space 146
33~ Totally dis eonneeted spaces
149
34. Lo c.al ly co nncctc d spa.c cs I 50

Chapter Senen

153

APPROXJM AT/(JN

35. The W cicrstra.ss a.p pro xima.t ion th eo rem l 5r1

36r The Sto nc-\\' cicrs t.rass thco rems 1S7
37. Locally compact H.a..11sdorff spaces 162
38. The extended Stone-Weierstrass t hcore ms 16.5

PART TWO:

Chapter Eioht

Chapter Nine

171

AWEBRAIC SYSTEJIS

39. Groups 172

:40~ Rings
181
41. The struct urc of rings l S4
4 2. Linear spaces 191
43. The dimension of a linear Ap&ee
44. Linc a.r tr a.nsformations 203
45 .. AlgP.hras 208

OPERATORS

196

211

BANACH SPACES

46. The dP.finition n.n<l some examples 212

47. Continuous linear t.ransforn1ations 219
48. The H ah n-Banac h theorem 22 4
49~ The natural imberl<ling of 1V in ]'{*
2~JI
50+ The open mapping theorc1n 235
51 ~ The conjugate of a.n o pPra. tor i39

Chapter Te-n

HII~BERT

243

SPACES

52. The clefinilion and some simple properties

53. Orthogona.1 complements 249
.5 4~ Orth o norm&! sets 251

244

Contents

KY

55. The oonjug&te space H 260

56. The adjoint of an operator 262
57. Self-adjoint operators 266
58~ Normal and unitary opera.tors 269
59. Projectiom 273

Chapter Eleven FINITE-Dll\.fENSIONAL SPECTRAL THEORY 278

60.
61.
62.
63.

Ma.trices 280
Dctermina.nta and the spectrum of an operator
The .s pectra.l theorem 290
A survc y of the sit ua.ti on 295

PART THREE:

287

ALGEBRAS OF OPERATORS

Chapter Twelve GENERAL PRELIMINARIES ON BANACH

AUJEBRAS
64~

65.
66+
67.
68.
69.

The definition and some examples 302

Regular and singular elem en ts 305
Topological divisors of zero 307
The spectrum 308
The formula. for the s pectra1 radius 312
The radical and aemi-simplicity 313

Chapter Thirteen THE STRUCTURE OF COMMUTATIVE

BANACH ALGEBRAS
70.
71.
7 2.
7 3~

301

The GeHand mapping 318

Applicationa of the form ul.a r (x)
lim IJ x" II 1 '~
Involutions in Ba.nae h a.lgebr.as 324
The Geliand-N eumar k theorem 325

318

323

Chapter Fourte.en SOME tSPECIAL COMMllTATIVE BANACH

AWEBRAS
327
74. Ide&ls in e (X) and the Ba.nae h-Stone theorem 327
75. The Stone~Cech com pa.eti fi cation (continued) 330
76. Comm u ta.ti ve C *-algebras 332

APPENDICES

ONE
TWO
THREE

Fixed po.int theorems and ~omc n.pplica.tions to aneJy19is 337

Continuous. curves and the Hahn-1\-f a..zurkicwicz theorem 341
Boo lean a]ge bras, Boole an rings, and Stone's theorem. 344

Bibliography

355

Index of Symbols

359

Subject Index

363

PART ONE

CHAPTER ONE

Sets and 1unetions

It is sometimes said that mathematics is the 5tudy of sets and func . .
tions.. Naturallyj this oversimplifies matters; hut it docs come a.s close
to the truth as an aphorism can.
r-fhe study of sets and functions leads two ways. One path goes
down, into the abysses of logic, philosophy, and the foundations of
mathematics~
The other goes up, onto the highlands of mathematics
itself" where these concepts are indispenBable in almost all of pure mathematics as it is today+ Xecdlcss to say1 we follow the latter course. We
regard sets and functions as tools of thought~ and our purpose in this
chapter is to develop these tools to the point v.here they are sufficiently
powerful to serve our needs through the rest of this book.
As the reader proceeds, he l\ill eome to understand that the words
set and fun..ction arc not as simple as they may seem. Jn a sense~ they
are simple; hut they are potent "~ords 1 and the quality of simplicity they
possess is that which lies on the far side of complexity~ 'They are like
seeds, 'vhich are primitive in appearance but have the capacity for vast
and intricate development~

1. SETS AND SET lNCLUSION

We adopt a naive point of view in our discussion of sets and assume

that the concepts of an element and of a set of elements a.re intuitively
clear~
By an element we mean an object or entity of some sort1 as~ for
example~ a positive integer~ a point on the real line ( = a real number)~
3

Topology

or a point in the complex plane ( = a complex number).. A set is a collection or aggregate of such elements, considered_ together or as a whole4
Some examples are furnished by the set of all even positive integers, the
set of all rational points on the real linei and the set of all points in the
complex plane who sc distance from the origin is 1 ( == the unit eirc le in the
plane). We reserve the word class to refer to a set of sets~ We might
speak1 for instance, of the class of all circles in a plan c (thinking of each
circle as a set of points). It will be useful in the work we do if we carry
this hierarchy one step further and use the term family for a set of classes4
One more remark: the words element, set, class, and family are not
intended to be rigidly fixed in their usage; we use them fluidly, to express
varying attitudes to'vard the mathematical objects and systems \Ve
study4. It is entirely reasonable, for instance, to thir:ik of a circle not as
a set of points, but as a single entity in itself, in which ease we might
justifiably speak of the set of all circles in a plane.
rhere are two standard .notations available for designating a particular set. Whenever it is feasible to do so, we can list its elements
bet,veen braces. Thus {1, 2, 3 \ signifies the set consisting of the first
three positive integers, {11 i, - lJ -i} is the set of the four fourth roots
of unity, and { + 1, + 3, + 5~ . ~ . } is the set of all odd integers. This
manner of specifying a set, by listing its elements, is unworkable in many
circumstances. We are then obliged to fall back on the 8econd method,
which is to use a property or attribute that characterizes the elements of
the set in question. If P denotes a certain property of elements, then
fx: P J stands for the set of all clements x for which the property P is
meaningful and true. For cxamplct the expression
1

~ x: x

is real and

irrational~,

which '\\'Te read the set of all x S:u.ch that xis real and irrational, denotes the
set of ul1 real numbers vrhirh cannot be written as the quotient of two
integers. The set under discussion contains all those elements (and no
others) which possess the stated property4 The three sets of numbers
described at the beginning of this paragraph can be written either way:

i 1, 2

and

is an integer and 0 < n < 4} ,

{1, i, -1 1 -i} = {z: z is a complex number and z' = 1},
{ 1, + 3 ~ + 5 ~ . . . } :;::= { n ~ n is an odd integer l .
1

3}

= {n : n

''le often shorten our notation.

For instance, the last two sets mentioned

might perfectly well be Vll'itten {z: z' = 1} and {n: n is odd i ~ Our purpose is to be clear and to avoid misunderstandings:t and if t.his can be
achieved with less notation~ so much the better~ In the sa1ne vein_ we can

Sets and F.unctions

write

and

the unit ci.rcle = {z ~ 'z~ ~ I},

the closed unit di'sc = {z: 'zl S 1},
the open unit disc ~ {z: ~z! < 1}
4

''re use a special system of notation for designating intervals of various

kinds on the real line. If a and b are real numbers such that a < bJ
then the following symbols on the left are defined to be the indicated sets
on the right:
[a, b]
(a, b]
[a,b)
(a~ b)

= {x : a < x
= {x : a < x
=

{x:a ~ x
{x : a < x

~
~

b} ,
b} ,

< bL
< b} .

We speak of these as the closed, the open-closedJ the closed-open, and

the open intervals from a to b. In particular, [O, I] is the closed unit
interval~ and (0, 1) is the open unit i'nte rval.
1.'"here are certain logical difficulties which arise in the foundations of
the theory of sets (see Problem 1). We avoid the~e difficulties by assuming that each discussion in 1vhich a number of sets are involved takes place
in the context of a single fixed set. This set is ca1led the universal set
It is denoted by U in this section and the next, and every set mentioned is
assumed to consist of elements in U.. In later ch~pters there will always
be on hand a given space within which we work, and this will serve
without further comment as our universal set+ 1 It is of ten convenient
to have available in U a set containing no elements whatever; we call this
the empty set and denote it by the symbol lt A set is said to be finite
if it is empty or consists of n elem en ts for some positive integer n; otherwise, it is said to be infinite.
We usually denote clements by small letters and sets by large letters+
If x is an element and A is a set~ the statement that x is an element of A
(or belongs to A, or is contained in A) is symbolized. by x e A.. We
denote the negation of this, namely, the statement that xis not an element
of A, by x ft A~
Two sets A and B are said to be equal if they consist of exactly the
same elements; we denote this relation by A = B and its negation by
A ~ B. We say that A is a subset of B (or is contained in B) if each element of A is also an element of B. This relation is symbolized by A kB.
We sometimes express this by saying that B is a mperset of A (or conThe words 8tt and space a.re of ten used in loose contrast to one a.no ther ~ A set i8
merely an amorphous collection of elements, without coherence or ronn.. '\\"hen some
kind of algebra.ie or geometric structure is imposed on a. .set, ao that its element.a a.re
organized in to a system.at ic whole,. t.hen it becomes e. space:.
1

Topology

tains A). A C B allows for the possibility that A and B might be equaL
If A is a subset of Band is not equal to B, we-say that A is a proper subset
of B (or is properly contained in B). This relation is denoted by A C B+
We can also express A C B by saying that B is a proper superset of A
(or properly contains A). The relation C is usually called set inclusion.
We sometimes reverse the symbols introduced in the previous paragraph. Thus A c B and A C Bare occasionally written in the equivalent forms B =:) A and B :J A .
It will of ten be convenient to have a symbol for logical implication~
and ~ is the symbol we useL .If p and q are statements, then p => q
means that p implies q, or that if pis true~ then q is also true~ Similarly,
~ is our symbol for two-way implication or logical equivalenceL
It
means that the statement on each side implies the statement on the other,
and is usually read if and only if, or is equitralent t(J.
The main properties of set inclusion are obvious. They are the
following:
(1) AC A for every A;
(2) A c B and B c A => A = B;
(3) A c B and B c C ~A c C.
It is quite important to observe that (1) and (2) can be combined into the
single statement that A :::;; B :::> A ~ B and B ~ A.. This remark
contains a useful principle of proofJ namely, that the only way to show
that two sets are equal~ apart from merely inspecting them~ is t-o show
that each is a subset of the other.
Problems

1..

Perhaps the most famous of the logical difficulties referred to in the

text is Russell1 s paradox.. To explain what this is, we begin by
observing that a set ran easily have elements which are themselves
setst e.g+, I 1, l 2,3 L 4}. This raises the possibility that a set might
well contain itself as one of its clements. We call such a set an
abnornial scti and any set which does not contain itself as an clement
we call a normal set+ Most sets are normal, and if we suspect that
abnormal sets arc in some way undesirable, we might try to ronfine
our attention to the set N of all normal sets. Someone is now sure
to ask, Is N itsell normal or abnormal? It is evidently one or the
other, and it cannot be both. Show that if ,,_,,.,.r is normal, then it
must be abnormal. Show also that if N is abnormalJ then it must
be normal. We see in this way that each of our two alternatives is
self~.ontradictory, and it seems to be the assumption that N exists
as a set which has.brought us to this impasse. Ii,or further discussion
of these matters~ we refer the interested reader to "\\7 ilder [42, p. ~5!l]

Sets and Functions

2.

3.

or Fraenkel and Bar-Hillel [ 10, p. 6]. Russell~ s own account of the

discovery of his paradox can be found in Russell [36, p . 75]~
The symbol we have used for set inclusion is similar to that used for
the familiar order relation on the real line: if x and y are real numbers,
x < y means that y - x is non~negativc+ The order relation on
the real line has all the pro pertics mentioned in the text:
(1') x ::; x for every x;
(2 ') x :S: y and y ::; x ~ x = y;
(3 f) x < y and y ::; z ==? x < z~
It also has an important additional property:
(4') for any x and y, either x < y or y < x+
Property (4') says that any two real numbers are comparable with
respect to the relation in questionJ and it leads us to call the order
relation on the real line a total (or linear) arder relation. Show by an
example that this property is not possessed by set inclusion. It is
for this reason that set inclusion is called a partial order relation.

(a)

(b)

(c)

(d)

Let U be the single-element set. { l}. There are two subsets,

the empty set 0 and {l l it.self. If _A_ and B arc arbitrary subsets
of U, there fU"e four posf;ible relations of the form A C B.
Count the nU;mber of true relations among these.
Let U be the set {1,2} ~ rrhere are four subsets+ List them .
If A and B are arbitrary subsets of U, there are 16 possible
relations of the form A ~ B~ Count the number of true ones .
Let U be the set { J~ 2, 3 l.. ~rhere a.re 8 subsets. What are
they? There are 64 possible relations of the form A c B.
Count the number of true oneR~
Let U be the set {1, 2~ ..
n} for an arbitrary positive
integer n. How many subsets are there? How many possible
relations of the form A ~ B are there? can you make an
informed guess as to how many of these are true?
+

2.. THE ALGEBRA OF SETS

In this section we consider several useful ways in which sets can be
combined with one another, and we develop the chief properties of these

operations of

combination~

As we emphasized above, all the sets we mention in this section are

assumed to be subsets of our universal set U. U is the frame of reference,
or the universe, for our present discussions. In our later work the frame
of reference in a particular context will naturally depend on what ideas
we happen to be considering. If \Ve find ourselves studying sets of real

Topology

numbers, then U is the set R of all real numben:t If we 'Wish to study

sets of complex numbers~ then v.re take U to be the set C of all complex
numbers. We sometimes want to narrow the frame of reference and to
consider (for instance) only subBets of the closed unit interval [O,l]J or
of the closed unit disc t z: !zl < 1 L. .and in these cases we choose U
accordingly. Generally speaking~ the universal set U is at our disposal,
and we are free to select it to fit the needs of the moment. For . the
present~ however, U is to be regarded as a fixed but arbitrary set. 1,his
generality a1lows us to apply the ideas we develop below to any situation
which arises in our later work.

Fig~

1~

Set inclusion.

14,ig. 2.

The union of A and B.

It is extremely helpful to the imagination to have a geometric

picture available in terms of which we can visualize sets and operations
on sets. A convenient way to accomplish this is to represent U by a
l"eetangular area in a plane, and the elements which make up U by the
points of this area. Sets can then be pictured by areas within this
rectangle, and diagrams can be drawn which illustrate operations on sets
and relations between them. For instance, if A and Bare setsJ then Fig.
1 represents the circumstance that A is a subset of B (we think of each
set as consisting of all points within the corresponding closed curve)
Diagrammatic thought of this kind is admittedly loose and imprecise;
nevertheless, the reader will find it invaluable. No mathematics, however
abstract it may appear, is ever carried on without the help of mental
images of some kind, and these a.re often nebulous,, personal, and difficult to describe .
The first operation we discuss in the algebra of sets is that of forming
unions. The union of two sets A and B, written AU B, is defined
to be the set of all elements which are in either A or B (including
those which are in both). AU B is formed by lumping together the
elements of A and those of B and regarding them as constituting a single
set. In Fig. 2.t A U B is indicated by the shaded areat The $hove
4

Sets and Functions

definition can also be expressed symbolically:

A V B ;:; { x : x e: A or x e B} .

The operation of forming unions is commutative and associative:

AVB=BUA

A U

and

(B U

C) = (A U B) V C.

It has the following additional properties:

A VA = A, A V 0 = A, and A V U = U.

We also note that

A~ B~AVB

B,

so set inclusion can be expressed in terms of this operation.

Our next operation is that of forming intersections. The intersection of two sets A and B, written A 0. B~ is the set of all elements which
are in both A and B. In symbols1

A f\ B = ix:x e: A and x

e B}.

A 0. B is the common part of the

sets A and B. In lt'ig. 31 A ('\Bis
represented by the shaded area. If
A n R is non-empt.y, we express thls
B
by saying that A infrsecta B. If,
on the other hand, it happens that A
and B have no common part, or
equivalently that A 0. B ~ 0, then Fig. 3. The in.terse ction of A and B ~
we say that A does not intersect B, or
that A and Bare disjoi.nt; and a class of sets in which all pairs of distinct
sets are disjoint is called a disjoint class of sets~ The operation of forming intersections is also commutati'7e and associative:
A('.B=Bf'.A

A 0. (B fl O) = (A fl B) ()

and

C~

It has the further properties that

A fl A;::;::; A , A fl

n.
'JI

= P,
n.

and An U =At

and since
A c B

;;>

A fl B = A,

we see that set inclusion can also be expressed in termB of forming

intersections.
We have now defined two of the fundamental operations on sets, and
we have seen how each is related to set inclusion4 The next obvious step
is to see how they are related to one another.. The facts here are given by

10

Topology

the di&tributive laws:

and

(B U C)
Av (B
C)
A

=
=

(A (\ B) U (A (\ C)
(Av B) n (Au C)~

These properties depend only on simple logic applied to the me.anings of

Fig. 4+

A U (B

C) - (A U B) f\ (A V C).

the symbols involved. For instance, the first of the two distributive
laws says that an element is in A and is in B or C precisely when it is in
A and B or is in A and C.. We can convince ourselves intuitively of the
validity of these laws by drawing pictures~ The second distributive
law is illustrated in Fig~ 4~ where A V (B 0 C) is formed on the left by
shading and (A U B) n (A V C) on the right by cross-shading.. A
momentJs consideration of these diagrams ought to convince the reader
that one obtains the same set in each
ease.
The last of our major operations
on sets is the foruiation of complements. The comple.ment of a set AJ
denoted by A' t is the set of all elements which are not in A. Since
the only elements we consider are
Fig. 5. The complement of Ar
those which make up U, it goes with..
out saying~but it ought to be said
~that A' consists of all those elements in U which are not in A.
Symbolically,
A 1 = {x:x J A}.
Figure 5 (in which A 1 is shaded) illustrates this operation. The operation
of forming complements has the following obvious properties:

(A')'
AU A'

= A, 0' == U, U' = JJ,

= U, and A ()A' = S.

Sets and functions

11

Further, it ia related to set inclusion by

~

$=::>

B 1 C A'

and to the formation of unions and intersections by

(A VB)' =

A'" B

and

(A fl B) 1

~ A~

U B'.

(1)

The first equation of (1) says that an element is not in either of two .sets
precisely when it is outside oi both:t and the second says that it is not in
both precisely when it i~ outside of one or th.e other.
The operations of forming unions and intersections are primarily
binary opPrations; that is, each is a process which applies to a pair of sets
und yields a third. V{e have emphasized this by our use of parentheses
to indicate the order in ,,... .hich the operations arc to be performed, as in
(A1 U A1) V A3i where the parentheses direct us first to unite Ai and
A 2', and then to unite the result of this v.ith A~- Associativity makes it
possible to dispense with parcn theses in an expression like this and to
write A1 V A2 U Aa, where we understand that these sets are to be
united in any order and that the order in which the operations are
performed is irrelevantL Similar remarks apply to Ai n As(\ A3r
Furthermore:1 if ~Al, A 2J
~
An l is any finite class of setsJ then we
can form
L

and
in much the same way without any ambiguity of meaning whatever.
In order to shorten the notation, we let l = ~ 1, 2, .... , n} be the set
of subscripts which index the sets under consideration. I is. called the
index set. We then compress the symbols for the union and intersection
just mentioned to U ~r Ai and (')i.I A-t. As long as it Li:; quite clear what
the index set is_, we can write this union and intersection even more
briefly~ in the form ViAt: and ni.._4;. For the sake of both brevity and
clarity, these sets are often written Vi,.. 1 A.,: and rl!. 1 Ai~
These extensions of our ideas and notations don't reach nearly far
enough. It is of ten necessary to form unions and intersections of large
(really large!) classes of setsT Let t A,\ be an entirely arbitrary class
of sets indexed by a set I of subscripts. Then

and

"141

= {x : x

e: A, for at least one i e: I}

Ai= {x:x e: Ai for every i el'

V ~r Ai

define their union and intersection~ As above, we usuaUy abbreviate these

notations to U..:A 1 and l'\iAi; and if the class {Ai} consists of a sequence
of sets, that is, if ~A,} = {A 11 A 2, A :i, ~ ~} J then their union and
intersection arc often written in the form U ;.. 1 A..: and ('\:_1 Ai. Observe
that we did not require the class [A.i} to be non-empty.. If it does

l2

Topology

happen that this class is emptyJ then the above definitions give (rememU. The
bering that all sets are subsets of U) V.A., = and rlla:
second of these facts amounts to the following statement; if we require
of an element that it belong to each set in a given class_, and if there are no
sets present in the class J then every element sa ti sfi es this requirement.
If we had not made the agreement that the only elements under consideration are those in U, we \vould not have been able to assign a meaning to
the intersection of an empty cla.ss of sets. A moment's consideration
makes it clear that Eqs. {I) are valid for arbitrary unions and intersections:

;:::=

and

(2)

It is instructive to verify these equations for the case in which the class
f Ai} is empty~
We conclude our treatment of the general theory of sets with a
brief discussion of certain special classes of sets which a re of con siderable importance in topology, logic, and measure theory. We usually

denote classes of sets by capital letters in boldface~

First, some general remarks which will be useful both now and
later, especially in connection with topological spaces~ We shall often
have occasion to speak of finite unionB and finite intersections, by which
we mean unions and intersections of finite classes of sets~ and by a
finite class of sets 've always mean one which is empty or consists of n
sets for some positive integer n. If we say that a class A of sets is closed
under the formation of finite unions_, we mean that A contains the
union of each of its finite subclasses; and since the empty subclass
qualifies as a finite subclass of A, we see that its uniont the empty set1
is necessarily an element of A. In the same way~ a class of sets which is
closed under the formation of finite intersections necessarily contains

the universal set~

Now for the special classes of sets mentioned above+

For the
remainder of this section we specifically assume that the universal set
U is non-empty~ A Boolean algebra of sets is a non-empty class A of
subsets of U which has the following properties:
(1) A and Be A=::::} A U B e A;
(2) A and B s A ~ A n B ! A-;
(3) A s A = } A' s AL
Since A is assumed to be non-empty, it must contain at least one set A.
Property (3) shows that A' is in A along with A_, and since A(\ A' = 0
and A U A' == U, (1) and (2) guarantee that A contains the empty set
and the universal set. Since the class consisting only of the empty set
and the universal set is clearly a Boolean algebra of sets:r these two
distinct sets are the only ones which every Boolean algebra of sets must

Sets and Functions

13

con ta.in. It is equally clear that the class of all subsets of U is also a
Boolean algebra of sets.. There are many other less trivial kinds, and
their applications are manifold in fields of study as diverse as statistics
and electronics~
Let A be a Boolean algebra of sets.. It is obvious that if
{A 1, At, . . . , An} is a non-empty finite subclass of A, then

and
are both sets in A; and since A contains the empty set and the universal
set, it is easy to see that A is a class of sets whlch is closed under the
formation of finite unions, finite intersections~ and complements. We
now go in the other direction, and Jet A be a class of sets which is closed
under the formation of finite unions, finite intersections, and complements. By these assumptions, A automatically contains the empty set
and the universal set, so it is non-empty and is easily seen to be a Boolean
algebra of sets. We conclude from these remarks that Boolean algebras
of sets can be described alternatively as classes of sets which are closed
under the formation of finite unions, finite intersections, and complements. It should be emphasized once again that when discussing
Boolean algebras of sets we always assume that the universal set is non ..
eDlpty.

One final comment. We speak of Boolean algebras of aets because

there are other kinds of Boolean algebras than those which consist of
sets, and we wish to preserve the distinction. We explore this topic
further in our Appendix on Boolean algebras .
Problems

1.

21

If {A..: l and {B;} are t\ro classes of sets such that {A~} ~ {B1 I ,
show that U1A1 ~ \J;B1 and r\jBj ~ fliAi.
The difference between two sets A and B, denoted by A - B ~is the set
of all clements in A and not in B; thus A ~ B = A fl B'~ Show
the following :

A - B

3..

=A

---. (A fl
(A - B) A - {B (A U B) A - (B V

B) = (A U B) - B;
C = A - (B U C);
C) == (A ~ B) U (A n C);
C ~ (A - C) U (B - C);
C) = (A - B) fl (A - C).

The symmetric difference of two sets A and B, denoted by A .6 B,

is defined by A AB ~ (A - B) V (B - A); it is thus the union of

14

Topology

their differences in opposite orders.

Show the f ollowi.ng:

A d (B l1 C) = (A a. B) d C;
A 6. 0 : : : : A ; . A 8 ~4. = 0;

AaB
A r\ (B Li C)

.4.

5.

6.

7.

BLlA;
(A r\ B) ~ (.ti 0. C) .
=

1\. ri11g of sets is a non-empty class A of sets such that if A and B a.re
in A'J then A ~ B and A /\ B are also in A. ShO\\r that A must also
contain the- empty set, A U B, and A - B. Show that if a nonempty class of sets contains the union and difierence of any pair of its
set.;:.;~ then it is a ring of sets. Shov{ that a Boolean algebra of sets is
a ring of sets.
Sho-w that the class of all finite subsets (including the empty set) of
an infinite set i~ a ring of sets hut. is not a Boolean algebra of sets+
Show that the class of a11 finite unions of closed-open interval8 on the
real line is a ring of sets but is not a Boolean algebra of setsT
Assuming that the universal set (I is non-empty, shoo,v that Boolean
algebras of sets can be described as rings of sets which contain U.

3. FUNCTlONS

J\1 any kinds of function f.; occur in topology~ in a great variety of

situations. In our \Vork "'-e ~hall need the full power of thP. general concept of a function)' and sin(e its modern meaning is much broader and
deeper than its cle mentary 1neaningt "~e di He uss this concept in considerable detail and develop its main abstract prop~rties.
Let us begin Vii'"it.h a brief i nspcction of some simple examples, Consider the elementary f nnction
y

== x2

of the real "\""ariablc x. \\'hat do "\Ve have in mind when we call this a
function a.nd say that. y is a function of x? In a nutshell_, we are drawing
attention to the fact. that each re.al number x ha.s linked to it a specific
real number y., "rhich can be calculated according to the rule (or law of
correspondence) given hy the formula. le have here a process "\vhich,
applied to any re.al number x, docH .something to it (squares it) to produce
another real number y (the square of x). Similarly,
y

x:i -..- 3x

and

are two other simple functions of the rea] variable. x, and each is given
by a rule in the form of an algebraic expression which specifics the exact
manner in 'vhich the value of y depends on the value of :r.

Sets and Functions

15

The rules for the functions we have just mentioned are expressed by
formulas. In general, this is possible only for functions of a very simple
kind or for those which are sufficiently import.ant to deserve special
symbols of their own.. Consider, for instance~ the function of the real
variable x defined as fol]O\"\fs: for each real number xJ vrritc x as an infinite
decimal (using the scheme of decimal expansion in which infinite chains
of 9~s are avoided~in whichJ for example, 74. is represented by 425000 ~ ...
rather than by .24999 .... ) ; then let y be the fifty-ninth digit after
the decimal point. There is of course no standard formula for this:P
but nevertheless it is a perfectly respectable function whose rule is given
by n verbal description. On the other hand, the function y = sin x of
the real Va.J4iab1e x is so important that its rule, though fully as complicated as the one just dcfinedJ is assigned the special symbol sin. When
discussing functions in gcneralt we \Vant to allow for all sorts of rules and
to talk about them all at once, so we usually employ noncommittal
notations like y = j(x), y = g(x), and so on.
Each of the functions mentioned above is defined for all real numbers
x. The ex.ample y = I/x shows that this restriction is much too severe,
for this function is defined only for non ... zero values of x. Similarly 1
y = log xis defined only for positive values of x~ and y = sin~t x only for
values of x which lie in the interval [ - I~ I] Wha tevcr our conception
of a function may bet it should certainly he broad enough to include
examples like these, which are defined only for some values of the real
L

variable x.

In real analysis the notion of function is introduced in the following

way.. Let X be any non-empty set of real numbers~ We say that a
function y = f(x) is defined on X if the rule f assoeiates a definite real
number y vrith eath real number x in X. The specific nature of the
rule f is totally irrelevant. to the concept of a function.. The set X is
called the domain of the given function:P and the set Y of all the values it
assumes is called its range. If we speak of complex numbers here
instead of real numberst l\~e have the notion of function as it is used in
.complex analysis.
This point of view toward functions is actually a bit more general
than is needed for the aims of analysis, but it isn't nearly general
enough for our purposes~ The sets X and Y above were taken to be
sets of numbers~ If we now remove even this restriction and allow X
and Y to be completely arbitrary non-empty setsJ then we arrive at the
most inclusive concept of a function+ By way of illustration, suppose
that X is the set of all squares in a plane and that Y is the set of all
circles in the same plane. We can define a function y == f(x) by requiring
that the rule f associate with each square x that circle y which is inscribed
in it. In general, there is no need at all for either X or Y to be a set of

16

Topology

numbent All that is really nece~sary for a function is two non-empty

sets X and Y and a i;-ule f which is meaningful and unambiguous in
assigning to each element x in X a specific element. yin Y~
With these preliminary descriptive remarks, v.:e now turn to the
rather abstract but very precise ideas they are intended to motivate.
A function consists of three objects: t\VO 11on-empty sets X and Y
(which may be equal, but need not be) and a rule f which assigns to each
element x in X a single fully determined element yin Y. The y which
corresponds in this \Vay to a given xis usuaJly 'vritten f(x)~ and is called
the image of x under the rule f, or the value off at the element x.

Fig. 6.

This

A \ray of visualizing mappings.

notation is supposed to be suggestive of the idea that the rule f takes the
element x and does something to it to produce the e1ernent y
/(x) .
The rule f is often called a mapping, or trans!ormation, or operator J to
amplify thls concept of it. We then think off as mapping x's to y's, or
transforming x's into y~s, or operating on x's to produce y'8~ The set X
is called the dmnain of the function, and the set of all f(x)'s for all x~s
in X iB called its range. A function whose range consists of just one
element is called a constant function.
We often denote by f: X-+ Y the function -..vith rule /, domain X,
and range contained in Y ~ 'fhis notation is uscf ul bee a use the esseu tia]
parts of the function are displayed in a manner vlhich emphasizes that it
is a composite object, the central thing being the rule or mapping f~
Figure 6 gives a convenient way of picturing this function. On the
left, X and Y are different sets, and on the right, they are equal~in which
case we usually refer to f as a mapping of X into itself. If it is clear
from the context what the sets X and Y are~ or if there is no real need to
specify them explicitly:- it is common practice to identify the function
f: X--+ Y v{ith the rule/, and to speak off alone as if it were the function
under consideration (without mentioning the sets X and Y).
It sometimes happens that two petfectly definite sets X and Y are
under discussion and that a mapping of X into Y arises which has no
natural symbol attached to it. If there is no necessity to invent a
;:::=

Sets and Functions

17

symbol for this mapping, and if it is quite clear what the map ping is, it is
often convenient to designate it by x ---t y. Accordingly, the function
y = x 2 mentioned at the beginning of this section can be written as
x ____,. x 2 or x - 4 y (where y is understood to be the square of x).
A function f is r.alled an extension of a function g (and g is called a
restriction of j) if the domain oi f contain8 the domain of g and f(x) = g(x)
for each x in the domain of g.
Most of mathematical analysis~ both classical and modern, deals
with functions whose values arc re~l numbers or complex numbers .

,-1
Fig~

7.

T'he j n verse of a mapping.

This is also true of those parts of topology- which are concerned with
the foundations of analysis4 If the range of a function consists of real
numbcrs 1 we call it_ a real function; similarly, a complex function is one
whose range consists of complex numbersa Obviously, every real function
is also complex. We lay very heavy emphasis on real and complex functions throughout our work.
As a matter of usage~ we generally prefer to reserve the term function
for real 9r complex functions and to speak of mappings when dealing
with fun;tions whose values are not necessarily numbers .
Consider a mapping /:X ~ Y. When we call f a mapping of X . .
into Y, we mean to suggest by this that the elements /(x)~as x varies
over all the clements of X-need not fill up Y; but if it definitely does
happen that the range off equals Y, or if we specifically want to assume
this, then we ca] 1 f a mapping of X onto Y. If two different elements
in X always have different images under f 1 then we call f a one-to-one
mapping of X into Y.. If f: X __., Y is both onto and one... to-onc 1 then
we can define its inverse mapping j- 1 : Y ~ X as follows: for each y in Y,
we find that unique element x in X such that /(x) = y (x exists and is
unique since f is onto and one-to-one) ; we then define x to be j~ 1(y).
The equation :c = 1- 1 (y) is the result of solving y = f(x) for x in just the
same way as x ==== log y is the result of solving y = e.z for .x. Figure 7
illustrates the concept of the inverse of a mapping4

18

Topology

If f is a one-to-one mapping of X onto Y~ it will sometimes be convenient to subordinate the conception of f as a mapping sending zts
over toy's and to emphasize its role as a link between x's and y}s.. Each
x has linked to it (or has corresponding to it) precisely one y = f(z);
and~ turning the situation aroundJ each y has linked to it (or has corresponding to it) exactly one x = f~ 1 (y). When we focus our attention
on th is asp~c t. of a mapping 'vhieh is onc-to--0ne onto, we usually call it
a nne-to-one correspondence~ Thus f is a one-to-one correspondence
between X and Y, ~i.nd 1~ 1 is a one-to-one correspondence between Y
and X.
Now consider an arbittary mapping f: X----? Y~ r'fhc mapping /,
which sends each clement of X over to an clement of Yi indu(~es the
following two important set mappings. If A is a subset of ..: thPn its
image f(A) is the subset of Y defined by
1

.f(A)

{f(x} :x AL

and our fir.st set mapping is that whi(' h sends each A over to its co rre
sponding f (A). Similarly~ if B i~ a subset of Y ~ then its inl.~erse inla(Jc
1~ 1 (B) is the subset of X defined by

J- 1 (B)

{ x :f(x) E.

BL

and the second set mapping pulls each B back to its corresponding
1~ 1 (B}+
It is often essential for us to know ho\V these set mappings
behave with respect to set inclusion and operations on sets. We develop
most of their significant features in the following t-..vo paragraphs.
The main properties of the first set mapping are:

/(0)

0;

f(X) C Y;
A 1 s; A 1 ::::::} f (A 1) C f (A~) ;
/(V,A1.) = UJ(Ai);

(I)

f(niAi) C rlif(At)
The reader should convince himself of the truth of these statements.
~For instancej to prove ( l) "\\,..e -y;rould have to prove first that f (U ,A 1) is a
subset of UJ(A1.), and second that UJ(Ai) is a subset of .f(U lAi).
A proof of the first of these sei inclusions might run as follows: an elcme n t
in /(UiA-.J is the image of some element in UiA1J therefore it i8 the image
of an element in some A..:t therefore it is in some /(A1), and so finally it is in
VJ(A-t). The irregularities and gaps which the reader will notice in the
above statements are essential features of this set mapping4 For example, the image of an intersection need not equal the intersect ion of the
images, because t \\i..o disjoint sets can easily have images whi cl1 are not

disjoint..

Furthermore~

without special assumptions (see Problem 6)

nothing can be said about the relation between /(~4) and /(A')~
1

Sets and Functions

The second set mapping is much better behaved.

are sa.tisfyi ngly complete~ and can be stated as fallows:
/ - 1(0) ~

1~ 1 (Y) =

O;

Its properties

X;

1-1(81) C 1~ (B2);
J- 1 (U,Bi) = Ujf-~ (B,);
J- 1 (rlJJi) = nif- (Bi);
J- 1 (B') = /~ (B)'.

B1 C

19

Bi;;;;}

(2)
(3)
(4)

Again, the reader should verify each of these statements for himself+
f

at
Fip;. 8.

:\! ultipJica.t)on of IIUtpping!'J.

We discuss one more concept in this sectiont that of the multiplication

(or co-m.posi,tion) of mappings. If y = f(x) = x 2
1 arid

g{y)

sin y,

then these two functions can be put. together to form a single function
defined by z = (gf) (x) = g(f(x)) ~ g(x 2
1) === sin (x 2 + I).. One of
the most important tools of calculus (the chain rule) explains how to differentiate functions of thls kind~ This manner of multiplying functions
together is of basic importance for us as well, and we formulate it in
general as follows.. Suppose that f: X ---7 Y and g: Y ~ Z are any two
mappings.. We define the prod1.tt-t of these mappings, denoted by
gf: X ~ Z~ by (g/){x) = g(f(x)). In words: an element x in X is taken
by f to the elementf(x) in Y~ and then g mapsf(x) to g(f(x)) in Z. :F~igure
8 is a picture of this process. We observe that the two mappings involved
here are not entirely arbitrary, for the set Y whieh contains the range of
the first equals the domain of the second. More generallyt the product
of two mappings is meaningf u) whenever the range of the first is contained in the domain of the second. We have regarded f as the first
mapping and g as the setondj and in forming their product gf, their
symbols have gotten turned around~ This is a rather unp1easant
phenomP.non, for which we blame the occasional pe-rversity of mathematical symbols~ Perhaps it will help the reader to keep this straight

20

Topotogy

in his mind if he will remember to read the product gf from right to left:
first apply f, then g.
Problems

Two mappings f: X ~ Y and a: X ~ Y are said to be equal (and we

vrrit-0 this f == g) ii f(x) = g(x) for every x in X. I.Jct f, Ut and h
be any three mappings of a non-empty set _)( into itself, and show
that multiplication of mappingR is associative in the sense that
f(gh) = (fg)h+
2t Let X be a non-empty set. The i"dentity mapping ix on X is the
mapping of X onto itself defined hy ix(x) = x for every x~ Thus ix
sends each clement of X to itself; that isJ it. leaves fixed each clement
of X. Show that fix :;::= ixf = f for any mapping f of X into itself.
If f is one-to-one onto~ so thut its inverse 1- 1 exists~ show that
11- 1 = 1~ 1! = ix~ Show further that 1-1 is the only mapping of _J{
into itself whir.h has this property; that isj sho,v that if g is a mapping
of X into it..self such that f g := a! = ix, t.hen g = J- 1 (h.int:
g = gix = g(ff- 1 ) ~ (gf)J~ = ixf- 1 === J- 1 ~ or

1~

g ~ ixg

3.

let X and Y be

= (J- 1J)g

non~mpty

J- 1 (jg) == f~

ix ~

J- 1}.

sets and.fa mapping of X into Yr

Shfnv

the f ollo\vi ng:

(a) f is one-to-one <=:::} there exists a mapping g of Y into X such
that gf = ix;
(b) f is on to {::::> there exists a 'mapping h of Y in to X sue h that

fh
4.

S.

6.

iv~

Lt X he a non-empty set and f a mapping of X into

it~elf.

Show
that f is one-to-one onto ~there exists a mapping g of X into itself
such that jg = gf = ix+ If there exists a mapping g with this
property, then there is only one sur.h mapping. Why~?
Let X be a non-empty set, and let f and g be onc~to-one mappings
of X onto itselL Show that Jg is also a one-to-one mapping of X
onto itself and that (fg)- 1 = g- 11- 1
Let X and Y be non-empty sets and fa mapping of X into Y. If
A n.nd B are, respectively, subsets of X and Y~ show the follo,ving:
(a) Jf- 1 (B) CB, and ff~ 1 (B) ~ Bis true for all B ~ f is onto;
(b) A C J- 1f(A), and A == f- 1f(A) is true for all A <=> f is one-to-one;
(c) f(A1 r\ A1) == f(A1) fl f(A2) is true for all A 1 and A2 ~ f is
one-to-one ;
(d) f(A) 1 f; f(A 1 ) is true for all A ~ f is onto;
(e) if f is onto-so t.hat f(A)' ~ f(A') is true for all A-then
f(A )' = f(A ') is true for all A <:;;;> f is also one-to-one.

Sets and Functions

21

A. PRODUCTS OF SETS
We shall often have occasion to \\:oeld together the sets of a given
class into a single new set called their product (or their Cartesian product).
The ancestor of this concept is the coordinate plane of analytic geometry,
that is, a plane equipped with the usual rectangular coordinate sys tern.
We give a brief description of this fundamental idea with a view to
paving the way for our discussion of products of sets in general..
First, a few preliminary comments about the real line. "\\,.c have
already used this term several timeH \\~ithout any explanation, and of
course what we mean by it is an ordinary geometric straight line (see
Fig. 9) whose points have been identified 'vith---or coordinatizcd by~the

-3

-2

-1

-1/3

Fig. 9.

V2

1/2

'IT

)r

The real line.

set R of all real numbers. We use the letter R to denote the real line
as well as the set of all real numbers, and we of ten speak of real numbers
as if they were points on the real line, and of points on the real line as if
they were real numbers. Let no one be deceived into thinking that the
real line is a simple thing, for its structure is exceedingly intricate. Our
present view of it~ however, is as naive and uncomplicated as the picture
of it given in ~,ig . 9. Generally speaking, 'vc assume that the reader is
familiar with the simpler properties of the real line--those relating to
inequalities (see Problem 1-2) and the basic algebraic operations of
addition, subtraction, multiplication, and division. One of the most
significant facts about the real number system is perhaps less well
known. This is the so-called least upper bound property, \V hi ch asserts
that every non-empty set of real numbers which has an upper bound has
a least upper bound~ It is an easy consequence of this that every nonempty set of real numbers which has a lo\ver bound has a greatest lower
bound. Al1 these matters can be developed rigorously on the basis of a
small uumber of axioms, and detailed treatments can of ten be found in
books on elementary abstract algebra.

To construct the coordinate plane, we now proceed as follows. We

take two identical replicas of the real line which \Ve cal 1 the x axis
and the y axis, and paste them on a plane at right angles to one another
in such a wa.y that they cross at the zero point on each. rfhe usual
picture is given in Fig. 10. Now let P be a point in the plane. We
project P perpendicularly onto points P% and Py on the axes~ If x and y
are the coordinates of P::i: and Pv on their respective axes, this process
j

22

Topology

leads us from the point P to the uniquely determined ordered pair (x y)

of real numbers, where x and y are <alled the x coordinate and y eoordinate of P. We can rc1.rerse the proress, andJ starting with the ordered
pair of real number8, 've can recapture the point. This is the manner in
i..vhieh we establish t.he fami]iar one~to-one correspondence between
points P in the plane and ordered pairs (x,y) of real numbers.. In fact~
\\re think of a point in the plane (\vhich is a geometric object) and its
rorresponding ordered pair of real numbers (which is an algebraic
objert) as being-to all intents and purposes-identical wi.t.h one another.
The essence of analytic geometry lies
in the possibility of exploiting this
y axis
identification by using algebraic
tools in geometric arguments and
giving geometric interpretations to
algebraic calculations.
- - - - -..P.!!!!!:f ~1Y)
I
The conventional attitude toI
I
\Vard the coordinate plane in ana~
I
I
lytir, geometry is that the geometry
I
is the focus of interest and the alge. .
1
I
bra of ordered pairs is only a con-0
~ axis
vcni en t too 1. llere \\l'e re verse this
point of view. :For us, the coordinate
Fig. 10. T11 e coor d innt c plane.
plane is defined to be the set of all
ordered pairs (xJy) of real numbers.
We can satisfy our desire for visual images by using :F'ig. 10 as a picture
of this set and by cailing such an o rdcred pair a point, but this geometric language is more a convenience than a necessity~
Our notation for the coordinate plane is R X R, or R 2 ~ This
symbolis1n reflects the idea that the coordinate plane is the result of
Hmultiplying together" two replicas of the real line R.
It is perhaps necessary to comment on one possible source of misunderstanding. When we speak of R 2 as a planeJ we do so only to
establish an intuitive bond with the reader's previous experience in ana~
lytic geometry. Our present attitude is that Ri is a pure set and has no
structure whatever, because no structure has yet been assigned to it.
We remarked earlier (with deliberate vagueness) that a space is a set
to which has been added some kind of algebraic or geometric structure.
In Sec+ 15 we shall convert the set R 2 into the space of analytic geometry
by defining the distance between any t\vo points (x1,y 1) and (x2,yi) to be
1

.,,,/ (x1 ._ X2) 2

(Y1 - Y~) 1

This notion of di.stance endows the set R 1 with a certain "spatiar' character~ which we shall recognize by calling the resulting space the Eucl.idean
plane instead of the coordinate plane.

Sets and Functions

23

We assume that the reader is fully acquainted with the way in

v.:hich the set C of all complex numbers can be identified (as a set)
with the coordinate plane R 2 .. If z is a complex number, and if z has
the standard form x
iy where ::t and y are real numbers, then we
identify z 'With the ordered pair (x,y)t and thus 'With an element of R 2
~rhe complex numbers, however, are much more than merely a set.
They constitute a number systemJ 'With operations of addition, multiplication, conjugation~ etc. When the coordinate plane ll 2 is thought
of as consisting of complex numbers and is enriched by the algebraic
st.rue ture it acquires in thi~ way,
it is called the complex plane,, The x2
letter C is u~d to denote either
the set of all complex numbers or
the complex plane. We shall make
~2
a space out of the complex plane
in Sec. 9.
Suppose now that X 1 and X 2
are any two non-empty sets. By
analogy with our above discussion,
their product X 1 X X 1 is defined to
xi
Xi
be the set of all ordered pairs (x1,Xt)~
Fig. 11. A way of visualizing X1 X X !
where x1 is in X1 and X2 is in X2.
In spite of the arbitrary nature of X1 and X2~ their product can be represented by a picture (see Fig. 11) which is loosely similar to the usual
picture of the coordinate plane. The term product is applied to this sett
and it is thought of as the result of "mu] ti plying to gcth cru X 1 and X ~, for
the following reason: if X 1 and Xi are finite sets vii.th m and n elementst
then (clearly) X1 X Xi has mn elements~ If /:X1 ~ X2 is a mapping
'"ith domain X1 and range in X"2, its graph is that subset of X1 X X2
which consists of all ordered pairs of the form (x1,f(x1))~ We observe
that this is an appropriate generalization of the concept of the graph of
a function as it occurs in elementary mathematics.
This definition. of the product of two sets extends easily to the case
of n sets for any positive integer n,, If Xh X2, . , . , Xn are non-empty
sets, then their product X 1 X X 2 X X Xn is the set of all ordered
n-tuples (xh Xi, ~ . . , xfl), where Xi. is in X1 for each subscript i. If the
X/s are all replicas of a single set X, that is, if

X1 = X2

= X'R.

= X,

then their product is usually denoted by the symbol Xn .

These ideas specialize directly to yield the important sets Ri1 and
Cn~ R 1 is just Rj the real lineJ and R 2 is the coordinate plane,, R 1-the
set of all ordered triples of real numbers-is the set which underlies
solid analytic geometry, and we assume that the reader is familiar with

24

Topology

the manner in which this set arises, through the introduction of a rectanguJru- coordinate system into ordinary three-dimensional .space. We
can draw pictures here just as in the case of the coordinate plane_, and we
can use geometric language as much as we please, but it must be understood that the mathematics of this set is the mathematics of ordered
triples of real numbers and that the pictures arc merely an aid to the
intuition~ Once we fully grasp this point of view, there is no difficulty
whatever in advancing at once to the study of the set Rft of all ordered
n-tuples (x1~ xf, . . . 1 x,q) of real numbers for any positive integer n.
It is quite true that when n is greater than 3 it is no longer possible to
draw the same kinds of intuitively rich pictures_, but at worst this is
mere]y an inconvenience. We can (and do) continue to use suggestive
geometric language1 so all is not lost.. The set Cn. is defined similarly:
it is t.he set of all ordered n-tuples (z1, z2, .... , Zn) of complex numbers+
Ea ch of the sets RR and en plays a prominent part i Il our later lYOr k.
We emphasized above that for the present the coordinate plane is to
be considered as merely a set, and not a space. Similar remarks apply
to R~ and c~.. In due course {in Sec. 15) we shall impart form and
con tent to each of these sets by su.i table definitions.. We shall convert
them into the Euclidean and unitary n-spaces whlch underlie and motivate
so many developments in modern pure mathematics, and we shall
explore some aspects of their algebraic and topological structure to the
very last pages of this book. But as of now-and this is the point we
insist on-neither one of these sets has any structure at all.
As the reader doubtless suspectsJ it is not enough that we consider
only products oi finit.e classes of sets.. 1~he needs of topology compel
us to extend these ideas to arbitrary classes of sets.
We defined the product X1 X X! X . . X Xn to be the set of
all ordered n-tup}es (X1 1 Xi, . , Xn) SUCh that X..: is in X~ for each
subscript i. To see how to extend this definition,, we reformulate it as
follows.. We have an index set /, consisting of the integers from 1 to n,
and corresponding to each index (or subscript) i we have a non-empty
set Xi. The n-tuple (x1 1 x2, ...... ~ xn) is simply a function (call it x)
defined on the index set I~ with the restriction that its value x(i) = Xi is
an element of the set X, for each i in /. Our point of view here is that
the function xis completely determined by, and is essentially equivalent
to, the array (x1, x2, ..... J Xn) of its values.
The way is now open for the definition of products in their full
generality. Let {X;.} be a non-empty class of non-empty sets, indexed by
the elements i of an index set I. The sets Xi need not be different
from one another; indeed, it may happen that they are all identical
replicas of a single sett distinguished only by different indices. The
product of the sets Xi._, written P 1J X,11 is defined to be the set of all
functions x defined on I such that z(i) is an element of the set Xl for

Sets and Fune ti ons

25

each index i. We call X 1 the ith coordinate set. When there can be no
misunderstanding about the index sct 1 the symbol P1J
is often abbrevia tcd to P i.X ,:~ The defini tiot1 we have just given requires that each
coordinate set be non-empty before the product can be formed. It vtill
be useful if we extend this definition slightly by agreeing that if any of
the X/s ar.e empty 1 then P1X" is also empty.
This approach to the idea of the product of a class of sets~ by means of
functions defined on the index setj is useful mainly in giving the definition .
In practice~ it is much more convenient to use the subscript notation
x, instead of the function notation x(i). We then interpret the product
p x~ as made up of elements x~ each of which is specified by the exhibited
array {x..:} of itB" values in the respective coordinate sets Xi. We call
Xi the ith co01dinate of the element x = {x,,}.
The mapping p1 of the product PX" onto its ith coordinate set X1
which is defined by Pi(x) = xi-that is~ the mapping whose value
at an arbitrary element of the product is the ith coordinate of that element
~is called the projection onto the ith coordinate set. The projection
p, selects the ith coordinate of each element in its domain. There is
clearly one projection for each element of the index set I, and the set of
all projections plays an important role in the general theory of topological
spaces.

x.

Problems
1.

2.

The graph of a mapping f :X -io Y is a subset of the product X X Y.

What properties characterize the graphs of mappings among all
subsets of X X Y?
Let X and Y be non-empty sets. If Ai and A2 are subsets of XJ and
B1 and B 2 subsets of Y, show the following:

(A1 X B1) n (A2 X B,.) ::::: (A1 r\ A2) X (B1 r\ B2);

(A1 X B1) - (A1 X Bi) = (Ai ~ A2) X (B1 - B2)
V (A1 n Ai)'X (Bi - B2)
V (A1 ~ A2) X (B1 ~ B'j).

3..

Let X and Y be non-empty sets, and let A and B be rings of subsets

of X and Y, respectively~ Show that the class of all finite unions of
sets o! the form A X B with A E. A and B e Bis a ring of subsets of
XX Y.

5. PARTITIONS AND EQUIVALENCE RELATIONS

In the first part of this section we consider a non-empty set X, and
we study decompositiona of X into non-empty subsets which fill it out

26

Topology

and have no elements in common with one another.. We give special

attention to the tools (equivalence relations) which are normally used to
generate such decompositions.
A partition of X is a disjoint class {Xi} of non-empty subsets of X
whose union is the full set X itself~ The Xls are called the partition sets~
E-xpressed somewhat differently, a partition of Xis the result of splitting
it, or subdividing it, into non.-empty subsets in such a way that each
element of X belongs to one and only one of the given subsets.
If X is the set {1, 2 ~ 3, 4, 5} , then {1, 3, 5 }~ {2 ~ 4 i and {1, 2, 3 }, {4, 5}
are two different partitions of X.. If X is the set R of all real numbers~
then we can partition X into the set of all rationals .and the set. of all
irrationals_, or into the infinitely many closed-open intervals of the form
[nj n + 1) where n is an integer~ If X is the set of all point.8 in the
coordinate plane1 then we can partition X in such a -.,vay that each
partition set consists of all points with the same x co or d inatc (vertical
Jines)j or so that each partition set consists of all points with the &~me
y coordinate (horizontal lines).
Other partitions of each of the~ sets l\rill readily occur to the readc r.
In generalt there are many different ways in which any given set can
be partitioned. These manufactured examples are admittedly rather
uninspiring and serve only to make our ideas more concrete~ I_jater in
this section we consider some others which are more germane to our
present purposes.
A binary relation in the set X is a mathematical symbol or verbal
phrase, which we denote by R in this paragraph_, such that for each
ordered pair (x,y) of elements of X the statement x Ry is meaningful~
in the sense that it can be classified definitely as true or false. For such
a binary relation, x R y symbolizes the assertion that x is related by R to
y" and x Jjt y the negation of this, namely~ the assertion that x is not
related by R to y. Many examples of binary relations can be given
some familiar and others less so, some mathematical and others not~
:For inst.anre, if X is the set of all integers and R is interpreted to mean
uis Jess than/ 1 which of course is usually denot-ed by the symbol <, then
we clearly have 4 < 7 and 5< 2r We have been speaking of binary
relations, which are so named because they apply only t.o ordeted pairs
of elements~ rather than to ordered triples, etc~ In our work we drop the
qua]ifying adjective and speak simply of a relation in X, since we shall
have occasion to consider only relations of this kind~ 1
We now assume that a partition of our non .. empty set X is given,
1

So me -Trit e:rs prefer to rega:rd a relation R in X aa a subset

of X X X. From
t.hiN point of view, x Ry and :t $. fJ are shnply equivalent ways of writing (x)y) e R
and (x~y) t R. This definition has the advan tsge of being more tangible than ours.
and the disadvantage th&t few people really think of a relation in this ws.y.
t

Sets and functions

27

and 've associate with this partition a relation in X. This relation is

de fined in the following way: 've say that x is eq:uivalent to y and write
th is x ~ y (the symbol
is pronounced u 'WiggleH) ~ if x and y be1ong to
the same partition set. It is obvious that the relation,......; has the following properties~
t-.,.J

(1) x ~ x for every x (reflexivity);

(2) x rov y -:;::::} y ~ x (gym metry) ;
(3) x rov y and y ~ z ;::::;) x ....v z (transitivity).
This particu1ar relation in X arose in a special way, in connection with a
given partition of X, and its properties are immecliate consequences of its
definition~ Any relation whatever in X which possesses these th.Iee
properties is called an equivalence relation in X.
We. have just seen that each partition of X has associated with
it a natural equivalence relation in X. We now reverse the situation
and show that a given equivalence relation in X determines a natural
partition of X.
Let ~ be an equivalence relation in X; that is} assume that it is
reflexive, symmetric, and transitive in the sense described above. If z
is a.n element of x~ the subset of
defined by [x] = {y: y ~ x} is called
the equivalence set of Xa The equivalence set of x is tb us the set of all
elements which are equivalent to x:. We show that the class of all
distinct equivalence sets forms a partition of X. By reflexivity~ x e: [x]
for each element x in X, so each equivalence set is non-empty and
their union is X. It remains to be shown that any two equivalence sets
{x1] and [x2] are either disjoint or identical. We prove this by showing
that ii [x1] and rx2] are not disjoint, then they must be identical. Suppose that [x1] and [x2J are not disjoint; that is~ suppose that they have
a common element z. Since z belongs to both equivalence sets, t ~ Xi
and 2 ~ x2J and by symmetry, x1 ~ z. Let y be any element of [x1],
so that y ~ x 1 ~ Since y ~ x1 and x:1 ~ z1 transitivity shows that
y ~ z. By another application of transitivity, y ,_,_, z and z "'-' Xt imply
that y ~ X2~ 80 that y is in rx~1r Since y was chosen arbitrarily in [x1], we
sec by this that Ix 1] ~ [x 21 The same reasoning shows that [x:2] ~ lx1J, and
from this we conclude (see the last paragraph of Sec. 1) that [x:J = [x2J.
The above discussion demonstrates that there is no real distinction
(other than a difference in language) between partitions of a set and
equivalence relations in the set~ If we start with a partition, we get an
equivalence relation by regarding elements as equivalent if they belong
to the same partition set, and if we start with an equivalence relation,
we get a partition by grouping together i.nto subsets all elements which
are equivalent to one another. We hav~ here a single mathematical
idea, which we have been considering from two different points of view,
and the approach we choose in any particular application depends entirely

28

Topology

on our own convenience~ In practicej it is almost invariably the case that

\\--e use equivalence relations (which are usually easy to define) to obtain
partitions (which are sometimes difficult to describe fully).
We now turn to several of the more important simple examples of
equival en cc rela t.io n s.
I..et 1 be the set of all integers. If a and b are clements of this sett
"\\~e write a = b (and say that a equals b) if a and b are the same integer.
Thus 2
3 = 5 means that the expressions on the left and right are
simply different ways of writing the same integer. It is apparent that =
used in this sense is an equi vale nee relation in the set I :
(1) a = a for every a;
(2) a = b ~ b ~ a;
(3) a = b and b = c ~ a = c.
Clearly~ each equivalence set consists of precisely one integer .
Another familiar example is the relation of c;quality commonly used
for fractions. We remind the reader that, strictly speaking~ a fraction
is merely a symbol of the form a/b, where a and b are integers and b is
not zero. The fractions % and % are obviously not identical, but
nevertheless we consider them to be equal. In general, we say that
two fractions a/b a.nd c/d are equal., written a/b = c/d, if ad and be are
equal a.s integers in the usua.1 sense (see the above paragraph). We leave
it to the reader to show that this is an equivalence relation in the set of
all fractions. An equivalence set of fractions is what we call a rational
number.. F:veryday usage ignores the distinction between fractions and
rational numbers~ but it is important to recognize that from the strict
point of view it is the rational numbers (and not the fractions) which
form part of the real number system.
Our final example has a deeper significance, for it provides us with
the basic t.ool for our work of the next two sections.
For the remainder of this section we consider a relation between
pairs of non-empty sets, and each set mentioned (whether we say so
explicitly or not) is assumed to be non-empty. If X and Y are two
sets, we say that Xis numerically equivalent to Y if thcr-c exists a one-toone correspondence between X and Y, i.e., if there exist8 a one-to-one
mapping of X onto Y. This relation is reflexive, since the identity
mapping ix: X ___,. Xis one-to-one onto; it is symmetric~ since if f :X---. Y
is one-to-one onto~ then its inverse mapping J- 1 : Y ~ X is also one-to-one
onto; and it is transitive, since if f: X --io Y and g: Y ----t Z are one-to-one
onto, then gf :X ~ Z is also one-to-one onto~ Numerical equivalence has
all the properties of an equivalence relation, and ii we consider it as an
equivalence relation in the class of all non-empty sub.sets of some universal
set u~ it groups together into equivalence sets all those subsets of u
which have the same number of elements. After we state and prove the

Sets and functions

29

fallowing very useful but rather technical theorem, we shall continue in

Secs. 6 and 7 with an exploration of the implications of these ideas.
The theorem we have in miitd-the Schroeder-Bernstein theorem-is
the following: if X and Y are two sds each of which is numerically eq1livalenl
lo a subset of th.e other~ then all of X is numerically cqu.ivalenl to all of Y.
There are several proofs of this classic theorem, some of \Vhir.h arc quite
difficult. T'he very elegant proof '"Te give is essentially due to Birkholf
and MacLane.
Now for the proof. We assume that f :X ~ Y is a one-to-one
mapping of X into Y, aud that g:Y----t X iB a one-to-one mapping of Y
into X. Our task is t-0 produce a mapping F:X ~ Y which is one-to-one
onto~
We may assume that neither f nor g is onto~ since if f is, we can
define F to be f 1 and if g is_, we can define F to be g~ 1 Since both f and g
are one-to-one_, it is permissible to use the ma.ppings 1- 1 and g- 1 as long
as we clearly understand that /- 1 is defined only on f (X) and g- 1 only on
g(Y)... We obtain the mapping F by splitting both X and Y into subset~
which we characteriz c in terms of the ancestry of their elements. I iet x
be an element of X. We apply g- 1 to it (if we can) to get the element
g- 1 (x) in Y. If g-- 1 (x) exist8, 've call it the first anrestor of x. The e1emcnt x i tgelf we eall the zeroth ancestor of x. "'\\re now apply 1- 1 to
y- 1 (x) if we ean~ and if (f- 1g- 1) (x) exists, we call it the second ancestor
of x. We now apply y- 1 to (f- 1g- 1) (x) if we can, and if (g- 1f- 1g- 1 ) (x)
exists,, we call it the third ancestor of xr As we continue this process of
tracing back the ancestry of x 1 it becomes apparent that there are three
possibilities. ( l) x has infini tcly many ancestors. We denote by X,
the subset of X which consists of all elements with infinitely many
ancestors~
(2) x has an even number of ancestors; this means that x
has a last ancestor (that is_, one which itself has no first ancestor) in x
We denote by X ~ the su hf::et of X consisting of all elemen ts ~Tith an even
number of ancestors. (3) x has an odd number of ancestors; this
means that x has a last ancestor in Y. We denote by X(J the subset of X
which consists of all elements with an odd number of ancestors. 'The
three sets Xi~ X~, Xo fonn a disjoint class whose union is X. We decompose Yin just the same way int-0 three subsets Yi, Ye, Ye. It is easy to
see that f maps Xi onto Y;: and X~ onto Y~, and that !(r 1 ma.ps Xo onto
Y~; and we complete the proof by defining Fin. the following piecemeal
manner:
r

if x X;: V X~:P
if x E Xo.

\Ve attempt to illustrate these ideas in Fig. 12. Jierc we present. t\Yo
replicas of the situation: on the 1efti X and Y arc represented by the
vertical lines, and f and g by the lines slanting down to the right and

30

Topology

left; and on the right, we schematically trace the ancestry oi three e}e . .
ments in X~ of which x1 has no first ancestor_, x2 has a first and second
ancestor~ and xa has a first_, second_, and third ancestor.

/(X)

g(Y)

FigT 12.

The proof of the Schroed!'r -llt"r rust cin t hcorem.

The Schroeder-Bernstein theorem has great theoretical and practical

significance. Its main value for us lies in its role as a toul by means of
which we can pr-ove numerical equivalence with a minimum of effort for
many specific sets. We put it to work in Sec. 7.
Problems
1..

Y be an arbitrary mapping~ Define a relation in X as

follows: X1
Xi means that f(x1) = f(x:z). Show that this is an
equivalence relation and describe the equivalence sets.
In the set R of all real numbers, let x
y mean that x - y is an
integer.. Show that this is an equivalence relation and describe the
equivalence sets.
Let I be the set of all integers_, and let m be a fixed positive integer.
Two integers a and b are said to be congruent modulo m-symbolized
by a ~ b (mod m)-if a - bis exactly divisible by m, Le.j if a - bis
an integral multiple of m. Sho\v that this is an equivalence rclationr
describe the equivalence sets, and state the number of distin(~t
equivalence sets.
Decide which ones of the three properties of reflexivity, symmetry,
and transitivity are true for each of the following relations in the set

Let f; X

r.J

2..

3t

4,

1-..t

Sets end Functions

5..

6.

31

of all positive integers: m < n~ m < n~ m divides n. Are any of

these equivalence relations?
Give an example of a relation which is (a) reflexive but not symmetric or transitive; (b) symmetric but not reflexive or transitive;
(.c) transitive but not reflexive or symmetric; (d) reflexive and
symmetric but not transitive; (e) re:O.exi vc and transitive but not
symmetric; (!) symmetric and transiti vc but not reflexive.
Let X be a non-empty set and ,......, a relation in X. The following
purports to be a proof of the statement that if this relation is symmetric and transitive, then it is necessarily reflexive: x ~ y ==> y ~ x;
x ~ y and y
x::::::} x ~ x; therefore x ,.._. x for every x~ In view
of Proble~ 5f, this cannot be a valid proof.. What is the fl.aw in the
reasoning?
Let X be a non-empty set. A relation ~ in X is called circular if
x ~ y and y
z ~ z '"'"' x, and triangular if x ,.._, y and x
z => y ,.._, z.
Prove that a relation in X is an equivalence relation :::;) it is refi.exi ve
and circular {::::} it is reflexive and triangular.
,.-..,_i

7..

,.-..,_i

,.-..,_i

6. COUNTABLE SETS
The subject oi this section and the next-infinite cardinal numberslies at the very foundation of modern mathematics. It is a vital instrument in the day-to-day work of many mathematicians, and we shall
make extensive use of it ourselves.. This theory, which was created by
the German mathematician Cantor, also has great aesthetic appeal, for
it begins with ideas of extreme simplicity and develops through natural
stages into an elaborate and beautiful structure of thought. In the
course of our discussion we shall answer questions which no one before
Cantor's time thought to askJ and we shall ask a question whlch no one
can answer to this day .
Without further ado, we can say that cardinal numbers are those used
in counting~ such as the positive integers (or natural nlUllbers) 1, 2,
a, . . iamiliar to us alt But there is mueh more to the story than this.
The art of counting is undoubtedly one of the oldest oi human
activities. Men probably learned to count in a crude way at about the
same time as they began to develop articulate speech~ The earlies~ men
who lived in communities and dome!3ticated animals must have found
it necessary to record the number of goats in the village herd by means of
a pile of stones or some similar device. If the herd was counted in each
night by removing one stone from the pile for each goat accounted forJ
then stones left over would have indicated strays, and herdsmen would
have gone out to search for them. Names for numbers and symbols for
I

32

Topology

them, like our 1, 2, 3, .... , would have been superfluous~ The simple
and yet profound idea of a one-to-one correspondence between the
stones an<l the goats would have fully met the needs of the situation.
In a manner of speaking~ we ourselves use the infinite set

1V = { I, 2, 3i . . ~ }
of all positive integers as a Hpile of stones." We carry this set around
with us as part of our intellectual equipmentr Whenever we want to count
a set, say, a stack of dollar bills~ we start through the set l'l and t..ally off
one bill against each positive integer a.s we come to iL The htst.. rtumber
we reach, corresponding to the 1ast bill, is what we rnll the 11 umber of
bills in the stack. If this last number happens to be 10, t.hen i~ 10" is
our symbol for the number of bills in the stack, as it also is for the number
of our fingers, and for the number of our toes~ and for the number of
elements in any set which can be put into one-to-one correspondence
with the finite set I 1~ 2, . . . , 10} . Our procedure is slightly more
sophisticated than that of the primitive savage. We have the symbols
1~ 2, 3~ . . .- for the numbers which arise in counting; we can record
them for future use, and communicate them to other ~oph~, and manipulate them by the operations of arithmetic. But the underlying idea,
that of the one-to-one correspondence, remains the same for us as it
probably was for him.
The positive integers are adequate for the purpose of counting any
non-empty finite set, and since outside of ma thPmat.ics all sets appear to
be of this kind, they suffice for all non-mathematical counting. But in
the world of mathematics we are obliged to consider many infinite sets,
such as the set of all positive integers itself, the set of all integers~ the
set of all rational numbers~ the set of all real nu mhers, the set of all
points in a plane, and so on. It is of ten import.ant to be able to count
such sets~ and it was Cantor,s idea to do this, and to develop a theory of
infinite cardinal numbers, by means of one-to-one correspondences~
In comparing the sizes of two sets1 the basic concept is that of
numerical eq ui valen re as defined in the previous scctionr We recall
that two non-empty sets X and Y are said to be numerically equivalent if
there exists a one-to-one mapping of one onto the other, or-and this
amounts to the same thing-if there can be found a one-to-one correspondence betlveen them. To say that two non-empty finite sets are
numerically equivalent is of course to say that they have the same number
of element8 in the ordinary sense. If we count one of them, we simply
establish a one-to-one correspondence between its clements and a set of
positive integers of the form I 1~ 2, .. ~ , n L and we then say that n is
the num.1Jer of elements possessed by bothi or the cardinal numbet of both.
The positive integers are the finite cardinal numbers. We encounter

Sets and Functions

33

many surprises as we follow Cantor and consider numerical equivalence

for infinite sets.
The set N = { 1, 2, 3, . . . } of all positive integers is obviously
rilarger" than the set {2~ 4, 6, . . . } of all even positive integers, for it
contains this set as a proper subset. It appears on the surface that N has
'~more" elements.
But it is very important to avoid jumping to conclusions when dealing with infinite sets, and we must remember that our
criterion in these matters is whether there exists a one~to--one correspondenee between the sets (not l\""hether one set. is or is not a proper
subset of the other). As a matter of fact, the pairing
1, 2' 3
n, ~ .
2, 4, 6, . . . , 2n, . . .
j

'

serves to establish a one-to-one correspondence between these sets, in

which each positive inte.ger in the upper row is matched with the even
positive integer (its double) directly below it, and these tv.ro sets must
therefore be regarded as having the. sa.m.e number of elements. This is a
very remarkable r.ircumstance~ for it seems to contradict our intuition
and yet is based only on solid conunon sense. We ITTiall see below, in
Problems 6 and 7-4, that every infinite set is numerically equivalent to a
proper subset of itself. Since this property is clearly not possessed by
any finite set, some writers even use it as the definition of an infinite set .
In much the same way as above, we can show that N is numerically
equivalent to the set of all e.ven integers:

1, 2t

3~ 4~

0, 2, -

2~

5, 6J
4, -4, 6J

7, . .
~ 6, . .

Here our device is to start 'With 0 and follow ea.ch even positive integer
as we come to it by its negative. Similarly~ N is numerically equivalent
to the set of all integers:
1, 2,
o~

3J 4,

5, 6,

-1, 2, - 2_, 3,

7'

..

-a, . .

It is of considerable historical in tercst to note that Galileo ob served in the

early seventeenth century that there are precisely as many perfect
squares (1, 4, 9, 16t 25, etc") among the positive integers as there are
positive integers altogether" 1~his is clear from the pairing
1,

2~

1', 2 2,

3, 4, 51
3~.i\

4\

5 21

.
.....

It struck him as very strange that this should be true, considering how

34

Topology

sparsely strewn the squares are among all the positive integersr But
the time appears not to have been ripe for the exploration of this phenomenonj or perhaps he had other things on his mind; in any case, he did not
follow up his idea.
These examples should make it clear that all that is really necessary
in 1;howing that an infinite set X is numerically equivalent to N is that we
be able to list the elements of X,. with a first 1 a second, a third~ and so on1
in such a way that it is completely exhausted by this counting off of its
elements. It is for this reason that any infinite set which is numerically
equivalent to N is said to be countably infinite~ We say that a set is
countable if it is non-empty and finite (in which case it can obviously be
counted) or ii it is countably infinite.
One of Can tor's earliest discoveries in his study of infinite sets was
that the set of all positive rational numbers (which is very large: it
contains Nanda great many other numbers besides) is actually countable.
We cannot list the positive rational n.um be rs in order of size~ as we can
the positive in tegers:P be gin.ni ng 'With the sm.al lest, then the next smallest,
and so onj for there is no smallest, and between any two there are infinitely many others. We must find some other way of counting them,
and following Cantor, we arrange them not in order of size,, but according
to the size of the sum of the numerator and denominator.. We begin
with all positive rationals whose numerator and denominator add up to 2:
there is only one, X ;:::= 1. Next we list (with increasing numerators) al1
those for which this sum is 3: Vij 31 = 2. N cxtj all those for which this
sum is 4:~, 7~ = 1, Ji = 3. Next~ all those for which this sum is
5: >i, %, %, ~ = 4. Next, all those for which this sum is 6: 7-Sj % = 72,
% ;:::= 1, % ;; 2, 91 = 5. And so on~ If we now list all these together
from the beginning, omitting those already listed when we come to them,

we get a sequence
1, J,1~ 2,

u, 3~ J~, .%, %, 4, 7-g, 5~

....

whlch contains each positive rational number once and only once.
Figure 13 gives a schematic representation of this manner of listing the
positive rationals~ In this figure the first row contains all positive
rationals with numerator 1, the second all with numerator 2:P etc.; and
the first column contains all Vlith denominator 1 ~ the second all -with
denominator 2J and so on~ Our listing amounts to traversing this array
of numbers as the arrows indicate, where of course all those numbers
already encountered arc left out~
It's high time that we christened the infinite cardinal number we've
been discussing, and for this purpose \Ve use the first letter of the Hebrew
alphabet (Nj pronounced "aleph~'} v.dth 0 as a subscript~ We say
that No is the number of elements in any countably infinite set. Our

Sets and Functions

35

complete list of cardinal numbers so far is

1, 2' 3

..

NO

We expand this list in the next section.

Suppose now that m and n are two cardinal numbers (finite or
infinite).. The statement that mis l.esB than n (written m < n) is defined
to mean the following: if X and Y are sets 'vith m and n elements, then

-11 -21

1
3

-41 ---51

.... .........

----

-52

.... ---

.-... .......

... _..-.._.

----

------

----

...

......--. ....

....._

2
l

-2
2

-23

-2

3
2

-33

-43

-34

-4

-4

5
3

-45

4
1

-51

-5

1
I

2
r

.-

.-......-

........... -

I
I

'
I

I
I

Fig. 13.

'

A listing of the positive

I
I

'

I
I
I

__

I
f
I

rationals~

( 1) there exists a one-to-one mapping of X into Y, and (2) there does not
exist a one-to-one mapping of X onto Y. UBing thjs concept, it is easy to
relate our cardinal numbers to one another by means of
1

<

< 3 < " . < No.

Vv. .i th respect to t.he finite cardinal numbers, this ordering corresponds

to their usua1 ordering as real n um hers.
Problems

1.

2.

Prove that the set of all rational numbers (positive, negative, and
zero) is countable. (Hint .~ see our method of showing that the set of
all integers is countable.)
Use the idea behind Fig. 13 to prove that if [ Xd is a countable class
of countable sets, then Vt.Xi is also countable.. We usually express
this by saying that any countable union of countable sets is countable.

36

3.
.44'

Topology

Prove that the .set of all rational points in the coordinate plane R 2
(i.er, all points whose coordinates arc both rational) is countable~
Prove that if X1 and X1 are countable, then X1 X X2 is also
countable~

5.

6.

7.
8.

Prove that if X 1, X 2, . . ~ X are coun tablet where n is any positive

integer, then X1 X X 2 X ~ X X n is also countab1e.
Prove that every countably infinite set is numerically equivalent to a
proper subset of itself.
Prove that any non-empty subset of a countable set is countable.
Let X and Y be non~mpt~y sets, and f a mapping of X onto Y.
If Xis countable, prove that Y is also countable.
fl

7. UNCOUNTABLE SETS
All the infinite sets lVe considered in the previous section were
countable, so it might appear at this stage that every infinite set is countable. If this were true~ if the end result of the analysis of infinite sets
were that they are all numerically equivalent to one another, then
CantorJs theory- would be relatively trivial. But this is not the case, for
Cantor discovered that the infinite set R of all real numbers is not countable---or, a8 we phrase it, ll is uncountable or uncountably infinite. Since we
customarily identify the elements of R with the points of the real line
(see Sec. 4) 1 this amounts to the asgcrtion that the set of all points on the
real line represents a. '~higher type of infinityn than that of only the
integral points or on]y the rational points~
Cantor,s proof of this is very ingenious, but it is actually quite
sim plP. In outline the procedure is a.s follows: V{IJ assume that all the
real numbers (in decimal form) can be listed~ and in fact have been
listed ; then Vt...e pro<l ure a real number Vt...hich cannot be in this list-th us
contradicting our initial assumption that a complete listing is possible .
In representing real numben~ by decimals, V{e use the scheme of decimal
expansion in which infinite chains of 9's are avoided; for instanre, we
write 72 as ..5000 ..... and not as .4999 . . . . In this way we guarantee
that each re.al number has one and only one decitnal representation.
Suppose no\v t.hat we can list all the real numbersj and that they have
been listed in a column like the one below (,vherc we use particular
numbers for the purpose of ii lustration).
lst number
2nd number
3rd number

+ .712983
~4 + .913572
13

0
+

.84326.5
..

'

.,

...
. ~

Sets ond functions

37

Since it is impossible actually to write down this infinite list of decimals~

our assumption that all the real numbers can be listed in this way means
that we assume that we have available some general rule according to
which the list is constructed, similar to that used for listing the positive
rationals~ and that every conceivable real number occurs somewhere
in this list. We now demonstrate that this assumption is false by exhibiting a decimal ~a 1 a2:a 3 . ~ which is constructed in such a way that it is
not in the list. We choose a1 to be 1 unless the first digit after the
decimal point of the first number in our list is 1, in which case l\...e choose
a 1 to be 2. Clearly, our new decimal will di ifer from the first n nm bcr in
our list regardless of how we choose its remaining digits. Next~ we
choose a2 to be 1 unless the second digit after the decimal point of the
second number in our list is 1, in which case we choose a~ to be 2. Just
as above~ our new decimal will necessarily differ from the second number
in our list. We continue building up the decimal .ala2a~ ~ .. in this
way_, and since the proces8 can be continued indefinitely, it defines a real
number in decimal form (.121 . . . in the case of our illn;;t.rative example)
which is different from each number in our list. 1~his rontradicts our
a.ssumption that we can list all the real numbers and completes our proof
of the fact that the set R of all real numbers is uncountable.
We have sec11 (in Problem 6-1) that the set of all rational points on
the real line is count.a blc, and 've have just proved that the set of all
points on the rea] line is uncountable~ We conclude at once from this
that irrational points on the real line (i.e.J irrational numbers) must
exist. In fact, it is very easy to see by means of Problem 6-2 that the
set of all irrational numbers is uncountably infinite. To vary slightly a
striking metaphor coined by E . rr. BellJ the rational numbers are spotted
along the real line Jike stars against a black sky, and the dense blackness
of the background is the firmament of the irrationals. The reader is
probably familiar \\Tith a proof of the fact that the square root of 2 is
irrational. This proof demonstrates the existence of irrational numbers
by exhibiting a specimen. Our remarks, on the other hand, do not show
that this or that particular number is irrational; they merely show that
such numbers must exist, and moreover must exist in overwhelming
abundance.
If the reader supposes that the set of all points on the real line R is
uncountable because R is infinitely long, then we can disillusion him by
the following argument_, which shows that any open interval on R~ no
matter how short it may be, has precisely as many points as R itself.
Let a and b be any two real numbers with a < b~ and consider the open
interval (a~b). Figure 14 shows how to establish a one-to-0ne correspondence between the points P of (a,b) and the points P' of R: we
bend (a,b) into a semicircle; we rest th.is semicircle tangentially on the

38

Topology

real line R as shown in the figure; and we link P and P' by projecting
from its center. If formulas are preferred over geometric reasoning of
this kind, we observe that y = a
(b - a)x is a numerical equivalence
be.tween real numbers x E (O, 1) and ye: (a_,b), and that z = tan r(x - ~)
is another numerical equivalence between {0,1) and all of R. It now
follows that (a,b) and R are numerically equivalent to one another.
We are now in a position to show that any subset X of the real line
R which contains an open interval I is numerica1ly equivalent to R, no
matter how complicated the structure
oI X may be.. The proof of this
fact is very simple, and it uses only
the Schroeder-Bernstein theorem and
-1
0
1
our above result that I is numeriFig. 14. A one-to-one corr es po n den ee
between an open interval and the reel cally equivalent to R.. The argu ...
ment can be given in two sentences.
lina.
Since X is nttmerically equivalent to
itself~ it is obviously numerical1y cquiva1ent to a subset of R; and R is
numerically equivalent to a subset of X~ namely~ to 1. It is now a direct
consequence of the Schroeder-Bernstein theorem that X and R are
numerically equivalent to one anothcra We point out that all numerical
equivalences up to this point have been established by actually exhlbiting
one-to-one correspondences between the sets concerned. In the present
situationJ howevert it is not feasible to do this~ _for very little has been
assumed about the specific nature of the .set X. V\'ithout the help of the
Schroeder-Bern8tein theorem it would be very difficult to prove theorems
of this type,.
\re give another interesting application of the Schroeder-Bernstein
theorem. Consider the coordinate plane Rt and the subset X of R 2
defined by X ::::= { (x,y) :0 ::5" x < 1 and 0 < y < 1 J. We show that Xis
nu1nerically cqui valcnt to the closed~open interval

I =

i (x, y) : 0 S x < 1 and

y ;:;;: 0}

which forms its base (see Fig. 15).. Since I is numerically equivalent to a
subset of X~ nameJyt to I itself_, our conclusion will follow at once from
the Schroeder-Bernstein theorem if we can establish a one-to-one mapping
of X into J. This we now do. Let (xJy) be an arbitrary point of X.
Each of the coordinates x and y has a Wlique decimal exw.tnsion which
does not end in an infinite chain of 9's. We form another decimal z from
these by alternating their digits; fo! examplet if x = .327 . . ,. and
y = .614 . ~ . ~ then z = .. 362174 . . . . We now identify z {which
cannot end in an infinite chain of 9ts) with a point of I. This gives the
require<l one-to-one mapping of X into I and yields the somewhat

Sets and Functions

39

startling result that there are no more points inside a square than there
are on one of its sides.
In Sec. 6 we introduced the i:;ymhol Mo for the number of elements
in any countably infinite set. At the
beginning of this section V{C proved
Y axis
that the set fl of all real nun1hers
(or of all points on the real line) is
uncountably infinite. We now in~
troduce the symbol c (called the
ca.rdinal number of the continuum)
for the 11 umber of elements in R.
c is the cardinal number of R and of
any set \vhich ]s numerically equivalent to ll~ In the above three
x axis
paragraphs we have demonstrated
that c is the cardinal number of any
Fig. 15
open interval, of any subset of ll
'vhich contains an open interval, and
of the subset X of the roordina te plaric "h ich is illustrated in Fig~ 15.
Our list of cardinal numbers has n O\V gro-\v n to
1, 2, 3, . ~ ~ ~ Ka~ c,.

and they are related to each other by

1

< 2 < 3 < < No < Cr

At this point \Ve encounter one of the most famous un~olved problemf; of
mathematics. Is there a cardinal number greater than Mn and leHS than
c? No one knows the answer to this question. Cantor himself thonght
that there is no such number 1 or in other \vords_, that c i8 the next infinite
cardinal number greater than N0 , and his guc~s has come to be kno\rn as
Cantor~s continuum hypothesis. 'fhe continuum hypothesis can also be
expressed by the assertion that every unrountable set of real numbers
has c as its cardinal number. 1
There is another question which arises naturally at this stage~ and
this one \Ve are fortunately able to an;swer. Are there any infinite
cardinal numbers greater than c? Yesi there are; for example, the
cardinal number of the class of all subsets of R. This ans,ver d.Ppends on
the following fact~ if Xis any non-empty set, then the cardinal number of
X is less than the cardinal number of the elass of all subsets of X~
We prove this statement as foHows. In accordance with the definition given in the last paragraph of the previous sectionJ we must show
1

For further information about the continuum hypothesis, see Wilder [42, p. 125]
and GOdel [I 2].

Topology

( 1) that there exists a one-to-one mapping of X in to the class of all its

subsets, a.nd (2) that there docs not exist such a mapping of X onto this
class. Tu prove {1)~ we have only to point to the mapping x ~ {x J,
whieh makes correspond to each clement x that set {x} which consists of
the element x alone. 'Ve prove (2) indirectly~ Let us assume that
there docs exist a one-to-one mapping j of X onto the class of all its
8ubset8. l\rc no\V deduce a contradiction from the assumed existence of
such a mapping. l~t A be the subset of X defined by A = fx : x t f (x)}.
Since our mapping f i8 unto~ there must exist an clement a in X such that
f(a) ~ A. V\~here is the element a? If a is in ~4 ~ then by the definition
of A we have at f(a)" and since f(a) = . d., at A. This is a contradiction,
so a cannot belong to A. But if a is not in .Li, then again by the definition
of A we have a E.f(a) or a c . .4, l\hich is another contradiction.. The
situation is impossible, so our assumption that such a mapping exists
must be false~
This result guarantees that given any cardinal number, there always
exists a greater one. If \ve ~tart \vi th a. set X1 = {1} containing one
clement, then there are t'vo subsets~ the empty set 0 and the set {I l
itself. If X1 = l 1,2 J is a set containing two elements~ then there are
four su bsc ts: 0J { 1 L {2 }J l 1,2 J If X 3 = {1~ Z, 3 } is a set con tai ni ng
three elements, then there are eight subsets: e~ i L f 2 Jt { 3}, { 1 ~2}, { 1 ~3 l ,
{2,3}, f 1, 2, 3 J . In general if X n is a set with n elements~ where n is any
finite cardinal number, then X n has 2ft subsets. If we now taken to be
any infinite cardinal numbert the above f a.cts suggPst that we define 2rt to
be the number of suhRets of any sPt i,.vith n clements. If n is the first
infinite cardinal numher" namely~ No, then it can he shown that

21(

=== C~

The simplest proof of this faet depends on the ideas dcvelope,d in the
foIJo,ving paragt'aph.
Consider t.he clo.se<l-opcn unit interval [0,1) .and a real number x in
this set. Our concp1n j~ 1vith the meaning of the decimal, binary1 and
ternary expansions of x. For the sake of clarity, let us take x to he }'4:.
llow do \Ve arrive at the derimal expansion of 7;1? First~ we split [O, 1)
in to the I 0 closed-open in terva.1s

use the 10 digits 0, 11 ~ , 9 to number them in order. Our

number~~ belongs to exactly one of these intervals, namely, to [71 0 ,~io).
We have labeled this interval with the digit 2, so 2 is the first digit after
the decimal point in the decimal expansion of ~-4::
and

\\re

Sets and Functions

41

NextJ we split the interval [3-{ 0 ,7} 0 ) into the 10 closed-open intervals

[71o, 2 Hoo), {2 ~f oo, 2~1 oo) ~

. ~ . , ~ 2 ;t1 oo~ r1 o) J

and '\\Te use the 10 digits to number these in order. Our nun1ber ~
belongs to (2j} 00 , 2 ~ 0 o), "'-hich is labeled VrTith the digit 5~ so 5 is the
second number after the decimal point in the decimal expansion of ~:
~~

= r25

~ ~

If we continue this process exactly as we started itt we can obtain the

decimal expansion of Yi to as many places as we wish. As a matter of
fact, if we do continue, we get 0 at each stage from this point on:

7.i =

~25000

..

The reader should notice th.at there is no ambiguity in this system as we

have explained it: contrary to customary usage, .24999 . . . is not to
be regarded as another decimal expansion of ~~ ,vhich is 'l(cquivalcntH to
~25000 ~ .. ~ In this system, eaeh real number x in [O, 1) has one and
only one decimal expansion which cannot end in an infinite chain of 9 1s .
'"!'here is nothing magical about the role of the number 10 in the above
discusRion. If at each stage \Ve split our closed-open interval into two
equal closed-open intervals~ and if 've use the t\vo digits 0 and 1 to
number thcmt 'vc obtain the binary expanRion of any real number x in
[0,1)+ rfhe binary expansion of J4 is easily seen t-0 be .01000
'The ternary expansion of xis found similarly~ at each stage we split our
closed-open interval into three equal closed-open in.tcrva}s, and we use
the three digits 0, 1, and 2 to number themr A moment's thought should
convince the reader that the ternary expansion of ~ is ~020202 . . . ~
~TuRt as (in our system) the decimal expansion of a number in [Oil) cannot
end in an infinite chain of 9i s 1 so also its binary expansion cannot end in
an infinite chain of l's,. and its ternary ~xpa1ision cannot end in an
infinite chain of 2' sL
We no\\T use this machinery to give a proof of the fact that
<I

2lh =

....

c.

Consider the two sets N = { l ~ 2, 3, . r} and I == ro,1 )~ the first with

cardinal number No and the, second with cardinal number c. If N denotes
the class of all subsets of N, then by definition N has cardinal number
2K. Our proof amounts to showing that there exi8ts a one-to-one correspondence between N and I~ We begin by establishing a one-to-one
mapping! of N into I~ If A is a subset of N~ thenf(A) is that real number x in I whose decimal expansion x ==- +d1d2d3 . ~ is defined by the
condition th.at d~ is 3 or 5 according as n is or is not in A. .Any other two
digits can be used here, as long as neither of them is 9. Next, we con&

42

TopoJogy

struct a one-to-one mapping g of I into N.

If xis a real nun1ber in J, and

ii x = rb1bibi .. ~ is its binary expansion {so that each bn is either 0 or IL
then g(x) is that subset A of N defined by A = {n: b~ = 1 J ~ We conclude the proof with an appeal to the Schroeder-Bernstein theorem, lvhirh
guarantees that under these conditions N and I are numericaily equivalent
to one another.
If we follow up the hint contained in the fact that 2MI} :.=. ct and
successively form 2c1 2~ 1 and so on 1 we get a chain of cardinal numbers
1

<

< 3 < < NQ < c <

2~

< 22r <

there are infinitely n1any infinite rarrlina] numbers. (~learlyJ

there is only one kind of countable infinity~ symbolized by No, and
beyond this there is an infinite hierarchy of uncoun tab1e jn fini ties \\rhieh
arc all distinct from one another.
At this point "~c bring our discussion of these matters to a close.
\Ve have barely touched on Cantor's theory and have left entirely to
one side~ for instance, all questions relating to the addition and multiplication of infinite cardinal numbers and the rules of arithmetic which
app1y to these operations. \Ve have developed these ideas, not for
thP"ir ovln sake, but. for the Rake of their app1ications in algebra and
topology,. and our main purpose throughout the ]ast t'vo section~ has
bPr11 t.o give the reader some of the nrrcssary in~ight into countable and
Ull("OUntablc sets and the distinction hetween thetn. 1
i11

"~hich

Problems
1.

Sho"r gcometrieaHy that the r.;et of all pointf-1 in the coordinate plane
Jlt is numcri(a1Jy equivalent to the suhset. X of R 2 illustrated in
:Fig. I;) and defined by X = { (xj y) : 0 < x < I and 0 .$ y < 1 }, and
1hat. therefore R~ has cardinal number c. [Hint: rest an open
he1niF-phrrical surf arc ( = a hemispherica.1 surface minus its bounda.ry)
1angentja1Jy on tl1c renter of
project from various points on the
Jine th rough its een ter and perpendirula.r to R-1. t and use the Sc~hroeder
Rernstcin theorem.J
Hholv that the subset X of R 3 defined by

x7

2..

X
has l~ ardinal

{ ( x 1,

TIU.pl ber

x 2, x s) : 0

:5

.t..:

<

1 for

:=

1, 2, 3 l

c.

For thr. riP-arler who -..vi.shes to learn something about the o.rit.hmetjc of infinite
('ard inHJ numbers~ we recommend Helrnos [16, sec~ 241, Kamke [24~ ch.ap. 2Jt Sierpinski 137, chaps~ 7-10], ot Fraenkel l9J ehap. 2]_
1

Sets and Functions

3.

.43

Let n be a positive integer and consider a polynomial equation of the

form

~lith

integral coefficients and an #- 0. Such an equation has precisely n complex roots (some of ~Thieh, of course 1 may be real). An
algebraic number is a complex number which i8 n. root of such an
equation. rrhe set of all algebraic numbers {~OntainH the set of all
rational numbers (e.g., % is the root of 3x - 2 ::: 0) and many
other n um be rs besides (the square root of 2 is a root of x 2 - 2 = O~
and 1 + i is a root of x 2 - 2x
2 = 0). Complex numbers which
are not algebraic arc called tra.nscendentaL The numbers e and rare
the best kno,vn transcendental numbers, though the fa.et that they
are transcendental is quite difficult to prove (t)ee Niven [33,, chap. 9]).
Prove that real transcendental numbers cxi~t (hint: see Problem 6-.5).
Prove also that the set of all real transcendent.al numbers is uncountably infinite.
Prove that every infinite set is numerirally equivalent to a proper
subset of itself (h.int: see Problem 6-6).
Prove that the set of all real functions defined on the closed unit
interval has cardinal number 2.:. [H inl: there a.re at least as many
such functions as there are chara-e.t-eristic functions (i.e~, functions
whose values are 0 or 1) defined on the closed unit intervaL]

4..

5.

8. PARTIALLY ORDERED SETS AND LATT,CES

There are two types of relations which often arise in mathematics:
order re1ations and equivalence relationEt We touched briefly on
order relations in Problem 1-2~ and in Section 5 VlC discussed equivalence
relations in some detail. We now return to the topic of order relations
and develop those parts of this subject which are necessary for our later
work. The reader will find it helpful to keep in mind that a partial order
relation (as vle define it below) is a generalization of both set inclusion
and the order relation on the real line.
l..iet P be a non-empty set. A partial order relation in Pis a relation
which is symbolized by ~ and assumed to have the following properties:
(1) x < x for every x (reflexivi.ty);
(2) x < y and y < x => x = y (antisymmetry);
(3) x < y and y < z ~ :c ~ z (transitivity).
We sometimes " rite x ::; yin the equivalent form y > x. A non-empty
Ret Pin \Vhich there is defined a partial order relation is called a partially_
1

Topology

ordered set.

It is cle.ar that any non-empty subset of a partially ordered

set is a partially ordered set in its o,vn right~
Partially ordered sets are abundant in all branches of mathematics.
Some arc simple and easy to gra~p, while others are complex and rather
inactessiblc. We give four examples which are quite different in nature
but possess in common the virtues of being both important and easily
described.
Example 1. Let P be the set of all positive integers, and let m
mean that m divides n.

Example 2. Let P be the set R of all real numbers, and let x

its usual meaning (see Problem 1-2)~
Example 3.

and let A

<

<

y have

Let P be the class of all subsets of some universal set U,

B mean that A is a subset of B.

Example 4. Let P be the set of all real functions defined on a noncm pty set X, and let f < g mean that /(x) < g(x) for every x.
'Two elements x and y in a partially ordered set are called comparable
ii one of them is less than or equal to the other, that is, if either x < y or
y < x. The word '~partially'' in the phrase '~partially ordered sctJ) is
intended to emphasize that t.he1e may be pairs of elements in the set
which are not comparable. In Example lt for instance1 the integers
4 and 6 are not comparable, because neither divides the other; and in

Example 3, if the universal set Uhas more than one element, it is always
possible to find two subsets of []neither of which is a subset of the other.
Some partial order relations possess a fourth property in addition to
the three required by the definition:
(4) any two elements are comparable.
A partial order relation with property (4) is called a total (or linear)
arder relation~ and a partially ordered set whose relation satisfies condition
(4) is called a totally ordered .set1 or a linearly ordered set~ or1 most frequently, a chain. Example 2 is a chain, as is the subset {2, 4, 8, . . . ,
2n~ . . . } of Example 1.
Let P be a partially ordered set. An element x in P is said to be
maximal if y > x ~ y = x 1 that is, if no element other than x itself is
greater than or equal to x~ A maximal element in P is thus an elementt of
P which is not less than or equal to any other element of P. Examples
11 2, and 4 have no maximal elements. Example 3 has a single maximal
element: the set U itself~
Let A be a non--empty subset of a partially ordered set P. An
element x in P is called a lower bound of A if x < a for earh a e A; and a
lower bound of A is called a greatest lower bound of A if it is greater than or

Sets ond Functions

45

equal to every lower bound of A. Similarly, an clement y in P is said

to be an upper bound of A if a < y for every a E A; and a least upper
bound of A is an upper bound of A \vhich is less than or equal to every
upper bound of A. In general, A may have many lower bounds and
many upper bounds 1 but it is easy to prove (&ec Problem 1) that a
greatest lower bound (or least upper bound) is unique if it exists+ It is
thcref ore legitimate to speak of the greatest lower bonn d and the least
upper bound if they exist~
We illustrate these concepts in some of the partially ordered isets
mentioned above.
In Example 1, let the subset A consist of the integers 4 and 6. An
upper bound of t 41 6 j is any positive integer divisible by both 4 and 6~
12, 24, 36, and so on, are all upper bounds of {4t6}. 12 is clearly its
least upper bound 1 for it is less than or equa.l to (Le.,. it divides) every
upper bound4 The greatest lower bound of any pair of integers in this
example is their greatest common divisor~ and their least upper bound is
their least common multiple-both of which arc familiar notions from
elementary arithmetic.
We now consider Example 2, the real line 'vith its -natural order
relation. The reader will doubtless recall from his stu<ly of calculus that
3 is an upper bound of the set {(1 + 1/n)n: n ~ 1, 2i 3, .... } and that
its least upper bound is the fundamental constant e = 2. 7182 . ~ ~ .
As we have stated before 1 it is a basic property of the real line that every
non-empty subset of it which has a lov{cr bound (or upper bound) has a
greatest lower bound (or least upper bound).. rrhere arc several items of
standard notation and t.erminology- \\'hich must be mentioned in connection with this example. Let A be any non-empty set of real numbers.
If A has a lower bound1 then its greatest lower bound is usually called
its injimum and denoted by inf A. -Correspondingly~ if A has .an upper
bound, then its least upper bound is called its supremum and written
sup A~ If A happens to be finite, then inf A and sup A both exist and
belong to A.. In this case, they are often called the minimum and
maximum of A and are denoted by min A and max A4 If A consists of
two real numbers a 1 and a2J then min A is the smaller of a1 and a2, and
max A is the larger4
Finally, consider Example 3, and let A be any non-empty class of
subsets of U.. A lower bound of A is any subset of U which is contained
in every set in A, and the greatest lower bound of A is the intersection
of all its sets. Similarly, the leMt upper bound of A is the union of
al1 its .sets.
One of our main aims in this section is to state Zorn 1 s lemma 1 an
exceedingly powerful tool of proof whlch is almost indispensable in
many parts of modern pure mathematics. Zorn1 s lemma asserts that

Topology

if P is a partially ordered set in which

every chain has an upper bf)Und, then

P possesses a maximal element It is not possible to prove this in the
usual sense of the word. However, it can be shown that Zorn~s lemma is
logically equivalent to the axiom of choice, which states the following:
given any non-empty class of non-empty setsJ a set can be formed which
contains precisely one element taken from each set in the given class.
The axiom of choice may strike the reader a$ being intuitively obvious,
and in fact, either this axiom itself or 80me other principle equivalent to

t
1

0
Fig. 16.

The geometric meaning of f A (} and f

v g,

it is usually po stula tcd in the logic with \vhich 1\,.e operate. We therefore
assume Zorn's lemma as an axiom of logic. Any reader wh-c;> is interested in these matters is urged to explore them further in the literature. 1
A lattice is a partially ordered set Lin which each pair of elements has
a greatest lower bound and a least upper bound~ If x and y are two
elements in I"', we denote their greatest lower bound and least upper
bound by x A y and xv y. These notations are analogous to (and are
_ intended to suggest) the notations for the intersection and union of t\\rO
sets~
We pursue this analogy even further, and call x A y and xv y the
meet and ioin of x and Y~ It is tempting to assume that all properties of
intersections and unions in the algebra of sets carry over to lattices, but
th is is not a valid assumption~ So me pro per ties do carry over (see
Problem 5)t but othersJ for instance the distributive laws, are fa]se in
some lat tires~
It is easy to sec that all four of our examples are lattices+ In
Example lt m An is the greatest common divisor of m and n, and 1n v n h:\
their least common multiple; and in Example 3, A AB = An B and
Av B ;:::;; A U B. In Example 2~ if x and y arc any two real num bers 1
then x A y is min {x~y} and xv y is inax {x~y J. In Example 4J f A !J is
for ex&mpleJ WiJder [42J pp. 129-1321~ Ho.lmos [l 6 1 secs~ 15---16]~ Birkhoff
[4t p. 42 J, Sierpinski [3 7t chap. 6 J, or Fraenk el and Bar-Hillel I l O, p. 44 J.
1 See"

Sels

and Functions

J,7

the real function defined on X by (f /\ g) (x) = min {f(x)~g(x} J, and

j v g is that defined by (f v g) (x) = ma..~ t f (x) ~g(x)}. Figure 16 ill ustra tes the geometric meaning of f /\ g and f v g for two real functions
f and g defi.ned on the closed unit interval [O,IJ.
Let L be a lattice. A sublatticc of Lis a non.-.empty subset L1 of L
wit.h the property that if x and y are in I.Ji, then x A y and xv y are also
in L1. If Lis the lattice of all real functions defined on the closed unit
interv-al, and if L1 is the set of all continuous functions in L 1 then L1 is
easily seen to be a sublattice of l.J.
If a lattice has the additional property that every non-empty subset
has a greatest lower bound and a least upper bound, then it is called a
complete lattice. Example 3 is the only complete lattice in our list.
There are many distinct types of lattices, and the theory of these
systems has a wide variety of interesting and significant applications
(see Birkhoff [41)~ We discuss some of these types in our Appendix on
Boolean algebras.
Problems

1.

2.
3.
4.

5..

Let A be a non-empty subset of a partially ordered set P. Show

that A has at most one greatest lower bound and at most one lea.st
upper bound+
Consider the set {1, 2:P 3, 4, 5 J. What elements are maximal if it is
ordered as Example 1? If it is ordered as Example 2?
Under what circumstances is Example 4 a chain?
Give an example of a partially ordered set whieh is not a lattice.
let L be a lattice. If x, y, and z arc clements of I"', verify the
following: x A x = x~ x v x = x, x ,., y = y /\ x~ x v y = y v x,
X /\

7..

A Z)

(x

y)

I\ ZJ

(y V Z) = ( X V y) V Z, { X A y) V X = X, ( X V y) I\ X = X ~
Let A be a c]ass of subsets of some non-empty universal set U. We
say that A has the finite inter section property ii every finite subclass of
A has non-empty intersection+ Use Zorn's lemma to prove that
if A has the finite intersection property, then it is contained in some
maximal class B with this property (to say that Bis a maximal class
'vith this property is to say that any class \Yhich proper]y contains B
fai1s to have this property). (Hint.~ consider the family of all
classes which contain A and have the finite intersection property:P
order this family by c1ass inclusion~ and show that any chain in the
family has an upper bound in the family.)
Prove that if X and Y are any two non-empty sets~ then there exists
a one-to-one mapping of one into the other. (Hint: choose an
X V

6.

(y

48

8.

9.

10.

Topology

element x in X and an element y in Y, and establish the obvious

one-to-one correspondence between the two single-element sets
{x} and {y} ; define an extension to be a pair of subsets A of X and
B of Y such that {x} ~A and {y~ ~ B 1 together "Nith a one-to-one
correspondence between them under which x and y correspond with
one another; order the set of all extensions in the natural way; and
apply Zorn's lemma+)
Let m and n be any two cardinal numbers (finite or infinite). The
statement that mis less than or eq:ual ton (written m S n) is defined
to mean the following: if X and Y are sets with m and n elements,
then there exists a one-to-one mapping of X into Y. Prove that
any non-empty set of cardinal numbers forms a chain when it is
ordered in th.iB way. The fact that for any two cardinal numbers
one is less t.han or equal to the other is usually called the comparability tkeorem for cardinal numbers.
Let X and Y be non-empty sets, and show that the cardinal number
of Xis less than or equal to the cardinal number of Y {::::=}there exists
a mapping of Y onto X.
Let {Xi} be any infinite class of countable sets indexed by the elements i of an index set I, and show that the cardinal number of
Vix, is less than or equal to the cardinal number of /4 (Hint if I
is only countably infinit.c 1 this fo1lows from Prob1em 6..2, and if I is
uncountable, Zorn's lemma can be applied to represent it as the
union of a disjoint class of countably infinite subsets.)

CHAPTER TWO

);fctrie Spaces

Classical analysis can be described as that part of math~Jl.ifcs

which begins with calculus a11d, in essentially the same spirit, Jlevelops
similar subject matter much further in many directions.. It is a great
nation in the world of mathematics, with many provinces, a few of which
are crdinary and partial differential equations, infinite series (especially
power series and Fourier series) 1 and analytic functions of a complex
variable~ Each of these has experienced enormous growth over a long
history, and each is rich enough in content to merit a lifetime of study.
In the course of its development, classical analysis became so complex
and varied that even an expert could find his way around in it only with
clifficulty~ Under these circumstances:P some mathematicians became
interested in trying to uncover the fundamental principles on which all
analysis rests. This movement had associated with it many of the great
names in ma.thematics of the last century: lliemann~ WeierstrassJ Cantor,
l.besgue 1 Hilbertt Ries.z, and others. It played a large part in the rise to
prominence of topology~ modern algebrat and the theory of measure and
integration; and when theM new ideas began to percolate back through
classical analysis, the brew which resulted was modern analysis.
As modern analysis developed in the hands of its creators, many a
major theorem was given a simpler.proof in a more general setting, in an
effort -to lay bare its inner meaning. Much thought was devoted to
a.nalyzing the texture of the real a.nd complex number systems, which are
the context of analysis. It we.s hoped~and these hopes were well founded
~that analysis could be clarified and simplified, and that stripping away
49

50

Topology

superfluous underbrush wouid give nev.,,. emphasis to what really mattered

from the point of view of the underlying theory. 1
Analysis is primarily concerned with limit prorcsses and continwty, so it is not surprising that mathematicians thinking along these
lines soon found themselves studying (and generalizing) two elementary
concepts: that of a convergent sequence of real or complex numbers_, and
that of a continuous function of a real or complex variable.
We remind the reader of the definitionR. :First, a sequence,.

f Xn J = {X 1,

X 2, . . , X ~ 1

of real numbers is said to be conitergent if there exists a rea] numher x

(called the fi1nit of the seque11ce) SU<~h that, given e > Qt a positive
integer no can be found v..ith the property that

n 2:: no ==>

lxn -- xf <

This condition means that. Xri must be '"close" to x for all

large'"' nJ and jt is usually 8ymbo1ized by
or

lim

Xn

'~sufficiently

and expressed by Raying that Xn approaches x or Xn corwerges to x. Second,. a real function f defined on a non-empty subset X of the real line is
said to be c<mtinuous at XD in X if for each it= > 0 there exists o > 0 such
that

x in X and jx - xol

<

.5 ~ l/(x) - f(xo)t

< t,

and f is said to be continuou.s if it is continuous at each point of X.

When X is an interval, this definition gives precise expression to the
intuitive requirement that f have a graph without breaks or gaps+ The
corresponding definitions for sequences of complex numbers and complex
functions of a complex variable arc word for word the same.
Our purpose in giving these definitions in detail here is a simple one+
We wish to point out explicitly that each is dependent for its meaning
on the concept of the absolute value of the difference between two real or
complex numbers. V./"e wish to observe aJso that this absolute value is
the distance bettoeen th.e numbers when they are regarded as points on the
real line or in the complex plane.
In many branches of mathematics-in geometry as \:Yell as analysisit has been found extremely convenient to have available a notion of
distance which is applicable to t.he elements of abstract sets. A metric
space (as we define it below) is nothing more than a non-empty set
We iUustrate these points in Appendix lJ where onf~ of the basic exist.enc!!:
theorem.a in the theory of differential equations il3 given a. brief and uncluttered proof
which depends only -0n the ide~ of thi3 chapter~
1

Metric Spaces

51

equipped 'With a concept of distiance which is suitable for the treatment

of convergent sequences in the set and continuous functions defined on
the set. Our purpose in this eh.apter is to develop in a systematic manner
the main elementary facts about metric spaces. These facts are important for their O\Yn sake 1 and also for the sake of the motivation they
provide for our later work on topological spaces.

9. THE DEFINITION AND SOME EXAMPLES

Let X be a non-empty set.

A metric on X is a real function d of

ordered pairs of elements of X which satisfies the f o llo 'Wing three
conclitions:
(1) d(x,y) > O, and d(x,y) = 0 <=> x = y;
(2) d(xty) = d(y,x) (symmetry);
(3) d(x,y) < d(x,z) + d(z,y) (the triangle inequality).
The ful}ction d assigns to each pair (x,y) of elements of X a non~ncgative
real number d(x~y)~ which by symmetry doPs not depend on the order
of the clements; d(x~y) is ca1led the distance between x and y. A metric
space consists of two objects: a non-empty set X and a metric d on X.
The clements of X are called the points of the metric space (X,d)a Whenever it can be done -without causing confusion, 've denote the metric
space (X,d) by the symbol X which is used for the underlying set of
points. One should always keep in mind, ho1vever, that a metric space
is not merely a non-empty set: it is a non-"-Cm pty set together with a
metric. It often happens that several different metrics can be defined
on a single given non-empty set, and in this case distinct metrics make the
set in to distinct metric spacesT
There are many different kinds of metric spaces, some of \vhich
play very signifiran t role8 in geometry and analysis. Our first example
is rat.her trivia.I, but it is often useful in showing that certain statements
we might wish to make are not true+ It also shov-.,.s that every nonempty set can be regarded as a metric space~
Example 1.

Let X be an arbitrary non-empty set, and define d by

d(x,y)

= {~

if x
if x

Y~

y.

The reader can easily see for himself that this definition yields a metric
on X.

Our next tv{O examples are the fundamental number systems of

ma them a tics.

52

Topotogy

Consider the real line R and the real function lxl defined
on R~ Three elementary propert.ies of this absolute value function are
important for our purposes;
Example 2~

lxl

(i)
(ii)
(iii)

~ 0~ and

jxl

~ 0

x = O;

1-xj = lxl;
Ix+ Yi < !xi + IYI

We now define a metric on R by

This is called the 'l:tsual metric on ll, and the real line, as a metric space, is
always understood to have this as its metric. The fact that d actually
is a metric f ollO\.\'"B from the th rce properties stated a hover 1'his is a
piece of reasoning v..~hich occur~ frequently in our work, so 1\,.e give the
details. By (i)_, d(x,y) = fX - YI is a non-negative real number which
equals 0 ~ x - y ~ 0 {::::> x ~ Y+ l3y (ii),
d(x,y)

And by

d(x,y)

YI

==

~x -

=:::

j(x - z)

~ (y - x) l =

xi ==

d(y,x),,

< Ix - zl

+ ~i

tY

(iii)~

= Ix - Yi

+ (z

- y)~

yj

= d(x,z)

d(z,y).

Example 3. Consider the complex plane C.. We mentioned C briefly in

Sec. 4, and we described the sense in which it can be identified as a set with
the coordinate plane R 1 We now give a somewhat. fuller discussion. If z
is a complex number t and if z = a
ib where a and b are real numbers~
then a and b are called the real part and the imaginary part of z and are
+

denoted by R(z) and J(2)4 Two complex numbers are said to be equal
if their real and imaginary parts are equal:

a + ib :.:::: c

+ id <=> a = c and b === d.

We add (or subtract) two complex numbers by adding (or subtracting)

thei~ real and imaginary parts~ and we multiply them by multiplying
them out as in elementary algebra. B.nd replacing i~ by -1 wherever it
appears:

and

(a+ ib) (c +id) = (a c) + i(b d)~

(a+ ib)(c +id) = a.c + iad + ibc + i 2bd
= (ac - bd) + i(ad + be).

Metric Spaces

53

Division is carried out in accordance with

a+ ib
(a+ ib)(c - id)
(ac
c + id = (c + id) (c - id) ~

+ bd) + i(bc

- ad)

+ d~
ac + bd
. be - ad
= ?-+ d2 + i cz + d2;'
non-.zero. If z ;;;::: a + ib is a. complex
c2

where c~ + dt is required to be
number 1 then its negative -z and its conjugate i are defined by

-z
and i
- ! ::::::::

(~a)+

i(-b)

a + i(-b), which are usually \Vritten more informally as

-a - ib and z =a - ib. It is easy to see that
~

R(2)

-z+z
--2

and

I (z) :;

2i .

The real line R is usually regarded as part of the complex plane:

R = {z:I(z) = Oi = {z:i

z}.

Simple calculations sho'v directly that

...

and

= z.

The origin, or zero 1 is the complex number 0 :;:: 0 + iO+

distance from z ~ a + ib to the origin is defined by
lzl = (a-2

lzl

lz'

ordinary

+ b2)H.

is called the absolute. V(llue of z1 and it i8 easy

!z'

1~he

lzl 2 =

and

to see that
zi.

The usual metric on C is defined by

d(z1,z~)

lz1 - z2 I

Exactly as in Example 2, the fact that this is a metric is a consequence

of the follo"=-ing properties of the. real function lzl:
(i)
(ii)
(iii)

]zl ~ OJ and

!-zl

[z1

+ z2!

lzl

= 0

=>

z :;; 0;

!zl;
< lz1[ + ii2l~
=

Properties (i) and (ii) are obvious~

is also a special case of the fact that

Since -z

::::=

(--

l)z, property (ii)

5.4
which

Topo,ogy
\VP

prove by means of

fz1z2\ 2 = z1z~z1z2 = z1z1z2Z'; = .\z1l 2 lz~r~ =

{!z1f

tz~l) 2

If we use the fact that 'R{z) f S fzf for any z, property (iii) follows directly
fron1
lz1 zd 2 ~ (z 1 z2) (zl z~) =::; (21 z2) (z:1 + Z';)

+
= Z1Z1 + z22; + :l1z; + Z1Z2
= lz1 I + lz2:l + (z1'Z; + z1'Z;}
= !z1l ~ + lz2f + 2R(z1"Z:)
::; lz1) + lz2l + 2lz1z~!
= lz1 I~ + }z2\ ~ + 2lz1\ 122\
= lz1l + lz~.d + 2lz1l \z:d
== (!z1( + fZ2\)
2

Wbenever the complex plane C is mentioned as a metric space,. its

metric is al '1t="a.ys assumed to be the usual metric defi ncd above.
The rernaining examples to be given in this sert.ion fit a con1n1on
pattcrnJ "'~hich we have t.ried to exhibit in our discussion of Examples 2
and 3. \Ve no\v point out several major features of this pattern, so that
the reader ran sec clearly how it applies in the slightly more complicated
examples that follow.
I. The elements of each space can be added and subtracted in a
natural 'vay, and every element has a negative. Each space
contains a special element, denoted hy 0 and cal1ed the origin1
or zero element.
II. In each space there is defined a notion of the distance from an
arbitrary Plement to the origin, that is 1 a notion of th~ "size'"'
of an arbitrary element. The size of an element x is a real
number denotPd belo\~/ by llxH and called its norm,. Our use
of the double vertical bars is intended to emphasize that the
norm is a generalization of the absolute value functions in
Examples 2 and 3, in the sense that it satisfies the following
three conditions: (i) llxH ~ 0, and flxH ~ 0 ~ x == 0; (ii)
(I -x~! = llx'.!; (ill) flx y\J ~ flx[I + lfyH ~
III. FinaHy, each metric arises as the norm of the difference between
t ""O elem en ts: d (x,y) = 11 x - y [l,. As in :E:xam p1c 2~ the fact
that thiH is a metric follows from the properties of the norm
ljst.ed in II. This metric is called the metric induced by the
normr
The kno\:vlcdgeable reader will see at once that we are describing here
(though incompletely and imprecisely) the concept of a normed linear
space. ?\.{ost of the metric spaces of major importance in analysis are

of this type.

Metric Spaces

55

Example 4.. Let f be a real function defined on the closed unit interval
[O,I]. We say that f is a boun<led function if there is a real number K
such that 1/(x}I < K for every x E {0 1 1]. This concept is familiar to
the reader from elementary analysis, as is that of the continuity off as
defined in the introduction to this chapter. The underlying set of points
in this example is the set of all bounded continuous real functions defined
on the closed unit intervat Actually, the boundedness of such a function
is a consequence of its other properties, but at this stage we assume it
explicitly. If f and g are two such functions~ we add and subtract
them, and form nega ti ves:P po in twise:

(f + g)(x)
(f - g)(x)

~
~

(~f)(x)

f(z)

+ g(x);

f(x) - g(x);
-f(x)~

The origin (denoted by 0) is the constant function which is identically

zero~

=0

O(x)

for all x; [O:Pl].

We define the norm of a function f by

llfll = / 0

lf(x) I dx,

and the induced metric by

d(f,g) =

II! - vii

= /

lf(x) - g(x) I dx.

The integral involved in this definition is the Riemann integral of elementary calculus~ Properties (i) and (ii) of the norm are easy to prove,
and {iii) follows from

J lf(x) + g(x)I
= J lf(x)I dx + /

Ill+ 1111 =

dx

1
0

<

J
0

(lf(x) I

+ lu(x)l) d:t

l11(x)I dx

= llfil + Hull.
Example 5..

The set of points in the preceding example-that is, the

set of all bounded continuous real functions defined on the closed unit
interval-has another metric which is far more important for our purposes. It is defined by means of

llfrl

= sup {IJ(x)I: x

e [O, 1] L

which we usually write more briefly as

and

d(f,g)

H!H = sup l/(x) l}

= Ill - uH = sup 'f(x)

- g(x)I.

56

Topology

Properties (i) and (ii) of the norm are obvious, and in Problem .5 we ask
the reader to prove (iii) in a slightly more general form. This example
is typical of a large class of metric spaces which will play a major role
in all our work throughout the rest of this book. V,.1"e denote this space
by e[0 1 l]+

So much for the present for specific examples. We nol\i turn to

several fundamental principles relating to metric spaces in general.
Let X be a metric space with metric d.. Let Y be an arbitrary nonempty subset of X. If the function d is considered to be defined only
for points in Y 1 then (Y,d) is evidently itself a metric spare. Y} 'With
d restricted in this way, is called a subspru:e of X.. This te<~hnique of
forming subspaces of a given n1etric space enables us to obtain an infinity
of further examples from the handful dcscribep above. }~or instanr.e
the closed unit interval [O, 1] is a su bspacc of the real line, as is the set
consisting of all the rational points; and the unit circle, the closed unit
disc 1 and the open uuit disc are subspaees of the complex plane. .A.lso_,
the re-al line it.self is a subspace of the complex planeA
It is dc8irable at this ~tage to introduce the extended real number
system, by v..~hich we mean the ordinary real number system R "'Tith
the symbols
1

adjoined~

symbols.

+CJ)

and

00

An extended real number is thus a real number or one of these

V{e say (by definition) that

-co<

+oo;

also, if x is any reaJ number, then

-oo<x<+oo.
1,he syn1bols - ~ and + oo add nothing to our understanding of the
real numbcrsA 'They are used mainly as a notational convenience, as "\Ye
see below.
I~t A be a non-empty set of real numbers which has an upper
bound. In SecA 8 we defined what is meant by the least upper bound
(or supremum} of A; sup A is the smallest upper bound of 11 t that is,
it is the smallest rPa1 Il umber y such that a ~ y for every a in Ar With
the stated assumptions about A, sup A always exists and is a real number.
If A is a non-empt.y sPt of real numbers which has no upper bound, and
therefore no least upper bound in R, we express this by writing
sup A

+ oo;

and if A is thl? empty subset of R 1 we put

~up

::::i.

oo ~

Metric Spaces

The greatest lower bound (or infimum) of A is defined similarly: if A

is non-empty and has a lower bound, inf A is the largest real number x
such that x :::; a for every a in A; if A is non-empty and hM no lower
bound, we put

inf A=

inf A

+ oo.

and if A is empty 1 we put

=

00

'

These remarks illustrate one advantage of the extended real number

system: it enables us to speak of sup A and inf A for subsets A of the
real line without any restrictions whatever on the natUl'e of A~

Another advantage of having available the symbols -- oo and + oo

is that they ma.ke convenient a reasonable extension of our concept of
an interval on the rea.l line~ The reader should refer to the definitions
given in Sec. 1 of the various kinds of int.ervals, for these are the definition~ whose scope 've are now ~Tidening~
Let a and b be any two real
numbers such that a :S b; then the closed interval from a to b is the
subset of the real line R defined by

{a,b]

= {x :a < x

b} .

This extends our previous notion in that a closed interval may now
consist of a single point (if a = b). If b is a real number and a is au
extended real number such that a < b1 then the open-closed interoal
from a to bis
(a, b] = {x : a

<x<

b}

This allows open-closed intervals of the form (- (() ~bJ. If a is a real

number and b is an extended real number such tha.t a < b, then the
closed-open interoal from a to b is
[a~b) ::;: ~ x: a

< z < b I.

rfhis permits [a,+ cc) to be considered a closed-open interval. If a and b

are extended real numbers such that a < b, then the open inf.eroal from
a to bis
(a,b) = {x~a < x < b}.

This adds to the previously defined open intervals those of the form
( - oc:i ,b) where b is real, (a,+ oo) where a is real, and (- a:i 1
oo).
Throughout the rest of this book, the term interval will always signify
one of the four types defined in this para graph. The extended real
numbers a and b are called the end-points of these intervals. We have
used the symbols - oo and + oo v. ith considerable freedom~ and it therefore seems desirable to emphasize that an interval in our present sense
is always a non-empty subset of the real number system~ it never actually

contains either of these symbols.

58

Topofogy

The very definition of a metric space presents us 'vi th the concept. of

the distance from one point to another. We now define the distanee
from a point to a set and the diameter of a set.
Let X be a metric space '\Vi th metric d~ and let 11 be a su bHct of ~Y.
If x is a point of X, then the distance from x to A is defined hy
d(x~A)

inf {d(x~a): a .A.};

that is, it is the greatest lower bound of the distances from x to the
points of A. The diameter of the set A is defined by
d(A)

sup {d(a1,a2): a, and a2 e A}.

The diameter of A is thus the least upper bound of the distances bct\veen
pairs of its points. .A is said to have finite diameter or infinite diameter
according as d(A) is a real number or ro .. We observe that the
empty set has infinite diameter, since d(0) = - ". A bounded set is
one whose diameter is finite. A mapping of a non-empty set into a
metric space is called a bounded mapping if its range is a bounded set.
Several of the Himpler facts about these concepts are brought out in the
follo"\\iing problems.
Problems

1.

2.

I Rt X be a metric space \vi th metric d~ Show that d 1, defined by

d1(x,y) = d(x,y)/[l + d(x,u)J, is also a metric on X. Observe that
X itself is a bounded set in the metric space ("',d1).
Let X be a non-empty set, and let d be a real function of ordered
pairs of elements of X "\vhich satisfies thP fo11owing t\vo eond1tions:
d(x,y) = 0 {::;> x ~ y, and d(x,y) < d(x 1 z)
d(y,z). Show t.hat d
is a metric on X.
Let X be a non-empty set 1 and let d be a real function of ordered
pairs of elements of X which satisfies the follo\\~ing three conditions:
d(x,y) ~ 0, and J: = y ~ d(x 1 y) = 0; d(x,y) ::::; d(y)x) _; and d(x,y) s;
d(x,z) + d(z,y)+ A function d \\it.h these properties i:5 ralled a
pseudo-metric on X+ A metric is obviously a pseudo-metric. Give
an exampl~ of a pgcudu~metric which is not a metric. Let d be a
pseudo-metric on X, define a relation .-v in X by means of

3.

x~y

{:::;} d(x,y)

= 0,

and show that this is an equivalence relation whose corresponding

class of equivalence sets can be made into a metric space in a natural

way.

A.

Let X 1 X 2, . . , X n be a fi.ni te class of metric space$ with me tries

di, dit ~ . ~ , d'*~ Show that each of the functions d and d defined
1

Mefric Spaces

59

5.

as follows is a metric on the product X1

Xt x ... x xH~
d( { :c;:} J ~Yd) = max: di(Xi,y,:); d( {xd' {yi}) = ~7 l di(Xi1Y").
Let X be a non-empty set and fa real function defined on X.. Show
that f is bounded in the sense of the definition given in the last
paragraph of the text~ there exists a real number K such that
lf(x)I < K for every x e X ~sup [f(x)~ < +co. Consider the set
of all bounded real functions defined on X, aud define the norm of a
function fin this set by

l1JH

is obvious that llfll is a

l /[[=08f=OJ and that
11! + oH < Ufll + [lgll.

It
1

6.

sup lf(x)l.
non~ncgative real number such

11-/!l=l~/ll-

that
Prove in detail that

Let I be a subset of the real liner Show that I is an interval {::::}it is

non-empty and contains ea<h point between any t'\\~o of its points
(in the sense that if x and z are in I and x; < y < z~ then y is in /)
If .{I.,;} is a non-empty class of intervals on the real line such that
tlJ; is uon-empty, show that U1.Ii is an interval.
Ik.t. X be a metric space ~ith metric d. If x is a point of X and
A a subset of X, show the follo,ving: if .A is non-empty~ d(xjA) is a
non-negative real number; and d(x,A) = +a) t=> A is empty+
Tiet X he a metric space with metric d and A a sub8et of X. Show
the following: if A is non-empty~ d(11) is a non-negative extended
real number-; d(A) = - ~ ::?A is empty; and if . A. is bounded, it is
non-empty .
+

7.

8.

10. 0 PEN SETS

Let X be a metric space with metric d. If xu is a point of X and r

is a po8itive real number, the open sphere Sr(XQ) with center xtl and -radius r
is the subset of X defined by
ST(xo)

= {x ~d(x,xo) < r 1~

An open sphere is alwa.ys non-emptyt for it contains its center~ In

Example 9-1~ an open sphere with radius I contains only its center.
S,.(xo) is often called the open sphere "With radius r centered on xo; intui ...
tively, it consists of all points in X which are '~close'' to xo, with the degree
of closeness given by r.
A few concrete examples aJ"e in order. It should be easy to visualize
the open sphere Sr(X<J) on the real line: it is the bounded open interval
(x(t - r, .x0
r) \vith mid-point xo and total length 2r. Conversely, it is
clear that any bounded open interval on the real line is an open s pherc,

60

Topology

eo the open spheres on the real line are precisely the bounded open
intervals.. The open sphere Sr(zo} in the complex plane (see Fig. 17)
is the inside of the circle with center zo and radius r. Figure 18 illustrates an open sphere in the space e[O,IJ: Sf'(fo) consists of all functions f
in efO,l] whose graphs lie within the shaded band of vertical width 2r
centered on the graph of frrA subset G of the metric space Xis called an open set if1 given any
point x in G, there exists a positive real number r such that Sr(x) ~ G,

T
0

Fig+ 17+
plane.

An open aphere in the complex

FigT

18~

An open sphere in

[O~IJ.

that is, if each point of G is the center of some open sphere contained in
G~
Loosely speaking, a set is open if each of its points is '~inside~' the
set, in the sense made precise by the defini t.ion. On the real line, a set
consisting of a single point is not open, for each bounded open interval
centered on the point contains points not in the set. Similarly, the
subset (01 1) of the real line is not open,. because the point 0 in [O, 1) has
the property ihat each bounded open interval centered on it (no matter
how small it may be) contains points not in {Ott), e.g., negative points~
If we omit the offending point O~ the resulting bounded open interval
(0, I) is an open set (this is very easy to prove and is a special case of
Theorem B below). :Further, it is quite clear that any open intervalbounded or not-is an open 8et, and also tha.t the open intervals are
the only intervals which are open sets.
Theorem A. In any metric space X, the empty Bet 9 and the full space X
are open sets.
PROOF'.
To show that 0 is open, we must show that each point in 0 is
the center of an open sphere contained in 0; but since there are no points
in 0, this requirement is automatically satisfied. Xis clearly open.1 since
every open sphere centered on each of its points is contained in X.

Metric Spaces

61

We have seen that i0,1) is not open as a subset of the real line.
However~ if we consider [0,1) as a. metric space X in its own rightt as a
subspace of the real line, then [O,l) is open as a subset of X, since from
this point of view it is the full space. This apparent paradox disappears
when we realize that points out.side of a given metric space have no
relevance to any iliscussion taking place within the context of that space~
A set is open or not open only with respect to a specific metric space containing it, never on its own.
Our next theorem justifies the adjective in the expression "open
sphere.J.,
The~rem

In any metric space X, each open sphere is an open set.

PROOF.
Let S.,.(xo) be an open sphere in x, and let x be a point in st'cx~).
We must produce an open sphere centered on x and contained in St'(x{t).
Since d(x,xo) < r, Ti = r ~ d(x,xo) is a positive real number. We show
that Sr"l(x) C S'"(xo).. If y is a point in Sr~.<x)., so that d{y,x) < r1, then
d(y,xo) < d(y,x)
d(x~xo) < r1 + d(x 1 xo) ~ [r - d(x,zo)]
d(x,xo) = r
shows that y is in S,.(xo).
B.

The folio-wing characterization of open sets in terms of open spheres

is a useful tool.
Theorem C.

Let X be a metric Bpace.

t.Lnion of open

spheres~

A subaet G of X is open

tj

u is a

We assume first that G is open, and we show that it is a union

of open spheres. If G is empty1 it is the union of the empty class of
open spheres. If G is non-empty, then since it is open 1 each of its
poinra is the center of an open sphere contained in it., and it is the union
of all the open spheres contained in it~
We now assume that G is the union of a cla~ S of open spheres. \\'"e
must show that G is open. If Sis empty, then G is also empty, and by
Theorem A, G is open. Suppose that S is non-empty~ G is also nonempty. Let x be a point in G. Since G is the union of the open sphcreB
in S:P x belongs to an open sphere S,.(xo) in S. By Theorem B, :r is the
center of an open sphere St" 1 (x) ~ Sr(xo)~ Since Sr(x:o) k. G, Sri.(x) CG and
we have an open sphere centered on x and contained in G. G is therefore open4
PROOF.

The fundamental properties of the open sets in a metric space are

those stated in
Theorem D. Let X be a metric 8'f)acc. Then {l) any union of open sets
in X is openj and (2) any finite intersection of open setB in X iB open.
PROOF. To prove {l), let {G.:} be an arbitrary class of open sets in X.
We must show that G ~ V,Gi is open. If {Gd is empty1 then G is

62

Topology

empty 1 and by Theorem A1 G is open. Suppose that {G.d is non-empty.

By Theorem CJ each Gi (being an open set) is a union of open spheres; G
is the union of all the open spheres which arise in this way; and by another
application of Theorem C, G is open.
To prove (2)) let {Gil be a finite class of open sets in X. We must
show that G = rt.{},, is open~ If {Gd is empty, then G = X; and by
Theorem A, G is open. Suppose that {G..:} is non. . empty and that
{G..:} = {G1 1 Gt, . . . ,. G~} for some positive integer n~ If G happens
to be empty, then it is open by Theorem A, so l\!e may assume that G is
non-empty. Let x be a point in G. Since xis in each (;i, and each G is
open 1 for each i there is a positive real number Ti such that Sr-i(x) ~ G~.
Let r be the smallest number in the set {ri, r~~ ~ ~ . 1 T n}. 1.,his number
r is a positive real number such that Sr-(x) ~ Sri(x) for each i, so S,(x) C
G, for each i, and therefore S,.(x) ~ G. Sinee J.'Sr{x) is an open sphere
ce11tered on x and contained in G, G is open+
The above theorem says that the class of all open sets in a metric
spa-re is closed under the formation of arbitrary unions and finite intersections. The reader should clearly understand t.hat 'theorem A is an
immediate consequence of this statement, since the e1npty set is the
union of the empty class of open sets and the full space is its intersection.
The limitation to finite intersections in this theorem is essential. To see
this, it suffices to consider t.he following sequence of open intervals on the
real line;

The intersection of these open sets is the set {O} consisting of the single
point 0 1 and this set is not open.
In an arbitrary metric space, the structure of the open sets can be
very complicated indeed. Theorem C contains the best information
available in the general case~ each open set is a union of open spheres.
In the case of the real line 1 ho\\'~evcr, a description can be given of the
open sets which is fairly explicit and rea8onably satisfying to the intuition~

Every non-ernpty open set on the real "line is the union of a

countable dis}oi.nt class of open intervals.
PnooF. Let G be a non-empty open subset of the real line~ Let x be a
point of G+ Since G is open, x is the center of a bounded open interval
contained in G. Define /:! to be the union of all the open intervals
which contain x and are contained in G. 1'he follo,ving three facts are
easily proved: lz i.s an open interval (by Theorem D a.nd Problern 9 ...6)
which contains x and is contained in G; lz contains every open interval
whlch contains x and is contained in G; and if y is another point in lz,
Theorem E.

Metric Spaces

63

then Is = I'JJ. We next observe that if x and y are any two distinct
points of G, then I and I JJ are either disjoint or identical; for ii they
have a common point z, then l:.t -= I. and Iv = I~, so I~ = I '1J Consider
the class I of all distinct sets of the form I~ for points x in G~ This is a
disjoint class of open intervals, and G is obviously its union+ It remains
to be proved that I is countable. Let Gr be the set of rational points in
G. G,, is clearly non-empty. We define a mapping f of Gr onto l as
follows: for each r in G~, let f(r) be that unique interval in I which con . .
tains r. G,. is countable by Problem 6-7, and the fact that I is countable
follows from Problem 6-8.
'Z;

A firm grasp of the ideas invo1ved in the theory of metric space$

depends on one's capacity to "see" these spaces with the mind's eye~
The complex plane is perhaps the
best metric space t-0 use as a model
from which to absorb this necessary
intnitive understanding. When we
consider an unspecified set A of complex numbers, we usually imagine
it as a region bounded by a cunrct
as in Fig. 19. We think of the point
z 1, which is completely surrounded
by points of A, as being '"inside~'
the set A, or in its ''interior,'' whll e
z2 is on the ''boundary,, of A. More
Fig. 19. A set A of complex numbers
precisely~ Z1 is the center of some
with interior point z1 and boundary
open sphere contained in A~ and point z:+
each open sphere centered on z~ intersects both A and its complement A'r We formulate these ideas for a
general metric space in the next paragraph and at the end of the next
section.
Let X be an arbitrary metric space, and let A be a subset of X. A
point in A is called an inwrior point of A if it is the center of some open
sphere contained in A; and the interior of A, denoted by Int(A), is the set
of all its interior points. Symbolically,.
Int(A) = {x:x

e A and

S~(x)

CA for some

rl.

The basic properties of interiors are the f ollovring:

(1) Int(A) is an open subset of A which contains every open subset
of A (this is often expressed by saying that the interior of A is
the largest open subset of A);
(2) A is open => A = Int (A);
(3) Int(A) equals the union of all open subsets of A~

64

Topology

The proofs of these facts are quite easy~ and we ask the reader to fill in
the details as an exercise (see Problem 8) ~
Problems

1.

2.

3.

4.

5.

6.

7.

8..
9..

10..

T.~t

X be a metric space, and show that any two distinct points of X

can be separated by open spheres in the following sense; if x and y

are distinct points in X 1 then there exist..q a disjoint pair of open
spheres each of which is centered on one of the points+
Let X be a metric space. If {x} is a subset of X consisting of a
single point,. show th.at its complement {x} / is open. More genera1ly~ show that A' is open if A. is any finite subset of X.
Let X be a metric space and S~(x) the open sphere in X with center
x and radius r+ Let A be a subset of X with diameter less than r
whieh intersec.tR Sr(X). Prove that A ~ s!b(x).
1J0t X be a metric space+ Show that every subset of X is open ~
each subset of X which consists of a single point i13 open.
Let X be a metric space with metric d, and let di be the metric
defined jn Problem 9-1. Show that the two metric spaces (X~d}
and (X,dI) have precisely the same open sets. (Hint: shov;r that
they have the same open spheres with one exception. What is this
exception?)
If X = X1 X X2 X X Xn is the product in Problem 9-4, and
if d and are the metrics on x defined in that problem, show that
the two metric spaces (X~d) and {X}d) have precisely the same
open sets. Observe that in this case the spaces do not have the
same open spheres.
Let Y be a subspace of a metric space X, and let A be a subset of the
metric space Y. Show that A is open as a subset of Y $:} it is the
intersection with Y of a set which is open in X.
Prove the statements made in the text about interiors.
Describe the interior of ca.ch of the following subsets of the re.a.I
line: the set of all integers; the set of all rationale; the set of all
inationals;. (0,1); [0}1]; [0,1) U {1,2J. Do the same for each of the
following subsets of the complex plane: {z: lzl < 1 } ; {z: Izl S 1 J ;
{z; l(z) = 0 J; {z: R(z) is rational l ~
Let A and B be two subsets of a metric space X 1 and prove the
following:
(a) Int(A) V Int(B) ~ Int(A U B);
(b) Int(A) ('\ Int(B) = Int(A ('\ B).
Give an example of two subsets A and B of the real line such that
Int(A) V Int{B) F Int(A VB).

Metric Spaces

65

11. CLOSED SETS

Let X be a metric space with metric d. If A is a subset of X, a

point x in Xis called a limit point of A if each open sphere centered on x:
contains at lea"Rt one point of A different from x~ The essential idea here
is that the points of A different from x get '~arbitrarily close" to x, or
'~pile up" at x.
The subset {l_, 72'~ ~$, . ~ . } of the real line has 0 as a limit point;
in fact, 0 is its only limit point. The closed-open interval ro,I) has 0 as a
1imit point which is in the set and 1 as a limit point which is not in the
set; f urthcr, every- real n um her .x sue h that 0 < .x < 1 is also a limit
point of this set. The set of all integral points on the real line has no
limit points at all, whereas every re al num her is a limit point of the set of
all rationals~ In Example 9-1, every open sphere of radius less than 1
consists only of its center J so no subset of this space has any limit points+
A subset F of the metric space X is called a clo.sed set if it contains
each of its limit points~ In rough terms, a set is closed if its points do not
get arbitrarily close to any point outside of it. Among the subsets of the
real line mentioned in the preceding paragraph, only the set of integral
points is closed~ In Example 9-1 ~ every subset is closed.
Theorem A. In any metric space X, the empty set
are clo Bed sets~

0 and the full space X

The empty set has no limit points, so it contains them all and is
therefore closed. Since the full space X contains all points, it automatically contains its ovrn limit points and thus is closed.

PROOF.

theorem characteri.z.es closed sets in terms of open sets~

We already know a good deal about open sets, so this characterization
provides us with a useful tool for establishing properties of closed sets~
'The

followi~g

Let X be a metric space.

complement F' i8 open .

Theorem B.

A subset F of X i.s closed ~ its

Assume first that F is closed~ We show that F1 is open. If

F1 is empty, it is open by Theorem 10-A, so we may suppose that F' is
non-empty. Let x be a point in F'a Since Fis closed and xis not in F, x
is not a. limit point of F~ Since xis not in F and is not a limit point of FJ
there exists an open sphere Sr(x) which iB disjoint from F. S1"(x) is an
open sphere centered on x and contained in F'~ and since x was taken to
be any point of F'T F' is open .
We now assume that F1 is open and show that Fis closed. The only
1vay F can fail to be closed is to have a limit point in F'. This cannot
PROOF.

66

TopoJogy

happen, for since F} is open each of its points is the center of an open
sphere disjoint from F 1 and no such point can be a limit point of F.
1

If x(] is a point in our metrir space X, and r is a non-negative' real

number, the closed sphere S.,-[;.to] with cent.er xo and radius r is the subset
of X defined by

Sr[xol contains its center, and when r = 0 it contains only its center.
The closed spheres on the real line are precisely the closed intervals. In
this connection, we observe that though open spheres on the real line are
open interval.st there are open intervals which are not open spheres, e.g.,
(~ooT+oo)~

The following theorem justifies the adjective in the phrase uclosed

sphere~''

Theorem C. In any metric space X 1 each closed sphere is a closed set.

PROOF.
Let St'[xo) be a closed sphere in X. By Theorem B, it suffices
to show that its complement S,..[xol' is open. S'"[xoY is open if it is empty~
so we may assume that it is non-empty. Let x be a point ;n Sr[xoY.
Since d(x,xo) > r, r1 == d(x,xo) - r is a positive real number. We take
r1 as the radius of an open sphere Sr-,(x) centered on x~ and we show that
S~f xvY is open by showing that J.'5rl(x) ~ Sr{xoJ'..
Let y he a point in
S,. (x), so that d(y,x) < r1. On the basis of this and the fact that
d(xo~x) < d(xo,y) + d(y,x), we see that
1

d(ytxa) ~ d(x 1 xo) ~ d(y,x)

>

d(x,xo) - ri

= d(x,xo)

- [d(x,xo) - rl

= r,

so tba t y is in S,.[ x 0] 1

The main general facts about closed sets are those given

~n

our next

theorem.
Theorem D. Let X be a m..etric space. Then (1) any interse.ction of closed
sets in X is closed,~ and (2) any finite union of closed sets in X is closed.
PROOF. By virtue of Eqs~ 2-(2) and Theorem B above, this theorem is
an immediate consequence of Theorem 10-D. We prove (1) as follows.
If f Pi) is an arbitrary class of closed subsets of X and F = ('.iF'i, then
by Theorem BJ Fis closed if F1 is open; but F' = U;.F.:' is open by Theorem
10-D, since by Theorem B each F/ is open.. The second statement is
proved similarly.

In Theorem E of the previous section, we gave an explicit characterization of the open sets on the real line.. We now consider the structure of its closed sets. Among the simplest closed sets on the real line
are the closed intervals (which are the closed spheres) and finite unions

Metric Spaces

67

of closed intervals. Finite sets are included among these, since a set
consisting of a single point is a c1osed interval with equal end-points.
'"7bat is the character of the most general closed set on the real line?
Since closed sets are the complements of open sets, Theorem lO~E gives
a complete answer to this question: the most general proper closed subset
of the real line is obtained by removing a countable disjoint class of open
..,.l
intervals. This process sounds innocent enough, but in fact it leads to
some rather curious and complicated o
______l/3
_____ . . . _~-----.1
examplesr One of these examples
is of particular importanceL It was
stuclied by Cantor aTid is usually --- ------ ........
called the Cant.or set.
To construct the Cantor set1 we
proceed as follows (see FigL 20)~ ...... ___ .,._..,. ______ .,._ .._.. _.,. .....

.....___________

First, denote the closed unit interval

Fig. 20r The Cantor set.
[0,1] by F 1 ~ Next, delete from Fi
the open interval (Ya,%) which is its middle third, and denote the
remaining c1osed set by Fi C lcarly 1

F2

[O,Ys] U [%, l].

Ncxt 1 delete from F 2 the open intervals (%~%) and (%,%), which are
the middle thirds of its two pieces1 and denote the remaining closed set
by Fi. It is easy to see that

F! == [O,}.t)] U [%~~1 U [%,~~] U [%1Il

If we continue this process, at each stage deleting the open middle third
of each closed interval remaining from the previous stage, we obtain a
sequence of closed sets F n, e acb of "'T hicb contains all its successors~ The
Cantor set J/ is defined by

and it is closed by Theorem D+ F consists of those points in the closed

unit interval IO 1 l] which "'ultimately re main'' after the removal of all the
open intervals (Vs,%), (7-9~79), c~~,5y)7 . ~ ~ . What points do remain?
F clearly contains the end-points of the closed intervals which make up
each set Fn:

o,

l~ ~~

%1

7~,

%, %1

~~J

Does F contain any other points? We leave it to the reader to verify

that~ is in F and is not an end ... point. Actually 1 F contains a multitude
of points other than the above end-points, for the set of these end ... points

68

Topotogy

is r lear1y countable~ w bile the cardinaJ number of F itself is c, the cardina]

number of the continuum. To prove t.his, it suffices to exhibit a one-toone mapping f of [0,1) into F. We construct such a mappi;ng as follo,vs.
Let x be a point in [O, l)' and let x :;: rb1btba ~ . be its binary expansion
(see Sec. 7). Each bR is either 0 or 1. Lett" = 2bn, and regard ~t1t2t 3
as the ternary expansion of a real number f(x) in [O, 1). The reader \vill
easily convince himself that /{x) is in the Cantor set F: since t1 is 0 or 2~
f (:t) is not in L~ r.%J ; since t"' is 0 or 2, f (x) is not in [76 ,,3-g) or [}~ t~~) j etea
Also, it is easy to see that the mapping f: [0_.1) ~ F is one-to-one.
According to this, F contains exactly as many points as the cntirP closed
unit interval [0,1]. It is interesting to compare this conclusion v..~ith the
fact that the sum of the lengths of all the open intervals removed is
precisely 1~ since
I

73

+ 79 + ;i7 +

= J

It js alw interesting to observe (by doing a lit.tie arithmetic) thRt /i'26 i~

the union of 16, 777,216 disjoint closed intervals of the same length \\rhich
n.rP rather irregularly distributed along {0,1]. 1.,hese facts may suffice to
iu(licate that the Cantor set is a very intricate mathematical object aud
is just the sort of thing mathematicians delight in. We shall encounter
this set again from time to time, for its pro per ties illustrate several
phc nomena discussed in lat.er sections.
We conclude this section by defining two additional concepts which
are of ten usefuL
Let X be an arbitrary metric space~ and let A be a subset of X. The
closure of A, denoted by A, is the union of A and the set of all its limit
pointsr Intuitively, A is A itself together with all other points in X
'vhich are arbitrarily close to A+ As an example, if A is the open unit
disc f z: !zJ < I } in the comp1ex plane, then A is .the closed unit disc
{ z: lzl ::; 1 J. The main facts about closures are the following:
(1) A is a closed super.set of A which is contained in every closed
superset of A (we express this by Sa)-ring that A is the smallest
c1osed superset of A);
(2) A is closed <=>A = A;
(3) A equals the intersection of all closed supersets of A.
It is a. routine exercise to pro vc these sta temcn ts, and we leave this task
to the reader (in Problem 6)~
Our second concept relates to the discussion of li"ig. 19 given at the
end of the previous sectionr Again 1 let X be a metric space and A a
subset of X. A point in X is called a boundary point of A if each open
Rphere centered on the point in terser.ts both A and A', and the boundary
of .A is the set of all its boundary points~ This concept possesses the
fol ]ov;ring propert.ies:
T

Metric Spec.es

the boundary of
the boundary of
(3) A is closed ~-it
We ask the reader to give
( 1)
(2)

69

A equals A (\ A
A is a closed set;
contains its boundary.
the proofs in Problem 11 ~
1

Problems

1.

2.

3.
A.

5..
6..
7.

Let X be a metric space, and extend Problem 10-1 by proving the

following statements:
(a) any point and disjoint closed set in X can be separated by open
sets, in the sense that if x is a point and F a closed set which
does not contain x~ then there exists a disjoint pair of open
sets G1 and Gz such that x e. G1 and F ~ G:2;
(b) any disjoint pair of closed sets in X can be separated by open
sets 1 in the sense that if Fi and F2 are disjoint closed sets, then
there exists a disjoint pair of open sets G1 and G1 such that
Fi ~ Gi and F, f; G2~
Let X be a metric space~ and let A be a subset of X~ If xis a limit
point of A, show that each open sphere cent.ered on x contains an
infinite number of distinct points of A. Use this result to show that
a finite subset of X is closed.
Show that a subset of a metric space is bounded ~ it is non-empty
and is contained in some closed sphere.
Give an example .of an infinite class of closed sets whose union is not
closed. Give an example of a set which (a) is both open and closed;
(b) is neither open nor closed; (c) contains a point which is not a
limit point of the set; and (d) contains no point which is not a limit
point of the set.
Describe the interior of the Cantor set.
Prove the stat.ements made in the text about closures.
Let X be a metric space and A a subset of X. Prove the following
facts:
Int{A');
(a) A1
(b) A = { x: d(x,A)
o} .
Describe the closure of each of the following aubsets o~ the real line:
the integers; the rationals; the Cantor set; (O, + oc ) ; ( -1,0) U (O, 1).
Do the same for each of the following subsets of the complex plane:
{z: izl is rational} ; {z: 1/R(z) is an integer} ; {z: izl < 1 and
I(z) <OJ.
Let X be a metric space, let x be a. point of X, and let r be a positive
real number.. One is inclined to believe that the closure of Sr(x)
must equal 8,.[x].. Give an example to show that this is not necessarily true~ (Hint: see Example 9-1.)
;:::=

;:::=

8.

9.

70
10.

Topology

Let X be a metric space, and let G be an open set in X.

G is disjoint from a set A

11.
12.

13.

~ G is disjoint from

Prove that

A.

Prove the fact.s about boundaries stated in the text .

Describe the boundary of each of the following subsets of the real
line: the integers; the rational~; {0,1]; (0, 1).. Do the same for each
of the following subsets of the complex plane: {z: lzl < 1}; {z: lzl :-: :;
1 } ; {z : I (z) > 0 }.
Let X be a metric space and A a subset of X. A is said to be dense
{or everywhere dense) if A = X.. Prove that A is dense~ the only
closed superset of A is X {=:::} the only open set disjoint from A is
0 ~ A intersects every non-empty open set {:::}- A intersects every
open sphere~

12. CONVERGENCE, COMPLETENESS, AND BAIRE'S THEOREM

As we emphasized in the introduction to this chapter, one of our

main aims in considering m'etric spaces is to study convergent sequences
in a context more general than that of classical analysis. The fruits of
this study are many, and among them is the added insight gained into
ordinary convergence as it is used in analysis~
~t X be a metric space with metric d, and let

{x~ }

= {X 1, X2J

..

,.

Xn~ ~

..

be a sequence of points in X. We say that {xn} is convergent if there

exists a point x in X such that either
(1} for each (; > ot there exists a positive integer no such that
n ~ no ===:> d (XnJX) < E; or equivalently,
(2) for each open sphere S,,;(.x) centered on x, there exists a positive
integer no such that Xn is in S~(x) for all n 2: n 0
The reader should observe that the first r.ondition is a. direct generalization of convergence for sequences of nurr1bers as defined in the introduction, and that the second can be thought of as saying that each open
sphere centered on x contains all points of the sequence from some place
on. If we rely on our knowledge oi what is meant by a convergent
sequence of real numbers, the statement that {xn J is convergent can
equally well be defined as fallows~ there exists u point x in X such that
d(x~,x) ~ 0. We usually symbolize this by ",.riting
x.,..~

x,

and we express it verbally by saying that Xn approaches x~ or that Xn

converges to x~ It is easily seen from condition (2) and Problem 10-1 that
the point x in this discussion is unique, that is, that Xn __,. y with y # x

Metric Spaces

71

is impossible. The point x is cnl1ed the limit of the sequence {xn L and
we sometimes write Xn ~ x in the form
lim

Xn

X.

and
..
mean P.xactly the same thing, namely, that I Xn \ is a convergent sequence
with limit x.
Every convergent sequence {xn} has the following property: for each
E > 0, there exists a positive integer n[) such that m 1 n > no :::::::> d(xm,Xn) <
f.
For if x~ ~ x:t then there exists a positive integer no such ths.t n ;::
no;;;;;;;;;), d(x1t 1 x) < f./2, and from this l\~e see that

The statements

m,n

~ no~ d(xm~X11)

< d(Xm,x) + d(x,x~) < f:/2 + ~12 ;:;: ;

f ..

A sequence with this property is called a Cauchy sequ.ence, and we have

just shown that every convergent sequence is a Cauchy sequence~
Loosely speaking, this amounts to the statement that ii the terms of a
sequence approach a lin1it, then they get close to one another~ It is oi
basic importance to understand that the converse of this need not be true,
that is~ that a Cauchy sequence is not necessarily convergent. As an
example~ consider the subspace X = (Otl] of the real line.
The sequence
defined by Xn = 1/n is easily seen to be a Cauchy sequence in this spaceJ
hut it is not convergent, since the point 0 (\vhich it wants to converge to)
is not a point of the space~ The difficulty which arises in this example
stems from the fact that the notion of a convergent sequence is not
intrinsic to the sequence itself, but also depends on the structure of the
space in which it liesa A convergent sequence is not convergent "on its
own"'; it must converge to some point in the space. Some writers
emphasize the distinction between convergent sequences and Cauchy
sequences by calling the latter 0 intrinsical1y convergentu sequences.
A complete metric space is a metric space in which every Cauchy
sequence is convergent. In rough terms, a metric space is complete if
every sequence in it which tries to converge is successful, in the sense that
it finds a point in the space to converge tor The space (0,1] mentioned
above is not complete, but it evidently can be made so by adjoining the
point 0 to it to form the slightly larger space {0,1]. As a matter of fact~
any metric space, if it isn't already complete, can be made so by suitably
adjoining additional points. Vle outline thls process in a problem at the
end of Se.~. 14 .
It is a fundamental fact of elementary analysis that the real line is a
complete metric spacea The con1plex plane is also complete, as we see
from the following argument.. I..et fZn L where z~ = a~ + ibn, be a
Cauchy sequence of complex numbers.. Then {an} and {b~} are them-

72

Topology

selves Cauchy sequences of real numbers, since

and

lllm - anr < lzm - z~l

fbm - bHI ~ lzm - znr.

By the completeness of the real linej there exist real numbers a and b such
that a,,, ---i- a and bn---+ b. If we now put z = a + ibt then we see that
z~ ~ z by means of

!zn - z! = I(an + ibft) - (a. + ilJ) I

= I(an - a) + i (bn -- b) I
< ran - al + lbn - bi
and the fact that both final terms on the right approach O~ The completeness of the complex plane thus depends directly on the completeness
of the real line. T"he metric space defined in Example 9-1 is also complete;
for in this space a Cauchy sequence must be constant (i.e., it must consist
of a sing1e point repeated) from some place on,. and ~t converges with that
point as its limit~
The first three of the five metric spaces given as examples in Sec. 9
are therefore complete. What about the last two?
We ask the reader to show in Problem 5 that Ex.ample 9-4 is not
complete.. The problem of completing this space leads to. the modern
theory of Lebesgue integration, and it would carry us too far afield to
pursue this matter to its natural conclusion.
On the other hand, the space e(O, lJ defined in Example 9-5 is complete.. We prove this in a more general form in Sec. 14. The complete-ness of this space:t and of others similar to it, is one of the major focal
points of topology and modern analysis.
The terms limit and limit point. are often a source of conlusion fQr
people not thoroughly accustomed to them.. On the real line, for
insta.nceJ the constant sequence t 1, 1, ~ .. ,. 1, . . . } is convergent
with limit 1 j but the set of points of this sequence is the set consisting of
the single element 1, and by Problem 11-2, t.he point 1 is not a limit point
of this set. The essence of the matter is that a sequence of points in a set
is not a subset of the set: it is a function defined on the positive integers
with values in the set, and is UBually specified by listing its values, as in
{ x~} =:::::; { x 1, Xi, . . ~ , x,,,, . . . ', where x. is the value of the function at
the integer n. A sequence may have a limit1 but cannot have a limit
point; arid the set of points of a sequence may have a limit pointt but
cannot have a limit~ 1"1hc following theorem relates these concepts to
one another and is a useful tool for some of our later work~

Theorem A.. If a convergent sequence in a metric space has infinitely mu.ny

diBtind pointst then ts limit i8 a limit pmnt of the set of points of the 8equencer

Metric Spaces

73

Let X he a metric space,, and let {x....} be a convergent sequence

in X 'With limit x. We assume that xis not a limit point of the set of
points of the sequence, and we show that it follows from this that the
sequence has only finitely many distinct points. Our assumption
implies that there exists an open Bphere S.,.(x) centered on x which con~
ta.ins no point of the sequence different from x~ However, since x is the
limit of the sequence, all Xn~s from some place on must lie in S-r(x),. hence
must coincide with .x~ From this we see that there are only finitely many
di~tinct points in the sequence.
PROOF+

Our next theorem guarantees the completeness of many metric:

spaces which arise as subspaces of complete m~tric spaces.

Theorem B.

Let X be a complete metric space, and let Y be a subspace of

X.. Then Y i8 complete ~it i.s closed.
PROOF.
We assume first that Y is complete as a subspace of X, and we
show that it is closed. Let y be a limit point of Y. For each positive
integer n,, S1 1 ~(y) contains a puint Y~ in Y. It is clear that l Ynl converges toy in X and is a Cauchy sequence in Y, and since Y is complete,
y is in Y. Y is therefore closed.
We now assume that Y is closed, and we show that it is complete.
Let i y7t;} be a Cauchy sequence in Y. It is also a Cauchy sequence in X,
and since Xis complete, {Ynl converges to a point x in X.. We show that
x is in Y. H { yft} has only finitely many distinct points~ then x is th&t
point infinitely repeated and is thus in Y. On the other hand, if {y"} has
infinitely many distinct points, then, by 1"heorem A, x is a limit point of
the set of points of the sequence; it is therefore also a limit point of Y, and
since Y is closed, xis in Y.
A sequence {A~} of subsets of a metric space is called a ckcreaaing
sequence if
Ai => A2 =1: As ~ .

The following theorem gives conditions under which the intersection of

such a sequence is non-empty.
Theorem C (Cantor, s lntersectiott Theorem). Let X be a complete metric
S']Jace, and let {F '4} be a decreasing .sequence of non-empty closed subseu of X
such th.al d(F#) ~ 0. Then F = r.:'_ 1 Fn contains exoclly one point.
PROOF.
It is first of all evident from the assumption d(F n) ___.. 0 that F
cannot contain more than one point, so it suffices to show that Fis non ..
empty4 Let Xn be a point in Fw.. Since d(Fn)--+ 0, {x.} is a Cauchy
sequence.. Since X is complete:1 {Xn} has a limit x. We show that x is in
F, and for this it suffices to show that.xis in F,.. for a fixed but arbitr.ary
11 o..
If {x.} baa only finitely many distinct points, then x is that point

74

Topology

infinitely repeated~ and is therefore in F tt.~ If l x~ l has infinitely many

distinct points:P then xis a limit point of the set of points of the sequence,
it is a limit point of the sub~et {Xn: n > no! of the set of points of the
sequence, it is a limit point of F"fJ' and th us (since F.,..fJ is closed) it iB in Fno.
A subset A of a metric space is said to be nowhere dense if its closure
has empty interior.. It is easy to see that A is nowhere dense {=}A does
not contain any non-empty open set ::) each non-empty open set has a
non-empty open subset disjoint from A (:::::} each non-empty open set bas a
non-empty open subset disjoint from A ~each non-empty open set con...
tains an open sphere disjoint from AL If a nowhere dense set is thought of
as a set which doesn't cover very much of the space, then our next
theorem. says that a complete metric space cannot be covered by any
sequence of such sets.
Theorem D. If {4nl is a sequence of n&Where dense sets in a complete metrio
8'pace X, then there exists a point in X which is not in any of the An"s.
PROOF~
For the duration of this proof we abandon our usual notations
for open spheres and closed spheres.. Since X is open and Ai is nowhere
dense, there is an open sphere 8 1 of radius less than l which is disjoint
from A 1 Let F 1 be the concentric closed sphere whose radius is one-half
that of S1,. and consider its interior. Since A 2 is now here dense, Int( F 1)
contains an open sphere S 2 of radius less than Yz whlch is disjoint from A2.
Let F 2 be the concentric closed sphere whose radius is one-half that of S2,
and consider its interior~ Since As is nowhere denseJ lnt(F2) contains an
open sphere 8 3 of radius less than~ which is disjoint from A.,. Let FJ be
the concentric closed sphere whose radius is one-half that of Sa. Con~
tinuing in this way,, we get a decreasing sequence iF~} of non-empty
closed subsets of X such that d(Fn) ~ 0. Since Xis completeJ Theorem
C guarantees that there exists a point x in X which is in all the F :tt.'s.
This point is c1early in all the Sn J s, and therefore (since S" is disjoint from
An) it is not in any of the AnJS.
j

For our purposes,. the follov..ing equivalent form of Theorem D is

often more convenient.
Theorem E.

If a complete metric space i.8 the union of a sequence of its

81.tbuts, then the closure of at least one set, in the sequence must have nonempty inwrior.

Theorems D and E are really one theorem expressed in two different

ways. We refer to both (or either) as Baire"s theorem. This theorem is
admittedly rather technical in nature, and the reader can hardly be
expected to appreciate its significance at the present stage of our work.

Metric Spaces

75

He will find~ ho\vever, that a need for it crops up from time to time, and
when this ne.ed ariscst Ba.ire's theorem is an indispensable tooL 1
Problems

1..

2.
3.

Let X be a metric space+ If {xn} and IYn} are sequence~ in X such

that Xn---+ x .and Yn ___.,. Y~ show that d(xn,Yn) ~ d(x,y).
Show that a Cauehy sequence is convergent {:::} it has a convergent
subsequence.
If X = X 1 X X 2 X ~ X Xn is the product in Problem 9-4, and if

each of the coordinate spaces X1, X'!., .... , Xn is complete, show

that X is complete with respect to each of the metrics d and d defined
in that probJemT
4. Let ""y be a.ny non-empty set. By Problem 9-5~ the set of all bounded
real functions defined on X is a metric space with respect to the
metric induced by the norm defined in that problem. Show that this
metric space is complete. (Hint. if {fn} is a Cauchy sequence, then
~ ffl.(x)} is a Cauchy sequence of real numbers for each point x in X.)
5.. In Example 9-4, show that the folJowing functions f" defined on {01 IJ
form a Cauchy sequence in this space which is not convergent:
fn(x) = 1 if 0 < X < Y2, fn(X) = -2n(x - 7'2) + 1 if 72' < X <
~ + (~)nJ and fn(x)
0 if 72 + ()-2)n :$; x < I.
6.. Give an example to sho'v that the set F in Cantor's intersection
theorem may be en1pty if the hypothesis d(Ji~n) ~ 0 is dropped.
7. ShO\V that a closed set is nolvhcre dense {j;:::} its complement is everyv..There dense.
8.. Show that the Cantor ~t is nowhere dense~
;:::=

13. CONTINUOUS MAPPINGS

In the previous section vcre extended the idea of convergence to the
context of a general metric space. We nov_.,- do the same for continuity .
Let X and Y he metric spaces 'With metrics d 1 and d2~ and let f be a
mapping of X int-0 Y. j is said to be continuous at a point xo in X if either
There is some rather undcscriptive terminology which is often used in connection
with Baire~s theorem. We shall not make use of it our~e1ve8, but the reader ought
to be acquainted with it.. A subset of a metric .space is caHed.a set of the first category
if it can be represented as the union of a sequence of nowhere d~nse sets, and e. set of
the second category if it is not a set of the first ~ategory. Ba.ire'"s thnorem-sometimes
called the Baire category theorem-can now be expressed ::...s follows~ any complete
metric ijpace (co nsidc.red a.s a. subset of itself) is a set of the second category.
1

76

Topology

of the following equivalent conditions is satisfied:

(1) for each E > 0 there exists a > 0 such that d1 (x,xo) < 6 ~
d:(f(x ),f(xrJ)) < f;
(2) for each open sphere Sl!(/(xo)) centered on /(xo) there exists an
open sphere Sa(xo) centered on xo such that f (Sa (xo)) ~ B~ (f (xo)).
The reader will notice that the first condition generalizes the elementary
definition given in the introduction to this chapter 1 and that the second
translates the first into the language of open spheres~
Our first theorem expresses continuity at a point in terms of
sequences which converge to the point.
Theorem A. Let X and Y be mttric spaces and fa mapping of X into Y.
Then f is continuous at xo if and only if Xn --t Zo t j f(x~) ~ f (xo).
PROOF..
W""e first assume that f is continuous at xo. If {xH} is a sequence
in X such that Xn ~ X.o, we must show that f(xn)--+ f(xo}. Let S~{f(xo))
be an open sphere centered on f(zo). By our assumption, there exists an
open sphere Ss(xo) centered on Xo such that f(Sa(xo)) h S~(f(xo)). Since
Xn ~ xo-~ all xn's from some place on lie in 86(Xo).
Since f(Sa(Xo)) C
S~{f(xo))" all f(xn)'s from some place on lie in S~(f(xo)). We see from
this that f (xn) ---.:; f (xo)
To prove the other half of our theorem, we assume that f is not
continuous at xo, and we show that X.i --i- Xo does not imply /{:en)__, f(x0:).
Hy this assumption" there exists an open sphere SE(J(xo)) 'With the
property that the image under f of each open sphere centered on x' is not
con taincd in it~ Consider the sequence of open spheres S 1 (xn) _, Sli(xo) t
~
S1 1n(xo),
Form a sequence {Xn} such that x~ e S11n(xo)
and f(xtj) t St (f(xf}) ).. It is clear that Xn converges to Xo and that f(xri)
does not converge tof(xo).
+

A mapping of one metric space in to another is said to be eontin11.ou ,'1

if it is continuous at each point in its domain. The following theorem is
an immediate consequence of Theorem .{\ and this definition.

Let X and Y be metric .spaces and f a mapping of X into Y

Then f is cmtinttous if and only if Xn --t x ==> f(xn)---.:; f(x).

Theorem B..

This result shows that continuous mappings of one metric space into
another are precisely those which send convergent sequ~nces into convergent sequences, or, in other words, which preserve convergence. Our
next theorem characterizes continuous mappings in terms of open sets+
Theorem C. Let X and }T be metric spaces and fa mapping of X into. Y .
Then f is continuaus <:::;> 1-~(G) is open in X whenever G is open in Y.

We first assume that f is continuous. If G is an open set in Y.,

we must show that f- 1 (G) is open in X~ 1-1 (G} is open if it is empty, so

PROOF.

Metric Spaces

77

we may assume that it is non-empty.. Let x be a point inf- 1 (G). Then

/(x) is in G, and since G is opent there exists an open sphere St(f(x))
centered on f(x) and contained in G. By the definition of continuity,
there exists an open sphere Sa(x) such that f(S~(x)) s S~(f(x)). Since
Sf!(f(x)) C G, we also have f(Sa(x)) ~ G, and from this we see that
Sa(x) ~ 1- 1(G). S,(x) is therefore an open sphere centered on x and
contained in J- 1 (G), so 1- 1 (G) is open .
We now assume that 1- 1 (G) is open whenever G is, and we show
that f is continuous~ We show that f is continuous at an arbitrary
point x in X. let Sf(f(x)) be an open sphere centered on f(x).. This
open sphere is an open set, so its inverse image is an open set which
contains x. By this, there exists an open sphere s~(x) which is contained
in this inverse image4 It is clear that j(Sa(x)) is contained in S~(f(:t)),
so f is continuous at x. Finally, since .x was taken to be an arbitrary
point in x~ f is continuous.

The f aet just established-that continuous mappings arc precisely

those which pull open sets back to open sets-will be of great importance
for all our work frotn Chap. 3 on~
We now <~ome to the useful concept of uniform continuity. In
order to explain v..~ha t this is~ we examine the definition of continuity
expressed in eondiiion (1) at the beginning of this section+ I..iet X and
Y be metric space~ with metrics dl and d2, and let f be a mapping of
X into Y. We asHume that f is continuous1 that is, that for each point
xo in X the following is true: given e > 0 1 a number o > 0 can be found
such that d1 (x .-i:o) < 6 ==> d2(f (x) ~ f (xo)) < t.. The reader is no doubt
fanriliar with the idea that if xo is held fixed and Eis made smaller, then~
in generalJ a has to be made correspondingly smaller. Thust in the case
of the real function f defined by f (x) = 2xt a can always be chosen as any
positive number < f./2, and no larger 8 will do. In general 1 therefore,.
6 depends on E+ Let us return to our examination of the defini tionr It
says that for our given ...:, a 6 can be found which "worksH in the above
sense at the particular point x 0 under consideration. But if we hold
f: fixed and move to another point x 0 , then it may happen that this 8 no
longer v.Torks; that is1 it may be necessary to take a smaller 8 to satisfy
the requirement of the definition. We see in this way that a may well
depend, in general, not only on t= but also on x 11 Uniform continuity is
essentially continuity plus the added condition that for each E we can
find a IJ which works uniformly over the entire spoce X, in the sense that
it does not depend on x0. The formal definition is as follows~ If X and
Y are metric spaces with metrics d 1 and d2, then a mapping f of X into
Y is said to be uniformly continuous if for each E > 0 there exists 8 > 0
such that d1 (x~x') < a==> d2(f(x),f(x')) < .:. It is clear that any uni1

78

Topology

formly continuous mapping is automatically continuous~ The reader

will observe that the above real function f defined on the entire real
line R by f(x) ~ 2x is uniformly continuous. On the other hand, the
function g defined on R by g(x). = x-i is continuous but not uniformly
continuous.. Similarly, the continuous function h defined on (0,1) by
h(x) = 1/x iB not uniformly continuous~
Uniformly continuous mappings-as opposed to those which are
merely continuous-are of particular significance in analysis. The following theorem expresses a property of these mappings which is 9ften

useful.
Theorem D. Let X be a metric space, let Y be a complzete -metric space, and
let A be a dense subspace of X. If f is a uni/ormly continuous mapping of
A into Y, thenf can be exlnde.d unit[Uely to a unifOTmly continuous mapping
g of X into Y.
PnooF.. Let di and d2 be the metrics on X and Y. If A = X~ the
conclusion is obvious. We therefore assume that A X. We begin
by showing how to define the mapping o~ If xis a point in A, we define
g(x) to be f(x). Now let x be a point in X - A4 Since A is dense, x
is the limit of a convergent sequence {Un l in A. Since ~an} is a Cauchy
sequence and/ is uniformly continuous, f j(~) J is a Cauchy sequence in Y
(see Problem 8)~ Since } is complete1 there exists a point in Y-we call
this point g(x)-such that /(an)~ g(x). We must make sure that g(x)
depends only on x 1 and not on the sequence {~}. Let {bn ~ be another
sequence in A such that bn ---i- x. Then d 1 (.un,b_.) ~ O~ and by the uniiorm
continuity of f, dt(f(~) 1 f(bn)) .....-? 0. It readily follows from this that
7

f(b~)--+ g(x).

We next show that g is uniformly continuous. Let ie > 0 be given,

and use the uniform continuity off to find a > 0 such that for a .and a'
in A we have di(a~a') < 8 =::::;. d2(f(a),f(a')) < E. Let x and x' be any
poin~ in X such that d1(x 1 x') < a.. It suffices to show that d2(g(x) 1
g(x'}) < f. Let {an} and {a~ J be sequences in A such that an ~ x and
x'. By the triangle in-equality, we see that d1(n,a:) :$; d1(an,x)
d1(x,x')
d1(x'~a:). This inequality 1 together with the facts that
d1(Un 7 X) ~ 0, d1(X~X 1 ) < 6, and di(tz:' 1 a:} ~ 0, implies that d1(aR,a:) < f5
for all sufficiently large n. It now follows that d2(f(~)~f(a~)) < e
for all sufficiently large n. By Problem 12-1,

a:-+

d2(g(x),g{x')) = lim dt(/(a.n),/(a~)),

and from this and the previous statement we see that d1 (g(x),g(x
< E.
All that remains is to show that g is unique, and this is easily proved
by means of Problem 3 below.
1

))

Metri~

Spaces

79

There is &n important type of uniformly continuous me..pping which

of ten arises in practice. If X and Y are metric spaces with metrics d 1
and di 1 a mapping f of X onto Y is called an i8&metry (or an isometric
mapping) if d1(x,x1) == d!(f(x) 1 f(x 1 ) ) for all points x and x' in X; and if
such a mapping exists1 we say that X is iBometric to Y. It is clear
that an isometry is necessarily one-to-one. If X is isometric to Y, then
t.he points of these spaces can be put into one-to-one correspondence in
such a way that the distances between pairs of corresponding points are
the same~ The spaces therefore cliffer on]y in the nature of their points,
and this is of ten unimportant. / e usually consider isometric spaces t.o
be identical 'With one another. It is often convenient t-0 be able to use
this terminology in the case of mappings which are not necessarily onto.
If f is a mapping of X into Y which preserves distances in the above
sense_, then '\\re call f a.n isometry of X into Y, ,or an isometry of X onto
the subspace f(X) of Y. In this situation, we often say that Y contains
an isometric image of X~ namely1 the subspace /(X)~
Problems
1..

2.

3.

Let X and Y be metric spaces and fa mapping of X into Y. If f is

a constant mapping~ show that f is continuous. Use this to show
that a continuous mapping need noi have the property that the
image of every open set is opena
Let X be a metric space with metric d1 and let Xo be a fixed point in X.
Show that the real function f~. defined on X by J~.(x) = d(x,x 0) is
continuous. Is it uniformly continuous?
Let X and Y be metric spaces and A a non-empty subset of X. If
f and g are continuous ma.ppings of X in to Y such that f (x) g (x)
for every x in A 1 sho'v thatf(x} = g(x) for every x in A .
Let X and Y be metric spaces and f a mapping of X into Y. Show
that f is continuous ~ /- 1 (F) is closed in X whenever Ji' is closed in
Y <=> f(A) C f(A) for every subset A of X.
Show that any mapping of the metric space defined in Example U-1
into any other metric space is continuous~
Consider the real function f defined on the real line R by f (x) = x 2
If bis a given positive real number, show that the restriction off to the
closed interval {0,b] is uniformly continuous by starting v,~ith an
f > 0 and exhibiting a 6 > 0 which satisfies the requirement of the
definition.
Determine which of the following functions are uniformly continuous
on the open unit interval (O, 1): 1/(1 - x); 1/ (2 - x); sin x; sin (l/x);
xH; x 1.. 'Which are uniformly continuous on the open interval
(0, +co)?
:;:::=

.4.

5.
6.

7t

80

8.

9.

Topology

In the proof of Theorem D we uRed the follo\\'i11g fact: the image of

a Cauchy sequence under a uniformly continuous mapping is agniu
a Cauchy sequence~ Give the drtail8 of the proof.
Let f be a continuous real function defined on It 'vhich satisfies the
functional equation f(x + y) = f(x) + f(y). Sho\v that this function
mUBt have the form f(x.) = mx for some real number m. (flint:
the subspace of rational numbers is dense in the metric space R+)

1-4. SPACES OF CONTINUOUS FUNCTIONS

In Example 9-5 we gave a brief description of the metric spare
e[Oj 1]. The reader v:ill recall that the points of this space are the
bounded continuous real functions defined on the closed unit interval
[O, I] nnd that its metric is defined by d(f1g) = sup lf(x) ~ g(x)j. \Ve
have two aims in this section: to gcncra]ize this very important example
hy considering functions defined on an arbitrary metric space, and to
place all function spaces of this type in their proper context by giving
the detuil8 of the st.ructUl'al pattern (discussed briefl~~t in Sec~ 9) whieh
they all have in common with one another~ We begin with the second 1
and define the algebraic systems 'vhirh are relevant to our present
interests.
Let L be a non-empty set,. and assume that ear.h pair of elements x
and yin L can be combined by a process ralJed addition to yield an clement z in L denoted by z = x + y. Assume also that this operation of
addition satisfies the following condition~:
(I) x + y = y
x;
(2) x + (y + z) = (x + y)
z;
(3) there exists in [.; a unique element, denoted by 0 and called
the zero element, or the origin, such that x + 0 ;:: x for every x;
(4) to each element x in L there corresponds a unique element
in L, denoted by -x and called the negative of x~ such that
x+ (-x) = 0.
We adopt the device of referring to the system of real numbers or to the
system of complex numbers as the scalars. We now assume that. each
scalar a and each element x in L can be combined by a process called
scalar multiplic.ation to yield an element y in L denoted by y = crx iu
such a way that
(5) a(x + y) = ax + a.y;
( 6) (a + f3)x = QX + ~x ;
(7) (o~)x = a (f3x) ;
(8) I ~ :t = x.

Metric Spaces

81

The algebraic system L defined by these operations and axioms is called

a linear 87JUU. Depending on the numbers admitted as scalars (only the
real numbers, or all the complex numbers)J we distinguish when necessary
between real linear spac.es and cmnplex linear spaces. For geometric
reasons discussed in the next section~ a linear space is of ten called a
vector space, and its elements are spoken of as vectors.
We are not concerned here -with developing the algebraic theory of
linear spaces+ Our only interest is in making available some pertinent
concepts and terminology which are useful as a background against which
to view the metric spaces we wish to study. With this in mind, we men..
tion a few simple facts which are quite essy to prove from the axioms
for a linear space: 0 + x == x for every x; z + z : :;: : y
z~x = y
(hint: add - z to both sides on the right); a 0 = 0 (hint: a . 0 + ax ==
a(O + x)
ax = 0 + ax); 0 x = 0 (hint 0 x + ax = (O
a)x =
ax = 0 + ax) ; and {-- 1) x ;:; - .z (hint: x + (- 1) x
1 x + (- I )x =
(1 + (- l)}x = 0 x = O)r The reader will notice that. in the relation
0 ~ x = 0 we have used the symbol 0 with two different meanings: as a
scalar on the left and as a vector on the right. Several other meanings
will be given to this single symbol~ but fortunately it is always possible to
avoid confusion h~r attending closely to the context in which it occurs.
It is convenient to introduce the operation of subtraction by using the
symbol x - y aei an abbreviation for x + (-y); x - y is called the
difference betv.~een x and y.
A no n---cm pt y subset Ai of a linear space L is called a linear subspace
of L if x + y is in Jl,f whenever x and y are and if ax is in M (for any
scalar a) whenever x is.. Since M is non-empty, 0 . x = 0 shows that
0 is in .llf. Since -x = (-l)x, -xis in M whenever xis. It will be
seen at once that a linear subspace of a linear space is it.self a linear space
with respect to the same operations4
A normed linear space is a linear space on which the1e is de.fined a
norm, Le . ~ a function which assigns to each element z in the space a real
number Hxfl in such a manner that
(1) 11x!I > 0, and [lxll = 0 t j , x :: O;

;:::=

;:::=

(2) llx
Y[! < "xll
!lYll;
(3) Haxl1 = la~ llxl].
In. general terms~ a normed linear space is simply a linear space in which
th ere is av aila blc a satisfaet ory notion of the distance from an arbitrary
element to the origin. From (3) and the fact that -x ~ ( - I)x~ we
obtain fl- xii = Hx[I As we saw in SecL gJ a normed linear space is a
metric space with respect to the induced metric defined by
r

d(x 1 y)

=z

llx -

y[~a

. .\. Banach space is a normed linear space which is complete as a metric

82

Topology

space.. By Theorem 12-.B, any closed linear subspace of a Banach space

is itself a Banach apace with respect to the same algebraic operations and
the same norm.
So much for the technical framework. We now tum to the metric
spaces 'vhich really concern us. They are all function, t;paces, in the
sense that they aJ"e linear spaces whose elements are functions defined
on some non-empty set X l\ith addition and scalar multiplication defined
pointwise, Le., by (f
g)(x) = f(x)
g(x) and (a/)(x) = af(x)r We
note that the zero element in such a Jinear space is the constant function
0 whose only value is the scalar 0 and that {~ f)(x) = -f(x).
Suppose, then 1 that X is an arbitrary non-empty sett and consider
the set L of all real functions defined on X. It is clear that L is a real
linear space with respect to the operations described above. We now
restrict ourselves to the subset B consisting of the bounded functjons in
L. B is obviously a linear subspace of L"' so it is a linear space in its
own right. Even moreJ if we define a norm on B by ti/II = sup lf(x)l.t
then Bis a Banach apace (see Problems 9..S and 12-4)~
We next assume that the underlying set Xis a metric space.. This
enables us to consider the possible continuity of functions defined on
X.. We define e(X,R) to be that subset. of B which consists of continuous functions.. e(X_,R) is thus the set of all bounded continuous
real functions defined on t.he metric space X, and itfis non-empty by
Problem 13-1.

Lemma. If f and g are continuo-s real functitmS defined &n a -metric space
X, then f
g and af are also continuous, where a is any real number~

Let d be the metric on~. We show that.f + g is continuous by

showing that it is continuous a.tan arbitrary point xo in X.. Let E: > 0
b~ given. Since f and g are continuous~ and thus continuous at xo"' we
can find 1'1 > 0 and .52 > 0 such that d(xJxo) < 81 =>- If (x) - f (xo) I < t/2
and d{:c"'Xo) < a~=>- ~g(z) - g(xo)] < f-/2~ Let o be the Smaller of the
numbers 31 and .52. Then the continuity of f + g at xo follows from
PROOF.

d(x1xo)

<

S~

+ g(xo)11

We leave it

l(f + g)(x) - (f + g)(xo)J = l[f(x)

;=-

llf(xJ - JCxo)J + la<x>

to the reader

+ g(x)]

- [f(xo)
- g(xo)l I ::; IJ(x} - J(xo) r
lg(x) - g(xo)I < E/2
E/2 =

to show similarly

E.

that af is continuous.

This lemma implies that e(X"'R) is a linear subspace of the linear

space B.. We next prove that it is closed a.s a subset of the metric
space B~

Metric Spaces

83

e(X ~R) is a closed 81lbset of the metric space B.

PROOF.. Let f be a function in B \\~hich is in the closure of e(X,R). We
show that f is continuousr and therefore is in e(X~R), by sho,ving that it is
continuous at an arbitrary point xo 1n X. Since a set 'vhich equals its
closure is closed, this win suffice to prove the lemma. I...ct d be the
metric on X, and let E > 0 be given. Sinre f is in the closure of e(X,Il)~
there exists a function fo in e(X~R) such that II! - foll < t=:/3, from whieh
it follows that If(x) - f o(x) I < ~f 3 for every point x in X. Since f o
is continuous, and hence continuous at x 0 ~ we ran find a ~ > 0 suc.h that
d(xjxo) < 8 ~ lfo(x) - fo(Xo)I < e/3. The fact that f is continuous at
xo now follows from
le mm a.

d(x,xo)

<

==> IJ(x) ~ f(x~)I ~ l[f(x) - J~(x)] + [f o(x) - f o(xi))]

+ [Jo(Xo) - f(xa)H < IJ(x) - f o(x) I + lfo(x) - fo(to) l
+ ~fo(xo) - f (xo) I < t/3 + f:/3 + t:/3
~

E.

Since a closed linear subspace of a Banach spare is itself a Banach

space, we can summarize the result of the above discuss.ion and lemmas
in the following theoremT
Theorem A.. The set e(X1 R) of all bounded continuous real functions
defined on a metric spac.e Xis a real Bana.ch space with respect to pointwise
addition and scalar multiplication and the norni defined by llfl\ = sup ~f(x)~.

It is desirable at this stage to makP. a rlcar distiuction between tl\ro

types of convergence for sequences of functions. J~t X be a metric
space, and let {!" l be a sequence of real functions defined on X~ If for
each x in X it happens that {fB(x)} is a Cauchy sequence of real numbers~
then by the completeness of the real number system we can define a limit
function f by f(x) = Hm fn(x).. We then say that fn converges pointwise
to f, or that f is the pointwise limit of fn~ It is oft.en important to know
\vhat properties of the functions fn carry over to the limit function f, but
unless we strengthen the mode of convergence, very little can be said
along these lines. The stronger type of convergence normally needed to
conclude anything of interest is called uniform convergence.. In order to
explain what this is, ~~e inspect a little more closely what is involved in
point\vise convergence. To say that f n converges pointvrise to f is to say
the fo1lowing: for each point x in X, if E > 0 is given, then a positive
integer no can be found such that lf,,,(x) - f(x)j < E for all n > no. In
general, the integer nG may depend on x as well as ~. If, howev-er~ for
each given tan integer n 0 can be found which serves for all pojntsx, then'\". e
say that fn converges uniformly to f, or that f is the uniform limit of J~.
The reader \vill observe that these concepts are quite independent of the
assumption that X is a metric space and that they are meaningful for
functions defined on an arbitrary non-empty .set.

84

Topofogy

It will be seen at once that convergence in the function space e(X.,R)

is precisely uniform convergence as y.te have just defined it. The fact
that e(X,R) is complete can be restated as fol1ovn~ in the language of
uniform convergence: if a bounded real function f defined on X is the
uniform 1imit of a sequence {f~ J of bounded continuous real functions
defined on X, then f is also continuous. In other words:t in the presence
of uniform convergence, continuity carries over from the frt's to the limit
function f.
A moment's thought ""'Till convince the reader that the entire discussion given above, beginning with our definition of the linear space L,
could perfectly well have been based on complex functions~ l\l"ithout
going again through all the details} V{C st.ate the following theorem and
consider it proved.
Theorem B. The set e(X, C) of all bounded conlin uous complex. functions
defined on a. nietric space X is a complex Banach space with respect to poin lwise addition and scalar multiplication and the norm defined by

fl!H

=:;=

sup lf(x)f

In summary, we associate v..ith each metric space X tvlo linear spaces

of continuous functions defined on X. The first-e(X 1 R)-contains
only real functionsJ and the second-e(X1 C}-consists of complex
functionsL Further, all functions considered are assumed to be boundcd 1
so that the norm defined by HiH = sup l/(x) I is always a real number.
In the special case in which Xis a closed interval [a,b] on the real line 1 we
write e(X,R) in the simpler form e[a,b] .
Problems

1.
2..

3.
4.

5.

Show that a non-empty subset A of a Banach space is bounded ~

there exists a real number K such that llx!I .:$ K for every x in A.
Construct a sequence of continuouR functions defined on [O, I J which
converges p oint,,1se but not uniformly to a coniinuous limit..
Construct a sequence of continuous functions defined on {O, 1] which
con verges poin twisc to a discontinuous limit.
Let X and Y be metric spaces ,vith metrics di and d2, and let {f n} be a
sequence of mappings of X into Y "~hich converges pointwise to a
mapping f of X into Y, in the sense that f s(x)--? f(x) for each x in X.
Define what ought to be meant by the statement that fn converges
uniformly to fj and prove that under this assumption f is continuous
if each f ~ is continuous.
In th.is problem we give a procedure for constructing the completion
X* of an arbitrary metric space X.. Denote by d the metric on X.

Metric Spaces

85

Let xo be a fixed point in X, and to each point x in X make correspond

i.hu real function J~ defined on X hy f~(JJ) = d(y~x) - d(y,xo).
(a) Show that fx is bounded. (Hint.- lf.1:(y) I S d(x~xo).}
(b) Show thatj;i; is contit1uous. (Hint: lJ~(yi) ~ f'.li(y'2)! ::::; 2d(Y1tY2) .)
By (a) and (1) ), the mapping F defined by F(x) ;::= I~ is a mapping of
X into e(X,R).
(c) Show that F is an isomctry.. (Hint: ~f~l (y) - f ~(y) I <
d(X1,X2) .)
F is thus an isometr:.Y' of X into the complete metric space
e(X}R). We define the co1npleUon X* of X to be the closure of
F(X) in e(X,Jl).
(d) Show that X* is a complete metric space which contains an
isometric image of X.
(e) Show that there is a natural isometry of X* into any complete
metric space Y "'Thich contains an isometric image of X (to say
that an isometry of X* into Y is unatural" means that the
image of a point 'in x+ which corresponds to a. point in X is the
point in Y v.rhich corresponds to this same point in X).
(j) Show that (d) and (e) characterize X* in thP following sense:
if Z is a complete metric space which contains an isometric
image of X, and if there is a natural isometry of Z into any
complete metric space Y '\r\'hich contains an isometric image of
X 1 then there is a natural isometry of Z onto X*.
(g) Show that if X occurs as a subspace of n. complete metric space,
then there is a natural iso1netry of the closure of x onto
(h) Show that there is a natural isometry of any complete metric
space V'1T hi ch contains X as a. dense subspace onto x~ 1

15.. EUCLlDEAN AND UNITARY SPACES

Let n be a fixed positive integer~ and consider the set. Rn of all ordered
n-tuples x = (x 1~ x2, . . ~ , xn) of real n um be rs. 2 \1/ e promised in Sec . 4
to make thiH set into a space, a.nd "\Ve are no\v in a position to do so.
The construction outlined in {a) to (c) clearly depends on the initial choice of the
fD.;.ed point Xih If another fixed point l'\.1 1~ chosen, then another iSometry F of X into
e{ XtR) is determined. It would see1n 1 therefon~, that there is 1itt.lc justifieation for
calling the particu1ar X* defined in this problem th completion of X. In practice.
however, we usually pursue the rea~onable course of regarding isometric spaces ss
essentia.lly identico.l. lf'rom this point of viewJ the X* defined here is a. complete
metric apace which con ta.ins X as a dense su bspac~; and since by ( h) it is the only
complete metric space with thie property J it is natural to call it the completion or X.
: From this point on 11 we omit the adjective aordered." It is to he understood
that an n-tuple is alway8 ordered.
1

86

Toporogy

We begin by defining addition and scalar multiplication in Rft.

If x = (x1, x~h . . . ~ x~) and y = (Yi, y2, . ~
t y,,.), then we define x + y
and ax (where a is any real number) by
4

+Y =

(x1

+ Yi, x~ + Yt

Xn

+ y.,.)

and

With the algebraic operations defined coordinatewise in this way, Rn is a

real linear space. 111.e origin or zero e1ement is clearly 0 :;;:; (0, O,
1 0) 1
and the negative of an element x = (x1, X2J . ~ Zn) is
4

-x = (-xl, -xi, . . .

-xn).

When we speak of Rn as an n-di~mensional space, all we mean at this stage

is that each element x = (x 1, x ~ ~ . . . , .x,,.) is the ordered array of its n
co&rdinates xl~ x2J . ~ ~ 1 xM.
The reader is probably familiar with vector algebra in the ordinary
three-dimensional space of our physical intuition. If so, then he is
.z axls

y
l

'

y axis

I I
I I
' ':::::jI I
_____

"

'

Fig. 21.

A vector (or poiot) in ordine.ry space .

.accustomed to regarding a point in this space as being essentially identical

with the arrow (or vector) from the origin to that point, in the sense that
given the point,. the vector is determined, and given the vector, the point
is determined4 This situation is illustrated in Fig. 21. The above definitions of addition and scalar multiplication in Rn correspond to vector
addition and the multiplication of a vector by a real number. A word of
warning must be given. In ordinary vector algebra, a vector is usually
allowed to have its tail at any point in the space and its head at any

Metric Spaces

87

other point. It should be c]ea.rly understood, however_, that for us a

vector always has its tail at the origin. In accordance with this intuitive
picture, we may think of the elements of the real linear space R" either
as points or as generalized vectors from the origin to those points. The
latter vie"T is often more fruitful and illuminating.
There is yet a third interpretation of the elements of Jln, of great
significance from the point of view of generalizations. An n ... tuple
X = (x1_, X2,
Xn) of real numbers can be thought of as a real function
f defined on the set {1, 2, . . . 1 n} of the first n positive integers. The
ith coordinate Xi of xis then just the value of this function at the integer
i (f(i) = xi)~ and the coordinatewise operations defined above become
pointwisc operations. This way of thinking about the elements of R,.
should help to allay any doubts which might be felt as to the feasibility
of visualizing n-dimcnsional .spaces for n ~ 4. The four-dimensional
space R'r for instance, is merely the space of all real functions defined on
the set consisting of the first four positive integers, and there is surely
nothing mysterious or incomprehensible about this. 'The advantages of
the function notation are so great that we shall often (but not al"rays)
use it in preference to the n-tuplc notation. 'The rPader v..Till find it
profitable to keep in mind all three aspects of the elements Qf Rn-as
points, as vcctors 1 and as functions-and he will train himself to use that
interpretation (and notation) which appears most natural in any given
situation.
Our next task is to define a .suitable norm on the linear space R'"' .
W c recall that in solid analytic geometry the usual distance from a point
(xt y 1 z) to the origin (see Fig~ 21) is given by the expression
+

Vx2 + y2 + z2.
If x = (x1, X2J
.P Xn) is an arbitrary element of R'fL, then it is natural
to define J!xlj~the distance from the point x to the origint or the length of
the vector x~by
a

llx~I

= V1x1! 2 +
n

lxil 2 )

lx~J 2

+ ~ + lxnl

~i.

i l

If we think of Rn as composed of real functions J defined on {1, 2,

n J, then this definition becomes

ft

11/H == ( 't

-1

J/(i) I2) ~I

This is called the Euclidean norm on Rn t and the real linear space R"'
normed in this ~~ay is called n-dimensional Euc.lidean space~ The

88

Topology

Euclidean.plane is the real linear space R~ with its Euclidean norm; that is,
it is the coordinate plane equipped with the above algebraic operations
and the above norm. :For refl.8ons \vhirh will appear a littlP. later, we
o bsPrve that our formula die.fining l~x Hcan be applied cqual1y well to
n-tnp1Ps of complex numbers.
'Ve have not yet proved~ of course, that the above expression for
fixlJ po~~esse~ the three properties required by the definition of a norm .
The first and third of these conditions are clearly satisfied+ The second,
namely, that

is another matter. v.:r c prove this by the following two lemmas, of whlcb
the fir.st is essentially a tool used in the proof of the second.
Lemmo (Cauchy's lnequa,ity). Let x = (x:1, X2,
Xn) and y
Y2t ~
~ y,.) be tu,,o n-tuples of real nr complex numbers.
Then
r

(Y~

or 7 in our notation~ 2!?,.,.. l IXiYi! < 11 r[I r1y11.

i'UOOl.4'.
'\\re first remark that if a and b arc any two non-negative real
numbers_, t.hr.n ahbH < (a+ b)/2; for on squaring both sides and rearrangingi this is Pquivalent. to 0 S (a - b) 2 , which js obviousJy true.
If x = 0 or y === 0, the .assertion of the lemma is clear+ We therefore
assume that x ~ 0 and y . 0. We define a1 a:nd bi by ai = {lx1i/rlxll) 2
and bi = ([ Yil I Hy JI) 2.. . By the above remark, \Ve obtain the following for
each i:

---~~Y;l 1. S

l!x!: l!Y:i

lxI 2/llxl! 2 +-lml /llY:l 2

2

Sun1ming these inequalities as i variPs from L to n yie1ds

fl!

I: lxiy~!~ < 1 + 1

i=I

[!xll llYll -

I
'

from t\rhich our conclusion foJlows at once+

Lemma (Minkowskir s Inequality). I~et x = (x 11 x 2 _, . ~ ~ Xm) and
Y = (Y1:r Y2i .
Yn.) be two n-tuples of re,al or com.plex numberB. Then
+

.,

or, in our notation_,

IJx + yl[ < IJxH

+ ljyJI+

Metric Spaces
PROOF4

89

Using Cauchy's inequality, we have: the following chain of

relations~
ft

l]x+ Yll

I
<I
=

~xj

+ Yil lxi + y,I

]x1

+ Yir([xil + 1YiD

i=l
n

i-1

.I Ix; + Yil !x"f + I

i-1

1=1

:5
=
4

or summar1z1ng,
If

11x +

the

y" : : :; 0,

lxi

+ Yil !Yir

ilx + YH llxlf + fix+ YH llYH

ilx + ylf (llxH + llYID,

"x + YIP~ < llx + Yfl CIJxH + !IY~I }.

-

our lemma is trivially true; other\visc, it

inequality last written on dividing through by

llx + Yil .

follo~~s

from

We are now in a position to state

The set Rn of all n-luples x = (x: 1, x~, ~ . ~ , Xn) of real
numbers is a real Banach space with respect to coordinalewise addition and
scalar m ultiplicalion and the norm d ~fl ned by 'Ix 11
2:7=-l iXi I2) ~ ~
PROOF. In view of the above discussions, all that remains is to prove
completeness. It will be convenient here to use the function notation, so
that a typical element of our space is regarded as a real function defined
on i I 1 2, . . . ~ n}
J. .et If~} be a Cauchy sequence in /~$. If f > 0
iB givcnJ then fol'" all suffic.iently ]arge m and m 1 we have I~/m ~ f mr [I < E~
11/m- f mr 11 2 < E2, and I-i"'"'t ]f"~(i) -- fm (i)l 2 < f 2 ; and from this WC see
that Ifm(i) ~ f$'(i)! < E for each i {and all sufficiently large tn and 1n').
The sequence {f m} the ref ore con verges pointwisc to a limit function f
defined by f(i) == lim f m(i}. Since the set {1, 2, .
l n l is finitri this
convergence is uniform. ''re can th us find a positive in t..eger 1no fiurh
that lf.(i) - f(i)I < E/nH! for all m ~mo .and every i. Squaring en.ch
of these inequalities and summing as i varies from 1 t.o n yields
I-:ICE1 ]fm(i) - f(i)l 1 < f; 2 or llf ~/fl < E for all m > m-o. rfhis shows
that the Cauchy sequence{ fm J converges to the limit/, so Rft is complete~

Theorem A.

;:==

ft

Just as in the previous section, virtua1ly every statement we have

ma.de about n-tuples of real numbers (or about real functions defined on
{1, 2~ . . . r n l) has its complPx analogue~ We therefore ronsider the
following theorem to be fully proved+
Theorem B. The set C11 of all n-tupleB z ~ (z1} z2, . . . , Zn.) of complex
numbers is a complex B an.ach space with respect to roordinate'W"ise addition and scalar niultiplication and the nor"1n defined by jlzlj = (Y7 i lz;l 2)H+

90

Topoiogy

The space C"', with these algebraic operations and this norm, is
called n-dimensional unitary space. Needless to say, it can equally well
be viewed as the set of all complex functions f defined on the set i 1, 2,
. . . , n}, with addition and scalar multiplication (by complex scalars)
defined pointwise and the norm defined by Hfrl = (:'.2:?_ 1 ff(i)~ 2 )~ 2
The four spaces defined and discussed in this and the previous section-e(X1R) and e(X,C), and RA and Cn_form the foundation for all
our future work. In Chaps. 3 to 7 we generalize the first two by loosening
the restrictions on the underlying space X. In Chaps . 9 to 11 we study
all four from a wider point of lriew, with special emphasis on ()n.. And
in the Jast three chapters we pull these lines of deve1opment together
in such a way that ea.ch aspect of our work sheds light on all the others.
Problems

1.

2..

3..

Show that a non-empty sub.set A of Rn is bounded~ there exists a

real number K such that for each x = (xi~ x 2, , Xn) in A \Ve
have fxii :S K for every subscript i.
Let X be the set { I, 2_, . . )' n}, equipped with the metric defined
in Example 9-1. Then e(X,R) and R" are two Banach spaces which
are essentially identical as real linear spaces but which have different
norms. Show that they have the same open sets~
Prove the following extension of Minkowski's inequality. If
x = {X1, X:r, . ' Xn, .. . J and y = { Y1, Yt, ~ ~ ~ 1 Yn~ . . } are
two sequences of real or complex numbers such that x;_1 Jxnl ! and
i::-=-1 !Yn.1 2 are convergcntj then 2!;_1 'Xn + v~I tis also convergent, and

.
( fxn +
~

Ynr 1 )~ <

n-1

4.

(I !xnl )H + ( ~ 'Yn! )h.

2

ft=]

il-1

This statement is also called ~linko~N~ki's inequaJity-for infinite

sums.
The set of all sequences x = {x 1, x2~ . . . ~ Xni J of reai numbers such that ~=-1 rxAt converges is denoted by R~A If addition
and scalar multiplication are defined coordinatewise (or tcrmwisc),
and if n. norm is defined by !JxH = (~:>C:El IXtt r 2) ~' sho'\\,. that Rm is a
real Banach space. Rf.IG is called infinite-di 1nensional n~uclidean
spo.ce. The infinite-dimensional unitary spact CfiO is de fined similarly,
and is a complex Banach space .

CHAPTER THREE

7:opologieal Spnees
In the previous chapter we defined the concept of a continuous
mapping of one metric space ir1to another, and this definition was formulated in term8 of the metrics on the spaces involved. It often happens~ hol\Tcver_, that it is convPnicnt~-even essential~to be able to
speak of continuous mappings in situations ''There no useful metrics are
defined, readily definablcT or capable of being defined~ In order to deal
effectively \vi th circumstances of this kind, it. is necessary for us to 1iberate
our concept of continuity from its dependence on metric spaces.
Theorem 13-C shows that the continuity of a mapping of one metric
space into another can be expressed solely in terms of open sets, 'vithout
.any direct reference t-0 metrics. frhis suggests the possibility of dis,...
carding mctries altogP.ther and of rep]acing them as the source of our
theory by open sets. With this in mind~ our attention is drawn to
Theorem 10-D, which gives the main internal properties of the class of
open r-;ets in a metric space. These t"\vO theorems provide the leading
hint on which we base out generalization of metric spaces to topological
spaces-a topological space being simply a non-empty set in which there
is given a class of subsets, called open sets, "\\Tith the properties expressed

in l'heorem I 0-D .
. .:;+ Our undcr1ying purpose in this and the next four chapters is to study

'I

topological spaces and continuous mappings of topological spares into

one another.
c shall see that these spaces provide the ideal context
for a tl1eory of continuity in its purest form.
1'hi~ chapter is devoted primarily to explaining the concept of a
91

92

Topology

general topologica.l space. We also construct some machinery which will

be usef u1 in the detailed study of these spaces.
Our main spPcial interest in the four chapters that follow -will he in
continuous real or co mp lex functions defined on p.arti cular types of
topological spaces, and V{e shall develop the point of ,de,v that there is a
constant illuminating interplay bet'i\'~ecn the structure of these spaces
and the properties of the continuous functioru; which they carry .

16.. THE DEFINITJON AND SOME EXAMPLES

Let X be a non-empty set. A class T of subsets of X is called a

topology on X if it satisfies the following two conditions:
(I) the union of every claSB of sets in T is a set in T;
(2) the intersection of every fini tc class of sets in T is a set in T.
A topology o_n Xis thus a class of subsct.s of X whieh is closed under the
formation of arbitrary unions and finite intersectionsr A topological space
consists of two objects: a non-empty set X and a topol~gy Ton X. The
sets in the class T are called the open sets of the topological space (X ,T) ~
and the elements of X are called its points. It is customary to denote
the topological space (X,T) by the symbol X which is used for its underlying set of points. No harm can come from this practice if one clearly
understands that. a topological space is more than merely a non-empty
set: it is a non-empty set together with a specific topology on that set.
We shall often be considering several topologies on a single given set,
and in these circums~~an('es distinct topologies make the set into distinct
topological spaces. (~re observe that the empty set and the full space
are always open sets in every topological space~ since they are the union
and intersection of the empty cla.ss of sets_, "Thich is a subcla_ss of every
topology .\:
We Ii.ow list several simple examples of topological spaces. In order
to exhibit a topological space, one must specify a non-empty set, tell
which subsets are to be considered the open sets, and v c-rif y that this
given class of sets satisfies conditions (1) and (2) above. In the examples which follow_, we leave this third step to the reader.
Example 1. Let X be any metric spaceJ and let the topology be the class
.of all subsets of X which are opP.n in the sen8e of the definition in Sec. 10.
Thls iB called the usual topology on a metric spacr~ and we say that these
sets are the open sets generated by the metric on the space. Metri~
spaces are the most important topological spaces, and wh~enevcr we
speak of a metric space as a topological space, it is understood (unless

Topo,ogical Spoces

93

we say something to the contrary) that its topology is the usual topology
described here.
Ex.amp1e 2. Let X be any non-empty set,. and let the topology be the
class of all subsets of X. This is called the di~crete topolr)(JY on X~ and
any topological space v..Those topology is the discrete topology is cal1ed a

discrete space.
Example 3. Let X be any non-empty set, and let the topology consist
only of the empty set 0 and the full space X. This topology is at the
oppo~ite extreme from that described in Example 2, but they coincide
when Xis a set with only one element .
Example 4. Let X be any infinite set, and let the topotogy consist of
the empty set 0 together with all subsets of X whose complements are
finiter
~

'

/ . Example 5.

Let. X be the three-element set {a, b,. c}, and let the topology
consist of the foUo,ving sub8ets of X: 0, {a}, {a,b i, {a,c l ~ X. Spaces of
this type serve mainly to illustrate certain aspects of the theory ,vhich
will emerge in later chap tcrs~ :

A mctrizabk .space is a topological space X with the property that

there exists at least one metric on the set X whose class of generated
open sets is precisely the given topology. A mctrizable space is thus a
topological space which is-so far as its open sets are concerned-essen. .
tially a metric spacer We shall encounter many important topological
spaces which are not metri.zable,. and it is the existence of such spaces
which gives our present theory a \i\dder scope than the theory of metric
spaces~ It is a problem of considerable interest to determine what
types of topological spaces are r.netrizable, and we shall return to this
question in Sec. 29.
Let X be a topological epacet and let Y be a non-empty subset of X.
Problem 10-7 suggests a natural way of making Y into .a topological space.
The relative topology on Y is defined to be the class of all intersections
with Y of open sets in X; and when Y is equipped v.ith its relative
topology~ it is caHcd a subspace of X.
Let X and Y be topological spaces and f a mapping of X into Y.
f is called a continuous map-ping if 1- 1 (G) is open in X whenever G is open
in Y, and an open mapping if f(G) is open in Y whenever G is open in X.
A mapping is continuous if it pulls open sets back to open sets~ and open
if it carries open sets over to open sets. Any image f (X) of a topological
8pace .}( under a continuous mapping f is called a continuous image of X.
A homeomorphism. is a one-to-one continuous mapping of one topological space onto another which is also an open mapping. Two topological

9-'

Topology

spaces X and Y are said to be home01norphic if there exists a homeo ..

morphism of X onto Y (and in this case, Y is called a homeomorphic image
of X). If X and Y are homeomorphic, then their points can be put into
one-to-one correspondence in such a way that their open sets also correspond to one another4 The two spaces therefore differ only in the nature
of their points, and can, from the point of view of topology, be considered
essentially identical.
We have just used the word topology in its primary sense, as the name
of a branch of mathematics. This word derives from two Greek "-'"Ords,
and its literal meaning is , . the science of positionr" How can we account
for this? Let us say that a topologkal property is a property which, if
possessed by a topological space X, is also possessed by every homeomorphic image of X. The subject of topology can now be defined as the
study of all topological properties of topological spaces. If, very roughly,
we think of a topological space as a general type of geometric configuration, aay, a. diagram drawn on a sheet of rubber,. then a homeomorphism
may be thought of as any deformation of thls diagram (by stretching~
bending, etc.) 'vhich does not tear the sheet. A circle can be deformed
in this way into an ellipse, a triangle, or a squareJ but not into a figure
eight, a horseshoe, or a single point. A topological property would then
be any property of the diagram which is invariant under (or unchanged
by) such a deformation. Distances, angles~ and the likeJ are not topological properties, because they can be altered by suitable l 'non-tearing''
deformations. What sorts of properties are topological? In the case of
the chcle, the fact that it has one "inside" and one uoutsideH (a point
has no inside, and a figure eight has two). Alsor the fact that when two
points are removed from a circle it falls into two pieces, whereas if only
one point is removedJ then the circle remains in one piece. These
remarks m.ay suffice to indicate why topology is often described to
non-mathematicians as urubber sheet geometry." For an excellent
nontechnical discussion of topology from this geometric point of view1 see

Courant and Robbins [6, chap. 5].

Problems
1.. Let Ti and T, be two topologies on a non-empty set X, and show
that Ti fl. T2 is also a topology on X.
2.. Let X be a non-empty set, and consider the class of subsets of X
-consisting of the empty set 0 and all sets whose complements are
3..

countable. Is this a topology on X?

Which topological spaces given as examples in the text are metriza~
ble? (Hint: if a space is metri.table, then its open sets must have

certain properties.)

Topological Spaces

4,

5..

6..
7~

8.

9.
10.
11.

12..

13..

17~

95

Show that if a topological space is metri zable, then it is metrizable

in an infinite number of different ways (i.e._, by means of an infinite
number of different me tries).
Show t.hat a subspace of a topological space is itself a topological
space.
Let X be a topological space_, and let Y and Z be subspaces of X such
that Y CZ~ Show that the topology which Y has as a subspace of
Xis the same as that which it has as a subspace of Z.
.
Let f be a continuous mapping of a topological space X into a
topological space Y. If Z is a subspace of X, show that the restriction off to Z is continuous.
Let X and Y he topological spaccst and f a mapping of X into Y.
Show thatf is continuous 8 it is continuous as a mapping of X onto
the subspace /(X) of Y.
Let X, Y, and Z be topological spaces. If f~X--+ Y and g: Y--+ ~
arc continuous map pi ngs 1 show that gf: X ~ Z is also continuous.
l...et f be a one-to-one mapping of one topological space onto another,
and show thatf is a homeomorphism~ both/ andf-:- 1 are -continuous .
Give an example to shov... that a one-to-one continuous mapping of
one topological spare onto another need not be a homeomorphism .
(Hint: consider Examples 2 and 3.)
Show that a to pol ogic al space X is metrizahle {::;;;} there exists a
homeomorphism of X onto a subspace of some metric space Y.
If X and Y arc topological spacesj let X ~ Y mean that X and Y are
homeomorphic. Show that this re1ation is reflexive, symmetric,
and transitive~

ELEMENTARY CONCEPTS

We have taken open sets as the starting point in our development of

topologyj and we now define a number of other basic concepts in terms of
open sets. Most of these 'A'il 1 be familiar to the re ad er from the previous
chapter, and he will ob~erve that in every case the definition given here is
a strict generalization of our earlier definition or some equivalent form
of it.
A closed set in a topological space is a set "\vhose comp]ement is open.
The f olJoY.dng theorem is an immediate consequence of Eqs. 2-(2) and the
assumed properties of open sets.
Theorem A. Let X be a topological S]Jate. 1 hen (1) any intersection of
closed sets in X is closed; and (2) any finite union of closed sets in X is closed .
1

By considering the empty class of closed setsi we see at once that the

96

Topo~ogy

e.mpty set and the full space-its union and intersection-are always
closed sets in every topological spaceT
If A is a subset of a topological space, then its closure (denoted by A)
is the intersection of all closed supersets of A.. It is easy to see that the
closure of A is a closed superset of A Y\'Thich is contained in every closed
superset of A, and that A is closed ~ A = A.. A subset A of a topological space Xis said to be dense (or everywhere dense) if A = X, and Xis
called a separable space if it has a countable dense subset~ For reasons
which will become clear at the end of this section, we summarize the main
facts about the operation of fonning closures in the follomng theorem.
Its proof is a direct application of the above .statements~

Let X be a topological space. If A and B are arbitrary

subsets of X 1 then the operation of forming closures has the following four
properties: {1) 0 = 0; (2) A ~A; (3) A =A; and (4) AU B =AV fJ.

Theorem B.

A neighborhood of a point (or a set) in a topological space is an open

set which contains the point (or the set). A class of neighborhoods of a
point is called an open base for the point (or an open base at the point) if
each neighbor hood of the point contains a neighborb ood in this class.
In the case of a point in a metric space, an open sphere centered on the
point is a neighborhood of the point 1 and the class of all such open
spheres is an open base for the point. Our next theorem gives a useful
characterization (in terms of neighborhoods) of the closure of a set.
Theorem C. Let X be a tO'pological space and A an arbitrary subset of X.
Then A ~ {x:each neighborhood of x interBects A} .
PROOF.
We begin by proving that A is contained in the given set (the
set on the right) by showing that any point not in the given set is not in A~
Let x be a point with a neighborhood which does not intersect A. Then
the complement of this neighborhood is a closed superset of A which
does not contain x 1 and since A is the intersection of all closed supersets of
A, xis not in A. In the same way, it can easily be shown that A contains

the given set.

Let X be a topological space and A a subset of X.. A point in A is
called an isolated point of A if it has a neighborhood which contains no
other point of A.. A point x in X is said to be a li1nit point of A ii each of its
neighborhoods contains a point. of A different from x. The derived set
of A..--denoted by D(A)~is the set of all limit points of A .
Theorem D. Let X be a topological space and A a Bubset of X.. '1 hen
(I) A = A V D(A); and {2) A is closed<=) A
D(A).
PROOF. To prove (I), we use Theorem C to show that any point not
in one side is also not in the other. If xis not in A, then it has a neigh1

=-

Topological Spaces

'11

borhood disjoint from A, so it is not in A or D(A); and if xis not in A or

D(A)~ then it has a neighborhood disjoint from A 1 so it is not in A.
We prove (2) as follows. If A is closed_, so that A = A, then hy
(1) A = A V D(A)~ from which we see that A :J D(A); and if A :J D(A )~
so that AV D(A) = A 1 then by (1) we have A = A, so A is closed.

By the above definitions, a point in a set is either an isolated point

of the set or a limit point of the set, but not both. This fact leads to the
following obvious but rather satisfying theorem.
Theorem E. Let X be a topological space. Th.en any closed subset of X
is the disjoint union of its set of isolated points and its set of limit pointst in
the sense t.hat it contains these sets, t.key are disjoint, and it is their union.

Let X be a topological space and A a subset of X. The interior of A

[denoted by Int (A) J is the union of all open subsets of A, and a point in
the interior of A is called an interior point of A. It is clear that the
. interior of A is an open subset of A \Vhich contains every open subset
of A, and that A is open {::::) A = Int(A). Also, a point in A is an
interior point of A q it has a neighborhood which is contained in A.
The boundary of A is A fl A> 1 and a point in the boundary of A is called
a boundary point of A. It f ono,vs at once f ram the dP.finition that the
boundary of A is a closed set, and that it consists of all points x in X with
the property that ea.ch neighborhood oi x intersects both A and A'~
It is easy to sec from the neighborhood characterizations of the
interior and boundary that a point in a set is an interior point of the set
or a boundary point of the set, but not both~ This immediately yields
the following theorem 1 which servP.S to validate our feeling about the
intuitive significance of interiors and boundaries.
Theorem F.. Let X be a topolagi.cal space. Then any closed subset of X
is the dis}oint union of its interior and. its boundary, in the sense that it
ctJnlai ns these sets, they are disjoint~ and it is their union.

In defining a topological .space~ we chose ''open set)~ as our primitive

undefined term. Our next theorem shows that "closed set" would have
served just as well~
Theorem G. I.Jet X be a non-empty set, and let there be gi.ven a cl.ass of subsets of X which is closed under the formation of arbitrary intersections and
finite unions. Then the class of all complements of these sets is a topology
on X whose closed sets are precisely those initially given.

This follows immediately from Eqs.

topology 1 and the definition of a closed set.
PROOF.

2-(2)~

the definition of a

As the following theorem shows, we could even have taken the term
"closu.ren as our undefined concept.

98

Topology

Theorem H. Let X be a non-empty sett and let there be given a "'closure''

operation which assiQns to each subset A of X a subset A of X in such a
manner that (1) 0 = 0, (2) A c AJ (3) A = A, and (4) A 0 B = A U /J.
If a "closedn set A is defined to be one for whi'ck A = A~ then the class
of all complements of such sets is a topology on X whose closure ope-ration is
precisely that initially given.
PROOF.
In view of Theorem G, it suffices to demonstrate two facts~
that the class of all ''closed" sets is closed under the formation of arbitrary
intersections and finite uruons; and that for any set A, A equals the
intersection of al1 ''closed, :t s u pcrse ts of A .
By (1) 1 the empty set is uclosed," and from this and (4) vle see that
any finite union of uclosed 1 J sets is uclosed." By (2) ~ the full space X iB
''closed/:t so all that remains in the first part of our proof is to show that
if {Ad is a non~empty class of sets such that A; = A, for every i~ then
n,Ai = niAi- By (2), it suffices to prove that fliAi c fl1A1. For this~
it suffices to show that A ~ B ~A ~ B (since ()iAi ~ A~ for each i~ it
will follow that fl 1A 1 C
= A, for each i, from which we see that
fl 1A..: ~ niAi). Assume that A CB. Then B = AU B, and by (4),
B = A u B = A U B or A c 13.
We now let A be an arbitrary subset of X, and we 8how that A equals
the intersection of all uclosedu supersets of A. By (2) and (3)~ .11 is a
uclosedn superset of A, so it suffices to sho,v that if A f;" B and B = B:t
then A CB.. Since.A C Bt B = AV B+ By (4) and our assumption
thatB = B, weobtainB = .B ==Au B =AV B =AU B,soA ~ B4

:r::

The four properties of the closure operation assumed in thls theorem

are called the Kuratowski closure axio1ns. The last two theorems show
that it is possible to approach the subject of topological spaces by taking
either closed sets or a closure opera ti on as the basic undefined concept.
A good deal of research was done along these lines in the early days of
topology~
It v.Ta8 found that there arc many different \vays of defining a
topological space, all of which are equivalent to one another. Several
decades of experience have convinced most mathematicians that the
open set approach is the simplest, the smoothest) and the most natural..
Problems
Y be a mapping of one topological sp~ce into another.
Sho\V that f is continuous ~ 1- 1(F) is closed in X \vhenever F is

1. Let f: X

2.

---i-

closed in Y ::::} f (A) ~ f (A) for every subset A of X.

Let X be a topological space, Y a metric space, and A a subspace
of X. If f is a continuous mapping of A into Y, show that f can be
extended in at most one way to a continuous mapping of A into Y.
(Hint. see Problem 13-3.)

Topological Spaces

3.

4.
5.

6..
7.
8.

9.

99

Show that a subset of a topological apace is dense $=? it intersects

every non-empty open set.
Let A be a non-empty subset of a topological space, and show that
A is dense as a subset of the subspace A.
A .subset A of a topological space is called a perfect set if
A == D(A)+ Show that a set is perfect ~ it is closed and has no
isolated points. Show that the Cantor set is perfect+
Show that Int(~4') = A' for every subset A of a topological space~
Show that a subset of a topological space is closed ~ it contains its
boundary.
Show that a subset of a topological space has empty bouudary ~ it. is
both open and closed. {Every topological space X has the property
that the empty set 0 and the full space X are both open and closed.
In Chap~ 6 we study the hypothesis that these are the only subsets of
X '\\~hich are both open and closed.)
A subset A of a topological space is said to be nowhere dense if A has
empty interior.
(a) Show that a set A is no\vhere dense~ every non-empty open set
has a non-empty open subset disjoint from 4.
(b) Show that a ciosed set is nowhere dense (:::;'> its complement is
everywhere dense. Is this true for an arbitrary set?
(c) Show that the boundary of a closed set is nowhere dense. Is
this true for an arbitrary set?

18. OPEN BASES AND OPEN SUBBASES

A special role is played in the theory of metric spaces by the class of

open spheres "\\Tithin the class of all open sets. The main feature of their
relationship is that the open sets coincide \vith all unions of open 8phercsi
and it iollov{s from this that the continuity of a mapping can be expressed
either in terms of open spheres or in terms of open sets, at our convenience.
'Ve no"T develop similar machinery for topological spaces.
J;et X be a topological space. An open base for X is a class of open
sets \Vith the property that every open set is a union of sets in this class.
Thls condition can also be expressed in the foilov.ring equivalent form: if
G is an arbitrary non-empty open se.t and x is a. point in G, thr.n there
exists a set Bin the open base such that x e B c G. The sets in an open
base are referred to as basic open sets ~ It is clear that the c Jass of open
spheres in a metric space is an open base, and also that any c1ass of open
sets \vhich contains an open base is itseif an open base+
Generally speaking~ an open base is useful only jf its sets are simple
in form or fel\. in number~ For instance, a space \vhich has a countable
open base has many pleasant properties. A space of this kind is said to

100

Topology

be a second countable space, or to satisfy the second axiom of countability . 1

It is easy to see that any subspace of a second countable space is also
second countable~ for the class of all intersections with the subspace of
sets in an open base is evidently an open base for the subspace. The
central fact about second countable spaces can be stated as follows.
Theorem A (Lindelofts Theorem). Let X be a second countable space.
If a non-empty open set Gin Xis represented as the union of a class {Gi} of
open sets 1 th en G can be represented as a countable uni an of G/ s.

Let {Bn} be a countable open base for X~ Let x be a point in G.

The point xis in some Gi 1 and we can find a basic open set Bn such that
x e Bn ~ Gi. If we do this for each point x in G, we obtain a subclass of
our countable open base whose union is G, and this subclass is necessarily
countable+ Further, for each basic open set in this subclass we can
select a G1 which contains it. The class of Gi's which arises in this way is
clearly countable~ and its union is G.
PROOF4

Most applications of Lindelof"s theorem depend more directly on

the follovring simple consequence of it.
Theorem B.. Let X be a second countable S'}JUe. Then any open base for
X has a countable subclass which is also an open base~

Let {Bnl be a countable open base and {B*} an arbitrary open

base. Since each Bn is a union of B/sj we see by Llndelof's theorem that
each non-empty BJJ is the union 0 a countable class of B./s. In this
way we obtain a countable fami]y of countable classes of B/s~ The
union of this family of classes is evjdently an open base which is a countable subclass of the open base f B.i}.
PR001'\

If a topological space X has a rountable open base {Bn}, then it also

has a countable dense subset. To .see this, we have only to select a
point in each non-empty Bn and to note that the set of all these points is
countable and dense in X. Thus every second countable space is
separable. This simple result admits the follomng partial converse~

Every separable m.etric space is second countable.

PROOF~ Let X be a separable metric space, and let A be a countable
dense subset~ If we consider the open spheres with rational radii
centered on .all the points of A, then the class of all these open spheres is a
countable class of open sets. We show that it is an open base. Let G be
an arbitrary non-empty open set and x a point in G. We must find an
open sphere in our class which contains x and is contained in G. Let
Theorem C..

1A

iB

ft:r 81 countable space-or a spa.ce w hie h Ba tisfiel9 the first axiom of cou ntahility---a topologies.I sp&ce which ha.s & countab]e open base at each of its points (seu

See4 17).

Topologicol Spaces

101

S,.(x) be an open sphere centered on x and contained in G, and consider the

concentric open sphere 81' 13 (x) ~Tith one-third its radius. Since A is
dense, there exists a point a in A which is in s,.,3(X)+ Let rl be a rational
number such that r/3 < r 1 < 2r/3. We conclude the proof by observing
that x E Sr (a) C S,.(x) ~ G.
1

In order to form the simplest

intuitive picture of our next concept,
we give a brief discussion of rectangles and strips in the Euclidean
plane R 2 Figure 22 is intended to
illustrate our remarks. If (a1 1 b1)
and (a2,b.2) are hounded open intervals-one on the x1 axis and the other
on the x2 axis- then their product
(a 1,Q 1) X { a2,b~) =

as

< x.

i (x l ~x2) :
< bi for i =

1~2 }

is ca.Bed an open rectangle in R 2 A

closed rectangle is defined similar1y, as Fig. 22. Open strips and an open
a product of tv{O closed intervalsr rectangle.
It is easy to prove (see Problem 8)
that the class of all open rectangles is an open base for the Euclidean
plane4 We now observe that each open rectangle is the intersection of
t\ro open strips 1 in the follovring sense. We call sets of the form
(a1,bl) X R

and

R X

(a2~b2)

= i {x1,X2) :a1 < .x1 < b1~ .x2 arbitrary}

= {(x1,.x2) :a2 < xi < b2, X1 arbitrary}

open. strips in R~. If we use closed intervals here, we get what we call
closed strips. It is plain that

Since every open strip in R 2 is clearly an open set_, the class of all open
strips is a class of open sets l\~hose finite intersections form an open base~
namely, the open base composed of the open strips~ the open rectangles,
the empty set, and the full space R 2
Now let X be a topological spacer An open 8ubbase is a class of open
subsets of X whose finite intersections form an open base. This open
base is called the open base generated by the open subbase~ We refer to
the sets in an open subbase as subbasic open setsr It is easy to see that
any class of open sets which contains an open subbase is also an open
subhase.. Since the bounded open intervals on the real line constitute
an open base for this BpaceJ it is clear that aU open intervals Of the type

102

Topology

(a,+~)

and (-oo,b), where a and bare real numbers, form an open

subbase. The open base generated by thls open subbase consists of all
open intervals of this kind 1 all bounded open intervals~ the empty set~ and
the ful] space R. The ideas in the previous paragraph show at once that
all open strips in the Euclidean plane form an open subbase for this space4
The practical value of open subbases rests mainly on the following
theorem.
Theorem D.. Let X be any non-empty setJ and kt S be an arbitrary class
of subsets of X. Then Scan serve as an open subbase /or a topology on X~ in
the sense that the clas8 of all uniona of .finite intersections of sets in S is a
topology.

If S is empty 1 then the class of all finite intersections of its sets

is the single-element class l XL, and the class of all unions of sets in this
class is the t"\\""o-element class {0}X l ~ Sinec this is the topology describ( d
in Example 16-3, we may assume that S is non-cmpty4 Let B be the
class of alJ finite intersections of sets in S~ and let T be t.he class of all
unions of srts in B.. We must show that T is a topology. T clear]y
contains 0 and X, and is closed under the formation of arbitrary unions.
All that remains is to show that if lG., G 2, ~
Gn} is a non-empty
finite class of ~ets in T, then G = f"\7=- 1 Gi is also in T~ Since the empty
set is in T, we may assume that G is non-empty+ Let x be a point in G.
Then x is in each Gi, and by the definition of T, for each i there is a set
Bi in B such that x .e Bi s;: Gi. Since each Bi is a fimtc intersection of sets
in 5 1 the intersection of all sets in S which arise in this 'vay is a set in B
which contains x and is contained in Gr We conclude the proof by
noting that this shows that G is a union of sets in B and is thus itself a
set in T~
PROOF.

\Ve speak of the topology in this theorem as the topology generated by

the class S. As v.,.~e shall see in later chapters, this theorem, though not
particularly valuable as an end in itself, is quite a useful tool. It is
Ilormally used in the following manner.. If X is a non-empty set, and if
\Ye have a class of subsets of X which we wish to regard as open sets~ all
we have to do is form the topology generated by this class in the sense of
Th eoren1 I).
Our next result of tel'.l makes much lighter the task of proving that a.
given specific mapping is either continuous or open.
Theorem Et Let f:X ---1' Y be a mapping of one topological SfJace i-nto
another~ and let there be given an open base in X. and an open subbasc with
its generated open base in Y. Then (1) f is continuous~ the inverse image

Topologi~al

Spaces

103

of each bask open set is open ~ the inverse image of eac.h BUbbasic open set is
open; and (2) f is open~ the image of each basic open set is open~
PROOF..
These statements arc immediate consequences of the definitions and~ respectively~ Eqs. 3-(2) and 3-(3) and Eq . 3-(1) .

We put these two theorems to \vork in the next section 1 where we

develop a fragment of lattice theory which is very useful in the applications of topology to modern analysis.

Problems

1..

2.
3.

4..

Let X be a topological space, and B an open base with the property

that each point in the space is contained in a basic open .set different
from X4 Show that ii 0 and X happen to be in B~ then the class
which results 'vhen these two sets arc dropped from B is still an open
base .
Under what circumstances is the metric space defined in Example
9-1 separable?
Show that the real line and the complex plane are separable.
Show also that Rn and C,. are separable. Show finally that Rw and
c~ a.re separable.
Let X be the metric space whose points arc the positive integers and
whose metric is that defined in Example 9-1, and show that e(X,R) is
not separable. (Hint: if {fn l is a sequence in e(X,R)t and if f is
the function in e(XtR) defined by f(n) === o if lfn(n)I > 1 and
f(n) = lfn(n)I
1 if l/ft(n)I < 1, then II/ - !~II > 1 for every n4)
Let X be any non-empty set with the metric defined in Example 9-1,
and show that e(X,R) is separable <=> Xis finite.
The following example demonstrates that a topological space with a
countable dense subset need not be second countable. Let X be the
set of all real numbers with the topology described in Example 16-4.
(a) Show that any infinite subset of Xis dense4
-(b) Show that X is not second countable.. (Hint: assume that
there exists a countable open base, let x 0 be a fixed point in X,
show that the intersection of all basic open sets which contain
x0 is the single~lemeot set {xo} J and conclude from this that the
complement of ~x 0 } is countable.)
Show that the set of all isolated points of a second countable space is
empty or countable. Show from this that any uncountable subset
A of a second countable space must have at least one point which is a
limit point of A4
Prove in detail that the open rectangles in the Euclidean plane form
an open base,

5,

6.

7.

8.

104

9"

Topology

Let f :X ~ Y be a mapping of one topological space into another.

f is said to be continuous at a point xo in X if for each neighborhood
H of f(xo) there exists a neighborhood G of xo such that f(G) C H.
(a) Sho"T that f is continuous=:? it is continuous at each point in . .X.
(b) If there is given an open base in Y J sho'v that f is continuous
at xo {::::} for each basic open set B "-hich contains f(xfJ) there
exists a neighborhood G of x 0 sur.h that /(G) c B.
(c) If Y is a metric space, show that f is rontinuous at xo ~ for
each open sphere Sr(/(xo)) centered on f(xr;) there exists a
neighborhood G of xo such that f(G) C Sf"(f(x<J))~

19. WEAK TOPOlOGIES

Let X be a non-empty set~ If T1 and T2 a.re topologies on X sur h

that T1 C T2, we say that T1 is weaker than Ti (or T~is slronger than T1).
In rough terms, one topology is weaker than another if it has fc'\\rcr open
sets, and stronger than another if it has more open sets. The topology
f 0,X] is the weakest topology on X, for it is weaker than every topology;
and the discrete topology is the strongest topology on X, since it is strongPr
than every topology. It is clear that the family of all topologies on Xis a
partially ordered set "\\Tith respect to the relation ''is weaker than+ ~
1

\\,.. e next show that this partially ordered set is a complete lattice.
In Problem 16-1 v..,.e asked the reader to prove that the intersect.ion of any
t\vo topologies T1 and T:2 on X is a topology on X. Since this topology is
evidently l\Teakcr than both T1 and r~ and stronger than any topology
\vhich is weaker than both~ it is the greatest lower bound of T1 and T~ It
is cq u ally easy to see that the in terser tion of any non-empty family of
topologies on X is a topology on X; and since it is "?t,.eaker than all ,;l;ht-~sc
and stronger than any topology which is weaker than all these, it is the
greatest lol\Ter bound of this family~ What nbout least upper bounds?
The si tua ti on here is a hit differen tt for the union of two topologies on
X need not be a topology. llowever, if \Ve have any non-empty family
of topologies T.:, then the discrete topology is a topology strongrr than
each Ti. \Ve can therefore appeal to our above remarks to conclude that
the intersection of nll topologies "?t'hich are 8tronger than each T..: is a
topology; and since it is stronger than each T:t and vireaker than any
topology \vhich is stronger than each Tf, it is t.he least upper bound of
our given family+
We summarize the results of this discussion in the following theorP.n1.

Theorem

on X

i.~

A. I.tet X be a non-empty set

1Jhen the fa niil y of all top oloy ie8

a com.plete latti.c.e with respect lo the relation "~is uieaker than.'~

Topological Spaces

105

Jl urtherrn.ore, this lattice has a least nu~ m.ber (the weakest topology on X) and
a greatest member (the discret.e topology on X) .
The reader vlill observe that if {T..:} is a non-empty family of topologies
on our set X, then the least upper bound of this family is precisely the
topology generated by the class Vil, in the sense of Theorem 18-D; that
is, the class ViTi is an open subbase for the least upper bound of the
family {T, l . l n the present contex.tJ therefore, Theorem 18-D can be
thought of as providing a mechanism for the direct construction of least
upper bounds in our lattice of topologies.
Let X be a non-empty set~ let {X 1} be a non-empty class of topological spaces~ and for each i let fi be a mapping of X into X... It is clear
that if X is given its discrete topology,, then all the f/s are continuous~
If 've look a little further, we may find other and weaker topologies on
X which also have this property. There is, in fact, a unique weakest
1.o p ology of this kind. The weak topology gene rated by the f / s is defined to
be the intersection of all topologies on X 'vith respect to each of vlhich all
the f/s are continuous mappings. This js clearly a topology on X v;hich
makes nll the f/s continuous, and it is weaker than any topology which
has this property. It V--'ill appear in later chapters that many a topology
which is used in practice is defined to be the weak topology generated by
some set of mappings of particular interest in a given situation.

Problems

1.

2.

I.Jet X be a non-empty set and {X j J a non-empty class of topological

spaces. If for each i there is given a mapping f;. of X into Xt:. denote
by T the weak topology on X generated by the //s~
(a) Shov_r that T equals the topology generated by the c]ass of all
inverse images in X of open sets in the X,/s.
(b) If an open subbasc is given in each Xi, show that T equals the
topology generated by the class of all inverse images in X of
subbasic open sets in the Xis.
(c) If Y is a subspace of the topological space (X,T), show that
the relative topology on Y is the weak topology generated by
the restrictions of the f/s to Y.
In each of the following '\\re specify a set ~ fi J of real functions defined
on the real line R ~ In each case give a complete description of the
v.,.eak topology on R generated by the j/s.
(a) {fi} consists of all constant functions.
(b) lfil consists of a single function f~ defined by f(x) = 0 if x ~ 0
and f (x) == I ii x > 0 ~

106

Topology
(c)
(d)
(e)

(f)

{ft:} consists of a siriglf, function Ii defined by /(x) = -1 if

z < 0,/(0) ~ 0 1 andf(x) = 1 if x > 0.
{/ii consists of a .single function f, defined by f(x) ~ x for all x+
tit} consists of all bounded functions which are continuous
'With respect to the usual topology on R.
{/,:} consists of all functions which are continuous with respect
to the usual topology on R.

20. THE FUNCTION ALGEBRAS e(X}R) AND e(X~C)

Let X be an arbitrary topological space.

\Ve generalize the notations established in Sec. 14 by defining e(X,R) and e(X)C) to be the sets
of all hounded continuous functions defined on X ~Thieh aret rcspectivclyJ
real and complex~
It is desirable to extend our discussion of the algebraic structure of
these set.s beyond that given in Sec. 14 by introducing the following
concepts. An algebra is a linear space \vhosc vectors can be multiplied
in such a vlay that
(1) x(yz) ~ (xy)z;
(2) x(y
z) = xy
xz and (x
y)z = xz + yz;
(3) a(xy) = (~x)y = x(a.y} for every scalar a.
\\:'" .e speak of a real algebra or a cnmple:c algebra according as the scalars
arc the real n uni bers or the complex n um hers. A commutative algebra is an
algebra v..-hosc multiplication satisfies the f ollov-.ing condition:
(4) xy = yx.
In the case of a commutative algebra~ the second part of {2) is clearly
redundant.. 1\n algebra with dentity is an algebra which possesses the
follow"ing property;
( 5) there exists a. non-zero clement in the algebra, denoted by I and
r.alled the identity element (or the identity)t such that
1 x == x 1 ~ x for every x.
Vt7 c speak of the identity because the identity in an algebra (if it. has one)
is unique; for if 11 is also a.n clement such that 1' x = x ~ l' = x for
Pvery x, then 1 ~ = 1' 1 = 1.. A subalgebra of an algebra is a linear
subspace "\\-Thi ch contains the product of each pair of its elements+ A
subalgcbra of an algebra is evidently an algebra in its own right.
In the case of a function space which is also an algebra, it is to be
underst-0od that multiplication is defined pointv,ise, that ist that the
product fg of t\vo functions in the spare is defined by (jg) (x) = f(x)g(x).
This point~vise multiplication off unctions should be clearly distinguished
from the mu1tip1icntion (or composition) of mappings discussed at the
end of Sec. 3. If such an algebra has an identity element 1, then Problem

Topological Spaces

107

1 shows that in all cases of interest to US this identity is the constant

function defined by l(x) == 1 for all x.
We prove two lemmas before going on to our main theorems~
If f and g are cantinuous real or complex functions defined on a
topological space Xt then j
gt af, and Jg are also continuous. Furthermore, if f and g are realJ then f "' g and f v g are continuous .
PROOF~ We illustrate the method by showing that fg and f v g are

lemma.

continuous.
V.le prove thatfg is continuous by sho,ving that it is continuous at an
arbitrary point xo in X (sec Problem l 8-9)+ Let E > 0 be given, and
find E1 > 0 such that ie1{lf(xD)l + lg(xo)1) + E1 '2 < ic;. Since f is continuous, and thus continuous at XD~ there exists a neighborhood G1 of
x 0 such that x e G1 ~ l/(x) - f(xo)! < t:i+ SimilarlyJ there exists a
neighborhood G2 of xo such that x e: G2 ;:9 fg(x) - g(x~)I < E1. The
continuity of fg at xo now follows from the fact that G = G 1 ('.. G2 ia a
neighborhood of xo such that
x

G ~ ~ (fg) (x) - (fg) (xo) I =:: If (x )g(x) ~ f (xo)g(xo) ~

= I[/(x )g(x) - /(x )g(xo) J
(x )g(x6) - f (xo)g (xo)] I
< If (x) i Ig(x) - g(xo) I lu(xo) 1If(x) - f(x6) r < fill(x) r Et lu(xo) j
= 11 {f(x) - f (xo)]
f (xo) I E1lg (xo) I ~ E1 If(x) - f (xo) I + E1l/(xo) I
2
if 1lg (xo) j < E1 (1 f (xo) I + Ig(x~) I) + f:1 <

+ [/

f.

We prove that! v g is continuous by recalling that all sets of the form

A == (a 1 +co) and B = {-co ,b) form an open sub base for the real line
and by showing that the inverse image of any such set is open (see
Theorem 18-E). All that is necessary is to observe that
(f v g)- 1(A) = {x :max {f(x),g(x) I

> al ==

{x :f(x)

> a'

Ix: g(x) > a I,

which is open since it is the union of two open sets, and that

(f v g)- 1( B) =- {x : max {f (x), g(x) l

<

b}

;::=

f x :f (x)

< b i ('.. {x : g(x) <

b} ,

which is open since it is the intersection of two open sets.

Let X be a topological space, and let {f n.1 be a sequence of real

or complex functions defined on X which converges uniformly to a function f
defined on X~ If all fhe !n's are continuous, then f is also continuous.
PROOF~ We show that f is continuous by sho'Wing that it is continuous
at an arbitrary point x 0 in X. Let E > 0 be given. Since/ is the uniform
limit of the f ,,,'sJ there exists a positive integer no such that 1f (x) - f ~r1(x) I <
oe/3 for nil points x in X. Since f R is continuous, and thus continuous at
Xo, there exist8 a neigJl borhood G of Xo such that x e G -==} If~ (x) Lemma~

108

Topology

f ~(zo)I < E/3. The continuity off at xo now follo"s from the fact that
x

e. G ~ !f(x) - f(xo) l
= I[f(x) - f (x)] + [/n.(x) 5 !J(x) - f (x) I + If~(x) a1

21 1

fn~(xo)]
f~ 0 (xc.) I

+ l/n~(xo)
+ IJn.a(xo)
<

- f(XG)] I
- f(xo);
.:/3 + oe/3 + t/3

This lemma is often stated more informally ns f olJo\vs: any uni/Mm

limit of continuou8 functions is continuous.
We are now in a position to give Theorem 14~A the f ollo,ving broader
and richer form.
Theorem A. Let e(X ~R) be the set()! all bounded continuous real functions
defined on a topological space X. Then (1) e(X,R) is a real Banach space
'With respect to pointwise addition and scalar multiplication and the norm
defined by l!fli = sup j/(z)!; (2) if multiplication is defined pointwiseJ
e(X,R) is a commutative real algebra with identity in which llfgtl :$ II/I) llgll
and H1 II = 1; and (3) if f =::;; g iB defined to mean that f(x) < g(x) for all x,

e (X ~R) is a lattice in which the greatest lower bound and lea.st upper bound
of a pair of functions f and g are given by (f" g)(x) :;:::: min lf(x) ,g(x)} and
max {/(x}Jg(x) i.
PROOF. In view of the above lemmasJ everything stated here is clear,
except perhaps the fact that ~]fgH ~ l!fll llgl[; and this follows from

(f v g) (x)

Hfgll

sup I(fg)(x)~ ~ sup lf(x)g(x)l ;;::; sup IJ(x)l 1u(x)I

< (sup if(x) I) (sup [g(x) I) ~

tiff I HY fl.

We also extend Theorem 14-B, but in a slightly different direction.

Theorem B.. Let e(X,C) be the set of all bounded continuous complex fun.ctions defined on a topological space X. Then (1) e (X ,C) is a com plc.r
Banach space with re8pect tn pointwise addition and scalar nlultiplication
and the norm defined by 11111 ~ sup lf(x) I; (2) if mulliplica.tion is de.fined
pointwise, e(X,C) is a commutative com.plex algebra u,ith identify in wlvich
ll!uH < fl/II !lall and 111 l ;::= 1; and (3) if is defined by ](x) = f(x) t then
f ----;) is a mapping of the algebra e( X ~ C) into itself which has the following
properties.~ f
g = J
g, af = a !1 fg = f ~ g1 j = f, and 11111 = l!JH.

This theorem is a. direct consequence of the background provided above. We do remark, however~ that the fact thatjis rontinuous
\Vhen f is follows from 1J(x) ~ ](xo) I = if(x) ~ f (xo) I+
PROOF.

The function J defined in this theorem is called the conjugate of the

function f, and the operation of forming J from f is called conjugation.
It will become clear in the later chapters of this book that the operation of
conjugation in the space e(X,C) iB one of t.he chief supporting pillars
of the theory we develop in those chapters.

Topological Spaces

109

We trust that the reader has noticed our insistence on assuming that
a topological space always has at lea.st one point~ Our reason for this is
that the empty set has no functions defined on it. If "Te were to allow a
topologir.al space X to be empty, then we would ha.ve t.o cope with the
fact that its corresponding e(XJ/l) and e(X 1 C) are also empty, and so
cannot b.c linear spaces, for a linear sp~e must contain at least one vector
(the zero vector).. Since constant functions are aJ\vays continuousJ "Te
avoid this difficulty by ta.king pains to assume that to polo gic al spaces
are non-empty .
Problems

1.

2..

3..

Let A be an algebra of real or complex functions defined on a nonempty set XJ and assume that for each point x in X there is a. function
f in A such that f(x) ~ 0. Sho"T that if A contains an identity
clement 11 than 1 (x) = 1 for all x.
I~t.f be a continuous real or complex function defined on a topological
space X 1 and assume that f is not identically zer:o, Le.~ that the set
Y = {x :f(x) 0 l is non-empty. Prove in detail that the function
1/f defined by (1//)(x) = 1/f(x) is continuous at ear.h point of the
subspace Y.
Let X be a topological space and A a subalgcbra of e(X,Jl) or e(XJC) .
Show that its closure A is also a subalgcbra. If 11 is a subalgcbra
of e(X,C) which contains the conjugate of each of its functionsJ
show that A also contains the conjugate of each of its functions.

CHAPTER FOUR

eompae!HCSS

Like many other notions in topology, the concept of compactness

for a topological space is an abstraction of an important property possessed by certain sets of real numbers.. The property vte have in mind
is expressed by the H einc-Borcl theorem~ which as..9erts the following:
if X is a closed and bounded subset of the real line R, then any class of
open subsets of R whose union contains X has a finite subclass whose
union also contains X. If we regard X as a topological space in its own
right~ as a subspace of RJ then this theorem can be thought of as saying
that any class of open subsets of X whose union is X has a finite subclass
whose union is also X.
The Heine-Borel theorem has a number of profound and far-reaching
applications in analysis. Many of these guarantee that continuous
functions defined on closed and bounded sets of real numbers are well
behaved~ For instance., any such function is automatically bounded
and unif onn ly con tin nous. In contrast to this satisfying be ha vi or
we note that the function f defined .on the open unit interval (Orl) by
f(x) = l/x is neither bounded nor uniformly continuous.
~ is of ten the case with crucial theorems in analysis~ the conclusion
of the Heine-Borel theorem is converted into a definition in topology.
This definition singles out for special attention what are called compact
topological spaces. Our main business in this chapter is to develop the
basic properties of these spaces and the continuous functions they carry,
and, in the case of metric spaces, to establish several equivalent forms of
compactneSB which are useful in applications.
t

110

Compactness
21~

111

COMPACT SPACES

Let X be a topological space. A class i Gi} of open subsets of X is

said to be an open cover of X if each point in X belongs to at least one Gi"
that is, if Vt.-Gi. = X~ A subclass of an open cover 'vhich is itself an
open cover is called a subcot~er. A conipact space is a topological space
in which every open cover has a finite subcover~ A compact subspace of a
topological space is a subspace "\ivhich is compact as a topological space
in its own right. We begin by proving two simple but widely used

theorems.

Any closed subspace of a compact space is compact

PROOF.
Let Y be a closed subspace of a compact space X, and let {G*}
be an open cover of Y~ Each Gi~ being open in the relative topology
on Y~ is the il)tcrscction with Y of an open subset Hi of X. Since Y is
closed_, the class composed of Y' and all the l//s is an open cover of X,
and since X is ~ompact, this open cover has a finite subcover. If Y'
occurs in this su bcovcr, we discard it~ What remains is a finite class of
H/s whose union contains X4 Our conclusion that Y is compact now
follows from the fact that the corresponding G/s form a finite subcover
of the original open cover of Y+

Theorem A.

Any continuous image of a compact space is compact

PROOF.
Let f:X ~ Y be a continuous mapping of a compact space X
into an arbitrary topological space Y. We must show that f(X) is_ a
compact subspace of Y. Let {G.d be an open cover of f(X). As in the
above proof, eacl1 Gi. is the intersection with f(X) of an open subset Hi
of Y.. It is clear that [f- 1(//1)} is an open cover of X, and by the compactness of X it has a finite suhcover~ The union of the finite class of 11/s
of which these arc the inverse images clearly contains f(X), so the class
of corresponding G/s is a finit.e subcover of the original open cover of
/(X), and f(.L) is compacL
Theorem

B.

It is sometimes quite difficult to prove that a given topological space

is com par.t by appealing directly to the definitionr The f ollovdng theorems give several equivalent forms of compactness which are often easier
to apply.
Theorem Ct A topological space is compact ~ every class of closed sets
with empty intersection has a fl ni te subclass with empty intersection+

This is a direct consequence of the fact that a class of open

sets is an open cover ~ the class of all their complements has empty
intersection.

PROOF+

112

Topology

We recall from Problem 8~6 that a class of subsets of a non-empty

set is said to have the finite intersection properly if every finite subclass
has non-empty intersection. This eoncept enables us to express Theorem C as follo\VS~
Theorem D.. A topological Sj)ace 1~s compac.t <=> every class of closed sets
wi.th the finite intersection property has non-empty intersection.

Let X be a topological space. An open cover of X whose sets are a.11

in some given open base is called a ba.sic open cover, and if they all lie
in some given open subbase_, it is called a subbasic open coter. \Ve observe
the trivial Iact that if X is compact, then every basic open rover has a
finite subcover+ Our next theorem asserts that compactness not only
implies this property, but is also implied by it.
Theorem E. A topological Bpac.e is co mp act if every basic open caver has a
fl n ite sub cover
PROOF
Let {Gd be an open cover and {Bi} an open base. Each G1 is
the union of certain B/s, and the totality of all such B/s is clearly a basic
open cover. By our hypothesis, this class of B/s has a finite subcover~
For each set in this finite subcover le can select a Gt which contains it.
The class of G/s whlch arises in this way is evidently a finite subcover
of the origin.al open cover.
L

We go one more step in this direction and prove a similar (and much
deeper) theorem relating to .subbasic open covers~ The proof is rather
difficult, and 've introduce the following concepts in an effort to make
it as .simple as possible.. They also make some of its applications considerably easier to handle. Let X be a topological space. A class of
closed subsets of X is called a closed base if the class of all complements
of its sets is an open base~ and a closed subbase if the class of a.11 eomplements is an open subbaseL Since the class of all finite intersections of
sets in an open subbasc is an open ba.seJ it follows that the class of all
finite unions of sets in a closed subbase is a closed base. 1"'his is called
the closed base generated by the closed subbase.

A topological space is cornpact if every subbasic open cover

has a fl n ite Bubcover, or equivalent! y, if every class of .su bbasic closed sets
uith the finite inter.section property has non-empty intersection.
PROOF.
The equivalence of the stated conditions is an easy consequence
of Theorems C and D4 Consider a closed subbase for our space 1 and
let ! Bt:} be its generated closed base, that is_, the class of all finite unions
of its sets. We assume that every class of subbasic closed sets with the
finite intersection property has non-empty intersection~ and we prove
from this that every class of B/s with the finit.e intersection property

Theorem F.

Compactness

also has non...empty intersection.

113

By Theorem E, this will suffice to

prove our theor.em~

Let I Bi} be a class of B/ s ~~i th the finite intersection property.
We must show that ('.iB; is non-empty+ We use Zorn's lemma to show
that {Bi} is contained in some class I Bk} of B/s which is 1naximal with
respect to having the finite intersect.ion property, in the sense that {Br.;}
has this property and any class of B/s which properly contains {Bk}
fails to have this property. The argument runs as follo,vs. Consider
the family of all classes of B, 1 s l\'~hich contain {Bi~ and have the finite
intersection rope rty. This is a partially ordered set '"d th respect to
class inclusion. If we consider a chain in this partially ordered set, the
union of all classes in it is a class of B./ s which contains every member
of the chain and has the finite intersection propertyJ as we see from the
-fact that every finite class of its ~ts is contained in some member of the
chain, and that member has the finite intersection property. We con-elude that every chain in our partia1ly ordered set has an upper bound,
.so Zorn's lemma guarantees that the partially ordered set has a maximal
element. This argument yields the existence of a class {Bl-} 'With the
properties stated above. Sin Ce nk Bk ~ n.iBj~ it no'v su:ffi ccs to show
that nkBk is non-empty.
Each B1i; is a finite union of sets in our closed subbase_, for instance1
B1 = 8 1 U 82 U U Sn. It now suffices to show th.at at least one
of the sets 81, s~t
~ ~ S,. belongs to the class {B1:-} ~ For if 'vc obtain
such a set for each Bt., the resulting class of sub basic closed sets will have
the finite intersection property (since it iB contained in {Bk l) ~ and therefore, by our hypothesis relating to the sub basic closed sets, it will have
non-empty intersection; and since this non-empty intersection 'Will be a
subset of nkB1:J we shall kn.ow that r\tB1c is itself non-empty~
\Ve finish the proof by showing that at lea.st one of the sets 81,
S2J .
~
.Sn does in fact belong to the class {Bk}. We assume that
each of the~ sets is not in this class, and we deduce a contradiction from
this assumption. Since 8 1 is a subbasic closed set, it is also a basic
closed set; and since it is not in the class {Bk L the class {Bk~S1} is a class
of B/s 'Which properly contains {Bt.}. By the maximality property of
{Bk} J the class {Bk~S 1} lacks the fi.ni te intersection property, so Si is
disjoint from the intersect.ion of some finite class of B1 1s+ If we do this
:for each of the sets S1~ S2, ~ . ~ ~ Sn, we see that B 1~the union of these
setg-..-is disjoint from the intersect.ion of the total finite class of all the
B1:.'s which arise in this l\ray.. This contradicts the finite intersection
property for the class {Bk} and completes the proof.

...

The great power of this theorem can be surmised from the complexity
of its proof~ It is really a tooC and we illustrate the manner in which

114

Topology

it can be used by applying it to give a simple proof of the classical Heine ...
Borel theorem stated in the introdur.tion to this chapter.
Theorem G (the Heine-Borel Theorem).

Every closed and bounded S'Ub-

space of the real line is compact.

PROOF~

A closed and bounded subspace of the real line is a closed subspace of some closed interval [a~b]. and by Theorem A it .suffices to show
that {aJb] is compact. If a = bJ this is clear, so we may assume that
a < b. By Sec. 18, we know that the class of all intervals of the form
[a~d) and (c,b], """here c and d are any real numbers such that a < c < b
a.nd a < d < bt is an open sub base for [a,b]; therefore the class of all
fa~c]'s a.nd all [d~brs is a closed sub base. Let S = { [a,ci], {dhb] l be a
class of these subbasic closed sets with the finite intersection property.
It suffices by Theorem F to show that the intersection of all sets in Sis
non-emptyr We may assume that Sis non-empty. If S contains only
intervals of the type [a~c-.:], or only intervals of the type [di 1 b}, then the
intersection clearly contains a or b. We may thus assume that S contains intervals of both types. We now define d by d == sup {di L and
we complete the proof by shov~dng that d < Ct: for every i. Suppose that
c~. < d for some irJ. Then by the definition of d there exists e. d;, such
that Ct:, < di,+ Si nee [a ~ci.ri] f\ (d;Q Jb] = 0, this contradicts the finite
intersection property for S and concludes the proof.

The reader should understand that there are elementary proofs of the
Heine-Borel theorem which do not use Theorem F or anything like it.
Theorem F wil1 render us its major sen.ice in connection with the proof
of the vital Tye honoff theorem of Sec. 23 ~
Problems

1.

2.

3.

A countably compact space is a topological space in which every count..

able open cover has a finite subcover. Prove that a second countable
space is countably conlpact ~ it is compact.
Let Y be a subspace of a topological space X
If Z is a non-empty
subset of YJ show that Z is compact as a subspace of Y ~ it is comp.act as a subspace of X.
l~et X be a topological space. If {Xd is a non-empty finite class
of compact .subspaces of X 1 show that ViX..: is also a compact sub ...
space of X. If {Xj} is a non-empty class of compact subspaces of X
each of which is closed, and if n.1 1 is non-empty 1 show that
is also a compact subspace of X .
Let X be a compact space.. We know by Theorem A that every
closed subspace of X is compact~ By considering Example 16-3,
show that a compact subspace of X need not be closed .
+

4.

n,.x,

Compactness

S.
6.

7.

8~

9t

115

Prove the converse of the Heine-Borel theorem~ every compact subspace of the real line is closed and boundedr
Generalize the preceding problem by proving that a compact subspace of an arbitrary metric space is closed and bounded. (It should
be carefully noted, as Secs. 24 and 25 will show, that a closed and
bounded subspace of an arbitrary metric space is not necessarily
compact.)
Show that a continuous real or complex function defined on a compact space is bounded+ More generally, sho-\v that a continuous
mapping of a compact space into any metric space is bounded4
Sho"; that a continuous real function f defined on a compact space X
attains its infimum and its supremum in the f olloVrTing sense: ii
a = inf ~ f(x) :x EX I and b = sup {/(x): x e X} J then there exist
points x1 and x2 in X such that f (x1) == a and f (x2) = b.
If X is a compact space) and if itti l is a monotone sequence of continuous real functions defined on X which converges pointwise to a
continuous real function f defined on X, sho'v that fn converges
uniformly to/. (The assumption that ifn l is a rnonotone sequence
means that either !1 :::; j,. ::; fa ~ or f 1 > !2 > fa > ~)

22. PRODUCTS OF SPACES

There are two main techniques for making new topological spaces
out of old ones+ The :first of t.hese 1 and the simplest~ is to form subspaces
of some given space. The second is to multiply together a nun1ber of
given spaces. Our purpose in this section is to describe the way in which
the latter process is carried out.
In Sec~ 4 we defined what is meant by the product P*X' of an arbitrary non-empty class of sets. We also defined the proj cction Pi of this
product onto its ith coordinate set X;:. The reader should n1akc certain
that these concepts are firmly in mind. If each coordinate set is a
topological spacet then there is a standard method of defining a topology
on the product. It is difficult to exaggerate the importance of this
definitionJ and we examine it with great care in the following disc-ussion~
Let us begin by recalling the discussion in Sec. 18 of open rectangles
and open strips in the Euclidean plane R'lr We observed there that the
open rectangles form an open base for the topology of R 2 ~ and also that
the open strips form an open subbase for this topology whose generated
open base consists of all open rectangles, all open strips, the empty set,
and the full space. The topology of the Euclidean plane is of course
defined in terms of a metricr If we vrish, howeverJ we can ignore this
fact and regard the topology of R' as generated in the sense of Theorem

116

Topo~ogy

18... D by the class of a.II open strips~ This situation provides the motivation for the more general ideas we now develop .
Let X 1 and X2 be topo1ogical spaces~ and form the product
X = X 1 X X 2 of the two sets Xi and X 2~ Consider the class S of all
subsets of X of the form G1 X X2 and X1 X G~, "'~here G1 and G! are
open subsets of X 1 and X 2~ respectively+ The topology on X generated
by this class in t.he sense of Theorem 18-D is called the product topology
on X~ The product topology therefore has S as an open subhase; in
fact, it is defined by the requirement that S be an open subbase~ The
open base generated by S, that is, the class of all finite interscr.tions of
its sets, is clearly the class of all sets o. the form G1 X G2~ and th.e open
sets in X are all unions of these
sets~ There are two projections
p1 and p~ of X onto its coordinate
spaces X1 and X2 1 and by definition they carry a typical element
(x1)x:i!) of X to xi and X2~ respectively.. We note that Sis prec.isely
the class of au inverse images
in X of all open subsets of X 1
and X2 under these projections:
G1 X X2 = p1- 1(G1) and X1 X G2 =
X1
p-i- 1 (0 2 ). The product topology is
Fig. 23. The product topology on
thus a topology on the product vr.ith
X 1 X X :.~
respect to which both projections
are continuous mappings~ and it jg
evidently the weakest such topology. In terms of the ideas discussed in
Sec4 19, the product topology- can be regarded as the weak topology
generated by the projections~ Figure 23 may assist the reader in vis-

ualizing some of these notionf;,.

With the above concepts to guide the way 1 one more step carries us
to the product topology in its iull generality. Let {X..:} be any nonempty class of topological spaces, and consider the product X = P1Xi
of the sets X,.. A typical element x in Xis an array x = {xd of points
in the coordinate spacest where each Xi be1ongs to the corresponding
space Xf; and for each index i, the projection pj is defined by Pi.(x) = x,:.
We now define the product topology on X to be the weak topology gen.crated by the set of all projections~ This means the product topology
is that generated by the class S of all inverse images in X of open sets
in the X/s~ that is, the class S of all subsets of X of the form S = Pi- 1 (Gi),
\vhere i is any index and Gt is any open subset of X;,. It is easy to see
that S can also be described as the c]ass of all product~ of the form
j_'; = Pt.Gi, vrhere Gt. is an open subset of X,, which equals X, for all i's but
one. The class S is called the defining open subbase for the product

Compactness

117

topology; and the class of all complements of sets in 5--namely, the

class of all products of the form P,.Fi, where F1 is a closed subset of Xi
which equals X.: for all iTs but one-is called the defining closed subbase~
The open b Me generated by S, that is, the cl ass of all finite inter.sections
of its sets, is called the defining open base for the :product topology; and
this is evidently the class of all products of the form Pt.G;, where G,: is an
open subset of Xi which equals X" for all but a finite number of i's. It
should be clearly understood that an unrestricted product of open sets in
the coordinate spaces need not be open
in the product topology. A convenient 1
way of thinking about this defining open
base is that a typical one of its sets consists of all points x = {Xt. 1 in the prod~
uct such that the ith coordinate Xi is
required to lie in an open subset Gi of
for a finite number of i'sJ all other
coordinates being unrestrf r,ted.
When the product of a non-empty
~
l
class of topological spaces is equipped
I
with the product topology defined in the
0
1
above paragraph, it is called a product
8pace, or more simply, the product of the Fig. 24. A set in the defining open
spaces involved~ 1 It should be clear base for a product space.
from Theorem 18-E and these definitions
that all projections of a product spaee onto its coordinate spaces are
automatical]y both continuous and open+
We conclude this section by analyzing an example which we hope vrill
increase the reader~s capacity to '(see" the structure of product spaces.
Let the index set I consist of all real numbers i in the closed unit interval
[0,1]. I is to be considered as a set without any structure4 Now let
each index i have attached to it a topological space xi, and let every
be a replica of the closed unit interval 10~1] with its usual topology. The
resulting product space X = P 1Xi is illustrated in Fig. 24. The base
of this figure is the index set I; and each vertical cross section represents
the coordinate space Xi attached to the index at its base4 An element
of the product space X is an array of points, one of which lies in each Xi~
Such an element is essentially a function-if we identify a function with
its graph-defined on the set I, with values in the closed unit interval~
We can now visualize as follows a typical set in the defining open base for
the product topology. We choose a finite set of indices, say {i lJ i2j i3 L
and Ior each of these we choose an open set on the vertical segment above

xi

xt

a topological space muat first of all be a. non-empty set, it is worth remarking here that the axiom of choice guarantees that the product of a non-empty clAea
of non-empty sets is non.empty.
1 Since

118

Topology

it. Our basic open set then consists of all functions in X whose graphs
cross each of these three vertical segments within the given open set on
that scgn1ent. In the figure, f belongs to our basic open set~ but g does
not. The product topology on any product space can be visualized in a
simila T way~ Al 1 one has to do is imagine the coordinate spaces as fibers,
each attached to a specific element of the index set. The resulting
mental image of the product space will then look something like a bundle
of fibers, or perhaps a hed of reeds growing in a pond~
Problems

1.

2.

3.

4.

5.

AU proj ections~ being open mappings, send open sets to open sets4
Use the Euclidean plane to sho\v that a projection need not send
closed sets to c1osed sets.
Show that the relative topology on a suhspace of a product space is
the v.r~eak topology generated by the restrictions of the projections
to that subspace.
1.JCt f be a mapping of a topological space X into a product space
P,X", and show that f is continuous ~ pJ is continuous for each
projection P'-
Con8ider the product space defined and discussed in the last para...
graph of the text, and show that this space is not second countable.
(Hint.+ recall Theorem 18-B, and observe that the index set is uncountably infinite~)
Let X, Y, and Z be topological spaccs 1 and consider a mapping
z = f(x y) of the product set X X Y into the set Z. We say that
f is continuous in x if for each fixed Yo the mapping of X into Z
given by z = f(x,yo) is continuous. The statement that f is continuous in y is defi11ed similarly. f is said to be }ointly con.tinuous in x
and y if it is continuous as a mapping of the product space X X Y
into the space Z.
(a) If all three spaces are metric spaces, show that j is jointly
continuous ~ x~---+ x and Yn---+ y implies f(xn~Yn) ~ f(xiy).
(b) Show that if f is jointly continuous, then it is con tin nous in each
variahle separately. Show that the converse of this statement
is false by considering the real function defined on the Euclidean
plane by f(x,y) = xy/(x 2
y 2} and f(O,O) = 0.
1

23. TYCHONOFf S THEOREM AND lOCALLY COMPACT SPACES

1

The main theorem of this section, to the effect that any product of
compact spaces is compact} is perhaps the most important single theorem

Compactness

119

of general topology.

We shall use it repeatedly throughout the rest of

this book, and the reader "\\iill come to see that its commanding position
is due largely to the fact that in the higher levels of our subject many
spaces constructed for special purposes tum out to be closed subspaces of
products of compact spaces. Such a subspace is necessarily compact, and
since compact spaces are so pleasant to work with, this makes the
resulting theory much cleaner and smoother than would otherwise be the
case~
1

The product of any non-empty cla88

Theorem A (Tychonoff s Theorem).

of compact spaces is compact.

PROOF. Let {X..:} be a non-empty class of compact spaces, and form the
product X = P1Xi4 Let {F1} be a non-empty subclass of the defining
closed subbase for the product topology on X.. This means that each
Fi is a product of the form Fi :;: ;:; P Fii, where Fti is a closed sub.set of Xi
which equals X1 for all i 1 s but one~ We assume that the class {F;l has
the finite intersection property, and by virtue of Theorem 21-F we conclude the proof by showing that rliFi is non-empty. For a given fixed
i) ~ FH} is a class of closed subsets of xi with the finite intersection
property; and by t~e assumed compactness of X, (and Theorem 21-D),
there exists a point x..: in X1 which belongs to fl 1Fii If we do this for
each (, we obtain a point x = {x.,:l in X which is in rl;Fj.
As our first application of Tychonoff's theorem, we prove an extension of the classical Heine-Borel theorem. We prepare the way for this
proof by defining v-.That we mean by open and closed rectangles in the
n-dimensional Euclidean space Rn. If (a..:-~bi) is a bounded open interval
on the real line for each i = 1, 2, . . . , n, then the subset of Rn defined by
Pt_1 (ai,b,:)

(x1,

X2,

is called an open rectangle in RJI ~

a product of n closed in tervalsa

Xn): a~

< Xi < bi for each ii

A closed rectangle is defined similarlyJ as

Theorem B (the Generalized Heine-Borel Theorem).

Every closed and

bounded 8ubspace of Rs is co1npact.

PROOF. A closed and bounded subspace of Rn is a closed subspace of
some closed rectangle, so by Theorem 21-A it suffices to show ~hat any
closed rectangle is compact as a subspace of Rn. Let X = P:_1 [Ui ,b,] be
a closed rectangle in R~. Each coordinate space [a1,b,J is compact by the
cl a.ssi cal Heine-Borel theorem, so by Tych onoff1s theorem it s nffices to
show that the product topology on Xis the same as its relative topology
as a .subspace of R
It is easy to see that the open rectangles in Rn form
an open hase for its usual topology~ that is~ for its metric topology 1 and
from this it follows that the product topology on Rn is the same as its
11

120

Topology

usual topology. By Problem 22-2, the relative topology on X is the

weak topology generated by its n projections onto the coordinate spaces
[a..:,b;]; but this is the product topology on xt so the proof is complete . 1
Then-dimensional Euclidean space R"' is the most important example
of a type of topological space which is of great significance in modern
analysis, especially in the theory of integration. A topological space is
said to be locally compact if each of its points has a neighborhood with
compact closure. It is easy to see by the above theorem that Rq actually
is local]y compact, because any open sphere centered on any point is a
neighborhood of the point whose closure 1 being a closed and bounded
8ubspacc of R"* 1 is com pact. It is trivial that any compact space is
locally compact~ for the full space is a neighborhood with compact
closure of every point in the space. We return to the study of locally
compact spaces in Sec.. 37, where '"e give a more detailed analysis of
their structure and properties.
Problems

1.

2..
3.
4~

Prove in detail that the open rectangles in n~ form an open base.

Show that every closed and bounded subspace of the n-dimensional
unitary space en is compact.
Show that a topological space is locally compact t=> there is an open
base at each point '". hose sets all have compact closures.
Observe that any discrete space is locally compact. Assuming that
there are topological spaces which are not locally compact (we
assure the reader that thls is true), show that a continuous image of a
locally compact space need not be locally compact.

24. COMPACTNESS FOR METRIC SPACES

In all candor~ V{C must admit that the intuitive meaning of compactness for topological spaces is somewhat elusive. This concept~ however,
js so vitally important throughout topology that we consider it worthVtThilc to devote this and the next section to giving several equivalent
forms of compactness for the special case of a metric space~ Some of
these a.re quite useful in applications and are perhaps more directly
comprehensible than the open cover definition. We hope they will help
It is worth remarking that the high-powered machinery used in this proof is
not rea.lly necess&ry for proving the theorem. There are other proofs which are more
cl~m en tary in n a. t ure, but we pref er the one given here be ca use it illustrates i:wme of
our current concepts and too ls.
1

Compactness

121

the reader to achieve a fuller understanding of the geometric significance

of compactness. 1
We begin by recalling the classical B olzano-Weierstras.s theorem: iI
X is a closed and bounded subset of t.he real line~ then every infinite
subset of X has a limit point in X 1'hls suggests that we consider the
property expressed here as one which a general metric space may or may
not possess.. A metric space is said to have the Bolzano-Weier.strass
property if every infinite sub set has a 1imi t point. Another property
closely allied to this is that of sequential compactness: a metric space is
said to be sequentially co1npact if every sequence in it has a convergent
subsequence4 Our main purpose in this section is to prove that each of
these properties is equivalent to compactness in the case of a metric
space~
The following is an outline of our procedure: \Ve first prove that
these two properties are equivalent to one another; next, th.at compactness implies the Bolzano-Weierstrass property; and finally, that
,. sequential
compactness implies compactness. The first two of these steps are
relatively simpl e 1 but the last involves several stages.
+

A metric space is Bequentially compact <==>it has tM BolzanoWeierBlrass properly.

Theorem A.
PROOF~

Let X be a metric spacc 1 and assume first that Xis sequentially

compact. We show that an infinite subset A of X has a limit point.
Since A is infinite, a sequence {Xn I of distinct points can be extracted
from A. By our assumption of sequential compactness, this sequence
has a subsequence which converges to a point x. Theorem 12-A shows
that x is a limit point of the set of points of the subsequence, and since
this set is a subset of A, xis also a limit point of A .
Vrl e now assume that every infinite subset of X has a Jimjt point1
and \Ve prove from this that X is sequentially compact. Let {Xn} be an
arbitrary sequence in X. If {Xn} bas a point which is infinitely repeatedJ
then it has a constant subsequence~ and this subsequence is clearly
convergent. If no point of ~ Xn} is infinitely repeated, then the set A of
points of this sequence is infinite. By our assumption~ the set A has a
li1nit point x~ and it is easy to extract from {xn I a subsequence which
converges to x~
Theorem B.

Every compact metric space ha.s the Bolzano-W eierstrass

property.
Let X be a compact metric space and A an infinite subset of X+
aSBume that A has no limit point, and from this we deduce a con-

PROOF+

' re

A solid case can be made for the pro position that com pa.et spaces a.re natural
g enera.li.z& tio n ~ of spa.c cs with only a finite num her of po in ts. For a discussion of
thia, a.nd of the significance of compactness in ana]yais, see Hewitt [19].
1

122

Topology

tradiction. By our assumption, each point of Xis not a limit point of A~

so each point of Xis the center of an open sphere \vhlch contains no point
of A different from its center. The class of all these open spheres is an
open cover, and by compactness there exists a finite subcover. Since
A is contained in the set of all centers of spheres in this subcovcr, A is
clearly finite. This contradicts the fact that A is infinite, and concludes

the proof.
Our next task is to prove that compactness is implied by sequential
compactness. We carry this out in several stages, the first of which can
be motivated by the fallowing considerations~ Let {Gt:} be an orw.n cover
of a metric space X. Then each point x in X belongs to at least one
G~,. and .since the G,/s are open~ each point x is the center of some open
sphere which is contained in at ]east one G;.+ lf we now move to another
point of X 1 we may be forced to decrease the radius of our open sphere
in order to squeeze it into a G1. Under special circumstances it may not
be necessary to take radii below a certain level as we move from point to
point over the en.tire space. The following concept is useful for handling
this sort of situation.. A real number a > 0 is called a Lebesgue number
for our given open cover {c.} if each subset of X whose diameter is less
than a is contained in at least one Gi~
1

Theorem C (lebesgue s Covering Lemmo)"

In a sequentially compact

-metric space, every open cover has a l..ebesgue number~

Let X be a sequentially compact metric space, and let {Gt} be

an open cover+ W c say that a su bsct of X is "big"' if it is not contained
in any Gi~ If there are no big sets, then any positive real number v:ill
serve as our J.&besguc number a. We may thus assume that big sets do
exist, and we define a1 to be the ~reat.cst lower bound of their diameters.
Clearly, 0 < a' :5 + oo. It will suffice to show t.hat a' > 0; for if
a' = + ~, then any positive real number ~~ill do for a, and if a' is real,
we can take a to be a1 We therefore assume that a' = O~ and "\\""e deduce
a contradiction from thls assumption. Since every big set. must have at
least two points, "'"e infer from a' = 0 that for each positive integer n
there exists a big set. Bn such that 0 < d(Bri) < I/n. We now choose a
point Xn in each Bn~ Since X is seq~entially compact, the sequence
~xn} has a subsequence which converges to some point x in X. The
point x belongs to at least one set G~ in our open cover, and since G4 is
open, x is the center of some open sphere S,.(x) contained in G;G. Let
Sf'r1.(x) be the concentric open sphere with radius r/2. Since our subsequence of fXn} converges to x~ there are infinitely many positive integers
n for which Xn is in S.,.n~(x). Let nn be one of these positive integers
which is so large that 1/n0 < r/2~ Since d(Bnt)) < l/n(l < r/2~ we see by
PROOF.

Comp ac.tness

123

Problem 10-3 that B~~ C Sr-(x) C G1t~ This contradicts the fact that
Bn.,, is a big set, and completes the proof.
The next stage requires the following concepts. Let X be a metric
space+ If e > 0 is givcnt a subset A of Xis called an f:-net if A is finite
and X = U m!A SE(a)~ that is~ if A is finite and its points are scattered
through X in such a way that each point of X is distant by less than
t from at least one point of A.
The metric space Xis said to be totally
bounded if it has an t-net for each E > 0. It is clear that if X is totally
bounded~ then it is also bounded; for ii A is an E-net, then the diameter of
A is finite (since A is a non-empty finite set) and d(X) < d(A)
2E.
'f utal boundedness is actually a much stronger property than boundedness1 as we shall see below.

Theorem D.. Every sequentially compact metric space 'is totally bounded.
PROOF.
Let X he a sequentially compact metric space~ and let E > 0
be given. Choose a point a1 in X and form the open sphere Sf(ai)~ If
this open sphere con ta.ins every point of X, then the single-element set
{ a1 ~ is an f-net~ If there are points outside of S~(ai)J let a 2 be such a
point and form the set S" (a1) V S" (a2).. If this union contains every
point of X, then the t\'vo--element set ia1Ja2} is an E-net+ If we continue in
this ~,.ay~ some union of the form SE(a1) U Se(a 2 ) U V S~(af)) will
necessarily contain every point of X; for if this process could be continued
indefinitely, then the sequence {a1, a!, . . . ~ an, ~ ~ l would be a
sequence -with no convergent subsequence~ contrary to the assumed
sequential compactness of X. We see by this that some finite set of the
form {a1, a2,
~ af)} is an E-net, so X is totally bounded.
+

We are now in a position to complete thiB line of thought by proving

that compactness is implied by sequential compactness~
Theorem E.. Every seq:uentially compact metric spa...ce i8 compact.
PROOF. Let X be a sequentially compact metric space1 and let fGi} be
an open cover. By Theorem C1 this open cover has a Lebesgue number a.
We put E = a/3, and use Theorem D to find an E-net

For each k = 1, 2, . . . , nJ we have d(S~(a,J) < 2E = 2a/3 < a.. By

the de_finition of a Lebesgue numberJ for each k we can find a Gi~ such
that S~ (a"") c r;t.._+ Since every point of X belongs to one of the 81 (a1) Js,
the class {Giu G,,u ~
Gi.} is a finite subcover of {G,}.. Xis therefore
compact.
+

12~

Topofogy

Our results so far can be summarized by the statement that if X is a

metric space, then the following three conditions are all equivalent to
one another:
(1) Xis compact;
(2) X is sequentially compact;
(3) X has the Bolzano-WeierstrasB property.
Also, of course~ l\'~e have as by-products the additional information that a
compact metric space is totally bounded and that every open cover of a
compact metric space has a Lebesgue number.. The latter fact has the
following useful consequence.

Theorem F. Any continuous m.apping of a compact ~tric s-pa.ce into a

metric space is 'Uniformly continuous.
PROOF+
Let f be a continuous mapping of a compact metric space X into
a metric space Y~ and let d1 and d'l be the metrics on X and Y. .Let
E > 0 be given~
For each point x in x~ consider its image f(x) and the
open sphere S~ 12 {f(x)) cent-ered on this image with radius r=/2+ Since j
is continuous, the inverse image of each of these open spheres is an open
subset of X~ and the cla.ss of all such inverse images is an open cover of X.
Since X is compact~ Theorem C guarantees that this open cover has a
Lebesgue number ~- If x a.nd x 1 are any t'~vo points in X for which
d1(XJXJ) < 3, then the set {x,x'} is a set with diameter Jess than a, both
points belong to the inverse image of some one of the above open spheres~
both f(x) and f(x') belong to one of these open spheres, and therefore
dtt(f(x), f(x')) < f:, which sho\vs that f is indeed uniformly continuous.

We continue our study of compact metric spaces in the next section.

Prob terns

1.
2.
3.

4.

Let A be a subspace of a metric space X, and shol\~ that A is totally

bounded~ A is totally bounded+
Show that a subspace of fltt is bounded (::::)it is totally bounded.
Prove the Ilol zano~ \Veierstrass theorem for Rfl. : if X is a closed and
bounded subset of fl-A~ then every infinite subset of X has ~ limit
point in X.
Show that a compact metric space is separable~

25t ASCOLr S THEOREM _

Our previous characterizations of compactness for a metric space

strongly suggest that this property is related to completeness and total

Compactness

boundedness in some \vay yet to be formu1ated.

theorem ~Thich clarifies this situation.

125

\Ve begin by proving a

Theorem A. A metric space is compact ~it is complete and totally bounded.

PR0011
IJet X be a metric spa~e. The first half of our proof is easy, for
if X is com par:t, then it is totally bounded by l'heure1n 24-D i and it is
complete by Proble1n 12-2 and the fact that every sequence (and therefore
every Cauchy sequence) has a convergent subsequence.
We now assume that X is complete and totally bounded 1 and we
prove that Xis compact by sho,ving that every sequence has a convergent
subsequence. Since Xis complcte:1 it suffices to shoy.. that every sequence
has a Cauchy subsequence. C-onsider an arbitrary sequence

The reason for this notation will soon be clear. Since X is totally
bounded, there exists ~finite class of open spheres 1 each with radius Yi~
\Vhosc union equals X; and from this we see that 81 has a suhf.5~quence
S2 = {X21t X2-i 1 Xi3~ . ~ . } a.11 of whose points lie in some one open sphere
of radius }1. Another appliration of the total boundedness of X sho\vs
similarly that S: ha8 a subsequence 83 = {Xa1, x~l2t XJa, ~ 1 all of
\Vhose points lie in some one open sphere of radius ?3. "\\'c continue
forming successive subsPquences in this manner~ and we let
+

S = {xu~ X22~ XaE3~ . . . }

be the '~diagonal"' subsequence of SJ~ By the nature of thls construction~

S is clearly a Cauchy subsequence of 81, and our proof is complete+
This theorem gives total boundedness a prominent part in dPtermining whether a metric space is compact or not. As we kno\v, many
metric spaces occur as closed subspaces of complete metric spaces} and
for these we can make the role of total boundedness even more striking~
Theorem Bt A closed subspace of a com pl ele 1netric space is co1npact <==> it
is totally bounded+
PROOF.
Since a cloRPd subspace of a complete metric space is automatically complete, t.his is a direct consequenr.e of Theorem . .\~
.

What sort of property is total boundedne.ss? We have seen that. it

always implies boundedness, and we know by l")roblem 24-2 that the
converse of this is true for subspaees of the finite-dilnensional Euclidean
space Rn. It is falsei however 1 that boundedness in1pHf"s total boundedness for .subspaces of the infinite-dimensional Euclidean 8pa('e R~. l n
fa.ct~ the closed unit sphere in RJ, defined by X = {x: llxll < 1 l., i8 not
totally bounded, though it is obviously bounded. To seP this, it suffices

126

Topology

to observe that the sequence {zn} in X defined by

= { 1, 0~ 0,

X1

X2

Xs

= {o, 1 ~ 0,.
= {o~ 0, I'

..

..

..

..

0, . . . },
07 . })

..

0,.

..

has no convergent subsequence,. for the distance from any point of the
sequence to any other is 2Ji. This shows that X is not compact, hence
not totally bounded. The following fact, which we cannot prove here
(see Sec. 47), may add to the reader's intuition about the relation between
boundedness and total boundedness: a Banach space is finite-dimensional
~ evecy bounded subspace is totally bounded.
We now turn to the problem of .characterizing compact subspaces
of e(X,R) or e(X,C). By Theorem B 1 we know at once that a closed
subspace of e(XJR)_ or e(X,C) is compact <=> it is totally bounded.
Unfortunately, however, this information is of little value in most applications to analysis. What is needed is a criterion expressed in terms of the
individual functions in the subspacer Furthermore, for most of the
applications it suffices to consider only the case in which X is a compact
metric space. We describe the relei... .ant concept as f ollo-..vs. Let X be a
compact metric space with metric d, and let A be a non-empty set of
continuous real or complex functions defined on X. If f is .a function in
A, then by Theorem 24-F this function is uniformly continuous; that is,
for each E > 0, there exists a > 0 such that d(x,x 1 ) < 8 ==} [f(x) /(x')I < 11.. In general, a depends not only on E but also on the functionf.
A is said to be equicontinuous if for each Ea acan be found which serves at
once for all functioDB fin A, that is, if for each E > 0 there exists a > 0
such that for every f in A d(x,x') < a;;;;} If(x) - f (x') I < E.
Theorem C (Ascoli's Theorem). If Xis a compact -metric space, then a
cloBed subapace of e(XJR) or e(X,C) is compact ~ it is bounded and
equicontinuous~

I.et d be the metric on X, and let F be a closed sub.space of

e(XtR) or e(X,C).
We first assume that Fis compact~ and. we prove that it is bounded
and equicontinuous. Problem 21-6 shows that Fis. bounded. We prove
that F is equicont.inuous as fallows. Let E > 0 he given.. Since F is
compact, and therefore totally bounded, we can find an (E/3)-net
{f1, f 2J . . . . , f n} in F. Each /1; is unifonnly continuous, so for each
k
IJ 2,
n, there exists ak > 0 such that d(x,x') < ai-=> !fk(X) fJ:(x')j < E/3. We now define a to be the smallest of the numbers
a1~ a,, ... , 8n. Hf is any function in F and /i is chosen so that llf PROOF.

=:::;

Compactness

fkll < .;/3, then

d(x,x') < a=} lf(x)

- f(x') ! ~ !f(x) - J~(x) !

+ Ifrc(x')

+ IJ~(x)

- f(x') I

127

~ fk(x~) f

< E/3 + ~/3 + t:/3 =

e.

This shows that Fis cquicontinuous.

We now assume that F is bounded and equicontinuous, and we
demonstrate that it is compact by shov;ring that every sequence in it has a
c..onvcrgcnt subsequence~ Since F is closed, and the refore complete,
it suffices to show that every sequence in it has a Cauchy su bsequcnce.
As \Ve proceed, the reader will see that our proof is similar in structure
to the last part of the proof of Theorem A. By.Problem 24-4, X has a
countable dense subset+ Let the points of this subset be arranged in a
sequence {x..: ~ = { x2, xi, . . . 1 x..:, . . ~}, ""here we start with the subscript 2 for reasons which will become clear below. Now let

s.

= {f u, f u.,

fl'~~

+}

'

be an arbitrary sequence in 1~.,. Our hypothesis that Fis bounded means

that there exists a real number K such that ~lfll < K for every fin FJ or
equivalently, such that lf(x) I < K for every fin F and every x in X.
Consider the sequence of numbers {f b(x2)}, j = 1, 2~ 3, . .
and
observe that since this sequence is bounded 1 it has a convergent subsequence. Let 82 = {f ~n, f22~ f ~:.r., ~ . J be a subsequence of 81 such that
{f2i(x: 2) l converges. Y-JT e next consider the sequence of num bcrs ff :u(xa) J,
and in the same Vt,.-ay we let SJ = {f:n, fa2, f-a,., ... l be a subsequence of
S,. such that {f li(x3)} converges. If "\\,.e continue this process 1 we get an
array of sequences of the form
+

Si ;;::: {f 11, f12, f u, .

S2
83
+

If21t f 22~ f <J.a,

{fa1,fs1,f:J31

.p

.p

..

..

'

l~
lt

..

in whlch each sequence is a subsequence of the one directly above it, and
for each i the sequence Si = {f~1r f1:1, fiJ,
l has the property that
{/";(xi) J is a convergent sequence of numbers. If 've define f ~ f'Jt f ~t
by f i = f u~ f-i. = f22, !J = f u~ . . . , then the sequence S
{f 1, f2, f3~ ~ .. J
is the ''diagonal'~ subsequence of 8 1 It is clear from this construction
that for each point xi in our dense subset of X 1 the sequence 1f~(xi) l is a
convergent sequence of numbers~ It remains only to sho'v that S, as a
+

;:::=

sequence of functions in e(X,R) or e(X,C), is a Cauchy sequence. Let

E= > 0 be given.
Since Fis equicontinuous~ there exists a > 0 such that
d(x,x') < a ; : : } Ifn(x) - f n(x')! < E/3 for all functions f'rl. in B~ We now

128

Topology

form the open sphere S-b(x~-) l\Tith radius a centered on each of the x~'s.
Since the x/8 are dense, these open spheres form an open cover of X~ and
sjnce X is r.ompact, X == U~~ 2 S~(xi) for some io. It. is easy to see that
there exists a positive integer no such that ni~ n > no~ ffm(Xi) - J'l(xi) I <
~/3 for all the points x2, x(I,
xi~- Our proof is completed by the
remark that if xis an arbit-rary point of X, then an i can be found in the
set {2, 3~ . . ~ , io} such that d(x,x. J < a, and that therefore
4

>

m,n

no~ lfm(X) ~ f n(x) I

< l/m(x) - f m(Xi) l + If"*(xi) - fn(Xi)~

+ I/n(Xi) - Jn(X) I < E/3 + E/3 + ~/3 =

E.

We observe th.at the total boundedness in Theorem B is repJacedJ

in Ascoli's theorem, by the weaker condition of boundedness, and that the
resulting deficiency is made up by the additional condition of equicontinuity+ 1 For several applications of Ascoli 1 s theorem (,vbich is sometimes called Arzela's theorem) to problems in analysist sec Goffman
[13~ pp. 151-156J or Kolmogorov and Fomin r26, vol. 1 ~ secs. 17-20].
Problems

1..
2.

3..
4..

Let A be a subspace of a complete metric space, and show that A is

compact ~ .11 is totally bounded.
Let X be a compact metric space and Fa closed subspace of e(X.,R)
or '2(X,C)~ Show that F is compact if it is equicontinuous and
F~ = {f(x) :J e F ~ is a bounded set of numbers for each point x in X.
Show that RfJC js not locally compact~
By considering the sequence of functions in e[O,IJ defined by
f,.(x) = nx

for 0
x < 1/n, f"'(x)
locally compact.

l for 1/n

< x =5

1, show that e{O~l] is not

The following terminology is often used with AscoliJs theorem. Let F be a.ny
non~mpty set of real or complex functions defined on a.n arbitrary non-empty set X.
The statement that a function fin Fis bounded mea.ns,. of course~ that there exists a
real number K such that ~/(x)I < K for every x in X. The functions in F a.rn often
said to be uniformly bounded (or F is called a unifnrmly bounded set offu.nction.r:;) if
th ere exists a single K which works in this w a.y for a 11 f 1s in ~"t ~ i.e T, if th ere is a K
such that If( x) r < K for every x in x and every I in F. If we ""Ter e tu USO this
expression 1 Aseoli's theorem wouJd take the following form: if Xis a compact metric
e:pace" th en a closed subspace of e( X,. R) or ~( X, C) is com pa ct (::::::} it is u niior m]y
bounded and ~quicontinuou@. The uniform boundedness here ]s merely boundedness
.as a sub.set of the metric spa.ee e(X.,R) or e(X;C)t
1

CHAPTER FlV'E

Separation

A topological space may be very sparsely endowed with open sets.

As we kno\v, some spaces have only two, the empty set and the full space .
In a discrete space, on the other hand, every set is open. Most of the
familiar spaces of geometry and analysis fall somewhere in between these
two artificial extremes.. The so-called separation properties enable us to
state with precision that a given topological space has a rich enough
supply of open sets to serve whatever purpose we may have in mind.
The separation properties are of concern to us because the supply
of open sets poMessed by a topological space is intimately linked to its
supply of continuous functions; and since continuous functions are of
central importance. in topology, we naturally wish to guarantee that
enough of them are present to make our discussions fruitful.. If, for
instance, the only open sets in a topological space are the f'IDpty set and
the full space, then the only continuous functions present are the constantsJ and very little of interest can be said about these. In general
termsjl the more open sets there areJ the more continuous functions a spa~e
has. Discrete spaces have continuous functions in the greatest possible
abundance, for all functions are continuous. However, few really
important spaces are discrete, so this goes a bit too far. The separation
properties make it possible for us to be sure that our spaces have enough
contin no us functions with out committing ourselves to the excesses of
disc rcte spaces.
129

130

Topology

26. Tr-SPACES AND HAUSDORFF SPACES

One of the most natural things to require of a topological space is

that each of its points be a closed set+ 1 The separation property "\vhich
relates to this is the follo"Vlrringr AT rspace is a topological spaec in which~
given any pair of distinrt points, each has a neighborhood which does
not contain the otherr 2 It is obvious that any subspace of a Ti-space
is also a Ti-space. Our first theorem shov-rs that Ti-spaces are precisely
those topological spaces in which points are closed~
point is a closed set
PROOF.
If X is a topological space, then an arbitrary~ point x in X is
closed <;::::;::> its complement is open ~ each point y different from x has a
neighborhood which does not contain x ~Xis a T1~space.

Theorem A.

A to'{Jological space is a 1'1-space

~each

Our next separation property is slight]y stronger. AH ausdorff space

is a topologiGal space in -.,vhich each pair of distinct points can he separated
by open sets, in the sense that they have disjoint neighborhoods. Every
Hausdorff spa.re is clearly a Ti-space, and every subspace of a Ilausdorff
space is also a Ila usdorff space.
Theorem B.. The product of any non-empty class of llausdorff spaces is a
Hausdorff space.

I . . et X = I'1Xi be the product of a non-empty class of Hausdorff

spaces X r. If x = { .ed and y = { Y1} are two distinct points in X 1 then
we must have xi, ~ Y~tJ for at least one index io. Since Xi~ is a Hausdorff
spa~e, X.;~ and Yi. can be separated by open sets in xi,
These t'vo
disjoint open subsets of X ~ give rise to two disjoint sets in the defining
open subb.ase for X each of '"~hich contains one of the points x and y.
PROOF.

Most of the important facts about Ilausdorff spaces depend on the

follo\ving theorem~

Theorem C. In a llausdorff space~ any point. and disjoint compact sub~

space can be separated by open sets, in the sense that they have dis}oint
neighborhoods.
be a Hausdorff space, x a point in x~ and c a compact
subspace of X which does not contain x. We construct. a disjoint pair of
PROOF.

Let

It is customary here to drop the distinction bet.ween a point x in a space and the
.set ~ x } w hi c: h contains only that point~ Th is eonvent ion of ten makes it possible to
avoid cumbersome mo<les of expression, and we sho.11 use it freely.
1
The Ti-space nomencla.ture 1 for i ~ 0, lt
J 5, was introduced by A1exandroff and Hopf in thPir famous trealise r21. The T refers to the German word
TTen n ung saxio m~ which mcs..ns ' fjepa.ra. tion a.xio m.'' The term T rspace is the only
one of these which is at.ill in genera.I use.
1

Separation

131

open sets G and H such that x e. G and C C H. Let y be a point in C.

Since Xis a Hausdorff space~ x and y have disjoint neighborhoods G~ and
H~+ If we allow y to vary over C, we obtain a. class of H 1/s whose union
contains C; and since C is compact, some finite subclass, which we denote
by l ll 1, H 21
H n l J is such that
~ u Z...1 Hi.. If G l t G41..,
~ G
are the neighborhoods of x which correspond to the H/s, we put
+

'

fi

G :;::: flf_1 G'

and H = VL1 H" and observe that these two sets have the required
properties.

In Theorem 21-A ~re proved that every closed subspace of a compact

space is compact, and in Problem 21-4 we saw that a compact subspace of
a compact space need not be closed+ We now use the preceding theorem
to show that compact subspaces of Hausdorff spaces arc always closed.

Ever-y comp act subspace of a H au:i/.orff space is closed.

PROOF.
Let C be a compact subspace of a Hausdorff space X. We prove
that C is closed hy shovling that its complement C' is open. C' is open
if it is emptyj Bo v.re may assume that it is non-empty. Let x be any
point inc~ By Theorem C, x has a neighborhood G such that x g G ~ C'~
This shows that C' is a union of open sets and is therefore open itself4

Theorem D.

One of the most useful consequences of this result is

Theorem E. A one-to-one continuous m.apping of a compact S'/)Me onto a

II ausdorff Rpace is a homeomorphism.
PUOOF
Let. f: X ~ Y be a one-to--0ne continuous mapping of a compact
space X onto a Hausdorff space Y. We must show that /(G) is open in
y "\Vhcnever G is open in x"! and for this it suffices to show that f(F) is
r.Josed in Y \vhenever Fis closed in X. If Fis empty, f(/l) is also empty
and therefore c]osed, so \Ve may a8sume that F is non-empty. By
1'heorcm 2 l-A} Fis a r.ompaet subspace of X; hy Theon.lJn 21-R1 f(F) is a
compact subspace of l""; and VLe complete the proof hy usiug the preceding
theorem to conclude that.. f(F) is a closed subspace of Y.
4

Compact Ilausdorff spaces are among the most important of all

topological spaces~ and in the f ollo\ving scction8 and chapters we shall
become thoroughly acquainted \\'i th their major properties.

Problems

1.

Show that the topological space defined in Example

T 1-spaee.

16~5

is not a

Topor ogy

132

2.

Show that the topological space defined in Example 16-4 is a T 1-space

but not a Hausdorff space.
3. Show that any finite T 1space is discrete~
A. If Xis a Ti-space with at least two points, show that an open base
which contains X as a member remains an open base if X is dropped+
5. Let X be a topological space, Y a Hausdorff space, and A a subspace
of X. Show that a continuous mapping of A into Y he..is at most one
continuous extension to a mapping of 1 into Y. Problem 17-2 is a
special case of thls statement.
6. If f is a continuous mapping of a topological space X into a Hausdorff
space Y 1 prove. that the graph off is a closed subset of the product
7.

8..

XX Y.
I.et X be any non-empty setJ and prove that in the lattice of all
topologies on X each chain has at most one compact Hausdorff
topology as a member. (It is interesting to speculate about v.-~ he th er
a compact Hausdorff topology can be defined on an arbitrary non-+
empty set.)
Let X be an ar bi tra.ry topological space and {x" } a sequence of
points in X. This sequence is said to be (!onvergent if there exists a
point x in X such that for each neighborhood G of x a positive integer
n-o can be found ''dth the property that Xn is in G for all n > n~.
The point x is called a limit of the sequence~ and we say that Xn
converges to x (and symbolize this by Xn ~ x)
(a) Show that in Example 16-3 any sequence converges to every
point of the space. This is the reason why the above point x
is called a limit instead of the limit.
(b) Ii Xis a Hausdorff space,. show .that every convergent sequence
in X has a unique limit.
(c) Sho,v that if f :X ~ Y is a continuous mapping of one topological space into another, thr.n Xn---+ x in X ~ f(xn) -~ f(x) in Y.
Prove that the converse of this is true if each point. in X has u
countable open base. 1
+

27.. COMPLETELY REGULAR SPACES AND NORMAL SPACES

Let X be an arbitrary topological space, and consider the set e(X,R)

of all bounded continuous real functions defined on X.. If for each pair
of distinct points x and y in X there exists a function fin e(X,R) such
that f(x) F f(y),. we say that e(X,R) aeparate8 pointB. It is easy to see
1

The f ac t.s brought out in t.h.iB problem a.re the main reasons w by the concept
of a couvergcnt sequenee is not very important in the general theory of topologica.1
sps.ecs~

Separation

133

that ii e(X,R) does separate points, then X is necessarily a Hausdorff

space; for assuming that/(x) < f(y) and that r is a real number such that
f (x) < r < f (y), then the sets {z :f(z) < r I and {z :f(z) > r i are disjoint
neighborhoods of x and y..
"
It is convenient to strengthen this separation property slightly by
allowing one of the pointr; to be an arbitrary closed subspace of X. A
completely regular space is a Trspacc X with the property that if xis any
point in X and F any closed subspace of X which does not contain x, then
there exists a function fin e(X~R), all of whose values lie in the closed
unit int~rvnl lOi l]J such that f(x) = 0 and /(F) = 1. It is worth noticing
that since constants arc continuous, 've could just as well have required
here that f be 1 at x and 0 on F, for the function g = 1 - f has these
properties. We may think of completely regular spaces as T 1-spa.ces in
\Vhich continuous functions separate points and disjoint closed subspaces.
Since points are closed in a completely regular space, it is permissible to
take the closed subspace F to be a point,, and it is clear by the above
paragraph that every completely regular space is a Hausdorff space.
It is also easy to see that every subspace of a completely re-gular space is
completely regu]arL In Sec+ 30 we gi\'~e an explicit characterization of all
completely regular spaces in terms of product spar.es.
Our next (and last) separation property is simi1 ar to that which
defines a Hausdorff space, except that it applies to disjoint closed sets
instead of merely to distinct points. A normal s-pa.ce is a Ti-space in
which each pair of disjoint closed sets can be separated by open sets, in
the sense that they have disjoint neighborhoods.. We shall see in the
next section (as a consequence of Urysohn~s lemma) that every normal
space is completely regular.
Figure 25 is intended to illustrate and clarify the relations among our
various separation properties: a topological space which possesses any one
property, in the order of their definition, also possesses all properties
which precede it; in other words, they have been defined in order of
increasing strength. The figure also indicates that metric spaces and
compact Hausdorff spaces are normat We established the first of these
facts in Problem 11-lb, and we now prove the second.

Every compact Hau8darff S'pace is normat

PROOF.
Let X be a compact Hausdorff space, and A and B disjoint
closed subsets of X4 We must produce a disjoint pair of open sets G and
H such that A ~ G and B ~ H ~ If either closed set is empty, we can
take the empty set as a neighborhood of it and the full space as a neighborhood of the other. We may therefore assume that both A and Bare
non-empty~ Since X is compact, A and B are disjoint compact subspaces of X ~ Let x be a point of A. By Theorem 26--C and our hypothcTheorem A.

13.4

Topology

sis that X is Hausdorff, x and B have disjoint neighborhoods G~ and

H B If we allow x to vary over A~ we obtain a class of G~'s whose union
contains A; and since A is compactJ some finite subclass, which we denote
by 1G lt G:, .
~ GR L is such that A ~ U ~ 1 G1.. If Hi~ H 2 ~ .. . , H ~
are the neighborhoods of B which correspond to the G/s, it is clear that
G = V7_ 1 Gi and H = n7_ 1 H, are disjoint neighborhoods of A and B .
4

Topological spaces

Hausdorff spaces

Completely regutar spaces

Normal spaces
Metric

Compact

spaces

Hausdorff
spaces

F ].(:!;. ')5
.....

The se pa.ration properties.

In Sec. 29 we investigate the manner in which normal spaces~

compact Hausdorff spaces 1 and metric spaces are related to one another.
Problems

1.
2.

3.

4.

Show that a closed subspace of a normal space is normal~

Let X be a T 1-space, and show that Xis normal ::::}each neighborhood
of a closed set F contains the closure of some neighborhood of F.
Assume, as Fig. 25 suggests~ that a compact Hausdorff space X is
completely regular and that therefore e(X,R) separates points.
Use this to prove that the V{eak topology generated by e(X~R)
equals the given topology+ Show further that if Sis any subset of
e(X 1 R) which also separates points, then the weak topology generated
by S also equals the given topology.
Let X be a completely regular space, and shov.T from the definition
that the weak topology generated by e(X~R) equals the given
topology.

Separation

135

28. URYSOHNt S LEMMA AND THE TIETZE EXTENSION THEOREM

As we suggested in the introduction to this chapter~ one of the ma.in
purposes served by assuming that a topo1 ogical space is rich in open sets is
to guarantee that it is also richin continuous functions. The following
is the fundamental theorem in this direction.

L.ct X be a normal space, and let A and

B be disjoi.nt closed subspaces of X. Then there exists a continuouB real
June.lion f defined on X, all of whose values lie in the closr,.d unit interval
[O,l], such lhat f(A) = 0 and f(B) = 1.
PROOF.
B' is a neighborhood of the closed set A~ so by the normality of
X and Problem 27-2t A bas a neighborhood UH such that
Theorem A (Urysohnts Lemma)..

A~ U~

CIH~

c B'.

U wand B' are neighborhoods of the closed sets A and

way there exist open sets Utt and C_I:n such t.ha t
A ~ Ui,t C llw ~ [Iu C

U}'.! C

[! }i}

so in the same

c B'.

{/H C llt.i

If we continue this process, for each dyadic rational number of the form
t = m/2n (where n = 1; 2, 3, . . . and m :;:::: l~ 2,
2n - 1) we have
an open set of the form llt,. such that
+

t1

< t2 ===>-A

--

Ut1 C Lrt c lit c rJi'!

3

B .

We now define our functionf by f(x) = 0 if xis in every f.It and

f(x)

sup {t:x i [Id

otherwise. It is clear that the values off lie in [O, I], and that J(A) = 0
and f(B) ~ L. All that remains to be proved i~ that f is continuous+
All intervals of the form {O,a) and (a,lL where 0 < a < 1, constitute an
open sub base for [OJ l 14 It therefore suffices to show that 1~ 1 ([O,a)) and
) 1 ((a~l]) are open. It is easy to see thatf(x) < a t j xis in some U, for
t < a; and from this it follows thatj- 1([0,a)) = {x :/(x) <a} = V,<a Ut,
which is an open set. Similarly, f(x) > a-:;::} x is outside of u; for some
t > a; and therefore 1- 1((a;11]) = tx:f(x) > al ~ v~>a Ul~ which is an
open set4
It is clear from this theorem that every normal space is completely
regular: all that is necessary is to take the closed subspace A to be a single
point and to observe that the function f is exactly what is required in the
definition of complete regularity.
The following slightly more flexible form of Urysohn's lemma will
be useful in applications.

136

Topology

Theorem B.. Let X be a normal space, and let A and B be disjmnt close.d
subspac.es of X~ If [a~b] is any closed interval on tke real line, then there
exists a continuous real function f defi.ned on XJ all of whose values lie in
[a~b1i such that f(A) = a and f(B) = b.
PROOF.
If a = b, we have only to define f by f(x) == a for every x> so we
may assume that a < b. If g is a function v.rith the properties stated in
U-rysohn's lemma, then f = (b - a)g
a has the properties required by
our theorem.

If there is given a continuous function defined on a subspace of a

topological space, Urysohn's lemma has an important bearing on the
interesting question of whether this function can be extended continuously
to the full space. The following is the classic theorem along these lines.
Let X be a normal space, F
a .closed subspace, and fa eontinuoua real function define.d on F whose values
lie in the closed int.erval [a,b]+ Then f has a continuous extension f' defined
on all of X whose values also lie in [aJb].
PROOF. If a = b:1 the conclusion of our theorem is obvious,. so we may
assume that a < b. We may clearly assume that [a,,b1 is the smallest
closed interval which contains the range off. :Furthermore, the device
UBed in the proof of 1"heorem B enables us to assume that a = -1 and
b = 1. We begin by defining f o to be f. The domain offo is our closed
subspace F, and we define two subsets Ao and Bo of F by

Theorem C (the Tietze Extension Theorem).

Ao== {x:fo(x) ~

-Yal

and Bo = {x ~fD(x) > 73' 1- Ao and Bo are disjoint, non-empty, and

closed in F; and since Fis closed, they are closed in X. Ao and Bo are
thus a disjoint pair of closed subspaces of X, and by Theorem B there
exists a continuous function go~ X ~ [-31~731 such that go(Ao) = - 73
and go(Bo) == %.. We next define/1 on F by f1 =fa - Yfh and we observe
that ~f1(x) r ~ %- If A1 = {x :/1(x) :::; ( ~ %) (~,~)} and

Bi == fx :f1(x) > CM)(%)}~

then in the same way as above there exists a continuous function
g1:X ~ [(- J.3)(%),(:!1j)(%)J such that g1( ..41) = (-~)(%)and
Y1(B1)

=== (M)(;3)~

We next define f2 on F by /1 = fi - g1 ;;;::: /o ~ (go + g1), and we observe

that Jf2(x)I < (3'~) 2 By continuing in this manner,. we get a sequence
{/o, /1,, /2,. ... } defined on F such that lfn(x)j < (_%)n, and a sequence
I go, U~ g';l~
~} defined on X such that lg~(x)I < C!i) (%)~1 with the
property that on F we have /. = fo - (go
a1
~
aA-1).. We
+

+ + +

Separation

137

+.

by Sn = go + U1
~ + OiL-1 1 and we regard the sn's as
the partial sums of an infinite seri~s of functions in e(X,R)~ We know
that e(XJR) is completeJ so by lgn(x)i ::;; {%) (%)n and the fact that
x===- 0 (7~) (%) = 1, we see thats~ converges uniformly on X to a bounded
continuous real function f' such that lf'(x)I < 1. We conclude our
proof by noting that since Iftt(x)I < (%)'\ s" converges uniformly on
F to fo = !1' and that therefore/' equals/ on F and is a continuous extension of f to the full space X Vlbich has the defiired property.
now define

Bn

11

It is of some interest to observe that this theorem becomes false if

we omit the assumption that the subspace F is closed. This is easily
seen by means of the following example. Let X be the closed unit
interval [O, 1], F the subspace (0, 11, and f the function defined on F by
f(x) = sin (l/x). X is clearly normal, F is not closed as a subspace
of X~ and f cannot be extended continuously to X in any manner
whatsoever .

Problems

1.

2..
3.

In the text we used lJrysohn's lemma as a tool to prove Tietze's

theorem. Reverse this process~ and deduce Urysohn 1s lemma from
Tietze's theorem.
State and prove a generalization of rriet.ze"s theorem "\\Thich relates to
functions whose values lie in Rn.
Justify the assertion in the last paragraph of the text that the function
defined there cannot be extended continuously to X.

29. THE URYSOHN IMBEDDiNG THEOREM

3~

we generalized metric spaces to topological spaces.

We now reverse this procedure and seek out simple conditions whlch
guarantee that a topological space is essentially a metric space 1 that is,
~Thich imply that it is metrizable.. Problem 16-12 shows that we must
look for properties of a topological space X which enable us to construct
a homeomorphism f of X onto a subspace of some metric space; for the
metric on this subspace can then be carried back by f to X~ and we can
infer that Xis metrizable. The simplest property of this kind is discreteness; for if Xis a discrete space, then its underlying set of points~ equipped
with the metric defined in Example 9-1, is a homeomorphic image of
X under the identity mapping. We can lift our discussion to a more
meaningful level by observing that since every metric space is normal,
normality must be among the properties assumed of X, or it must be
implied by them .
In Chap..

138

Topology

As motivation for our main theorem, we note that since the metric
space R 110 is second countable, every subspace of it is also second countable4
It turns out that second countability, in addition to normality, suffices
to guarantee that a topological space is homeomorphic to a subspace
of R~. In effect~ we imbcd such a space homeomorphically in R~.
If Xis a second countable normal space, then there exists a hom.eomorphismf of X onto a subspace
of R~ 1 and X is therefore metrizable
PROOF. We may assume that X has infinitely many points~ for other-wise it would be finite and discrete 1 and clearly homcomorphlc to
any subspace of Rec with the same number of points. Since X is
second countable, it has a countably infinite open base {G1, G2, GJ, . . . }
each of whose sets is different from the empty set and the full space.
If G1 and x a Gi are given~ then by normality there exists a G;: such that
x e Gi C G; c Gj. '"fhe set of all ordered pairs (G1,G;) such that G1 C Gj
is countably infinite, and we can arrange t.hem in a sequence P1j
P 2 ~ . . . . , P n .. . By Urysohnts lcmrna~ for each ordered pair
P ~ = ( Gi~ G1) there exists a continuous real function f n: X ~ [O, 1] such
that f ,.(G;.) = 0 and fn(G/) ==== 1. :For each x in X we define /(x) to be
the sequence given by f(x) = {f 1(x) 1 f2(x)/2,
~ fn(x)/n, ~ .. }~ If
we recall that the infinite series ~== 1 l/n'2 convergesJ it is ca.sy to see that
f is a one-to-one mapping of X into Ri ~ It remains to be proved that
f and J- 1 arc continuous.
To prove that f is continuous, it suffices to show that given x~ in
X and E > 0 1 there exists a neighborhood H of Xo such that y E H ==>
l]f(y) - f(xo) ll < E. Since an infinite series of functions converges
uniformly if its terms are bounded by the terms of a convergent
infinite series of constants, it is easy to see that there exists a positive
integer no such that for every y in X we have
Theorem A (the Urysohn lmbedding Theorem).
+

nJ(y) - f(xo)H 2

=:;

I=-=1 ~[f~{y) -

/n(xo)l/nl'"
< :z::..1 I[/n(Y)

f ~(xo)J/nl !

+ E2/2.

By the continuity of the /n.'s, for each n = 1, 2, ..

nc. there exists a
neighborhood Hn of Xo such that y E Hn ~ llfs(y) - fn(xo)]/nl 2 < E'2/2no.
If we define H by H ;::;: r'i:-_ 1 H ni it is clear that H is a neighborhood of xo
such that y EH=> 11f(y) - f(xo) 1! 1 < t 2 => llf(y) - f(xo) ll < E.
We conclude our proof by showing that 1- 1 is continuous aa a mapping of /(X) onto X. It suffices to show that given xo in X and a basic
neighborhood Gi of Xe~ there exists E > 0 such that 11/(y) - f(xo) ~I <
E t j ye: Gi4
Gj is the second member of some ordered pair Pn~ = (Gi,Gj)
such that xo .e Gi C fl C G;. If we choose E < 1/2no, then we see
that ]lf(y) - f(xo)ll < E ==> ~::E 1 I[f(y) - f tt(xo)]/nl 2 < (1/2no) 2 ~ Ifn:t(Y)
4

139

Separation

- f nt{xo)I < 72. Since Xo is in a,~ fn11(xc.)

-==::

0, and therefore l!n.a(Y)[

<

~~

Since /n,(G/) -;: : : 1, we see that y is in G1.

Thls theorem puts us in a position to answer several natural questions
whlch arise in connection with the inner portions of Fig~ 25~ We ask
the reader to deal with these matters in the iollowing problems.
Probtems

1..

2.

3..

We know that every metric space is normalJ and also that a normal
space) if second countable~ is metrizable.. Give an example of a
normal space which is not metrizable (hint.: see Problem 22-4)~ Thls
shows that metrizable spaces cannot be characterized among topologi ..
cal spaces by the property of normality~
Among normal spaces~ second countability implies metrizability.
Give an example of a metric space which is not second countable.
This shows that metrizable spaces cannot be characterized among
normal spaces by the property of second countability.
Show that a compact Hausdorff space is metrizable <=> it is second
countable~ 1

30. THE STONE-CECH COMPACTIFICATION

In the preceding section we sho\ved that if a normal space is second
countable 1 then it can be imbedded as a subspace in the familiar metric
space R~. We now develop a similar imbedding theorem for completely regular spaces.
In order to motivate this theorem properly~ we remark that if X
is a topological space which occurs as a subspace oi a compact Hausdorff
space Y~ then since Y is completely regularJ Xis also completely regular~
and is a dense subspace of the compact Hausdorff space X. \Ve see in
this v;ray that many completely regular spaces are dense subspaces of
compact Hausdorff spaces. Our purpose in this section is to show that
any completely regular space X can be imbedded as a dense subspace in
a special co1npact Hausdorff space denoted by {3(X), and that f3(X) bas
the remarkable property that every bounded continuous real function
defined on X has a unique extension to a bounded continuous real
function defined on fJ(X).
,,;

The results of this se.r:tion have chara.ct.erizcd metrizable spaces among second
oounlaLle ~pa.ces (by normality) and among com pa.ct Hausdorff spaces (by second
countahi1ity ), but not among topological spaces in genera..l. This more difficult
problem 'vas solved by Smirnov [38]. For an exposition of his solution~ see Kelley
[25, pp. 126~ 130J.

140

Topology

How truly remarkable this extension property is can be Been by

considering the example given at the end of Sec. 28 .. Here the completely
regular space X is the interval (0 1]. This space is clearly a dense
subspace of the compact Hausdorff space [O, 1]. The function f defined
on (0,1] by f(x) = sin (1/x) is a bounded continuous real function
defined on X ~ but it cannot be extended continuously to [0, I l. rfhe
space [O~ lL though it is a compact I-Iausdorff space which contains (0, 1J as
a dense subspace, is evidently not the space {3(X). The latter is much
too complicated for any simple description of jt to be possible.
Before 've start our discussion 1 we rec a.11 t'vo items from the p rcvious
1

sections;
(1) if X is a completely regular space, then the 'veak topology
generated by e(X)R) equals the given topology;
(2) the relative topology on a subspace oi a product space equals
the weak topology generated by the restrictions of thr. projections
to that subspace .
These facts (they are Problems 27-4 and 22-2) are the basic principleH on
v.Thlch the following analysis rests.
We begin 'vith an arbitrary topological space X and the set e(X,R)
of all bounded continuous real functions defined on X. Let the fun('tions in e(X 1R) be indexed by a set of ind ices i, so that e (X ,ll) == {/i l
For each index it lr.t [; be the smallest closed interval '"'hich contain8 the
range of the function f,, ~ Each Ii: is a com pact Hausdorff spacei: so their
product P = Pili is also a compact Hausdorff space, and every subspace
of P is completely regular. We next define a mapping f of X into this
product space by means of f(x) = {fi(~)}, that ist in such a v,ray that
f(x) is that point in the product. space P whose ith coordinate is the real
number fi(x)r By ProblPm 22-3 and the fact that p,f = fi for each projection Pi~ it is clear t.ha t f is a continuous mapping of X in to P.
We now assume that e(X,R) separates the points of X~ This is a
weaker assumption than complete regularity and is exactly the requirement that f be a one-to-one mapping. At this stage, 1ve use f to replace
f (X) as a set by X; that is, we imbed X in P as a set. X is thus a sub~
set of P l\:rhich has two topologies: its own, and the relative topology
\Vhich it has as a subspace of Pr We observe two features of this situation. First~ since f is continuous~ the given topology on X is stronger
than its relative topologyr Sccondj e(X,R) is precisely the set of all restrictions to X of the projections pi of Ponto it~ coordinate spaces Ii. It is noYl
clear that if X is completely regular, so that by state1nent (1) its given
topology equals the weak topology generated by e(X,R), then by stutement (2) its given topology equals its relative topology, and X ean be
regarded as a subspace of P .
In accordance with these ideas, we now assume that X is con1pletely

Separation

l 41

regular~

and we fully identify it, both as a set and as a topological space,

with the subspace f(X) of P.. It is easy to see that the closure X of X
in P is a compact Hausdorff space in which X is imbcdded as a dense
subspace. Furthermore, each f,, in e(X,R)-that is, each projection p"
restricted to X-has an extension to a bounded continuous real function defined on X ; this extension is Pi restricted only to X 1 and it is
unique by Problem 26-5. The space Xis commonly denoted by {:J(X)~
We summarize these results in the follo-wing theorem.
Theorem A.. Let X be an arbitrary completely regular space~ Then there
exists a com pact Hausdorff space fJ (X) tDith the following properties 4. (1)
X is a di-nse subspace of fJ ( X); (2) ei-ery bounded continuous real f uncti(Yll,
defi-rn;d on X has a uni.que exf.ensi.an to a bounded continuous real function
defiru;d on fJ(X).

We shall prove in Chap. 14 that the space fj(X) is essentially unique~

in the sense that any other compact Hausdorff space with properties (1)
and (2) is homeomorphic. to fj(X). It is called the Stone-Cech compactijication of the given comp1etely regular space. 1
Even before our VtTork in thls section, it was clear that every sub
space of a product of closed intervals is completely regular. It is " ..orthy
of special emphasis that the above discussion shows, conversely 1 that
every completely regular .space is homeomorphic to a subspace of such a
product.
Prob1ems

1.

24>

3.

If X is completely regular, show that every bounded continuous complex function defined on X has a unique extension to a bounded
continuous complex function defined on t3(X).
Every closed subspace of a product of closed intenTals is a compact
Hausdorff space. Show, conversely, that every compact Hausdorff
space is homeomorphic to a closed subspaee of such a product.
Prove the following generalization of the rietze extension theorem.
If X is a normal space, F a closed subspace of X, and fa continuous
mapping of F into a completely regular space Y, then f can be
extended continuously to a mapping f' of X into a compact Hausdorff
space Z whlch contains Y as a subspace.
1

For Stoneta own version of these ideas. 1 aec his paper (39J.

CHAPTER SIX

eo11Jtcetcd11cs1

From the intuitive point of view, a cminected space is a topological

space \vhich consists of a single piece. 1,his propr.rty is perhaps the
simplest which a topological space may have, and yPt it is one of the
most important for the applications of topology to analy8iS and geometry.
On the real line~ for instance, intervals are connected subspacesJ and
we shall see that they are the only connected subspaces. Continuous
real functions are of ten d cfincd on intervals, and functions of this kind
have many pleasant properties. }'or example, such a function assumes as
a value every number between any two of its values (the Weierstrass
intermediate talue theorem); furthermore, itB graph is a connected subspace
of the Euclidean plane+ Connect-edncss is also a basic notion in complex
analysis, for the regions on which analytic functions are studied are
generally taken to be connected open subspaces of the complex plane.
In the portion of topology- l\Thich deals "\\ri th continuous curves and
their properties, connectedneBB is of great significancet i or whatever e1se a
continuous curve may be, it is certainly a connected topological space.
We describe some of the central ideas of this field in Appendix 2.
Spaces v.rhich a.re not connected are also interesting. One of the
outstanding characteristics of the C-antor set is the extreme degree in
'"hich it f ai]s to be connected~ Much the same is true of the subspace of
the real line which consists of all rational numbers. These spaces are so
badly disconnected that they are almost granular in texture.
Our purpose in this chapter is to convert these rather vague notions
into precise mathematical ideas, and also to establish the fundamental
facts in the theory of connectedness which rests upon them.
1.42

Connectedness

1"3

31 .. CONNECTED SPACES
A connected space is a topological space X whlch cannot be represcn tcd as the union of two disjoint non-empty open sets. If X ~ A VB,
where A and B are disjoint and openJ then A and B are also closed, so
that X is the union of two disjoint closed sets 1 and conversely. We .see
by this that Xis connected ~it cannot be represented as the union of two
disjoint non-empty closed sets. It is also clear that the connectedness
of X amounts to the condition that 0 and X are its only subsets which
are both open and closed~ A connected subspace of X is a subspace Y
which is connected as a topological space in its own right. By the definition of the relative topology- on Y 1 this is equivalent to the condition that
Y is not contained in the union of two open subsets of X whose intersections 'With Y a.re disjoint and non-empty.
Our space Xis said to be disconnected if it is not connected~ that is~
if it can be represented in the form x = A U B, where A and B are
disjointi non-empty, and open; and if Xis disconncctcd 1 a representation
of it in this for1n (there may be many) is called a disconnection of X .
We begin by proving a theorem which supports a considerable part
of the theory of connectedness.
Theorem A. A 8Ub8pace of the real linR R is connected <==>it is an interval.
In particulari. ll is connected.
PROOF..
Let X be a subspace of R. We first prove that if Xis connected 1
then it is an interval. We do this by assuming that Xis not an interval
and by using this assumption to show that X is not connected. To say
that Xis not an interval is to say that there exist real numbers x, y, z such
that x < y < zJ x and z arc in X, and y is not in X~ It is easy to see from
this that X ~ [X r\ (- oo ,y)] U [X
(y,
~ )J is a disconnection of
X 1 so X is disconnected.
We complete the proof by shol\Ting that if Xis an intenral, then it is
necessarily connected. Our strategy here is to assume that X is disconnectr.d and to deduce a contradiction from this assumption~ Let
X = AV B be a disconnection of X. Since A and Bare non-emptyt we
can choose a point x in A and a point z in B. A and B are disjoint, so
x r:E 2 1 and by altering our notation if necessary, we may assume that
x < z. Since X is an interval, [x,z] C X, and each point in [x,z] is in
either A or B.. We now define y by y == sup ([x,z]
A). It is clear
that x ~ y :::; z, so y is in X. Since A is closed in X, the definition of
11 shows that y is in A. From this we conclude that y < z.. Again by the
definition of y 1 y
E is in B for every e: > 0 such that y
~ :$; z, and
since Bis closed in X, y is in B. We have proved that y is in both A and
B, which contradicts our assumption that these sets are disjoint.

144

Topology

Our next theorem asserts that the property of connectedness is

preserved by continuous mappings~
Theorem B~ Any continuous image of a connected SJ)ace i8 connected.
ruooF. Let f:X ~ Y be a continuous mapping of a connected space
X into an arbitrary topological space Y+ We must !-3how that /(X) is
connected as a subspace of Y. Assume that f (X) is diseo nnected. As l\e
have seen1 this means that there exist two open suhsets G and H of Y
whose union contains f(X) and whose intersections \vi th f(X) arc disjoint
and non-empty~ This implies, ho,vever, that X = 1- 1(G) U 1~ 1 (H) is a
disconnection of x~ which contradicts the connectedness of

x.

As a direct con.sequence of the two theorems just proved~ v.,..e have the
f ollov,.ring gen~ra1iza tion of the W cierstrass in term~d ia te value theorem.
Theorem C.. 77u range of a continuous real function defined on a connected
space is an interval.

It is a trivial observation that any tv"ro discrete spaces \vith the same
number of points arc e~s~ntially id en tic al; for any one-to-one mapping of
one onto the other (there is at least one) is a homeomorphism, and \ve may
think of them as differing only in the symbols used to designate th(-~ir
points. It is in thls sense that t.here is only one discrete space \vith any
given number of points. The discrete two-pO'int space, '\\,..ruch is obviously
disconnected. is a usPful tool in the theory of connectedness~ V{e denote
its points by the symbols 0 and I, and we think of them as rca] numbers_
A topological S'pace X is disconnected ~ there exislB a continuou8 mapping of X onto the discrete two~point space {0,1 l
PROOF. If Xis disconner.ted and X = A VB is a disconnPetiont then v.~e
define a continuous mapping .f of X onto {O, 1 l by the requirement that
f(x) = 0 if x is in A and f(x) -== 1 ii xis in B. This is a valid definition
by the fact that A and Bare disjoint and their union is X. Since A and
B are non-Pmpty and open, f is clearly onto and continuous.
On the othPr hand, if there exists such a mapping, then X is disconnected; for if X were connected~ Theorem B ",.ould impJy that {O~ 1} is
connected, and this wou]d be a contradiction.
Theorem D.

Thjs result is a useful tool for the proof of our next theorem.
Theorem E.

The, product of any non-.empty cl.ass of connected sp aceB is

conn.ected4
I-Rt ~ X1 l be a non-empty class of connected spaces~ and form
their product X == P~i- We assume that X is disconnected,. and we
deduce a contradiction from this assumption.. By Theorem D, thPre
exists a continuous mapping f of X onto the discrete two-point space
PROOF~

Connectedness

1'5

{0,1]. Let a = la,} be a fixed point in X, and consider a particular

index i1.. We define a mappingf-.: of Xi. into X by means of /~1 (xt:1 ) = {y,} 1
where y, = Ui for i i1 and y.:l ~= Xii This is clearly a continuous
mapping, so ff" 1 is a continuous mapping of Xi 1 into {OJI'. Since Xii is
connected~ we see by Theorem D that ff.;,i is constant and that
1

for every point x1 1 in Xi

This shows that f(x) = f(a) for all x's in X
which equal a in aJl coordinate spaces except X1 1 By repeating this
process ",..ith another index i'2~ etc., we see that f(x) = f (a) for all x's in X
which equal a in all but a finite number of coordinate spaces. The set of
all x~s of this kind is a dense subset of X, so by Problem 26-5, f is a constant n1apping. This contradicts the assumption that f maps X onto
{0~1 J, and completes the proof.
1

As an appJication of this result, ""~c show that alJ finite-dimensionaJ

Euclidean and unitary spaces are connected.
Theorem F. The spaces Rn and en. are ronnected.
PROOF. We shol\'ed in the proof of Theorem 23-B that Rn, as a topological space~ can be regarded as the product of n replicas of the real
line R. We have seen in Theorem A that R is connected, so R"" is connected h;r Theorem R. We next prove that Cn and R 23 are essentially
the same as topological spaces by exhibiting a homeomorphism f of
C'$ on to R~~. Let z ;::::. (z 1,. Z.27 ~ ' Zn) be an arbitrary element in cf*-,
and let each coordinate z,,.. be written out in the form z1r. ::::; ak
ibt, where
ak and bk arc its real and imagjnaiy parts. We define f by

j(z) = (a1, bi~ a2, b'2, ~ ... ,

anJ

b").

is clearly a one-to-one mapping of cft onto R2n, and if we observe that

!if(z) (] = [lz1L it is ea.sy to see that f is a homeomorphism. The fact that
R2n is connected now shows that Cta is also connected.

The techniques developed in the next section will make it possible

to give an easy proof of a much more general theorem than this, to the
effect that any Banach space is co nncctcd.
Problems

2.

Show that a topological space is connected~ every non-empty proper

subset has a non..-empty boundary.
Show that a topological space X is connected ~ for every two points
in X there is some connected subspace of X whir.h contains botha

146

Topology

Prove that a subspace of a topological space Xis disconnected ~it

can be represented as the union of two non-empty sets each of which
is disjoint from the closure (in X) of the other.
44 Show that the graph of a continuous real function defined on an
interval is a connected subspa.ce of the Euclidean planer
5~ Show that if a connected space has a non-constant continuous
real function de.fined on it, then its set of points is lUlcountably
infinite.
6. If X is a completely regular space, use Theorem D to prove that X is
connected ~ fj(X) is connected~

34

32. THE COMPONENTS OF A SPACE

If a. space is not itself connected, then the ne~t best thing is to be

able to deyompose it into a disjoint class of maximal connected subspaces.
Our presen~ objective is to show that this can always be done.
A maxima.I connected subspace of a topological space)' that is)' a
connected subspace which is not properly contained in any larger connected subspace, is called a component of the space. A connected space
clearly has on1y one component~ namely, the space itself. In a discrete
space, it is easy to see that each point is a component.
The fallowing two theorems vrill be useful in obtaining the desired
decomposition for a general space.

Theorem A~ Let X be a topologict.il S'pace. If {A_.} is a non-empty cl.ass

of connected sub:rpaces of X such that nf, is non-empty 1 then A = U{, is
also a connected subspace of X~
PROOF. Assume that A is disconnected. This means that there exist two
open subsets G and H of X whose union contains A and whose intersections with A are disjoint and non-empty. All the A .:'s arc connected)'
and each lies in G U H)' so each Ai lies entirely in G or entirely in H and is
disjoint from the other. Since f\iAi is non..-empty, either all the A /s lie
in G and are disjoint from H, or all lie in Hand are disjoint from G. We
see by this that A itself is disjoint from either G or HJ and this contradiction .shows that our assumption that A is disconnected is untenable.
Theorem B. Let X be a topological spau and A a connected subspace of X.
If B is a subspace of X such th.at A k B C At then B is cannecud; in

particular,

A is connectoo.

Assume that Bis disconnected, that is, that there exist two open
subsets G and H of X whose union contains B and whose intersections
with Bare di~oint and non-empty. Since A is connected and contained
PROOF.

Connectedness

147

in GU H, A is contained in either G or Hand is disjoint from the other.

Let us say, just to be specific~ that A is disjoint from Ha This implieliJ
that A is also disjoint from H~ and since R c A, Bis disjoint from H.
This contradiction shows that B cannot be disconnected, and proves our
theorem.
We are now in a position to state and prove the main facts about
components~

Theorem C. If X is an arbitrary topological space, then we have the following: (1) each point in Xis contained in exactly one component of X; (2)
each r.onnected 811hspace of X is contained in a component of X; (3) a
connected subspace of X which is both open and close.d is a component of X;
and (4) each component of Xis closed.
PROOF.
To prove (1), let x be a point in X. Consider the class {Cd of
all connected subspaces of X which contain x.. This class is non-empty,
since x itself is connected~ By Theorem A, C = UiCi is a connected
subspace of X which contains x. C is clearly ma...ximal, and therefore a
component of X, because any connected subspace of X which contains
C is one of the C./s and is thus contained irr C. :Finally, C is the only
component of X which contains x. For if
is another, it is clearly
among the C/s and is therefore contained in C, and since C* is maximal a.s
a connected subspace of X~ we must have C* = C.
Part (2) is a direct consequence oi the construction in the above
pa.ragraphj for by this construction, a connected subspace of X is contained in the component which contains any one of its points.
We prove (3) as follows~ Let A be a connected subspace of X which
is both open and closed+ By (2))1 A is contained in some component C.
If A is a proper subset of C 1 then it is easy to see that

C = (Cf\ A) U (C

A')

is a disconnection of C. This contradicts the fact that C, being a component, is connected, and \Ve conclude that A = C.
, Part (4) follo,vs immcdiat~]y from Theorem B; for if a component
C is not closed, then its closure C is a connected subspace of X \Vhich
properly contains C, and this contradicts the maximality of C as a connected subspace of X.

In view of parts (3) and ( 4) of this theorem~ it is natural to ask if a

component of a space is necessarily open. The answer is no, as the
following example shov.rs. Let X be the subspace of the real line which
consists of all rational numbersa We observe two facts about X. :First~
if x and z are any two distinct rationals, and if x < z, then there exists
an irrational y such that x < y < zj and therefore

[X

n (-

a:i

,y)]

[X n (y,+

~ )l

14

Topology

is a disconnection of X 'i\ hich separat~s x and za It is easy to see from

this that any subspace of X with more than one point is disconnected, so
the components of X arP its points+ Second, the points of X are not
open, for any open subset of R ~Thich contains a given rational number
also contains others diffPrent from it. HerP 1 then~ is a space whose
components are its points and VrThose points are not openL This example
also sh o"\\~s that a space 1\eed not be discrete in order that each of its
points be a component~
1

Problems

1.

If A 1, A 2~ , An, ~ .. is a sequence of connected subspaces of a

topological space each of which intersects its succussor~ show that

u:_

2.
3.

4.

5..

An is connected4

Show that the union of any non-empty class of connected subspaces

of a topological space each pair of 1\...hich intersects is connected.
In 'Theorem 31-E we proved that a product space is connected if its
coordinate spaces are. Devise a di ff cren t proof of this fact, based
on Theorem A, for the case in which there are only two coordinate
spaces.
Use Theorem A to prove that an arbitrary Banach space B is connected. (Hint: if xis a vector, sho'v that the set of all sca1ar multiples of x is a connected subspace of B.)
Let B be an arbitrary Banach space. A eo1u_,1ex set in Bis a non-empty
subset S \Vith the property that if X and y are in 8:1 then
z

6..
7~

BL

+ t(y -

x) = (1 - t)x

+ ty

is also in S for every real number t such that 0 < t < la Intuitively~ a
convex set is a non-empty set which contains the segment joining any
pair of its points.. Prove that every convex subspace of B is connected. Prove also that every sphere (open or closed) in B is
con vex, and is therefore connected.
Show that an open subspace of the complex plane is connected ~
every two points in it can be joined by a polygonal line:
Consider the union of two open discs in the complex plane which are
externally tangent to each other. State whether this subspace of the
plane is connected or disconnected, and justify your answer. Do
the same when one disc is open and the other closed~ and when both
are closed.
Consider the following subspace of the Euclidean plane: {(x, y) ~ x ~ 0
and y = sin (1/x) J. Is this connected or disconnected? Why?
Answer the same questions for the subspace {(x,y) :x r!E 0 and
y ~ sin (I/x) l V {(x,y) ;x = 0 and -1 < y < 1}.

Connectedness

149

33. TOTALLY DISCONNECTED SPACES

We have seen that a connect.Pd space is one for which no disconnection is possible. We now consider spaces which have a great many
disconnections, and \Yhich therefore lie,. in a manner of speaking, at the
opposite end of the connectivity spectrum.
A totally disconnected space is a topological space X in which every
pair of distinct points can be separated by a disconnection of X. This
means that for every pair of points x and y in X such that x r!E y, there
exists a disconnection X ;;::: A V B vrith x e A and y e B. Such a space
is evidcn tly a Hausdorff space~ and if it has more than one point,. it is
disconnected. Oddly enough, a one-point space is both connected and
to tally disc onnccted
The discrete spaces are the simplest totally disconnected spaces.
A more interesting example is the space ...discussed at the end of the
previous section~ that is, the set of all rational numbers considered as a
subspace of the real line. The set of all irrational numbers is also a
totally disconnected subspace of the real line~ and this is proved in much
the same way, from the fact that. there exists a rational number between
any two irrationals.. The Cantor set is yet another totally disconnected
subspace of the real line, this time one which is compact.
Our first theorem should not rome as a surprise to anyone.
+

eomponent.s of a totally diseonnected Bpace are its point.a.

PROOF. If X is a totally disconnected spaceJ it suffices to show that
every .subspace Y of X which contains morr. than one point is discon..
nected. Let x and y be distinct points in Y, and let X :=:::: A U B be a
disconnection of X with x L4. and ye. B~ It is obvious that

Theorem A..

Tf~

(Y l\ A) U (Y

B)

is a disconnection of Y.
Tot.al disconnectedness is closely related to another interesting
property.

Let X be a Hausdorff space~ If X has an open base whose

sef8 are also closed, t.hen Xis totally disconnected .
PROOF..
Let x and y be distinct points in X. Since X is Hausdorff,
x has a neighborhood G \\rhich does not contain y. By our assumption,
there exists a. basic open set B 'vhich is also closed such th at x e: B C G~
X ;:;;: BUB' is clearly a disconnection of X 'vhlch separates x and y.

Theorem B.

If the space X in this theorem is also compact, then the implication

can be reversed, and the two conditions are equivalent.

1.50

TopoJogy

Theorem C. Let X be a compact Hau!3dorff s-pace. Then Xis totally di!3connected {::::} it has an open base 'l.l)hose sets are also closed+
PROOF..
In view of Theorem BJ it suffices to assume th.at X is totaUy
disconnected and to prove that the class of all subsets of~ v.,..hlch are both
open and closPd forms an open base. Let x be a point and G an open set
which contains it+ We must produce a set B which is both open and
closed such that x e: B C G. We may assume that G is not the full
space~ for if G :;:::: X, then we can satisfy our requirement by taking
B = X. G' is thus a closed subspace of XJ and since Xis compa:ct)' G' is
also compact. By the assumption that X is tot.any disconnected~ for
each pointy in G1 there exists a set H,, which is both open and e1o.sed and
contains y but not x~ G' is compact, so there exists some finite class of
H 1/s, which we denote by {Hh H2~ ...
Hn}, with the property that its
union contains G' but not x. We define H by H = u: 1 Hf~ and we
obscnre that since this is a finite union and all the H/s are closPd as "~en
as open, Il is both ~pen and closed~ it contains G', and it does not contain
x. If vle now define B to be H', then B clearly has the properties required
of it.
+

Totally disr..onnccted spaces arc of considerable significance in

several parts of topologyJ notably in dimension theory (see Hure\vicz and
Wallman [21}) and in the classic representation theory for Boolean
algebras given in Appendix 3. 1

Problems

1.
2.

Prove that the product of any non..empty class of totally disconnected spaces is totally disconnected.
Prove that a totally disconnected compact Hausdorff space is
homeomorphic to a closed subspace of a product of discrete two-point
spaces~

3.4. LOCALLY CONNECTED SPACES

Sec~

23 we encountered the concept of a locally comp.act space,

that is, of a space which is compact around each point but need not be
compact as a whole4 We now study another .i~local" property lvhich a
In

The reader should be made a.ware of the fact that several different definitions
of tota.1 disconnectedne.ss a.re commonly found in the liters.lure. The definition
given above seems to the present writer to have the logic of language behind it; and
Theorem C show.s that th.is definition is equivalent (in the case of R compact Hausdorff
spa.ce) to the moat important of the.se alternative de.finitionsr
l

Connectedness

151

topological space may have, that of being connected in the vicinity of

each of its points .
A locally connected space is a topological space with the property that
if x is any point in it and G any neighborhood of x 1 then G_ contains a
connected neighborhood of x.. This is evidently equivalent to the condition that each point of the space have an open base whose sets are all
1-

A
I

I .x
\

I
I
I
I

,_

,,,,

I
I
I

1/r

.-..1-

Fig4 264

2/-;r

- - - .......... ._,.,__ __

A U B is connected but not locrdly connected.

connected subspaces. Locally connected spaces are quite abundant, for~

as \vc have seen in Problem 32-5)' -every Banach space is locally connectedL
We know that local compactness is implied by compactness. Loca]
connectedness)' however, neither imp1ics, nor is implied by, connectedness.
The union of tiNo disjoint open intervals on the real line is a simple
example of a space which is locally ~onnectcd but not connected. .A. space
can also be connected without being locally connected, as the following
example shows. Let X be the subspace of the Euclidean plane defined
by X =A UB)' where . 4. = {(x,y)~x = 0 and -1 < y < ll and
B ~ {(.x,y) :0 < .x < 1 and y = sin (l/x) I (see Fig. 26). Bis the image
of the interval {O, l] under the continuous mapping f defined by
f(x)

;z::!

(x~

sin (1/x)),

so B is connected by Theorem 3 l~D; and since X :c:: B, X is connected by

Theorem 32-B Xis not locally connected, however~ for it is reasonably

152

Topology

easy to see that each point x in A has a neighborhood which does not
contain any connected neighborhood of x.
We know by Theorem 32-C that the components of an arbitrary
topological space X arc always closed sets, and from this we see at once
that the components of any closed subspace of X are also closed in X.
The reader may feel, 'vith some justification~ that the components of a
w cll-beha ved space ought to be open sets. This is true for locally

connected spaces .
Let. X be a locally connecte.d space. if Y is an open subBpace
of X, then each component of Y is open in X+ In particular, each component
of Xis open.
PROOF.
I_jet C be a component of Y. We wish to show that C is open
in X. Let x be a point in C+ Since X is locally connected and Y is
open in X, Y contains a connected neighborhood G of x~ It suffices to
sho"T that G C C. This will follow at once from the fact that C is a
component of Y if we can show that G is connected as a subspace of Y.
But this is clr.ar by Problem 16-6, according to which the topology of
Gas a subspace of Y is the same as its topology as a subspace of X; for
G is connected "\\'i th rcspe ct to the lat tcr topology

Theorem A.

The principal applications of local connectedness lie in the theory

of continuous curves (see Appendix 2) .
Problems

1.
2.

3.

44
5.

6.

7~

Prove that a topological space X is locally connected if the components of every open subspace of X are open in X ~
A connected subspace of a locally connected space X is locally
connected ii Xis the real line. "Wliv?
. Is this true if X is an arbitrary locally connected space?
Show that a compact locally connected space has a finite number of

components .
Show that the image of a local~y connected space under a mapping
which is both continuous and open is locally connected+
Prove that the product of any non-empty finite class of locally
connected spaces is locally connected .
Show that the product oi an arbitrary non-empty c]ass of locally
connected spaces can fail to be locally connectedL (Hint consider a
product of discrete two-point spaces+)
Prove that the product of any non-empty class of connected }or.ally
connected spaces is locally connected.

CHAPTER SEVEN

Approrimation
Our work in the present chapter centers around the famous theorem
of Weierstrass on the approxhnation by polynomials of continuous real
functions defined on closed intervals. This theorem, important as it is
in classical analysis, has been overshadowed in recent years by a generalized form of it discovered by 1' tone~ The latter rela tcs to con tjn uous
functions defined on compact Hausdorff spaces, and has become an
indispensable tool in topology and modern analysis.
W.e prove the 1\7 eierstrass theorem and then the t'vo forms of the
Stone-Weierstrass theorem which deal separately with real and complex
functions+ Finallyt after an excursion into the theory of locally compact
I!ausdorff spaces, we extend the Stone-Weierstrass theorems to this
context .

35~

THE WEIERSTRASS APPROXIMATION THEOREM

Let us consider a closed interval [a,b] on the real line and a polynomial
p(x} ~ ao

+ a1x +

+ anxn,

with real coefficients, defined on [a,bj. 1 Every such polynomial is

obviously a continuous real function, and as a consequence of the- second
lemma in Seca 20,. we know that the limit of any uniformly convergent
This polynomial can of wurse be re~arded &S a. (unction defined on the en tire
real line. We ignore this fact and consider only .x's which lie in [a,b J.
1

153

154

Topology

sequence of such polynomials is also a continuous real function. The

Weierstrass theorem states that the converse of this is also true, that is,
that any continuous real function defined on [a,b] is the limit of some
uniiormly convergent sequence of polynomials. This is clearly equivalent to the statement that such a function can be uniformly approximated by polynomials to within any given degree of accuracy.. Many
proofs of this classic theorem are known,, and the one we give is perhaps
.as concise and elementary as most.
Theorem A (the Weierstrass Approximation Theorem). Let f be a continuous real functian defined on a closed interval {a,b], and let t= > 0 be
given. Thn there exists a polyno-mial p toith real coefficients such that
rf(x) ~ p(x)I < f. for all x in la,bJr
PRO OF. As a first step, we show that it suffices to prove the theorem f Ol'
the special case in which a = 0 and b == la If a = b~ the conclusion
follows at once on taking p to be the constant polynomial--dcfined by
p(x) = f(a)~ We may thus assume that a < b+ We next obscnre that
x = [b - aJx'
a gives a continuous mapping of [O, l] onto [a,b], so that
the function g defined by g(x 1 ) = f([b - aJx'
a) is a continuous real
function defined on [0,1]. If our theorem is proved for the case in which
a = 0 and b = 1, then there exists a polynomial p' defined on [O, I} sueh
that Jo(x') ~ p 1 (x 1) r < E for all x' in [O~l]. If we now express this
inequality in terms of x~ WC obtain rJ(x) ~ p 1 ([x - a]/[b - aD I < E fol'
all x in [a,bJ; and defining a polynomial p by p(x) = p'([x - aJ/[b - a])
yields our theorem in the general case4 Accordingly, we may assume
that a = 0 and b = 1.
We next reca11 that if n is a positive integer and k an integer such

that 0

< k < n, then the binomial

coefficient (;) is defined by

(;) = n!/k!(n - k)!.

The polynomials Bn-=one for each n-defined by

are called the Bermtein polynomials associated with /. We prove our

theorem by finding a Bernstein polynomial 'With the required property.
Several identities will be needed for this. rfhe first is a special case
of the binomial theorem:

t (n)

I: mi:O

xi(l -

x)K-~

[x

+ (1

-- x)]R

(1)

Approximation

155

If we differentiate ( l) with respect to x, we get

and multiplying through by x{l - x) gives

1~ ( ; ) xk{l

x)"~(k -

nx)

= 0,

(2)

On differcntiating (2) with respect to x and considering xi (1 - x) n 4 as

one of the two factors in applying the product rule, v.. .e get

Applying (1) t.o (3) gives

(n) zk-

iO

(1

~ x)'f;~~ 1 (k

- nx)z = n;

and on multiplying this through by x(l - x), we find that

(n)

i z:::Q

xl'(l - x)"-k(k - nx) 2

nx(l - x),

or, on dividing both sides by n 2,

(n)

l:ccO

x-1(1 - x)n-'"

(x ~ nk)2 ~ .x(l n- x).

(4)

Identities (1) and (4) will be our main tools in showing that Bn(x) is
uniformly close to f(x) for all sufficiently large n.
Now for the proof of the fact just stated. By using (I), we see that
f(x) - B8{x) =

10

~ (~) x (1
1

- x)-k [ f(x) -

(!)}

so that
lf(x) - B .. (x)I

10

< ~ (~) xi.(1

- x) 8 - t f(x) -

f (:)

(5)

Since f is uniformly continuous on [01 IJ, we can find a lJ > 0 such that
Ix - k/nj < a~ ~f(x) ~ f(k/n)I < f./2. We now split the sum on the

156

Topology

right of (5) into tvt'"o parts, denoted by ~ and X':1 where Z is the sum of
those terms for which jx - k/nf < o (we think of x as fixed but arbitrary) and where %' is the sum of the remaining terms. It is easy to see
that :2; < t:/2. We complete the proof by showing that if n is taken
sufficiently large, then l;' can be made less than f/2 independently of x.
Since f is bounded, there exists a positive real number K such that
jf(x}[ < K for al1 x in (0,1]+ l~rom this it follows that

~t < 2K ! (~) xk(l

- x)"-k,

\vhere the sum on the right-denote it by ~"-is taken over all k such
that Ix - k/n r 2: 8. It now suffices to sho''{ that if n is taken sufficiently
large, then lo" can be made less than t:/4K independently of x. Identity
(4) sho1~ls that
11
x(l - x)

-<

~2

"""''' <

so

'

x(l ~ x )__
~tn

The maximal value of x(I - x) on [0,1] is }i, so

,,
I
~

<
.
- 4o~n

If we take n to be any integer greater than K/ 0 2 ~, then ~ < f/4Kt

Z' < f/2, IJ(x) ~ Bn(x)! < E for all x in [0,1], and our theorem is fully
proved.
11

The W cierstrass theorem clearly a.mounts to the assertion that for

any closed interval [a~bJ on the real line~ the polynomials arc dense in the
metric space e(a~bJT This i~ the form of the theorem "~hich 'vc shaU
generalize in 1.he next section to e(X,R), where X is an arbitrary compact
Hausdorff spare.
The slightly restricted statement that thP. polyno1nia1s are dense in
e[Ot 1J has another generalization, in a different direction. 1,his result
is so remarkable that \ve state it because of its intrinsic interPst.} though
we give no proof. The Weierstrass theorem for e[0,1] says, in effcct:1
that all renl linear combinatious of the functions
l, x,

x 2~ ,

xn. t

are dense in efO, 1], u~hcrc by a real linear comhina.tion of these functjons
we mean the result of choosing any finite set of them, multiplying each by
a real number~ and adding. Instead of working with all positive pov;rers
of x, 1et us pennit gaps to occur, and consider the infinite set of functions
1'

X 'H

x"' .... , x.,. .., ..... ,

'.,
t

Approximation

157

the n.1::'B being positive integers for which ni < n2 < < n.1:: <
The result we have in mind is called Munh'.s theorem, and asserts that aH
rea1 linear combinations of these functions are dense in e[O,l] -=> the
series ~: 1 1/n.t divergesr For a proof, we refer the interested reader to
Lorentz [29, pp. 46--48J or Achieser [I~ PP~ 43--46].

Problems

<t

2.

3.

Prove that e[a~bJ is separable.

l.J:t f be a continuous real function defined on [O, 1].

The moments
of f are the numbers
f (x)x" dx, where n = 0, I , 2, . . . . Prove
that two continuou8 real functions defined on f0, 1] are identical if
they have the same sequence of moments.
lTse the Weierstrass theorem to prove that the polynomials arc
dense in e(X,ll) for any closed and bounded subspace X of the real
line.

Jo

36. THE STONE-WEIERSTRASS THEOREMS

Our purpose in this section is to lay bare the true nature of the
Weierstrass approximation theorem. W-e achieve this end by generali.~ing the theorem in such a manner that its inessential features are

stripped away~
Our starting point is the fact that the polynomials are dense in
e[a,b] for any closed interval [a,b]. We ~rish to replace [a,b] by an
arbitrary compact Hausdorff space X and to make a similar statement
about e(X~R). The most obvious difficulty in this program is that it is
meaningle&13 to speak of polynomials on Xr This obstacle will disappear
when we take a closer look at what polynomials are~
Consider the two functions I and x defined on [a,b].. The set P of
all polynomials on {a,b] is identical with the set of all functions which
can be built from these two by applyjng the following three operations:
multiplication~ multiplication by real numbers, and addition. P is an
algebra of real functions on [a~bJ, for it is cJosed with respect to these
three operations. Even more~ it is a subalgebra of ~[a,bJ. We say that
P is the subalgebra of efatbl generated by {l ,x} 1 for it is a subalgebra
containing {l ,x} which ii:~ contained in every subalgebra with this property. We knov-.,.. by Problem 20-3 that the closure of a subalgebra of
e(X~R)-.---,-for any topological space X-is also a subalgcbra of e(X,R).
We may therefore speak of the closure P of P as th.e closed subalgebra of
e[a,b J generated by {1,x} . As above, this means that P is a closed subalgebra containing {I 1x} which is contained in every closed subalgebra

158

Topoiogy

with this property4 These ideas make it possible for us to state the
Weierstrass theorem in the following equivalent forms~
(1) the elosed subaJgebra of e[a,b] generated by ~ 1,x I equals
era,b];
(2) any closed subalgebra of e[a,b] which contains {1,x} equals

e[aJbJ.

..

These are potent statements, aaying 7 as they do~ that the very small set of
functions {1,x} suffices to generate the much more extensive set e[a,b].
As our theorems below will show, these statements depend only on the
fact that a closed subalgebra of e[a~b] which contains the set : I~x l
separates points in the sense of Sec. 27 (for it contains the function x)
and contains all constant functions (for it contains the non-zero constant
function 1).

Before we go further)' it is 'vorth observing that ~tatement (1) is not

true in general if either 1 or xis omitted from the generating set.. If xis
ornitted 7 then the closed subalgebra generated by i I} consists of the
constant f unctions 1 and this is not equal to e[ a ,b] unles,s a = b. On the
other hand? if 1 iB omitted, then the closed subalgebra generated by {x}
contains only functions which vanish at 0, and if 0 is in [a,.bJ, then the nonzero constant functions~ among others, are not in this closed subalgebra~
Our theorems rest on two lemmas, both of 'vhich have to do with the
fact that e(X,R) is a lattice for any topological space X. If f and g are
functions in e(X,R), we recall that their join and meet are d~.fined by

U v g)(x) =
and

(f A g)(x)

max {/{x)~g(x)}
min {/(x) 1 g(x) 1-

Our first lemma states conditions which guarantee that a closed sublattice of e(X,R) equals e(X 7 R)~

lemma. Let X be a compact Hau8durff space tvith more than one pO'int,
and let L be a closul sublattice of e(X 1R) with the following P'fOperty if
x and y are distinct points of X and a and b any two real numbers, then there
exists a junction fin L Bt1.Ch that f(x) = a and f(y) = b. Then L equals
e(X,R). 1
PROOF. Let f be an arbitrary function in e(X,R).. We must show that
j is in L. Let E > 0 be given. Since Lis closed1 it suffices to construct a
function g in L such that f(z) - E < g(z) < f(z) + E for an points z in
4..

If X has only one point J then a. single eonst&n t function constitutes a closed
iu blat tice of e( X ~R) with the stated property which does not equal e{ X, R). It is
therefore nece.ssa.ry to assume th.a. t X has more than one point+ Further, the reader
wi1l notice that the proof given below makes no use of the assumption th.at X is
Hausdorff. HoweverJ if there existe a closed sublattice of e(X,R) with the stated
property then X iB necesap.rily Hq.usdorff r eo tl;lere .is nothing to be gained by omitting
t

J'

this

886\lm ptioJl~

Approximation

X, for it will follow from this that

Ill - gll < E+

159

We now construct such

a function.
I,,et x be a point in X '\\Thich is fixed throughout this paragraph)' and
let y be a point different from x~ By our assumption about L, there
exists a function !'JI in L such that fv(x) = f(x) and fu(Y) = f{y)~ Now
consider the open set G11 = {z :f11 (z) < f(z) + ~d. It is clear that both
x and y belong to G"~ so the class of Gi/s for all poiuts y different from
x is an open cover of X. Since X is compact? this open cover has a
finite subcover, which we denote by {Gh G12)' ~ . , Gft}. If the corre..sponding functions in l.J are denoted by /1, f2i "
In, then
I

9~ =

f 1 A /i

'

"

I\

fn

is evidently a function in l..i such that g.z(x) ~ f(x) and g.z(z) < f(z) + e
for all points z in X.
We next consider the open set H ~ = {z: g.:r:(i) > /(z) - E 1- Since
x belongs to H~t the class of H 2's for all points x in Xis an open cover of
X~
The c.ompactness of X implies that this open cover has a finite subcover, which we denote by {111,. H 2 ~ . ~ ~ H m}.. We denote the corresponding functions in l..1 hy gi~ !J2J . ~ . , gm, and we define g by
g = {}1 v 02 v . v Ym~ It is clear that g i~ a function in L with the
property that f (z) - --= < g(z) < f (z) + E for all points z in X 7 so our
proof is completer
In our next lemma \ve make use of the concept of the ab~olute value
of a function. If f is a real or complex function defined on a topological
space X, then the function !fl-called the absolute value of /-is defined by
I/[ (x) == lf(x)I. If f is rontinuousJ then ]fl is also continuous. We
observe that the lattice operations in e(X~R) are expressible in terms of
addition, scalar multiplication, and the formation of absolute values:
fv g

and

f "0

=f +
=

+ If - ol
2

+ g - If - gl .
2

These identities show that any linear subspace of e(XiR) which contains
the absolute value of each of its functions is a sublattice of e(XjR).

Let X be an arbitrary topological ~~pace. 'Phen every closed subalgebra of fB(X,R) is also a closed sublattice of e(X ~Il).
PROOF.
Let A be a closed subalgebraof e{X,R). By the above remarks,
it suffices to show that if f is in ..4, then ]/' is al:oo in A. I_Jet ~ > 0 be
given. Since lti is a rontinuous function of the real variable t~ by the
Weierstrass approximation theorem there exists a polynomial p 1 with
the property that ~ ltl - p'(t)I < ~;2 for every t on the closed interval

lemma.

160

Topology

[- H!H ,Ir/JI]~

If p is the polynomial Vt,.hich results when the constant

term of p 1 is. replaced by O, then pis a polynomial "'Tith 0 as its constant
term which has the pr<?perty that I J ij - p (t) r < E for every t in
[ -1~/ll, rlfll ]. Since A is an algebra, the function p(f) in e(X,R) is in A.
By the stated property of p, it is easy to see that f lf(x)f - p(f(x))I < f:
for every point x in X, and from this it follows that 11 I/I -- p(f) H< f:.
We conclude the proof by remarking that since A is closed 7 the fact that
lff can be approximated by the function p(f) in A shows that j iE in A.

rt

We are no\v in a position to prove the Stone-Vl eierstrass theorems.

Theorem A (the Real Stone-Weierstrass Theorem).

Let X be a compact
Hausdorff space., and let A be a closed subalgebra of e(X:1ll) which separate8
pointa and contains a non-zero constant function. 1lhen A equals e(X 7 R) .
enoo~~. If X ha8 only one point, then e(X,R) contain~ only constant
functions; and since A contains a non-zero constant function and is an
algebra, it contains a.ll constant fun-etions and equals e(X,R). '\\',.. e may
thus assume that X has more than one point. By the above lemmas, it
suffices to sho\V' that if x and y arc distinct points of X, and if a and bare
any two real numbers, then there exists a function f in . ..4. such that
f(x) = a and f(y) = b. Since ..4 separates points, there exists a function
g in A such that g(x) ~ g(y). If we no'\,. define f by
f(z) == a g(z) - g(y)
g(x) - g(y)

+ b g(z)

~ g(x)'
g(y) ~ g(x)

then f clearly has the required properties+

V{e next turn our attention to the complex case, that is, to conditionf
which guarantee that a closed suba1gcbra of e(X,C) equals ~(X,C).
It is first of all necessary to understand that t.he conditions of Theorem A
will not suffice. 1;he simplest example which shows this requires a little
kno\vledge of the theory of analytic functions, and the reader without
such knowledge may skip at once to the next paragraph. Let X be the
closed unit disc {z ~ lzl < 1} in the complex plane. Xis clearly a compact
Hausdorff space. Consider the set A of all functions in e(X,C) ~~hich
are analytic in the interior of X. A is evidently a subalgebra of e(X,C}J
and one secs that it is closed by using IVIorera's the~rem. A separates
points, for it rontains the function f defined by f(z) = z. It also contains
all constant functions. In spite of this, A docs not equal e(X~C); for
the function g defined by g(z) ~ z iR in e(X,C) but is not in A, since it is
not. differentiab]e at any point.
What can be done to salvage Theorem A in the complex case?
".fhe answer lies in the operation of conj l lga tion discussed at the end of Sec.
20. If f is a complex function defined on a topological spaee X, \Ve

App roxi motion

161

remind the reader that its conjugate! is defined by f(x) ~ f(x). It wiU
also be convenient for us to define the real part and the imaginary part off:
R(f) =

f +J
2

and

l(f)

f-. J.
2i

(1)

We observe that if a complex function f has different values at two distinct

points of Xj then at least one of the functions R(f) and I(/) also has
different values at these points.
Theorem B (the Comp,ex Stone-Weierstrass Theorem). Let X be a
compact Hausdorff space, and let A be a closrAl subalgebra of e(X,C) whi'ch
separates points~ contains a non-zero constanl function, and contains the
conjugate of each of its functions.. Then A equals e (X JC).
PROOF~

The real functions in A clearly form a closed subalgebra B of

e(X 1R). Let us assume for a moment that B equals e(X,R). If f is an
arbitrary function in e(X,C), then R(f) and I(f) are in ~(X,R), and are
thus in B. llut since/= R(f) ~ il(f) and A is an algobra, f is in A and
A equals e(X,C). It therefore suffices to shov,,. that B equals e(X,R) .
We prove this by applying 'fheorem A.
We begin by showing that B separates points. Let x and y be
distinct points in X~ Since A separates points, there exists a function
f in A which has different values at x a.nd y. As '"Te saw in the above
remarks, R(f) or I(f) also has different values at x and y. Since A is an
algebra Vt,..hich contains the conjugate of each of its functions~ formulas (1)
show that both Il(n and I(f) are in B, so B separates points. We next
show that B contains a non-zero constant iunctionL :By our hypothesis~
A contains some non-zero constant function g.. A is an algebra which
contains the conj ugatc of each of its functions, so g{j == IYP~ L'; a non-zero
constant function in B. Theorem A no\\,.. implies directly that B equals
e(X ~R), and our proof is complete.

The two Stone-Weierstrass theorems are among the most important

facts in modP.rn analysis.. The theory developed in the last three chapters
of this book could hardly exist without them, and they have many other
applications as well. 1
Problems

1. Prove the t:~vo~variahle Weierstrass approximation theorem: if

f(x,y) is a real function defined and continuous on the closed rer.t.angle
X ~ [a~b] X fr,d] in the Ruclidean plane R 2J then f can be uniformly
approximated on X by polynomials in x and y with real coefficients.
1

See Stone [40 J.

1&2
2.

3.

Toporogy

Let X be the closed unit disr, in the complex plane, and show that any
function f in ~(X,C) can be uniformly approximated on X by polynomials in z and z with complex coefficients.
Let X and Y be compact Hausdorff spaces, and f a function in
~(X X Y,C)~ Show that f can be uniformly approximated by
functions of the form "X7rac1 f q.;,, where the j/s are in e(X,C,Y) and the
g/s are in e(Y,C).

37 .. LOCALLY COMPACT HAUSDORFF SPACES

In Sec. 23 we defined a locally compact space to be a topological space

in which each point has a neighborhood vrith compact closure. Locally
compact spaces often arise in the applications of topology to geometry
and ana.1ysis~ and since those v..~hich do are almost always Hausdorff
spacesj we restrict our attention in this section to localiy compact
Hausdorff spaces .
The main fact a.bout Buch a space is that it can be converted into a
compact Hausdorff space by suitably adjoining a single pointL The
reader is perhaps familiar from analysis "\\,..ith the prototype of this process)'
in which the complex plane C is enlarged by adjoining to it an uideal
point~' called the point at infinity and denoted by oo. This ideal point
can be thought of as any object not inc)' and we denote by c~ the larger
set C U { ~ l - C~ is called the extended complex plane when the neighborhoods of cc (other than Coo itself)
are taken to be the complements in
C~ of the closed and bounded subsets
(i.e.)' the compact subspaces) of C.
These ideas add nothing to our understanding of the complex plane, but
they do clarify many proofs and simplify the statements of many theorems, and they are valuable for this
reason. Figure 27 gives an easy way
c
of visua1izing the extended complex
plane. In this figure~ the surface S
Fig. 27. The Riemann sphere~
of a sphere of radius Yi is rested tangentially on C at the origin. It is
-customary to call the point of contact the south pole and the opposite
point the north pole. The indicated projection from the north pole
estab1ishes a homeomorphism between s minus its north pole and c~ so

Approximation

163

from the topologiea] poi11t of view, S minus its north pole can be regarded
as essentially identical with the complex plane C~ The north pole of 8
can be considered to be the point at infinity, and passing from C to
CCC) amouuts to using the point Ct; to plug up the hole in C at the north
pole. When S is identified in this manner with the extended romplex
plane)' it is usually called the llieniann sphe-reL In summary)' the locally
compact Hausdorff space C has been made into the compact Hausdorff
space S by adding the single point ~ .
We now generalize the construction outlined above to the ra~c of' an
arbitrary locally compact Hausdorff space X. Let o:J be an ohje('t not
in X 1 and form the set Xd"J = XU l i } .. We define a topology on Xri'.i by
specifying the following as open sets: (i) the open subsets of X, regarded
as subsets of X "; (ii) the complements in X ~ of the compar.t subspaces
of x; and (iii) the full space
If lf\-Te keep in mind the fact that a
rompact subspace of a Hausdorff space is closed, then it is ea8y to sho'v
that this class of sets actually iH a topolog~y on X ~, and also that the given
topology on X equals .its relative topology as a subspace of x~. The
fol1o\ving are the main propert.iP.s of the topological space X ~
(1) X~ is co1npact. To prove this~ let t Gd be an open rover of X110.
\V c must produce a. finite subcover. If X !10 occurs among the (i./s~ then
{(_T.i} clearly has a finite sub-cover, namely~ {X t;t]). We may therefore
assume that each G1 is a set of type (i) or type (ii). At least one Gi, say
Gi@J must contain the point ~ )' and this set is necessarily of type (ii).
Its complement G1.a' is thus a compact subspace of X v.,Thieh is contained in
the union of some class of open subsets of ..:Y of the form (Ji{\ X, so it is
contained in the union of some finite subclass of these sets, say {G1 f\
X, G2 (\ X, ..
G1J. n X}
It is no'v easy to see that the class
{Gi~, Gh G2~ . ~
Gfl J is a finite subcover of the original open cover of
xi"JJ)' so lit is compact.
(2) X~ is Hausdorffr X is Hausdorff, so any pair of distinct points
in X ~ both of v-.,..hich lie in X can be separated by open su hsets of X, and
, th us can be separated by open subsets of X ~ of type {i). It therPf ore
suffices to sho\v that any point x-in X and the point oo can be separated by
open subsets of Xr:o-r X is locally compact, so x has a neighborhood G
"lvhosP closure G in X is compact. It is 110-...v clear that (; and (;1 are
disjoint open subsets of X r10 such that x .e G and O'J e G', so X ~ is Hausdorff.
The compact Hausdorff space X~ associated with the locally rompact
Ilausdorff space X in the manner described above is called the one-point
compactification of X 1 and the point ~ is called the point at infinity.
'\\Te kno'v that compact spaces are locally compact, so thPsc ideas apply
Vt'"it.hout change when X is a compact Hausdorff spaceL It is easy to see
that the locally compact Hausdorff space Xis compact <i=-:> oo is an isolated

x.

16'

Topology

point of X... It may seem useless to consider the one-point compactification of a compact Hausdorff space:1 but we shall see in the next section
that it enables us to weaken the hypotheses of the Stone. Weierstrass
theorems.
The one..point compactification is useful mainly in simplifying the
proofs of theorems about locally compact Hausdorff spaces~ As an
example, any space X of th.is kind is easily seen to be completely regular;
for X is a subspace of Xll0 1 which is compact HallBdorfi and therefore
completely regular, and every subspace of a completely regular space is
completely regular. Accordingly~ if x is a point of X, and G a neighborhood of x which does not equal the full space, then there exists a continuous real function f defined on X, all of whose values lie in the closed
unit interval [0 1 IJ, such that f(x) = 1 and f(G') = 0. This fact can
easily be generalizedJ again by using the one-point compactification, to
the case in which the point x is replaced by an arbitrary compact subspace of X.
Theorem A. Let X be a locally compact Hausdorff spaceJ kt C _be a compact subspace of X, and let G be a neighborhood of C which does not equal
the full space~ Then there exists a continuous real junction f defined on X,
all of whose values lie in the closed unit interval [O,l]J such that f(C) == 1
and f (G') = 0.
PROOF.
Let
be the one-point compactmcation of X. Then c and

xd',j

G' are disjoint closed subspaces of X ~, and by Urysohn's lemma there

exists a continuous real function g defined on Xr;0 1 all of whose values lie
in [0,1] 1 such tha.t g((}) = 1 and g(G 1 ) = OL If f is the restriction of g to
X, then f has the required propertiesr
This result is an important tool in the theory of measure and integration on locally eompact IIausdorff spaces.

Problems
1.

2..
3.

4.

Let X be a locally compact Hausdorff space, and C 1 and C 2 disjoint

compact subspaces of X. Show that C1 and C2 have disjoint neighborhoods whose closures are compact.
Show that a Hausdorff space is locally compact ~ each of its points
is an interior point of some compact subspace.
Let f be a mapping of a locally compact space X onto a Hausdorff
spa~e Y.
If f is both continuous and open, show that Y is also
locally compact.
Show that if the product of a non-empty class of Hausdorff spaces i~
loca.Jly compart 1 then each coordinate space is also locally compact .

Approximation

165

38t THE EXTENDED STONE.. WEIERSTRASS THEOREMS

Let X be a locally compact Hausdorff space. Our present purpose is
to generalize the theorems of Sec~ 36 to this context.
A real or complex function f defined on X is said to vanish at infinity
if for each t > 0 there exists a compact subspace C of X such that
If(x) r < If for every x outside of
On the real line, for instanceJ the
functions f and g defined by f(x) = e-~i and g(x) = (x 2
1)- 1 have this
property~ but the non-zero constant functions do not~ It is easy to see
that if X is compact, then every real or complex function defined on X
vanishes at infinity, so in this case the requirement that a function vanish
at infinity is no restriction at all.
We denote by eo(X,R) the set of all continuous real functions defined
on X which vanish at infinity~ eo(X,C) is defined similarly.. If f is a
function in one of these sets~ then since lf(x) I < e outsidR of some compact
subspace C of X, and f ~ bounded on C ~ f is necessarily bounded on all of
X. It followsirom this that ec(XJR) ~ e(X,R) and eo(X,C) c e(X,C).
~"'lurtherJ the remark in the prere-Oing paragraph sho\vs that when X is
compact we have equality in each ease .

c.

lemma..
e{X,C).

eo(X,R) and ee(X,C) are closed subalyebras of e(X,R) and

We first show that eo(XtR) is a closed subset of e(X,R). It

suffices to show that if f is a function in e(X ,R) which is in the closure of
eo(X 1 R) ~ then f vanishes at infinity. Let f > 0 be given. Since f is in
the closure of eo(X,R) 1 there exists a function g in eo(X,R) such that
Hf~ u1' < E/2~ and this implies that lf(x) - g(x) I < t/2 for all x. The
function g vanishes at infinity, so there exists a compact subspace C of
X such that jg(x)i < E/2 for all x outside of C. It now follows at once
that
PROOF.

l/(x)I = !{f(x) - g(x)J

+ g(x) l < ]f(x)

- g(x) l

+ jg(x)! < t/2 + t/2

==

t:

-for all x outside of CJ so f vanishes at infinity. The same argument shows

that eo(X JC) is a closed subset of e(X ,C).
We next show that if f and g are in eo(X 7 R)t then f
g is also in
eo(X,R), that isJ that f
g vani~hcs at infinity~ Let E > 0 he given.
Since f vanish.es at infinity, there exists a compact subspace c1 of X
outside of which If(x) I < .:./2. Similarly, there exiHts a compact subspace
C 2 of X outside of ~~hich lg(x)I < 1;/2. C : : : : C1 U C-i is then a compact
subspace of X outside of which

I(f
so

+ g)(x)!

;;::: l/(x)

+ g(x) I < ~f(x) I + !g(x) I < E/2 + ~;2 =

+ g vanishes at infinityr

t:1

We can prove in much the same way that

166

Topology

.
e 0{X,R) is also closed with reepect to scalar multiplication and multi...
plication; and since eo(X,R) is non..-empty (it contains the function which
is identically zero), it is clearly a subalgebra of e(X,R). Similarlyt
eo(X~C) is a subalgebra of e(X,C).

This lemma permits ua to regard eo(X,R) and e0 (XiC) as algebra.a of

functions in their own right. We next establish a natural and useful
connection between continuous functions defined on X which vanish at
infinity and continuous functions defined on Xm which vanish at the
point oo, Vtrhpre of course X. is the one-point compactification of X.
It is important here to be quite clear about the distinction between these
concepra. ~.,or a function on X to vanish at infinity means precisely what
the above definition says. Such a function need not have 0 as a value.
On the other hand~ to Bay that a function on x' vanishes at the point 00
is to say that this function assumes the value 0 at the point ct)
4

Lemma.

e(J (X ~R) eq:ual8 the set of all restrictions to X of those functions

in e(Xccitll) which vanish at the point -xi. Similarly~ eo(XJC) equals the
set of all restrictions to X of the functions in e(Xtl),C) which vanish at the
point

co ..

Let g be a function in e(Xm)R) which vanishes at the point a:i.

Since g is continuous at .;.o , for each E > 0 there exists a neighborhood
G of oo such that ~g(x) I < E for all x in G. By the definition of a neighborhood of DO given in Sec. 37, G is the iull space Xco or the complement
in X oo of a compact subspace of X.. In either easer there cl early exists a.
compact subspace C of X such that lu(x)i < E for every poh;it x in X and
outside of C~ In other words, the restriction f of g to X vanishes at
infinity:t so is a function in eo(X~R). We must also show, conversely,
that every function fin eo{X,R) arises in this way from some function
g in e(X~,R) which vanishes at the point oo . All that is necessary is to
define g by g(x) = f(x) for every x in X and g( oo) = O~ and to observe
that the condition that f vanishes at infinity is precisely what is needed
to guarantee that g is continuous at oo. The proof oI the second statement of the lemma is exactly the same.
FROOF.

The machinery given above is intended to m~ke the proofs of the

follo,ving t~,.o theorems relatively simple. They are called the extended
Stone-TVeierstrass theorems4
Let X be a locally compact Hausdorff spaceJ and let A be a
closed subalgebra of Go(X,R) which separate8 points and for each poi.nt in
X containB a function which does not vanish there4 Then A equals eo(X,R).
PnooF. Let XfJO be the one-point compactification of X~ By our second
lemma~ we can extend every function in A to a function in C!!.(X~,R) which
Theorem A..

App roximotion

167

vanishes at co
We denote the set of all these extensions hy Ao. Our
hypotheses imply that Ao is a closed suhalgcbra of e(XG!;l,R) w~hich separates points and has the property that all its functions vanish at ;.(). Let
A 1 be the set of f unct.ions obtained by adding all constant fun ct.ions to
each function in Ao~ It is easy to see that .A.1 is a closed subalgebra of
e(Xc:o~R) which separates points and contains a non-zero constant func~
tion, so by Theorem 36-A, Ai equals e(XCJ), !~)~ It follows from this that
Ao consists of all functions in e(X~tR) which vanish at ct), and another
application of our second lemma shows that A equals eo(X~R),.
4

Theorem B. Let X be a locally compact llausdorff space, and let A be a

closed subalgebra of eo(X,C) whU:h separates points, for each point in X
eontains a function. which does not vanish there, and contains the conjugate of
eac.h of its functions. Then A equals. e[)(X,C).
PROOF~
The proof of Theorem A will serve here almost word for v-.,.ord~

We observe that when Xis assumed to he compact in the above two

theorems~ so that eo(X,R) = e(X,R) ai:id ei)(X~C) = e(X,C), then
they constitute slightly stronger forms of the Stone-,Veierstrass theorems,
for they yield the same conclusions under slightly \veaker asi:;umptions,.
Problems

1.
2~

3.

If X is a locally compact Hausdorff space_. prove that Go{X,R) is a

subla ttice of e (X ~R).
Let X be a locally compact Hausdorff space, and show that the weak
topology generated by '2u(X,ll) equals the given topology.
Let X be a locally compact Hausdorff space and Sa subset of <3o(X,R)
which separates points and for each point in X contains a function
which does not vanish there. Sho\v that the weak topology generated by 8 equals the given topology.

PART TWO

Operators

CHAPTER EIGHT

.Algcbraie S11stcJ1ts
We have seen in the preceding chapters that one of the basic aims of
topology a11d modern analysis is the study of the bounded real and complex functions which are defined and continuous on a topological space X
These functions can of course be studied individuallyJ but this doesn't
carry us very far. It is desirable to consider the sets e(X,Jl) and e(X,C)
of all such functions as mathematical sys terns with a high level of internal
organization, and this program compels us to give serious attention to
their structural features. It is at this point that algebra enters the
picture; for modern algebra is essentially the result of crystallizing into
abstract form~ and studying for th~ir own sake, a f cw simple patterns of
structure which underlie many diverse po.rts of mathematics.
In Secs. 14 and 20 we defined what is meant by a linear space and
an algebra, but we did not develop the theory of these systems to any
appreciable degree. We used them only descriptively, as a convenient
means of calling attention to the fact that the points in the spaces R~ and
c~ can be added and multiplied by numbersJ and those in e(XJR) and
e(X,C) can be multiplied together as well. Our work in t}).e rest of this
book requires a deeper understanding of these systems and several
.others_. and the purpose of this chapter is to provide a concise but reasonably complete exposition of this r;iecessary background material.
The algebraic systems we discuss below~groupsJ ring8, linear spacesJ
and algebras-have been the subject of many books and innumerable
research articles. In the few page.a we devote to each, we clearly can do
little more than explain what each system isJ mention several outstanding
examples, and develop the theory to the limited extent required by our
+

171

172

Operators

later work. If the reader finds it desirable to ampliiy our abbreviated

treatment by consulting additional sources, we suggest McCoy [31J and
Halmos !17}.

39. GROUPS
We begin by considering two familiar algebraic systems~ each of
which is a group, with a view to pointing out those features common to
both which are set forth abstractly in the general concept of a group.
We firstobserve that the set R of all real numbers) together with the
operation of ordinary addition, has the following properties: the sum
of any two numbers in R is a number in R (R is closed under addition);
if x, y, z are any three numbers in R, then x
(y
z) = (x
y)
z
{addition iB associative); there is present in Ra special number 1 namely 0,
with the property that x
0 = 0
x = x for every :t in R (R contains
an additive identity element); and to each number x in R there corresponds another number in R, its negative -x, with the property that
:t
(-x) = (-x)
x ~ 0 (R contains additive inverses)~
It is equally clear that the set P of all positive real numbers, together
with the operation of ordinary multiplication, has the following corresponding properties: the product of any two numbers in P is a number
in P (Pis closed under multiplication); if x, y 1 z are any three numbers in
P, then x(yz) =- (xy)z (multiplication is associative); there is present in
P a special number, namely 1, with the property that xl = lx = x for
every x in P (P contains a multiplicative identity element); and to each
number x in P there corresponds another number in P, its reciprocal
1/x = x~j with the property that xx- 1 = r- 1x = 1 (P contains multip] ica ti ve in verses).
Each of these systems plainiy possesses many properties other than
those we have mentioneL We ignore all such properties and concentrate
our attention solely on the ones we have listed. Let us now consciously
disregard the concrete nature of the elements composing the ahove sets
and the familiar character of the algebraic operations involved4 What
remains in each case is a non...empty set which is clo~ed under an operation
possessing certain f onnal properties~ and apart from notation and
tenninology, these properties are identical in the two systems~ The
concept of a group is a distillation of the common structural form of
these and many other similar systems.
The definition is as follows. A group is a non--empty set G together
with an operation (called multiplication) which associates with each
ordered pair xt y of clements in G a third element in G (called their
product and written xy) in such a manner that

+ +

Algebraic Systems

173

multiplication is associative~ that is, if x1 Y,~ z are any thrf>c

elements in GJ then x(yz) = (xy)z;
(2) there exists an element e in G, called the identity element (or
.simply the identity), \Vith the property that xe = ex = x for
every x in G; and
(3) to each clement x in G there corresponds another element in G1
called the inverse of x and written x- 1, "vith the property that
x.x- 1 = x-lx = e.
It should be carefully noted that we do not assume that xy = yx for alJ
elements x and y. A group which satisfies this additional condition is
called a eonlmutative group or an Abelt'an group (after the Norwegian
mathematician Abel). If G consists of a finite number of elements~ then
it is called a finite group and this number is called its arderJ otherwiseJ it
is called an infinite group.
In axiom (2) vlc speak of the identity, as if there v-.Tere only one
identity element in G. This is indeed the case~ for if e' is also an element
in G such that xe' == e1 x = x for every x, then e' = ere ::::: e shows that
e' necessarily equa1s e. Similarlyi in axiom (3) we speak of the inverse of
x, as if each element had only one inverse~ This also is true 1 for if x' is
another clement in G such that xxr = x'x = e, then
(1)

.:t'

== x' e = x r ( xx- 1) = {x 'x) x- 1

ex- 1

= x- 1

shows that x' equals x- 1..

If we have succeeded in disengaging ourselves from our intuitive
ideas1 we must admit that we know nothing 'vhatever about the actual
nature of either the set G or the operation. Both arc completely abstractJ
and it is essential to understand that our knowledge of G and its operation
is strictly confined to the information contained in the above axioms. 1
As our examples below will showJ the elements of a group need not be
numbers at all, and its operation can perfectly well be some bizarre rule
of combination which bears no relation to the usual operations of elementary algebra~ In its essence, the study of groups is the study of a
single a1gehraic operation in its purest form 1 and the theory of groups is
the body of theorems-together with their applications~"'"hich can be
deduced from the given axioms. This theory is richer in content than
can easily be imagined by anyone who has not delved into it for himself,
and the applications reach throughout mathematics and even beyond~
into such strikingly diverse fields as geometry, the theory of the solvaSome writers ero phasi ze this by ref erring to G as an abBlract groupL ThiE,
concept is then contrasted with th.at of a concrete grQl.lp, .such e..s the group of all rea.}
numbers with addition. The abstra.~t character of an .abstract group is sometime3'.
further emphasized by using a noncommittal symbol like x yin place of xy and by
speaking of the 81.ar operal.ion~ or the groop operation~ instead or multiplication.
1

174

Operators

bility of algebraic equations, crystallography, quantum theory, and the

theory of relativity~
Example 1. \Ve have seen that the real numbers form a group ~7 ith
respect to addition. In the present example we place this group in a
context of sevPral groups of the same type, that is, groups whose elements arc numbers, \vhose operation is ordinary addition, and in which
the identity is 0 and the inverse of a number is its negative.
(a) The single-element set consisting only of the number 0.
(b) The set I of all integers. Observe that the set of all nonnegative integers meets every requirement for a group except axiom (3),
so it is not a group.
(c) The set of all even integers. The odd integers do not constitute a group 1 for the sum of two odd integers is even.
(d) The set of all rational numbers.
(e) The set R of all real numbersT
(f) The set C of all complex numbers.
(g) 1.,he set of all complex numbers 'vhose real and imaginary parts
a.re both integers.
Exampte 2. \Ve also saw at the beginning of this section that the positive real numbers form a group v..ith respect to multiplication. Again"

thiR group is only one among many of a similar kind, some of v..~hich arc
listed below. In aU t.hesc the elements arc numbers, the operation is
ordinary multiplication, the identity is 1, and the inverse of a number is
its reciprocaL
(a) The single-element set consisting only of the number 1.
(b) The two-element set {1, - 1 }.
(c) The set of all positive rational numbers.
{d) 1'he set of all positive real numbers. Observe that the negative
real numbers do not form a group, for thls set is not closed under
multiplies tion.
(e) The set of all non-zero real numbers~ The set R of all real
numbers contains a numberJ namely 0 1 which has no reciprocal, .so it is
not a group 'With respect to the operation considered here4
(f) The set of all non-zero complex numbers.
(o) The set {l, i, - l, -i} of the four fourth roots of unity.
(h) The set {z ~z7; ~ 1} of the n nth roots of unity for a fixed but
arbitrary positive integer n. Groups (a), (b), and (g) are the special
cases which correspond to choosing n equal to 1, 2, and 4. We see by this
that there exists a finite group of order n for each positive integer n~
(i) The unit circle {z: lzl ~ 1} in the complex plane~ Thia is called
the circle group.

A,gebraic Systems

175

The groups in the next two examples a.re closely related to our work
in the previous chapters. They differ from the groups described above
in that their elements are not numbers~
Excrnple 3. (a) The set Rn of all n-tuples of real nun1bers 1 the operation being coordinatcvlise addition. The identity here is

o=

(07

and the inverse of the element. x

-x

o~
~

. . . , O) ~
(x1, x2 1

= ( ~x1:ii -X2 7

Xn) is

~ ~xn) .

(b) The set Cm of all n-tup]es oi complex numbers

coordinate,visc addition.

v..~ith

respect t-0

{a) The set e (X tR) of aH bounded continuous real f unctions defined on a topological space X. The operation here is pointv..ise
addition, the identity is the f l1T1ction 'vhieh is identically zero~ and the
inverse of a function f is the funr.tion - f defined by (-f) (x) = -f(x).
(b) The set e(X~C) of all hounded continuous complex functions
defined on a topological space xj the operation again being pointwise
addition.
Example 4.

The follo,ving examples are some\\'hat misceUaneous in charart.er.

They should be illuminating to the read~r \vho is not already familiar
with these ideas, for several have nothing \vhatever t.o do ""~ith numbers.
Example 5.

(a) The class of all sub:;ets of a set. l]i the operation being
the formation of svmmetric
di ff r.ren ces. 'The re a.dcr \vill rec all that t.h e
....
symmetric d ifferen ce of t V{O sets A and B is dP fined by
A

L).

(A - B) V (B - A);

and in Problem 2-3 it was sho\vn that this operation is associ.ativc 1 that
the idcn tity is the empty set 0~ and th.at the in verse of a set is the set
itselL It is interesting to note that if [] is non-emptyi theu this class
of sets doe!f not constitute a group \vith respect to the formation of either
unions or intersections.
(b) Ai:-y ring of subsrts of a set lI (see Problem 2-4), the operation
again being the formation of symmetric differences.

Let m be a positive integer a.nd define Im to be the set of all

non-ncgati ve integers less than m : Im == r0, It
Ht 1 }. If a and b
arc two numbers in Im, we define their ''sum'' a + 1; to be the remainder
obtained when their ordinary sum is divided by m~ If 1n i~ 7, for instance~
then /7 = {07 1, 2, 3, 4~ 5, 6l and we have 2 + 3 == 57 5 + 2 ;;- o~ and
Example 6t

..

'

176

Operators

4
5 = 2. Figure 28 is a complete addition table for I 1: to find the
sum of any two numbers in the set, Jook for the first number down the
left-hand side, look for the second across the topj and observe t.heir sum
in the corresponding placP within the table4
0 1 2 3 4
0
1
2
3
4

5
6

5 6

1 2 3 4 5
2 3 4 5 6
3 4 5 6 0
4 5 6 0 1
5 6 0 1 2
6 0 1 2 3
6 0 1 2 3 4

0
1
2
3
4
5

Fig. 28.
for I 1.

6
0
1
2
3

4
5

Example 7. The set of all one-to-one

mappin;s of a non-empty set X onto itself.
The opPration hPre is the multiplication of
mappings defined at the end of Sec. 3: if f
and g are two such mappings and x is an
arbitrary element in X, then (fg)(x) =
f(g(x)). The fact that this system forms
a group was sho,vn in Problems 3-1, 3-2;p
and 3-5.

The add1t.ion table

Example 8. Consider the special case of

the previous example in which the set Xis
taken to be a finite set lrith n elements~ e.g., the set {1, 2,
n].
A one-to-one mapping of this set onto itself is usually called a permutationt for it can be regarded as a rearrangement of the elements of the set.
If n is 4~ for instance, the permutation 'vhich sends 1 to 3, 2 to I, 3 to 4,
and 4 to 2 can be written in the convenient form
+

1234)

= ( 3142

'

lvhere be]ow each of the integers 11 2, 3, 4 is placed its image under the
mapping P~ If

(1234)
4132

is another such permutation, then their product pq (first q, then p) takes

1 to 4 then 4 to 2, 2 to 1 then 1 t.o 3 3 to 3 then 3 to 4, and 4 to 2 th en
2 to I . This result can be writ ten
j

pq

1234)
( 2341 .

The group of all permutations of n elements is denoted by S~ and called

the symmetric group of degree n. The detailed structure of symmetric
groups is of fundamental importance in the theory of the solvability of
algebraic equations.
Example 9.. Our final example is the group of symm.etries of a square4
Imagine that Fig. 29 represents a cardboard square placed on a plane
with fixed axes in such a way that its center is at the origin and its sides
are parallel to the axes. This square is carried onto itself by the follow-

Algebraic Sys fems

177

ing rigid motions: the identity motion I, which leaves fixed each point
of the square; the counterclockVtrise rotations R1 R', and R" about the
center through angles of 90, 180~ and 270 degrees; the reflections 11 and V
about the horizontal and vertical axes; and the reflections D and D' about
the indicated diagonals. Eaeh of these rigid mot.ions is related to a
certain aspect of the symmetry of the square, and they are therefore
called symmetries. We multiply t'vo symmetries by performing them
in succession~ beginning v..--ith the one on
the right. Acrordingly, RV is the result
v\_f.}
oi first reflecting the square about the
1
vertical axis, then rotating it counter- ~
I
/;
1
clockwise through 90 degrees. If \\""C D ',.....,2~""""""'"""'
~~~
/
. ::.: . .:.. .......... -
........ . . . .. ......
.
b
.....................
.... .
trace t h e e ff ect o f t h ese motions Y
}?;_~'f:\lll!!)l!l!!i!ii2!41Ji
following the manner in which the num.:i.~+0:~+:.:1.::_:,::/~::m
)~~ ~ ~~ ~ ~ ~ ~ ~ ~ ~ ~::: :~ ~:: ~~~~: ~ ~ ~~~c~:
bered vertices are shifted about, we see - - ~ ::0:2:T0,. <7:717
~~~= ~:-:::::: :~~:~ ~:::;;: ~~~~~~~~::..--:::::-

:;t _R~. h;h;:: ::: ;;:~!e~;ic~' t8;_

~~~'.~);,:~--~~~1

gcther with the operation we have de~ n' / 3

1
scribed, are easily seen to form a group. ~
:
Associativity is a special case of Prob1
lem 3-1; I is evidently the identity; and
1
it is clear that H, V ~ D ~ and D' arc their Fig. 29. The syn1metrics of a
own inverses and that R- 1 = R'' ,. square.
R'-1 = R'~ and R 11 - 1 = R. In much
the same way 1 we can define the group of symmetries oi an isosceles
trianglc 1 an equilateraJ triangle, a rectangle, a regula1 pentagon, etc. 1
and in each case the group describes in a precise fashion the ''symmetry
charactcristicsu of the figure+ We can go even further and consider
the group of symmetries of a regular solid in ordinary three-di me nsional
space. Groups of this kind have interesting and important applications in crystallography+
We now return briefly to the consideration of a general group G.
One of the more elementary facts about G is that certain simple equations
are always solvable.
(4) If a and b arc any two clements of G~ then the equations
ax = b and ya = b have solutions x and y in Gr
To prove (4)~ we have only to observe that x = a- 1b and y = ba- 1 are
in fact solutions, since a(a- 1b) = (aa- 1)b = eb
b and
;::=

(ba- 1)a == b(a-1a) = be

b.

Not only are the equations in {4) solvable in G, but their solutions are
unique. This is a direct consequence of the f o11owing cancellation law
r

178

Operators

If a is any element in G, then ax = ax'~ x = z' and

ya .=. y'a ~ y ::;; y'.
We prove the first of these statements by multiplying ax = ax' by
a- 1 on the left. This gives a~ 1 (ax) ~ a~ 1 (ax'), from which we get
(a- 1a)x = (a- 1a)x1 , ex ~ ex', and x = x'~ The second is proved similarly~
It is sometimes useful to know that (4) is capable of replacing axioms
(2) and (3) in the definition of a group. This amounts to the assertion
that n G is a non-empty set which is closed under an associative multiplication 'With property (4), then G is a group. To prove thisJ we must
show that G has an identity element and that each element in G has an
inverse~ We reason as follows.
Let c be an element in GJ and e a solu...
tion of ye ;; c. If a is any element in G and x is a solution of ex .:.w:. aJ
then ea ;;::: e(cx) = (ec)x = ex = a~ so e acts as an identity on the left.
We still must show that ae = a. For any clement bin G1 denote a solution of yb e by b- 1 and call it a left inverse of bL In particular,. a- 1a = eL
It is clear that (a-1a)a~ 1 = ea- 1 = a- 1, so a- 1(aa- 1) == a- 1 ; and ii we
multiply both sides of this on the lcf t by a left inverse of a-i, we get
aa- 1 = e. It is now easy to see that ae = a, for
(5)

;:::=

ae

===

a(a- 1a) = (aa- 1)a

ea

;:::=

a.

As the reader will observe, we have not only shown that e is an identity
element~ but we have also shown that each element has an inverse in the
required sense.
A subgraup of a group G is a non-..empty subset H of Q which is itself
a group with respect to the operation in G~ It is easy to sec that the
identity e' in II equals the identity e in G; for e' e' = e.' = e' e, and by the
cancellation law in G we have e' = e. Also~ if xis an element in H and x'
is its inverse in H~ so that xx' = x'x ::::;; e, then x' equals the inverse x- 1
of x in G; for xx' = e and xx- 1 ~ e yield xx' = xx- 1, and another application of the cancellation law in G gives x' = x- 1 By these remarks, we
sec that a non-empty subset H of G is a subgroup of G {::::)it is closed
under multipliration 1 it contains the identity e of Gj and it contains the
inverse x- 1 of each of its elements x. Since xx- 1 = e, it is equally clear
that a non-empty subset H of G is a subgroup of G {::::> it is closed under
mul tip lica tion and the formation of inverses
Many of the groups described above are subgroups of other groupsL
For instance, in Example 1 it is easy to see that (a) is a subgroup of (c),
(c) of (b)~ (b) of (d) 1 (d) of (e) 1 and (e) of (f). The subgroups of Example 7 arc of particular importance, and are called transformation groups.
If the underlying set is finite, as in Example 8, a transformation group
is often called a permutation group. In the case of Example 9, for
instance, each symmetry can be regarded as a permutation of the numbers which label the vertices of the square; e.g.J the reflection H about the
L

Algebraic Systems

179

horizontal a.xis interchanges 1 .and 4t and also 2 and 3, so we may put

H =

G~~i).

The group of symmetries of a squareJ being a subgroup of the symmetric

group
of degree 4t is thus a permutation groupr It is obvious that
any group G ha8 I el and (}itself as trivial subgroupsr
Every group in our list of exampl~s is Abelian~ with the exception
of the last three+ "\Ve ask t.hc reader t-0 sho\v in Problem 9 that Examp1e 7
is non-Abelian 'vhenever the set X contains more than t\VO elements~
It will follo'\' from this that the symmetric group S,, is non-..A.beJian Vt,..henever n > 3~ T'his can easily be seen for "-q-t hy computing the product
qp~ ~There p and q are the permutations given in Example 8:

s_.

1234)
qp ::;: ( 3421 .
Since pq - qp~ . S-t is non-_A_beHatL We sa\v in Example 9 t.hat R l' = D~
If 'ye nov.r compute VR~ \Ve gPt l7R = D'~ so Rl" :#- v'" ll and the group
of sy1nmetries of a square is a1so non-Abelia.1L
'Vhen the theory of groups is studied for its o\Yn sake, thr. Pmphasis
is usually placed on non-A helian groups. 'rhe prrscnt section ho,vever,
is intended ma.inly to provide a proper foundation for our \vork in ~he
rest of this chapter, and AbP1ian groups are the ones of grPatPst in1port.anee for us. In the A.belian ease, t.he multip]icative notation usrd above
is often replaced by additive notation, in ,vhich the product xy is v.rritten
x + !J and ralled the .~um of x and y. CorrPspondingly, thP idrnt.it.y is
dPnotcd by 0 instead of e and i~ caH~d the zero f.! enient (or simp1y zero);
and the inverse of x is denoted hy -x instr.ad of x- 1 and is r.al1rd 1he
negative of x. AJ so, the operation of subtraction ii:i definPd by
1

and the e]ement x - y is r.allcd the difference bet\\~cen .r and !I An

Abe1ian group in v.hith this additive notation is nRed is raUed an additive
A belian group. It is clear that a subgroup of an additivP . A. belian group
is a non-empty subset ,,hich is closed under addition and thP formation
of negatives.
Problems

1.

Let G be a group, and show that (xy)- 1 = y- 1x- 1 for any t\vo e]ements x and y in G. Show also that (x- 1) - 1 = x for any e]ement
x in G.

180

2.

3..

4...
5.

6+

Operate rs

Let G be a finite non-empty set -...vhich is c1osed under an associative

multiplication with property (5). Show that G is a group~ (Hint:
prove property (4) by considering the mappings of G into itself
defined by f (x) == ax and g (y) = ya.)
Prove that a group with the property that x~ = e for every element
x is necessarily Abclian (needless to say, x 1 is the convcntionaJ
symbol for the product of x with itself: x 2 = xx).
Prove that a group of order n with n 5 4 is necessarily Abelian.
Let H be a non-empty subset of a group Gt and show that H is a
subgroup of G ~ xy~ 1 is in H whenever x and y are. We see from
this that a non-empty subset of an additive Abe]ian group is a
subgroup~ it is closed under subtraction.
Let G be a groupJ and let C be the subset of G defined by
C == [a : ax

7.

9.
10.

11.

xa for every x

E G} .

Prove that C is a subgroup of G. C is called the center of

Let ni he a positive integer~ consider the set

Im=

8.

{0~1~

+.

G~

tm~IL

and define the uproductH of any two numbers in it to be the remainder left \Vhen their ordinary product is divided by m. Construct a multiplication table similar to ~~ig. 28 for the non-zero
e1ements of I b Does this set ,vith this operation form a group?
Compute a similar table for the non-zero elements of / 7. Does
this system form a group?
Introduce a symbol for e.ach element of the symmetric group 83
of degree 3, and eonstruct a multiplication table for this group.
Show that n ! is the order of SB~
Show that the group of all one-to--0nc mappings of a non-empty
~et X onto itself is non-Abelian if X has more than tv..,.o elements.
Construr.t a multiplication table :for the group of symmetries of a
square.
I Jet G and G' be groupsr A mapping f of G into G' is ea11ed a homomorphism if f(xy) = f(x)f(y) for all elements x and yin G~ Assume
thatf is a homomorphism of G into G', and prove the following facts:
(a) f (e) = e', where e and e' are the identity elements in G and G';
(b) /(x~l) = f(x)-1;
(c) /((;) is a subgroup of G';
(d) 1- 1 ( { e' l) is a subgroup of G.
If a homomorphism is one-to-one, it is called an isomorphism. If
there exists an isomorphism of G onto G' j then G is said to be isomorphic lo G'. To say that one group is isomorphic to another is to say
that they have the same number of elements and the same group

Af geb rai c Systems

181

stmcture 1 and differ only with respect to such inessentials as notation

and terminology. The reader 'Will o bscrve that the function f defined
on the real line by f(x) = a:r:t where a is a fixed real number greater
t.han lJ is an isomorphism of the group of all real numbers with addi~
tion onto the group of positive real numbers with multiplication, so
that these two systems as groups are abstractly identical. Now let G
be an arbitrary group, and let f be the mapping defined on G by
l(a) = Ma, where M ~ is the mapping of G into itself given by
M 0 (x) = ax. Show that/ is an isomorphism of G into the group of
one-to-one mappings of G onto itself. This fact is ca.II ed Cayley ~ s
the.arcm, and it shows that from the abstract point of view the theory
of groups is coextensive with the theory of transformation groups~

40. RlNGS
We have seen that the set I of alJ integers is an additive Abelian
group with respect to the opcratio n of 01d ii tary addition. It is just as
important to observe that Ii~ also cJosed under ordinary multip1icat.ion
and that multiplication is linked to addition in a way which enriches t.he
structure of the system as a whole.. The theory of rings is the theory
of such systems.
A ring is an additive Abelian group R \vhich is closed under a second
operation called multiplication-the product of two element.a x and y
in R is written x~in such a manner that
(1) multiplication is associative, that is1 if x, y~ z are any three
elements in R, then x(yz) ~ (xy)z; and
(2) multiplication is distributive, that isJ if x, y, z are any three elements in RJ then x(y
z) = xy
xz and (x + y)z = xz + yz.
In other words, a ring ia an additive Abelian group whose elements ean
be multiplied as well as added 1 and in which multiplication behaves
reasonably with respect to itself and addition. We note particularly
that multiplication is not assumed to be commutative,
Many of the additive Abelian groups listed in the previous section
a.re also rings with respect to natural multiplications.

Each of the following rings consists of numbers, and addimultiplication are understood to have their ordinary meanings.
The single-element set containing only the number 0.
The set I of all integers.
The set of all even integers.
The set of all rational numbers.
The set R of all real numbers.

Example 1.

tion and
(a)
(b)
{c)
(d)
{e)

182

Operators

{/)

The set C of all complex numbers.

T'he set of all complex numbers whose real and imaginary parts

(g)
are both integers.

Example 2. (a) e(X,R), with point~Tise addition and 1nultiplication.

(b) e(X~C), \\=-ith point.wise addition and multiplication.

Example 3. Any ring of subsets of a set ll, \Vith addition and multiplication defined by A + B = A 6. B and AB ==== A n B (see Problems
2-3 and 2-4). The fact that a ring of set~ is a ring in our present sense
is the reason :for the name ring of sets.
Example .4. Let m be a positive integer~ and Im the set of all non-nega ..
tive integers less than m: Im = {O~ l, .
~ m - 1 l ~ If a and b are
tv10 numbers in I"'~ \Ve define their l~sum" a
and Hproductn ab to be
the remainders obtained 'vhcn their ordinary sum and product are
divided by m. If 1n is 6,, for instance,, then I = { O~ 1, 2 1 3, 4, 5}, and
v.,.c have 3
4 = 1 and 2 3 = 0.. Im "~th these operations is called
the ring of integers mod m~
T

+ {)
I

\Ve now consider a general ring ll~ l\.1any fami1iar facts from ele
mentary algebra are valid in R. Nevertheless, c~ch must be proved on
its o\vn merits from the axioms or previous theorems~ for one never knows
\\,.hen something \\'hich appears to be uohviousi~ will turn out to be false.
\VP have a]ready defined subtraction in any additive Abelian group
by x - y = x + {-y)i and it is easy to shov.,. that such statements as
~ (x - y) = y - x and x ;:::=- y {::::;} x - y ;::;; 0 arc true~
Problem 39-1
assures us that - (-x) = x. And so orL Properties of this kind relate
to the additive structure of R and are comparatively trivial. It is only
'vhen we consider multiplication, and its relation to addition 1 that \Ve
begin to encounter some interesting situ_ations.
We i1lustrate this by proving that xO = 0 for any element x in R~
First~ we

have
xO

+ xO =

x(O

+ 0)

= xO.

Our next step is to add ~xO (the negative of xO) to both sides of this on
the right, ''Thich gives

(xO

+ xO} + (-xO)

and by the associativity of addition

xO

\Ve

xO

+ (-xO);

can "',.rite this in the form

+ (xO + (-xO)) = xO + (-xO).

Since the sum of any element and its neg&:tive is 0, this collapses to

xO

'?.thich yields

+0 =
xO

0,,
04

Algebraic Systems

183

Similarly, Ox ~ 0 :for any x~ We see in this "Tay that the product of two
elementrs in a ring is zero whenever either factor is zeror We have given
the detai1 s of the proof of this seemingly obvious fact because, surprisingly
enough~ its converse is false~ It can perfectly well happen (and it often
does happen) that the product
two non-zero elements in a ring is .zero .
The simplest examples of this phenomenon are found in the rings of
integers mod m \\"here m is greater than 1 and is not a prime number.
We have already sccn:t for instance, that in /6 the product of the two
non-zero elements 2 and 3 is 0. An element z in a ring such that either
zx = 0 for some non-zero x or yz = 0 for some non-zero y is called a
divisor of zero. In any ring with non-zero elements, the clement 0 itself
is a divisor of zero.
By using distributivity and the fact that the product of two elements in the ring R is zero when either factor is zero, it is easy to verify
the following familiar rules oi calcu] ation: x (- y) = ( - x )y = - xy,
(-x)(-y) = xv:t x(y - z) = xy ~ xzj (x - y)z = xz ~ yz~ As a simple consequence of the last t\\.'O of these rules, vte have the following
rancellation la.v.r: if a is not a divi8or of zero, then either of the relations
ax =:; ay or xa = ya implies tha.t x ;;:::: y.
R is called a co1nmutafive ring if xy = yx for all clements x and y.
Every ring in the above Jist of examples is commutative. We shall
encounter some non-commutative rings of very great importance
in Secs4 44 and 45.
If the ring R contains a non-zero element 1 with the property that
xl = Ix = x for every x, then 1 is called an identity element (or an identity) 1 and R is calJed a ring with identity~ I:f a ring has an identity, then
it has only one. In Example 11 only (a) and (c) have no identity4
In
both rings described in Example 2J the identity is the function which is
identically 1. A ring of subsets o:f a set U has an identity ~ there exists
a non-empty set in the ring which contains every set in the ring; in particular 7 if ll is non-empty and the ring is a Boolean algebra of subsets
of U, then the set lJ itself is the identity. rrhe ring Im has an identity
<=>m > 1.
Let R be a ring \~:i th identity+ If x is an element in ll 1 then it may
happen that there is present in R an element y such that xy = yx ~ I~
In this case there is only one such clcment 1 and it i~ ,vritten x- 1 and
caned the inverse of X+ If an element x in R has an inverse~ then x is
said to be regular. Elements ~Thich are not reguJar are <~aUed singular.
Regular clements are often called invertible elementsp or 1ion-8'ingular
elements. The Plement 0 is alv... ays singu]ar in a ring "~ith identity~ and
the clement 1 is always regular~ In Example lb~ 1 and ~ 1 are the only
regu1ar elemPnts; in ld to 1/1 aU non-zero elements are regular; and in
lg, the regular elements are 1~ i, -1, and -i.

or

184

Operators

A ring with identity is called a division ring if all its

ments arc regular~ A field is a commutative division ring.
numbers constitute a field, as do the real numbers and
numbers. Roughly speaking, fields are the ''number
mathematics.

nonzero eleThe rational

the comp]ex
aystemsJ' of

Problems
l~
2.

3.

4.

5..
6.

Jn the ring of even integers, why is 2 neither regular nor singu] ar?
Consider the ring of all subBets of a non-empty set lJ., "'-it.h the
operations dPfined in Example 3. What are the regular clements in
this ring? "\Vhat are the singular clements? What are the divisors
of zero? llnder what conditions is this ring a field?
In each of the following rings of functions defined on the closed unit
interval [O, t]J descrjbe the regular clements~ the singular elements~
and the divisors of zero:
(a) an real functions;
{b) all continuous real functions;
{c) all bounded continuou!:l real function~~
What changes are Ilecessary in these descriptions if [0~11 is replaced
by (Orl)?
I...et R be a ring with iderttity, and shov{ that any divisor of zero in
R is singuln.r~
Let R be a ring ,vith identity} and show that Risa division ring<=> the
non-zero elements of R form a group \vith respect to multiplir.at.ion~
Show that the ring Im is a field {:::::> 1n is a prime number. (If int: in
sho\\ring t.hat Im is a field if m is primc:t shO\V first that in this ca.se
Im has no non-zero divisors of zero 1 so that the non-zero elements
of I"' are closed under multiplication and the cancellation law
ax ~ ay ==> x = y holds for these elements; now apply Problem 39-2
and Problem 5 above.)

41. THE STRUCTURE OF RINGS

A non.-empty subset S of R is called a subring

of R if the elements of S form a ring vrith respect to the operations
defined in Rr This is equivalent to the requirement that S be closed
under the formation of sums, negatives, and products.
We concentrate our attention on a special type of subring. An ideal
in R is a su bring I of R which has the f 'J:llowing further property:

J..ict R be a rjng.

i E I==> xi and ix

s I for every e]ement x e R.

Algebroic Systems

185

It is in this sense that an ideal in R can be described as a subring of R

which is closed ~~ith respect to multiplication on both sides by every
element of R. If the ideal I is a proper subset of R, then it is called a
proper ideal. T'he trivial ideals in R are the zero icleal f 0} consisting of
the zero element alone~ and the full rh1g R itself We see from this that
every ring with non~zero elements has at least two distinct ideals.
In order to clarify the concept of an ideal_, we mention a fe,v specific
examples. We begin by considering the ring of all integersL 1~he even
integers (i.e. 1 all integral multiples of 2) obviously form an ideal in this
ring. So also do all integral multiples of 3, of 4, and so on~ In genera],
if m is any positive integer~ then the set
L

m ===

{. ~ ~ -

2mJ -m, 0' m:r 2m~ . . . ~

of all integral multiples of m is a non-zero ideal. We next con8ider the

ring e[0,1] of all bounded continuous real functions definPd on the closed
unit interval. If X is a subset of [O, I}~ then the set

I(X) = {f:f(x)

0 for every x c X}

is an ideal in this ring~ It is easy to see that I(X) equals the full ring
'vhcn X is the empty set and equals the zero ideal when X = [O, l].
As a final example, we consider the ring of all subsets of an infinite set UJ
and we o bservc that the class of all finite su bscts of U is a proper ideal in
this ring.
Some rings have a multiLude of non-.trivial ideals, while others have
none at alL In gPneral, the structure of a ring is very closely ronnected
with the ideals in it+ The following theorem illuRtrates this point.
Theorem A. If R is a commutatitte ring with identity~ then R is a field ~
it has no non-triv1"al ideals.
PROOF. We first assume that ll is a field, and we show that it has no
non-trivial ideals~ It suffices to sho-.,v that if I is a non-zero ideal in R~
then I = R. Since I is non-zerot it must contain some element a ~ 0.
R is a field, so a ha::; an inverse a- 1, and I (being an ideal) contains
1 == a~ 1 a. Since I contains 1, it also contains x = xi for every x in R 1
and therefore I = R.
We now assume that R has no non-trivial ideals, and we prove that
Risa field by shomng that if xis a non-zero element in R, then x has an
inverse. The set I == {yx :y e. R} of all multiples of x by elements of
R is easily seen to be an ideal.. Since I contains x = Ixi it is a non-zero
ideal, and it consequently equals R. We conclude from this that I
contains 11 and therefore that t.here is an element yin R such that yx = l+
This shov{S that x has an inverse, so R is a field.

The real significance of the ideals in a ring is that they enable us to

186

Operators

construct other rings which are associated with the first in a natural
way. We explain how this is done~
Let I be an ideal in a ring R. We use I to define an equivalence
relation in R as follows: two elements x and y in R arc said to be con ...
gruent 1nodulo I~ written x = y (mod I), ii x - y is in I. Since only one
ideal is under consideration~ we abbreviate this symbolism to x
y.
It is easy t.o verify that we actually do have an equivalence relation here,
that iB, that the follo\\dng three conditions are satisfied:
(1) x = x for every x;
(2) x = y ===9 y = x;
(3} x ~ y and y = z =:- x
z.
Furthermore_, congruences can be added and multiplied, as if they were
ordinary equations:
(4) x1 = Zi and Y1 = Yi~ xi+ Yi
x~ +Yi and X1Y1 = X2Y2
The hypothesis of (4) is that x1 - x2 and Yi - Y2 are e]cmcnts of I, and
since I is an ideal, the conclusions follow at once from

(xi

and

+ Y1)
X1Y1 -

- (x2

+ y,.)

X~Y! =

;;::: (x1 - x~:)

+ X1Y2 z1(Y1 ~ y1} + {x1 -

X1Y1 -

XfY':l.

(y1 - y'J.)
X1Y2

x'A)Y2~

It will be necessary to use property (4) at a critical stage in our discussion

below~ and the reader will see there that this property is the main reason
why the ideals in a ring are so mu<~h more important than its subrings.
According to the general theory of Sec. 5i this equivalence relation
has associated mth it a partition of R into equiv alenec sets-r,al 1ed
coset./J in this context-\vhich are non-empty and disjoint and \vhose union
is the full ring fl. What is the structure of these cosets? In order to
anfi.nver this question~ \.Ve Jet x be an element of ll. The coset [xJ containing x is by definition the set of all elements y such t.hat y == x; that
is, [xJ = fy: y = x I . But

{y;y

xl

fy:y {y :y -

x e:

Jl

x ;:::= i for some i e: I J

~ {y: y = x + i for some i e: 11
- ~x + i: i EI l
A natural notation for the set In.st written is x + I, which we understand
to signify the set of all sums of x and clements of I~ The structure of the
coset (xJ containing x is fully exhibited by the far..t that [x] = x
1.
Sometimes it is convenient to denote this r.oset by [x 1 and sometimes
by x + I~ We recall that the sam~ r.oset can perfectly "~ell arise from
another element_, say x1, and that [xJ = [xd means that x = r1:' that is,
that x - x1 is in IL Thr c]ements x and x1 are called representati'l/es of the
coset which contains them~

Algebraic Systems

187

Our next step is to construct a ne\v ring_, which \ve denote by RII and
caH the quotient ring oi R with respect to 14 The elements of the ring
R/ I arc the distinct cosets of the form [.r] (or x
I)+ All that remains
is to define the manner in \Vhich these cosets arc to be added and multi..
plied and to verify the fact that. \\fe do indeed obtain a ring. The
definitions arc as follo\vs:

[x]

+ {y] =

ix] ' [y]

and

[x
y]
[xy]a

In other \Vords, we add and multiply t\vo cosets [x] and [y] by first adding
and multiplying the representatives x and y, and then by forming the
cosets \vhich contain x + y and xy+ It is necessary to make certain that
these arc legitimate definitions, that is, that the resulting cosets {x + y]
and [xy] do not depend on the particular representatives x and y chosen
for the cosrts [x1 and fy]~ ~ro this end, we take t\ {0 other representatives
of the same t~~o c:osct~, Lr.t t\\~o other elements X1 an<l y i of R such that
xi - x and Y1 - y. \Ve \Vant to satisfy ourselves that
11

rx + y .J

+ yJ
{xy], or equivalently~ that x1 + Yl
1

rx

and [x1y1] ==
- x
y and X1Y1 = XY~
Since this is prerisely the content of property (4) above, \Ve do have
valid dcfini t.ions for our ring opera t.ions in R /I. \Ve omit the detailed
vcrifir.at.ion of t.hP fact that Ii I I \vi th these operations actually is a ring,
rernarking only that the zero element of this ring is {01 = 0 + I = I and
that the negative of a typical element ~x] = x + I is [ -x] = (-x)
I.
It is easy to see that ll/ I is eommutative ii fl isi and that if R has an
identity 1 and I is a proper ideal~ then ll/ I has an identity 1 + I.
Wr summarize t.he results of this discussion in the follo\ving theorem.

Let 1 be an ideal in a ring fl, and let the coset of an elrnu~nt x

in R be defined by x + I = { x + i: i ~ I}
Then the dist-i net co set.~ JOT rn. a
partition of R; and if addition and multiplication are defined by

Theorem B.

(x +I)

and

(x

(y +I) = (x

+ l)(y + J)

+ y)

+I

xy +I,

then these cosets constitute a ring dennted by R/ I and called the quotient

ring of R with respect to I, in u. hich the zero elcnu~nt is 0 + I = I and the

negative of x + I is ( ~x) + I. f/urther~ if R t's corn1'nutative~ then R/ I is
also conunutative; and if R has an identity 1 and I is a proper ideal" then
R/ I ha .~ an identity I + I+
1

\Ve no\\~ give a brief account of homomorphisms and of the rnanner

in which ideals, quotient rings, and homomorphisms are all re]atcd to one
another.

188

Operators

Let R and R' be two rings. A homomorphism of R into R' is a

mapping f of R into R' '"ith the follo\\~ing two properties:
f(x
and

+ y)

f(x)
f(y)
f(xy) = f(x)f(y).
=

A homomorphism of one ring into another is thus a mapping of... the first
ring into the second ,vhich preserves the ring operations. It is easy to
see that f preserves zero in the sense that f(O) = 0 ;1 for

f (0)

+ f (0)

= f (0

+ O) = f (0),

and subtracting f (0) from both sides yields our result.

preserves negatives, for j(-x) = ~ /(x) foJlows from
f(x)

+ f( ~x)

= f(x

+ (-x))

= /(0)

Similarly,

O~

The image f(R) of ll under f is clearly a subring of R'. This subring

/(R)-which is R~ itself \vhcn f is onto~ ~-is eaHed a honiomorphic irna.ge
of R~ If the homomorphism f is one-to-one~ ri1en it is called an isomorphism~ and the suhring J(ll) is caUed an isoniorphic irna.ge of R.
An
isomorphic image of R {~an be thought of as a ring \vhieh is essentially
identical '"Tith ll, for it differs from R only in the matter of not.at.ion.
The properties of Rare reflected with complete precision in an isomorphie
image and with somev.r~hat less prr.cision in a homomorphic image+
Let f be a homomorphism of R in to R'. The kernel K of this
homomorphism is the inverse in1age in R of the .zero ideal in R':

{x:x e R an<lf(x)

=OJ.

It is easy to see that K is an ideal in R, and also that K is the zero ideal
in R ~ f is an isomorphism. W c leave these verifications to the reader.
In a sense, the size of the kernel K is a measure of the extent to which
f fails to be an isomorphism.
What is the real significance of homomorphism8~ homomorphic
images, and kernc1s? A full ans\ver t.o this questjon v;~ould carry us into
the utmost reaches of the general theory of rings~ """"here \\c havr. no
intention of treading. 'Ve ~will, hotrever, attempt a brief and ncce8sariJy vague partial ansv{er. Suppose that Risa ring ,~those f e.atures are
unfamiliar, v..Those structure is unknown. The question confronting us is,
'Wbat is the nature of fl? Andj as is often the r.asc in mathematics, if
we can state adequately \vhat this question means~ vire -aill have taken
a long step toward ans,vering it. Suppose now that Il' is a homomorphic
image of R and that R 1 is a wcll~kno'"~n ring which is intuitive]y familiar
We use the s ym ho I 0 here to designs. t.c two different elem en ts: the z oro in /~
and the zero in R". This convention is customary, and it aavea mor.c trouble than it
1

ca.uses~

Algebraic Systems

189

and thoroughly understood. R' then provides a picture of the structure

of R. The details of this pieture may be blurred and fragmentary~ but
'"Te can usually g1ean from them a :few hints as to the nature of R itself.
If v.Te have available many homomorphic images of ll 1 it i.'i3 of ten possible
to corte]ate the hint-s we get from these many sources in such a \\~ay as to
build up a fully detailed and completely precise pirture of the original
ring R~ This js the overall strategy in
the structure theory (or representation
~ 1ffiD~
theory) of rings~ 1 The relevance of
/
Ai
ideals to th is str ategie pattern dcp ends
on the following fact: all poRsible homomorphic images of R can be constructed
by means of the ideals in R. We next
describe how th is is ar.compl ished~
Let R be a ring~ and let f be a
homomorphism of R onto a ring R'~
Let K be the kernel of f. Since K is
an idea1 in R, we can fonn the quotient ring R/ K. W c now observe
Fig. :lO
that R/ K is a homomorphic image
of R under the homomorphism g--called the natural hornomorphism
-defined by
g(x) = x

+ K~

The fact that g is a hom-0morphism follo\vs directly from the definition

of the ring operations in R/K:

g(x

and

+ y) = (x + y) + K
g(xy) = xy

+K

+ K) + (y + K)
== g(x) + g(y)
= (x + K)(y + K)
;;::: (x

g(x)g(y).

Finally, we show that R/K and R 1 are essentially identical by producing

an isomorphism of R/K onto R'. Let a mapping h be defined on RIK by
h(x + K) :::::: f(x). We ]eave it to the reader to verify that h is a v.rclldefined mapping of R/K onto R' and is also an i.~omorphi.sm. Figure 30
gi:Ye~--~ schematic representation of this situation+ Since R/ Kand R' are
iBomorphie, we can replace R' in any discus.sion by its replica R/K. It is
The procedure described here is loosely similar to a. familiar technique from
three-dimensional s.nsly tic g_eom et"ry J in which the form -or a curved surf ace is studied
by means of itB crose sections. The information obtainable from any given cross
.section is meagerJ but a.n intel1igent consideration 0 s.ll the euccessive cross sections
c&n yield a Mtisfaetory mental image of the aurface ss a whole.
l

190

Operators

therefore unnecessary to go beyond the ring R to find a11 its homomorphic

images.
If I is an ideal in a ring R, then its properties relative to all of R are
reflected in corresponding properties of the quotient ring R /I, and many
aspects of the study of R depend on the presence in it of ideals whose
corresponding quotient rings are simple and familiar.
In order to illustrate this fundamental principle, we introduce the
following concept. An ideal I in a ring R is said to be a m.aximal ideal if
it is a proper ideal which is not properly (',ontained in any other proper
ideal. Our next theorem is an immediate consequence of this concept and
Theorem A.

C.

If R is a commutative ring u ith identity~ then an ideal I in

R is maximal ~ R/ I is a field.
PROOF.
We first observe that if I is maximal~ then R/ I is a. commutative
ring with identity in v,.Thich there are no non-trivial ideals, so by Theorem
A it follows that RI I is a field. 'Ve no\v assume that I is not maximal,
and we show that Rf I is not a field. rrherc are two possibilit..ie~; (a)
that I = R, and (b) that there exists an ideal J such that I C J C R.
In case (a)~ ll /l has no non-zero elements) so it canno i. be a field. In
case {b) 1 R/ I is a commutative ring \vith idcnt.it.Y ~rhirh eontains the
non-trivial ideal J /I, so again it cannot be a field.
Theorem

The commutative rings we study in lat.er chapters have a great

many distinct maximal ideals~ and this theorem will serve us v-.TeH in our
program of analyzing the structure of these rings.
Probfems

1.

2..

3~

Let R be a ring with identity v-.'"hich is not necessarily commutative.

In vie\\"" of --rheorem A, it is natural to .conjecture that R is a division
ring ~ it has no non~trivial ideals. Try to prove this conjecture by
the method used in the proof of Theorem Ar At what preeise point
docs this attempted proof break do\vn~~ lio\v much of the conjecture can you prove?
Let R be the ring of all real functions defined on t..he closed unit
interval [O, I]+ 'If X is a subset of [O_, 1J, sho\\' that the ideal 1 (X) in
R defined hy 1 (X) = { f :f(x) = 0 for every x s X l is maximal {::=} X
consists of a single point.
Let l be the ring of integers and m a positive integer~ It is easy to
see that if x is any integer, then x can be represented uniquely in the
form x = qrn + r,. \vhere q and r arc integern and r is in t.he set
{O, I} .
~ 1n - 11 ~ lT se this fact to show that a non-zero ideal
in I is necessarHy of the form fft = f.
~ -2m, -m 1 0, m, 2nil . . . j
+

Algebraic Systems

191

for some positive integer m~ Show that the mapping f defined on I

by f(x) = r is a homomorphism of I onto the ring I. of integers mod
m. Show that the kernel of this homomorphism is the ideal m, so
that the quoticn t ring I/ fli is isomorphic to I "'J and conclude from
this that mis maximal ~mis a prime number.

42. LINEAR SPACES

We introduced linear spaces in Sec. 14:, and we also mentioned a few
of their simpler propertiefL Our present purpose is to develop the
theory of these systems in some"'That greater detail.
We begin by restating the definition in terms of concepts now available to us~ The reader v... ill recall that by the scalars we mean either the
system of real numbers or the system of complex numbers. A linear
t~pace (or vector spacP) is an additive Abelian group I~ ("hose elements
are called vectors) 'vith the property that any scalar o. and any vector x
can be combined by an operation called scalar multiplication to yield a
vector ax in such a manner that
(l) a(x
y) = ax
ay;
(2) (a + fJ)x = ax
{3x;
(3) ( atl)x = a(fjx) ;
(4) 1 . x ;;; x.
A linear space is thus an additive Abelian group whose elements can be
multiplied by numbers in a reasonable way~ but not neces.sarily by one
another (as in the case of rings). The two primary operations in a
1inear space-addition and scalar multiplication~are called the linear
operations, and its zero element is usually referred t.o as the origin.
A linear space is called a real linear space or a complex linear space
according as the scalars are the re.al numbers or the complex numbers ..
The advantage of calling the numerical coefficients scalars is that we
avoid committing ourselves to either the real ease or the complex case
and are free to develop the theory for both simultaneously . 1 In later
chapters \Ve shall be concern~xl exclusively '"Tith complex linear spaces,
but for the presPn t ""rP. prPfer to lea vc the door opPn.
Ref ore proceeding to the general theory of linear ~paces, we Jist a
f ev-.T examp 1es~

+
+

. . -.~ In some approaches to th~ theory of linear spaees, the system of sc:U.a.rs is
allo""~d to be a.n arbitrary fieldT This degree or generality is unneeessn.ry for our
purposes; and since unuecesaa..ry generality is und~ira.Lle, we Emit ourselves accord~
ing]y. Even further, the system of scalars can be taken to be a.n arbitrary ring. In
this case!, one spea.ks of a maduh in~ten.d of & lilleat space. Modules are of great
importance in the st ru ctu re theory of rings.

192

Operators

The set R of all real num hers~ "'. i th ordinary addition and
multiplication taken as the linear operations, is a real linear space ..
Ex.amp( e 1

Example 2. 'TJ1c ~et R"' of all n-t.uples of real numbers is a real Jinen.r
space under the f ollo,\ring eoordina.tewisc linear operations: if
and

then

and

= (xl + Yi, X2 + Y2 1 ~
ax = (ax1, ax2, ~ .
ax&).

+Y

Xn

+ Yn)

This reduces to Example l 'vhe-n n : : ; ; l.

Example 3. The set e(Xi/l) of all bounded continuous real functions
defined on a topological space X is a real linear space under the f ollo,ving
pointwise 1inear operations: if f and g are functions in e(X,R)~ then
f + g and af are defined by

Cf+ g)(x) = f(x)

(aj) (x) = af(x)r

and
Example

+ g(x)

..t. The ~ct (} of all complex

numbers is a complex linear space

under ordinary addition and 1nultiplica.tion.

The set Cn of all n-tuplcs of complex numbers is a complex
linear space "Tith respect t-o the coordinatewisc linear operations defined
in Example 2. This reduces to Example 4 when n == I.
Example 5.

6.

rfhe set e(X,C') of all bounded continuous complex functions defined on a topological space X is a complex linear space with
respect to the pointwise linear operations defined in Example 3.

Example

Let P be the set of all polynomials, with real coefficients,

defined on the closed unit interval [O~ll~ We specifically include all
non-zero constant polynomials (which have degree 0) and the polynomial
which is identically zero (this has no degree at all)~ If the linear operations are taken to be the usual addition of two polynomials and the
multiplication of a polynomial by a real number, then P is a real linear
Example 7.

space~

Example 8.. For a given positive integer nt let P n be the subset of P

consisting of the polynomial which is identically zero and all polynomials
of degree less than n. P ~is a real Iinca.r space with rc~pect to the linear

operations defined in P.
Exompre 9. A linear space may consist solely of the vector O~ with
scalar multiplication defined by a o = o for all ar We refer to this as
the zero space 1 and we aJways denote it by {0 J.

Argebraic Systems

193

These examples are typical of the spaces which will concern us and
give ample scope for the illustration of all the important phenomena.
There are a number of other linear spaces of great interest, and we mention some of these from time to time in later chapters+ For the present,
however~ the above list will suffice.
We saw in Sec+ 14 that in any linear space "?{e have a~ 0 = o~
0 . .x = o~ and ( -1).x = - XT It is also easy to sho"r that ax = 0 ~a = 0
or x = 0; for if a ~ 0, then multiplying both sides of ax
0 by a- 1
yields a- 1 (ax) = a- 1 o~ {a- 1o:)x =:::: o, lx === o~ and finallyt x = 0.
We now turn to the general theory oi an arbitrary linear space L~
A non-empty subset M of Lis called a Bubspace (or a linear subspace)
of L if Mis a linear space in its o'~n right with respect to the linear operations defined in L.. This is clearly equivalent to the condition that ;.l{
contain all sums, neg.a tivl1st and scalar mnltiples of its elem.en ts; and
since -x = (- l)xt this in turn is equivalent to the condition that M be
closed under addition and scalar multiplication. If the subspace Mis a
proper subset of L, then it is called a proper subspace of L+ The zero
space [0} and the full space L itself are always subspaces of L+ Among
our examples, P n is a subspace of P for each positive integer n, and
P is a subspace of e[O~l]~ Also, the following are easily seen to be sub ...
spaces of R1 :
;:::=

and

= i (.x 1,

0, 0)}, M "i
M ~ ~ ~ (0, X2, xa)}, M 6

Mi

== i (0,
=

{ (x1~

x2~ 0)} 1

Ma

0, Xa) L, M 6 ==

{ (0, 0, xi) l ~
{(x1, X2~ O)}.

The subspaces Mi, Mi, Ml are usually called the coordinate axes in solid
analytic geometry, and M ,1 M 6J M 6 are called the coordinate planes~
The most general non-zero proper subspace of R' is a line or a plane
through the origin.
If M is a sube;pace of L, then-just as in the case of an ideal in a
ring~wc can use M to define an equivalence relation in L as follows~
x ~ y (mod . ~)
. means that x - y is in AI. The discussion leading up to
Theorem 41-B can be repeated \vithout essential change (but with considerahle simplification) to yield the concept of the quotient spac.e L/ M of
L v.,..ith respect to M. We give a formal statement of the basic facts in
the following theoremL
Theorem A. Lel M be a subspace of a linear .Bpace L~ and let the coset of
an element x in I,J be defined by x
M = {x
m:m e, M}. Then the
distinct cosets form a partition of L; and if addition and scalar multiplication are defined by

(x

and

+ M) + (y + M)
a(x + M)

(x + y)
= ax + M,

+M

194

Operators

then these coset8 cunstituu a linear space denoted by L/ M and called the
quotient space of L with respect to M. The origin in L/ M is the coset
0 + M
M, and the negative of x + Af is (-x)
+ M.

;:;;:r;

The proof of this theorem is routine, and we leave the details to the
reader.
It is worth remarking that the concept of a quotient space has a
simple geometric intPrprcta.tion~ To bring this out most clearly, we let
L be the linear space R'}. and A-f the suhspacc indicated in Fig. 31. If we
(~+M)+(y+M)
=ii: (

Fig_ 3 I.

x+y) +M

Addition in a quotient

spae~.

think of the vectors in L as the heads of arro,vs ~Those tails are at the
origin, then the non-zero proper subspace .i1J is a ~tra.ight. Jine through the
origint a typical coset x + kl is a line parallel to M ~ and L/ "ftf consists of
all lines parallel to j,f We add tVo coscts x +Mandy + M by adding
x and y and hy forming the line (x
y)
M through the head of x
y
and paraHe1 to !f. Scalar multiplication is carried out similarly
The subspaces of our linear space L can he characterized conveniently
as follo\\'S. If l Xt, x~, . ~ . , Xn} is a finite non-empty set of vectors in
L~ then the vector
x : : : :; a1X i + a2X'1 +
a~x"
r

is called a linear combination of xi, x 2 J , Xs. It is evident that a

subspace of L is simply a non-empty subset of L \vhich is closed under
the formation of Hnea.r combinations~ If S is an arbitrary non-empty
subset of L, then th-r set of all linear combinations of vectors in S is
clearly a subspace of L; Vr""e denote this subspace by [SJ, and we call it.
the subspace spanned by S. Since [S] is a subspa.ce 'vhich contains ~sand
is contained in every subspace "'~hich contains S~ "\\Te may think of [S] as
the smaJlest subspace which contains S~ If Jf is a subspace of L, then
a non-empty subset S of M is said to span M if {S] ;; Al+
Suppose now that M and N are subspaces of L, and consider the set

A1gebraic Systems

19.S

M
N of all sums of the form x + y, where x e: M and y E N4 Since
M and N are subspaces, it is easy to see that M + N is the subspace
spanned by all vectors in !JJ and N together, i.e., that ill + N = {MU NJ.
If it happens that M
N ~ L~ then ~.ve say that L is the sum of the
subspaecs M and N. This means that each vector in L is expressible as
the sum of a vector in Al and a vector in N 'The case in v-""hich even
more is truc--namely~ that each vector ~ in L is expressible un-iquely in
the form z ::=: x + y, with x M and y f N-will be of particular importance for us. In this rase V{C say that Lis the direct sum of the subspaces
Af and N _, and we syn1b0Jizc this statement by v{riting L = ltf N ~

Theorem B. Let a linear space L be the sum of two subspaces M and N,

so that L = M + N. Then L ~ M E9 N <=> M f\ N = {0} .

We begin by assuming that L = Jf E9 N, and we deduce a

contradiction from the further assumption that there is a non-zero
vector z in .J.\1
N ~ It suffices t-0 observe that z is exprrssi ble in t'vo
different \vays as the sum of a vector x in .Jf and a vector y in N, for
z = z + 0 (here x ~ z and y = O) and z = 0 + 2 (here x = 0 and y = z).
This contradicts the uniqueness required by the assumption that L =
Jf e Na
We no\\~ assume that Af n N = {O}, and we shovl that it follows
from this that L = 21! + N can be strengthened to L = llf EB N.
Sinr.e I-1 = "'~1 + N., each z in[.; can be written in the form z = x + y v.Tith
x e: AI and ye N~ \V c wish to shov.,. th.at this decompo~ition is unique.
If "re have t~~o sueh decompositions of 2, 80 that z = x1 + y 1 = x2 + Y2,
then x1 - x2 = Y?. - Y1 The left side of this i~ in M, the right side i~
in N, and they are Pqual; it therefore follov.Ts from Mn. N = {O l that
both sides are o~ that X1 == r2 and Y1 ~ Y2~ and that the decomposition
of z is unique+
PROOF~

The condition in this theorem-that the subspaces Al and N have

only the origin jn common-is often exprPssed by 8aying that. Jf and N
are disjoint There is fortunately little danger of confusing thi~ \vith the
set-theoretical notion of disjointness, for a subspace of a linear space
allays contains the vector 0, so the intcrscr.t.ion of any two must also
contain this vector and they can never be disjoint in the set-theoretical
sense~

The concept of a direct sum can easily he broadened to allow for three
or more subspaceEL If M 1, M -i, ~ ~ . , M ~ (n > 2) are subspaces of L,
then the statement that Lis the direct :ium of the Jf /S:-written
L=M1Ea1f!EB
~means

~.

Mn

that each vector z in L can be represented uniquely in the form

z = x1 + X2 + + x"., where Xie Mt: for every i. The reader will

196

Operators

observe that R 1 can be represented in various ways as direct sums of the

coordinate axes and coordinate planes m~ntioned above:

R3

:;:

M1

e Mi m Ma=

M1 EB M,

= M2 EB M.5 = Ma

fl} MB~

We shall often hav~ or.casion in the follovfing chapters to study problems

which are intimately concerned with the representation. of a linear space
as the direct sum of certain of its subspaces.
Problems
1..

2.,

3.

Each of the following conditions determines a subset of the real

linear space R 3 of all triples x ;: : : (x 1, x2, x:1) of real numbers: (a) x1 is
an integer; (b) x 1 = 0 or x 2 = 0; (c) x 1 + 2x2 = 0 ; (d) xi + 2x ~ = 1.
'Wb.ich of these subsets are subspaees of /l 3?
Each of the following conditions determines a subset of the real
linear space e[ - l~IJ of all hounded continuous real functions y = f(x)
defined on {~ 1~11: (a) f is differentiable; (b) f is a polynomial of
degree 3; (c) f IB an even function~ in the sense that f( ~x) == f(x) for
all x; (d) f is an odd function, in the sense that f(-x) = -f(x) for
all x; (e) f(O) = 0; (f) f(O) = 1; (g) f(x) > 0 for all x. Which of
these subsets are subspaces of e(-1 1 1]?
Jn the preceding problem, show that e[ --1 ~11 is the direct sum of the
subspaces defined by conditions (c) and (d). (Hint: observe that
f(x) = If(x) + f{ -x)]/2
[f(x) - f(-x)]/2.)
Let a linear space l.i be the sum of certain subspaces 1'J 11 M 2, .
M tL (n > 2) ~ and show that L is the direct sum of these subspaces
~ each Mi is disjoint from the subspace spanned by .all the othet!:li.
The latter condition clearly implies that each .Lllr, is disjoint from each
of the others. Show that the converse of this statement is false by
exhibiting three subspaces i.W1~ M 2, M8 of R 1 such that

4.

n A-fj =Mi(') bf.,

(M2 + M;j) ~ {01.

M1

and M1 fl

== M2

n M3

== {O}

43.. THE DIMENSJON Of A LINEAR SPACE

Let L be a linear space, and let S = {x1~ x2, . . ~ , Xn} be a finite
non-empty set of vectors in La Sis said to be linearly dependent if there
exist scalars ai, a~, . . . , a,0 not all of which are 0, such that
(1)

lf

s is not linearly dependent,

then it is caned linearly independen(; and

Al gebroic Systems

197

this clc.arly means that if Eq. (I) holds for certain scalar coefficients
cq, 02, . . , ani then a11 these scalars are necessarily 0. In other
words, S is linearly independent if the trivial linear combination of its
vectors (vrith all scalar coefficients equal to 0) is the only one ,vhich equals
0 ~ and it is linearly de.pendent if some non-trivial linear com bina t.ion of its
ve(~tors equals 0. In either cu.se, as 'VP know, the vectors in the subspace
lSJ spanned by S are precisely the linear combinations
(2)

of the x/s. 1.,hc significance of the linear independence of S rests on the

f aet that if S is linearly independent, then each vector x in lS] is uniquely
expressible in this form; for if "'Te also have
X

= f31X1 + f12x~ +

(3)

then subtracting (3) from (2) yields

(a1 - tj1)X1

(a~

- /3~)x2

+ (Un

t'~) Xn ~ o~

from l\hich-by the linear independence of S-,ve ol~tain ai

/3i = 0 or
a.~ = /3i for every i. l~,urther, the linear independence of i..~ not only
implies this uniqueness~ but is a1so implied by it, for the statement that
the vector 0 in [S] is uniquely expressible in the form
m

XI

+0

X2

+0~

Xn

is exactly what is meant by the linear independence of Sa

It is necessary to extend these concepts to cover the case of an
arbitrary non-empty set of vectors in L. We shall say that such a set is
ltnearly independent if every finite non-empty subset is lineu.rly independent in the sense of the above paragraph; othernTise~ it is said to be
linearly dependent. Just as in the finite case, an arbitrary non-empty
subset S of L is linearly indepe11dent ~ each vector in the subspace {SJ
spanned by S is uniquely expressible as a linear com bi nation of the vectors
in S. We are particularly interested in linearly independent sets which
span the "'Thole 8pacc L+ Sur..h a set is called a basis for L. It is impor...
tant to observe that if S is a linearly independent subset. of JJ, then S is a
basis for L (::::}it is niaximal v-.Tith respect to being linearly independent, in
the sense that every subset of l.1 ,vhich properly contains S is linearly
dependent.
Our first theorem assures us that if a linearly independent set is not
already a basis, then it can aJ,vays be enlarged to form a hasis.
If S
L, then there exists
PROOF. Consider
which contain S.

Theorem A.

i"s a linearly independent set of vectors in a linear space

a basis B for L such that S ~ B.
the class P of all linearly independent subsets of L
P is clearly a partially ordered set ,\~ith respect to set

198

Operators

inclusion~

It suffices to sho\\~ that P contains a maximal set B 1 for such a

maxima1 set will automatically be a basis for L such that S s; B. By
Zorn's lemma, it suffices to sho\v that every chain in P has an upper
bound in P.. But this is evident from the fact that the union of all the
sets in any chain of linearly independent sets which contain S is itself a
linearly independent set 'vhich contains S~
A linearly independent set is non~empty by definition, and it clearly
cannot contain the vector 0. We sec from this that if our linear space
Lis the .zero space {OJ, then no subset off..; is linearly independent and L
has no basjs+ On the other hand, if L # ~ 0} and x is a non-zero vector
in L, then the single-element set {x l
is linearly independent and Theorem
.4... guarantees that L has a basis \\;hi{~h
contains [x} . This proves

Theorem B.
h.a."J a basis.

Etery non-zero linear

.~pace

Since any single-clement set consi8ting of a non-zero ver.tor can he

enlarged to form a basis, it i8 evident
that any given non-zero linear 8pace
has a great many different bases+
Figa 32+ Two ba~c:. khe:d an<l
In R 2 ~ for instance~ the vectors
{/1,/:d for R2.
e1 = (1,0) and ei = (O~ 1) form a hasis,.
as do fl = ( l , 1) and f z == (o~ - 1)
(sec :Fig. 32)~ If we think of a ,..-cctor as an arro'v v.Thosc tail is the origin~
it is fairly clear on geometrical grounds that in this space any two nonzero vectors form a basis if they are not collinear. We bring order out of
this apparent chaos by proving in sevcra] stages that any two bases in a
non-.zero linear space have the same number of elements. ()ur next
theorem is the first step in this process.
Theorem C. Let ~') = {.r h x2~ . . . , x~ J be a finite non-e1npty set of vectors in a linear space L. If n = 1 ~ then Sis linearly dependent "7 x1 = O+
If n > 1 and x 1 #- 0, then S is linearly de pendent {:=:? .some one of the vectors
x2, . . . ~ Xn ia a linear conibination of the fectors in S whU:h precede it.
PROOF. The first statement is obvious, so \\e assume that n > 1 and
that xi - 0. It is easy to sec th.at if one of the vectors x~, . . . ~ Xn is a
linear ~ombination of the preceding ones, then the equation expressing
this fact can be re\vritten in the form of Eq. (1) in such a \Vay that the
coefficient of the vcct<>r in question is 1, so Sis linearly dependent.. We
now assume that S is linearly dependent, so that Eq. (I) holds with at
least one non-zero coefficient. If a-.: is the last non-zero coefficient, then

Algebraic Systems

199

>

1 (since x 1 '#- O) and Eq. (I) can be rewritten in such a way that Z:i is
exhibited as a linear combination of xh ~ . . , Xt-1 (with coefficients

-a1/a1, ~ . ~ , - CWi-J/a;)~

We next prove the following restricted form of our main theoremr

Let L be a non-zero linear space+ If L has a finite basis

B 1 = {ed = ~ eit ei, . . . , en l 'UJith n elements,. then any other basis
B-i = (/; l is also finite and also ha.s n elements.
PROOF.
To show that B2 is finite, vle assume that it is not, and we deduce
a contradiction from this assumption. We first observe that each Ci is a
linear combination of certain f/s and that all the //s v-.Thich occur in this
"Tay ronstitute a finite subset S of B-iT Since B2 is assumed to be infinite,
there exists a vector fj in Bi which is not in S. But fio is a linear combination of the els~ and therefore of the vectors in S. This 8hows that
~S U IfifJ ~ is a linearly dependent subset of B ~, Vt-~hich contradicts the fa.ct
that B2 is a basis.
Since the basis B'J. is fiuitP., it can be written in the form

Theorem D..

B 'J

f I = I f h f 'l,

{ 1

for some positive integer m. We must now shov.T that m and n are equal,
nnd this ~Tc do as follo\VSr Since the e/s span J,J, f 1 i~a1inear combination
of the e/s, and the ~et sl ;:::= { f b e !) e2' ~ ~ en l is 1incarly dependent.
'Ve know by Theorem(~ that one of the e/s, say ei:~~ is a linear combination
of the vectors in S1 'Which precede it. If '"'Tc delete e1 from 81,, then the
remaining set Ssi. = If 1 ~ e 1~ . . . , e-i~-1, e,o+ 1~ . ~ . ~ e"} still spans 4
Just as before~ f?. is a linear combination of the vectors in '"-9'-i~ so the set
Si = {f 1, f 2~ ei, ~ . . , e.-(1-1, ei~+ 1~ . ~ , en} is linearly dependent.
Another application of Theorem C shows that some vector in 8 3 is a
linear combination of the preceding ones; and since the // s ate linearly
independent~ this vector must be one of the e/s. If 've delete this vector,
then the remaining set again spans L. If we continue in this way, it is
clear that 've cannot run out of e/s before the f/s are exhausted; for if we
do, then the r~maining f/s are linear combinations of those already used,
which contradicts the linear independence of the //s~ This shows that
n is not less than m, or equivalently, that m < n. If we reverse the
ro1es of the e/s and f/s, then precisely the same reasoning yields n 5 m~
from which we conclude that m = n~
r

We are now in a position to prove our main theorem in its full

generality~

Theorem Et LetLbeanon . . zerolinearspace. If B1 = {ed andB'J = {/;)

are any two bases for L, then B 1 and B 2 have the sanu number of elements

(that is, the .same cardinal

number)~

200

Operators

If either B1 or B1 is finite, then by the preceding theorem t.he

other is also finite~ and they have the same number of elements. We may
t.herefore confine our attention to the case in which both are infinite.
Since B.,, is a basis, eaeh ei can be expressed uniquely as a linear
combination (with non-zero coefficients) of certain f/s:
PROOF.

-Bi

+ a2fit +

a.1fi1

+ aR/i..

Fu rt.her r every f; occurs in at least one such expression, for if a certain o net
say fj(j, does not, then sine e B 1 is a basis, L. is a linear combination of cer
tain e/s, and therefore of certain f/s ~ / 1.-which contradicts the fact
that the // s are 1ine.arly independent. This process associates with
each ei a finite non-empty set F i1r: of //st and in such a way that B2 =
Uo:l1B 1 F~i
Let n1 and n2 be the cardinal numbers of B1 and Bt, and let
n be the cardinal number of the indicated union. It follows from the
above set equality that nt = n, and Problem 8-10 shows that n :::; ni,
so n2 s n1. If we reverse the roles of the e/s and fjJs, then in the same
manner we obtain n1 < n2, from which we conc]ude that n1 ~ n2.
These theorems enable us to define the dimension of an arbitrary
linear space L~ If L = {O} then it is said to be 0--dimensional, or to
have dimensian 0; and if L . {0 L then its di1nension is the number ol
elements in any basis. A linear space is called finite-dim-ensonal if its
dimension is 0 or a positive integer, and i'nfi.nt"te-dimensionaI.. otherwise+
We can now justify the usual practice of calling R" and Cn n-dimensional
spaces by exhibiting the following n vectors as a bas.is for both spaces:
j

6t

= (1,.

es

~ (O~

o, o, . . , 0)

1, O, . . . , 0),
(O, O, 1, . .
0),

=
t

ei

e. = (0, 0, 0,

1).

It is easy to see that P.. is also n-dimensional, for the polynomials 1, x~

x'' ~ ~ J x~- 1 co nsti tu te a basis f0 r this spacer Similarly, the set
{1, x, x1 t ~ x"J . . . J is a basis for PJ so this space is infinite-dimensional. More precisely, the dimension of Pis ND.
The existence of a be.sis for an arbitrary non-zero Iin ear space, and
the fact that the number of elements in a basis is a constant determined
only by the spacet can also be used to give a simple but complete structure
theory for these spaces. We proceed as follows+
Let L and L' be linear spaces with the same system of scalarea An
isomorphism of L ontv L' is a one-to-one mapping f of L onto L' such
that f(x
y) = f(x)
f(y) and /(ax) = af(z); and if there exists such
an isomorphism, then L is said to be iaom.orphic to L'. To say that one
I

Algebraic Systems

201

linear space is isomorphic t.o another js to say~ in effect, that they are
abstractly identical with respect to their structure as linear spaces.
NO\\r Jet L be a non-zero finite-ditnensiona] linear space of dimension
n, and let. B = {e1 e2 , . . , e1l l be a basis for L \\'ho8e elemPnts are
written in a. definite order as indi,~ated by the subscripts. Each vector
x in L is uniquely expressibh.l in the form

so then-tuple (a1, a2~ . , . , on) of s_ca1ars is uniquely determined by x.

If we define a mapping f by /(x) = (a1i a~, .
an), then it is easy to
see that f is an isomorphism of L onto Il~ or Cfl according as L was real
or complex to begin "',..ith. It should be recognized that the isomorphism
f is by no means unique, for if some other basis is chosen for l.J" or if the
order of the elements in the basis B is altered, then the resulting isomorphism of L onto Ji"- or Cn will clear1y. be diffcren t from /T \Ve summarize these remarks in
T

Theorem F. Let L be a non-zero finite-dimensional linear .~pace of ditrten,.-..

S'ion n. If L is real, then it is isomorphic to R~ j and if it is coJnplex, then
it is isomorph i'c to Cm
T

This theorem can easily be extended to the case of an .arl)itrary nonzero Jinear space~ We begin by deserihing the concrete JinPar spaces
which wiH replace Rn and Cn in our ge11eralized for1n of Theorem F.
Let X be an arbitrary non~cmpty sct 1 and de-note by J...(X) the set of a.11
scalar-valued functions defined on X which vnnish oulsidr finite sets~
Addition and scalar multiplication for such functions a.re understood to
be defined point.wise, and T-'(X) is obvioufdy a non~zrro linear Rpacc which
is real or complex according 38 the functions con~idered a.re real or complex~
Our purpose is to shovl that these spaces are universal model~ for
non-~ero linear spaces, in the sense that an arbitrary non-zero Jineat
space L is isomorphic to some L(X). We start by choosing a basis
B = i ed for Lr and "\ve let B be the set Xr We next establish an isomorphism of f.; onto l.i(B) by making correspond to each vector x in L a
scalar-valued function. f~ defined on B. If x = 0, then f'J; is defined by
f :t!(ei) = 0 for every ei in B If x r=: O, it is uniquely expressible in the
4

form

with non-zero coefficients; and f~ is defined by fx(e1) ~ 0 outside the set

{ ei1' e.:
i
e.:A J and by f~(eii) = ai inside this set. It is trivial to
verify that the mapping we have described is an isomorphism of L onto
L (B). This discussion yields
1,

202

Operators

Theorem G. Let L be a non-zero Ii-near space. If Bis a basis for L, then

Lis i8omorphic to the linear space L(B} of all scalar-valued functions defined

on B which vanish outside finite sets+

Theorems 1 1~ and G arc of considerable interest in that they reveal
~~hat simple things linear space~ rpally arP.
The reader may \Veil f ecl, in
t..he light of t.h~se results, that the ~onccpt of an abstract linear spa.ce has
scrv~d its purpose a.nd should no\v be abandoned, and that all further
study of linear spaces should be directed specifieally at the L(X)'s, or
in thP. finite~dimensional case~ at Rtt. and Cft. There are at lea8t two
reasons v.r""hy this is not a useful point of vic\v ~ One of thPsc lies ifl the
fact thnt the ahove isomorphisms 'verc established by arbitrarily choosing
one particular basis B for Lin preference to all the others, whereas most
of the important ideas in the theory of Jinear spaces are inde~ndent of
any specially chosen basis and are best trcatedJ v.~hen this is pos8iblej
\vithout ref erP.nce to any basis vlhat.ever A second reason is that almost
all the linear spaces of greatest interest carry additional algebraic or
topologirn.l :structure, which need not be related in any significant manner
to the above isomorphisms.
+

Problems
l.

2.

3.

Let L he a non-z:ero finite-dimensional linear space of dimension n.

Shovir that every set of n + 1 vectors in L is linearly dcpendipnt.
Sho\V that a Ret of n vectors in L is a basis (..:..} it is lineu.r]y independent .q it spans L.
Show that the vectors (1, O~ O), {1 1 ~ 0), (I, 1, 1) form a basis for n~ 3
ShoVir that if ~ c1, e2, ea l h:; a basis for Ii 3, then {ei + c2.' e1 + eJ, ez + el l
i

is also a b.asig ~
Let ill he a subspace of a linear space I-', and show that there Pxists a
subspace N such that L = ~M" ffi N. Give an example for the case
in which L = R'l. to show that N need not be uniquely determined by

M4
4.

5.

If A! and N are subspaces of a lin car space l.1 ~ and if I.J :::; ill ffi NJ
sho'" that the mapping y ~ y + ~f which sends each y in N to
y +Min L/1'1 is an isomorphism of 1V onto L/lf.
Denote the dimension of a linear space L by d(l-1)~ If L is finitedimensiunal, and if M and N are subspaces of I..1, prove the fallowing:
(a) d(ftf) < d(L)t and d(M) = d(L) ~ !f = L;
{b) d(M) + d(N) = d(Al
N)
d(M f\ N);
(c) if L = M + N~ then L = J1 ffi N => d{L) = d(M) + d(N);
(d) d(L/ M) = d(L) - d(M)~
If L and L' are linear spaces, show that L is isomorphic to L' ~ they
have the same scalars and the same dimension~

6..

Algebraic Systems

203

44. LINEAR TRANSFORMATIONS

Let L and L' be linear spaces with the same system of scalarsr
ma.pping T of L into L' is r.alled a linear transfornlation if
T(x

+ y)

+ T(y)

= 7,(x)

or eq uivalcntly, if
'l~(ax

+ {3y)

and

:;::: aT(x}

T(ax)

aT(x),

+ {3T(y)~

A linear transformation of one linear space into another is thus a homomorphi~1n of lhe first spare into the secondt for it is a mapping which
preservE's the linear operations. 7' also preserves the origin and negatives, for '.f(O) = T(O O) ~ 0 7'(0) ~ 0 and
+

T(-x)

T({-J)x)

(-l)T(x)

-T{x)T

The importance of linear spaces lies mainly in the linear f..ransforma-tions thry carry, for vast tracts of algebra and analysis, ,vhen placed in
their proper contiext~ reduce to the study oi linear tran~f or ma tions of one
linear space into another. The theory of matrices, for .instance, is one
small corner of this subject~ as are the theory of certain types of differentia1 and integral equations and the theory of int~gration in its most elegant modern form.
ln the f ollo,ving examples \Ve leave l t to the reader to shOl that eaeh
mapping dP.srribed actually is a linear transformation.

Example 1. ,V. e co nsidrr the linear space R~ t and each Ji near t.ran.sf orma ti on mentioned is a. mapping of ll 2 into itself.
(a) T1((x1~x2)) = (ax1iax2), 'vhere a is a real number. The efir.ct.
of T 1 is to multiply each vector in R 2 hy the scalar a.
(b) 1\i( (x1,x2))
(xz,x1). T2 reflects ll 2 about thP. diagonal Jlnip,
X1

X2.

(c)
(d)

1,a( (x1,x2))
T t ( (x1,x2))

T'(. projects R 2 onto the x 1 axis~

T 4 projects R 2 onto the x 2 axis.

Ex amp Ie 2. Consider the linr.ar space P of a.U polynomial.s p(x) ~ '\\Tith

real coefficients, defined on fOi l J. ~rhe mapping D defined hy
D( ) - dp
p - dx

is

ch~ar!y

a linr.a.r transf ormat.ion of P in to itselL

Example 3.

The rnar>ping l defined by

I (f) = /

1
1

f(x) dx

i8 easily seen to be a linear transformation of e[O, 1] into the real linear

space R of a.H real numbers.

2a.

Operators

We return to our consideration of the linear spaces Land L 1 and of

the linear transformations of L into L'. If T and U are two such
transformations~ then they can be added in a natural way to yield T + U,
which is defined by
(T

+ U) (x)

.:=

T(x)

+ U(x)~

(I)

Similarly, any such transformation T can be mu1t.i plied by any sr.alar

in accordance with .
(aT) (x)

aT(x)~

a~

(2)

Simple computations show readily that T

U and aT are them.selves
linear transformations of ].; into L 1 , and it is easily provr.d that these
definitions convert the set of all such linear transformations into a linear
space.. The zero transformation 0 (i.e., the zero clement of this linear
~pace) and the negative -T of a transformation 1 are defined by
O(x) = 0 a.nd (-T) (x) = -T(x). In summary, we have
1

Theorem A. Let L and [;, be two linear S'paces with the sam.e system of
.~cal.ars.
7 hen the set of aU linear transformatians of L into I/ is itself a
linear space u;ilh respect to the linear operations defined by b'q8L (I) and (2).
1

The most interesting and significant applir.ations of these ideas occur

in the special cases in "~hich (1) /./ equals L., and (2) Lr equals the linear
space of a11 scalars of L. We now develop a fe,v of the simpler concept8
which a.rise in case (I). Case (2) will be treated in some detail in the

next

chapter~

\Ve assume, then~ that V{e have a single linear space L, and we consider the Ji near space of all linear transformn tions of I~ in to itself. W P
usually speak of these as linear transformations on I..J. The most important feature of this situation is that if T and U are any t.wo linear transformations on Lt then we r.an dPfine their prnduct TCI by means of
(TU) (x)

T( Cl(x)).

(3)

This is precisely the mu1tiplication of mappings discussed at the end

of Sec. 3~ and Problem 3-1 assures us that this opera.tion is associative:
T(UV)

(TU)V.

(4)

Furthermore, multiplication is related to addition by the distributive

Jaws

and

+
+

T( U
V)
(T
U) V

= TU + TV
= TV+ UV,

(5)
(6)

and to scalar multiplication by

a(TU)

(aT) U

T(aU).

(7)

Algebraic Systems

The proofs of these facts are

the following computation;
{(T

easy~

205

As an illustration. we prove (6) by

+ U)V](x) = (T + U)(V(x))
T(V(x)) + U(l'(x))
= (TV)(x} + (UV)(x}
~ (TV + UV)(x).
;;=

Examples can readily be found to show that multiplication is in general

non-commute..tive~ For instancet if we define a linear transformation M
on the space P of polynomials p{x) by M(p) = xp~ then
(MD)(p) = M(D(p))

and

= xD(p) = x dp
dx

(DM)(p) = D(M(p)) = D(xp)

= x dp + p,
dx

so MD ~ DM. Also, it is quite possible for the product of two non-zero

linear transformations to be 0. Examples le and Id demonstrate this,
for the transformations TJ and T~ are both different from OJ and yet
TJT . . = 0.
We have so far seen only one specific linear transformation on the
arbitrary linear space LJ namely, the zero transformation O~ Another
is the identity transformation I, defined by l(x) = x. We observe that
I :. 0 ;;> L ~ t0}, and that

7'!

(8)

IT = T

for every linear transformation T on L. If a is any scalar, then the

linear transformation a.I is called a scalar multiplication, for
(al)(x)

al(x)

===

ax

shows that the effect of al is to multiply each vector in L by a.

A linear transformation 'J' on Lis called non-singular if it is one~to
one and onto, and singular othen,~se. If T is non-singular, then by
Sec~ :3 its inverBe T-l exists 8,S a mapping and satisfies the following
equation:
(9)

It is not difficu;I t to sbo w that when '1 is non-singular, then the mapping
T- 1 is also a linear transformation on L~
A particularly important type of linear transformation on L arises
as follows. Let TJ be the direct sum of the subspaces Al and N~ so that
L = M El) N. This means, of course, that each vector z in L can be
written uniquely in the form z = x
y ,vitb x in M and yin N. Since
x is uniquely determined by z, Vt'"e can define a mapping E of L into itself
1

206

Operators

by E(z)
z~ E is easily seen to be a linear t.ransforrnation on L, and it
is called the projection on ft! along N. I~""igure 33 indicates the geometric
reason for this terminology. The most significant property of Eis that
it is ide1npotent, in t.he sense that E 2 = I!); for since x = x + 0 is the
representation of x as the sum of a
R2
vector in JI and a vector in N, 1\... c
have
-;:::=

Et{z)

(EE) (z)

n~(n~(z))

~ n~(x)

= x

E(z).

This property of idempotenee is char ...

actcristic of projections, as our next
theorem sho,\s.

}4,ig. 33. ThP projet ions E on .i.\J

along l'fT and 1 ~ R on JV along Jf.

Theorem B. If E is a linear transfor...

1nation on a linear space L, then b' is
idcrnpotcnt R there erisl subspaces 111
and 1V of TJ such ihal L = ill ffi N
and /~T is the projection on Jl along J.V.

In vie'"' of the above remarks, it suffices to sho\v that if E is

idempotent, then it is the projection on Jf along "A7 for suitable JI and lV.
We define Al and Jrrl by i.1/ = !E (z) : z e L} and 1V ~ f z : E (z) = 0 l .
Both arc clearly subspaces, and "Tc must show that L === Jf ffi l'v. By
Theorem 42~13_, it. suffices t.o sho1'. . that Jf and Jtl span L a.nd are disjoint.
That M and N span L follo"~s from the fact that each z in L can be
written in the form
(10)
z = R (z)
(I - E) (z) ;
PROOF.

for E (z) is obviously in ill, and

E((l - E}(z)). = (E(l - E)) (z)

(l!J - _bt2) (z)

(I~'

- I!}) (z)
=

O(z)

shows that (I - E)(z) is in N~ To see that M and N are disjoint 1 v,"e

have only to notice that if a vector E(z) in ]J.f is also in N, ~o that
E(E(z)) = o~ then E(E(z)) ~ E 2 (z) == E(z) shO\VS that E(z) = 0. 'fhis
proves that TJ = J-/ EB N, and it follo,vs from Eq. (10) that Eis precisely
the projection on Art aJong N.
The unsymmctric vlay in which M and N are treated in this discus . .
sion can easily be balanced by considering the mapping \vhich makes
correspond to each z ~ x
y the vect-0r y (instead of x) ~ "fhh~ linear
transformation (it is clearly I - E) is the projection on JV aJong llf
ln
the light of our theorem, we define a projection on L to be an idempotent

A! gab ra i e Systems

linear transformation on 4
the equation

'2<J'1

If E is a linear transformation on L, then

(I - E) 1 = (I - E)(I - E) =I - E - E

+E

shows that Eis a projection =>I - Eis a projection; and ~,.e know that
ii E is the projection on M along NJ then I - Eis the projection on N
.along M~ and conversely. We make one further comment on these
matters: if M is a given subspace of L, then by Problem 43-3 there cer...
tainly exists a projection on Jl.f and along some N; but since there may
be many different subspaces N such that L = l\f E9 Ni there may also
be many different projections on JI (and along various N's).
Problems

1.
2.

3.

A.

5.

6.
7~

8.

9.

10.

Show that the mappings defined by Eqs. (1) to (3) are linear
transformations+
Sho\v that the linear transformations T2 and T.a defined in Exam.ple 1
do not commute; that is, shovl that T2T3 ~ TaT2.
If D and J..f are the linear transformations on the space P defined
in the wxt, show that Dllf = MD +I and (MD) 2 = 1lf 2D 2 + 1\lD.
Let T be a linear transformation on a linear space ~and show that T
is non-singular{::::} there exists a. linear transformation T' on L such
that TT' = T'T == I.
Let T be a linear transformation on a linear space L 1 and prove
that T is non-singular # T (B) is a basis for l.1 whenever B is~
Prove that a linear transformation on a finite-dimensional linear
space is non~singular (::::}it is one~to--0ne ::::>it is onto4
Show that the set of all non-singular linear transformations on a
linear space L i" a group with respect to multiplication. If L is
finite-dimensional with dimension n > o~ this group is called the
full linear group of degree n.
If L and L 1 arc non-zero linear spaces (both real or both complex),
prove that there exists a non-zero linear transformation of L into L'.
Let I"' be a linear space, and let x and y be vectors in L such that
x r== 0.. Prove that there exists a linear transformation T on L such
that T(x) = y. If y is not a scalar multiple of x, prove that there
exists a linear transformation 'Jl' on L such that T 1 (x) === 0 and
T'(y) ~ 0.
Let Land L' be linear spaces with the same scalars, and let T be a
linear transformation of L into L 1 The null space of T, namely1
{x: T(x) = 0 L and its range, { T(x) :x e L}, are clearly subspaces
of L and L'. The nullity of T, denoted by n(T), is the dimension

208

Operators

of its null space, and its rank r(T) is the dimension of its range. If
Lis finite-dimensional~ prove that n(1 + r('.f) == d(L)~
If E is a projection on a linear spare L, sho\v that its r:t.nge equals
the set of all vectors \vhich arc fixed under E; ie~, shO\\' that
(E (z) :z e L 1 = {z : .K (z) = z J .
1

11.

45. ALGEBRAS
A linear space A is tailed an algebra (see Sec. 20) if its vectors can
be multiplied in such a way that A is also a ring in which scalar multipli ...
cation is related to multiplication by the follo,ving property:

a(xy)

(ax)y

x(ay)~

The concept of an algebra is therefore a natural combh1ation of the con~

cepts of a linear space and a rjng. !Figure R4 illustrates the manner in
Groups
Additive Abe !ia n groups
R~ngs

Linear

Algebras

spaces

Fig.

!34~

The major ulgebrajc systems.

which the major algebraic systems defined in this chapter are related to
one anotherr
Since an algebra is a linear space 1 all the ideas devcl oped in Secs. 4 2
and 43 arc immediately app1icable. Some algebras arc real and some
are complex 1 .and every algebra has a well-defined dimension. Furthermore1 since an algebra is also a ring, it may be commutative or non-commutative, and may or may not have an identity; and if it does have
an identity, then we can speak of its rcgu1ar and singular e1ements. A
division algebra is an algebra 'vith identity which~ as a ring, is a division
ring~ A subalgebra of an algebra A is a non-empty subset Ao of .A. 'vhich
is an algebra in its Ol-Vn right with respect to the operations in A. This
condition evidently means that A 0 is closed under addition, scalar multiplication, and multipli<.:ation .

Algebraic Systems

209

Example 1
(a) The real linear space R of all re.al numbers (see Example 42-1) is a commutative real algebra \vith identity if multiplication is
defined in the ordinary way. 1;he reader vml observe that 8C alar
multiplication is indistinguishable from ring multiplication in this system.
(b) rfhe complex linear spuce (} of all complex numbers defined in
Example 42-4 is a commutative complex algebra with identity if multiplication is defined as usual. Again we see that scalar roul ti plication
and ring multiplication are the same.
to

Example 2. (a) The real linear space e(XJR) (see Example 42-3) is a
com mutative real algebra with identity if multiplication is defined
poin tu. ise.
(b) The complex linear space e(X~C) defined in Example 42-6 is a
commutative complex algebra with identity with respect to pointwise
multiplication.

Example 3.. Let L be a linear space~ We kno\v by Theorem 44-A that

the set of all linear transform.a tio ns on L is a linear space with respect
to the linear operations defined by Eqs. 44-(1) and 44-(2). rrhe disc ussion following this theorem can be summed up as follo\vs. If multiplication is defined by Eq. 44-(3)~ then this linear space is an algebra which
is real or complex according as Lis real or complex~ This algebra has an
identity (the identity transformation)-<==> L # {O} J and in general it i~
non-commutative and has non-zero divisors of zero+
An ideal 1 in an algebra A is a non-empty subset of A which is both a
subspace ~Then A is considP.red as a linear space and an ideal when A is
considered as a ring. . By 1'heorem..~ 42-A and 41-B, A/ I i~ both a linear
space and a ring. It is easy to see that A/ I is actuaHy an algebra, called
the quotient algebra of A "\i\rith respect to I. An ideal in A in our present
sense is sometimes c ailed an algebra ideal, as opposed to -..vb at we might
call a ring ideal, that is, an ideal in A when A is considered as a ring.
By our definitionJ an algebra ideal is a ring ideal v..-.hich is also a subspace.
In the cases of interest to us, the distinction bet ween these two types of
ideals disappears. fi,or if A has an identity 1, and if I is a ring ideal in A,
then the f aet that i e I ~ ai = a( 1i) = (al )i e I ior every scalar a
shol\rs that I is closed under scalar multi pl ica tio n J and is thcrcfore a.n
aJ ge bra ideal.
We shall return to the subj cct of algebras and their idea1s in later

chapters+
Problems

1.

Let 1, be a non-singular linear transform.a tion on a linear space L,

and let A be the algebra of all linear transformations on L. Show

210

2.

3.

4~

Operators

that T is non-singular (i 4e 4, regular) as an element of the algebra

A~ L "# {O}~
If A is an algebraJ show that the subset of A defined by C = {x:xy
= yx for every ye. Al is a subalgebra of A. C is called the center
of A (see Problem 39-6).
Let A be an algebra of linear tranBforroations on a linear space L.
If A contains the id entity transforma t.ion, prove that the center of A
contains all scalar multiplications~ If A is the algebra of all linear
transformations on L, prove that the center of A consists precisely
of the scalar multiplications (I/int: see Problem 44~9).
Let A and A' be algebras which are both real or both complex. As
usual, we define a homomorphism of A into A' to be a mapping/ of A
into A' which preserves all the operations, in the sense th.atf(x
y)
== f(x) f(y),f(ax) = af(x), andf(xy) ~ f(x)f(y). An isomorphism
is a one-to-one homomorphism, and A is said to be isomorphic to A'
if there exists an isomorphism of A onto A'. Now let A be an
axbitrary algebra with identity~ and prove that the mapping f defined
on A by f(x) = M ;i;, where M J;(y) = xy, is an isomorphism of A into
the algebra of all linear transformations on A. This fa-ct is analogous
to Cayley's theorem (see Problem 39-11). The isomorphism f is
called the regular representation of A (by linear transformations on
itself).

CHAPTER NINE

Ha11aclt Spaces
We have already seen~ in Sec. 14, that a Banach space is a linear
space which is also, in a special way, a complete metric space. ..fhis
combination of algebraic and metric structures opens up the possibility
of studying Ji near transformations of one Banach space into another
v-.,..hich have the additional property of being continuous.
1\1ost of our work in this chapter centers around three fundamental
theorems relating to continuous linear transformations. The HahnBanach them-em guarantees that a Banach space is richly supplied with
continuous linear functionalst and ma.kes posHiblc an adequate theory of
conjugate spaces. The open mapping theorem enables us to give a satis
factory description of the projections on a Banach space, and has the
important closed graph theorem as one of its consequences. We use the
uniforni boundedness thertrem in our discussion of the conjugate of an
operator on a Banach space, and this in turn provides the setting for our
treatment in the next chapter of the adjoint of an operator on a Hilbert
space.
Virtually all this theory had its origins in analysis~ Our present
interest, however~ lies in the study of form and structure, not in exploring the many applications of these ideas to specific problemsL This
chapter is therefore strongly oriented t.oward the algebraic and topological aspects of the matters a.t hand~
21T

212
46~

Operators

THE DEFINITlON AND SOME EXAMPLES

We bgin by restating the definilion of a Banach space.

A normed linear space is a linear space N in which to each vector x
there corresponds a real number~ denoted by llxfl and called the norm
of x, in such a. manner that
(I) l~x \l > 0~ and ii x l = 0 {::::} x = 0;
(2) Hx + YH < llx~I + llY~I;
(3) Haxll = la! llxtl
The non-negative real number ~lxll is to be thought of as the length
~ of the vector x. If \\~c regard 11 x H as a real function defined on NJ this
function is called the norm on N. It is easy to verify that the normed
linenJ" space N is a metric space ~Ti th respect to the metric d defined by
d(x~y) = Hx - Yll ~ A Banach space is a complete normed linear space.
Our main interest in this chapter is in Banach spac.es, but there are
several points in the body of the theory at which it is convenient to have
the basic definitions and some of the simpler facts formulated in terms
of normed linear spaces. :For this reason, and also to emphasize the
role of completeness in theorems which require this assumption, we work
in the more general context v.--henevcr possible. The reader will find that
the deeper theorems, in which completeness hypotheses arc nec.essary,
of ten make esH-ential use of Haire' s theorem.
Several simple hut important facts about a normed linear space are
based on the follo"Ting inequality:
r

~lxll

- HY fl I ::; tlx -- Yll ~

(1)

To prove thisJ it suffices to prove that

l\xll - llYH < llx - Yll;

(2)

for it follo,vs from (2) that we also have

-(l]x~I

- 11YID

!IYll -1ixH < llY - xll - 11-(x -

y)[j ==

nx -

Yll,

which together 'with (2) yields (1). We now prove (2) by observing that
Hx!I = I (x ~ y) u!J < ~Ix - Yll + \lYl1- The main conclusion we draw
from (1) is that the norm is a continuous function~

z ~ Hx.. I~ ~ 11xU.
that I l x~ll - HxH t < Hxn

x" ~

This is clear from the fact

~ xH, since X11 "'""7 x
means that if xn ~ xii ---. O~ In the same vein, v.rc can prove that addition
and scalar multiplication are jointjy continuous (see Problem 22-5)J for

and

x~ ~ x and Yn
a,. ~ a and

-4

y ~ x"

Xn --t

+ Ya ~ x + y

x ====> a"X11 ._,;;. ax.

Banach Spaces

213

These assertions follow from

l[(xn

+ Y~)

- (z

+ y)I[

~ fl(x~ - x)

+ (Yn

- y)[]

5 flx11 - x[j

+ ~ly. - y!f

and
l!a~x~

axl[ ~

IJan(Xn

~ x)

(a~ ~ a)xH

< lanJ llxn - xfl + la" - al llxl!

Our first theorem exhibits one of the most useful ways of forming
new normed linear spaces out of old ones.
Theorem A. Let M be a ~losed linear subspace of a nor-med linear space N.
If the norm of a coset x
M in the quotient space N / }.{ is defined b-y

+
Hx + Ml1

=:::;

inf {llx

then NJ M is a normed Zin.ear sprue.

then so is N / M.

+ml! :m 5 ML

(3)

Further, if N is a Banach space,

We first verify that (3) defines a norm in the required sense.

It is obvious that l~x + Afll ~ 0; and since M is closed~ it is easy to see
that 11 x
M H = 0 <=> there exists a sequence {m'} in li-I such that
J[x + m.1:H--+ 0 =!P xis in Mt=> x
M == the zero clement of N/M.
Next1 we have H(x + M) + (y + M)H = ll(x + y) + ll!ll =inf {fix+
y+mlf~msMJ =inf ilJx+y+m+m'fl:mandm'eM} =inf {l[(x
+ m) + (y + m') I~ :m and m 1 e: M} < inf {fix + mil + UY m'll: m and
m' ~ MJ = inf rrlx + ml(~m E 1lfl +inf r~IY + m'll~m' e-M} = IJx
~!fl + HY + M II~ 1--he proof of i a(x + 111) ![ = lal Ux + Mii is similar4
Finally~ we .assume that N i8 complctP, and \Ve shov{ that N/M is
al.so complete~ If \'\re start 'vith a Cauchy sequence in N /Mt then by
Problem 12-2 it suffices to show that thi8 sequence has a convergent subsequP.nce. It is clearly possible to find a subsequence ix~ + Jf} of the
original Cauchy sequence such that 11 (x1 + lrf) - (x2 + M) H < 7~ 1
H(x~ + M) - (x3 + M) l < 7i, and, in general, [I (xn + M) - (x~+ i +
M)~I < 1/2tt. We prove that this sequence is convergent in N /M. We
begin by choosing any vector y 1 in x1 + bf, and wc select Y2 in x2 + M
such that 11111 - Y2ll < 7f. \Ve next select a vector u~ in X;J
M such
that rlY~ - y3fl < ~. Continuing in this "\vay1 've obtain a sequence
l Yn J in N such that I] Yn - Yn+I [I < 1/2n. If m < n, then
PRoOlt'\

+ }.{:;::::;

IJYm

Ynli == [I (ym

+ '"

_.. Ym+1)
(Ym+l - Jlm+2)
(Y~-1 - Yn) I~ ::; HYm ~ Ym+l ~!
llYm+l - Ym+2!f
llYn-1 - Ynl] < l/2m + 1/2m+l
~ ~ + 1/2n-l

< 1/2tn- 1

so {y,., I is a Cauchy sequence in N. Since N is completet there exists a

vector y in N such that Ya_,, y. It now follows from l~(Xn + M) (y + M)ll <HY" - Yll that Xn +kl-+ y + M, so N/M is complete~

21.4

Operators

In the following sections and chapters, we shall often have occasion

to consider the quotient space of a normed linear space with respect to a
closed linear subspace. In accordance with our theorem~ a quotient
space of this kind can always be regarded as a normed linear space in its
own right~
Vrlc no'v describe some of the main examples of Banach spaces+
Tn each of these, the linear operations are understood to be defined
either eoordinate~d.se or pointwise, whichever is appropriate in the
ci re u ffi8 t..a.n ces.
The spaces R and C-the real numbers and the complex
numbers-are the simplest of all normed linear spaces. The norm of a
number x is of course defined by llxll = JxL and each space is a Banach
space.
Example 1..

2.. 1.,he linear spaces Rn and CH, of all n-tuplcs x

(x1, x2,
~ . , X-n) of real and complex numbers can be made into normed linear
spaces in an infinite variety of ways, as we shall sec below. If the
norm is defined by
Example

:;;!!;

(4)

then we get the n-dimcnsional Euclidean and unitary spaces familiar to

us from our earlier work. We denoted these spaces by RB and G"of~ in
Part 1 of this bookJ and we know by the theorems of SccT 15 that both are
Banach

spaces~

Each of the following examples consists of n-tnples of scalarst

sequences of scalars, or scalar-valued functions defined on some nonempty sett ,vhere the scalars are the real numbers or the complex numbersT
We do not normally specify which system of scalars is to he used, and it
should be emphasized that both possibilities are allowed unless the
contrary is clearly stated. Also 1 \Ve make no distinction in notation
between the real case and the complex case. When it turns out to be
necessary to distinguish these two cases~ we do so vcrbal1y~ by referring,
for instanceJ to uthe complex space ~.'~ These conventions are in
accord with the standard usage preferred by most nmt..hcm.aticians~
and they enable us to avoid a good deal of cumbersome notation and
many unnecessary case distinctions.
Let p be a real num her such that 1 < p < co. "\Ve denote
by l; the space of all n-tuples x == (x 1 ~ x~~
~ xJ'l;) of scalars, with the
norm defined by
Examp1 e 3.

(5)

Banach Spaces

215

Formula (4) is obviously the special case of (5) which corresponds to

p z: 2~ so the real and complex spaces l; are then-dimensional Euclidean
and UD itary spaces Rft and c~. It is easy to see that (5) satisfies conditions (1) and (3) required by the definition of a norm. In Problem 4 we
outline a proof of the fact that (5) also satisfies condition (2) 1 that is,
that llx + y]!P < ]lxllP + HYHJ)+ The completeness oi l; follows from
substantially the same reasoning as that used in the proof of Theorem
15-A, so 1; is a Banach space .
Exomp,e 4. We again consider a real number p ~Tith the property that
l .$ p < co, and we denote by lp the spnce of all sequences
X =

of scalars such that 2: ; ... 1

{ X 1,. X :;i.,

!xn IP <

..

Xn' . }

"\\,.ith the norm defined by

(TJ ,

'IO

llxllP ~ (

n.=l

lxftl P) 11 P

(6)

The reader Vr"'ill observe that the real and complex spaces l2 arc precisely
tl1e infi.nite4limcnsional Euclidean and unitary spaces R~ and C' defined
in Problem 15-4r The proof of the fact that l,, actually is a Banach
space requires arguments similar w those used in Problems 15-3 and 15-4+
The Banach spaces discussed in these examples are all special cases of
the important LP spaces studied in the thco:ry of measure and integration.
A detailed treatment of these spaces is out.side the scope of this book, but
we can describe them loosely as fo11ows. An LP space essentially consists
of all measurable functions f defined on a measure space X with measure
m which arc such that. lf(x)I P is integrable, v..~ith

H!HP ==

(j lf(x)]

dm (x)) lip

(7)

taken as the norm. In order to include the spaces l; and lp VtTithin the
theory of LP spaces1 we have only to consider the sets [ 1, 2j ~ .. , n l
and I I~ 2., .... , n 1 i as measure spaces in 'vhich each point has
measure I, and to regard n ~tuples and sequences of scalars. as functions
dcfin ed on these sets+ Since integration is a generalized type of summation, formulas (5) and (6) are special cases of formula (7) . 1

Example 5. Just as in Example 3,. wc start "'Tith the linear space of all
n-tuples x = (x 1,, x 2 ~ J Xn) of scalar8, but this time we define the
Several remarks and example.s relating to LP spaces n.re scatt-Prr.d about in this
and the next chapter. T"his fragmentary material is not essential for an understanding of the.se chapters, and may Le disregarded by any reader without. the necessary
backgrouud. Brief ~ketches of the relevant idc&S can be found in Taylor [41 ~ chap. 7J
and T..001nis [27 chap. 3J. For more ext.ended treatments 1 !'Jee l!a.lmos [ 18], Za.anen
[45j, or Kolmogorov and Fomin [26i- vol. 21.
1

216

Operators

norm by
(8)

This Banach space i~ rommonly denoted by l~, and the symbol llxllco is
oreasiona1ly used for the norm given by (8). The reason for this practice
lies in the interesting fact that

!lxll~

:;;:

Hm

Hxll,,

that is, that

Jt

max

111\! J = lim (I lxif

11 )

IP

asp~

(9)

oo

i=I

We briefly inspect the case n = 2 to sec why this is true. Let x = (xh X2)
be an ordered pair of real numbers with x1 and x2 > 0. It is clear
that l!x ii~ = max ~ x1, X2} :::; (x1P + x2P) 11 P = llxlLp+ If x1 = X!b t.hen
lim llxl',, = lim (2x2P) 11P = lim 2 1 ' 11 x2 = X-:? = ]!xl[~. And if x1 < x2t then

Ii m 11 x lIp

I im (x 1,,

+x

2P) 1f P =

Jim ([(x 1Ix2)'"' + 1] x ~ii) 1J P

= Hm [(x1/x~)i> + 1]1 1Px2

x2

llx:!~

Consider the linear space of all bounded sequences x ;:::=- ~ x ~

x"2, ~ ... , x~, . . . J of scalars. By analogy with Example 5, -.,ye define
the norm by
Example

6..

~lxll == sup lx~I,

(10)

and we denote the resulting Banach space by l~., The set c of all con~
vergen t seq uen cc s is easily seen to be a closed linear subspace of l ~ and
is thereforP itself a Banach space. Another Banach space in this family
is the subset co of c which consists of all convergent sequences ,dt.h
Hmit O+
Examp1e 7~ The Banach space of primary interest to us js the space
e(X) of all bounded r.ontinuous scalar~valucd functions defined on a
topologiral space X, "1.th the norm gi vPn by

11111 = sup

(11)

lf(x)!L 1

r'fhis norm is sometimes called the uniform norm-, because the statement
that fn converges to f \Vith respect to this norm means that J~ converges
tof uniiorm.ly on X. The fact that this space is complete amounts to the
fact that if f is the uniform limit of a sequence of bounded cont.inuous
functions~ then f itself is bounded and continuous. If, as above 1 \ve
consider n-tuples and sequences as functions defined on f I ~ 2~
n.}
and ~ l, 2,
n, . . . }, then the spaces l! and l~ are the special CUSPS
of e(X) which correspond to choosing X to be the sets just mentioned.,
each \vi th the discrete topology .
r

...

The real apac~ e{ X) and t.he complex space e( X) are, of course, the spaces
previously denoted by e(XJR) a.nd e(X~C).
1

Banach Spaces

217

Many important properties of a Banach space arc closely linked to

the shape of its closed unit sphere, that is~ the set S = {x ~ llxll < 1}.
One basic property of S is that it is always coni,ex, in the sense (see
Problem 32-5) that if x and y are any t~To vectors in S, then the vector
z = ax + {3y is also in S, where a and {J are non-negative real numbers
such that a+ /3 = 1; for Hzll = llnx
'3Yll < o:\lxll ~llYll < a ~ l.
In this connection, it. is illuminating to con~ider the shape of S for certain
simple examples. Let our underlying linear space be the real linear

Fig. 35.

=:::;

&me olosed unit ~phere~T

space R 2 of all ordered pairs x = (z1:Px1) of real numbers. As we have

seen, there are many different norms which can be defined on R 2 ~ among
which are the follo-"ving: HxH 1 = lx1I
lz2!; flxl!2 ~ (lx1j 2 + lx21 2) }i; and
11x11 ~ = max {lx1l Jx2D. ~'igure 35 illustrates the closed unit sphere
\vhich corresponds to each of these norms. In the first case~ S is the
square with vertices ( 1,0), (O, l) 1 ( - 11 0), (O:P - 1) ; in the second~ it is the
circular disc of radius 1; and in the thlrd, it is the square with vertices
(1,1), ( - IJ l) J (~Ii - l):P (I~ -1). If we consider the norm defined by

(12)

where 1 :=; p < oo, and if we a11ow p to increase from 1 to ~ ~ then the
corresponding S's swell continuously from the first square mentioned to
the second. We note that S is truly "'spheric.al'' ~ p = 2. These
considerations also show quite clearly why we always assume that
p > 1; for if we \Vere to define llxll,. by formula (12) with p < I~ then
S = t x: Bx11 p 5 1 J would not be con vex (see the star-shaped inner portion
of Fig. 35). For p < 11 tberef ore, formula ( 12) does not yield a norm4

218

Operators

In the above examples, we have exhibited several different types of

Banach spaces, and there are yet others which we have not mentioned.
Amid this diversity of po.~sipilities, it is well to realize that any Banach
space can be regarded~from the point of view of its linear and norm
structures alone----a.s a closed linear sub space of e (X) for a suitable
compact Hausdorff space X. We prove this below, in our discussion
of the natural imbedding of a Banach space in its second conjugate space.

Problems

2t

3..
4.

Let N be a non-zero normed linear space, and prove th.at N is a

Banach space ~ {x : 11 x H ::;; 1 } is complete.
Let a Banach space B be the direct sum of the linca.r subspaces JI and
N, so that B = M N. If z ;; x + y is the unique expression of a
vector z in B as the sum of vectors x and yin .Af and lv'"., t..hen a ne\\T
norm can be defined on the linear space B by l.lzl1' = l~xll + l!Y!i.
Prove that this actually is a norm. If B' symbolizes the linear space
B equipped -with this nccw norm, prove that B' is a Banach space if
M and N are closed in B.
Prove Eq . (9) for the case of an arbitrary positive integer n.
In this problem we sketch the proofs-and \Ve ask the reader to fill
in the details-of some important inequalities relating t-0 n-tuples
x ~ (x 1, X..-!~
Xn) and y =:::::: (y1, Y<J,
Yn) of scalars. \1lhenever p occurs alone, and nothing is said to the contrary~ we assumP
that 1 < p < oo; and whenever p and q occur togethcrt we assum~
that both are greater than L and that 1/p + 1/q = 1.
(a) Show that a and b > 0 ===} allPblri < a/p + b/q. (If a = 0 or
b = Ot the conclusion is cleart so assume that both are positive+
If k e (0, I), define f(t) for t ?: 1 by f(i) = k(t - 1) - tk + 1~
Note that f (1) = 0 and f' (l) > 0, and conclude thn t tfl= < /.,:t
+ (1 - k) . If - a 2:. b~ put t == alb and k = 1/ p ; if a < b,
put t = b/a and k = 1/q; and in each case, dra"\v the required
r

(b)

'

conclusion.)
Prove II alder's inequality: 2:~ 1 [x1Yil < [lx[IPllY~I q (If x = 0 or
u = 0, the inequality is obvious~ so assume that both arc 0.
Put a; = (lxi[/llxll ,)P and bi = (IY1!/llvllq)q, and use part (a)
to obtain lxt:Yil/llxH,,llYI[ q < ai/p + bi/q. Add these inequali-+
ties for i = 11 2, . . . , n 1 and conclude that
n

( ~ [x&il)/llxHP!IYllq <

1/p

1/q

1.)

i=l

(c)

Prove Minkowski's inequality: llx + ylJP ~ llxllv + !IY~lp {The

inequality is evident v.fhPn p = I, so aHsume that p > l.

Banach Spaces

219

Use Holder's inequality to obtain

If z

+ yl[J'P =

ixi + Yt:~p

i-1

<

i-l

Jx1I ~x, +

~ ~

lXi + Yl /x, + y.;JP- 1

i 1
K

Y1IP- 1

IYir ~x.: + y,lp-- 1

i=l

< ([lxlJP

+ llyJI,.) llx + Yli,"''.)

When p = q = 2, Holder's inequality becomes Cauchyts

inequality as stated and proved in Sec. 15. The llolder and
JV1inko~,.ski inequalities can easily be extended from finite sums
to series+ For readers with some knowledge of the theory of
measure and integration~ we remark that these inequalities can
also be stated in the following much more general forms: if f is
in L,. and g is in Lu then their point'Wise product fg is in L1 and
and if f and g are both in Lr>, then f

+ g is also in LP and

It is to be understood that f and g ate measurable functions

defined on an arbitrary measure space and that the norms
occurring in these inequalities are those defined by formula {7).

47. CONTINUOUS LINEAR TRANSFORMATIONS

Let N and N' be normed linear spaces with the same scalars, and
let T be a linear transformation of N into N 1 l- \Vhen we say that T is
continuous, \Ve mean that it is continuous as a mapping of the metric
space N in to the metric space N'. By Theorem 13-B~ this amounts to the
condition that x~ ---t x in N ::::::} T(x~) ---i- 7"'(x) in N'. Our main purpose
in this section is to convert the rcquiremr.nt of continuity into several
more useful equivalent forms and t-0 show that the set of all continuous
linear transformations of N into N' can itself be made into a normed linear
space in a natura] way .
Theorem A. Let N and N' be normed linear spaces and T a linear trans-formation of N into N'. Then the following condition3 on Tare all equivalent to one another:
(1) Tis continuous;
(2) T 1~s continuous at the origin, in the sense that Xn ~ 0 ~ T(xn.) ___. 0;
In the futurcJ whenever we mention two normed linear spaces with a view to
~-onsirlc~Ting linear transformations of one into the other . we shall always e.asume-v.-T 1thou t n ec cssarily say jng so ex plici tly-tha t they have the aam e scalars.
1

220

Operators
(3)

there exist.a a real number K

> 0 with !he property

that

If

T (x) 11

< Kllxl[

for every x EN;

{4) if S = {x: [lxH < l} is the closed unit sphrt! in N, then its
image T(S) is a bounded set in N'.
PnOOF. (1) ~ (2)~ If T is continuous, then since T(O) ~ 0 it is certainly continuous at the origin~ On the other hand~ if Tis continuous at
the origin, then Xn "--"-;lo x <=;> x11- _... x """""7- 0 ==> T (xn - x) ___.., 0 ~ T (xn) - T (x)
--+ 0 ~ T(xn) ___.., T(x), so T is continuous~
(2) {::::> (3). It is obvious that (3) => (2), for if such a K exists~ then
Xn ~ 0 clearly implies that T(xft) ___, 0.
To show that (2) ~ (3), we
assume that there 'is no such K. It follows from this that for each
positive integer n we can find a vector Xm such that l T(x,.) l > n[lxnH, or
equivalently 1 such that llT(xn/n[lxn1Dll > 1~ If we now put

Yw. == Xn/n[lxn![,
then it is e-asy to see that y'*' ~ 0 but T(yn) -H- 0, so Tis not continuous
at the origin.
(3) -=> ( 4). Since a non-empty subset of a normed linear space is
bounded ~ it is contained in a closed sphere centered on the origin, it is
evident that (3) ~ (4); for if Hxll < 1, then [IT(x)~I < K~ To show that
(4) ~ (3), we assume that T{S) is contained in a closed sphere of radius
K centered on the origin. If x == O~ then '1. (x) = O, and clearly 11 T (x) li
::; K~lxll; and if x ~ 0, then x/llxll e S, and .therefore l T(x/~lxll) II < K,
so again ~re have 11 T{x) [I < K!lxll
1

If the linear transformation Tin this theorem satisfies condition (3),

so that there exists a real number K > 0 with the property that

HT(x)B ::; Kflxl1

for every x 1 then K is called a bound for T, and such a T ii often referred.
to as a bounded li'.near transformationr According to our theorem, T is
bounded 8 it is continuous, so these two adjectives can be used interchangeably. We now assume that T is continuous, so that it satisfies
condition (4), and we define its norm by
11 T 11 =

SU p

{uT (x) 11

: Hx 11

:5

1 J.

(1)

When N #- iO], this formula can clearly be written in the eqwvalent form

l T~I = sup {11 T(x) ~I : llxll

== 1}.

(2)

It is apparent from the proof of Theorem A that the set of all bounds for
T equals the set of all radii of closed spheres centered on the origin which
contain T(S). This yields yet another expression for the norm of T,

Banach Spaces

221

namely,

[1T[l

=inf {K:K

> 0 and llT(x)I[ < Kllxll

for all x};

(3)

and from this we see at once that

JI T(x) II

<

~I T[j

HxU

(4)

We now denote the set of all continuous {or bounded) linear transformations of N into N' by ffi(N,N'), where the letter Hffi" is intended to
suggest the adjective ''boundcd~n It is a routine matter to verify that
this set is a linear space with respect to the poin twise linear operations
defined by Eqs. 44-(1) and 44-(2) and to show that formula (1} actually
does define a norm on this linear space. we summarize and extend these
remarks in
Theorem B. If N and N' are nor~d linear spaces, then fM set ffi(}l,N 1 )
of all .continuous linear transformations of N into N' is itself a normed
linear space wi,th TC 8pect to the poinlwisc linear Operati&n3 and the narm
defined by ( 1). Further, if N' is a Banach space~ then CB ( N, N') U; also a
Banach space .
PnooF. We leave to the reader the simple task of showing that ffi(N,N')
is a normed linear space~ and we prove that this space is complete when
Nt is.
Let {Tn} be a Cauchy sequence in m(N,N')~ If xis an arbitrary
vectorinN,thcn j[Tm(x) - Tn{x)ll = ll(Tm - Tn)(x)1! < llTm - Tnll 11xH
shows that ~ Tn(x) J is a Cauchy sequence in N 1 ; and since N' is complete,
there exists a vector in_ N'~we denote it by T(x)---such that T11(x) ~
7'(x). This defines a mapping T of N into N' ~ and by the joint continuity of addition and scalar multiplicationJ T is easily seen to be a linear
transformationL To conclude the proof, we have only to show that T is
continuous and that T" --4 T "tith respect to the norm on ru(N,N').. By
the inequality 46-(1)~ the norms of the terms of a Cauchy sequence in a
normed linear space form a bounded set of numbers, so
11 T(x)

II

fllim

T n(x) l

;::=

1im

!ITn(x) I~

~ sup

([ITnll llxl[)

===

(sup

!IT~II) ~lxtl

shows that T has a bound and is therefore continuous. It remains to be

proved that llTn - Tll ~ 0. Lett> 0 he given, and let no be a po~itive
integer such that m, n > n4J ~ l)Tm - T~ll < E. If llxll < 1 and mi
n > no, then

ltTm(x) ~

Tn(X)[I

=:::;

ll(Tm - Tn)(x)!I < llTm - Tnll llzll

< If Tm -

T~~I

< E.

We now hold m fixed and allow n to approach oo, and we see that
[ITtft(x) ~ Tn(x)" ~ [IT.(.z) - T(x) 11, from which we conclude that

Operators

222

HT.(x)

- T(x) II ::5 '- for all m 2::. no and all x such that llx!I :5 1. This
shows that HT. - Tll S E for all m > no, and the proof is complete.
Let N be a normed line.ar spaee. We call a continuous linear
transf orma tio n of N in to itself an operalor on N, and 'vc dcno tc th c
normed linear space of all operators on N by ffi (N) instead of ffi (1V :tN) .
Theorem B sho~rs that <B(N) is a Banach space when N is. Ii,urthcrmore,
if operators are multiplied in accordance with formula 44-(3), then
fB(N) is an algebra in which multiplication is related to the norm by
(5)

This relation is proved by the f ollo'Wing computation:

HTT'H

<

sup 111 (TT') (x) ll: Hxl]

sup {II TH II T'(x) l ~ llxH

< 1}
< 1}

=
=

: llxli < 1 i
ll T(I sup [[IT' (x) ll: llx~I < 1}
== 11 TH 11 T' IJ.
sup { l T(7''(x)) [I

We know from the previous section that addition and scalar multiplication in <B (N) are jointly continuous, as t.hey are in any normed linear
space~ Property (5) permits us to conclude that multiplication is also
jointly continuous:
T~

---i-

T and T~---+ T' ~ Tn1 1 ~ ~TT'.

This follows at once from

II T l)T~ -

TT'll

IJ Tn(T: - T')

(Tn - T) T'll :$

li T n 1l

r: - T' 11

I]
HTfi -

Tll [l T'l1

We also remark that when N rfi { 0 l t then the identity transformation I

is an identity for the algebra <B (N). In this case~ we clearly have

l[J[I = I;

(6)

for II I u = SUp 11) I (x} ri : 11 x ll ~ 1 } ;;::: sup {!Ix " : Hx tI :::; 1 }

1.
We complete this section with some definitions which will often be
useful in our later work. Let N and N' be normed linear spaces. A:
isometric isomorphism of N into N' is a one-to-one linear transformation
T of N into N' such that l 1'(.x) 11 = !lxrl for every x in N; and N is said to
be isometrically iBomorphic to N 1 if there exists an isometric isomorphism
of N onto N'. This terminology enables us to give precise meaning to
the statement that one normed linear space is essentially the same as
another.
=:;;

Problems

1..

If M is a closed linear subspace of a normed linear space N, and if

Tis the natural mapping of N onto N/M defined by T(x) == x + M,
show that T is e. continuous linear transformation for which 11 TI! < 1.

Banach Spaces

2..

223

If T is a continuous linear transformation of a normed linear space

]\/Tinto a normed linear space N', and if Mis its null space, shov"'" thnt
T induces a natural Jinear transformation T' of N / M into N' and
that. r_ T'll ~ 11 T~l
Let N and N' be normed linear spaces Vt~ith the same scalars. If
N is infinite-dimensional and N 1 ~ { 0}, show that there exists a
Jine.ar transformation of N into N' \\hich is not continuous. (We
shall sec in Problem 7 that if N is finite-dimensional, then every
Jinear transformation of N into N' is automat]cally continuous~)
Let a line-ar space I-' be made into a normed linear space in t1vo 'vays,
and let the two norms of a vector x be denoted by l~xl[ and l]x[! '.
--rhcse norms are said to be equivalent if they generate the same to pol . .
ogy on L. Show that this is the case ~ there exist two posit.ive real
numbers K1 and K2 such that K1llxll ~ Hx!I' < K211xll for all x.
(Jf L is finite-dimensional, then any two norms defined on it a.re
equivalent. See Problem 7r)
If n is a fixed positive integer, the spaces l; (1 < p < C(J) consist
of a single underlying linear space v-,rith different norms defined
on it. Shov.r,. that these norms are all equivalent to one another.
(flint: sho,-_r that convergence vrith respect to each norm amounts to
roordinate\\rise convergence.)
If N is an .arbitrary normed linear space~ show that any linear transformation '1' of
(1 < p < co) into N is continuousr (Hint: if
{ei, e2, . . . , en} is the natural basis for 1;~ V.!here e, is the n-tuple
with 1 in the ith place and OJs elsewhere, then an arbitrary vector x
in l; can be written uniquely in the form
+

3.

4.

5.

6~

z;

74'

and from this 've get T(x) = aiT(e1) + a~ T(e2) + .

att T(ett);
novl apply the bin t given for Problem 5r)
Let N be a finite-dimensional normed linear space \vith dimension
n > O~ and let {e1, e2, ..
e,, 1 be a basis for N+ Each vector x
in N ran be "\\'"ritten uniquely in the form
4

If T is the one-to-one Jinl'ar transformation of N onto l7 defined by

7 (x) = (a1, a 2 ~ _, an), then 1,- 1 is continuous by Pro bl em Q_
(a) Prove that T is continuous. (/lint: if 1, is not rontinuous,
then for some E: > 0 there exists a sequence ~ Yn} in N such that
Yn--+ 0 a.nd II T(yn;) l ~ t; if Zn = Yn/ l T(y,i) I!, then z~-+ 0 and
Ii 7'(z:1';) 11 = 1; the subset of l~ consisting of all vcetors of norm 1
is compact, so [ T(z~)} has a subsequence \vhich converges to a
1.)
vector vlith norm 1; now use the continuity of
1

r-

224

Ope rotors

Show that every linear transformation of N into an arbitrary

normed linear space N' is continuous.
(c) Show that any other norm defined on N is equivalent to the
.
given norm+
(d) Show that N is complete, and inier from this that every finite . .
dimensional linear subspace of an arbitrary normed linear space
is closed .
It is a simple consequence of Problem 7 that every finite-dimensional
normed li11 ear space is locally compact~ Prove the converse, that is,
that a locally compact normed linear space N is finite-dimensional.
Hint: the cloRed unit sphere S of N is compact~ so there is a finite
subset of S, say {x1, x'2,
xft} J \.vith the property that each point
of S Lq dist.ant by less than 72" from one of the x./s; let !f be the
linear subspace of N spanned by the x/s; and show that M = N
(w do this~ assume that there exists a vector y not in M, use the fact
that M is closed to infer that d ~ d(y,Jf) > Oi find m 0 in M such
that d ::; llY - m-oll ~ 3d/2, and deduce the contradietion that the
vector Yo in S defined by YfJ = (y - mo)/ 11 y - me>~I is distant f ram M
by at least %) .
(b)

8.

48 . THE HAHN-BANACH THEOREM

One of the basic principles of strategy in the study of an abstract
mathematical system can be stated as follo,vs: consider the set of all
structure-preserving mappings of that system into the simplest system
of the same type. This principle is richly fruitful in the structure theory
(or representation theory) of groups, rings, and algebrasi and v.re shall see
in the next section ho\\"' it V{orks for normed linear 8paces.
We have remarked that t.he spaces R and C are the simplest of all
normed linear spaces. If N is an arbitrnry normed linear space, the
above principle lead~ us to form the set of all continuous linear transformations of N into R or C, according as N is real or complex. This setit is IB(N,R) or IB(NJ(_})--is denoted by N* and is called the conjugate
spUe of N+ The elements of N* are called continuous linear Junctionals 1
or more briefly, funclionals. 1 It follov-.rs from our "''ork in the previous
section that if these functionals are added and multiplied by scalars
The noun '{functional" seem~ to have originated in the theory of integral
equations. It was used to di~tin guish bet ween a function in the elem en t.ary sense
d cfin cd on a set of numbers and a function (or fun ct ionel) defined on e set of rune lions~
In thia book, ~~e always use the word to mea.n & scalar--va.lued continuoua linear function defined on a normed linear spa.ce.
1

Bana(:h Spaces
pointwise~

and if the norm of a functional

ll!H

= sup {lf(x)I: llxll

225

f is defined by

S 1}

== inf {K: K > 0 and !f(x)i :$ Kllxl~ for all x I,

then N* is a Banach space.

Wb.en we consider various specific Banacn spaces~ the problem arises
of determining the concrete nature of the functionals associated with
these spaces. It is not our aim in this section to explore the ample body
of theory which r.enters around this problem, and in any case, the machinery necessary for such an enterprise (mostly i,he t~eory of measure and
integration) is not available to us. Nevertheless~ for the reader who may
have the required background, we mention some of the main facts without proof.
Let X be a measure spaice ~ith measure m, and let p be a given real
number such that 1 < p < Cl:J.. Consider the Banach space L 11 of all
measurable functions f defined on X for which If(x) IP is integrable. If
g is a function in Lq, v-.There 1/p
1/q = 1, we define a function Ffl on
LP by

Fa(!J =

Jf(x)g(x) dm(x).

The Holder inequality for integrals mentioned at the end of Problem

4.6-4 shOVlS that

IFQ(f)I

I/ f(x)g(x) dm(x)I

<

J lf(x)g(x)~ dm(x)

::=; I! !~I ~II gH~

We conclude from this that F{/is a well-defined scalar-valued linear function on LP with the property that HFall ~ l\gH 11 ~ and is therefore a functional on LP. It can be shown that equality holds here~ so that

HF.:11 = 11Yll ~
It can also be shown that every function al on L 11 f.lJ'i.ses in this way, so the
mapping g-----+ Fa (which is clearly linear) is an isometric isomorphism of
L.q onto L'P *. This statement is usually expressed by writing

L,,*

L"-J

(1)

where t.he equality sign is to be interpreted in the sense just explained.

If we sper,ialize these considerations to n-tuples of scalars, we see
that (I) becomes
(2)

226

Operators

FurtherJ it can be shown that

and that

(l~) =

l!

(3)

(l!)* =

l~ ~

(4)

We sketch proofs of (2), (3), and (4) in the problems.. When we consider
sequences of seal ars, then for 1 < p < oo we have the f oUowing special
case of (1):
ljj.
If p

l(J_~

(5)

1, we obtain s natural .extension of (3):

l1* =

l~.

(6)

The corresponding extension of {4) is another matter, for it is false that

l.i * = l 1 ~ Instead, we have
(7)
What is lfR *? We saw in Sec. 46 that l~ is a special case of e(X), so this
question leads naturally to the problem of determining the nature of the
conjugate space ~ (X). The classic solution of this problem for a space
X which is compact Hausdorff (or even normal) is kno\vn as the Riesz
representation theorem, and it depends on some of the deeper parts of the
theory of measure and integration (sec Dunford .and StJ1,vartz [8, pp.
261 ~265 J). The situation is somewhat simpler for the case in v,Thich X is
an interval [a~b] on the real line~ but even here an adequate treatment
requires a knov.Tledge of Stieltjes integrals (see Riesz and Sz.~N agy
{35~ secs~

49-51 ]) ~
Most of the theory of conjugate spaces rests on the Hahn-Banach
theorem, which asserts that auy functional defined on a linear subspace of
a normed linear space can be extended linearly and continuously to the
whole space without increasing its norm. The proof is rather complicated, so we begii1 with a lemma \Vhich serves to isolate its most difficult
parts.
Lemma. !Jet }rf be a linear subspace of a normed linear space N, and let
f be a functional defined on M. If XQ is a vector not in M, and if

Mo= M

+ [x-0]

is the linear aubspace spanned by M and Xo, then f can be extended ta a

functional f (J defined on M 0 such that 11f0 rI == if fl I
PROOF. We first prove the lemma under the assumption that N is a
real normed linear space. We may assume, without loss of generalityJ
that ll!H =:; 1.. Since xo is not in MJ each vector y in Mo is uniquely
expressible in the form y = z + a.z 0 with x in M. It is clear that the
r

Banach Spaces

271

definition fo{x + axo) = fo(x) + afo(xo) = /(x) + aro extends f linearly

to Mo for every choice of the real number ro = f o(xo). Since we are
trying to arrange matters so that Hfol! = 11 our problem is to show that
To can be chosen in such a way that ifo(x + a:r:et)I < llx + axo!I for
ever:y x in M and every a 0. Since f o(z + axo) ~ f(x) + aroJ this
inequality can be v...-ritten as

~ Hx
axo![
-f(x) ~ !Ix+ ax{dl

or

< J(x) + aro < llx + axolf

< aro < ~ /(x) + Hx + ax(]I],

which in turn is equivalent to

\Ve now observe that for any two vectors x 1 and x!I in .llf we have

f (x2)
so

f (xi)

llx"l

f (x2 - x J) ::::; If(x~ - xl) J < l !JI ~Ix~ - x. H

x1ll = 'I (x~ + Xo) - (x. + xo) H < llx2 + xo!I + llxr
-

/(x1) - l:x1 + xolJ < -

f(x:.)

+ ilx! + XD~I ~

+ xol!,
(9)

If \Ve define t\vo real numbers a and b by

and

a ~ sup { -f(x) - llx

b = inf { ~ f (x)
11 x

+ xoB ;x J_~J
+ Xo I~ : x C. Jf J ,

then (0) shows that a ~ b. If "Te nov{ choose ro to he any real number
such that a < r 0 ~ bJ then the required inequali(y (8) is satisfied and this
part of the proof is complete.
lVc next use the result of the above paragraph to prove the lemma
for the case in which N is complex. Here f is a complex-valued func~
tional defined on ~! for \Vhich flfll = lr \\re begin by remarking that a
complex linear space can be regarded as a real linear space by simply
restricting the scalars to he real numbers~ If g and h are the real and
imaginary parts of fi so t.hat f(x) = g(x)
ih(x) for every x in L~f 1 then
both g and h are en.sily seen to be real-valued functionals on the real
spae.e 11!; and since 11f11 ;:; 1, we have ~I g 11 < 1 ~ 'fhe eq ua ti on

f (ix)
together \Vith /(ix) = g(ix)

===

if (x)

+ ih(ix) and

if (x) ;:;;: i (g (x) + ih (x))

ig (x) - h (x) i

shol\rs that h(x) = -g(ix), so we can write /(x) = g(x) - ig(ix). By

the aboYe paragraph 1 we can extend g to a real-valued functional go on
the real space J.~f o in such a way that lluoll = If gll ~ and we define fu for

228

Operators

x in Mo by fo(x)

= go(x) - ig(l(ix).

It is easy to see thatfo is an extension

ofjfrom M to Mo 1 thatfo(x + y) = frJ(x) + fo(y), and thatfo(a.r) = afo(x)
for an real a~S+ The fact that the property last stated is also valid for
all comp]ex ats is a direct consequence of
j 0 (ix)

;::=

gfJ (ix) - ig fJ ( i 2 x) = go (ix)

+ ig o(x)

;;::=

i (go ( x) - ig o(ix)) = ifo(x) J

so f c. is linear as a complcx~valued function defined on the complex space

Mo. All that remains to be proved is that ~I foll = 1 1 and we dispose of
thi.B by showing that if x is a vector in Mo for which llxjl = I, then~
lfo(x)j < 1.. If f-0(x) is real~ this follows from fo(x) = go(x) and l]goll < 1 .
If fo(x) is complex, then we can write f.o(x) = re~ 6 with r > 0, so
f fo(x)I

= r =

ri 8fo(x)

and our conclusion now follows from

that fG(e-i 6x) is real~

fo(ri'x);

lle-'xl] ;:;: ; llx[I

= 1

and the fact

Let M be a linear subspace of

a normed linear space N, and let f be a functional defined on M. 'J~hen f
can be extended to a functi-Onal fo defined on the whok space N Buck that
Theorem A (rhe Hahn-Banach Theorem)t

11/oll

llfl].

The set of all extensions of f to functionals g with the same

norm defined on subspaces which contain Mis clearly a partially ordered
set with respect to the follol\dng relation: g 1 < 02 means that the domain
of g1 is contained in the domain of Y2i and g-,{x) = g1(x) for all x in the
domain of g1. It is easy to see that the union of any chain of extensions
is also a.n extension and is therefore an upper bound for the chaina
Zorn.J:s lemma now implies that there exists a maximal extension f 0 We
complete the proof by observing that the domain of fo must be the entire
space N 1 for othe1vlisc it cguld be extended further by our lemma and
would not be maximal.
PROOF.

As we stated in the introduction to this chaptert the main force of

the Hahn-Banach theorem lies in the guarantee it provides that any
Banach space (or normed linear space) has a. rich supply of functionals.
This property is to be understood in the sense of the following two
theorems, on which most of its applications depend.

If N is a normed linear 8'f'UCe and x 0 is a non-zero vector in N,

then thRre exist8 a functional f o in N * such t.hat f o(x-0) = Uxo I! and I~ f o1 = 1.
PROOF.
Let M = { ax 0 i be the linear subspace of N spanned by xo1
and define f on M by f(axo) = ailxo[I It is clear that f is a functional on
Theorem 8.

M such that f(xo) = lixoll and 11/ll = l~ By the Hahn-Banach theorem,

f can be e_xtended to a functional f 0 in N* with the required properties.

Banach Spaces

229

Among other things, this result shows that N* separates the vectors
in N, for if x and y are any two distinct vectors, so that x - y # 0 1 then
there exists a functional fin N such ~that f (x - y) ~ 0, or equivalently,
f(x) ~ f(y).
Theorem C.

If M is a closed linear subS'])ace of a norme.d linear space N

and Xo is a vector nat in Mt then the.re exists a f unclional f o in N * such that
f o(M) = 0 and frlxo) ~ 0.
PROOF.
The natural mapping T of N onto N / M (sec Problem 47-1) is a
continuous linear transformation such that T(M) = 0 and
T(x(l) === Xo

+M

~ 0.

By Theorem B~ there exists a functional fin (N / M) * such that

f (XG
Ii we now defin~ f\J by fo(x)
desired properties~

+ jf)

~ Q.

= f(T(x))~

then f-0 is easily seen to have the

These theorems play a critical role in the ideas developed in the

following sections, and their significance will emerge quite clearly in the
proper context.
Prob[ems

1.

2.

Let M be a closed linear subspace of a normed linear space N t and

let Xo be a vector not in lv.l~ If dis the distance from xo to M~ show
that there exists a functional / 0 in N* such that JQ(M) = 0, fo(xo) = I J
and U!ol~ = 1/d.
Prove that a normed linear space N is separable if its conjugate space
N* is. (Hint: let {fn l be a countab1e dense set in N* and {Xn l a
corresponding set in N such that llxfil:
I and I fn(Xu) r > llfn[l/2;
let M be the set of all linear combinations of the xri's whose coefficients are rational or~if N is complex~have rational real and
imaginary parts; and use rfheorcnt C to show that M = N~) We
remark that N* need not be separable when N is, for l1 is easily
proved to be separable, l 1 * = li, and Zm",) is not separable (see Problem

s:

3.

18 4).
In this problem we ask the reader to convince himself of the vu.lidity
of Eqs. {2) to (4)~ Let L be the linear space of aU n-tuples
X

(x l~

X2,

..

Xn)

of scalars. If feh e~, . . . , en} is the natural basis described in

Problem 47~6~ then x = x 1e + Xit + . + x~en; and if f is

230

Operators

any scalar-valued linear function defined on L, then the equation

f(x) = x1f(e1) + x,,f(e2) + ~ + x..f(e.,.,) shows that f determines,
and is determined by1 the n scalars Y = f(e1). The mapping
Y

(Y11

Y21

Yn) --+- f,

"'here f(x) = ~;_ 1 x,.yr:, is clearly an isomorphism of L onto the

linear space L' of all JJ s~ Wilen the space IJ of all x 1 s is made in to
l; (1 < p < ~) by suitably defining its norm1 then by Problem 47-6
the space L' of all f s equals its conjugate space (l;) * 1 "\vhere the norm
off is understood to be given by

llfl:

inf

{K: K

~ 0 and lf(x) I

< Kl]x]f

for all x}.

All that remains is to see 'vhat norm for the y~s makes the mapping
y ~fan isometric isomorphism.
(a) If 1 < p < o: 1 th.en (l;) * =
The norm in this caHe is
defined by Hxll = (Zi-1 lriJP) llPi and it fo1lov{S from

1:.

lf(x)I = I,~ x,y,'

:$ ;~

1x,y,r < (;~ jx;f')-1'.!' (;~ :y,:) i1~

<

(2:ic.1 IYil q) 1l q. Sho'v that i~J: I == (~~zml r y;] q) l/ q by

considering the vector x defined by Xi = 0 if y1 :::s: 0 and

that I! f 11

oth erVtrise.
(b)

Herc we have !lxll ~ ~~ -1 lxil, and it follo\vs from

lf(x) I = 1i:.c1 XjYif < 2;~ - t ~xii [y;j -S max UYil J (:E7=--1 lxil) that
11 f I! < rnax {!Y1d J . Shovl that 11f11 = max {~ Yif J by considering
the vector x defined by Xi = jyil/Yi if ~Yil = max {lYil l and
Xi = 0 othcrv.:ise.
(l~) * =. l~.
In thls case, the norm is defined by
(l~) * = l~ ~

(c)

If xi] = max

Uxil},

and it f oHo"'~s from

n

[f(x) I ~ I~ xiyi~

<

i~l

that 11 f ll < l;~~l

4.

~ jxil
il

Yi!.

IYif 5 max { ]xi]l

(I IY1il)
i~I

Show that 11 ill == !.7_1 IY,;! by considering

the vector x defined by x. = 0 if Yi = 0 and Xi = IYil/Yi
o th cr\v ise.
The follo\\:ing generalized form of part of the IIahn-Banach theorem
is uSl"f ul in certain problemB of measure theory. Prove it by gui tab]y
modifying the arguments given in the text+ If p is a real f unrtion
r

Banach Spaces

231

defined on a real linear space L such that p(ax) = crp(x) for a: > 0
and p(x
y) < p(x) + p(y) 1 and if f is a real linear function defined
on a linear subspace JI such that f{x) :::; p(x) for all x in ~f 7' then
f can be extended to a real linear function f rJ d cfincd on l.J such that
fo(x) ,:::; p(x) for all x in L .

.49. THE NATURAL IMBEDDING OF N IN 1'l**

Since the conjugate space N* of a normed linear space }l is itself a
normed linear Bpace, it is possible to form the conjugate space (N*) =- of
N*. \\,..c. denote this space by N**, and \\~c call it the second conjugate
space of N~
The importanee of N** rests on the fact that each vector x in N
gives rise to a functional F ~ in N **. If \\~e denote a typical element of
N* by f, then F~ is defined by

Fzj(f) = f(x).
In other wordB, we invert the usua1 pract.ir..e by regarding the symbol
f(x) as specifying a function off for eaeh fixed xJ and v...-e emphasize this
point of view by writing f(x) in the form F~(f)+ A simple manipulation
of the definition shows that F ~ is linear:

F~(af

+ {jg)

(af + /3g) (x)

== af(x) + ~g(x)
= aF~(fJ + l'F%(g).

If we now compute the norm of F~, we see that

HF~ll

= sup {IF:.{/)!: llfll ~ 1}

= sup {lf(x) I: Hill :::; 1}
< sup {[Iii! llxH: 11![1 < l}
~

ltxl!

Theorem 48-B is exactly wh&t is needed to guarantee that equality

holds herc7' so for each x in N we have

llF~II

[]xJI.

It follows from theae observations that x-+ F~ is a norm-preserving

mapping of N into N**. F~ is called the functional on N* induced by
the vector x, and we refer to f unc tio nals of thls kind as indmed functionals~
We next point out that the mapping x--+ F~ is linear and is therefore an
isometric isomorphism of N into N**.. To verify this, we must show that
F>HJ(f) = (F2
F11 )(f) and F{U(f) = (aF~)(f) for every fin N. The

232

Operators

first of these relations follows from

F~+v(f)

+
+

= f(x
y)
= f(x)
f(y)
= Fr(f) + Fu(f'J
= ( F :t + F !J) (f),

and the second is proved similarly. 1'hc isometric isomorphism x ~ F%

is called the natural -imbedding of Nin N** for it allo\vs us to regard N as
part of N** "\\rithout altering any of its structure as a normed linear space.
7
"'\\ c write
N CN**
~
j

where this set inclusion is to be understood in the sense just explained .

A normed linear space N is said to be reflexive if N = N**.. The
spaces lp (and Lp) for 1 < p < co are reflexive~ for lp * = lq and

l p **

--

l q*

--

l p

z:

It fo1lows from Problem 48-3 that the spaces for 1 < p < co arc also
reflexive. Since N** is complete, N is necessarily complete if it is
reflexive+ If N is complete~ however, it is not necessarily reflexive, as we
see from c{t* ~ 11 and co** = l1* = l~. If X is a compact Hausdorff
space, it can be sho"'"'n that e(X) is reflexive RX is a finite set+
'rherc is an interesting criterion for reflexivity, which depends on the
co nee pt of the wealc topology on a 11 or med linear space N ~ This is
de.fined to be the weak topology on N generated by the functions in N*
in the sense of Sec. 19; that is 1 it is the weakest topology on N with
respect to whirh all the functions in N* remain continuous~ The criterion
referred to is t.he f ollowing: if B is a Banach space, and ifs = { x: nxH < 11
is its closed unit sphere, then B is reflexive-RS is compact in the weak
topology. J.'his fact is something one should kno-.,v about Banach spaces,
but "\\Tc shall have no need for it ourselves, so we state it "\\ithout proof . 1
Far more important for our purposes is the weak topology on N*,
lNhich is defined to be the weak topology on N* generated by all the
induced functionals F~ in N**. This .situation is rather complicated, so
\\Te shall try to make clear just what is going OIL
1:irst of all, N* (like N) is a normed line~r space 1 and it therefore
has a topology derived from its character as a metric space~ Thia is
called the strong topology. N** is the set of all scalar-valued linear
functions defined on N* \i\Thich arc continuous ,vith respect to its strong
topology.. The weak topology on N* (like the weak topology on N) is the
weakest topology on N* ~rith respect to which all the functions in N* a.re
continuous 1 and clearly this is weaker than its strong topology. So farj as
See Hille ~nd Phillips [20, p. 38] or Dunford and Sehwartz

l8:.

p. 425].

Banach Spaces

233

we have indicated, these concepts apply ~qnally to N and N*~ However,

since N* is the conjugate space of N, the natural imbedding enables us to
consider N as part of N*.. "\\'~e nO\V form the weakest topology on N*
with respect to which all the functions in N~regarded as a subset of
N**-remain continuous~ This is the weak* topologyt and it is evidently
weaker than the weak topology. The weak* topology- can be given a
more explicit description~ in which its defining subbasic open sets are
displayed.. Consider a vector x in N and its induced functional F:z in N**
The weak topology on N is the weakest topology u1:1der which all such
F~ts are continuous.. If fo is an arbitrary element in N*, and if E > 0 is
given 1 then the set
S(x,

fo,

E)

;:::=

f :f N* and IF~<n

- {f :f N* and l/(x) -

F;:(fo)~

<

f n(x)I <

~:l

f}

is an open set (in fact, a neighborhood of Jo) in the ~,.eak* topology .

Furthermore~ the class of all sets of this kind 1 for all x s~ Jo's., and E'st jg
the defining open subbase for the weak* topology. All finite intersections
of these sets constitute an open base for this topology, and the open sets
themselves a.re all unions. of these finite intersections~
We remark at this point that N is a Hausdorff space with respect
to its weak* topology.. This follows at once from the fact that if f and g
are distinct functionals in N*, then there must exist a -vector x in N such
that f(x) r== g(x); for if we put E = lf(x) - g(x) l/3, then S(x, f, t:) and
S(x, g~ E:) are disjoint neighborhoods off and gin the weak* topology.
Let us now consider the closed unit sphere S* in N*, that is~ the set
S* = f f:f e N* and 11/I] < l}r 1 It is an easy consequence of Problem 2
that S* is compact in the strong topology ~ N is finite-dimensional~ so
the strong compactness of S* is a very stringent condition. If N is
complete,, it follows from Problem 3 and our unproved criterion for
reflexivity that S* is compact in the "\i\;Peak topolo~l RN is reflexive, so
the "Teak compactness of s is still a fairly substantial restriction. We
state these facts to emphasize that the situation is quite different with
the weak* topology, for here s is always compact.
1

Theorem A,. If N is a normed limar spac-et then the closed unit sphere S*
in N* is a compact Hausdorff space in the weak* topology.

We already know that S* is a Hausdorff space in this topology t

so we confine our attention to proving compactness~ With each vector
x in N 've associate a compact space C~, where C~ is the closed interval
[- llxll ~llxllJ or the closed disc {z :lzl < llxll} 1 according as N is real
or complex. By Tychonoff~s theorem, the product C of all the G\/s is
PROOF.

1 When we use the adjective "closedH in referring to S*, we int.end only lo emphasize the ineq u.a.lity 1 /II ~ 1, as con tm.sted with 1! /H <1.

234

Operators

also a compact space. For each x, the values f(x) of all fs in s lie in
C~~ This enables us to imbed S* in C by regarding each f in S* as
identical with the array of all its values at the vectors x in N~ It is clear
from the definitions of the topologies concerned that the weak* topology
on S* equals it::i topology as a subspace of C; and since C is compact.i it
suffices to sho,v that s is closed as a subspace of C. We shov{ that if g
is in S*, then g is in S* ~ lf we consider g to be a function defined on the
index set Nt then since g is in C we have lu(x) I < Hxfl for every x in N.
It therefore suffices to show that g is linear as a functjon defined on N.
Let ~ > 0 be given~ and Jet x and y be any t'\\~o vectors in N. Every
basic neighborhood of g intersects J.~, so there exists an f in 8* 8uch that
fg(x) ~ f(x)I < f/3, lg(y) - /(y)i < f/3, and Jo(x + y) - f(x + y)I <
t=/3. Since f is linearJ j(x + y) - f(x) - j(y) = 0, and we therefore have

jg(x

+
+ +

+
+

+ y)

- g(x) - g{y)[ = l[g(x

y) - f(x
y)] ~ [g(x) ~ f(x)]
- [g(y) - /(y)Jt < rg<x
y) - J(x
y)f
~g(x) - f(x)I

~g(y) - f(y)J

< E/3 + t/3

+ t:/3

f;.

The fact that this inequality is true for every E > 0 nov-r implies that
g(x + y) = g(x) + y(y)r We can show in the same way that
g(ax)

ag(x)

for every scalar a 1 so g is linear and the theorem is proved.

\Ve arc no\\" in a position to keep the promise made in the last
para.graph of Sec. 46~ for the f ono,ving result is an obvious conscq uence
of our preceding \\.Tork.

B.

I.ct ~r be a normed linear space J and let S* be the corripact

H au.s<lorff space obtained by imposing the weak* topology on the closed unit
.ftJphere in N*. Then the mapping x ---+ F x, tDhere F x (j) = f (x) for each f in
i..'::*, is an isotnetric isoni.orphisni of N into e(1..~*). If N is a Banach space. 1
this 1napping i~~ an isonietric isomorphism of N onto a closed linear subspace
of e{S*).
Theorem

This theorem sho,vs 1 in effect 1 that the most general Banach space is
rssent.ially a closed linear subspa~e of e(X), where X is a eompaet
llausdorff r.;pacc. The purpose of representation theorems in abstract
mathema.tif's is to revca] the structures of complex systems in terms of
simpler onP:.-:i} and from this point of view, Theorem B is satisfying to a
dc-grec. It must be pointed outJ however~ that ,,.e know next to nothing
about the r.losed linear subspaces of e(X), though we know a good deal
about e( . Y)
. it.self
Theorem R is thcrefo1e some'"That less revealing than
appears at first glance~ We shall see in Chaps. 13 and 14 that the
L

Banach Spaces

23.S

corresponding representation theorem for Banach algebras is much more

significant and useful.
Problems

1.

2.
3.
4.

5.

Let X be a compact IIausdorff space 1 and justify the assertion that

e(X) is reflexive if X is finite.
If N is a :finite-dimensional normed Ii near space of di men sio n n, show
that N* also has dimension n. Use this to prove that"/'{ is reflexive.
If B is a Banach space} prove that R is reflexive ~ B* is reflexive.
Prove that if B is a reflexive Banach spacei t.hen its closed unit
sphere Sis Vtreakly compact.
Show that a linear subspace of a normed linear space is closed <=>it is
\Veakly closed+

50. THE OPEN MAPPING THEOREM

In this section ~Te have our first encounter 'vith basic theorems which
require that the spaces concerned be complete. The follo.ring rather
technical lemma is the key to these theorems.
Lemma. If B and B' are Banach Bpaces, and if T is a contin11ouB linear
transformation of B onto B 1 1 then the image of each open sphere centered on
the origin in B contains an open sphere centered on the origin in B 1

We denote by Sr ands; the open spheres 'vith radius r centered

on the origin in B and B'. It is easy to see that
PROOF~

T(S,.) ~ T(rS1) = rT(S1),

so it suffices to show that T(S1) contains some s;~

We begin by proving that T(,.,9 1) contains some s;. Since Tis onto,
we see that B' =
B 1 is completc 1 so Baire~s theorem
1 T(S,.).
implies that some 1'{8,.,,J has an interior point yo 1 which may be assumed
to lie in T(Sn~). 1.,he mapping y ~ y ~ Yo is a homeomorphism of B'
onto itsclf:r so T(Sn~J ~ Yo has the origin as an interior point.. Since Yo
is in T(Sn 0) , \Ve have T(S~ 0) ~ Yo ~ 11 (82na); and from this we obtain
T(Sna) - yf) = T(Sn(J) - Yo C T(S 2ri,), which sho~~s that the origin is an
interior point of T(S 2 niJ.. :Lvf ultip}ication by any non-zero scalar is a
homeomorphism of B, onto itself, so '1'(82~~) ~ 2n.GT(S1) = 2nnT(S1); and
it follo~s from this that the origin is also an interior point of T(S 1), so
s; C T(S 1 ) for some positive number f~
\Vc conclude the proof by sho,ving that
S::: T (S l) 1 which is clearly
1
equivalent to s; 1a ~ '.f (81)~ Let y be .a vector in B such that 11Y~: < f.

v:r=

s:

236

Operators

is in T(S 1), there exists a vector x1 in B such that Hx1ll < 1 and
llY - Y1 II < E/2)' where Yi = T(x1). We next observe that
s;
T(Sh), so there exists a vector x2 in B such that llx2!I < ~ and
if (y ~ Y1) - Y2fl < f/4, where Y'l = T(x2)~ Continuing in this way, we
obtain a sequence {xn} in B such that IJxmlf < 1;2~~ 1 and l!u -- (Yt + Y2
+ Yn) 'I < E/2~ 1 where y,,.. = T(xn). If we put

Since

s:12

then it follows from

which

Jlxnil <

l/2"- 1 that {sn} is a Cauchy sequence in B for

B is complete, so there exists a vector x in B such that sn ..._... x; and

llxll ~ Ulim snll == Iim U8nU ~ 2 < a shows that x is in
All that
remains is to notice that the continuity of T yields

s3.

T(x)

T(lim sn) = lim T(sn) = fun (Yi

+ Yi + + Y")

y,

from which we see that y is in T(Sa).

This makes our main theorem easy to prove.

If B and Br are Banfich
spaces1 and if T is a continuous linear trans/ormation of B on.to B 1 , then
T i8 an open mapping.
PROOF4
We must show that if G is an open set in B, then T(G) is also
a.n open set in B 1 If y is a point in T(G), it suffices to produce an op~n
sphere centered on y and contained in T(G). Let x be a point in G
such that T(x) = y. Since G is open, xis the center of an open sphere-which can be written in the form x +Sr-contained in G.. Our lemma
now implies that T(ST) contains some s;l.. It is clear that y
s;~ is an
open sphere centered on y, and the fact that it is contained in T(G) follows
at once from y
~y
T(S,.) = T(x) + T(S,) = T(x +Sr) ~ T(G)~
Theorem A (the Open Mapping Theorem).

+ s:.

Most of the app1ications of the open mapping theorem depend more

directly on the following special case, which we state separately for the

sake of emphasis.
A one-to-one continuous linear transformation of Qne Banach
space onto another is a homeomorphism. In particular, if a one-to-one
linear transformation T of a Banach space onto it.JJelf is continUOUB, then its
inverse ~ 1 is automatically continuous .
Theorem B.

As our first application of Theorem B, we give a geometric characterization of the projections on a Banach space+ The reader will rece.ll
from Sec. 44 that a projection E on a linear space L is simply an idem-

Banach Spaces

237

potent (E1 = E) linear tra.nsformation of L into itself~ He will also

recall that projections on L can be described geometrically as follows:
(1) a projection E determines a pair of linear subspaces M and N
such that L = M E9 N, where M = f E}(x) :x e L} and N =
{x :E(x) = O} are the range and null space of E;
(2) a pair of linear subspaces M and N such that L = M EB N
determines a projection E whose range and null space are
M and N (if i
x + y is the unique representation of a
vector in L as a sum of vectors in M and N t then E is defined by
::=:;:

E(z) = x)~
These facts show that the study of projections on Lis equivalent to the
study of pairs of linear subspaces which are disjoint and span L.
In the theory of Banach spaces1 however, more is required of a
projection than mere linearity and idempotence. A pra}edion on a
Banach space Bis an idempotent operator on B; that is, it is a projection
on B in the algebraic sense which is also continuous. Our present task
is to assess the effect of the additional requirement of continuity on the
geometric descriptions given in (1) and (2) aboveL The analogue of (1)
.
lS easy.
Theorem C1 If P i8 a projection on a Banach space B 1 and if Mand N
are its range and null space, then Mand N are closed linear subspaces of B
such that B
M EB N .
PROOF. - P is an algebraic projection 1 so (1) gives everything except
the fact that M and N a.re closed. The null space of any continuous
linear transformation is closed J so N is obviously closed; and the fact
that M is also closed is a consequence of
:;:::=

M = {P(x):xEBt = {x:P(x) = xJ == {x:(J

~ P)(x)

= O},

whlch exhibits M as the null space of the operator I - P.

The analogue of (2) is more difficult, for Theorem B is needed in its
proof.

Let B be a Banach space, and let 11.f and N be clo8e.d linear

subspaces of B 8Uch that B = M EB N+ If z = x
y is the unique
representation of a t. ecwr in B as a sum of vectors in M and N ~ then the
mapping P defined by P(z) = xis a projection on B whose range and null
space are Mand N .
Fiioo F~ Everything stated is clear from (2) except the fa ct that P is
continuous, and this we prove as follows. By Problem 46-2J if B'
denotes the linear space B equipped with the norm defined by
Theorem D..

I zII' :::;

Jlx[[

+ HYU,

238

Operators

then B' is a Banach space; and since l! P(z)jl ~ llxll S Hxfl + l!-u[I = flzll'1
P is clearly rontinuous as a mapping of B' into B. It therefore suffices
to prove that B' and B have the same topology. If T denotes the
identity mapping of B' onto B, then
~I T(z)[I =

[liH

IJx + Yll <

l]x~I

+ llyff

llzl!'

sho\vs that T is continuous as a one-to-one linear transformation of B'

onto B. 1-rheorem B now implies that T is a homeomorphism, and the
proof is complete.
This theorem raises some interesting and Hignificant questions~ Let
J-f be a closed linear subspace of a Banach space B. As "'~e remarked at
the end of Sec. 44 1 there is a},.,ays at least one algebraic projection defined
on B whose range is ~it and there may be a great many+ Howcver 1 it;
might \vell happen that none of these are continuous, and that conse<1 uen t.ly none are projections in our presr.nt sense. In the light of our
theorems, this is equivalent to saying that there might not exist any
closed Jinear subspace N such that B = .Jf N. \\that sorts of Banach
spaces have the property that this a\,:kv{ard 'situation cannot occur?
'Ve shall sec in the next chapter that a Hilbert space~,vhich is a sperial
type of Banach space--has this property. \Ve shail also see that this
property is closely linked to the satisfying geometric structure ,vhich
sets Hilbert spaces apart fron1 general Banach spaces.
\Ve no\v turQ to the closed graph theorem. Let Band B' be Banach
spaces. If " . .c define a metrir. on the product B X B' by
j

then the resulting topology is easily seen to be the same as the product
topology 1 and convergence "Tith respect to this metric is eq ui valen t to
coordinate,visc convergence+ Now let 1' he a linear transformation of
B into B'~ '\Ve recall that the graph of T is that subset of B X B 1
which consists of all ordered pairs of the form (x/T(x))
Problem 26-6
shows that ii T is continuous1 then its graph is closed a.s a sub:.-:;et of
B X B ~ In the present context, the converse is also true.
+

Theorem E (the Closed Graph Theorem). If B and B, are Banach

Bpaces, and if Tis a linear transformation of B into B' 1 then Tis continuous
~ its gra pk is closed.

In view of the above remarks~ we may confine our attention

to proving that Tis continuous if its graph is closed. 'Ve denote by B1
the linear space B renormed by 11xfl 1 = IJxH + II T(x) !I.. Since

PROOF.

HT(x) 11

<

~lxl]

+ ll T(x) [I

= llxH h

Banach Spaces

239

Tis continuous as a mapping of B1 into B'. It therefore suffices to sbo\V

that B and B1 have the same topology.. The identity mapping of 81
onto Bis clearly continuousJ for llxll < llxll 11 T(x) l ~ !lxll 1~ If we can
show that B1 is completet then Theorem B will guarantee that this
mapping is a homeomorphlsm, and this ";11 conclude the proof. I.Rt
i Xn} be a Cauchy sequence in Bl It follows that {x~} and f T (xn) } are
also Cauchy sequences in B and B'; and since both of these spaces are
complete, there exist vectors x and yin B and B~ such that Hxn - xll ---+ 0
and l T(xn) - YH ---+ 0. Our assumption that the graph of T is closed
in B X B~ implies that (x,y) lies on this graph, so T(x) ::= y. The
completeness of B1 now follows from

llxn - xll = llxtJ - xll + l T(xn

1

- x) H=
=

xH + l T(xn) - T(x) [I
llx1) - xii + llTCxn) - Yll ~ 0.

llx"

The closed graph theorem has a number of interesting applications

to problems in analysis~ but since our concern here iH m.ainly \Vith matters
of algebra and topology, we do not pause to illustrate its uses in this
direction . 1
Problems

Let a Banach space B be made into a Banach space 8 by means of &

new norm, and show that the topologies generated by these norms
a.re the same if either is stronger than the other. _
2.. In the text we used Theorem B to prove the closed graph theorem.
1

1.

3..

Show that Theorem B is a consequence of the closed graph theorem.

Let T be a linear transformation of a Banar.h space B into a Banar.h
space B 1 If {f.d is a set of functionals in Bt* whirh separates the
vectors in B' ~ and if f,,T is continuous for each f i~ prove that T iH
continuous.

51. THE CONJUGATE OF AN OPERATOR

We shall see in this section that each opera.tor Ton a normed linear
space N induces a corresponding operator~ denoted by T* and called the
conjugate of TJ on the conjugate space N*. Our first task is to define T* ~
and our second is to investigate the properties of the mapping T ~ T*.
We base our discussion on the follov-.. ing theorem.
Theorem A (the Uniform Boundedness Theorem). Let B be a Banach
apace and N a normed linear space. If {'I\ l is a non-empty set of con1

See Taylor

[41~

pp.

181-1851~

240

Operators

tinuOU$ linear transforrnations of B inw N with the -property that {Ti(x) }

is a bounde.d sub8et of N for each vector x in BJ then f!I Till} is a bounded set
of numhers 11~ that is, {T1d is bounded as a subset of Gl(B,N)~
PROOF.

For each positive integer nJ the set

F~ = {x :x

s Band If Ti(x)Jl

n for all i}

is clearly a closed subset of B 1 and by our assumption we have

Since B is complete 1 BaireJs theorem sho-v-.Ts that one of the Fn 1s, say
F nfJ~ has non-empty interior~ and thus contains a closed sphere So with
Center Xo and radius ro > 0.. rfhis Says, in e ffectJ that each veetor in
every set Tt(So) has norm less than or equal to no; and for the sake of
brevityt "\\Te express this fact by writing II Ti(~S'o) I < no. It is clear that
S 0 - x o is the closed sphere with radius r o centered on the origin 1 so
(So - xo)/ro is the closed unit sphere B+ Since Xo is in So, it is evident
that ll"Ti(So - xfr) H ~ 2no. This yields IJ Ta:(B) H < 2no/ro, so fl T,JI <
2n 0/ r 0 for every i, and the proof is complete.

This theorem is often called the Banach-Steinhaus theorem 1 and it

has several significant a p plic e.tio ns to an n.l ysis. See~ for e:xa m ple,
Zygmund [46, vol. 1, pp~ 165--168] or Gal [11]. For the purposes we
have in view, our main interest is in the following simple consequence of it4
Theorem

B. A non-empty subset X of a norm..ed linear space N -la bounded

f(X) is a bounded Bet of numbers for each f in N.

PROOF~ Since lf(x) j < Bill llxl~, it is obvious that if Xis bounded, then
f (X) is also bounded for each f.
In order to prove the other half of the theorem1 it is convenient to
exhibit the vectors in X by writing X = {x;:}. We now use the natural
imbedding io pass from X to the corresponding subset {F~J of N**~
Our assumption that f{X) = {f(xi) ~ is bounded for each f is clearly
equivalent to the assumption that {F:;.(/)I is bounded for each f, and
since N is con1plete, Theorem A shows that IF.1=J is a bounded subset
of
We know that the natural imbedding preserves norms, so X is
evidently a bounded subset of N.
8

N4

We now turn to the problem of defining the conjugate of an operator

on a normed linear apace N.
Let L be the linear space of all scalar-valued linear functions defined
on N. The conjugate space N* is -clearly a linear subspace of 4 Let
T be a linear transformation of N into itself which is not necessarily
continuous. We use T to define a linear transformation T' of L into

B-ana ch Spaces
itself, as follows.

241

If f is in L'I then T'(f) is defined by

(1)

!T'(f)J(x) = f(T(x)).

We leave it. to the reader to verify that T 1 (f) actually is linear as a func ..
tion defined on N, and also that 1 is linear as a mapping of L into itself.
The following natural question now presents itself. 1-7 nder what
circumstances does T, map N* into lv*? Thi:s question has a simple and
elegant answer: T' (N*) C N * <=> '~f i8 continuous. If "~e keep Theorem
B in mindJ t.he proof of this statement is very easy; for if ~g is the closed
unit sphere in N 1 then '1' is continuous<=> T(S) is bounded 8 f(T(S)) is
bounded for each fin 1V* ~ CT' (f) ](S) is bounded for each f iu .i:V* ~ T'(f)
is in N* for each f in N* .
We now assume that the linear transformation T i~ continuous and is
therefore an operator on N. The preceding dcve1opmcnts allow us to
consider the restriction of T' to a mapping of N* into itself.. We denote
t.his restriction by T*~ and we call it the conjugate of T. The action of
T* is given by
1

'

[T*(f)J(x) = f(T(x)),

(2)

in which-in contrast to (1)~/ is understood to be a functional on N,

and not merely a scalar-va!ued linear function+ T* is clearly lincar 1 and
the following computation shows that it is continuous:
If

T*ll

= sup UI T(f) fl: JIJH < 1}

= sup {llT*(f)](x)j: ll/U and rf xlj ~ l l
= sup f l/(T(x))I: llfrl and llxll < 1}
< sup f II/ii l Tll rlxll ~ Hfn and !lxH < I J

:5 If Tit

Since HTU = sup {fIT(x) H~ fIx [I < 1}, we see at once from Theorem 48-B
that equality holds here 1 that is, that
If

T*ll

l! Tll

(3)

The mapping T ~ T* is thus a norm... preserving mapping of <B(N) into

m(N*).
We continue in this vein by observing that the mapping T ~ T*
also has the following pleasant algebraic properties:
(aT1

+ ~T,.)* =

+ PT2*~

= T2*T1,
I* = I.

(T1T2)

and

aT1*

(4)
(5)
(6)

The proofs of these faets are easy consequences of the definitions. We

illustrate the principles involved by proving (5)4 It must be shown that

242

Operators

(TJT~,) *(f) =

(TzT1 *) (f) for each fin N*, and this means that
[(Tl T2)*(/)J (x) = [(1\~*T 1*) (f)J (x)

for each fin N and each .:t in N.

r(1\T2)*(j)](x) = f((T1Ti)(x))

A simple computation nuv..-

shO\YS

that

f(1,1(1't2(x)}) = [Tl*(j)]('.Fi(x))
= [ 1' 2 * ( 7\ ( f) ) ] (x) = [('.t 2 * T 1 *) (f) ] (x) .

It may be helpful to the reader to have the

the results of this discussion.

follo"Vl-~ing

summary of

Theorem C. If 1T is an. opera.tor on a normed linear space i.\r, then its

conjugate T* defined by Eq. (2) is an operator on
J and the tnapping
T ~ 'J,+ is an i~'fometric isomurphis1n of ffi(N) into ffi(ftl*) u~hich reverses
produets and preseroes the identity lransform ation.

J.v

1...,he general significance of the ideas developed here ran be understood 011ly in the light of the theory of operators on IliJbert spaccsr
Some preliminary comments on these matters are given in the introduction to the next. r hapter
+

Problems

1.

2~

Let B be a Ranach spat,e and N a normed linear space. If {T l is a

sequence in ffi(B~/v"') such that.. 1,(x) ;:::: lim T ~(x) exists for each x in
B 1 prove that Tis a continuous linear tran~forn1a.tion.
J_.eti 'J1 be an operator on a normed linf'ar 8pR{'P N. If Pl i~ considered
to bP part of N** hy ineans of the uat.ura] imbedding, shcnv that. 7, **
is an xtPn!-iion of 1' ~ ()bservc that if .i.V is rcftcxi ve thr.n 1 ** ==== T
Let T l)(~ au operator on a 1~anach spare B~ Shov.. that. 1~ has an
inverse 1'- 1 t j '1 * has an inverse (T*)- 1, and that iu this case
(7,*)- = (T-t)* .
1)

3.

CHAPTER TEN

Jlil/Jert Spaces
One of the principal app1ications of the theory of Banach algebras
developed in Chaps+ 12 to 14 is to the study of operators on Hilbert
spar.es. Our purpose in this chapter is to present enough of the elementary theory of Hilbert spaces and their operator~ to provide an
adequate foundation for the deeper theory discussed in these Jatcr
chapters.
We shall see from the formal definition given in the next section that
a IIilbert space is a special type of Banach space, one \\Thich possesses
additional structure enabling us to tell when 1Nf0 vectors arc orthogona1
(or perpt~ndicular) T The first part of the chapter is concerned solely "'Tith
the geometric jmplica.t.ions of this additional structure.
As "\\"'e _have said before, the objects of greatest interest in -r..onnection
with any linear spaee are the linear transformations on that space+ In
our treatment of llanach spacest we took advantage of the metric structure of such a space by focusing our attention on its operators~ A
Banach space, however, is still a hit t<>o genera] to yield a really rich
theory of operators+ One fact that did emPrgP., 'vhich is of great significance for our present work~ is that \vith each operator T on a Banach
space B there is associated an operator T* (its r,onjugatc) on the r.onjugate space B. We shall sec bclo\v that onP of the central properties of
a Hilbert space H is that there is a natural correspondence between II
and its tonjugate space H*. If T is an operator on H, this correspondence makes it possible to regard the conjugate T* as acting on JI i tsc1f
(instead of H*), vrhere it can be compared "ith T. These ideas lead to
the concept of the adjoint of an operator on a Hilbert space, and they
2A3

244

Operators

make it easy to understand the importance of operators (such as selfadj oint and normal operators) which are related in simple ways to their
adjoin ts .

52 .. THE DEFlNITtON AND SOME SIMPLE PROPERTIES

The Banach spaces studied in the previous chapter are little more
than linear spaces provided 'With a reasonable notion of the length of s.
vector. The ma.in geometric concept miBsing in an abstract space of this
type is that of the angle between t~~lo vectors. The theory of Hilbert
spaces does not hinge on angles in general, but rather on some means of
telling when two vectors are orthogonal.
In order to see how to introduce this concept 1 we begin by considering
the three-dimensional Euclidean space R 3 A vector in R:t. is of course
an ordered triplex = {x1, x2j X3) of real numbers, and its norm is defined
by
ll x 11 = {[xi I1
l x j J ~ + Ix! [2) w.

In elementary vector algebra~ the inner product of

Y = (y1, Y2~ y3) is defined by
(x,y)

=::

X1Y 1

X2Y2

;t

and another vector

+ XsY3,. 1

and this inner product is related to the norm by

(x,x) ~

Uxll ~-

We assume that the reader is familiar with the equation

(x,y)

llxll llYll cos 6,

where fJ is the angle between x and y, and also with the fact that x and y
are orthogonal precisely when (x,y) = o~
Most of these ideas can readily be adapted to the three-dimensional
unitary space C 1 ~ 14~or any two vectors x = (x1, Xi, xa) s.nd y = (y1, Y2J ya)
in thls space 1 we define their inner product by
(x,y) ~

X1Y1

+ X2Y1 + X&:V3

(1)

Complex conjugates are introduced here to guarantee that the relation

(xix) ;;

!lxU2

remains true~ It is clear that the inner product defined by ( 1) is linear

as a function of x for each fixed y, and is also conjugate-symmetric, in the
The term dot produ.d and the notation ~ 11 are used in most introductory
treatments of v ectora,
1

Hi rbert Sp aces
sen~ that (x,y)

245

(y,x)~

In this case~ it is no longer possible to thlnk

of (x,y) as representing the product of the norms of x and y and the
cosine of the angle between them, for (x,y) is in general a complex num ...
ber4 Nevertheless, if the c.ondition (x,y) = 0 is taken as the definition
of orthogonality, then this concept is j ust as uscf ul here as it is in the real
case.
With these ideas as a backgroundJ we are now in a position to give
our basic definition. A Hilbert space is a complex Banach space v.~hose
norm arises from an inner productJ thn. t is, in which there is defined a
complex function (x~y) of vectors x and y with the following properties:
(I) (ax +Py, z) = a(x,z)
fJ(y,z);
(2) (x 1 y) = (y,x);
(3) (xJx) = Hxll :t
It is evident that the further relation

(xJ ay

+ ~z)

= a(x,y)

+ S(x,z)

is a direct con.sequence of properties (1) and {2)~

The reader may wonder why we restrict our attention to complex
spaces. Why not consider real spaces as well? As a matter of fact,
we could easily do so, and many writers adopt this approach~ There
are a few places in this chapter where complex scalars are neces~ary, but
the theorems involved are not crucial, and we could get along v.-ith real
scalars without too much difficulty. It is only in the complex caseJ
however, that the theory of operators on a Hilbert spaee as~umes a really
satisfactory form. This will appear with particular cla~ity in the next
chapterJ where we make essential use of the fact that every polynomial
equation of the nth degree with con1plex coefficients has exactly n
complex roots (some of which, of course 1 may be repeated). For this
and other reasons, we limit ourselves to the complex case throughout the
rest of this book.
The following are the main examples of Hilbert spaces.. In accord ..
e.nce with the above remarks, the scalars in each example are understood

to be the complex numbers.

Example 1.

The spaoe z;, with the inner product of two vectors

x = (z1,

X2i ,

Xn) and Y

=:::

(y1 1 Y2, . , y,.)

defined by

It is obviou.s that conditions ( l) to (3) are satisfied.

246

Operators

E~omple

The space

2.

= {x1

:t

x 2' ~

l~, ~~ith

' Xn '

the inner product of the vectors

~

and y ;:::; { y 11 y :2' ~

yn J ~

defined by
om)

I.

(x,y) =

XnY~

n 1

1,hc fact that this series convergcs-----and thu~ defines a -complex number-for ~ach r and y in l'l. is an easy eonHequencc of <:anehy'H inequality.
Example 3. --rhc space ?. a~sociatcd . . vith a rncn.t{ure :::.pace . Y
. "=-ith
measure rn, "~ith the inner product of t\vo functions f and g defined by

(J ,g)

J1(x) 9 ( x) d n <x) .
l

This llilbert space is of course not part of the offieial content of this book,
but \VC mention it any,vay in c.ase the reader has some knov... ledge of these
matte-rs~

As our first theoremj 've prove a futuJamental relation kno'\vn as the

A-c:;chwarz inequality.
Theorem A.. If x and y are any tu,o t.'ectors in a
Hlberl space, then I(xJy)I < llxH l~Y~!
PHOOF.
\V'hen y = 0, the result is clear, for both
ti id PH vanish.
\\,..hen y ~ 0, the iueq uality is equiv~
a hint to I(x,y/ liYIDI < llxl~.
may therefore confinP our attention to proving that if [lyl] = l, then
'"Te have ~ (x,y)I < :i!rli for all x. This is a direct
conHequencc of the fact that

' re

Fig
afi. Sch ~"!:Lr z' H
inc4 uality.

0 ~ ~Ix - (x~y)y\! 2 = (x - (x,y)y, x - (x,y)y)

= ( x ~x) - (x, y) ( x, y) - (r, y) ( x ~ y)
x, y) (x y) (y 1 y)
= (x 1 x) - (x,y)(x,11) = l\xll 2 - l(x,y)]!.

+{

An inspection of :Fig. 36 will reveal the geometric motivation for this

computation.
lt follO\\TS easily from Schvl'arz's inequality that the inner product
in a. H ii he rt ~pace is jointly continuous:
x and

Xn ----;

Yn

---+

===?

(xn,Yn) - t (x,y).

1,o prove this 1 it suffices to observe that

I(xn,Yn)

- (x,y)J

1(Xn~Yn)

- (Xn 1Y)

+ Ctn~Y)

(xJy)I < I(xn,y")

+ I (x~~y) ~ (x,y)] ~ I(x~, Yn ~ y)!

+ I (x" - X1 y)I ~ \lxn I! llY" - yl[ + llxn

(:t1i,Y)I

11

llYll

Hirbert Spaces

247

A well-known theorem of elementary geometry states that the sum

of the squares of the sides of a parallelogram equals the sum of the squares
of its diagonals. This fact has an analogue in the present context, for
in any Hilbert space the so-called parallelogram law holds:

This is readily proved by writing out the expression on the left in terms
of inner products:

l!x + Yll:t + llx - Yll 1 =

=

(x,x)

+ (x,y) +

+ Y1 x + y) + (x

(x

(y,x)

(y,y)

~ y, x -

(x,x) - (x,y)
= 2(x,x)
2(y,y)

y)
(y7x)

+ (y,y)

= 2Hxll 2

+ 2llYll

The parallelogram law has the follovnng important consequence for

our work in the next section.
Theorem Bt A closed convex subset C of a llilbert S]JUCe H contains a
u niqm vector of smallest norm~

Tlle recall from the definition in Problem 32-5 that since C is

convex, it js non-empty and contains (x + y)/2 whenever it contains
x and y.. Let d = inf l l!x H: x E CJ. There clearly exists a sequence
{x.. } of vectors in C such that llxn[I ~ d~ By the convexity of C,
(xm + Xn)/2 is in(] and [l(x. + x~)/2B > d, SO llxm + Xn" > 2d.. Using
the parallelogram law, we obtain
PROOF.

llxm -

x~ll 2

= 2l]xmlP' + 2l[xnll 2

l~x.

< 2!lxml]' + 2llxnll! -

+ Xmil 1

4d!;

and since 2llx..ll 2

21f x"Jl 2 - 4d2 ---:; 2d1 2dt ~ 4d 2 = ot it follows that
fXn} is a Cauchy sequence in C. Since His complete and C is closed,
C is complete 1 and there exists a vector x in C such that x~---+ x.. It is
clear by the fact that l!xll == IUim Xnfl = lim llxnH = d that xis a vector
in C with smallest norm. To see that x is unique~ suppose that x' is a
vector i~ C other than x "Thich a1so has norm d. Then (x + x')/2 is
also in C, and another application of the parallelogram law yields

+ x'
2

<
~rhich

llxlP
2

HxH 2
2

ilx'r[ 2
2

--..

x - x'
2

+ Hx2 H -d,
1

contradictH the definition of d.

1,hc parallelogram law has another interesting application, which

depends on the fact that in any Hilbert space the inner product is related

2..t 8

Operators

to the norm by the following identity:

4(x,y)

= llx + Yrl' - HJ:: - y[I 2 + illx

+ iy!f 2 -

il[x - iylf 1.

(2)

This is easily verified by co nvcrting the expression on the right into inner
products.
If B is a complex Banach space whose norm obeys the parallelogram lawt and if an inner product is defintrl on B by (2) 1 then B is a
Hilbert space.
PROOF~ All that is necessary is to make sure that the inner product
defined by (2) has the three properties required by the definition of a
Hilbert space~ 'fhis is easy in the case of properties (2) and (3). Property (l) is best trr.atcd by splitting it into two parts:
Theorem

C.

(x

+y

z) = (x,z)

+ (yjz),

and (ax,y) = a(x,y)~ The first requires the parallelogram law, and the
second follovrrs from the first. We ask the reader (in Problem 6) to ~rork
out the detai1s.
1.,his result has no implications at all for our future \vork. However,
it does provide a satisfying geometric insight into the place Hilbert
spaces occupy among all complex Banach spaces: they are precisely those
in which the parallelogram la'v is true4

Problems
1..

2.

3.
4.

5.

6..

Show that the series \vhich defines the inner product in Example 2
is convergent.
The Hilbert cube is the su bsc t of l 2 consisting of all sequences

such that jxli.I ::; 1/n for all n. Show that this set is compact as a
subspace of 12.
l1~or the special Hilbert space li, use Cauchy's inequality to prove
Sch '\va rz 's inequality.
Sho\V that the parallelogram law is not true in z~ (n > 1)4
In a Hilbert space, show that if Uxll = 11YH ::::; 1, and if= > 0 is given,
then there exists & > 0 such that Jl(x + y)/2[1 > I - a~ IJx - Yll
< E~ A Banach space with this property is said to be uniformly
conz:ex4 See Taylor [41, p. 231].
Give n detailed proof of Theorem C.

Hilbert Spaces

249

53, ORTHOGONAL COMPLEMENTS

Two vectors :c and y in a Hilbert space H are said to be orthogonal

(written x .1 y) if (x,y) = 0. The symbol ..l is often pronounced uperp."
Since (x,y) ~ (y~x), we have x .l y <=> y .l x. It is also clear that x .l 0
for every x~ and (xJx) = llxll ~ shows that 0 is the only vector orthogonal
to itself~ One of the simplest geometric facts about orthogonal vectors
is the Pythagorean theorem:
x .l Y-::=}

11x

+ Yll = Jlx - Yll

1

== l~xlf 2

+ flYll

A vector x is said t.o be orthogonal to a non-empty set S (written x .l S}

if x .l y for every y in S, and the orthogonal complement of S-denoted
by Sj_-is the set of all vectors orthogonal to S. The follofilng statements are easy oonseq uences of the definition:

{OJ J.

H; HJ_ ;::;: {O};

Sr1Sl.~{OJ;
S1CS2~S11- 282..L;

S 1- is a closed linear subspace of H ~

It is customary to write (S 1-) 1- in the form S -1 1-. Clearly, S C S ..L J_.

Let M be a closed linear subspace of H. We know that kl 1- is also
a closed linear subspacc 1 and that M and M .L are disjoint in the sense
that they have only the zero vector in common~ Our aim in this section
is to prove that II = M @ M 1-, and each of our theorems is a step in
th is direction.

Let M be a closed linear subspace of a Hilbert space H, let

.x be a vector not in Mj and let d be the distance from x to M~ Then there
Theorem A.

cxiRla a unique vector Yo in M 8Uch that llx - Yoll = d.

PROOF.
The set C ;:::= x
Mis a closed convex set, and dis the distance
from the origin to C (see Fig. 37). By Theorem 52-B, there exists a
unique vector zfJ in C such that Hzol[ = d. The vector y 0 = x - zo is
easi1y seen to be in M ~ and II x - Yo I~ = [lzo 11 = d~ The uniqueness of Yr
follows from the fact that if Y is a vector in M such that Y1 - yo and
11x - Y1ll = d~ then z1 ~ x - Yi is a vect-0r in C such that z1 - zo and
Uz1!I = d, which contradicts the uniqueness of zo.

We use this resllJ.t to prove

If Mis a proper closed linear subspace of a Hilbert space H,
then there exists a non-zero vector zo in H such that z 0 .l A! .
PROOF.
J.,et x be a vector not in M, and let d be the distance from x to M.
By Theorem A, there exists a vector Yo in M such that l~x - Yo~I = d.
Theorem B..

250

Operators

We define z~ by zo = x - Yo (see :F'iga 37), and we observe that since

d > O~ zo is a non-zero vector4 We conclude the proof by sho,ving that
if y is an arbitrary vector in M~ then z 0 l. y. },or any scalar a~ we have

!lzo - ayll
so

and

If we put

o:

ay)ll > d = l]zolL

Hzo - Y~l 2 - [lzoll 2 > 0
a(zo,y) ~ a(zo,y) + Jal 2 11Y 1; 2 > 0.
~

llx -

(yo+

(1)

/3(zo.,y) for an arbitrary real number {3, then (1) becomes

-2Pl (zo,y)I i

+ f3 i(zo,Y)! 2llYll
1

> 0.

If we now put a= l(zo,y)~ 2 and b = llYil 2, we obtain

- 2/3a

+ p~ab > 0

f3a(/jb - 2)

so

for all real

P~

However, if a

>

(2)

> O,

then (2) is obviously false for al

sufficiently s.mall positive fl. We see
from this that a = 0 1 which means
that zo ..l Y~

This proof of Theorem B may

strike the reader as being excessively
dependent on ingenious computations.
If so T he will be pleased to learn that
the ideas developed in the next section can be used to provide another
proof which is free of computation.
Fig. 37
In order to state our next theorcmT we need the following additional
concept. 1\vo non-empty subsets J..r.;1 and S,, of a Hilbert space are said
to be orthogonal (written 8 1 ..l S,:) if x l. y for all x in 81 and y in 82.
Theorem C. If Mand N are cl08ed linear subspaces of a Hilbert E!]Jace JI
su.ch that M 1- N, then the linear subspace M
N is al!Jo closed.

Let z be a limit point of M + N. It suffices to show that z is

in M + N. There certainly exists a sequence {znl in M + N such that
~---+ z. By the assumption that M l_ N, we see that M and N arc
disjoint, so each Zn can be written uniquely io the form Zn = Xn + Yn,
where Xn is in Mandy$ is in N~ The Pythagorean theorem shows that
llz. ~ z~H 2 "= llXm - Xnll' + l]y,,. - yn,]1 2 , so {Xn} and {Yn} are Cauchy
sequences in M and N. M and N are closed, and thcref ore complete, so
there exist vectors x and y in Jf aod N such that Xn.---+ x and Y~--+ y.
Since x + y is in M + N, our conclusion follo'"'s from the fae.t that
z ~ lirn Zn = lim (x~ + y,,.) = lim x~ + lim Yn ~ x + y.
PROOF4

The way is now clear for the proof of our principal

theorem~

Hilbert Spaces

2.Sl

Theorem D. If M is a closed linear subspace of a Hilbert space H, then

H = M $ M1-.

Since M and M 1- are orthogonal closed linear subspaces of H,

Theorem C shows that M + M _i_ is also a closed linear subspace of H.
We prove that M + M .L equals H. If this is not so, then by Theorem B
there exists a vector zo ~ 0 such that zo l. (M
M ..1 ). This non-zero
vector must evidently lie in M .L r'i M .L .1; and since this is impossible, we
infer that 11 == M + M ..L. To conclude the proof, it suffices to observe
that since M and M 1- are disjoint, the statement that H = I'll
M _i_
can he strengthened to H = M M J..
PROOF..

The main efiect of this theorem is to guarantee that a Hilbert space

is always ricl;t in projections. In fact, if M is an arbitrary closed linear
subspace of a Hilbert space H, then Theorem 50-D shows that there
exists a projection defined on II whose range is M and whose null space
is M 1-~ This satisfactory state of affairs is to be contrasted with the
Bituation in a general Banach space 1 as explained in the remarks following
Theorem 50-D.

Problems

1.

2.

1: Sis a non-empty subset of a Hilbert space, show that S 1- = S 1.. 1- 1-.

If M is a linear subspace of a Hilbert space, show that M is closed
{::}M = M..L.L~

3.

A.

If S is a non-empty subset of a llilbert space ll, show that the set

of all liflcar combinations of vectors in Sis dense in H ~ S l. = {O}.
If Sis a non-empty subset of a Hilbert space H, show that S 1- 1- is the
closure of the set of a11 linear com bina tio ns of vectors in S. Tiiis is
usually expressed by saying that S 1- .L is the smallest closed linear
subspace of H which contains S4

SA. ORTHONORMAL SETS

An orthonormal set in a Hilbert space His a non-empty subset of H
which consists of mutually orthogonal writ vectors; that is, it is a non
empty subset {e~} of H with the following properties:
( 1) i ~ j ~ ei .l e1 ;
(2) !lc,11 = 1 for every i.
If H contains only the zero veL~or, then it has no orthonormal sets. If
H contains a non-zero vector x~ and if we normalize x by considering
e ~ x/IJxJI, then the single-element set {e J is clearly an orthonormal seta
More generally~ if ~xi} is a non-empty set of mutually orthogonal non-

252

Operators

zero vectors in H~ and if the x/s are normalized by replacing each of them
by ej = Xi/llx,:ll, then the resulting SE~t. fed is an orthonormal set.
Example 1. The HUbset {ei, _et, . . . , en} of l;, ":rhere Ci is the n-tuple
'"Tith 1 in the ith place and O's clse\v here, is evidently an orthonormal
set. in this space.
Example 2. Similarly, if en is the sequence \vith 1 in the nth place and
OJ s clse\vhere ~ then {e1, e2,. . . . , en, . . J is an orthonormal set in l2~
L

At the end of this section, \\re give some additional ex.rimples taken
from the field of ana}vsiH+
....
Every aspect of the theory of orthonormal sets dependr.; in one way
or another on our first theoremr
Theorem A. T"'et {e1, ez, ~ .. T e~} be a
space II. If x is any vector in H~ then

finite orthonormal set in a Hilbert

I (x,e,)l 2

< If xiii;

(1)

1-1
'l

/urtherJ

x -

(x,ei)e.: 1-

(2)

Ci

i= l

for ear.h j.
PRoott~.

The inequality ( l) follows from a computation similar to that

used in proving Schu~arz~s inequality~
n.

~ (x~ei)ei

< r1 x -

11

l=l

- (r -

(x~e,)ei~ x ~ ~ (x,ei)ei)

l=l

~ (x,x) -

j=l

I
I

(x,eJ(x,e1) ~ ~ (x,ej) (xl~i)

i=-l

1"=1

I.:-1

~ (x,ei) (x,e;) (ei~ei)

j::::ol

1l

- llxlf 2

I (x~e.:)l 2T

i=J

'fo conclude the proof, v-.Tc observe that.

~

( x - ~ {x 1 ei)e1}ej) ~ (x,eJ) il

(x,eD (ei,ei) -

(x_,eJ) - (x,eJ) = OJ

i=l

from which statement (2) iolloVt;S at once.

1~hc

reader should note that the inequality (I) can be given the
following loose but illuminating geometric interpretation: the sum of the
squares of the components of a. vector in various perpendicular directions

Hilbert Spaces

253

docs not exceed the square of the length of the vector itself. Thls is
usually ca lied Bessel's i.rn.equality., though., as we shall see be1o w, it is only
a special case of a more general inequality with the same namer In a
similar vein, rclat.ion (2) says that if we subtract from a vector its components in several perpendicu1ar directions, then the result has no component left in any of these directions.
Our next task is to prove that both parts of Theorem .A. gencraliie
to the case of an arbitrary orthonormal set. The main problem here
is to shov.r that the sums in (1) and (2) can be defined in a reasonable way
'vhen no restriction is placed on the number of e/s under consideration~
The key l-0 this problem Hes in t.he following theorem.
Theorem B. If {ei} is an orthonormal set in a llilbert space II~ and if x
is any vector in HJ then the set S = {Ci; (x~.e-i:) - 0} is either empty or
countable.
PROOF~ },or each positive integer n, consider the set

By Bessel's incq ua1i ty ~ i..~n con tu ins at most n - 1 vectors.

cl usion nov{ f ollo\,TS from the fart that 1~ = U :_ 1 j.S n

The con ...

As our first appHration of this result, we prove the genera.] form of

Bessel's jncquality~
Theorem C (Besser' s lnequality)t
Hilbert space H, then

If [ed is an orthonarmal set in a

(3)

for every vector x in H.

PROOF.
Our basic obJigation here is to explain what is meant by the
sum on the left of (3). Once this is clearly understood, the proof is
easy.. As in the preceding theorem' we '"'~rite == {et: (x~et:) " 0 I If
is empty, we define :ti (x,e.-)j 2 to be the number O; and in this ca.se, (3) is
obviously true4 We now assume that Sis non-empty, and we see from
Theorem B that it must be finite or countably infinite. If 8 is finite, it
can be "\\. . titten in the form S = {e 1J e2 , ~ , e~] for some positive
integer n. In this rase, we define ];i (x~e~)i 2 to be 2:;__ 1 [{x,e,)r 1 , which is
clearly independent of the order in which the elements of S are arranged.
The inequality (3) now reduces to (1)~ lNhich has already been proved.
All that remains ]s to consider the case in which S is countably infinite.
Let the vectors in S be arranged in a definite order:

S = f e1 e2,
11

~ . ~

:11 eA:11

.. }

By the theory of absolutely con vcrgent series, if x:=1 j (x~en) i2 con verges,
then every series obtained from this by rearranging its terms also con-

254

Operators

vergest and all such series have the same sum. We therefore define
l:[ (x,ei)I t to be ~:- 1 I(x,eliL)l1, and it follows from the above remark that
:E !(x,e;:) I2 is a non-negative extended real number which depends only
on 8 1 and not on the arrangement of its vectors. We conclude the proof
by observing that in this case~ (3) reduces to the assertion that
OD

I (xten)l 2 ~

tlxl! 2 ;

(4)

n.-1

and since it follows from (1) that no partial sum of the seriPs on the left
of (4) can exceed l[xll 2, it is clear that (4) itself is true.
The second part of Theorem A is generalized in essentially the
same way.
If fed is an orthonormal set in a /filbert space H, and
is an arbitrary vectOT in H ~ then

Theorem Dt

if x
(5)

for eaih j.
PROOF.

As in the above proof, we define I: (x,ei)et.~ for each of the various

eases, and we prove {5) as we go

along~

We again write

S = fe,< (x,e,) #- 0}.

When S is emptyJ we define l;(xie,.)ei to be the vector 0, and we observe
that (5) reduces to the statement that x - 0 = xis orthogonal to each eiJ
which is precisely what is meant by saying that S is empty. When S
is non-empty and finite, and can be written in the form

we define 2; (i:,e,)e.: to be 2;:m:o 1(x~et:)ei; and in this case~ (5) reduooe to (2),
which has already been proved.
By Theorem B, we may assume for the remainder of the proof that
Sis countably infinite4 Let the vectors in S be listed in a definite order:
s : : ; t ei, e2, ~ ~ eHT } we put s~ = ~~-=.J (xfei)ei, and "'~e note
that form > n we have
1

!Is. -

s~H

;;;

"

i=~+I

(x~et:)e,:tl ~
2

(x~ei) I2.

i:aq1+l

Bessel's inequality shows that the series z;... 1 !(x,en) I2 converges, so

{Sn} is a Cauchy sequence in II; and since II is complete, this sequence
converges to a vector s, which we write in the form s = :2:; 1 (x~en)en.
We now define :!;(x,ei).e" to be ~=-l (x,e~)en~ and--dcferring for a moment
the question of what happens when the vectors in S are rearranged-we

Hilbert Spaces

255

observe that (5) follows from (2) and the continuity of the inner product:
(x - ~ (x,Bi)e1, e1)

(x -- st ei) ~ {x Jei) - (s,e3) = (x 1e,) - (lim sn, e;)

= (x ,ej) - lim (s&~e,.) = (x,e;) - (xJei) = 0.

A.ll that remains is to show that this definition of ~(x~e")et: is valid, in the
sense that it does not depend on the arrangement of the vectors in S~
Let the vectors in S be rearranged in any manner:

i/1,/2,

,f~,

"

.}.

We put s: = ~~ 1 (x,fi)fi, and we see--as above-that the sequence

{s:} converges to a limits', 'vhich vte 911'ite in the forms' = ~: 1 (x,f~)fn
We conclude the proof by showing thats' equals s. Let 'f > 0 he given,
and let no be a positive integer so large that if n > no, then l!sn - sll < ~~
11 s: s' 11 < If:, and ~~ 0 + 1 I (xtei) l 2 < E 2 . For some posit.i ve in tcgcr
m 0 > no, all terms of Sn~ occur among those of s~ 01 so s~ ~ Sn~ is a finite
sum of terms of the form (xreDc,: for i ::=::: no
1, no
2, .
This
yields lls~o - 8n~ll 2 < };;:no+l '(x,e~)[ 2 < ~ 2 , so l!s:o ~ Sn() H< ~ and

!Is'
Sinee

~
i

sl1 < Hs' - s:c-li + lls:o - 8no!I + l!sno - sll < t: + tr + E

is arbitrary 1 this

sho~,..s

3~~

thats' == s.

Let H be a non-zero Hilbert space, so that the class of all its orthonormal sets is non-empty. 1-~his class is clearly a partially ordered set
with respect to set inclusion. An orthonormal set {ei i in H is said to be
complete if it is maximal in this partially ordered sett that isJ if it is
impossible to adjoin a vector e to {ed in such a way that {eiJe J is an
orthonormal set which properly contains ~ ei} .
Theorem E.

Every non-zero Hilbert space contains a complete orthcmortnal

aet.
The statement follov_.,-s at once from Zorn 1 s lemma, since the
union of any chain of orthonormal sets is clearly an upper bound for the
chain in the partially ordered set of all orthonormal sets.

PnooF.

Orthonormal sets are truly interesting only when they are complete.
The reasons for this arc presented in our next theorem.
Theorem F. Let H be a Ililhert space, and let {ed be an ortho-normal set in
H. Then the following conditions a re all equi()alent to one another:
(1) le l is complete;
(2) x ..l {e..: l :=9 x = 0 ;
(3) if xis an arbitrary vector in H~ th-en x ~ ~(x,ei)ei;
(4) if x is an arbitrary vector in H, then ~lxll 2 = ~I (x,edl 2 .
PROOF.
We prove that each of the conditions (1), (2), and (3) implies
the one following it and that. (4) implies (1).

256

Opera tors

(2).. If {2) is not true, there exists a vector x ~ 0 such that

x j_ fed~ We now define e by e = x/nxfl, and we observe that {ei,el
is an orthonormal set which properly contains {ei~. This contradicts t.hc
completeness of {ei l .
(2) ~ (3). By 1~heorem D, x ~ l: (x,e,;)ei is orthogonal to {ei} , so
(2) implies that x ~ l:;(x,e;:)e~ = 0, or equivalently 1 that x = l:(x,ei)ei.
{3) ==> (4). By the joint continuity of the inner product, the expression in (3) yields
(1)

::=::}

tlx112 =

(x,x) ;::; (:Z(xtei)eiJ %(x,e1)e;)

l;(x,ei) (x~ei)

=:::;

)1

~j (xte .

(4) ~ (1)4 If {cd is not complete, it is a proper subset of an

orthonormal set fe..:1 e}. Since e is orthogonal to all the r"'~s~ (4) yields
lle!l 2 = ~I (e,e;:) I2 0)1 and this contradicts the fact that e is a unit vector
=:::;

There is some standard terminology which is often usPd in connection

with thiB theorem. Let {ei} be a complete orthonormal set in a Hilbert
space H, and let x be an arbitrary vector in H. '"fhe numbers (x,ei-) are
called the Fourier coefficients of x, the expression x = ~(x,ei)e;: is called
the Fourier expansion of xJ and the equation [lxH'
~j (x,ei)j 2 is called
Paruval 1 s equation-all with respect to the particular complete orthonormal set {e,;} under consideration. These terms come from the
classical theory of Fourier series, as indicated in our next example4
=:::;

Example 3. Consider the Hilbert space L~ associated \vith the measure

space [01 2..-], where measure is Lebesgue measure and integrals ate
Lebesgue integralsL 1 This space essentially consists of all complex
functions f defined on [0,21r] which are Lebesgue measurable and squareintegrable, in the sense that

Jo

27

lf(x)j 2 dx <

co ,

Its norm and inner product are defined by

JIJll =
and

(f,g) =

(/oh lf(x)/' dx )Ji

J.

2.-

dx~

f(x)g(x)

A simple computation shows that the functions e..:7U, for

1i
1

== 0, + I, + 2,

'

In order to under.s t..an d thia and the next exam plet the reader should ha.ve some
knowledge of the modern theory of measure and integra tionr We wjsh to emphasize
once a.gain that these examp1es a.re in no way e.ssentia.l to the structure of the book,
and may be skipped by a.ny reader without the necessary background.. We
advise such a reader t.o ignore these examples s.nd to proceed at once to the discussion
of the Gram-Schmidt process.

HUbert Spaces

257

are mutually orthogonal in L2:

m n
m = n.
It f ollo\vs from this that the functions eR (n == 0, + 1, + 2, . . . ) defined
by en(X) = ef~.z;-/y'21r form an orthonormal set in l.i~. For any function
f in L21 the numbers
l ~.2!11'"

{6)
c,. = (f,e,.} = y-f (x)e-""'
dx
211" 0

a.re its classical Fourier coefficients, and Bessel 1 s inequality takes the form

I""'

t.1:=-o11;1

lc,.1 2 < / 0 " lf(x)I? dx.

It is a fact of very great importance in the theory of Fourier se rics that

the orthonormal set {en} is ~..omplete in L2. As \ve have seen in Theorem
F 1 the completeness of {en} is equivalent to the assertion that for every
fin f,;2t Bessel's inequality can be strengthened to Parsevars Ct[llation:
~

n ""' -

lcnl 2

= /

2':

lf(xW dx.

1'heorem Falso tells us that the completeneBs of tefl J is equivalent to the

statement that Pach fin L'l has a Fourier expansion:
(7)

It must be emphasized that this expansion is not to be interpreted as

saying that the series converges pointwisc to the function. The meaning
of (7) is that the partial sums of the series, that is, the vectors f n in L2
defined by
!.,..(X) --

~ I~

v 21f

.., Cl;:C
k -~

ik~ J

(8)

converge to the vector fin the sense of L2:

llf

1i

Ill~

o.

-This situation is often expressed by saying that f is the limit in the mean
of the ft/s.. We add one final remark to our de.script.ion of this portion of
the theory of liou ricr series~ Ii f is an arbitrary f unet.io n in l..12 ,vi th
Fourier coefficients cfl defined by (6)~ then BPssP1's inequality tells us that
the series E; -~ ~cnl 2 converges. 1.,he celebrated lliesz-Fischer theorem
assert.H thP converse : if ctl (n ;::;:; 0 1 + 1J + 2, ~ . . ) are given complex
numbers for which
Icn. I2 converges, then t.he re ex ists a function

i;;

-CCI

258

Operators

fin L 2 whose Fourier r.oefficients are the c~'s. If we grant the completeness of L2 as a metric space, this is very easy to prove. All that is
necessary is to use the en's to define a sequence of /,/sin accordance with
(8). The functions ein'&/ V21r form an orthonormal set~ so for m > n we
have

11/m - J~l! 2 =

t
[cA:! ~
ikl=n+l
2

(9)

By the convergence of 1:=--~ 1cn[ 21 the sum on the right of (9) can be made
as small as we please for a 11 suffi cicntly large n and all m. > nr Thls
tells us that the /1/s form a Cauchy sequence in L~; and since L'}. is complete, there exists a function fin L2 such that f n--+ f.. 1,_his function f is
given by (7)t and the cl1's are clearly its Fourier coefficients. It is
apparent from these remarks that the essence of the Riesz-Fischer
theorem lies in the completeness of 2 as a metric space.
We shall have use for one further item in the general theory of
orthonormal sets 1 namely, the Gram-Schmidt orthogonalization proces8.
Suppose that {x1 1 x~, ~ ~ . t Xn, . . . . J is a linearly independent set in a
Hilbert space H. The problem is to exhibit a constructive procedure for
converting this set into a CO rrespo ndi ng orthonormal set {Ct~ e'2 ~
en 1 J -..vith the propert~~t that for each n the linear Bubspaee of II
spanned by fe1T e2", . . , en} is the same as that spanned by {x1~ x2 1
. . . , xfl l ~ Our first step is to normalize .x 1-which is necessarily non ..
zero-by putting
r

The next step is to subtract from x2 its component in the direction of e1 to

obtain the vector x2 ~ (.x 2,e1)e1 orthogonal to eh and then to normalize
this by putting
x2 -

et

r1 X1 -

(x1~e1 )el

(X2,e1)e111

We observ~ that since x 2 is not a scalar multiple of x 1, the vector x2 . . . . .

(x2,e1)e1 is not zero 1 so the definition of e~ is valid. Also, it is clear that
e2 is a linear combination of x 1 and Xi, and that x~ iB a linear combination
of ei and e2. The next step is to subtract from xa its components in the
directions of ei and e2 to obtain a vector orthogonal to ei and e,, and then
to normaljze this by putting
X; -

ea

~ Hx 3

(.xa,e1)e1 - (xhe2)e2
(xs,e1)e1 - (X.:1,e2)e1 If

If this process is continued in the same way, it clearly produces s.n

orthonormal set {e1, er, . . . 1 et:t 1
Mth the required property.
4

Hilbert Spaces

259

Example ..4. Many orthonorrnal sets of great interest and importance

in analysis can be obtained conYeniPntly by applying the Gram-Schmidt
process to sequences of simple fu11etions.
(a) In the space L2 asso<~iated v:ith the interval [- t ~IL the functions xn (n = 0, l ~ 2,. . . . ) are linearly independent. If \Ve take these
functions to be the Xn's in the Gram-Schmidt process, then the en's are
the normalized Legendre polynomialsr
(b) Consider the space Lf). over the entire real lineT If the x/s
here arc taken to b~ the functions xne-~t.1 2 (n ~ O~ I, -2, .
then the
corresponding ents are. t.hP. normalized Herniite functions.
(c) Consider the space L2 associated with the interval (O, +co)+
If the Xn"s are the functions xne-.z (n = o~ I~ 2,
~), then the e,/~ are
the normalized Laguerre functions.
+

.) ,

Each of the orthonormal sets described in the above example -can be

sho,vn to be complete in its eorresponding Hilbert space. The analy:::;iH
involved in a detailed study of these matters is quite complieated and ha~
no proper place in the present book. The reader should recognizP,
however-and this is our only reason for mentioning the mat.eria.l in
Examples 3 and 4-that the theory of Hilbert spaces does have significant
contacts "rith many solid topics in analysis.
Problems
1.

Let (ei, e~, . .

en} be a finite orthonormal set in a Ililbcrt space
H, and let x he a vector in H. If ah a"2~
~ a'l arc arbitrary
scalarf-', sho\v that Ux - 2;~ 1 ate.:a attains its minimum value p
I

'

ai ~

(xte,)

for each i. (Hint.: expand llx ~ ~~ 1 aieil: 11 add and subtract ~7=-= 1
I Cx,e-i)l 2J and obtain an expn:1ssion of the form
1 r (x,eJ -- ail 2 in
the rcsultr)
Show that the orthonormal sets described in Examples 1 and 2 are
complete.
Show that every orthonormal set in a Hilbert space is contained in
some complete orthonormal set,. and use this fact to give an alternative proof of Theorem 5:i-B.
Prove that a llilbert space// is separable$===) every orthonormal set in
H is countable+
Show that an orthonormal set in a Hilbert space L" linearly independent~ and use this to prove that a Hilb~rt space is finite-dimensional <=> every complete orthonormal set it':l a basis+
Prove that any two complete orthonormal sets in a Hilbert space H
have the same cardinal number. This cardinal number is callr.d the

1::

2~

3..

4.

5.

6.

260

Operotors

cwtlwganal dimension of H {if H has no complete orthonormal sets, its

orthogonal dimension is said to be 0).
7. If H and H' are Hilbert spaces, prove that H is isometrically isomorphic to H' ~ they have the same orthogonal dimension. (Hint:
by Eq. 52-(2), an isometric isomorphism T preserves inner products~
in the sense tha.t (T(x),T(y)) = (x,y).)
8. Let S be a non-empty setJ and let li(S) be the set of all complex
functions f defined on S with the following t\vo properties~
( 1) is :f(s) ~ 0 I is empty or countable;
(2) 2:ff(s)l 2 < ao.
These functions clear1y form a complex linear space with respect to
pointwise addition and scalar multiplication4 Sho\v that l2(S)
becomes a HHbert space if the norm and inner product are defined by
lli~I = (l;[f(s)r 2) li and (f~g) = 2:;j(s)y(s) ~ Sho\v also that the set of
all functions defined on S 'vhich have the value I at a single point
and are 0 else~"'herc is a complete orthonormal set in l:.(S). We shall
see in the next problem that Hilbert .spaces of the type described
he.re are universal models for all non-zero Hilbert spaces.
9~ Let S = {e-.:} be a complete orthonormal set in a Hilbert space /Jr
Each vector x in H determines a function f defined on S by

and rrheorems B and C tell us that f is in l2(S)~ Sho'v that the

mapping x ~ f is an isometric isomorphism of H onto l2(S)+

55. THE CONJUGATE SPACE H*

We pointed out. in the introduction to this chapter that one of the
fundamental properties of a Hilbert space // is the fa~t that there is a
natura1 correspondence bet,veen the vect-0rs in ll and the functionals in
11*. Our purpose in this settion is to develop the features oft.his correspondence v.~hich are relevant to our work with operators in the rest of the
chapter.
Let y be a fixed vector in H, and consider the function fv defined
on H by f1J(x) = (XtY)~ It is easy to sec that fu is linear, for
f~(x1

+ xz)

(x1
x2, y)
= (xitY) + (x21Y)
;:::= f JJ(x1)
f 11(X2)
f 11 (ax) = (ax,y)
= cr.(x,y)
::== af11(X)
=

and

Hilbert Spaces

261

lturther ~ f 11 is continuous and is therefore a functional, for Schwarz's

inequality gives

!f11(x)j

~ (x,y)!

< l!xrl l1Yrr,

\vhich shows that

that is,

IJ/

11 IJ

llYll.

HfuH < 11YH.

Even more, equaJity is attained here 1

This is clear ii y = O; and if y 0, it follows from
llJ~rl

>

sup {[/11(x) I: lrx[j

1}

1~ C1~!1)

- C1~11' y)

llYI/.

To summarize, l\,.e have seen that y ---7 fv L~ a norm-preserving mapping of

Hinton. 1.,his observation \\='OUld be of no more than passing interest
if it were not for the fact that every function al in II arises in just this VtTay.

Let H be a Hilbert space, and let f be an arbitrary functional

Then the-re exists a unique. ve.ctar y in H such thal

Theorem A.

in H .

f(x)

= (x,y)

(1)

for every x in H4
PROOF
It is easy to see that if such a y exists, th en it is necessarily
unique. For ii we also have f (x) = (x~y') for alt x, then (x,y') = (x,y)
and (x, y 1 - y) :;: : ;" 0 for all Xj and since 0 is the only vector orthogonal
to every vector, this implies that y' - y
0 or y' ::::; y.
We now turn to the problem of showing that y does exist. If f = 0 1
then it clearly suffices to choose y = 0. We may therefore assume that
f F- 0.. The null space M off is thus a proper closed linear subspace of H,
and by Theorem 53-B, there exi8ts a non-zero vector Yo \vhich is orthogonal to M.. We sho'v that if a is a suitably chosen scalar~ then the vector
y ;: a:yo meets our requirements. We first observe that no matter what
a may be:1 (1) is true for every x in M; for f(x) = 0 for such an x, and
since x is orthogonal to yr,, 've also have {x,y} .;: : : 0. "fhis allows us to
focus our attention on choosing a in such a way that (1) is true for
x = Yo- The condition this imposes on a is that
4

=::::;

J(y~) = (yo,<XYo)

atlYolt 24

We therefore choose a to be f(y 0)/HYo" 2, and it follows that (1) is true for
every x in M and for x = Yo- It is easily seen that each x in H can be
written in the form x = m + /jy 0 with m in M ~ all that is necessary is
t-0 choose /3 in such a way that j(x - {Jyo) = f(x) ~ f1f(y 0) = 0, and this
is accomp1ished by putting (3
f(x)/f(y 0 ). Our conclusion that ( l) is
;::;!;

262

Operoto rs

true for every x in H now follows at once from

f(x) ~ f(m

/Jy{t)

;:::=

f(ni)

+ fJf(yo)

(m-,y)

+ fJ(yo,y)
= (m

+ /3yoJ y)

;:::=

(x,y).

This re~ult tells us that the norm-preserving mapping of H into H*

defined by
Y --:., f v~ where J. . (x)

(x~y ),

(2)

is actually a mapping of H onlo //*. It \vould be pleasant if (2) were

also a linear mapping. This is not quite true 1 however, for
and

(3)

It is an easy coMequcnce of (3) that the mapping (2) is an isometry, for

II!~ - fl'H = ~lfz~l1 = llx - Yll. We state several interesting addit.iona.l
facts about this mapping (and v.-hat it enables us to do) in the problems,
and "Te leave their verifi~..ation to the reader. It should be remembered,
ho,,Tever 1 that the real significance of this entrre circle of ideas lies in its
influence on the theory of the operators on H. We begin the treatment
of these matters in the next section.
Problems
1..

Verify relations {3).

2.

Let H be a Hilbert space, and show that H* is also a Hilbert space

with respect to the inner product defined by U~J"/}) = (y~x). In just
the same way~ the fact that H* is a Hilbert space implies that H* is
a Hilbert space whose inner product iS" given by (F1,F 0 ) = (aJ>~
Let H be a Hilbert space. We have two natural mappings of II
into H*, the second of which is onto: the Banach space natural
imbedding x--:., Fz, where F~(f) = f(x)J and the product mapping
x--+ f~--:., F1 ~, 'vhcre f%(y) = (y,x) and F1;,;(f) = (f,f~)~ Show that
these mapph1gs are equal, and conclude that H is reflexive. Show
also that (F' ~,F v) = (x 1 y).

3.

56. THE ADJOINT OF AN OPERATOR

Throughout the rest of this chaptcrt we focus our attention on a
fixed but arbitrary Ililbert space H, and unless we sperifically state
other"Tise, it is to be understood that H is the cont.ext for all our discussions and theorems.
Let T be an operator on H~ We .sa\v in Sec. 51 t.hat T gives rise to

Hirbe rt Spaces

263

an operator T* (its conjugate) on H*, where T* is defined by

(T*f)x

= f(Tx). 1

We also saw that the mapping T __.,. T* is an isometric isomorphism of

IB (H) into ffi (H*) which reverses products and preserves the identity
tra.nsf ormation. In the same way, T*
gives rise to an opera tor
on //* ;
and since H is reflexive~ it follows that
T** = T when H** is identified with
H*
H by means of the natural imbedding.
These statements depend only on
the fact that H is a reflexive Banach
space. We now bring its llilbert
space character into the picture~ and
we use the natural correspondence
between H and H* discussed in the
previous section to pull T* do,vn to II~
'fhe details of this procedure are as
H
follows (see Fig~ 38). Let y be a vector in J/J and /y its corresponding
functional in H*; operate vdth T* on Fig. 38. The conjugate and the
f JJ to obtain a functional fs = T*fv; .and adjoint of T.
return to its oorres ponding vector z
in H. 1.nere are three mappings under consideration here, and we are
forming their product:

(1)

We write z = T*y~ and we call this ne'v mapping 1?* of H into itself the
adjoint of T,. The same symbol is used for the adjoint of T as for its
conjugate because these two mappings are actually the same if II and H*
are identified by means of the natural ~..orrespondence. It is easy to
keep track of whether T* signifies the conjugate or the adjoint of T by
noticing whether it operates on functionals or on vect-0rs. 1.~hc action
of the adjoint can be linked more closely to the structure of H by observing that for every vector x we have (T*f21}x = f"sJ(Tx) - (Tx~y) and
(T~f,,)x ;:::= f.(x) = (x,z) = (x, T*y), so that
( Tx,y)

for all x and Y~

(x, T*y)

(2)

Equation (2) is much more than merely a property of the

In working with operators, it~ common practice to omit parenthese.s whenever

it seeme convenient. There is evidently no impairment of clarity in writing (T~J):t ~
j( T.x) instead or l T *( j)]( x) - f( T( x)) 1 and th ere w Hl be a considerable ga.1 n when we
consider operators and inner products togethert as \Ye do below ..
l

26'

Operators

adjoint of T, for it uniquely determines this adjoint. The proof is

simple~ ii T' is any mapping of JI into itself such that (1Tx,y) == (x.~T'y)
for all x and Yi then (x 1 T'y) __: (x, 'J'*y) for all x~ so T'y = 1'*y ; 1 and
since the latter is true for all y, T' = 1~.
Our remarks in the above paragraph have shoVt~n that to each
operator T on H there corresponds a unique mapping T* of H into itself
(cal1ed the adjoint of T) which satisfies relation (2) for all x and y.
There is a more direct but less natural approach to these ideas, one \.Yhich
avoids any reference to the conjugate of T. If y is fixed, it is clear that
the expression (Tx 1y) is a scalar-valued continuous linear function of x.
By Theorem 55-A, there exists a unique vector z such that (Tx y) = (x,z)
for all x. We now write z = T*y, and since y is arbitrary, 've again have
relation (2) for all x and y. The fact that T* is uniquely determined by
(2) folloVt,.s just as before.
The principal value of our approach to the definitiou of the adjoint
(as opposed to that just mentioned) lies in the motivation it provides for
considering adjoints at alL We can express thi~ by emphasizing that an
operator on a Banach space alv.."'ays has a conjugate which operates on the
conjugate space; and when the Banach space happens to he a Hilbert
epace, then 1 as we have seen, the natural correspondence discussed in the
previous section makes it almost inevitable that we regard the conjugate
as an operator on the space itself. Once the definition of the adjoint is
fully understood, however, there is no further need to mention conj uga. tes. All our future work with adjoints will be based on Eq.. (2)t
and from this point on, the symbol
-..vill always signify the adjoint of T
(and never its conjugate).
As our first step in exploring the properties of adjoints, we verify
that T* actually is an operator on H (a.ll we know so far is that it maps
H into itself). For any y and z 1 and for all z, we have
1

(x, T*(y

+ z))

~ (Tx, y

+ z)

~ (Tx~y)

+ (Tx,z)

= (x,T*y)
80

T(y
The relation

+ z)

+ (x,T*z)

- (x 1 Ty

+ T*z)i

T*y
Tz4
T*(ay) == aT*y

is proved similarlyJ so T is linear. It remains to be eee n that T is

continuous; and to prove this, we note that

The re&eoning here depends on the fact that if y, and y 1 are veetora such tha. t
(ztfll) ..... (:t,1/2) for all :rji then {2; 'II - Y:1) :m 0 for all x:, so YL - :Yt = 0 or 1/i ~ f/J.

Hi Ibert Sp aces

implies the.t I T*yH

:=; HT[j HYll

for all Y1

265

so

llT$1l < llTll.

T ---t T * is a mapping of ffi (H)

These facts tell us that

mapping is called the adjoint ope.ration on IB(H).

in to itself.

This

The adjoint operation '1'--=; 7'* on ffi(H) ha8 the following

Theorem A.
pro periies:

(1)
(2)

{T1
T2)* = T1*
(a.T) * ~ aT*;

(3)
(4)
(5)
(6)

(TlT~)* ~

+ T2*;

1'i*1 1*;
1

T*t. == T;
ll T* II = [l T [I ;

"T*7'11

l Tll !_

The arguments used in proving (1) to (4) are all essentially

the sarne. As an illustration of the method 1 \\"'e observe that (3) follows
from the fact that for all x and y we have

PROOF+

(x~(T1TiY'*y) =

(T1T2x,y)

(Tzx, T1 *y)

(x 1 T2*'1.\*y).

To prove (5),wcnotcthatwealrea.dyhave llT~I < HTH;and ifweapply

this to T* instead of T and use (4) ~ ,ve 0 btain ll Tl[ :::;; 11T**11 < ll T* Hr
Ilalf of (6) follo,vs from (5) and the inequality 47 ~(5), for

and the

fact that "Tl1:!:!

!I Txll 2

(Tx1Tx)

:=;;
=

llT*Tll is an immediate
(T*Tx,x)

consequence of

< llT*Txll [lxll < llT*T!l 1lxll 2

rfhe presence of the adjoint operation is what distinguishes the

theory of the operators on H from the more general theory of the operators
on a re tlexi ve Banach space. 1 In the next three sections, we use this
operation as a tool by means of which VtTc single out for special study
certain types of operators on H whose theory is particularly complete and

satisfying+
Problems

1.
2t

Prove parts (1), {2), and (4) of 'rb-eorem A.

Show that the adjoint operation is one-to-one onto as a mapping of
ffi ( H) in to itself.
l

See Ke.ku tani find Mae key [23 J.

266
3.

4..

Operators

0 and I* = I+ lJse the latter to show that if Tis

non-singular,. then T* is also non-singular, and that in this case
Show that O*

( T*)-1 ;; ( ']_,-1) .
Show that 11 TT*ll

HTH 2

57. SELF-ADJOINT OPERATORS

There is an interesting analogy between the set ffi(H) of all operators

on our Hilbert space Hand the set C of all complAx numbers. This can be
summarized hy observing that each is a complex algebra together with a
mapping of the algebra onto itself (T ----t T* and z ____, z) and that these
mappings have similar properties. We shall see that this analogy is
quite useful as an intuitive guide to the study of the operators on H.
The most significant difference between these systems is that multiplication in the algebra ffi(H) is in genera.I non-commutative, and it will
become clear as we proceed that this is the primary source of the much
greater structlll'al complexity of IB(//).
The most important subsystem of the complex plane is the real
line_, which is characterized by the relation z ~ 2+ By analogyi we
consider thosc operators A on H ~rhich equal their adjoints, that is,
which satisfy the condition A = A*. Such an operator is said to be
self-adjoint. The self-adjoint operators on H are evidently those which
are related in the simplest possible way to their adjoints.
We kno'v that 0'* = 0 and [* = I, so 0 and I are self-adjoint. If
A 1 and A .2 are self-adj ointj and if a and {j are real n umbcrs 1 then
(aA1

+ PA2)*

aA1*

+ PA1*

=a Ai+ PA2

shows th.at a A 1 + p. .4 ~ is also self-adjoin ta Further, if {An I is a sequence

of self-adjoint operators ~\rhich converges to an operator A, then it is
easy to see that A is also self-adjoint ; for

l!A -

< [IA - An JI + HAA - An *~I + l]An * - A *[I = JIA - A~U

+ IJ(A1' - A)*U = HA - AnH + rlAn - Afl ~ 2!rAn - All -4 0

A *!I

shows that A - A
theorem.

= Oi

so A

= A*.

These remarks yield our first

Theorem A. The self-adjoint operators in rn(H) form a closed real linear

subspace of &(H)~and therefore a real. Banach s'{NJCe-which cant.aim the
identity transform.ationr

The reader will notice that we have said nothing here about the
product of two self-adjoint operators. Vecy little is known about such

267

Hilbert Spaces

. products, and the following simple result represents almost the extent of
our information.
Theorem B.. If Ai and A 2 are self-adjoint operators on H, then their
product A 1A2 is self-adjoint=> A lA ~ = A2A i
PROOF4

11iis is an obvious consequence of

The order properties of self-adjoint opera tors are more interesting~

and we devote the remainder of the section to establishing some of the
simpler farts in this dire(~tion~
If T is an arbitrary operator on Hi it is easy to sec that
T = 0

=}

(Tx,y) = 0

for all x and Y~ It is also clear that T == 0 ==> (Tx,x)

shall need the converse of this implication.

0 for all x.

We

Theorem C. If T is an operator on Hf or which (Tx,x) = 0 for all xi

then T = OT
enooF~ It suffices to show that (Tx,y) = 0 for any x and y~ and
the proof of this depends on the follo\ving easily verified identity:
(T(ax

+ JJY)~ ax + /3y)

lal

(Tx~x) -

l~l 2(7'YiY)
~ aP(Tx,y)

+ a{:J(Ty~x).

(1)

We first observe that by our hypothesis~ the left side of (1)-and therefore
t.he tight side as we11-eq uals 0 for all a and {J. If we put a ;;; 1 and
(j ~ 1, then (1) becomes
(Tx,y)

+ (Ty,x)

(2)

O;

and if we put a = i and (3 = 1, we get

i(Tx,y) - i(Ty,x)

:;==

(3)

0.

Dividing (3) by i and adding the result t-0 (2) yields 2{ Tx, y)
(Tx,y) == 0 and the prooi is complete+

O~

so

It is worth emphasizing that this prooi makes essential use of the

fact that the scalars are the complex numbers (and not me1ely the real
numbers).
We now apply this result to proving our next theorem, which
indica tcs that self-adjoint opera tors are linked to real numbers by stronger
tics than might be suspected from the loose analogy that led to their
de fin i tio n.

268

Operators

An operator Ton H iB self-adioint

If Tis self-adjoint, then

Theorem D.
PBOOF~

(Txtx)

= (x, Tx) =

(x~ Tx)

(Tx,x) is real for all x.

(Tx,x)

sho\~ls

that (Tx~x) is real for all x~ On the other hand, if (Tx,x) is real
for all x 1 then ( Tx,x) = (Tx"x) = (x, T*x) = ( T* x,x) or
([T -

for all x..

By Theorem

C~

T*]x" x)

=0

this implies that 1' - T*

OJ so T

T*,.

This theorem enables us to define a respP.ctablc and useful order

relation on the set of all self-adjoint operators. If A 1 and A 2 are selfadjointJ 've "\Vritc A 1 < A2 if (A .xjx) < (A ~x,x) for all x. The main
elementary facts about this relation are summarized in
1

Theorem E. _ 1 he real Banach sprue of all

self~adjmnt

operators rm H is a

partially ordered set whose linear structure and order structure are related by
the following properties:
(l) if A 1 < A2t then Ai + A. :=;; A~ + A far eyery A;
(2) if Ai < A2 and a 2:: 0, then a A1 :5'. qA2.
PROOF.
The relation in question is obvious.Ir reflexive and transitive
(see Sec. 8) To shov-r that it is .also antisymmetric, \ve assume that
A 1 < Ai and A2 < A 1 ~ This implies at once that ([Al - A ~Jx, x) :.== 0
for all x, so by Theorem C, A1 - Az ~ 0 and Ai= A1~ The proofs of
properties (l} and (2) are easy. For instance, if A 1 < A2, so that
(A1x,x) S ( A!X,X) for all x 1 then (A 1xJx) + (Axtx) < (A iX,x) + (Ax,x)
or ([A1 + A]x, x) ~ ((Aft+ A]x 1 x) for all x, so A1 +A < A2 +A+
The proof of (2) is similar.
+

A self-adjoint operator A is said to be po8itive if A 2:: 0, that is, if

(Ax,x) > 0 for all x. It is clear that 0 and I are positive, as are T*T
and TT* for an arbitrary operator T.

If A is a positive opera/.or on II, then I + A is non-singular.

In particularJ I + T*T and l + TT* are non-singular for an arbitrary
operator Ton H.
PROOF.
We must show that I
A is onc... to-one onto as a mapping
of H into itself. First, it is one-to-one~ for

Theorem F..

(I+ A)x

==

O~Ax = -x~ (Ax~x)

(~x~x) == -l~xll 2

>

O~x

= O~

We next show that the range M of I +A is closedr It follows from

]l(I + A)xll 2 = rlxll' + l!AxU 2
2(Ax,x)-a.nd the assumption that A is
positive---that llxrl ~ U(J + A)xl]. By this inequality and the completeness of H, M is complete e.nd therefore closed. We conclude the

HiJbert Spaces

269

proof by observing that M = H; for otherwise there \vould exist a nonzero vector Xo orthogonal to M, and this would contradict the fact that
(xoJ [l + A]xo) = 0 ~ llxoll 2 == - (Axo,x-o) S 0 ~ Xo = 0.
If the reader wonders why we fail to show that the partially ordered
set of all self-adjoint operators is a lattice, the reason is simple: it isn't
true. As a matter of fact, this system is about. as far from being a lattice
as a partially ordered set can be 1 for it can he shown that t \\~o opera tors
in the set have a greatest lower bound {=} they are comparable. This
whole situation is intimately related to questions of commutativity for
algebras of operators and is too complicated for us to explore here+ For
further de ta il.s, see Kadi.Bon I22].
Probtems

1..

Define a new operation of "multiplicationn for self-adjoint operators

by Ai 0 A2 = (A1A~ + A2A1)/2, and note that A 1 0A2 is al\vays
8elf-adjoint and that it e<1ua]s A 1A "l 'vhenever A 1 and A2 commute.
Show that this operation has the following properties:
A 1 A 2 = A~ 0 A i 1
Ai~ (A2 + Aa) = Ai 0 Ai+ A1 o A3,
a(A 1 o A 2) = (aA i) 0 At -= A 1 o ( aA 2) ~

== Io A = A4 Show also that A 1 c (A 2 o A 3 )

whenever A 1 and A 3 commute~

snd Ao I

= (A 1 o A-:i:) o A3

2.. If T is any opera tor on Ht it is clear that l( Tx,x) r < HTx [I II x II <
II TU l]x]1 2 j so if H ~ f0} ~ we have sup ~I (Tx~x)I /llxl] 2 :x r! 0} < [I Tl].
Prove that if T is self-adjoint~ then equality holds here. (Hint
write a = sup f I (Tx~x)[ /!lx[l 2 :x r! 0} = sup rl (Tx,x)!: HxH = 1 l, and
show that l Txll ~a whenever llxll = 1 by putting b = llTxl]H~if
Tx ~ o--and considering

4ll Tx[l 2

~ (T(bx

+ b~ Tx), bx + b-irx)
1

- (T(bx - b-irx), bx - b-irx)

< aillbx + b-irx[l 2

+ ltbx

- b- 1 TxW~J

= 4a[I Txll4)

58. NORMAL AND UNITARY OPERATORS

An operator N on H is said to be normal if it commutes with its
adjoint, that iB, if NN* = N* N. The reason for the importance of
normal operators will not become clear until the next chapter. We shall
see that they are the most genera.I operators on H for which a simple and
revealing .structure theory is possible. Our purpose in this section is to

270

Operators

present a few of their more elementary properties which are necessary

for our later work.
It is obvious that every self-adjoint opera tor is normal, and that if
N is normal and a is any scalart then a.N is also normal. Furthert the
limit N of any convergent sequence {N~} of normal operators is normal;
for we know that N 11; * ----+ N*, so
UNN* -

N*Nll

~ UNN* - N1N1!ll
HNk*N1: ~ N*N[] = !INN* -

+ HNkNk* - Nk*NlU
NkNk*ll + llNk*NJt; -

NNH ~

o,

These remarks prove

which implies that N N - N* N = 0.

Theorem A.. The set of all norm.al operator.s on H i8 a closed subset of

ffi(H) which cantains the set of all aelf-adjoint operators and is closed under
scalar multiplication.

It is natural to wonder whether the sum and product of two normal

opera tors are neceHSarily normal. P"rhey are not, but neverthelessJ we can
say a 1i ttl e in th is direction.
Theorem B. If N 1 and J.l ~ are normal operators on 11 with the property
that either commutefi with the adjoint of the other, then Ni
N 2 and N iN 2
are nor1nal.
PRO OF. It is clear by taking adj oints that

so the assumption implies that each commutes with the adjoint. of tha
other. To shov{ that Ni+ N2 is normal under the stated conditions, we
have only to compare the results of the follo\ving computations:

(N1

+ N2){N1 + N2)*

and

(N1

+ N-i.)*(N1 + N2) =

(N1 + N2){N1* + N?.*)

N1N1* + N1N2*. + N.,.Ni*
(N1* + N2*)(N1 + N2)

= N1*N1 + N1*N2

+ NJV2*

+ N'l.*N1 + N"J*N2~

The fact that N 1N 2 is normal follows similarly from

N1N2(N1N2)* = N1N2N2*N1*

= N1Ni*N2N1*
=

N,,*N1N1*N.,,
Ni*N1*N1N2 = (N1N.,,)"'NJN2.
=

By definition, a .self-adjoint operator A is one which satisfies the

identity A *x = Ax. Many properties of self-adjoint operators do not
depend on this, but on1y on the weaker identity HA *xii = l!Axll~ Our
next theorem shows tha.t all such properties are shared by normal
operators4

Hj Ibert Spaces

271

Theorem

c.

PROOF.

In view of Theorem 57-C, this is implied by the fact that

llT*xll

An operator Ton His norm.al<=>

!ITxrl ~ ltTxrl 2

:::::=

IJT~xu

==

r1rxrr for every X+

!ITxH 2 Q

(Tx, Tx) (::::} (TT*x,x)

(T*x,T*x)
(T*Tx,x) <=>([TT* -

T*T]x,x) = OL

The following consequence of this result v.dll be useful in our later

\\rorkL
Theorem D. If N is a normal operator on H, then l1J-~l 2 fl
PROOF.
The preceding theorem shows that

for every x 1 and this impiies that ll..L~,l 2 ll = llN*NI~have [IN*NH = [! 1'l !I2, so the proof is com plcte.

= llNfl'l.

By Theorem 56-A, we

W c kno\v that any complex number z ran be expressed uniquely in

the form z ~ a + ib '"~here o and bare real numbe1s, and that the~ real
numbers are called the real and imaginary parts of z and are given by
a = (z
z)/2 and b = (z - z)/2i. The analogy -betv.~ccn general
operators and eomplex numbers,, and between self-adjoint opera.tors and
real numbers, suggests that for an arbitrary opera.tor T on H "\Ve form
Ai = { T + T*) /2 and A 2 = (T ~ '.P*) /2i4 A 1 and .A. 2 are el early sclfadjoin t, and they have the property that T = A 1
i'A z. The unique ...
ness of this expression for T follo,vs at once from the fact that

The self~adjoint operators Ai and A~ are ca11ed the real part and the
imagi-nary part of 1'.
W c emphasized earlier that the complicated structure of ill (JI) is
due in large part to the fact that operator multiplication is in gcneraJ
non-commutative. Since our future '"Tork ~~in be focused mainly on
normal operatorsi it is of interest to scc--as the following theorem show.s---that the existence of non-normal operators can be traced directly to the
non-commutativity of self-adjoint opera tors.
Theorem E. If T is an operator on II, then T is norm.al
imaginary parts cmnmute.

its real and

If Ai and A 2 arc the real and imaginary parts of T 1 so that

Ai
iA2 and T* = A 1 - iA~ 1 then

PROOF.

'l.'

.;;>

TT*

= (Ai

+ iA2)(A1 -

iA2)

= A1 2 + A2~ + i(A2A1

- A1A2)

and

T*T ~ (A1 - iA2)(A1

+ iA2) = Ai~+ A'l 2 + i(A1A2 -

A2A1)~

272

Ope rotors

It is clear that if A1A2 ~ A 2A 1 _, then TT* = T*T. Conversely, if

TT = TT, then AiA2 - A2A1 = A2A1 ~ A1A2, Eo 2A1A2 = 2A2A1
a.nd A 1A2 = Az_A 1+

Perhaps the most important subsystem of the complex plane after

the real line is the unit circle, which is characterized by either of the
equivalent identities lzl = 1 or zz = Zz = 1. An operator U on H
which satisfies the equation U U* = U* [] ~ I is said to be unitary.
l"';"nitary opcrators-v..Thich are obviously normal- arc thus the natural
analogues of complex numbers of absolute value 1. It is clear from the
definition that the unitary operators on Hare precisely the non~singular
operators whose inverses equal their adjoiuts. The geometric significance of these operators is best understood in the light of our next theorem.

If Tis an operator on Ii, then the f oUowing conditions are all

equit~alent to one another:
(1) T*T == I;

Theorem f.

(Tx~Ty)

(x~y)

for all x and y;

II TxH = llx Hfdr all X~
PROOF. If (1) is true_, then ('1 *1.tx,y) = (x,y) or (Tx,Ty) ~ (x~y)
for all x and Y~ so (2) is true; and if (2) is true, then by taking y = x we
obtain (Tx, T:c) = (x,x) or II 1'xli 2 ~ 11xU 2 for all ~, so (~) L~ t.1ue. The
fact that (3) implies (1) is a conseqne.nce of ~rheorem 57-C and the following cha.in of implications:
(2)
(3)

HTxlj =

llxll ~ II Txll 2 = Hxll 2 ~ {Tx_, Tx) =

=

(x~x) ==:. (T*Tx,x)

(x,x) ==> ([T*T - l]x,x)

0.

An operator on H with property (:J.) of t.his theortm is simply an

isometric isomorphh'3m of H into itself. 1.,h~t t. an opera tor of t.his kind
need not be unitary is -easily ::-IBen by considering the operator on l-i
defined by

T {X1,

Xi,

J =

which preserves norms but has no

o, xi, X2J

inverse~

i~

These ideas lead at once to

An operator T on II is unitary ~ it is an iBometric isomorphism of H onto itself.

PROOF. If T is unitaryJ then vle know from the definition that it is
onto; and since by Theorem F it preserves norms, it is an isometric
isomorphism of H onto itself. Conversely~ if T is an isometric isomorphism of H on to itself1 then ']_'- exists~ and by Theo rem Ii'' vte have
7. *T = I. It no,,T folloVr'"S that (T*T) r- 1 JT- 1, 80 T* == 7t-l and
7 T* ;;; T *'1 = I, v..T hi ch shovts that T is unitary

Theorem G.

:;:=

Hilbert Spaces

273

This theorem makes quite clear the nature of unitary operators:

they are precisel3t those one-to-one mappings of H onto itself which
preserve all structure---the linear operations~ the norm, and the inner
product.

Problems
1.

2.
3~

4..

If T is an arbitrary operator on 11, and if a and fl are scalars such

that lal ~ rPlr show that aT + IJT* is normal~
If H is finite-dimensional~ show that every isometric isomorphism of
H int-0 itself is unitary.
Show that an operator Ton His unitary~ T({ei}) is a complete
orthonormal set Vr~hcnever [e, 1 i'3.
Show that the unitary opera.t-0rs on H form a group.

59.. PROJECTIONS
Aecording to the definition given in Sec. 50t a projection on a Banach
space B is an idempotent operator on B~ that iBt an operator P with the
property that P 2 = P. It "~as proved in that sect.ion that each projection P determines a pair of closed linear subspaces Ai and N~the range
and null space of P~sucb that B == Al $ N, and 3.lso~ conversely, that
each such pair of closed linear subspaces .ll1 and N determines a projection
P with range J.lJ. and null space N. In this way~ there is established a
one-to-one correspondence between projections on B and paLn; of closed
1inear subspaces of B ,,..-hich span the whole space and have only t.he zero
vector in common.
The context of our present \\"Ork~ however, is the llilbert space H,
and not a general Banach space~ a.nd the structure \Yhich H enjoys in
addition to being a Banach spacr rnableB us to single out for special
attention those projections whose range and nun space are orthogonal.
Our first theorem gives a convenient characterization of these projections~
Theorem A,, If P is a proiectinn on H with range M and null space N,
tlten M .l N ~ P i8 self-adjoint, and in this case, N = M 1- ~
PROOF.
Each vector z in H can be written uniquely in the iorm
z ~ x
y with x arid y in M and N. If M .l N, so that x .l. y, then
p = P will follow by Theorem 57-C from (P*z,z) = (Pz,z); and this
is a consequence of
4

(P*z,z)

(z,Pz)

(z~x) -

(x

+y

x)

(x,x)

+ (y,x)

(x,x)

27.4

Operators

and (Pz,z) = (x,z) = (x, x + y) ~ (x~.:t)

(:t,y) = (x,x)~ If, conversely,
p)Jl = P, then the conclusion that M J.. N follows from the fact that for
any x and yin Mand N we have
{x,y)

(Px,y)

(x,P*y) ~ (x,Py) ;:::;: (x 1 0) = 0.

All that remains is t.o see that if ~f .L N~ then N = M 1-+ It is clear

that N ~ M 1-; and if N is a proper subset of Ml..~ and therefore a proper
closed linear subspace of the Hilbert space Af 1-, then Theorem 53-B
implies that there exists a non-zero vector z 0 in ~! .i such that z~ ..L N+
Since zo l. Mand zo .L N, and since H == M $ N~ it follov.~s that zo ..L H.
This is im.poBsible, so we conclude that N = Af 1-,

A projection on H whose range and null space arc orthogonal is

sometimes called a perpendicular projection. 1-.,he only projections considered in the theory of Hilbert spaces are those \vhich arc perpendicular,
so it is customary to omit the adjective and to refer to them simply as
projections~ In the light of this agreement and Theorem A,. a projection
on II can be defined as an opera.tor P which satisfies the conditions
pi = P and P* ~ P~ The operators 0 and I are projections, and they
are distinct~ H ~ f0}4
The great importance of the projections on H rests mainly on
rrheorem 53-D, which allows us to set up a natural one-to....one correspondence between projections and closed linear subspaces~ 'To each
projection P there corresponds its range M = { Px: x E HI, which is a
closed linear subspace; and conversely, to each closed linear subspace jl..f
there corresponds the projection P lvith range M defined by P(x
y) = x 1
where x and y are in M and M J... Either way, we speak of P as the
projection mi J.V.
It is clear that P is the projection on M $=::}I - P is the projection
on M 1-, AlsoJ if Pis the projection on M~ then

x E. M ~ Px

= x {=} ![Pxl1

llx11:

The first equivalence here was proved in Problem 44-11; and since for
every x in H we have

1lxl! 1 =

UPx

+ (I -

P)x!i 2

= \!Pxli:a + H(I -

P)xlli,

(1)

the non-trivial part of the second is given by the following chain of

implications:

llPxH :;:: llxll => 1!PxU 2

1lxH 1 ~ 1'CI - P)zn! = O ~Px = x.

HPxl1 ~ Hxll for every x, so 11PU < 1.

Relation (1) also shows that

is an arbitrary vector in H, it is easy to see that
(Px~x)

= (PPxix) = (Px,P*x) = (Px,Pz) = HPxB 2 2. OJ

If z
(2)

Hirbert Spaces

275

so Pis a positive operator (0 $ P) in the sense of Sec. 57. Since I - P

i.s also a projection, we also have 0 < I ~ P or 1~ 5 I, so 0 :::; P < I.
Let T be an operator on II. A closed linear subspace M of H iR
said to be invariant under T if TC.~) C M ~ When t.his happe1ls, the
restriction of T to M can be regarded as an operator on A1 alone, and
the action of T on vectors outside of M can be ignored. If both M and
M 1- are invariant under T, we say that J.1 reduceB T~ or that Ti reduced
by M. This situation is much more interesting, for it allows us to
replace the ~tudy of T as a \vhole by the study of its restrictions to M
and "AI .L, and it invites the hope that these restrictions will turn out to
be operators of some particularly simple type. In the following four
theorems, we translate these concepts into relations between T and the
projeetion on M
4

Theorem B. A closed linear subS']Jace Jf of 11 iB inuarian~ under an

operator T ~ M .!. is invariant under T ~
,

Since M ..L..L = M and 'J. * = T~ it suffices by symmetry to

prove that if M is invariant under T" then M 1- is inva.riant under T*+
If y is a vector in M 1-, our conclusion '\vill follow from (x, T*y) = 0 for
all x in ML But this is an easy consequence of (x,T*y) = (Tx,y), for
the invariance of M under T implies that (Tx,y) = 0.
1

PROO!t,L

Theorem C. A cl-Osed linear Bubspac.e Jf of ll reduces an operator T ~ M

is invariant under both T and 7'*,.
PROOF.

This is obvious from the definitions and the preceding theorem.

Theorem D. If P is the proiection on a closed linear sub,~ace ..f"\tf of H,

then Mis int.ariant under an operator T <:=:-? TP = PTP.

If Jf is invariant under T and x is an arbitrary vector in Hr

then TPx is in M, so PTPx = '1 Px and PTP;:::: TP+ Conversely, if
1. P = P?.TP and x is a vector in M, then '1,x = T Px = PT Px is also in
M, so M is invariant under T.
PROOF.

Theorem E. If Pis the projection on a cl.osed linear 1tttb81Jace M of H, then

M reduces an operator 7 <::=> TP = PT.
1

reduces T ~Ji is invariant under T and T* ::;. TP = PTP

and T*P = PT*P <=> TP == PTP and PT = PTP~ The last statement
in this chain clearly implies that TP = PT; it also follov.r~ from it, as
we see by multiplying 1,P = PT on the right and left by P.

PROOF.

~/

Our next theorem shows how projections can be used to express the
statement that tl\;rO closed linear subspaces of Hare orthogonal.

If P and Q are tM projections on clo8ed linear 8'Ubspaces "'"\/

and N of HJ then M L N ~ PQ ~ O <;;;;;:; QP = 0.

Theorem F.

276

Operators

We first remark that the -rquivalrnce of PQ = 0 and QP = 0

iB clear by taking adjoints. If Af J_ ~v, so that N ~ kl l , then the fact
that Qx is in N for every x implies that PQx =::: O, so PQ = 0. If, conversely, PQ = 0, then for every x in N we have Px = PQx == O, so
N c M..L and J-! .l N.
PROOF..

Motivated by this result, we say that two projections P and Q are

orthogonal if PQ = 0 ~
Our finn.l theorem describes the circumstances under which a sum
of projections is also &. projection.
P ~ are the projections on closed linear sub..,,
spac,es M1, M!, . . . , Mn of H, then P =Pl+ Pi+ ~ .
Pn. is a
projection~ the Pis are pairwise ort.hogonal (in the sense that P.Pi = 0
whene-i~er i '# J) ; and in th is case 1 P is the. proje.c.tion on

Theorem G.

If Pi,

P2t

... ,

M == M 1

+ Jf +
!

+ MB

Since P is clearly self-adjoint, it is a projection ~ it is idempotent. If the P/s are pair,vise orthogonal~ then a simple compu t.ation
shows at once that Pis idempotent. To prove the converse, we assume
that P is idempotent. Let x be a vector in the range of p,.j so that
x ~ P"x. 1'"hen
PROOF.

TC

[]xB~

]~PtTi[~

<I

;'=-l

HP;xllr

(PJx,x)

(Px1x) ~ []Pxli 2

::;

Hxll 2

j=l

We conclude that equality must hold all along the line here, so
n

2
llP1xllt
=
llPall
j=l

and

Thus the range of Pi is contained in the null spa.cc of P1J that is~M, ~ Mi1-,
for every j ~ i. This means that Mi ...L M, whenever i # j, and our
conclusion that the P/s are pairwise orthogonal now follows from the
preceding theorem. We prove the final statement in two steps. First,
we observe that since HPxU = Uxl] for every x in M,, each Mi is contained
in the range of P ~ and therefore M is also contained in the range of P ~
Second, if xis a vector in the range of P, then
z = Px -== Pix
is evidently in M

+ P2z + + P,,.x

There are many other ways in which the algebraic structure of the
set of all projections on H can be related to the geometry of its closed
linear subspaces, and several of these are given in the problems below.

Hilbert Spaces

277

11ie significance of proje etions in the genera.I theory of operators on H

is the t.heme of the next chapkr. A8 we shall see, the essence of the
matter (the spectral theorem) is that. every norm.al operator is made of
projections in a way lvhich ch~arly reveals the geometric nature of its
action on the vectors in H.
Problems

1.

2.

If P and Q are the projections on closed linear sttbspaccs M and N

of II, prove that PQ is a projection<=> P<J ~ QP. In this caseJ show
that PQ is the projection on Jf rt N.
If I) and Q are the projections on closed linear subspaces Llf and N
of H, prove that the following statements are alt equivalent to one

another:
(a)
(b)

< Q;
HPx!I < HQxl1

for

every x;

(c) .tlf c N;
(d) PQ = P;
(e) QP = P.
(Hint the equivalence of (a) end (b) is easy to prove, as is that of
(c), (d), and (e); prove that (d) implies (a) by using

(Px,x) ~

3.

4.

IJPxll 2

= ~1r~Qxlr 2

< 11QxH 2 =

(Qx}x);

and prove that {b) implies (c) by observing that if x is in M, then

llxU = l!Pxl[ < 11Qx1[ :$ ilx!I+)
SbO"?."" that the projections on H form a complete lattice with respect
to their natural ordering as self-adjoint operators. (Compare this
situation with that described in the last paragraph of Sec. 57 ~)
If P and Q are the projections on elosPd linear subspaces M and N
of H, prove that Q - P is a projection~ P ~ Q. In this case,
show that Q - P is the projection on N n M J.,

CHAPTER ELEVEN

1inite-dimensional Spectral ?:lteorv

If T is an operator on a Hilbert space H ~ then the simplest thing T
can do to a vector xis to transform it into a scalar multiple of itself:

Tx == AX.

(1)

A non-zero vector x such that Eq. (1) is true for some scalar A is called
an eigeniector of T, and a scalar X such that ( 1) holds for some non~zero x
is called a11 eigenvalue of T. 1 Each eigenvalue has one or more eigenvectors a.swcia ted 'vi th it~ and to ea eh eigenvector there corrcspo nds
precisely one eigenvalue. If JI has no non-zero vectors at alli then T
certainly has no eigenvectors~ In this case the ~\thole theory collap~es
into triviality, so we assume throughout the present chapter that
H ~ ~O}.
Let X be an eigenvalue of T, and consider the set M of all its corre..
sponding cig-cnvect.ors together with the- vector 0 (note that 0 is not an
eigenvector). Mis thus the set of all vectors x which satisfy the equation
(T - J,,l)x

= 0,

and it is clearly a non-zero closed linear subspace of II. ~le call M the
eigenspace of T corresponding to A+ It is evident that ~Y is invariant
under T and that the restriction of 'T to ~Y is a very simple operator~
namely~ scalar multiplication by k.
In order to place the ideas of this chapter in their proper framework,
1

The equivalent terms ckaraden'slie vector a.nd charac.t.eristie iuluei and proper
vect-Or and pr-oper value, a.re UBed by many write~.
278

Finite-dimensional Spectral Theory

279

l\~e

lay down several rather sweeping hypotheses, whose vaJidity we

examine later:
(a) Tactually has eigenvalues~ and there are finitely many of them,
say A.1, A.~, ~ ~ , Xm..--L-which are understood to be distinct'Vl-'ith corresponding cigcnspaces A-I h M ~, . . . , Mm;
(b) the Mt:'s are pairwise orthogonal~ that is, i "#- j t j Mj .l M 1 ;
(c) the ~1/s span H~
Putting aside for a moment the question of whether these statements are
true or not} V1-,. e investigate their implications. By (b) and (c), every
vector x in H ran be expressed uniquely in the form
=

\\rhere

+ X2 +

X1

~ ~

. + X~,

(2)

is in M..: for each i and the x/s are pairwise orthogonaL

follows from (a) that
xi

Tx

+ Tx2 + + Txm
A1X1 + A2X~ + + AmXm

It now

Tx 1

(3)

This relation exhibits the action of T over all of H in a manner which

renders its structure perfectly clear from the geometric point of view.
It will be convenient to express this result in terms of the projections
P1 on the eigcnspaces Mi. By T~heorcm 59-F1 (b) is equivalent to the
follov,dng statement:
the P/s are painvise orthogonal.

Also_, since for each i and for every

i we have M 1

(4)
~

M,1., Eq. (2}

yields
and it follows at once from this that

Ix = x =
=

for every

xin H, so
I

+ x~ + +
P1X + P'JX +
+Pm::c
(P1 + P2 + + Pm)x
X1

Xm

+ P2 + +Pm.

P1

(5)

Re1a.tion (3) now tells us that

+ X2X2 + ~ + Amxm
== A1P 1X + A2P 2X + ~ ' + ).,_mpm,X

Tx = A.1x1

(X1P1

+ A2P2 + '

~ ~

+ XmP .,,.)x

for every x, so
(6)

280

Operators

The expression for T given by (6)~whcn it exists-Ls ca.lied the spectral

reaolution of T. Whenever this term is used, it is to be understood that
the X/s are distinct and that the P/s are non-zero projections which
satisfy conditions (4) and (5). W c shall see later that the Bpectral
resolution of T is unique \vhen it exists.
All our inferences from (a), (b), and (c) are perfectly rigorous 1 but
the status of these three hypotheses remains entirely up in the air. First
of all, \vith reference to (a), does an arbitrary operator Ton H necessarily
have an eigenvalue? The answer to this is no 1 as the reader will easily
verify by considering the operator T on l, defined by

1
[

X1_, X2~ . .

= {

o, X1, X2,

i.

On the other hand, if H is finite-dimensional, then we shall see in Sec. 61

that every operator has an eigenvalue. :For this reason~ we assume for
the remainder of the chapter-unless we specifically state ot.hen~lisc
that H is fi ni te-dimcn.sio nal 'v ith dimension n~
We have seen that if T sa tisfics conditions (a), (b) and (c) _, then it
has the spectral resolution (6). It is too much to hope that every operator on H meets these requiremcnts 1 so the question arises as to what
restrictions they impose on T~ This question is easy to a.ns\ver: T must
be normaL For it f otlo''ts from (6) that
l

T*

X1P1

+ 'X;P:i + + h:Pm,

and by using (4) we readily obtain

TT*

= C\1P1 + XJ'i

+ " " ~ + Am.Pm)(~P1 + "A;P2 + " + A:"Pm)

= IA1l 2P1 + P\2J ~pi + " + f Am.I :pm
4

and, similarly,

This entire circle of ideaB will be completed in the neatest possible way
if we can show that evecy normal operator on H satisfies conditions (a),
{b), and (c), and therefore haB a spectral resolution. Our aim in the
present chapter is to prove this as..~rt.ion~ which is known as the spectral
theorem, and the machinery treated in the follo,ving sections is directed
exclusively toward this end+ We emphasize once again that His understood to be finite-dimensional with dimension n > 0.

60 . MATRICES
Our first goal is to prove that every operator on H has an eigenvalue_,
and in pursuing this we make use of certain elementary portions of the

Finite-dimensiona1 Spectra I Theory

281

theory of matrices. We adopt the view that the reader is probably

familiar lith this theory to some degree and that it suffices here to give a
brief sketch of its basic ideas. Our discussion in this section is entirely
inde~nden.t of the Hilbert space character of Hand applies equally well
to any non-trivial finite-dimensional linear space~
J.~t B = {ei, e2, . . . , en} be an ordered basis for H, so that each
vector in H is uniq ucly c xpressible as a lin1?..ar com bina tio n of the e/ s.
If Tis an operator on II, then for each e3 \Ye have
1iL

Te;

~Jet..

-1

(1)

'vhich a.re determined in this \Vay by T form t.he tnatri:.r.

of T relative to the oI"dered basis B~ ,,,..e symbolize this matrix by [~t],
or if it seems desirable to indicate the ordered basis un<ler conf-iidP.ra.tion,
by [T]B. l t is customary to write out a matrix as a square arrJ..y;
rrhe n 2 scalars

Ciij

[T] =

(2)

The array of scalarA (ai1 1 a;21

~ ain) is the ifh rotD of the matrix (T],
and (a1h a2;.,
Cin:J) is its jth column+ As this terminology shows,
the fir~t subscript on the entry Cli.:i a],vays indicates the ro\V to \vhich it
belongs, and the second the column. In our work, '"re generally write
(2) more concisely in the form
L

4'

(3)

The reader should make sure that he has a perfectly clear understanding
of the rule according to VtThich the matrix of Tis eonstn1ctcd: lA:rrite 1'fei
as a linear combination of el~ e~, . . . , ei-i, and use the resulting coefficients to form the jth column of [1']~
We offer several comments on the above pa.ragTaph. First:t t.hP
term mairix has not been defined at all, but only ~"'t.he matrix of an
operator relative to an ordered basis/' .l'1. matrix~defined simply as a
square array of scalars-is sometimes regarded as an object 'vorthy of
interest in its own right. For the most part, however, we shall consider
a matrix to be associated with a definite operator relative to a particular
ordered basis, and \.Ve shall regard matrices as little more than romputa~
tional devices which are occasionally useful in handling operator~+ Next,
the matrices we work with are an square matrices. R.cctangular matrices
occur in connection with linear transformations of one linear space int.a
another and are of no interest to us here. Finally_, 've took B to be an

282

Ope rotors

ordered basis rather than merely a basis, because the appearance of the
array (2) clearly depends on the arrangement of the e/s as well as on the
e./ s themscl ves. In most theoretir,al considerations, however t the order
Qf the rows and columns of a matrix is as irrelevant as the order of the
vectors in a basis. For this reason, we usually omit the adjective and
speak of ''the matrix of an opera tor relative to a basis.''
By using the fixed basis B ~ {e;}J we have assigned a matrix:
[T] = [rui] to each operator T on H, and the mapping T-4 {T] from
opera to rs to matrices is described by Tei = ,;711:11 1 a.ijCi+ The importance
of matrices is based primarily on two facts: T --4 [T] is a one-to-one
mapping of the set of all operators on H onto the set of all mat.rices; and
algebraic operations can be defined on the set of all ma trices in such a
manner that the mapping T ---+ [ T] preserves the algebraic s lruc t ure
of IB(H).
1;he first of these st.a temen ts is easy to prove. If vle know that
[aiil is the matrix of. T, then this information fully determines 1 x for
every x; for if x = ~j i t31-eh then
1

Tx

I: {3;Te5
I: f3i cI Uijei)
I <I: O:ijfjj)et: .

j=l

Pl

j=]

~=I

.; ... 1

.::i:i:-t

This shov{s that T---:,, [TJ is onc-to~one. We see that this mapping is
onto by means of the follo1,dng reasoning~ if [ai1] i8 any matrix~ then
Tei = !~_ 1 auei defines T for the vectors in B 1 and when T is extended
by linearity to all of H, it i8 clear that the resulting operator has [aii] as
its matrix.
To establish the second statement, it suffices to discover ho"Pr~ to add
and multiply two matrices and how to multiply a matrix by a scalar, in
such a way that the follov.Ting matrix equations arc true for all operators
T1 and T2 on H: [T1 + T~d = [T 1] + [T2], faT1] ~ a[T1]J and

fT1T2J

Let

[~;]

[T1UT2J+

and [p#] be t.he matrices of Ti and 2,2.

(T1

+ T2)e;

--rhe computation

+ T2ei
= I: ai;e i + I: f31J-ei
= T1e1

1t

-1

i=-=l

..: :o::i:l

(ai;

+ /3i;)ei.

Finite-dimensional Spectral Theory

283

shows that. if we define addition for matrices by

[lli11 + [tji11
[T1 + T2J

then we obtain

= [ai:j + PiiL
= [T 1J + [T2].

(4)

Similarly, if we multiply a matrix by a scalar in accordance with

a[.:liJ]
(aT1]

then

= [aa.ij],
= a[T1]~

(5)

Finally, the computation

n

(T1T2)e, = T1(Tf].e1)

T1

(I: ~k,~k)
k==l

I {3k-;T
= I
(I: airre1)
~ I (I aa~k;) ei
tC.k

t-= I
n

t'.ti

k=I

i=-=1

i=l

k=l

shovls that if we define multiplication for matrices by

n

{ai;] [tft:;] = [

o:ud31t:i ~

(6)

k= l

then we get

{T1T2] =

(T1HT~J.

The operations defined by (4), (.5), and (6) are the standard algebraic
operations for matricesa In v_rords, i,.vc add t\vo matrices by adding
corresponding entries~ and we multiply a matrix by a scalar by multiplying each of its entries by that scalar. The verbal description of (6) is
more complicated, and is often called the ro1D-by-column rule: t.o find the
entry in the ith row and jth column of the product [aij][.8..:i], take the ith
row (ail, 0-~2, . . ~ , "in) of the first factor and the ith co1umn (ft1J, /j2;,
. . , fJni) of the second, multiply corresponding entries, and add:
n

~ 12iJ:{jkj

~J3Ii

O!.f"J./32;

+ r:l1nf3ni

k e. l

It is worth noting that the image of the zero operator under the mapping
T--+ [TJ is the zero matrixJ all of '"~hose entries are 0. Further, it is
equally clear that the image of the identity operator is the identity matrix i
which has 1is down the main diagonal (v..Thc re i = j) and O's el sew here.
If we introduce the standard Kronecker delia 1 which is defined by
aij

0
{ 1

if i ~ :i
if i = j J

then the identity matrix can be written

[a1JJ~

284

Operators

We no\V reverse our point of view for a moment (but only a. moment)
and consider the set An of all n X n.matrices as an n.lgehraic system in its
ovrn right, with addition, sca.lar multiplication, and mu1tiplication defined
by (4) 1 (5), and (6). It can be verified directly from these definitions
that An is a complex algebra with identity (the ide..ntit~{ matrix), called
the total matrix algebra of degree n. If we ignore the ideas leading to
(4) (fl), and (6)~ then the structure of An is defined, and can be studied,
'Yithout any reference to its origin as a representing system for the operator~ on H~
This a.pproar.h \Vou1d make very little sense, ho1vever, because
the primary reason for considering matrices in the first place is that they
provide a {~ompulational t-001 'vhich is usP-ful in treating certain aspects
of the theory of t..hesc opera tors~
Let us return to our original position and observe two facts: that
IB(H) is an algebra; and that the structure of An is defined in just such a
\Vay as to guara ntce that the one-to-one mapping 7r ---+ [ 7'1 of ~ (H) on to
An prescrveH .addition, scalar multiplica.tioni and multiplication. It IlOl\r
fol1o\vs at once that An is an algebra, and that T ~ [T] is an isomorphism
(see Problem 4f>-4) of C(H) onto A 11
\'le give the following formal summary of our work so iar.
t

Theorem A. lj B = {ed is a basis for 111 then the mapping T ~ 111,

v'hich assigns io ea.ch operator T its 1n.atrix relatii~e ta B, is an isomorphism
of the algebra ill ( R) onto the total matrix algebra .4nL
If T is a non-singular opera.tor whose matrix relative to B is {a.H],,
then T- 1 clearly has a matrix whose entries arc determined in some Vlay
by the Cli/S. The formulas involved here are rather clumsy and complicated~ and since they have no importance for us, we shall say nothing
further about them.
It is necessary, hov{ever 1 to know what is meant by the inverse of
a matrixi when it is considered purely as an element of An and without
reference to any operator '\vhich it may represent. We first remark that
the id Pnti ty matrix [rhi1 is easily seen by dire ct matrix mu] tip li cation to
be an identity clement for the algebra An~ in the sense that we have
( Oij]( 8iJ-}

= [8ii][ O:ij] == [UiJJ

for every matrix {~.'.il; and by the theory of rings, this identity is unique.
A matrix [001] is said to be non-singular if there exists a matrix [t3i1] such
that

1O.ij JLB~; J =

LB.;] [Orij]

= [8iJ] ;

and, again by the theory of rings~ if such a matrix exists~ then it is unique,
it is denoted by [aiJ]- 1 , and it is called the inverse of [a#].
These ideas are connected with operators by the following considerations. Suppose that [ai 1] is the matrix of an operator T relative to B.

Finite-dimensional Spectral Theory

285

We know that the non-singularity of Tis equivalent to the existence of an

operator T- 1 such that
TT- 1 == '1 1T = I.
1
-

The isomorphism of Theorem A transforms this operator equation into

the matrix equation

which is equivalent to
[~iHT- 1 ] =

[T- 1][aii} = [ai;].

We therefore have
Lei B be a basis fur H, and 1' an operator tDhose m.-atrix
relative ta R is [aij]. Then Tis non-singular (::::} [~i] is non-singu.la.r, and in
this case [at;}- 1 = [T- 1]~

Theorem B.

There is one further issue which requires discussion. If T is a fixed

opera t-0r on H, then its matrix [TJn relative to B o bviouf3ly depends on
the choice of B. If B changes, how~ does [TJu r..hange? J\1orc spccifi-+
cally, if B' = {/1, /2i . ~ . , f n l is a1so a ba8is for ll, Vt'"hat is the relation
bct,vccn (T]B and (1"'lsj? 'fhe answer to thi8 question is best given in
terms of the non-.singu1ar operator A defined by Ae; = f(. Let [aii]
and LBi.i] be the matrices of T relative to B and B 1, so that
n

Te; : : :;.

I. a.i,-ei
i-1

and Tf1 = Z7==1 /31ff1T J...ct ['Yi;] be the matrix of A relative t.o B, so that
Ae; = 2;~_ 1 'Yii~- By Theorem ll, [~ii] is non-singular. -~le now compute
Th in two different ways:
n

Th

f31c1fi;

I. ~qAl\t
I <I. 'Yilt\)
~ ( I: 1'ikf1~;) ei;

==

l::cw:l

.1::-1
a

n.

fl1t:i

i:-1
n

i=l

it>;ICLl

i=l

and

Th

===

TAei

(I. 'Y;et)
~ :!: Y1c~Te1i;
~ I,
{I: ai:te1)
I (I aik'Yk;) ei.

=T

k=l

:cot}

n.

;r:i

Yki

1:=1
n

i::::!:El

,:-1

k:::ml

286

Operators

A comparison of these results shows that

for all i and j 1 so

[-r ii][Pi;] = [a i; ][1\iJ

[t'i1] = [1'ij]- 1 [ aij] ['Yi.?1

or

(7)

If we no \V VtTritc this in the form

['T]n1

=:;

[AJ B- 1[T]B{A ]B,

then it becomes quite clear how the matrix of T changeR \vhen B is

replaced by B 1
Two ma trices (afil and [t3'..:i1 are said to be similar if there exists a
non-singular matrix [l'iiJ such that (7) is true. The analysis given above
proves half of the following theorem (\ve leave the proof of the other half
to the reader).
4

Theorem Ct

Two matrices in An are similar 8 they are the rn..(1..trices of a

single operator on H relative to (possibly) different bases.

We. arc now in a position to formulate the fundamental problem of

the classical theory of matrices. A given operator on H may have many
different matrices relative to different bases~ and Theorem C sho\vs in
purely matrix terms hovir these matrices are related to one another+ The
question arises as to "\\Thethcr it is possible to find~ for each operator (or for
each opera tor of a special kind) 1 a basis relative to which its matrix
assumes some particularly simple form+ 11lis is the canonical forrn
problem of matrix theory, and the most important theorem in this direction is the spectral theorem, which \Ve state in the language of matrices
in Sec. 62r In the classical approach to these ideas, it \Vas customary to
work exclusively with matrices~ Ilol\rever, the great advances in the
understanding of algebra which have taken place in recent years have
made it plain that problems of this kind are best treated intrinsica11y,
that is,. directly in terms of the linear spaces and linear transformations
involved. As matters now stand~ it is possiblc----and preferable--t-0 state
the main canonical form theorems of matrix theory without mentioning
matrices at al1. N evcrtheless, matrices remain useful for some purposes,
notably (from our point of view) in the problem of proving that an
arbitrary operat-0r on H has .an eigenvalue.
Problems

1~
2.

Show that the dimension of ffi(H) is n 2

A scalar matrix in An is one which has the same scalar in every position on the main diagonal and O:ts elsewhere. Show that a scalar

Finitedimensionol Spectral Theory

3.

4~

5.

matrix commutes with every matrix, and that a matrix which

commutes with every- matrix is necessarily scalar. Vihat does this
imply about IB(//)? (Sec Problem 45~3+)
A diagonal m.atrix in An is one which has arbitrary scalars on the main
diagonal and 0 1s elsewhere~ Show t.hat all diagonal matrices commute \vith one another, and that a matrix is necessarily diagonal if it
'
commutes
with all diagonal matrices.
Complete the proof of "fheorcm (;.
Let 8 be a fixed real number~ and show that the following two ma.trices
in A 2 are similar:

8 - cos
sin 8]
()

c?s
[ sin fJ

6.

287

and

[eift
0

O ].
e-t8

(/lint: let T be the operator on l: \Vhosc matrix relative to the basis

B ::::;: {e1 ,e 2]--'\\hcrc e1 ::::;: ( 1,0) and e2 = {O _, 1)-is the first of those
given~ and find another basi~ B' = {f1,/2} such that T/1 = ei 11f 1 and
Tf2 = e-tt1/2.)
Let T 1 and T2 be operators on H 1 and show that there exi~t bases
B and R 1 such that [T ds = [Tdn' ~ the-re cxi~t~ a. non-singular
operator A such that T ~ = A TlA- 1 (/lint: if l T dn = f1~~]B 1 1 let
A be the operator which carries B onto B 1 ; and if T 2 = ..tl T1A ~ 1 , let
B be any basis and B 1 its image under A.)

61. DETERMINANTS AND THE SPECTRUM OF AN OPERATOR

Determinants are often advertised to students of elementary
:mathematics as a computational device of great value and efficiency for
~l ving numerical pro bl ems involving systems of linear equations. 1;hi.s
is somcl\rh at mislca ding, for their val uc in pro blems of this kind is very
limited. On the other hand, they do have definite importance as a
th~oretical tool.
Briefly) they provide a numerical means of distinguishing between singular and non-singular matrices (and operators)~
This is not the place for developing the theory of determinants in any
detail~ Instead 1 we assume that the reader a] ready knows something
about them, and we confine ourselve8 to listing a few of their simpler
properties which are relevant to our present interests.
Let (aij] be an n X n matrix. The determinant of this matrix, \vhich
we denote by det([ai1]), is a sca1ar associated with it in 8uch a \\'ay that
(1) dct ai1]) :;::: 1 ;
{2) dct er a,;;] [~if]) = det([ a~jJ) det([.8,j]) ;
(3) deter lli1D r!- o {::=? ( aiiJ is non-singular; and
(4) det(rO.Ji] - ~[3;;]) is a polynomial, "With complex coefficients, of
degree n in the variable A.

e[

288

Operators

The determinant function det is thus a scalar-valued function of matrices

which has certain properties. In elementary work, the determinant of a
matrix is usually written out with vertical bars, as follows,

..

ar~

a1n

:i!:'2

a2n

a~2

. .
. ann
~

'

and is evaluated by complicated procedures which arc of no concern to

us here.
We now consider an operator Ton JI. Jf Band B 1 are ha.Res for H~
then the matrices [aiJ] and [i3i;l of T relative to Band B' m~1y be entirely
different, but nevertheless they have the &l.me determinant-. For '"Tc
know from the previous section that there exists a non-singular matrix
[1'ii] such that
[t3iA = [--r11]- 1[ai;"] ["Yi1J;
and therefore, by properties (1), (2) . and (3), we have
det([.BiJ-J)

dct({1'ii]-l[ni.'1[1'i.il)
= det (['Yii 1- 1) det ([ lliiD <let ([ 'Yi1])
= det([ "Yii}- 1) det 1'ii]) det([ O:i;])
= dct(["'f~11- 1 !'YiJD dct{{ 1J)
=== dct({ 61;}) det([ aii1)
= <let{[aiJ]) ~

This result allo\V~ us to speak of the detern1inant of lke operator T, meaning,

of course, the determinant of its matrix relative to any basis; and from
this point on, \Ve shall regard the dctcrmin ant function primarily as a
scalar-valued function of the operat-0rs on H. We at once obtain the
following four properties for this function, which a re simply trans la tio ns
of those stated above:
(I') det(J) = 1;
(2') det(T1T2) = det(T1) det{T2);
(3') det{T) ~ 0 <==> Tis non-singular; and
(41 ) det(T - >..I) is a polynomial, with complex eoefficients, of
degree n in the variable h,
We arc now in a position to take up once again, and to sett.let the problem
of t.hc existence of eigenvalues.
Let T be an operator on H. If we recall Prohlem 44-6, it is clear
that a scalar ). is an eigenvalue of T ~ there exists a non-zero vector x
such that (T - }...f)x = 0 <=> T ~ )J is singular ~ det(T - XI) ~ 0.
The eigenvalues of 7t are therefore precisely the distinct. roots of the
equation
(l)
det(T - Al) = _O,

Finite-dimensional Spectral Theory

289

which is called the characteristic equalt'on of T ~ It may illuminate ma ttcrs

somevlhat if we choose a basis JJ f(Jf H, find the matrix [a:;j] of T rela.tivc
to B, and write the characteri"tic equation in the extended form

. .
~

..

..

afl i

..

..

..

'r

,.

.,

.,

Onn -

..

Our search for eigenvalues of Tis reduced in this \Vay to a ~earth for roots
oi Eq. ( l). Propr:rty (4 1) tells us that this is a po]ynomial equation,
\vit.h cornplex coefficients, of degree n in the c.omp]ex variable X. \'1e nov{
appeal to the f!1ndame11tal theorem of a]gebra, \vhich guarantee~ that an
equation of thi8 kind al\vays has exactly n complex routs+ Some of these
roots may of course be repeated, in \vhich -ease there are fe,ver than n
distinct roots. In summary, lre have
If 'I' is an arbitrary operator on H, then the eigenvalues of T
constitute a non-em.pty fin1~te subsel of the com.plex planea F'uriherntore, the
num.ber of P'n~nts in thi.~ set does not excee.d th.P dim.ension n of the space !Ir
Theorem A.

1~he

set of eigenvalues of 1~ is caUed its spectr1t1n, and is denoted Ly

(J"(Tj. Ji""or future rcfercncci \\:e observe that u(T) is a compact subspace
of the romplcx plane.
It should no\V be reasonably clear w-hy v.Te required in the definition
of a Hilbert space that its scalars be the complex numbers+ The reader
~'idll easily convince himself that in the Euclidean plane the operation of
rotation about the origin through 90 degrees is an operator on this real
Banach space which has no eigenvalues at all, for no noo-zero vector is
transformed into a real multiple of itself. The existence of eigenvalues is
therefore linked in an essential way to properties of the complex numbers
which are not enjoyed by the f('al numbers, and. the most significant
of these properties is that stated in the fundamental theorem of algebraT
The mechanism of matrices and determinants turn8 out to be simply a
dcVice for making effective use of this theorem in our basic problem of
proving that eigenvalues exist. We also remark that Theorem A and its
proof remain valid in the case of an arbitrary linear transformation on
any complex linear space of finite dimension n > 0.
Problems

1.

Let T be an operator on H, and prove the following statements~

(a) T is singular <====> 0 E t:r( T);
(b) ii T is non-singular, then A e: q(T) :::} x- 1 E u(T- 1) ;
(c)

if A is non-singular, then a(AT A- 1)

:;;-

a(T);

290

Operators
if X e: u( T), and if p is any polynomial, then p(~) an (p( T));
if T~ = 0 for some positive integer k, then cr(T) = {O}.
Let the dimension n of H be 2, let B = fe 1~e2} be a basis :for H 1 and
assume that the determinant of a 2 X 2 matrix [ai.1] is given by
(d)
(e)

2.

a11a22 -

(a)

(b)

auau.

Find the spectrum of the operator Ton H defined by Tei = e2

and Te2 = -e1.
Ii Tis an arbitrary operator on H whose matrix relative to B is
[a.iiL show that T 2 - (a11
a22) T
(a11a~2 - aua21)I ;:;;;- 0.
Give a verbal statement of this result.

62. THE SPECTRAL THEOREM

We now return to the central purpose of this chaptcri namely, the

statement and proof of the spectral theorem~
Let T be an arbitrary, operator on H. \Ve know by Theorem 61-A
that the distinct cigen1.ralucs of T form a non-empty finite set of complex
numbcr~L I-Rt A1, X.2, .... , )l.m be these eigenvalues; let .Jf., ,J.'f 2, ~ . ,
Mm be their corresponding cigenspaces; and let P1 1 P1 1 ~ , P. be the
projections on these cigenspo. ces. We consider the foJlowing three
statements.
I~ The Al /s are pairwise orthogonal and span H.
II. The P/s are pairwise orthogonal, I = ~~ 1 Pi~ and T =
1 X"P;..
III. T is normal.
We take the spectral tlworem to be the assertion that these statements arc
all equivalent to one another+ It \Vas proved in the introduction to this
chapter that I===} II===} III. We now complete the cycle by showing that
III :;:;;;} I.
The hypothesis that T is normal plays its most critical role in our
first theorem.

i:.

Theorem At If T is norm.al, then x is an eigenvector of T with eigenvalue

A 8 x is an eigenvector of T * 1.JJith eigmvalue ;:. ~
PROOF.
Since T is normal, it is easy to sec that the operator T - A.I
(whose adjoint is T* - XI) is also normal for any scalar A. By Theorem
58-C~ we have

llTx - Xxll

!IT*x - Xx[I

for every vector x, and the statements of the theorem follow e..t once from
this.
The way is now clear for
Theorem B.

If T is normal, then tlu M/s are pairwise orthogonal.

Finite-dime n si ona I Sp eel r a J Theory

291

Let x..: and x1 be vectors in M1 and M; for i ~ J~ so that

Tx, = )\izt: and Tx1 = 'A;x;~ The preceding theorem shows that

PROOF.

Ai(Xi, x1)

and since Ai

~ Xh

(J\ai, Xj)

= (Txi Xi)
1

=:::;

= (Xi,~X;) =

it is clear that we must have

~i, T *xt)

Xi(Xi,Xj);
(xi~x1 ) =

0.

Our next step is to prove that the M/s span H when Tis norma1, and
for this we need the following preliminary fact.
Theorem C.

If Tis normal, then each kfi reduces T~

PROOF. It is obvious that each Af..: is invariant under Tt so it snfliccs,
by Theorem 59-C, w show that each Mi is also invariant under T*.
This is an immediate consequence of 1~heo rem ~i\., for if x. is a vector in
Mi~ so that Txi = ~Xit then 7 x-i: = X:zi is also in Af..:.
1

:Finally, '"Te have

Theorem D.

then the M*'s span H.

If Tis normal,

"fhe fact that the JI.ts are parrwisc orthogonal implies, by

Theorems 59~F and 59-G, that Af = M1 + M2
+ Mm is a
closed linear subspace of H~ and that its associated projection is
PROOF.

p = P1

+ P2 +

+Pm,.

Since each M" reduces T, we see by Theorem 159-E that TP,; = P..:T for
each Pi. It follows from this that TP ~PT, so M also reduces T, and
consequently 21! 1- is invariant under 7ra If i.lf 1- ~ 0}, then~ since all the
eigenvectors of T arc contained in M, the restriction of T to Ml. is an
opera tor on a non-trivial finite-dimension.al H i1 bcrt space which has no
eigenvectors, and hence no eigenvaluesa Theorem 61-A shows that this
is impossible~ We therefore conclude that M 1- = {0}) so M = H and

the .LY/s span H.

This completes the proof of the spectral theorem and, in particular,
of the fact that if T is normal!t then it ha.s a spe.ctral resolution

T ;: h1P1

+ AIJP2 +

4
'

+ AmP~~

(1)

We no,~l make several observations which will be useful in carrying out

our promise to show that this expression for Tis unique. Since the P.:'s
are pairnisc orthogonal, if we square both sides of (1) we obtain
m

T2

)..,/P1~

oja.1

More generally, if n is any positive integer, then

m

Tn

==;

.I:

1-1

')...i'* Pl.

(2)

292

Operators

If ,vc make the customary a-greemcnt that T 0 = I, then the fact that
I = :z 1: 1 P"' shows that (2) is also valid for the case n = O~ Next, let
p(z) be any polynomial, \vit.h eomplex cocffieicnts 1 in the complex va1ia.ble
z. Ry taking linear combinations_, (2) ca.n c'idently be extended to
m.

p(T)

(3)

p(Xi)P1.

t .1::11" 1

\\re 'would like to find a polynomial p such that the right side of (3)
col.lapses to a specified one of the P/s, say Pi ''rhat is nccd~d is a
polynomial p1 '\vith the property that Pi(X~) = 0 if i ~ J and PJ(X1) = 1.
\1.,le delinc Pi as f ollo\VS:

pz
( )
)

(z -:- X1)
(z - Ai-t) (z - AJ+t) (z - Am)
.
(Aj ~ h.1) ' ('J...j - AJ-1) (AJ - A1+1) ~ ' (A - ~m)
+

Since Pi is a polynom1al, and since p1 (Xi)

=:.:

Otn (3) yield~

(4)

Pi= P1(T).

In order to interpret these remarks to our advantage, we point out that

only three facts a.bout (I) have been used in obtaining (4): the X/s are
distinct complex numbers; the Pis are painvisc orthogonal projectionH;
and I = ~f: 1 Pi. By using these properties of ( 1), and these a] one, "\Ve
havP. sho\vn that t..he J)/8 arP uniquely determined as specific polynomials

in T.
We no'v as8nmc that

"\VP

have another expression for T similar to (I)>

(5)

and that this is also a Hpectral resolution of T, in the sense that the a/s
are distinct complex numbersi- the Q/s arr. non-zero pairwise orthogonal
projections, and I =
";-e wish to show that (5) is actually
1 Qj.
identical ":-ith (1), except for notation and order of terms. We begin by
proving, in t\vO steps,. that the a/s are precisely the eigenva1ues of T .
.First, since Qi. ~ O, there exists a non-2'ero vector x in the range oi Q*; and
since Q~-x = x and Qix = 0 for j ~ i, we see from (5) that 71:c = atX, so
each ai. is an eigenvalue of '1'. Next, if A is an eigenvalue of T, so that
Tz = Xx for some non-zero x, then

2::

Tx = Xx = A.Ix = A

!r Qa I
===

~=l

il

and
so

Tx =

i (X i=-1

!r a1Qat

il

lti)Qix

0.

XQa

Finite dime nsi ona I Spectra I Theory

293

Since the Q.-x' s a.re pairwise orthogonal, the non-zero vectors among
them~there

is at le.a.st one, for x ~ 0---are linearly independent, and

this implies that A = a..: for some i. These arguments show that the set
of a/s equals the set of X./s, and therefore, by changing notation if
necessary, we can write (5) in the form
(6)
The discussion in the preceding paragraph now applies to (6) and gives
(7)
for every j+ On comparing (7) \vith (4), vlc see that the Q/s equal the
P/s. rrhis shows that (5) is exactly the game as (l)~except for notation
and the order of terms-and completes our proof of the fact that the
~per lra1 re solution of T is unique.
We conclude with a brief look at the matrix inlerpreta.tion of statements I and 11 at the beginning of this section. Assume that I i8 t.ruet
that is~ that the eigenspaces ~M 11 1JJ2, . . . , ."41
. m of '1 a.re pair,vise orthogonal and span H. For ea.eh Mi, choose a basis "\vhieh consists of mutually
orthogonal unit vector8. rrhis can a.lv;ays be done, for a basis of this
kind-called an orthono-rrnal bas-is-is prceisciy a complete ort.honol'mal
set for JI,.. It is Pasy to sep, tl1at the union of thesP. little bases is an
orthonormal basis for all of ff; and relative to thist the matrix of T has
the follo\ving diagonal form (all entries off the main diagonal are understood to be 0):
1

(8)

294

Operators

We next assume that II has an orthonormal basis relative to which the

matrix of Tis diagonal. If we rearra11gc the basis vectors in such a \vay
that equal matrix entries adjoin one.another on the main diagonal,, then
the matrix of T relative to this new orthonormal basis will have the form
(8). It is easy to see from this that T can be written in the form

\\here the -X/s are distinct complex numberst the P/s are non-zPro pairwise
orthogona1 projPr.tions, and I = :!:f' 1 Pi. rrhc uniqucnes8 of the spectral
resolution no\v guarantees that the A./s are the distinct eigenvalues of T
and that thP P/s are the projections on the corresponding eigenspaces.
The spectral thPorPm tPlls us that statements It II, and Ill are equivalent
to one ano t.her. 1~he a hove remarks earry us a bit f urthcr i for they
constitute a proof of the fart that these statements a.re also equivalenl to
IV. There exists an orthonormal basis for H relative to \vhich the
matrix of 1T is diagonal.
It i8 interesting to realize that the implication III ===:>IV, which '\Ve
proved by sho\\"i ng that Ill ==? I a n<l 1 ~ I\!~ can be made to depend
more direet]y on matrix computations. 'This proof is outlined in the
]ast three problems below.
Problems

l.

24

Show that an operator Ton His normal=> its adjoint T* is a polynomial in T.

Let 'I' be an urbitrary operator on H~ and N a normal operator.
Sho\v that if T commutes \vith .l\r, then T also commutes v.."'ith N*
L{~t 'l' he a normal opera tor on H ,vi th spectrum {X1, A. 2 , . . , Ami ,
and use the spcetral resolution of T to prove the follo,ving statements:
(a) 1~ is self-adjoint ~ each X" is real; (b) Tis positive <=} Ai > 0 for
each i; (c) T is unitary <:=:> rxij = 1 for each i.
Sho\v that a positive operator T on H has a unique positive square
root; that is, show that there exists a unique positive operator . .4 on
II such t.hat A~ = T .
Let. R ~ {e1t e2t
en} be an orthonormal basis for H ~ If T is
an operator on H \vhose matrix re]ative to B is [ooJ], show that the
ma. lrix of T* relative to B is ['31)J, ,vhere f3ii = -;_;;;~ [p~i] is of ten called
thr, conjugate transpose of [ai1].
Let T be an arbitrary operator on H, and prove that there exist n
closed linear subspaces M 1, M 2,
M ~ such that
r

3.

4.

5.

64

{ 0}

C M 1 C ltf ~ C

C !Jf" == II 1

the dimension of cac h M, is i. nnd each nl i is in variant under T

finite-dimensional Spectral Theory

1, the statement is clear; and if n > 1, assume it for all

Hilbert spaces of dimension n - 1, and prove it for II by using
1.'heorem 59-B and the fart that T* has an eigenvector.)
Let The an arbitrary operator on Hi and use the previous problem
to show that there exists a basis B re1ative to which the matrix
[a,;i] of T is triangular, in the sense that i > j ===> aii = 0. If T is
normal, show that there exists an orthonormal basis B' relative to
which the matrix of T is diagonal. (Hint." generate B' hy applying
the Gram-Schmidt process to B, observe that the matrix of T relative
t-0 B' is still triangular, and use Pro blem 5 to show that this matrix is

(Hint.. if n

7.

295

actually diagonal.)

63. A SURVEY OF THE SITUATION

Tiie spectral theorem is often stated in a somc-u~hat more restricted
form than that given in the previous scctionr The usual version is that
each normal operator Non H has a spectral resolution~ that is~ that there
exist distinct complex numbers Xb ~~t , Am and non-zero painvise
orthogonal projections P1, P?.,
~
pm such that ~:.l pi = I, with the
property that
I

'

"'

IA?t

(1)

:i= 1

In our version, we attemptred to give equal emphasis to both the geometric and the algebraic sides of the matter. Most v-.""riters~ however,
con fine their statement of the theorem to that given above~ and for a
very good reason: it is ( 1) tba t generalizes to the infinite-dime nsiona I case~
There are two ways of carrying out this generalization, and we give a
brief description of each.
First, there is the analytic approach. For the sake of simplicity t "\Ve
consider a self-adjoint operator A, and we write (1) in the form
(2)

Our reason for making this assumption is that the eigenvalucB of A are
real numbers and a.re therefore ordered in a natural way~ We further
assume that the notation in (2) is chosen so that A1 < x, < . < ~m,
and we use the PlB to define new projections:

EA. = O;
E>.1 = P1;
E).I :=: P1 + P2;
ii

..

..

E~=P1+P,+

..

..

..

..

. +Pm.

296

Operators

(The subscript Xo iB introduced solely for notational convenience and has

no significance beyond this.) ThcE)i,.'s enable us to rev.,..rite (2) asfollo"\\rs:

+ A2P2 + ~ + "AmP~
X1(b').l - E).(}) + X2(E>..* - EAJ + ... ~ + A-m(Ex.,.

A = X1P1

- E'}.,1J'o_J

?Pl-

Ai(E}.j -

EAi-1).

t: =I

If we denote E\i - E-,..;_, by l!J.E~.., then "\\'e ean compress this to

m

A ::::::

I..:-1 'A\ tJ.E~),,iJ

'v hi ch suggr sts an in t.egral re prcsen tatio n

(3)
In this form, the spectral resolution remains valid for self-adjoint operators on infinite-dimensional Ililbe.rt spaces. ..A. similar result holds for
normal operators,
(4)
N
XdE,.,.

=I

There are many difficulties to be surmounted in rc-a~hing the level of (3)

and (-i). \\re haYe alteu.dy met one of these, namely, the fact ihat. an
operator 1' on an arbitrary llilbert space 11 r!- {0} need not have any
eigenval ucs at all.. In this general case, th c spectruui of T is defined by
<1(T) = {X: T - XI is singular}.

lhcn II is finite-dimen8ional, \Ve have seen that D'(7') consists entirely of

eigenvalues. This made our "'ork in the present chapter relatively easy~
but it is not true in general. What is true is that <1(T) is a.l\vays nonempty, closed, and bounded~ and is thus a compact subspace of the
complex plane. Once this difficulty is dealt with, there remain substantial problems in giving meaning to integrals like those in (3) and (4)
and in proving the validity of these relatio ns4 1
The second approach to generalizing (1) is essentia11y algebraic and
topological in nature. I ts starting point li3 the o bse rva tion made in the
previous section that the spectral resolution
.,.,.,

N =

).J',

1=1

For a. general discuBBion of t.he spectral thoor~m from this .an.a.lytie point of
view, see LoTch [28). A full trep.tment can be found in Riesz and Sz.-Na.gy [3..3,
chap. 7].
1

Fi nife..dimensiona 1 Spec:tra1 Theory

2'17

1ead.s to, and is actually part of, the fart that

m

p(1"l) -= ~ p("A~)Pi

(5)

i=l

for any polynomial p. The set of al I polynomials in N is evirlcn tly an

algebra of operators, a sub.algebra of IB(H). Let us no\v consider t.he
corresponding alg~bra of all polynomial functions defined on the set of
A./s. "\Ve have seen that this algebra contains polynomials Pi surh that
p;CA.,-) = aiii and it therefore con8ists of all complex functions defined on
the set of X/s~ If \\'e denote the latter set by X for a moment and think
of it as a compact su bspacc of the complex plane, then, since X is finite,
the algebra in quP.stion is precisely e(X), the algebra of all continuous
complex functions defined on x. rrhe mappiug

p(1V)

~ p,

(G)

\vhich make8 eorrPspond to c-ach p(N) the function p in e(X), is easily

seen by the properties of (5) to preserve all algebraic operations+ \"\7 e
kno'\v from the previous section that the eigPnvalues of p(l'l) are-with
possible repetitions~t.he p(X..)'s on the right of (5); and 'vc shall see
later, as an nnexpPcted dividend, that the norm of a normal operator
allvays equals the maximum of the absolute values of its eigenvalues.
It follo'\'"S from these remarks that the mapping (6) is an isometric
isomorphism of the algebra of all p(N)~s onto e(X). '"fhr.se ideas constitute an extended ver~ion of (5) and arc thus! in a sense, a generalization
of the spectral resolution oi N. They apply virtually \vithont change to
the ca8B of a normal operator on an infinite-dimensional Ililbcrt space,
a.nd one of our aims in the next three chapters is to treat them in detail.

PART THREE

Algebras o. Operators

CHAPTER TWELVE

(jeJteral
Preliminaries ou !!auack Algebras
Our work in Part 1 of this book was primarily concerned with
topological spaces and the continuous functions (~arrie<l by them.
The ideas of Part 2, on the other hand, \Vere essentially algebraic in
naturcr The function spaces \Ve encountered earlier led us to begin a
study of Banach spaces for their ov.rn sake~ and as \ve procccde<l 1 we
found our attention focusing more and more closely on the properties
of their operators. Except for a few elementary notions about metric
spacesJ v,re used very little genuine topology in Part 2. As a mat.ter of
fact, our treatment of spectral theory in Chap. 11 'vas completely independent of topology~ for in the fi ni te-dimen sional case, all Ii near trans-

formations are continuous~

In the following three chapters, these two apparently diverse trains
of thought-the topological and the algebraic-are united by a single
elegant concept: that of a Banach algebra. Our remarks in the last
section of the previous chapter suggested that there may be important
links between algebras of operators on Hilbert spaces and algebras of the
type e(X), where Xis a compact Hausdorff space. Banach algebras are
the svstems
which enable us to establish these connections on a firm
... .
footing. They are also interesting in that they constitute a field of study
in 'vhich a wide variety of mathematical ideas meet in significant contact+
Our main task in the present ch.apter is to provide a number of
miH ccllaneo us tQols \vhich are ne cc ssa. ry i or the structure theory de velo pcd
later.
301

302

Algebras of Operators

64. THE DEFINITION AND SOME EXAMPLES

A BaTUJ.ch algebra is a complrx Banach space which is also an algebra
with identity 1~ and in ~Thi ch the mul tiplica ti ve structure is rcla ted to the
norm by the following requirements:
(1) Hxyll < Hxll llYll;
(2) U1 [I = 1~
It follows from (1) that multiplication is jointly continuous in any
Banach algebra, that IB 7 that ii Xn----+ x and Yn ~ y,. then XnY"' --+ xy
(proof:
UxnYn - xy!I

[lx~(YR

- y)

+ (Xn -

x)ylj S If x,.H llYn ~ YH

+ Uxn -

x[I HY ti)~

A Banach aubalgebra of a Banach algebra A is a closed subalgebra of A

v. hich contains I. The Banach subaigebras of A arc precisely those
subsets of A which are themselves Banach algebras with respect to the
same algebraic operations 1 the same identity, and the same norm.
The definition of a Banach algebra is sometimes given without the
restriction that the scalars are the complex num1?ers. The complex case,
however, is the only one that concerns us, and by framing the definition
as we do, vre a void the necessity of treating the additional complications
\vhich arise in the real case.. We have further assumed, for the sake of
simplicity, that every Banach algebra has an identity~ It is possible,
at a considerable sacrifice of clarity~ to develop most of the important
ideas without this assumption, and this is done whenever the primary
purpose of the theory is the study of group algebras of locally compact
but not discrete groups+ Since our attention will be clirected chiefly to
the structure of operator algebrasi there is no need for us to strain for the
added generality obtained by not requiring the presence of an identity.
The Banach algebras of principal interest to us are described in the
following examples. The reader will notice that they all consist of
functions or operators and that the linear operations in all of them are
defined pointwise. They can be classified in a general way into function
algebraR,. operaror algebras, or grou.p algebras, according as multiplication
is defined pointwise, by composition, or by convolution .
1

One of the most important Banach algebras is the

set e(X) of all bounded continu~us complex functions defined on a
topologi~l space X. The case in whlch X is a compact 1Iausdorff space
will have particular significance for our later work4 If X has only one
point, then e(X) can be identified with the simplest of all Banach algebras~
the algebra of complex numbers.
(b) Consider the closed unit disc D ~ {z: lzl < 1} in the complex
plane. The subset of e(D) whlch consists of all functions analytic in the
Example 1..

(a)

General Preliminaries on Banach Algebras

303

interior of D is obviously a subalgebra which contains the identitya A

simple application of Morera's theorem from complex analysis sho"\\TS that
it is closed and is therefore a Banach subalgebra of e{D).. This Banach
algebra is called the disc alge"bra+ It has a number of interesting properties, which are, of course, intimately related to the special character of
its functions.

Example 2. (a) If B is a non-trivial complex Banach space, then the

set ffi(B) of all operators on Bis a Banach algebra.. We assume that Bis
non-trivial in order to guarantee that the identity operator is an identity
in the algebraic sense4
(b) If we consider a non-trivial Hilbert space H~ then IB(H) is a
Banach algebra. This is a special case of <B(B)~ and it is important to
observe that additional structure is present here, namely, the adjoint
operation T--+- T* ~
(c) A subalgebra of <B{H) is said to be self-adjoint if it contains the
adjoint of each of its operators4 Banach subalgebras of &(H)'s v.rhich
a re self-adjoint are called C*-algebras. We shall return to the subject
of commutative C*-algebras in Chap . 14.
(d) The weak ope-ratM topology on CB(H) is the weak topology gen..
erated by all functions of the form T--+- (Txty); that is, it is the weakest
topology with respect to ~~hich all these functions are continuous~ It
is easy to see from the inequality I(Tx,y} - (ToX 1 y)~ < I T - ToH llxB ~IYll
that this topology is weaker than the usual norm topology~ so that its
closed sets are also closed in the usual sense~ A C*-algebra with the
further property of being closed in the weak operator topology is called
a W*-algebra~ Algebras of this kind are also called rings of operal.mst
or von Neumann algebrasr They arc among the most interesting of all
Banach algebras, but their theory is quite beyond the scope of this book.'
Example 3. (a) If G ~ ~U~ g2, . . . _, Yn l is a finite group 1 then its
group algebra Li(G) is the set of all complex functions defined on G.
Addition and scalar multiplication are defined pointwisc, and the norm
by 11 fl~ = :Z7_ 1 If(gi) ~ . In order to see what underlies the definition of
multiplirationt it is convenient to regard a t.ypical element f of L1(G}
as a formal sum ~~ 1 aifJi~ where ~ is the value off at U~ With this
interpretation_, we use the given multiplieation in G to define multiplication in L1(G), as follows:
ta.

<I

i=-=1

where

~tui)

CI

,, I: ====

"

/jj(Jj)

Jflffl

'Ykulq

(1)

k-1
g"j

(J-;-fl~ aJ3,.

(2)

The meaning of the sum in (2) is that the summation is to be extended

1

s~e

Dixm ier !7l.

304

Algebras of Operators

over all subscripts i and j such that g~-g1 = (}1t In effect, thereforeJ we
formally multiply out the sums on the left of (1), and we then gather
together all the resulting terms which contain the same element of G ~
With these ideas as an intuitive guide, we revert to our first point of
view, in which the clements of L1{G) are functionsJ and we see that our
definition of multiplication can be expressed in the following way. If
two functions f and g in JJ 1(G) are given~ then their product, which is
denoted by j * g and ca]led thPir convolution7' is that function whose
value at g1: is
(f

* g) (g,J

~Oi!Jj=tHr f(gi)g(gi)

n.

(3)

ffrh:YJ-l)g(gi).

j=l

We note that if each element of G is identified with the function whose

value is l at that element and 0 elsewhere, then G becomes a subset of
L1 (G). l~urthor, multiplication in G agrees with convolution in L1(G) ~
and thr. element of L1(G) which eorrcsponds to the identity in G is an
identity for L1(G). \\le con.elude this description by observing that every
element of G has norm 1, so t.hat 111 H = I J and that the basic norm
inequality for a Banach algebra i~ satisfied ~
f!

IJ/ * o~I ~
==

t (f * o)(gk)I
t I t f(gkg.i~)g(g;)I
t

k =-= l
71

1-1

j:c::l

t t

tL

ft

<

lf(glg;- 1)1 lg(UJ)I

k-li-1
n
n

I.
t
If (gkgj- 1)I Iu(gj)
1-1
= I lg(g1)I t lf(OlUii-1
= I. Jg(g;)[ l!/H .
:i
=

k~l

)[

i:l

-n.

= 11/U

i =-l

IY<ui)I

= 11111 lluH-

f...

0, lJ 2, . ~ . } be the additive group

of integers. Its group algebra L1(G) is the set of all complex functions f
defined on G for which x=--cio 1f(n)' converges. The linear operations
(b)

Let G

~ -2, -17'

General Prelimjncries on Banach Algebras

305

are defined pointwise, the norm by ~I/II = I::~ ~f(n)', and the convolu.tion off and g---see Eq. (3)~by
~

(f * g) (n) ~

I.

m= -

f(n - m)g(m).
llJ

.Just as in (a), G is r.ontained in L1(G) in a nature.1 wayJ and /.;1(G) is a

Banach algebra~ A..ny attempt to discuss the group algebra of a nondiscret.e topological group like the re--al line must clearly be based on a.n
adequate theory of integration. It should also have available a theory
of Banach algebras in which no identity is assumed to be present. These
ideas constitute a rich and beautiful field of modern analysis. They are,
however, outside the scope of this work~ 1
The llanach algebras described above are many and diverse~ and
there are yet others which ,,. e have not mentioned. Our attention in
the following chapters \Vill be centre.red on e(X)'s and commutative
C*-algebrasJ hut the general theory we develop is equally applicable to
alL It is worthy of notice that an arbitrary Banach algebra A can be
regarded as a Banach subalgcbra of CB(A)~ In a senscJ therefore, Example 2a and its BaHach suba]gebras includP an possible Banach algebra~.
'fo see this, 've recall from Problem 45-4 that a ---io A11 a, where M a(x) = ax,
is an isomorphism of A into ill(A). It is easy to see that Ml is the identity
operator on . .4, so all that remains is to observe that Ila Jr = ~IM4H for every
a (proof: llM u(x) 11 = HaxlJ ~ lla~I llxl! shows that ~!.W Q;tl < JlaH, an<l the
fart that IJall ::; l]Mal1 follows from

IjMB Il

SU p

{ lIM4 ( x) lI : Hx 1!

<

I ~

> [I}/a (I) I~

:=:.

~I a 11)

1'1he mapping a-----=.. JI a is thus an isometric isomorphism of . .4 onto a

Banach subalgebra of IB(A), and it allows us to identify the abstract
Banach algebra A \vith a concrete Banach algebra of operators on A.

65 .. REGULAR AND SINGULAR ELEMENTS

Let .11 be a Banach algebra. We denote the set of regular elements

in A by G, and its compJement, the set of singular clements, by S. It is
c1ear that G contains I and is a group, and that S contains 0. Several
important issues depend on the character of G and S. Our first result
afong these lines is
1

Loomis [27j is t.he standard reference in thls subject. For a general exposition
of the main id,eas~ see Mackey [30]4 A brief treatment or the classical analyeie which
underlies the modern theory can be found in Goldberg fl4JT

306

Algebras of Operators

Theorem A. Every element x for which llx - I 11 < 1 is regular, and the
inverse of such. an element iB given by the formula X- 1 = 1 +
l ~ x) n ..
PHOOF.
If we put r = llx ~ l[I, so that r < 1. then

i;=- (

shows that the partial sums of the series 2;:=- 1 (1 ~ x)~ form a Cauchy
sequence in A. Since A is complete~ these partial sums converge to an
element of A~ which we denote by ~~::::ic 1 (l - x)n4 If we define y by
y = 1 + ~::.. 1 (1 - x)tt, then the joint continuity of multiplication in A
implies that
OD

JI -

xy =

(1 - x)y

(1 - x)

.m;i.

+k

(1 - x)n

n=2

so xy = 1.

Similarly, yx

:k (1 -

x)ft

y -

1,

n l

I.

We no~ use this as a tool to prove

G is an open set1 and there/are Sis a closed set.
PROOF.
Let Xo be an element in G, and let x be any clement in A such
that !Ix - Xo~I < I/Uxo- 1 ll.. It is clP'.ar that

Theorem B.

so we see by Theorem A that xo- 1x is in G.

follows that x is also in G, so G is open.

Since x = xo(xo- 1x), it

It was shown in Problem 32-5 that every Banach space is locally

conn cctcdJ so A is also 1ocally connected~ A dire ct a pplica ti on of
Theorem 34...A yields the fact that the components of Gare themselves
open sets .
As our final resultJ we have
Theorem C. The -mapping x --4 x- 1 of G into G is continuous and is
lherefore a homeomorphism of G onto itaelf.
PROOF..

that

Let

Xo

Hx - xoH <

element of G, and x another element of G such

1/(2llxo-1 ll). Since

be

an

II xo- 1x ~ 1U = ljxo~ 1 (x - xo) ~l

<

IJ Xo~ 1 II ll.x - x(}ll

we see by 'Theorem A that x 0- 1x is in G and

ICI

X- 1Xo =

(xo- 1x)- 1 = 1

n=l

(1 - xo- 1x)~.

< ~2

General Preliminaries on Banach Atgebras

307

Our ooncl usion now follows from

l!x- 1

xo- 1 H

l (x- 1xo

~ l)xo- 1 11

< Hxo- ~r llx-xo - 1 fl

1

II ~

n~l

==-

J!xo- 11
1

(1 -

xo- 1 x)~t1 ::; l!xo- 1 l1 ~II I -

XQ-

n~l

x[I"

[j 1 ~ xo- 1xl1

llxo- 1 11

Hxn-1 l
-

1~

<

I; H1 ~ xo- xlt"
1

ncmO

1ll -

Ul

xo- 1xU

~ xo~ 1 xl1

2llxo~ 1 I! 111 - xo- 1x[I

< 2Hxo~ 1l 2l!x - xoll

1

If ~t is an element in AJ it should always be kept in mind that the

regularity or ~ingularity of x depends on A as well as on x itself. If x
is regular in A, and if we pass to a Banach subalgcbra A 1 of A which also
contains x, then x may lose its inverse and become singular in A 1 By
the same token, if x is singular in A, and if A is regarded as a Banach
subalgebra of a larger Banach algebra A''t then x may acquire an inverse
and become regular in A''- In the next section, \Ve study certain elements in A which arc singular and remain singular 'vith respect to all
possible enlargements of A.

66. TOPOLOGICAL DIVISORS Of ZERO

.A.n element z in our Banach algebra A is called a lopo logical divisor

of zero if there exists a sequence {z"} in A such that llznll = I and either
zz.,. ~ 0 or ZnZ--.... 0.. It is clear that every divisor of zero is also a topological divisor of zeroL We denote the set of all topological divisors of
zero by Z.
Theorem A.. Z is a subset of S.
PROOF~ Let z be an element of Zand {zn} a sequence such that j[zfJlt = 1
and (say) zzft ~ 0.. If z were in G~ then by the joint continuity of multiplication we would have z:-l (zz") ;:;:: Zn -----t ot contrary to 11 z~U = l ~
Our next theorem relates to the manner in which Z is distributed
"\Yithin 8 .

Theorem B. Th boundary of Sis a subset of Z.

PROOF. Since S is closed, its boundary consists of all points in S which
ate limits of convergent sequences in G. We show that if z is such a
poin t7' that is, if z is in S and there exists a sequence {r,, } in G such that
r~ ~ z~ then z is in Z. First, we see from r,.-- 1z - 1 -::;: ; Tn~ 1 (z - r,.) that

308

Algebros of Operators

the sequence ~ rn- 1 } is unbounded; for otherwise, we """ould have

llrn- 1Z

1!I

<1

for some n, so that r n -lz~ and the refore z = r n (r n -lz) J would be regular.
Since {r "- 1 J is un bounded 1 we may assume that Hr ft.- 1 II ~ co If Zn i~
now defined by z1J = rn - 1/l]rn - 1 then our conclusion follows from the
observations that Hzfll! = 1 and

a,

zz~ =

ZTtt-l

llru H=

l+(z - rn}rn-l

-1

l!r.-111

- Ur"-1 [! + {z ~

rn)Zm ~ o~

In order to understand the significance of these facts, let us suppo8e

that A is imbedded as a Banach subalge bra in a larger Banach algebra A 1
As we remarked in the previous section~ an element which is singular in A
may cease to be so in A'. However, if it is a topologir.al divisor of zero
in A t then it is in A' as well, so it is singular in A'. The topological
di visors of zero in A are th us ~'perm.a nently singular~,~ in the sense that
they are .singular and remain so with respect to every posfilble enlargf"ment of the containing Banach algebra. Theorem B tells us tha.t no
matter what happens to Sas a whole in such a proccss 1 its boundary is
'"permanentn in this sense.

67. THE SPECTRUM

Let T be an operator on a non-trivial Hilbert space.

chapter, we defined the ~pectrum of T to he the set
a(T)

J'}-..:

In the previous

T - XI is singular} J

and we devoted a good deal of attent1un tu the geometric ideas leading

to this concept. We found~at least in the finite-dimensional case-that
a number in u(T} is a value assumed by T 1 in the sense that T actR on some
non-zero vecto1 as if it were scalar multiplication by that number. \Ve
shall see later that this formulation of the meaning of the spectrum haR a
much '\\"'ider significance than we might at first suspect.
Let us now consider an element x in our general Banach a1gcbra A.
By analogy with the above, we define the spectrum of x to be the following
subset of the complex plane:
a(x) = fX :x -

Xl is singular}+

'Vhenevcr it is desirable to express the fact tba.t the spectrum of x depends

on A as well as xj we use the notation uA{x)a It is easy to sec that x - hl
is a continuous function of A with values in A; and since the set of singular
e1ements in A is closed~ it follows at once that q(x) is closed. We further

Generar Pretimino ries on Banach Algebras

309

observe that a(x) i...~ a subset of thP eloscd disc fz: lzl < llxl~} t for if X is a
complex number such that IXi > ~lxll ~ then ~Ix/All < 1t I! 1 - (1 - x/X) II
< 1, 1 --- x/X is regular, and therefore x - Xl is regular.
Our first task is to establish the fact that a(x) is al,vays nou-empty,
a.nd for this we need a few preliminary notions~ The resolvent set oi x,
denoted by p(x)t is the complement of a(x); it is clearly an open subset
of the complex plane which contains {z: lz! > lix~l l. The resolvent of x
is the function with values in A defined on p(x) by

Theorem 65-C tells us that x (X) i8 a continuous function of >.. ; and the
fact that x(X) = x~ 1 (x/X - I)- 1 for X ~ 0 jmplies that x(k) ~ 0 as
A ---i- ~. If "J.. and are both in p(x)~ then
x(X) ;; x(X}[x - ,ul ]x()
= x(X)[x - 1\1 (>-. - ) l]x()
= [1
(J\ - )x(X)]x()

x("J...)

+ (A -

)x(X)x(.u) ~
x (.u) = (A - )x (;\)x ()

= x()

so

Th is relation is ealled the resolvent equation.

o-(x) is non-empty.
PROOF.
Let j be a functional on A--that is, an element of the conjugate
space A *-and define/(">..) by /(X) = f(x(X) )T It is clear that/('>..) is a
complex function v.'hich is defined and continuous on the resolvent set
p(x)+ The resolvent equation shows that
Theorem A.

f (A)

f (If~ = f (x(A}x())

A-

and it follo,\fs from this that

lim /(X) -- /{.)
A~~

/(x() 2 }J

A - .

so f{X) has a derivative at each point of

p(x)~

Further 1

0 as X --+ oo . We now assume that u(x) is empty 1 so that p(x)

is the entire complex plane~ Liouville's theorem from complex analysis
allows us to conclude that /(X) -;;::; 0 for all A~ Since f i9 an arbitrary
functional on A, Theorem 48-B implies that x(X) = 0 for all Xa This is
impossible, for no inverse can equal 0 1 and therefore it cannot be true
that d'(x) is empty.
so j(X)

--+

310

Algebras of Operators

If the reader is surprised by the appearance of Liouvillc's theorem

in such a context, he should recall two factsa .First, our proof of Theorem
61 ~A~ which is a special case of the above result~ required the use of the
fundamental theorem of algebrar And second 1 the fundamental theorem
of algebra is most commonly proved as a simple consequence of Liouville~s
theorem. It is therefore only to be expected that somP. tool from analysis
comparable in depth with Liouville~s theorem shouid b~ necessary for
the proof of Theorem A.
Now that we know that a(x) is non-empty, \Ve also kno\ll that it is a
compact subspace of the complex plane. rfhe number r(x) dPfined by
r(x)

sup

rl :\I :)\ e: a(x}}

is called the spectral radius of x. It is clear that 0 < r(x) < l~x\I
The
eon ce pt of th c spec t. ra1 radius \Vill be uscf ul in r,e rtain parts of our later
a

"Tork.

We recall that a division algebra. is an algebra with identity in which

each non-zero element is regular.
t1uencc of Theorem A is

The most important singte conse-

Theorem B. If A is a divisimi algebra:r thr,n it equals tJw set of all scalar

niulti ples of the i.dentity ~
PROOF~
We must sho,N that if x is an element of A, then x eq ualt5 A. 1 for
some scalar X. Suppose, on the contrary, that x #- XI for every "A.
Then x ~ X1 =F- 0 for every }..., x - Xl is regular for every X, and the refore
q(x) is empty. This contradicts Theorem A and completes the prooL
The mapping Xl ~ X is clearly an isometric isomorphism of the set
of all scalar multiples of the identity onto the Banach algebra C of ali
complex numbers. We may therefore identify this set with C; and in
terms of thif.; identification, Theorem R says that any Banach algebra
which is a division algebra equals C~ This fact is the foundation on which
'~Te build the ~tructurc theory presented in the next r.ha pter
It is obvious that C itsclft which is the simplest of all Ranach algebras, is a division algebraJ so Theorem B characterizes (l as thP" only
Banach algebra with this property.. In the next two theorems~ 've give
some other in tcrcsti ng characteri.za tions of C among al1 pos~i b le Banach
r

algebras.
Since 0 is a divisor of zero, it is a topological divisor of ze10 in every
Banach algebra. In the Banach algebra G, 0 is plainly the only topologieal di visor of zero. Conversely, \Ve have

If 0 is the only topological divisor of zero in A then A == C~

PROOF.
I.ret x be an element of A~ Its spectrum u(x) is non-emptyt so
it. has a boundary point X; and x - Xl is easily seen to be a boundary
Theorem C.

General Preliminaries on Banach Algebras

311

point of the set S of all singular elements. By Theorem 66-B, x - Xl

is a topological divisor of zeroj so it follows from our hypothesis that
:t - Al = 0 or .z = Al.
The basic link between multiplication in A and the norm is given
by the inequality l]xylJ :5 llxtl UY ti, and when A = C, this inequality can
be reversed. The following result shows to what extent this reversibility
is true in general .
Theorem D. If the norm in A satisfies the inequality
for 8ome posi.tive constant K, then A ~ C~

IJxy II >

KJlx It I~ y H

PROOF~

In the light of Theorem c7' it suffices to observe that the

hypothesis here implies that 0 is the only topological divisor of zero.

We next look into the question of what happens to the spectrum

of an element x in A when A is enlarged .
If A is a Banae.h subalgebra of a Banach algebra A 1 , then
the spectra of an element x in A with respect to A and A 1 are related as f ollou;s:
(I) ctA'(x) C ttA(x); (2) each boundary point of aA(x) is alBo a boundary
point of trA'(x).
.
1
PROOF. If x - Xl is singular in A J then it is certainly singular in A,
so {l) is clear~ To prove (2), we let}... be a boundary point of crA(X)~ It
is easy to see that x - Xl is a boundary point of the set of singular elements in A, so by Theorem 66.. B7' it is a topological divisor of zero in A .
It is therefore a topological divisor of zero in A' as well7' so it is singular
in A' and X is in <$.it~(x). The fact that >.. is actually a boundary point
of (!Ar (z) is immediate from (l), so the proof of (2) is completer

Theorem E.

This result shows that in general the spectrum of an element shrinks

when its containing Banach algebra is enlarged, and further, ~hat since
its boundary points cannot be lost in this process7' it must shrink by
"hollowing out . H An illuminating example of this phenomenon is
provided by the disc algebra A of all complex functions which are defined
and continuous on D = {z: lz! :S l} and analytic in the interior of this
set. If f is a function in A, then the maximum modulus theorem from
complex analysis implies that

Ii/II

= sup
=

sup

1lf(z)1: lzl
f l/(z)I: lzl

< lI
~ 1}

This allows us to identify A with the Banach algebra of all the restrictions
of its functions to the boundary of D, which is a Banach subalgcbra of
A I = e( r z : j zl = l J).. If we now consider the element f in A defined by
f (z) = zJ then it is easy to see th& t u A(/) equals D and that uA, (f) equals
the boundary of D.

312

Algebras of Operators

68. THE FORMULA FOR THE SPECTRAL RADIUS

Let x be an element in our general Banach algebra A, and consider

its spectral radius r(x), which is defined by
r (x) ;::; sup

UXI : A e: uA ( x) }.

Now let A' be the Banach sub.algebra of A generated by x, that is, the
closure of the set of all polynomials in x~ Theorem 67-E shows that r(x)
has the same value if it is computed with respect to A':
r(x)

===

sup { 1xr: A E UA (x)}.

1

This suggests qrnte strongly that r(x) depends only on the sequence of
powers of x. The formula for r(x) is given in Theorem A below, and our
purpose in this section is to prove it~ It is convenient to begin with the
following preliminary result.
Lemma..

a(xn)

= q(,x)ta.

Let X. be w non-zero complex number and

distinct nth roots 1 so that
PROOF.

x" - XI

X~l)

(x - A1 l)(x -

~lJ

>..2 1

..

Xn its

(x - X.l).

The statement of the lemma follows easily from the fact that x" - Xl
is singular <=> x - '>..~I is singular for at least one i~
Theorem A. r(x) = lim l[xnll lln.
PROOF.
Our Jemma shows that r(xft) ;:::;: r(x)'\, and since r(.x") <
l!x~ll, we have r(x)~ < jlx~I[ or r(x) < l~xnU 1 !n for every nr To conclude
th~ proof, it suffices to show that if a is any real number such that
r(x) < a, then .llxH 1 1~ ::;; a for all but a finite number of n,s, and this we
now do.
It follows from Theorem 65-A and our work in Sec. 67 that if IJ..I >
llxl~, then

x(X)

(z - >.1)~ 1

-~~1

-).-1

:;:

x - I
>..- 1 ( A

)-1

x)-1
(
1- A

[1 + ~] .
n-1

(1)

Genera I Prefiminaries on Banach Algebras

313

Hf is any functional on A, then (1) yields

f (x(>.))

= -

>.- 1 [ f (1)

+ .. f (~:))
1

+I

l!IG

-A- 1 [/(I)

f(xn)x-n]

(2)

n=l

for all IX! > llxl!~ We saw in the proof of Theorem 67-A that /(x('A)) is
&n analytic function in the region IXI > r(x); and since (2) is its Laurent
expansion for !Xl > Hxll, \ve know from complex analysis that this
expansion is valid for IXI > r(x). If we now let a be any real number such
that r(x) < a < a~ then it follo,vs from the prcc~ding remark that the
series
1 f(x~/an) converges, so its terms form a bounded sequence.
Since this is true for every fin A*, an application of Theorem 51-B shows
that the element8 xft/an form a bounded sequence in A. Thus

i;::_

fl znll 11,. < K 1 fna for some positive constant K and every n. Since
Kr.Ina < a for every sufficiently large n we have llxB ll
< a for an but a
or

1Jn

finite number of n,s, and the proof is complete.

The applications we make of this formula will appear in the next

chapter4

69. THE RADICAL AND SEMI-SIMPLICITY

Our final preliminary task is to reach a clear understanding of what

ia meant by the statement that our Banach algebra A is semi-simple.
For this, it iB necessary to give an adequate definition of the radical of A,
and this in turn depends on a detailed analysis of its ideals.
We recall that an ideal in A was defined in Sec. 45 to be a subset I
with the fallowing three properties:
(I) I is a linear subspace of A ;
(2) i e I=? xi e I for every element z e: A;
(3) i e I -==>ix E I for every element z e A4
If I is assumed only to satisfy conditions (I) and (2) [or conditions (I)
and (3)] ~ it is called a left ideal (or a right ideal). For the sake of clarity
and emphasis, an ideal in our previous sense------one which satisfies all
three of these conditions-is often called a two-sided ideal. In the
commutative case1 of course, these three concepts coincide with one
another .

314

Algebras of Operators

The properties of the ideals in A are closely related to the properties

of its regular and singular elements~ In our work so far;p the statement
that an element x in A is regular has meant that there exists an element
y such that xy = yx = 1.. For our present purposes, it is useful to refine
this notion slightly, as folio ws. We say that x is left regular if there
exists an clement y such that yx == I ; and if x is not lei t regular~ it is
called left wingidar4 The terms right regular and right singular are defined
similarly. If x is both left regular and right regular, so that there exist
elements y and z such that yx = 1 and xz ~ 1, then the relation
y = yl = y{xz) = (yx)z

;:::=

li = z

shows that xis regular in the ordinary sense and that x- 1 ~ y = z.

The concept of maximality for two-sided ideals was introduced in
Sec. 41.. By analogy, we define a maximal left ideal in A to be a proper
left ideal which is not properly contained in any other proper left ideal.
A straightfor\ivard application of Zorn's lemma shows that any proper
left ideal can be imbedded in a maximal left ideal; and since the zero
ideal i 0} is a proper left idea11 maximal lcf t ideals certainly exist~ We
now define the radical R of A to be the intersection of all ita maximal
left ideals. It will be convenient to abbreviate this definition by vrriting
R = fl M LI. R is clearly a proper left idea.I .
These ideas can he formulated just as easily for right ideals as for
left idealsJ and there is no reason for giving preference to either side over
the other~ The purpose of the following chain of lemmas is to shovl that
R iB also the intersection of all the maximal right idea!s in A, that is, that
R = rlMRI.
Lemma.
PROOF~

If r is an element of R, then 1 - r is left regular~

'Ve assume that 1 - r is left singular, so that
I~

= A ( l - r) ::::: ~ x

xr : x

Al

is a proper left ideal '"Thich contains I - r.. We next imbed L in a

maximal left ideal ~ft which of course also contains I ~ r. Since r is in
R~ it is also in llf, and therefore 1 = {1 - r) + r is in M. This implies
that M ;::; A, which is a contradiction~
If r is an element of R, then 1 - r is regular
PRoot...
By the lemma just proved, there exists an element 8 such
that s(l - r) = 11 so s is right regular and s = 1 - (-s)r. The fact
that R is a lcf t ideal implies that ( - s )r is in R along 'vi th T J and another
application of the preceding lemma WO\VS that 1 - { ~.s)r ~ 8 is left
regular. Since 8 is both left regular and right regular, it is regular with
in verse 1 - r, so 1 - r .is also regular .
Lemma.

General PreUminaries on Banach Algebras

315

If r is a-n eltnnent of Rt th.en 1 - J:r is regular f fl'r every x .

.PROOF. R is a left ideal, so xr is in R and t.he statement follows from
the lemma just proved~
Lemma~

If r iB an ekmenl of A with the property that I - xr is regular

for every xt then r i.s in R.
paooFr We assume that r is not in R, so that r is not in some maximal left ideal M. It.is easy to see that the set
Lemmot

M +Ar

;:::=

{m

+ xr~m e. Mand x e: A}

is a left ide~l which contains both Mand r1 so M +Ar

m +xr

;::=

A and

for some m and x+ It now follows that 1 ~ xr = mis a regular element

in M 1 and this is impossible, for no proper ideal can contain any regular
element.
The effect of these lemmas is to establish the equality of two sets:

flMLI

= {r: 1 - .xr is regular for

every x}.

(1)

Precisely the same argumentsJ when applied to maximal right ideals,

show that

r\Af RI = fr: 1 - rx is regular for every x J.

(2)

We now prove that all four of these sets arc the same by showing that the
two sets on the right of (l) and (2) are equal to one another. By symmetry~ it evidently suffices t-0 prove the

xr is regular, then 1 - rx is also regular.

We assume that 1 ~ xr is regular with inverse

Lemma. If 1 PROOF.

s = (1 -

xr)- 1 ~

This means, of course, that (1 - xr).s = s(l - xr) = L

it to the reader to show, by a simple computation~ that
(1 - rx)(l

+ rsx)

(1

+ rsx)(l

- rx)

We leave

1,

so that 1 - rxisregularwith inverse l

rsx+ (The formula for (1 - rx)- 1
is less mysterious than it looks, as the reader can see by inspecting the
meaningless but suggestive expressions

(1 - xr)- 1

+ + (xr) + ~
XT

and

+ rx + (rx)~ + (rx) + ~ ~ 1 + rx + rxrx

+ ~ = I + r(l + xr + xrxr + )x = 1 + rsx.)

(1 - rx)- 1 ~ 1
rxrxrx

A~gebras

316

of Operators

We summarize our results in

Theorem A. The radical R of A equa/.5 ('ach of the four sets in ( 1) and (2)
and is there/ore a proper two-side.d de al.

A is said to be semi-aimple if its radical equals the zero ideal {O},

that is1 if each non-zero element of A is outside of some maximal left ideal.
It will be obsPrved that the ideas discussed above are purely algebraic
in nature~ They can be applied not only to our Banach algehra A~ but
also to any algebra or ring -with id entity~ Our interest, hov..Tever J is in A ~
and 1ve no~. bring t-0 bear upon these notions the rPsuJts of Sec. 65,
notably 1 the fact that the set S of all sjngular clcmen tg in A is closed.
begin by noting that if I is any ideal in A (left, right, or t\vosidr.d)~ then by the joint continuity of the algebraic operations~ its
rlosurr I is a.n ideal of the same kind. Next, since any proper ideal is
ronta.inrd in thP proper rlosed set 8~ the closure of any proper id-+;al is a
proper ideal of thP .same kind. It is an easy step from these facts to

' re

l~z:ery

1naximal left ideal in A is clasedr

PROOF.
If any n1axima1 left ideal L is not closed, then L is a proper
suh~.~t of the propPr Jeft ideal L; and this canuot happenJ for it contradicts
thP maximality of I.J.

Theorem B..

Taken
Theorem

together~

C.

the above two theoremR yield

T"he radical ll of A is a proper closed two-sided ideal.

'':re 8hall also nrPd

If I is a proper clo~~ed two-si.ded ideal in A~ then the quotient

algebra -~/I is a Banach algebra.
PttoOF.
T'hcorcm 46-A te11s us that A/I is a. non-trivial complex Banach
~pa rp with respect to the norm defined by

Theorem D.

+ i 11 :i e I J.
is clearly an algebra Vr"'ith identity I + I,
I! 1 + I I! ~ jnf { I~ I + if I :i e I} < I! 1 ;1 =
l~x

Further, A/ I

The

11 (.r

muHipli<~ativc

I} (y

+ 1} I:
<
<

+ JI~

inf 111 x

and
1.

inequality for the norm is easily proved as

follo~~s:

= rIxy + I [I = inf { 11 xy
i 1.: : i e I J
inf {IJ ( x
i r) (y
i ~) 11 :i 1 1 i 2 e: I J
inf l I": z
i 1 !I I: y
i 2 I! : i 1, i 1 e: I J
[inf { 11 x + i i 1 e I } l ii nf ~ 11 y i 2 I~ :i 2 e.. I H

+
+

ti: :

+
+

:;:;; llx

+ /j[

l]Y +Ill.

General Preliminaries on Banach Algebras

317

All that. remains is to show that lll +II~ = I; and since we already have
~I l + Ill < 1, this is an immediate consequence of the fact that
Ill +Ill ~ 1~(1 + I) 2 H < Ul + lllz implies 1 < Ill+ 1[1.
As a final result, we state

A/R i8 a semi-simple Banac.h algebra4

PROOF. It suffices to observe that the natural homomorphism x---+ x + R
of A onto A/ R induces a one-to-one correspondence between the maximal left ideals in A and those in A/R.

Theorem E..

In the following chapters, we shall be concerned almost exclusively

with commutative Banach algebras~ An algebra of this kind is of
course much easier to handle than one 'vhich is not commutative 1 for all
its ideals are two-sided and its radical is simply the intersection of its
maximal ideals~ Our reason for studying the general case here is that
when it becomes necessary to assume commutativity~ as it will in the
next section, 1ve \vant the force of this assumption, and the issues that

depend on it, to be quite clear.

CHAPTER THIRTEEN

?:lte Structure of

etJ111m11tative Hauaclt Atge!JrtlS

The set e(X) of all bounded continuous complex functions defined
on a topological space X is the simplest of the really interesting Banach
algebras. Our purpose in this chapter is to prove the famous Gelfand ...
N eumark l.heorem, '\Vhich says that every commutative Banach algebra A
of a certain type is essen tiaUy identical with e(X) for a suitable compact
Hausdorff space X~ More precisely, we shall prove that a compact
Hausdorff space X can be built out of the inner structure of A, that Xis
accompanied by a natural mapping of A into e(X)J and that this mapping
is one-to-one onto and preserves all the structure SBsumed to be present
in A.

70. THE GELFAND

MAPPING

Let A be an arbitrary commutative Banach algebra~ Our first

theorem below is the principal source of the structure theory of A~ and
the remainder of the chapter will be devoted cnt.ire]y to shaping its
consequences into the elegant form of the Gelfand-Neumark theorem~
If M i8 a maximal ideal in A, then the Banach algeltra A/111
is a dittision algebra~ and therefore equals the Ba'n-ach al.gebra C of complex
numbers. Tht natural honiomorphism x......, x + M _of A onto A/M = C
assigns to each element x in A a complex number x(M) defined by

Theorem A..

x(M) ~
311

x + M,

The Structure of Commota tive Banach Alg ebros

319

and the mappitt-g x-----+ x(M) has the following propertiu:

(1) (x
y)(M) ~ x(M)
y(M);

(2)
(3)
(4)
(5)
(6)

(ax)(M) = ax(M);
(xy)(M) = x(M)y(M);
x(M) = 0 ~ x e M;

l(M) =I;

ix(M) l < tlxU

PROOF~ Theorems 69..-B and 69-D tell us that A/M is indeed a
Banach algebra. Since A contains an identity 1 M is maximal as a ring
ideal (see the comments on this matter in Sec~ 45) j and therefore, by
Theor~m 41-C, A/Mis a division algebra. We now appeal to Theorem
67 ...B to conclude that A/Al equals C4 (Actually, of course, A/M equals
the set of all scalar multiples of its own identity, but we identify this set
\\1th C in accordance with the rcmarksfollowingTheorem 67-B~) Finallyt
properties (1) to (5) arc obvious consequences of the nature of the homomorphism under discussionJ and {6) follows from

1x(M)! =

lx + M[ = l[x + 1\!~I = inf {llx +ml] :me M}

:::;

ilxU.

It is interesting to observe that this proof depends 1 either directly

or indirectly, on virtually every major theorem in the previous chapter.
We also note that the ultimate reason for assuming that A is commutative
lies in Theorem 41A 1 whlch is definitely not true in the non-commutative
case (see Problem 41-1)4
The language of Theorem A is oriented toward the idea that x(M)
is a function of x for each fixed M. The notation~ however, suggests that
we reverse this point of view and that for each fixed x we regard x(M) as
a complex function defined on the set ~ of all maximal ideals in A .
This is the direction in which we now proceed.
If xis a given element of A, we denote. by x the function defined on
m by i(M):: x(M), and we put A= tx:xEA}. Our next step is to
define a topology for mi in such a manner that every function in A is
continuous. The most natural way of doing this is to introduce the weak
topology generated by A. It Vtill be recalled that this is the weakest
topology on 91l relative to which e_very function i is continuous and that e.
typical subbasic open set has the form
A

S(t,

Mo,~) =

{M :Me ml and ]i(M) - t(Mo)I

< E}.

We call the topological space ~ the -space of maximal ideals 1 or the

maximal ideal space~ and the mapping x-----+ of A onto A will be ref erred
to as the Gelfand

mapping~

We are now in a position to reformulate Theorem Aj and to extend

it, in such a way that the Gelfand mapping is displayed as the object of
central importance.

320

Algebras of Operators

The Gelfand mapping x--. t is a norm-decrea.sing (and

therefore continuous) homomorphi1tm of A into e(m) with the follounng
Theorem B..

properties:

A.

of A iB a subalgebra of e{mt.) which ~eparaies the

points of ~ and contains the identity of e(~);
(2) tke radical R of A equalti the set of all elements x far which~ = 0,
so .x ___,. :f is an isomorphism ~ A is semi-Bimple;
(3) an element x in A is regular~ it does not belong to any maximal
ide.al ~ x(M) -F 0 Jar every M;
(4) if x is an element nf A, then its spectrum equals the range of the
function and its spectral radius equals the norm af ~' that is~ q(X) ~ (mi)
and r(x) = sup lf(M)1 == ~1ll 4
PROOF.
The definition of the topology on mt guarantees that each
function x is continuous, and part (6) of Theorem A shows that x is
bounded and that llfH ::::: sup ~x(M)i < Hxll, so x ~xis anorm-decreasing
(1)

the image

mapping of A int~ e(~)- The fact that this mapping is a homomorphism

is immediate from parts (1), (2) 1 and (3) of Theorem A.
Since x ~ f. is .a. homomorphism, A is obviously a subalgebra of
e(~). The stated properties of A follow readily from parts (4) and {5)
of Theorem A~ if M 1 ~ M 2~ and if (say) x is in Mi but not in M 2, then
i:(M i) === 0 and l(Jlf ~) ~ O; and i (M) = l for every M.
If we rec all that R is the intersection of all the 11-f 's, then the proof of
(2) is easy: we hnve only to notice that part (4) of 1,heorem A tellB us
that x(M) ~ 0 for every M ~xis in every!! .
To prove (3), it suffices-in view of part (4) of Theorem A-to show
that xis regular~ it does not belong to any M4 It is elementary that a
regular element cannot lie in any proper ideal, so we confine our attention
to showing that if xis singular~ then it does belong to .some -M. We prove
this by observing that the singularity of x implies that Ax ;::= { yx : y e A i
is a proper ideal which contains x and can therefore be imbedded in a
maxima1 ideal Af which also contains x.
I~,inally, we use (3) to prove (4). By the definition of the spectrum of

x, we have Xe a(x) ~ x - Al lil singular <;::=}(;::,.}:1)(M) = 0 or at least

one M ~ (i - A.l) (M) = 0 for at least one M ~ i(M) ~ X for at least
one Jf 1 so u(x) equals the range of i. The rest of (4) follows from this
Atatemen t and the definition of the spectral radiUB .

the final touch to this portion of the theory by showing that

~ is a compact Hausdorff space.. The reader ril recall that if A is
the conjugate space of A 1 then its closed unit sphere
"VifT e add

S*

= { f~f

EA* and

llf1!

:$ 1}

is a compact Hausdorff spare in the -y.,Teak* topology (see Theorem 49-A) .

The Structure of Commutative Banach Algebras

321

Our strategy is to id en tify fill 7' both as a set and as a topologi ca.I space,
with a closed subspace of s.
A multiplicative functional on A is a functional! in the ordinary sense
~that is, a.n element of the conjugate space A *-which is non-zero and
satisfies the additional condition f(xy) = f(x)f(y).. Theorem A shows
that to each M in ~ there corresponds a roultiplicalive functional f M
defined by fM(X) =-= x(M). It is important for us to know that M ~ f Mis
a one-to-one mapping of mi onto the set of all multiplicative funet.ionals.
It will facilitate our \vork if we begin by proving the
Lemma. If fl and f 2are multiplicative functionals on A with the sanie null
apace M 7' then f1 = /2~
PnOoF. We first show that f1 = af2 for some scalar a. JJet Xo he

an element of A which is not in M. If xis an arbitrary clement of ~4, it h~

ea.sy to see that x can be expressed uniquely in the form x ~ ni
{3xo
\vit.h tn in ft{ (set f3 = f,,(x)/f2(x 0), put m == x - {3xoJ and observe that
f42\ni) = O). It no\V follows t.hat

fi(X) = f1(rn)

+ f3f1(Xo)

f3f1(Xo) = rf1(Xo)ff~(XQ))f2(x),

so f 1 = af2 with a ~ f1(xr:i)/f2-(xo). We complete the proof by sho"ring

that a equals l. Let x he an element not in MJ so that f2(x) "#:- O~ "fhen
af2 ( x) t ;::::: af2 ( x 2 ) :::- f 1 ( x !) = f i ( x) 2 = {a/2 ( x) J1 ;:; a 'lj2 ( x) 2 implies that

so

a=

0 or a= l.

Sincef1 #: 07' Vle conclude that a =

1~

\Ve now use this to prove

M ~ fM is a one-to---one m.appi,ng Df the set mi of all maximal

ideal-s in A onto the set of all its m.ultiplicatiiie functionals.
PROOF.
The mapping is ea.sily seen to be one-to~onc} for if 1'! 1 r! M 2,
and if (say) x is in llf 1 and not in 111"~~ then f MJx) ~ 0 and f Mt.(x) ~ 0.
'fo prove that it is onto, let f be an arbitrary multiplicative functional,
and consider its null ~pace lrf = f x :f(x) ~ 0} . It is clear by the
assumed properties off tha.t M is a proper closed ideal in A~ Furthermore, .JJ is maximal, for if it were properly cont.ained in a proper ideal I,
then f(l) would be a non-trivial ideal in C~ contrary to 1~heorern 41-A.
Since f and f~v are multiplicative functionals with the same null space,
the lemma just proved implies that f = f Mi and our proof is complete.
Theorem C.

In some of its more concrete applications, this theorem is used to

rep1ace the algebraic problem of determining the maximal ideals in A by
the analytic problem of finding its multiplicative functionals. Its
importance for oltr e11rrent task of sho\ving that~ ii3 a compact Hausdorff

322

Algebras of Operators

space is that it enables us to regard mi as a subset of A. We can say

even more than this, for parts (5) and (6) of Theorem A tell us that
every multiplicative functional f" bas norm 1, so ffi'l is a subset of the
closed unit sphere S*.. We recalled earlier that S* is a compact Hausdorff space with respect to the weak* topology, which is (see Sec~ 49)
the weak topology generated by all the functions F~ defined on S* by
F~(f) = f(x). We now observe that when F~ is restricted to mr,, it is
precisely .f, for
F~(fM) ~ f M(X) :;:= x(M) = x(M).

Therefore, by Problem 19-lc, the topology which mi has as a subspace of

is exactly its topology as the space of maximal ideals.. These considerations permit us to regard~ as a subspace of

s.

Theorem 0.. ThlJ maximal ideal 8'pate mt ia a compact Hau8d()Tjf space.

PROOF~ In view of the above discussion, it suffices to show that ~ is
a closed subspace of
We acc_omplish this by forming the subspace
X of S* defined by

s.

X = ('\ .ttKA

l f: f e. S

It is evident that X is simply

since we have

== r\~,V'A {f :f e. S*
~ flz~urA { f :f e
~ n~.11tA {f :f

Es

and f (xy)

:;=

f (x )/(y) }

together with the zero functional; and

and f(xy) - f(x)f(y) = 0 J

and F ~(f) - Fr&(f)Ft1(j) = 0 J
and (F ~ - F 'ZF1J) (/) = 0}'

it is easy to see that X is closed in s (note that each of the sets last
written has this property).. We next remark that F1 is continuous on X
and equals 1 on $land 0 at the .zero functional.. It follows from this that
~ is closed in X and is therefore closed in S* ~

It is worthy of notice that the topology we imposed on ~ is the only

one which makes it into a compact Hausdorff space on which all the
functions x are continuousJ for by Theorem 26-E, any stronger co1npact
Hausdorff topology must equal the given one.
When Theorems B and D are taken together t the result is often
called the Gelfand representation thRorem. In essence, thls tells us that
every commutative semisimple Banach algebra is isomorphic to an
algebra of continuous complex functions on a suitable compact Hausdorff
space. In general, the norm is not preserved by this isomorphism and
the representing algebra does not exhaust the continuous :functions on the
underlying space. We shall remove these deficiencies in the following
sections by assuming that additional structure is present in the Banach
algebra under discussion.

The Structure of Commutative Banach Algebras

71. APPLICATIONS OF THE FORMULA r(x)

323

= lim llxnjl um

We continue our study of an arbitrary commutative Banach algebra

A and of the Gelfand mapping x ~ i of A onto the subalgebra A of
e (;rrt) . Our first theorem provides a shnple way of gun.ran teeing that
this mapping preserves norms.
Theorem A. The following cunditions
(1) llx'll = llxll 2 jOT every x;
(2) r(x) = Hx i for every x;
(3) 11~1 = ltxl~ for every x.
PRO OF.

on A are all equivalent to one another:

It foil ows from condition ( 1) that

and, in general 1 that l!x~*ll = llxll~ for every positive integer k.

formula for the spectral radius now yields
r(x) = lim

IJxnll 11n

lim If xt"ll1 12k = lim

llxll

The

llxlL

so (1) implies (2)~ The fact that (2) implies (1) is immediate from
l~:x 2 ll = r(x 2 ) = r(x) 2 = if xii~~ In view of the equation t(x) = Jixll (see
Theorem 70-B), the equivalence of (2) and (3) is obvious.
Our next problem is to devise a VtTay of making Slll'e that the representing algebra A comes as close as it can to exhausting ~(mi) J and we
accomplish this by introducing the follomng property~ A is said to be
self-adjoint if for each x in A there exists an element y in A such that
ii (2Jl) ~ x(.tlf) for every M.
Theorem

B.

If A is self-adjoint~ then A is dense in e(mi) .

By part (1) of Theorem 70-B, we know that A is a suhalgcbra of e(mr) which separates the points of mi and contains th~
identity function~ Problem 20-3 and our hypothesis now tell us that the
closure of A is a closed subalgebra of e(mt) which separates points~ contains the identity function, and contains the conjugate of each of its
functions. Theorem 3&-B {the complex Stone-Weierstrass theorem)
shows that this closure equals e{ mt), so A itself is dense in e( m) .
PnOOF.

If we put together the results obtained in the above two

we have

theorems~

If A is self-adjoint, and if Jlxi l = II x Ii 2 for every x, then the

Gelfand mapping x ~ i is an isometric isomorphism of A onto e(~).
PROOF.
By Theorem A, the mapping x ~ i preserves norms. It is
therefore an isometric isomorphism of A onto A, and we see from this

Theorem C.

324

Atgebrcs of Operators

that .A is closed in e(31t). Since A is dense in e(mr) by Theorem B~ it

follows that A equals e(mr), and the proof is complete.
This theorem lacks a certain simplicity which it ought to havet for
the condition of self-adjointness is rather far removed from the intrinsic
structure of A~ Our work in the next two sections v:i.11.remedy this defect
and at the same time 'Will establish closer connections vii.th the operator
algebras to which we apply our final result.

72. INVOLUT10NS IN BANACH ALGEBRAS

A Banach algebra A is called a Banach *-algebra if it has an int olution 1

that L"'J if there exists a mapping x ---+ x of A into itself -with the foll owing
properties :
(I) (x+y)=x*+y;
(2) (ax) = ax*;
(3) (xy) = yx;
(4) x** = x.
It is an easy consequence of (4) that. the involution x---+ x* i~ actually a
one~to~one mapping of A onto itself.
\\'Te also note that o = 0 and
I* = 1:t a.s \Ve sec from 0
x* = x* === (0 + x)* = o~
x* and
i = 11 * ~ 1**1 * = (11 *) * = (1 *) * == 1 ** = 1. ..rhe element x* is
called the adjoint of x., and a subalgcbra of A is said to be self-adjoint if it
contains the adjoint of each of its clements. If At ~ also a Banach
-algebra, and if f is an isomorphism of . .4. onto ...4_ ' 1 then f is called a
*-isomorphism if it preserves the involution in the sense tha.tf(x*) = f(x)*.
We naturally want the involution in a Banach lk-algebra to be linked
in some useful way to the norm. The property llx*ll = Hxll clearly
implies that t.be involution is continuous; for if Xn ~ x~ then
1

shows that Xn *---+ x*. A much stronger relation between the involution
and the norm is given by the condition

llx*xll

llxlP\

and any Banach -algebra \vhicb satisfies it is called a B* ~algebra.

easy to see that we have llx*ll == Hxll in every B*-algebra; for

Hxll 2 ==

l~x*x]l

<

It is

llx*~I ~lxll

llxH : : ; llx*ll for every xi so llx*~I :$ Hx* I~ = flxH, and therefore

= flxll. It follows from this that the relation 11x*xll = flx*11 llx~I is

shows that
Ux*I~
also true.

Several of the Banach algebras described in Sec. 64 are also Bannch

The Structure of Commutative Banach Algebras

325

-algebras with respect to natural involutions. If X is any topological

space, then e(X) is clearly a commutative B*-algebra relative to the
involution defined by f*(x) = f(x). 1'\he disc algebra, however, is not,
for ii f is the function defined by f(z) = z, then f*(z) = z is not analytic
at any point.. If H is a non-trivial Hilbert space, then ffi(H) is a B*algebra with the adjoint operation T ~ T* taken as the involution (see
Theorem 56-A). Since C*-algebras are the self-adjoint Banach subalge bras of IB(l/) ~ s, they too are B-algebras. Finally t the group
algebra L1 (G) of a finite group G is a Banach *-algebra with respect to the
involution defined by f*(g1) = /(giH 1 )i and it is easy to see that llr 11 = 11/H .
It should be reasonably clear that Banach -algebras (and especially B*-algebras) are modeled a1ong lines suggested by ffi(H).. We have
already called the element x* in such an algebra the adjoint of x.. By
analogy, we say that xis self-adjoint if x = x*, normal if xx* = x*x, and a
pro}ection if x = x and x! = x.
If xis a normal element in a B*-algebra, then l!x 2 ~1 = Hxll 2
PROOF.
It is obvious that !lx!!I < jfx~I 2 .. The inequality in the other
direction is a consequence of the follo-wing computation:

Theorem A.

llx *II ~u xii~ ;:; (l]x II !Ix 11) 2

= llxx*xxH

::;=

IJx*x ~j :r = 11 (x*x) *x*xH = llx*xx*x II

II {x*)!x 2 B~ l (x 2)*x2'll = I~ (x~r~H llx 2 l
1
= I[ (x*)ill l~xil! < llx* IJ llx

~l-

This result suggests more strongly than ever that there are close
connections between B-algebras and algebras of operators on Hilbert
spaces (see Theorem 5S-D).. We describe the true state of affairs in this
matter at the end of the next section.

73~

THE GELFAND-NEUMARK THEOREM

We are now in a position to give Theorem 71-C its final

form~

Theorem A. If A iB a commut,ative B*-algebra, then the Gelfand mapping

x--+ ~is an isam..etric *-isomorphism of A onto the commutative B-algebra
e(~)-

PnooF. Since A is commutativeJ each of its elements is normal, and

it follows from Theorem 72-A that Hx 2 H== llxl; 2 for every x. By Theorem
71-Cj it now suffices to show that ;t:(M) = i(M) for each x in A and M

in mt.
Our first step is to prove that if x is seJf . .adjoint, then i(M) is real
for every M. We assume the contrary, namely, that there exists an Jr!
such that (M) = a
i/3 with fj ~ 0. Since x is self-adjoint,

y ~ (x - al)/P

Algebras of Operators

326

is. also self-adjoint~ We further note that y(M) = i, soy - il is in

It is obvious from the properties of the involution in A that
M*

M~

{m:mEMI

is a maximal ideal; and since it contains (y - ii)* = y

y(M*) = -i. If K is any positive numberJ then
,........,.........
(y - iKl)(M) = -i(l
K)

+ ii

we see that

...,.......,.. .........

+ iKl) (M)

and,....,,..........__
(JJ

lly -

iKll! ~

= i(I

+ K)~

ll:v -

iKI!I and, similarly~ that 1

multiplying these two inequalities, we obtain
(I+ K)t ~

iKllJ ~IY
j!(y + iKI)*(y

llY -

+K

It follows from this that I

+ K :s; llY + iKlll~

s;:
On

+ iKlrl = l~(y + iKl)*ll llY + iKlil

+ iKl}li = tl(y - iKl)(y + iKl)ll
= 1iY + K~l IJ < liy IJ + K
2

so I + 2K < llY 2 H~ Since K is arbitrary~ this is impossible, and this

portion of our proof is complete~
We now conclude the proof by showing that if xis any element of A,
then '?(M) = :t(M) for every M. It is clear that y ~ (x
:r*)/2 and
z === (x - x*)/(2i) are self-adjointJ and that x = y
iz; and therefore,
by the result of the above paragraphJ we have

...-....

x*(M)

...,.,...,..,, .........

= (y

a..-.

iz)(M) == y(jjf) ___. iZ(M) == g(M) - i2(1lf)

= y(ilf) + iz(M)

x(M).

We already know that if X is any compact Hausdorff space, then

e(X) is a commutative B*-algebra. The theorem just proved-it is
called the Gelfaml-..Neumark representation theorem-tells us that commutative B*-algebras are simply abstract (g(X)~B, in the sense that every
such algebra is abstractly identical vnth e{X) for a suitable compact
Hausdorff space X~
There is another Gelfand-Neumark theorem of great interest,
which applies to arbitrary B*-algebra.sa We observed in the previous
section that every C*-algebra is a B*-algcbra. The converse of this is
also true, for if A is a B*-algebra~ then there exists a Hilbert space H
with the property that A is isometrically *-isomorphic to a C*-algebra
of operators on H .1 General B-algehras are therefore abstract
algebras. The proof of this theorem evidently requires that a suitable
Hilbert space be constructed out of the given structure of A.. The
details of this construction are beyond the scope of this book, and "re
content ourselves with merely stating the facts.

c-

See Ric kart [34, p. 244].

CHAPTER FOURTEEN

Some Speeia/
Commutative Ranae/J Algebras
Our discussion in Sec. 63 foreshadowed a generalized form of the
spectral theorem~ and the principal purpose of the present chapter is to
formulate and prove this rcsultr We begin with some additional material
relating to Banach algebras of continuous functions. In particular, we
keep the promise made in Sec. 30 by showing that the Stone.-Cech
compactification of a completely regulu space is essentially unique.

74.. IDEALS IN e(X} AND THE BANACH-STONE THEOREM

Let X be a compact Hausdorff space, and consider the commutative

B*-algebra e(X)~ If mt is the space of maximal ideals in e(X), then. the
developments of the previous chapter lead us to expect that ~ can be
identified 'With X and that the Gelfand mapping is the identity mapping
of e{X) onto itself.
In order to substantiate this conjecture, we begin by observing
that to each point x in X there corresponds a proper ideal M~ in e(X),
defined by
M~ = {f:fee(X) and/(x) = 0}.
M~ is easily seen to be maximal and is thus an element of mi, for

it is the

null space of the multiplicative functional f M. defined by I Mlt(f) :;;; f(x),

which assigns to each function in ~(X} its value at x. Since Xis compact
Hausdorff, and therefore normal, Urysohn's lemma tells us that for ee-ch
point y ~ z there exists a function f in e(X) such that f(x) ~ 0 and
327

328

f (y)

Algebras of Operators
~

0. This shows that x -+ M ~ is a one-to-one mapping of X into

51!. Our next step is to prove that this mapping is onto~ and for this it
clearly suffices to show that if Mis any maximal ideal in e(X)~ then there
exists a point in X at which every function in M vanishes. We assume
the contrary, namely, that for each point x in X there exists a function
fin M such that f{x) '#: O~ Since f is continuoust x has a neighborhood
at no point of which f vanishes. We now vary x to obtain an open cover
for X, and we use compactness to infer that this open cover has a finite
subcover. Let flt f 2~ . . J f ~ be the corresponding functions in M.
M is an ideal, so the function g = ~~ 1 fi.T. = ~i- 1 I/, I2 is also in M; and by
the manner of its construction~ it clearly bas the property that g(x) > 0
for every x. It follows that g is a regular element of e (X), and this
contradicts the fact that it lies in the proper ideal M. We therefore
conclude that x--+ M:t is a one-to-one mapping of X onto grr,
'rhese considerations enable us to identify the set ;m: with the set X,
and in terms of thls identification, we regard X and mi as two possibly
different compact Hausdorff spaces built on the same underlying set of
points~ By our work in the previous chapter, we know that the Gelfand
mapping f ~ f is a one-to-one mapping of e(X) onto e(~). If we use
the notation established there, then we find that
f(M~) = f(M~)

= fM~(f) =

f(x),

so J = f and e(mT) === e(X)~ We now recall that any compact Hausdorff
topology on a non-empty set is uniquely determined as the weak topology
generated by the set of all its continuous complex functions (see Problem
27-3). It follows from thie that X and ml are equal a~ topological spaces.
We fflmmarize the results of thls discussion in
Theorem A.. Let X be a compact Hausdorff space and mi the space of
niaximal ideals in the commutative B*-algebra e(X). Then to each point
x in X tkere corresponds a maximal ideal M:.: defined by

Ms= {f:/ee(X) afld/(x)

= 0},

and z --+ M. is a om-to-one map'[ling of X onto mt~ If this mapping is used

t.o identify mi with X, then mt and X are equal a.s topological spaces, e(;m:)
equals e(X), and the Gelfand mapping f--+ /is 'the identity mapping of e(X)
onto itself.

The main idea of this theorem is that the maximal ideals in e(X)
correspond in a natural way to the points of X. Our next step is to
extend this idea and to obtain a similar characterization of the proper
closed ideals in e(X).
We again consider a compact Hausdorff space X, and we begin our
discussion with the observation that to each non-empty closed subset F

Some Special Commutative Banach Algebras

329

of X there corresponds a proper closed ideal l(F) in e(X), defined by

I(F)

= l f :f e e(X)

and f(F) = O} ~

If x is any point not in F, then it follows from the complete regularity of

X that there exists a function fin e(X) such that f(x) # 0 and f(F) == 0~
This shows that F ~ I(F) is a one-to-one mapping of the class of all nonempty closed subsets of X into the set of all proper closed ideals in e(X).
We shall prove that this mapping is onto} that is~ that every proper closed
ideal in e(X) arises in this way from some F.
Let I be a proper closed ideal in e(X). We may assume that I is
not the zero ideal, for this ideal clearly arises from the full space X.

We define F by

F=

{x~f(x)

= 0 for every f e.IJ.

It is easy to see that Fis a proper closed subset of X; and since I is contained in some maximal ideal, it follows that F is non-empty. Our
task is to prove that I(F) = I, and since it is obvious that I c I(F),
the real problem is to prove that l(F) C I~ If f is any function which
vanishes on F, we must show that flies in I~ We may evidently assume
that/~ 0, so that {x:f(x) = O} is a proper subset of X~
In the first part of our proof~ we assume that f vanishes on some open
set G which contains F. Since f ~ 0, G' is non-empty and is thus a
compact subspace of X. For each point x in G', there exists a function
gin I such that g(x) 'F 0. The technique used in the proof of Theorem A
can now be applied again~ to yield a finite number of functions g1, g,,
~ ~ ~ J Un in I with the property that at least one is non-zero at every
point of G'. We next define a function go by go = ~i-1 g1g ~ i;:_.1 1o,I z,
and we observe that g0 is in I and that go(x) > 0 for every x in G'. By
the Tietze extension theorem, the function whose valuee on G' are given
by 1/ go(x) can be extended to a function h in e(X). It is easily seen that
goh is in I, that it equals 1 on G', and that f = fgoh, so f is in I.
We now turn to the general case~ For each E > 0 1 the sets Kand L
defined by K = {x: 1/{x) I < ~/2 J and L = {x ~ lf(x) i ~ t} are disjoint
closed subsets of X. K is clearly non-empty, and since j ~ Ot Lis also
non-empty for every sufficiently small ie. We assume that E bas been
chosen at least this small, so that K and L constitute a disjoint pair of
closed subspaces of X~ By Urysohn's lemma, there exists a function
gin e(X) such that g(K) = 0, g(L) = 1, and 0 < g(x) < 1 for every x.
We now define a function k in e (X} by h == f g, and we note that

I I - hll = llf(l -

g) !I ::;

E.

It is evident that h vanishes on the set G = {x: lf(x)I < t/3}; and since
G is an open set which contains F, it follows from the preceding para-

330

Algebras of Operators

graph that his in I. This shows that for every sufficiently small positive
number e there exists a function h in I such that Ill - hll < E, and since
I is closed~ we conclude that f is in /.
We give the following "formal statement of our result.
Theorem B. Let X be a .compact Hausdorff space. Then to eac,h no'Tl. . .
empty closed set F in X there corresponds a proper closed ideal I (F) in
e(X), defined by l(F) = {f :/ e e(X) andf(F) == 0 J; and further, F--+ I(F)
i8 a one...to-one mapping of the class of all non-empty closed subsets of X onto

the set of all proper closed ideals in e(X).

As an easy consequence of this, we have
Theorem C. If X is a compact Hausd&rjf space, then every closed ideal.
in e(X) is the int.erBe.ction of the maximal ideals which conmin it
PROOF. Since the intersection of the empty set of maximal ideals is
e(X) itself,. we may confine our attention to a proper closed ideal I.
By Theorem B, I = l(F) for some non-empty closed set F. It is clear
that the maximal ideals which contain I are precisely those associated
"With the points of F. It therefore suffices to observe that a function in
e(X) vanishes on F-<==> it vanishes at each point of F.

We have seen in Theorem A that the points and the topology o:f a
compact Hausdorff space X r.an be recovered from the maximal ideals in
e(X)~ Since the maximal ideals in ~(X) are objects of a purely algebraic nature~ it follo\YS that the compact Hausdorff space X is fully
determined, both as a set and as a topological space, by the algebraic
structure of e(X)
These observations lead us directly to
I

Theorem D (the Banach-Stone Theorem). 7 wo compact Hausdorff spaces

X and Y a.re homeomorphic ~ fhei.r corresponding function algebraa e(X)
and e(Y) are isomorphic .
1

...

75. THE STONE-CECH COMPACTIFICATJON (continued).

It is natural to wonder what can be said along the lines of Theorem

74-A in the case of a topological space X which is not necessarily compact
Hausdorff. Regardless of the properties of X, we know from our previous work that e(X) is a commutative B-a)gebra, that its maximal
ideal space 51l is a compact Hausdorff space~ and- that x---+ M~ is a
mapping of X into m. Our difficulty is that vlithout restrictions of some
kind on
we know practically nothing about the properties of the map ...
ping x ---ry M :t If it happens that this mapping is one-to-one and is also
a. homeomorphism of X onto a subspace of mi, then we observe that X

xt

Some Speciaf Commutative Banach Algebras

331

is necessarily completely regular. It is therefore reasonable to assume at

the outset that X is comple t.ely regular, and we shall see that several
interesting conclusions follow from this hypothesis.

Let X be a completely regular space and 3lt the space of

maximal i.deal8 in the commutative B*-algebra e(X). Then the mapping
x........:,. M:t: is a homeomorphism of X onto a subspace of :m. Furthermore, if
this mapping is used to identify X with its image in ml~ then (1) X is a
dmse subspace of~; (2) each function in e(X) has a unique exttrn.mcm to a
function in e(~); and (3) if Y is a compact HaUBdorff space with t,he
'/)Toperties of 'tll1 stated in (l) and (2)~ then there exi8t8 a homeomorphism of
mr onto Y which leaves the points of X fixed.
PROOF. The fact that x........,. M~ is one-to-one is immediate from the
complete regularity of X, so we may identify X as a set vlith its image in
gn:_ The subset X of mi has two topologies: its own, and its relative
topology as a subspace of mt. The following arguments show that these
topologies are equal. We know that the Gelfand mapping f ---io J is an
isomorphism of e(X) onto e(~)- Also 1 just as in the proof of Theorem
74~A, we have /(x) == f(x) for each fin e{X) and each x in X. These
observations imply that e(X) is precisely the set of a11 restrictions to X
of functions in e(~); and since both topologies are completely regular,
it follows from Problems 19-lc and 27-4 that each is the weak topology
generated by e(X), so they are equal and X can be regarded as a subspace
of mi. These observations also show that X is dense in ;rrI-for if J
vanishes on X, then f ;:::= Or/ = 0, and/ also vanishes on mt-and that
each function f in e(X) has a unique extension / in G('tlIT)a All that
remains is to prove (3)r We know that /----:,. f is an isomorphism of
e(:m) onto e(X); and by the assumptions about Y, the mapping f-:,. f',
which assigns to each fin e(X) its extension f' in e(Y), is an isomorphism
of e(X) onto e(Y). Thus J ~ f ~ f 1 is an isomorphism of e(mt) onto
e( Y). If x is a point of X t then th is isomorphism clearly carries the
maximal ideal in e(3lt) corresponding to x over to the maximal ideal in
e(Y) corresponding to x; so by the Banach-Stone theorem, it induces a
homeomorphism of mi onto Y which leaves the points of X fixed.
Theorem A,.

On comparing this result with Theorem 30-A~ we ......~ee that mi is

homeomorphic, in the manner described, to the Stone-Cech compactifica tion /3(X). In this sense, therefore, mt and (3 (X) can be considered
equal to one another~ and also to any other compact Hausdorff space
which contains X as a dense subspace and has the required extension
property. In effect, we have shown that the Stone-Cech compactification of a completely regular space X is unique and can equally well be
regarded as the maximal ideal space of e(X) .

332

Algebras of Operators

76. COMMUTATIVE C*-ALGEBRAS

In this :final section, we apply the results of the preceding two chapters to the theory of operators on a non-trivial IIilbert. space H~ \Ve
know that ffi{H) and all its self~adjoint Banach subalgebras (that isJ all
c-algebras of operators on H) are B*-algebras~ As a special case of the

GeHand-Neumark theorem, we therefore have

Theorem A. Let A be a C-Ommutative C*-al.gebf'a of operators on H, and
~its space of maximal ideals.. Then the Gelfand mapping T ....._. fr ia an
isometric *-isomorphism of A onto e(5Tl).

f Tt:}

is a non-empty set of operators on H, then the smallest

Banach subalgebra of ffi(H) which contains every r . . is called the Banach
subalgebra of (B (H) generated by the T" ~ s. It is easy to see that this
Banach subalgebta of ffi(H) is the closure of the set of all polynomials
in the T/s~ If N is a normal opera.tor on H, then the Banach suhalgebra of <B(H) generated by N and N* is c]early a commutative C*alge bra, and is called the comm uta.ti~)e C*-algebra generated by N. We no"\v
If

specialize Theorem A to
Theorem B.. I.Jet N be a normal operator on H, and A the commutalive
C*-algebra generated by N. If mt is the space of maximal ideals in A, then
the Gelfand mapping T---+ Tis an isometric *--isomorphism of A onto e(~)~

As it stands1 this result is only a beginning. In order to exploit it

effectively, our first task is to show that the spectrum of an operator in
A~whicb is understood to be its spectrum as an element of ffi.(H)~equa.ls
its spectrum as an element of A. In proving this~ we shall need the
following preliminary fact~

Let X be a compact Hausdorff space and A a Banach subalgebra

of e(X). If f is a real function in A u~hich is regular in e(X), Oien it is
also regular in A.
PROOF.
The range of f is r.learly a compact subspace of the real line
which does not contain 0. If ~ > 0 is given, then by the V{ eierstrass
Lemma.

approximation theorem (see Problem 3~3) there exists a polynomial

p such that lp(t) - I/ti < ie for every tin /(X). It follows from this that
[p(f(x)) - 1//(x)I < t: for every x, so llp(f} - 1//11 < E. Since p(f) is
in A and A is closed, we conclude that I/f is in A.
Theorem C. Le.t A be a commutative C*-a"lgebra of operators on H. If
an operator T in A is regular in m(H), then it is also regular in A, and
"therefore the spectrum of T as an operatqr on H equals its spectrum as an

element of A .

Some Special Commutative 8dnach Algebras

333

\Ve begin by considering the special case in which T is

assumed to be self-adjoint. I.Jet B be the Banach subalgebra of m(H)
generated by T and 7 1 ~ Since T and r- 1 are self-adjoint and commute
\vith one another, it is evident that Bis a commutative C*... a.lgebra; and if
mi. is its space of maximal ideals~ then Bis isometrically *-isomorphic to
e (mt) and T is represented by a real function in e(~). l" ~ A f"l B is
a Banach subalgcbra of B and is therefore isomorphic to a Banach
subalgebra of e(mi). Since T is in C and is regular in B~ our lemma
shows that T- 1 is also in G"' and therefore lies in A .
We now turn to the general caset in which Tis not assumed to be
self-adjoint.. It is clear that U = TT* is a self-adjoint operator in A, and
since it has an inverse u- 1 = (TT)- 1 = (T*)~ 1 r- 1 = (T- 1 ) *T-1 in
ffi(H)J Vle know from the preceding paragraph that u- 1 is in A. We now
make use of the commutativity of A to v;rite the relation U u- = I in
the form T(TU- 1 ) = (T*U- 1 )T ~I. This shows that T- 1 = T*U- 1 ~
so r- 1 lies in A and the proof is complete.
PROOF.

1
-

This result tells us, in particular, that if N is a normal operator on H~

then its spectrum cr(.lv) equals its spectrum as an clement of the commutative C*-algebra generated by N. Our next step is to provide a concrete
rcpre sentation for the space of maximal ideals in this algebra.
Theorem D. Let 1"-l be a normal operator on H, A the commutative C*-algebra generated by NJ and mt the space of maximal ideals in A. 7 hen the
function G in e(mz) 'lDhich corresponds to N under t.he Gelfand mapping
is a homeomorphism of mt onto (J'(N) .
1

PROOF~

It follo\YS from Theorem C and part (4) of Theorem 70-B that

er (N) is precisely the range of the continuous function N defined on ~.
Since both ml and a-(N) are compact Hausdorff spaces~ it suffices by
Theorem 26-E to show that ~r is one-ta,.-.one. Let M 1 and M :t be points
of mz such that f:(M1) = J\r(M2). Then we also have

= R(M1} =
functions fl and N*
i?(M1)

J:i(M2)

N*(M2),

so each of the
takes equal values at Mi and M2.
Since A is the closure of the set of all polynomials in N and N*, every
....
............
function in e (~) is a uniform Jilnit of polynomials in !l and N*, and
therefore every function in <3($,) takes equal values at M 1 and M2. We
conclude the proof by observing that since e(mt) .separates the points
of ml, it follows that M 1 = M 2..

In accordance with this result~ we may identify mt with the compact

subspace u(N) of the complex plane; and when this identification is
carried out, it is easy to see that "fl(z) ;:= : for every z in ;rrt. We sum...
mariz c our con cl u.sious in

334

Algebras of Operators

Theorem E. Let N be a normal operator r.m H with spectrum a(N), and

let A be tAe commutative C*-algebra generated by N. Then the space mi
of maximal ideals in A e.qual8 a(N), and the Gelfand m.a'fJPing T---+ P of
A onto e(ITTl) iB an is<nnetric *-iBomarphism which carrits N into thefunction whose values are given by $(z) = z for every z in mi.

This theorem has a number of simple consequen~s~ of which the

following are only a few: (1) N = 0 ~ u(N) = {O}; (2) N is singular
~ a(N) contains O; (3) u(N*) ~ u(N); (4) N is self-adjoint<=> a-(N) lies on
the real line; (5) N is unitary =} a(N) ~ ~z: rzl = 1 ~ ; (6) N is R projection
{;;;;} u(N) ~ io~I I. If it happens that H is finite-dimensional 1 so that
u(N) consists of a finite number of distinct complex numbers A. 1, A.2~
. . . , X.J then we can write

P.

where
is the function in e(~) defined by ';(>..,.)
from this that

&..;.

It is evident

where the P/s are non-.zero pairwise orthogonal projections in A such that
?:t,1 Pi = I. This is precisely the spectral resolution of N treated in
Chap. 11, so Theorem R actually contains the finite-dimensional spectral
theorem. We therefore have solid grounds for regarding Theorem E
as the generalized form of the spectral theorem discussed in the last paragraph of Sec.. 63, and all our promises are fulfilled.

Appendices

APPENDIX ONE

liKetl Point ?:ltcorems

aud Some Applicatit11ts to Auatvsis

f be a continuous

mapping of the closed interval [-1,1] into

itseli+ liJ.gure 39 suggests that the graph off must touch or cross the
indicated diagonal, or more precisely, that there must exist a point Xa in

Let

r------

11
I

I
I

--...--....-. ...... J

-1

Fig. 39

[ ~ l J 1J vnth the property that f (xc) = Xo~ The proof is easy+ We con~
aider the continuous f llllction F defined on {- 1, I] by P' (x) :;::: f(x) - x,
and we observe that F(-1) > 0 and that F(l) :5 0. It now follows
from the Weierstrass intermediate value theorem (see Theorem 31-C
and the introduction to Chap. 6) that there exists a point Xo in [ ~ 1 tlJ
such that F(xo) = 0 or f(xn) = XoIt is convenient to describe t.his phenomenon by means of the follow~
ing terminology" A topological space X is called a fixed point space if
337

338

Appendices

every continuous mapping f of X into itself has a fixed point, in the sense
that /(xo) = Xo for some xo in X. The remarks in th~ above para.graph
show that {-1,1] is a fixed ~~nt space.. Furthermore, the closed disc
{ (x,y) :zt
y 1 < 1 l in the Euclidean plane R" is also a fixed point space
(for a lucid elementary proof of this~ see Courant and Robbins [6, PP~ 251255]) ~ Both of these facts are special cases of

BrOtJwer s fixed Point Theorem.

in Rn is a fixed point 8pace.

The closed unit spM.re S =

tx ~ llx I <

1}

T'here are several proofs of this classic result~ but since they &11
depend on the methods of algebraic topology-, we refer the reader to Bers
(3, p. 86}. Brouwer' s theorem itself U3 a Bpecial case of
1

Schauder s Fixed Point Theorem..

Banach space is a fixed point space+

Every convex compact subspace of a

For a proof, together with a discussion of other related results, see

Bers [3, pp. 93-97}. Schauder's theorem was foreshadowed by the work
of Birkhoff and Kellogg [5] on existence theorems in analysis. We
illustrate the relevance of these ideas to such problems by giving a full
treatment of Picard's theorem on the existence and uniqueness of solutions of first order differential equations .
We begin by considering an arbitrary metric space X with metric d.
A mapping T of X into itself is called a contracti&n if there exists a positive
real number r < 1 with the property that d(Tx, Ty) < r d(xJy) for all
points x and y in Xr It is obvious that such a mapping is continuous.
We shall need the following

Lemma.. If Tis a cootraction defined on a complete nietri-c space X, then

T haa a unique fixed point.
PROOF.
Let xo be an arbitrary point in X, and WTite X1
Txa,
;:=

and, in general~

d(xm 1Xn)

Xn

==

rnxo = Tx~-1

If m

< n,

then

= d ( Tw'x o, Tnx o)

== d( Ttnx D~ T T~xo)
S r"" d(xo, Tw~o) ~ r- d(xo,x--.)
::; T"' (d(Xo,X1)
d(x1,x~)
~
d(z,. 1,X,.-w1)]
< r d{xoJX1)ll + r
+ r~ 1 ]
< r d(xo,X1).
1- r

+
+

Since r < 1, it is evident from this that {xn} is a Cauchy sequence, and
by the completeness of X, there exists a point x in X such that x.--+ x.

fixed Point Theorems

and Some Applications to Ancdysis

339

We now use the continuity of T to infer that xis a fixed point:

Tx = TOim. Xs) = lim Tx = fun

X.+t =

z.

We conclude the proof by sho'Wing that x is the only fixed point.. If y

is also a fixed point~ that is, if Ty ~ Y~ then d(x,y) = d(Tx,Ty) 5 r d(x,y);
and since r < 1, this implies that d(x,y) = 0 or y = x.
Thia result is the key to

Picard's Theorem.. If f(x,y) and afjay are continuous in a cl-OBed Tectangle R == f(x,y):a1 < x <at and b1 Sy~ b,J, and if (xo,Yo) i8 an
interior point of R, th.en tJu differential equation
dy

dx

haa a unique solution y

= f(x,y)

(1)

g(x) which passeB th-rough (xotYo) .

PROOF. Since f(x,y) and fJf jay are continuous in R~ they are bounded,
and consequently there exist constants K and M such that
~

l/(x,y)[
and

:Y f(x,y)

(2)

<M

(3)

for all points (x,y) in R~ We next observe that if (x;y1) and

in R, then the mean value theorem guarantees that

l/(:i:,y1) - /{:i:,y2) I =
for some 8 such that 0

IY - y,I aya f(x, Yi+

<8<

If(x,y1)

I.

(x,y~)

9(y, - 111))

are

(4)

It now follows from {3) and (4) that

- f(x,y2) I 5 M IY1 - Y2I

(5)

for all (x,y1) and {x,y2) in R. 1

It is convenient at this stage to replace our problem by an equivalent
problem relating to an integra! equation. If y ;:::= g(x) satisfies (I) and
has the property that g(x o) = ya~ then integrating (1) from xo to x yields
g(x) - g(xo) =

g(x) = Yo

or

J.,: f(t,g(t)) dt

+ J~ f(t,g(t)) dt.

(6)

ConverselyJ iI y = g(x) satisfies (6), then it is clear that g(xo) = Yo, and
on ditTerentiating (6) we obtain ( 1). It therefore suffices to show that
the integral equation (6) has a unique solution .
The only uBe we nta.ke of the hypothesis that af/ay exists and is continuous in
R is to derive the so-ealled Lipuh.itz condition ( 5)~
i

:WO

.Appendices

To accomplish this, we choose a positive number a such that Ma < 1

and the closed rectangle R' determined by Ix - xol < a and IY - Yol < Ka
is contained in R. We now let X be the set of all continuous real functions y ~ g(x) defined on the closed interval Ix - XQI ~ a such that
lg(x) - Yol s; Ka. Xis clearly a closed subspace of the complete metric
space e[xo ~ a 1 xo + a] and is therefore itself a complete metric space.
Our next step is to consider the mapping T of X into itself defined by
T g = h, v.There
h(x) = Yo+
z f(t,g(t)) dt.

::f""

The fact that T maps X into itself is evident from (2), for
lh(x) -

y~J =

J,: f(t,g(t)) dt

< Ka.

Furthermorct it f ollow.s from (5) that

lh1{x) - h,(x}I =

/"': [f{t,g1(t))

f(t,g~(t))J dt

<Ma sup lg1(x) - g2(x)I;

and since Ma < lJ this shows that T is a contraction on XL We now
appeal to our lemma to conclude that the equation Tg ~ g has a unique
solution~ Since this amounts to saying that the integral equation (6)
has a unique solution, our proof is complete.
The ideas in this proof ha vc a much 1vider scope than might be
suspectedt and can be applied to establish many other existence theorems
in the theory of differential and integral equations.

APPENDIX TWO

CoHfiHllOllS tJlltVCS aHti t/te

Jlalt1t-Mozurkiewicz ?:lteore111
A continuous curve is usually thought of as "the path of a continuously moving point~H and this rather vague notion is often felt to carry
v.ith it the even vaguer at tribu tc of 'lthinness/' or "one-dimensionality."
For the case of plane curves, Jordan (in 1887) gave precise expression
to this intuitive geometric concept by means of the following definition~
if f is a continuous mapping of the closed unit interval I = [O,lJ into the
I
I

l
J

-- --+-_..

......

l
Il

+--+~ ~.!.- -

--tI
!

l~

.J__

- +--T--+
I

I
J

11 {l) I

12 (I}
Fig. 40

Euclidean plane R21 then the subsetf(/) of R 2 is called a continuous curve.

The fame of Jordan's definition rests mainly on Pean(/s discovery (in
1890) of a continuous curve which passes through every point of a closed
square~ Curves of this type have come to be called space-jilling curves.
In Fig. 40, we show the first three stages in the construction of a
particularly simple example known as Hilherts space-filling curve. If
the square under consideration is S == { (x ,y) : 0 < .x < I and 0 < y :::;; I L
341

342

Appendices

then the respective curves are the images of I under continuous mappings
f 1~ f 2l and f ;j of I into S.. The process of constructing these curves can
be continued in the same way, and it yields a sequence of continuous
mappings f. of I into S. By the manner in which ea.ch curve is constructed from its predecessort it is clear that the sequence {/'A} converges
pointwise to a mapping f of I into S; a.nd since this convergence is evi
dently uniform, f is continuous (see Problem 14-4) andf(l) is a continuous
curve in the sense of Jordan~ Furthermore, each point of S 1ies in f (l),
so f{/) is a space-filling curve.
Peano's discovery of space-filling curves was a shock to many mathematicians of the time, for it violated all their preconceived ideas of what
a continuous curve ought to be. To a few of the others~ however~ it
presented an opportunity. It suggested the very interesting problem
of determining what a continuous curve actually is, or in other words, of
finding intrinsic topological properties of a subset X of R 2 which are
equivalent to the existence of a continuous mapping of I onto X~
Before describing the solution of this problemJ we plar.e it in a wider
context by extending Jordan's definition. A topological space X is
called a continuous curve if X is a Hausdorff space and there exists a
continuous mapping of I onto X~ 1 We know that I is compact and
connected, so by Theorems 21-B and 31-B, any continuous curve is also
compact and connec.ted. Jn the lemmas below, we give two additional
properties which every continuous curve must have.
It is convenient to begin by introducing the following concept. A
mapping f of one topological space into a.nother is said to be closed if it
carries closed sets into closed sets, that is 1 if f(F) is closed whenever Fis
closed.. We shall use the fact that a continuous mapping of a compact
space into a Hausdorff Bpace is automatically closed (see the proof of
Theorem 26-E).
Lemma.

Every cantinuous curt e is second countable.

4

Let f he a continuous mapping of I onto a Hausdorff space X.

We must show that Xis second countable, that is, that it has a countable
open base. I is a separable metric space~ so it has a countable open
base IB..:}, and it is easily seen that the class i Gd of all finite unions of the
B;.'s is also a countable open base for I. Since I is compact and X is
Hausdorff~ f is closed, and therefore each set f(G/) is closed. The class of
all sets of the form f(G./)' is thus a countable class of open subsets of X,
so it suffices to shovl that these sets r.onstit ute an open base for X.
Let x be a point of X ,vith neighborhood G. rrhe set 1-1 ( {x]) is
closed and is therefore a compact subspace of I with neighborhood 1~ 1 (G).
For each pointy inf- 1 ( {x}) J there exists a set in {Gil which contains y and
PROOF.

A continuous curve in the sense of th is definition is of ten called a Pea no space.

Continuous Curves and the Hahn-Mazurkiewicz Theorem

3.43

is contained inf-1 (G). By the compactness of J- 1 ({ x}) and the fact that
{Gil is closed under the formation of finite unions, there exists a G, such
that

On talcing complements1 we obtain

1-1 ({x} )J

G/

1-1 (G)

I,

(1)

and since the complement of an inverse image equals the inverse image of
the complementJ we can -write (1) in the form

1- 1( lx} ')

G/

~ 1~ 1 (G')~

(2)

If we now apply f to all members of (2), we get

jJ-1({xJ') :) f(G/) ~ ff- 1 (G')
{x l I -;::) f (G/) =:. G' J
tx} c j(G/) 1 c G,

or
so

and the proof is

Lemma.

complete~

Every continuous curve is locally connected~

Let f be a continuous mapping of I onto a Hausdorff space

X. We must show that X is locally connected~ By Prnblem 34-It it
suffices t-0 show that if C is a component of an open subspace G of X, then
C is open.
Let A be a component of l- 1{G). Then A is connected, and therefore
f(A) is connected in G; and since C is a component of G, we see thatf(A) is
either disjoint from C or contained in C. It follows from this that/-1 (C)
is a union of components of J-l(G). Since 1~ 1 (G) is open and I is locally
connected, Theorem 34-A tells us that the components of 1- 1 (G) are open,
so ;i (C) is open and l-1 (C) 1 = 1- 1 (ly 1 ) is closed. We conclude the proof
by observing that since the mapping f is closed, the set jf- 1 (C') =:;: C' is
closed, so C is open.
PHOOF.

The above remarks and lemmas estab1ish the easy half of the following famous characterization of continuous curves.
The Hahn-Mazurkiewicz Theorem.. A topological Bpaee Xis a continu()U8
curve ~ X iB a compact Hausdorff space which is second countable, connecteil,.

and locally oonn.ected.

For the remainder of the proof, we refer the reader to Wilder [43,
p. 76J.. Additional discussions of a descriptive and historical nature -can
be f omid in Wilder [44] and Hahn [15] .

APPENDIX THREE

Nooleax Algebras,
Hoo!eax /(ixgs, 11Hd Stone$ 7:1teore111
We saw in Sec. 2 that a Boolean alge1Jra of sets can be defined as a
cle..ss of subsets of a non-empty set which is closed under the formation of
finite unions 1 finite intersect ions~ and com pl em en ts+ Our purpose in
this appendix is threefold: to define abstract Boolean algebras by means
of lattices; to show that the theory of these systems can be regarded as
part of the general theory of rings; and to prove the fa mo us theorem of
Stone, whlch asserts that every Boolean algebra is isomorphic to a Boolean
algebra of sets~
The render will recall that a la.ttice is a partially ordered set in which
each pair of elements x and Y. has a greatest lower bound x A y and a least
upper bound xv yJ and that these elements are uniquely detcnnined by
x and y. It is easy to show (see Problem 8-5) that the operations A and

v have the following properties:

Xl\X=X

XAy=yAX

x /\ (y /\ z)

(x r-.. y) /\ z
(x A y) V X = X

and
and
and
and

xvx

::::=

.x;

xv y :::;- yv x;
x v (y v z) :;: (x v y) v
(x Vy) AX = X.

z;

(1)
(2)
(3)
(4)

We shall see in the next paragraph that these properties are actually
characteristic of lattices. Bcfore proceeding f urthcr, however, we remark
that

x < y <=> x I\ y = x.

This fact serves to motivate the following discussion.

Let L be a non-empty set in which two operations A and v are defined,
and assume that these operations satisfy the above conditions. \\1 e
3.U

Boolean

Stone~ s Theorem

Atgebras, Boolean Rings, and

345

shall prove that. a partial order relation ::::; can be de fined in L in such a
way that L becomes a lattice in \vhich x" y and xv y are the greatest
lower bound and least upper bound of x and -y. Our first step is to notice
that x A y = x a.nd xv y = y are equivalent; for if x A y ;::::; x, then
xv y = (x " y) v y = (y A x) v y = yt' and similarly xv y = y implies
x /\ y = x. \\re now define x < y to mean that either x" y ;: : :. x or
xv y :;;;::; y. Since x Ax = x, '"Te have x < x for every x. If x < y and
y < x, so that x A y = x and y,... x ~ y, then x = x A y = y f\ x = y.
If x < y and y < z, so that x A y = x and y A z = !J 7 then
x

z ~ (x

y)

(y

z)

x:P

so x < z. This completes the proof that < is a partial order relation+
\"\le no'\v sho\Y that x A y is the greatest lower bound of x and y. Since
(x A !J) v x = x and (x A y) v y = (y A x) v y = y :t '\\"'e see that x A y < x
and x A y ~ Y~ If z < x and z < y 1 so that z Ax ==- z and z A y = z, then
z /\ (x A y} = (z A x) A y = z A y = z, so z < x I\ y It is easy to pro vc,
by similar arguments, thu t x v y is the 1east upper bound of x and y.
This chara<~terization of lattices brings the theory of these systems
somewhat closer to ordinary abstrart a1gebra, in which operations
(instead of relations) are usually pl aced in the foreground.
A lattice is said to be distributii\e if it has the follov.ring properties:
+

x " (y v z} =
and

X V

(y /\

Z)

(x

y) v (.t

= (X

y)

z)

A ( X V Z)

It is useful to know that (5) and (6) are equivalent to one another..
if (5) holds, then
(xvy)t-.(xvz)

(5)
(6)

For

[(xvy)Ax}v[(xvy) AZ]
=xv [(xvy) AZ]
~ xv [(x A z) v (y A z)]
= [X V (X A Z)) V (y AZ)
= xv (y f\ z) J

and a similar computation shows that (6) implies (5)~ We shall say that a
lattice is complemented if it contains distinct elements 0 and 1 such that

(7)
for every x (these elements are clearly unique when they exist), and U
each element x has a complement x' v.ith the property that
X A X

=0

and

xv x' : : : :;: 1.

(8)

We now define a Boolean algebra to be a complemented distributive lattice.

It is quite possib1e for an element of a complemented lattice to have
many different complements~ In a Boolean algebra, however, each

3"'6

Append ices

element has only one complement. To prove thls, we suppose that x* is

also an element with the property that x " x ==-= 0 and z v x == 1.
Then
z = x " 1 = x " (x v x') = (x "' x) v ( x A x')
= 0 v (x* A x') = x A :1/ J
If we now reverse the roles of x' and .x*, we obtain x1 < x,
In the light of this result, it is evident from (8) that xis the
complement of x':

so x < x'.
so x ;;:: x'.

(9)

Furthermore 1 it follows from (7) that 0

O'

0 and 0 v 1

and

l'

y)' = x' v y'

and

(X V

11 so we have

0.

(10)

The identities
(x

I\

y)'

X'

y'

(11}

are also true in every Boolean algebra. We shall prove the first part of
(11). Our principal tool will be the fact that
(12)

To establish (12), it suffices to show that x ~ y im plieB y'

proof of this is easy: if X < y, then :t A y' < y A y 1 = Qj 80
y'

y1

1 : : : ; y' /\ (x v x') = (y'

x) v (y 1

x I)

::;::;;

0 v (y' /\

< :.t' ~ and the

X

1
)

y'

x;'

and therefore y' < x'. We now turn to the proof of (x A y) 1 = x' v y' .
Our first step is to observe that if x ~ < z and y' :::; z, so that z' 5 x and
2/ < y, then z' ~ x A y or (x A y)' ::;; z. This sho~,..s that (x" y) 1 is less
than or equal to any upper bound of x' and y'~ so (x A y)' < x' v y'. We
conclude the proof by showing that x' v y 1 < (x A y)'. Thls follows at
once from the rcla tions x' ::;; (x /\ y )1 and y 1 < (x A y) ', ",..hlch, since they
are equivalent to x A y < x and x 1-. y < Y~ are evidently true. The
second part of (11) can be proved in essentially the same way.
One of the basic facts about Boolean algebras is that these systems
can be identified with a certain class of rings. Thls enables us to study
Boolean algebras by means of powerful techniques which are already
available in the general theory of ringsr
A Boolean ring is a ring with identity in which every element is
idempotent (i.e. 1 x 1 = x for every x)~ It is a surprising fact that multiplication in a Boolean ring is automatically commutative and that
x + x = 0 (or equivalently, x = -x) for every x. The proof of these
statements rests on the relation
x

+y =

(x

+ y) =
1

(x

+ y)(x + y)

zt

= x

+ xy + yx + y
+ xy + yx + yj

Bootecn Algebras, Boolean Rings, and Stone*s Theorem

347

which implies tha.t xy

y:r == 0, so xy ~ -yx. If we put. y = x 1 this
yields x 2 ~ ~xi~ so x = ~ x; and from this we obtain xy = - yx = yx.
Jn order to make a Boolean algebra A into a Boo1ean ring R, we
define addition and m ultiplira.tion by
X

+y

(X A

y1 )

V (X

y)

and

= x" y.

xy

(13)

(For the motivation behind these definitionst see Example 40-3.) To

verify that I~ actually is a Boolean ring~ \\re proceed as foIJo~,..s. It is clear
that
1
1
1
1
x
y = (x A y ) v (x A y) = (y " x) v (y " x )
= (y A X t) V (y 1 A X)

=
that

+0

(x

+ x~

O') v (xf " 0)

(x

1) v 0

=xAl =x.,

and that
x

+x =

(x A x')

v (x' " x) == 0 v 0 ~ O~

The proof that addition is associa ti vc is more com pli('atc<l.

venient to begin v.:-i th the o bscrvat.io n that
(X

+ y)

= [ (X
~ [(X

= [(X

= (X

y ')
V Y)

/\

V (X

/\

A X] V [ ( X V

X) V ( y A

JJ)

r=

y)

(X,

.r) j
y')+

(X1

y)

1f) /\ y 1]
{ { X t A y ~)

V (

(.T
y

It is con-

y')

y1) j

Now, using this~ we have

x

+ (y + z}

[x" (y + z) ] v {x " (y + z)l

= (x A ((y" z) v (y~ A z'))] v f:r' A ((y A z1 ) v (y' A z))]
= (x A y /\ z) v (x " y' A Z 1 ) v (x I " y A Z1 ) v (X 1 A y 1 A: z).
1

It is clear by inspection that the expression last written is unaltered by

interchanging x and 2, so x + (y + z) = z + (y + x); and since:r by
comrnutativity1 we have z + (y + x) == (.r + y) + z, it follows that
addition is associative. 1,he relevant properties of multiplication are
fairly easy to establish+ It. is immediate from the definition that
x(yz) = (xy)z:r

x, and that 1 is an identity~ In view of the fact that multiplication is obviously commutative~ a11 that remains is to verify that
x(y + z) = xy
xzJ and this is a consequence of th_e following computations:

that x 2

;;::

X (y

+ Z)

X A

(y

+ Z)

= X A { (y A z') V (y' AZ)]

= (X I\ y A Z. 1 ) V ( X A y' /\ Z) ,

348

Appendices

and
xy

+ xz =

(x A y)
r(x J\ y)

+ (x A z)

z) 1 l v [Cr A y) t A ( x A z) 1
[Cr A y) A ( x, v z') ] v [(x' v y') A ( x A z) J
~ ( X I\ y A X ') V ( X A y A Z ') V ( X t A X /\ Z) V ( y;
= (x A y A z') v (x A y' ;., z)
A ( x J\

X !\

Z)

Thus R is a Boolean ring.

\Ve no\r reverse t.his process; that is, 'vc start with a Boolean ring Rt
and we sho\\r that t.he definitions
x

and

y = xy

xv y = x

+ y + xy

(14)

convert it into a Boolean algebra . cl. If \Ve keep in mind the fact that
multiplication in R is commutative and that for every x \\ e have x 2 = x
and :r. == ~ :r, then (1) and (2) are evident. Property (3) follO"\VS from
the a.ssociativity of multiplication and the computations
7

:r v (y v

z)

x v (y + z + y z)
= x + y + z + y.z + xy

(xv y) v z = (x
~ x

and

Property (4) is a1so

(x" y) v x

true~

+ xz + xyz

+ y + xy) v z
+ y + xy + z + xz + yz + xyz+

for

xy v x = xy

+ x + xyx

+ xy + xy =

and
(xv y) Ax

+ y + xy)

(x

Ax = x 2

These remarks show that A is a lattice+

tive~ for

(y v z)

=p

x(y

+ z + yz)

yx

+ xyx

+ xy + xy

x.

F"'"urther, this lattice is distribu-

+ xz + xyz
xy + xz + xyxz

= xy

=
==

xz
(xAy)v(XAZ)r

xy v

It is easy to see that the elements 0 and 1 have the property that 0 <
x < I for every x and that x' = 1 + x acts as a complement for x, so A is
1

a Boolean algebra4
It is worth noting that the two processes we have described are
inverses of one another. Suppose we start \vith a Boolean algebra A and
use (13) to make it into a Boolean ring R:

+y

= (x

y') v (x'

y)

and

xy

==

x " y.

Boolean Algebras, Boofean Rings, and Stones Theorem

Nextt we use (14) t' convert R back into a Boolean algebra

It is apparent that x

xv y

and

xl..y=xy

"A y = xy =

===

349

A:

+ y + xy.

y; and since

+ x = (l A x') v (I' " x)

~ (1 A x

1
)

v (0

x)

= x'vO == x',
we also have

xv y = x + y + xy = 1 +

(1
x) (1
= 1 x'y'
= (x' " y')'
~xv y.

+ y)

This shows that the operations in A coincide with those in A. Conversely, if we start '\i\dth a Doo1ean ring ll, make it into a Boolean algebra
.11, and then convert A back into a Boolean ring ll, then the operations in
R coincide with those in R. "\Ve leave the verification of this to the
reader .
The ideas developed above show that Boolean algebras are essen
tialJy identical with Boolean rings. ~rhe practical effect of this is a
considerable sa.ving of labor1 for it allows us to transpose our study of
Boolean algebras into the more familiar cont.ext of the theory of rings 1
where many standard tools-ideaJst homomorphisms~ etc.--Jie ready at
hand. We illustrate this principle by proving the basic representation
theorem for Boolean algebras in t\vo steps: first, \Ve prove the corresponding theorem for Boolean rings; and second t we t.ra.ns1ate this result
back into the language of Boolean algebras.
Before entering into the details, we give a brief descri pt.ion of the type
of represent.at.ion we are aiming a.t+ If X is a compact Hausdorff space~
then each of the sets 0 and X is both open and closed (or more briefly,
open-closed)~ and the class A of an such sets is a Doo1ean algebra of subsets
of X. If X is disconnected 1 then A contains a.t least three sets; and if X is
totally disconnected, then A may contain a great many sets) for, by
Theorem 33-C, it is an open base for the topology of X.. :i-~urthermorP~
\\~e kno'v that A becomes a Boolean ring of sets if addition and multiplication are defined by

A+ B

(An B 1)

(A;

B)

and

AB= An B.

Our basic representation theorem states that ei.rery Boolean algebra

(Boolean ring) is isomorphic to the Boolean algebra (Boolean ring) of aU
open-closed subsets of some totally disconnected cqmpact Hausdorff
space.
Now for the details.. The simplest of all Boolean rings is the ring
{0, l} of integers mod 2 1 and this ring is evidently a field. Conversely,

350

Appendices

any Boolean ring which is a field necessarily equals {0 1 1} ~ 1,o see this, it
suffices to observe that if x is a n on-z r.ro e1em en t in such a ring~ th en

xx- 1

x 2x- 1 = .r(xx- 1 ) = x l = x.

If I is a proper ideal in a Boolean ring Rt then the quotient ring fl/ 1 is

also a Boo]ean ring; for Il/ I clearly has an identity, and
(x

+ 1) = x
2

+I

=x+I

for every :r in R. Thus, by Theorem 4 I-C, R/ I .==:. IO~ l l {=::>I is maximal.

Since every homomorphism of R arises from a.n ideal in R, this tells us
that the homomorphisms of R onto 10~1 ~ are precisely those of t.hc form
R ~ RI ~l, where Mis a maximal ideal in R. A standard app]icatiun of
Zorn's lemma shows that R has maximal idcals:t so there do exist homomorphlsms of R onto lO, l} ~ We shall need the follo,ving stronger
statement.
If x is a non-zero element in a Boolean ring ll, then there exists
a homarnorphism h of R onto ~0,1 J such that h(x) ~ 1.
PROOF.
By the above remarks, it suffices to show that there exists a
maximal ideal in R which does not contain x. Since x s-= 0, there clearly
exists at least. one ideal which does not .contain x. If Zorn's lemma is
applied to the set of all ideals which do not contain x, we obtain an ideal
M which is maximal with respect t.o the property of not containing x~
We conclude the proof by sho~dng that M actually is a maximal idcalr
To prove this, it suffices to show that M contains 1
x (for it will then
follow that any strictly larger ideal contains both x and 1
x, and so
contains 1). We therefore assume that Jrl does not contain 1
xj and v-.Te
deduce a contradiction from this assumption. It is clear that
Lemma.

+
+

I= {m+r(l+x):ms.MandreRl

is the smallest ideal containing both M and 1 + x so I properly contains

M. However, I does not contain x; for if it. did 1 we would have
1

+ r(l + x)

ior some m and r 1 and this implies that

+ r(l + x)]x
= mx + r(x + x~)
~ mx + r(x + x)

= x2 =

[m

== mx

conirary to the fact that xis not in Af. This contradicts the maximality
property of M, and the proof is complete.

We are now in a position to prove our principal theoremt

Boolean Algebras. Boolean Rings. and Stonets Theorem

351

The Stone Representation Theorem. If R is a Boolean ring, then there

exists a totally disconnected compact Ilausdorff Bpace H such that R is
iso1norphic to tlie Boolean ring of all open~closed subsets of ll.
Pn.001
Let H* be the set of all mappings of R into the Boolean
ring {0,1]. If for each x in R we define II~ by H~ = [OTl L then H* is the
product set P:p.H. H ~ \V. . e now impose the discrete topology on each H Z.j
and thus convert it ~to a totally disconnected compact Hausdorff space.
This permits us to regard H* as a product space~ and it is also a totally
disconnert..ed compact Ha.usdorff space. 14,or use in the next paragraph,
vlc note that if xis any given element of R 1 then each of the sets
1, .

it:f(x)

0}

and if :f(x) = 1} is open-closed. This follows at once from the fact that
each is the inverse image of an open-closed set in H:., under the projection
of H* onto H~.
We now pass to the subspace II of H* whirh consists of .all homorn orphisms of R on to I 0, 1 J . It is cl ear that H is a totally disconnected
Hausdorff space. To prove that it is also compact, it suffices to show that
it is closed in H*} and thls v-.,.e do as f ollows4 A homomorphism of ll onto
{Otl J is of course a mapping fin H such that f(x
y) = f(x)
f(y)
and f(xy) = f(x)f(y) for all x and y a.nd such that /(1) = lr It is evident
from this that H is the intersection of the following three subsets of H*;

r\~i~R{f:f(X
f\:z;,JJtR

+ y) = f(x) + f(y)J,

l f :f(xy)

{f:j(l) =

and

f (x )j(y)} ~

(15)
(16)

114

{17)

We know from our remark in the preceding paragraph that (17) is closed;
and if we can show that the other two sets are a.I so closed, th en it will
follow at once that H is closed. We inspect (15). If x and y are any
given elements of R~ then it is easy to see that

{/ :f(x

+ y)

f(x)

+ f(y) l

is the union of the fallowing four sets:

and

{f:f(x) =
{/ =f(x) :=::
{f:f(x) ===
{j:f(x) =

0, f(y)
0, J(y)
1, f(y)
1, f(y)

== 0, and f(x
= 1, and j(x
= O, and f(x
= IJ and f(x

+ y)

+ y)

+ y)

0L

1L

= 1L
+ y) = O}.

Each of these sets, being itself the intersection of three closed setst is
closed, so {/ :f(x
y) == f(x) + /(y)} is closed, and consequently (15)
is also closed~ A similar argument shows that (16) is closed, so 11 is
closed and therefore compact.

352

Appendices

Our next step is to exhibit an isomorphism T of R into the Boolean

ring R of all open-closed subsets of H~ We define T by

T(x)

tf~f

It is clear that T maps R into R.

T(x

+ y) =

1 }.

T is also a homomorphism, for

lf:f(x + y) = 1}
iJ:f(x)
f(y) = 1}
{f :f(x) = l} + {/ :f(y) == I I
T.(x) + T(y)
l f :f(xy) = 1}
{/ :f(x)f(y) = 1}
tf :f(x) = 1} n i I :f(y) = 1}
T(x) T(y)4

=
=

T(xy) ;:;;

and

e Hand /{x)

;;

OW' lemma tells us that T(x) is non-empty whenever x rf 0 1 so T is an

isomorphism of R into Rr It vrill be useful in the next paragraph if we
also note here that

= {/ :/(1) =

T(l)

1}

H,

for it follows from this that

+ x)

T(l

== T(I)

+ T(x) = H + T(z)

= T(x)'

for every x in R.
Finally~ we show that T maps R onto R.
We begin by observing
that the topology of H is defined by means of basic open sets of the form
B = ~ f :f (xt) ==

Ei 7

i = I, ~ . ~ ~ n ~ ~

where {x 1, . . , x~ l is an arbitrary finite subset of Rand each ti equals

0 or 1 1.'hese sets arc e vide ntl y closed as well as open. FurthcrmoreJ
every set of this kind is in the range of T ; for since
+

if we define Yi to be x1 or 1

+ x* according as t..: equals 1 or 0. then

B =
=

n:=-1 ff :f(x1)
r\L1 tf :f(yi}

=
-;::=

i1J
11

= n:._1 T(yi)

:::::: T(y1 y~)+

We now consider an arbitrary open-closed set Sin R. Since Sis compact
and the B 1 s constitute an open base, S is the unioI1 of a finite number
of B 1 s1 say B1,
B.; and by the above result~ each 8 1 is expressible
r

Boolean Algebras,, Boorean Rings. ond

in the form Bi = T(z,) for some element z; in R.

s = Vi-1 Bi

en;. . \ B/)

Stone~s

Theorem

353

It now follows that

= (rl~l 7'(zt)')'

(r\;:l T[l + Zj])'

= (T([l
Z1] [l + .zm]))'
= T(l
[I
Z1] ' [l
ZmJ).

+
+ +

This shows that Tis an isomorphism of R onto

R~

so the proof is complete~

Y.le now conclude our theory by translating Stone's theorem into the
language of Boolean algebras.
Let A and A be Boolean algebras+ A mapping h of .ti into A* is
called an isornMph isrn (or a Boolean algebra isomorphism) if it is one-to-one
and has the following three properties: h(x A y) = h(x) A h(y) 1
h(x v y)

h(x) v h(y),

and h(x') = h(x)'. A is said to be isomotphic to A* if there exists an

isomorphism of A onto A*. If A and A* are converted into Boolean
rings R and R* ~ then it is easy to show that every Boolean algebra
isomorphism of A onto A is a Boolean ring isomorphism of R onto R*,
and con vcrsely. l\r e leave the details to th c reader.
1'hese ideas make it possible for 1is to state the follol\Ting equivalent
form of StoneTs theorem: If ..4. is a Boolean algebra, then there exists a
totally disconnected co-nl.pact II ausdorff space 11 such ihat A is isom.orphic to
the Boolean algebra of all open-closed subsets of H.

l!ibliograplt11
1. Achieser, X. Ir: "Vorlesungen tiber Apptoximationstheorie,n Akadcmie, Berlin~ l 953.
2. Alexandroff ~ P. ~ and 11. Ilopf ~ "Topologie ,' ~ Springer, Berlin, 193b.
3. Bers, I~~: l'Topology,'' 1ecture notes,. Xe,v "'Y~ork University Institute
of lvf athcmatical Sr.iP.nces~ Kc'\v York, 19fi7.
4. Bir kho.ff ~ G. : '-~Lattice Theory,'' American Ma the ma ti cal Society
Colloquium Pub1ica tions, vol. 25, N e\v r~ ork 1 1948.
5. Rirkhofi, G. 11., and 0. D. Ke Hogg: Invariant Points io Function
Spare, Trans. A Jner. kl ath. J...';or.~ 23 (1922), pp. 96-1 lfi~
6+ Conran t~ R.. ~ and H. Rob bins: 'i \\That Is l\.la thematics?" Oxford,
I~ondon and Ne'v 1. . ork, l941.
7. Dixmicr~ J.: "LP.s a.1gebres d~opcrateurs dans Fespace hilbertien
(A1gebrcs de von X eumann),'' GauthiP.r-\'.iHars, Paris, 19t57.
Sr Dunford, X., and J T ~ Sch"\vartz: ~~I "'inear Operators, Part I;
General Theory,'' I ntcrscience,. Ke\v \' ork, l 958+
9~ Fracnkel 1 A. A.: ''Abstra.ct Set Theory,~~ Korth-Holland 1 Amsterdamt
1953.
10. Fraenkel 1 A. A.~ and Y. Bar-IIiUel: i~Foundations of Set Theory,''
North-Holland, Amsterdam, 1958r
11~ Ga.1 1 I. S.: On Sequences of Operations in (~01nplete Vector Spaces,
A mer. Math. Monthly~ 60 ( 1953) 1 pp. 527-538+
12. GOdel, K.: What Is Cantor~s Continulllll Problcm~r Amer. J1 athr
Monthly 1 54 (194 7) ~ pp. 515-525.
13. Goffm~1n, C.: Preliminaries to Functional Analysis, in ustudies in
Mathematics," vol. 1 i l\.Iathematica] Association of America, 1962.
14. Goldberg~ R+ R.: 'lFourier Transforms/' Cambridge~ New York, 1961.
15. Hahn, H. : The Crisis in Intuition, in l'The World of :rvrathematics,' ~
Simon and Schuster1 New r~ork 1 1956.
16+ Halmos~ P.R.: uNaivc Set Theory/ 1 \Tan Nostrand, Princeton 1 N.J.,
T

1960.
: ".Finite-dimensional Vector Spaccs,'J ,ran N ostrandt Prince . .
ton~ K .J., 1958.
18.
: ul\leasure Theory/' Van Nostrand, Princeton, N .J., 1950 .
17.

. 355

356

Si bliogf a phy

19+ Ilewitt~ E.: The Role of Compactness in Analysis, A mer. Math.

J.f onthly 1 67 (1960), pp. 499-516.
20. Ilille, E.~ and R.. S. Phillips: l"F'unctional Ana1ysis and Semi-groups,,'
American l\fu thematical Society Colloquium Publications~ vol. 31,
Providence, R.I., 1957.
21. Hurev-ricz~ W.~ and H. Wallman: uDimension Theory," Princeton,
Princeton, N .J ., 1941
22. ICadison, R. \r.: Order Properties of Bounded Self-adjoint Operators 1
Proc ..Am-er. ltf athr Soc., 2 (1951), pp. 505-510.
2-~. Kakutani, S~~ and G. W. ivlackey: Ring and Lattice Characterizations
of Complex Hilbert Space, Bull. Am.er. Math . Soc.j 52 (1946), pp.
r

727-733~

24. Kamke, E.: '"'fheory of Setsj" Dover, New York, 1950.

25~ Kelley, J~ L~: (~General Topology/' Van Nostrandj Princeton, NaJ.~
1955.
26+ Kolmogorov, Aa N., and S. V . Fomin: uElcments of the Theory of
functions and Functional Analysis," 2 vols+, Graylock~ Rochester
and Albany, 1957 and 1961.
27 J-'oomis, L. H. ~ '~An Introduction to Abstract Harmonic Analysis, n
\Tan Nostrand, Princeton t N .J., 1953.
28. IJorch, E. R+: The Spectral Theorem, in "'Studies in ~Ia thematics/'
vol. 1 ~ 1lath cmatical Association of America, 1962.
29. Lor~ntz, Gr GT: "~Bernstein Polynomials~'' l.Jniversity of Toronto
Press~ Toronto~ 1953.
30. 1-.-l ackey, G. W. : ~,unctions on Locally Compact Groups 1 Bull. A mer+
1lf ath. ~';ocT, 56 (1950)~ pp. 385--412.
31. 1IcCoy, N. II~: ~'Rings and Ideals/" Carus Mathematical Monographs~ no. 8, Mathematical Association of America, l948.
32. N nimar k~ :r\if. A. : ~'Normed Rings,'~ N oordhoff, Groningen, N ethcrlands, 1959.
3~. Niven~ I.: ''Irrational Numbers 1 '~ Carus :Mathematical Monographs,
no. 11, 11athematical Association of America, 1956.
3.J:. Rickart, C. E.: ''General Theory of Banach Algebras~'~ \ran Nostrand,
Princeton, N.J.~ 1960.
35 Riesz, Fr J and B~ Sz.-N agy: I:~ Functional Analysis," Frederick u ngar'
New York, 1955+
36. Russell 1 R.: uMy.Philosophical Development," Simon and Schuster,
N P"\V "\r ork~ 1959.
37 ~ Sierpinski, W. ~ '"Cardinal and Ordinal Numbers,u Monografie
I\.latematyczne, voL 34, Warszawa, 1958.
38~ Smirno'r, "\.... M. : A Necessary and Su:ffi.cien t C-ondi tion for !v[etri zability of a Topological Spa.ce 1 Dokl. A kad. N auk SSSR, 77 (19.51),
pp. 197-200.
I

Bibi iog raphy

357

39. Stone 1 M. H+: On the C-ompactification of Topological Spaecs, A n.n.

Soc. Pol. Math.~ 21 (1948), pp. 153-160.
40.
: A Gencra1ized \\t eierstrass Approximation rrheoremJ in
''Studies in 1-Iathema.tics,' ~ vol. I, ivlathcma.tical Association of
America, 1962~
41. Taylor, A. E.: HJntroduction to }. . unctional Analysis," \\~iley, Ne"'
York~ 1958.
42. Wilder, R.. L+: 1 'In troduction to the :Foundations of Ma t.hemati(s,''
)\liley, New york, 1952.
43.
: l'Topology of Manifolds/~ American Mathematical Society
(~olloquium Publicationsr vol. 32:P New 1 . ork, 1949.
44.
: The Origin and Gro,vth of 1-lathcmatical (~oncepts, Rull.
Atner ~ llf alh. {)oc., 59 (195~~) ~ pp. 423-448.
4.5. Zaanen, .!:l. C.: . , J\n Introduction to the Theory of Integration/)'
N ort..h-lfolland, . Amsterdam, 1958.
46. .Zygmund, A. : l .cTrigonometric Series, ~ 2 vols., Cambridge, Nev{ i" or k,
1

1959~

JndeK of S!fmbo!s

A= B, A ;:e B
ACB
ACB
AUB

AnB
A'

A-B
AAB
a= b (mod m)
la,bl, (a.~bl~ etct

No
A

Ai< A2
An

A
CB(N~N')

m(N)
/3(X}

Equality and inequality for sets, 5

Set inclusiont 5
Proper set inclusion, 6
Union of two sets, 8
Intersection
t WO sets, 9
Con1ple1nent of a set, 10
Difference of t '\'\"O sets, 13
Symmetric difference of two sets, 13
Congruence modulo m for integers 1 30
Intervals on the real 1inet 5~ 57
Cardinal number of a countably infinite set, 34
Closure of a set 1 68, 96
Ord er relation for self-adjoint opera tors, 268
1-"'otal matrix algebra of degree n, 284
Function algebra representing a commutative
Banach algebra~ 319

of

Space of bounded (or continuous) linear tra.nsformations of N into N', 221

Algebra of operators on N, 222
Stone-Cech compaetification of X 1 139, 141

Cardinal number of the continuum, 39

e[a,b1

e(X,R)

Complex number system~ 23, 52-54, 214

n-dimcnsional unitary spacet 23~24 1 89"-90, 214
In finite-dimensional unitary spaceT 90
Extended complex planeJ 162-163
Space of bounded continuous real functions on
[0,1], 56
Space of bounded> continuous real functions on
{a,b}, 84
Space of bounded continuous real functions on X,
82J 106

360

Index of Symbols
Space of bounded continuous complex functions
on X~ 84, 106
Spaces of continuous functions on X which vanish
at infinity~ 165
Space of bounded continuous scalar-valued functions on X, 216

e(X,C)

e(X)

Distance from one point to another~ 51

Distance from a point to a set~ 58
Diameter of a set, 58
Derived set, 96
Kronecker delta, 283
Determinant of a matrix, 287

d(x,y)
d(x,A)
d(A)

D(A)
,.ii

det ({ai;])

Empty set, 5
J:x~

J-1: y __,., x
j(A)
J- 1 (B)
J~g
F~

fM
G
af:X--+ Z

Int (A)
.

i.r

==>, ~
r\..:Ai etc~
1

~cat

+co

inf A

Im.
I(F)

lim
lfJ.P
l,,

Xn

Function (or mapping) with domain X a.nd range

in Yr 16
Inverse function (or mapping)~ 17
Image of a set under a mapping, 18
Inverse image of a set under a mapping, 18
Equality for mappings, 20
Induced function a I on a conj ugatc space, 231
Multiplicative functional induced by a maximal
ideal, 321
Group of regular elements in a Banach algebra~ 305
Product of two mappings f: X---+ Y and g: Y --t Z 1
19
Interior of a set, 63 ~ 97
Identity mapping on a set, 20
Implication and logical equivalence, 6
Intersection of a class oi sets 1 11
Infinity (minus and plus), 56
Infimum {or greatest lower hound) of a set of real
numbers, 45, 57
Integers modulo m, 182
Ideal associated l\Tith a closed set~ 329
Limit of a sequence, 50~ 71
Banach space of n-t.uples~ 214
Banach space of sequences~ 215

Index of Symbols

Lp

zn

i;l(I

lot:i, c, Co
L/~l

grr

m]n A, max A
M~

J.11

+N

~f

.J.\l
m < n, m :::; n

N*
N**
Jtl/ }.{

361

Banach space of measurable functions, 215

Banach space of n-tuples_, 216
Banach spacPs of sequences, 216
Quotient space of a. Jinear space with respect to a
subspace~ 193-194
Group algebra of a finite or discrete group 1 303-305
Space of maximal ideals, 319
Minin1u1n and maximum of a finite set of real
numbe'rs, 45
11aximal ideal associated \\Tit.h u point, 327
Sum of i.\vO subspaces of a linear space~ 19S
Direct sum of tviro subspaces of a linear space, 195
Order relation for cardinal numbers, 35, 48
Conjugate space of a normed linear space~ 224
Second conjugate space of a nor1ncd linear space,
231
Quotient. ~pace of a normed linear space with

respect to a closed subspace, 213

Product of a class of sets~ 25
Projection of a product onto a coordinate setj 25
R
R X R (or [i,'!1.)
n~

R.,;J
r(x)
p(x)
Ill!
'

s
S*
SJ_

[SJ
Sr(Xet)
S,{xo]
sup A

c-( T)
O'(X) 1 CTA(X)

Real num her system, 21, 52, 214

Coordinate plane, 22
n-dimenHional Euclidean space, 23-24, 8fi-89_, 214
I nfini te-d irnr.nsiona.l Euclid can .space, 00
Spcctra1 radius of x, 310
Resolvent set of x, 309
Quotient ring of a ring \vith respect to an ideal, 187

Set of singu1ar ele1nents in a. 13.anarh algebra_, 305

Clo:=;r.d unit 8phere in a conjugate spaeej 233~234
()rthogonal coin plcrn en t, 2,19
Subspace spa.nnc? by ""' 194
Opeu sphere \Vith.ra.dius r and center Xut 59
Closed sphere \\rilh rRdius r a.nd ccnt~r xo, Go
Su prcrn u1n (or least upper hound) of a set of renl
nurnhers, ,15, 56
Spectrun1 of an operator, 289~ 2'06
Spectrum of an element in a Banach algebra,
308

362

Index of Symbols

Conjugate of an operator, 241

Adjoint of an operator, 263
Norm of an operatorJ 220-221
Matrix of an operatorJ 281

T*

T
ll Tl!
{T}, [T]B

u
u,.A1.t etc.
x--+ y
X11 _..,, .X

x""

X1 X X'J,

[x]

!lxl1
Ux + MH
x(M)
~

x().)

A y,

(X1Y)
x e: A 1 x
x _l y
x .-....- y
x<y
x ~

Vy

tA

y (mod I)
x - y (mod M)

Universal set, 5
Union of a class of sets1 11

Mapping notation, 17
Convergent sequenceJ 50, 70-71, 132
One-point compactification of X, 163
Product of two sets1 23
Equivalence set associated with x, 27
Norm of x 1 54, 81, 212
N arm of coset x -t- ]{, 213
Function on marximal ideals, 318-319
Function on maximal idcalst 319
Rcsolvent of x, 309
Meet and join of x and y, 46
Inner product of x and y, 245
x is (is not) an element of A 1 5
x is orthogonal to y, 24 9
xis equivalent to yJ 27
Order relation for real numbers and partially
ordered sets~ 7J 43
Congruence modulo an ideal in a ring 1 186
Congruence modulo a subspace in s linear spscet

193

Adjoint of x, 324

Set of topological divisors of a.ero in a Banach

algebra, 307

811/Jjeet J11tler
Absolute value, on complex plane, 53
of a !unetionj 159
on real lin et 52
.A.chieser, N. I., 157
Adjoint~ of element in Banach ~a.JgebraJ
324

of operator, 263
Adjoint operation (involution), on ffi(H),
265
on Ba.na.ch 111 -aJgebnt,. 324
AtexandrotI, P ., 130n.
Alge bra, I 06) 208
B*~,

324

Banach, 302
Banach *-, 324

lloolesn {see Boolean algebra)

center~

210
commuta.tivei 106

corn pl ex, l 06

C*-, 303
disc, 303
division,. 208
~roup, 303-305
hon1orr1orphism, 210
ideal jn, 209~ 313
\\ith i<lentity, 106
ison1orphism, 210
quotient aJJ?;ebra of, 209
radical, 314
r.egu1ar representation, 210
sr:1nj-.":Simplc~ 316
sub algebra of J l06i 208
tot a] matrix r 28'.l
von Neun1ann, 303

TV*-, 30:1.
Antisymmetry, 43
Arzela.)s theoremr 128

Ascoli's theorem~ 1261 128

Axiom of choiee, 46

B *-algl'.!bra J 324
representation, 32.&-.326
Baire~s theoremt 7 4, 75n .
Banach n.lgcbra., 302

Banach subalgeb.ra ofJ 302

representationt 305
Banach *-algebra~ 324
*-isomorphism~ ~24

Banach apace, 82, 212

clo:sed unit sphcrcj 217, 232
represcntation 1 234
un.if ormly convex, 248
Ba.nach~Steinhaus theorem, 240
Banach-Stone theorem, 330
Bar~llillclt Y., 11 46n.
Basc-:i- c1osed, 112
generated by subbs.aeJ 101, 112
openj 99
Baais, 197
orthonormal) 293
Beilj E. T.J 37
Bernstein polynomi.alsJ 154
Berst L., 338
Bessel's inequality, 252-253, 257
Birkhoff, G+t 29, 46n+, 47
Birkhoff, G. D . , 338
Bolzano-Weierstrass property, l 21
Bolza.no-Weierstrass theorem,. 121
Boolean algebra, 345
a.s Boolean ring, 347-349
isomorph ism, 3 .53
representation i 353
~

of sets J 12* 344

363

364

Sub;ect Index

Bool ea.n ring, 34 6

. as ll-Oolean algebra . 348-349
ag fi cld~ 349-350
maxim al ideals in J 3 50
representation) 351
se1ni-sim p1irity, 300
llou n d ary i 68~ 97
Boundary pointi 68~ 97
Bounded functiont 55
Bounded linea.r trRnsf ormation~ 220
Bounded 1nn.pplng1 .58
Bounded set 1 58
Brou"".er' :s fixed point th eore Ill, 3 38

C 111 -al~cbra~

303
oom1nutative 1 33Z...334
Canonical form problem for matrices~ 286
Cant.or~ G. i .31-4:3.i 49
C.an tor continuum hypothesis, 39
Cantor intersection theorem~ 73
C.a.n tor .set, 67
Cardina.1 nun1bf'r(s )~ 31
comparability t.neorem, 48
of con tinuurn, .39
finite, 32
Carte~i.un product (product u sets),
23--25
Cau-chy seqnoence, 71
Cauchy"s inequa.J ity, 88, 219
Cs.yleyts theorem, 181
Chain, 44
Charaetcristic cq u a tion, 288-289
Char acteristi e v a] u e1 2 7 8n~
Char act ieristie vector, 27 8n.
Choice, axiom of, 46
Class, 4
disjoint, 9
Closed baseT 112
generated by closed subbase,, 112
Cl oscd graph th core m ~ 238
Closed mapping~ 342
Closed rectangle . 101, 119
Closed set~ 65, 9.5
Closed eph eru Ou
Closed strips, 101
Closed ~u b base~ 112
Closed unit ~phere, 217~ 232
Closure,, 68, 96
Compact subspace, 111
Com pa.ct to po] ogic a1 space~ 1 l 0 ~ 111
j

Compactification, one-pointi 163

Stone-Cech~ 14 lio 331
Com par ability theorem for cardinal
numbers~ 48
Comparable elements . 7 7 44
Com pl et.e me t..r ic sp.a.ce.. 71
Complete orthonormai set, 25~5
Coin p1 ct cJy regular space~ I 33
Comp] etion of metric space~ 81-85
Com pl ex pl an e ~ 23 ~ 52-54
e.xtendedi 162
Component of a. space~ 146
Congruent n1odulo, &n idcal 1 186
o. llu ear subs pace J 193
a po~l ti ve in tc ger ~ 30
Conjugate, of a function, l OS, l 61
of operator 1 241
Conjugate space~ 224
Connected sptt.cei- 142, 143
Connected subspace, 143
Continuous curve, 341-342
Continuous fun etio n, .'10
Cou tin uous im a~e ~ "93
Continuous Jin Par f un1.."':tiona.l, 224
Continuous linear trn.nsf onnation ~ 219220

Continuous mapping, 76, 93

jointly, 118
at point, 75-76t l 04
in sing]p, variable,, 118
Continuum~ cardinal number f 39
hypothesis 1 39
Contraction, 338
Convergence of fun rtions 1 point.\\ise, 83
t~niform, 83
Convergent sequence, limit~ 50, 71, 132
of numbers, 50
in a spacet 70, 132
C-on vex se(, 148
Convolution, ~04--30S
Coordinate plane 1 22
Cose ts, 186-187:r 193-194
Countably com pact 5pa.cc, 114
Courant:t R.j 94 1 338
Cu rvP, cont in uo u :s, 34 l-3-! 2

Dense ( e.very\\here. dense) set, 70t 96

Derived set, nfi
Determinant, of matr ix, 2.8 7
of operator~ 288

Subject Index
Dim enaion, of a linear spaceJ 200
orthogonal t of a Hilbert spaee, 259-260
Disc algebra, 303
Disconnected epa.ce, 143
DiBconn ec tion of a. space~ 143
Discrete space~ 93
D isc.rete topology~ 93
Discrete two-point spac c) l 44
Disjoint linear suhsp.aces) 195
Disjoint sets 1 9
I) is tanc e, from point to set J 58
bet ween tvito point.BJ 51
Diatrib u tive 1a ws ~ for In tticcs;11 345
for sets;11 10
D ivi.slon alge hr a 1 208
Diisor of :zero, 183
topologies.I, 307
DixrnierJ J.) 303n.
Dunford~ N., ix, 226~ 232n.
Eigcnspac e, 278
Eigenvalue, 27 8
Eigenvector~ 278
Element(s)~ 3
com pa.re.bl c ~ 7 ~ 44
maJ[imal, 44
in ring, regu]ar~ 183J 314
singular1 183~ 314
Empty set, 5
rnetJ 123
Equicontinuous functions, 126
Equivalence rela.tionJ 27
Equivalence set~ 27
Euclidea.n plane~ 22~ 87-88
Eu cl jd ean space, in fini tc-dim en sion alt 90
n-dimensional, 24, 87 ~ 214
Everywhere dense seti 70, 96
Extended complex plane, 162
Extended real number system~ 56

Family_, 4
Field, 184
Fini tc intersection property) 47 ~ 112
Fir.st countable space, 1OOn.
Fixed point1 338
Fixed point space, 337~338
Fixed point theorem~ Brouwer is, 338
Schauder'ij, 338
FominJ S. V., 128, 215nT
F-0urier coefficients~ 256 257
1

365

Fourier expa.nsion, 256,. 257

Fraenkel., A. A., 7J 42n. 4.6n.
Full linear group, 207
Function{s );p &bsol ute value, 159
bounded 1 55
complex, 17
conjugate, 1OSJ 161
constant J 16
continuous, 50
at point, 50
in contrast to mapping, 17
convergence~ pointwise., 83
uniform, 83
definition J 16
domaint 1.5, 16
cquicontinuous1 126
ex tension 1 17
gener ali tics, 14-16
imaginary part, 161
moments) 157
range, 15_, 16
rea1, 17
rea.l part J 161
restriction J 17
uniformly bounded 1 12Sn..
vanishing &t infinity, 165
( See al.so Ma. ppin g)
Function ap-a.ces, 82
Functional(s)~ 224
extension~ 226--228
on Hilbert apace) representation, 261
induced) 231
tn ultiplic a ti v e ~ 321
Funda.menta1 theorem of algebra:1 245 1
289~

CAI) 1.

310

s.)

Gali!eo~

240

33
Gelf&nd ma.pping, 319
Gclfa.n d rcpre.sen ta tion theorem, 322
Gclj~nd-N eumark th eoremst 318~ 325326
G Od-nl ~ K~ ~ 39n.
Goffman, C.;11 128
Goldberg~ R. R., 305-n.
Gram-Schmidt proceSB~ 258, 295
Graph of ma. pping~ 23
Group~ 172-173
Abelie.n (commutative)) 173, 179
additive) 179

366

Subiect Index

Group, abstract vs. concrete,

center of~ 180

173n~

circ1eii 174
finiteii 173
full lineart 207
homomorphism) 180
identity in 1 173J 179
infinitet 173
inverses,, 173, 179
isomorph ie J 180
isomorphism,. 180
order:t 173
perm u ta ti on~ 178
regular represe nta. ti on, 181
subgroup or, 178
:symmetric~ 176
of symmetries of square;11 176-177
transformation~ 178, 181
(iro up alge bra 1 303--305

IInhn,, H.,, 343

Hahn-Barra.ch

theorem~

2llJ 228
gen cral iz ed form~ 230-231
Hahn~ Mazurkie w ic z theorem J 343
llalmos, P. R~, 42n~, 46n~,. 172, 215n.
Ho. usd o :r II space~ 130
H cine-Borel theorem, 11 O~ 114
con verse t 115
gen er ali.zed 7 119
Her.in ite run ctiona~ 259
HewittJ E. ~ l21n~
Hilbert~ D~J 49
Hilbert cube,. 248
Hilbert space( s) ~ 245
s.mong com ptex Banach spacesjl 248
inner prod ue t ~ 245
orthogonal eomplementeJ 249

orthogonal dimension, 259-260

orthogonal au bspa.ces J 250
ort hogona.l vectot"i!, 249
orthonorm&l set, 251
oomplete,, 255
parallelogram law, 24 7
Pythagorean theorem;11 249
re pre sen ta. tion, 260
representation of runction&le on~ 261
Hilbert's apace-filling eurve, 341-342
Hille, E~, ix, 232n.
Holder~ s in equality~ 218
general form, 219

Holder's inequality., in relation to

Cauchy's, 219
Homeomorphic image, 94
Homeomorphie spac.es~ 94.
Homeomorphism, 93
Hopf t H. ~ 130n.
Hurewicz, W., 150

Ideal J in a.lgebrat 209 i 313

con tra.sted with ring ideal ii 209
ma:ximal" 314
in ring, 184, 185
general significance, 1&8--190
maximal,, 190
lrn age J continuous,. 93
ho.in eomor phic, 94
Induced f u nctiona.ls, 231
In equality, Bessel's, 2 52-253t 257
Cauchy's, 88~ 219
Ho1der'sj 218 7 219
!\finkowskii~, 88,, 90, 218J 219
Sch'\\Tartz.'s~ 246
lrin.ngl e J 5 I
lnlimum, 45
Infinite-dimensional Euclidean spaceJ fJ
In fi ni t-e-d i 111e nsio n o.l uni ta.ry sp~e ~ 90
Inner product, 245
Interior~ 63J 97
Interior poin ti. 63, 9 7
In terva 1st 5, 57
Involution J 324
lsol e. ted point t 96

Isom etrie isomorphism 1 222

Isom etry" 79

Join~ 46

Joint continuity, 118

Jordan;p C~ ~ 341-342

KadiBon, R~ V., 269

Kaku tani, S., 265n.~

K&m.k:e, E., 42n.

KelleyJ J. L., 139nr
Kelloggt

O~

D., 338
Kolmogorov~ AT N.~ 12BJ 215n.
Kronecker delta 1 283
Kuratowski closure axiomsjl 98

Sub ieel Index

Laguerre functions. 259
La. t tic et 46, 344
char.a.cteriza tion, 344--345
complemented,. 34.5
com pie te, 47
dis tr ibu tive, 345
sublattice or~ 47
Laurent expansionf 313
Least upper bound property, 21t

45
Lebesgue, H.;p 49
Lebesgue covering 1emma., 122
Lebesgue numbert 122
Legendre polynomials, 259
Limit, in the mean, 257
of sequence~ 50, 70--7lf 132
Limit point~ 65~ 96
contrasted with 1im it 1 72
Lindel~r~ s theorem, 100
Linear space1 81 1 191
baeis for 1 197
dim cnsion f 200
isomorph ism, 200
linear comb in at ion in, 194
linear dependence in 1 196-1 Q7
linear in depend en ee in, 196--197
linear operations in, 191
linear subspace(s)J 81 1 193
diejoint, 195
sum ofJ 195
direct~ 195
normed, 54~ 81, 212
quotient space, 193--194
repre sen ta.tion, 201-202
Lineat tre.nsforma.tion(e), 203
bound for~ 220
bounded, 220
continuous~ 219, 220
idempote ntt 206
identity, 205
inverse, 205
negative 1 204
non-singular, 205
norm, 22Q--.221
null spa.ee, 207
nullity t 207-208
product, 204
range~ 207
rank~ 208

aca.lar mul tiple;p 204

erot 204

36i

Lin early ordered eet, 44

Liou vill e1s theorem, 309-310
Li psc h j ti condition~ 339n.
Locally eom pact space, 120~ 162
Locally connected space~ 15 l
Loon1is, L. H.~ ix~ 215n. 1 30.5n.
Lorcht E. R. ~ 296nr
Lorcnt.zJ G. G., 157
Lo?ter bound~ 44
greatest, 44--45
LP space, 215

~1cCoy ~

N. H.~ 172
~IackP.y ~ (;. W ., 265nq 305n~
h1ac Lane~ S., 29
Mappjng(s)~ 16
bounded~ 58
closedJ 342
~omposition (multiplication) or, 19
continuoua.t 76, 93
jointly, 118
at point, 7 5-761 104
in single variable, 118
uniformly~ 77
contrasted with function~ 17
equality for, 20
Gelfand~ 319
graph~

23
identity, 20
into 1 17
inverse~

17
isometric t 79
one-to-one, 17
onto 7 17
open, 93
productJ 19
of sets, 1g.....19
(See also Function)
M.atriecs, ennonical form problem, 286
. __operations for 7 282-283
sirn il a.r, 286
Matrix~ conj uge. te transpose, 294
determinant, 287
diagonal, 287
identity, 283
as in dependent entity~ 28l 1 284
inverse~ 284
non-singular~ 284
of operator, 281
sca.1.a.r J 2 86

368

Subiecl Index

Ma tr i~ i triangular 1 295
zero:t 283
:tit1 ax i mal e1em en t 1 44
!\faxim..a.l ideal space~ 319
Maximum. 45
Maximum modulus theoremt 311
MeetJ 46
M etrie, 51
Metric spaee, 50, 51
comp]ete, 71
comp]etion~ 84--85
con tr action in 1 338
seq uen t.ial] y com pa.ct~ 121
SU hspace of, 56
tota.lly bounded, 123
M etri zabl c spa~c J H3
~[injmumt

45

r--.:tinkowskrs jnequaJity~ 88, 90, 218, 219

Module 1 l 9 l n.
r--.foments of o. function 1 157
r--.-r o rcr s.. ta theorem r 160J 303
T\.Iultiplicativc functional;P 321
,.f fin t.z 'B theorem , 157

n-dimensional Euclidean spa.cc, 87

n-dimcnsional unitary space, 90
N aim ark {or N cum ark )i MT A t ix
(See al8o (}e]fand~Neumark theorems)
N atura.l imbedding 1 232
Neighborhood, 96
N cumsrk (see N .aim ark}
Niven~ I "J 43
Norm(s), 54, 81 1 212
eq uival en t t 223
metric induced by, 54, 81, 212
uniform 1 216
N onnal opera tor 269
N orm.a.l space 7 133
Normed linen.r spacej 54, 81, 212
11on jugate space of, 224
isometric isomorph ism 1 222
locally com pa.ct~ 224
natural im bedding, 23 2
reflex iv c, 232
rep.rcsentation:t 234
second conjugate space or 1 231
strong topology t 232
weak topology, 232
weak~ topoJogy, 232-233
Now here dens-e. set, 7 4 1 99
j-

Numbers~

algebraic J 43
cardinal~ 31
com parab ili ty th eorcm, 48
finite, 32
eomplex~ 23, 52-54
real 1 21
transcendental, 43

One-t0-0nc eorrespondenee, 18
One-point eompactific.ation, 163
Open base,. 99
generated by op~n subbase, 101
for point (at point)~ 96
Open-dosed set, 349
Open cover 1 II I
basic, 112
suhLasict ] J 2
su bcovcr of~ 111
Open ma ppingJ 03
Open mapping theorem~ 211 1 23&
Open rectangles, 10 l~ 119
Open set, 60t 91~ 92
basieJ 09
subbasie, 101
Open sphere, 59
Open strips 1 101
Open suhbMe~ 101
Operator( s) t 222
adjoint, 263
characteristic eq nation, 288--289
conjugate 1 241
detennina.n t, 288
eigr:nspac-n J 27 8
eigenvalueJ 278
eigenvector~ 278
imaginary pa.rt, 271
matrix, 281
norm al, 269
projection(s)J on Banach space, 237
on Hilbert space i 27 4
orthogonal, 276
ree.1 part, 271
reduced by subspace, 275
ring, 303
self-adjoint, 266
ord ering 1 268
posi tiire~ 268
ap-ectra1 resolution 1 280, 291
uniqueness~ 291-293
~pectr al theorem, 280t 290, 295--297

Subject Index
Operato.r(s) ~ epeetrum, 289~ 296
sq ua.r e roott 294
subspace invariant under, 275
unifa1ry, 272
Order r.e latio n ~ partial, 7, 43
on real 1ine1 7
tota.1 (or linear) J 7, 44
()rigin 1 54, 80, l 9 l
Orthogonal eomp1e.ment1 249
Orthogonal dimension, 259-2 60
Orthogon a] vectors J 24 9
Ort hon onn a1 basis, 293
Ort hon onn al set, 25 l
comp1ete1 255

Parrt.11 elogr.am law J 24 7

P.arscval's equation, 250t 2.57
Partial orde!" re1ation_, 7t 43
Partially ordered set~ 43-44
Partition, 26
Partition sets, 26
Penno~ GT, 341-342
Pea.no space, 342n+
Perfect set1 99
Perm u ta.ti-0n ~ 176
Phillips, RT S., ix 232n~
Picard's theorem,, 339
Plane~ com pl ex 1 23 J 52-54
coordinate, 22
Euclideant 22~ 87-88
Point, boundary, 68J 97
fixed~ 338
at infinity 1 162, 163
interior, 63 ~ 97
isola. tcd, 96
limit, 65, 96
nceighborhood of~ 96
in a. spa.ce, 51 ~ 92
Point wise conve.rgence, 83
Poin twise opero.tions, 55J 82, I 06
Pioduct of BP.ts~ 23--25
Product space) 117
Product topology, 116
closed sub baset 117
open ba.se 1 117
open aubba.&e:, 116
Projection, 25
on Bane.ch space J 237
on Hilbert space, 274
on linear space~ 20~207
j

369

Proper va]ue, 278n.

Pro per vector~ 27 811.
Psti:u <lo-metric, 58
Pyt h agoresn theorem, 249

Quotient, algebra.,, 209

ring, 187
sp.a.c P 1 l 93--194

Radical, 314
Re.al line~ 21
absolute va.Jue on 1 52
cxten ded, 56
Jeast up per bound property, 21 _, 45
usual ru et.r ic on, .52
llec tangle 1 closed, 10 l ~ 11 ~
open, 101J 119
Reflexivity, of normed linear space,, 232
of re]ation, 27 ~ 43
R e1 a tion (binary), 26
antisymmetric~ 43
circu] s..r, 31
equivalen ee,, 27
partial order, 7J 43
reflexive 1 2 7 1 43
symmetric, 27
transitive, 21, 43
triangular~ 31
Re1 a tive topology, 93
Representation,, of algebrat 210
of B *-algebra1 325-326
or B a.ns..ch el ge bra, 305
of Banach space, 234
of Boolean aJgebra, 353
or Bool can ring t 3 51
or commutative C*-olgebra 1 332--334
of commutative semi-siro p1 e Banach
alge br & 1 322
or grou P~ 181
of Hll b ert space~ 260
of linear space, 201~202
of ringi 188--100
Reso1v en t of an el em ent. ~ 309
Resolvent eq ua ti on, 309
Re.solvent set, 309
Rickart~ C+ E., b.: 1 326n.
Riemann, B., 49
Riemann sphere, 163
Riesz~ F ., 49~ 226, 296n.

370

Subiect Index

Riea.z r epre.scn ta tion theorem J 226

Riesz-F isch er th eorcrn, 25 7-258
Ring, 181
corn m ute.tive, 183
cosct in J 186
div iBion, 184
di visor of zero,, 183

e]ements in, invertible, 183

non-singular, 183
regu] ar, 183
singular, 183
homomorphismJ 188
ideal in 1 184
with IdentityJ 183
-0f in teg~rs mod m~ 182
in verses, 183
isomorph ism j 188
kernel,. 188
of opera tors, 303
quotient 1 187
or setsJ 14, 182
.l:iUbring of, 184
Ilob binfl, H., 94J 338

Russe1l 1 R., 7
Russcl1 '.s paradox~ 6

Scalarsj 80 1 191, 214

Schauder~s fixed point theorem, 338
Schroeder-Bernstein theorem t 29
Sch,vn.rtz, J. T~, ix, 220, 232n.
Sehv..T.artz~s in~qua]jty :1

246

Second conjugate space, 231

Second eo un ta.bl e spac c, 99- l 00

Seit-adjoint opera tor, 2G6

8elf-adjoint su baJ g ebra. of <B{ H), 303
Scmi-simp]c a1gcbra 7 316
Sr.parable space, 96
Sequence, Cauchy :t 71
convergent, 50, 70 1 132
Jim it of, 50, 71 ~ 132
S equen tiaJ Jy com pact m etrie apace:1 121
Set{s}, 4

a bnorma.l, 6
Boolean a. lge bra j 12,, 344
boundary J 68, 97
boundary point, 68~ 97
bounded~ 58

Cantor, 67
(~art -r.:sia n pro<l u ct'" 21
cJosPd ~ 65, 95

Set (B );11 closure, 68, 96

complement., 10
contrasted with space, 5n.
convex t 148
countable, 34
countably in:fi n ite:1 34
d ensc (everywhere dense)~ 7 O,
derived, 96
diagrams, 8
diameter, 58
ditT eren ce 0, 13
disjoint 1 9
distnn ee from point to, 58
empty, 5
equality, .S
equivalen.r.e 1 27
finite:t 5

fini tu intersection, 12
finite union 1 12
of first category1 7 5n.
inclusion~ 6
index~ 11
infinite, 5
interior of, 63, D7
interior pointt 63i 97
intersection,, 9, 11
1in early ordered, 44
neighborhood or' 96
normal~ 6
nowberP densei 7 4, 99
n UII1 eric al equiv al en cc, 28, 32
open 1 60'" 91, 92
ba.sic 1 99
SU bha:sic' 10 I
open-('Josed 1 349
orthonormal, 2 51

eo1n pl ete 7 2 55
partially ordered, 43-4 4
partition 26
perfect, 99
product, 23--25
proper au bset, 6
pro per su perset 7 6
ring~ 14
of aeoon d category~ 75-n.
subsctJ 5
i-

;au per-set,, 5

symmetric di:ff er ence 1 13

toto.lly ordered, 44
un coun table:1 36
union1 S~ 11

Subiect Index
Set (s) ~ universal 1 5, 7-8
Set mappings, 18-19
Sierpinski, W . 1 42n+, 46n.
Similarity for ma trice.s~ 286
Smirnov~ Y. M., 139n.
Space, contrasted with set, 5n.
E:uclidean~

24, 87, 90, 214

fixed point1 337-338
of maximal idea.] st 319
m etrizabl e ~ 93
unitary, 24, 90, 214
(See al.so Banach space; Hilbert epa.ce;
Linear spa(! e ; ~f e tr ic a pa.c e ;
Normed linear space)
Space-filling curve( s }1 341
HilbcrtJs, 341-342
Spectral radius, 310
formulat 312
Spectral resolution, 280
Spectral theorem, 280 ~ 290
gen er &li.z ed f orms 7 29 5-- 297, 3!l 4
Spectrum_, of el em en t in Ba.naeh a1 ge bra'"
308

of op-erator, 289t 296

Sphere~ closed, 66
closed unit, 217, 232
open~ 59
Stonet 1\-L H.t 14ln., 153, 161n+
(Sec also B anach-S tone theorem)
St-one repr csc n ta. tion theorem J or Boolean
algebras, 353
for Boolean rings, 351
Stone-Cech c-ompactification~ 141J 331
Stone- \Veicrstra.ss theorem(s)J oomplexJ

161
extended~

371

Topologies.[ space(~), 91, 92

eompaot:11 110, Ill
countably, 114
locally, 120, 162
oom pact subspace, 111
completely regular, 133
corn pon entst 14 6
connected, 1421 143
connected subspace, 143
dbco nn cc ted ~ 143
disconnection, 143
discrcteii 93
discrete two-point 1 144
first co u.n ta bl ct l OOn..
Hau:!i dorff 1 130

homcomorphici 94
locally connected, 151
mctrizable;1: 93
normal 133
open base 1 99
open :!iubb!:u3c1 101
j

Pea.not 342n..
prodn~t, 117

second countable~ 99-100

separa bl er 9 6
subspaee:t 93
1\-~ 130
tota.1Iy disconnected, 149
Topology, 92
as branch of mathematics 1 04
discrete~ 93

generated by given c1 ass of sets, 102

open basei 99
open subbase, 101
product, 11 fi-. l 17

166--167

rerul 160
Strips~ 101
Strong topology, 232

re]ative, 93
strong 1 on normed linear space, 232
strongest~

l 04

usual, 92

Strongest topology) 104

Subbasct closed, 112
open, 101
Subcover 1 111

weak~ generated

Supremum,. 45

weakest~ l 04

by set of ma.ppingaJ 105

on norm cd linear space 1 232
weak*~ on conj uga. te space~ 23 2-233
w es.k operator, 3 03

Symmetryt 27, 51

Total matrix algebra, 284

Sz.-Na.gyt B., 226~ 296n.

Totally bounded metric space, 123

Totally dis connected spacer 149
Totally ordered set~ 44
Transitivity, 27 ~ 43
Triangle inequality, 51
Tychonoff s theorem, 119

Taylor,, A+ E., 215nq 239n.,. 248

Tietze extension theorem_, 136
Topological divisor of zero~ 307

372

Subiect Index

Unifor1n boundedness theorem, 211.

23g.......240
Uniform continuity, 77
Cniforro convergence~ 83
Uniforn1 norm, 216
{Jniformly bounded functions, 12Sn.
U niforrnly convex Ba.nach space, 248
Un it circh~~ 5
LT nit disc, closPd 1 .5
open, 5
LTnitary operator, 272
Unitary spac.P-j infinite-dimensionn.1 1 90
n..diinensional1 24, 90~ 214
Universal set 5, 7-8
Upper bound 1 lea.st, 45
Uryso hn 1s im bedding th eore rn 1 138
U r-ysolu1 i :::J 1enun a, 135
Usu.al topology on n1etric space"' 92

Vectors 1 pro per~ 27 Sn.

v on N" eume.nn .alge br&. J 303

Vector- s p a.ce (see Line sr space)

Ve.ct.errs, 81, 86--87, 101
eha ra.e teristic, 2 7Sn.
ei~un-, 278
linearly dcpcn<lent, 106-197
linnarly indcpcndcntJ 196-197
orthogona.IJ 249

Zaanen, A. C~, 215n.

Zero 1 54 1 80~ 179
divisor, 183
topological~ 307
ZP.ro space, 192
Zorn~ s 1e uuua, 45--46
Zygmundt A.J 240

W *-algebra~ 303
allrnan, H t l.50

7
"'

'V.-7 eak operator topology i 303

\Veak topo]ogy, generated by set or rnappiugsi 105

on nonnc<l lincar space, 232
Weak* topology on conjugate space 1
232-233
Weakest topology, 104
Weierstrass, K. 1 49
(S~e also Bolz.ano-1\rcicrstr-ass; StoneWeierstrass)
\Veierstrass .appr-oxima.tion theoremt l54t
161
Vleierstrass intermediate value theorem
142, 144
WildP.r~ R~ L., 6~ 39n.~ 46n., 343
~