Professional Documents
Culture Documents
Toaz - Info Simmons George Introduction To Topology and Modern Analysis 1st Ed 1963 PR
Toaz - Info Simmons George Introduction To Topology and Modern Analysis 1st Ed 1963 PR
GEORGE F. SIMMONS
Associate Professor of Mathematics
Colorado College
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 :sub-
sldiaries.
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
to
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 com-
plexes~ 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 funda-
mental 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
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 1
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 analy-
sis.. 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 excep-
tions, 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 func-
tion8 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 + I • .. ~ 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. Mathe-
matics 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 mathe-
n1a tica.l values.
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'\\r-
ing 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
Preface vu••
•
A Note to the Reader Xt
APPENDICES
ONE Fixed po.int theorems and ~omc n.pplica.tions to aneJy19is 337
TWO Continuous. curves and the Hahn-1\-f a..zurkicwicz theorem 341
THREE Boo lean a]ge bras, Boole an rings, and Stone's theorem. 344
Bibliography 355
Index of Sy·mbols 359
Subject Index 363
PART ONE
CHAPTER ONE
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 1
3} is an integer and 0 < n < 4} ,
= {n : n
{1, i, -1 1 -i} = {z: z is a complex number and z' = 1},
and { ± 1, + 3 ~ + 5 ~ . . . } :;::= { n ~ n is an odd integer l .
''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 pur-
pose 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 5
write
the unit ci.rcle = {z ~ 'z~ ~ I},
the closed unit di'sc = {z: 'zl S 1},
and the open unit disc ~ {z: ~z! < 1} 4
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 para-
graph. Thus A c B and A C Bare occasionally written in the equiva-
lent 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 7
u u
u u
Fig. 4+ A U (B n 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 dia-
grams 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 comple-
ments. The comple.ment of a set AJ
denoted by A' t is the set of all ele-
ments 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')' = A, 0' == U, U' = JJ,
AU A' = U, and A ()A' = S.
Sets and functions 11
can form
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
V ~r Ai = {x : x
e: A, for at least one i e: I}
and "141 Ai= {x:x e: Ai for every i el'
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 (remem-
e
bering that all sets are subsets of U) V.A., = and rlµla: U. The
;:::=
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 comple-
ments. 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 comple-
ments. 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. 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.
21 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 =A ---. (A fl B) = (A U B) - B;
(A - B) - C = A - (B U C);
A - {B - C) == (A ~ B) U (A n C);
(A U B) - C ~ (A - C) U (B - C);
A - (B V C) = (A - B) fl (A - C).
3. FUNCTlONS
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 compli-
cated 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
L
f f
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) .
;:::=
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 func-
tions 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
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 exam-
ple, 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 19
f g
at
Fip;. 8. :\! ultipJica.t)on of IIUtpping!'J.
in his mind if he will remember to read the product gf from right to left:
first apply f, then g.
Problems
1~ 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
g ~ ixg = (J- 1J)g = J- 1 (jg) == f~ 1
ix ~ J- 1}.
3. let X and Y be non~mpty 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 = iv~
4. L€t X he a non-empty set and f a mapping of X into Show it~elf.
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~?
S. 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•
6. 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
• 0•
1/2
• •l •
V2
•2 -·
e
••
3
'IT
)r
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 non-
empty 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
j
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
22 Topology
of real numbers, where x and y are <·alled the x coordinate and y eoordi-
nate 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
y axis in the possibility of exploiting this
identification by using algebraic
tools in geometric arguments and
giving geometric interpretations to
- - - - -..P.!!!!!:f ~1Y) algebraic calculations.
I The conventional attitude to-
I
I
I \Vard the coordinate plane in ana~
I lytir, geometry is that the geometry
I
1 is the focus of interest and the alge. .
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 geo-
metric 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 mis-
understanding. 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
.,,,/ (x1 ._ X2) 2 + (Y1 - Y~) 1 •
This notion of di.stance endows the set R 1 with a certain "spatiar' char-
acter~ which we shall recognize by calling the resulting space the Eucl.idea·n
plane instead of the coordinate plane.
Sets and Functions 23
the manner in which this set arises, through the introduction of a rec-
tanguJru- 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 under-
stood 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 P1·J x. is often abbrevi-
a 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.
Problems
1. 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?
2. 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.
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 ordet·ed 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,
t n
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.
Sets and functions 27
the same partition set. It is obvious that the relation,......; has the follow-
ing properties~
(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 x 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. Sup-
pose 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
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 y x y
g(Y) /(X)
FigT 12. The proof of the Schroed!'r -llt"r rust cin t hcorem.
Problems
1.. Let f; X ~ 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
r.J
of Proble~ 5f, this cannot be a valid proof.. What is the fl.aw in the
reasoning?
7.. 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
,.-..,_i z => y ,.._, z.
,.-..,_i
6. COUNTABLE SETS
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 manipu-
late 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 corre-
spondence 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
1, 2' 3 j n, • ~ .
• • I '
2, 4, 6, . . . , 2n, . . .
1, 2t 3~ 4~ 5, 6J 7, . . .
0, 2, - 2~ 4, -4, 6J ~ 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, 3J 4, 5, 6, 7' .. I •
1, 2~ 3, 4, 51 ~ . •
1', 2 2, 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 phenom-
enonj 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 infi-
nitely 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
1, 2' 3 j .. • ~ J 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 .... .........
.-
----
2 -2 -23 -2 -52 .... ---
.-... ....... __
l 2 4
3 3
1
3
2
-33 -43 .-......-
5
... _..-.._. ----
~
4
1
4
~
2
-34 4
-4 -4
5
------ ----
-51 -5
2
~
5
3 -45 5
5
........... - ... ......--. ....
1 I I r
I
i
r
s
I
'
I
I
I
•
I
I
I
I
I
I
I
f
I
I
Fig. 13.
I
( 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
Problems
1. 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.)
2. 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 Topology
3. 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~
.44' Prove that if X1 and X1 are countable, then X1 X X2 is also
countable~
5. Prove that if X 1, X 2, . • . ~ X are coun tablet where n is any positive
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 counta-
ble. 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 counta-
ble---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).
+ • • + .. ' ., + I •
Sets ond functions 37
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 numeri-
Fig. 14. A one-to-one corr es po n den ee ·
between an open interval and the reel cally equivalent to R.. The argu ...
lina. ment can be given in two sentences.
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
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 equiv-
alent to ll~ In the above three
paragraphs we have demonstrated x axis
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 concp1·n 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
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 ~ • ~
'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
2lh = c.
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 hemispheric·a.1 surface minus its bounda.ry)
1angentja1Jy on tl1c renter of x7
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
2.. Hholv that the subset X of R 3 defined by
Ret Pin \Vhich there is defined a partial order relation is called a partially_
Topology
t
0 1 0 1
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 inter-
ested 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
1 See" 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.
Sels and Functions J,7
Problems
1. 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+
2. 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?
3. Under what circumstances is Example 4 a chain?
4. Give an example of a partially ordered set whieh is not a lattice.
5.. 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 /\ (y A Z) = (x A y) I\ ZJ
X V (y V Z) = ( X V y) V Z, { X A y) V X = X, ( X V y) I\ X = X ~
6. 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.)
7.. 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
48 Topology
);fctrie Spaces
This condition means that. Xri must be '"close" to x for all '~sufficiently
large'"' nJ and jt is usually 8ymbo1ized by
or lim Xn = X
= {~
if x = Y~
d(x,y)
if x ¢ 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
Example 2~ 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;
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 ·w·hich
equals 0 ~ x - y ~ 0 {::::> x ~ Y+ l3y (ii),
d(x,y) == ~x - YI = r ~ (y - x) l = tY ~ xi == d(y,x),,
And by (iii)~
-z+z
z ~ z
R(2) ~ --·- and I (z) :; 2i .
2
== (!z1( + fZ2\) 2
•
O(x) =0
for all x; [O:Pl]. We define the norm of a function f by
1
llfll = / 0 lf(x) I dx,
and the induced metric by
1
d(f,g) = II! - vii = /
0
lf(x) - g(x) I dx.
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 non-
empty 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 1
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
- 00 and +CJ)
adjoined~ An extended real number is thus a real number or one of these
symbols. 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 A ::::i. - oo ~
Metric Spaces
Problems
1. 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).
2. 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.
3. 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
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
1
di, dit ~ . ~ , d'*~ Show that each of the functions d and d defined
Mefric Spaces 59
as follows is a metric on the product X1 x Xt x ... x xH~
d( { :c;:} J ~Yd) = max: di(Xi,y,:); d( {xd' {yi}) = ~7 l di(Xi1Y").
5. 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 ~ sup lf(x)l.
It is obvious that llfll is a non~ncgative real number such that
l /[[=08f=OJ and that
1
1 11-/!l=l~/ll- Prove in detail that
11! + oH < Ufll + [lgll.
6. 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 /) +
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 illus-
trates 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 frr-
A 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,
0 1
Fig+ 17+ An open aphere in the complex FigT 18~ An open sphere in e [O~IJ.
plane.
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 interval-
bounded 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 con-
taining it, never on its own.
Our next theorem justifies the adjective in the expression "open
sphere.J.,
The~rem B. 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).
The folio-wing characterization of open sets in terms of open spheres
is a useful tool.
Theorem C. Let X be a metric Bpace. A subaet G of X is open tj u is a
t.Lnion of open spheres~
PROOF. 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 non-
empty. 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 there-
fore open4
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
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 inter-
sections. 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~
Theorem E. 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,
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
'Z;
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. 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+
2. 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 gen-
era1ly~ show that A' is open if A. is any finite subset of X.
3. 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).
4. 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.
5. 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?)
6. If X = X1 X X2 X · · · X Xn is the product in Problem 9-4, and
a
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.
7. 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.
8.. Prove the statements made in the text about interiors.
9.. 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 ~
10.. 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
Theorem A. In any metric space X, the empty set 0 and the full space X
are clo Bed sets~
PROOF. 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 auto-
matically contains its ovrn limit points and thus is closed.
'The followi~g
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~
happen, for since F} is open each of its points is the center of an open
1
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~''
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 remov-
ing a countable disjoint class of open °.....___________ ..,.l
intervals. This process sounds inno-
cent 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
(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
73 + 7·9 + ;i7 + O L • = J +
the boundary of
( 1) A equals A (\ A 1
;
Problems
1. 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~
2. 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.
3. Show that a subset of a metric space is bounded ~ it is non-empty
and is contained in some closed sphere.
A. 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.
5.. Describe the interior of the Cantor set.
6.. Prove the stat.ements made in the text about closures.
7. Let X be a metric space and A a subset of X. Prove the following
facts:
(a) A1 ;:::= Int{A');
(b) A = { x: d(x,A) ;:::=o} .
8. 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.
9. 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 neces-
sarily true~ (Hint: see Example 9-1.)
70 Topology
10. Let X be a metric space, and let G be an open set in X. Prove that
G is disjoint from a set A ~ G is disjoint from A.
11. Prove the fact.s about boundaries stated in the text .
12. 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 }.
13. 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~
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.
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.
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.. 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).
2. Show that a Cauehy sequence is convergent {:::} it has a convergent
subsequence.
3. 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.
;:::=
1
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.
76 Topology
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.
Theorem B.. Let X and Y be metric .spaces and f a ·mapping of X into Y +
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
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 uni-
78 Topology
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
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
a:-+ 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,
and from this and the previous statement we see that d1 (g(x),g(x < E.
1
))
All that remains is to show that g is unique, and this is easily proved
by means of Problem 3 below.
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.. 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
2. 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?
3. 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):;:::=
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~
PROOF. 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
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 l-
wise addition and scalar multiplication and the nor·m defined by
Problems
1. 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.
2.. Construct a sequence of continuouR functions defined on [O, I J which
converges p oint,,1se but not uniformly to a coniinuous limit..
3. Construct a sequence of continuous functions defined on {O, 1] which
con verges poin twisc to a discontinuous limit.
4. 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.
5. 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 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.
1
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.
86 Toporogy
.z axls
y
l I y axis
'I I
I I"
' ':::::jI I
_____ ' I
Vx2 + y2 + z2.
If x = (x1, X2J a.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
llx~I = V1x1! 2 + lx~J 2 + ~ · · + lxnl 2
n
= ( ~ lxil 2 ) ~i.
i• l
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 r •• j = (Y·~
Y2t ~ • ~ y,.) be tu,,o n-tuples of real nr complex numbers.
+ Then
l!x!: l!Y:i 2
Sun1ming these inequalities as i variPs from L to n yie1ds
fl!
I: lxiy~!~ < 1 + 1
i=I I
[!xll llYll - 2 '
from t\rhich our conclusion foJlows at once+
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
1
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
T •
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 sec-
tion-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. 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.
2.. 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~
3.. 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 )~ <
~
2
~
7:opologieal Spnees
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 topologi-
cal space onto another which is also an open mapping. Two topological
9-' Topology
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
countable. Is this a topology on X?
3.. 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 95
closed sets in X is closed; and (2) any finite union of closed sets in X is closed .
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 topo-
logical 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~
Theorem B. 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.
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.
Problems
1. Let f: X Y be a mapping of one topological sp~ce into another.
---i-
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.
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 defini-
tions 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 applica-
tions of topology to modern analysis.
Problems
1.. 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 .
2. Under what circumstances is the metric space defined in Example
9-1 separable?
3. 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.
4.. 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)
5, 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.
6. 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.)
7. 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
8. Prove in detail that the open rectangles in the Euclidean plane form
an open base,
104 Topology
\\,.. 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 A. I.tet X be a non-empty set 1Jhen the fa niil y of all top oloy ie8
on X i.~ a co·m.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) .
Problems
1. 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.
2. 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
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~
lemma. 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. Further-
more, 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
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 con-
tinuous, 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
(f v g)- 1(A) = {x :max {f(x),g(x) I > al == {x :f(x) > a' U 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} ,
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) -
a1 fn~(xo)] + l/n~(xo) - f(XG)] I
5 !J(x) - f (x) I + If~(x) -
21 1 f~ 0 (xc.) I + IJn.a(xo) - f(xo);
< .:/3 + oe/3 + t/3 = £~
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. Let A be an algebra of real or complex functions defined on a non-
empty 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.
2.. 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.
3.. 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
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.
110
Compactness 111
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 cot•er. \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
L
PROOF L 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.
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 con-
siderably 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 eomple-
ments 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.
Theorem F. 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
u·ith 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
Compactness 113
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 sub-
space 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 con-
tains 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. 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.
2. 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 com-
p.act as a subspace of X.
3. 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
x
each of which is closed, and if n.1 1 is non-empty 1 show that n,.x,
is also a compact subspace of X .
4. 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 .
Compactness 115
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 arbi-
trary 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 motiva-
tion 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 defini-
tion they carry a typical element
(x1)x:i!) of X to xi and X2~ respec-
tively.. 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 non-
empty 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
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. AU proj ection·s~ 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.
2. 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.
3. 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'-·
4. 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 uncount-
ably infinite~)
5. 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
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
1
Theorem A (Tychonoff s Theorem). The product of any non-empty cla88
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 con-
clude 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 exten-
sion 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, ~ • + 1 Xn): a~ < Xi < bi for each ii
is called an open rectangle in RJI ~ A closed rectangle is defined similarlyJ as
a product of n closed in tervalsa
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
120 Topology
Problems
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.
Compactness 121
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 compact-
ness 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.
Theorem A. A metric space is Bequentially compact <==>it has tM Bolzano-
WeierBlrass properly.
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 .
V•rl 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.
PROOF+ Let X be a compact metric space and A an infinite subset of X+
' re aSBume that A has no limit point, and from this we deduce a con-
1A 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].
122 Topology
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 bounded-
ness1 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
+
compact.
12~ Topofogy
Prob terns
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
• • • 4 I .. .. • .. I I • • • .. I I •
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 applica-
tions 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~
PROOF. 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, I •• } 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 -
Compactness 127
{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
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 + • +
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
4 •• ,
Problems
1.. Let A be a subspace of a complete metric space, and show that A is
compact ~ .11 is totally bounded.
2. 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.
3.. Show that RfJC js not locally compact~
4.. By considering the sequence of functions in e[O,IJ defined by
f,.(x) = nx
for 0 s
x < 1/n, f"'(x) = l for 1/n < x =5 1, show that e{O~l] is not
locally compact.
1The 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
CHAPTER FlV'E
Separation
droff 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.
Separation 131
This shows that C' is a union of open sets and is therefore open itself4
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 ho·meomorphism.
PUOOF 4 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 VL·e complete the proof hy usiug the preceding
theorem to conclude that.. f(F) is a closed subspace of Y.
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
Topological spaces
Hausdorff spaces
Normal spaces
Metric
Compact spaces
Hausdorff
spaces
F ].(:!;. ')5
..... • The se pa.ration properties.
Problems
1. Show that a closed subspace of a normal space is normal~
2. 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.
3. 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.
4. 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
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.
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 exten-
sion of f to the full space X Vlbich has the defiired property.
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. 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.
2.. State and prove a generalization of rriet.ze"s theorem "\\Thich relates to
functions whose values lie in Rn.
3. Justify the assertion in the last paragraph of the text that the function
defined there cannot be extended continuously to X.
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~.
Theorem A (the Urysohn lmbedding Theorem). If Xis a second count-
able 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
4 • 4
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 R«i ~ 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
- 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.. 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~
2. 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.
3.. Show that a compact Hausdorff space is metrizable <=> it is second
countable~ 1
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
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 pro-
jection 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 require-
ment 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 situa-
tion. 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 restric-
tions 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 stute-
ment (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 func-
tion 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 compacti-
jication 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. If X is completely regular, show that every bounded continuous com-
plex function defined on X has a unique extension to a bounded
continuous complex function defined on t3(X).
24> 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.
1
3. 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
31 .. CONNECTED SPACES
A connected space is a topological space X whlch cannot be repre-
scn 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 defini-
tion 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 inter-
sections '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.
for every point x1 1 in Xi 1• 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 con-
stant n1apping. This contradicts the assumption that f maps X onto
{0~1 J, and completes the proof.
Problems
1 + Show that a topological space is connected~ every non-empty proper
subset has a non..-empty boundary.
2. 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
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~
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:_ 1 An is connected4
2. Show that the union of any non-empty class of connected subspaces
of a topological space each pair of 1\...hich intersects is connected.
3. 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.
4. Use Theorem A to prove that an arbitrary Banach space B is con-
nected. (Hint: if xis a vector, sho'v that the set of all sca1ar multi-
ples of x is a connected subspace of B.)
5.. 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 = x + 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 con-
nected. Prove also that every sphere (open or closed) in B is
con vex, and is therefore connected.
6.. Show that an open subspace of the complex plane is connected ~
every two points in it can be joined by a polygonal line:
7~ 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.
BL 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
Y = (Y l\ A) U (Y n B)
is a disconnection of Y.
Problems
1. Prove that the product of any non..empty class of totally discon-
nected spaces is totally disconnected.
2. Prove that a totally disconnected compact Hausdorff space is
homeomorphic to a closed subspace of a product of discrete two-point
spaces~
1-
A•• I
·I
I
/ I
I \ I
I .x I
\ ,_ ,,,,
L
I
I
I
I
I
0 1/r 2/-;r l
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 .
Theorem A. 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
The principal applications of local connectedness lie in the theory
of continuous curves (see Appendix 2) .
Problems
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 gen-
eralized 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 .
t (n)
I: mi:O k
xi(l - x)K-~ = [x + (1 -- x)]R = L (1)
Approximation 155
If we differentiate ( l) with respect to x, we get
±(n) zk-
i•O k
1 (1 ~ x)'f;~~ 1 (k - nx)z = n;
±(n)
i z:::Q k
xl'(l - x)"-k(k - nx) 2 = nx(l - x),
±(n)
l:ccO k
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) - f (!)}
so that
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 arbi-
trary) 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
\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 -< n '
so """''' < x(l ~ x )__
~ - ~tn
Z' < f/2, IJ(x) ~ Bn(x)! < E for all x in [0,1], and our theorem is fully
proved.
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
1<t Prove that e[a~bJ is separable.
2. l.J€:t f be a continuous real function defined on [O, 1].The moments
of f are the numbers Jo 1
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.
3. 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.
Our purpose in this section is to lay bare the true nature of the
Weierstrass approximation theorem. W-e achieve this end by general-
i.~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 prop-
erty. 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 sub-
algebra 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 non-
zero 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) = max {/{x)~g(x)}
and (f A g)(x) ~ min {/(x) 1 g(x) 1-
Our first lemma states conditions which guarantee that a closed sub-
lattice 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 4..
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
If X has only one point J then a. single eonst&n t function constitutes a closed
t
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
J'
X, for it will follow from this that Ill - gll < E+ 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 A ' ~ " 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 sub-
cover, which we denote by {111,. H 2 ~ . • ~ ~ H m}.. We denote the corre-
sponding 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 =f + g + If - ol
2
and f "0 = f + 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).
lemma. Let X be an arbit·rary topological ~~pace. 'Phen every closed sub-
algebra 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
160 Topology
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)
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 Toporogy
2. 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 poly-
nomials in z and z with complex coefficients.
3. 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).
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 r·a~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 x®. 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, .. L G1J. n X}
, + It is no'v easy to see that the class
{Gi~, Gh G2~ . ~L Gfl J is a finite subcover of the original open cover of
,
x
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
16' Topology
Problems
I(f + g)(x)! ;;::: l/(x) + g(x) I < ~f(x) I + !g(x) I < E/2 + ~;2 = t:1
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 ..
Problems
1. If X is a locally compact Hausdorff space_. prove that Go{X,R) is a
subla ttice of e (X ~R).
2~ Let X be a locally compact Hausdorff space, and show that the weak
topology generated by '2u(X,ll) equals the given topology.
3. 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 gen-
erated 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 com-
plex functions which are defined and continuous on a topological space X +
39. GROUPS
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 opera-
tion being coordinatcvlise addition. The identity here is
o= (07 o~ . . . , O) ~
and the inverse of the element. x ~ (x1, x2 1 • ~ • , Xn) is
-x = ( ~x1:ii -X2 7 ••• ~ ~xn) .
(b) The set Cm of all n-tup]es oi complex numbers v..~ith respect t-0
coordinate,visc addition.
Example 4. {a) The set e (X tR) of aH bounded continuous real f unc-
tions 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.
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). B = (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.
1234)
p = ( 3142 '
lvhere be]ow each of the integers 11 2, 3, 4 is placed its image under the
mapping P~ If
q
= (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
j
1234)
pq = ( 2341 .
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
vertical axis, then rotating it counter- ~
1
I /;
clockwise through 90 degrees. If \\""C D ',.....,2~""""""'"""' ~~~
1
. 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·
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
H = G~~i).
The group of symmetries of a squareJ being a subgroup of the symmetric
group s_. 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,..hen-
ever 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:
1234)
qp ::;: ( 3421 .
Problems
1. Let G be a group, and show that (xy)- 1 = y- 1x- 1 for any t\vo e]e-
ments x and y in G. Show also that (x- 1) - 1 = x for any e]ement
x in G.
180 Operate rs
40. RlNGS
We have seen that the set I of alJ integers is an additive Abelian
group with respect to the opcratio n of 01·d 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 ·w·hich 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 ele-
ments 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.
xO +0 = 0,,
'?.thich yields xO = 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 or 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 ele-
ments 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 sim-
ple 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 iden-
tity) 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 par-
ticular 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.
184 Operators
A ring with identity is called a division ring if all its non·zero ele-
ments arc regular~ A field is a commutative division ring. The rational
numbers constitute a field, as do the real numbers and the comp]ex
numbers. Roughly speaking, fields are the ''number aystemsJ' of
mathematics.
Problems
l~ Jn the ring of even integers, why is 2 neither regular nor singu] ar?
2. 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?
3. 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)?
4. I...et R be a ring with iderttity, and shov{ that any divisor of zero in
R is singuln.r~
5.. 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~
6. 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.)
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
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 + Y1) - (x2 + y,.) ;;::: (x1 - x~:) + (y1 - y'J.)
and X1Y1 - X~Y! = X1Y1 - + X1Y2 -
XfY':l. X1Y2
= z1(Y1 ~ y1} + {x1 - x'A)Y2~
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] = [x y] +
and ix] ' [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
11
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
rx + y .J
1 + yJ
= rx
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
1
and the zero in R". This convention is customary, and it aavea mor.c trouble than it
ca.uses~
Algebraic Systems 189
g(x + y) = (x + y) + K + K) + (y + K)
;;::: (x
== g(x) + g(y)
and g(xy) = xy +K = (x + K)(y + K)
~ g(x)g(y).
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.
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. 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 con-
jecture can you prove?
2.. 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.
3~ 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
··. ··. -.~ 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
Ex.amp( e 1 The set R of all real num hers~ "'. i th ordinary addition and
t
Example ..t. The ~ct (} of all complex numbers is a complex linear space
under ordinary addition and 1nultiplica.tion.
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 men-
tion 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
;:::=
Mi = i (.x 1,
0, 0)}, M "i == i (0, x2~ 0)} 1 Ma = { (0, 0, xi) l ~
and M ~ ~ ~ (0, X2, xa)}, M 6 = { (x1~ 0, Xa) L, M 6 == {(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 con-
siderahle 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
(x + M) + (y + M) = (x + y) +M
and a(x + M) = ax + 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)
;:;;:r; + M.
•
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: ( x+y) +M
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 tV¥·o coscts x +Mandy + M by adding
r
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 ~. r ©Mn
~means 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
Problems
1.. 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?
2., 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]?
3. 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.)
4. Let a linear space l.i be the sum of certain subspaces 1'J 11 M 2, • . a ,
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~ + I • •
(3)
then subtracting (3) from (2) yields
(a1 - tj1)X1 + (a~ - /3~)x2 + I ~ • + (Un - t'~) Xn ~ o~
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
0 = Q a XI +0 1
X2 +• 4 1
+0~ Xn
i >
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
Theorem D.. 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 com-
0
bination 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
B 'J = f I = I f h f 'l,
{ 1 - . • , f $ }
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.
r •
'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 0
• + 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.
6t = (1,. o, o, . . , 0) J
es ~ (O~ 1, O, . . . , 0),
=
ei (O, O, 1, . . 0),
4 ,
•t•···········
e. = (0, 0, 0, 4 ~ • , 1).
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
form
Problems
l. 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 inde-
pendent .q it spans L.
2. Show that the vectors (1, O~ O), {1 1 ~ 0), (I, 1, 1) form a basis for n~ 3 •
i
ShoVir that if ~ c1, e2, ea l h:; a basis for Ii 3, then {ei + c2.' e1 + eJ, ez + el l
is also a b.asig ~
3. 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. 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.
5. Denote the dimension of a linear space L by d(l-1)~ If L is finite-
dimensiunal, 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)~
6.. If L and L' are linear spaces, show that L is isomorphic to L' ~ they
have the same scalars and the same dimension~
Algebraic Systems 203
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
1
I (f) = / 1
f(x) dx
linear space u;ilh respect to the linear operations defined by b'q8L (I) and (2).
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 con-
sider 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 impor-
tant feature of this situation is that if T and U are any t.wo linear trans-
formations on Lt then we r.an dPfine their prnduct TCI by means of
(TU) (x) = T( Cl(x)). (3)
= (TV)(x} + (UV)(x}
~ (TV + UV)(x).
E((l - E}(z)). = (E(l - E)) (z) (l!J - _bt2) (z) (I~' - I!}) (z)
= O(z) = O
(I - E) 1 = (I - E)(I - E) =I - E - E +E 2
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. Show that the mappings defined by Eqs. (1) to (3) are linear
transformations+
2. Sho\v that the linear transformations T2 and T.a defined in Exam.ple 1
do not commute; that is, shovl that T2T3 ~ TaT2.
3. 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.
A. 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.
5. 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~
6. Prove that a linear transformation on a finite-dimensional linear
space is non~singular (::::}it is one~to--0ne ¢::::>it is onto4
7~ 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.
8. 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'.
9. 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.
10. 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)~
1
)
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)~
Groups
R~ngs
Linear
Algebras spaces
-·
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. Further-
more1 since an algebra is also a ring, it may be commutative or non-com-
mutative, 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 multi-
plication, and multipli<.:ation .
Algebraic Systems 209
Example 1 to (a) The real linear space R of all re.al numbers (see Exam-
ple 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 multi-
plication is defined as usual. Again we see that scalar roul ti plication
and ring multiplication are the same.
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 us-
sion following this theorem can be summed up as follo\vs. If multiplica-
tion 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 Operators
Ha11aclt Spaces
x~ ~ x and Yn -4 y ~ x" + Ya ~ x + y
and a,. ~ a and Xn --t 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
~ .• , 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
(4)
norm defined by
(5)
Banach Spaces 215
of scalars such that 2: ; ... 1 !xn IP < (TJ , "\\,.ith the norm defined by
'IO
llxllP ~ ( t
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)] P 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 summa-
tion, 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
1 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 understand-
ing 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
j
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
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:!~·
Example Consider the linear space of all bounded sequences x ;:::=- ~ x ·~
6..
x"2, ~ ... , x~, . . . J of scalars. By analogy with Example 5, -.,ye define
the norm by
~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 ·µ,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 lf(x)!L 1 (11)
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.} r + • ,
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
Problems
1 t 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.
2t 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.
3.. Prove Eq . (9) for the case of an arbitrary positive integer n.
4. 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..-!~ • r Xn) and y =:::::: (y1, Y<J,
• j I Yn) of scalars. \1lhen-
I + '
If z + yl[J'P = i-1
~ ixi + Yt:~p ~ ~
i• 1
lXi + Y•l /x, + y.;JP- 1
n K
< I
i-l
Jx1I ~x, + Y1IP- 1 + I
i=l
IYir ~x.: + y,lp-- 1
< ([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
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 equiva-
lent to one another:
(1) Tis continuous;
(2) T 1~s continuous at the origin, in the sense that Xn ~ 0 ~ T(xn.) ___. 0;
1 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.
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 sph£rt! 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 cer-
tainly 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
1
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
1
When N #- iO], this formula can clearly be written in the eqwvalent form
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 trans-
formations 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 linea·r 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 con-
tinuity 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
222 Operators
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] < 1} = sup { l T(7''(x)) [I : llxli < 1 i
< sup {II TH II T'(x) l ~ llxH < 1} = ll T(I sup [[IT' (x) ll: llx~I < 1}
== 11 TH 11 T' IJ.
We know from the previous section that addition and scalar multipli-
cation 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 1lI] r: - T' 11
+ HTfi - Tll [l T'l1 4
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 223
3. 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~)
4. 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)
5. 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.)
6~ If N is an .arbitrary normed linear space~ show that any linear trans-
formation '1' of z; (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
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 %) .
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)
\Ve now observe that for any two vectors x 1 and x!I in .llf we have
and our conclusion now follows from lle-•'xl] ;:;: ; llx[I = 1 and the fact
that fG(e-i 6x) is real~
Theorem A (rhe Hahn-Banach Theorem)t 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
11/oll = llfl].
PROOF. 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 a
complete the proof by observing that the domain of fo must be the entire
space N 1 for othe1·vlisc it cguld be extended further by our lemma and
would not be maximal.
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.
f (XG + jf) ~ Q.
Ii we now defin~ f\J by fo(x) = f(T(x))~ then f-0 is easily seen to have the
desired properties~
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
X = (x l~ X2, . r .. ~ Xn)
of scalars. If feh e~, . . . , en} is the natural basis described in
Problem 47~6~ then x = x 1e + Xi€t + . · · + x~en; and if f is
230 Operators
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 is1:.
defined by Hxll = (Zi-1 lriJP) llPi and it fo1lov{S from
n n n n
lf(x)I = I,~ x,y,' :$ ;~ 1x,y,r < (;~ jx;f•')-1'.!' (;~ :y,:•) i1~
that I! f 11 <
(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
oth erVtrise.
(b) (l~) * = l~ ~ 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. 1
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 .
part of N** "\\rithout altering any of its structure as a normed linear space.
7
"'\\ c write
N CN**
- ~
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
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)
\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.
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.selfL Theorem R is thcrefo1·e some'"That less revealing than
appears at first glance~ We shall see in Chaps. 13 and 14 that the
Banach Spaces 23.S
Problems
1. Let X be a compact IIausdorff space 1 and justify the assertion that
e(X) is reflexive if X is finite.
2. 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.
3. If B is a Banach space} prove that R is reflexive ~ B* is reflexive.
4. Prove that if B is a reflexive Banach spacei t.hen its closed unit
sphere Sis Vtreakly compact.
5. Show that a linear subspace of a normed linear space is closed <=>it is
\Veakly closed+
Since yis 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; s:12
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
then it follows from Jlxnil < l/2"- 1 that {sn} is a Cauchy sequence in B for
which
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
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
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 +
Problems
1. Let a Banach space B be made into a Banach space 8 by means of & 1
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
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 con-
1 See Taylor [41~ pp. 181-1851~
240 Operators
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.
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.
1
'
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
[T*(f)J(x) = f(T(x)), (2)
(TJT~,) *(f) = (Tz•T1 *) (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. A simple computation nuv..- shO\YS that
r(1\T2)*(j)](x) = f((T1Ti)(x)) = f(1,1(1't2(x)}) = [Tl*(j)]('.Fi(x))
= [ 1' 2 * ( 7\ • ( f) ) ] (x) = [('.t 2 * T 1 *) (f) ] (x) .
Problems
1. Let B be a Ranach spat,e and N a normed linear space. If {T l is a 1)
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.
2~ 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
j
1
+
(7,*)-• = (T-t)* .
CHAPTER TEN
Jlil/Jert Spaces
treatments of v ectora,
Hi rbert Sp aces 245
Example 1. The spaoe z;, with the inner product of two vectors
x = (z1, X2i • • • , Xn) and Y =::: (y1 1 Y2, . · · , y,.)
defined by
defined by
om)
(x,y) = I.
n• 1
XnY~·
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~
This is readily proved by writing out the expression on the left in terms
of inner products:
2 = 2 + 2 --.. 2
+ Hx2 H -d,
1
HxH 2 2
- 2
< 2
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.
Theorem If B is a complex Banach space whose norm obeys the paral-
C.
lelogram 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). Prop-
erty (l) is best trr.atcd by splitting it into two parts:
(x +y 1 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.. Show that the series \vhich defines the inner product in Example 2
is convergent.
2. 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.
3. l1~or the special Hilbert space li, use Cauchy's inequality to prove
Sch '\va rz 's inequality.
4. Sho\V that the parallelogram law is not true in z~ (n > 1)4
5. 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].
6.. Give n detailed proof of Theorem C.
Hilbert Spaces 249
If we put o: = /3(zo.,y) for an arbitrary real number {3, then (1) becomes
-2Pl (zo,y)I i + f3 i(zo,Y)! 2llYll
1
i > 0.
If we now put a= l(zo,y)~ 2 and b = llYil 2, we obtain
- 2/3a + p~ab > 0 1
Problems
1. 1:£ Sis a non-empty subset of a Hilbert space, show that S 1- = S 1.. 1- 1-.
2. If M is a linear subspace of a Hilbert space, show that M is closed
{::}M = M..L.L~
3. 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}.
A. 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
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 finite orthonormal set in a Hilbert
space II. If x is any vector in H~ then
n
I
1-1
I (x,e,)l 2 < If xiii; (1)
'l
/urtherJ x - I
i= l
(x,ei)e.: 1- Ci (2)
for ear.h j.
PRoott~. The inequality ( l) follows from a computation similar to that
used in proving Schu~arz~s inequality~
n.
0 < r1 x - ~ (x~ei)ei 11
2
l=l
n n
- (r - I
l=l
(x~e,)ei~ x ~ ~ (x,ei)ei)
j=l
n ~ n n
~ (x,x) - I
1"=1
(x,eJ(x,e1) ~ ~ (x,ej) (xl~i)
i=-l
+ I.:-1 ~ (x,ei) (x,e;) (ei~ei)
j::::ol
1l
- llxlf 2 - I
i=J
I (x~e.:)l 2T
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 com-
ponents in several perpendicu1ar directions, then the result has no com-
ponent 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. The con ...
cl usion nov{ f ollo\,TS from the fart that 1~ = U :_ 1 j.S n·
As our first appHration of this result, we prove the genera.] form of
Bessel's jncquality~
Theorem C (Besser' s lnequality)t If [ed is an orthonarmal set in a
Hilbert space H, then
(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
s
easy.. As in the preceding theorem' we '"'~rite == {et: (x~et:) "¢ 0 I If r s
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
n.-1
I (xten)l 2 ~ tlxl! 2 ; (4)
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.
Theorem Dt If fed is an orthonormal set in a /filbert space H, and if x
is an arbitrary vectOT in H ~ then
(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
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
1 r
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:
s = i/1,/2, ~ • • ,f~, a " .}.
yields lls~o - 8n~ll 2 < };;:no+l '(x,e~)[ 2 < ~ 2 , so l!s:o ~ Sn() H< ~ and
!Is' ~ sl1 < Hs' - s:c-li + lls:o - 8no!I + l!sno - sll < t: + tr + E = 3~~
Sinee i is arbitrary 1 this sho~,..s thats' == s.
Let H be a non-zero Hilbert space, so that the class of all its ortho-
normal 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.
PnooF. 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.
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
(1) (2).. If {2) is not true, there exists a vector x ~ 0 such that
::=::}
Jo
27 lf(x)j 2 dx < co ,
1
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
I""'
t.1:=-o11;1
2
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:
~
I
n ""' - m
lcnl 2 = /
0
2': lf(xW dx.
(7)
!.,..(X) -- ~ I~
1 ~
.£., Cl;:C
ik~ J (8)
v 21f k• -~
llf1i - 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 li·ou 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 i;; Icn. I2 converges, then t.he re ex ists a function
-CCI
258 Operators
fin L 2 whose Fourier r.oefficients are the c~'s. If we grant the complete-
ness 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
m
11/m - J~l! 2 = t
ikl=n+l
[cA:! ~ 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 com-
plete, 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 ~ r • • ,
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; - (.xa,e1)e1 - (xhe2)e2
ea ~ Hx 3 - (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
Problems
1. Let (ei, e~, . . I '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
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::
1 r (x,eJ -- ail 2 in
the rcsultr)
2~ Show that the orthonormal sets described in Examples 1 and 2 are
complete.
3.. Show that every orthonormal set in a Hilbert space is contained in
some complete orthonormal set,. and use this fact to give an alter-
native proof of Theorem 5:i-B.
4. Prove that a llilbert space// is separable$===) every orthonormal set in
H is countable+
5. Show that an orthonormal set in a Hilbert space L" linearly inde-
pendent~ and use this to prove that a Hilb~rt space is finite-dimen-
sional <=> every complete orthonormal set it':l a basis+
6. Prove that any two complete orthonormal sets in a Hilbert space H
have the same cardinal number. This cardinal number is callr.d the
260 Operotors
f~(x1 + xz) = +
(x1 x2, y)
= (xitY) + (x21Y)
+
;:::= f JJ(x1) f 11(X2)
and f 11 (ax) = (ax,y)
= cr.(x,y)
::== af11(X) •
Hilbert Spaces 261
> 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.
Theorem A. Let H be a Hilbert space, and let f be an arbitrary functional
in H •. Then the-re exists a unique. ve.ctar y in H such thal
f(x) = (x,y) (1)
for every x in H4
PROOF 4 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. =::::;
and (3)
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>~
3. 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).
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 con-
j uga. tes. All our future work with adjoints will be based on Eq.. (2)t
and from this point on, the symbol r•
-..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
(x, T*(y + z)) ~ (Tx, y + z) ~ (Tx~y) + (Tx,z)
= (x,T*y) + (x,T*z) - (x 1 T•y + T*z)i
80
T•(y + z) =
T*y T•z4+
The relation T*(ay) == aT*y
1 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 265
Theorem A. The adjoint operation '1'--=; 7'* on ffi(H) ha8 the following
pro periies:
(1) {T1 +
T2)* = T1* + T2*;
(2) (a.T) * ~ aT*;
(TlT~)* ~ 1'i*1 1*; 1
(3)
(4) T*t. == T;
(5) ll T* II = [l T [I ;
(6) "T*7'11 = l Tll !_
PROOF+ 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
Problems
1. Prove parts (1), {2), and (4) of 'rb-eorem A.
2t Show that the adjoint operation is one-to-one onto as a mapping of
ffi ( H) in to itself.
The reader will notice that we have said nothing here about the
product of two self-adjoint operators. Vecy little is known about such
Hilbert Spaces 267
. 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
T = 0 ¢=} (Tx,y) = 0
for all x and Y~ It is also clear that T == 0 ==> (Tx,x) = 0 for all x. We
shall need the converse of this implication.
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:
+ JJY)~ ax + /3y)
~
(T(ax ~ lal 2
(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
Dividing (3) by i and adding the result t-0 (2) yields 2{ Tx, y) = O~ so
(Tx,y) == 0 and the prooi is complete+
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 - T*]x" x) =0
for all x.. By Theorem C~ 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 self-
adjointJ 've "\Vritc A 1 < A2 if (A .xjx) < (A ~x,x) for all x. The main
elementary facts about this relation are summarized in
self~adjmnt
1
Theorem E. _ 1 he real Banach sprue of all 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.
Theorem F.. 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
(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 com-
pleteness of H, M is complete e.nd therefore closed. We conclude the
HiJbert Spaces 269
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) ~
snd Ao I == Io A = A4 Show also that A 1 c (A 2 o A 3 ) = (A 1 o A-:i:) o A3
whenever A 1 and A 3 commute~
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
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:
for every x 1 and this impiies that ll..L~,l 2 ll = llN*NI~- By Theorem 56-A, we
have [IN*NH = [! ·1'l !I2, so the proof is com plcte.
W c kno\v that any complex number z ran be expressed uniquely in
the form z ~ a + ib '"~here o and bare real numbe1·s, 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 sclf-
adjoin 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 .¢;;> its real and
imaginary parts cmnmute.
PROOF. If Ai and A 2 arc the real and imaginary parts of T 1 so that
'l.' = Ai +
iA2 and T* = A 1 - iA~ 1 then
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
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.1·ue. The
fact that (3) implies (1) is a conseqne.nce of ~rheorem 57-C and the follow-
ing cha.in of implications:
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, • • I J = { o, xi, X2J • • I i~
which preserves norms but has no inverse~ These ideas lead at once to
Problems
1. 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~
2. If H is finite-dimensional~ show that every isometric isomorphism of
H int-0 itself is unitary.
3~ Show that an operator Ton His unitary~ T({ei}) is a complete
orthonormal set Vr~hcnever [e, 1 i'3.
4.. 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 projec-
tion 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~
M, so M is invariant under T.
Theorem E. If Pis the projection on a cl.osed linear 1tttb81Jace M of H, then
M reduces an operator 7 <::=> TP = PT.
1
PROOF. ~/
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.
Our next theorem shows how projections can be used to express the
statement that tl\;rO closed linear subspaces of Hare orthogonal.
M == M 1 + Jf + ! I • • + MB•
PROOF. Since P is clearly self-adjoint, it is a projection ~ it is idem-
potent. 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
TC n
I
j=l
llP1xllt = llPall 2
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 + P2z + · · · + P,,.x
is evidently in M 4
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
Problems
1. 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.
2. 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) p < Q;
(b) 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) ~ 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+)
3. 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 ~)
4. 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
Tx == AX. (1)
A non-zero vector x such that Eq. (1) is true for some scalar A is called
an eigeni·ector 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 eigen-
vectors 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
X = X1 + X2 + ~ ~ . + X~, (2)
\\rhere is in M..: for each i and the x/s are pairwise orthogonaL
xi It now
follows from (a) that
Tx = Tx 1 + Tx2 + · · · + Txm
= A1X1 + A2X~ + · · • + AmXm• (3)
Also_, since for each i and for every i ¢ i we have M 1 ~ M,1., Eq. (2}
yields
Ix = x = X1 + x~ + · · · + Xm
= P1X + P'JX + +Pm::c
I ••
= (P1 + P2 + · · · + Pm)x
for every xin H, so
I = P1 + P2 + · · · +Pm. (5)
11
[ X1_, X2~ • . . l = { o, X1, X2, I I • 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 l
has the spectral resolution (6). It is too much to hope that every opera-
tor 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
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
theore·m, and the machinery treated in the follo,ving sections is directed
exclusively toward this end+ We emphasize once again that His under-
stood 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-dimensiona·1 Spectra I Theory 281
. ~ .
[T] = • • • • r •
(2)
The array of scalarA (ai1 1 a;21 ~ ain) is the ifh rotD of the matrix (T],
L • •
and (a1h a2;·., •• Cin:J) is its jth column+ As this terminology shows,
4' ,
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
(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 coeffi-
cients 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 1
= I: f3i cI Uijei)
j=] ~=I
n n
= I <I: O:ijfjj)et: .
.; ... 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 = [T1UT2J+
Let [~;] and [p#] be t.he matrices of Ti and 2,2. --rhe computation
(T1 + T2)e; = T1e1+ T2ei
1t ~
= I: ai;e i + I: f31J-ei
•-1 i=-=l
~
= I
..: :o::i:l
(ai; + /3i;)ei.
Finite-dimensional Spectral Theory 283
n n
= I (I: airre1)
t'.ti
k=I i=-=1
n n
~ I (I aa~k;) ei
i=l k=l
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 multiply-
ing 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 ~ • + 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
0 if i ~ :i
aij = { 1 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,
t
'Yithout any reference to its origin as a representing system for the opera-
tor~ 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.
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 compli-
cated~ 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 considera-
tions. Suppose that [ai 1] is the matrix of an operator T relative to B.
Finite-dimensional Spectral Theory 285
which is equivalent to
[~iHT- 1 ] = [T- 1][aii} = [ai;].
We therefore have
Theorem B. 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-sing·u.la.r, and in
this case [at;}- 1 = [T- 1]~
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 n
Th = I
l::cw:l
f31c1fi; == I. ~qAl\t
.1::-1
a n.
= I <I. 'Yilt\)
i:-1
fl1t:i
i=l
n n
= ~ ( I: 1'ikf1~;) ei;
i=l it>;ICLl
n
and Th === TAei =T (I. 'Y•;et) k=l
n
~ :!: Y1c~·Te1i;
k :cot}
;r:i n.
~ I,
1:=1
{I: ai:te1) °Yki
i::::!:El
n n
I (I aik'Yk;) ei.
,:-1 k:::ml
286 Operators
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).
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 direc-
tion 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
c?s 8 - cos
[ sin fJ
sin 8]
()
and [eift
0
O ].
e-t8
«:i!:'2
4 ~
• a2n
.. . • ~ 4
• • . . • ~
'
a~2
. • . ann
. . ~
.. .. 4 • .. .. I .. + I 'r • ,. ., • • ., • .. •
afl i Onn - A
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
Theorem A. If 'I' is an arb·itrary 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
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
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 va1·ia.ble
z. Ry taking linear combinations_, (2) ca.n c'idently be extended to
m.
p(T) = k
t .1::11" 1
p(Xi)P1. (3)
\\re 'w·ould 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:
(z -:- X1) · + (z - Ai-t) (z - AJ+t) · · · (z· - Am)
•
p·z
( ) = .
) (Aj ~ h.1) ' · • ('J...j - AJ-1) (AJ - A1+1) · ~ ' (A§ - ~m)
Since Pi is a polynom1al, and since p1 (Xi) =:.: Otn (3) yield~
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 = 2::
1 Qj. ";-e wish to show that (5) is actually
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
k k
Tx = Xx = A.Ix = A !r Qa I
i•l
===
~=l
XQa
k
and Tx = !r a1Qat
i•l
•
i (X -
•
so lti)Qix = 0.
i=-1
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 state-
ments 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 orthog-
1
onal 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 under-
stood to be 0):
•
•
•
.
•
•
(8)
•
.
•
•
•
•
294 Operators
\\··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. Show that an operator Ton His normal¢=> its adjoint T* is a poly-
nomial in T.
24 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* r
{ 0} C M 1 C ltf ~ C ~ · r 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 295
property that
"'
N = IA?t··
:i= 1
(1)
(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 • • .. • 4 .. • • • I .. t 4 .. • .. ..
E~=P1+P,+ · . ·+Pm.
296 Operators
= I
t: =I
Ai(E}.j - EAi-1).
(3)
In this form, the spectral resolution remains valid for self-adjoint opera-
tors on infinite-dimensional Ililbe.rt spaces. ..A. similar result holds for
normal operators,
N XdE,.,. =I (4)
For a. general discuBBion of t.he spectral thoor~m from this .an.a.lytie point of
1
view, see LoTch [28). A full trep.tment can be found in Riesz and Sz.-Na.gy [3..3,
chap. 7].
Fi nife..dimensiona 1 Spec:tra1 Theory 2'17
Algebras o. Operators
CHAPTER TWELVE
(jeJteral
Preliminaries ou !!auack Algebras
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
71 n
== t I t f(gkg.i~•)g(g;)I
1-1 j:c::l
t t
ft tL
n n
= I lg(g1)I t lf(OlUi-
1
)[
i-1 i:•l
-n.
= I. Jg(g;)[ l!/H .
:i• l
n
= 11/U t
i =-l
IY<ui)I
= 11111 lluH-
(b) Let G = f... ~ -2, -17'
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
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= - llJ
f(n - m)g(m).
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 .. i;=-· (
PHOOF. If we put r = llx ~ l[I, so that r < 1. then
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 .m;i.
so xy = 1. Similarly, yx = I.
We no~ use this as a tool to prove
Theorem B. 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
ICI
l!x- 1 - xo- 1 H =
~
l (x- 1xo ~ l)xo- 1 11 < Hxo- ~r llx-•xo - 1 fl
1
~
==- J!xo- 11
1
n~l n~l
i»
-
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 ele-
ments in A which arc singular and remain singular 'vith respect to all
possible enlargements of A.
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 µare both in p(x)~ then
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
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.
PROOF. Let X. be w non-zero complex number and ~lJ >..2 1 .. • • , Xn its
distinct nth roots 1 so that
The statement of the lemma follows easily from the fact that x" - Xl
is singular <=> x - '>..~I is singular for at least one i~
~ -).-1 [1 + ±~] .
n-1
(1)
Genera I Prefiminaries on Banach Algebras 313
+I
l!IG
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 i;::_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
or 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
1
< a for an but a 1Jn
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 sym-
metry~ it evidently suffices t-0 prove the
and
(1 - rx)- 1 ~ 1 + rx + (rx)~ + (rx) + · ~ · ~ 1 + rx + rxrx
1
l~x + JI~ = + i 11 :i e I J.
inf 111 x
Further, A/ I is clearly an algebra Vr"'ith identity I + I, and
I! 1 + I I! ~ jnf { I~ I + if I :i e I} < I! 1 ;1 = 1.
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
Theorem E.. 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 maxi-
mal left ideals in A and those in A/R.
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
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
and, in general 1 that l!x~*ll = llxll~ for every positive integer k. The
formula for the spectral radius now yields
r(x) = lim IJxnll 11n = lim If xt"ll1 12k = lim llxll = 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 repre-
senting 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) .
PnOOF. By part (1) of Theorem 70-B, we know that A is a suh-
algcbra 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~ con-
tains 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) .
If we put together the results obtained in the above two theorems~
we have
Theorem C. 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
324 Atgebrcs of Operators
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)• = y•x•;
(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
*-iso·morphism 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
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. It is
easy to see that we have llx*ll == Hxll in every B*-algebra; for
shows that llxH : : ; llx*ll for every xi so llx*~I :$ Hx•* I~ = flxH, and therefore
Ux*I~ = flxll. It follows from this that the relation 11x*xll = flx*11 llx~I is
also true.
Several of the Banach algebras described in Sec. 64 are also Bannch
The Structure of Commutative Banach Algebras 325
Some Speeia/
Commutative Ranae/J Algebras
f (y) ~
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 ffl•mmarize 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},
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 th·at 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 alge-
braic 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) I These observations lead us directly to
Theorem D (the Banach-Stone Theorem). 7 wo compact Hausdorff spaces
1
...
75. THE STONE-CECH COMPACTIFICATJON (continued).
Theorem A,. 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.
If 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 suh-
algebra 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
specialize Theorem A to
\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(T•U- 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.
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 commuta-
tive 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*-alge-
1
bra 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) .
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
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 par-
agraph of Sec.. 63, and all our promises are fulfilled.
Appendices
APPENDIX ONE
1
r------
11
I
I
J
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) = Xo-
It 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~ 251-
255]) ~ Both of these facts are special cases of
1
BrOtJwer s fixed Point Theorem. The closed unit spM.re S = tx ~ llx I < 1}
in Rn is a fixed point 8pace.
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.. Every convex compact subspace of a
Banach space is a fixed point space+
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 solu-
tions 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
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
l/(x,y)[ ~ K (2)
for all points (x,y) in R~ We next observe that if (x;y1) and (x,y~) are
in R, then the mean value theorem guarantees that
l/(:i:,y1) - /{:i:,y2) I = IY• - y,I aya f(x, Yi+ 9(y, - 111)) (4)
for some 8 such that 0 <8< I. It now follows from {3) and (4) that
If(x,y1) - 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
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
i
The fact that T maps X into itself is evident from (2), for
Jlalt1t-Mozurkiewicz ?:lteore111
I I I
I I I
I - +--+~ ~.!.- -
l I I I
J I I
-- --+--
l
·- --t-
I
r
~
.J__
f
_..
Il ! I
...... T
I
l~ - +--T--+
I
I
J
11 {l) I 12 (I}
Fig. 40
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 con-
structed 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 mathe-
maticians 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 e·xistence 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
Nooleax Algebras,
Hoo!eax /(ixgs, 11Hd Stone$ 7:1teore111
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 A z ~ (x A y) A z = x A (y A z) = x A y = 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, +
= x + x·y + yx + yj
Bootecn Algebras, Boolean Rings, and Stone*s Theorem 347
= (y A X t) V (y 1 A X)
= Y + x~
that
x +0 = (x A O') v (xf " 0) = (x A 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. It is con-
venient to begin v.:-i th the o bscrvat.io n that
(X + y) t = [ (X Ay ') V (X
1
/\ y) r= (X1 V y) A (.T V y')
~ [(X 1f) /\ y 1]
1 1
V Y) A X] V [ ( X V
= [(X 1
/\ X) V ( y A .r) j V { { X t A y ~) V ( y A y1) j
= (X A JJ) V (X, A y')+
Now, using this~ we have
x + (y + z} = [x" (y + z) ] v {x " (y + z)l
1 1
and
xy + xz = (x A y) + (x A z)
r(x J\ y) z) 1 l v [Cr A y) t A ( x A z) 1
A ( x J\
[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; A X !\ Z)
= (x A y A z') v (x A y' ;., z) T
x A y = xy and 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 7
and :r. == ~ :r, then (1) and (2) are evident. Property (3) follO"\VS from
the a.ssociativity of multiplication and the computations
:r v (y v z) = x v (y + z + y z)
= x + y + z + y.z + xy + xz + xyz
and (xv y) v z = (x + y + xy) v z
~ x + y + xy + z + xz + yz + xyz+
Property (4) is a1so true~ for
(x" y) v x = xy v x = xy + x + xyx = x + xy + xy = x
and
(xv y) Ax = (x + y + xy) Ax = x 2 + yx + xyx = x + xy + xy = x.
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:
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
1 = 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) = x2 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 homo-
morphlsms of R onto lO, l} ~ We shall need the follo,ving stronger
statement.
Lemma. 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
I= {m+r(l+x):ms.MandreRl
is the smallest ideal containing both M and 1 + x so I properly contains
1
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
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 homo-
rn 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*;
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:
{f:f(x) = 0, f(y) == 0, and f(x + y) = 0L
{/ =f(x) :=:: 0, J(y) = 1, and j(x + y) = 1L
{f:f(x) === 1, f(y) = O, and f(x + y) = 1L
and {j:f(x) = 1, f(y) = IJ and f(x + 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
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
+
in the form Bi = T(z,) for some element z;· in R. It now follows that
s = Vi-1 Bi = en;. . \ B/) 1
= (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 iso·motphic 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
Empty set, 5 •
u Universal set, 5
u,.A1.t etc. Union of a class of sets1 11