Yosida 1 Functional - Analysis

DIE GRUNDLEHREN DER
MATHEMATISCHEN
WISSENSCHAFTEN
IN EINZELDARSTELLUNGEN MIT BESONDERER
BER U CKSI CHTI G UNG D ER ANWEND UN GS GE BlE TE
HERAUSGEGEBEN VON
J.L.DOOB E.HEINZ. F.HIRZEBRUCH

E. HOPF H. HOPF W. MAAK. S. MAC LANE
W. MAGNUS. F. K. SCHMIDT. K. STEIN
GESCHAFTSFUHRENDE HERAUSGEBER
B. ECKMANN UND B.L.VAN DER WAERDEN

ZURICH
BAND 123
SPRINGER-VERLAG BERLIN HEIDELBERG GMBH

1966
FUNCTIONAL ANALYSIS
BY
KOSAKU YOSIDA
PROFESSOR OF MATHEMATICS
AT THE UNIVERSITY OF TOKYO
SPRINGER-VERLAG BERLIN HEIDELBERG GMBH

1966
Geschftsfiihrende
Herausgeber:
Prof. Dr. B. Eckmann

EidgenOssische Technische Hochschule Ziirich
Prof. Dr. B. L. van der Waerden

Mathematisches Institut der
Universitt
Ziirich
All rights reserved, especially that of translation into foreign languages. It is also
forbidden to reproduce this book, either whole or in part, by photomechanical means
(photostat, microfilm and/ormicrocard) or any other means without written permission
from the Publishers
ISBN 978-3-642-52816-3
ISBN 978-3-642-52814-9 (eBook)
DOI 10.1007/978-3-642-52814-9
Springer-Verlag Berlin Heide1berg 1965

Originally pub1ished by Springer-Verlag Berlin Heide1berg New York in 1965
Softcover reprint of the hardcover 1st edition 1965
Library of Congress Catalog Card Number 64-8025
Titei Nr. 5106
Preface
The present book is based on lectures given by the author at the
University of Tokyo during the past ten years. It is intended as a
textbook to be studied by students on their own or to be used in a course
on Functional Analysis, i.e., the general theory of linear operators in
function spaces together with salient features of its application to
diverse fields of modern and classical analysis.
Necessary prerequisites for the reading of this book are summarized,
with or without proof, in Chapter 0 under titles: Set Theory, Topological Spaces, Measure Spaces and Linear Spaces. Then, starting with
the chapter on Semi-norms, a general theory of Banach and Hilbert
spaces is presented in connection with the theory of generalized functions
of S. L. SoBOLEV and L. ScmVARTZ. While the book is primarily addressed
to graduate students, it is hoped it might prove useful to research mathematicians, both pure and applied. The reader may pass e.g. from
Chapter IX (Analytical Theory of Semi-groups) directly to Chapter XIII
(Ergodic Theory and Diffusion Theory) and to Chapter XIV (Integ-ration
of the Equation of Evolution). Such materials as "Weak Topologies
and Duality in Locally Convex Spaces" and "Nuclear Spaces" are
presented in the form of the appendices to Chapter V and Chapter X,
respectively. These might be skipped for the first reading by those who
are interested rather in the application of linear operators.
In the preparation of the present book, the author has received
valuable advice and criticism from many friends. Especially, Mrs.
K. HILLE has kindly read through the manuscript as well as the galley
and page proofs. Without her painstaking help, this book could not
have been printed in the present style in the language which was
not spoken to the author in the cradle. The author owes very much
to his old friends, Professor E. HILLE and Professor S. KAKUTANI of
Yale University and Professor R. S. PHILLIPS of Stanford University for
the chance to stay in their universities in 1962, which enabled him to
polish the greater part of the manuscript of this book, availing himself
of their valuable advice. Professor S. ITo and Dr. H. KoMATSU of the
University of Tokyo kindly assisted the author in reading various parts
VI
Preface
of the galley proof, correcting errors and improving the presentation.

To all of them, the author expresses his warmest gratitude.
Thanks are also due to Professor F. K. ScHMIDT of Heidelberg UniYersity and to Professor T. KATO of the University of California at
Berkeley who constantly encouraged the author to write up the present
book. Finally, the author wishes to express his appreciation to SpringerVerlag for their most efficient handling of the publication of this book.
Tokyo, September 1964
KosAKu YosmA
Contents
0. Preliminaries
1.
2.
3.
4.
Set Theory
Topological Spaces
Measure Spaces .
Linear Spaces
I. Semi-norms . . .
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
Semi-norms and Locally Convex Linear Topological Spaces.

Norms and Quasi-norms . . . . . . . .
Examples of Normed Linear Spaces . . .
Examples of Quasi-normed Linear Spaces
Pre-Hilbert Spaces . . . . . . . .
Continuity of Linear Operators . . . . .
Bounded Sets and Bornologic Spaces
Generalized Functions and Generalized Derivatives
B-spaces and F-spaces
The Completion . . . . .
Factor Spaces of a B-space
The Partition of Unity. . .
Generalized Functions with Compact Support.
The Direct Product of Generalized Functions .
TJ. Applications of the Baire-Hausdorff Theorem

1. The Uniform Boundedness Theorem and the Resonance
Theorem . . . . . . . . . . . . . . . . . . . . . .
2. The Vitali-Hahn-Saks Theorem . . . . . . . . . . . . .
3. The Termwise Differentiability of a Sequence of Generalized
Functions . . . . . . . . . . . . . . . . . .
4. The Principle of the Condensation of Singularities
5. The Open Mapping Theorem . . . . . . . . . .
6. The Closed Graph Theorem. . . . . . . . . . .
7. An Application of the Closed Graph Theorem (Hormander's
Theorem) . . . . . . . . . . . . . . . . . . . . . .
lii. The Orthogonal Projection and F. Riesz' Representation Theorem . . . . . . . . . . . . .
1. The Orthogonal Projection .
2. "Nearly Orthogonal" Elements
.1
1
3
lfj
20
23
23
30
32
38
39
42
44
46
52
56
59
60
62
65
68
68
70
72
72
75
77
80
81
81
84
VIII
Contents
3. The Ascoli-Arzela Theorem . . . . . . . . . .
85
4. The Orthogonal Base. Bessel's Inequality and Parseval's
Relation . . . . . . . . . . . .
86
5. E. Schmidt's Orthogonalization . .
88
6. F. Riesz' Representation Theoren:
90
7. The Lax-Milgram Theorem . . . .
92
8. A Proof of the Lebesgue-Nikodym Theorem
93
9. The Aronszajn-Bergman Reproducing Kernel .
95
10. The Negative Norm of P. LAx . . . . .
98
11. Local Structures of Generalized Functions
100
IV. The Hahn-Banach Theorems . . . . . . . .

1.
2.
3.
4.
5.
6.
7.
8.
9.
The Hahn-Banach Extension Theorem in Real Linear Spaces

The Generalized Limit . . . . . . . . . . . . . . . .
Locally Convex, Complete Linear Topological Spaces
The Hahn-Banach Extension Theorem in Complex Linear
Spaces . . . . . . . . . . . . . . . . . . . . . . .
The Hahn-Banach Extension Theorem in Normed Linear
Spaces . . . . . . . . . . . . . . . . . . . . . . . . .
The Existence of Non-trivial Continuous Linear Functionals
Topologies of Linear Maps . . . . . . . .
The Embedding of X in its Bidual Space X"
Examples of Dual Spaces . . . . . .
V. Strong Convergence and Weak Convergence
102
102
103
104
105
106
107
110
112
114
119
1. The Weak Convergence and The weak* Convergence

120
2. The Local Sequential Weak Compactness of Reflexive Bspaces. The Uniform Convexity . . . . . . . . . .
126
3. Dunford's Theorem and The Gelfand-Mazur Theorem
128
4. The Weak and Strong Measurability. Pattis' Theorem
130
5. Bochner's Integral . . . . . . . . . . . . . . .
132
Appendix to Chapter V. Weak Topologies and Duality in Locally
Convex Linear Topological Spaces
136
1.
2.
3.
4.
Polar Sets . . . . . . . . .
Barrel Spaces . . . . . . .
Semi-reflexivity and Reflexivity
The Eberlein-Shmulyan Theorem
136
138
139
141
VI. Fourier Transform and Differential Equations
145
1.
2.
3.
4.
The Fourier Transform of Rapidly Decreasing Functions

The Fourier Transform of Tempered Distributions. . . .
Convolutions . . . . . . . . . . . . . . . . . . . .
The Paley-Wiener Theorems. The One-sided Laplace Transform . . . . . . . . . . . . . .
. . . . . .
5. Titchmarsh's Theorem . . . . .
6. Mikusiflski's Operational Calculus .
7. Sobolev's Lemma . . . . . . . .
146
149
156
161
166
169
173
Contents
8.
9.
10.
11.
12.
Garding's Inequality
Friedrichs' Theorem
The Malgrange-Ehrenpreis Theorem
Differential Operators with Uniform Strength.
The Hypoellipticity (Hormander's Theorem)
VII. Dual Operators

l.
2.
3.
4.
5.
Dual Operators . .
Adjoint Operators
Symmetric Operators and Self-adjoint Operators
Unitary Operators. The Cayley Transform
The Closed Range Theorem . . . . . . . . .
VIII. Resolvent and Spectrum. . . .

1.
2.
3.
4.
5.
6.
7.
8.
IX
175
177
182
188
189
193
193
195
197
202
205
209
209
The Resolvent and Spectrum . . . . . . . .
211
The Resolvent Equation and Spectral Radius
213
The Mean Ergodic Theorem . . . . . . . .
Ergodic Theorems of the Hille Type Concerning Pseudo215
resolvents . . . . . . . . . . . . . . . . .
218
The Mean Value of an Almost Periodic Function
224
The Resolvent of a Dual Operator. . . .
225
Dunford's Integral . . . . . . . . . .
228
The Isolated Singularities of a Resolvent
IX. Analytical Theory of Semi-groups. . . .

1. The Semi-group of Class (C0}
2. The Equi-continuous Semi-group of Class (C0 ) in Locally
Convex Spaces. Examples of Semi-groups . . . .
3. The Infinitesimal Generator of an Equi-continuous Semigroup of Class (C0 )
4. The Resolvent of the Infinitesimal Generator A . . .
5. Examples of Infinitesimal Generators . . . . . . .
6. The Exponential of a Continuous Linear Operator whose
Powers are Equi-continuous . . . . . . . . . . . . . .
7. The Representation and the Characterization of Equi-continuous Semi-groups of Class (C0 ) in Terms of the Corresponding Infinitesimal Generators . . . . . . . . . .
8. Contraction Semi-groups and Dissipative Operators . .
9. Equi-continuous Groups of Class (C0 ). Stone's Theorem
10. Holomorphic Semi-groups . . . . . . . . . . . . .
11. Fractional Powers of Closed Operators . . . . . . .
12. The Convergence of Semi-groups. The Trotter-Kato Theorem
13. Dual Semi-groups. Phillips' Theorem . . . . . . . . . .
X. Compact Operators . . . . . . . . . .
1. Compact Sets in B-spaces . . . . .
2. Compact Operators and Nuclear Operators .
231
232
234
237
240
242
244
246
250
251
254
25B
2GB
272
274
274
277
Contents
3.
4.
5.
6.
The Rellich-Garding Theorem

Schauder's Theorem
The Riesz-Schauder Theory
Dirichlet's Problem . . . .
281
282
283
286
Appendix to Chapter X. The Nuclear Space of A. GROTHENDIECK 2H!l

XI. Normed Rings and Spectral Representation
1.
2.
R.
4.
5.
6.
7.
H.
H.
10.
11.
12.
lR.
14.
15.
16.
Maximal Ideals of a Normed Ring. . .

The Radical. The Semi-simplicity . . .
The Spectral H.esolution of Bounded Normal Operators
The Spectral Resolution of a Unitary Operator . .
The Resolution of the Identity . . . . . . . . . . .
The Spectral Resolution of a Self-adjoint Operator
Real Operators and Semi-bounded Operators. Friedrichs'
Theorem
. . . . . . . . . . . . . . . . . . . . . .
The Spectrum of a Self-adjoint Operator. Rayleigh's Principle and the Krylov-\Veinstein Theorem. The Multiplicity
of the Spectrum. . . . . . . . . . .
The General Expansion Theorem. A Condition for the
Absence of the Continuous Spectrum
. . . . .
The Peter-Weyl-Neumann Theorem . . . . . . .
Tannaka's Duality Theorem for Non-commutative Compact
Groups . . . . . . . . . . . . . . .
Functions of a Self-adjoint Operator
. . . . . .
Stone's Theorem and Bochner's Theorem. . . . . . . . .
A Canonical Form of a Self-adjoint Operator with Simple
Spectrum . . . . . . . . . . . . . . . . . . . . . .
The Defect Indices of a Symmetric Operator. The Generalized
Resolution of the Identity . . . . . . . . . . . .
The Group-ring D and \Viener's Tauberian Theorem
XII. Other Representation Theorems in Linear Spaces .

1.
2.
3.
4.
5.
G.
Extremal Points. The Krein-Milman Theorem

Vector Lattices . . . . . . . . .
B-lattices and F-lattices . . . . . . . . . .
A Convergence Theorem of BANACH . . . . .
The Representation of a Vector Lattice as Point Functions
The Representation of a Vector Lattice as Set Functions
XIII. Ergodic Theory and Diffusion Theory . . . . . . . .

l.
2.
3.
4.
5.
6.
7.
8.
The Markov Process with an Invariant Measure

An Individual Ergodic Theorem and Its Applications
The Ergodic Hypothesis and the H-theorem . . . .
The Ergodic Decomposition of a Markov Process with a
Locally Compact Phase Space . . . . . . . . . . . .
The Brownian Motion on a Homogeneous Riemannian Space
The Generalized Laplacian of W. FELLER
An Extension of the Diffusion Operator
Markov Processes and Potentials . . . .
2!li
2!);)
2!1Fi
302
30G
30!1
R1R
Rlfi
31!1
R23
32fi
332
R38
345
R4 7
34!l
354
362
R62
364
369
370
372
R75
379
379
38R
389
393
398
403
408
410
Contents
XIV. The Integration of the Equation of Evolution
in L 2 (Rm)
2. Integration of Diffusion Equations in a Compact Riemannian Space . . . . . . . . . . . . . . . . . . . .
3. Integration of Wave Equations in a Euclidean Space R"'
4. Integration of Temporally Inhomogeneous Equations of
Evolution in a Reflexive B-space . . .
5. The Method of TANABE and SoBOLEVSKI
1. Integration of Diffusion Equations
XI
Jll
413
419
421
424
431
Bibliography
43H
Index . . . .
4:i3
0. Preliminaries
It is the purpose of this chapter to explain certain notions and theorems used throughout the present book. These are related to Set Theory,
Topological Spaces, Measure Spaces and Linear Spaces.
1. Set Theory
Sets. x E X means that x is a member or element of the set X; x E:: X
means that x is not a member of the set X. \Ve denote the set consisting of all x possessing the property P by {x; P}. Thus {y; y = x} is
the set {x} consisting of a single element x. The void set is the set with
no members, and will be denoted by 0. If every element of a set X is also
an element of a set Y, then X is said to be a subset of Y and this fact
will be denoted by X s;; Y, or Y ;::? X. If I is a set whose elements are
sets X, then the set of all x such that x EX for some X E I is called the
union of sets X in I; this union will be denoted by U X. The interXEX
section of the sets X in I is the set of all x which are elements of every
X E I; this intersection will be denoted by n X. Two sets are disxEx
ioint if their intersection is void. A family of sets is disjoint if every

pair of distinct sets in the family is disjoint. If a sequence {Xn}n~l, 2 , ...
00
of sets is a disjoint family, then the union U Xn may be written in

n~l
the form of a sum
00
n~l
Xn.
Mappings. The term mapping, function and transformation will be

used synonymously. The symbol f: X_.,.. Y will mean that f is a singlevalued function whose domain is X and whose range is contained in Y;
for every x E X, the function f assigns a uniquely determined element
f(x) = y E Y. For two mappings f: X_.,.. Y and g: Y _.,.. Z, we can
define their composite mapping gf: X_.,.. Z by (g/) (x) = g(f(x)). The
symbol f (M) denotes the set {/ (x); x E M} and f (M) is called the image
of M under the mapping f. The symbol j-1 (N) denotes the set {x; f(x)EN}
and j-1 (N) is called the inverse image of N under the mapping f. It is
clear that
Y1 = j(f-1 (Y1)) for all Y1

1
Yosida, Functional Analysis
s;; /(X),
and X 1
s;; j-1 (t(X1))
for all X 1
s;; X.
0. Preliminaries
Iff: X-+ Y, and for each y E /(X) there is only one x EX with f(x) = y,
then f is said to have an inverse (mapping) or to be one-to-one. The inverse
mapping then has the domain f (X) and range X; it is defined by the
equation x = j-1 (y) = j-1 ({y}).
The domain and the range of a mapping f will be denoted by D (/) and
R(f), respectively. Thus, iff has an inverse then
j-1 (f(x))
x for all xED (f), and /(j-1 (y))
y for ally E R (!).
The function f is said to map X onto Y iff (X) = Y and into Y iff (X) ~ Y.
The function f is said to be an extension of the function g and g a restriction
of f if D (f) contains D (g), and f (x) = g (x) for all x in D (g).
Zorn's Lemma
Definition. Let P be a set of elements a, b, ... Suppose there is a
hinary relation defined between certain pairs (a, b) of elements of P,
expressed by a
b, with the properties:
-<
a-< a,
if a
b and b a, then a = b,
if a-< band b-< c, then a-< c (transitivity).
-<
-<
-<.
Then P is said to be partially ordered (or semi-ordered) by the relation

Examples. If P is the set of all subsets of a given set X, then the set
inclusion relation (A ~ B) gives a partial ordering of P. The set of all
complex numbers z = x + i y, w = u + i v, ... is partially ordered by
defining z
w to mean x < u and y ::;: v.
Definition. Let P be a partially ordered set with elements a, b, ...
If a-< c and b-< c, we call can upper bound for a and b. If furthermore
c-< d whenever dis an upper bound for a and b, we call c the least upper
bound or the supremum of a and b, and write c = sup(a, b) or a V b.
This element of P is unique if it exists. In a similar way we define the
greatest lower bound or the infimum of a and b, and denote it by inf (a, b)
or a 1\ b. If a V b and a 1\ b exist for every pair (a, b) in a partially
ordered set P, Pis called a lattice.
Example. The totality of subsets M of a fixed set B is a lattice by
the partial ordering M 1
M 2 defined by the set inclusion relation
-<
M1 ~M2.
-<
Definition. A partially ordered set Pis said to be linearly ordered (or

totally ordered) if for every pair (a, b) in P, either a-< b or b-< a holds.
A subset of a partially ordered set is itself partially ordered by the relation which partially orders P; the subset might turn out to be linearly
ordered by this relation. If Pis partially ordered and Sis a subset of P,
an mE P is called an upper bound of S if s-< m for every s E S. An
mE P is said to be maximal if p E P and m
p together imply m = p.
-<
2. Topological Spaces
Zorn's Lemma. Let P be a non-empty partially ordered set with the

property that every linearly ordered subset of P has an upper bound
in P. Then P contains at least one maximal element.
It is known that Zorn's lemma is equivalent to Zermelo's axiom of
choice in set theory.
Open Sets and Closed Sets
Definition. A system -c of subsets of a set X defines a topology in X
if -c contains the void set, the set X itself, the union of every one of its
subsystems, and the intersection of every one of its finite subsystems.
The sets in -c are called the open sets of the topological space (X, r); we
shall often omit T and refer to X as a topological space. Unless otherwise
stated, we shall assume that a topological space X satisfies Hausdorff's
axiom of separation:
For every pair (x1 , x2) of distinct points x1 , x2 of X, there exist disjoint
open sets G1 , G2 such that x1 E G1 , x2 E G2 .
A neighbourhood of the point x of X is a set containing an open set which
contains x. A neighbourhood of the subset M of X is a set which is a
neighbourhood of every point of M. A point x of X is an accumulation
point or limit point of a subset M of X if every neighbourhood of x contains at least one point mE M different from x.
Definition. Any subset M of a topological space X becomes a topological space by calling "open" the subsets of M which are of the form
M (\ G where G's are open sets of X. The induced topology of Miscalled
the relative topology of Mas a subset of the topological space X.
Definition. A set 1\f of a topological space X is closed if it contains
all its accumulation points. It is easy to see that M is closed iffl its
complement Me = X - M is open. Here A - B denotes the totality of
points x E A not contained in B. If M ~X, the intersection of all closed
subsets of X which contain Miscalled the closure of M and will be denoted
by Ma (the superscript "a" stands for the first letter of the German:
abgeschlossene Hiille).
Clearly Ma is closed and M ~ Ma; it is easy to see that M = Ma iff
M is closed.
Metric Spaces
Definition. If X, Y are sets, we denote by X X Y the set of all ordered
pairs (x,y) where xEX and yE Y; XxY will be called the Cartesian
product of X and Y. X is called a metric space if there is defined a func1
1*
iff is the abbreviation for "if and only if".
0. Preliminaries
tion d with domain X X X and range in the real number field R 1 such
that
d (x1 , x2} > 0 and d (xv x2} = 0 iff x1 = x2 ,
d (xv x 2) = d (x 2 , x1),
d (x1 , x3 ) < d (x1 , x2) + d (x2 , x3 ) (the triangle inequality).
d is called the metric or the distance function of X. With each point x 0
in a metric space X and each positive number r, we associate the set
S (x0 ; r) = {x E X; d (x, x0 } < r} and call it the open sphere with centre x 0
and radius r. Let us call "open" the set M of a metric space X iff, for
every point x 0 EM, M contains a sphere with centre x 0 Then the totality
of such "open" sets satisfies the axiom of open sets in the definition of the
topological space.
Hence a metric space X is a topological space. It is easy to see that a
point x 0 of X is an accumulation point of Miff, to every s > 0, there exists
at least one point m = x 0 of M such that d (m, x 0 ) < s. Then-dimensional
Euclidean space R" is a metric space by
d (x, y)
(.i (xi <~1
Yi) 2) 112 , where x
= (x1 ,
xn) and y
= (y 1 ,
... ,
y,.).
Continuous Mappings
Definition. Let f : X --? Y be a mapping defined on a topological
space X into a topological space Y. f is called continuous at a point x 0 EX
if to every neighbourhood U of f (x0 ) there corresponds a neighbourhood
V of x 0 such that f(V) ~ U. The mapping I is said to be continuous if it is
continuous at every point of its domain D (f) =X.
Theorem. Let X, Y be topological spaces and f a mapping defined
on X into Y. Then f is continuous iff the inverse image under f of every
open set of Y is an open set of X.
Proof. If I is continuous and U an open set of Y, then V = I-1 (U)
is a neighbourhood of every point x 0 EX such that f (x0 } E U, that is,
Vis a neighbourhood of every point x 0 of V. Thus Vis an open set of X.
Let, conversely, for every open set U::;f(x0 } of Y, the set V=f- 1 (U)
be an open set of X. Then, by the definition, f is continuous at x 0 EX.
Compactness
Definition. A system of sets GIX, ex E A, is called a covering of the set
X if X is contained as a subset of the union U<>EA GIX. A subset M of a
topological space X is called compact if every system of open sets of X
which covers M contains a finite subsystem also covering M.
In view of the preceding theorem, a continuous image of a compact set
is also compact.
Proposition 1. Compact subsets of a topological space are necessarily
closed.
Proof. Let there be an accumulation point x0 of a compact set M of a

topological space X such that x0 EM. By Hausdorff's axiom of separation, there exist, for any point mE M, disjoint open sets Gm,x, and Gx,m
of X such that mE Gm,x,, x0 E Gx,,m The system {Gm,x,; mE M} surely
covers M. By the compactness of M, there exists a finite subsystem
{Gmx;
i
t' 0
1, 2, ... , n} which covers M. Then
i=l
Gx0' m does not

t
intersect M. But, since x0 is an accumulation point of M, the open set

n
i=l
GxOl mt 3 Xo must contain a point mE M distinct from Xo This is a
contradiction, and M must be closed.

Proposition 2. A closed subset M 1 of a compact set M of a topological
space X is compact.
Proof. Let {Ga} be any system of open sets of X which covers M 1
M 1 being closed, Mf =X- M 1 is an open set of X. Since M 1 ~ M,
the system of open sets {G,,} plus Mf covers M, and since M is compact, a
properly chosen finite subsystem {Ga;; i = 1, 2, ... , n} plus Mf surely
covers M. Thus {Ga;; i = 1, 2, ... , n} covers M 1 .
Definition. A subset of a topological space is called relatively compact
if its closure is compact. A topological space is said to be locally compact if
each point of the space has a compact neighbourhood.
Theorem. Any locally compact space X can be embedded in another
compact space Y, having just one more point than X, in such a way that
the relative topology of X as a subset of Y is just the original topology
of X. This Y is called a one point compactification of X.
Proof. Let y be any element distinct from the points of X. Let {U} be
the class of all open sets in X such that uc =X-U is compact. We
remark that X itself E {U}. Let Y be the set consisting of the points of X
and the point y. A set in Y will be called open if either (i) it does not
contain y and is open as a subset of X, or (ii) it does contain y and its
intersection with X is a member of {U}. It is easy to see that Y thus
obtained is a topological space, and that the relative topology of X
coincides with its original topology.
Suppose {V} be a family of open sets which covers Y. Then there must
be some member of {V} of the form U0 V {y}, where U0 E {U}. By the
definition of {U}, ug is compact as a subset of X. It is covered by the
system of sets V (\ X with V E {V}. Thus some finite subsystem:
V 1 (\X, V 2 (\X, .. . , Vn (\X covers Uf Consequently, V 1 , V 2 , , Vn
and U0 V {y} cover Y, proving that Y is compact.
Tychonov's Theorem
Definition. Corresponding to each <X of an index set A, let there be given
a topological space X". The Cartesian product IT X" is, by definiaEA
0. Preliminaries
tion, the set of all functions f with domain A such that f(iX) E XIX for
every IX EA. We shall write f = IT f(iX) and call /(IX) the IX-th coordio:EA
nate of
f. When
A consists of integers (1, 2, ... , n), ll Xk is usually

k~l
denoted by X 1 xX2 x xXw We introduce a (weak) topology in the

product space rr X"' by calling "open" the sets of the form II GIX,
o:EA
1XEA
where the open set G" of XIX coincides with X"' for all but a finite set of IX.
Tychonov's Theorem. The Cartesian product X= II XIX of a
o:EA
system of compact topological spaces X" is also compact.

Remark. As is well known, a closed bounded set on the real line R 1 is
compact with respect to the topology defined by the distanced (x, y) =
lx- y I (the Bolzano-Weierstrass theorem). By the way, a subset M
of a metric space is said to be bounded, if M is contained in some sphere
S (x0 , r) of the space. Tychonoff's theorem implies, in particular, that a
parallelopiped:
(i
= 1, 2, ... , n)
of the n-dimensional Euclidean space Rn is compact. From this we see

that Rn is locally compact.
Proof of Tychonov's Theorem. A system of sets has the finite intersection property if its every finite subsystem has a non-void intersection.
It is easy to see, by taking the complement of the open sets of a covering,
that a topological space X is compact iff, for every system {M"; 1XE A}
of its subsets with finite intersection property, the intersection n
o:EA
M:
is non-void.
Let now a system {S} of subsets S of X= II XIX have the finite
<>EA
intersection property. Let {N} be a system of subsets N of X with the

following properties:
(i) {S} is a subsystem of {N},
(ii) {N} has the finite intersection property,
(iii) {N} is maximal in the sense that it is not a proper subsystem of
other systems having the finite intersection property and containing
{S} as its subsystem.
The existence of such a maximal system {N} can be proved by Zorn's
lemma or transfinite induction.
For any set N of {N} we define the set N"' = {f(iX); fEN} ~XIX.
We denote then by {N"'} the system {N"; N E {N}}. Like {N}, {N,,}
enjoys the finite intersection property. Thus, by the compactness of X"''
there exists at least one point p"' EX"' such that p" E n N~. We have
to show that the point p = II piX belongs to the set
cxEA
NE{N}
NE{N}
Na.
n N~'o any open set G,. o

But since Prxo belongs to the intersection ~m
of x"o which contains Pao intersects every N"o E {Nrxo}. Therefore the open
set
JI X a with X O:o E G1Xo}
G("o) = {x; X = 1XEA
of X must intersect every N of {N}. By the maximality condition (iii)
of {N}, G(rxo) must belong to {N}. Thus the intersection of a finite number
of sets of the form G(rx,) with a 0 E A must also belong to {N} and so such a
set intersect every set N E {N}. Any open set of X containing p being
defined as a set containing such an intersection, we see that p = II Prx
must belong to the intersection
NE{N}
cxEA
Na.
Urysohn's Theorem
Proposition. A compact space X is normal in the sense that, for any
disjoint closed sets F 1 and F 2 of X, there exist disjoint open sets G1 and
G2 such that F 1 ~ G1 , F 2 ~ G2
Proof. For any pair (x, y) of points such that x E Fl> y E F 2 , there
exist disjoint open sets G(x,y) and G(y,x) such that xEG(x,y),
y E G(y, x). F 2 being compact as a closed subset of the compact space X,
we can, for fixed x, cover F 2 by a finite number of open sets G(y 1 , x),
G(y2 , x), ... , G(Yn(x) x). Set
n(x)
n(x)
Gx = U G(yj, x) and G(x) = (1 1 G(x, Yi).
;=1
)T
Then the disjoint open sets Gx and G (x) are such that F 2 ~ Gx, x E G (x)
F 1 being compact as a closed subset of the compact space X, we can cover
F 1 by a finite number of open sets G (x1 ), G (x2), . . , G (xk). Then
k
G1 = U G (xj)
J=l
and
G2 =
j=l
Gx
'
satisfy the condition of the proposition.

Corollary. A compact space X is regular in the sense that, for any
such that
non-void open set G~ of X, there exists a non-void open set
(G;)a ~ G~.
Proof. Take F 1 = (G~f and F 2 = {x} where x E G~. We can then take
the open set G2 obtained in the preceding proposition.
for
Urysohn's Theorem. Let A, B be disjoint closed sets in a normal space
X. Then there exists a real-valued continuous function f (t) on X such
that
0 < f(t) "'S 1 on X, and f(t) = 0 on A, f(t) = 1 on B.
c;
c;
Proof. We assign to each rational number r = k/2n (k = 0, 1, ... , 2n),

an open set G (r) such that (i) A ~ G (0), B = G (1)c, and (ii) G (rt ~ G (r')
whenever r < r'. The proof is obtained by induction with respect to n.
0. Preliminaries
For n = 0, there exist, by the normality of the space X, disjoint open sets
G0 and G1 with A ~ G0 , B ~ G1 . We have only to set G0 = G(O). Suppose
that G (r)'s have been constructed for r of the form k/2n.-l in such a
way that condition (ii) is satisfied. Next let k be an odd integer > 0.
Then, since (k- 1)/2" and (k + 1)/2" are of the form k' /2"- 1 with
0 < k' S: 2"-1, we have G ((k -- 1)/2"t ~ G((k + 1)/2"). Hence, by the
normality of the space X, there exists an open set G which satisfies
G((k --1)/2"t ~ G, Ga ~ G((k + 1)/2"). If we set G(k/2") = G, the induction is completed.
Define f (t) by
f(t)
= 0 on G(O), and f(t) = supr whenever tE G(Of.

tEG(r)
Then, by (i), f(t) = 0 on A and f(t) = 1 on B. We have to prove the

continuity of f. For any t0 EX and positive integer n, we take r with
f (t0 ) < r < f (t0 ) + 2-"- 1 Set G = G (r) (\ G (r- 2-ntc (we set, for
convention, G(s) = 0 if s < 0 and G(s) =X if s > 1). The open set G
contains to- For, f (t0) < r implies t0 E G (r), and (r- 2-"- 1 ) < /(t0 )
implies t0 E G (r- 2-"- 1 f ~ G (r - 2-"tc. Now t E G implies t E G (r)
and so f (t) < r; similarly t E G implies t E G (r- 2-ntc ~ G (r- 2-")c so
that r- 2-n < f(t). Therefore we have proved that lf(t)- f(t0 ) I< 1/2"
whenever t E G.
The Stone-Weierstrass Theorem
Weierstrass' Polynomial Approximation Theorem. Let f (x) be a realvalued (or complex-valued) continuous function on the closed interval
[0, 1]. Then there exists a sequence of polynomials Pn (x) which converges,
as n ~ oo, to f(x) uniformly on [0, 1]. According to S. BERNSTEIN, we
may take
(1)
Proof. Differentiating (x
+ yt = 2: ,.Cp
P=O
x and multiplying by x, we obtain nx(x
xP yn-p with respect to

n
+ yt- 1 =.I p
P=O
,.Cp xPyn-P.
Similarly, by differentiating the first expression twice with respect to x and

multiplying by x 2 , we obtain n (n- 1) x2 (x
y"-P. Thus, if we set
(2)
we have
n
" p (p- 1) ,.Cp xP

+ yt- 2 = P=O
2:
.I rp(x) = 1, .I prp(x) = nx, .I p(p-1) rp(x) = n(n-1) x2 (3)

P=O
P=O
P=O
Hence
n
2: (p -nx) 2 rp(x) = n 2 x2 2: rp(x)- 2nx 2: prp(x)

p~o
p~o
p~o
= n 2 x2 - 2nx nx + (nx
2: p 2 rp(x)
p~o
+ n (n- 1) x2 )
(4)
=nx(1-x).
We may assume that lf(x) I < M < CXJ on [0, 1]. By the uniform
continuity of f(x), there exists, for any 8 > 0, a b > 0 such that
lf(x)- f(x') I<

We have, by (3),
I=
PJ0j(pjn) rp(x) IP-#o (f(x) -f(pjn)) rp(x)

IIP-~:;;dnl + IIP-~>dnl
f(x)-
<
For the first term on the right, we have, by rp (x)

I
~
IP-.iXT:;;dn
(5)
whenever lx- x' I< b.
<
>
0 and (3),
2: rp(x) =
p~o
8.
For the second term on the right, we have, by (4) and lf(x) I< M,
I
2:
IP-nxl>dn
<2M
2:
IP-nxl>dn
2Mx(l-x)
nc52
=
rp(x) <
<
2M
2 _. 2
n u
2: (p- nx) 2 rp(x)
p~o
2c52n___,.. 0 (as n ____,. CXJ).
The Stone-Weierstrass Theorem. Let X be a compact space and C (X)

the totality of real-valued continuous functions on X. Let a subset B of
C (X) satisfy the three conditions: (i) if /, g E B, then the function product f g and linear combinations tx/ + {3g, with real coefficients tx, {3,
belong to B, (ii) the constant function 1 belongs to B, and (iii) the uniform
limit / 00 of any sequence {/,.} of functions E B also belongs to B. Then
B = C (X) iff B separates the points of X, i.e. iff, for every pair (sv s2) of
distinct points of X, there exists a function x in B which satisfies
x (s1 ) #- x (s2).
Proof. The necessity is clear, since a compact space is normal and so,
by Urysohn's theorem, there exists a real-valued continuous function x
such that x(s1) #- x(s 2).
To prove the sufficiency, we introduce the lattice notations:
(f V g) (s)
max(f(s), g(s)), (f 1\ g) (s)
min(f(s), g(s)), Ill (s)
lf(s) I
By the preceding theorem, there is a sequence {Pn} of polynomials such that
llt:-Pn(t)l<lfn for
-n<t<n.
10
0. Preliminaries
Hence 11/(s) 1-PnCf(s))l < 1/n if -n < f(s) < n. This proves, by (iii),
that I/ IE B iff E B, because any function f(s) E B ~ C (X) is bounded on
the compact space X. Thus, by
fVg=ftg+lJ2gl
and
fl\g=I--+;FI_If----;___d,
we see that B is closed under the lattice operations V and (\.

Let hE C (X) and Sv s2 E X be arbitrarily given such that s1 # s2
Then we can find an /s,s, E B with /s,s, (s1 ) = h (s1) and /s,s, (s2) = h (s 2).
To see this, let g E B be such that g (s1 ) # g (s 2), and take real numbers <X
and f3 so that /s,s, = ,xg + f3 satisfies the conditions: /s,s, (s1) = h (s1 )
and /s,s, (s 2) = h (s 2).
Given e > 0 and a point t E X. Then, for each s E X, there is a neighbourhood U(s) of s such that fst(u) > h(u) - s whenever uE U(s). Let
U (s 1), U (s 2), . . . , U (sn) cover the compact space X and define
ft = fs,t V V fsnt
Then ft E B andft(u) > h(u)- e for all u EX. We have, by fs;t(t) = h(t),
ft(t) = h(t). Hence there is a neighbourhood V(t) oft such that ft(u) <
h(u) + e whenever uE V(t). Let V(t 1 ), V(t2 ), , V(tk) cover the compact space X, and define
Then fEB and f(u)
>
I = It, 1\ 1\
ftk
h(u) -s :!'or all uE X, because ft;(u)
>
h(u) -e
for u EX. Moreover, we have, for an arbitrary point u EX, say u E V (t;),
f(u) < !t;(u) < h(u) +e.
Therefore we have proved that lf(u)- h(u) I< eon X.
We have incidentally proved the following two corollaries.
Corollary 1 (KAKUTANI-KREIN). Let X be a compact space and C (X)
the totality of real-valued continuous functions on X. Let a subset B
of C(X) satisfy the conditions: (i) if/, gE B, then f V g, f (\ g and the
linear combinations ,xf + (Jg, with real coefficients <X, (3, belong to B,
(ii) the constant function 1 belongs to B, and (iii) the uniform limit / 00
of any sequence {In} of functions E B also belongs to B. Then B = C (X)
iff B separates the points of X.
Corollary 2. Let X be a compact space and C (X) be the totality of
complex-valued continuous functions on X. Let a subset B of C (X)
satisfy the conditions: (i) if /, g E B, then the function product f g and
the linear combinations ,xf + (Jg, with complex coefficients <X, (3, belong
to B, (ii) the constant function 1 belongs to B, and (iii) the uniform
limit / 00 of any sequence {In} of functions E B also belongs to B. Then
B = C (X) iff B satisfies the conditions: (iv) B separates points of X
and (v) if f(s) E B, then its complex conjugate function f(s) also belongs
to B.
11
Weierstrass' Trigonometric Approximation Theorem. Let X be the

circumference of the unit circle of R 2 It is a compact space by the usual
topology, and a complex-valued continuous function on X is represented
by a continuous function f(x), -oo < x < oo, of period 2:n:. If we take,
in the above Corollary 2, forB the set of all functions representable by
linear combinations, with complex coefficients of the trigonometric
functions
einx (n = 0, 1, 2, ... )
and by those functions obtainable as the uniform limit of such linear

combinations, we obtain Weierstrass' trigonometric approximation theorem: Any complex-valued continuous function f(x) with period 2:n: can
be approximated uniformly by a sequence of trigonometric polynomials
of the form I en einx.
n
Completeness
A sequence {xn} of elements in a metric space X converges to a limit

point x E X iff lim d (xn. x) = 0. By the triangle inequality d (xn, xm) <
.......oo
d (x, Xm), we see that a convergent sequence {xn} in X satisfies

d (xn, x)
Cauchy's convergence condition
(1)
Definition. Any sequence {xn} in a metric space X satisfying the above
condition (1) is called a Cauchy sequence. A metric space X is said to be
complete if every Cauchy sequence in it converges to a limit point EX.
It is easy to see, by the triangle inequality, that the limit point of
{xn}, if it exists, is uniquely determined.
Definition. A subset M of a topological space X is said to be nondense in X if the closure Ma does not contain a non-void open set of X.
Miscalled dense in X if Ma =X. M is said to be of the first category if M
is expressible as the union of a countable number of sets each of which is
non-dense in X; otherwise M is said to be of the second category.
Baire's Category Argument
The Baire-Hausdorff Theorem. A non-void complete metric space is of
the second category.
Proof. Let {Mn} be a sequence of closed sets whose union is a complete
metric space X. Assuming that no Mn contains a non-void open set, we
shall derive a contradiction. Thus Mf is open and Mfa =X, hence Mf
contains a closed sphere 5 1 = {x; d (xv x) ~ r 1} whose centre x1 may be
taken arbitrarily near to any point of X. We may assume that 0 < r 1 <1/2.
By the same argument, the open set M~ contains a closed sphere
12
0. Preliminaries
5 2 = {x; d(x 2 , x) < r 2} contained in 5 1 and such that 0 < r2 < 1/22
By repeating the same argument, we obtain a sequence {5,.} of closed
spheres sn = {x; d(xn, x) < r,.} with the properties:
0 < r,.
<
1/2", Sn+l ~ S,., Sn (\ M,.
=0
(n
1, 2, ... ) .
The sequence {x,.} of the centres forms a Cauchy sequence, since, for any
n < m, Xm E sn so that d (xn, Xm) < rn < 1/2". Let Xoo EX be the limit
point of this sequence {xn} The completeness of X guarantees the existence of such a limit point X00 By d (xn, X 00) < d (x,., Xm) + d (xm, x00 ) <
rn + d (xm, x00 ) -+ rn (as m-+ ex:>), we see that X 00 E 5,. for every n. Hence
00
x 00 is in none of the sets Mn, and hence X00 is not in the union U Mn =X,
n=I
contrary to X00 E X.
Baire's Theorem 1. Let M be a set of the first category in a compact
topological space X. Then the complement Me =X- M is dense in X.
Proof. We have to show that, for any non-void open set G, Me interco
sects G. Let M = U Mn where each M,. is a non-dense closed set. Since

n=l
M 1 = Mi is non-dense, the open set Mf intersects G. Since X is regular

as a compact space, there exists a non-void open set G1 such that
Gi ~ G (\ Mf. Similarly, we can choose a non-void open set G2 such that
G~ ~ G1 (\ Mf Repeating the process, we obtain a sequence of non-void
open sets {Gn} such that
G~+l ~ G, (\ M~+l
(n = 1, 2, ... ) .
The sequence of closed sets {G~} enjoys, by the monotony inn, the finite
intersection property. Since X is compact, there is an x EX such that
00
x E nr-:2 1 G~.
(n
x E Gi implies
x E G, and from x E G~+l ~ G, (\ M~+l
0, 1, 2, ... ; Go= G), we obtain X E
00
M~ =Me. Therefore we
have proved that G (\Me is non-void.
Baire's Theorem 2. Let {Xn (t)} be a sequence of real-valued continuous
functions defined on a topological space X. Suppose that a finite limit:
lim Xn (t) = x (t)
n=l
n-+00
exists at every point t of X. Then the set of points at which the function
xis discontinuous constitutes a set of the first category.
Proof. We denote, for any set M of X, by M- the union of all the
open sets contained in M; Mi will be called the interior of M.
Put Pm(8)
00
= {t EX; Jx(t)- xm(t) J < 8, 8 > 0}, G(8) = U P;,.(8).

m=l
Then we can prove that C =
00
n G (1/n)
n=l
coincides with the set of all
points at which x (t) is continuous. Suppose x (t) is continuous at t
t0
13
n G(1/n).
n=l
exists an m such that Ix (t0 ) - Xm (t0 ) I <
We shall show that t0 E
00
Since
lim xn(t) = x(t), there
e/3. By the continuity of x (t) and

Xm (t) at t = t0 , there exists an open set U1, 3 t0 such that Ix (t)- x (t0) I< e/3
lXm (t) - Xm (t0 ) I < e/3 whenever t E U1,. Thus t E U1, implies
Xm (t0 ) I + IXm (t0 ) - Xm (t) I < e,

which proves that t0 E P~(e) and so t0 E G(e). Since e > 0 was arbitrary,
Xm (t) I <
IX(t) -
IX(t) -
we musthavet0 E
X(t0 ) I + IX(t0)
00
n G(l/n).
n=l
00
n G (1/n). Then, for any e > 0, t0 E G (e/3) and

n=l
so there exists an m such that t0 E P~ (e/3). Thus there is an open set
U1, 3 t0 such that t E U1, implies Ix (t) - Xm (t) I < ef3. Hence, by the
continuity of Xm (t) and the arbitrariness of e> O,x(t) must be continuous
at t = t0
After these preparations, we put
Let, conversely, t0 E
Fm(e) = {t EX; lxm(t)- Xm+k(t) I< e
(k = 1, 2, ... )}.
00
This is a closed set by the continuity ofthe Xn (t)'s. We have X= U F m(e)

m=l
by lim xn(t) = x(t).Again by lim Xn(t) = x(t), wehaveFm(e) ~ Pm(e).
~
00
U F:.O(e) ~ G(e). On the other hand, for

P:.O(e) and so m=l
.
00
.
any closed set F, (F- P) is a non-dense closed set. Thus X- U F:,. (e)
m=l
.
00
= U (F m (e) - F:,. (e)) is a set of the first category. Thus its subset
m=l
G(e)c =X- G(e} is also a set of the first category. Therefore the set
of all the points of discontinuity of the function x(t), which is expressible
Thus F:.O(e)
00
n G(l/n)
n=l
00
U G(l/nf, is a set of the first category.

n=l
Theorem. A subset M of a complete metric space X is relatively compact iff it is totally bounded in the sense that, for every e > 0, there exists
a finite system of points ~. m2 , , mn of M such that every point m of
M has a distance< e from at least one of mv m2 , , mn- In other words,
M is totally bounded if, for every e > 0, M can be covered by a finite
system of spheres of radii< e and centres EM.
as X-
Proof. Suppose M is not totally bounded. Then there exist a positive

number e and an infinite sequence {mn} of points E M such that d (mi, mj)
> e fori -=1= 7'. Then, if we cover the compact set Ma by a system of open
spheres of radii < e, no finite subsystem of this system can cover Ma.
For, this subsystem cannot cover the infinite subset {m.} ~ M ~ Ma.
Thus a relatively compact subset of X must be totally bounded.
14
0. Preliminaries
Suppose, conversely, that M is totally bounded as a subset of a complete metric space X. Then the closure M" is complete and is totally
bounded with M. We have to show that M" is compact. To this purpose,
we shall first show that any infinite sequence {p,.} of M" contains a subsequence {Pcnr} which converges to a point of M". Because of the total
boundedness of M, there exist, for any e > 0, a point q. EM" and a subsequence {P,.} of {p,.} such that d (p,.., q.) < e/2 for n = 1, 2, ... ; consequently, d(p,.., p,...) < d(p,.., q.) + d(q., p,...) < e for n, m = 1, 2, ... We
set e = 1 and obtain the sequence {Po}, and then apply the same reasoning as above with e = 2-1 to this sequence {p..}. We thus obtain a
subsequence {p,...} of {p,..} such that
d (p,.., p,...)
<
1, d (p,... , p,..,)
<
1/2
(n, m = 1, 2, ... ) .
By repeating the process, we obtain a subsequence {p,.(k+I>} of the sequence
{p,.(kl} such that
d (p,.(~o+I>, p,..(k+I>)
<
1f2k
(n, m = 1, 2, ... ) .
Then the subsequence {Pc..r} of the original sequence {p,.}, defined by the
diagonal method :
Pcnl=P,.(n>,
surely satisfies
lim
n,m-->00
d (Pcnl' Pcmy) = 0. Hence, by the completeness of
M", there must exist a point
p E Ma such that lim d (p(nl' p)

n-+00
= 0.
We next show that the set Ma is compact. We remark that there

exists a countable family {F} of open sets F of X such that, if U is any
open set of X and x E U (\ M", there is a set FE {F} for which x E F ~ U.
This may be proved as follows. Ma being totally bounded, M" can be
covered, for any s > 0, by a finite system of open spheres of radii e
and centres E Ma. Letting e = 1, 1/2, 1/3, ... and collecting the countable family of the corresponding finite systems of open sphere's, we
obtain the desired family {F} of open sets.
Let now {U} be any open covering of Ma. Let {F*} be the subfamily
of the family {F} defined as follows: F ~ {F*} iff F ~ {F} and there is
some U E {U} with F ~ U. By the property of {F} and the fact that
{U} covers M", we see that this countable family {F*} of open sets covers
M". Now let {U*} be a subfamily of {U} obtained by selecting just one
U E {U} such that F ~ U, for each FE {F*}. Then {U*} is a countable
family of open sets which covers Ma. We have to show that some finite
subfamily of {U*} covers M". Let the sets in {U*} be indexed as U1 ,
U2 ,
Suppose that, for each n, the finite union .U U; fails to cover

j=l
Ma. Then there is some point x,. E (M- U uk). By what was proved
k=l
15
3. Measure Spaces
above, the sequence {xn} contains a subsequence {x(n)'} which converges

to a point, say X 00 , in Ma. Then X 00 E UN for some index N, and so
Xn E UN for infinitely many values of n, in particular for ann> N. This
contradicts the fact that
Xn
was chosen so that
Xn
Hence we have proved that 111a is compact.
(M- uuk).
k=l
3. Measure Spaces
Measures
Definition. Let 5 be a set. A pair (5, 58) is called a a-ring or a aadditive family of sets ~ 5 if 58 is a family of subsets of 5 such that
5E 58,
(1)
BE 58 implies Be = (5- B) E 58,

Bi E 58 (j
00
= 1, 2, ... ) implies that .U Bi E 58 (a-additivity).

J=l
(2)
(3)
Let (5, 58) be a a-ring of sets E 5. Then a triple (5, 58, m) is called a
measttre space if m is a non-negative, a-additive measure defined on 58:
m (B) > 0 for every B E 58 ,
mC~ Bi) = i~ m(Bj)
(4)
for any disjoint sequence {Bj} of sets E 58
(countable- or a-additivity of m),
(5)
S is expressible as a countable union of sets Bi E 58 such that m (Bj)

(j = 1, 2, ... ) (a-finiteness of the measure space (5, 58, m)).
(6)
This value m (B) is called the m-measure of the set B.
<=
Measurable Functions
Definition. A real- (or complex-) valued function x (s) defined on 5 is
said to be 58-measurable or, in short, measurable if the following condition
is satisfied:
For any open set G of the real line R 1 (or complex
(7)
plane C1), the set {s; x(s) E G} belongs to 58.
It is permitted that x (s) takes the value=
Definition. A property P pertaining to points s of 5 is said to hold malmost everywhere or, in short m-a. e., if it holds except for those s which
form a set E 58 of m-measure zero.
A real- (or complex-) valued function x(s) defined m-a.e. on 5 and
satisfying condition (7) shall be called a 58-measttrable function defined
m-a.e. on 5 or, in short, a 58-measurable function.
16
0. Preliminaries
Egorov's Theorem. If B is a \8-measurable set with m (B) < oo and if

{/n(s)} is a sequence of \8-measurable functions, finite m-a. e. on B, that
converges m-a. e. on B to a finite \8-measurable function /(s), then
there exists, for each e > 0, a subset E of B such that m(B- E) < e
and on E the convergence of In (s) to f (s) is uniform.
Proof. By removing from B, if necessary, a set of m-measure zero,
we may suppose that on B, the functions fn(s) are everywhere finite, and
converge to /(s) on B.
The set Bn =
Bn
Bk if n
<
00
l/(s) -/k(s) I< e} is \8-measurable and
n {s E B;
k=n+l
k. Since lim fn(s) = /(s) on B, we have B =

n--+00
Thus, by the a-additivity of the measure m, we have
00
U Bn-
n=l
+ (B2 - B 1) + (B3 - B 2) + }
m (B 1 ) + m (B 2 - B 1 ) + m (B3 - B 2) +
m(B) = m{B1
=m(B1 )
(m(B 2)-m(B 1))
(m(B 3)-m(B 2))
lim m(Bn).
-00
Hence lim m (B- Bn) = 0, and therefore, from a sufficiently large k0

n--700
on, m(B- Bk) < 1J where 1J is any given positive number.

Thus there exist, for any positive integer k, a set Ck ~ B such that
m(Ck) < ef2k and an index Nk such that
lf(s) -fn(s) I< 1/2k for n
>
Nk and for s E B-
ck.
00
Let us set E = B- U Ck. Then we find

k=l
00
m(B- E)< I
k=l
00
m(Ck) < I
k=l
ef2k = e,
and the sequence /n(s) converges uniformly on E.

Integrals
Definition. A real- (or complex-) valued function x (s) defined on S
is said to be finitely-valued if it is a finite non-zero constant on each of
a finite number, say n, of disjoint \8-measurable sets Bi and equal to zero
n
on S- .U Bi.
J=l
Let the value of x(s)
on Bi be denoted by xi.
Then x (s) ism-integrable or, in short, integrable overS if

n
and the value .I xi m (Bj) is defined as the

J=l
.I Ixi Im (Bj) < oo,
J=l
integral of x (s) over S with
respect to the measure m; the integral will be denoted by
Jx(s) m(ds)
17
3. Measure Spaces
x(s) or simply by x(s) when no confusion can be

s
expected. A real- (or complex-) valued function x(s) defined m-a. e. on 5 is
said to be m-integrable or, in short, integrable over 5 if there exists a
sequence {xn(s)} of finitely-valued integrable functions converging to
x (s) m-a. e. and such that
or, in short, by
lim
J Jxn(s)- xk(s)l m(ds) =
n,~s
0.
x,. (s) m (ds) exists and the value

s
of this limit is independent of the choice of the approximating sequence
x (s) m (ds) over 5 with respect to the
{ Xn (s) }. The value of the integral
s
xn(s) m(ds). We shall
measure m is, by definition, given by lim
,.._.,.00 s
sometimes abbreviate the notation x(s) m (ds) to x(s) m(ds) or to
s
x(s).
Properties of the Integral
It is then proved that a finite lim
,.._.,.00
i) If x(s) and y(s) are integrable, then <Xx(s) + {Jy(s) is integrable

and (<Xx(s) + {Jy(s)) m(ds) =<X x(s) m(ds) + {3 y(s)m(ds).
s
s
s
ii) x (s) is integrable iff Ix (s) I is integrable.
x(s) m(ds) > 0,
iii) If x(s) is integrable and x(s) > 0 a. e., then
s
and the equality sign holds iff x(s) = 0 a. e.
X (s) m (ds) is
iv) If X (s) is integrable, then the function X (B) =
i3
a-additive, that is, X
c~ Bi) = i~ X (Bi)
sequence {Bj} of sets E ~.Here
for any disjoint
J x (s) m (ds) = SJ C
(s) x(s) m (ds),
where Cn(s) is .the defining function of the set B, that is,

C B (s) = 1 for s E B
and C B (s) = 0 for s E 5- B.
v) X (B) in iv) is absolutely continuous with respect to m in the sense

that m (B) = 0 implies X (B) = 0. This condition is equivalent to
the condition that lim X (B) = 0 uniformly in BE ~m(B)->-0
The Lebesgue-FatouLemma. Let {xn(s)} be a sequence of real-valued

integrable functions. If there exists a real-valued integrable function
x(s) such that x(s) 2: xn(s) a. e. for n = 1, 2, ... (or x(s) ~ xn(s) a. e.
for n = 1, 2, ... ), then
J (lim Xn (s)) m (ds) > lim sJXn (s) m (ds)

(or [ (!!_: x (s)) m(ds) < ~ [ Xn(s) m(ds)),
s
n->-00
n->-00
11
18
0. Preliminaries
under the convention that if~ Xn(s) (or!!, Xn(s)) is not integrable,
weunderstandthat lim
JXn(s) m(ds) =-oo(or lim Jxn(s)m(ds) =oo).
~s
~s
Definition. Let (5, 58, m) and (5', 58', m') be two measure spaces. We
denote by 58 X 58' the smallest a-ring of subsets of 5 X 5' which contains
all the sets of the form B X B', where B E 58, B' E 58'. It is proved that
there exists a uniquely determined a-finite, a-additive and non-negative
measure m X m' defined on 58 X 58' such that
(mxm') (BxB') = m(B) m'(B').
mxm' is called the product measure of m and m'. We may define the
58x58'-measurable functions x(s, s') defined on 5x5', and the mxm'integrable functions x (s, s'). The value of the integral over 5 X 5' of an
m X m'-integrable function x (s, s') will be denoted by
JJx(s, s') (mXm') (ds ds')
sxs
JJx(s, s') m(ds) m' (ds').
or
sxs
The Fubini-Tonelli Theorem. A 58 X 58'-measurable function x(s, s') is
m X m' -integrable over 5 X 5' iff at least one of the iterated integrals
JJ
J Jx(s, s')/ m(ds)}m'(ds') and

sls
is finite; and in such a case we have
JJ x(s, s') m(ds) m' (ds')
SxS'
=
=
J{Js Jx(s, s')/ m'(ds')}m(ds)
J{J x(s, s') m(ds) 1m'

(ds')
f
S'
f{J x(s, s') m'(ds')lm(ds).

l
Topological Measures
Definition. Let 5 be a locally compact space, e.g., ann-dimensional
Euclidean space Rn or a closed subset of Rn. The Baire subsets of 5 are
the members of the smallest a-ring of subsets of 5 which contains every
compact G6 -set, i.e., every compact set of X which is the intersection of
a countable number of open sets of 5. The Borel subsets of 5 are the
members of the smallest a-ring of subsets of 5 which contains every
compact set of 5.
If 5 is a closed subset of a Euclidean space Rn, the Baire and the Borel
subsets of 5 coincide, because in Rn every compact (closed bounded)
set is a G6-set. If, in particular, 5 is a real line R 1 or a closed interval on
R 1 , the Baire (=Borel) subsets of 5 may also be defined as the members
of the smallset a-ring of subsets of 5 which contains half open intervals
(a, b].
Definition. Let 5 be a locally compact space. Then a non-negative Baire
(Borel) measure on 5 is a a-additive measure defined for every Baire
19
3. Measure Spaces
(Borel) subset of S such that the measure of every compact set is finite.
The Borel measure miscalled regular if for each Borel set B we have
m(B) = inf m(U)
u;;,s
where the infimum is taken over all open sets U containing B. We may
also define the regularity for Baire measures in a similar way, but it
turns out that a Baire measure is always regular. It is also proved that
each Baire measure has a uniquely determined extension to a regular
Borel measure. Thus we shall discuss only Baire measures.
Definition. A complex-valued function f(s) defined on a locally
compact space S is a Baire function on S if j-1 (B) is a Baire set of S for
every Baire set B in the complex plane C1 . Every continuous function
is a Baire function if S is a countable union of compact sets. A Baire
function is measurable with respect to the a-ring of all Baire sets of S.
The Lebesgue Measure
Definition. Suppose S is the real line R 1 or a closed interval of R 1 .
Let F (x) be a monotone non-decreasing function on S which is continuous
from the right: F (x) = inf F (y). Define a function m on half closed
x<y
intervals (a, b] by m ((a, b]) = F (b)- F (a). This m has a uniquely determined extension to a non-negative Baire measure on S. The extended
measure m is finite, i.e., m(S) <=iff F is bounded. If m is the Baire
measure induced by the function F(s) = s, then miscalled the Lebesgue
measure. The Lebesgue measure in Rn is obtained from then-tuple of the
one-dimensional Lebesgue measures through the process of forming the
product measure.
Concerning the Lebesgue measure and the corresponding Lebesgue
integral, we have the following two important theorems:
Theorem 1. Let M be a Baire set in Rn whose Lebesgue measure IM I
is finite. Then, if we denote by B e C the symmetric difference of the
set B and C: B e C = B V C- B (\ C, we have
lim
lhl-+0
I(M
+h) 8 M
Here m + h
(hl, ... 'hn) and
I= 0, where M
+ h
{xE Rn; x
(m1 + hv ... , mn + hn) for m
IhI =
C.Ji hjr
=m+
(m1 ,
... ,
h, mE M}.
mn), h
Theorem 2. Let G be an open set of Rn. For any Lebesgue integrable

function f (x) in G and s > 0, there exists a continuous function C, (x) in
G such that {xE G; C,(x) #-Otis a compact subset of G and
Jlf(x)- C,(x) Idx <
2*
s.
20
0. Preliminaries
Remark. Let m be a Baire measure on a locally compact space S.

A subset Z of S is called a set of m-measure zero if, for each e > 0, there
is a Baire set B containing Z with m(B) <e. One can extend m to the
class of m-measurable sets, such a set being one which differs from a
Baire set by a set of m-measure zero. Any property pertaining to a
set of m-measure zero is said to hold m-almost everywhere (m-a. e.).
One can also extend integrability to a function which coincides m-a. e.
with a Baire function.
4. Linear Spaces
Linear Spaces
Definition. A set X is called a linear space over a field K if the
following conditions are satisfied:
X is an abelian group (written additively),

A scalar multiplication is defined: to every element!
x E X and each iX E K there is associated an element of
X, denoted by iXX, such that we have
+ y) =
(iX + {J) x =
iX (x
+ iXY
iXX + {Jx
iXX
(iX E K; x, Y E X),
(1)
(2)
(iX, {J E K; x EX),
(iX{J) x = iX ({Jx) (iX, {J E K; x EX),
1x
x (1 is the unit element of the field K).
In the sequel we consider linear spaces only over the real number
field R 1 or the complex number field C1 . A linear space will be said to be
real or complex according as the field K of coefficients is the real number
field R 1 or the complex number field C1 . Thus, in what follows, we mean
by a linear space a real or complex linear space. We shall denote by
Greek letters the elements of the field of coefficients and by Roman
letters the elements of X. The zero of X (= the unit element of the
additively written abelian group X) and the number zero will be denoted
by the same letter 0, since it does not cause inconvenience as 0 x =
(iX- iX) x = iX x - iX x = 0. The inverse elementofthe additivelywritten
abeliangroupXwill be denoted by-x; it is easy to see that-x = (-1)x.
Definition. The elements of a linear space X are called vectors (of X).
The vectors xv x 2 , , Xn of X are said to be linearly independent if the
n
equation ..I iXJ x1 = 0 implies iX1

J=l
iX2 =
0. They are linearly
dependent if such an equation holds where at least one coefficient is

different from 0. If X contains n linearly independent vectors, but every
system of (n + 1) vectors is linearly dependent, then X is said to be of
21
4. Linear Spaces
n-dimension. If the number of linearly independent vectors is not finite,

then X is said to be of infinite dimension. Any set of n linearly independent vectors in an n-dimensional linear space constitutes a basis for X
and each vector x of X has a unique representation of the formx= ..I()(.; Y;

J=l
in terms of the basis Yv y2 , , y,.. A subset M of a linear space X is

called a linear subspace or, in short, a subspace, if whenever x, y EM,
the linear combinations ()(.X+ {Jy also belong toM. M is thus a linear
space over the same coefficient field as X.
Linear Operators and Linear Functionals
Definition. Let X, Y be linear spaces over the same coefficient field
K. A mapping T : x-+ y = T (x) = T x defined on a linear subspace D
of X and taking values in Y is said to be linear, if
T(()(.x 1 + {Jx2) = ()(.(Tx 1)
+ {J(Tx2).
The definition implies, in particular,

TO=O, T(-x)=-(Tx).
We denote
D =D(T), {YE Y;y= Tx,xED(T)} =R(T),{xED(T); Tx= 0} =N(T)
and call them the domain, the range and the null space of T, respectively.
T is called a linear operator or linear transformation on D (T) ~ X into
Y, or somewhat vaguely, a linear operator from X into Y. If the range
R (T) is contained in the scalar field K, then Tis called a linear functional
on D (T). If a linear operator T gives a one-to-one map of D (T) onto
R (T), then the inverse map T-1 gives a linear operator on R (T) onto
D(T):
T- 1Tx = x for xE D(T) and T T- 1 y = y for yE R(T).
T-1 is the inverse operator or, in short, the inverse of T. By virtue of
T (x1 - x2) = T x 1 - T x2 , we have the following.
Proposition. A linear operator T admits the inverse T- 1 iff T x = 0'
implies x = 0.
Definition. Let T 1 and T 2 be linear operators with domains D(T1)
and D (T2) both contained in a linear space X, and ranges R (T1 ) and
R(T2) both contained in a linear space Y. Then T 1 = T 2 iff D(T1) =
D(T2) and T 1x = T 2 x for all xE D(T1) = D(T2). If D(T1) ~ D(T2) and
T 1x = T 2 x for all xE D(T1), then T 2 is called an extension of Tv and T 1
a restriction of T 2 ; we shall then write T 1 ~ T 2
Convention. The value T (x) of a linear functional T at a point
xED (T) will sometimes be denoted by <x, T), i.e.
T(x) = (x, T).
22
0. Preliminaries
Factor Spaces
Proposition. Let M be a linear subspace in a linear space X. We say
that two vectors x1 , x2 E X are equivalent modulo M if (x1 - x2) E M and
write this fact symbolically by x 1
x2 (mod M). Then we have:
(i) x _ x (mod M) ,
(ii) if x1 = x 2 (mod M), then x2
(iii) if x1
x1 (mod M),
x3 (mod M), then x1 = x3 (mod M) .
x2 (mod M) and x2
Proof. (i) is clear since x- x = 0 EM. (ii) If (x1 - x2) EM, then
(x2 - x 1) = - (x1 - x 2) EM. (iii) If (x1 - x2) EM and (x2 -- x3 ) EM,
then (x1 - x3 ) = (x1 - x2) + (x 2 - x3) EM.
We shall denote the set of all vectors E X equivalent modulo M to a
fixed vector x by ~x Then, in virtue of properties (ii) and (iii), all vectors in
~x are mutually equivalent modulo M. ~x is called a class of equivalent
(modulo M) vectors, and each vector in ~xis called a representative of the
class ~x Thus a class is completely determined by any one of its representatives, that is, y E ~x implies that ~Y = ~x Hence, two classes ~"' ~Y
are either disjoint (when y ~~,.)or coincide (when y E ~,.).Thus the entire
space X decomposes into classes ~x of mutually equivalent (modulo M)
vectors.
Theorem. We can consider the above introduced classes (modulo M)
as vectors in a new linear space where the operation of addition of classes
and the multiplication of a class by a scalar will be defined through
;x
+ ;y =
;x+y CX~x = ~ax
Proof. The above definitions do not depend upon the choice of representatives x, y of the classes ~"' ~Y respectively. In fact, if (x1 - x) E M,
(y1 - y) EM, then
(x1
+ y 1) -
(x
+ y) =
(x1 - x)
+ (y1 -
y) E M,
(ax1 - ax)= a(x1 - x) EM.
We have thus proved ~x,+y, = ~x+y and ~ax,= ~ax, and the above definitions of the class addition and the scalar multiplication of the classes are
justified.
Definition. The linear space obtained in this way is called the factor
space of X modulo M and is denoted by XfM.
References
Topological Space: P.
J. L. KELLEY
ALEXANDROFF-H. HoPF
[1], N.
[1].
Measure Space: P.R.
HALMOS
[1],
S. SAKS
[1].
BouRBAKI
[1],
1. Semi-norms and Locally Convex Linear Topological Spaces
23
I. Semi-norms
The semi-norm of a vector in a linear space gives a kind of length for
the vector. To introduce a topology in a linear space of infinite dimension
suitable for applications to classical and modern analysis, it is sometimes
necessary to make use of a system of an infinite number of semi-norms.
It is one of the merits of the Bourbaki group that they stressed the
importance, in functional analysis, of locally convex spaces which are
defined through a system of semi-norms satisfying the axiom of separation. If the system reduces to a single semi-norm, the corresponding linear space is called a normed linear space. If, furthermore, the
space is complete with respect to the topology defined by this seminorm, it is called a Banach space. The notion of complete normed linear
spaces was introduced around 1922 by S. BANACH and N. WIENER independently of each other. A modification of the norm, the quasi-norm in
the present book, was introduced by M. FRJ~CHET. A particular kind of
limit, the inductive limit, of locally convex spaces is suitable for discussing
the generalized functions or the distributions introduced by L. ScHWARTZ,
as a systematic development of S. L SoBOLEV's generalization of the
notion of functions.

As was stated in the above introduction, the notion of semi-norm is of
fundamental importance in discussing linear topological spaces. We shall
begin with the definition of the semi-norm.
Definition 1. A real-valued function p (x) defined on a linear space X
is called a semi-norm on X, if the following conditions are satisfied:
p (x
+ y) <
p (x)
+ p (y)
p(cxx) =
(1)
(subadditivity),
lcxl p(x).
(2)
Example 1. The n-dimensional Euclidean space Rn of points x =

(x 1 , . . , xn) with coordinates x 1 , x 2 , . , xn is an n-dimensional linear
space by the operations:
(xl + Yv X2 + Y2, , Xn + YnL
cxx = (cxxv cxx2 , . , cxxn).
In this case p (x) = max lxj I is a semi-norm. As will be proved later,
X
+Y=
P (x) = CillxilqY'q
;:~:;;; > 1 is also a semi-norm on Rn.
Proposition 1. A semi-norm
p(x)
satisfies
(3)
p(O)=O,
P(xl-x2 )
> IP(x 1) -p(x2 ) !, in
particular, p(x)
>
0.
(4)
24
I. Semi-norms
Proof. p (0) = p (0 x) = 0 p (x) = 0. We have, by the subadditivity,

p (x1 - x2) + p (x2) > p (x1) and hence p (x1 - x2) > p (x1) - p (x2). Thus
P(x1 - x2) = I-ll P(x2 - x1) > P(x2) - p(x1) and so we obtain (4).
Proposition 2. Let p (x) be a semi-norm on X, and c any positive
number. Then the set M = {x E X; p (x) < c} enjoys the properties:
M30,
(5)
M is convex: x, y EM and 0 <<X< 1 implies

<XX
(1- <X) y E M,
(6)
M is balanced (equilibre in Bourbaki's terminology):
xE M and I<XI
< 1 imply <XXE M,
(7)
M is absorbing: for any x EX, there exists
<X
>
such that <X-1 x EM,
p (x) =
inf
a>O,x- 1xEM
<X c
(8)
(in
infimum
the greatest lower
bound).
(9)
Proof. (5) is clear from (3). (7) and (8) are proved by (2). (6) is proved
by the subadditivity (1) and (2). (9) is proved by observing the equivalence of the three propositions below:
[<X- 1 x EM]~ [p(<X- 1 x)
< c];;? [p(x) <<XC].
Definition 2. The functional
PM (x) =
in
cx>O,a.- 1xEM
(9')
<X
is called the Minkowski functional of the convex, balanced and absorbing

set M of X.
Proposition 3. Let a family {Pr (x); y E F} of semi-norms of a linear
space X satisfy the axiom of separation:
For any x0 # 0, there exists Pr. (x) in the family such

that Pr. (x 0 ) # 0.
(10)
Take any finite system of semi-norms of the family, say Pr (x), Pr. (x), ... ,
. , P'Yn (x) and any system of n positive numbers 81> 8 2 , , 8n, and set
1
{xE X; Py;(x) <
8j
(j = 1, 2, ... , n)}.
(11)
U is a convex, balanced and absorbing set. Consider such a set U as a

neighbourhood of the vector 0 of X, and defined a neighbourhood of any
vector x 0 by the set of the form
x0
U = {y E X; y = x0
+ u, u E U}.
(12)
25
Consider a subset G of X which contains a neighbourhood of each of its

point. Then the totality {G} of such subsets G satisfies the axiom of open
sets, given in Chapter 0, Preliminaries, 2.
Proof. We first show that the set G0 of the form G0 = {xE X; p"' (x) < c}
is open. For, let x0 E G0 and py(x0 ) = {3 <c. Then the neighbourhood
of x0 ,x0 + U where U={xEX;py(x) < 2- 1 (c-{3)}, iscontainedinG0 ,
because u E U implies py (x 0 + u) < pY (x0 ) + py (u) < {3 + (c- {3) = c.
Hence, for any point x 0 EX, there is an open set x 0 + G0 which contains x 0 It is clear, by the above definition of open sets, that the union
of open sets and the intersection of a finite number of open sets are also
open.
Therefore we have only to prove Hausdorff's axiom of separation:
If x1 =;I= x 2 , then there exist disjoint open sets G1 and G2 such that
(13}
x1 E Gv x 2 E G2 .
In view of definition (12) of the neighbourhood of a general point x 0 ,

it will be sufficient to prove (13) for the case x1 = 0, x2 =;I= 0. We choose,
by (10), Py, (x) such that Py, (x2) = iX > 0. Then G1 = {x EX; Py, (x) < iX/2}
is open, as proved above. Surely G1 3 0 = x1 . We have to show that G1
and G2 = x 2 + G1 have no point in common. Assume the contrary and
let there exist a y E G1 (\ G2 . y E G2 implies y = x 2 + g = x 2 - (-g) with
some g E G1 and so, by (4), Py, (y) 2 Py, (x2 ) - p (-g) > i X - 2- 1 iX = iX/2,
because -g belongs to G1 with g. This contradicts the inequality
P"'z (y) < iX/2 implied by y E G1 .
Proposition 4. By the above definition of open sets, X is a linear
topological space, that is, X is a linear space and at the same time a
topological space such that the two mappings X X X__,.. X: (x, y) __,.. x + y
and KxX __,..X: (iX, x) __,.. iXX are both continuous. Moreover, each seminorm p"' (x) is a continuous function on X.
Proof. For any neighbourhood U of 0, there exists a neighbourhood
V of 0 such that
V
+V=
{w E X;
= v1
+ v2 where v1 , v2 E V} ~
U,
since the semi-norm is subadditive. Hence, by writing
(x
+ y) -
(x0
+ y0) =
(x- x0)
+ (y -Yo),
we see that the mapping (x, y) __,.. x + y is continuous at x = x0 , y = y0

For any neighbourhood U of 0 and any scalar iX =;I= 0, the set iX U =
{xE X; x = iXU, uE U} is also a neighbourhood of 0. Thus, by writing
iXX-
iX0 X 0 =
iX
(x- x0 )
(iX- 1X0 ) x0 ,
we see by (2) that (iX, x) __,.. iXX is continuous at iX = iX0 , x = x0

The continuity of the semi-norm py (x) at the point x = x 0 is proved
by !Py(x) -py(x0 ) I< py(x-x0 ).
26
I. Semi-norms
Definition 3. A linear topological space X is called a locally convex,

linear topological space, or, in short, a locally convex space, if any of its
open sets 3 0 contains a convex, balanced and absorbing open set.
Proposition 5. The Minkowski functional pM (x) of the convex, balanced and absorbing subset M of a linear space X is a semi-norm on X.
Proof. By the convexity of M, the inclusions
imply
PM(x)
xf(PM(x) +e) EM, yf(PM(Y) +e) EM for any e > 0
x
PM(Y) + c
Y .
M
+ 2c PM(x) + c + PM(x) + PM(Y) + 2c PM(Y) +eE
PM(x + y) ~ PM(x) + PM(Y) + 2e. Since e > 0 was arbitrary,
p;yi(x)
c
PM(Y)
and so
we obtain the subadditivity of PM(x). Similarly we obtain PM(1Xx) =
[IX pM (x) since M is balanced.
We have thus proved
Theorem. A linear space X, topologized as above by a family of seminorms py (x) satisfying the axiom of separation (10), is a locally convex
space in which each semi-norm py (x) is continuous. Conversely, any
locally convex space is nothing but the linear topological space, topologized as above through the family of semi-norms obtained as the Minkowski functionals of convex balanced and absorbing open sets of X.
Definition 4. Let f(x) be a complex-valued function defined in an open
set.Q of Rn. By the support (or carrier) off, denoted by supp(/), we mean
the smallest closed set (of the topological space .Q) containing the set
{ x E .Q; f (x) =I= 0}. It may equivalently be defined as the smallest closed
set of .Q outside which f vanishes identically.
Definition 5. By Ck (.Q), 0 < k < oo, we denote the set of all complexvalued functions defined in .Q which have continuous partial derivatives
of order up to and including k (of order< oo if k = oo). By C~(.Q), we
denote the set of all functions E Ck (.Q) with compact support, i.e., those
functions E Ck (.Q) whose supports are compact subsets of .Q. A classical
example of a function E Cg" (Rn) is given by
J
f(x)
exp((lxl 2 -1)- 1) for lxl
= 0 for Ix I > 1.
l(xv .. . , Xn)l
=C~ xJY' 2 <
1, (14)
The Space ~ (.Q)

Ck (.Q) is a linear space by
(/1 + /2) (x) = /1(x) + /2(x), (1X/) (x) = IX/ (x).

For any compact subset K of .Q and any non-negative integer m
(m < oo when k = oo), we define the semi-norm
PK,m(f) =
sup
isl;ii;m,xEK
ID 5 /(x)
I,
/E Ck(.Q),
< k
27
where sup= supremum= the least upper bound and
Is I =
(sv Sz, ... , sn) I =
.I" Sj.
;~1
Then Ck (.Q) is a locally convex space by the family of these semi-norms.

We denote this locally convex space by (Zk (.Q). The convergence
the uniform convergence
lim lh = I in this space czk (.Q) is exactly
"'
h---7>00
lim Dslh(x) = Dsl(x) on every compact subset K of .Q, for each s
k-+00
with lsi< k (lsi<= if k = =).

Proposition 6. (Zk (.Q) is a metric space.
Proof. Let K 1 s;; K 2 s;; ~ Kn ~ be a monotone increasing
00
U Kn. Define, for

sequence of compact subsets of .Q such that Q = n~1
each positive integer h, the distance
k
dh (!,g) = ~O 2-m PK,.,m (f- g) (1

Then the convergence lim
S->00
00
Is
I in
(Zk (.Q)
d(f, g)= .l: 2-h dh(f, g). (1

h~1
+ PK11,m (f- g))- 1

is defined by the distance
+ dh(f, g))- 1 .
We have to show that dh(f, g) and d(f, g) satisfy the triangle inequality.
The triangle inequality for dh (!, g) is proved as follows: by the subadditivity of the semi-norm PK11 ,m (f), we easily see that dh(f, g)=
satisfies the triangle inequality dh(f, g)< dh(l, k) + dh(k, g), if we can
prove the inequality
+ i(X--pJ)-1 < [tX-y[ (1 + [tX-y[)-1

+ Jy -- p (1 + Jy _ p [)-1
for complex numbers tX, p andy; the last inequality is clear from the
equality valid for any system of non-negative numbers 1X, p and y:
(1X + Pl (1 + 1X + p)-1 < 1X(1 + 1X)-1 + /J(1 + m-1.
[1X-PI (1
m-
The triangle inequality ford(!, g) may be proved similarly.

Definition 6. Let X be a linear space. Let a family {X"} of linear
subspaces XIX of X be such that X is the union of X,.'s. Suppose that each
XIX is a locally convex linear topological space such that, if X,., ~ X"'
then the topology of X,, is identical with the relative topology of XIX,
as a subset of Xcx, We shall call "open" every convex balanced and
absorbing set U of X iff the intersection U f\ X,. is an open set of XIX
containing the zero vector 0 of X"', for all XIX. If X is a locally convex
28
I.
Semi-norms
linear topological space whose topology is defined in the stated way.

then X is called the (strict) inductive limit of X,,/s.
Remark. Take, from each XIX, a convex balanced neighbourhood U"'
of 0 of X"'. Then the convex closure U of the union V = U UIX, i.e.,
IX
U = {u EX; u =
.i {Jj Vj, vj E V, {Jj > 0 (i =
J=l
with arbitrary finite
n}
1, 2, ... , n),
,i {Jj =
J=l
surely satisfies the condition that it is convex balanced and absorbing in

such a way that U f\ X"' is a convex balanced neighbourhood of 0 of XIX.
for all X"'. The set of all such U's corresponding to an arbitrary choice of
U"''s is a fundamental system of neighbourhood of 0 of the (strict) inductive
limit X of
s, i.e., every neighbourhood of 0 of the (strict) inductive
limit X of
s contains one of the U's obtained above. This fact justifies
the above definition of the (strict) inductive limit.
x:
x:
The Space ~(D)
e3" (.Q)
is a linear space by
(/1
+ /2) (x) = /1 (x) + /2 (x),
(ex f) (x)
= ex f (x).
For any compact subset K of .Q, let ~K(.Q) be the set of all functions
~ K. Define a family of semi-norms on
':I)K(.Q) by
f E ego (.Q) such that supp (/)

PK,m(f) =
sup
IDf(x)
II;S;m,~EK
1. where m <
oo.
'l)K (.Q)
is a locally convex linear topological space, and, if K 1 ~ K 2

then it follows that the topology of 'DK, (.Q) is identical with the relative
topology of 'DK, (.Q) as a subset of 'DK, (.Q). Then the (strict) inductive
limit of 'DK(.Q)'s, where K ranges over all compact subsets of .Q, is a
locally convex, linear topological space. Topologized in this way, ego (.Q}
will be denoted by 'l) (.Q). It is to be remarked that,
P(f) =
sup
:<ED
If (x) I
is one of the semi-norms which defines the topology of 'l) (.Q). For, if we
set u = {/ E ego (.Q) ; p (f) < 1}' then the intersection u (\ '1) K (.Q) is given
by UK={/ E ~K(.Q); PK(f) =sup l/(x) I< 1}.
:<EK
Proposition 7. The convergence lim j,. = 0 in '1) (.Q) means that the
h-->oo
following two conditions are satisfied: (i) there exists a compact subset
K of .Q such that supp {/~s) ~ K (h = 1, 2, ... ) , and (ii) for any differential
operator n, the sequence {Dj,.(x)} converges to 0 uniformly on K.
29
Proof. We have only to prove (i). Assume the contrary, and let there
exist a sequence {x<k}} of points E Q having no accumulation points in
Q and a subsequence {lh; (x)} of {lh (x)} such that lh; (x<il) =/=- 0. Then the
semi-norm
00
P(l) =;!; 2
k~1
sup
xEKk-Kk_,
il(x)/lhk(x<kl) j, where the mono-
tone increasing sequence of compact subsets Ki of

Q
satisfies
00
U Ki = Q
j~1
(k = 1, 2, ... ) , K 0
and
x<k) E Kk - Kk_ 1
=0
defines a neighbourhood U = {I E C~ (Q) ; p (f) < 1} of

However, none of the lhk's is contained in U.
Corollary. The convergence lim
h-+OO
lh
of :tl (Q).
=I in ;tl (Q) means that the
following two conditions are satisfied: (i) there exists a compact subset
K ofQ such that supp (Ih) s;; K (h = 1, 2, ... ) , and (ii) for any differential
operator D 5 , the sequence Dslh(x) converges to D 5 l(x) uniformly on K.
Proposition 8 (A theorem of approximation). Any continuous function
j E cg (R") can be approximated by functions of C~ (Rn) uniformly on R".
Proof. Let 81 (x) be the function introduced in (14) and put
Oa(x) = h-;; 1 Ot(xfa), where a> 0 and ha
J Oa(x) dx =
>
0 are such that

(15)
1.
R"
We then define the regularization fa off:
Ia (x)
f(x- y) ea (y) dy
f(y) ea (x- y) dy, where
R"
R"
(16)
The integral is convergent since f and 8,. have compact support. Moreover,
since
l(y)Oa(x-y)dy,
la(x)=
supp(/)
the support of Ia may be taken to be contained in any neighbourhood of

the supp(/) if we take a> 0 sufficiently small. Next, by differentiating
under the integral sign, we have
D5 la(x)
= D~la(x) =
J f(y) D:Oa(x-y) dy,
R"
(17)
30
I.
Semi-norms
and soIa is in C0 (Rn). Finally we have by J Oa(i- y) dy = 1,

R"
lfa(x)-f(x)l
< JJf(y)-f(x)JOa(x-y)dy
<
Jf(y)-f(x)JOa(x-y}dy
Jf(y)-f(x)JOa(x-y)dy.
Jf(y)-/(x)l ;;>s
1/(y)-/(x)l >
The first term on the right is < e; and the second term on the right
equals 0 for sufficiently small a> 0, because, by the uniform continuity
of the function f with compact support, there exists an a> 0 such that
II (y) - I (x) I > e implies Iy - x ! > a. We have thus proved our Proposition.
2. Norms and Quasi-norms

Definition I. A locally convex space is called a normed linear space,
if its topology is defined by just one semi-norm.
Thus a linear space X is called a normed linear space, if for every
xE X, there is associated a real number JJxJJ, the norm of the vector x,
such that
IJxJJ > 0 and JJxJI = 0 iff x = 0,
(1)
llx + Yll < llxll + IIYII
(triangle inequality),
JIXxll = IXIIIxll
(2)
(3)
The topology of a normed linear space X is thus defined by the distance
d(x, y) = Jlx -YII
(4)
In fact, d(x, y) satisfies the axiom of distance:

d (x, y)
> 0 and d (x, y) = 0 iff x =
d (x, y)
< d (x, z)
+ d (z, y)
y,
(triangle inequality),
d(x, y) = d(y, x).

For, d(x, y) = jjx- y II= IIY- x II= d(y, x) and d(x, y) = llx- Yll =
llx-z + z-yll < Jjx-zll + llz-yll =d(x,z) + d(z,y) by (1}, (2),
(3) and (4).
The convergence lim d (xm x) = 0 in a normed linear space X will be
denoted by s-lim Xn
~
x or simply by
xn--+
x, and we say that the se-
quence {xn} converges strongly to x. The adjective "strong" is introduced

to distinguish it from the "weak" convergence to be introduced later.
31
2. Norms and Quasi-norms
Proposition 1. In a normed linear space X, we have
lim
n---+00
llx,ll = llxll
if s-lim Xn
n---+00
= x,
(5)
s-lim <Xn x, =<XX if lim <X,= <X and s-lim x,

n---+oo
n~
n----K:"O
s-lim (x,
n~oo
+ Yn) = x + y
if s-lim x
n---KX)
(6)
x,
and s-lim Yn = y. (7)
=x
n---+00
Proof. (5), (6) and (7) are already proved, since X is a locally convex
space topologized by just one semi-norm p(x) = II x II However, we shall
give a direct proof as follows. As a semi-norm, we have
llx-yll~ lllxii-IIYIII
and hence (5) is clear. (7) is proved by II (x
(8)
+ y) - (x, + y,) I =
II<XX-<XnXnll S
From
IIY-Ynll
+
llx-xnll
<
(Y-Yn)ll
+
ll(x-x,)
II X- x, II and
I
<Xn
I
+
1\
X
II
I
<Xn
<X
I
<
II
x,
<Xn
X
<Xn
II
+
II
X
<Xn
II <X X
the boundedness of the sequence {<X,.} we obtain (6).
J
Definition 2. A linear space X is called a quasi-normed linear space,

if, for every x E X, there is associated a real number II xI J, the quasi-norm
of the vector x, which satisfies (1), (2) and
lim II<X,x\1=0
\1-xll= llxl\, "n-->0
and
lim
Jlxnll-->0
II<Xx,II=O.
(3')
Proposition 2. In a quasi-normed linear space X, we have (5), (6)

and (7).
Proof. We need only prove (6). The proof in the preceding Proposition shows that we have to prove
lim
n---+00
lim II<X x, II = 0 uniformly

IIx, II = 0 implies that n---+oo
in <X on any bounded set of <X.
(9)
The following proof of (9) is due to S. KAKUTANI (unpublished). Consider

the functional Pn (<X) = !J <X Xn II defined on the linear space R 1 of real
numbers normed by the absolute value. By the triangle inequality of
p,(<X) and (3'), p,(<X) is continuous on R 1 Hence, from lim Pn(<X) = 0
n->00
implied by (3') and Egorov's theorem (Chapter 0, Preliminaries, 3. Measure Spaces), we see that there exists a Baire measurable set A on the
real line R 1 with the property:
the Lebesgue measure IA I of A is
uniformly on A .
>
0 and lim
n->00
p, (<X) = 0
(10)
Since the Lebesgue measure on the real line is continuous with respect to
translations, we have, denoting by B e C the symmetric difference
BVC-Bf\C,
I(A + a) e A I~ 0 as a~ 0.
32
I. Semi-norms
Thus there exists a positive number a0 such that
/a/:::;;: a0 implies J(A+ a)
A J</A J/2, in particular, !(A+ a)!\ A[> 0.

Hence, for any real number a with J a [ < a0 , there is a representation
a= <X- <X' with <X E A, <X' EA.
Therefore, by Pn (a) = Pn (<X- <X')
< Pn (a)
+ Pn (<X'), we see that
lim Pn (a) = 0 uniformly in a when
n-->00
[a [ < a0
Let M be any positive number. Then, taking a positive integer k > M fa0
and remembering p,.(ka) < kpn(a), we see that (9) is true for /<X[< M.
Remark. The above proof may naturally be modified so as to apply

to complex quasi-normed linear spaces X as well.
As in thecaseofnormedlinearspaces, the convergence lim [[x-xnll =0
n-->00
in a quasi-normed linear space will be denoted by s-lim xn

n-->00
x,
or
simply by Xn ~ x; we shall then say that the sequence {xn} converges

strongly to x.
Example. Let the topology of a locally convex space X be defined by a

countable number of semi-norms Pn(x) (n = 1, 2, ... ). Then X is a
quasi-normed linear space by the quasi-norm
For, the convergence lim Pn (xh) = 0 (n = 1, 2, ... ) is equivalent to

s-lim xh
h---+00
h-oc
0 with respect to the quasi-norm /1 x
II above.
3. Examples of Normed Linear Spaces

Example I. C (S). LetS be a topological space. Consider the set C (S)
of all real-valued (or complex-valued), bounded continuous functions
x (s) defined on S. C (S) is a normed linear space by
(x
+ y)
InC (S), s-lim

to x(s).
(s) = x(s)
Xn
n---+00
+ y(s),
(<Xx) (s) = <Xx(s), llxl/ = supx(s).

sES
= xmeans the uniform convergence of the functionsxn (s)
Example 2. LP(S, )S, m), or, in short, LP(S) (1 < p <=).Let LP(S)
be the set of all real-valued (or complex-valued) 5S-measurable functions
x(s) defined m-a. e. on S such that [x(s) JP ism-integrable overS. LP (S) is
a linear space by
(x
+ y)
(s)
x(s)
+ y(s),
(ax) (s)
<Xx(s).
33
For, (x(s) + y(s)) belongs to Li'(S) if x(s) and y(s) both belong to U(S),
as may be seen from the inequality lx(s) + y.(s)!P < 2P (I x(s)IP + iy(s)iP).
We define the norm in U(S) by
llxll
= (/ lx(s)IP m(ds)tp.
(1)
The subadditivity
( / lx(s)
+ y(s)IP m(ds)Y'P :s;: (/ lx(s)IP
m(ds)tp
+ (/ iy(s)IP m(ds)Y'P,
called Minkowski's inequality, is clear for the case p
general case 1 < p < =, we need
Lemma 1. Let 1 < p

defined by
<=
(2)
1. To prove the
and let the conjugate exponent p' of p be

1
p + p' =
(3)
1.
Then, for any pair of non-negative numbers a and b, we have

aP bp'
ab<p+p;-,
where the equality is satisfied iff a
= b1i(P- 1l.
Proof. The minimum of the function f(c)
=; + ~~
-c for c
(4)
> 0
is attained only at c = 1, and the minimum value is 0. By taking

c = ab- 1/(P- 1) we see that the Lemma is true.
The proof of (2). We first prove Holder's inequality
J lx(s) y(s)i <

(for convenience, we write
(f lx(s)IP) 11P (f iy(s)IP') 11P'
J z(s) for j
(5)
z(s) m(ds))
To this end, we assume that A= (f lx(s)jP) 11P and B = (f iy(s)IP') 11P'

are both=!= 0, since otherwise x (s) y (s) = 0 a.e. and so (5) would be true.
Now, by taking a= lx(s)i/A and b = iy(s)i/Bin (4) and integrating, we
obtain
1 BP'
1 AP
------::1B-- <pAP+ p' BP' = 1 which implies (5).
Jlx(s) y(s) I
Next, by (5), we have
Jix(s)
+ y(s)IP- 1 lx(s)i
+ J lx(s) + y(s)IP-I iy(s)i
< (f lx(s) + y(s)!P'<P-1)JI!P' (f lx(s)iP)lfP
+ (J (ix(s) + y(s)!P'<P-1))1/P' (f (iy(s)iP)liP),
+ y(s)!P <
J ix(s)
which proves (2) by p' (p- 1)

3
p.
34
I.
Semi-norms
Remark 1. The equality sign in (2) holds iff there exists a non-negative
constant c such that x(s) = cy(s) m-a.e. (or y(s) = cx(s) m-a.e.). This is
implied from the fact that, by Lemma 1, the equality sign in Holder's
inequality (5) holds iff Jx(s)l = c ly(s)J 11<~->- 1 > (or Jy(s)l = c lx(s)J 11~-'- 1 )
are satisfied m-a.e.
Remark 2. The condition llxll = (J lx(s)IP) 11P = 0 is equivalent to the
condition that x(s) = 0 m-a.e. We shall thus consider two functions of
LP (S) as equivalent if they are equal m-a.e. By this convention, LP (S)
becomes a normed linear space. The limit relation s-lim x,. = x in LP (S)
1>-->00
is sometimes called the mean convergence of p-th order of the sequence of

functions x,.(s) to the function x(s).
Example 3. D'0 (S). A 58-measurable function x(s) defined on S is

said to be essentially bounded if there exists a constant .x such that
Ix (s) I < .x m-a.e. The infimum of such constants .x is denoted by
vraimax lx(s)l or essential sup lx(s)l.
~s
~s
L 00 (5, 58, m) or, in short, L 00 (S) is the set of all 58-measurable, essentially
bounded functions defined m-a.e. on S. It is a normed linear space by
(x
+ y) (s) =
x(s)
+ y(s),
(.xx) (s) = .xx(s), llx II = vrai max lx(s) I,

sES
under the convention that we consider two functions of L 00 (S) as equivalent if they are equal m-a.e.
Theorem 1. Let the total measure m(S) of S be finite. Then we have
lim (flx(s)IPm(ds)) 11P=vraimaxlx(s)l
p-+oo S
sES
for x(s)EL00 (S). (6)
Proof. It is clear that (f lx(s)IP m(ds))l/P < m(S) 11P vraimax lx(s)!
S
so that lim
f>-KX>
(f lx(s)IP) 11P < vraimax lx(s)l.

S
sES
sES
By the definition of the
vrai max, there exists, for any e > 0, a set B of m-measure > 0 at each
point of which Ix (s) I 2 vrai max Ix (s) 1- e. Hence ( Ix (s)IP m (ds)) 11P
sES
> m(B) 11P (vraimax lx(s)l-e). Therefore lim (J lx(s)IPJIIP > vrai max
sES
lx(s)l-e, and so (6} is true.
boo
sES
Example 4. Let, in particular, S be a discrete topological space consisting of countable points denoted by 1, 2, ... ; the term discrete
means that each point of S = {1, 2, ...} is itself open inS. Then as linear
subspaces of C({1, 2, ... }),we define (c0 ), (c) and (lP), 1 < p < oo.
(c 0): Consider a bounded sequence of real or complex numbers {~..}Such a sequence{~,.} defines a function x(n) =~.. defined and continuous
on the discrete space S = {1, 2, ... } ; we shall call x = {~..} a vector
35
with components ~n- The set of all vectors x = {~n} such that
lim ~n = 0 constitutes a normed linear space (c0 ) by the norm
n---+00
llxll =sup lx(n) I= sup l~nl

n
(c): The set of all vectors x =
{~n}
such that finite lim ~n exist,

n---+oo
constitutes a normed linear space (c) by the norm llxll =sup lx(n)l
n
=sup l~nl
n
(ZP), 1 <
00
.I
n=l
l~n IP
p<
< oo
00
oo: The set of all vectors x =
{~n}
such that
constitutes a normed linear space (ZP) by the norm

VIP
llxll = ( n~ l~niPJ
. As an abstract linear space, it is a linear subspace
of C({l, 2, ... }). It is also a special case of LP(5, \8, m) in which

m({l}) = m({2}) = = 1.
(l 00 ) = (m) : As in the case of L 00 ( 5), we shall denote by (l 00 ) the
linear space C({1, 2, ... }), normed by llxll =sup lx(n)l =sup l~nl
n
(l 00 ) is also denoted by (m).
The Space of Measures. Let \8 be a a-ring of subsets of 5. Consider

the set A (5, \Bj of all real- (or complex-) valued functions rp(B) defined
on \8 such that
lrp (B) I =F oo for every BE \8,
(7)
rpc~ Bj)= j~rp(Bj)
for any disjoint sequence {Bj} of sets E \B.
(8)
A (5, \B) will be called the space of signed (or complex) measures defined
on (5, \B).
Lemma 2. Let rp E A (5, \B) be real-valued. Then the total variation of
rp on 5 defined by
(9)
v (rp; 5) = (rp; 5) + I!:':_ (rp; 5) I
is finite; here the positive variation and the negative variation of rp over
B E \8 are given respectively by
V (rp; B) = sup rp (B 1) and

~8
~ (rp; B) =
inf rp (B 1).
~B
(10)
Proof. Since rp(0) = 0, we have V(rp; B):::.:: 0 > ::(rp; B). Suppose
that V (rp; 5) = oo. Then there exists a decreasing sequence {En} of sets
E \8 such that
V(rp; En)= oo, lrp(Bn) I> n -1.
The proof is obtained by induction. Let us choose B 1 = 5 and assume
that the sets B 2 , B 3 , .. , Bk have been defined so as to satisfy the above
conditions. By the first condition with n = k, there exists a set BE \8
3*
36
I. Semi-norms
such that B ~ Bk, !<J? (B) I > !<J? (Bk) I + k. We have only to set Bk+l = B
in the case V (<p; B) = CXJ and Bk+l = Bk- Bin the case V (<p; B) < CXJ.
For, in the latter case, we must have V (<p; B k -B) = CXJ and
j j- k which completes the induction.
By the decreasing property of the sequence {B,}, we have
5-
n B, =.I (5-B,)
= (5-Bl) + (Bl-B2) + (B2-Ba) + + (Bn-Bn+l) +
00
00
n~l
n~l
so that, by the countable additivity of <p,

<p ( 5-
"0 Bn) = <p (5- B

1
[<p (5) -
+ <p (B1 - B 2 ) + <p (B 2 - B 3 ) +

<p {B1)] + [<p (B1) - <p (B 2 )]
+ [<p(B2)-<p(B3 )] +
1)
= <p (5)- lim <p (En) =
CXJ
11-+00
or -
CXJ,
which is a contradiction of (7).

Theorem 2 (Jordan's decomposition). Let <pEA (5, \B) be real-valued.
Then the positive variation V (<p; B), the negative variation V (<p; B)
and the total variation V (<p; B) are countably additive on B. Moreover,
we have the Jordan decomposition
<p (B) = V (<p; B)
+ V (<p; B) for any BE \B.
(11)
Proof. Let {En} be a sequence of disjoint sets E \B. For any set BE ~
00
00
00
n=l
n~l
n~l
such that B ~ .I En, we have <pc~ en)
=n!<p(Cn)andso v(<p;n_;Bn)> ,!v(<p;Bn) Hence we have proved the countable additivity of V (<p; B) and those of V (<p; B) and of
V (<p; B) may be proved similarly.
To establish (11), vve observe that, for every C E )B with C ~ B, we
have <p(C) = <p(B) -<p(B- C)< <p(B)- V(<p; B) and so V(<p; B)<
<p(B)- V(<p; B). Similarly we obtain V(<p; B)> <p(B)- V(<p; B). These
inequalities together give (11).
Theorem 3 (Hahn's decomposition). Let <pEA (5, \B) be a signed
measure. Then there exists a set P E ~ such that
- 0 for every B E )B with B
P,
<p (B) :S:: 0 for every B E )B with B ~ pc
= 5 - P.
37
The decomposition 5 = P
of 5 pertaining to cp.
(5- P) is called the Hahn decomposition
Proof. For each positive integer n we choose a set Bn E \B such that

cp (Bn) > 17 (cp; 5) - 2-n. Hence by (11), we have
V(cp;Bn)>-2-n
V(cp;5-Bn)<2-".
and
{12)
The latter inequality is obtained from V (cp; 5- Bn) = V (cp; 5) - V (cp; Bn)
and V (cp; Bn) > cp (Bn). We then put
lim Bn =
= ........co
5 - P = lim (5- Bn) =
Then
........co
00
00
00
n Bn .
k=l n=k
U
00
00
n U (5- B,.)
k=1 n=k
U (5- Bn)
n=k
for
every k, and therefore, by the a-additivity of V (cp; B),

00
17(cp; 5- P) < .I V(cp; 5- Bn) <

n=k
2-(k-1},
which gives 17 (cp; 5 - P) = 0. On the other hand, the negative variation

V (cp; B) is a non-positive measure and so, by (12) and similarly as above,
which gives V (cp; P) = 0. The proof is thus completed.

Corollary. The total variation V (cp; 5) of a signed measure cp is defined
by
(13)
V(cp;5)= sup jfx(s)cp(ds)j
supJx(s)J ;::;;1 S
where x (s) ranges through )B-measurable functions defined on 5 such that

sup Ix (s) I :S: 1.
s
Proof. If we take x (s) = 1 or = - 1 according as s E P or s E 5- P,

then the right hand side of (13) gives V (cp; 5). On the other hand, it is
easy to see that
lx(s)l J V(cp; ds) =sup lx(s) I V(cp; 5)

J x(s) cp(ds)l <sup
s
s
s
:s
and hence (13) is proved.

Example 5. A (5, \B). The space A (5, \B) of signed measures cp on \B
is a real linear space by
+ <X2 cpz) (B) =
(B)
It is a normed linear space by the norm
(,x1 Cf!1
<X1 Cf!1
l/cpi/=V(cp;5)=
sup
+ <X2 Cf!2 (B),
B E lB
\fx(s)cp(ds)j.
supJx(s)J ;::;;1 S
(14)
I. Semi-norms
38
Example 6. The space A (5, iB) of complex measures cp is a complex

linear space by
(cx1 cp1
+ cx2 cp2) (B) =
+ cx2 cp 2 (B), B E iB with complex cx1 , cx 2 .
cx1 cp1 (B)
It is a normed linear space by the norm

llffJII=
(15)
[fx(s)g;(ds)[,
sup
supJx(s)i ~1 S
where complex-valued iB-measurable functions x(s) defined on 5 are

taken into account. We shall call the right hand value of (15) the total
variation of cp on 5 and denote it by V (cp; 5).
4. Examples of Quasi-normed Linear Spaces

Example 1. Cff!' (!2). The linear space ~k (.Q), introduced in Chapter I, 1,
is a quasi-normed linear space by the quasi-norm II x II = d (x, 0), where
the distance d (x, y) is as defined there.
Example 2. M(5, iB, m). Let m(5) < = and let M(5, iB, m) be the
set of all complex-valued iB-measurable functions x (s) defined on 5 and
such that lx(s)J < = m-a.e. Then M(5, iB, m) is a quasi-normed linear
space by the algebraic operations
(x
+ y) (s) =
x(s)
+ y(s),
(cxx) (s) = cxx(s)
and (under the convention that x = y iff x(s) = y(s) m-a.e.)
lx(s)l (1 + lx(s)l)- 1 m(ds).

s
The triangle inequality for the quasi-norm II x! I is clear from
llxfl
lex+ PI <
+lex+ PI =
(1)
lex!+ IPI <_lex_!_ __IP_I_

+lex!+ IPI = 1 +lex!+ 1 + IPI.
The mapping {ex, x} ~ cxx is continuous by the following
Proposition. The convergence s-limxn

n->OO
=x
in M(5, )B, m) is equi-
valent to the asymptotic convergence (or the convergence in measure) in 5

of the sequence of functions {xn (s)} to x (s):
For any e > 0, lim m {s E 5; Jx(s)- Xn(s)J2
n->00
e}
= 0.
{2)
Proof. Clear from the inequality

6
1+6m(B.,)
<
JJxl! ~ m(B.,)
+ 1 + 6 m(5-B.,), B.,={sE5;
Jx(s)J>b}.
Remark. It is easy to see that the topology of M(5, )B, m) may also
be defined by the quasi-norm
l!xl! = inf tan-1 [e
>0
+ m{sE 5;
Jx(s).J2 e}].
(1')
39
5. Pre-Hilbert Spaces
Example 3. 'l) K (.Q). The linear space 'l) K (.Q), introduced in Chapter I, 1,
is a quasi-normed linear space by the quasi-norm [lxll = d(x, 0), where
the distance d(x, y) is defined in Chapter I, 1.
Definition 1. A real or complex normed linear space X is called a
pre-Hilbert space if its norm satisfies the condition
(1)
Theorem 1 (M. FRJ3:CHET-J.
a real pre-Hilbert space X,
VON NEUMANN-P. JoRDAN).
(x, y) = 4-1 (11x
We define, in
+ Y 11 2 -llx- Y 11 2)
(2)
(aER 1),
(3)
+ (y, z),
(4)
Then we have the properties:
(ax,y)=a(x,y)
+ y, z) =
(x
(x, z)
(x, y) = (y, x),
(5)
llxll 2
(6)
(x,
x)
Proof. (5) and (6) are clear. We have, from (1) and (2),
(x' z)
+ (y' z) =
4-1 (II x
+ z 112 - II x -
z 112
+ II y + z 112 - II y -
z I\2)
(7)
=2(~tY,z).
= 0, we obtain (x, z) = ( 2 ~ , z) , because (0, z) = 0 by (2).

Hence, by (7), we obtain (4). Thus we see that (3) holds for rational
numbers a of the form a = mf2n. In a normed linear space, !lax + y II
and !lax- y II are continuous in a. Hence, by (2), (ax, y) is continuous in
a. Therefore (3) is proved for every real number a.
If we take y
Corollary (J. VON NEUMANN-P.

normed linear space X,
(x, y)
where i =
(x, y) 1
p,
jORDAN).
We define, in a complex
+ i(x, iy)v
(x, y) 1 = 4-1 (llx
+ y l\ 2 -llx- yj[ 2).
(8)
Then, we have (4), (6) and
(ax, y) =a (x, y)
(x, y)
(a E C1),
(y, x) (complex-conjugate number).
(3')
(5')
40
I. Semi-norms
Proof. X is also a real pre-Hilbert space and so (4) and (3') with real ex
hold good. We have, by (8), (y, xh = (x, Y)v (ix, iy) 1 = (x, y) 1 and hence
(y, ix) 1 = (-iiy, ixh = - (iy, x) 1 = - (x, iy) 1 . Therefore
(y, x) = (y, x) 1 + i(y, ix) 1 = (x, Yh- i(x, iy) 1 = (x, y).
Similarly, we have
(ix, y) = (ix, y) 1 + i (ix, iy) 1 = - (x, iy) 1 + i (x, y) 1 = i (x, y),
and therefore we have proved (3'). Finally we have (6), because
(x,x) 1 = llxll 2 and(x,ix) 1 =4-1 (11+il 2 -l1-il 2) llxll 2 =0.

Theorem 2. A (real or) complex linear space X is a (real or) complex
pre-Hilbert space, if to every pair of elements x, y EX there is associated
a (real or) complex number (x, y) satisfying (3'), (4), (5') and
(x, x) > 0, and (x, x) = 0 iff x = 0.
(9)
Proof. For any real number ex, we have, by (3'), (4) and (5')
(x + cx(x, y) y, x + cx(x, y) y) = llxll 2 + 2cx l(x, Y)l 2

+ cx 2 l(x, y)I 2 IIYII 2 > 0, where llxll = (x, x) 112
so that we have I(x, y) 14 - llx 11 2 I(x, y) 12 IIY 11 2 < 0. Hence we obtain
Schwarz' inequality
(10)
l(x, Y)l < llxiiIIYIJ,
where the equality is satisfied iff x and y are linearly dependent.
The latter part of (10) is clear from the latter part of (9).
We have, by (10), the triangle inequality for llxJJ:
JJx + yJJ 2 = (x + y, x + y) = JJxJJ 2 + (x, y) + (y, x) + !JyJJ2
< (JJxJJ + JJyiJ)2.
Finally, the equality (1) is verified easily.
Definition 2. The number (x, y) introduced above is called the scalar
product (or inner product) of two vectors x andy of the pre-Hilbert space
X.
Example 1. P(S, ~' m) is a pre-Hilbert space in which the scalar
product is given by (x, y) =
x(s) y (s) m(ds).
s
Example 2. The normed linear space (12) is a pre-Hilbert space in
which the scalar product is given by ({~n}, {nn}) = 2: ~n tJn

00
n=l
Example 3. Let D be an open domain of Rn and 0 < k

totality of functions I E ck (D) for which
II/ Ilk= ( 1 .2:

1
J
J JDij(x)J
~k !J
dx) 112
<
< oo. Then the
oo, where dx = dx 1 dx 2
is the Lebesgue measure in Rn,
dxn
(11)
41
constitutes a pre-Hilbert space fik (.Q) by the scalar product
(f,gh= ..I fDij(x)Djg(x)dx.
(12)
IJI::;;k D
Example 4. Let .Q be an open domain of Rn and 0 < k
<
oo. Then
C~ (.Q) is a pre-Hilbert space by the scalar product (12) and the norm (11).
We shall denote this pre-Hilbert space by fl~(.Q).

Example 5. Let G be a bounded open domain of the complex z-plane.
Let A 2 (G) be the set of all holomorphic functions f(z) defined in G and
such that
II! II= (~f l/(z)l2 dxdyY'2 < oo, (z = x + iy).
(13)
Then A 2 (G) is a pre-Hilbert space by the norm (13), the scalar product
(!,g)=
JJ f(z) g(z) dxdy
(14)
and the algebraic operations
(!+g) (z)
f(z) + g(z), (rxf) (z)
rx/(z).
Example 6. Hardy-Lebesgue class H-L2. Let H-L2 be the set of

all functions f (z) which are holomorphic in the unit disk {z; Iz I < 1}
of the complex z-plane and such that
sup (J"If(rei 0 )l 2 d8)<oo.
O<r<1
(15)
00
Then, if f(z) = .I enzn is the Taylor expansion of/,

n=O
oo2"
.
J If (r e'o) [2 d() = -2n1 n,m=O
_I J en em yn +m e(n-m)O d()
12".
F (r) = -
2n 0
= .I Ienf2 r2n
00
n=O
is monotone increasing in r, 0 < r

it is easy to see that
<
1, and bounded from above. Thus
1 ( 2n
)]1/2 =
II/II = O<r<1
sup [ 2n J lf(reio) 12 d()
0
( .I len 12 )1/2
00
n=O
is a norm which satisfies condition (1), since (l2) is a pre-Hilbert space.

Remark. Let a sequence {en} E (l2) be given, and consider
00
00
n=O
n=O
f(z) = /(reiO) = _I enzn = _I en yn einO,
lz I< 1.
By Schwarz' inequality, we have

oo
In~
en znl <
1
oo
n~k
oo
)112
lenl2 )112 ( n~
y2n '
(16)
42
I. Semi-norms
00
and so .I CnZn is uniformly convergent in any disk Iz I< e with 0 <

n=O
e < 1.
Thus f (z) is a holomorphic function in the unit disk Iz I < 1 such that
(15) holds good, that is, f(z) belongs to the class H-L2.
Therefore we have proved
Theorem 3. The Hardy-Lebesgue class H-L 2 is in one-to-one correspondence with the pre-Hilbert space (l 2) as follows:
00
H-L2 3 f(z) = .I CnZn-++ {en} E (l2 )

n=O
in such a way that

f(z)
f(z)
00
= .I CnZn-++ {en}, g(z) =

n=O
+ g(z) -++{en + dn},
00
_I dnzn-++ {dn} imply
n=O
c.f(z) -++{c. en} and
I[/ II =
C! \cn\ Y'
2
Hence, as a pre-Hilbert space, H-L 2 is isomorphic with (l2).
6. Continuity of Linear Operators

Proposition 1. Let X and Y be linear topological spaces over the
same scalar field K. Then a linear operator Z on D (T) s; X into Y is
continuous everywhere on D (T) iff it is continuous at the zero vector
X= 0.
Proof. Clear from the linearity of the operator T and T 0 = 0.
Theorem 1. Let X, Y be locally convex spaces, and {p}, {q} be the
systems of semi-norms respectively defining the topologies of X and Y.
Then a linear operator Ton D (T) s; X into Y is continuous iff, for every
semi-norm q E {q}, there exist a semi-norm p E {P} and a positive number
{3 such that
q(Tx)<f3p(x) forall xED(T).
(1)
Proof. The condition is sufficient. For, by T 0 = 0, the condition
implies that Tis continuous at the point x = 0 ED (T) and so Tis continuous everywhere on D (T).
The condition is necessary. The continuity of Tat x = 0 implies that,
for every semi-norm q E {q} and every positive number e, there exist a
semi-norm p E {P} and a positive number !5 such that
xE D(T) and p(x)
< !5 imply q(Tx)
<e.
Let x be an arbitrary point of D (T), and let us take a positive number )..
such that A.p (x) < !5. Then we have p (A.x) < !5, AxED (T) and so
q(T(h)) <e. Thus q(Tx) < efA.. Hence, if p(x) = 0, we can take )..
arbitrarily large and so q (T x) = 0; and if p (x) =I= 0, we can take)..= 15fp (x)
and so, in any case, we have q(Tx):::::;: f3p(x) with {3 = ej15.
43
6. Continuity of Linear Operators
Corollary 1. Let X be a locally convex space, and fa linear functional

on D (f) ~X. Then f is continuous iff there exist a semi-norm p from the
system {P} of semi-norms defining the topology of X and a positive
number {3 such that
I/ (x) I <
PP (x)
(2)
for all xED (f).
Proof. For, the absolute value I X I itself constitutes a system of seminorms defining the topology of the real or complex number field.
Corollary 2. Let X, Y be normed linear spaces. Then a linear operator
T on D (T) ~X into Y is continuous iff there exists a positive constant
fJ such that
(3)
II T x II < fJ II x II for all x E D (T) .
Corollary 3. Let X, Y be normed linear spaces. Then a linear operator
T on D (T) ~ X into Y admits a continuous inverse T-1 iff there exists a
positive constant y such that
II T x II
> y II x II for every x E D (T) .
(4)
Proof. By (4), Tx = 0 implies x = 0 and so the inverse T-1 exists.

The continuity of T- 1 is proved by (4) and the preceding Corollary 2.
Definition 1. Let T be a continuous linear operator on a normed linear
space X into a normed linear space Y. We define
!IT II =
~r~ fJ, where B = {{J; II T x II
< fJ II x II for all x E X}.
(5)
By virtue of the preceding Corollary 2 and the linearity of T, it is easy

to see that
(6)
II Til= sup IITxll = sup IITx!l
llzll =1
llzll ;:;;1
liT II is called the norm ofT. A continuous linear operator on a normed
linear space X into Y is called a bounded linear operator on X into Y,

since, for such an operator T, the norm IITxll is bounded when x ranges
over the unit disk or the unit sphere {xE X; llxll < 1} of X.
Definition 2. Let T and S be linear operators such that
D(T) and D(S)
~X,
and R(T) and R(S)
Y.
Then the sum T + S and the scalar multiple X T are defined respectively
by
(T + S) (x) = Tx + Sx for x E D(T) f\ D(S), (tXT) (x) = X(Tx).
Let T be a linear operator on D (T) ~X into Y, and Sa linear operator
on D(S) ~ Y into Z. Then the product ST is defined by
(ST) x = S(Tx) for xE {x; xE D(T) and TxE D(S)}.
T
+ S, XT and STare linear operators.
44
I. Semi-norms
Remark. 5 T and T 5 do not necessarily coincide even if X = Y = Z.
V(
= tx(t), 5x(t) =
1) :tx(t) considered as
linear operators from L 2 (R 1 ) into L2 (R 1 ). In this example, we have the
commutation relation (5 T- T 5) x (t) =
x (t).
Proposition 2. If T and 5 are bounded linear operators on a normed
linear space X into a normed linear space Y, then
An example is given by Tx
FJ:
liT+ 511 < IITII + 11511, llcxTII = lcxiiiTII

(7)
If T is a bounded linear operator on a normed linear space X into a
normed linear space Y, and 5 a bounded linear operator on Y into a
normed linear space Z, then
(8)
II5TII < II5IIIITII
Proof. We prove the last inequality; (7) may be proved similarly.
II5Txll < II5IIIITxll < 115/IIITII Jlxll and so IJ5TJI:::;; JI5IIIITII
Corollary. If Tis a bounded linear operator on a normed linear space

X into X, then
(9)
where T" is defined inductively by T" = T rn- 1 (n = 1, 2, ... ; yo = I

which maps every x onto x itself, i.e., I x = x, and I is called the identity
operator).
7. Bounded Sets and Bornologic Spaces

Definition 1. A subset B in a linear topological space X is said to be
bounded if it is absorbed by any neighbourhood U of 0, i.e., if there exists
a positive constant ex such that cx-1 B ~ U. Here cx-1 B = {x EX; x=cx- 1 b,
bE B}.
Proposition. Let X, Y be linear topological spaces. Then a continuous
linear operator on X into Y maps every bounded set of X onto a bounded
set of Y.
Proof. Let B be a bounded set of X, and V a neighbourhood of 0 of Y.
By the continuity of T, there exists a neighbourhood U of 0 of X such
that T U = {Tu; uE U} ~ V. Let ex> 0 be such that B ~ cxU. Then
T B ~ T (ex U) = ex (T U) ~ex V. This proves that T B is a bounded
set of Y.
Definition 2. A locally convex space X is called bornologic if it satisfies
the condition:
If a balanced convex set M of X absorbs every bounded
set of X, then M is a neighbourhood of 0 of X.
(1)
Theorem I. A locally convex space X is bornologic iff every seminorm on X, which is bounded on every bounded set, is continuous.
45
7. Bounded Sets and Bornologic Spaces
Proof. We first remark that a semi-norm p(x) on X is continuous iff

it is continuous at x = 0. This we see from the subadditivity of the seminorm: p (x- y) > Ip (x) - p (y) I (Chapter I, 1, (4)).
Necessity. Let a semi-norm p(x) on X be bounded on every bounded
set of X. The set M = {xE X; p(x) < 1} is convex and balanced. If B
is a bounded set of X, then sup p (b) = lX < oo and therefore B ~ lXM.
bEB
Since, by the assumption, X is bornologic, M must be a neighbourhood

of 0. Thus we see that p is continuous at x = 0.
Sufficiency. Let M be a convex, balanced set of X which absorbs

every bounded set of X. Let p be the Minkowski functional of M. Then p
is bounded on every bounded set, since M absorbs, by the assumption,
every bounded set. Hence, by the hypothesis, p (x) is continuous. Thus
M 1 = {x E X; p (x) < 1/2} is an open set 3 0 contained in M. This proves
that M is a neighbourhood of 0.
Example 1. Normed linear spaces are bornologic.
Proof. Let X be a normed linear space. Then the unit disk S =
{x EX; II xll < 1} of X is a bounded set of X. Let a semi-norm p(x) on X

be bounded on S, i.e., sup p (x) = lX < oo. Then, for any y o:F 0,
xES
p(y)
p (1\Y\1 11;1!) =
1\Y\1 p (~) <
lX
1\Y\1
Thus p is continuous at y = 0 and so continuous at every point of X.

Remark. As will be seen later, the quasi-normed linear space M (S, ~)
is not locally convex. Thus a quasi-normed linear space is not necessarily
bornologic. However we can prove
Theorem 2. A linear operator Ton one quasi-normed linear space into
another such space is continuous iff T maps bounded sets into bounded
sets.
Proof. As was proved in Chapter I, 2, Proposition 2, a quasi-normed
linear space is a linear topological space. Hence the "only if" part is already proved above in the Proposition. We shall prove the "if" part.
LetT map bounded sets into bounded sets. Suppose that s-lim xk = 0.
Then lim
II xk II =
k~
0 and so there exists a sequence of integers {nk}
lim nk = oo
such that
k-->oo
while
lim nk II xk 1\ = 0.
We may take, for instance, nk as follows:

nk =the largest integer<
if
xk
0.
1\xk\1-1'2
if xk # 0,
46
I. Semi-norms
Now we have link xk II= llxk + xk + + xk II< nk llxk II so that

s-lim nk xk = 0. But, in a quasi-normed linear space, the sequence
k~
{nk xk}, which converges to 0, is bounded. Thus, by the hypothesis,

{T(nk xk)} = {nk Txk} is a bounded sequence. Therefore
s-lim T xk =s-lim n;;- 1 (T (nk xk)) = 0,
k-+oo
k~oo
and so T is continuous at x = 0 and hence is continuous everywhere.

Theorem 3. Let X be bornologic. If a linear operator T on X into a
locally convex linear topological space Y maps every bounded set into a
bounded set, then Tis continuous.
Proof. Let V be a convex balanced neighbourhood of 0 of Y. Let p be
the Minkowski functional of V. Consider q (x) = p (T x). q is a semi-norm
on X which is bounded on every bounded set of X, because every bounded
set of Y is absorbed by the neighbourhood V of 0. Since X is bornologic,
q is continuous. Thus the set {x EX; Tx Eva}= {x EX; q(x) < 1} is a
neighbourhood of 0 of X. This proves that Tis continuous.
8. Generalized Functions and Generalized Derivatives

A continuous linear functional defined on the locally convex linear
topological space 'Il (Q), introduced in Chapter I, 1, is the "distribution"
or the "generalized function" of L. SCHWARTZ. To discuss the generalized
functions, we shall begin with the proof of
Theorem 1. Let B be a bounded set of 'Il (Q). Then there exists a
compact subset K of Q such that
supp (rp) s;; K for every rp E B,

sup IDj rp (x) I <
xEK,q;EB
for every differential operator Dj.
(1)
(2)
Proof. Suppose that there exist a sequence of functions {rp;} s;; B and
a sequence of points {p;} such that (i): {P;} has no accumulation point
in Q, and (ii): rp; (p;) # 0 (i = 1, 2, ... ). Then
00
p(rp) = i~i lrp(p;)l/lrp;(p;)l

is a continuous semi-norm on every 'IJK(rp), defined in Chapter I, 1.
Hence, for any e > 0, the set {rp E DK (Q); p (rp) < e} is a neighbourhood
of 0 of 'IJK(Q). Since 'Il (Q) is the inductive limit of 'IJK(Q)'s, we see that
{rp E 'Il (.Q); p (rp) < e} is also a neighbourhood of 0 of 'Il (.Q). Thus p is
continuous at 0 of 'Il (.Q) and so is continuous on 'Il (.Q). Hence p must be
bounded on the bounded set B of 'Il (.Q). However, p (rp;) > i (i = 1, 2, ... ).
This proves that we must have (1).
47
We next assume that (1) is satisfied, and suppose (2) is not satisfied.
Then there exist a differential operator Dj, and a sequence of functions
{cp;}s;;B such thatsupJDj'cp;(x)J>i(i=1,2, ... ). Thus, if weset
xEK
p(cp) =sup JDj'cp(x)J
for
cpE ~K(.Q),
xEK
p(cp) is a continuous semi-norm on ~K(.Q) and p(cp;) > i (i = 1, 2, ... ).

Hence {cp;} s;; B cannot be bounded in ~K(.Q), and a fortiori in ~(.Q).
This contradiction proves that (2) must be true.
Theorem 2. The space ~ (.Q) is bornologic.
Proof. Let q(cp) be a semi-norm on ~ (.Q) which is bounded on every
bounded set of~ (.Q). In view of Theorem 1 in Chapter I, 7, we have only
to show that q is continuous on~ (.Q). To this purpose, we show that q is
continuous on the space ~K (.Q) where K is any compact subset of .Q.
Since ~(.Q) is the inductive limit of ~K(.Q)'s, we then see that q is continuous on ~ (.Q).
But q is continuous on every ~K (.Q). For, by hypothesis, q is bounded
on every bounded set of the quasi-normed linear space ~K (.Q), and so, by
Theorem 2 of the preceding section, q is continuous on ~K(.Q). Hence q
must be continuous on ~ (.Q).
We are now ready to define the generalized functions.
Definition 1. A linear functional T defined and continuous on ~ (.Q)
is called a generalized function, or an ideal function or a distribution in .Q;
and the value T (cp) is called the value of the generalized function T at
the testing function cp E ~ (.Q).
By virtue of Theorem 1 in Chapter I, 7 and the preceding Theorem 2,
we have
Proposition 1. A linear functional T defined on ~ (.Q) is a generalized
function in Q iff it is bounded on every bounded set of ~ (.Q), that is,
iff Tis bounded on every set BE~ (.Q) satisfying the two conditions (1)
and (2).
Proof. Clear from the fact that T (cp) is continuous iff the semi-norm
J T (cp) J is continuous.
Corollary. A linear functional T defined on C8" (.Q) is a generalized
function in .Q iff it satisfies the condition:
To every compact subset K of Q, there correspond a
positive constant C and a positive integer k such that
JT(cp)J < C sup JDjcp(x)J whenever cpE ~K(cp).
(3)
Jjl :;>k,xEK
Proof. By the continuity of T on the inductive limit ~ (.Q) of the
~K(.Q)'s, we see that T must be continuous on every ~K(Q). Hence
the necessity of condition (3) is clear. The sufficiency of the condition
48
I. Semi-norms
(3) is also clear, since it implies that Tis bounded on every bounded set
of :i:l (.Q).
Remark. The above Corollary is very convenient for all applications,
since it serves as a useful definition of the generalized functions.
Example 1. Let a complex-valued function f(x) defined a.e. in .Q
be locally integrable in .Q with respect to the Lebesgue measure
dx = dx dx dx in Rn, in the sense that, for any compact subset K
of.Q,
lf(x)l dx <ex>. Then
T 1(rp) =
f(x) rp(x) dx, rpE :i:l(.Q),

(4)
n
defines a generalized function T1 in .Q.
Example 2. Let m (B) be a a-finite, a-additive and complex-valued
measure defined on Baire subsets B of an open set .Q of Rn. Then
Tm(rp)= jrp(x)m(dx), rpE:i:l(.Q),
(5)
defines a generalized function T min .Q.

Example 3. As a special case of Example 2,
Thv(rp) = rp(p), where Pis a fixed point of .Q, rpE :i:l(.Q),
(6)
defines a generalized function T hv in .Q. It is called the Dirac distribution

concentrated at the point p E .Q. In the particular case p = 0, the origin
of Rn, we shall write Tb or~ for Tho
Definition 2. The set of all generalized functions in .Q will be
denoted by :tl (.Q)'. It is a linear space by
(T + S) (rp) = T(rp) + S(rp), (~T) (rp) = ~T(rp),
(7)
and we call :i:l (.Q)' the space of the generalized functions in .Q or the dual
space of :i:l (.Q).
Remark. Two distributions T1, and T 1, are equal as functionals
(T1, (rp) = T1, (rp) for every rp E :i:l (.Q)) iff / 1 (x) = / 2 (x) a.e. If this fact is
proved, then the set of all locally integrable functions in .Q is, by f ++ T1 ,
in a one-one correspondence with a subset of ;tl (.Q)' in such a way that
(11 and / 2 being considered equivalent iff / 1 (x) = /2 (x) a.e.)
T 1,
T1, = T1,+1,,
~T1
=Tat
(7')
In this sense, the notion of the generalized function is, in fact, a generalization of the notion of the locally integrable function. To prove the
above assertion, we have only to prove that a locally integrable
function f is = 0 a.e. in an open set .Q of Rn 'if
f (x) rp (x) dx = 0 for
n
all rp E C0 (.Q). By introducing the Baire measure fl (B) =
f (x) dx,
the latter condition implies that
Jrp (x) fl (dx) = 0 for
all rp E C0 (.Q),
which further implies that
f rp (x) t-t (dx) =
49
0 for all rp E cg (.Q), by virtue
!}
of Proposition 8 in Chapter I, 1. Let B be a compact G6-sct in ,Q:

B
00
= n Gn, whereGn is an open relatively compact set in .Q. By applying

n~l
Urysohn's theorem in Chapter 0, 2, there exists a continuous function

ln(x) such that
0
< l,(x) < 1 for xE.Q, ln(x) = 1 for xE G~+ 2 and ln(x) = 0
for X E G~- Gn+l (n = 1, 2, ... ) ,
assuming that {Gn} is a monotone decreasing sequence of open relatively

compact sets of ,Q such that G~+2 ~ Gn+l Setting rp =In and letting
n ~ex:>, we see that ft (B) = 0 for all compact G6-sets B of .Q. The Baire
sets of ,Q are the members of the smallest a-ring containing compact G6 sets of .Q, we see, by the a-additivity of the Baire measure ft that ft
vanishes for every Baire set of .Q. Hence the density I of this measure {t
must vanish a.e. in .Q.
We can define the notion of differentiation of generalized functions
through
Proposition 2. If T is a generalized function in ,Q, then
S(rp) = - T
(:~),
(8)
rp E 'IJ(.Q),
defines another generalized function S in ,Q.

Proof. S is a linear functional on 'IJ (.Q) which is bounded on every
bounded set of 'IJ (.Q).
Definition 3. The generalized functional S defined by (8) is called the
generalized derivative or the distributional derivative of T (with respect to
x1 }, and we write
8
5=-T
8x1
so that we have
(9)
'
(10)
Remark. The above notion is an extension of the usual notion of the
derivative. For, if the function I is continuously differentiable with
respect to Xv then we have
8 T1 (rp) = - T1 (8q;)
= -1. . .
8x
8x1
!}
=f ... f 8! l(x). rp(x) dx

1
8q;
f(x) -dx
1 . . dxn
8x1
dxn
Toflax,
(rp),
!}
as may be seen by partial integration observing that rp (x) vanishes identically outside some compact subset of .Q.
4
50
I. Semi-norms
Corollary. A generalized function T in Q is infinitely differentiable

in the sense of distributions defined above and
(DiT) (rp) = (-1)1il T(Dirp), where
.i Ji,
li 1=
'= 1
Di = - ._81 '~1 _ .
ox~'
{11)
. . . ox!,n
Example 1. The Heaviside function H(x) is defined by
H(x) = 1 or= 0 according as x

Then we have
> 0 or x < 0.
(12)
(12')
where T 6, is the Dirac distribution concentrated at the. origin 0 of R 1

In fact, we have, for any rp E 'l) (R1),
Cx T
d
J H (x) rp' (x) dx = - J rp' (x) dx = - [rp (x)]g" =

00
00
(rp) = -
-00
rp (0).
Example 2. Let I (x) have a bounded and continuous derivative in the

k
open set R 1 - .U xi of R 1 . Let si = l(xi

J=l
+ 0) -f(xi- 0)
be the saltus
or the jump of l(x) at x =xi. Since
(dxd Tf) (rp) = - J I (x) rp' (x) dx = ~ rp (xj) sj + J f' (x) rp (x) dx,
00
00
-oo
-oo
we have
(12")
where
Ox;
is defined by (6).
Example 3. Let I (x) = I (xv x2 , , Xn) be a continuously differentiable function on a closed bounded domain Q ~ Rn having a smooth
boundary S. Define I to be 0 outside Q. By partial integration, we have
! Tf) (rp) = - JI (x) o!; rp (x) dx
(8
I (x) rp (x) cos (v, xj) dS + 0~1 rp (x) dx,

n
s
where v is the inner normal to S, (v, xi) = (x1 , v) is the angle between v
and the positive xraxis and dS is the surface element. We have thus
=
!.' T = Toffox; + T
5,
whereT5 (rp) =
(12")
l(x) cos (v, xi) rp(x) dS.
s
Xn) is C2 on Q and is 0 outside,
Corollary. If l(x) = l(xv x2 , . ,

then, from (12") and 88 = ~ -88 cos (xi, v) we obtain Green's
X;
J
V
integral theorem
(LI T1) (rp) = T LJf(rp)
..
+ f: rp (x) dS- f I (x) ~~ dS,
51
(12"')
where Ll is the Laplacian -~ 82f8xj.

J=l
Proposition 3. If T is a generalized function in Q and I E C00 (Q), then

S(rp)
T(lrp), rpE 'l)(Q),
(13)
defines another generalized functionS in Q.

Proof. S is a linear functional on 'l) (Q) which is bounded on every
bounded set of 'l)(Q). This we see by applying Leibniz' formula to lrp.
Definition 4. The generalized function S defined by (13) is called
the product of the function I and the generalized function T.
Leibniz' Formula. We have, denoting Sin (13) by IT,
_!__(IT)
OX;
T +I oT
=.fl..
OX;
OX;
(14)
because we have
- T
rp)- T (_!__ (lrp))

(1 OX;orp)= T (.fl..
OX;
OX;
by Leibniz' formula for 8(lrp)f8xi. This formula is generalized as follows.

Let P (~) be a polynomial in ~I> ~2 , ~... and consider a linear
partial differential operator P(D) with constant coefficients, obtained
by replacing ~i by i- 1 8f8xi. The introduction of the imaginary coefficient
i- 1 is suitable for the symbolism in the Fourier transform theory in
Chapter VI.
Theorem 3 (Generalized Leibniz' Formula of L. HoRMANDER). We
have
P(D) (IT) =
where, for oc:
(oc:v oc:2 ,
p(e<)(~)-
and
(oc:)
1) DO<I.
L -(
cx.
(e<);;:;;o
oc:k) with 1
0~0<,. 0~"' o~"'k'
(D) T'
(15)
< oc:; < n,
o"'l+e<z++O<k
k, and p(O) (~)
p(e<)
- I Ii -ox"'i'
0<-
(16)
j=l
P(~). and D 0 =I for (oc:)
0.
(17)
Proof. Repeated application of (14) gives an identity of the form

P(D) (IT) =
~ D"'l Q"'(D) T,
(e<);;:;;o
(18)
where Q"'(D) are differential operators to be determined. We may assume

that they are so chosen that they are invariant under permutation of the
indices oc:; the invariance may be attained by rearrangement and summa4*
52
I. Semi-norms
tion over the indices a with the same (a), divided by the number of
summands. Since (17) is an identity, we may substitute in (17)
f(x)
= ei(x,t;)
and T
= ei(x,TJ>, where
<x,
~)=,I Xj~j
J~l
Then, by the symbolism

P(D)
ei(x,l;)
= P(~)
(19)
ei(x,l;),
we obtain
On the other hand, we have, by Taylor's formula,
P(~
and so we get Qa('Y})
+ 'Y})
= 2; _1 ~a p(a) ('YJ)
a (ex)!
'
=-(~!p(a)('Yj).
9. B-spaces and F-spaces
n-+oo llxn-xll = Oimplies, by the

llxn-xmll < llxn-xll + llx-xmll, that {xn} is a
Inaquasi-normedlinearspaceX, lim
triangle inequality
Cauchy sequence, i.e., {xn} satisfies Cauchy's convergence condition
lim llxn-
n,tn--7-00
Xm
II= 0.
(1)
Definition 1. A quasi-normed (or normed) linear space X is called an

F-space (or a B-space) if it is complete, i.e., if every Cauchy sequence {xn}
of X converges strongly to a point x00 of X:
lim llxn-Xooll
n-+oo
= 0.
(2)
Such a limit x00 , if it exists, is uniquely determined because of the triangle

inequality llx- x' II< llx- Xn II + Jlxn- x' II A complete pre-Hilbert
space is called a Hilbert space.
Remark. The names F-space and B-space are abbreviations of Frechet space and Banach space, respectively. It is to be noted that BouRBAKI uses the term Fnkhet spaces for locally convex spaces which are
quasi-normed and complete.
Proposition 1. Let Q be an open set of Rn, and denote by @; (Q) =
coo (Q) the locally convex space, quasi-normed as in Proposition 6 in
Chapter I, 1. This @: (Q) is an F -space.
Proof. The condition lim Ilin -fm II= 0 in @:(Q) means that, for
n,m~oo
any compact subset K of Q and for any differential operator Da, the
sequence {D"fn(x)} of functions converges, as n-+oo, uniformly bn
K. Hence there exists a function f (x) E C00 (Q) such that lim D"' /,. (x) =
n-+oo
53
D"'l(x) uniformly on K. D"' and K being arbitrary, this means that

lim II In- I II = 0 in ~(.Q).
n--+00
Proposition 2. LP(S) = LP(S, ?S, m) is a B-space. In particular, L 2 (S)
and (l2 ) are Hilbert spaces.
Proof. Let lim II Xn - Xm II = 0 in LP (S). Then we can, choose a
n m--+00
'
subsequence {x~k} such that
llxnk+t- Xnk II
< =
Applying the tri-
angle inequality and the Lebesgue-Fatou Lemma to the sequence of

functions
we see that
j(l-~Ye(s)P)m(ds) <~~ IIYeW<(IIxn,ll + ~llxnk+t-Xnk\lf

Thus a finite lim y1 (s) exists a. e. Hence a finite lim Xn 1 (s)
1--+00
t--+OO
a.e. and x 00 (s) E LP(S), since Jxn 1 (s)

+1
+t
x 00 (s) exists
I< lim y (s) E LP(S).

t~oo
Applying
again the Lebesgue-Fatou Lemma, we obtain
IJ X
00 -
Xnk W=
J (lim IXnt (s)- Xnk (s) lp) m (ds) :::::;: (i II Xn +t -
l=k
l-->00
Xn1 ll)p
Therefore lim llx00 -Xnk II= 0, and hence, by the triangle inequality and
k--+00
Cauchy's convergence condition lim llxn- Xm IJ

n,~oo
lim
H----700
II X
00 -
Xn II
lim II X00 - Xnk II

< k--7>00
lim
k,n--+00
0, we obtain
II Xnk -
Xn II = 0.
Incidentally we have proved the following important

Corollary. A sequence {xn} E LP (S) which satisfies Cauchy's convergence condition (1) contains a subsequence {xnk} such that
a finite lim Xkn(s)
k-->00
= X00
s-lim Xn =
n-->00
(s) exists a.e., X00 (s)ELP(S)
and
(3)
X 00
Remark. In the above Proposition and the Corollary, we have assumed

in the proof that 1 < p < = However the results are also valid for the
and the proof is somewhat simpler than for the case
case p =
1 < p < = The reader should carry out the proof.
Proposition 3. The space A 2 (G) is a Hilbert space.
Proof. Let {ln(z)} be a Cauchy sequence of A 2 (G). Since A 2 (G) is a
linear subspace of the Hilbert space L2 (G), there exists a subsequence
{Ink (z)} such that
=.
a finite lim Ink (z) = loo (z) exists a.e., loo E L2 (G)
k-->00
lim
J lloo(z)- ln(z)
n->OOG
dx dy = 0.
and
I. Semi-norms
54
We have to show that / 00 (z) is holomorphic in G. To do this, let the sphere

iz-z0 l <(!be contained in G. The Taylor expansion fn(z)- fm(z) =
00
..J: ei (z- z0 ) 1 implies
J=O
llfn-fmli 2 >
1-1 :;;;~
lf,.(z)-fm(z)i 2 dxdy
+ 2t1 >
00
2n i~ (!2H2 led (2j
= n
e2 l!,.(zo)- fm(zo)
n leo 12 (!2
(4)
12
Thus the sequence {/,.(z)} itself converges uniformly on any closed sphere
containedinG./n(z)'s beingholomorphicinG, weseethatf00 (z) = lim In (z)
n->OO
must be holomorphic in G.
Proposition 4. M (S, 58, m) with m (5) < = is an F-space.
Proof. Let {x,.} be a Cauchy sequence in M (5, 58, m). Since the convergence in M(5, 58, m) is the asymptotic convergence, we can choose a
sub-sequence {x,.k(s)} of {x,.(s)} such that
m(Bk) < 2-k for Bk = {sE 5; 2-k < lx"k+~(s) -x,.k(s) !}.
The sequence
k-1
x,.k (s) = x,., (s) +
i~
(x"i+ (s) - x,.; (s)) (k = 1, 2, ... )

00
is s-convergent to a function EM (5, 58, m), because, if s E U Bi,

j=t
we
and
00
00
J=l
J=l
< ,I m(Bj) < .I 2-i < 21-
t-
CXJ,
tion
X 00
consequently
1;
we
see,
(gt Bj)
by letting
that the sequence {x,.k(s)} converges m-a.e. to a func(s)EM(5,58,m). Hence lim Jlx,.k-xooii=O and so, by
lim llx,.-xmll
n,m-+oo
k->00
0, we obtain lim Jlx,.-x00 JI

~oo
The Space (s). The set (s) of all

med by
11{~,.}11
00
sequences{~,.}
.
= ..J:2-J
J=l
l~jl/(1
0.
of numbers quasi-nor-
+ l~jl)
constitutes an F-space by{$,.}+ {17,.} = {~,. + 1],.}, ,x{~,.} ={<X~,.}. The

proof of the completeness of (s) may be obtained as in the case of
M(5, 58, m). The quasi-norm
II{$,.} II= inftan-1 {e + the number of ~,.'s which satisfy l~n I> e}
>0
also gives an equivalent topology of (s).
55
Remark. It is clear that C (S), (c0 ) and (c) are B-spaces. The completeness of the space (lP) is a consequence of that of.LP (S). Hence, by Theorem 3 in Chapter I, 5, the space H-L2 is a Hilbert space with (l2).
Sobolev Spaces Wk,p (.Q). Let .Q be an open set of R", and k a positive
integer. For 1 < p < =, we denote by Wk,p (.Q) the set of all complexvalued functions I (x) = I (x1 , x 2 , , xn) defined in .Q such that I and
its distributional derivatives
to LP (.Q).
Wk,p (.Q)
ns I of order Is I = J=1
..I ISj I <
k all belong
is a normed linear space by
+ /2) (x) = l 1 (x) + l 2 (x),
= O(.l(x) and
l/1/lkP
= ( is~kn
2: J IDsl(x) IP dx) 11P, dx = dx 1 dx 2 dx,
,
under the convention that we consider two functions 11 and 12 as the same
vector of Wk,p (.Q) if 11 (x) = 12 (x) a.e. in .Q. It is easy to see that Wk 2 (.Q)
(/1
(0(.1) (x)
is a pre-Hilbert space by the scalar product
(.I j Dsl(x) Dsg(x) dx).
(/, g)k,2 =
isl;;i;k D
Proposition 5. The space Wk,p (.Q) is a B-space. In particular, Wk (.Q) =

is a Hilbert space by the norm II I Ilk = II I !/k, 2 and the scalar
product (I, gh = (/, gh,2
Proof. Let {ih} be a Cauchy sequence in Wk,p (.Q). Then, for any differential operator ns with Is I < k, the sequence {ns In} is a Cauchy sequence
in LP (.Q) and so, by the completeness of LP (.Q), there exist functions
l(s) E LP (.Q) (/ s I < k) such that lim
ID~ In (x) - /s) (x)/P dx = 0. By
h-oon
virtue of Holder's inequality in Chapter I, 3, applied to compact sets
of Q, we easily see that lh is locally integrable in Q. Hence, for any function rp E C0 (.Q),
Wk 2 (.Q)
Tns11,(rp) =
and
lim
so,
again
JDsl,.(x) rp(x) dx =
applying
J /l,.(x)- / 0l (x) IP dx =
h~n
Holder's
0,
(-l)lsl
J l,.(x) D rp(x) dx,

5
inequality,
we
obtain,
by
lim Tnth(rp) = (-l)lsl Tf(OJ(Dsrp) = nsrf(OJ(rp).
h-oo
Similarly we have, by lim
J /Dsl,.(x)- /s)(x)/P dx =
"~n
0,
T nth (rp) = T1(sJ (rp).
Hence we must have ns Tf(O) = Tt(S), that is, the distributional derivative Ds l(o) equals l(sl. This proves that lim Ill" - / 0l 1/k,p = 0 and
Wk,p (.Q)
is complete.
h-oo
56
I. Semi-norms
10. The Completion

The completeness of an F-space (and a B-space) will play an important
role in functional analysis in the sense that we can apply to such spaces
Baire's category arguments given in Chapter 0, Preliminaries. The following theorem of completion will be of frequent use in this book.
Theorem (of completion). Let X be a quasi-normed linear space which

is not complete. Then X is isomorphic and isometric to a dense linear
subspace of an F-space X, i.e., there exists a one-to-one correspondence
x -. ~ of X onto a dense linear subspace of X such that
(:--+--;;) = x + y, (~) = cxx,
(1)
llxll = llxll
The space X is uniquely determined up to isometric isomorphism. If X

is itself a normed linear space, then X is a B-space.
Proof. The proof proceeds as in Cantor's construction of real numbers

from rational numbers.
The set of all Cauchy sequences {xn} of X can be classified according
to the equivalence {xn} . . . . , {Yn} which means that lim II Xn - Yn II = 0.
n--->00
We denote by {xn}' the class containing {x11}. Then the set

classes = {xn}' is a linear space by
{xn}'
+ {Yn}' =
{xn
+ Yn}',
X of all such
cx{xn}' = {cxxn}'
We have lllxn[l-[[xm[[[<[lxn-xm[[ and hence lim llxn[l exists.

n->00
We put
It is easy to see that these definitions of the vector sum {x,.}'
+ {y
11 }',
the scalar multiplication cx{x,.}' and the norm ll{xn}' II do not depend
on the particular representations for the classes {xn}', {Yn}', respectively.
For example, if {xn} . . . . , {x~}, then
lim II Xn II :S: lim II x~ II
n---+00
t~OO
and similarly
+ lim
n---+00
II< -
Xn
II
:S lim
n-KX:l
II x~ II
lim llx~ll :Slim Jlxnll, so that we have !l{xn}' II=
n-+oo
n-+00
ll{x~}'ll.
To prove that ll{xn}' !I is a quasi-norm, we have to show that
lim iicx{xn}' II= 0

a~o
and
lim
JJ{.nYIJ~o
iicx{x11} ' II= 0.
The former is equivalent to lim lim [lcxx,.! I = 0 and the latter is

C:--*0
n---+oo
equivalent to ll~ llcxxn!I=O. And these aretruebecause llcxx[liscon

tinuous in both variables ex and x.
57
10. The Completion
To prove the completeness of X, let {xk} = {{x~l}} be a Cauchy

sequence of X. For each k, we can choose nk such that
llx~!l-x~211<k- 1
Then we can show that the sequence

the Cauchy sequence of X:
if
(2)
m>nk.
{xk} converges to the class containing
{x~:l, x~:l, . . . , x~2, . . .} .
(3)
To this purpose, we denote by x~2 the class containing

(4)
{x~2, x~2, . . . , x~~, . . .} .

Then, by (2),
(5)
and hence
llx~2 -x~~ II= llx~2- .x;::; II< llx~2
<
-.Xkll + [j.Xk-xmll + iixm-x~~ II
iixk-xmii + k-1 + m-1.
Thus (3) is a Cauchy sequence of X. Let

Then, by (5),
!lx-.Xkll <
II.X-x~211
x be the class containing (3).
+ llx~2-.Xkll ~
II.X-x~211
+ k- 1
Since, as shown above,
jj.X- x~211 <

lim
we prove that k-+00
lim
j>->00
jjxjf,l- X~~ II<
lim
j>->00
ilx- x~211 = o, and so
jj.Xp- xk II + k- 1
lim
k-+OO
ilx- .xk II= o.
The above proof shows that the correspondence

X 3x
++
x=
{x, x, ... , x, ... }' =:X
is surely isomorphic and isometric, and the image of X in X by this

correspondence is dense in X. The last part of the Theorem is clear.
Example of completion. Let Q be an open set of Rn and k < oo. The
completion of the space C~ (Q) normed by
II/ Ilk= Cs~ki jDs f(x) [2 dxY' 2
will be denoted by H~ (Q); thus H~ (Q) is the completion of the preHilbert space H~ (Q) defined in Chapter I, 5, Example 4. Therefore H~ (Q) is a
Hilbert space. The completion of the pre-Hilbert space Hk (Q) in Chapter I, 5, Example 3 will similarly be denoted by Hk (Q).
The elements of H~ (Q) are obtained concretely as follows: Let {fh} be a
Cauchy sequence of C~ (Q) with regard to the norm II f Ilk Then, by the
58
I. Semi-norms
completeness of the space L 2 (Q), we see that there exist functions
(x) E L 2 (Q) with Is I= ..I Sj < k such that

J~l
lim
J 1/s) (x)- D' lh (x) 1 dx =

2
~D
(dx = dx 1 dx 2
dxn).
Since the scalar product is continuous in the norm of L 2 (Q), we see, for
any test function rp (x) E C3" (Q), that
Tt<l(rp) =lim <Dsfh, rp) =lim (-1)11 Tth(D 5 rp)

h~
h~
= (-1)1s11im <lh,D5 rp) = (-1)1sJ <l(o),D 5 rp) = (D 5 T1<ol) (rp).

h~
Therefore we see that l(s) E L 2 (Q) is, when considered as a generalized

function, the distributional derivative of 1(0}: l(s) = D 5 I(O).
We have thus proved that the Hilbert spaceH~ (Q) is a linear subspace
of the Hilbert space Wk (Q), the Sobolev space. In general H~ (Q) is a
proper subspace of Wk (Q). However, we can prove
Proposition. H~ (Rn) = Wk (Rn).
Proof. We know that the space Wk (Rn) is the space of all functions
l(x) E L 2 (Rn) such that the distributional derivatives D 5 l(x) with lsi=
n
..I Sj < k all belong to L 2 (Rn)' and the norm in
J~l
Let
IE
wk
(Rn) is given by
II I Ilk = ( .l:' Ins I (x) 12 dx)l'z .
Jsj;;;;k
wk (Rn) and define IN by
IN(x) = OiN(x) l(x),

where the functions OiN (x) E Cg" (Rn) (N = 1, 2, ... ) is such that
IXN(x) = 1 for lxJ <Nand
sup
nRn;ls/:i?,k:N~l,2,
...
ID 5 1XN(x)J
< =
Then by Leibniz' formula, we have
D5 l(x)-D 5 IN(x)=0 for lxJ<N,

= a linear combination of terms
D1 IXN(x). nu l(x)with luI+ It!< k for lx I> N.

Hence, by D 5 IE L 2 (Rn) for Is I < k, we see that lim
and so lim II IN- I Ilk = 0.
N-+oo
II D IN-ns lllo =
5
N~
Therefore, it will be sufficient to show that, for any IE wk (Rn) with

compact support, there exists a sequence {Ia (x)} ~ C3" (Rn) such that
lim lila- I Ilk= 0. To this purpose, consider the regularization of I
a-+oo
(see (16) in Chapter I, 1):
la(x) =
Jl(y) Oa(x-y) dy,
Rn
a>
0.
59
11. Factor Spaces of a B-space
By differentiation, we have
D 5 f,.(x)
ff(y)D~O,.(x-y)dy= (-1)1sl ff(y)D~O,.(x-y)dy
R"
R"
= (D Tj) (O,.,x) (where o,.,x(Y) = O,.(x- y))

5
JD f(y) O,.(x- y) dy
5
(for is I< k).
R"
Hence, by Schwarz' inequality,
J iDsf,.(x)- D'f(x) 12 dx
< ( JO,.(x- y) dy) J [ J !D~f(y)- D~f(x) 2 O,.(x- y) dy] dx
R" R"
R"
= J [ J iD~f(y) -D~f(y +e) 1 dy] O,.(e) de, where
R"
il:;i;"
R"
Y + e = (Yl + ev Y2 + e2, . . , y,. + e,.)

inner integral on the extreme right tends to 0 as e ~ 0
the
that
We know
ID 5 f,.(x) -D5 f(x)! 2dx
(seeTheorem 1inChapterO, 3),andhence lim
a--+0 R"
= 0. Thus lim II/,.- !Ilk= 0. Therefore, the completion H~(R") of C~(R")

a-->0
with regard to the norm II !lk is identical with the space Wk (R").
Corollary. H~(R") = Hk(R") = Wk(R").
11. Factor Spaces of a B-space

Suppose that X is a normed linear space and that M is a closed linear
subspace m X. We consider the factor space XfM, i.e., the space whose
elements are classes modulo M. In virtue of the fact that M is closed,
all these classes~ are closed in X.
Proposition. If we define
!lxll.
II~ II= inf
xEo
(1)
then all the axioms concerning the norm are satisfied by !1~11
Proof. If~= 0, then~ coincides with M and contains the zero vector
of X; consequently, it follows from (1) that II~ II = 0. Suppose, conversely,
that II~ II= 0. It follows then from (1) that the class contains a sequence
{x,.} for which we have lim llx,.ll = 0, and hence the zero vector of X
n--+00
belongs to the closed set~ of X. This proves that~= M and hence is

the zero vector in X fM.
Next suppose ~. 'YJ E XfM. By definition (1), there exists for any e > 0,
vectors x E ~. y E 'YJ such that
llxll <II~ II+ e, IIYII <lin II+ e.
60
I. Semi-norms
Hence llx + Yll < llxll + IIYII < 11;11 +lin II+ 2e. On the other hand,
(x + y} E (; + n}, and therefore II;+ nil< llx + y II by (1}. Consequently, we have 11; +nil< 11; II + lin II + 2e and so we obtain the
triangle inequality II;+ n!l < 11;11 + linll
Finally it is clear that the axiom II IX; II = IIX Ill; II holds good.
Definition. The space XfM, normed by (1}, is called a normed factor
space.
Theorem. If X is a B-space and M a closed linear subspace in X, then
the normed factor space XfM is also a B-space.
Proof. Suppose{~..} is a Cauchy sequence in XfM. Then{;,.} contains
a subsequence {~n~c} such that ll~..k+~
r. II< 2-k-2 Further, by definition
(1} of the non11 in XfM, one can choose in every class (~..k+ - ~..r.)
a vector Yk such that
-e..
IIYk II< ll~..

~ II + 2-k- < 2-k-l.
Let x.., E ~..,. The series x,., + y 1 + y2 + converges in norm and consek+t-
..k
quently, in virtue of the completeness of X, it converges to an element x

of X. Let ~ be the class containing x. We shall prove that ~=s-lim~...
Denote by sk the partial sum x,., + y1 + y2

+ Yk of the above
series. Then lim II x - sk II = 0. On the other hand, it follows from the
~
k-->00
relations x .., E ~..,, Yp E (~..,+~ - ~..,) that sk E ~"k' and so, by (1),
- ~.. II < II x -
II-+ 0
Therefore, from the inequality II~-;,. II <
II~
sk
as
k -+ oo.
II~- ~..k II + II ~..l c - ~.. II and

the fact that {~..} is a Cauchy sequence, we conclude that lim II~-~.. II
~
=0.
12. The Partition of Unity

To discuss the support of a generalized function, in the next section,
we shall prepare the notion and existence of the partition of unity.
Proposition. Let G be an open set of R". Let a family {U} of open
subsets U of G constitute an open base of G: any open subset of G is
representable as the union of open sets belonging to the family {U}.
Then there exists a system of open sets of the family {U} with the properties:
the union of open sets of this system equals G,
(1}
any compact subset of G meets (has a non-void intersection with) only a finite number of open sets of this
system.
(2}
Definition 1. The above system of open sets is said to constitute a
scattered open covering of G subordinate to {U}.
61
12. The Partition of Unity
Proof of the Proposition. G is representable as the union of a countable

number of compact subsets. For example, we may take the system of all
closed spheres contained in G such that the centres are of rational coordinates and the radii of rational numbers.
Hence we see that there exists a sequence of compact subsets K,
such that (i) K, ~ K,+l (r = 1, 2, ... ), (ii) G is the union of K,'s and
(iii) each K, is contained in the interior of K,+l Set
U, = (the interior of K,+l) - K,_ 2 and V, = K,- (the interior of K,_ 1 ),

where for convention we set K 0 = K_ 1 = the void set. Then U, is open and
00
U V,. For any point x E V, take an open
V, is compact such that G =
r=l
set U(x; r) E {U} such that xE U(x; r)
U,. Since V,is compact, there

hr
exists a finite system of points x<ll, x<2l, ... , x<hrl such that V, ~ i~l U (x<iJ; r).
Then, since any compact set of G meets only a finite number of U,'s, it is
easy to see that the system of open sets U (x<il; r) (r = 1, 2, ... ; 1 :::=:: i < h,)
is a scattered open covering of G subordinate to {U}.
Theorem (the partition of unity). Let G be an open set of Rn, and let
a family of open sets {Gi; i E I} cover G, i.e., G = U Gi. Then there
exists a system of functions {.xj (x); f E ]} of

for each
ego (Rn)
iEI
such that
f E J,
(3)
(4)
(5)
supp (.x) is contained in some Gi,

foreveryfEJ, O<.xj(x)<1,
.I .x,-(x) == 1 for xE G.
jEJ
Proof. Let x<0 l E G and take a Gi which contains x<oJ. Let the closed
sphere S (x<0l; r) of centre x<0l and radius r be contained in Gi. We
construct, as in (14), Chapter I, 1, a function p<~~l (x) E ego(~) such that
r.
x - x<OJ I ~
(x) = 0 for
<> (x) > 0 for x - x<OJ < r ' p<>
{Jx(O)
x(O)
We put u<~~l
x
(x) =l=
= {x; p<~~)
x
0}. Then u<~~) ~ Gi and

"
"(O)EG,r>O
u<~~l
x
= G,
and, moreover, supp ({J~~l) is compact.

There exists, by the Proposition, a seattered open covering { Uj ; i E ]}
subordinate to the open base {U~~l ; x<oJ E G, r > 0} of G. Let {Jj(x) be
any function of the family {[:J<~~l (x)} which is associated with Uj.
Then, since {Uj; i E]} is a scattered open covering, only a finite number
of {Jj(x)'s do not vanish at a fixed point x of G. Thus the sum s(x) =
..I R1(x) is convergent and is> 0 at every point x of G. Hence the func-
jEJ t'
tions
<Xj (x)
Pi (x) /s (x)
satisfy the condition of our theorem.
(f E ])
62
I. Semi-norms
Definition 2. The system {lXj (x); j E l} is called a partition of unity

subordinate to the cuvering {G;; i E J}.
13. Generalized Functions with Compact Support

Definition 1. We say that a distribution T E :il (.Q)' vanishes in an
open set U of .Q if T (g;} = 0 for every g; E :il (.Q) with support contained
in U. The support of T, denoted by supp (T), is defined as the smallest
closed set F of .Q such that T vanishes in .Q- F.
To justify the above definition, we have to prove the existence of the
largest open set of .Q in which T vanishes. This is done by the following
Theorem 1. If a distribution T E :il (.Q)' vanishes in each U; of a family
{ U;; i E J} of open sets of .Q, then T vanishes in U = U U;.
iEI
Proof. Let g;E :Il(.Q) be a function with supp(g;) ~ U. We construct

a partition of unity {lXj (x); j E l} subordinate to the covering of Q
consisting of { U;; i E J} and .Q - supp (g;). Then g; =
lXj(/J is a finite
sum and so T (g;) =
i1J T (lXjf!J).
itJ
If the supp (lXj) is contained in some U;,
T(lXjf!J) = o by the hypothesis; ifthesupp (1Xj) is containedin.Q- supp (g;),

then lXjf/J = 0 and so T (lXjf!J) = 0. Therefore we have T (g;} = 0.
Proposition 1. A subset B of the space ~ (.Q) is bounded iff, for any
differential operator Di and for any compact subset K of .Q, the set of
functions {Di f (x); f E B} is uniformly bounded on K.
Proof. Clear from the definition of the semi-norms defining the topo-
logy of ~ (.Q).
Proposition 2. A linear functional T on ~ (.Q) is continuous iff T is
bounded on every bounded set of ~ (.Q).
Proof. Since ~ (.Q) is a quasi-normed linear space, the Proposition is
a consequence of Theorem 2 of Chapter I, 7.
Proposition 3. A distribution T E :il (.Q)' with compact support can
be extended in one and only one way to a continuous linear functional T(}
on ~ (.Q) such that T 0 (f) = 0 if f E ~ (.Q) vanishes in a neighbourhood
of supp(T).
Proof. Let us put supp (T) = K where K is a compact subset of .Q.
For any point x0 E K and e > 0, we take a sphere S (x 0 , e) of centre
x 0 and radius e. For any e > 0 sufficiently small, the compact set K is
covered by a finite number of spheres S (x0 , e) with x0 E K. Let {lXj (x) ;j E l}
be the partition of the unity subordinate to this finite system of spheres.
Then the function tp (x) =
I;
rxi (x), where K' is a compact
supp(<X;)nK'9=void
neighbourhood of K contained in the interior of the finite system of

spheres above, satisfies:
0 (.Q)
tp (x) E C
and
tp (x) = 1 in a neighbourhood of K.
13. Generalized Functions with Compact Support
63
We define T0 (/) for /E C""(.Q) by T0 (/) = T('lfJf). This definition is independent of the choice of 'IJJ. For, if 1p1 E ego (.Q) equals 1 in a neighbourhood
of K, then, for any f E C"" (.Q), the function (1p -1p1) f E 'Il (.Q) vanishes
in a neighbourhood of K so that T(1p/)- T(1fJd) = T((1p-1p1) f)= 0.
It is easy to see, by applying Leibniz' formula of differentiation to
1pj, that {1p/} ranges over a bounded set of 'Il (.Q) when {/} ranges over
a bounded set of ~(.Q). Thus, since a distribution T E 'Il (.Q)' is bounded
on bounded sets of 'Il (.Q), the functional T 0 is bounded on bounded sets
of ~(.Q). Hence, by the Theorem 2 of Chapter I, 7 mentioned above, T 0 is
a continuous linear functional on~ (.Q). Let f E ~ (.Q) vanish in a neighbourhood U (K) of K. Then, by choosing a 1p E ego (.Q) that vanishes in
(.Q- U(K)), we see that T0 (/) = T(1pj) = 0.
Proposition 4. Let K' be the support of 1p in the above definition of T 0
Then for some constants C and k
IT0 (f)I<C sup IDif(x)l forall /EC""(.Q).
ljj;:>;k,xEK'
Proof. Since Tis a continuous linear functional on 'Il (.Q}, there exist,
for any compact set K' of .Q, constants C' and k' such that
IT(rp)I<C' sup IDirp(x)l forall <pE'IlK'(!J)
Iii ;:>;k',xEK'
(the Corollary of Proposition 1 in Chapter I, 8). But, for any g E C00 (.Q),
we have ;k',xEK'
Iii ;:>;k',xEK'
with a constant C" which is independent of g. Setting g

we obtain the Proposition.
= f and k =
k',
Proposition 5. Let S0 be a linear functional on C"" (.Q) such that, for

some constant C and a positive integer k and compact subset K of.Q,
IS0 (/) I < C
sup
ljf;:>;k,xEK
Then the restriction of S 0 to

tained inK.
IDi f (x) I for all f E C"" (.Q).
ego (.Q) is a distribution
T with support con-
Proof. We observe that S0 (/) = 0 iff vanishes identically in a neighbourhood of K. Thus, if 1p E ego (.Q) equals 1 in a neighbourhood ofK, then
S0 (/)
S0 (1p /) for all
f E C"" (.Q) .
It is easy to see that if{/} ranges over a bounded set of 'Il (.Q), then, in
virtue of Leibniz' formula, {1pf} ranges over a set which is contained in
a set of the form
{gE C""(.Q); sup jDig(x) I= ck < oo}.
Ul;;>;k,xEK
64
I. Semi-norms
Thus 5 0 (1p f) = T (f) is bounded on bounded sets of 'Il (.Q) so that T is

a continuous linear functional on 'Il (.Q).
We have thus proved the following
Theorem 2. The set of all distributions in Q with compact support is
identical with the space ~ (.Q)' of all continuous linear functionals on
~ (Q), the dual space of ~ (.Q). A linear functional T on C00 (Q) belongs to
<r (.Q)' iff, for some constants C and k and a compact subset K of .Q,
IT (f) I :S::: C
sup
Iii ;:;;;k,xEK
IDi f (x) I for all
f E C00 (.Q) .
We next prove a theorem which gives the general expression of distributions whose supports reduce to a single point.
Theorem 3. Let an open set .Q of Rn contain the origin 0. Then the
only distributions T E 'Il (.Q)' with supports reduced to the origin 0 are
those which are expressible as finite linear combinations of Dirac's
distribution and its derivatives at 0.
Proof. For such a distribution T, there exist, by the preceding Theorem 2, some constants C and k and a compact subset K of .Q which contains the origin 0 in such a way that
IT(!) I<
sup
lil;:;;;k,xEK
IDjf(x) I for all /E C00 (.Q).
We shall prove that the condition
Dij(O)
0 for all
with
If I< k
implies T (f) = 0. To this purpose, we take a function

equal to 1 in a neighbourhood of 0 and put
f,(x)
1p
E C00 (Q) which is
f(x) 1p(xje).
We have T(f) = T(f,) since f =f. in a neighbourhood of the origin 0.

By Leibniz' formula, the derivative of f. of order < k is a linear combination of terms of the form lel-iDiVJ Dif with Iii+ If I< k. Since,
by the assumption, Di f (0) = 0 for IiI ;;;;;; k, we see, by Taylor's formula .
that a derivative of order Is I of f. is 0 (i+Hsl) in the support of
1p(xje). Thus, when e,} 0, the derivatives off, of order< k converge to 0
uniformly in a neighbourhood of 0. Hence T (f) = lim T (!,) = 0.
e.j,O
Now, for a general f, we denote by fk the Taylor's expansion off up

to the order k at the origin. Then, by what we have proved above,
This shows that T is a linear combination of linear functionals in

the derivatives of f at the origin of order < k.
65
14. The Direct Product of Generalized Functions

We first prove a theorem of approximation.
Theorem 1. Let x = (x 1 , x2 , . , Xn) ERn, y = (Yv y 2 , , Ym) E Rm
and Z=xXy=(x 1 ,x2 , ,Xn,Yl,y2 , ,ym)ERn+m. Then, for any
functionrp(z) = rp(x, y) E CZO (w+m), we can choose functions U;j (x) E CZO (R")
and functions V;j (y) E CZO (Rm) such that the sequence of functions
rp; (z) = rp;, (x, y)
k;
oo,
.I U;j (x)
V;j
J~l
(1)
(y)
to rp (z) = rp (x, y) in the topology of 'Il (Rn+m).

tends, as i--+
Proof. We shall prove Theorem 1 for the case n = m = 1. Consider
>
00
00
J J rp(~, 'f)) exp(-((x-~) 2 + (y-n) )/4t) d~d'f},

-oo -oo
= (2 Vnt)- 2
l/J(x, y, t)
(2)
0; l/J(x, y, 0) = rp(x, y).
We have, by the change of variables ~1 = (~- x) /2 Vt, 'f}1 = ('f)- y) /2 Vt,
l/J(x, y, t)
00
(V~)- 2
00
J J rp(x + 2~1 Vt,y + 21]1 Vt) e-ol-'lld~1 d'f) 1 .
-00-00
J J e-<l-rJld~1 dn 1 =
n,
00
00
00
Hence, by
00
-00-00
llP(x,y,t)-rp(x,y)l< (V~)- 2
J J lrp(x + 2~Vt,y + 2nVf)-rp(x,y)l
-00-00
X e-e'-n' d~d'f}
"S n-l t,+[~T'
+ N[,~T'}.
Since the function rp is bounded and e-o'-rJ' is integrable in R 2 , we see

The second term
that the first term on the right tends to zero as T
on the right tends, for fixed T > 0, to zero as t.} 0. Hence we have
proved that lim l/J (x, y, t) = rp (x, y) uniformly in (x, y).
too.
tj,O
Next, since supp(rp) is compact, we see, by partial integration,
= foof (2 Vnt)-2_am+krp(;,
'YJ)
e-[(x-o)'+(Y-'1)']i4t d~d1
am+k'P (x, y)
. am+kf/J (x, y, t)
11m
= oxm oy k k
oxm oy
tj,O
I . ( )
.f
um orm y m x, y .
cr+kfP(x, y,_tl
oxm oyk
-oo
a;"' or/
7'
t> 0
'
o"'+krp (x, y)
- - - - - , f=O.
8xm8yk
Thus we see as above that
(3)
It is easy to see that l/J (x, y, t) fort> 0 given by (2) may be extended
<
as a holomorphic function of the complex variables x andy for
IxI oo,
66
Semi-norms
I.
IY I < = Hence, for any given y > 0, the function 0 may be expanded into Taylor's series
oo
~
<P(x, y, t) =
2: c5 (t) xym-s
m~o s~o
which is absolutely and uniformly convergent for jxj < y,

may be differentiated term by term:
om+kiP(x, y, t)
_ ___,____ - oxm oyk
00
m,
IYI:::;: y
and
om+kxym,-s
oxm oyk
2: c. (t) ----'----;
m,~o s~o
Let {ti} be a sequence of positive numbers such that tit 0. By the above
we can choose, for each ti, a polynomial section P; (x, y) of the series
oo
m~o s~o
C8
(t) x ym-s such that

_lim Pi (x, y) = tp (x, y) in the topology of
,-700
~ (R 2 ),
that is, for any compact subset K of R 2 , lim nP;(x, y) = nrp(x, y)

i-700
uniformly on K for every differential operator n. Let us takee (x) E C0 (R 1 )

and a(y) E C0 (R 1) such that e(x) a(y) = 1 on the supp(tp (x, y)). Then
we easily see that Ti (x, y) = e(x) a (y) P; (x, y) satisfies the condition of
Theorem 1.
Remark. We shall denote by ;I) (Rn) X ;I) (Rm) the totality of functions
E ;I) (Rn+m) which are expressible as
k
..I Ti (x) "Pi (y)
J=l
with
Ti (x) E ;IJ (Rn), "Pi (y) E ;IJ (Rm) .
The above Theorem 1 says that ;I) (Rn) X ;I) (Rm) is dense in ;I) (Rn+m)
in the topology of ;I) (Rn+m). The linear subspace ;I) (Rn) X ;I) (Rm) of
;I) (Rn+m) equipped with the relative topology is called the direct product
of ;I) (Rn) and ;I) (Rm).
We are now able to define the direct product of distributions. To indicate explicitly the independent variables x = (xv x2 , . . . , xn) of the
function tp(x) E ;IJ(Rn), we shall write (;Ilx) for ;Il(Rn). We also write
(;Ily) for ;IJ(Rm) consisting of the functions 1p(y), y = (Yv y2 , . . , Ym)
Likewise we shall write (;Ilxxy) for ;I) (Rn+m) consisting of the functions
X (x, y). We shall accordingly write T(x) for the distribution T E ;I) (Rn)' =
(;Ilx)' in order to show that T is to be applied to functions tp (x) of x.
Theorem 2. Let T(x) E (;Ilx)', S(y) E (;Ily)'. Then we can define in one
and only one way a distribution W = W(xxy) E (;Ilxxy)' such that
W(u(x) v(y))
T(x)(u(x)) S(y)(v(y))
for
uE (;Ilx), vE (;Ily),
(4)
W(cp(x,y)) = S(y) (T(x)(tp(x, y))) = T(x) (S(y)(cp(x, y))) for cpE (;Ilxxy) (5)
(Fubini's theorem).
67
Remark. The distribution W is called the direct product or the tensor

T(x) and S(:vl and we shall write
product of
(6)
W = T(xJXS<:vl = S(:vJXT(xJ
Proof of Theorem 2. Let >B = {cp (x, y)} be a bounded set of the space
('Ilxx:v) For fixed y<0 l, the set {cp (x, y< 0l); tp E >B} is a bounded set of
('!>,.). We shall show that
{rp (y(O));
1j! (y<Ol)
T(x)
(cp (x, y<Ol)), tp E >B}
(7)
is a bounded set of ('Il:v<oJ). The proof is given as follows.

Since >B is a oounded set of ('Ilxx:v), there exist a compact set K.,ER"
and a compact set K:v E Rm such that
supp(cp) ~ {(x, y) E R"+m; xE Kx, yE K:v}
Hence y<Ol f. K:v implies rp (x, y< 0 l)
Thus
supp (1J!) ~ K:v
whenever
cpE >B.
= 0 and 1J! (y<0 l) = T(x) (cp (x, y<0l)) = 0.

whenever
cp E >B.
(8)
We have to show that, for any differential operator Dy in Rm,

supjDy1J!(Y)\ <oo where 1J!(Y)=T(x)(tp(x,y)), cpE>B.
:v;p
To prove this, we take, e. g., D:v, = 8f8y 1 . Then, by the linearity of
1J'(YI
_ T
+ h, Y2 , Ym) -1J'(YI Y2 , Ym)
Jcp(x, Y1
(x)
m,
(9)
T(x)
h, Y2 , Ym)- cp(x, Y1 Y2. ., Ym)}
the functions E { } of X, with parameters y E Rm

When tp ranges over
and h such that jh I < 1, constitute a bounded set of ('Ilx) This we see
from the fact that >B is a bounded set of ('Ilxx:v) Hence we see, by letting
h-+ 0 and remembering Proposition 1 in Chapter I, 8, that (9) is true.
Therefore, by the same Proposition 1, we see that the set of values
{S<:vl(T<xl(cp(x,y))); cpE )S}
(10)
is bounded. Consequently, the same Proposition 1 shows that we have

defined a distribution w<ll E ('Ilx x:v)' through
w<1 l (cp) = S<:vl (T<xl (cp (x, y))).
Similarly we define a distribution w< 2l E ('Ilxx:v)' through
(11)
w< 2l(cp) = T<xl(S<:vl(cp(x,y))).

Clearly we have, for u E ('Ilx) and v E ('Il:v),
(12)
(13)
w<1l(u(x) v(y)) = w< 2l(u(x) v(y)) = T<xl(u(x)) S<:vl(v(y)).
Therefore, by the preceding Theorem 1 and the continuity of the
distributions w<1l and w< 2l, we obtain w<1l = w< 2l. This proves our
Theorem 2 by setting W = w< 1l = w< 2l.
5*
68
II. Applications of the Baire-Hausdorff Theorem
References for Chapter I

For locally convex linear topological spaces and Banach spaces, see
N. BouRBAKI [2], A. GROTHENDIECK [1], G. KoTHE [1], S. BANACH [1],
N. DuNFORD-]. ScHWARTZ [1] and E. HILLE-R. S. PHILLIPS [1]. For
generalized functions, see L. ScHWARTZ [1], I. M. GELFAND-G. E. SILOV
[1], L. HORMANDER [6] and A. FRIEDMAN [1].

The completeness of a B-space (or an F-space) enables us to apply
the Baire-Hausdorff theorem in Chapter 0, and we obtain such basic
principles in functional analysis as the uniform boundedness theorem, the
resonance theorem, the open mapping theorem and the closed graph theorem.
These theorems are essentially due to S. BANACH [1]. The termwise
differentiability of generalized functions is a consequence of the uniform
boundedness theorem.
1. The Uniform Boundedness Theorem and the Resonance

Theorem
Theorem 1 (the uniform boundedness theorem). Let X be a linear
topological space which is not expressible as a countable union of closed
non-dense subsets. Let a family {Ta; a E A} of continuous mappings be
defined on X into a quasi-normed linear space Y. We assume that, for
any aE A and x, yE X,
/ITa(x
+ y) II< II Tax II+
IITaYII and IITa(ax) II= IIa Tax II for a> 0.
If the set {Tax; aE A} is bounded at each xE X, then s-lim Tax= 0

x---+0
uniformly in a E A.
Proof. For a given s > 0 and for each positive integer n. consider
Xn = fx EX; sup {lln-1 Tax II+ lln-1 T,. (- x)ll}::::;;: sf\. Each set Xn is
aEA
closed by the continuity of Ta. By the assumption of the boundedness of

00
{iiTaxll;aEA},wehaveX= U Xn. Hence, by the hypothesis on

n~l
X, some Xn. must contain a neighbourhood U = x 0 + V of some point

x 0 EX, where V is a neighbourhood of 0 of X such that V = - V.
Thus x E V implies sup II n01 Ta (x0 + x) II < s. Therefore we have
aEA
II T,.(nol x) II= II Ta(nol (xo
+ llno
+ x- xo)) II<
llnolTa (xo
+ x) II
Ta(-xo) II< 2s for xE V, aE A.
Thus the Theorem is proved, because the scalar multiplication ax in a

linear topological space is continuous in both variables a and x.
1. The Uniform Boundedness Theorem and the Resonance Theorem
69
Corollary 1 (the resonance theorem). Let {Ta; a E A} be a family of

bounded linear operators defined on a B-space X into a normed linear
spaceY. Then the boundedness of {/I Tax II; a E A} at each x EX implies
the boundedness of {!!Tall; aE A}.
Proof. By the uniform boundedness theorem, there exists, for any
e>O, a b>O such that llxll<b implies supiiTaxll<e. Thus

aEA
sup/ITall<ejb.
aEA
Corollary 2. Let {T,.} be a sequence of bounded linear operators defined

on aB-space X into a normed linear spaceY. Suppose that s-lim T,.x = Tx
n---+oo
exists for each x EX. Then Tis also a bounded linear operator on X into
Y and we have
(1)
liT II< lim IIT,.II
n-->00
Proof. The boundedness of the sequence {II T ,.x II} for each x E X is
implied by the continuity of the norm. Hence, by the preceding Corollary, sup IIT,.II<oo,and so IIT,.xll<supiiT,.IIIIxll (n=1,2, ... ).
n;;;;l
n;;;;l
Therefore, again by the continuity of the norm, we obtain
which is precisely the inequality (1). Finally it is clear that T is linear.

Definition. The operator T obtained above is called the strong limit
of the sequence {T,.} and we shall write T = s-lim T,..
n---+oo
We next prove an existence theorem for the bounded inverse of a

bounded linear operator.
Theorem 2 (C. NEUMANN). Let T be a bounded linear operator on a
B-space X into X. Suppose that III- Til< 1, where I is the identity
operator: I x = x. Then T has a unique bounded linear inverse T- 1
which is given by C. Neumann's series
T- 1 x =s-lim (I+ (I-T)+ (I -T) 2 ++(I- T)")x, xE X.

n---+oo
(2)
Proof. For any x EX, we have
II,.~ (I-T)" xll < n~ II (I- T)"xll < ,.~ II(I- T)"llllxll
s
00
.I III- Til"
n=O
The right hand side is convergent by
llxll.
III- Til< 1. Hence, by the comk
pleteness of X, the bounded linear operator s-lim .I (I-Tt is defined.

lv--KJo n=O
70
It is the inverse of T as may be seen from
T s-lim
1: (I-Tt x =s-lim (I- (I-T)) ( 1: (I-Tt x)
n=O
n=O
k-+00
s-lim (I - T)k+l
X -
and the similar equation s-lim
k->oo
.J: (I-
n=O
X,
rt) T x = x.
'
2. The Vitali-Hahn-Saks Theorem

This theorem is concerned with a convergent sequence of measures,
and makes use of the following
Proposition. Let (5, QJ, m) be a measure space. Let Q30 be the set
of all BE Q3 such that m (B) < oo. Then by the distance
d(B1 , B 2)
m(B1
B 2), where B 1
B2
= B1
V B 2 - B 1 f\ B 2 ,
(1)
}80 gives rise to a metric space (Q30 ) by identifying two sets B 1 and B 2
of }80 when m(B1 e B 2) = 0. Thus a point B of (Q30 ) is the class of sets
B 1 E }80 such that m(B e B 1) = 0. Under the above metric (1), (Q30 )
is a complete metric space.
Proof. If we denote by CB (s) the defining function of the set B:

C B (s)
1 or 0 according as
we have
s E B or s E B,
IC B, (s) - C 8 , (s) I m (ds).

(2)
s
Thus the metric space (Q30) may be identified with a subset of the B-space
Ll(5, QJ, m). Let a sequence {CBn(s)} with BnE Q30 satisfy the condition
d (BI> B2)
lim d(Bn,Bk)=lim
n,k--+00
fiC 8 n(s)-CBk(s)lm(ds)=0.
n,k--+00 5
Then, as in the proof of the completeness of the space Ll (5, QJ, m), we can
choose a subsequence {CBn,(s)} such that J~ CBn,(s) = C(s) exists
m-a.e. and
J IC (s) -
n~s
CBn (s) I m (ds) = 0. Clearly C (s) is the
defining function of a set B 00 E Q30 , and hence lim d (B00 , En) = 0.

1>-+00
This proves that (Q30 ) is a complete metric space.

Theorem (VITALI-HAHN-SAKS). Let (5, QJ, m) be a measure space, and
(A.n (B)} a sequence of complex measures such that the total variations
[A.n [(5) are finite for n = 1, 2, ... Suppose that each An (B) ism-absolutely
continuous and that a finite lim An (B) = A(B) exists for every B E Q3.
n-+00
Then the m-absolute continuity of An (B) is uniform in n, that is,

lim m (B) = 0 implies lim An (B) = 0 uniformly in n. If m (5) < oo, then
A.(B) is a-additive on }8.
71
2. The Vitali-Hahn-Saks Theorem
Proof. Each An defines a one-valued function Xn (B) on (580 ) by

ln (B) = },n (B), since the value An (B) is, by the m-absolute continuity
of An(B), independent of the choice of the set B from the class B. The
continuity of X" (B) is implied by and implies them-absolute continuity of
An (B).
Hence, for any e > 0, the set
Fk (e)
IS
lfjj; sup
nGl
jJk (B)- J.k+n (B) j < e~

J
closed in (580 ) and, by the hypothesis lim An (B) = A(B), we have

~
00
(580 ) = U Fk (e). The complete metric space (58 0 ) being of the second
k=l
category, at least one Fk, (e) must contain a non-void open set of (580).
This means that there exist a B0 E (580 ) and an 'YJ > 0 such that
d (B, B 0 )
<
'YJ
implies sup [lk, (B)- ik,+n (B) [ <e.

nGl
On the other hand, any B E 580 with m (B) < 'YJ can be represented as
B = B 1 - B 2 with d(B 1 , B 0 ) < 'YJ, d(B 2 , B 0 ) < 'YJ We may, for example,
take B 1 = B V B 0 , B 2 = B 0 -B (\ B 0 Thus, if m(B) < 'YJ and k > k0 ,
we have
I + [Ak, (B)- Ak (B) I

< IAk, (B) I + IAk, (Bl) - Ak (Bl) I + IAk, (B2) -
[Ak (B) I < [Ak, (B)
Ak (B2) I
< [Ak,(B) [ + 2e.
Therefore, by the m-absolute continuity of Ak, and the arbitrariness

of e > 0, we see that m (B) ---+ 0 implies An (B) ---+ 0 uniformly inn. Hence,
in particular, m (B)--+ 0 implies A(B)--+ 0. On the other hand, it is clear
that A is finitely additive, i.e.,
lim A(B)
m(B)-..0
m(S)
<
.i A(Bj)
I Bi) = j=l
A(. J=l
Thus, by
0 proved above, we easily see that A is a-additive if
oo.
Corollary 1. Let {An (B)} be a sequence of complex measures on S such

that the total variation [An [(5) is finite for every n. If a finite lim An (B)
n-..0()
exists for every BE 78, then the a-additivity of {An(B)} is uniform inn,
in the sense that, lim An (Bk) = 0 uniformly in n for any decreasing
lv-+OO
sequence {Bk} of sets E 78 such that

Proof. Let us consider
00
n Bk = 0.
k=l
72
The a-additivity of m is a consequence of that of the A/s, and we have

0 < m (B) < 1. Each flj and hence each Aj is m-absolutely continuous.
Thus, by the above Theorem, we have lim A, (Bk) = 0 uniformly in n,
k---+00
because lim m (Bk) = 0.

k---+00
Corollary 2. A(B) in the Theorem is a-additive and m-absolutely

continuous even if m(S) = oo.
3. The Termwise Differentiability of a Sequence

of Generalized Functions
The discussion of the convergence of a sequence of generalized functions is very simple. We can in fact prove
Theorem. Let {Tn} be a sequence of generalized functions E :tl (.Q)'.

Let a finite lim Tn (tp) = T (tp) exist for every tp E :tl (.Q). Then Tis also a
~00
generalized function E ;tl(.Q)'. We say then that Tis the limit in :tl(.Q)'
of the sequence {T,} and write T = lim Tn(:tl (.Q)').
n---+OO
Proof. The linearity of the functional Tis clear. Let K be any compact
subset of .Q. Then each Tn defines a linear functional on the F-space
:tlK(.Q). Moreover, these functionals are continuous. For, they are bounded
on every bounded set of :tlK(.Q) as may be proved by Proposition 1 in
Chapter I, 8. Thus, by the uniform boundedness theorem, T must be a
linear functional on :tlK (.Q) which is bounded on every bounded set of
'l)K (.Q). Hence Tis a continuous linear functional on every :tlK (.Q). Since
:tl (.Q) is the inductive limit of :tlK (.Q)'s, T must be a continuous linear
functional on :tl (.Q).
Corollary (termwise differentiability theorem). LetT= lim T n (:tl (.Q)').
n---+00
Then, for any differential operator Dj, we have Dj T = lim Dj T n (:tl (.Q)').
Proof.
n---+00
lim Tn = T(:tl(.Q)') implies that lim Tn ((-1)ijiDjtp) =
n---+00
n---+00
T((-1)ijiD;tp) for every tp E :tl (.Q). Thus we have
(D;T) (tp)
lim (DjTn) (tp)

n---+00
for every tp E :tl (.Q).

The Baire-Hausdorff theorem may be applied to prove the existence of
a function with various singularities. For instance, we shall prove the
existence of a continuous function without a finite derivative.
Weierstrass' Theorem. There exists a real-valued continuous function

x (t) defined on the interval [0, 1] such that it admits a finite differential
quotient x' (t0 ) at no point t0 of the interval [0, 1/2].
73
Proof. A function x (t) admits finite upper- and lower-derivatives on

the right at t = t0 iff there exists a positive integer n such that
sup h-1 lx(t0 +h)- x(t0 } I< n.
2- 1 >h>O
Let us denote by M,. the set of all functions x(t) E C [0, 1] such that
x (t) satisfies the above inequality at a certain point t0 of [0, 1/2]; here
the point t0 may vary with the function x (t). We have only to show that
00
eachM,.is a non-dense subset of C [0, 1]. For, then, the set C [0, 1]- U M,.
n=l
is non-void since; by the Baire-Hausdorff theorem, the complete metric

space C [0, 1] is not of the first category.
Now it is easy to see, by the compactness of the closed interval [0, 1/2]
and the norm of the space C [0, 1], that M,. is closed in C [0, 1]. Next,
for any polynomial z (t) and e > 0, we can find a function y (t) E C[O, 1] -M,.
such that sup lz(t)- y(t) I= liz- y II :S::: e. We may take, for exo~t~1
ample, a continuous function y (t) represented graphically by a zig-zag

line so that the above conditions are satisfied. Hence, by Weierstrass' polynomial approximation theorem, M,. is non-dense in C [0, 1].
S. BANACH [1] and H. STEINHAUS have proved a principle of condensation of singularities given below which is based upon the following
Theorem (S. BANACH). Given a sequence of bounded linear operators

{T,.} defined on a B-space X into a normed linear space Y ,.. Then the set
B ={xEX; ~IIT,.xll <oo}

either coincides with X or is a set of the first category of X.
Proof. We shall show that B =X under the hypothesis that B is of
the second category. By the definition of B, we have lim sup II k-1 T,. x II
Hoc> n;;;;l
= 0 whenever x E B. Thus, for any e > 0,

00
B ~ U Bk where Bk = flx EX; sup llk-1 T,.xll < e}

n:;;;1
k=1
Each Bk is a closed set by the continuity of the T ,.'s. Hence, if B is of the

second category, then a certain Bk. contains a sphere of positive radius.
That is, there exist an x0 E X and a <5 > 0 such that 1Jx-x0 JJ < <5
implies supJJk01 T,.xJJ <e. Hence, by putting x-x0 =y, we obtain,
n;;;>:1
for I!YII
thus
< <5,
Jlk01 T,.yJJ < JJk01 T,.xl!

sup
n:;;;l,ll:ll ~ko'"
and soB= X.
+ l!k01 T,.x0 JJ <
IJT,.zi!:S:::2e,
2e. We have
74
Corollary (the principle of the condensation of singularities). Let

{Tp,q} (q = 1, 2, ... ) be a sequence of bounded linear operators defined
on a B-space X into a normed linear space Yp (p = 1, 2, ... ). Let, for
Then the set
each p, there exist an Xp E X such that lim Jl Tp qXp II =
q-->OO
oo.
'
E={xEX;q~IITp,qxll=oo forall P=1,2, ...

is of the second category.
Proof. For each
p,
the set Bp
= lxEX;
l
lim IITpqxll
q-->00
'
< oo}
is,
by the preceding Theorem and the hypothesis, of the first category.

00
Thus B = X- U Bp must be of the second category.

P=1
The above Corollary gives a general method of finding functions with

many singularities.
Example. There exists a real-valued continuous function x (t) of
period 2:n: such that the partial sum of its Fourier expansion:
/q(x; t)
.I (ak cos kt
k=O
where Kq (s, t) = sin((q
+ bk sin kt) = J x(s) Kq(s, t) ds,

n _,
n
+ 2-1) (s- t))/2 sin 2-1 (s- t),
(1)
satisfies the condition that

lim J/q(x; t) I=
q-->OO
oo
on a set P
s:;
[0, 2:n:] which is of the
power of the continuum.
(2)
Moreover, the set P may be taken so as to contain any countable sequence

~ [0, 2:n:].
{tj}
Proof. The totality of real-valued continuous functions x (t) of period

2:n: constitutes a real B-space C2,. by the norm II x II = sup I x (t) I
0~1~2n
As may be seen from (1), /q(x; t) is, for a given t0 E [0, 2:n:], a bounded
linear functional on C2,. Moreover, the norm of this functional /q(x; t0 )
is given by
! JIKq(s, t
0)
Ids= the total variation of the function
-:n
(3)
-;; -nJ Kq (s, t0 ) ds.

I
It is also easy to see that, for fixed t0 , (3) tends, as q--+ oo, to oo.
Therefore, if we take a countable dense sequence {ti} ~ [0, 2:n:], thet;,
by the preceding Corollary, the set
B=lxEC2,; lim l/q(x;t)l=oo
q->00
for
t=tvt2 ,
.1
75
5. The Open Mapping Theorem
is of the second category. Hence, by the completeness of the space C2 ,.,

the set B is non-void. We shall show that, for any x E B, the set
p
lim l/q(x; t) I= oo'tf.

= ftl E [0, 2n]; q___,.oo
is of the power of the continuum. To this purpose, set
Fm,q = {tE [0, 2n]; l/q(x; t) I< m},
Fm =
00
q~l
Fm,q
By the continuity of x(t) and the trigonometric functions, we see

that the set F m,q and hence the set F m are closed sets of [0, 2 n]. If we
00
can show that U Fm is a set of the first category of [0, 2n], then the
m~l
set P
=(co, 2n] -mQl Fm) 3g}
would be of the second category.
Being a set of the second category of [0, 2n], P cannot be countable and
so, by the continuum hypothesis, P must have the power of the continuum.
Finally we will prove that each F m is a set of the first category of
[0, 2n]. Suppose some Fm, be of the second category. Then the closed
set Fm, of [0, 2n] must contain a closed interval [<X, {J] of [0, 2n]. This
implies that sup l/q(x; t) I :S m 0 for all t E [<X, {J], contradicting the fact
q~l
that the set P contains a dense subset {tj} of [0, 2n].

Remark. We can prove that the set P has the power of the continuum without appealing to the continuum hypothesis. See, for
example, F. HAUSDORFF [1].
5. The Open Mapping Theorem

Theorem (the open mapping theorem of S. BANACH). LetT be a continuous linear operator on anF-spaceX onto anF-space Y. Then T maps
every open set of X onto an open set of Y.
For the proof, we prepare
Proposition. Let X, Y be linear topological spaces. Let T be a continuou~ linear operator on X into Y, and suppose that the range R (T)
of Tis a set of the second category in Y. Then, to each neighbourhood
U of 0 of X, there corresponds some neighbourhood V of 0 of Y such that
V <;;; (TUt.
Proof. Let W be a neighbourhood of 0 of X such that W = - W,
W + W <;;; U. For every x EX, xjn--?- 0 as n--?- oo, and so x En W for
00
large n. Hence X= U (nW), and therefore R(T)

n~l
00
U T(nW). Since
n~l
R (T) is of the second category in Y, there is some positive integer n0 such

that (T (n0 W)t contains a non-void open set. Since (T (n W)/ = n (T (W)t
16
and since n(T(W)Y and (T(W)Y are homeomorphic 1 to each other,

(T(W)Y also contains a non-void open set. Let y0 = Tx0 , x0 E W, be
a point of this open set. Then, since the mapping x--+- x0 + x is a
homeomorphic mapping, we see that there exists a neighbourhood V
of 0 of Y such that V s;;- y0 + (T(W)y. The elements of -y0 + T(W)
are expressible as -y0 + Tw = T(w- x0 ) with wE W. But w- x0 E
W + W s;; U, since W is a neighbourhood of 0 of X such that
W = - W and W + W s;; U.
Therefore, -y0 + T(W) s;; T(U) and hence, by taking the closure,
-y0 + (T(W)t s;; (T(U)t and so V s;; -y0 + (T(W)t s;; (T(U)t = (T Ut.
Proof of the Theorem. Since Y is a complete metric space, it is of the
second category. Thus, by the preceding Proposition, the closure of the
image by T of a neighbourhood of 0 of X contains a neighbourhood of 0
of Y.
Let X., Y, denote the spheres in X, Y respectively, with centres

at the origins and radii s > 0. Let us set si = sf2i (i = 0, 1, 2, ... ).
Then by what we have stated above, there exists a sequence of positive
numbers {17i} such that
(i = 0, 1, 2, ...) .
(1)
Let y E yflo be any point. We shall show that there is an X E x2., such
that T x = y. Frorr~ (1) with i = 0, we see that there is an x0 E X,, such
that IIY- Tx0 II< 171 . Since (y- Tx0 ) E Y 11,, we see again from (1)
with i = 1 that there is an x 1 EX,, with IIY- Tx0 - Tx1 ll < 17 2 Repeating the process, we find a sequence {xi} with xi EX,; such that
(n = 0, 1, 2, ... ) .
We have
II k=m+l
i: xkll
so the sequence
< k=m+l
i: llxk II< k=m+l
i: sk < (k=m+l
i: 2-k) s0 ,
and
t~ xk} is a Cauchy sequence. Hence, by the complete-
ness of X, we see that s-lim
n--->00
k=O
xk = x EX exists. Moreover, we have
llxll = lim IIi:

xk II< n--->OOk=O
lim i: llxk II< (i
2-k)so = 2s0
k=O
k=O
n--->00
We must have y = Tx since Tis continuous. Thus we have proved that

any sphere X 2,, is mapped by T onto a set which contains a sphere Y 11,.
After these preliminaries, let G be a non-void open set in X and
x E G. Let U be a neighbourhood of 0 of X such that x + U s;; G. Let V
1 A one-to-one mapping M of a topological space 5 1 onto a topological
space 5 2 is called a homeomorphic mapping If M and M-1 both map open
sets onto open sets.
6. The Closed Graph Theorem
77
be a neighbourhood of 0 of Y such that T U ~ V. Then T G ~ T (x + U) =

Tx + TU ~ Tx + V. Hence TG contains a neighbourhood of each one
of its point. This proves that T maps open set of X onto open sets of Y.
Corollary of the Theorem. If a continuous linear operator T on an
F -space onto an F -space gives a one-to-one map, then the inverse operator
T-1 is also a linear continuous operator.
Definition 1. Let X and Y be linear topological spaces on the same
scalar field. Then the product space X X Y is a linear space by
{xv Y1} + {x2, Y2} = {xl + X2, Yt + Y2}, X{x, Y} ={XX, iXY}
It is also a linear topological space if we call open the sets of the form
G1 XG2 ={{x, y}; xE Gv yE G2},
where Gv G2 are open sets of X, Y respectively. If, in particular, X and

Y are quasi-normed linear spaces, then Xx Y is also a quasi-normed
linear space by the quasi-norm
(1)
Proposition 1. If X andY are B-spaces (F-spaces), then Xx Y is also
a B-space (F-space).
Proof. Clear since s-lim {xn, Yn} = {x, y} is equivalent to s-lim Xn = X
n->00
s-lim Yn
n->00
y.
Definition 2. The graph G (T) of a linear operator T on D (T) ~ X into

Y is the set {{x, Tx}; x E D(T)} in the product space Xx Y. Let X, Y be
linear topological spaces. Then T is called a closed linear operator when
its graph G (T) constitutes a closed linear subspace of X X Y. If X and Y
are quasi-normed linear spaces, then a linear operator T on D (T) ~ X
into Y is closed iff the following condition is satisfied:
{xn}
D(T), s-lim Xn = x
and
s-lim Txn = y
xE D(T)
and
Tx
n->00
n->00
imply that
y.
(2)
Thus the notion of a closed linear operator is an extension of the notion

of a bounded linear operator. A linear operator Ton D (T) ~ X into Y is
said to be closable or pre-closed if the closure in X X Y of the graph G (T)
is the graph of a linear operator, say S, on D (S) ~ X into Y.
Proposition 2. If X, Y are quasi-normed linear spaces, then T is
closable iff the following condition is satisfied:
{xn}
D(T), s-lim Xn = 0 and s-lim Txn = y imply that y = 0.

n--+00
n->00
{3)
78
Proof. The "only if" part is clear since the closure in Xx Y of G(T)
is a graph G (5) of a linear operator 5 and soy = 5 0 = 0. The "if" part
is proved as follows. Let us define a linear operator 5 by the following
condition and call 5 the smallest closed extension of T:
xED (5) iff there exists a sequence {xn}

s-lim
Xn
n->00
= x and s-lim
n->00
D (T) such that
T Xn = y exist; and we define
(4)
5x=y.
That the value y is defined uniquely by x follows from condition (3).
We have only to prove that 5 is closed. Let wn ED (5), s-lim wn =wand
s-lim 5wn = u. Then there exists a sequence {xn}
n->00
n->00
D (T) such that
llwn-xnll < n-1 , IISwn- Txnll < n-1 (n = 1, 2, ... ). Therefore

s-lim Xn =s-lim wn = w, s-lim T Xn =s-lim 5wn = u, and so w (: D (5),
~
n-KX)
n-KX)
~00
SW=U.
An example of a discontinuous but closed operator. Let X = Y =
C [0, 1]. Let D be the set of all x (t) E X such that the derivative x' (t) E X;
and let T be the linear operator on D (T) = D defined by T x = x'. This
Tis not continuous, since, for xn(t) = tn,
llxnll=1,jjTxnll= sup lx~(t)!= sup jntn- 1 l=n

0~1~1
(n=1,2, ... ).
0;?;;1~1
ButT is closed. In fact, let {xn} t;;;: D (T), s-lim Xn = x and s-lim T xn = y.
n--+00
n->00
Then x~ (t) converges uniformly to y (t), and xn(t) converges uniformly

to x(t). Hence x(t) must be differentiable with continuous derivative
y(t). This proves that xE D(T) and Tx = y.
Examples of closable operators. Let Dx be a linear differential operator
Dx =
.
_I ci(x) Di
JjJ~k
(5)
with coefficients cj(x) E Ck(.Q), where.Q is an open domain of Rn. Consider

the totality D offunctions f (x) E L 2 (.Q) f\ Ck (.Q) such that Dxf (x) E L 2 (.Q).
We define a linear operator T on D (T) = D ~ L 2 (.Q) into L2 (.Q) by
(Tf) (x) = Dxf(x). Then T is closable. For, let {ih} ~ D be such that
s-limfn=O, s-limDxfn=g. Then, for any IJ!(x)EC~(.Q), we have, by
h-+00
h->00
partial integration,
JDxf(x) IJ!(X) dx = J f(x) D~IJ!(x) dx,
ll
ll
(6)
where D: is the differential operator formally adjoint to Dx:
D:IJ!(X)
= .2:
iJ~k
(-l)lil Di(cj(x) IJ!(x)).
(7)
79
The formula (6) is obtained, since the integrated tPrm in the partial
integration vanishes by rp (x) E C~ (Q). Hence, by the continuity of the
scalar product in P(Q), we obtain, by taking f = fh and letting h ___,. =
in (6),
(8)
g(x) rp(x) dx =
0 D~rp(x) dx = 0.
Since rp(x) E C~(Q) was arbitrary, we must have g(x) = 0 a.e., that is,
= 0 in P(Q).
Proposition 3. The inverse T-1 of a closed linear operator on D (T)

into Y, if it exists, is a closed linear operator.
~X
Proof. The graph of T-1 is the set {{T x, x}; x E D (T)} in the product
space Y X X. Thus the Proposition is proved remembering the fact that
the mapping {x, y} ___,. {y, x} of Xx Y onto YxX is a homem:wrphic one.
We next prove Banach's closed graph theorem:
Theorem 1. A closed linear operator defined on an F-space X into an
F-space Y is continuous.
Proof. The graph G (T) of Tis a closed linear subspace of the F-space
X X Y. Hence, by the completeness of X X Y, G (T) is an F-space. The
linear mapping U defined by U{x, Tx} =xis a one-to-one, continuous
linear transformation on the F-space G(T) onto the F-space X. Hence,
by virtue of the open mapping theorem, the inverse u-1 of u is continuous.
The linear operator V defined by V{x, Tx} = Tx is clearly continuous
on G (T) onto R (T) ~ Y. Therefore, T = V U-1 is continuous on X into Y.
The following theorem of comparison of two linear operators is due to
L. H6RMANDER:
Theorem 2. Let Xi (i = 0, 1, 2; X 0 = X) be B-spaces and Ti (i = 1, 2)
be linear operators defined on D (Ti) <;;;X into Xi. Then, if T 1 is closed
and T 2 closable in such a way that D(T1 ) <;;; D(T2 ), there exists a constand C such that
/IT2 xiJ<C(jjT1 xJI2+ llxll 2 ) 1' 2 forall xED(T1 ).

(9)
Proof. The graph G(T1 ) of T 1 is a closed subspace of XxX1 . Hence
the mapping
(10)
is a linear operator on the B-space G(T1 ) into the B-space X 2 We shall
prove that this mapping is closed. Suppose that {x,., T 1 x,.} s-converges
in G (T1 ) and that T 2 x,. s-converges in X 2 Since T 1 is closed, there is an
element xED (T1 ) such that x =s-lim x,. and T 1 x =s-lim T 1 xw By the
,__,.00
hypothesis, we have x E D (T2), and, since T 2 is closable, the existing

s-lim T 2 x,. can only be T 2 x. Hence the mapping (10) is closed, and so,
~
by the closed graph theorem, it must be continuous. This proves (9).
80
7. An Application of the Closed Graph Theorem

(Hormander's Theorem)
Any distribution solution u E L 2 of the Laplace equation
Llu = /E L2
is defined by a function which is equal to a C00 function after correction
on a set of measure zero in the domain where f is C00 This result is known
as Weyl's Lemma and plays a fundamental role in the modern treatment
of the theory of potentials. See H. WEYL [1]. There is an abundant
literature on the extensions of Weyl's Lemma. Of these, the research
due to L. H6RMANDER [1] seems to be the most far reaching. We shall
begin with his definition of hypoellipticity.
Definition I. Let.Qbe an open domain ofRn. A function u(x), xE.Q,
is said to belong to Lfoc {.Q) if I u (x) 12 dx < = for any open subdomain
Q'
.Q' with compact closure in .Q. A linear partial differential operator P (D)
with constant coefficients:

1 -88 , ---o1 8 , ... ,--:-1 88) , where
P(D) = P ( ---o8X2
~
Xl
~
~
Xn
(~)
= P (~v ~2 ,
, ~n)
is a polynomial in
(1)
~v ~ 2 , , ~n,
is said to be hypoelliptic, if every distribution solution u E Lfoc (.Q) of

f is C00 after correction on a set of measure zero in the subdomain where f is C00
P (D) u =
Theorem {H6RMANDER). If P (D) is hypoelliptic, then there exists, for

any large positive constant cl> a positive constant c2 such that, for any
solution C= ~ + i r; of the algebraic equation P (C) = 0, we have
ICI =
C4' !Cjl
r
2
~ cl
lr;l
if
c~ lr;jl 2
r
2
<
C2.
(2)
Proof. Let U be the totality of distribution solutions u E L2(.Q') of

P (D) u = 0, that is, the totality of u E L2 (.Q') such that
f u P' (D) rp dx =
for all
rp E C~ (.Q'),
(3)
[)'
where the adjoint differential operator P' (D) of P (D) is defined by
P' (~)
P (-~v -~2
.,
-~n)
(4)
We can prove that U is a closed linear subspace of L 2 (.Q'). The linearity

of U is obvious from the linearity of the differential operator P(D). Let
a sequence {uh} of U be such that s-lim uh = u in L2(.Q'). Then, by the
h--->00
continuity of the scalar product in L2(.Q'), we have

0 =lim
Juh P'(D) rp dx = Ju P'(D) rpdx =
h--+00 [)'
[)'
0, i.e., uE U.
81
1. The Orthogonal Projection
Thus U is a closed linear subspace of L2 (Q'), and as such a B-space.

Since P (D) is hypoelliptic, we may suppose that every u E U is C00
in Q'. Let Q~ be any open subdomain with compact closure in Q'. Then,
for any u E U, the function oufoxk is C00 in Q' (k = 1, 2, ... , n). By the
argument used in the preceding section, the mapping
U ;:Ju~ 88u E L 2 (Qi_)
(k = 1, 2, ... , n)
xk
is a closed linear operator. Hence, by the closed graph theorem, there

exists a positive constant C such that
J.I I-
.Q~
n f)u
k=l 8xk
12 dx < C
Jlul
dx,
for all u E U.
!J'
If we apply this inequality to the function u(x) = ei(x,C>, where C=

+ i'YJ = (~1 + i"'l ~2 + i'YJ2 ... , ~n + i'Y}n) is a solution of P(C) = 0
and <x, C) = j~ <xj, Cj), we obtain
Therefore, when I'YJ I is bounded, it follows that
CI must be bounded.
Remark. Later on, we shall prove that condition (2) implies the
hypoellipticity of P(D). This result is also due to H6RMANDER. Thus, in
particular, we see that Weyl's Lemma is a trivial consequence of Hormander's result. In fact, the root of the algebraic equationsatisfies (2).
.I CJ
J=l
= 0
References for Chapter II

S. BANACH [1], N. BOURBAKI [2], N. DUNFORD-}. SCHWARTZ [1],
E. HILLE-R. S. PHILLIPS [1] and L.HORMANDER (6].
III. The Orthogonal Projection and F. Riesz' Representation

Theorem
In a pre-Hilbert space, we can introduce the notion of orthogonality
of two vectors. Thanks to this fact, a Hilbert space may be identified
with its dual space, i.e., the space of bounded linear functionals. This
result is the representation theorem of F. Riesz [1], and the whole theory
of Hilbert spaces is founded on this theorem.
6 Yosida, Functional Analysis
82
III. The Orthogonal Projection and F. Riesz' Representation Theorem
Definition 1. Let x, y be vectors of a pre-Hilbert space X. We say that

xis orthogonal to y and write x ..l y if (x, y) = 0; if x j_ y then y .l_ x,
and x ..l x iff x = 0. Let M be a subset of a pre-Hilbert space X. We
denote by M .L the totality of vectors E X orthogonal to every vector m
ofM.
Theorem 1. Let M be a closed linear subspace of a Hilbert space X.
Then M .L is also a closed linear subspace of X, and M .L is called the
orthogonal complement of M. Any vector x EX can be decomposed
uniquely in the form
x= m
+ n,
(1)
where mE M and n EM .L.
The element min (1) is called the orthogonal projection of x upon M and
will be denoted by PM x; PM is called the projection operator or the
projector upon M. We have thus, remembering that M ~ (M.l).L,
x
PMx +PM .LX, that is, I= PM+ PM.L.
(1')
Proof. The linearity of M .L is a consequence of the linearity in x of

the scalar product (x, y). M.L is closed by virtue of the continuity of the
scalar product. The uniqueness of the decomposition (1) is clear since a
vector orthogonal to itself is the zero vector.
To prove the possibility of the decomposition (1), we may assume
that M =I= X and x EM, for, if x EM, we have the trivial decomposition
with m = x and n = 0. Thus, since M is closed and x EM, we have
d
= mEM
inf /1 x -
II >
0.
Let {mn} ~ M be a minimizing sequence, i.e., lim llx- mn II= d. Then

n-co
{mn}isaCauchysequence. For, by /Ia + b 11 2 +I la-b 11 2 = 2([1a 11 2 +II b 11 2 )

valid in any pre-Hilbert space (see (1) in Chapter I, 5), we obtain
II mk -
mn 11 2
II (x -
mn) - (x- mk) 11 2
= 2 (II X -
mn 11 2
+ II X -
mk 11 2)
-ll2x-mn-mki! 2
= 2(llx- mnl/ 2 + 1/x-mk 1/ 2) - 4llx- (mn

< 2 (II x - mn 11 2 + II x- mk 11 2 - 4d2)
(since (mn
~
2(d2
+ mk)/2 W
+ mk)/2 EM)
+ d2)- 4d2 =
0 as k, n
oo.
By the completeness of the Hilbert space X, there exists an element mE X

such that s-lim mn = m. We have mE M sinceM is closed. Also, by the
n-co
continuity of the norm, we have II x - m II = d.

Write x = m + (x- m). Putting n = x - m, we have to show that
n E M.L. For anym' EM and any real number <X, we have (m + ,xm') EM
83
and so
d2
<
Jlx- m -1Xm'JI 2 =(n-IX m', n-IX m')
= jjnjj 2 -1X(n,m')-1X(m',n)
+1X2Jim'll2.
This gives, since Jlnll = d, 0 <- 21XRe(n, m') + IX2 jjm'JI 2 for every
real IX. Hence Re(n, m') = 0 for every m' EM. Replacing m' by im',
we obtain Im (n, m') = 0 and so (n, m') = 0 for every m' EM.
Corollary. For a closed linear subspace M of a Hilbert space X, we
have M = MJ.J. = (M.l.)J..
Theorem 2. The projector P =PM is a bounded linear operator such
that
P
(2)
P 2 (idempotent property of P),
(Px, y)
(3)
(x, Py) (symmetric property of P).
Conversely, a bounded linear operator P on a Hilbert space X into X

satisfying (2) and (3) is a projector upon M = R (P).
Proof. (2) is clear from the definition of the orthogonal projection.
We have, by (1') and PMx j_ PMJ.Y,
(PMx, y) = (PMx, PMJ.Y
=
(PMx
+ PMy) =
+ PMJ.X, PMy) =
(PMx, PMy)
(x, PMy).
Next let y = x + z, xEM, zE MJ. and w = u + v, uEM, vEMJ.. then

y + w = (x + u) + (z + v) with (x + u) EM, (z + v) E M J. and so,
bytheuniquenessofthedecomposition (l),PM(Y + w) = PMy + PMw;
similarly we obtain PM(1Xy) = IXPMY The boundedness of the operator
P.u is proved by
llx 11 2
=
=
+ PMJ.x /1 2 = (PMx + PMJ.X, PMx + P.uJ.x)

JIPMxJI 2 + ![PMJ.x[l 2 > fiPMxfl 2
IIPMx
Thus, in particular, we have
(4)
The converse part of the Theorem is proved as follows. The set
R (P) is a linear subspace, since Pis a linear operator. The condition
x EM is equivalent to the existence of a certain y EX such that x = P y,
and this in turn is equivalent, by (2), to x = Py = P 2 y = Px. Therefore
x EM is equivalent to x = Px. M is a closed subspace; for, Xn EM,
s-lim Xn = y imply, by the continuity of P and Xn = Pxn, s-lim Xn =
-00
s-lim Pxn
n->00
= Py
so that y = Py.
n->00
We have to show that P =PM. If x EM, we have Px = x =PM x;

and ifv EM J., we have PMy = 0. Moreover, in the latter case, (Py, Py)=
6*
84
(y, P 2 y) = (y, Py) = 0 and so Py = 0. Therefore we obtain, for any

yEX,
Py
=
=
P(PMy
PMY
+ PMJ_Y) =
+ 0,
i.e., Py
p PMy
+ p PMJ_Y
PMY
Another characterization of the projection operator is given by

Theorem 3. A bounded linear operator P on a Hilbert space X into X
is a projector iff P satisfies P = P 2 and II P II < 1.
Proof. We have only to prove the "if" part. Set M = R(P) and
N = N (P) = {y; Py = 0}. As in the proof of the preceding Theorem 2,
M is a closed linear subspace and x EM is equivalent to x = Px. N is
also a closed linear subspace in virtue of the continuity of P. In the

decomposition x = Px +(I -P)x, we have PxE M and (I -P)xE N.
The latter assertion is clear from P (I- P) = P - P 2 = 0.
We thus have to prove that N = M j_. For every xE X, y = Px-xE N
by P 2 = P. Hence, if, in particular, xE N.l., then Px = x + y with
(x, y) = 0. It follows then that llxll 2 > 11Pxll 2 = llxll 2 + IIYII 2 so that
y = 0. Thus we have proved that xE Nj_ implies x = Px, that is,
Nj_ ~ M = R(P). Let, conversely, zE M = R(P), so that z = Pz. Then
we have the orthogonal decomposition z = y + x, yEN, x EN j_, and
so z = Pz = Py + Px = Px = x, the last equality being already
proved. This shows that M = R(P) ~ Nj_. We have thus obtained
M = N.l., and so, by N = (NJ_)j_, N = Mj_.
2. "Nearly Orthogonal" Elements

In general, we cannot define the notion of orthogonality in a normed
linear space. However, we can prove the
Theorem (F. Rmsz [2]). Let X be a normed linear space, and M be
its closed linear subspace. Suppose M =!= X. Then there exists, for any
E > 0 with 0 < E < 1, an x, EX such that
llx. II= 1
and
dis(x., M) =
inf
mEM
llx,- mII> 1- s.
(1)
The element x, is thus "nearly orthogonal" toM.

Proof. Let yE X -M. Since Mis closed, dis(y, M) =
inf lly-m II=
mEM
a> 0. Thus there exists an m,EM such that IJy-m, II< a ( 1
+1e
e).
The vector x, = (y- m,)/IIY- m,ll satisfies IJx, II = 1 and

llx.-mll = lly-m,II-1 11Y-m,-lly-m,ll mil> l!y-m,ll-la
>
(1-=- f;r11 -c.
85
3. The Ascoli-Arzela Theorem
Corollary 1. Let there exist a sequence of closed linear subspaces Mn

of a normed linear space X such that _1\dn ~ Mn+ 1 and Mn # M,.+l
(n = 1, 2, ... ). Then there exists a sequence {Yn} such that
YnE Mn, llYn II= 1 and
dis(Yn+l Mn) > 1/2 (n
= 1, 2, ... ).
(2)
Corollary 2. The unit sphere S = {x E X; II x II < 1} of a B-space X

is compact iff X is of finite dimension.
Proof. The "if" part is a consequence of the Bolzano-Weierstrass
theorem that a bounded closed set of Rn is compact. The "only if" part
is proved as follows. Suppose X is not of finite dimension. Then there
exists, by the preceding Corollary 1, a sequence {Yn} satisfying the conditions: II Yn II = 1 and II Ym- Yn II > 1/2 form> n. This is a contradiction
to the hypothesis that the unit sphere of X is compact.
3. The Ascoli-Arzela Theorem

To give an example of a relatively compact infinite subset of a
B-space of infinite dimension, we shall prove the
Theorem (Ascou-ARZELA). Let S be a compact metric space, and
C (S) the B-space of (real- or) complex-valued continuous functions x (s)
normed by II x II = sup Ix (s) I Then a sequence {xn (s)} ~ C (S) is relatively
sEs
compact in C (S) if the following two conditions are satisfied:
(1}
Xn(s) is equi-bounded (inn), i.e., sup sup lxn(s) I<=,
n~1
xn(s) is equi-continuous (in n), i.e.,

lim
sup
IXn (s') -
d t 0 n.;;;;1,dis(s',s") ;:;;;d
s ES
(2)
Xn (s") I = 0.
Proof. A bounded sequence of complex numbers contains a convergent

subsequence (the Bolzano-Weierstrass theorem). Hence, for fixed s, the
sequence {xn (s)} contains a convergent subsequence. On the other hand,
since the metric spaceS is compad, there exists a countable dense subset
{sn} ~ S such that, for every e > 0, there exists a finite subset
{sn; 1 -::::: n < k (e)} of {sn} satisfying the condition
sup
inf
s E S 1;:;;;j:o;;k(e)
dis(s,sj)<e.
(3)
The proof of this fact is obtained as follows. Since S is compact, it is

totally bounded (see Chapter 0, 2). Thus there exists, for any b > 0, a
finite system of points E S such that any point of S has a distance < b
from some point of the system. Letting b = 1, 2-1 , 3-1 , . . and collecting
the corresponding finite systems of points, we obtain a sequence {sn}
with the stated properties.
We then apply the diagonal process of choice to the sequence {xn(s)},
so that we obtain a subsequence {xn' (s)} of {xn (s)} which converges for
86
s = Sv s2 , , sk, ... simultaneously. By the equi-continuity of {x,. (s)},

there exists, for every e > 0, a (J = b(e) > 0 such that dis(s', s") < b
implies Ix,. (s') - x,. (s") I < e for n = 1, 2, . . . Hence, for every s E S,
there exists a f with j < k (e) such that
Ix,.. (s) -
Xm (s) I < Ix,.. (s) - x,.. (sj) I + lx,.. (sj) - Xm (sj) I
+ IXm (sj) -
Xm (s) I < 2e + IX,. (sj) - Xm (sj) I

Thus lim maxlx,..(s)-Xm(s)l<2e,andso lim llx,.-Xm!I=O.
n,~
n,m-+00
4. The Orthogonal Base. Bessel's Inequality and

Parseval's Relation
Definition 1. A set S of vectors in a pre-Hilbert space X is called
an orthogonal set, if x j_ y for each pair of distinct vectors x, y of S. If,
in addition, llxll = 1 for each xES, then the setS is called an orthonormal set. An orthonormal set S of a Hilbert space X is called a complete
orthonormal system or an orthogonal base of X, if no orthonormal set of X
contains S as a proper subset.
Theorem 1. A Hilbert space X (having a non-zero vector) has at
least one complete orthonormal system. Moreover, if S is any orthonormal set in X, there is a complete orthonormal system containing S as
a subset.
Proof (by Zorn's Lemma). LetS be an orthonormal set in X. Such a set
surely exists; for instance, if x =F 0, the set consisting only of xf II x II is
orthonormal. We consider the totality {S} of orthonormal sets which
contain S as a subset. {S} is partially ordered by writing 5 1
5 2 for
the inclusion relation 5 1 ~ 5 2 Let {S'} be a linearly ordered subsystem
of {5}, then U S' is an orthonormal set and an upper bound of
-<
S'E{S'}
{5'}. Thus, by Zorn's Lemma, there exists a maximal element 50 of {5}.

This orthonormal set 50 contains Sand, by the maximality, it must be
a complete orthonormal system.
Theorem 2. Let S = {x,.; IX E A} be a complete orthonormal system
of a Hilbert space X. For any IE X, we define its Fourier coefficients
(with respect to S)
(1)
/,. = (/, Xa)
Then we have Parseval's relation
11111 2 = ~ l/,.1 2
(2)
Proof. We shall first prove Bessel's inequality
,.t: ''" 1 < 11111

2
(2')
4.
The Orthogonal Base. Bessel's Inequality and Parseval's Relation
87
Let ex 1 , ex 2 , . . . , ex,. be any finite system of .x's. For any finite system
of complex numbers c"'' c"' ... , c"n' we have, by the orthonormality of
{x"'}'
\f- J=l
.i ccx; x,Ji'z = (t- J=l
.I c"J x"'J' f- J=1
.I c"i x,
1)
"
Hence the minimum of

attained when c"'1 =
Ia;
II f 11 2 - j~ If"'i 12 + j~ Ifet; -
w,
lit- j~ c"'1 x.,1
Ca;l 2
for fixed ex1 , ex 2 ,
. ,
ex,., is
(f = 1, 2, ... , n). We have thus
llt-j~/.,1 X"'1 1 2 = ll/11 2 -j~l/"'1 1 2 ,
and hence
j~l/.,1 l 2 <llf!l 2
{4)
By the arbitrariness of ex1 , ex2 , , ex,., we see that Bessel's inequality (2')
is true, and f"' =F 0 for at most a countable number of ex's, say
exv ex 2 ,
sequence
of {x.,},
ex,., ... We then prove that
f = s-lim .~" /.,1 x"'J' First, the

n-+00}=1
{if.,.
x,..} is a Cauchy sequence, since, by the orthonormality
j=1 ' '
which tends, by(4)provedabove, to Oas k~oo. We sets-lim .~" f"',x"'. = f',

n-+00}=1
'
and shall prove that (f- f') is orthogonal to every vector of S. By the
continuity of the scalar product, we have
=
and, when ex =I= exj (f =
(f- f',
Xa 1)
(/ - t', x"') =
lim
11-+00
(1- i
k-1
f.,k
x.,,, x.,1) =
/.,1 -
/.,1
0,
1, 2, ... ),
lim
11-+00
(t - Zf.,, x"'"' x"') = 0 k-1
0.
Thus, by the completeness of the orthonormal system S = {x"'}' we

must have (/- f') = 0. Hence, by (4) and the continuity of the norm,
we have
0 =lim
~00
[It- J=1.Zf.,
1
1 Xco1 11
= 11111 2 - fi-+OOJ=l
lim .i l/.,1 12 = 11111 2 - o:EA
~ 11"'12
Corollary 1. We have the Fourier expansion

00
"
I= j=1
~ f,.. x"'1 = s-lim ~ f"'. x"'l"
'
n-+ooj=1 '
(5)
88
Corollary 2. Let l 2 (A) be the space L2(A, )8, m) where m({cx}) = 1 for
every point ex of A. Then the Hilbert space X is isometrically isomorphic
to the Hilbert space l 2 (A) by the correspondence
(6)
X 3/ ++ {/~} E l 2 (A)
in the sense that
!l/11 2 )= !1{1"}11 2 = o:EA

.I llxl 2 (7)
0, 1, 2, ... J is a complete orthonormal
(f+g)++{l,+g,,}, {3/++{(3/r,J and

1
.
Example. { ~ e'nt;
system in the Hilbert space L2 (0, 2n).

Proof. We have only to prove the completeness of this system. We
have, by (3),
11
/-.i
J=-n
_2_cieiitl!
2:n:
~ll't-.i
J=-n
2
_2_fieiitll
=
2:n:
1
11111 2
-.i
J=-n
l!il 2 ,
where fi = (/, eiit).

If f E L2 (0, 2n) is continuous and with period 2n, then the left hand
side of the inequality above may be taken arbitrarily small by virtue of
Weierstrass' trigonometric approximation theorem (see Chapter 0, 2).
Thus the set of all the linear combinations ~ ci eiit is dense, in the sense
J
of the norm, in the subspace of L2 (0, 2n) consisting of all continuous

functions with period 2n. Such a subspace is also dense, in the sense of
the norm, in the space P(O, 2n). Therefore, any function /E P(O, 2n),
orthogonal to all the functions of {
vh
eint} must be a zero vector of

1
L 2 (0, 2n). This proves that our system offunctions V2 :n: e'nt is a complete
orthonormal system of P(O, 2n).
5. E. Schmidt's Orthogonalization
Theorem (E. Schmidt's orthogonalization). Given a finite or countably
infinite sequence {xi} of linearly independent vectors of a pre-Hilbert
space X. Then we can construct an orthonormal set having the same
cardinal number as the set {xi} and spanning the same linear subspace
as {xi}
Proof. Certainly x 1 =/= 0. We define y1 , y2 , . and u 1 , tt 2 , recurrently as follows:
Yt
Xl,
Y2 = X2- (x2, ul) ul,

Yn-t-1
Xn+l -
.Il (xn+l'
J=
.
Uj} Uj,
U1
U2
YI/IIY1II.
Y2/IIY2II.
5. E. Schmidt's Orthogonalization
This process terminates if {xj} is a finite set. Otherwise it continues
indefinitely. Observe that Yn # 0, because x 1 , x 2 , . . , x,. are linearly
independent. Thus un is well defined. It is clear, by induction, that each
un is a linear combination of Xv x 2 , , Xn and that each xn is a linear
combination of u 1 , u 2 , , u,.. Thus the closed linear subspace spanned
by the u's is the same as that spanned by the x's.
We see, by jju1 jj = 1, that y 2 j_ u 1 and hence u 2 j_ u 1 . Thus, by
jju1 jj = 1, y3 j_ u 1 and hence u 3 ..L u 1 . Repeating the argument, we see
that u 1 is orthogonal to u 2 , u 3 , . , un, . .. Thus, by !lu2 !1 = 1, we have
y3 j_ u 2 and so u 3 j_ u 2 Repeating the argument, we finally see that
uk ..l um whenever k > m. Therefore {uj} constitutes an orthonormal set.
Corollary. Let a Hilbert space X be separable, i.e., let X have a dense
subset which is at most countable. Then X has a complete orthonormal
system consisting of an at most countable number of elements.
Proof. Suppose that an at most countable sequence {ai} of vectors
EX be dense in X. Let x1 be the first non-zero element in the sequence
{ai}, x2 the first ai which is not in the closed subspace spanned by x1 ,
and Xn the first a; which is not in the closed subspace spanned by
x1 , x 2 , . . . , xn-l It is clear that the a's and the x's span the same closed
linear subspace. In fact, this closed linear subspace is the whole space X,
because the set {ai} is dense in X. Applying Schmidt's orthogonalization
to {xi}, we obtain an orthonormal system {uj}which is countable and spans
the whole space X.
This system {uj} is complete, since otherwise there would exist a
non-zero vector orthogonal to every uj and hence orthogonal to the
space X spanned by u/s.
Example of Orthogonalization. LetS be the interval (a, b), and consider the real Hilbert space L2 (S, 55, m), where 55 is the set of all Baire
subsets in (a, b). If we orthogonalize the set of monomials
1, s, s2 , s3 ,
... ,
s'", ... ,
we get the so-called Tchebyschev system of polynomials

P 0 (s) =constant, P 1 (s), P 2 (s), P 3 (s), ... , Pn(s), ...
which satisfies
b
J P;(s) Pj(s) m(ds) =
b;j (= 0 or 1 according as i# j or i = j).
In the particular case when a= -1, b = 1 and m(ds) = ds, we obtain

the Legendre polynomials; when a= -oo, b = oo and m(ds) = e-s' ds,
we obtain the Hermite polynomials and finally when a= 0, b = oo and
m (ds) = e-s ds, we obtain the Lagtterre polynomials.
90
It is easy to see that, when - oo < a < b < oo, the orthonormal
system {Pj{s)} is complete. For, we may follow the proof of the completeness of the trigonometric system {see the Example in the preceding
section 4); we shall appeal to Weierstrass' polynomial approximation
theorem, in place of Weierstrass' trigonometric approximation theorem.
As to the completeness proof of the Hermite or Laguerre polynomials,
we refer the reader to G. SzEGO [1] or to K. YosiDA [1].
6. F. Riesz' Representation Theorem

Theorem (F. Riesz' representation theorem). Let X be a Hilbert space
and f a bounded linear functional on X. Then there exists a uniquely
determined vector y1 of X such that
= (x, y1)
for all x EX, and IIIII = IIYtll

{1)
Conversely, any vector y EX defines a bounded linear functional /, on
X by
f(x)
t,(x)=(x,y) forallxEX, and
11/,II=IIYII
{2)
Proof. The uniqueness of y1 is clear, since (x, z) = 0 for all x EX

implies z = 0. To prove its existence, consider the null space N = N (f)=
{xEX;f(x) = 0} of f. Since f is continuous and linear, N is a closed
linear subspace. The Theorem is trivial in the case when N = X; we
take in this case, y1 = 0. Suppose N =1=- X. Then there exists a y0 =1=- 0
which belongs to N.l. (see Theorem 1 in Chapter III, 1). Define
(3)
We will show that this y1 meets the condition of the Theorem. First,
if x EN, then f (x) = (x, y1) since both sides vanish. Next, if x is of
the form x = .xy0 , then we have
(x, Yt)
(.xyo, Yt)
= (.xyo, rf~fa Yo)= .xf(Yo) = f(.xyo)
= f(x)
Since f(x) and {x, y1) are both linear in x, the equality f(x) = (x, y1),
x E X, is proved if we have proved that X is spanned by N and y0 To
show the last assertion, we write, remembering that f(y1) =1=- 0,
X=
f(x) )
( X - f(y,)
'Yj
f(x)
+ f(y,) Yt
The first term on the right is an element of N, since

f(x)
I ( x- f(Yt) Yt =
f(x)
f(x)- f(Ytl f(yf)
We have thus proved the representation f(x)
0.
(x, y1).
6. F. Riesz' Representation Theorem
91
Therefore, we have
Ill !I= sup il(x) I =
11~11 :i>l
sup
11%11 :::>1
I(x, Yt) I< II~sup

II ::;;;1
llx II !IYtll = IIYtll
sup ll(x)l > lf{Yt/IIYtiDI = ( -,,Y:tll , Yt) = IIYtii

\ Yt
Hence we have proved the equality IIIII = IIYtll
Finally, the converse part of the Theorem is clear from ll,(x) I=
j(x, y) I::;; llxiiIIYII
Corollary 1. Let X be a Hilbert space. Then the totality X' of bounded
linear functionals on X constitutes also a Hilbert space, and there is a
norm-preserving, one-to-one correspondence I-. Yt between X' and X.
By this correspondence, X' may be identified with X as an abstract set;
but it is not allowed to identify, by this correspondence, X' with X as
linear spaces, since the correspondence I-. Yt is conjugate linear:
and also IIIII =
11%11~1
(M
(tx1l1 + tx2/2) -. (iilYt, + li2Yt,),
where~. tx2 are complex numbers.
Proof. It is easily verified that X' is made into a Hilbert space by
defining its scalar product through (/1> / 2) = (y1,, y1,), so that the statement of Corollary 1 is clear.
Corollary 2. Any continuous linear functional T on the Hilbert space
X' is thus identified with a uniquely determined element t of X as
follows:
T (f) = I (t)
for all I E X'.
(5)
Proof. Clear from the fact that the product of two conjugate linear
transformations is a linear transformation.
Definition. The space X' is called the dual space of X. We can thus
identify a Hilbert space X with its second dual X" = (X')' in the above
sense. This fact will be referred to as the reflexivity of Hilbert spaces.
Corollary 3. Let X be a Hilbert space, and X' its dual Hilbert space.
Then, for any subset F of X' which is dense in the Hilbert space X', we
have
(6)
sup lf(x0)j, x0 EX.
llxoll =
/EF,JJfiJ;S;l
Proof. We may assume that x0 =F 0, otherwise the formula (6) is trivial. We have (x0 , x0fllx0 II)= llx0 11, and so there exists a bounded linear
functional/0 onXsuch that llloll = 1,/0 (x0) = llxoll Sincel(x0) = (x0 , Yt)
is continuous in y1, and since the correspondence f-. y1 is norm-preserving
we see that (6) is true by virtue of the denseness ofF in X'.
Remark. Hilbert's original definition of the "Hilbert space" is the
space (l 2). See his paper [1]. It was J. VON NEUMANN [1] who gave an
U2
axiomatic definition (see Chapter I, 9) of the Hilbert space assuming

that the space is separable. F. Rmsz [1] proved the above representation
theorem without assuming the separability of the Hilbert space. In this
paper, he stressed that the whole theory of the Hilbert space may be
founded upon this representation theorem.
7. The Lax-Milgram Theorem

Of recent years, it has been proved that a variant of F. Riesz' representation theorem, formulated by P. LAX and A. N. MILGRAM [1], is a
useful tool for the discussion of the existence of solutions of linear partial
differential equations of elliptic type.
Theorem (LAX-MILGRAM). Let X be a Hilbert space. Let B(x, y) be a
complex-valued functional defined on the product Hilbert space X X X
which satisfies the conditions:
Sesqui-linearity, i.e.,
B (iX1 x1
+ iX 2 x 2, y) =
iX1 B (x1 , y)
+ iX 2B (x 2, y)
and
(1)
B(x,f3IYI +/32 Y2) =fJ1B(x,y1) +fJ2B(x,y2),

Roundedness, i.e., there exists a positive constant y such that
IB(x, Y) I <y l!xiiIIYII.
Positivity, i.e., there exists a positive constant
B(x, x)?:
(2)
o such
that
o [[x[[ 2
(3)
Then there exists a uRiquely determined bounded linear operator S with

a bounded linear inverse s-1 such that
(x, y) = B (x, Sy) whenever x andy EX, and II S II<
o-1 , I!S-1 [[ <
y. (4)
Proof. Let D be the totality of elements y EX for which there exists

an element y* such that (x, y) = B (x, y*) for all x EX. D is not void,
because 0 ED with 0* = 0 y* is uniquely determined by y. For, if w
be such thatB(x, w) = Oforallx, thenw = ObyO = B(w, w) > o llw!/ 2
By the sesqui-linearity of (x, y) and B(x, y), we obtain a linear operator
S with domain D(S) = D: Sy = y*. S is continuous and IISYII <
o-l IIYI!. yE D(S), because
0 IISYII 2 < B(Sy, Sy) = (Sy, y) < IISYIIIIYII

Moreover, D = D (S) is a closed linear subspace of X. Proof: if Yn ED (S)
and s-lim Yn =Yeo then, by the continuity of S proved above, {Syn}
n-->00
is a Cauchy sequence and so has a limit z =s-lim Syn. By the continuity

of the scalar product, we have lim (x, Yn) = (x, Yeo) We have also, by
n-+oo
93
(2), lim B(x,Syn) =B(x,z). Thus, by (x,yn) =B(x,Syn), we must

n-+00
have (x, Yeo) = B (x, z) which proves that Yeo ED and Sy00 = z.
Therefore the first part of the Theorem, that is, the existence of
the operator S is proved if we can show that D (S) = X. Suppose
D (S) =F X. Then there exists a w0 EX such that w0 =F 0 and w0 ED (S)l_.
Consider the linear functional F (z) = B (z, w0 ) defined on X. F (z) is
continuous, since IF(z) I = IB(z, w0 ) I < y liz II llw0 II Thus, by F. Riesz'
representation theorem, there exists a w~ EX such that B (z, w0 ) =
F (z) = (z, w~) for all z EX. This proves that w~ ED (S) and Sw~ = w0
But, by bllwoll 2 < B(w0 , w0 ) = (w0 , w~) = 0, we obtain w0 = 0 which
is a contradiction.
The inverse S-1 exists. For, Sy = 0 implies (x, y) = B(x, Sy) = 0
for all x E X and soy = 0. As above, we show that, for every y E X there
exists a y' such that (z, y') = B (z, y) for all z EX. Hence y = Sy' and so
s-1 is an everywhere defined operator, and, by I(z, S-1 y) I = IB (z, y) I :S
y liz II IIY II, we see that IIS-1 11 < y.
Concrete applications of the Lax-Milgram theorem will be given in
later chapters. In the following four sections, we shall give some examples
of the direct application of F. Riesz' representation theorem.

This theorem reads as follows.
Theorem (LEBESGUE-NIKODYM). Let (S, 58, m) be a measure space,
and v (B) be a a-finite, a-additive and non-negative measure defined on 58.
If v is m-absolutely continuous, then there exists a non-negative, mmeasurable function p (s) such that
v(B)=JP(s)m(ds)
forall
BE\Bwithv(B)<oo.
(1)
Moreover, the "density" p (s) of v (B) (with respect to m (B)) is uniquely

determined in the sense that any two of them are equal m-a.e.
J. VON NEUMANN [2]). It is easy to see that e(B) =

is a a-finite, a-additive and non-negative measure defined
Proof (due to
m (B)
+ v (B)
on 58. Let {En} be a sequence of sets E 58 such that S =
co
U En, En<;;; Bn+l
n=l
and (En) < oo for n = 1, 2, ... If we can prove the theorem for every
B <;;;En (for fixed n) and obtain the density Pn (s), then the Theorem is
true. For, we have only to take p (s) as follows:
p(s) = p1 (s)
for s E Bv and P (s) =Pn+I (s) for s E Bn+l- B,.
(n = 1, 2, ... ) .
94
Therefore we may assume that e(S) <ex> without losing the generality. Consider the Hilbert space L2(S, )8, e). Then
f x(s) v(ds), X E L2(S, )8, e)'

s
gives a bounded linear functional on L 2 (S, )8, e), because
f(x) =
!f(x) I~
jx(s) I v(ds)
2 v(ds)Y'2 ( / 1 v(ds)t2
< (/ jx(s) 1
< llxlle. v(S)l/2,

where llxlle=(jlx(s)i 2 e(ds)t2
Thus, by F.Riesz' representation
theorem, there exists a uniquely determined y E L2 (5, )8, e) such that
I x(s) v(ds) = sI x(s) y(s) e(ds) = sI x(s) y(s) m(ds) + sI
x(s) y(s) v(ds)
holds for all x E L2(S, )8, e). Taking x as non-negative functions and considering the real part of both sides, we may assume that y (s) is a
real-valued function. Hence
x(s) (1- y(s)) v(ds) =

x(s) y(s) m(ds) if x(s) E P(S, )8, e) (2)
s
s
is non-negative.
WecanproveO <y(s) < 1 e-a.e. Tothispurpose,setE1 ={s;y(s) < 0}
and E 2 = {s; y(s) > 1}. If we take the defining function CE,(s) of E 1
for x (s) in (2), then the left hand side is > 0 and hence
y (s) m (ds) > 0.
E,
Thus we must have m(E1) = 0, and so, by them-absolute continuity

of v, v(E 1 ) = 0, e(E 1) = 0. We may also prove e(E 2 ) = 0, by taking
the defining function CE, (s) for x(s) in (2). Therefore 0 < y(s) < 1
e-a.e. on s.
Let X (s) be m-measurable and > 0 e-a.e. Then, by e(S) < CX>' the
"truncated functions" Xn(s) = min(x(s), n) belong to L2(S, )8, e)
(n = 1, 2, ... ), and hence
Xn(s) (1- y(s)) v(ds) =

Xn(s) y(s) m(ds) (n = 1, 2, ... ) .
s
s
Since the integrals increase monotonely as n increases, we have
lim
J xn(s) (1- y(s)) v(ds) =lim JXn(s) y(s) m(ds) =
~s
~s
Since the integrands are

Lemma,
L >
L>
>
lim (xn(s) (1- y(s)) v(ds) =
J~im (xn(s)y(s))m(ds)
<ex>. (4)
0 e-a.e., we have, by the Lebesgue-Fatou
(3}
Jx(s) (1-
= Jx(s)ym(ds),
y(s)) v(ds),
(5~
95
9. The Aronszajn-Bergman Reproducing Kernel
under the convention that if x(s) (1- y(s)) is not v-integrable, the right
hand side is equal tooo; and the same convention for x(s) y(s). If x(s)y(s)
ism-integrable, then, by the Lebesgue-Fatou Lemma,
L<
lim (x,. (s) y (s)) m (ds)

s .......
00
J x (s) y (s) m (ds).
(6)
This formula is true even if x (s) y (s) is not m-integrable, under the convention that then L = oo. Under a similar convention, we have
S,
Therefore, we have
J x(s) (1- y(s)) v(ds).
(7)
x(s) y(s) m(ds) for every x(s) which

x(s) (1- y(s)) v(ds) =
s
s
is 58-measurable and > 0 e-a.e.,
under the convention that, if either side of the equality is
other side is also = oo.
Now we put
x (s) (1- y (s))
z (s), y (s) (1- y (s))-1
(8)
oo then the
= p (s).
Then, under the same convention as in (8), we have
Jz(s) v(ds) = sJ z(s) p(s) m(ds) if z(s) is 58-measurable
and
>
(9)
0 e-a.e.
If we take the defining function CB(s) of BE 58 for z(s), we obtain
v (B)
J p(s) m (ds)
for all B E 58.
The last part of the Theorem is clear in view of definition (1).

Reference. For a straightforward proof of the Lebesgue-Nikodym theorem based upon Hahn's decomposition (Theorem 3 in Chapter I, 3) see
K. YosmA [2]. This proof is reproduced in HALMOS [1], p. 128. See also
SAKS [1] and DuNFORD-SCHWARTZ [1].

Let A be an abstract set, and let a system X of complex-valued functions defined on A constitute a Hilbert space by the scalar product
(!,g) = (f(a), g(a))a
(1)
A complex-valued function K (a, b) defined on A xA is called a reproducing kernel of X if it satisfies the condition:
For any fixed b, K (a, b) EX as a function of a,
f(b)=(f(a),K(a,b))a andhence f(b)=(K(a,b),f(a))a
(2)
(3)
96
As for the existence of reproducing kernels, we have

Theorem 1 (N. ARONSZAJN [1], S. BERGMAN [1]). X has a reproducing
kernel K iff there exists, for any y0 E A, a positive constant CY' depending
upon y0 , such that
/f(y0)/~Cy.//f//
forall
/EX.
(4)
Proof. The "only if" part is proved by applying Schwarz' inequality

to /(y0 ) = (l(x), K(x, y0 )},:
/ f(Yo) / <
I// (K(x, Yo), K(x, Yo))!12
I 1/ K(yo, Yo) 112
(5)
The "if" part is proved by applying F. Riesz' representation theorem to
the linear functional Fy. (f) = f(y 0 ) off EX. Thus, there exists a uniquely
determined vector gY (x) of X such that, for every f E X,
/I
f(y 0 )
1/
= Fy.(/) = (f(x), gy.(x))x,
and so gy, (x) = K (x, y 0 ) is a reproducing kernel of X. The proof shows

that the reproducing kernel is uniquely determined.
Corollary. We have
sup /I (Yo)/
ll/ll:;;l
= K (Yo Yo) 112 ,
(6)
the supremum being attained by
fo (x)
= eK (x, Yo)/K (Yo Yo) 112 Ie I =
1.
(7)
Proof. The equality in the Schwarz' inequality (5) holds iff f(x) and
K (x, y0 ) are linearly dependent. From the two conditions f(x) = IXK (x, y0 )
and /If//= 1, we obtain
1 = /IX/ (K (x, y0 ), K (x, y 0 ));/ 2 = /IX /K (y0 , y0 ) 112 , that is, /IX/= K (y0 , y0)- 112
Hence the equality sign in (5) is attained by / 0 (x).
Example. Consider the Hilbert space A 2 (G). For any /E A 2 (G) and
z E G, we have (see (4) in Chapter I, 9)
/f(z0 )J 2 < (nr2 )- 1
!f(z)/ 2 dxdy (z=x+iy).

Jz-z.J;:;;r
Thus A 2 (G) has the reproducing kernel which will be denoted by KG (z, z').
This KG(z, z') is called Bergman's kernel of the domain G of the complex
plane. The following theorE-m of Bergman illustrates the meaning of
KG(z, z') in the theory of conformal mapping.
Theorem 2. Let G be a simply connected bounded open domain of the
complex plane, and z0 be any point of G. By Riemann's theorem, there
exists a uniquely determined regular function w = f(z; z0 ) of z which
gives a one-to-one conformal map of the domain G onto the sphere
Iw / < (!G of the complex w-plane in such a way that
/0
(z0 ; z0)
0, (d/0 (z; z0 )jdz)z=z.
1.
97
Bergman's kernel KG(z: z0 ) is connected with l 0 (z; z0 ) by
l0 (z; z0 ) = Kc(z0 ; z0 )-1
J Kc(t; z
0)
(8)
dt,
Zo
where the integral is taken along any rectifiable curve lying in G and
connecting z0 with z.
Proof. We set
Ai (G) ={l(z); l(z) isholomorphicin G,f' (z) EA 2 (G), l(z0 )= Oandf' (z0 )=1},
and consider, for any IE Ai (G), the number
111'1!2 =
= cp (w)
If we denote by z
any IE Ai(G),
(9)
flf'(z)l 2 dxdy, z=x+iy.
the inverse function of w = lo (z; z0 ), then, for
111'11 2 = fj j/'(cp(w))j2 jcp'(w)j 2 dudv,

1wr<11a
w=u+iv.
For, by the Cauchy-Riemann partial differential equations

we have
dx dy
= 8o((x,u, y))v du dv = (x,. Yv- Yu xv) du dv = (x~ + y~) du dv

= jcp' (w) j2 du dv.
= l(cp (w)) be expanded into power series:
Let IE Ai (G), and let F (w)

F(w)
/(q;(w))
= w
00
2: en w" for [wj < I!G
n~2
00
= f' (cp (w)) q;' (w) = 1 + 2:
Then F' (w)
n~2
dr {
r(r
i n2
+ n=2
Therefore minimum
/EA!(G)
! Cn
Ill' II=
j2
2"- 1)
vn ec.
n c, wn- 1 and so
~
ao} = e~ + n-2
'Jt
A2(G)
and
7
= {g(z); g(z) is holomorphic in
1\g' II= 1},
Yos!da, Functional Analysis
n I c,. 12
e~".
and this minimum is attained iff
= f(cp(w)) = w, that is, iff f(z) = /0 (z; z0).

We set, for any f E Ai(G), g(z) = f(z)/111' II Then
F(w)
'Jt
llg' II =
1. If we put
G, g(z0 ) = 0, g' (z0 )
>
98
then the above remark shows that

maxi_mumg'(z0) = 1/ll/~11 = (V~
gEA'(G)
eG)-1,
and this maximum is attained iff g (z) is equal to

go(z) = fo(z; Zo)/11/~ II= fo(z; Zo)
I Vn eG
Hence, by (7), we obtain

(z; Zo}
vr.:n eG)-1 d/0 dz
=
(1
go, ( z) =
II
K G(z'. Zo )JKG(Zo'. Zo)1/2 ' 1A1 = 1.
Hence, by putting z = z0 , we have
vn
eGt 1 = KG(zo; Zo)/KG(zo; Zo) 112 = KG(zo; Zo) 112 '

and so we have proved the formula
(A.
d/0 dz
(z;z0)
= K G(z'. Zo)JKG(Zo,. Zo )
10. The Negative Norm of P. Lax
Let H 0(Q) be the completion of the pre-Hilbert space
with the scalar product (rp, w) and the norm llfPIIs:
(rp, 'lfJ)s = ,I
.f Dj rp(x) D;w(x) dx,
!JJ :;;;s .Q
b Hg (Q) =
Any element E
fb on H~ (Q) by
C~(Q)
endowed
llrp lis= (rp, rp)}/ 2
(1)
L2 (Q) defines a continuous linear functional
!b(w)
(2)
(w, b) 0 , wE H~(Q).
For, by Schwarz' inequality, we have

l(w, b)ol < Jlwllo llbllo < llwlls llbllo
Therefore, if we define the negative norm of bE
(Q) = L2 (Q) by
Hg
llbll-s=
then we have
sup
wEH~(D),JiwJI 8 :;;;1
l!b(w)l=
llbJI-s
and so, by llbll-s
>
Hence we may write
<
sup
wEH~(D),I!wJI 8 :;;;1
l(w,b) 0 l,
IJbllo
(3)
(4)
l(w/llwlls b)ol
l(w,b)ol
<
llwllsllbll-s
llbiJ-s= 11/bll-s= sup l(w,b) 0 1 forany bEH8(Q).

We shall prove
[[wl!s:;;;1
(5)
(3')
Theorem 1 (P. LAX [2]). The dual space H 0(Q)' of the space H 0(Q)
may be identified with the completion of the space Hg (Q) = L2 (Q) with
respect to the negative norm.
10. The Negative Norm of P. LAx
99
For the proof we prepare

Proposition. The totality F of the continuous linear functionals on
H 0(.Q) of the form lb is dense in the Hilbert space H 0(.Q)', which is the dual
space of H 0(.Q).
Proof. F is total on H~ (.Q) in the sense that, for a fixed w E H~ (.Q),
the simultaneous vanishing lb(w) = 0, bE H8(.Q), takes place only if
w = 0. This is clear since any element wE H~ (.Q) is also an element of
H8(.Q).
Now if F is not dense in the Hilbert space H 0(.Q) ', then there exists
an element T # 0 of the second dual space H 0(.Q)" = (H~ (.Q)')' such
that T (!b) = 0 for all lb E F. By the reflexivity of the Hilbert space
H 0(.Q), there exists an element t E H~ (.Q) such that T (f) =I (t) for all
IE H 0(.Q)'. Thus T(fb) = lb(t) = 0 for all bE H8(.Q). This implies, by the
totality of F proved above, that t = 0, contrary to T # 0.
Corollary. We have, dually to (3'),
llw!ls=
sup
bEHg(.O),JJbJJ_,:;;l
!(w,bJ0 I for any wEH0(.Q).
(6)
Proof. Clear from Corollary 3 in Chapter III, 6, because F =

{lb; bE H8 (.Q)} is dense in Hg (.Q)'.
Proof of Theorem 1. Clear from the facts that i) F is dense in the
dual space H 0(.Q)' and ii) F is in one-to-one correspondence to the set
H8(.Q) = L2(.Q) preserving the negative norm, i.e.,
F 3 lb +? b E Hg (.Q) and II lb li-s = II b li-s

We shall denote by H0 s (.Q) the completion of H8 (.Q) with respect to
the negative norm II b li-s Thus
H 0(Q)' = H 0 s (Q) .
(7)
For any continuous linear functional I on H 0(.Q), we shall denote by

<w, I> the value of I at wE H 0(.Q). Thus, for any bE H8(.Q), we may write
lb(w) = (w, b) 0 = <w, lb> = <w, b), wE HO(Q),
(8)
and have the generalized Schwarz' inequality
l<w,b)l < ilwlls llbll-s,
(9)
which is precisely (5).

Now we can prove
Theorem 2 (P. LAX [2]). Any continuous linear functional g(b) on
Hr;s (.Q) can be represented, by a fixed element wE H0(.Q), as
g(b) = gw(b) = <w, b).
(10)
H0(.Q)' = H0 s (.Q), H 0 s (.Q)' = H 0(.Q).
(11)
We have, in particular,
7*
100 III. The Orthogonal Projection and F. Riesz' Representation Theorem

Proof. If bEH8(Q), then <w,b)=fb(w)=(w,b)0 Since F=
{fb; bE H8 (Q)} is dense in the Hilbert space H 0(Q)', we know, by (9),
that <w, b) = (b, w) 0 with a fixed wE H 0(Q), defines a linearfunctional gw
continuous on a dense subset F of H 0(Q)'. The norm of this functional
gw on F will be denoted by /lgwlls Then, by (6),
llgwlls=
sup
llbiJ-s~l
l(b,w)0 l=
sup
llbll-s~l
l(w,b) 0 l=llwlls
(12)
We may thus extend, by continuity, the functional gw on F to a continuous linear functional on the completion of F (with respect to the
negative norm), that is, gw can be extended to a continuous linear functional on H 0(Q)' = H 0 s (Q) ; we denote this extension by the same letter
gw. We have thus
(13)
l[gwll = sup lgw(b) [ = llwlls
llbll-~1
Hence, in view of the completeness of the space H 0(Q), the totality G of

the continuous linear functionals gw on H 0 5 (Q) may be considered as
a closed linear subspace of H 0 5 (Q)' by the correspondence gw ++ w. If this
closed subspace G were not dense in H 0 5 (Q) ',then there exists a continuous
linear functional f# 0 on H 0 5 (Q)' such that f(gw) = 0 for all gwE: G. But,
since the Hilbert space H 0 s (Q) is reflexive, such a functional f is given
by f(gw) = gw(/0 ), / 0 E H 0 5 (Q), and so by (13) / 0 must be equal to 0, contrary to the fact f # 0. Therefore we have proved H 0 5 (Q)' = H 0(Q).
Remark. The notion of the negative norm was introduced by P. Lax
with the view of applying it to the genuine differentiability of distribution
solutions of linear partial differential equations. We shall discuss such
differentiability problems in later chapters. It is to be noted that the
notion of the negative norm is also introduced naturally through the
Fourier transform. This was done by J. LERAY [1] earlier than LAX.
We shall explain the point in a later chapter on the Fourier transform.

A generalized function is locally the distributional derivative of a
function. More precisely, we can prove
Theorem (L. ScHWARTZ [1]). Let T be a generalized function in
Q ~ R". Then, for any compact subset K of Q, there exist a positive
integer m0 = m0 (T, K) and a function f(x) = f(x; T, K, m0 ) E L 2 (K)
such that
T(ffJ) =
Jf(x)
a"m ( )
cp x
dx whenever
ox':_' ox;" ox;:''
qJ
E 'l)K(Q).
(1)
101
Proof. By the Corollary in Chapter I, 8, there exist a positive constant C and a positive integer m such that
IT(<p)I<C sup /Dj<p(x)l whenever <pE:fJx(D).
(2)
liJ~m,xEK
Thus there exists a positive number b such that

Pm(<p)
sup
JjJ;;;m,xEK
IDj<p(x) I< b, <p E :FJx(D), implies IT(<p) I< 1. (3)
We introduce the notation
(4)
and prove that there exists a positive constant e such that, for m0 =
m+ 1,
J l8m<p(x)j8xm'/
dx < e, <pE :FJx(D), implies Pm(<p) <b.
(5)
[}
This is proved by repeated application of the following inequality

/"P(x)j<
<( J
Kf\(-oo,x;)
Kr,(-oo,x;)
= t112 (
/8"P(x 1 ,
dy) 112 (
Kf\(-oo,x;)
Kf\(-oo,x;)
/8"1' (x1 ,
,x;_ 1 ,y,xi+l>xn)J8yjdy
/8"P(x1
. . , X;_ 1 ,
,x,_vy,xi+I>xn)J8yJ2 dy) 1' 2
y, xi+l ... , Xn)f8y /2 dy) 112 ,
where t is the diameter of K, i.e., the maximum distance between two

points of the compact set K.
Consider the mapping <p (x)----;. "''(x) = 8m'<p (x)j8xm defined on :FJx (Q)
into :FJx (D). As may be seen by integration, "''(x) = 0 implies <p (x) = 0.
Hence the above mapping is one-to-one. Thus T(<p), <p E :FJx(D), defines
a linear functional 5('1/'), '1/'(x) = 8m'<p(x)Joxm, by 5('1/') = T(<p). By (3)
and (5), 5 is a continuous linear functional on the pre-Hilbert space X
(x) /2 dx 2
IJ =
consisting of such "'''sand topologized by the norm
II"'' (jl"''
Y'
Thus there exists, by F. Riesz' represerutation theorem, a uniquely determined function f (x) from the completion of X such that
T(<p) = 5('1/') =
J (8m'<p(x)J8xm) f(x) dx for all <pE :fJx(D).
Actually, the completion of X is contained in L2 (K) as a closed linear

subspace, and so the Theorem is proved.
References for Chapter III

For general account of Hilbert spaces, see
1. M. GLASMAN [1], N. DUNFORD- J. SCHWARTZ [2],
F. Rmsz-B. Sz. NAGY [3] and M. H. STONE [1].
N.
I. AcHIESER[1],
B. Sz. NAGY
102
IV. The Hahn-Banach Theorems

In a Hilbert space, we can introduce the notion of orthogonal coordinates through an orthogonal base, and these coordinates are the values
of bounded linear functionals defined by the vectors of the base. This
suggests that we consider continuous linear functionals, in a linear topological space, as generalized coordinates of the space. To ensure the existence of non-trivial continuous linear functionals in a general locally
convex linear topological space, we must rely upon the Hahn-Banach
extension theorems.
1. The Hahn-Banach Extension Theorem in Real Linear Spaces

Theorem (HAHN [2], BANACH [1 ]) . Let X be a real linear space and
let p (x) be a real-valued function defined on X satisfying the conditions:
p (x
+ y) <
+ p (y)
p (x)
p (~Xx) =
~XP (x)
(subadditivity),
for
lX
(1)
> 0.
(2)
Let M be a real linear subspace of X and / 0 a real-valued linear functional

defined on M:
/0 (~Xx
+ {3y) =
~X/0 (x)
+ {3/0 (y)
for x, y EM and real
~X,
{3.
(3)
Let / 0 satisfy /0 (x) < p (x) on M. Then there exists a real-valued linear
functional F defined on X such that i) F is an extension of /0 , i.e.,
F (x) = / 0 (x) for all x EM, and ii) F (x) < p (x) on X.
Proof. Suppose first that X is spanned by M and an element x0 E M,
that is, suppose that
+ lXX0 ;
mE M, lX real}.
Since x0 EM, the above representation of x E X in the form x
is unique. It follows that, if, for any real number c, we set
X= {x = m
F(x) = F(m
+ lXX0 ) =
+ ~X Xu
/ 0 (m) +~XC,
then F is a real linear functional on X which is an extension of f0 V.le

have to choosecsuch thatF(x) < p(x), thatis,/0(m) +M< p(m + ~Xx0).
This condition is equivalent to the following two conditions:
/ 0 (mj~X)
+ c < P(x0 + mj~X)
for
/0 (m/(-~X))-c<p(-x0 +mf(-~X))
lX
>
0,
for ~X<O.
To satisfy these conditions, we shall choose c such that

/ 0 (m')- p(m'- x 0 ) < c < p(m" + x 0 ) - / 0 (m") for all m', m" EM.
Such a choice of c is possible since
/ 0 (m') + /0 (m") = /0 (m! + m'') < P(m' + m") = P(m'- x 0 + m" + .Y-0 )
~ p(m' -x0) + plm" + x0):
103
2. The Generalized Limit
we have only to choose c between the two numbers

sup [/0 (m')- p (m'- x0)]
~M
and
inf [p (m"
~M
+ x0) - l 0 (m")J.
Consider now the family of all real linear extensions g of lo for which
the inequality g (x) < p (x) holds for all x in the domain of g. We make
gto mean that h
this family into a partially ordered family by defining h
is an extension of g. Then Zorn's Lemma ensures the existence of a
maximal linear extension g of lo for which the inequality g (x) < p(x)
holds for all x in the domain of g. We have to prove that the domain
D (g) of g coincides with X itself. If it does not, we obtain, taking D (g) as
M and g as 10 , a proper extension F of g which satisfies F (x) < p (x) for
all x in the domain ofF, contrary to the maximality of the linear extension g.
Corollary. Given a functional p (x) defined on a real linear space X
such that (1) and (2) are satisfied. Then there exists a linear functional I
defined on X such that
(4)
-p(-x) < l(x) < p(x).
>-
Proof. Take any point x0 E X and define M = {x; x = exx0 , ex real}.

Set lo (exx0 ) = exp (x0). Then lo is a real linear functional defined on M.
We have l 0 (x) < p(x) on M. In fact, exp(x0 ) = p(exx0) if ex> 0, and
if ex< 0, we have exp (x0) :S - exp (-x0) = p (exx0) by 0 = P (0) :S P (x0) +
p (-x0 ). Thus there exists a linear functional I defined on X such that
l(x) = l 0 (x) on M and l(x) < p (x) on X. Since -l(x) = 1(-x) < p(-x),
we obtain -p(-x) :S f(x) < p(x).
2. The Generalized Limit

The notion of a sequence {xn} of a countable number of elements x,.
is generalized to the notion of a directed set of elements depending on a
parameter which runs through an uncountable set. The notion of the
limit of a sequence of elements may be extended to the notion of the
generalized limit of a directed set of elements.
Definition. A partially ordered set A of elements ex, {J, ... is called a
directed set if it satisfies the condition:
For any pair ex, fJ of elements of A, there exists ayE A

such that ex-< y, fJ-< y.
(1)
Let, to each point ex of a directed set A, there be associated a certain set

of real numbers /(ex). Thus /(ex) is a, not necessarily one-valued, real
function defined on the directed set A. We write
limf(ex) =a
aEA
(a is a real number)
104
-<
if, for any e > 0, there exists an <X0 E A such that <X0
<X implies
J/(<X)- a J < e for all possible values off at <X. We say, in such a case, that
the value a is the generalized limit or Moore-Smith limit of /(<X) through
the directed set A .
Example. Consider a partition Ll of the real interval [0, 1]:
0 = t0 < t1 < < t,. = 1.

The totality P of the partition of [0, 1] is a directed set A by defining
the partial order Ll
Ll' as follows: If the partition Ll' is given by
0 = t~ < t~ < < t~ = 1, then Ll
Ll' means that n < m and that
every t, is equal to some tj. Let x (t) be a real-valued continuous function
defined on [0, 1], and let f(LI) be the totality of real numbers of the form
-<
-<
n-1
..I (tH 1 - ti) x(tj), where tj is any point of [tj, tH 1].
J=O
Thus f(LI) is the totality of the Riemann sum of the function x(t) pertaining to the partition Ll. The value of the Riemann integral
Jx (t) dt
is nothing but the generalized limit of f(LI) through P.

As to the existence of a generalized limit, we have the
Theorem (S. BANACH). Let x(<X) be a real-valued bounded function
defined on a directed set A. The totality of such functions constitutes a
real linear space X by
+ y)
+ y(<X), (f3x) (<X)= f3x(<X).

We can then define a linear functional, defined on X and which we shall
denote by LIM x (<X), satisfying the condition
(x
o:EA
(<X)= x(<X)
limx(<X) < LIMx(<X) < limx(<X),

o:EA
where
o:EA
o:EA
lim x (<X) =sup inf x (/3), lim x ({3) = inf sup x (/3).
"'EA
"'
"'>fJ
aEA
"'
a>fJ
Therefore LIM x (<X) =lim x (<X) if the latter generalized limit exists.
aEA
aEA
Proof. We put p(x) =limx(<X}. It is easy to see that this p(x)

aEA
satisfies the condition of the Hahn-Banach extension theorem. Hence,

there exists a linear functional f defined on X such that -p (-x) <
f(x) < p(x) on X. We can easily prove that lim x(<X) = - p(-x) so
aEA
that we obtain the Theorem, by puttingLIMx(<X)

aEA
= f(x).
3. Locally Convex, Complete Linear Topological Spaces

Definition. As in numerical case, we may define a directed set {x"'} in
a linear topological space X. {x,.} is said to converge to an element x of
4. The Hahn-Banach Extension Theorem in Complex Linear Spaces 105

X, if, for every neighbourhood U(x) of x, there exists an index <Xo such
that x.. E U (x) for all indices <X >- cX0 A directed set {x..} of X is said to
be fundamental, if every neighbourhood U (0) of the zero vector 0 of X
is assigned an index <X0 such that (x.. -x11)E U(O) for all indices<X,f/>-<Xo
A linear topological space X is said to be complete if every directed
fundamental set of X converges to some element x EX in the sense
above.
Remark. We can weaken the condition of the completeness, and
require only that every sequence of X which is fundamental as a directed
set converges to an element x E X; a space X satisfying this condition is
said to be sequentially complete. For normed linear spaces, the two definitions of completeness are equivalent. However, in the general case, not
every sequentially complete space is complete.
Example of a locally convex, sequentially complete linear topological
space. Let a sequence {/,. (x)} of :tl (.Q) satisfy the condition
lim (j,. -fk) = 0 in :tl (.Q). That is, by the Corollary of Proposition 7
h,k->00
in Chapter I, 1, we assume that there exists a compact subset K of .Q for

which supp(j,.) ~ K (h = 1, 2, ... ) and lim (Dsj,.(x)- D 5 fk(x)) = 0
h,k->00
uniformly on K for any differential operator D 5 Then it is easy to see,

by applying the Ascoli-Arzela theorem, that there exists a function /E :tl (.Q)
for which lim D 5 j,. (x) = D f (x) uniformly on K for any differential
h->00
operator D 5 Hence lim j,.

h->00
=f
in :tl (.Q) and so :tl (.Q) is sequentially
complete. Similarly, we can prove that~ (.Q) is also sequentially complete.

As in the case of a normed linear space, we can prove the
Theorem. Every locally convex linear topological space X can be
embedded in a locally convex, complete linear topological space, in
which X forms a dense subset.
We omit the proof. The reader is referred to the literature listed in
J. A. DIEUDONNE [1]. Cf. also G. KoTHE [1].
4. The Hahn-Banach Extension Theorem in Complex Linear Spaces

Theorem (BoHNENBLUST-SOBCZYK). Let X be a complex linear space
and p a semi-norm defined on X. Let M be a complex linear subspace of
X and f a complex linear functional defined on M such that I/ (x) I< p(x)
on M. Then there exists a complex linear functional F defined on X
such that i) F is an extension of/, and ii) IF(x) I< p(x) on X.
Proof. We observe that a complex linear space is also a real linear
space if the scalar multiplication is restricted to real numbers. If f(x) =
g (x) + i h (x), where g (x) and h (x) are the real and imaginary parts of
f(x) respectively, then g and hare real linear functionals defined on M.
106
Thus
lg(x) I<
lf(x)
I< p(x)
and
lh(x)
I<
lf(x)
I< p(x)
on M.
Since, for each x EM,

g(ix)
+ ih(ix) = f(ix) = if(x) = i(g(x) + ih(x)) =- h(x) + ig(x),
we have
h(x)=-g(ix)
forall
xEM.
By the Theorem in Chapter IV, 1, we can extend g to a real linear functional G defined on X with the property that G (x) < p (x) on X. Hence
-G(x) = G(-x) ::S: p(-x) = p(x), and so IG(x) I< p(x). We define
F(x)
G(x)- iG(ix).
Then, by F(ix) = G(ix)- iG(-x) = G(ix) + iG(x) = iF(x), we easily

see that F is a complex linear functional defined on X. F is an extension
of f. For, x EM implies that
= G(x)- iG(ix) = g(x)- ig(ix) = g(x) + ih(x) = f(x).

To prove IF(x) I< p(x), we write F(x) = re-; 8 so that IF(x) I=
e; 8 F(x) = F(e; 8 x) is real positive; consequently IF(x) I= IG(e; 8 x) I<
F(x)
p(e; 8 x)
li 8 l p(x)
p(x).
5. The Hahn-Banach Extension Theorem in Normed Linear Spaces

Theorem 1. Let X be a normed linear space, M a linear subspace of X
and / 1 a continuous linear functional defined on M. Then there exists a
continuous linear functional f defined on X such that i) f is an extension
of /1, and ii) llt1ll =II! II
Proof. Set p(x) = ll/111 ilxll Then p is a continuous semi-norm

defined on X such that I/1 (x) I < p (x) on M. There exists, by the Theorem
in the preceding section 4, a linear extension f of / 1 defined on the whole
space X and such that lf(x) I< p(x). Thus II/II II /1 11
An Application to Moment Problems

Theorem 2. Let X be a normed linear space. Given a sequence of
elements {xn} ~X, a sequence of complex numbers {<Xn} and a positive
number y. Then a necessary and sufficient condition for the existence of
a continuous linear functional f on X such that f(x;) =<X; (i = 1, 2, ... )
and II f II < y is that the inequalities
I/~ /J;<X;I < y iii~ /J;x;ll
6. The Existence of Non-trivial Continuous Linear Functionals
107
hold for any choice of positive numbers n and complex numbers
fJv fJ2' ' f3n

Proof. The necessity of this condition is clear from the definition of
j [ I II We shall prove the sufficiency. Consider the set
X1 =
{z; z = i~ {J;x;
where nand {3 are arbitrary}.

n
For two representations z = .2: {J;x;

>=1
= .2: {3;X; of the same element

>=1
z E X 1 , we have, by the condition of the Theorem,
.i: {J;cx;- i
l=1
>=1
{J;IX;l < y
[[.i: {J;X;- i
=1
=1
{J;X;ll. = 0.
Thus a continuous linear functional 11 is defined on X 1 by 11 c~ {J;x;)
.2: {J;cx;. We have only to extend, by the preceding Theorem 1, 11 to a
=1
continuous linear functional
on X with
II I II = II 11 1!
Remark. As will be shown in section 9 of this chapter, any continuous

linear functional I on C [0, 1] is representable as
l(x)
J x(t) m(dt)
with a uniquely determined Baire measure m on the interval [0, 1]. Thus,
if we take xi(t) =ti-l (i = 1, 2, ... ), Theorem 2 gives the solvability
condition of the moment problem:
1
J ti-
1 m(dt)
= cxi (j
= 1, 2, ... ).

Theorem 1. Let X be a real or complex linear topological space, x 0 a
point of X and p(x) a continuous semi-norm on X. Then there exists a
continuous linear functional F on X such that F (x 0 ) = p (x0 ) and
IF(x) I< p(x) on X.
Proof. Let M be the set of all elements cxx0 , and define I on M by
= cxp (x0 ). Then I is linear on M and l1(cxx0 ) I= lcxP (x0 ) I= P (cxx0 )
there. Thus there exists, by the Theorem in Chapter IV, 4, an extension
(x) < p (x) on X. Hence F (x) is continuous at x = 0
F of I such that
with p (x), and so, by the linearity of F, F (x) is continuous at every
point of X.
f(cxx 0)
IF I
Corollary 1. Let X be a locally convex space and x 0 =F 0 be an element

of X. Then there exists a continuous semi-norm p on X such that
108
p (x0) =1=- 0. Thus, by Theorem 1,

tional f0 on X such that
/0
(x0) =
p(x0) =1=- 0
there exists a continuous linear funcand
l/0 (x) I< p(x)
on X.
Corollary 2. Let X be a normed linear space and x 0 =1=- 0 be any element of X. Then there exists a continuous linear functional / 0 on X such
that
= llxoll and llfoll = 1.

for p (x) in Corollary 1. Thus II /0 II < 1 from
But, by / 0 (x0 ) = llx0 jl, we must have the equality
fo(xo)
Proof. We take
l!0 (x) I< llxll.
lifo II= 1.
II x II
Remark. As above, we prove the following theorem by the Theorem

in Chapter IV, 1.
Theorem 1'. Let X be a real linear topological space, x 0 a point of X
and p(x) a real continuous functional on X such that
p(x+y)<p(x)+p(y)
and
p(1Xx)=1XP(x)
for
IX>O.
Then there exists a continuous real linear functional F on X such that

F(x0 ) = p(x0 ) and -p(-x) < F(x) < p(x) on X.
Theorem 2. Let X be a locally convex linear topological space. Let M
be a linear subspace of X, and f a continuous linear functional on M.
Then there exists a continuous linear functional F on X which is an
extension of f.
Proof. Since f is continuous on M and X is locally convex, there exists
an open, convex, balanced neighbourhood of 0, say U, of X such that
x EM(\ U implies lf(x) I < 1. Let p be the Minkowski functional of U.
Then pis a continuous semi-norm on X and U = {x; p(x) < 1}. For
any x EM choose IX> 0 so that IX> p (x). Then p (x/IX) < 1 and so
lf(x/1X) I< 1, that is jf(x) I< IX. We thus see, by letting IX t p(x), that
lf(x) I< p (x) on M. Hence, by the Theorem in Chapter IV, 4, we obtain
a continuous linear functional F on X such that F is an extension off and
IF(x) I< p(x) on X.
Theorem 3 (S. MAZUR}. Let X be a real or complex, locally convex linear
topological space, and M a closed convex balanced subset of X. Then,
for any x 0 EM, there exists a continuous linear functional f on X such
that / 0 (x0 } > 1 and l/0 {x) I< 1 on M.
Proof. Since M is closed, there exists a convex, balanced neighbourhood
V of 0 such that M (\ (x0 + V) = 0. Since V is balanced and convex,
we have ( M
+ :) (\ (x 0 + :) =0.
hood of x 0 , the closure U of ( M
The set ( x 0
+ :) being a neighbour-
+ :) does not contain x0 Since M
3 0,
the closed convex balanced set U is a neighbourhood of 0, because U
109
contains ~as a subset. Let p be the Minkowski functional of U. Since U

is closed, we have, for any x0 E U, p (x0 ) > 1 and p (x) < 1, if x E U.
Thus there exists, by Corollary 1 of Theorem 1, a continuous linear
functional / 0 on X such that / 0 (x0 ) = p (x0 ) > 1 and I/0 (x) I < p (x) on X.
Hence, in particular, l/0 {x) I< 1 on M.
Corollary. Let M be a closed linear subspace of a locally convex
linear topological space X. Then, for any x 0 EX- M, there exists a
continuous linear functional / 0 on X such that / 0 (x0 ) > 1 and / 0 (x) = 0
on M. Moreover, if X is a normed linear space and if dis (x0 , M) > d,
then we may take II /0 II < 1/d.
Proof. The first part is clear from the linearity of M. The second
part is proved hy taking U = {x; dis (x, M) < d} in the proof of
Theorem 3.
Remark. As above, we prove the following theorem by virtue of
Theorem 1'.
Theorem 3' (S. MAZUR). Let X be a locally convex real linear topological space, and M a closed convex subset of X such that M 3 0. Then,
for any x0 EM, there exists a continuous real linear functional / 0 on X
such that / 0 (x0 ) > 1 and / 0 (x) :S:: 1 on M.
Theorem 4 (S. MAZUR). Let X be a locally convex linear topological
space, and M a convex balanced neighbourhood of 0 of X. Then, for
any x0 E M, there exists a continuous linear functional / 0 on X such that
fo (xo) :? sup Ifo (x)
xEll!
Proof. Let
p be the Minkowski functional of M.
Then
p (x0 ) >
1 and
p (x) < 1 on M. p is continuous since M is a neighbourhood of 0 of X.
Thus there exists, by Corollary 1 of Theorem 1, a continuous linear functional / 0 on X such that
/ 0 {x0 )
= p(x0 ) > 1 and
l/0 (x) I< p(x) :S:: 1
on M.
Theorem 5 (E. HELLY). Let X be a B-space, and fv /2 , . . . , In be a

finite system of bounded linear functionals on X. Given n numbers
av IX2 , , 1Xn- Then a necessary and sufficient condition that there exists,
for each e > 0, an element x. E X such that
fi(x.)
IXi
(i
1, 2, ... , n)
and
llx.ll ~ y
+e
is that the inequality
holds for any choice of n numbers fJv {3 2, ... , fJnProof. The necessity is clear from the definition of the norm of a
continuous linear functional. We shall prove the sufficiency. We may
110
assume, without losing the generality, that f's are linearly independent;
otherwise, we can work with a linearly independent subsystem of {/;}
which spans the same subspace as the original system{/;}.
We consider the mapping X-'rtp(x)=(/1 (x),/2 (x), ... ,fn(x)) of X
onto the Hilbert space l 2 (n) consisting of all vectors x = (~v ~2 , , ~n)
normed by
!IX II
c~ I~j 2) 112
1
By the open mapping theorem in
Chapter II, 5, the image rp(S,) of the sphereS,= {xE X; llxll < y + e}
contains the vector 0 of l2 (n) as an interior point for every e > 0. Let
us suppose that (cxv cx 2 , , cxn) does not belong to rp(S,). Then, by
Mazur's theorem given above, there exists a continuous linear functional
F on l 2 (n) such that
F((cx1 ,cx 2 ,
..
,cxn))>
sup
llxll:;>y+
IF(rp(x))l-
Since l 2 (n) is a Hilbert space, the functional F is given by an element

n
({31, {32, ... , f3n) E l 2 (n) in such a way thatF((cxv cx2, ... , cxn)) = ..I cxjf3j
J=l
Thus
icxf3
>[.it-(x)fJI
J J
j=l J
J
j=l
for
IJxll S:y +e.
But the supremum of the right hand side for II x II < y
+ e is =
Iii~ /jf3j[\, and this contradicts the hypothesis of the Theorem.
(y
+ e) X
7. Topologies of Linear Maps
Let X, Y be locally convex linear topological spaces on the same scalar

field (real or complex number field). We denote by L (X, Y) the totality
of continuous linear operators on X into Y. L (X, Y) is a linear space by
(cxT+f3S)x=cxTx+f3Sx,
where
T,SEL(X,Y)
and
xEX.
We shall introduce various topologies for this linear space L (X, Y).
i) Simple Convergence Topology. This is the topology of convergence
at each point of X and thus it is defined by the family of semi-norms
of the form
P(T) = P(T; x1 , x2 , .. , Xn; q) = sup q(Txj),
l;;ij;;in
where x1 , x2 , . . . , xn are an arbitrary finite system of elements of X and q

an arbitrary continuous semi-norm on Y. L (X, Y) endowed with this
topology will be denoted by Ls (X, Y). It is clearly a locally convex
linear topological space.
ii) Bounded Convergence Topology. This is the topology of uniform
convergence on bounded sets of X. Thus it is defined by the family of
111
7. Topologies of Linear Maps
semi-norms of the form

p(T) =P(T;B;q)
= supq(Tx)
xEB
where B is an arbitrary bounded set of X and q an arbitrary continuous

semi-norm on Y. L (X, Y) endowed with this topology will be denoted by
Lb (X, Y). It is clearly a locally convex linear topological space.
Since any finite set of X is bounded, the simple convergence topology
is weaker than the bounded convergence topology, i.e., open sets of
Ls (X, Y) are also open sets of Lb (X, Y), but not conversely.
Definition 1. If X, Y are normed linear spaces, then the topology of
Ls (X, Y) is usually called the strong topology (of operators); the one of
Lh(X, Y) is called the uniform topology (of operators).
Dual Spaces. Weak and Weak* Topologies
Definition 1'. In the special case when Y is the real orc:omplexnumber
field topologized in the usual way, L (X, Y) is called tht dual space of X
and will be denoted by X'. Thus X' is the set of all continuous linear
functionals on X. The simple convergence topology is then called the
weak* topology of X'; provided with this topology, X' will sometimes be
denoted by X~. and we call it the weak* dual of X. The bounded convergence topology for X' is called the strong topology of X'; provided
with this topology, X' is sometimes denoted by x; and we call it the
strong dual of X.
Definition 2. For any x E X and x' EX', we shall denote by <x, x')
or x' (x) the value of the functional x' at the point x. Thus the weak*
topology of X', i.e., the topology of X~. is defined by the family of seminorms of the form
p (x') = p (x'; xv x 2 ,
xn)
= sup I<xj, x') I,

l:;;;j:;;;n
where xv x2 , , Xn are an arbitrary finite system of elements of X.

The strong topology of X', i.e., the topology of x; is defined by the family
of semi-norms of the form
p(x') = p(x'; B)= sup l<x, x') I
xEB
where B is an arbitrary bounded set of X.

Theorem 1. If X is a normed linear space, then the strong dual space
x; is a B-space with the norm
l/(xllIIIII= 11sup
.. 11:;;;1
Proof. Let B be any bounded set of X. Then sup II b II = f3 < oo,
and hence
II/II <ex
bEB
implies p(f;B) =supl/~b)J <sup J/(x)J <cxf3.

bEB
l!xii:;;;P
112
On the other hand, the unit sphere 5 = {x; jj x JJ < 1} of X is a bounded

set, and so JJ/11 = p(f; 5). This proves that the topology of X~ is equivalent to the topology defined by the norm II! II
The completeness of x; is proved as follows. Let a sequence {/,.}
of x; satisfy lim Ilin- fm II= 0. Then, for any X EX, lfn(x)- fm(x) I<
n,........oo
JJfn- fmJJ Jjx jj-+ 0 (as n, m-+ =),and hence a finite lim fn(x) = f(x)
~
exists. The linearity of f is clear. The continuity of f is proved by observing that lim fn (x) = f (x) uniformly on the unit sphere 5. Incidentn->00
ally we have proved that lim Jlfn- f II = 0.

~
Similarly we can prove

Theorem 2. If X, Yare normed linear spaces, then the uniform topology (of operators) Lb (X, Y) is defined by the operator norm
JIT!I =sup JITxll
!lxll:;;;l
Definition 3. We define the weak topology of a locally convex linear

topological space X by the family of semi-norms of the form
p(x) =P(x;x~,x;, ... ,x~) =
<
sup l<x,xj)l,
l:;;;j:;;;n
where x~, x;, ... ,

are an arbitrary finite system of elements of X'.
Endowed with this topology, X is sometimes denoted by Xw.
8. The Embedding of X in its Bidual Space X"

We first prove
Theorem 1 (S. BANACH). Let X be a locally convex linear topological
space, and X' its dual space. A linear functional f (x') on X' is of the form
f (x') = <x0 , x')
with a certain x0 E X iff f (x') is continuous in the weak* topology of X'.

Proof. The "only if" part is clear since j<x0 , x') j is one of the seminorms defining the weak* topology of X'. The "if" part will be proved
as follows. The continuity of I (x') in the weak* topology of X' implies
that there exists a finite system of points xv x 2 , , Xn such that
lf(x') I< sup l<xi, x') I Thus
l;;;;js;n
l.(x') =<xi, x')
0 (i = 1, 2, ... , n) implies l(x') = 0.
Consider the linear map L: X'-+ l2 (n), defined by

L(x') = (II(x'), lz(x'),. ., ln(x')).
= L (x;) impliesL (x~ -x;) = 0 so thatli (x~ -x;) = 0 (i = 1, 2, ... , n)

and hence I (x~- x;) = 0. Hence we may define a continuous linear mapF
L (x~)
113
8. The Embedding of X in its Bidual Space X"
defined on the linear subspace L (X') of l 2 (n) by
F (L (x'))
= F (fr(x'), ... , In (x')) = f (x').
This map can be extended to a continuous linear functional defined on

the whole space l 2 (n); the extension is possible since l 2 (n) is of finite
dimension (easier than using the Hahn-Banach extension theorem in
infinite dimensional linear spaces). We denote this extension by the same
letter F. Writing
(Yl, Yz, ... , Yn)
.2: Yiei,
= J=l
where
e;
(0, 0, ... , 0, 1, 0, 0, ... , 0)
with 1 at the j-th coordinate,

we easily see that
F (yl, Yz, , Yn)

Therefore
f(x')
on
.2: Yi
= J=l
IXj, IXj
.2:n 1Xj/;(x') = .2:n IX; <xi, x') =

= J=l
J=l
F(ej).
< >
n IXjX;, x'
,I
J=l
Corollary. Each x0 E X defines a continuous linear functional / 0 (x')

<x0 , x'). The mapping
x; by /0 (x') =
Xo--+ fo
=f
Xo
of X into (X~)~ satisfies the conditions
J(x1 + x2) =
]x1
]x2 , ](1Xx) = 1X](x).
Theorem 2. If X is a normed linear space, then the mapping J is

isometric, i.e., IIJxiJ = !lxll
Proof. We have J/0 (x')J = J<x0 ,x')J < JJx0 JJJJx'JJ so that JJ/oll <
Jlx0 II On the other hand, if x0 =F 0, then there exists, by Corollary 2 of
Theorem 1 in Chapter IV, 6, an element x~ EX' such that <x0 , x~) =
Jlx0 II and Jlx~ II= 1. Hence/0 (x~) = <x0 , x~) = l!x0 II so that lifo II> l!xoll
We have thus proved IJJxiJ = JJxll
Remark. As the strong dual space of x;, the space (X;)~ is a B-space.
Hence a normed linear space X may be considered as a linear subspace
of the B-space (X;); by the embedding x--+ ] x. Therefore the strong
closure of ] X in the B-space (X;); gives a concrete construction of the
completion of X.
Definition 4. A normed linear space X is said to be reflexive if X may
be identified with its second dual or the bidual (X;); by the correspondence x *"*] x above. We know already (see Chapter III, 6) that a Hilbert
space is reflexive. As remarked above, (X;); is a B-space and so any reflexive nonned linear space must be a B-space.
8
114
Theorem 3. Let X be a B-space and x~ any bounded linear functional on x;. Then, for any s > 0 and any finite system of elements
IJ, 12, ... 'In of x;, there exists an Xo EX such that
jjxojj<jjx~!j+s
and
lj(x0)=x~(fj) (j=1,2, ... ,n).
Proof. We apply Helly's theorem 5 in Chapter IV, 6. For any system

of numbers f3v {32 , . , f3n, we have
.i: {3iet.il =I j: Pix~ (li) I=lx~ (,i Pi 1i) I
.
l J=l
J=l
< y
J=l
lli~Pdi/1
wehre y =
jjx~lj, Ct.j = x~(fj),
and hence, again by Helly's theorem, we obtain an x0 EX with the

estimate jjx0 jj < y + s = jjx~ll + s and x0 (1j) = Ct.j (i = 1, 2, ... , n).
Corollary. The unit sphere 5 = {x EX; jjx jj < 1} of a B-space X is
dense in the unit sphere of (X;); in the weak* topology of (X;)'.
Example 1. (c0 )'
9. Examples of Dual Spaces

(ll).
To any IE (c0 )', there corresponds a uniquely determined y1 =

{1Jn} E (ll) such that, for all x = {~n} E (c 0 ),
(x,
I>=
00
.I~.. 1Jn and
n~l
111!1 =
And conversely, any y = {1J,.} E (/1) defines an

any x = {~..} E (c 0 ),
IIYtll
(1)
lr E (c0 )'
such that, for
00
(x,ly) = ..~~n1Jn and lllrl! = jjyjj.
(1')
Proof. Let us define the unit vector ek by

k-1
ek
(0, 0, ... , 0, 1, 0, 0, ...) (k
1, 2, ... ) .
k
For any x = {~..} E (c0 ) and IE (c0 )', we have, by s-lim ~ ~.. e,. = x.
(x, I)= lim
k-+oo
(1 ~. e.. , 1) =lim 1 ~.
n=l
k-+oo n=l
k-+oo
1],.,
n~l
1Jn = l(e,.).
Let 1Jn = s,. inn j for 1Jn # 0, and s,. =

for 1Jn = 0. Take x<n,) =
{~..} E (c0 ) in such a way that~.. = s;;- 1 for n < n 0 , and~.. = 0 for n > n 0
n,
ljx<n,) II< 1 and so IIIII = sup l<x, I> I> l<x<n,l, I> I= .I ln.. l
llxll~l
n=l
Thus, by letting n 0 ~ =, we see that Yt = {1Jn} E (ll) and jj Yt j! =
Then
00
n~ Inn I< IIIII
115
then\ ..~ ~.. 1],.1
for all
< llxiiIIYII
x = {~..}E (c0 ), and soy defines an /,E (c0 )' and II!, II< IIY!!
If, conversely, y
= {?J,.}E (ll),
Example 2. (c)' = (ll).

For any x = {~..} E (c), we have the representation
.I (~.. = ~0 e0 + s-lim
k--o-oo n=l
~0 )
e,., where ~0
lim ~... e0 = (1, 1, 1, ... ) .
Thus, for any f E (c)', we have
lim< l (~.. -~0) e,., t) = ~0 1]~ + Z (~,.-~0)1],.,

~0 (e0 , /) + k--o-oo
( )
n=l
n=l
(x, /) =
where 1]~ = (e0 , /) and 1],. = (e,., /) (n = 1, 2, ...). As above, we may

take x<n,) = {~..} E (c0) ~ (c) which satisfies
...
.I I?Jnl
llx(n,) II < 1, ~0 = lim ~.. = 0 and (x<n,), /) = n=l
by f<x<nol,f>l < llx<nolllll/11. we see that {?J,.} E(ll). We
~
Hence,
1]~-
set
00
00
.I 1Jn = 1Jo Then, by (2), we have
n=l
00
(2')
.I~.. 1],., where x = {~..} E (c) and ~0 = lim ~...
+ n=l
1Jn = e,. I?Jn I for 1Jn # 0, and e,. = oo for 1Jn = 0 (n = 0, 1, 2, ... ).
(x, /) = ~0 1Jo
Let
Take x
= {~..}E (c)
such that
~.. = e;;- 1 if n
< n0 , and ~.. = e01 if n > n0
...
~ ln.. l +
+ n=l
lim ~.. =e0\ and (x,/) = I?Jol
'
.I ln.. I< II/I!
e0 1 .I 'YJn Hence, we must have I?Jol + n=l
n=no+l
.I !n.. l < oo,
If, conversely, y = {n..}g" is such that II y I! = Ino I + n=l
Then
llxll < 1,
~0 =
00
00
00
then
1Jo lim ~..
~
.I~.. 17,.,
+ n=l
00
defines an/, E(c)' such that
where x = {~..}i E (c),

00
.I ln.. I
II/, II < I?Jo I + n=l
Therefore, we have proved that (c)' = (ll) as explained above.

Example 3. LP(S, ~. m)' = Lq(S, ~. m) (1 < p < oo and p-1
q-1 = 1). To any /E LP(S)', there corresponds a y1 E Lq(S) such that
(x, /) = Jx(s) YJ(s) m(ds) for all xE LP(S) and II/II= IIY!II.
s
and conversely, any y E Lq (S) defines an/, E LP (S)' such that
(x,f,) =
8*
Jx(s) y(s) m(ds) for all xE LP(S) and II!, II= IIYII
+
(3)
(3')
116

oo
Proof. LetS= .U Bi with 0 < m(Bj) <ex:> and set B(n) = .U Bi.
J=l
J=l
For a fixed n, the defining function CB (s) of the set B ~ B(nJ is E LP (S).
Thus the set function tp (B) = <C B I> is a-additive and m-absolutely
continuous in B ~ B(n). By the differentiation theorem of LebesgueNikodym (see Chapter III, 8), there exists a Yn (s) E V (B(n), 58(n), m)
such that
tp(B) =
Yn(s) m(ds) whenever B ~ B(n),
the family 58(n) of sets being defined by 58(n)= {B (\ B(n); BE 58}.

Therefore, by setting y (s) = Yn (s) for s E B(n), we have
J y(s) m(ds)
<CB,I>=
for BE 58(n) (n = 1, 2, ... ).
Hence, for any finitely-valued function x,
<x, I>=
Let x E LP (S) and put
J x(s) y(s) m(ds).
(4)
Xn(s) = x(s) if lx(s) I< n and s E B(n),

= 0 otherwise.
We decompose the set {z; lz I < n} of the complex plane into a finite
number of disjoint Baire sets Mn,k,t (t = 1, 2, ... , dk,n) of diameters
< 1Jk. Set, for the Xn(s) E L 00 (S, 58, m),
xn,k (s) =a constant z, such that zE (the closure M~,k, 1 ) and Iz I=

whenever Xn(s) EMn,k,t
inf Iw I
wEMn,k,t
Then lxn,k(s) I< lxn(s) I and lim Xn,k(s) = Xn(s) and so, by the Lebek->oo
sgue-Fatou Lemma, s~ Xn,k = Xn (n = 1, 2, ... ). Thus, again by
the Lebesgue-Fatou Lemma,
<xn, I>= k->oo

lim <xnk
I>= k->oo
lim s Xnk(s)
y(s) m(ds)
.
=
(5)
J lim Xn,k(s) y(s) m(ds) = sJ Xn(s) y(s) m(ds).
k->oo
Since s-lim Xn = x, we see that <x, I> = lim <xno I> We put, for any
n->00
n->00
complex number z, a(z) = e-io if z = rei 0 and a(O) = 0. Then llxll >
II (ixn I a(y)) II and so
II lllllxll ~ <lxnl a(y), I>= sJ lxn(s)l
Thus, by the Lebesgue-Fatou Lemma,
ly(s) I m(ds)
IIIII llxll >
lx(s)lly(s)l m(ds)
s
and so the function x(s) y(s) belongs to E Ll(S). Therefore, letting
117
n-+ CXJ in (5), we obtain
(x, I> =
Yn(s)
=
=
x (s) y (s) m (ds) whenever x E LP (5).

s
We shall show that y E Lq (5). To this purpose, set
y(s) if /y(s)/ < n and s E B(n),
0 otherwise.
Then Yn E Lq (5) and, as proved above,
/Ill/ 1/x/1 > </x/ a(y),l)
>
J /x(s)//y(s)/ m(ds)
/x(s)/1 Yn(s)J m(ds).

s
If we take x(s) = /Yn(s)/qfp and apply HOlder's equality, we obtain
/x(s)jjy,.(s)l m(ds) =
(j /x(s)jP m(ds)tp (j /yn(s)/q m(ds)tq.
Hence /Ill/> 1/Ynll = ( / /Yn(s) /q m(ds)tq with the understanding

that, when
p=
1 we have /11/1
>llYn!!= essential sup
/Yn(s)/.
sES
Therefore, by letting n-+ =and applying the Lebesgue-Fatou Lemma,

we see that y E U (5) and /Ill!> 1/Y II On the other hand, any y E Lq (5)
defines an IE LP(5)' by (x, I>= x(s) y(s) m(ds) as may be seen by
s
Holder's inequality, and this inequality shows that /II /1 < IIY II
Remark. We have incidentally proved that
Wr = W)
(1 <
p<
= and
p-1 +
q-1
= 1).
Example 4. Let the measure space (5, \B, m) with m (5) <=have the
property that, for any BE \B with 0< m(B) = (j <=and positive integer n, there exists a subset Bn of B such that (j (n + 1)-1 < m (En)< bn-1 .
Then no other continuous linear functionals EM (5, \B, m)' than the zero
functional can exist.
Proof. Any xE V(5, \B, m) belongs to M(5, \B, m) and the topology
of V(5, \B, m) is stronger than that of M (5, \B, m). Thus any IE M(5, \B,
m)', when restricted to the functions of L 1 (5, \B, m), defines a continuous
linear functionali0 E V(5, \B, m)'. Thus there exists ayE L 00 (5, \B, m)
such that
(x, I>
= (x, lo> =
J x (s) y (s) m(ds)
whenever xE V (5, \B, m).
Since Ll(5, \B, m) is dense in M(5, \B, m) in the topology of M(5, \B, m),
the condition I =I= 0 implies that lo =I= 0. Thus there exists an e > 0
such that B = {s; /Y (s) I > e} has its measure m(B) = (j > 0. Let Bn ~ B
118
be as in the hypothesis, and let y (s) = re'0 for s E B. Set x,. (s) = e-o
for s E Bn and Xn (s) = 0 otherwise. Then z,. (s) = nx,. (s) converges to 0
asymptotically, that is, s-lim z,. = 0 in M(S, ~. m). But
11-+00
lim (z,., /) = lim (z,., / 0)
11-+00
11-+00
lim
11-+00
J z,. (s) y (s) m (ds) >
~e
>
0,
f.
contrary to the continuity of the functional
Example 5. L (S, ~. m)'.

Let an /E L 00 (S, ~. m)' be given and set, for any BE~. f(C 8 )
= 1p(B) where C8 (s) is the defining function ofthe set B. We have then:
00
B 1 f\ B 2 =
0 implies 1p(B1 + B 2)
= 1p(B1)
+ 1Jl(B2),
(6)
that is 'If' is finitely additive,

the real part 1p1 (B) and the imaginary part 1p2 (B) of
1p(B) are of bounded total variation, that is, sup
<
00
(i = 1, 2).
'If' ism-absolutely continuous, that is m(B)
!'lf'(B)I
0 implies
'IJ'(B) = 0.
(7)
(8)
The condition (6) is a consequence of the linearity off, and (7) and (8)
are clear from 11Jl(B)I < IIIII IICBII
For any x E L00 (S, ~, m), we consider a partition of the sphere
{z; Iz I < II x II} of the complex plane into a finite system of disjoint Baire
sets A 1 , A 2 , , A., of diameters <e. If we set B, = {s E S; x(s) E A,},
then, no matter what point tXi we choose from Ai (i = 1, 2, ... , n), we
have
and so
\t(x)- ~~tXi'IJ'(B,)\ < IIIII e.
Thus, if we let n ~ oo in such a way that e
f (x)
.
lim ..I
=1
t 0, we obtain
(9)
tXi'IJ' (Bi),
independently of the manner of partition {z; Iz I <
II x II} =
.
..I A, and
-1
choice of points o.:'s. The limit on the right of (9) is called Radon's integral
of x(s) with respect to the finitely additive measure m. Thus
f(x) =
and so
f x(s) 1p(ds) (Radon's integral) whenever xE L

IIIII =
sup
00
(S, ~. m). (10)
[ x(s) 'IJ'(ds)[.
(11)
ess.sup!z(sll ;:;;1 S
119
Conversely, it is easy to see that any 1p satisfying (6), (7) and (8) defines
an f E L 00 (S, ~. m)' through (10) and that (11) is true.
Therefore, we have proved that L 00 (S, ~. m)' is the space of all set
functions 1p satisfying (6), (7) and (8) and normed by the right hand side
of (11), the so-called total variation of "P
Remark. We have so far proved that LP(S, ~. m) is reflexive when
1 < p <=However, the space V (S, ~. m) is, in general, not reflexive.
Example 6. C (S)'.
LetS be a compact topological space. Then the dual space C (S)' of the
space C (S) of complex-valued continuous functions on S is given as
follows. To any f E C (S)', there corresponds a uniquely determined
complex Baire measure f.l on S such that
f (x) =
and hence
111!1 =
sup
supJx(s)! :;;;1
s
Jx (s) f.l (ds)
J x(s) fl,(ds)
whenever x E C (S),
=the total variation of f.l on S.
(12)
(13)
Conversely, any Baire measure f.l on S such that the right side of (13)
is finite, defines a continuous linear functional f E C (S)' through (12)
and we have (13). Moreover, if we are concerned with a real functional/
on a real B-space C (S), then the corresponding measure f.l is real-valued;
if, moreover f is positive, in the sense that f(x) > 0 for non-negative
functions x (s), then the corresponding measure f.l is positive, i.e.,f.l (B)> 0
for every B E ~.
Remark. The result stated above is known as the F. Riesz-A. MarkovS. Kakutani theorem, and is one of the fundamental theorems in topological measures. For the proof, the reader is referred to standard text
books on measure theory, e.g., P.R. HALMOS [1] and N. DUNFORD]. SCHWARTZ [1].
References for Chapter IV
For the Hahn-Banach theorems and related topics, see BANACH [1],
BouRBAKI [2] and KoTHE [1]. It was MAZUR [2] who noticed the importance of convex sets in normed linear spaces. The proof of Helly's theorem
given in this book is due toY. MIMURA (unpublished).

In this chapter, we shall be concerned with certain basic facts pertaining to strong-, weak- and weak* convergences, including the comparison of the strong notion with the weak notion, e.g., strong- and weak
measurability, and strong- and weak analyticity. We also discuss the
120
integration of B-space-valued functions, that is, the theory of Bochner's

integrals. The general theory of weak topologies and duality in locally
convex linear topological spaces will be given in the Appendix.
1. The Weak Convergence and The Weak* Convergence

Weak Convergence
Definition 1. A sequence {xn} in a normed linear space X is said to
be weakly convergent if a finite lim f (xn) exists for each f E x;; {xn} is
~
said to converge weakly to an element x00E X if lim / (xn) = / (x00 ) for
x;.
all IE
In the latter case, Xoo is uniquely determined, in virtue of
the Hahn-Banach theorem (Corollary 2 of Theorem 1 in Chapter IV, 6); we
shall write w-lim Xn = X00 or, in short, Xn.....,. X00 weakly. X is said to be
~
sequentially weakly complete if every weakly convergent sequence of X

converges weakly to an element of X.
Example. Let {xn(s)} be a sequence of equi-bounded continuous functions of C [0, 1] which is convergent to a discontinuous function z(s) on
[0, 1]. Then, since C [0, 1]' is the space of Baire measures on [0, 1] of
bounded total variation, we see easily that {xn(s)} gives an example of
a weakly convergent sequence of C [0, 1] which does not converge weakly
to an element of C [0, 1].
Theorem 1. i) s-lim Xn = x00 implies w-lim Xn = X00 , but not conver~
sely. ii) A weakly convergent sequence {xn} is strongly bounded, and, in

particular, if w-lim Xn = X00 , then {jJxnll} is bounded and Jjx00 J! <
lim llxnll
~
n->OO
Proof. i) The first part is clear from If (xn) -I (xoo) I < II/II II Xn- Xoo II
The second part is proved by considering the sequence {xn} in the Hilbert
space (l2):
Xn ={~~)}where~~) = c5nm (= 1 or 0 according as n = m or not).
For, the value of a continuous linear functional E (l 2)' at x = {~n} is given
by ]
n=l
~n ijn with some {1Jn} E (l 2); consequently, w-lim Xn = 0 but {xn}

~
does not converge strongly to 0 because II Xn II = 1 (n = 1, 2, ... ) .

ii) Consider the sequence of continuous linear functionals Xn defined on
the B-space
by Xn(f) = <xn, /),and apply the resonance theorem in
Chapter II, 1.
Theorem 2 (MAZUR}. Let w-lim Xn = x00 in a normed linear space X.
x;
n->00
Then there exists, for

( !Xj
>
0,
i:
J=l
rXj
any e > 0,
a convex combination
1) of x/s such that l[x

I
00 -
.i:
rXjxjl <e.
J=l
I
.I rXjXj
J=l
121

n
Proof. Consider the totality M 1 of elements of the form ..I IXjXj
with IXj
J=1
> 0, ..I IXj

)=1
= 1. We may assume that 0 E M 1 by replacing x 00
and Xj by (x00 -x1) and (xj-x 1 ), respectively. Suppose that llxoo -u II> e
for every u E M 1 . The set M ={vEX; llv- ull < e/2 for some u E M 1}
is a convex neighbourhood of 0 of X and llx00 -v II> e/2 for all v EM.
Let p(y) be the Minkowski functional of M. Since x00 = {3-1 u0 with
p (u0) = 1 and 0 < f3 < 1, we must have p.(x00 ) = {3-1 > 1. Consider a
real linear subspace xl = {x EX; X= y Uo,- CXl < y <ex:>} and put
11 (x) = y for X = y Uo E xl. This real linear functional/! on x1 satisfies
I (x) < p(x) on X 1 . Thus, by the Hahn-Banach extension theorem in Chapter IV, 1, there exists a real linear extension I of / 1 defined on the real
linear space X and such that I (x) < p (x) on X. M being a neighbourhood
of 0, the Minkowski functional p (x) is continuous in x. Hence f is a continuous real linear functional defined on the real linear normed space X.
We have, moreover,
sup I (x) < sup I (x) < sup
xEM
xEM 1
xEM
P(x)
= 1 < {3- 1 =
I ({3-1 u 0 )
I (x
00 ).
Therefore, it is easy to see that x 00 cannot be a weak accumulation point

of Mv contrary to X 00 = w-lim xw
~
Theorem 3. A sequence {xn} of a normed linear space X converges

weakly to an element x 00 E X iff the following two conditions are satisfied: i) sup llx.,ll <ex:>, and ii) lim l(xn) = f(xoc,) for every I from any
n~l
x;.
strongly dense subset D' of

Proof. We have only to prove the sufficiency. For any g Ex; and
e > 0, there exists an IE D' such that llg- I II< e. Thus
lg(x,;)- g(x00 ) I< lg(xn)- l(xn) I + ll(xn) -f(xoo) I + ll(xoo)- g(xoo) I

<sllxnll + ll(xn)-l(xoo)l +sllxooiJ,
and hence lim lg(xn)-g(x00 )1 < 2s sup llxnll This proves that
lim g (xn)
~
oo<;:n~l
g (x00).
Theorem 4. A sequence {xn} in V (5, >8, m) converges weakly to an elex.. (s) m(ds)
ment xEV(S, )B,m) iff {llx.. JI}isboundedand a finite lim
~B
exists for every B E )8.

Proof. The "only if" part is clear since the defining function C B (s)
of BE )8 belongs to L 00 (S, )8, m) = V(S, >a, m)'.
Xn (s) m(d,s)
The proof of the ''if'' part. The set function 1p (B)= lim
~B
BE )8, is a-additive and m-absolutely continuous by the Vitali-Hahn-Saks
122
theorem. Hence, by the differentiation theorem of LEBESGUE-NIKODYM,

there exists an x00 E V(5, 58, m) such that
lim J Xn(s) m(ds) = J X00 (s) m(ds) for all BE 58 .
..-oo B
Thus, for any decomposition 5

lim
..-oo 5
Xn(s) y(s) m(ds)
= .I B; with Bi E 58, we have

J~l
JX
00
(s) y(s) m(ds), y(s)
= J:!XjCB;(s) .
j~l
Since such functions as y (s) constitute a strongly dense subset of the

space L 00 (5, 58, m) = V (5, 58, m)', we see that the "if" part is true by
Theorem 3.
Theorem 5. Let {xn} converge weakly to x 00 in V(5, 58, m). Then
{xn} converges strongly to x 00 iff {xn(s)} converges to x00 (s) in m-measure
on every 58-measurable set B such that m(B) <ex:>.
Remark. {xn(s)} is said to converge to x00 (s) in m-measure on B, if,
for any e > 0, them-measure of the set {sE B; lxn(s) -x00 (s) I>- e}
tends to zero as n-+ ex:> (see the Proposition in Chapter I, 4). The space
(ll) is an example of V (5, 58, m) for which 5 = {1, 2, ... } and m ({n}) = 1
for n = 1, 2, ... In this case, we have (ll)' = (l00 ) so that the weak convergence of{xn}, Xn = (~\."), ~~n), . . . , ~hn), ... ), tox00 = (~iool, ~~oo), ... , ~h00 l, ... )
implies that lim ~hn) = ~hoo) (k = 1, 2, ... ), as may be seen by taking
n-->00
IE (ll)' in such a way that f(x) = <x, /) = ~k for x ={ME (ll). Thus,
in the present case, {xn} converges to x 00 in m-measure on every 58measurable set B of finite m-measure. In this way we obtain the
Corollary (I. ScHUR). In the space (ll), if a sequence {xn} converges
weakly to an X00 E {tl), then s-lim Xn = X00
n~co
Proof of Theorem 5. Since the strong convergence in V(5, 58, m)

implies the convergence in m-measure, the "only if" part is clear. We shall
prove the "if" part. The sequence {xn- x00 } converges weakly to 0, and so
lim
J (xn (s) -
n-->OOB
x 00 (s)) m (ds) = 0 for every BE 58.
(1)
Consider the sequence of non-negative measures
"Pn(B) = Jixn(s)-x00 (s)im(ds), BE 58.

B
Then we have
lim "Pn(Bk) = 0 uniformly in n, for any decreasing
k~
sequence {Bk} of sets E 58 such that
00
k~l
Bk
= 0.
(2}
If otherwise, there exists an e > 0 such that for each k there exists some

nk for which}~ nk =
CXJ
and
'~Pnk(Bk)
123
>B. Consequently, we must have
J 1Re(xnk(s)-x00 (s)) lm(ds)>sfV2 or J lim (xnk(s) -x00 (s)) Im(ds)

Bk
Bk
>
s/(2, and so there must exist some B~ ~ Bk such that
11
(xnw (s)- X 00 (s)) m (ds)l
>
s/2 {2 (k
1, 2, ... ) ,
-contrary to the fact that, in virtue of (1), the m-absolute continuity ol

(xn (s) - Xoo (s)) m (d)s is uniform
the sequence of measures lfn (B) =
iJ
inn (see the proof of the Vitali-Hahn-Saks Theorem in Chapter II, 2}.
Next let B 0 be any set of 5B such that m(B0 ) <=We shall show that
lim 1fln (B0 ) = 0.
Suppose there exist an
>
(3)
n-->00
0 and a subsequence {1fJn} of {1fJn} such that
>
1fln(B0 )
(n
(4)
1, 2, ... ).
By the hypothesis that {(xn (s)- X 00 (s))} converges to 0 in m-measure on

B 0 , there exist a subsequence {(xn" (s)- X 00 (s))} of {(xn' (s)- X 00 (s))} and
somesetsB~ ~ B 0 such thatm(B~) < 2-nand lxn(s)-x00 (s) I <s/m(B0 )
00
on (B0 - B~). We put Bk = U B~. Then {Bk} is a decreasing sequence

such that
m(
k~l
Bk):S
n~k
_i m(B~) :S 2-k+l
n~k
(k= 1, 2, ... ) and so m
(n
k~l
Bk) = 0.
Hence, by (1) and the Corollary of the Vitali-Hahn-Saks Theorem referred

to above, lim 1fln (Bk) = 0 uniformly inn. Therefore
k-->00
1fln" (B0 )
<
1fln" (B~)
+ B m (B0)-l m (B 0 -
B;;)--*
(< s) as
n--*
CXJ,
-contrary to (4). This proves (3).

Now we take a sequence {B~} of sets E 5B such that m (B~) < =
00
(k = 1, 2, ... ) and S = U B~. Then

k~l
J +
Jlxn(s)-x 00 (s)lm(ds)=
U B'
k~l
Jt
5- U B'
k~l
1c
By (3) the first term on the right tends to zero as n--* =for fixed t, and
the second term on the right tends, by (2), to zero as t--* CXJ uniformly
inn. Therefore we have proved that s-lim Xn = x00 in V(S, 58, m).
n-->00
A similar situation in the case of the space ~ (.Q)' is given by

Theorem 6. Let {Tn} be a sequence of generalized functions E ~ (.Q)'.
If lim Tn =Tin the weak* topology of 'l)(.Q)', then lim Tn = T m
n-->00
the strong topology of
n-->00
~ (.Q)'.
124
Proof. The strong topology of the space ~ (.Q)' is defined (see Definition 1 in Chapter IV, 7) through the family of semi-norms
Ps(T) =sup IT(cp) I, where )8 is any bounded set of ~(.Q).

~pE$
The weak* topology of the space

semi-norms
PF(T) =sup IT(cp) I, where
(.Q)' is defined through the family of
iJ is any finite
set of ~(.Q).
~pEl)'
Thus lim T" = T

n-HX>
in the weak* topology of
(.Q)' is precisely
lim T,. = T (~ (.Q)') defined in Chapter II, 3.
n-+00
Let )8 be any bounded set of ~ (.Q). Then there exists a compact

subset K in .Q such that supp (rp) ~ K for any cp E )8 and sup IDicp (x) I
xEK,<pE$
< oofor any differential operator Di (Theorem 1 in Chapter I, 8). Thus, by

the Ascoli-Arzela Theorem, )8 is compact in~ K (.Q). We apply the uniform
boundedness theorem to the sequence {T,.- T} to the effect that, for any
e > 0, there exists a neighbourhood U of 0 of ~ K (.Q) such that
sup I(T,.-T)(cp)l<e.
n;<pEU
The compact subset )8 of ~K(.Q) is covered by a finite system of sets of

the form Cf?i + U, where rpi E )8 (i = 1, 2, ... , k). Hence
I(T.,-T) (cpi
+ u)l <
I(T,.-T) (rpi)l
< I(Tn-T) (rpi)
+ I(T,.-T) (u)J
+ e for any u E U.
Since lim (T,.- T) (rpi) = 0 fori= 1, 2, ... , k, we have

n-+oo
lim (T,.- T) (cp) = 0 uniformly in cp E $8.
n-+oo
This proves our Theorem.

Theorem 7. A reflexive B-space X is sequentially weakly complete.
Proof. Let a sequence {x..} of X be weakly convergent. Each x,. defines
a continuous linear functional X,. on x; by X,.(x') = <x,., x'). Since
x; is a B-space (Theorem 1 in Chapter IV, 8), we may apply the resonance
theorem. Thus a continuous linear functional on x; is defined by a finite
lim X,. (x') which exists by hypothesis. Since X is reflexive, there exists
n-..oo
an
X 00
EX such that <x00 , x') = lim X,. (x') = lim <x,., x'), that is,
n-+oo
n-+00
Theorem 8. Let X be a Hilbert space. If a sequence {x.,} of X converges

weakly to X 00 EX, then s-lim Xn = X 00 iff lim llx., II= JJxoo 11n-+oo
n-..oo
12f,
Proof. The "only if" part is clear from the continuity of the norm. The
"if" part is clear from the equality
1/ Xn J/ 2 - (xn, Xoo) - (xoo, Xn)
+ II Xoo 11 2
In fact, the limit, as n ~ CXJ, of the right hand side is llxoo 1/ 2 -1/xoo 1! 2 1/xoo 11 2 + 1/xoo 11 2 = 0.
VVeak* Convergence
Definition 2. A sequence {/n} in the dual space
x;
of a normed
linear space X is said to be weakly* convergent if a finite lim
11---+00
fn (x) exists
x;
for every x E X; {In} is said to converge weakly* to an element / 00 E

fn=/00
if lim In (x) = /00 (x) for all xE X. In the latter case, we write w*-lim
,._,.00
n--->00
or, in short, In~ /00 weakly*.
In = / 00 , but not conIn= /00 implies w*-lim

11---+00
n--->oo
versely. ii) If X is a B-space, then a weakly* convergent sequence
{In} <;;; converges weakly* to an element /00 Ex; and llfoo II < !!m Ilin II
n->oo
Proof. (i) The first part is clear from lin (x) - /00 (x) I< Ilin -foo 111/x II
The second part is proved by the counter example given in the proof of
Theorem 1. (ii) By the resonance theorem, we see that /00 (x) = lim In (x)
Theorem 9. i) s-lim
x;
is a continuous linear functional on X and llfoo II
<
11---+00
lim
n-->00
llfn II
Theorem 10. If X is a B-space, then a sequence {In} ~
x; converges
lim fn(x) =
weakly*toanelement/00 EX;iff(i){ll/nll}isbounded,andii) n->OO
/ 00 (x) on a strongly dense subset of X.
Proof. The proof is similar to that of Theorem 3.
Strong and VV eak Closure
Theorem 11. Let X be a locally convex linear topological space, and M

a closed linear subspace of X. Then M is closed in the weak topology of X.
Proof. If otherwise, there exists a point x 0 EX- M such that x 0 is
an accumulation point of the set Min the weak topology of X. Then,
by the Corollary of Theorem 3 in Chapter IV, 6, there exists a continuous linear functional/0 on X such that / 0 (x0) = 1 and / 0 (x) = 0 on M.
Hence x 0 cannot be an accumulation point of the set M in the weak
topology of X.
126
2. The Local Sequential Weak Compactness of Reflexive B-spaces.

The Uniform Convexity
Theorem 1. Let X be a reflexive B-space, and let {x,.} be any sequence
which is norm bounded. Then we can choose a subsequence {xn} which
converges weakly to an element of X.
We will prove this Theorem under the assumption that X is separable,
since concrete function spaces appearing in applications of this Theorem
are mostly separable. The general case of a non-separable space will be
treated in the Appendix.
Lemma. If the strong dual x; of a normed linear space X is separable,
then so is X.
Proof. Let {x~} be a countable sequence which is strongly dense on the
surface of the unit sphere {x' Ex;; Jlx' II= 1} of x;. Choose Xn EX so
that II x,. II = 1 and I<x,., x~) I > 1/2. Let M be the closed linear subspace
of X spanned by the sequence {xn} Suppose M =F X and x0 EX- M.
By Corollary of Mazur's Theorem 3 in Chapter IV, 6, there exists an
x~ Ex; such that IJx~ II= 1, <x0 , x~) =F 0 and <x, x~) = 0 whenever
xE M. Thus <x.. , x~) = 0 (n = 1, 2, ... ), and so 1/2 < J<x,., x~) I<
I<x.. , x~)-<xn.x~>l + J<x.. , x~>l which implies that 1/2 < llx,.l! JJx~-x~l[
= Jl x~- x~ II This is a contradiction to the fact that {x~} is strongly
dense on the surface of the unit sphere of x;. Thus M =X, and so linear
combinations with rational coefficients of {x..} are dense in X. This proves
our Lemma.
Proof of Theorem 1. As we have remarked above, we assume that X
is separable and so (X;); = X is separable also. By the preceding Lemma,
x; is separable. Let {x~} be a countable sequence which is strongly dense
in x;. Since {x..} is norm bounded, the sequence {<xn, x~)} is bounded.
Thus there exists a subsequence {x...} for which the sequence {<xn, x~)}
is convergent. Since the sequence {<x,.,, x;)} is bounded, there exists a
subsequence {x..,} of {xn,} such that {<x,.,, x;)} is convergent. Proceeding
in this way, we can choose a subsequence {xn,+J of the sequence {x,.J
such that the sequence of numbers {<x,.iH' xj)} converges for j =
1, 2, ... , i
1. Hence the diagonal subsequence {x....} of the original
sequence {x,.} satisfies the condition that the sequence {<xnn xj)} converges for j = 1, 2, ... Thus, by Theorem 3 in the preceding section,
lim <x..,., x') exists and is finite for every x' EX'. Hence, by Thea-
rem 7 of the preceding section, we see that
w~
x.... exists.
Milman's Theorem
We owe to D.P. MILMAN a theorem that a B-space is reflexive when

it is uniformly convex in the sense that, for any e > 0, there exists a
2. The Local Sequential Weak Compactness of Reflexive B-spaces
127
b = b(e) > 0 such that JJxJJ < 1, JJyJJ < 1 and JJx-yJI > e implies
llx + y II< 2(1- b). A pre-Hilbert space is uniformly convex as may be
seen from the formula
y 11 2 = 2(II X 11 2 + II y 11 2 )
valid in such a space. It is known that, for 1 < p < CXJ, the spaces LP and
(lP) are uniformly convex (see J. A. CLARKSON [1]).
II X + y 11 2 + II X -
Theorem 2 (MILMAN [1]). A uniformly convex B-space X is reflexive.

Proof (due to S. KAKUTANI). Given an x~ E (X~)~ with llx~ II= 1.
Then there exists a sequence {1,.} ~ X~ with II In II = 1, x~ (fn) > 1-n- 1
(n = 1, 2, ... ). By Theorem 5 in Chapter IV, 6, there exists, for
every n, an Xn EX such that
(i=1,2, ... ,n)
l;(xn)=x~(f;)
llx,.ll<llx~ll+n- 1 =1+n- 1
and
Since
1-n- 1 < x~ (/,.) = l,.(x,.) 00 '
If the sequence {x,.} does not converge strongly, there exists an e > 0
and n1 < m1 < n2 < m2 < . < nk < mk < such that e <
llx,.k-xmkll (k = 1, 2, ... ). Thus, by lim llxnll = 1 and the uniform
n-->-00
convexity of X, we obtain lim llxnk

since nk
<
k--->00
mk,
+ Xmkll <
2(1- b(e))
<
2. But,
Ink (xnJ = Ink (xmk) = x~ (Ink) and so
2(1- nj; 1 ) ::=;: 2x~ Unk) =Ink (x,.k + XmJ < Ilink II llxnk + Xmk II
Hence, by Jllnkll = 1, we obtain a contradiction lim llxnk + Xmkll >
k--->00
We have thus proved the existence of s-lim xn

n--->00
2.
= x0 , and x0 satisfies
(1)
We show that the solution of the above equations in (1) is unique. Otherwise, there exists an 0 =F x0 which satisfies the same equations. By the
uniform convexity, 1Jx0 + x0 ll < 2. We also have l;(x0 + x0 ) = 2x~(f;)
(i = 1, 2, ... ). Thus
2(1-i-1) < 2x~(/;) = l;(x0 + x0) < III;IJIIxo + xoll = llxo + xoiJ,
and so llx0 + x0 II >_lim 2 (1- i- 1 ) = 2 which is a contradiction.
Finally let lo be any point of x;. If we show that lo (x0 ) = x~ (10),
then (X;);~ X and the reflexivity of X is proved. To prove that lo (x0 ) =
--+00
we take 10 , lv ... , In, .. in place of

hence we obtain 0 EX such that
x~ (/0 ),
llxoll =
1, l;(xo)
x~(f;) (i
ft, 12 ,
0, 1,
. ,
.,
In, ... above, and
n,
.);
128
we must have 0 = x 0 by virtue of the uniqueness just proved above,

and so the proof of Theorem 2 is completed.

Definition 1. Let Z be an open domain of the complex plane. A mapping x (C) defined in Z with values in a B-space X is called weakly holomorphic in Cin the domain Z if, for each f E X', the numerical function
of
Cis holomorphic in Z.
f(x(C))
<x(C), /)
Theorem 1 (N. DuNFORD [2]). If x(C) is weakly holomorphic in Z,

then there exists a mapping x' (C) defined in Z with values in X such that,
for each Co E Z, we have
s-lim h- 1 (x(C0 + h)- x(Co)) = x' (Co).
h---+0
In other words, the weak holomorphic property implies the strong holomorphic property.
Proof. Let C be a rectifiable Jordan curve such that the closed bounded
domain l: enclosed by C lies entirely in Z and Co E l:- C. Let Z 0 be any
open complex domain 3 Co such that its closure lies in the interior of l:.
Then, by Cauchy's integral representation, we have
f(x(Co)) =
2 ~i J~(x(~~ dC.
c
Hence, if both ' 0 +hand ' 0 + g belong to Z 0 ,
(h _ g)-1 f(x (~o
= 2~i
+ h~- f(x(~o)) _
f(x(~o
+ g~- f(x(~0))}
Jf(x(m{a;-~o-h) (~-=-~o-g) (~-~oJac.
c
By the assumption, the distance between Z 0 and Cis positive. Hence, for
fixed f E X', the absolute value of the right hand side is uniformly bounded when Co ,C0 + h and Co + g range over Z 0 Thus, by the resonance
theorem, we have
-II
sup
__
1
~x(~0 +h)- x(~0 ) _ x(~0 +g) -x(~0 )}11 < oo.
~ ~,+h,C,+gEZ, jh-gj l
h
g
Therefore, by the completeness of the space X, x (C) is strongly differentiable at every Co E Z.
Corollary 1 (Cauchy's integral theorem). The strong differentiability
of x (C) implies its strong continuity in C. Thus we can define the curvilinear integral x (C) dC with values in X. Actually we can prove that
c
x(C) dC = 0, the zero vector of X.
c
129
Proof. We have, by the continuity and the linearity of/EX',
l(j x(C) dC) = j
f(x(C))dC.
But the right hand side is zero, because of the ordinary Cauchy integral
x(C) dC = 0 by
theorem. Since/EX' was arbitrary, we must have
c
Corollary 2 of Theorem 1 in Chapter IV, 6.
From the above Corollary, we can derive other Corollaries, as in the
ordinary theory of functions of a complex variable.
Corollary 2 (Cauchy's integral representation).
x(C0 ) =
2 ~i c Cx(C~
dC
for any interior point Co of c.
Corollary 3 (Taylor's expansion). For any point Co which is in the

interior of the closed domain c, the Taylor expansion of x(C) at C=Co
converges strongly in the interior of the circle with centre at C0 , if this
circle does not extend outside of c:
x(C)
00
=.I (n!)- 1 (C-Cotx(n)(Co). where

11=0
x(n)(C)o -
__1Z__!__
J~-)-dC
2ni c
(C- Col"+l
Corollary 4 (Liouville's theorem). If x(C) is (strongly) holomorphic in

the whole finite plane: ICI < CXJ, and sup Ix (C) I < CXJ, then x (C) must
ICI<oo
reduce to a constant vector x(O).
C, then, as r =,
curve
the
for
r
=
C
J
Proof. If we take
J
llx(n) (0) !I=;' sup llx(C) II

n
ICI<oo
J :il-+ 0
-+
(n = 1, 2, ... ) .
Hence the Taylor expansion of x(C) at C = 0 reduces to the constant

term x(O) only.
We shall now apply Corollary 4 to the proof of the Gelfand-Mazur theorem. We first give
Definition 2. A commutative field X over the field of complex numbers
is called a normed field, if it is also a B-space such that the following
conditions are satisfied:
II eII =
1, where e is the unit of the multiplication in X,
JlxyJI
< JlxJIJiyJI, wherexyisthemultiplicationinX.
(1)
Theorem 2 (GELFAND [2]-MAZUR [1]). A normed field X is isometrically isomorphic to the complex number field. In other words, every element x of X is of the form x = ; e where ; is a complex number.
9
130
Proof. Assume the contrary, and let there exist an x EX such that
(x- ~e) =F 0 for any complex number~- Since X is a fitld, the non-zero
element (x- ~e) has the inverse (x- ~e)-1 EX.
We shall prove that (x- A.e}-1 is (strongly) holomorphic in ). for
[A.[ < oo. We have, in fact,
h- 1 ((x- ().
h)e)- 1 - (x-A.e)- 1 )
= h- 1 (x- (A. + h) e)- 1 {e- (x- (A.+ h) e) (x- A.e)- 1}

= h- 1 (x- ().+h) er 1 {e- e + h(x- A.e)- 1}
= (x -- (). + h) e)- 1 (x- A.e)- 1.
On the other hand, forsufficientlysmall[h [,the seriesy- 1
(e + n~/hy- 1 t),
where y = (x- ).e), converges by (1), and it represents the inverse

(y- he)-1 = y-1(e- hy-1)-1, as may be seen by multiplying the series
by (y- he). Hence, by the strong continuity in h of the series, we can
prove that (x-). e)-1 is (strongly) holomorphic in ). with the strong
derivative (x- ).e)-2.
Now, if [A.[> 2 !lx!l, then, as above, (x-A.e)- 1 = -).-1(e-).-1x)-1
=-A.-1 (e+
(A.- 1xt) and so
[[(x-A.e)- 1 J[ < [A.- 1 J(1 +
n~ (1/2t)~ 0
as
[A.[~oo.
Moreover, the function (x- A.e)-1, being continuous in A., is bounded on

the compact domain of).: [A./ S:: 2[[x[[. Hence, by Liouville's theorem,
(x- A.e)-1 must reduce to the constant vector x-1 = (x- Oe)-1. But,
since s-lim (x- A.e)-1 = 0 as proved above, we have arrived at a contral-'1-+>
diction x-1 = 0, e = x- 1 x = 0.
4. The Weak and Strong Measurability. Pettis' Theorem

Definition 1. Let (S, ~. m) be a measure space, and x(s) a mapping
defined on S with values in a E-space X. x (s) is called weakly ~-mea
surable if, for any/EX', the numericalfunction f(x(s)) = <x(s), /)of s
is ~-measurable. x(s) is said to be finitely-valued if it is constant =F 0 on
each of a finite number of disjoint ~-measurable sets Ei with m (Ej) < oo
and x (s) = 0 on S -l) Ei. x (s} is said to be strongly ~-measurable if
J
there exists a sequence of finitely-valued functions strongly convergent

to x (s) m-a.e. on S.
Definition 2. x (s) is said to be. separably-valued if its range {x (s); s E 5}
is separable. It is m-almost separably-valued if there exists a ~-measurable
set E 0 of m-measure zero such that {x(s); s E 5- E 0 } is separable.
4. The Weak and Strong Measurability. Pettis' Theorem
131
Theorem (B. J. PETTIS [1]). x(s) is strongly ~-measurable iff it is

weakly ~-measurable and m-almost separably-valued.
Proof. The "only if" part is proved as follows. The strong ~-measu
rability implies the weak ~-measurability, because a finitely-valued function is weakly ~-measurable, and, by the strong ~-measurability of
x(s), there exists a sequence of finitely-valued functions x,.(s) such that
s-lim x,. (s) = x (s) except on a set B 0 E ~ of m-measure zero. Thus the
11-+00
union of the ranges of x,. (s) (n = 1, 2, ... ) is a countable set, and the
closure of this set is separable and contains the range {x(s); s E S- B0}.
The proof of the "if" part. Without losing the generality, we may
assume that the range {x (s); s E S} is itself separable. So we may assume
that the space X is itself separable; otherwise, we replace X by the
smallest closed linear subspace containing the range of x(s). We first
prove that II x (s) II is itself ~-measurable. To this purpose, we shall make
use of a lemma, to be proved later, which states that the dual space X'
of a separable B-space satisfies the condition
there exists a sequence{/,.}~ X' with
II!,. II< 1 such
that, for any / 0 EX' with lifo II< 1, we may choose a
(1)
subsequence {/,.} of {/,.} for which we have lim /,. (x)

11-+00
= / 0 (x) at every x EX.
Now, for any real number a, put
A= {s; llx(s) II< a}
and
00
.n
If we can show that A =
J=1
A1 = {s; lf(x(s)) I< a},

AIJ, then, by the weak
where
/EX'.
~-measurability
of
~ n AI.
x(s), the function llx(s) II is ~-measurable. It is clear that A -llfll:;;l
But, by Corollary 2 of Theorem 1 in Chapter IV, 6, there exists, for
fixed s, an / 0 E X' with lifo II= 1 and / 0 (x(s)) = llx(s) II Hence thereverse inclusion A ::2 n A 1 is true and so we have A = n A1. By
ll!ll s;l
- llfll :>;1
the Lemma, we obtain
00
n A 11
n A 1 = i=1
ll!ll:;o;1
and so A =
00
n Air
j-1
Since the range {x(s); s E S} is separable, this range may, for any
positive integer n, be covered by a countable number of open spheres
S;, .. (i = 1, 2, ... ) of radius < 1fn. Let the centre of the sphere S;,n be
x;,n As proved above, llx(s)- x;,n II is ~-measurable in s. Hence the
set B;,..
= {s E S; x(s) E S;,,.}
x,.(s)
is ~-measurable and S
x1 ,. if s E s;,.
'
'
= ;'d1 B;,n
00
We set
l-1
U B3,..
= B,,.i-l
'
J
Then, by S = ~~ s;,,., we have lx (s) - x,. (s) I < 1/n for every s E S.
00
9*
132
Since B;,. is ?.8-measurable, it is easy to see that each x,. (s) is strongly ?.8measurable. Therefore x(s), which is the strong limit of the sequence
{x,.(s)}, is also strongly ?.8-measurable.
Proof of the Lemma. Let a sequence {x,.} be strongly dense in X.
Consider a mapping I-+ q;,. (f) = {I (x1 ), I (x2 ), . . . , I (x,.)} of the unit
sphere 5' = {IE X'; II I II < 1} of X' into an n-dimensional Hilbert space
)1/2
l 2 (n) of vectors (~v ~2 . , ~..) normed by II (~1 ~2 .. , ~..)II =v~ l~il 2

The space l 2 (n) being separable, there exists, for fixed n, a sequence
{ln,k} (k = 1, 2, ... ) of 5' such that {q;,. (f,.,k); k = 1, 2, ...} is dense in the
image q;,. (5') of 5'.
We thus have proved that, for any lo E 5', we can choose a subsequence
{l,.,mJ (n = 1, 2, ... ) such that lln,m,. (x;)- lo (x;) I < 1/n (i = 1, 2, ... , n).
Hence lim lnm (x;) = lo (x;) (i = 1, 2, ... ), and so, by Theorem 10 in Chapn---xJO
' n
(
ter V, 1, we obtain that lim ln,m,. (x)
= lo (x)
for every x E X.
5. Bochner's Integral
Let x(s) be a finitely-valued function defined on a measure space
(S, ?.8, m) with values in a B-space X; let x(s) be equal to x; =I= 0 on
B; E l8 (i = 1, 2, ... , n) where B/s are disjoint and m (B;) < oo for
i
= 1, 2, ... , n, and moreover, x (s) = 0 on
define them-integral
J x(s) m(ds)
(s- .i B;). Then we can

~1
of x(s) over S by
i:
X;
m(B;). By
i~1
virtue of a limiting procedure, we can define the m-integral of more

general functions. More precisely, we have the
Definition. A function x(s) defined on a measure space (S, ?.8, m) with
values in a B-space X is said to be Bochner m-integrable, if there exists
a sequence of finitely-valued functions {x,.(s)} which s-converges to x(s)
m-a.e. in such a way that
lim
11-->00
J llx(s)- x,.(s) II m(ds) =
0.
(1)
For any set BE ?.8, the Bochner m-integral of x(s) over B is defined by
Jx(s) m(ds) =s-lim J C8 (s) x,.(s) m(ds),
n---xJO
where C8 is the
(2)
defining function of the set B.

To justify the above definition, we have to verify that the s-limit on
the right of (2} exists and that the value of this s-limit is independent of
the approximating sequence of functions {x,. (s)}.
Justification of the Definition. First, x(s) is strongly ?.8-measurable
and consequently the condition (1) has a sense, since 1/x(s) -- x,.(s) II is
133
m-measurable as shown in the proof of Petti's Theorem. From the inequality
Ill Xn(s) m(ds)
-1 xk(s) m(ds) /I= Ill (xn(s)- xk(s)) m(ds) I
< J llxn(s)- xk(s) II m(ds) < J llxn(s)- x(s)ll m(ds)

S
J llx(s) -xk(s)
s
II m(ds)
and the completeness of the space X, we see that s-lim J xn(s) m(ds)
.......coB
exists. It is clear also that this s-lim is independent of the approximating

sequence, since any two such sequences can be combined into a single
approximating sequence.
Theorem 1 (S. BocHNER [1 ]) . A strongly m-measurable function X (s) is
Bochner m-integrable iff llx(s) II ism-integrable.
Proof. The "only if" part. We have llx (s)jj < llxn (s)/1 + llx(s) -Xn (s) II
By the m-integrability of II Xn (s) II and the condition (1), it is clear that
II x (s) II is m-integrable and
Jllx(s)llm(ds)< Jllxn(s)llm(ds)
B
Jllx(s)-xn(s)llm(ds).
Moreover, since
J !llxn(s) II -1/xk(s) Ill m(ds) < J 1/xn(s)- xk(s) II m(ds),

B
we see from (1) that lim J 1/xn(s) II m(ds) exists so that we have
.......coB
J /lx(s) II m(ds) <lim J llxn(s) II m(ds).

.......coB
The "if" part. Let {xn (s)} be a sequence of finitely-valued functions

strongly convergent to x(s) m-a.e. Put
Yn(s) = Xn(s)
if
if
II (1 + n- 1),
/lxn(s) II> ljx(s) II (1 + n- 1).
/lxn(s) II< /lx(s)
Then the sequence of finitely-valued functions {Yn (s)} satisfies llYn (s) II <
/lx(s) II (1 + n- 1) and lim llx(s)- Yn(s) II= 0 m-a.e. Thus, by the
.......00
m-integrability of llx(s) /1, we may apply the Lebesgue-Fatou Lemma to

the functions llx(s)- Yn(s) II< 2!lx(s) II (1 + n- 1) and obtain
lim J llx(s)- Yn(s) II m(ds) = 0,
.......cos
that is, x(s) is Bochner m-integrable.

Corollary 1. The above proof shows that
l/lx(s) /1 m(ds) >
//1 x(s) m(ds)//,
134
and hence
Jx(s) m(ds) ism-absolutely continuous in the sense that

s-lim J x(s) m(ds) = 0.
. J x(s) m(ds)
m(B~B
The finite additivity
x(s) m(ds) = .~
"
E B;
1=1
so, by virtue of the a-additivity of
J x(s) m(ds)
is clear and
J=1 B;
J llx(s)
IJ
m(ds), we see that
is a-additive, i.e.,
00
B = .~ B; with m(B;)
J=1
<
c:x:>
implies
.
= s-lim .~ J x (s) m (ds).
J x(s) m(ds)
00
E B;
1=1
~J=1B;
Corollary 2. Let T be a bounded linear operc:j.tor on a B-space X into a

B-space Y. If x(s) is an X-valued Bochner m-integrable function, then
Tx(s) is a Y-valued Bochner m-integrable function, and
J Tx(s) m(ds) = T Jx(s) m(ds).

B
Proof. Let a sequence of finitely-valued functions {y,. (s)} satisfy

II y,. (s) II < II x (s) II (1
+ n-1)
and s-limy,. (s) = x (s) m-a.e .

..-co
Then, by the linearity and the continuity ofT, we have

T
J Ty,.(s) m(ds)=
J y,.(s) m(ds). We have, moreover, by the continuity ofT,

B
IITy,.(s) II< IITIIIIYn(s) II< IITIIIIx(s) II (1

s-lim Ty,.(s) = Tx(s) m-a.e.
+ n- 1)
and
Hence T x (s) is also Bochner m-integrable and
JTx(s) m(ds) =s-lim J Ty,.(s) m(ds) =s-lim T Jy,.(s) m(ds)
11-+00
= T
Jx(s) m(ds).
n->00
Theorem 2 (S. BocHNER [1]). Let S be an n-dimensional euclidean

space, ~ the family of Baire sets of S, and m(B) the Lebesgue measure
of B. If x(s) is Bochner m-integrable, and if P(s0 ; ~X) is the parallelopiped
with centre at s0 E S and side length 2X, then we have the differentiation
theorem
s-lim (2X)-"
x(s) m(ds) = x(s0) for m-a.e. s0
"'tO
P(s0 ;<>)
135
Proof. Put
(2X)-"
x(s)m(ds)=D(x;s0 ,X).
P(s0 ;"'}
If {x,.(s)} is a sequence of finitely-valued functions such that /lx,.(s)

/lx(s) /1 (1 + n-1) and s-lim x,.(s) = x(s) m-a.e., then
II<
D (x; s0 , X)- x(s0 ) = D (x- xk; s0 , X)

and so
+ D (xk; s0 , X)- x(s0 ),
lim IID(x; s0 , X) -x (s 0 ) II< lim D(llx- xk 1/; s0 , X)

"'+O
"'+O
+lim liD (xk; s0 , X) - xk (s0) II + l/xds0 ) - x(s0 ) II.

"'+O
The first term on the right is, by Lebesgue's theorem of differentiation of
numerical functions, equal to llx (s0 ) - xk(s0 ) II m-a.e. The second term
on the right is = 0 m-a.e., since xk(s) is finitely-valued. Hence
lim
"'-+0
II D (x; s0 , X)- x(s0 ) II <
2llxk (s0) -x(s0)
II for m-a.e. s0
Therefore, by letting k --Too, we obtain Theorem 2.

Remark. Contrary to the case of numerical functions, a B-spacevalued, a-additive, m-absolutely continuous function need not necessarily
be represented as a Bochner m-integral. This may be shown by a counter
example.
A Counter Example. Let S = [0, 1] and ~ the family of Baire sets
on [0, 1], and m(B) the Lebesgue measure of BE~. Consider the totality
m [1/3, 2/3] of real-valued bounded functions ; = ; (()) defined on the
closed interval [1/3, 2/3] and normed by II;//= sup I;(O) I We define
an m [1/3, 2/3]-valued function x (s)
= ; ((); s) defined on [0, 1] as follows:
the graph in s-y plane of the real-valued function y = y 6 (s),

which is the {}-coordinate ; ((); s) of x (s), is the polygonal line
connecting the three points (0, 0), ((), 1) and (1, 0) in this order.
Then, if s =I= s', we have Lipschitz' condition:
II (s- s')-1 (x(s)- x(s')) II= sup I(s- s')-1 (;(O; s)- ;((); s')) I< 3.
8
Thus, starting with the interval function (x(s)- x(s')) taking values in
m [1/3, 2/3], we can define a a-additive, m-absolutely continuous set
function x(B) defined for Baire set B of [0, 1].
If this function x (B) is represented as a Bochner m-integral, then,
by the preceding Theorem 2, the function x(s) must be strongly differentiable with respect to s m-a.e. Let the corresponding strong derivative
x' (s) be denoted by 'YJ ((); s) which takes values in m [1/3, 2/3]. Then, for
136
Appendix to Chapter V. Weak Topologies and Duality
every () E [1/3, 2/3] and m-a.e. s,

0=1imllh-1 (x(s +h) -x(s))-x'(s)ll >limlh-1 (~(0; s
11-+0
11-+0
+ h)-~(O;s))
-n(O;s)l.
This proves that ~ (0; s) must be differentiable in s m-a.e. for all () E

[1/3, 2/3]. This is contradictory to the construction of~((); s).
References for Chapter V

S. BANACH [1], N. DUNFORD-J. SCHWARTZ [1] and E. HILLE-R. S.
PHILLIPS [1].

in Locally Convex Linear Topological Spaces
The present book is so designed that the reader may skip this appendix in the first reading and proceed directly to the following chapters.
1. Polar Sets
Definition. Let X be a locally convex linear topological space. For
~X, we define its (right) polar set M 0 by
any set M
M 0 = {x' EX'; sup I<x. x') I :S:: 1}.
(1)
xEM
Similarly, for any set M'

0M'
X', we define its (left) polar set 0M' by
= {xE X; sup l<x, x') I< 1} =X(\, (M') 0 ,

>fEM'
(2)
where we consider X to be embedded in its bidual (X;)'.

A fundamental system of neighbourhoods of 0 in the weak topology
of X is given by the system of sets of the form 0M' where M' ranges over
arbitrary finite sets of X'. A fundamental system of neighbourhoods of
0 in the weak* topology of X' is given by the system of sets of the form
M 0 where M ranges over arbitrary finite sets of X. A fundamental
system of neighbourhoods of 0 in the strong topology of X' is given by
the system of sets of the form M 0 where M ranges over arbitrary bounded
sets of X.
Proposition. M 0 is a convex, balanced set closed in the weak* topology
of X'.
Proof. For any fixed x EX, the linear functional f(x') = <x. x') is
continuous in the weak* topology of X'. Thus M 0 = n {m}0 is closed
mEM
in the weak* topology of X'. The balanced convexity of M 0 is clear.
1.
137
Polar Sets
An Application of Tychonov's Theorem

Theorem 1. Let X be a locally convex linear topological space, and A
a convex, balanced neighbourhood of 0 of X. Then A 0 is compact in the
weak* topology of X'.
Proof. Let p (x) be the Minkowski functional of A. Consider, for each
xE X, a sphere S:e = {zE C; lzl < p(x)} and the topological product
S = II S". S is compact by Tychonov's theorem. Any element x' EX'
:eEX
is determined by the set of values x' (x) = (x, x'), x EX. Since x E
(p(x)+e)A for any e>O, we see that x'EX' implies (x,x')=
( (p (x) + e) a, x') with a certain a E A. Thus x E A 0 implies that
Ix' (x) I < p (x) + e, that is, x' (x) E S". Hence we may consider A 0 as a
subset of S. Moreover, it is easy to verify that the topology induced on
A 0 by the weak* topology of X' is the same as the topology induced on
A o in the Cartesian product topology of S = ..,{I Sx.
Hence it is sufficient to prove that A 0 is a closed subset of S. Suppose
y = II y (x) is an element of the weak* closure of A 0 in S. Consider any
:e~X
e > 0 and any x1 , x 2 E X. The set of all u = II u (x) E S such that

xEX
lu(x1 ) -y(x1 ) I< e, lu(x2) -y(x2 ) I< e and lu(x1
+ x 2) -y(x1 + x2) I< e
is a neighbourhood of yin S. This neighbourhood contains some point

x' E A 0 and, since x' is a continuous linear functional on X, we have
ly(x1 + x2 ) - y(x1 ) - y(x2 ) I < ly(x1 + x2) - (x1 + x2 , x') I
+ I(xv x') - y (x1) I + I(x2 , x') - y (x 2) I < 3e.
This proves that y(x1 + x 2) = y(x1 ) + y(x2 ). Similarly, we prove that
y({Jx) = {Jy(x), and soy defines a linear functional on X. By the fact
that y = II y(x) E S, we know that ly(x) I< p(x). Since p(x) is conxEX
tinuous, y (x) is a continuous linear functional, i.e., y EX'. On the other

hand, since y is a weak* accumulation point of A 0 , there exists, for any
e > 0 and a E A, an x' E A 0 such that ly(a)- (a, x') I< e. Hence
ly(a) I< l<a, x') I+ e < 1 + e, and so ly(a) I< 1, that is, yE A 0 .
Corollary. The unit sphereS*= {x' EX'; llx' II < 1} of the dual space
of a normed linear space X is compact in the weak* topology of X'.
x;
An Application of Mazur's Theorem

Theorem 2. Let M be a convex, balanced closed set of a locally convex
linear topological space X. Then M = 0 (MO).
Proof. It is clear that M ~ 0 (M0). If there exists an x 0 E 0 (M0) - M,
then, by Mazur's theorem 3 in Chapter IV, 6, there exists an x~ EX'
such that (x0 , x~) > 1 and I(x, x~) I < 1 for all x E M. The last inequality
shows that x~ E M 0 and so x 0 cannot belong to 0 (M0).
138
2. Barrel Spaces
Definition. In a locally convex linear topological space X, any convex,
balanced and absorbing closed set ic:; called a barrel (tonneau in Bourbaki's
terminology). X is called a barrel space if each of its barrels is a neighbourhood of 0.
Theorem 1. A locally convex linear topological space X is a barrel
space if X is not of the first category.
Proof. Let T be a barrel in X. Since Tis absorbing, X is the union of
closed sets nT = {nt; tE T}, where n runs over positive integers. Since
Xis not ofthefirst category, atleast one of the (nT)' s contains an interior
point. Hence T itself contains an interior point x0 If x0 = 0, Tis a neighbourhood of 0. If x0 =I= 0, then -x0 E T by the fact that Tis balanced.
Thus -x0 is an interior point of T with x0 . This proves that the convex
set T contains 0 = (x0 - x0 )/2 as an interior point.
Corollary 1. All locally convex F-spaces and, in particular, all Bspaces and (;l; (Rn) are barrel spaces.
Corollary 2. The metric linear space 'l) K (Rn) is a barrel space.
Proof. Let {'f?k} be a Cauchy sequence with respect to the distance
d"IS ('f?, 1p)
Pm(rp-!p)
= _~2-m 1 +
p-(-=-) , wh ere p.m ()
'f? =
m~O
m
'P
q;
sup
lil;;;m,xEK
[Di 'f? (x)[ .
For any differential operator Di, the sequence {Di'f?k(x)} is equi-continuous and equi-bounded, that is,
lim
lx1-x'l+ 0
This
sup [Di'f?k(xl)- Di'f?k(x2) J = 0 and

k~l
we
see
sup [Di'f?k(x) [<co.

xEK,k~l
from
the
fact
that,
for
any
coordinate
X5 ,
xE~~~l Ie~. Di'f?k (x)[
<co. Hence, by the Ascoli-Arzela theorem, there

exists a subsequence of {Di 'f?k' (x)} which converges uniformly on K.
By the diagonal method, we may choose a subsequence {'f?k" (x)} of {'f?k (x)}
such that, for any differential operator Di, the sequence {Di 'f?k" (x)}
converges uniformly on K. Thus
lim Di 'f?k" (x)
k"-+-00
= Di 'f? (x)
where <p (x) = lim 'f?k" (x),

k"--K:JO
and these limit relations hold uniformly on K. Hence the metric space
<f) K (Rn) is complete and so it is not of the first category.
Remark. (i) The above proof shows that a bounded set of 'l) (Rn) is
relatively compact in the topology of 'l) (Rn). For, a bounded set B of
<f) (R .. ) is contained in some 'l)K (Rn) where K is a compact set of R~, and,
moreover, the boundedness condition of B implies the equi-boundedness
and equi-continuity of {Di'f?; 'f? E B} for every Di. (ii) Similarly, we see
that any bounded set of (;l; (Rn) is a relatively compact set of (;l; (Rn).
139
3. Semi-reflexivity and Reflexivity
Corollary 3. 'l:l (Rn) is a barrel space.

Proof. 'l:l (R,.) being an inductive limit of {ilK (Rn)} when K ranges
over compact subsets of Rn, the present corollary is a consequence of
the following
Proposition. Let a locally convex linear topological space X be an
inductive limit of its barrel subspaces X.,, iX EA. Then X itself is a barrel
space.
Proof. Let V be a barrel of X. By the continuity of the identical
mapping T"': x--+ x of X"' into X, the inverse image T;; 1 (V) = V (\X"'
is closed with V. Thus V (\ XC< is a barrel of XC<. X"' being a barrel space,
V (\X"' is a neighbourhood of 0 of XC<. X being an inductive limit ofXtx's,
V must be a neighbourhood of 0 of X.
Theorem 2. Let X be a barrel space. Then the mapping x --+ ] x of X
into (X;);, defined in Chapter IV, 8, is a topological mapping of X onto
J X, where the topology of J X is provided with the relative topology
of] X as a subset of (X;);.
Then the polar set (B') 0 =
Proof. Let B' be a bounded set of
of 0 of (X;);,
neighbourhood
{x" E (X;)'; sup I<x', x") I < 1} of B' is a
x;.
11'EB'
and it is a convex, balanced and absorbing set closed in (X;);. Thus

(B') 0 (\ X = 0 (B') is a convex, balanced and absorbing set of X. As a
(left) polar set, 0(B') is closed in the weak topology of X, and hence
0 (B') is closed in the original topology of X. Thus 0 (B') = (B')O (\ X is
a barrel of X, and so it is a neighbourhood of 0 of X. Therefore, the
mapping x-+ ] x of X into (X;); is continuous, because the topology of
(X;); is defined by a fundamental system of neighbourhoods of 0 of the
form (B') 0 , where B' ranges over bounded sets of X'.
Let, conversely, U be a convex, balanced and closed neighbourhood of
0 of X. Then, by the preceding section, U =o(Uo). Thus]U =]X(\ (UO)o.
On the other hand, U0 is a bounded set of x;, since, for any bounded set
B of X, there exists an IX> 0 such that iXB ~ U and so (1XB) 0 ~ U0
Hence (U0 ) 0 is a neighbourhood of 0 of (X;);. Thus the image J U of the
neighbourhood U of 0 of X is a neighbourhood of 0 of ] X provided with
the relative topology of ] X as a subset of (X;);.
3. Semi-reflexivity and Reflexivity

Definition 1. A locally convex linear topological space X is called
is given by
semi-reflexive if every continuous linear functional on
x;
<x, x'), with a certain x EX.
(1)
Thus X is semi-reflexive iff
(2)
140
Definition 2. A locally convex linear topological space X is called

reflexive if
X= (X;);.
(3)
By Theorem 2 in the preceding section, we have

Proposition 1. A semi-reflexive space X is reflexive if X is a barrel space.
It is also clear, from Definition 2, that we have
Proposition 2. The strong dual of a reflexive space is reflexive.
Theorem 1. A locally convex linear topological space X is semireflexive iff every closed, convex, balanced and bounded set of X is
compact in the weak topology of X.
Proof. Let X be semi-reflexive, and T a closed, convex, balanced and
bounded set of X. Then, by Theorem 2 in Section 1 of this Appendix,
T = 0 ( TO). T being abounded set of X, T 0 is a neighbourhood of 0 of x;.
Thus, by Theorem 1 in Section 1 of this Appendix, (T0) 0 is compact in the
weak* topology of (X;)'. Hence, by the semi-reflexivity of X, T = 0 (TO)
is compact in the weak topology of X.
We next prove the sufficiency part of Theorem 1. Take any x" E (X;)'.
The stcong continuity of x" on x; implies that there exists a bounded
set B of X such that
I(x', x") I < 1 whenever x' E B 0 , that is, x" E (B0)0.
We may assume that B is a convex, balanced and closed set of X. Thus,

by the hypothesis of Theorem 1, B is a compact set in the weak topology
of X. Hence B = Bwa where Bwa denotes the closure of B in the weak
topology of X. Since Xw is embedded in (X;)~. as a linear topological
subspace, we must have (B 0) 0 ~ Bwa =B. Therefore we have to show
that x" is an accumulation point of Bin (X;)~. Consider the mapping
x _,. q; (x) = {(x, x~) .. ., (x,x~)} of X into / 2 (n), where x~; ... . , x~E X'.
The image q;(B) is convex, balanced and compact, since B is convex,
balanced and weakly compact. If {(xi. x"), ... , (x~. x")} does not belong
to q;(B), then, by Mazur's theorem, there would exists a point
{c 1 , . , en} E/2 (n) such that sup
ci (b, xi)/ ~ 1 and (I: ci (x~. x")) > 1,
proving that
~B
/.Ei
I: cixi E B 0 and x" cannot belong to B 00
Theorem 2. A locally convex linear topological space X is reflexive

iff it is a barrel s.pace and every closed, convex, balanced and bounded set
of X is compact in the weak topology of X. In particular, ~ (Rn) and
Cf (Rn) are reflexive.
Proof. The sufficiency part is proved already. We shall prove that
the first condition of Theorem 2 is necessary.
141
4. The Eberlein-Shmulyan Theorem
LetT be a barrel of X. We shall prove that T absorbs any bounded

set B of X so that B 0 ~<X ro, <X~. B 0 being a neighbourhood of 0 of
x;, we see that TO is a bounded set of x;. By the Proposition and Theorem 2 in Section 1 of this Appendix, we have T = 0 (TO). By the hypothesis
that X is reflexive, we have 0(T 0) = (T0) 0 and so T = (T0) 0 Therefore we
have proved that the barrel T is a neighbourhood of 0 of X= (X;);.
Hence X is a barrel space.
By hypothesis, the closed, convex and balanced set =Conv ( U .x a
1<>1::;>1
B)
is compact in the weak topology of X. Here we denote by Conv(N)a the

""'
closure in X ofthe convex closure (see p.28) Conv(N) ofN. Set Y = n~ 1nK
and let p (x) be the Minkowski functional of K. Then since K is w-compact in
Y, p (x) defines a norm of Y. That is, the system {!XK} with IX> 0 defines a
fundamental system of neighbourhoods of the normed linear space Y, and
Y is a B-space since K is w-compact. Hence Y is a barrel space. On the
other hand, since K is a bounded set of X, the topology of Y defined by the
norm p (x) is stronger than the relative topology of Y as a subset of X.
As a barrel of X, Tis closed in X. Hence T (\ Y is closed in Y with respect
to the topology defined by the norm p (x). Therefore T (\ Y is a barrel
of the B-space Y, and so T (\ Y is a neighbourhood of 0 of the B-space
Y. We have thus proved that T (\ Y and, a fortiori, T both absorb
K~B.

This theorem is very important in view of its applications.
Theorem (EBERLEIN-SHMULYAN). A B-space X is reflexive iff it is
locally sequentially compact; that is, X is reflexive iff every strongly
bounded sequence of X contains a subsequence which converges weakly
to an element of X.
For the proof we need two Lemmas:
Lemma 1. If the strong dual x; of a B-space X is separable, then X
itself is separable.
Lemma 2 (S. BANACH). A linear subspace M' of the dual space X' of a
B-space X is weakly* closed iff M' is boundedly weakly* closed; that is,
M' is weakly* closed iff M' contains all weak* accumulation points of
every strongly bounded subset of M'.
Lemma 1 is already proved as a Lemma in Chapter V, 2. For Lemma 2,
we have only to prove its "if" part. It reads as follows.
Proof. (E. HILLE-R. S. PHILLIPS [1]). We remark that M' is strongly
closed by the hypothesis. Let x~ EM'. Then we can prove that, for each
142
constant C satisfying the condition 0 < C <
inf llx' -x~ il, there exists
x'EM'
an x0 E X with llxo!l < 1/C such that
= 1 and (x0 , x') = 0 for all
(x0 , x~)
x' E M'.
(1)
Thus the strongly closed set M' must contain all of its weak* accumulation points.
To prove the existence of x 0 , we choose an increasing sequence of
numbers {Cn} such that C1 = C and lim Cn = = Then there exists a
n->00
finite subset a 1 of the unit sphere S = {x E X; II x II < 1} such that

llx'-x~II::;;:C 2
and
supl(x,x')-(x,x~)I<C 1
;tEa1
implies
x'E.M'.
If not, there wonld exists, corresponding to each finite subset a of S, an
x: E M
such that
llx:-x~ll <
c2
and
sup l<x,X:>-<x,x~>l <

xEa
cl.
We order the sets a by inclusion relation and denote the weak* closure of
the set {x:; a' > a} by
It is clear that
enjoys the finite intersection
property. On the other hand, since M' is boundedly weakly* closed,
the Corollary of Theorem 1 in this Appendix, 1, implies that the set
N:.
N:
M;, ={x'EM'; llx'll < C'}

is weakly* compact. Hence N: ~ M;, for C' = C2 + llx~ jj, and so there
exists an x~ E n N: ~ M'. Hence we have sup I(x, x;)- (x, x~) I < C1
a
~s
and so II x~ - x~ II < C1 , contrary to the hypothethis that 0 < C1 <

inf II x' - x~ II
x'EM'
By a similar argument, we successively prove the existence of a sequence of finite subsets a 1 , a 2 , . . of S such that
llx'-x~ll < Ck
and sup
sEal
<x, x'>- <x, x~> 1 < ci

(i = 1, 2, ... ' k- 1)
do not imply x' E M' .
Thus, since lim Ci

i-+OO
=ex:>, we see that x' EM' if
I(x, x')- (x, x~) I <
C for all
x E (C /Cj) ai
(j = 1, 2, ... ) .
Let {xn} be a sequence which successively exhausts the sets (C/Cj)aj

(j = 1, 2, ... ). Then lim Xn = 0 and soL (x') = {(xn, x')} is a bounded
n->00
x;
linear transformation of
into the B-space (c0). We know that the
point {(xn, x~)} E (c0 ) lies at a distance > C from the linear subspace
L (M'). Thus, by Corollary of Theorem 3 in Capter IV, 6, there exists a
143
continuous linear functional {ex,.} E (c0 )'
jj{cx,} II= I
00
n=l
00
I ex,. (x,, x')
n=l
The element x0
= I
00
n=l
(ll) such that
00
jcx,. I < 1/C, I cx,.(x,., x~)

n=l
=1
and
x' E M'.
0 for all
.x,x, clearly satisfies condition (1).
Corollary. Let (x', x~) = F(x') be a linear functional defined on

the dual space X' of a E-space X. If N (F)= N (x~) = {x' EX'; F (x') = 0}
is weakly* closed, then there exists an element x0 such that
F (x')
(x', x~)
(x0 , x')
for all
x' EX'.
(2)
Proof. We may assume that N(F) #-X'. Otherwise we can take

x 0 = 0. Let x~ E X' be such that F (x~) = 1. By (1) of the preceding
Lemma 2, there exists an x0 E X such that
(x0 ,
xci> =
and
(x0 , x')
0 for all
x' E N (F).
(3)
Hence, for any x' EX', the functional
x' -F(x')x~
= y' EX'
satisfies F(y') = 0, i.e. y' E N(F). Therefore, by (3), we obtain (2).

Proof of the Thorem. "Only if" part. Let {x,} be a sequence of X
such that II x,. II = 1. The strong closure X 0 of the subspace spanned by
{x,.} is a separable E-space. Being a E-space, X 0 is a barrel space. We
shall show that X 0 is reflexive. Any strongly closed, bounded set E 0 of
X 0 is also a strongly closed, bounded set of X and hence E 0 is compact
in the weak topology of X by the reflexivity of X. But, as a strongly
closed linear subspace of X, X 0 is closed in the weak topology of X (see
Theorem 3 in Chapter IV, 6). HenceE0 is compact in the weak topology
of X 0 Thus X 0 is reflexive by Theorem 2 in the preceding section. We
have thus X 0 = ((X0 ););. By Lemma 1 above, (X0 ): is thus separable. Let
{x:} be strongly dense in (X0 );. Then the weak topology of X 0 is defined
by an enumerable sequence of semi-norms Pm (x) = I(x, x~) I (m =
1, 2, ... ). Hence it is easy to see that the sequence {x,.}, which is compact
in the weak topology of X 0 , is sequentially weakly compact in X 0 and
in X as well. We have only to choose a subsequence {x,..} of {x,.} such
that finite lim (x,.., x~) exists for m = 1, 2, ...
n->-oo
"lj" part. Let M be a bounded set of X, and assume that every infinite sequence of M contains a subsequence which converges weakly to
an element of X. We have to show that the closure M of Min X in the
weak topology of X is weakly compact in X. For, then the barrel space
X is reflexive by Theorem 2 in the preceding section.
144
Since Xw ~(X;)~., we have M = M f\ Xw, where M is the closure

of Min the weak* topology of (X:)'. LetS~ be the sphere of
of radius
r > 0 and centre 0. By the correspondence
x:
M 3m+->- {<x', m); llx' II< 1} E
Ix' = {z;
lz\ <
JJ/,_, where
sup J<x', m>l},

mEM
M may be identified with a closed subset of the topological product
II I"'. By Tychonov's theorem, II Ix' is compact and hence M is
x~E:Sl
x'ES~
compact in the weak* topology of (X~)'. Hence we have only to show

M~Xw.
Let x~ E (X~)' be an accumulation point of the set M in the weak*

topology of (X~)'. To prove that x~ E Xw we have only tc show that the
set N(x~) = {x' EX'; <x', x~) = 0} is weakly* closed. For, then, by the
above Corollary, there exists an x 0 E X such that <x', x~) = <x0 , x') for
all x' EX'. We shall first show that
for every finite set x~, x;, ... , x~ of X', there exists a z E M
such that
<x;, x~) = <z, xj) (f =
The proof is as follows. Since

element Zm EM such that
x~
1, 2, ... , n).
(4)
is in the weak* closure of M, there is an
I<zm, xj) -<xi, x~> I <
lfm (f
= 1, 2, ... , n).
By hypothesis there exists a subsequence of {zm} which converges weakly

to an element z E X and so z EM, since the sequential weak closure of M
is contained inM. We have thus (4).
Now, by Lemma 2, N (x~) is weakly* closed if, for every r > 0, the
set N (x~) f\ s; is weakly* closed. Let y~ be in the weak* closure of
N(x~) f\ S~. We have to show that y~ E N(x~) f\ S~. To this purpose, we
choose an arbitrary s > 0 and construct three sequences {z,.} ~ M,
{x,.} ~ M and {y~} ~ N(x~) f\ S~ as follows: By (4) we can choose
a, z1 EM such that <z1 , y~) = <y~, x~). z1 being in the weak closure
of M, there exists an x1 EM such that I<x1 , y~)- <z1 , y~) I < sj4. y~ being
in the weak* closure of N (x~) f\ s;_, there exists a y;_ EN (x~) f\ s;_ such
that I<xv Yi>- <x1 , y~) \ < sf4. Repeating the argument and remember-
145
ing (4), we obtain {zn} ~ M, {xn} ~ M and {y~} ~ N(x~) (\ S~ such that
(z1 , y~)
=
=
(y~, x~),
= 0
I(xn, y;..)- (zn, y;..) / :'S e/4
I(xi, y~)- (xi, y~) I < e/4
(zn, y;,.)
(y;.., x~)
(m = 1, 2, ... , n- 1),
(5)
2, ... , n).
(6)
(m = 0, 1, ... , n- 1),
(i
1, 2, ... , n).
Thus we have
I(y~, x~)- (xi, y~) I <
e/4
+ e/4 =
e/2 (i
1,
Since {xn} ~ M, there exists a subsequence of {xn} which converges weakly

to an element x EM. Without losing the generality, we may assume that
the sequence {xn} itself converges weakly to x EM. Hence, from (5),
I(x,y;..) I< ef4. Fromw-lim xn = xandMazur's theorem 2inChapterV, 1,
~
.i:
there exists a convex combination u = .

such that
llx- ull <
J~l
()(jXj
((Xi >
.i:
0, .
t~l
()(j
= 1)
ef4. Therefore, by (6),
and hence
I<y~, x~> I < I<y~, x~>- <u, Y~> I + I<u, Y~>- <x, Y~> I + I<x, Y~> I
<:;; e/2 + llu-xiiiiY~II + e/4 <e.
As e was arbitrary, we see that (y~, x~) = 0 and soy~ EN (x~). Combined
with the fact that S~ is weakly* closed, we finally obtain y~ EN (x~) (\ S~.
Remark. As for the weak topologies and duality in B-spaces, there
is an extensive literature. See, e.g., the references in N. DUNFORD]. SCHWARTZ [1]. Sections 1, 2 and 3 of this Appendix are adapted
and modified from N. BouRBAKI [1] and A. GROTHENDIECK [1]. It is
remarkable that necessary tools for proving this far reaching theorem of
EBERLEIN [1]-SHMULYAN [1] are found, in one form or other, in the book
of S. BANACH [1].

The Fourier transform is one of the most powerful tools in classical
and modern analysis. Its scope has recently been strikingly extended
thanks to the introduction of the notion of generalized functions of
S. L. SoBOLEV [1] and L. SCHWARTZ [1]. The extension has been applied
successfully to the theory of linear partial differential equations by
L. EHRENPREIS, B. MALGRANGE and especially by L. HoRMANDER [6].
146
1. The Fourier Transform of Rapidly Decreasing Functions

Definition 1. We denote by 6 (Rn) the totality of functions f E C"" (Rn)
such that
sup
.ER"
lxll D"'f(x) I< oo (xll = ij xff;)
(1)
J-l
for every a = (av a 2 , . . . , 1Xn) and fJ = (fJv {3 2 , , fJn) with non-negative

integers iXj and fJk Such functions are called rapidly decreasing (at oo).
2 ) and functions f E C0 (Rn) are rapidly decreasExample. exp (-1 x 1
ing.
Proposition 1. 6 (Rn) is a locally convex linear topological space by
the algebraic operation of function sum and multiplication of functions
by complex numbers, and by the topology defined by the system of
semi-norms of the form
p (f) =sup IP (x) D"' f (x)
1. where P (x)
denotes a polynomial.
(2)
xER"
Proposition 2. 6 (Rn) is closed with respect to the application of

linear partial differential operators with polynomial coefficients.
Proposition 3. With respect to the topology of 6 (Rn), C0 (R") is a
dense subset of 6 (R").
Proof. Let f E 6 (R") and take 'IJl E ego (R") such that 'IJl (x) = 1 when
lxl < 1. Then, for any e > 0, /,(x) = f(x) 'IJl(ex) E Cg"(Rn). By applying
Leibniz' rule of differentiation of the product of functions, we see that
D"'(f.(x)- f(x)) = D"'{f(x) ('IJl(ex) -1)}
is a finite linear combination of terms of the form

Dilj(x)(e)IYI{D"'IJl(Y)}y=w where
lfJI+ lrl=lal
with
lri>O,
and the term D"'f(x) ('IJl(ex) -1). Thus it is easy to see that f.(x) tends
to f (x) in the topology of 6 (Rn) when e -). 0.
Definition 2. For any f E 6 (Rn), define its Fourier transform f by
/(~)
= (2n)-n/ 2
Je-i(o,:r) f(x) dx,
(3)
R"
where ~
dx
(~v ~2 , ~.. ), x
dx 1 dx2
..
= (xv x2 , . , xn),
x) = .-I
J=l
~ixi
and
dxn. We also define the inverse Fourier transform
g of
(~.
gE 6(R") by
(4)
Proposition 4. The Fourier transform: f-+ maps 6 (R") linearly

and continuously into 6 (R"). The inverse Fourier transform: g-+ g also
maps 6 (R") linearly and continuously into 6 (R").
1. The Fourier Transform of Rapidly Decreasing Functions
147
Proof. Differentiating formally under the integral sign, we have

D"'f(~)
= (2n)-n/Z J e-i{<,x)
(-i)l"'l x"'l(x)dx.
(5)
The formal differentiation is permitted since the right hand side is, by
(1), uniformly convergent in r Thus f E C00 (Rn). Similarly, we have, by
integration by parts,
(6)
Thus we have
(i)llll+ia\ ~~~ D"} (~) = (2n)-n/Z
J e-i(t;,x) Dil(x"'l(x))dx,
(7)
and (7) proves that the mapping I--+ J is continuous in the topology of
6(Rn).
Theorem 1 (Fourier's integral theorem). Fourier's inversion theorem
holds:
(8)
(x) = (2n)-n/Z ei(x.O j(~) d; = l(x), i.e., we have
f = I, and similarly f = I.
(8 1 )
Therefore it is easy to see that the Fourier transform maps 6 (Rn) onto
6 (Rn) linearly and continuously in both directions, and the inverse Fourier
transform gives the inverse mapping of the Fourier transform.
Proof. We have
Jg (;) i(;) e*'0
d; =
J g(y) I (x + y) dy (I and g E 6 (Rn)).
(9)
In fact, the left hand side is equal to
I (y) dy} e i(x,nd;

Jg (;) {(2n) -n/2J
d;} I (y) dy
= (2n)-n/2 J{j g (~)
= J g (y - x) I (y) dy = J g(y) I (x + y) dy.
e-i(t;,y)
e-i(l;,y-x)
If we take g(s;) for g(;), s
(2n)-nl 2
e-i(y,t;)
Hence, by (9),
>
0, then
g(s;)d; = (2n)-n/ 2
en J g(z)e-i(y,zfe) dz =e-ng (yjs).
f g (s;) 7(~) ei<x,t;> a;= f g(y) 1(x + sy) dy.
We shall take, following F. Rmsz, g(x) =e-lxi'!Z and lets-). 0. Then
J f (;) ei(x,n d~ = l(x) Jg(y) dy.

(8), since g(O) = 1 and Jg(y) dy = (2nt12
g(O)
This proves
known facts:
Je-lx\'/2 e-i(y,x) dx = e-IY\'/2'

(2n)-n/2 J e-lxl'/2 dx = 1.
(2n)-n/2
10*
by the well-
(10)
(10
148
Remark. For the sake of completeness, we will give the proof of (10).
We have
A
(2:n;)-1/2 J e-1'/2
e-iul
dt
e-u'/2 ( 2 :n;)-1/2
-A
e-(l+iu)'/2
dt.
-A
Let u > 0, and integrate the function e-z'/2 , which is holomorphic in z

+ iu, along the curve consisting of the oriented segments
=T.1,
A, A + iu, A + iu, -A
+ iu
and -A
+ iu,- A
in this order. By Cauchy's integral theorem, the integral vanishes. Thus

(2:n:)-1/2 J
dt = (2:n;)-1/2 J e-t'/2 dt
e-(t+iu)'/2
-A
-A
+ (2:n;)-l/2
+ (2:n:)-l/2 J"
e-(-A+iu)'/2
e-(A+iu)'/2
idu
idu.
The second and third terms on the right tend to 0 as A_. oo, and so, by
(10'),
00
(2:n;)-1/2 J e-1'/2
e-ilu
dt =
e-u'/2 ( 2 :n:)-112
-oo
004~d
-~
e
t= e
-oo
We have thus proved (10) for the case n = 1, and it is easy to prove the
case of a general n by reducing it to the case n = 1.
Corollary (Parseval's relation). We have
J i(~) g(~) d~ = J l(x) g(x) dx,
(11)
/I(~) g(~) d~ = JJ(x) g(x) dx,
(f *g)= (2:n;)n12 I
/'-..
I/'-..
(12)
~
and (2:n:t 2 (/-g)= l*g,
* g is defined through
JI (x- y) g(y) dy = J g(x- y) I (y) dy.
(13)
where the convolution I
(/ *g) (x)
(14)
Proof. (11) is obtained from (9) by putting x = 0. (12) is obtained

from (11) by observing that the Fourier transform of g is equal tog. We
next show that
(2:n:)-n/2 (f *g) (x) e-i(;,,.) dx
= (2:n:)-n/2
Jg(y) e-<e,y) {f I (x- y) e-i(e,,.-y) dx} dy
(lo)
:n:t12 i (~) g(~) .

product f. g of functions 1 and gE @5 (Rn) is again a function
= (2
Since the
of @5 (Rn), we know that the right hand side of (lo) belongs to @5 (~).
2. The Fourier Transform of Tempered Distributions
149
It is easy to see that, the convolution I* g oftwo functions of 6 (R11 ) belongs

also to 6 (R11). Thus we have proved the first formula of (13). The second
formula may be proved similarly to (9) by (15).
Theorem 2 (Poisson's Summation Formula). Let q; E 6 (R1) and
rfo E 6 (R1) its Fourier transform. Then we have
00
.I q; (2nn) =
11=-00
00
.I q;(x
Proof. Set l(x) =
n=-oo
00
.I rfo (n).
(16)
11=-00
+ 2nn). This series is absolutely conver-
gent, E C00 and I (x + 2 n) =I (x), as may be proved by the fact thattp (x)
is rapidly decreasing at oo. In particular, both sides of (Hi) are convergent.
We have to prove the equality.
The Fourier coefficients ck of l(x) with respect to the complete orthonormal system {(2n)-1 ' 2 e-ikx; k = 0, 1, 2, ... }ofL2(0, 2n) are given
by
2n
00
q;(x
1' 2
00
= .I (2n)- 112
2n(11+1)
n=-oo
Thus, by
2n
J l(x) e-ikx dx = n=-00

.I (2n)- I
ck = (2n)- 1' 2
2nn
q; (x) e-ikx dx
+ 2nn) e-ikx dx
= ;p (k).
IE L2(0, 2n), we have

00
f (x) = 2: q; (x + 2nn) = l.i.m. .I ;p (k) eikx.

s too k=-s
n=-oo
However, since
00
qi (x) E 6 (Rn), the series .I ;p (k) eikx converges absoluk=-oo
tely. Hence
00
00
.I q; (x + 2nn) = .I
n=-oo
k=-00
and so we obtain (16) by setting x

Example. We have, by (10),
(2n)-l/2
;p
(k) eikx,
0.
J e-tx e-'*" dx = (2n)-1/2 J e-x'/2 e-i%)!/V2t (2t)-l/l! dx

00
00
-oo
-oo
= (2t)- 1' 2 e-"''41, t >
0.
Hence, by (16}, we obtain the so-called 0-lormula:

00
.I e-4tn'n' =
n=-00
.I (2t)-l/2 e-11 /41, t > 0.

00
n=-00
(17)
Definition 1. A linear functional T defined and continuous on 6 (Rn)

is called a tempered distribution (in Rn). The totality of tempered distri-
150
butions is denoted by ~ (R")'. As a dual space of ~ (R"), ~ (R")' is a

locally convex linear topological space by the strong dual topology.
Proposition 1. Since COO (R") is contained in ~ (R") as an abstract set,
and since the topology in ;t) (R") is stronger than the topology in ~ (R"),
the restriction of a tempered distribution to COO (R") is a distribution in
R". Two different tempered distributions define, when restricted to
ego (R"), two different distributions in R", because ego (R") is dense in
~ (R") with respect to the topology of ~ (R"), and hence a distribution
E ~ (R")' which vanishes on ego (R") must vanish on ~ (R"). Therefore
(1)
Example 1. A distribution in R" with a compact support surely belongs to ~ (R")'. Therefore
(2)
Example 2. A a-finite, non-negative measure p, (dx) which is a-additive
on Baire sets of R" is called a slowly increasing measure, if, for some nonnegative k,
(3)
(1 + lxl 2)-k p,(dx) < oo.
R"
Such a measure p, defines a tempered distribution by

T"(rp) = frp(x)p,(dx), rpE ~(R").
(4)
R"
For, by the condition rpE ~(R"), we have rp(x) = 0((1 + lxl 2)-k) for
large lxl.
Example 3. As a special case of Example 2, any function IE LP (R"),
p ?': 1, defines a tempered distribution
T1 (rp)
Jrp (x) I (x) dx,
rp E ~ (R").
(4')
R"
That an 1E LP (R") gives rise to a slowly increasing measure p, (dx) =

\f(x) I dx may be proved by applying Holder's inequality to
J (1 + lxl2)-k \l(x)
dx.
Definition 2. A function f E C00 (R") is called slowly increasing (at

oo), if, for any differentiation Di, there exists a non-negative integer N
such that
(5)
lim Jx\-N IDi l(x) I= 0.
jxj--..oo
The totality of slowly increasing functions will be denoted by () M (R").

It is a locally convex linear topological space by the algebraic operations
of function sum and multiplication of functions by complex numbers,
151
and by the topology defined by the system of semi-norms of the form

{6)
P(f) = p,.,vs(/) =sup lh(x) Di f(x) 1. IE OM(R"),
:eER"
where h (x) is an arbitrary function E 6 (R") and Dian arbitrary differentiation. As may be seen by applying Leibniz' formula of differentiation
of the product of functions, h (x) Di f (x) E 6 (R") and so Ph,Dj (f) is finit~
for every f E () M (R"). Moreover, if Pn,Dj (!) = 0 for all h E 6 (R") and D 3 ,
0 as may be seen by taking D =I and hE i)(R").
then f(x)
Proposition 2. ego (R") is dense in () M (R") with respect to the topology
of ()M(R").
Proof. Let f E () M (R"), and take tp E ego (R") such that tp (x) = 1 for
IX I < 1. Then Is (x) = I (x) tp (ex) E ego (R") for any e > 0. As in Proposition 3 in Chapter VI, 1, we easily prove that fs (x) tends to f (x) in the
topology of () M (R") when e t 0.
Proposition 3. Any function IE () M (R") defines a tempered distribution
(7)
TJ(rp) = f I (x) rp (x) dx, rp E 6 (R").
]lft
Definition 3. As in the case of a distribution in R", we can define the

generalized derivative of a tempered distribution T by
(8)
DiT(rp) = (-1)1iiT(Dirp), rp E 6(R"),
since the mapping rp (x) ---+ Di rp (x) of 6 (R") into 6 (R") is linear and
continuous in the topology of 6 (R"). We can also define a multiplication
by a function f E () M (R") to a distribution T E 6 (R")' through
{9)
(IT) (rp) = T(frp), rpE 6(R)",
since the mapping rp (x) ---+ f (x) rp (x) of 6 (R") into 6 (R") is linear and
continuous in the topology of 6 (R").
The Fourier Transform of Tempered Distributions
Definition 4. Since the mapping rp (x) ---+ (j; (x) of 6 (R") onto 6 (R")
is linear and continuous in the topology of 6 (R"), we can define the
Fourier transform T of a tempered distribution T as the tempered
distribution T defined through

T(rp) = T((j;), rp E 6(R").
Example 1. If f E L 1 (R"), then
T1 =
TJ, where i(x) = (2n)-"'2
J e-i(x,~) f(~) d~,

R"
as may be seen by changing the order of integration m

f(x) cp(x) dx = (2n)-"' 2 f(x) { e-i(x,O rp{~) d~} dx.
R"
R"
(10)
(11)
T1 (rp)
152
Remark. In the above sense, the Fourier transform of a tempered

distribution is a generalization of the ordinary Fourier transform of
functions.
Proposition 4. If we define
f(x) = /(-x),
(12)
then Fourier's integral theorem in the preceding Section 1 is expressed by
i = 1,
IE6
(13)
(R") .
Corollary 1 (Fourier's integral theorem). Fourier's integral theorem

is generalized to tempered distributions as follows:
..,
..,
..,
T = T, where T(cp) = T(cp).
(14)
In particular, the Fourier transform T
6 (R")'.
T maps 6 (R")' linearly onto
Proof. We have, by definition,

~
f(cp)
= T(~) =
T((p)
= T(cp) for all cpE 6(R").
Corollary 2. The Fourier transform T ~ f and its inverse are linear

and continuous on 6 (R") onto 6 (R") with respect to the weak* topology
of 6(R")':
lim T" (cp) = T (cp) for all cp E 6 (R") implies that
lim f" (cp)
= f
(cp) for all cp E 6 (R").
(15)
Here the inverse of the mapping T ~Tis defined by the inverse Fourier
transform T ~ T given by
T(cp)= T((j}), cpE 6(R").
(10')
Example 2.
fd
Proof. Td(cp) = Td(~)
(2nt"12 T 1 (cp), and
T~ =
T1 =
(2nt"12 T 1 ,
= ~(0) =
(2n)'112 Td.
(2nt"12
T8 = T8 =
(16)
J 1 cp(y) dy =
R"
(2n)-"12
T1 .
Example 3.
/~
(8Tf8xj)
/'-...
ixiT,
~
(ixiT) = - (8Tf8xj).
(17)
(18)
153
Proof. We have, by (5) in Chapter VI, 1,

./'-....
./'-....
(8Tf8x1) (tp) = (8Tf8x1) (tP) = - T(arpf8x1) = - T(-ix;tp (x))

_.............
= T(ix1q;) = (ix;T) (tp).
We also have, by (6) in Chapter VI, 1,

"'
....
...................
,................
(ix1T) (tp) = (ix1T) (tp) = T(ix;tp) = T(8tpf8x1) = T(8tpf8x1)

=-(8Tf8xi) (tp).
Plancherel's Theorem. If IE L 2 (R"), then the Fourier transform
of T1 is defined by a function f E L 2 (R"), i.e.,
T1 =
and
JJ/JJ
TJ with [E L 2 (R"),
(19)
(J Jf(x)J dxY' = (fn Jl(x)J dxr = IIIII

2
f1
(2o)
Proof. We have, by Schwarz' inequality,
!Tt(q;)J = !Tt(q)}J =
lfn l(x) <P(x) dxl < lllllllo/11
IIIIIII~PII (21)
The equality above JI<PII = !liP I! is proved in (12) of the preceding section
1. Hence, by F. Riesz' representation theorem in the Hilbert space L 2 (R"'),
there exists a uniquely determined {E L 2 (R"') such that
T1 (q;) =
Jtp(x) {(x) dx =
Tj (tp), that is,
R"'
J /(x) tp (x) dx = Jf (x) rP (x) dx
for all q; E 6 (R"') .
(22)
Moreover, we have, from (21), IIIII ~ IIi II since 6 (R") is dense in L 2 (R"')
in the topology of L 2 (R"). We have thus IIi II ::;; IIi II::;; IIIII On the other
hand, we have, by (13) and (22),
Jf(x}tp(x)dx= Jl(x)rP(x)dx= Jf(-x)tp(x)dx forall

R"
Rn
R"'
q;E6(R"),
that is,
i(x) = f(-x) = i(x} a.e.

Thus
(20).
IIlii = IIIII and so, combined with IIlii < llfll < IIIII. we
(23)
obtain
Definition 5. The above obtained f (x) E L 2 (R"') is called the Fourier

transform of the function f(x) E L 2 (R").
154
Corollary 1. We have, for any /E L 2 (R"),
j (x) = l.i.m. (2 :n) -n/2

htoo
jxj;;;;h
e-i(x,y) f (y) dy.
(24)
Proof. Put
fh(x) = f(x) or= 0 according as jxj
< h or jxj >h.
Then lim llfh- !II= 0 and so, by (20). lim llh
f (x)
h-+00
= l.i.m.
~00
h-+OO
fh (x) a. e. But, by (22),
-111 =
0, that is,
Jh. (x) rp (x) dx = Jfh (x) ~ (x) dx
R"
R"
J f(x) { (2:n)-"12 J e-i(x,y)rp(y) dy}dx,
jxj;;;;h
R"
which is, by changing the order of integration, equal to
J (2:n)-nf
R"
2{
J e-i(x,y) f (x) dx} rp (y) dy,
jxj;;;;h
since fh (x) is integrable over Ix J ,~ h as may be seen by Schwarz' inequality. Thus f,. (x) = (2:n)-"' 2
e-i(x,y) f(y) dy a.e., and so we obtain
(24).
jxj;;;;h
Corollary 2. The Fourier transform

in a one-one-manner such that
f--+ f maps L 2 (R") onto L 2 (R")

(25)
Proof. Like the Fourier transform

form f--+ defined by
{(x) = l.i.m. (2:n)-"' 2

h..,.oo
f--+ j, the inverse Fourier trans-
J ei(x,y) f(y) dy
jyj;;;;h
(26)
mapsL2 (Rn) into L 2 (Rn) in such a way that llfll = IIIII Hence we see
that the Fourier transform f--+ J maps L 2 (R") onto L 2 (R") in a one-one
way such that 1/fll = IIIII Hence, by the linearity of the Fourier transform and
(x,y) = 4-1 (llx
+ Yll2-/lx-yll2) + 4-li(llx + iy/12- llx-iy/12).
we obtain (25).
Parseval's Theorem for the Fourier Transform. Let / 1 (x) and / 2 (x)
both belong to L 2 (R"), and let their Fourier transforms be f~(u) and / 2 (u),
respectively. Then
J{ (u)/ (u) du = J / (x) /

1
2 (-x)
dx,
(27)
155
and so
J f1 (u) f2 (u)
e-i(u,x)
Jft (y) /2 (x- y) dy.
du =
(28)
Thus, if j1 (u)tz(u) as well as both its factors belong to L 2 (Rn), it is the

Fourier transform of
(2n)-n/2
J /1 (y) / 2 (x- y) dy.
(29)
This will also be true if / 1 (x), / 2 (x) and (29) all belong to L 2 (~).
Proof. It is easy to see that
(2n)-n12 /2(-x)
e-i(u,x)
dx = f2(u)~
and so, by (25), we obtain (27). Next, since the Fourier transform of the
function of y, f2 (x-Yf, is h(u)ei(u,x>containing a parameter x, we
obtain (28) by (25). The rest of the Theorem is clear from (28) and j =f.
The Negative Norm. The Sobolev space Wk 2 (.Q) was defined in
Chapter I, 9. Let /(x) E Wk 2 (Rn). Since /(x) E L 2 (Rn), f gives rise to a
slowly increasing measure lf(x) I dx in R". Thus we can define the Fourier
transform T1 of the tempered distribution T1. We have, by (17),
/"---
II
D"'T1 = (i)"'
T1 .
JIn x~s
J
A
J=l
By the definition of the space Wk 2 (R")' D"'

Thus, by Plancherel' s theorem for L 2 (R"),
/"---
IID"'Ttllo = IID"'Ttllo. where
Iillo
r, E L2 (Rn)
for
IX I <
k.
is the L2 (R..)-norm.
Hence we see that (1 + J x j2 )kf2 T1 E L 2 (R"), and so it is easy to show that

ID"' Tl dx)112 is equivalent to the norm
the norm ll!llk = ( ~
l"'~kRn
+ lxl 2 )k12 t, llo =
(30)
IIIII~.
we
which
for
c
and
c
in the sense that there exist two positive constants 1
2
have
c1 < JI/IJk/IJ/11~ < c2 whenever /E Wk 2 (~).
We may thus renorm the space Wk 2 (R") byJJ/11~; Wk 2 (~) may thus
be defined as the totality of f E L 2 (Rn) such that 11111~ is finite. One advantage of the new formulation of Wk 2 (Rn) is that we may also consider the
case of a negative exponent k. Then, as in the case of L 2 (R") pertaining
to the ordinary Lebesgue measure dx, we see that the dual space of the
renormed space Wk 2 (R") is the space w-k, 2 (Rn) normed by !l!ll~k
This observation due to L. ScHWARTZ [5] is of an earlier date than the
introduction of the negative by P. LAX [2].
IJ(1
156
3. Convolutions
We define the convolution (Faltung) of two functions f, g of C (R"), one
of which has a compact support, by (cf. the case in which f, gE 6(R")
in Chapter VI, 1)
(f*g) (x)
= J f(x-y) g(y) dy = f
f(y) g(x-y) dy
= (g*/) (x). (1)
Suggested by this formula, we define the convolution of aTE ;l) (R")' and
a rp E ;l) (R") (or aTE ~(R")' and a rp E ~(R")) by
(T*rp) (x) = Tey 1 (rp(x-y)),
(2)
where T[yJ indicates that we apply the distribution T on test functions

of y.
Proposition 1. (T
supp(rp), that is,
supp(T
* rp) (x) E C
00
(R") and supp (T
* rp) ~ supp (T) +
* rp) ~{wE R"; w = x + y, x E supp(T), y E supp(rp)}.
Moreover, we have
(3)
Proof. Let rp E ;l)(R") (orE ~(R")). If lim x" = x, then, as funch->O
tions of y, lim rp(x"- y) = <p(x- y) in ;l) (R") (or in ~(R")). Hence
h->0
T[Yl (rp (x- y)) = (T rp) (x) is continuous in x. The inclusion relation
of the supports is proved by the fact that T[yJ (rp (x- y)) = 0 unless the
support of T and that of rp (x- y) as a function of y meet. Next let e1 be
the unit vector of R" along the XJ-axis and consider the expression
T[yJ ((rp (x
+ he1 -
y)- rp (x- y))/h).
When h ~ 0, the function enclosed by the outer parenthesis converges.

as a function of y, to (
proved
!:J
f)
fJxl (T
(x- y) in ;l) (R") (or in
* cp) (x) =
~ (R")).
Thus we have
fJrp)
* fJxl
(x).
Moreover, we have
(T
tp
fJrp)
* fJxl
(x) =
T[y] -
fJrp(x-y))
EJyl
FJT[11]
EJyl
(rp (x- y)) =
(FJT
EJxl
* rp) (x).
Proposition 2. If rp and tp are in ;l) (R") and T E ;l) (R")' (or rp E ~ (R"),
E ;l) (R") and T E ~ (R")'), then
~*~*tp=T*~*~
3. Convolutions
157
Proof. We approximate the function (fl! * VJ) (x) by the Riemann sum
f,. (x)
h-n ~ fP (x- kh) 1p (kh),

h
where h > 0 and k ranges over points of R" with integral coordinates.
Then, for every differentiation D"' and for every compact set of x,
D"'f,(x)
h-n ~ D"'ff!(X- kh) 1p(kh)

II
converges, ash t 0, to ((D"'ff) * VJ) (x) = (D"'(ff! * 1p)) (x) uniformly in x.

Hence we see that lim/,= fP * 1p in 'Il(R") (or in ~(R")). Therefore, by
h.j,O
the linearity and the continuity of T, we have

(T * (fl! * VJ)) (x) =lim (T * /,.) (x) =lim h-n ~ (T * fl!) (x- kh) 1p (kh)
h.j,O
= ((T
* fl!) * VJ) (x).
h+O
Definition. Let fP E 'Il (R") be non-negative,
supp(ff!) ~ {xE R";
lxl <
JfP dx =
1 and such that
R"
1}. We may, for instance, take
fP (x) = exp (1/(lx 12 -1))/
J exp (1/(lx 12-1)) dx if IxI< 1,
JxJ<l
=Oiflxl>l.
We write ff!,(x) for s-"fl!(xjs), s> 0, and call T*ff the regularization
of T E 'Il (R")' (orE ~ (R")') through ff, (x)'s (Cf. Chapter I, 1).
Theorem 1. Let T E 'Il (R")' (or E ~ (R")'). Then lim (T * ff,) = T in
e.j,O
the weak* topology of 'Il(R")' (or of ~(R")'). In this sense, ff, is called
an approximate identity.
For the proof we prepare a
Lemma. For any 1pE 'Il(R") (orE ~(R")), we have lim 1p *ff = 1p
+O
in 'Il(R") (or in ~(R")).

Proof. We first observe that supp(VJ * cp,) ~ supp(VJ) + supp(ff!,) =
supp(VJ) + s. We have, by (3), D"'(VJ*ff,) = (D"'VJ) *cp. Hence we
have to show that lim (VJ * ff,) (x) = 1p (x) uniformly on any compact set
+O
Jff!e (y) dy = 1,
(VJ * ff,) (x) -1p (x) = J {1p (x- y) -
of x. But, by
1p (x)} ff!e (y) dy.
R"
Therefore, by fl!s (x) > 0,
Jfl!e (y) dy =
1 and the uniform continuity of
R"
1p(x) on any compact interval of x, we obtain the Lemma.

Proof of Theorem 1. We have, by -i/J(x) = 1p(-x),
T(VJ) = (T
* ip) (0).
(5)
1.1)8
Hence we have to prove that lim ((T

as proved in (4),
(T t.p,) {1p) = T ((t.p.
e.j,O
* t.p,) * ip) (0) =
(T
* ip) (0). But,
(T*t.p,)*ip=TtE-(t.p.*lP) and so, by (5),

Hence we obtain lim (T t.p,) (1p) =
* ip)
v ).
e.j,O
T((ipf) = T(w) by the Lemma.

We next prove a theorem which concerns with a characterization of
the operation of convolution.
Theorem 2 (L. ScHWARTZ). Let L be a continuous linear mapping on
~ (R") into a; (R") such that
L-r11 t.p = -rhLt.p for any hE Rn and t.p E ~ (R"),
where the translation operator
Th
Th'f!
(6)
is defined by
(x) = t.p (x- h).
(7)
Then there exists a uniquely determined T E ~ (Rn)' such that Lt.p =-=
T t.p. Conversely, any T E ~ (Rn)' defines a continuous linear map L
on ~ (Rn) into a; (Rn) by Lt.p = T t.p such that L satisfies (6).
Proof. Since t.p--+ ip is a continuous linear map of ~ (Rn) onto ~ (R"),
the linear map T: ip--+ (Lt.p) (0) defines a distribution T E ~ (Rn)'. Hence,
by (5), (Lt.p) (0) = T (ip) = (T t.p) (0). If we replace t.p by Th'f! and make
use of condition (6), then we obtain (Lt.p) (h)= (T t.p) (h). The
converse part of Theorem 2 is easily proved by (2), Proposition 1 and (5).
Corollary. Let T 1 E ~ (Rn)' and T 2 E ~ (Rn)'. Then the convolution
T 1 T 2 may be defined, through the continuous linear map L on
'1) (Rn) into (; (Rn) as follows:
(Tl
* T2) *'P =
L(t.p)
Tl
* (T
*t.p),t.pE
(8)
~(Rn).
Proof. The map t.p--+ T 2 t.p is continuous linear on ~ (Rn) into

~ (Rn), since supp(T2 ) is compact. Hence the map t.p--+ T 1
(T2 t.p)
is continuous linear on ~ (Rn) into ~ (R"). It is easily verified that condition (6) is satisfied for the present L.
Remark. We see, by (4), that the above definition of the convolution
T 1 T 2 agrees with the previous one in the case where T 2 is defined by
a function E ~ (Rn). It is to be noted that we may also define T 1 T 2 by
(Tl
* T2) (t.p)
= (TI(xJXT2(yJ) (t.p(x
+ y)),t.pE ~(Rn),
(8)'
where T 1 (x) X T 2 (y) is the tensor product of T 1 and T 2 See L. ScHWARTZ

[1 ].
Theorem 3. Let T 1 E ~ (Rn)' and T 2 E ~ (Rn)'. Then we can define
another "convolution" T 2 B3 T 1 through the continuous linear map L:
t.p--+ T 2 (T1 t.p) on ~ (Rn) into ~ (Rn)
as follows. (T2 00 T 1) (*t.p) = L (t.p), rp E ~ (Rn). Then we can prove T 2 00
T 1 = T 1 T 2 so that the convolution is commutative if it is defined
either as T 1 T 2 or a8 T 2 B3 T1
3. Convolutions
159
Proof. The map rp-+ T 1 rp is linear continuous on 'II (Rn) into

Hence the map rp-+ T 2 (T1 rp) is linear continuous on
'!l (Rn) into (; (Rn). Thus T 2 ffil T1 is well-defined. Next, we have for any
fP1 fP2 E 'II (Rn),
(J; (Rn).
(Tl * T2) * (rpl * fP2)
= Tl * (T2 * (rpl * fP2)) = Tl * ((T2 * fP1) * fP2)

= Tl * (rp2 * (T2 * fP1)) = (Tl * fP2) * (T2 * fP1)
by the commutativity of the convolution of functions and Proposition 2,

observing that supp (T2 rp 1 ) is compact because of T 2 t: (; (Rn)'.
Similarly we obtain
(T2 00 Tl) * (rpl * fP2) = T2 * (Tl * (rpl * fP2)) = T2 * ((Tl * fP2) * fP1)
= T2 * (rpl * (Tl * fP2)) = (T2 * fP1) * (Tl * fP2)

Thus (T1 * T 2 ) * (rp1 * rp 2) = (T 2 00 T1) * (rp1 * rp 2).. and so, by (5)
and the Lemma above, we obtain (T1 T 2) (rp) = (T2 00 1'1) (rp) for all
rpE 'Il(Rn), that is, (T1 T 2 ) = (T200 Ttl
Corollary
T1 * (T2
*T
(T1 * T 2 ) *T3
ifallTjexcept one havecompactsupport,
D'"(T1 * T 2 )
3)
(D'"T1 ) * T 2
= T1 *
Proof. We have, by (5) and the definition of T 1

(T1 * (T 2 * T3 )) (rp)
=
=
(9)
(10)
(D'"T2 ).
*T
2,
((T1 * (T 2 * T 3)) * ip) (0)

(T1 * ((T2 * T 3) * i/J)) (0)
= (T1 * (T2 * (T3 * ip))) (0)

and similarly
((T1 * T 2 ) * T 3) (rp) = (T1 * (T2 * (T3 * ip))) (0),

so that (9) holds.
The proof of (10) is as follows. We observe that, by (3),
(D"'T~)
* rp
T~
* (D"'rp)
D"'(T~
* rp)
(11)
D"'rp,
which implies
(D'"T) * rp
T * (Darp) = T * ((D"'Ta) * rp)
that is, by (5),

D'"T
(D"'T~)
(T * D'"Ta) * rp,
* T.
(12)
Hence, by the commutativity (Theorem 3) and the associativity (9),
DIX(Tl * T2)
(D'"TIJ) * (Tl * T2)
(D'"T~)
= ((D'"TIJ)
* Tl) * T2
(D"'Tl) * T2
* (T2 *T1 ) = ((D"'T4)*T2)* T 1 = (D'" T 2)*T1 .
160
The Fourier Transform of the Convolution. We first prove a theorem

which will be sharpened to the Paley-Wiener theorem in the next section.
Theorem 4. The Fourier transform of a distribution T E @; (Rn)' is
given by the function
T (~)
= (2n)-nf2 T[x] (e-i(x,n).
(13)
Proof. When s,} 0, the regularization T. = T cp. tends to Tin the

weak* topology of @:(Rn)' and a fortiori in the weak* topology of ~(Rn)'.
This we see from (T cp.) (tp) = (T (cp. ?p)) (0) = T[xJ((cp. ?p) (-x))
and the Lemma. Thus, by the continuity of the Fourier transform in the
weak* topology of
~ (Rn)'.
* *
* cp.) =
/"-...
(Rn)', lim (T
e.j.O
T in the weak* topology of
Now formula (13) is clear for the distribution defined by the

function (T cp,) (x). Hence
(2n)nf2
('I~.) (~)
= (T
* cp.)[.,l (e-i(x,<))'
which is, by (5), = (T[x]

(cp. e-i(xJ))) (0) = T[x] (~. e-i(xJ>).
The last expression tends, ass,} 0, to T[xJ (ei(x,;)) uniformly in~ on any
bounded set of ~ of the complex n-space. This proves Theorem 4.
Theorem 5. If we define the convolution of a distribution T E ~ (Rn)'
and a function cp E ~ (Rn) by (T cp) (x) = T[yJ (cp (x- y)), then the
linear map L on ~ (Rn) into @; (Rn) : cp ___,.. T cp is characterized by the
continuity and the translation invariance -rh.L = L-rh.
Proof. Similar to the proof of Theorym 2.
Theorem 6. If T E
~ (Rn)'
and cp E ~ (Rn), then
/'---.
(T*cp)
If T 1 E
= (2nt12 (j;T
(14)
(Rn)' and T 2 E @: (Rn)', then
(15)
which has a sense since, as proved above in Theorem 4, T2 is given by
a function.
Proof. Let tp E ~ (Rn). Then the Fourier transform of({; tp is, by (13)
of Chapter VI, 1, equal to (2n)-nf 2 ~ fJ = (2n)-nf 2 ip ip. Thus
*
(T * cp) (tp) = (T * cp) (ip) = ((T * cp) * q;) (0) = (T * (cp * fJ )) (0)
= T((cp * ~) y) = T(~ * fJ) = T((2n)nf
tp()
,/"-...
2 (({;
(2n)n 12
which proves (14).
T (q;tp) =
(2n)nf 2 9JT(tp),
4. The Paley-Wiener Theorems. The One-sided Laplace Transform
T1
161
Let tp. be the regularization T 2 * cp. Then the Fourier transform of

* tp. = T * (T * cp.) = (T * T * cp. is, by (14), equal to
ip. = (2nt12 T (2n)n12 T2 fi;. = (2n)n12 (T;;;T2) cp,.
(2nt12
1
2)
f'1
Hence we obtain (15), by letting e t 0 and using lim cP.(x)

s.j.O
1.

The Fourier transform of a functionE ego (Rn) is characterized by the
Paley-Wiener Theorem for Functions. An entire holomorphic function
F (C) = F (Cv C2 , , Cn) of n complex variables Ci = Ei + i 'YJi (i =
1, 2, ... , n) is the Fourier-Laplace transform
F(C)
(2n)-n12
J e-i(C,x>j(x)dx
(1)
Rn
of a function f E ego (Rn) with supp (f) contained in the sphere IxI < B
of Rn iff there exists for every integer N a positive constant eN such that
IF(C) I<
eN (1 +
(2)
ICI)-N eBIImCI.
Proof. The necessity is clear from
.IT (iCj)Pi F(C) =
(2n)-nf2
J e-i(C,x) DPf(x) dx
~SB
J=l
which is obtained by partial integration.

The sufficiency. Let -:IS define
f(x) = (2n)-n12
J ei(x,t;) F(E) dE.
(3)
Rn
Then, as for the case of functions E 6 (Rn), we can prove that the Fourier
transform j (E) is equal to F (~) and f E C"" (Rn). The last assertion is
proved by differentiation:
DPf(x)
(2n)-nf 2
.IT (i~j)Pi F(~) d~,

f e'<x.~> J=l
Rn
(4)
by making use of condition (2). The same condition (2) and Cauchy's
integral theorem enable us to shift the real domain of integration in (3)
into the complex domain so that we obtain
f(x)
(2n)-nf2
J e'<x.HirJ) F(~ + irj) d~
for arbitrary real 'YJ of the form 'YJ =

by taking N = n + 1,
l/(x) I<
xxfl xI with x > 0.
eN eBlrJJ-(x,fJ) (2n)-n/2
If lxl > B, we obtain f(x)

{xE Rn; lxl < B}.
(3')
Rn
J (1 +
We thus obtain,
1~1)-N d~.
Rn
0 by letting x
t + oo.
Hence supp(/) ~
162
The above theorem may be generalized to distributions with compact

support. Thus we have
ThePaley-WienerTheorem forDistributions Eli(R")' (L. ScHWARTZ).
An entire holomorphic function F (C) = F ((1 , ... , C,.) of n complex variables Ci = ~i + i rJi (j = 1, 2, ... , n) is the Fourier-Laplace transform
of a distribution T E ~ (Rn)' iff for some positive constants B, N and C
IF(C) I< C(1 + ICI)N eBIImCI.
(5)
Proof. The necessity is clear from the fact (Theorem 2 in Chapter I, 13)
that, if TE ~(Rn)', there exist positive constants C, Nand B such that
IT(r:p)I<C..E supiD11 r:p(x)l
whenever r:pE~(R").
jtif~N jxj ~B
For, we have only to take tp (x) = e-i(xJ) and apply (13) of the preceding
section.
The sufficiency. F (~) is in ~ (Rn)' and so it is the Fourier transform
of a distribution T E e> (Rn)'. The Fourier transform of the regularization
T, = T
r:p, is (2nr1 2 f. (j;, by (14) of the preceding section. Since the
supp (r:p,) is in the sphere IxI < 8 of Rn, we have, by the preceding theorem,
l(j;,(~)
1~ C'.
eIImCI.
Moreover, since f is defined by the function F (~).we see that (2nt/ 2 i' (j;,
is defiend by a function (2n)n/Z F (~) (j;, (~) which, when extended to the
complex n-space analytically, satisfies the estimate of type (5) with
B replaced by B +e. Thus, by the preceding theorem, T, = T * r:p.
has its support in the sphere lx I :S B + 8 of Rn. Thus, letting 8 t 0 and
making use of the Lemma in the preceding section, we see that the supp (T)
is in the sphere jx I< B of Rn.
Remark. The formulation and the proofs of the Paley-Wiener theorems
given above are adapted from L. H6RMANDER [2].
The Fourier Transform and the One-sided Laplace Transform. Let
g(t) E L2(0, =).Then, for x > 0,
g(t)
e-tx
E L 1 (0, =) f\ L 2 (0, =)
as may be seen by Schwarz' inequality. Hence, by Plancherel's theorem,

we have, for the Fourier transform
f(x
+- iy) =
J g(t) e-tx e-ity dt

00
(2n)- 112
= (2n)- 112
00
-00
00
(x
>
(6)
0),
the inequality
J lf(x+iy)l
J g (t) e-t(x+iy; dt
00
dv=
J lg(t)l
0
00
e- 21xdt<
J lg(t)l
0
dt.
(7)
163
The function f(x + iy) is holomorphic in z = x + iy in the right halfplane Re (z) = x > 0, as may be seen by differentiating (6) under the
integral sign observing that g ( t) t e-t as a function oft belongs to L 1 (0, oo)
and to L2 (O,oo) when Re(z) = x > 0. We have thus proved
Theorem 1. Let g (t) E L 2 (0, oo). Then the one-sided Laplace transform
f(z) = (2 nt 112
J g(t) e-tz dt
00
(Re(z)
>
(6')
0)
belongs to the so-called Hardy-Lebesgue class H 2 (0), that is, (i) f(z) is
holomorphic in the right half-plane Re (z) > 0, (ii) for each fixed x > 0,
f(x + iy) as a function of y belongs to L2(-oo,oo) in such a way that
~~~ (_[1 f (x + i y)
12
dy)
<
(7')
oo.
This theorem admits a converse, that is, we have

Theorem 2 (PALEY-WIENER). Let f(z) E H 2 (0). Then the boundary
function f(iy) E L2(-oo,oo) of f(x + iy) exists in the sense that
00
J lf(iy)- f(x + iy) 2 dy =
lim
(8)
J f(iy) eity dy
N-+00 -N
(9)
x.j-0 _ 00
in such a way that the inverse Fourier transform

g(t)
= (2n)- 112 I.i.m.
vanishes for t < 0 and f(z) may be obtained as the one-sided Laplace
transform of g (t).
Proof. By the local weak compactness of L 2 (-oo, oo), we see, that
there exists a sequence {xn} and an f(iy) E L 2 (-oo, oo) such that
xn-J.-0 and weak-limf(xn+iy)=/(iy).

For any
o>
n-+00
0, there exists a sequence {Nk} such that

6
lim Nk = oo
k-+00
and
lim
k~oo
J Jf(x iNk) 2 dx =
0+
0,
as may be seen from
_N{ 6Jf(x + iy)

0
12
} < o6{
dx dy
oo
Jf(x + iy)
dy dx
<
oo
(for all N
Thus, by Schwarz' inequality,
>
0).
lim
J lf(x iNk) I dx =
k~ooo+
11*
0.
(10)
164
From this we obtain
J j(itjtdt
00
f(z) = (2n)-1
(Re(z)
>
(11)
0).
-00
The proof may be obtained as follows. By Cauchy's integral representation, we obtain
JJ''lz
d?;, (Re (z) > 0),

c
where the path C of integration is composed of the segments
f(z)
(2ni)-1
x 1 +iNk, x 0
<
(x0
Re(z)
+ iNk,
<
Xv
(12)
x 0 +iNk, x 0 - iNk
-Nk < Im(z)
<
Nk)
so that the closed contour C encloses the point z. Hence, by letting k-+ oo
and observing (10), we obtain
J
00
f(z) = (2n)-1
-oo
f(xo +it~ dt
z- (x0 + tt)
J
00
+ (2n)-1
-oo
f(xl -:f-it) dt.

(x1 + tt) - z
We see that the second term on the right tends to 0 as x 1 -+ oo, because
of (7') and Schwarz' inequality. We set, in the first term on the right,
x0 = Xn and letting n-+ oo we obtain (11). Similarly, we obtain
J (t(it) dt
00
o=
(2nr 1
>
0,
(for x
<
Let us put, for Re (z)

h(x)
-z
-oo
0);
h(x)
<
(Re(z)
0).
(for x
e-"'
(13)
>
0).
We then have
J h (x) eixt dx = J eixt-zx dx =

00
00
-oo
(z- it)- 1 ,
and so, by Plancherel's theorem,
J
N
l.t.m. (2n)-1
N-+00
Similarly we have
l.t.m.
N-+00
-N
J
N
(2 n)-1
e-4 1"'
{
0 (for x
--.tdt= _.,. (f
z-t
e
or x
e-1"'
- - .t
-N Z - t
dt =
{-e-"
0
(for x
(f
<
>
<
Or X>
0)
O) if Re(z)
0)
O)
if Re (z)
>
0.
(14)
<
0.
(14')
Therefore, by applying the Parseval theorem (27) of Chapter VI, 2

to (11), we obtain the result that /(z) is the one-sided Laplace transform
165
of g(t) given by (9). By applying the Parseval theorem to (13), we see also
that g(t) = 0 fort< 0.
We finally shall prove that (8) is true. By adding (11) and (13), we
see that f(z) admits Poisson's integral representation
J
=-;00
f(z) =f(x
+ ~y)
By virtue of
-oo
/(it)
(t-y)z +xadt
whenever
x>
0.
(15)
00
~
n
we have
-oo
<; J
dt
(t - y) 2
+ xz --1 '
(16)
00
Jf(x
+ iy)- f(iy)
-oo
and so
lt(x
if+(s
~/~ ~f+(y) Ids,
+ iy)- f(iy) I' dy < (:)' _[

<
where f+(y) =f(iy),
Ll'f+l ;. ~~,f+IYI'
'"'}'y
12 ds)d
Joo if+(s +52y)-f+(y)
z2 Joo( Joo~)(
Y
+ xz
52 + xz
n2
~:
i ,. ~ ..(_[If+~<+
-oo
-oo
-oo
JttU+;+
y)
-f+(y) l'dyl
00
= ~
n
-oo
sz
s)
xz
ds
'
where 0 < p,(f+; s) < 4JJf+ JJ 2 and p,(f+; s) is continuous in s and has
value 0 at s = 0.
To prove that the right side tends to zero as x t 0, we take, for any
s > 0, a ~ = ~(s) > 0 such that p,(f+; s) < s whenever JsJ < ~- We
decompose the integral
:i ~,II_;;~'"' ~ : Lt _f ll ~
+ +
I, +I, +I,
We have JI2 J< s by (16), and Jiil < 4n-1 llf+IJ 2 cot-1 (~/x) (j = 1, 2).
Hence the left integral must tend to zero as x t 0. This proves the Theorem.
Remark 1. For the original Paley-Wiener theorem, see PaleyWiener [1]. For the one-sided Laplace transform of tempered distributions, see L. SCHWARTZ [2].
166
Remark 2. M. SATO [1] has introduced a happy idea of defining a

"generalized function" as the "boundary value of an analytic function".
His idea may be explained as follows. Let $B be the totality of functions
<p (z) which are defined and regular in the upper-half and the lower-half
planes of the complex z-plane, and let \R be the totality of functions
which are regular in the whole complex z-plane. Then $B is a ring with
respect to the function sum and function multiplication, and \R is a subring of $B. Sato calls the residue class (mod \R) containing <p (z) as the
"generalized function" 1J (x) on the real axis R 1 defined through <p (z).
The "generalized derivative d1} (x)fdx of the generalized function 1J (x)"'
is naturally defined as the residue class (mod \R) containing drp (z)fdz.
Thus the "delta function <5 (x)" is the residue class (mod ffi) containing
-(2ni)-1 z-1 Sato's theory of "generalized function of many variables"
admits the following very interesting topological interpretation. Let M be
ann-dimensional real analytic manifold and let X be a complexification
of M. Then the n-th relative cohomology group Hn(X mod (X- M))
with the germ of regular functions in X as coefficients gives rise to a
notion of the "generalized function on M". That is, the relative cohomology class is a natural definition of the "generalized function".
Remark 3. For more detailed treatment of the Fourier transform of
generalized functions, see L. ScHWARTZ [1] and GELFAND-SILOV [1].
In the latter book, many interesting classes of basic functions other than
~ (Rn), @) (Rn) and () M (Rn) are introduced to define generalized functions;
the Fourier transform of the corresponding generalized functions
are also discussed in GELFAND-SILOV [1]. Cf. also A. FRIEDMAN [1]
and L. H6RMANDER [6].
5. Titchmarsh's Theorem
Theorem (E. C. TITCHMARSH). Let f(x) and g(x) be real- or complexvalued continuous functions. 0 < x <ex:>, such that
%
(/*g) (x) =
J f(x-y) g(y) dy = 0J g{x -- y) f(y) dy =

0
~
(g *f) (x)
(1)
vanishes identically. Then either one off (x) or g (x) must vanish identically.
There is a variety of proofs of this important theorem, ~uch as by
TITCHMARSH [1] himself and also by CRUM and DuFRESNOY. The following proof is elementary in the sense that it does not appeal to the theory
of functions of a complex variable. It is due to RYLL-N ARDZEWSKI [1] and
given in the book by J. MIKUSI:NSKI [1].
Lemma 1 (PHRAGMEN). If g(u) is continuous for 0:::::; u:::::; T, and
0 :S t :::::; T, then
lim
%-->00
i; t"- 1(~ f
k~l
k.
ekx(t-u)
g(u) du =
g(u) du.
(2)
167
5. Titchmarsh's Theorem
Proof. We have
00
k=l
(-1)k- 1 (k!)- 1 ekx(t-u) -.1- exp(-ex(t-u)), and,
for fixed x and t, the series on the left converges uniformly in u for
0 < u < T. Thus the summation in (2) may be carried out under the
integral sign and we obtain (2) by the Lebesgue-Fatou Lemma.
Lemma 2. If f(t) is continuous for 0:::::: t < T and
II
en 1 f(t) dtl
~M
for n = 1; 2, ... , where M is a positive constant which is independent

of n, then f(t) must be = 0 for 0 < t < T.
Proof. We have
Ii; J ( -k.l r l in(t-u) f(T -u) du'i:::;;; i
k=l
k=l 0
<
If t
~ekn(t-T) iJin(T-u) f(T -tt) du

k.
M (exp(e-n(T-t)) -1).
T, the last expression tends to zero as n-+ oo. Hence, by Lemma 1,
with g(u)
= f(T - u), we see that
J f(T -- u) du =
0 for 0 :'S t
< T.
Since f is continuous, it follows that f(t) = 0 for 0 < t < T.

Corollary 1. If g(x) is continuous for 1 < x <X and if there exists
p~sitive
number N such that
I/
xng(x) dx/
<
N (n = 1, 2, ... ),
then g(x) = 0 for 1 < x <X.

Proof. Putting x = ~. X= eT and xg(x) = f(t), we obtain, by
Lemma 2, that f(t) = 0 for 0 < t < T. Thus xg(x) = (I for 1 < x <X,
and so g(x) = 0 for 1 < x <X.
Corollary 2 (Lerch's theorem). If f(t) is continuous for 0 ~ t < T and
T
J tnf(t) dt =
0 (n
1, 2, ... ), then f(t)
0 for 0
< t < T.
Proof. Let t0 be any number from the open interval (0, T), and put
T = t0 X, f(t) = g(x). We then obtain
= t0 x,
to+l
J xng(x) dx =
0 (n = 1, 2, ... ),
and so I
j xng(x) dxi =IloJxng(x) dxi < o/ig(x) Idx =N
(n = 1, 2, ... ).
Hence, by Corollary 1, we obtain g (x) = 0 for 1 < x < X, and so

f(t) = 0 for t0 :::;::; t < T. Since t0 was an arbitrary point of (0, T), we
must have f(t) = 0 for 0:::;;; t < T.
The proof of Titchmarsh's theorem. We shall first prove the Theorem
for the special case when f = g, that is, if /(t) is continuous and(/*/) (t)
J f(t- u) f(u) du =
0 for 0 < t
< 2T, then f(t) = 0 for 0 < t < T.
168
We have
/ 2T en( 2 T-t)
l(u) l(t- 1~) du) dt
and, by the change of variables u =T-v, t
0,
2T- v- w, we obtain
J J en(v+w) I(T- v) I(T- w) dv dw
..1
where .1 is the triangle v + w > 0, v < T, w <Tin the v- w plane.

Let .1' be the triangle v + w ~ 0, v >- T, w >- T. Then the join
.1 + .1' is the square -T < v, w < T. The above equality shows that
the integral of en(v+w) I(T- v) I(T- w) over .1 + .1' is equal to the integral over .1'. The integral over L1 + .1' is the product of two single integrals, and in the integral over .1' we have en(v+w) < 1. Thus
enu I(T- u) dul =
1-T
I J J en(v+w) I(T- v) I(T- w) dv dw I

d+LI'
< Jfli(T-v) I(T-w)j dvdw :S:
2T2 -A 2 ,
Ll'
where A is the maximum of ll(t) I for 0 < t < 2T, and 2T2 is the area
of .1'. We thus have
j_f enui(T-u)dul< V2TA,
and, moreover,
If
j_f enu I(T- u) dul< T A. Therefore
enu I(T-u) dul =
Lf- _/I<
(1
V2) TA
(n
= 1, 2, ... ),
and so, by Lemma 2, I (t) = 0 for 0 ;;:;; t ;;:;; T.

We are now ready to prove Titchmarsh's theorem for the general case.
I
Let
I
J l(t-u) g(u) du =
0 for 0 < t< ex:>. Then we have, for 0 :S: t< ex:>,
J (t-u)l(t-u)g(u) du + J l(t-u)ug(u)du=t J l(t-u)g(u)d~t=O.
This may be written as
(11 *g) (t)
where
(I* g1) (t) = 0 (0
11 (t) = tl(t),
Thus
and so
[f * {g1 * (11 * g
g1 (t)
< t <ex:>),
tg (t).
+ I* g1)}] (t) = o
W *g)* (l1 * g1)J (t) + [(! * g1) *(I* g1)J (t) = o.
6. M:ikusi.D.ski's Operational Calculus
169
Thus, by (f *g) (t) = 0, we have [(/ * g1) * (f * g1)] (t) = 0, and so,
by the special case proved above, (f g1) (t) = 0, that is,
Jf(t-u) ug(u) du =
0 (0
<t<
oo).
From this we obtain, similarly to above,

I
Jf(t-u) u g(u) du =
2
0 (0 ~ t
< oo).
Repeating the argument, we find that
Jf(t-u) u"g(u) du =
0
0 (O<t< oo; n = 1, 2, ...).
Hence, by Lerch's theorem proved above, we obtain
f(t-u)g(u)=O
for
o:::;;u<t<oo.
If a u0 exists for which g (u0 ) =I= 0, then f (t- u 0) = 0 for all t > u0 ,
that is, f(v) = 0 for all v > 0. Therefore, we have either l(v) = 0 for
all v > 0 or g (v) = 0 for all v > 0.
6. Mikusmski's Operational Calculus

In his "Electromagnetic Theory", London (1899), the physicist
0. HEAVISIDE inaugurated an operational calculus which he successfully
applied to ordinary linear differential equations connected with electrotechnical problems. In his calculus occured certain operators whose
meaning is not at all obvious. The interpretation of such operators as
given by HEAVISIDE himself is difficult to justify. The interpretation
given by his successors is unclear with regard to its range of validity,
since it is based upon the theory of Laplace transforms. The theory of
convolution quotients due to J. MIKUSINSKI provides a clear and simple
basis for an operational calculus applicable to ordinary differential equations with constant coefficients, as well as to certain partial differential equations with constant coefficients, difference equations and integral
equations.
The convolution quotients. We denote by C the totality of complexvalued functions I (t) defined for 0 < t < oo. In this section, we write such
functions by {f(t)} or simply by I; whereas, f(t) will mean the value at
t of the function l(t). We write {l(t)} {g(t)} or simply I g for the convolution function{/ l(t - s) g(s)
ds}. i.e.,
{f(t)} {g(t)} ={(/*g) (t)} = {/ f(t
-s) g(s) ds}.
(1)
170
We have, as proved in Chapter VI, 3,
f g = g f (commutativity),
f (g h)
= (/ g) h (associativity).
We can define, beside the convolution product
{f(t)}
+ {g(t)} =
{f(t)
(2)
(3)
f g, the sum through
+ g(t)},
(4)
and we have the distributive law
h . (/
+ g) =
h 1 + h g.
(5)
Hence C is a ring with respect to the addition f + g and the multiplication f g. The zero of this ring is represented by the function which is
identically zero; we shall thus denote this function by 0. This ring C is
without zero factor, that is, in C, f g = 0 implies either f = 0 or g = 0.
This is a consequence of Titchmarsh's theorem. Hence, by introducing
the convolution quotient ffg = j_ of two functions f, g E C with g:j= 0,
g
we obtain a commutative field Q:
afb = cjd is equivalent to ad= be, and, in particular,

ajb = cis equivalent to a = be,
a
(6)
a c
(7)
7)"-;[=l).d,
+d =
ad+hc
--b- -d-~'
ajb =a cfb c (c=i= 0).
(8)
(9)
The operator. We shall call the quotient ajb an "operator". Any

a E C gives an example of operators, since, by (9), we can identify a with
a bfb (b =i= 0).
The unit- or ()-operator. The operator cjc (c =i= 0) is the unit of the
multiplication in the field Q. For, we have, by (7) and (9),
(10)
We have, by (6), cjc = bfb. We call cjc the unit- or ()-operator, and we shall
henceforth denote this operator by 1:
a
1 b
=b
=b
(11)
The operator 1 = cjc does not belong to C, for, if we take for c the function {1}, then {1}/{1}
= { / f(s)
d,.,} =
= {f(t)} E C implies that {1} {f(t)} = {/ 1 f(s) ds}
{1}, which is surely a contradiction.
171
6. Mikusinski's Operational Calculus
The operator of integration. We shall denote by h the operator defined
by the function {1}:
(12)
h = {1},
and call h the operator of integration, since, for any f E C, we have, as

above,
h {f(t)}
= {1} {f(t)} = {/ f(s) ds}.
(13)
Remark. In his book, referred to above, MIKUSINSKI uses the symbol

l for {1}. We here use the symbol h in honour of Heaviside and for typo-
graphical convention. A locally integrable function {f(t)}, t

identified with the operator {
jf
> 0, may be
(s) ds} fh. Thus the convolution quotient
is another kind of "generalized function".

The scalar operator. Let X be any complex number, and {X} the function which is identically equal to X, Then the operator
(14)
is called a scalar operator, because we have
[1X]
Proof.
[1X]
[{J] = [1X
+ [{J] =
[X J [{J] --
+ {J],
{~}
[1X]. [fJ] =
+ {!} =
{1X
[1X{J], [1X]. {f(t)} = {1X/(t)}. (15)

{3} = [1X
+ {J],
{1X}. {{3}- {1X{3t}- h. {1X{3}- {1X{3}- [ {J]

h - X
h2
h2 h2
{ / iXf(s)
h
[1X]{f(t)}={1X}y(t)}= 0
ds}
=h{~f(t)}={1X/(t)}.
Remark. As a corollary, we obtain

{a (t)}
{1Xa (t)}
[X J. {b (t)} = {b(t)f,
(16)
so that the effect of the operator [X] is exactly the X- times multiplication. We can thus identify the scalar operator{1}/{1}with the unit operator 1, and the operator [X] with the number X,
The operator of differentiation. We shall denote the operator 1/h by
the symbol s:
= 1/h =
1/{1}.
(17)
s is called the operator of differentiation, since, if f = {I (t)} E C has a

continuous derivative f' = {/' (t)}, then
(18)
sf= f' + /(0), where /(0) ='-' {f(O)}fh.
172
Proof. Multiply both sides of the equation
= {/{0)} + {/ f' (s)
{f(t)}
by s and use the fact s h

Corollary 1. If
{/(n) (t)}, then

/(n)
Sn .
hs
ds}
h /{0)
+ h {/' {t)}
1.
= {/ (t)} has a continuous n-th derivative
f- Sn-1 f (0) -
sn-2 .
f' {0) - ... -
s /(n-2)
{0) -
/(n-1)
jfnl
{0)'
where jUl (0) is the operator of the jUl (0)-times multiplication,
(19)
that is, jUl (0) = {jUl (O)}jh.

Proof. For n
s2 f
s (s f)
= 2, we have
s (/'
+ f(O)) =
s-/'
+ s /{0) =
f"
+ f' (0) + s /{0).
The general case may be proved by induction.

Corollary 2. We have
{20)
and more generally,
1/(s- <X)n
Proof. s . {e"'1}
1j(s- <X) 2
f tfl-1
l
= \(n
_ 1)! e"'1f
= {<Xe"'1}
+1=
<X{e"'1}
(n = 1, 2, ... ) .
(21)
+ 1 by (18). Thus (20) holds.
= {e"'1} {e"'1} = {/ e"(t-s) e"'
ds}
{e"' 1/ ds}
= {:1 e"'1} .
Application to the integration of linear ordinary differential equations

with constant coefficients. We shall show the applications by examples.
Example 1. Solve the equation
x" (t)- x' (t)- 6x(t)
2, x(O)
1, x' (0)
0.
Solution. We write the equation in the operator form
x" (t) - x' (t) - 6x (t)
Then, by (19),
s2 x - s x(O)- x' (0)- s x
2/s.
+ x(O)- 6x =
2/s,
and so, by substituting the initial condition, we obtain

s2 x - s - s x
H~nce,
+ 1- 6x =
by (20),
x
2/s, i.e., (s2 - s - 6) x
s -1
+ 2/s.
+2
1 1
8
1
4
1
+ 2) = - 3 s + 15 s- 3 + 5 s + 2
f 1 8 31 4 -2tl
= 1-3 + 15 e + 5 e f
s2 -
s (s- 3) (s
17B
7. Sobolev's Lemma
Example 2. Let .?. be a constant =F 0. Solve the equation
x" (t)
+ .?.2 x(t)
= 0, x(O) =ex, x' (0) = {3.
Solution. The corresponding operator equation is

s2 x-cxs -{3
+ .?.2 x =
0, i.e., (s 2
+ .?.2) x =
+ {3.
cxs
Hence, expanding into partial fraction, we obtain

<5
y
cxs+P
x = sz + i2 = s + iA + s - i A.
with iXS
+ {3 =
y(s- iJ.)
+ O(s + iJ.) SO that
iP)
ip) 0=2-1 (cx-;:.
y=2-1 ( ex+;:,
Hence
ip
iP
x=2(cx+ ;:)s +iA. +2(cx-T)s-iA.

- f_!_ (ex+ ip) e-iAt +_!_(ex- ip) ei.<el
J
A.
2
A.
-l2
=lex
, t + Tp.,t
sm ~~.tf.
cos~~.
Example 3. Solve the system of equations
x'(t)-cxx(t)-{3y(t) =f3e"'1, y'(t) +f3x(t)-cxy(t) = 0,

under the initial condition x (0)
0, y (0) = 1.
Solution. Solving the operator equations
s x- cxx- f3y = {3/(s- ex), s y- 1

we obtain,
X
2P
+ pz'
1 . l
s-ex +~PI
(s- cx)Z
+ {3x- cxy =
(s-cx)z-pz
(s-ex) ((s- cx) 2
0.
+P
2) '
from which we obtain
X=
Y
J 1 . 2:-~ l_s-cx-~P
2:-~ {e<cx+iPJI -
e<cx-ip)e}
{2 ecxt sin {3t}
'
1
1
1
1
2 (s-ex)
(s- cx) 2 + p2 - s-ex= s-ex- iP +s-ex +ip-s- ex
{e(cx+iP)t
+ e<cx-ip)e _
e"'t}
{ecxt (2 cos {3t _ 1)}.
For further development and applications, see MIKUSINSKI's book, [1].

We also refer the reader to a book by A. ERDELYI [1].
7. Sobolev's Lemma
A generalized function is infinitely differentiable in the sense of
the distribution (see Chapter I, 8). So the differentiability in the generalized sense has no bearing on the ordinary differentiability. However,
174
we have the following result which is of fundamental importance m

modern treatment of partial differential equations.
Theorem (Sobolev's Lemma). Let G be a bounded open domain of
Rn. Let a function u (x) belong to Wk (G) fork> 2- 1 n + a, where a is an
integer ~ 0. We thus assume that the distributional derivatives of
u (x) of order up to and including k all belong to L2 (G). Then, for any open
subset G1 of G such that the closure G~ is a compact subset of G, there
exists a function u 1 (x) E ca (G1) such that u (x) = u 1 (x) a. e. in G1
Proof. Let ex (x) be a function E CZO (Rn) such that
G1 <;;; supp(ex) ( G, 0 < ex(x) :S 1 and ex(x) = 1 on G1 .
We define a function v (x) defined on Rn by
v (x) = ex (x) u (x) for x E G; v (x) = 0 for x E Rn - G.
Then v (x) = u (x) whenever x E G1 . Since v (x) is locally integrable over
Rn, it defines a distribution E ~ (Rn)'. By the assumption that u E Wk (G),
the distributional derivatives ns v (x) E L 2 (Rn) when Is I < k. For example'
the distributional derivative
8x1
(v)
8x1
(exu)
a~
8x1
+ ex 8x1 u
belongs to L 2 (Rn) by the fact that u, 8uf8xi both belong to L2 (G) and
that the infinitely differentiable function ex (x) has its support contained
in a compact subset of the open domain G. By the Fourier transformation
v (x) --+ v(y), we obtain
/'---..
(Ds v) (y)
= (i)lsl Yi' Y2' ... y~n v (y).
Since, by Plancherel's theorem, the L 2 -norm is preserved by the Fourier

.
/'---..
transformation, we have (D 5 v) (y) E L 2(Rn) for lsi< k. Thus

v(y)yi'Y~' ... y~"EL 2 (Rn) for lsl<k.
In particular we have
(1)
(1')
Let q = (qv q2 , , qn) be a system of non-negative integers. Then
by (1), we can prove that
(y) Yi' y~' ... y~" is integrable over Rn whenever Iq I + ; < k. (2)
For the proof, take any positive number C. We have, by Schwarz' inequality,
j lv (y) Yi' y~ . .. Y~"l dy
< (
l:vi;;;;C
l:vi;;;;C
IYi' Y~' Y~" /2 dy j
l:vi>C
i:vj;;;;c
lv(y) /2 dy) 1' 2 <
CXJ,
lv(y) Yi' y~ ... y~"l ay
<( f /( 1 + IY/2)-kl2yi'Y~' ... y~"l2dy

jyj>C
jyj>C
lv(y) (1
+ IY/2)k/2l2dy,112.
j
175
8. Garding's Inequality
The second factor on the right of the last inequality is < oo by (1). The
fir:;t factor is finite by
dy = dYI dy 2 dy, = r"- 1 dr dQ,,
where dQ, is the hypersurface element of the surfacer'
(3)
of tlre unit sphere of R" with the origin 0 as its centre,

provided
2lql--2k+n-1<-1, that is, if
k>;
+ lql.
Now, by Plancherel's theorem, we have

v(x) = l.i.m. (2:nr"12
h->00
[y[ :;;,h
v(y) exp(i (y, x)) dy
and so, as in the proof of the completeness of L 2 (R"), we can choose a

subsequence {h'} of positive integers h such that
v(x) =
Fm (2n)-nfZ
h -+00
But, since
equal to
J v (y) exp(i (y, x)) dy
[y[ :;i;h'
for a.e. xE R".
v(y) is integrable over R" as proved above, the right side is

v (x) = (2n)-"12 J v(y) exp (i (x, y)) dy,
1
Rn
that is, v (x) is equal to v1 (x) for a.e. x E R". By (2), the differentiation of
v1 (x) under the integral sign is justified up to the order a; and the result
of the differentiation is continuous in x. By taking u 1 (x) = v1 (x) for
x E G1, we have proved the Theorem.
Remark. For the original proof, see S. L. SoBOLEV [1], [2] and
L. KANTOROVITCH-G. AKILOV [1].
8. Garding's Inequality
Consider a quadratic integral form defined for C00 function u (x) =
u (x1 , x2 , , X 11) with compact support in a bounded domain G of R":
B [u, u]
[s[,fir:;;,m
(c 51 D 5 u, D 1u) 0 ,
(1)
where the complex-valued coefficients c51 are continuous on the closure

Ga of G, and (u, v) 0 denotes the scalar product in L2(G).
Then we have
Theorem (L. GARDING [1]). A sufficient condition for the existence of
positive constants c, C so that the inequality
llu II~< c ReB [u, u] + C llu II~
(2)
holds for all u E ego (G), is that, for some positive constant Co,
Re
~.
[s[,ltf=m
C51
~s ~~
2m for all
> c0 I~ 1
(~v ~2
x E G and all real vectors

~m) .
(3)
176
Remark. The inequality (2) is called Carding's ineqttality. If condition (3) is satisfied, then the differential operator
L =
2:
lsl=lll~m
(-1)1 11D 1c51 D 5
(4)
em
ca.
is said to be strongly elliptic in G, assuming that Cst is

on
Proof. We first prove that, for every e > 0, there is a constant e (e)>O
such that for every ego (G) function u,
(5)
To this purpose, we consider u to belong to ego (Rn), defining its value as
0 outside G. By the Fourier transformation we have
II (D/'-....u) II~= [ I;lJn yjiu (y) 12 dy,
I!D5 u II~=
in virtue of Plancherel's theorem. Thus (5) is a consequence from the fact

that
I(
(1~-1jft yj'i) e + ltrfmjft y;e;)(ls I = j~ s;, It I = j~ t;)

tends to zero uniformly with respect to y = (YI, y
Yn) as e too.
2 , ,
Suppose that the coefficients c51 are constant and vanish unless
Is I = It I = m. By the Fourier transformation u (x) --+ u(~) and Plancherel's theorem, we have, by (3),
ReB [u, u]
J 2:
c51 ~s~t lu(~) 1 d~
s,t
;;;:: J Co I~ 12 !u(~) 12 d~ > C1 (!I u 11;.- II u 11;.-1)'
=
Re
where c1 > 0 is a constant which is independent of u. Hence, by (5), we see

that (2) is true for our special case.
We next consider the case of variable coefficients c51 Suppose, first,
that the support of u is sufficiently small and contained, say, in a small
sphere about the origin. By the preceding case, we have, with a constant
c~ > 0 which is independent of u,
c~
llu II;.< ReB [u, u] + Re lsl=if=m J(cse(O)- Cse(x)) D u. D udx

- Re 2: JCse(x) D u. D1udx + e(e) llull~lsi+Tir<2m
5
If the support of u is so small that c51 has very small oscillation there,
we see that the second term on the right may be bounded by 2- 1 c~ II u 11;..
The third term on the right is bounded by constant times II u lim II u llm-1
Hence we find that, constants denoting positive constants,
2- 1 c~llull;.
<
ReB[u,
u]
+constant llullm
llu[lm-1 +
e(e) [lu[l~
177
9. Friedrich's Theorem
Thus, by
2 j,x I IP I < e Jcx /2 + e-1
e>
IP j2 which is valid for every

(6)
0,
we obtain !lull!< constant ReB [u, u] +constant llull!-1+ C(e)

ll~tll~. and so, by (5), we obtain (2).
Next we consider the general case. Construct a partition of unity in G:
N
1 = j~ wj,
Wj
E CC0 (G)
and
Wj
(x)
> 0 in G,
such that the support of each Wj may be taken as small as we please.

Then, by Leibniz' rule of differentiation of the product of functions,
Schwarz' inequality and the estimate of the case obtained above, we have
ReB [u, u]
Re ~
s,t
Jc, D'uD udx = Re ~ ~ Jwjc, D'uD udx

1
s,t
= Re ~ ~ c,1D'(wju) D1 (wju) dx
J
s,t
+ O(llu lim jJullm-1)
llwiu[J!-1) + O(llul[m llullm-1)

O(llullm llullm-1)
>~constant (llwjull!J
>constant !lull! +
(2) by (5). We remark that the constants c, C in (2)
obtain
Thus we
depend upon c0 , c81 and the domain G.
9. Friedrichs' Theorem
Let
(1)
~ D'c'1 (x) D1
ll,ltf:!Om
be strongly elliptic with real C00 coefficients c81 (x) in a bounded open domain
of R". For a given locally square integrable function l(x) in G, a locally
square integrable function u (x) in G is called a weak solution of
L =
(2)
Lu=f,
if we have
(u, L*cp) 0 = (/, cp) 0 , L* =II~
s ,!lf:;>m
(-1)11+1t1 D1 c81 (x)
n,
(3)
for every cp E ego (G). Here(/, g) 0 denotes the scalar product of the Hilbert
space L 2 (G). Thus a weak solution u of (2) is a solution in the sense of the
distribution. Concerning the differentiability of the weak solution u, we
have the following fundamental result:
Theorem (K. FRIEDRICHS [1]). Any weak solution u of (2) has square
integrable (distributional) derivatives of order up to (2m + p) in the
domain G1 ~ G where I has square integrable (distributional} derivatives
of order up top. In other words, any weak solution u of (2) belongs to
WP+ 2m(G1) whenever I belongs to JVi'(Gl).
178
Corollary. If p = oo, then, by Sobolev's lemma, there exists a function u0 (x)E C"'(G1 ) such that u(x) = u0 (x) for a.e. xE G1 Thus, after a
correction on a set of measure zero, any weak solution u (x) of (2) is CC"'
in the domain ~ G where I (x) is C00 ; the corrected solution is hence a
genuine solution of the differential equation (2) in the domain where l(x)
is C00
Remark. When L = Ll, the Laplacian, the above Corollary is Weyl's
Lemma (see Chapter II, 7). There is extensive literature concerning
the extensions of Weyl's Lemma to general elliptic operator L;
such extensions are sometimes called the Weyl-Schwartz theorem.
Among an abundant literature, we refer to the papers by P. LAx [2],
L. NIRENBERG [1] and L. NIRENBERG [2]. The proof below is due
to the present author (unpublished). A similar proof was given by L. BERS
[1]. It is to be noted that a non-differentiable, locally integrable function
I (x) is a distribution solution of the hyperbolic equation
of_= 0
oxoy
'
as may be seen from

0 = _[ {_[ l(x)
8 ~t~:) dy}dx
(<p(x,y)ECgo(R 2 )).
Proof of the Theorem. We shall be concerned only with real-valued

functions. Replacing, if necessary, L by I + rxL with a certain constant
rx =I= 0, we may assume that the strongly elliptic operator L itself satisfies
Garding's inequality
>
I(<p, L*VJ)o I <
(<p, L*<p) 0
b ff
0),
(4)
y [[ 0), whenever <p, ' E ego (G)
The latter inequality is proved by partial integration. We assume here
that each of the derivatives of the coefficients est (x) up to order m is
bounded in G, so that the constants b and y are independent of the
test functions VJ, <p E ego (G).
Suppose that G1 is a periodic parallelogram
0
(f = 1, 2, ... , n),
(5)
and that the coefficients of L and I periodic with the period 2n in each
Xj. Under such assumptions, we are to deal with functions <p (x) defined
on a compact space without boundary, then-dimensional torus G1 given
by (5), and the distribution E C00 (G 1)' associated with the space of test
functions <p E coo (G1) consisting of coo functions <p (x) = <p (x 1 , x2 , . . . , xn)
periodic with period 2n in each of the variables xj. It is to be noted that,
since G1 is without boundary, "'e need not restrict the supports of our
test functions <p(x).
179
The condition v E Wq (G1) thus means that the Fourier coefficients v,.
of v (x) defined by the Fourier expansion
v (x),....., .I v,. exp(ik x)
"
(k = (kv ... , kn), x = (xv ... , xn),
and
k x
/!i. kjxj)
6
( )
satisfy
(7)
For, by partial integration, the Fourier coefficients of the distributional
derivative nv satisfy
(D 5 v (x), exp (ik . x)) 0 = (-1)11 (v (x),
-- (- Z")11
IIn
J=l
n exp (ik x)) 0
kY
J v,.,
S --
(Sv
S2, , Sn ) ,
and so, by Parseval's relation for the Fourier coefficients of Dqv E L 2 (G1 ),
we obtain (7).
It is convenient to introduce the space Wq (G1 ) with integer q ~ 0, by
saying that a sequence {w,.; k = (kv k2 , , kn)} of complex numbers
belongs to Wq (G1) if (7) holds. This space Wq (G1) is normed by
with
!I{ w,.} llq = ( Iw,. 12 (1 + Ik I2 2 By virtue of the Parseval relation
w,.=w_,.
w,.
)T'
with respect to the complete orthonormal system {(2n)-nf2 exp (ik x)} of
ID5 v(x) l2 1' 2
L2(G1), we see that, whenq> 0, the norm !lvllq= ( ~
lsf~q G1
dx)
is equivalent to the norm ll{v,.} llq where v(x),....., .I v,. exp(ik x).
"
The above proof of (7) shows that, if IE WP (Gl). then n IE wP-11 (Gl)'
and q; f E WP (G1) for q; E coo (G1). Hence
iff E WP (G1), then, for any differential operator N of
(8)
order q with coo (Gl) coefficients, N IE wP-Iql (Gl).
To prove the Theorem for our periodic case, we first show that we may
assume that the weak solution uE L 2 (G1) = WO(G1) of (2) belongs to
Wm(G1 ). This is justified as follows. Let
u(x) ,....., .I u,. exp(ik x), v(x) ,....., .I uk(l
k
+ lk 2)-m exp(ik x).

1
Then it is easy to see that v (x) E W 2 m (G1) and that v is a weak solution
n
of (I-- L1)m v = u, where L1 denotes the Laplacian ..I 82J8xj. Hence vis
J=l
a weak solution of the strongly elliptic equation of order 4m:

L(I -L1)m v
=f.
(2')
actually belongs to
If we can show that this weak solution v E
W 4m+p (G1 ), then, by (8), u == (I- L1)m v belongs to W 4 m+P- 2m (G1) ==
W2m(G1)
12*
180
WP+2"'(G 1 ). Therefore, without losing the generality, we may assume

that the weak solution u of (2) belongs to W"'(G1), where m is half of
the order 2m of L.
Next, by Garding's inequality (4) for Land that for (I- A)m, we may
apply the Lax-Milgram theorem (Chapter III, 7) to the following effect.
The bilinear forms on coo (G1) :
(rp,1p)' = (rp, L*1p) 0 and (rp,1p)" = (rp, (I -L1)"'1p) 0
(9)
can both be extended to be continuous bilinear forms on W"' (G1) such

that there exist one-one bicontinuous linear mappings T', T" on W"'(G1)
onto wm (G1) satisfying the conditions
(T' rp, 1p)'
(rp, 1p),., (T" rp, 1p)"
(rp, 1p),. for rp, 1p E W"' (G1 ).
Therefore, there exists a one-one bicontinuous linear map T,. = T" (T')-1
on W"'(G1) onto wm(G1 ) such that
(rp, 1p)' = (T ,.rp, 1p)" whenever rp, 1p E wm (G1).
(10)
We can show that

for any
In fact, we have
i>
1, T,. maps W"'H (G1 ) onto wm+i (G1) in
a one-one and bicontinuous way.
(11)
(rp, L*(I -L1)1 "P)o = (T,.rp, (I -L1)"'H"P)o for rp, 1pE C00 (G1 ).
On the other hand, there exists, by the Lax-Milgram theorem as applied
to the strongly elliptic operators (I -A)i L and (I -A)"'+i, a one-one
bicontinuous linear map T m+i of wmH (G1 ) onto wmH (G1) such that
(rp, L *(I- L1) 1"P)o = (T m+i rp, (I- A)m+i "P)o for rp, 1p E C00 (G1).
Therefore, the function w = (TmH- T ,.) rp is, for any rp E C00 (G1), a
weak solution of (I -A)m+i w = 0. But, such a w(x) is identically zero.
For, the Fourier coefficient w,. of w (x) satisfies
0 =((I -L1)"'+i w(x), exp(ik x)) 0 = (w(x), (I -A)"'+i exp(ik x)) 0
=
(1
+ Jk J2)"'+i (w(x), exp(ik x))0 =
(1
+ Jk J2)m+j w,.,
and so w,. = 0 for all k. Thus (Tm+i- T ,.) is 0 on C00 (G1 ). The space
C00 ( G1 ) is dense in W"'+i (G1) ~ wm (G1), since trigonometric polynomials
~ w,. exp (ik x) are dense in the space wm+i (G1). Hence T m+i = T,.
ikJ<oo
on W"'+i (G1).
We are now ready to prove the differentiability theorem for our
periodic case. We have, for 1pE C00 (G1 ),
(f, 1p)0 = (u, L *1p) 0 = (u, 1p)' = (T,.u, 1p)" = .(T,.u, (I- L1)"'1p) 0
181
Hence, for
T mU
"'~ ck
k
exp (ik x), tp (x) ,_,
~ 'IJ'k
k
exp (ik x),
we obtain, by Parseval's relation,

fk'iiik
+ lk l2)m ~k = ~
k
the arbitrariness of tp E C"" (G1), we have ck (1 + Ik l2)m = fk, and so,
(T mu, (I -Ll)m 'IJ')o = ~ ck(1
k
By
by fEWP(G 1 ), we must have TmuEWP+Zm(G1 ). Hence, by (11),
uE WP+Zm (G1 ) .It is to be noted that the above conclusion uE WP+Zm (G1)
is true even in the case 0 > p > (1-m), i.e., the case {ik} E WP (G)
with 0 > p > (1-m). For, by p + 2m> m + 1, we can apply (11).
We finally shall prove, for the general non-periodic case, our differentiability theorem. The following argument is due to P. LAx [2].
We want to prove, for the general non-periodic case, the differentiability theorem in a vicinity of a point x0 of G. Let {3 (x) E C~ (G) be identically one in a vicinity of x 0 . Denote {3u by u'. u' is a weak solution of
Lu' = {31
+ Nu,
(12)
where N is a differential operator of order at most equal to (2m- 1)

whose coefficients are, together with {3 (x), zero outside some vicinity V
of x 0 , and the operator N is to be applied in the sense of the distribution.
We denote the distribution {3 f + N u by f'.
Let the periodic parallelogram G1 contain V, and imagine the coefficients of L so altered inside G1 but outside V that they become periodic
without losing their differentiability and ellipticity properties. Denote
the so altered L by L'. Thus u' is a weak solution in G of
L'u' = f', where f' = {3/ + Nu.
(13)
We can thus apply the result obtained above for the periodic case to
our weak solution u'. We may assume that the weak solution u' belongs
to wm (G1 ). Thus, since N is of order < (2m -1) and with coefficients
vanishing ontside V, f' ={31 + Nu must satisfy, by (8),
f'E WP' (G1) with p' = min(p, m- (2m-1)) = min(p, 1-m);:;:::: 1-m.
Therefore, the weak solution u' of (13) must satisfy
u' E WP" (G1) with p" =min (p + 2m, 1-m + 2m)
= min(p +2m, m + 1).

Hence, in a certain vicinity of x 0 , u has square integrable distributional
derivatives up to order p" which is ? (m + 1).Thus f' = {31 + Nu
has, in a vicinity of x 0 , square integrable distributional derivatives of
order up to
p"' = min(p, p"- (2m- 1)) 2:": min (p, 2-m).
182
Thus again applying the result already obtained, we see that u' has, in a
certain vicinity of x0 , square integrable distributional derivatives up to
order
p( 4) = min(p +2m, 2-m+ 2m)= min(p +2m,
m-+
Repeating the process, we see that u has, in a vicinity of

integrable distributional derivatives up to order p + 2m.
2).
x0 ,
square
10. The Malgrange-Ehrenpreis Theorem

There is a striking difference between ordinary differential equations
and partial differential equations. A classical result of PEANO states
that the ordinary differential equation dyfdx = f(x, y) has a solution
under a single condition of the continuity of the function f. This result
has also been extended to equations of higher order or to systems of
equations. However, for partial differential equations, the situation is
entirely different. H. LEWY [1] constructed in 1957 the equation
. au
-z
axl
au
. ) -au =
2 (x 1 + 1x
+-2 axa
ax2
I( x3) ,
which has no solution at all iff is not analytic, even iff is C00 Lewy's
example led L. H6RMANDER [3] to develop a systematic method of
constructing linear partial differential equations without solutions. Thus
it is important to single out classes of linear partial differential equations
with solutions.
Let P (.;) be a polynomial in .;I> .;2 , , .;.. , and let P (D) be the linear
differential operator obtained by replacing .;j by Dj = -1 8/oxj. P (D) may
be written as
P (D) =
DIX
.I
m~(1X)~O
c!XD,., where, for
<X
= (<X 1 , <X 2 ,
. ,
<X.,) with 1 :s;;
<Xj
< n,
=.II DIX1, and D0 =I and, moreover, (<X)= k.

J=l
Definition 1. By a fundamental solution of P(D) we mean a distribution E in R" such that

P(D) E = <5 = T~.
The importance of the notion of the fundamental solution is due to the

fact that
u = E f, where f E ego (R") ,
gives a solution of the equation
P(D) u =f.
In fact, the differentiation rule (10) in Chapter VI, 3 implies that
p (D) u = (P (D) E) f = <5 *I = f.
183
Example. Let P (D) be the Laplacian L1
Then the distribution
'i; ax,a: in R" with n > 3.
j= 1
E = Tg, where g (x) = ( 2 _ n) S,. Ix 12 - " and S,. = the area of the
surface of the unit sphere of R",
is a fundamental solution of Ll.
Proof. We have, in the polar coordinates, dx = lxl"- 1 d
so the function g(x) is locally integrable in R". Hence
L1 r,~,-n (I]J) = lim
e.j.O 1~1;:;:;
I X 12 -n
lxl dS,.,
and
LlqJ dx' IP E ;1) (R") .
Let us take two positive numbers e and R (> e) such that the supp (I]J)
is contained in the interior of the sphere IxI < R. Consider the domain G:
e < Jx I< R of R" and apply Green's integral theore~, obtaining
J(lx 12-n.
LlqJ- L1
lx 12-n qJ) dx =
J(lx 12-n. ~~- al~,_,.
qJ) dS
where S = 8G is the boundary surfaces given by !xl = e and lxl = R,

and v denotes the outwards normal to S. Since qJ vanishes around IxI = R,
2 -" = 0 for
=I= 0 and thatwe have, remembering that L1 lx 1
= e,
[xI
at the points of the inner boundary surface
f Jxl
2- "
LlqJ dx=- ( e2 -n aa~ dS
Rn
l~i=e
I I
+ (
I~J=
:v = al: I
(2- n) e1-n 1J! dS.
When e t 0, the expression 81]J/8Ix I = ..I (xjflx I) 81]!/8xj is bounded

J=1
and the area of the boundary surface IxI = e is S,.e"- 1 . Consequently
the first term on the right tends to zero as e t 0. By the continuity of IP
and a similar argument to above, the second term on the right tends, as
e t 0, to (2- n) S,. IP (0). Thus Tg is a fundamental solution of Ll.
The existence of a fundamental solution for every linear partial differential equation with constant coefficients was proved independently by
B. MALGRANGE [1]andL. EHRENPREIS [1]around 1954--55. The exposition
of the result given below is due to L. HoRMANDER [4].
Definition 2. Set
(1)
184
We say that a diffuential operator Q(D) with constant coefficients is

weaker than P (D) if
Q(~) < CP(~), ~ERn, C being a positive constant.

(2)
Theorem 1. If Q is a bounded domain of Rn and IE L2 (!J), then there
exists a solution u of P (D) u = I in !J such that Q (D) u E L2 (!J) for all Q
weaker than P. Here the differential operators P (D) and Q(D) are to be
applied in the sense of the theory of distributions.
The proof is based on
Theorem 2. For every s > 0, there is a fundamental solution E of
P(D) such that, with a constant C independent of u,
I(E * u) (0) I< C sup

lu(~ + i1)) 1/P(~) d~, u E C(f(Rn).
1'11;:;; Rn
Here is the Fourier-Laplace transform of u:
u(C) =
(2n)-nf2
J e-i(x,o u (x) dx,
Rn
C= ~
(3)
+ i TJ,
and the finiteness of the right side of (3) is assured by the Paley-Wiener
theorem in Chapter VI, 4.
Deduction of Theorem 1 from Theorem 2. Replace u in (3) by
Q(D) tf, * v, where u and v are in C(f (Rn). Then, by (10) in Chapter VI, 3,
I(Q(D) E *
u * V) (0) I= I(E * Q(D) u * v) (0) < CN(Q(D)

J
tJ
* v),
sup
lit(~ + i 1)) J/P (~) d~.
1'11;:;; Rn
The Fourier-Laplace transform of Q(D) u *vis, by (17) in Chapter VI, 2
and (15) in Chapter VI, 3, equal to (2nt' 2 Q(C) (C) (C). Since, by
Taylor's formula,
where
N (u)
u v
Q(~ + i1)) =~(<X)! (-n)"' D"'Q(~). where (-1))"'= I.I (-1)"';), (4)
we have, by (2),
"'
/Q(~+i1))//P(~)<C' for /TJJ<s and ~ERn,
where the constant C' may depend on s. Thus
N(Q(D) ti* v) < (2nt12 C'. sup
J Iu(~ +
j1)j;;;e Rn
i1))
v(~ +
i1)) I d~.
By Parseval's theorem for the Fourier transform, we obtain, denoting

by 1/ /I the L2(Rn)-norm,
J lu(~
Rn
+ i1)) 12 d~ =
flu (x) 12 e2 <x,1J) dx :::=: Jlu(x) elxiiJz when 111 < e,

J
Rn
v. Thus, by Schwarz' inequality,

v) < C" II u (x) elxl II 1/ v (x) elxl Jl whenever u and vE C(f (Rn),
and a similar estimate for

N ( Q(D) u *
C" denoting a constant which may depend on s.
185
10. The J.\llalgrange-Ehrenpreis Theorem
J(Q(D) E *
u) (x) v (-x) dxll ~
(e e") II uelxlllllvelxlll (u, vE e:;o (R")).

(fi)
Rn
We shall write L; (Rn) for the Hilbert space of functions w (x) normed by
(i
lw(x) 12 elxl dxt2 = llw(x)eelxlll
Since ego (R") is dense in L; (Rn) and, as may be proved easily, L':.. (Rn)
is the conjugate space of L; (R,.), we obtain, by dividing (5) by !!v (x) elxfll
and taking the supremum over v E ego (Rn),
II (Q (D) E * u) (x) e-efxlll < (ee") llu (x) e*'JI,
u E ego (Rn).
Hence the mapping

u~
(6)
Q(D) E*u
can be extended by continuity from ego (Rn) to L; (Rn), so that it becomes

a continuous linear mapping on L;(Rn) into L':..(R,.). Thus, to prove
Theorem 1, we have to take 11 =I in.Q and 11 = 0 in Rn -Q and define
u as equal to u = E 11 .
For the proof of Theorem 2, we prepare three Lemmas.
Lemma 1 (MALGRANGE). If I (z) is a holomorphic function of a complex

variable z for Iz I < 1 and p (z) is a polynomial in which the coefficient
of the highest order term is A, then
,.
!A 1(0) I< (2n)-1 _,.J !l(e6 ) p(e6 ) IdO.
(7)
Proof. Let z/s be the zeros of p (z) in the unit circle !z I < 1 and put
p(z)
q(z)
JI -z:z
z,l
Then q(z) is regularin the unit circle and !P(z) I= !q(z) I for lz I = 1. Hence
we have
,.
(2n)-1
-n
ll(e16)
p(e6 )
I dO=
,.
(2n)-1
J !l(e6) q(e'6) I dO
-n
> (2n)-I[_l l(e' 6 ) q(e 6 ) dO
I=
!1(0) q(O) I
Thus Lemma 1 is proved since Jq(O)/A I is equal to the product of the

absolute values of zeros of p (z) outside the unit circle.
Lemma 2. With the notations in Lemma 1, we have, if the degree of
p(z) is < m,
!1(0) p<kl (0) I< (m m!k)T (2n)-1
_[
ll(e"1 p(e'6 ) I dO.
(8)
186
Proof. We may assume that the degree of
p(z) ism and that
p (z) = .II (z- zj).

J=l
Applying the preceding Lemma 1 to the polynomial

m
.I/ (z-zi)
J=l
and the
holomorphic function I (z) . II (z- zj), we obtain

J=k+l
11(0)
i=R/i~~ (2n)-1 _[
ll(eiB) p(eiB) I d().
A similar inequality will hold for any product of (m- k) of the numbers
zion the left hand side. Since p(k) (0) is the sum of m!f(m-k)! such terms,
multiplied by (-1)m-k, we have proved the inequality (8).
Lemma 3. Let F (C) = F {~1 , ~2 , . . , ~n) be holomorphic for ICI =
Cli ~j r C2, ... ' Cn)
a polynomial of degree
< m. Let f/J (C) = f/J (~v C2 , , Cn) be a non-negative integrable function
with compact support depending only on 1~1 1, IC21, ... , ICnl Then we
have
IF(O) D"P(O) I
ICI("')f/>(C)d~< (m m<~IT'
ICI<oo
. ICI<oo
IF(3) P(C) I f/>(C) dC,
(9)
where d' is the Lebesgue measure d~1 dr11 d~n dr;n (Ck = ~k + i 'YJk).
Proof. Let I (z) be an entire holomorphic function and apply {8) to the
functions I (r z) and p (r z), where r > 0. Then we obtain
11(0) p(k)(O) I. rk <
(m m!k)f
{2n)-ll ll(reiB) p(reiB) I d().
Let 1p (r) be a non-negative, integrable function with compact support.

Multiplying the above inequality by 2n r 1p(r) and integrating with
respect to r, we obtain
11 (o) p(k) (O) 1f 1t lk 1p ( 1t!) dt ~ (m m 1kff f 11 (t) p (t) 11p ( 1t 1)
dt
{10)
where dt = rdrd() and the integration is extended over the whole complex
t-plane. Lemma 3 is obtained by applying {10) successively to the variables Cl> C2 , , Cn, one at a time.
Proof of Theorem2. Put P(D)u=v, where uECif(Rn). Then
P (C) (C) = (C). Apply Lemma 3 with F (C) = (~ + C), with P (C)
replaced by P (~ + C) and with If/> (C) I = 1 when ICI < e and = 0 otherwise. Since P(~) <.I ID" P{$) I, we obtain, from (9),
"'
u(~).P(;)I<cl
JCI ;:;;;.
lu(~+C)P($+C)IdC=cl
ICI ;:;;;.
lv(~+~)lac.
187
Hence, by Fourier's integral theorem,

lu(O) 1< (2n}-"12
e~ f
J lu(~) 1d~ <
( f
~'+'l'':;;e
e;_
J (J lv(~
ICI;>;
+ C) 1/P(~) d~) d~
lti (~ + ~~ + i 17') 11.P (~) d~' d17') d~.
On the other hand, we have

because
D"' P(~
+ ~') =
f \~~~ D"'+P P(~).
so that J D"' P (~ + ~') I/P (~) is bounded when

lu(O)I<e{e2f(
E''+'l'';>i;'
If
S e. Therefore we have
lv(~ + t + ifJ')I/P(~+~')~'dfJ')d~
(11)
< eallv II', where

llvll'
I'll;:;;;;.
(f lv(~ + i7J} J/P(~) d~) dfJ
(uE eij(R,.)),
e 3 denoting a constant depending only on e.

By the way, the finiteness of II vII' is a consequnece of the Paley-Wiener
theorem in Chapter VI, 4. Consider the space Cgo(R") which is the
completion of ego (Rn) with respect to the norm II vII' Then, by the HahnBanach extension theorem, the linear functional L:
P (D) u ~ u (0)
(where u E ego (R,.))
can be extended to a continuous linear functional L defined on Co0 0 (R").

As in the case of the space L 1 (R,.), we see that there exists a Baire function k(~ + i17) bounded a.e. with respect to the measure P(~)-1 d~d'I'J
such that the extended linear functional Lis represented as
(12)
f (J v(~ + i17) k(~ + ifJ}/F(~) d~) dfJ.
1'11;:;0
When v11 (x} E ego (Rn) tends, as h ~ oo, to 0 in the topology of i) (R"),
v11 (x) e<x.'l > also tends to zero in the topology of i) (R"), uniformly with
respect to 17 for 117 I s e. Hence, as in Chapter VI, 1, we see easily that
11 (~ + i 17), as function of ~, tends to zero in the topology of 6 (R"),
uniformly in 17 for 117 I< e. Therefore, by (12), L defines a distribution
TE i)(R")'. Thus, by (5) in Chapter VI, 3.
L(v) =
L (v)
(T
* v) (0) =
(f
We have thus proved Theorem 2 by taking E
* v) (0).
=f. (3) is clear
(13)
from (11).
188
11. Differential Operators with Uniform Strength

The existence theorem of the preceding section may be extended to a
linear differential operator
P(x,
D)=~
IX
(1)
a.,(x) D"
whose coefficients a., (x) are continuous in an open bounded domain Q of

Rn.
Definition. P (x, D) is said to be of uniform strength in Q, if
(2)
where P(x, $)is defined by (
f IP'"') (x, $) I Y'

2
considering x as para-
meters.
Examples. The differential operator P(x, D)=
Dsas 1 (x) D1
'
with real, bounded C"" coefficients as,t (x) = a1,s (x) in Q is strongly elliptic
in Q (see Chapter VI, 9) if there exists a positive constant <5 such that
L
JsJ,Tif=m
$sas,t(X)
$>
1
<5 (
JsJ,JtJ~m
.i $J)m in Q.
V=l
(3)
In such a case, P(x, D) satisfies the condition (2). Next let P(x, D) be
strongly elliptic in an open bounded domain Q of Rn-t. Then
8
(4)
ex,. -P(x, D)
is said to be parabolic in the product space Qx{xn; 0 < xn} It is easy

to see that the operator (4) is of uniform strength in the above product
space.
Theorem (H6RMANDER [5]). Let P(x, D) be of uniform strength in an
open bounded domain Q of Rn. For any point x 0 , there exists an open
subdomain Q 1 of Q such that x 0 E Q 1 and the equation P (x, D) u = f has,
for every f E L2 (Q1), a distribution solution u E L2 (Q1) for which, moreover, Q(D) u E L2 (Q1) for every Q(D) weaker than P (x, D) for any fixed
xEfl1 .
Proof. Write P(xO, D)= P0 (D). The set of all the differential operators with constant coefficients weaker than P 0 (D) is a finite dimensional
linear space. For, the degree of such operators Q(D) cannot exceed that
of P0 (D). Thus there exist P 1 (D), P 2 (D), ... , PN(D) which form a basis
for the differential operators weaker than .?0 (D). Hence we can write
P(x, D)= P0 (D)
+ J=l
.~ bj(x) Pi(D),
bi(xO) = 0,
with uniquely determined bj (x) which are continuous in Q.
(5)
12. The Hypoellipticity (Hormander's Theorem)
189
By the result of the preceding section, there exists a bounded linear

operator T on L2 {Q1) into L2 (Q1) such that
=I for all /E L2{Q1),
P0 (D) Tf
(6)
and that the operators Pi (D) T are all bounded as operators on L2 (Q1)
into L2(Q1). Here!J1 is any open subdomain ofQ. We have only to take
Tf as the restriction to Q 1 of E
1 where / 1 =I in Q 1 and / 1 = 0 in
Rn -Q1.
The equation P(x, D) u = f is equivalent to
*/
(7)
We shall seek a solution of the form u

obtain, by (6),
= Tv.
Substituting this in (7), we
.~ bi(x) Pj(D) Tv= f.

+ J=l
(8)
Let the sum of the norms of the bounded linear operators Pj (D) T on
L2 (Q1 ) into L2 {Q1) be denoted by C. Since bj (x) is continuous and
bj(x0 ) = 0, we may choose Q 1 3 x 0 so small that
C jbj(x)
I< 1/N whenever
xE!J1 (j
1, 2, ... , n).
We may assume that the above inequalities hold whenever x belongs

to the compact closure of Q 1 . Thus the norm of the operator
N
.~
J=l
bj (x) Pj (D) T is less than 1, and so the equation {8) may be solved
by Neumann's series (Theorem 2 in Chapter II, 1):

v
= (1
+.I bjPj(D) T)- ! = Af,

1
;=1
where A is a bouned linear operator on L2 (Q1 ) into L 2 (Q1). Hence

u = TAf is the required solution of P(x, D) u =f.

We have defined in Chapter II, 7 the notion of hypoellipticity of
P(D) and proved Hormander's theorem to the effect that, if P (D) is
hypoelliptic, then there exists, for any large positive constant C1 , a
positive constant c2 such that, for all solutions = ~ irJ of the algebraic equation P (C)= 0,
(1)
c +
To prove conversely that (1) implies the hypoellipticity of P (D),

we prepare the
Lemma (H6RMANDER [1]), (1) implies that
1 J 0 jP<'">(~)I 2 / 1 ,.~0 !P''">(~)I 2 -+0 as
~ERn, !~j-+oo.
(2)
190
Proof. We first show that, for any real vector g E R", we have
+ 0){ P (;) ~ 1 as ;E R", j;j ~ oo.

(3)
Wemayassumethatthecoordinatesaresochosen that e = (1, 0, 0, ... , 0).
P(;
We have, by (1),
+ i'Y}) :f= 0 when
I'YJI < c1 and [;I > c2.

Then the inequality j;- C' I > c1 holds i" j;j > c1 + c2 and p (C') = 0.
For, setting C' = ;' + i'Yj', we have either I'YJ'I:;::: c1 or else WI < c2 so
that j;- ;'j > C1 . Giving fixed values to, ; 2 , ; 3 , ... , ;n we can write
P(;
P(;) = C II
k=1
(~ 1 - tk),
C :f= 0,
where (tk, ; 2 , ... , ;n) is a zero of P. Hence we have jtk- ; 1 j

!;j > C1 C2 Thus
~(~+ 6>) = rr~1 +.1-tk =

P(~)
k=1
~1- tk
rr(l +
k=1
> C1 if
_1 ~)
~1- tk
satisfies
I p~~) O)
-11 < mC1 1(1 + C1 1)m- 1 if [;I> c1 + C2.
As we may take C1 arbitrarily large by taking C2 sufficiently large, we

have proved (3).
We have, by Taylor's formula,
P(~ + 'YJ) =~(<X~! p<cx> (~)
'YJ"', p<cx> (~) =
(i)<"'>D,.P(~),
!1
(cx)
D,. =
D,.1
and so
(4)
where 'YJ(i) are arbitrary real vectors and t; arbitrary complex numbers.
k
The coefficients ..:1: t; ( 'YJ(il)"', (.x) < m, can be given arbitrary values,
=1
which are symmetric in .x, by a convenient choice of k, t; and 'YJ(il. If

otherwise, there would exist constants C"', (.x) < m, symmetric in <X and
not all equal to zero and such that .:1: Cex 'YJ"' = 0 for every 'YJ. Thus
k
p(cx) (~) = ..J: t;P (~

=1
+ 'YJ(i))
(cx)
with real vectors 'YJ(i).
Since the principal part on the right must cancel out when (.x) :f= 0, we
k
must have ..J: f;

=1
0. Hence, by (3), we obtain (2).
Corollary. Suppose that P 1 (~) and P 2 (;) satisfy (2). Then P (;) =
P 1 (;) P 2 (;) also satisfies (2). Moreover, if Qj (D) is weaker than Pj (D)
(f = 1, 2), it follows that Q1 (D) Q2 (D) is weaker than P (D).
191
Proof. By applying Leibniz' formula of differentiation of product of

functions, we see that p<cxJ (.;) is a linear combination of products of
derivatives of P 1 (.;) and P 2 (.;) of the order sum < (<X). Hence (2) holds
for P(.;). The latter part of the Corollary may be proved similarly.
We are now ready to prove
Theorem (H6RMANDER [1]). P(D) is hypoelliptic iff the condition (2)
is satisfied.
Proof. The "only if" part is already proved (Chapter II, 7 and what
was proved above). We shall prove the "if" part.
Let D be an open subdomain of Rn. A distribution u E ;tl (D)' is said
to belong to Hfoc (D) if, for any q;0 E C;) (D), the Fourier transform 0
of u0 = q;0 u satisfies (see Chapter VI, 2)
j(1+
Rn
lc;l 2)klu0 (.;)1 2 d.;<oo,
that is, if u 0 =q;0 uEWk 2 (Rn).
(5)
By virtue of Sobolev's lemma in Chapter VI, 7, the "if" part is

proved by the following proposition:
Let P(.;) satisfy (2). If a distribution u E ;tl (.Q)' satisfies
P (D) u E Hfoc (D) with a positive s, then u belongs to
(6)
fltoc(D)
For, if P(D) uE C00 in D, then P(D) uE Hfoc(D) for every positive s
because of Leibniz' formula of differentiation.
The following proof of (6) is based upon two Lemmas:
Lemma 1. Let f E W 2 (Rn) and 'IJ1 E CQ' (Rn), s
0. Then 'IJ1 f E W 2 (Rn).
Lemma 2. Let P(~) satisfy (2). Then there exists a positive constant
fl such that Ip(cx) (.;) .;"'I/ IP (.;) I ~ 0 as .; E Rn, 1.; I~ oo for every <X =I= 0.
The proof of Lemma 1 will be given later, and the proof of Lemma 2
will not be given here (for the latter, we refer the reader to L. H6RMANDER [6] or to A. FRIEDMAN [1]).
Next, letD1 andD0 be arbitrary open subdomains ofD such that their
closures D~ and D 0 are compact and D~ ~ D 0 , .Q0 ~D. By Schwartz'
Theorem in Chapter III, 11, the distribution u E ;tl (D)' is, when considered as a distribution E ;tl (.Q0 )', a distributional derivative of the form
D 1v of a function v(x) E L2(D0 ). Let q; f Cg"(D0) be such that q;(x) = 1
in D1 . Then u = u 0 = D 1q;v as distributions E ;tl (D1) '. q; v being E L2 (Rn),
we see that there exists a (possibly) negative integer k such that
(7)
p(cx) (D) Uo = p(cx) (D) D1q;v E wk,Z (Rn) for every <X.
Hence, by Lemma 1 and the generalized Leibniz formula (see Chapter l, 8)

P(D) q;1 u 0 = q;1 P(D) u 0
1) Da(/Jl p(a)(D) u
L -(
(<><)>0
<X
0,
(8)
192
we see by p (D) Uo E Hfoc (Q) that, whenever cpl E ego (.Ql),
P(D) cp1 u 0 E Wk,, 2 (Rn) with k1 = min(s, k).
(9)
Thus the Fourier transform ul(~) of ul(x) = cpl(x) Uo(x) satisfies
RR
P (~) ul (~) 12 (1
~ l2)k, a~ < oo
(10)
and hence, by Lemma 2,
J IP(a:) (~) ul (~) 12 (1 +
Rn
p(a:)
I~ l2)k,+p d~
<
oo, that is,
(11)
(D) u1E Wk,+p,Z (Rn) for every IX# 0.
Let .Q2 be any open subdomain of .Q1 such that its closure~ is compact
and contained in.Q1 Then, for anycp 2 E ego(.Q2), we prove, by (8) and (11)
as above, that
P (D) cp 2 u 1 E Wk 2 (Rn) with k2 =min (s, k1 + p) and hence
P("') (D) cp2u 1 E Wk,+p,Z (Rn) for every IX# 0.
Repeating the argument a finite number of times, we see that, for any
open subdomain .Q' of .Q such that its closure is compact and contained
in .Q,
p(a:) (D) cpu E ws,Z (Rn) for all IX # 0 whenever cp E ego (Q').
Thus,
p(a:)
=constant =l= 0 gives cpu E ws,z (Rn).
(~)
Proof of Lemma 1. The Fourier transform of "PI is

(2n)-nf 2
J fJ ('Y/) 1(~- 'YJ) d'YJ (see Theorem 6 in Chapter VI, 3)
RR
and thus we have to show that, for s
Rn
(1
0,
+ I~ 12f IRnf .;p (nll (~- n) dnjz d ~ < oo.
By Schwarz' inequality, this can be estimated from above by
R[ (1 + 1~\ 2) 8 [J \.fP(n) I dn R[ l.fP(n) 11/(~- n) l2dn] a~
l.fP(n) I dn
[i i
(1
+ 1~1 2 ) 8 l.fP(n) I 11 (~- n) 12 a~ dn}
We then make use of the inequality

(1 + ~~ 12)s < 4lsl (1 + I'YJ !2)lsl (1
which may be proved by
1
+ I; Ia
+ l$-nla<
4(1
lnl2), 4(1
~~
1~12) >
By (13), the right side of (12) is estimated by
_ 'YJ 12)'
1
+ I;- nIa
1
+ lnla .
J I~ ('Y/) Idn-times
Rn
(12)
(13)
193
1. Dual Operators
This integral converges since /E W 8 2 (Rn) and .fP('YJ) E 6(Rn).

We have thus proved our Theorem.
Further Researches
1. A linear partial differential operator P(x, D) with C00 (Q) coefficients is said to be formally hypoelliptic in Q ~ Rn if the following two
conditions are satisfied: i) P(x0 , D) is hypoelliptic for every fixed x0 EQ
and ii) P(x0 ,~) = O(P(x',~)) as ~ERn, 1~1-+= for every fixed x0 and
x' EQ. L. HoRMANDER [5] and B. MALGRANGE [2] have proved that,
for such an operator P (x, D), any distribution solution u E 'Il (Q)' of the
equation P(x, D) u =I is coo after correction on a set of measure zero
in the open subdomain ~ Q where I is C00 The proof above for the constant coefficients case may be modified so as to apply for the formally
hypoelliptic case, see. e.g., J. PEETRE [1].
2. It was proved essentially by I. G. PETROWSKY [1] that all distribution solutions u E 'Il (Rn) of P(D) u = 0 are analytic functions in Rn
of the highest degree m does not
iff the homogeneous part p m (~) of p
vanish for ~ E Rn. If this condition is satisfied then P (D) is said to be
(analytically) elliptic. It is proved that in such a case the degree m is
even and P (D) is hypoelliptic. It is to be noted that the hypoellipticity
of an analytically elliptic operator P (D) can also be proved by Friedrichs'
Theorem in Chapter VI, 9. For, by the non-vanishing of P m (~), we easily
see, by the Fourier transformation, that P (D) or -P (D) is strongly
elliptic. For the proof of Petrowsky's Theorem, see, e.g., L. HoRMANDER
[6], F. TRiNES [1] and C. B. MORREY-L. NIRENBERG [1].
VII. Dual Operators

1. Dual Operators
The notion of the transposed matrix may be extended to the notion of
dual operator through
Theorem 1. Let X, Y be locally convex linear topological spaces and

x;, Y; their strong dual spaces, respectively. Let T be a linear operator on
D (T) ~X into Y. Consider the points {x', y'} of the product space X~ X Y~
satisfying the condition
<T x, y') = <x, x')
for all
xED (T).
(1)
Then x' is determined uniquely by y' iff D (T) is dense in X.

Proof. By the linearity of the problem, we have to consider the condition:
<x, x') = 0 for all xED (T) implies x' = 0.
13
194
VII. Dual Operators
Thus the "if" part is clear from the continuity of the linear functional
x'. Assume that D (Tt =F X. Then there exists, by the Hahn-Banach
theorem, an x~ =F 0 such that <x, x~) = 0 for all xED (T); consequently,
the "only if" part is proved.
Definition 1. A linear operator T' such that T' y' = x' is defined
through (1) iff D (Tt =X. T' is called the dual or conjugate operator of
T; its domain D (T') is the totality of those y' E Y; such that there exists
x' E x; satisfying (1), and T' y' = x'. Hence T' is a linear operator defined
on D (T') ~ Y; into x; such that
<T x, y')
(x, T' y') for all xED (T)
y' E D (T').
and all
(2)
Theorem 2. If D (T) = X and T is continuous, then T' is a continuous linear operator defined on Y~ into X~.
Proof. For any y' E Y;, <Tx, y') is a continuous linear functional of
x E X and so there exists an x' E x; with T' y' = x'. Let B be a bounded
set of X. Then, by the continuity ofT, the image T B = {Tx; xE B}
is a bounded set of Y. Thus, by the defining relation <Tx, y') = <x, x'),
the convergence to 0 of y' in the bounded convergence topology of Y'
(given in Chapter IV, 7) implies the convergence of x' in the bounded
convergence topology of X'. Thus T' is a continuous linear operator on
Y~ into X~.
Example 1. Let X = Y be n-dimensional euclidean spaces normed by
the (l2)-norm. For any continuous linear operator TEL (X, X), set
Tx = y,
where
(x 1 , x2 ,
.. ,
xn)
and
= (YI,
y2 ,
.. ,
Then Yi = .2:1 t;jXj (i = 1, 2, ... , n) and so, for z = (zv z2 ,

)=
<Tx, z)
Thus T' z
= <y, z) =
~
J
yjzj =
w is given by wi
~ (~
>
t;jXj) Z; =
~
J
Xj
.. ,
(~
'
Yn)
zn),
t;jZ;) .
= .2: tiizi (j = 1, 2, ... , n). This proves

=1
that the matrix corresponding to T' is the transposed matrix of the matrix
corresponding to T.
Example 2. Let X = Y be the real Hilbert space (l 2 ), and let
Tn E L (X, X) be defined by
Then from
<Tn(X1,
we obtain
X2, . ),
(zv Z2, .. .)) = XnZ1

n-1
+ Xn+1Z2 + Xn+2Z3 + ,
195
2. Adjoint Operators
Proposition 1. The mapping T-+ T' of L (X, Y) into L (Y;, x;) is

not, in general, continuous in the simple convergence topology of operators, that is, lim Tnx = Tx for all xE X does not necessarily imply
~
lim T~y' = T' y' for ally' E Y', in the strong dual topology of
n->00
x;.
Theorem 2'. Let T be a bounded linear operator on a normed linear

space X into a normed linear space Y. Then the dual operator T' is a
bounded linear operator on
y; into x; such that
(3)
IITI/=IIT'll
Proof. From the defining relation <Tx, y') = <x, x'), we obtain
liT'y'JI = llx'JI =sup J<x,x')l

llxll:51
=sup J<Tx,y')l<lly'll sup IITxJI<IiY'IJIITJI,

l!xll :51
I! xi! :51
and so II T' II < II T II The reverse inequality is proved as follows. For
any x0 E X, there exists an / 0 E Y' such that II /0 II = 1 and / 0 (T x0 ) =
<Tx0 , /0 ) = !1Tx0 JI Thus/~= T'/0 satisfies <x0 ,/~) = JJTx0 IJ and so
IITxoll = <xo, T'fo> < JIT' IJJifoiJIJxoJI = JIT'JIIIxoll, i.e.,
IITIJ< !IT'll
Theorem 3. i) If T and 5 are E L(X, Y), then (exT+ {35)' =
(exT'+ {35'). ii) Let linear operators T, 5 be such that D(T), D(5),
R (T) and R (5) are all contained in X. If 5 is E L (X, X) and D (Tt = X,
then
(4)
(5T)' = T' 5'.
4
If, moreover, D (T 5) =X, then
(5)
(T 5)' ~ 5' T', i.e., (T 5)' is an extension of 5' T'.
Proof. i) is clear. ii) D (5 T) = D (T) is dense in X, and so (5 T)'
exists. If yE D((5T)'), then, for any xE D(T) = D(5T), <Tx, 5'y) =
<5Tx,y)=<x,(5T)'y). This shows that 5'yED(T') and T'5'y=
(5 T)' y, that is, (5T)' ~ T' 5'. Let,conversely,y ED (T'5'),i.e.,5' yED (T').
Then, for any xE D(T) = D(5T), <5Tx, y) = <Tx, 5'y) = <x, T' 5'y).
This shows that y ED ((5 T)') and (5 T)' y = T' 5' y, that is, T' 5' ~ (5 T)'.
We have thus proved (4).
To prove (5), let y ED (5' T') = D (T'). Then, for any xED (T 5),
<T5x,y)=<5x,T'y)=<x,5'T'y). This shows that yED((T5)')
and (T5)'y = 5'T'y, that is, 5'T' ~ (T5)'.
2. Adjoint Operators
The notion of transposed confugate matrix may be extended to the
notion of adjoint operator through
13*
196
VII. Dual Operators
Definition I. Let X, Y be Hilbert spaces, and T a linear operator

defined on D (T) ~ X into Y. Let D (Tt = X and let T' be the dual
operator ofT. Thus (Tx,y') = (x, T'y') for xED(T), y'ED(T'). If
we denote by]x the one-to-one norm-preserving conjugate linear correspondence x; 3 I++ y1 E X (defined in Corollary 1 in Chapter III, 6), then
(Tx, y') = y'(Tx) = (Tx, ]yy'), (x, T'y') = (T'y') (x) = (x,JxT'y').
We have thus
(Tx,
J yy') =
(x,] xT' y'), that is, (Tx, y)
= (x,] xT' T;}y).
In the special case when Y =X, we write
T* = fxT'Tx?
and call T* the adjoint operator of T.
Remark. If X is the complex Hilbert space (l2 ), we see, as in the Example in the preceding section, that the matrix corresponding to T* is the
transposed conjugatP- matrix of the matrix corresponding to T.
As in the case of dual operators, we can prove
Theorem I. T* exists iff D (Tt =X. It is defined as follows: y EX
is in the domain of D (T*) iff there exists a y* E X such that
(T x, y)
(x, y*)
holds for all
xED (T),
(1)
and we define T* y = y*.

We can rewrite the above theorem in terms of the graph G (A) of the
linear operator A (the graph was introduced in Chapter II, 6):
Theorem 2. We introduce a continuous linear operator V on X X X
into XXX by
V{x, y} = {-y, x}.
(2)
Then (V G (T))j_ is the graph of a linear operator iff D (Tt =~ X, and, in
fact, we have
G(T*) = (VG(T))j_.
(3)
Proof. The condition {-Tx, x} _L {y, y*} is equivalent to (Tx,y) =
(x, y*). Thus Theorem 2 is proved by Theorem 1.
Corollary. T* is a closed linear operator, since the orthogonal complement of a linear subspace is a closed linear subspace.
Theorem 3. Let T be a linear operator on D (T) ~ X into X such that
D(Tt =X. Then T admits a closed linear extension iff T** =(T*)*
exists, i.e., iff D (T*t =X.
Proof. Sufficiency. We have T** ~ T by definition, and T** = (T*)*
is closed by the above Corollary.
Necessity. Let 5 be a closed extension of T. Then G(S) contains
G (Tt as a closed linear subspace, and so G (Tt. is the graph of a linear
operator. But G(Tt = G(T)j_j_ = (G(T)j_)j_ by the continuity of the
3. Symmetric Operators and Self-adjoint Operators
197
scalar product, and, moreover, by VG(T*) = G(T)_j_, we obtain

(VG(T*))_!_ = G(T)_j__j_. Therefore, (VG(T*))_!_ is the graph of a linear
operator. Thus by Theorem 2 D (T*t =X and T** exists.
Corollary. Under the condition that D (Tt =X, Tis closed linear iff
T = T**.
Proof. The sufficiency is clear. Necessity is proved by observing the
formula G (Tt = G (T**) obtained above. For, G (T) = G (Tt implies
that T = T**.
Theorem 4. An everywhere defined closed linear operator Tis a continuous linear operator.
Proof. Clear from the closed graph theorem.
Theorem 5. If Tis a bounded linear operator, then T* is also bounded
linear and
(4)
IITII =liT* II
Proof. Similar to the case of dual operators.
A Hermitian matrix is a matrix which is equal to its transposed

conjugate matrix. It is known that such a matrix can be transformed
into a diagonal matrix by a suitable (complex) rotation of the vector
space on which the matrix operates as a linear operator. The notion of the
Hermitian matrices is extended to the notion of self-adjoint operators
in a Hilbert space.
Definition 1. Let X be a Hilbert space. A linear operator on D(T) ~X
into X is called symmetric if T* ~ T, i.e., if T* is an extension ofT. Note
that the condition of the existence of T* implies that D (Tt = X.
Proposition 1. If T is symmetric, then T** is also symmetric.
Proof. Since Tis symmetric, we have D (T*) ~ D (T) and D (Tt =X.
Hence D(T*t =X and soT**= (T*)* exists. T** is surely an extension of T and so D (T**) ~ D (T). Thus, again by D (Tt =X, we have
D(T**t =X and so T*** = (T**)* exists. We have, from T* ~ T,
T** ~ T* and hence T*** ;{ T** which proves that T** is symmetric.
Corollary. A symmetric operator T has a closed symmetric extension
T** = (T*)*.
Definition 2. A linear operator T is called self-adjoint if T = T*.
Proposition 2. A self-adjoint operator is closed. An everywhere defined
symmetric operator is bounded and self-adjoint.
Proof. Being the adjoint of itself, a self-adjoint operator is closed.
The last assertion is proved by the fact that an everywhere defined closed
operator is bounded (closed graph theorem).
Example 1 (integral operator of the Hilbert-Schmidt type). Let
-oo <a< b < oo and consider L2(a, b). Let K(s, t) be a complex-
198
VII. Dual Operators
valued measurable function for a< s, t < b such that

b
J J jK(s, t) j2 ds dt
a a
< CXJ.
For any x(t) E P(a, b), we define the operator K by

b
(K x) (s) = J K(s, t) x(t) dt.

a
We have, by Schwarz' inequality and the Fubini-Tonelli theorem,

b
jj(Kx)(s)j2ds-::;;.J JjK(s,t)j 2 dtdsJjx(t)\ 2 dt.
Hence K is a bounded linear operator on L2(a, b) into L2(a, b) such that
bb
)1/2
II K II < (j j j K (s, t) j2 ds dt . It is easy to see that the operator K*
b
is defined by (K*y) (t)
= J K(s, t) y(s) ds.

a
Hence K is self-adjoint iff
K (s, t) = K (t, s) for a.e. s, t.

Example 2 (the coordinate operator in quantum mechanics). Let
X= L2 (- CXJ, CXJ). Let D = {x(t); x (t) and t x(t) both E P(- CXJ, CXJ)}.
Then the operator T defined by Tx(t) = t x(t) on Dis self-adjoint.
Proof. It is clear that Da =X, since the linear combinations of
defining functions of finite intervals are strongly dense in P(-CXJ,CXJ).
Let y ED (T*) and set T*y = y*. Then, for all xED= D (T),
00
00
J tx(t) y(t) dt = J x(t) y* (t) dt.
-00
-00
If we take for x(t) the defining function of the interval [eX, t0 ], we have
t0
t0
J t y (t) dt = J y* (t) dt, and hence, by differentiation, t
y (t0 )
y* {t0 )
"' a.e. t Thus"' y ED and T*y (t) = t y (t). Conversely, it is clear that
for
0
y ED implies that y ED (T*) and T* y (t) = t y (t).

Example 3 (the momentum operator in quantum mechanics). Let
X= L 2 (-CXJ, CXJ). Let D be the totality of x(t) E P(-CXJ, CXJ) such that
x (t) is absolutely continuous on every finite interval with the derivative
x' (t) E L 2 ( - CXJ, CXJ). Then the operator T defined by Tx(t) = i-1 x' (t)
on D is self-adjoint.
Proof. Let a continuous function x,. (t) be defined by
x,. (t) = 1 fortE [eX, t 0 ],
x,. (t)
0 for t -::;;.
eX -
n-1 and for t
>
t0
+ n-1 ,
x,.(t) is a linear function on [iX- n-1 , eX] and on [t0 , t0
+ n-1 ].
199
Then the linear combinations of functions of the form xn (t) with different
values of c:x, t0 and n are dense in L 2 (-oo, oo). Thus Dis dense in X.
Let y E D(T*) and T*y = y*. Then for any xED,
00
J i- x' (t) y (t) dt = J x (t) y* (t) dt.

00
-oo
-oo
If we take Xn (t) for x (t), we obtain

t,+n-t
"'
J i- y(t)dt-n tJ
.:-n-
oo
J Xn(t)y*(t)dt,
-oo
and so, by letting n -+ oo, we obtain i-1(Y (c:x) - y (t = J' y* (t) dt for
n
i-1 y(t)dt=
0 ))
"'
a.e. c:x and t0 It is clear, by Schwarz' inequality, that y* (t) is integrable
over any finite interval. Thus y (t0) is absolutely continuous in t 0 over
any finite interval, and so we have i-1y' (t0 ) = y* (t0 ) for a.e. t0 . Hence
yE D and T*y(t) = i-1 y'(t). Let, conversely, yE D. Then, by partial
integration,
b
J i- x' (t) y(t) dt = i-1 [x(t) y(t)]~ + aJ x(t) (i-1y' (t)) dt.
a
1
over (-oo,oo), we see that
By the integrability of x(t)y(t)
00
00
I[x(t) y(t)]~ I= 0, and so-ooJ i- x' (t) y(t) dt = -ooJ x(t) (i-1y' (t)) dt.
aj,-oo,btoo
lim
Thus y ED (T*) and T*y (t) = i-1y' (t).

Theorem 1. If a self-adjoint operator T admits the inverse T-1 , then
T-1 is also self-adjoint.
Proof. T = T* is equivalent to (VG(T)).l. = G(T). We have also
G(T-1) = VG(-T). Hence, by (-T)* = -T* = -T, (VG(-T))..L =
G(-T) and so
(VG(T- 1 ))1.
= G(-T)..L = (VG(-T)).l...L =
VG(-T)
= G(T-
1 ),
that is, (T-1)* = y-1.

We have used, in the above proof, the fact that (VG(-T)/ = VG(-T)
in virtue of the closedness of (-T).
Corollary. A symmetric operator T in a Hilbert space X is self-adjoint if D(T) =X or if R(T) =X.
Proof. The case D (T) =X was proved already. We shall prove the
case R(T) =X. Tx = 0 implies 0 = (Tx, y) = (x, Ty) for all
y ED (T), and so, by R (T) = X, we must have x = 0. Therefore the
inverse T- 1 exists which is surely symmetric with T. D (T- 1) = R (T) =X,
and so the everywhere defined symmetric operator T-1 must be self-adjoint. Hence T = (T-1)-1 is self-adjoint by Theorem 1.
200
VII. Dual Operators
We can construct self-adjoint operators from a closed linear operator.

More precisely, we have
Theorem 2 (J. VON NEUMANN [5]). For any closed linear operator T
in a Hilbert space X such that D(Tt =X, the operators T*T and IT*
are self-adjoint, and (I+ T* T) and (I+ TT*) both admit bounded
linear inverses.
Proof. We know that, in the product space XxX, G(T) and VG(T*)
are closed linear subspaces orthogonal to each other and spanning the
whole product space XxX. Hence, for any hE X, we have the uniquely
determined decomposition
{h, 0} = {x, Tx}

Thus h
= x- T*y,
+ {-T*y, y}
= Tx
with
+ y. Therefore
xED (T* T)
and
xED (T), y ED (T*).
T* T x =h.
(1)
(2)
Because of the uniqueness of decomposition (1), x is uniquely determined by h, and so the everywhere defined inverse (I + T* T)-1 exists.
For any h, k E X, let
X= (I+ T*T)- 1h, y =(I+ T*T)- 1k.

Then x and y E D (T* T) and, by the closedness of T, (T*)* = T. Hence
(h, (I+ T*T)- 1k) =((I+ T*T) x, y) = (x, y) + (T*Tx, y)
= (x, y) + (Tx, Ty) = (x, y) + (x, T*Ty)
= (x, (I+ T*T) y) =((I+ T*T)-1 h, k),
which proves that the operator (I + T* T)-1 is self-adjoint. As an everywhere defined self-adjoint operator, (I + T* T)-1 is a bounded operator.
By Theorem 1, its inverse (I + T* T) and hence T* Tare self-adjoint.
Since Tis closed, we have (T*)* = T, and so, by what was proved
above, TT* = (T*)* T* is self-adjoint and (I+ TT*) has a bounded
linear inverse.
We next give an example of a non-self-adjoint, symmetric operator:
Example 4. Let X= P(O, 1). Let D be the totality of absolutely
continuous functions x(t) E P(O, 1) such that x(O) = x(1) = 0 and
x'(t)EL2(0, 1). Then the operator T 1 defined by T 1x(t) =i-1x'(t) on
D = D (T1) is symmetric but not self-adjoint.
Proof. We shall prove that Ti = T 2 , where T 2 is defined by:
T 2 x(t)=i-1x'(t) on D(T2 )={x(t)EL2(0,1);x(t) is
absolutely continuous such that x' (t) E L2 (0, 1)}.
Since D = D (T1) is dense in L2 (0, 1), the operator Ti is defined.
Let y ED (Ti) and set Tfy = y*. Then, for any xED= D (T 1),
1
J i- x' (t) y (t) dt = J x (t) y* (t) dt.

1
201
By partial integration, we obtain, remembering x(O) = x(1) = 0,

t
J x (t) y* (t) dt = - J x' (t) Y* (t) dt,
Hence, by x(1) =
J y* (s) ds.
where Y* (t) =
J x'(t) dt =
0, we have, for any constant c,
J x' (t) ( Y* (t) -
i-1 y (t) -
C) dt =
for all
x E D (TJ .
J z(t) dt surely belongs
Jz (t) dtI
On the other hand, for any z (t) E L2 (0, 1), the function Z (t) =
to D{T1). Hence, taking Z(t) for the above
x (t), we obtain
j {z(t)- j z(t) dt} (Y*(t)
- i -1 y(t)- C) dt = 0.
1
If we take the constant c in such a way that
then
1
z (t) ( Y* (t) - i-1 y (t)
J (Y* (t)- i-
y(t) -c) dt = 0,
-C) dt =
0,
and so, by the arbitrariness of zEL 2 (0, 1), we must have Y*(t)
Jy* (t) dt= iI
y (t)
+ c.
Hence y E D (T2) and T 2 y = y*. This proves
that T'f ~ T 2 It is also clear, by partial integration, that T 2 ~

so T 2 = Tt.
Theorem 3. If H is a bounded self-adjoint operator, then
Tt
and
(3)
IIHJI =sup I(Hx,x)l.
11 .. 11:;3;1
Proof. Set sup I(Hx,x)l =y. Then, by I(Hx,x)l < JIHxJIJixJI,
11 ..11:;3;1
y < JIHII. For any real number A., we have
I(H(y
lcz), y lcz) I=
I(Hy, y)
<y
Hence
I4A Re(Hy, z) I:::;; y(IIY
2/c Re(Hy, z) + lc2(Hz, z) I
I!Y 1czll2
+ 1czll2 + !ly-/czll2) =
2y(IIYII2
+ )c211zJI2).
By taking A= IIYII/IIzll, we obtain IRe(Hy,z)l <y IIYIIIIzll. Hence,

by substitutiag zei 9 for z, we obtain I(Hy, z) I< y IIY Illizi I, and so
(Hy, Hy) = !IHyJI2 < y !IYIII!Hyll, i.e., [[H[[ < Y
202
VII. Dual Operators
4. Unitary Operators. The Cayley Transform

A symmetric operator is not necessarily a bounded operator. Various
investigations of a symmetric operator H may be made through the
continuous operator (H- il) (H + ii)-1 called the Cayley transform
of H. We shall begin with the notion of isometric operators.
Definition 1. A bounded linear operator T on a Hilbert space X into
X is called (bounded) isometric if T leaves the scalar product invariant:
forall
(Tx,Ty)=(x,y)
x,yEX.
(1)
If, in particular, R (T) = X, then a (bounded) isometric operator T

is called a unitary operator.
Proposition 1. For a bounded linear operator T, condition (1) is
equivalent to the condition of the isometry
II T x II = II x II
for all x E X.
Proof. It is clear that (1) implies (2). We have, by (2),
4Re(Tx, Ty) = IIT(x + y) 11 2 -IIT(x-y) 112
=
II x + y 11 2 - II x -
(2)
y 11 2 = 4 Re (x, y) .
By taking iy in place of y, we also obtain 4Im(Tx, Ty) = 4Im(x,y),

and so (2) implies (1).
Proposition 2. A bounded linear operator on a Hilbert space X into X
is unitary iff T* = T-1.
Proof. If Tis unitary, then T-1 surely exists in virtue of condition (2),
and D(T-1) = R(T) =X. Moreover, by (1), T*T =I and soT*= T-1.
Conversely, the condition T* = T-1 implies T* T = I which is the condition of the invariance of the scalar product. Moreover, T* = T-1
implies that R (T) = D (T-1 ) = D (T*) = X and hence we see that T
must be unitary.
Example 1. Let X= L 2 (-=, =).Then, for any real number a, the
operator T defined by Tx(t) = x(t +a) on P(-=, =)is unitary.
Example 2. The Fourier transform on L 2 (R") onto L 2 (Rn) is unitary,
since it leaves the scalar product (/,g)=
f(x) g(x) dx invariant.
Rn
Definition 2. Let X be a Hilbert space. A linear operator T defined

~ X into X such that D (Tt = X is called normal if
on D (T)
TT*
T*T.
(3)
Self-adjoint operators and unitary operators are normal.

The Cayley Transform
Theorem 1 (J. voN NEUMANN [1]). Let H be a closed symmetric
operator in a Hilbert space X. Then the continuous (but not necessarily
203
+ ii)-1 exists, and the operator

UJl = (H -ii) (H + ii)-1 with the domain D (UJI) = D((H + il)-1)
is closed isometric (I IUJI x II = II x II), and (I- UJI)-1 exists.
everywhere defined) inverse (H
We have, moreover,
H
i(I
+ UJI) (I- UJI)-1.
(5)
Thus, in particular, D (H) = R (I- UJI) is dense in X.

Definition 3. UJI is called the Cayley transform of H.
Proof of Theorem 1. We have, for any xE D(H),
((H if) x, (H if) x) = (Hx, Hx)
(Hx, ix)
(4)
(ix, Hx)
+ (x, x).
The symmetry condition for H implies (H x, ix) = - i (H x, x) =

-i(x,Hx) =-(ix,Hx) and so
II (H if) xll2 = 11Hxll2 + llxll2
(6)
Hence (H + if) x = 0 implies x = 0 and so the inverse (H + iJ)-1

exists. Since II (H + ii) x II > II x II, the inverse (H + i J)-1 is continuous.
By (6), it is clear that II U HY II = II y II, i.e., U JI is isometric.
UJI is closed. For, let (H + if) Xn = Yn--')- y and (H- ii) Xn = Zn--')- z
as n-')-oo. Then we have, by (6), 11Yn-Ymll 2 = IIH(xn-Xm)ll 2 +
jIXn- Xm 11 2 , and so (xn- xm) --')- 0, H (xn- Xm)--')- 0 as n, m--')- oo. Since
H is closed, we must have x = s-lim Xn E D (H) and s-lim H Xn = H x.
1>---700
1>---700
Thus (H + ii) Xn ->- y = (H + ii) x, (!l- if) Xn--')- z = (H- ii) X and
so UJIY = z. This proves that UJI is closed.
y = (H + ii) x and UHY = (H- ii) x, we obtain
From
2-1 (I- Uu) y = ix and 2-1 (I + Uu) y = Hx. Thus (I- Uu) y = 0
implies x = 0 and so (I+ UJI) y = 2Hx = 0 which implies y =
2-1 ((1- UJI) y +(I+ UJI) y) = 0. Therefore the inverse (I- Uu)- 1
exists. By the same calculation as above, we obtain
Hx = 2-1 (1
+ UJI) y =
H =
+ UJI) (I- Uu)-1 x,

i(I + Uu) (I- Uu)-1 .
i(I
that is,
Theorem 2 (J. VON NEUMANN [1]). Let U be a closed isometric operator such that R (I- U)a = X. Then there exists a uniquely determined
closed symmetric operator H whose Cayley transform is U.
Proof. We first show that the inverse (I- U}-1 exists. Suppose that
(I- U) y = 0. For any z =(I- U) wE R(I- U), we have, by the
isometric property of U, (y, w) = (Uy, Uw) as in Section 1. Hence
(y, z) = (y, w)- (y, Uw) = (Uy, Uw)- (y, Uw) = (Uy- y, Uw) = 0.
Hence, by the condition R(I- Ut =X, y must be= 0. Thus (I- U)-1
exists. Put H = i (I + U) (I- U}-1 . Then D (H) = D ((I- U)-1) =
204
VII. Dual Operators
R (I- U) is dense in X. We first prove that H is symmetric. Let

x.,yE D(H) = R(I- U) and put x =(I- U) u,y =(I- U) w. Then
(Vu, Uw) = (u, w) implies that
(Hx, y)
+ U) u, (1- U) w) = i((Uu, w)- (u, Uw))

=((I- U) u, i(I + U) w) = (x, Hy).
=
(i(I
The proof of UH = U is obtained as follows. For x = (I- U) u, we have

Hx = i(I + U) u and so (H + ii) x = 2iu, (H -ii) x = 2iUu. Thus
D(UH) = {2iu; u E D(U)} = D(U), and UH(2iu) = 2i Uu = U(2iu).
Hence UH = U.
To complete the proof of Theorem 2, we show that H is a closed
operator. In fact, His the operator which maps (I- U) u onto i (I+ U) u.
If (I- U) Un and i (I + U) Un both converge as n ~ oo, then Un and
U un both converge as n ~ oo. Hence by the closure property of U, we
must have
Un~ U,
(I- U)
Un~
(I- U)
U,
i(I
+ U) Un~ i(I + U)
U.
This proves that H is a closed operator.

For the structure of the adjoint operator of a symmetric operator,
we have
Theorem 3 (J. voN NEUMANN [1]). Let H be a closed symmetric
operator in a Hilbert space X. For the Cayley transform U H =
(H- ii) (H + ii)-1 of H, we set
(7)
Then we have
Xfi = {xE X; H*x = ix}, XJi = {xE X; H*x = -ix},
(8)
and the element x of D (H*) is uniquely expressed as
+ x1 + x 2 , where x 0 E D (H), x1 E Xfi, x 2 E XJi

H*x = Hx0 + ix1 + (-ix2).
x = x0
so that
(9)
Proof. xE D(UH)j_ = D((H + ii)- 1)j_ implies (x, (H + ii) y) = 0

for all yE D(H). Hence (x, Hy) = - (x, iy) = (ix, y) and so xE D(H*),
H* x = i x. The last condition implies (x, (H
i I) y) = 0 for ally E D (H),
i.e. xED((H
iJ)- 1)j_ =D(UH)j_. This proves the first half of (8);
the latter half may be proved similarly.

Since U H is a closed isometric operator, we see that D (U H) and
R(UH) are closed linear subspaces of X. Hence any element xE X is
uniquely decomposed as the sum of an element of D(UH) and an element
of D (UH) j_. If we apply this orthogonal decomposition to the element
(H* + ii) x, we obtain
(H*
+ ii) x =
(H
+ ii) x 0 + x'
where x0 E D(H), x' E D(UH)J.
205
But we have (H + il) x 0 = (H* + il) x 0 by x0 E D(H) and H

We have also H* x' = ix' by x' ED (U H)l. and (8). Thus
x'
and so
(H*
(H*
+ il) x
+ iJ) x 1 ,
=
(H*
x1
H*.
= (2i)- 1 x' E D(UH)l.,
+ il) x 0 + (H* + il) x 1
where
x 0 ED(H), x 1 ED(UH)l.
Therefore (x- x 0 - x 1 ) E R (U H)l. by H* (x- x 0 - x 1 ) = - i (x- x 0 - x 1 )

and (8). This proves (9). The uniqueness of the representation (9) is
proved as follows. Let 0 = x 0 + x1 + x2 with x 0 ED (H), x 1 ED (UH) 1,
x 2 E R(UH)l.. Then, by H*x 0 = Hx0 , H*x 1 = ix1 , H*x 2 = - ix2 ,
0 = (H*
+ il) 0 =
(H*
+ il) (x0 + x1 + x 2) =
(H
+ il) x 0 + 2ix1 .
But by the uniqueness of the orthogonal decomposition of X as the sum

of D (U H) and D (UH)l., we obtain (H + il) x 0 = 0, 2ix1 = 0. Since the
inverse (H + iJ)-1 exists, we must have x 0 = 0 and so x 2 = 0- x 0 -
0 - 0 - 0 = 0.
Corollary. A closed symmetric operator H in a Hilbert space X is selfadjoint iff its Cayley transform U His unitary.
Proof. The condition D (H) = D (H*) is equivalent to the condition
D (U H)l. = R (U H)l. = {0}. The last condition in tum is equivalent to the
condition that U H is unitary, i.e. the condition that U H maps X onto
X one-one and isometrically.
x1
5. The Closed Range Theorem
The closed range theorem of S. BANACH [1] reads as follows.

Tl.eorem. Let X and Y be B-spaces, and T a closed linear operator
defined in X into Y such that D (Tt = X. Then the following propositions are all equivalent:
R (T) is closed in Y,
(1)
R (T') is closed in X',
(2)
R(T) = N(T')l. = {yE Y; (y, y*) = 0 for all y* E N{T')},
(3)
R (T') = N (T)l. = {x* EX'; (x, x*) = 0 for all x EN (T)}.

(4)
Proof. The proof of this theorem requires five steps.
The first step. The proof of the equivalence (1) ++ {2) is reduced to
the equivalence (1) ++ (2) for the special case when T is a continuous
linear operator such that D (T) =X.
The graph G = G (T) of Tis a closed linear subspace of X X Y, and so
G is a B-space by the norm ll{x, y} II= [[x II + fly II of XX Y. Consider a
continuous linear operator S on G into Y:
S{x, Tx} = Tx.
206
VII. Dual Operators
Then the dual operator S' of S is a continuous linear operator on Y' into
G', and we have
<{x, Tx}, S'y*)
<S{x, Tx}, y*)
<Tx, y*)
= <{x, Tx}, {0, y*}), xED (T), y* E Y'.

Thus the functional S' y*- {0, y*} E (XX Y)' =X' X Y' vanishes at
every point of G. But, the condition <{x, Tx}, {x*, yi}) = 0, xE D(T).
is equivalent to the condition <x,x*) = <-Tx,y~), xED(T), that is,
to the condition - T' Yi = x*. Hence
S' y*
= {0, y*} + {-T'yi, yi} =
{-T'yi, y*
+ yi},
y* E Y'.
By the arbitrariness of y*, we see that R (S') = R (- T') X Y' = R (T') X Y'.
Therefore R (S') is closed in X' X Y' iff R (T') is closed in X', and, since
R(S) = H(T), R(S) is closed in Y iff R(T) is closed in Y. Hence we
have only to prove the equivalence (1) -. (2) in the special case of a
bounded linear operatorS, instead of the original T.
The second step. Let X and Y be B-spaces, and T a bounded linear
operator on X into Y. Then (1) -7 (2).
We consider T as a bounded linear operator T 1 on X into the B-space
Y1 = R(Tt = R(T). We have to prove that (2) is true. T~yi, Yi E Y~,
is defined by
<T1 x, Yi) = <T x, Yi) = <x, T~yf), xED (T1 ) = D (T) =X.
By the Hahn-Banach theorem, the functional yf can be extended to a

y* E Y' in such a way that <Tx, Yi> = <Tx, y*), xE D(T) =X. Hence
T~ yf '= T' Yi and so R (T~) = R (T'). Thus it suffices to assume that
R (T) = Y. Then, by the open mapping theorem in Chapter II, 5, there
exists a c > 0 such that for each y E Y, there exists an x E X with
Tx = y, llx II < c IIY II Thus, for each y* in D (T'), we have
l<y, y*> I= I<Tx, y*> I=
Hence
l<x, T'y*> I
< II X II . II T' y* II < c II y II . II T' y* II

IIY*II =sup l<y,y*>l < c IIT'y*ll
JJyJJ~l
and so (T')- 1 exists and is continuous. Moreover, (T')-1 is a closed linear

operator as the inverse of a bounded linear operator. Hence we see that
the domain D((T')-1) = R(T') must be closed in X'.
The third step. Let X and Y be B-spaces, and T a bounded linear
operator on X into Y. Then (2) -+ (1).
As in the second !"tep, we consider T as a bounded linear operator
T 1 on X into Y 1 = R(Tt. Then T~ has the inverse, since T~yi = 0
implies
<T1 x, yf) = <Tx, Yi> = <x, T~yf) = 0, xE D(T1 ) = D(T) =X,
207
5. The Closed Range Theorem
and so, since R (T1 } = R (T) is dense in Y1 = R (Tt, Yi must be 0. Therefore, the condition that R (T') = R (T~) (proved above) is closed, implies
that T~ is a continuous linear operator on the B-space (R (Tt)' = Y~
onto the B-space R (T~) in a one-one way. Hence, by the open mapping
theorem, (T~)- 1 is continuous.
We then prove that R(T) is closed. To this purpose, it suffices to
derive a contradiction from the condition
there exists a positive constant e such that the image

{T1 x; llx II < e} is not dense in all the spheres II y ll<n- 1
(n = 1, 2, ...) of Y 1 = R(Tt = R(T1 }a.
For, if otherwise, the proof of the open mapping theorem shows that
= R (T) = Y 1 . Thus we assume that there exists a sequence
{Yn} ~ Y1 with
R (T1}
s-lim Yn
1!r-->OO
0, YnE {T1 x; llxll
< e}a
(n
1, 2, ... ).
Since {T1 x; II x II < e}a is a closed convex, balanced set of the B-space Y1
there exists, by Mazur's theorem in Chapter IV, 6, a continuous linear
functional In on the B-space Y1 such that
fn(Yn)
Hence II T~fn II <
e-1
sup lfn(Tlx) I (n = 1, 2, ... ) .

llxll :::;;e
II In 1111 Yn II, and so, by s-lim Yn = 0, T~ does not have
>
1!r-->OO
a continuous inverse. This is a contradiction, and so R (T) must be closed.

The fourth step. We prove (1} ---* {3). First, it is clear, from
(Tx,y*)=lx,T'y*), xED(T), y*ED(T'),
thatR(T) c;;,N(T')j_. We show that (1) impliesN(T')j_ c;;, R(T). Assume
that there exists ay0 E N(T')j_ withy0 E R(T). Then, by the Hahn-Banach
theorem, there exists a Yri E Y' such that (y0 , Yri> # 0 and (T x, Yri) = 0
for all x E D (T). The latter condition implies (x, T' Yri> = 0, x E D (T),
and hence T' Yri = 0, i.e., y0 E N (T') j_. This is a contradiction and so we
must have N(T')j_ s;; R(T).
The implication (3} _,.. (1) is clear, since N(T')l_ is closed by virtue
of the continuity in y of (y, y*).
The fifth step. We prove (2)--* (4). The inclusion R(T') ~ N(T)l_ is
clear as in the case of (3). We show that (2) implies that N(T)l_ ~ R(T').
To this purpose, let x* EN (T) 1_, and define, for y = T x, the functional
/ 1 (y) of y through / 1 (y) = (x, x*). It is a one-valued function of y, since
Tx = Tx' implies (x-x') EN(T) and so, by x*EN (T)l_, ((x-x'),x*) = 0.
Thus / 1 (y) is a linear functional of y. (2) implies (1), and so, by the open
mapping theorem applied to the operator 5 in the first step, we may
choose the solution x of the equation y = T x in such a way that s-lim
208
VII. Dual Operators
y = 0 implies s-lim x = 0. Hence / 1 (Y) = <x, x*) is a continuous linear

functional on Y1 = R(T). Let /E Y' be an extension of / 1. Then
f(Tx) = /1 (Tx) = <x, x*).
This proves that T'f = x*. Hence N(T)J. ~ R(T').
That (4) implies (2) is clear, since <x, x*) is a continuous linear functional of x.
Corollary 1. Let X andY be B-spaces, and T a closed linear operator
on D (T) ~ X into Y such that D (Tt = X. Then
R(T) = Y iff T' has a continuous inverse,
(5)
R (T') =X' iff T has a continuous inverse.
(6)
Proof. Suppose that R(T) = Y. Then, from <Tx, y*) = <x, T'y*),
xED (T) and T' y* = 0, we obtain y* = o; that is, T' must have the
inverse (T')-1 Since, by R(T) = Y and (2), R(T') is closed, the closed
graph theorem implies that (T')-1 is continuous. Next let T' admit a
continuous inverse. Then N (T') = {0} and also (2) holds since T' is
closed. Thus, by (3), R (T) = Y.
Suppose that R(T') =X'. Then, from <Tx, y*) = <x, T'y*),
y* E D (T') and T x = 0, we obtain x = 0, i.e., T must have the inverse
y-1. Since, by R(T') =X' and (1), R(T) is closed, the closed graph
theorem implies that T-1 must be continuous. Next let T admit a continuous inverse. Then N(T) = {0} and also (1) holds since Tis closed.
Thus, by (4), R(T') =X'.
Corollary 2. Let X be a Hilbert space with a scalar product (u, v), and
T a closed linear operator with dense domain D (T) ~ X and range
R (T) ~ X. Suppose that there exists a positive constant c such that
Re(Tu,u)>cJJuJJ 2 forall uED(T).
Then R(T*) =X.
Proof. By Schwarz' inequality, we have
(7)
JITullllull > Re(Tu, u) > c llull 2 for all uE D(T).

Hence II T u II > c II u II, u E D (T), and so T admits a continuous inverse.
Thus, by the preceding Corollary, R (T') =X. Hence R (T*) = R (T') =X.
Remark. A linear operator T on D (T) ~ X into X is called accretive
(the terminology is due to K. FRIEDRICHS and T. KATo) if
Re(Tu,u) > 0 for all uED(T).
(8)
T is called dissipative (the terminology is due to R. S. Phillips) if -T
is accretive.
References for Chapter VII
For a general account concerning Hilbert spaces, see M. H. STONE
[1], N. I. ACHIESER-1. M. GLASMAN [1] and N. DUNFORD- J. SCHWARTZ
[5]. The closed range theorem is proved essentially inS. BANACH [1].
1. The Resolvent and Spectrum
209
VIII. Resolvent and Spectrum

Let T be a linear operator whose domain D (T) and range R (T) both
lie in the same complex linear topological space X. We consider the linear
operator
T;. =U -T,
where Ais a complex number and I the identity operator. The distribution
of the values of A for which T;. has an inverse and the properties of the
inverse when it exists, are called the spectral theory for the operator T.
We shall thus discuss the general theory of the inverse of T;..
1. The Resolvent and Spectrum

Definition. If Ao is such that the range R (T;..) is dense in X and T;.,
has a continuous inverse (A.0 I - T)-1 , we say that Ao is in the resolvent
set e(T) of T, and we denote this inverse (A 0 I - T)-1 by R (Ao; T) and
call it the resolvent (at A0 ) of T. All complex numbers A not in e(T) form
a set a(T) called the spectrum ofT. The spectrum a(T) is decomposed
into disjoint sets Pa(T), Ca(T) and Ra(T) with the following properties:
Pa (T) is the totality of complex numbers A for which T;. does not have
an inverse ;Pa (T) is called the point spectrum of T.
Ca (T) is the totality of complex numbers A for which T;. has a discontinuous inverse with domain dense in X; Ca (T) is called the continuous spectrum of T.
Ra (T) is the totality of complex numbers A for which T;. has an inverse
whose domain is not dense in X; RG (T) is called the residual
spectrum of T.
From these definitions and the linearity of T we have the
Proposition. A necessary and sufficient condition for Ao E P (T)
is that the equation Tx =A-ox has a solution x # 0. In this case A.o is
called an eigenvalue of T, and x the corresponding eigenvector. The null
space N (A 0 l - T) ofT;., is called the eigenspace ofT corresponding to the
eigenvalue Ao of T. It consists of the vector 0 and the totality of eigenvectors corresponding to A0 The dimension of the eigenspace corresponding to Ao is called the multiplicity of the eigenvalue A0 .
Theorem. Let X be a complex B-space, and T a closed linear operator
with its domain D (T) and rangeR (T) both in X. Then, for any A.o E e(T),
the resolvent (A0 I - T)-1 is an everywhere defined continuous linear
operator.
Proof. SinceA.0 isin the resolvent set e(T), R (A 0 I - T) = D((lol- T)-1)
is dense in X in such a way that there exists a positive constant c for
which
II (loi - T) x II > c II x II whenever x E D (T) .
(1
210
We have to show that R().0 I - T) =X. But, if s-lim (). 0 1- T)

n--->00
Xn
= y
exists, then, by the above inequality, s-lim xn = x exists, and so, by the
n--->00
closure property of T, we must have ().0 1 - T) x = y. Hence, by the

assumption that R (Aol- Tt = X, we must have R (Aol- T) = X.
Example I. If the space X is of finite dimension, then any bounded
linear operator T is represented by a matrix (tij) It is known that the
eigenvalues of T are obtained as the roots of the algebraic equation,
the so-called secular or characteristic equation of the matrix (tij):
det (Hii- tij) = 0,
(1)
where det(A) denotes the determinant of the matrix A.

Example 2. Let X = L 2 ( - =, =) and let T be defined by
T . X (t)
t X (t) '
that is, D(T) = {x(t); x(t) and tx(t) E L2(-=, =)} and Tx(t) = tx(t)
for x(t) E D(T). Then every real number ).0 is in Ca(T).
Proof. The condition (Aol- T) x = 0 implies (Ao- t) x (t) = 0 a.e.,
and so x(t) = 0 a.e. Thus ().0 1- T)-1 exists. The domain D((Aol- T)-1)
comprises those y (t) E L2 (- =,=)which vanish identically in the neighbourhood of t = ).0 ; the neighbourhood may vary with y(t). Hence
D((Aol- T)-1) is dense in L2 (- =,=).It is easy to see that the operator
(Aol- T)-1 is not bounded on the totality of such y (t)'s.
Example 3. Let X be the Hilbert space (! 2), and let T0 be defined by
To (.;1, .;2, .)
(0, .;1, .;2, .)
Then 0 is in the residual spectrum of T, since R (T0 ) is not dense in X.

Example 4. Let H be a self-adjoint operator in a Hilbert space X.
Then the resolvent set e (H) of H comprises all the complex numbers).
with Im(A) # 0, and the resolvent R().; H) is a bounded linear operator
with the estimate
(2)
JJR().;H)JJ < 1/Jim(A)J.
Moreover,
Im((H-H)x,x)=Im().) JJxJJ 2 , xED(H).
(3)
Proof. If xED(H), then (Hx,x) is real since (Hx,x) = (x,Hx) =
(Hx, x). Therefore we have (3), and so, by Schwarz' inequality,
II (H- H) xJJ Jjxjj > J((H- H) x, x) I> Jim().) I Jlxll 2

(4)
which implies that
JJ(H -H) xlj >Jim().) IJJxjj, xE D(H).
(5)
Hence the inverse ().I- H)-1 exists if Im (A) # 0. Moreover, the range
R (H- H) is dense in X if Im ().) # 0. If otherwise, there would exist a
y # 0 orthogonal to R(H- H), i.e., ((H- H) x, y) = 0 for all xED (H)
and so (x, (XI- H) y) = 0 for all xED (H). Since the domain D (H) of a
2. The Resolvent Equation and Spectral Radius
211
self-adjoint operator H is dense in X, we must have (.H- H) y = 0,

that is, H y = ~y, contrary to the reality of the value (H y, y).
Therefore, by the above Theorem, we see that, for any complex
number A with Im (A) =I= 0, the resolvent R (A; H) is a bounded linear
operator with the estimate (2).
2. The Resolvent Equation and Spectral Radius

Theorem 1. Let T be a closed linear operator with domain and range
both in a complex B-space X. Then the resolvent set e(T) is an open
set of the complex plane. In each component (the maximal connected
sets) of e(T), R (A; T) is a holomorphic function of A.
Proof. By the Theorem of the preceding section, R (A; T) for A E e(T)
is an everywhere defined continuous operator. Let A0 E e(T) and consider
S(A) = R(Ao; T)
{I+"~ (A0 -A)" R(A0 ; T)"}.
(1)
The series is convergent in the operator norm whenever IA0 - A I

II R (A0 ; T) II < 1, and within this circle of the complex plane, the series
defines a holomorphic function of A. Multiplication by (AI- T) =
(A - A0) I + (Aoi- T) on the left or right gives I so that the series
S (A) actually represents the resolvent R (A; T). Thus we have proved
that a circular neighbourhood of Ao belongs to e(T) and R (A; T) is
holomorphic in this neighbourhood.
Theorem 2. If A and fl both belong to e(T), and if R (A; T) and
R (fl; T) are everywhere defined continuous operators, then the resolvent
equation holds:
(2)
R(A; T)- R(ft; T) = (fl- A) R(A; T) R(ft; T).
Proof. \Ve have
R(A; T) = R(A; T) (fti- T) R(fl; T)
= R(A; T) {(ft-A) I+ (AI- T)} R(ft; T)
= {fl-A) R().; T) R(fl; T) + R(fl; T).
Theorem 3. If Tis a bounded linear operator on a complex B-space X
into X, then the following limit exists:
(3)
lim IIT"IIlfn=ra(T).
1>-->00
It is called the spectral radius of T, and we have

If IAI >
senes
(4)
r" (T), then the resolvent R (A; T) exists and is given by the
00
R(.A.; T) = .l: A-"T"- 1

"=1
which converges in the norm of operators.

14*
(5)
212
Proof. Set r = inf JJT'J/ 11". We have to show that lim JJT'II 11"<r.
,.::;,1
1>->00
For any e > 0, cho-;se m such that IITmqttm < r +e. For arbitrary n,
write n = pm + q where 0 < q :S (m-1). Then, by IIAB/1 <//A II
liB JJ, we obtain
JJT'JJ1fn < IITmW/" IIT/Jqfn < (r
+ e)mpfn liT WI".
Since pmfn--+ 1 and qfn--+ 0 as n--+ oo, we must have lim
n--K>O
+ e. Since e was arbitrary, we have proved lim II T" WI" <

Since
II T" II < II T II",
1>->00
we have lim
1>->00
II T' WI"< II T II
II T' win<
r.
The series (5)
is convergent in the norm of operators when A> ra(T). For, if J.A. J >
ra(T) + e with e > 0, then, by (3), II.A.-"T'IJ < (ra(T) +e)-"
(ra(T) + 2- 1 et for large n. Multiplication by (U- T) on the left or
right of this series gives I so that the series actually represents the resolvent R(/.; T).
Corollary. The resolvent set 12 (T) is not empty when Tis a bounded
linear operator.
Theorem 4. For a bounded linear operator TE L(X, X), we have
ra(T) =sup ji.J.

(6)
I.Ea(T)
Proof. By Theorem 3, we know that ra(T) > sup J!. J. Hence we
I.E<1(T)
have only to show that ra(T) sup J!.

.\Ea(T)
Thus it admits a uniquely determined Laurent expansion in positive and
non-positive powers of A convergent in the operator norm for IA >
sup JA. J. By Theorem 3, this Laurent series must coincide with
I.Ea(T)
00
.IA.-"T'-1. Hence lim IJ!.-"T'I/=0 if J!.J> sup IA.J, and so, for
J.
n=l
I.Ea(T)
(e
1>->00
I)"
any > 0, we must have II T' II<

+ sup J!. for large n. This
proves that
.tEa(T)
r <1 (T) = lim II T' win ->00
Proof. Let r
.\Ea(T)
00
2: .A_-nyn- 1 diverges if J!. J < ra(T).

n=l
be the smallest number > 0 such that the series
Corollary. The series

00
2: .A_-nyn-t converges in the operator norm for J!. J >
n=l
r. The existence
of such an r is proved as for ordinary power series in ,A.-1 Then, for

/.A. J > r, lim 1/1.-"T' II= 0 and so, as in the proof of ra(T) eo
.\Ea(T)
we must haye lim IIT'IJlfn < r. This proves that ra(T) < r.
1>->00
213
3. The M;ean Ergodic Theorem
3. The Mean Ergodic Theorem

For a particular class of continuous linear operators, the mean ergodic
theorem gives a method for obtaining the eigenspace corresponding to the
eigenvalue 1. In this section, we shall state and prove the mean ergodic
theorem from the view point of the spectral theory, as was formulated
previously by the present author. The historical sketch of the ergodic
theory in connection with statistical mechanics will be given in
Chapter XIII.
Theorem 1. Let X be a locally convex linear topological space, and T
a continuous linear operator on X into X. We assume that
the family of operators {T"; n = 1, 2, ... } is equicontinuous in the sense that, for any continuous seminorm q on X, there exists a continuous semi-norm q' on
X such that sup q (T" x) < q' (x) for all x E X.
(l)
n~1
Then the closure R (I- Tt of the range R (I- T) satisfies

R(I- Tt
= lfxE X; n-->oo
lim Tnx =
0, Tn
= n- 1
J: ym},
m~1
(2)
and so, in particular,

R(I- Tt f\ N(I- T) = {0}.
(3)
Proof. We have Tn (I-T) = n- 1 (T- T"+ 1). Hence, by (1),

wE R(I- T) implies that lim Tnw = 0. Next let zE R(I- Tt. Then,
n-->00
for any continuous semi-norm q' on X and8> 0, there exists awER(I -T)
such that q'(z-w) < 8. Thus, by (1), we have q(Tn(z-w)) <
n
n- 1 ~
m~1
q(Tm(z-w))
q(Tn(z- w))
< q'(z-w) <
< q(Tnw)
+ 8,
8.
Hence
q(Tnz)
< q(Tn w) +
and so lim Tnz = 0. This proves that

n-->00
R(I- T)a ~ {xE X; ,!i~ Tnx = 0}.

Let, conversely, lim Tnx = 0. Then, for any continuous semi-norm
n-->00
q on X and 8 > 0, there exists an n such that q(x- (x- Tnx))

q(Tnx) < 8. But, by
x-Tnx=n- 1 ~ (I-Tm)x
m~1
n- 1 ~
m~l
(I- T) (I
+ T + T 2 + ... + ym- 1 ) x,
(x- T nx) E R (I- T). Hence x must belong to R (I- T)a.

Theorem 2 (the mean ergodic theorem). Let condition (l) be
satisfied. Let, for a given x EX, there exist a subsequence {n'} of {n}
214
such that
weak-lim Tn'x = x0 exists.
(4)
n'--.co
Then Tx0
x0 and lim Tnx

1>--700
x0
Proof. We have TTn- Tn = n- 1 (P+l- T), and so, by (1),

lim (TTnx- Tnx) = 0. Thus, for any IE X', lim <TTn,x,l) =
n--.oo
lim <Tn, x, T' I> exists and
n--.oo
n--.oo
= n--.oo
lim <Tn' x, I) = <x0 , 1). Therefore
<x0 , I>= <Tx0 , I> and so, by the arbitrariness of IE X', we must have
Tx 0 = x0 .
We have thus T"x = Tmx0 + T"(x-x0 ) = x0 + Tm(x-x0 ) and
so Tnx = x0 + Tn(x-x0 ). But, (x-x 0 ) =weak-lim (x- Tn,x) and,
n--.oo
as proved above, (x- Tn'x) E R (I-T). Therefore, by Theorem 11 in

Chapter V, 1, (x-x 0 )ER(I -T)a. Thus, by Theorem 1, lim Tn (x-x0 ) = 0
and so we have proved that lim Tnx
n--.oo
n--.oo
= x0 .
Corollary. Let condition (1) be satisfied, and X be locally sequentially weakly compact. Then, for any x EX, lim Tnx = x 0 exists, and
n--.oo
the operator T 0 defined by T0 x = x0 is a continuous linear operator such

that
(5)
T 0 = T5 = TT0 = T 0 T,
R(T0 ) = N(I- T),
(6)
N(T0 ) = R(I- Tt = R(I- T 0 ).
(7)
Moreover, we have the direct sum decomposition
X= R (I-Tt Ell N (I-T),
(8)
i.e., any x E X is represented uniquely as the sum of an element E R (I-Tt

and an element EN (I-T).
Proof. The linearity of T0 is clear. The continuity of T0 is proved by
the equi-continuity of{Tn} implied by (1). Next, since Tx0 = x 0 , we have
TT0 = T 0 and so PT0 = T 0 , TnTo = T 0 which implies that T5 = T 0
On the other hand, Tn- TnT= n- 1 (T- yn+I) and (1) imply that
T 0 = T 0 T. The equality (6) is proved as follows. Let Tx = x, then
Tnx = x, T,.x = x and so T 0 x = x, that is, x E R (T0 ). Let conversely,
xE R(T0 ). Then, by T5 = T0 , we have T 0 x = x and so, by TT0 = T 0 ,
Tx = TT0 x = T 0 x = x. Therefore, the eigenspace of T corresponding
to the eigenvalue 1 of T is precisely the range R (T0 ). Hence (6) is proved.
Moreover, we have, by Theorem 1, N (T0) = R (I-Tt. But, by T5 = T 0 ,
we have R(I- T0 ) ~ N(T0 ), and if xE N(T0 ), then x = x- T 0 xE
R(I- T 0 ). Thus N(T0 ) = R(I- T 0 ).
4. Ergodi<; Theorems of the Hille Type Concerning Pseudo-resolvents
215
Therefore, by I= (I-- T0 ) + T0 and (6) and (7), we obtain (8).

Remark. The eigenspace N (AI - T) of T belonging to the eigenvalue
A with [A I= 1 may be obtained as R (T(A)), where T(A) x
lim n-1 .I
n-?00
(T jA)"' x.
m=l
The Mean Ergodic Theorem of J. von Neumann. Let (5, \8, m) be a

measure space, and P an equi-measure transformation of 5, that is, P
is a one-one mapping of 5 onto 5 such that P BE \8 iff B is E \8 and
m(P B)= m(B). Consider the linear operator Ton L2(5, \8, m) onto
itself defined by
(Tx) (s) = x(Ps), x E L2(5, \8, m).
(9)
By the equi-measurable property of P, we easily see that the operator T

is unitary and so the equi-continuity condition (1) is surely satisfied by
II P If = 1 (n = 1, 2, ... ). Therefore, by the sequential weak compactness
of the Hilbert space L2(5, \8, m), we obtain the mean ergodic theorem of
]. von Neumann:
tt
For any x E L 2 (5, \8, m), s-lim n- 1 .I Tmx

exists and T x0 = x0
n-?00
m=l
= x0 E L 2 (5, \8,
m)
(10)
Remark. Theorem 1 and Theorem 2 are adapted from K. YosiDA [3].

Cf. also S. KAKUTANI [1] and F. Rmsz [4]. Neumann's mean ergodic
theorem was published in J. VON NEUMANN [3].
4. Ergodic Theorems of the Hille Type Concerning Pseudo-resolvents

The notion of resolvent is generalized to that of pseudo-resolvent by
E. HILLE. We can prove ergodic theorems for pseudo-resolvents by a
similar idea to that used in the proof of the mean ergodic theorems in the
preceding section. See K. YosmA [4]. Cf. T. KATO [1]. These ergodic
theorems may be considered as extensions of the abelian ergodic theorems of E. HILLE given in E. HILLE-R. S. PHILLIPS [1], p. 502.
We shall begin with the definition of the pseudo-resolvent.
Definition. Let X be a locally convex complex linear topological
space, and L (X, X) the algebra of all continuous linear operators defined
on X into X. A pseudo-resolvent ];. is a function defined on a subset
D (]) of the complex A-plane with values in L (X, X) such that
];. - ],_, = (p- A)];.],_, (the resolvent equation).
(1)
Proposition. All ];,, AED(]), have a common null space, which we

denote by N (]), and a common range which we denote by R (]). Similarly, all (I- A];.), A ED(]), have a common null space, which we denote
216
by N (I - f), and a common range which we denote by R (I - f). Moreover, we have the commutativity:
];.],_. = ],_.];. (A, t-t ED(])).

Proof. By interchanging A and f-l in (1), we obtain
(2)
],_.- ];. = (A- t-t) ],_.];. = - (t-t- A) ]p];.,
and hence (2) is true. The first part of the Proposition is clear from (1)
and (2}. The second part is also clear from
(I- A];.) =(I- (A- t-t) ];.) (I- t-tfp)
(1')
which is a variant of (1).
Theorem 1. A pseudo-resolvent];. is a resolvent of a linear operator A
iff N(J) = {0}; and then R(J) coincides with the domain D(A) of A.
Proof. The "only if" part is clear. Suppose N (]) = {0}. Then, for
any AE D(J), the inverse Ti 1 exists. We have
AI--Ti 1 =t-ti-T; 1 (A,f-lED(J)).
(3)
For, by (1) and (2},

J;.Jp(u- r;l- t-tl+
r;ll =(A- t-t) Jd!J- Jd/Ju-;1- r;l)

=(A- t-t) ];]p- (Jp- ];.)
We thus put
= 0.
(4)
Then h.= (AI- A)-1 for A ED(]).

Lemma 1. We assume that there exists a sequence {A.n} of numbers
ED(]) such that
~An=
0 and the family of operators {An];.n} is equi-continuous.
Then we have
and hence
R(I- Jt
(5)
= 1.fx EX; lim An];. x = Ol,

n--+co
n
J
(6)
N(I- f) r\ R(I- Jt = {0}.
(7)
Proof. We have, by (1),

A];. (I- t-tlp)
= (1- f-l (t-t- A}-1)}.];.- A (A- t-t)- 1 t-tfw
(8)
Hence, by (5}, the condition x E R(I- f-l]p) = R(I- f) implies that

lim An];. x = 0. Let y E R (I- ft. Then, for any continuous semi-norm
n--+co
n
q on X and e > 0, there exists an x E R (I- f) such that q (y- x)

By (5), we have, for any continuous semi-norm q' on X,
<
e.
q' (An];.n (y- x)) < q (y- x) (n = 1, 2, ... )

with a suitable continuous semi-norm q on X. Therefore, by An];.nY =
AnfAnx + An];.n (y- x), we must have lim AnfAnY = 0.
n--+co
4. Ergodic Theorems of the Hille Type Concerning Pseudo-resolvents
217
lim A,];.n x = 0. Then, for any continuous semiLet, conversely, n---->00
norm qon X ande> 0, there exists a A, such that q(x-(x -A,];.nx))< e.

Hence x must belong to R (I- A,];.nt = R (I- ])a.
Lemma 1'. We assume that there exists a sequence {A,} of numbers
E D (]) such that
lim iAn I= ex:> and the family of operators {A,];.n} is equi-continuous. (5')
n--->00
Then we have
and hence
R(]t = {xE X;~ A,.];.nx = x},
(6')
N(]) f\ R(])a = {0}.

1
"'
A];.],.=;:A];.]
1,-;:A];.
(7')
+ ],..
Hence, by (5'), the condition xE R(]p) = R(]) implies that

lim A,];. x = x. Let y E R (])a. Then, for any continuous semi-norm q
n--->00
on X and e > 0, there exists an x E R (]) such that q (y- x)

(5'), we have, for any continuous semi-norm q' on X,
q' (A,];.n (y- x))
<
q(y- x)
<
e. By
(n = 1, 2, ... )
with a suitable continuous semi-norm q on X. Therefore, by (5') and

An];.nY- y = (A,];.nx- x)
we must have lim An];.nY

n--->00
+ (x- y) + A,];.n (y- x),
lim An];.n x = x. Then, for any continuous semiLet, conversely, n---->eo

norm q on X and e > 0, there exists a An such that q(x- A,];.nx) <e.
Hence x must belong to R (];.nt = R (ft.
Theorem 2. Let (5) be satisfied. Let, for a given x EX, there exist a
subsequence {n'} of {n} such that
A,];.n ,x = xh exists.
weak-lim
n---->00
(9)
lim An];.n x and xh EN (I-]), Xp = (x- xh) E R (I- JtThen xh = n---->eo

Proof. Setting fl =An' in (1') and letting n'--+ ex:>, we see, by (5),
that (I- A];.) x = (I- A];.) (x- xh), that is, (I- A];.) xh = 0. Hence
xh E N (I - ]) and so
(10)
A,.];.nX = Xh + An];.n (x- xh).
Therefore we have only to prove that lim A,];.n (x- xh) = 0, or, by
n--->00
Lemma 1, (x -- xh) E R (I- ft. But (x- An];.nx) E R (I-]), and so,
by Theorem 11 in Chapter V, 1, we must have (x- xh) E R (I- ItCorollary 1. Let (5) be satisfied, and let X be locally sequentially
weakly compact. Then
(11)
X= N(I-]) EB R(I- ])a.
218
Proof. For any x EX, let xh
= n-+00
lim Anli.nx and Xp = (x- xh) be the
components of x inN (I-!) and R (I- Jt, respectively.

Theorem 2'. Let (5') be satisfied. Let, for a given x EX, there exist
a subsequence {n'} of {n} such that
weak-lim
An];.n ,x = xh' exists.
(9')
n---+00
Then xh'
~ Anfi.nx and xh'
E R(Jt, Xp = (x- Xh) E N(J).
Proof. Setting fl =An in (8) and letting n' ~ oo, we see, by (5'), that
J.]J. (x- xh') = 0, that is, (x- xh.) EN(]). Hence
(10')
Therefore we have only to prove that n---+00

lim A,.];.n xh' = x 11., or, by
Lemma 1', xh' E R(J)a. But Anf;.n,xh' E R(J), and so, by Theorem 11 in
Chapter V, 1, we must have xh' E R (ft.
Corollary 1'. Let (5') be satisfied, and let X be locally sequentially
weakly compact. Then
(11')
X = N (]) E9 R Ut .
Proof. For any x E X, let xh'
= n---+00
lim An];. n x and Xp = (x- x 11 .) be
the componerts of X in R(Jt and N(J), respectively.

Remark. As a Corollary we obtain: In a reflexive B-space X, a pseudoresolvent !J. satisfying (5') is the resolvent of a closed linear operator A
with dense domain iff R (Jt =X. This result is due to T. KATO, loc.
cit. The proof is easy, since, by Eberlein's theorem, a B-space X is
locally sequentially weakly compact iff X is reflexive.
5. The Mean Value of an Almost Periodic Function

As an application of the mean ergodic theorem we shall give an
existence proof of the mean value of an almost periodic function.
Definition 1. A set G of elements g, h, ... is called a group if in G a
product (in general non-commutative) gh of any pair (g, h) of elements
E G is defined satisfying the following conditions:
(1)
(2)
ghEG,
(gh) k = g(hk) (the associativity),
there exists a unique element e in G such that eg =
g e = g for all g E G; e is called the identity element
of the group G,
(3)
for every element g E G, there exists a uniquely determined element in G, which is denoted by g-1 , such
that gg-1 = g-1 g = e; the element g-1 is called the
inverse element of g.
(4)
219
5. The M:ean Value of an Almost Periodic Function
Clearly, g is the inverse of g-1 so that (g-1 )-1 = g. A group G is said to be

commutative if gh = hg for all g, hE G.
Example. The totality of complex matrices of order n with determinants equal to unity is a group with respect to the matrix multiplication;
it is called the complex unimodular group of order n. The identity of this
group is the identity matrix and the inverse element of the matrix a is
the inverse matrix a-1 . The real unimodular group is defined analogously.
These groups are non-commutative when n > 2.
Definition 2 (J. VON NEUMANN [4]). Given an abstract group G. A
complex-valued function f (g) defined on G is called almost periodic on G
if the following condition is satisfied:
the set of functions {g. (f, h); s E G}, where fs (g, h) =
f(gsh), defined on the direct product GxG ;s totally
bounded with respect to the topology of m orm convergence on G X G.
(5)
Example. Let G be the set R 1 of all real numbers in which the group
multiplication is defined as the addition of real numbers; this group R 1
is called the additive group of real numbers. The function f(g) = ei"'g,
is almost periodic on R 1 This
where (X is a real number and i =
we easily see from the addition theorem f(gsh) = ei"'g ei"'s ei"'h and the
fact that {ei"'1 ; t E R 1} is totally bounded as a set of complex numbers
of absolute value 1.
Proposition 1. Suppose f(g) is an almost periodic function on G. If
we define, following A. WEIL,
V-1,
dis(s,u)
then
(6)
suplf(gsh)-/(guh)l,
g,hEG
(7)
dis(s, u) = dis(asb, au b).
Proof. Clear from the definition of the group.

Corollary 1. The set E of all elements s which satisfies dis(s, e) = 0
constitutes an invariant subgroup in G, that is, we have
if ct. e2 E E, then cte2 E E and ae1 a-1 E E for every a E G.
(8)
Proof. Let dis (e1 , e) = 0, dis (e2 , e) = 0. Then, by (7) and the triangle
inequality, we obtain
dis(cte2 , e)< dis(e1 e2 , e1 e)
+ dis(cte. e)= 0 + 0 =
0.
Similarly we have dis(ae1 a-1 , e)= dis(ae1 a-1 , aea-1 ) = 0 from

dis (ev e) = 0.
u (mod E) when su-1 E E, then s = u
Corollary 2. If we write s
(mod E) is equivalent to dis (s, u) = 0.
Proof. Clear from dis(su-1 , e)= dis(s, eu) = dis(s, u).
220
Corollary 3. The concept s _ u (mod E) has all the general properties

of equivalence, namely
s _ s (mod E),
s - u (mod E), then u-s (mod E),
if s1 s2 (mod E) and s2 - s3 (mod E), then s1 _ s3
(mod E)
(9)
(10)
(11)
Proof. Clear from Corollary 2 and the triangle inequality for the
dis(s, u)o
Hence, as in the case of the factor space in a linear space, we can
define the factor group or residue class group GfE as follows: we shall
denote the set of all elements E G equivalent (mod E) to a fixed element
x E G by~"'' the residue class (mod E) containing x; then the set of all
residue classes ~"' constitutes a group G/E by the notion of the product
(12}
To justify this definition (12) of the product, we have to show that
gx~y
~xy
if x1 - x2 (mod E), y1 - y2 (mod E), then x1y1
x2 y2 (mod E)
(13)
This is clear, since we have by (7) and Corollary 2,

dis (x1y1, x2 y2) < dis (x1y1, x2 YI) + dis (x 2 yv x2 y2)
= dis(x1, x2) + dis(y1, y2) = 0 + 0 = Oo

Since the function f (x) takes the same constant value on the residue
class~"'' we may consider f(x) as a function F(~.,) defined on the residue

class group GfEo
Corollary 4. The residue class group is a metric space by the distance
dis(~.,, ~y)
Proof. x -
= dis (x, y)
(14)
x1 (mod E) and y - y1 (mod E) imply
dis (x, y) <dis (x, x1) +dis (xv y1) +dis (Yv y) = 0 +dis (xv y1) + 0
and dis(x1, YI) < dis(x, y) to the effect that dis(x, y) = dis(x1, y1)o Thus
(14) defines a distance in G/Eo
Corollary 5. The group GfE is a topological group with respect to the
distance dis(~.,, ~y), that is, the operation of multiplication ~x~y is continuous as a mapping from the product space (G/E) X (G/E) onto GfE,
and the operation ~;; 1 is continuous as a mapping from GfE onto GfE.
dis (su, s' u') < dis (su, s' u) + dis (s' u, s' u') =dis (s, s') + dis (u, u')
and
dis (s-1, u-1) = dis (s s-1u, su-1u) = dis (u, s) = dis (s, u).
We have thus proved the following
221
Theorem 1 (A. WErL). The topological group GfE, metrized by (14),

is totally bounded, and the function f (x) gives rise to a function F (~x)
( = f (x)) which is uniformly continuous on this group Gf E.
Proof. The uniform continuity of the function F (~J is clear from
IF (~x) - F (~y) I = I/ (x) -I (y) I < dis (x, y) = dis (~x ~y).
The almost periodicity of the function f(x) implies, by (7) and (14), that
the metric space GfE is totally bounded.
By the above theorem, the theory of almost periodic functions is
reduced to the study of a uniformly continuous function f (g), defined on a
totally bounded topological group G, metrized by a metric dis (g1 , g2)
satisfying condition (7). By virtue of this fact, we shall give a proof for
the existence of the mean value of an almost periodic function.
Since G is totally bounded, there exists, for any e > 0, a finite
system of points gl> g2 , . , gn such that min dis (g, gi) < efor any gE G.
l;;>>;;>n
Hence, collecting these finite systems of points corresponding to e =

1, 2-1 , 3-1 , . . . , respectively, we see that there exists a countable
system {gj} of points E G such that {gi} is dense in G. We take a sequence
co
of positive numbers
aj
such that .2: ai = 1. Let C (G) be the set of all

J~l
uniformly continuous complex-valued functions h (g) defined on G. C (G)

is a B -space by the operation of the function sum and the norm II h II =
sup lh(g) I We define an operator T defined on C(G) into C(G) by
gEG
(15)
By the uniform continuity of h (g) on G, there exists, for any e > 0, a

I')> 0 such that dis(g, g') 
0 and
co
i~ ai =
1, that Tis a bounded linear opera-
tor on C (G) into C (G). By the same reasoning we see that the set of
functions hn (g) defined by
n
hn (g) = n- 1 2: (Tm h) (g), which is of the form

m~l
hn(g)
co
= .2: (Jjh(gjg)
J~l
with
(Jj
>
0, .2: (Jj
J~l
(16)
1,
is equi-bounded and equi-continuous with respect to n. Hence, by the

Ascoli-Arzela theorem, the sequence {hn (g)} contains a subsequence which
is uniformly convergent on G.
Therefore, by the mean ergodic theorem, there exists an h* (g) E C (G)
such that
(17)
lim sup lhn(g) -h*(g) I= 0 and Th* = h*.
n->00
222
Proposition 2. h* (g) is identically equal to a constant.

Proof. We may assume, without losing the generality, that h(g) and
h* (g) are real-valued. Suppose that there exist a point g0 E G and a positive constant b such that
h* (g0 ) < fJ- 215, where fJ =sup h* (g).

gEG
By the continuity of h* (g), there exists a positive number e such that

dis (g', g") :S E implies Ih* (g') - h* (g") I < b; in particular, we have
h* (g") < fJ- b whenever dis (g0 , g") < e. Since the sequence {g;} is dense
in G, there exists, for any e > 0, an index n such that, for any g E G, we
have m~n dis (g. gj) <e. Hence, by (7). we have, for any g E G,
l~J~n
min dis(g0 ,gjg) <e.
l~J~n
Let the minimum be attained at
= fo- Then
00
h* (g)= (Th*) (g)= j~ cx;h* (gjg)
~cxj,({J-b)
+ (1-cxj,){J =fJ-cxj,b<fJ,
contrary to the assumption that g was an arbitrary point of G. Therefore

h* (g) is identically equal to a constant.
Definition 3. We shall call the constant value h* (g) the left mean
value of h(g) and denote its value by M~(h(g)):
M~ (h (g)) = lim n- 1 i: (Tm h) (g).

m=l
(18)
n--KX)
Theorem 2 (J. voN NEUMANN). We have

M~(cxh(g))
M~(hJ{g)
+ h2 (g)) =
cxM~(h(g)),
(i)
+ M~(h2 (g)),
M~(ht(g))
M~ (1) = 1,
(ii)
(iii)
if h(g) ~ 0 on G, then M~(h(g)) > 0;

if, moreover, h(g) $0, then M~(h(g))> 0,
IM~(h(g)) I< M~(lh(g)
1).
(iv)
(v)
M~(h(g)) = M~(h(g)),
(vi)
M~ (h (ga)) = M~ (h (g)) 1
(vii)
M~(h(ag)) = M~(h(g))~
(vii')
M~(h(g- 1 ))
= M~(h(g)).
(viii)
Proof. By definition (18) it is clear that (i) (ii) (iii) the first part of
1
(v) and (vi) are true. The truth of (vii) is proved by Proposition 2.
(vii') is proved as follows.
(iv)~
223
Starting with the linear operator T' defined by

n
(T' h) (g)= ..I IX;h(gg;)'

J=1
we can also define a right mean M;(h(g)) which, as a functional of h(g),

satisfies (i), (ii), (iii), the first part of (iv), (v), (vi) and (vii'). We thus
have to prove that the left mean M~(h (g)) coincides with the right mean
M;(h(g)). By its definition of the left mean, there exist, for any e > 0,
a sequence of elements {kj} ~ G and a sequence of positive numbers pi
00
with ..I fJ;

J=1
1 such that
sup
g
/.i
J=1
fJ;h(k;g)- M!(h(g))l <e.
Similarly, there exist a sequence of elements {s;}

00
positive numbers Y; with ..I Y;
/.i
J=1
sup
g
J=1
(19)
G and a sequence of
1 such that
y;h (g s;) - M; (h (g)) I <e.
(20)
We have, from (19) and (vii)

sup ~ YfJ;h (k;g si) - M~(h (g)) I < e,
I
, ,J
1
and, similarly from (20) and (vii'),

sup (~ YfJ;h (k;gsi)- M; (h (g)) I <e.
g
t,j
Hence we must have M~(h(g)) = M;(h(g)).

We next remark that a linear functionalMg(h(g)) defined on C(G) is
uniquely determined by the properties (i), (ii), (iii), the first part of (iv),
(v), (vi) and (vii) (or (vii') as well). In fact, we have, by (20),
00
M~(h(g))
-e < _.Iy.h(gsi) < M;(h(g))

=1
+e
for real-valued h(g).
Hence, for real-valued h(g), Mg(h(g)) must coincide with the right mean
M;(h(g)) and hence with the left mean M~(h(g)) as well. Therefore we
see that we must have Mg = M; = M~. Being equal to the right mean,
Mg must satisfy (vii'). Moreover, since M~(h(g-1 )) satisfies, as a linear
functional, (i), (ii), (iii), the first part of (iv), (v), (vi) and (vii'), we must
have M~(h(g-1)) = M;(h(g)) = M~(h(g)).
Finally we shall prove the last part of (iv). Suppose h (g0) > 0. For
any e > 0, there exists, by the total boundedness of G, a finite system
of elements s11 s2 , , sn such that
m_in sup lh(gs,)- h(gs) I< e
1~::=;;n
224
for all s E G. This we see by the uniform continuity of h (g) and the fact
that dis(gsi, gs) = dis(si, s). Hence, fore= h(g0 )/2, we obtain, for any
s E G, a suffix i, 1 h(g0 )/2.
Thus, by the non-negativity of the function h(g), we obtain
n
..~h(g0 sj 1 s)>h(g0 )/2>0 forall sEG.
=1
Therefore, by taking the right mean of both side,, we have
M;(~ h(g0 sj 1 s)) =
n M;(h(s)) > h(g0 )/2 > 0.
Remark. The happy idea of introducing the distance (1) is shown in

A. WEIL [1]. The application of the mean ergodic theorem to the existence proof of the mean value is due to the present author. See also
W. MAAK [1].
6. The Resolvent of a Dual Operator

Lemma 1. Let X and Y be complex B-spaces. Let T be a linear
operator with D (Tt =X and R (T) s;; Y. Then (T')- 1 exists iff R (Tt = Y.
Proof. If T'
= 0, then
yt
<x, T' Yri> = <T x, Yci> = 0 for all
xED (T),
and hence Yci (R (Tt) = 0. Thus R (Tt = Y implies yci = 0 and so T'
has an inverse. On the other hand, if y0 E R (Tt, then, the Hahn-Banach
theorem asserts that there exists a continuous linear functional Yci E Y'
such that Yci (y0 ) = 1 and Yci (R (Tt) = 0. Hence <T x, Yri> = 0 for all
xED (T), and so Yci ED (T') and T' Yri = 0, whereas Yci (y0 ) =I= 0, i.e.,
yci =I= 0. Therefore, the condition R (Tt =I= Y implies that T' cannot
have an inverse.
Theorem 1 (R. S. PHILLIPS [2]). Let T be a linear operator with an
inverse and such that D(Tt =X and R(Tt = Y, where X andY are
B-spaces. Then
(T')-1 = (T-1)'.
(1)
T- 1 is bounded on Y iff T is closed and (T')- 1 is bounded on x;.

Proof. (T-1)' exists because D (T-1) = R (T) is dense in Y. (T')-1
exists by Lemma 1. We have to show the equality (1). If yE R(T) and
y* E D (T'), then
<y, y*) = <TT-1 y, y*) = <T-1 y, T'y*).
Hence R (T') s;; D((T-1 )') and (T- 1 )' (T' y*) = y* for ally* ED (T'). Thus
(T-1)' is an extension of (T')-1 . Next, if xE D(T), then
<x, x*) = <T-1 Tx, x*) = <Tx, (T-1)' x*) for all x* E D((T-1 )').
225
7. Dunford's Integral
Hence R((T-1 )') s:; D (T') and T' (T-1 )' x* = x* for all x* E D((T-1)').
Thus (T-1 )' is a contraction of (T')-1 . Therefore we have proved (1).
If, in addition, T-1 is bounded on Y, then (T-1 )' is also bounded.
Conversely, if (T')-1 is bounded on x;, then, for all X E R (T) and
x* EX', we have, by (1),
I<T-1 x, x*) I =
<
<x, (T- 1)' x*) I =
I<x, (T')-1 x*) I
II (T')-1 11 llx* II llxll
Since T-1 is closed and R(T)a = Y, T-1 must be bounded.

Lemma 2. Let T be a linear operator with D (T)a = X and R (T) s:; Y,
where X and Yare B-spaces. If R (T') is weakly* dense in X', then T has
an inverse.
Proof. Suppose that there exists an x0 # 0 such that Tx0 = 0. Then
<x0 , T' y*) = <T x0 ,
0 for all y* E D (T').
y*~
This shows that the weak* closure of R (T') is a proper linear subspace
of X', contrary to the hypothesis.
Theorem 2 (R. S. PHILLIPS [2]). Let X be a complex B-space, and T a
closed linear operator with D (T)a = X and R (T) s:; X. Then
e(T) = e(T')
and R(A; T)' = R(A; T')
for
AE e(T).
(2)
Proof. If AEe(T), then, by Theorem1, AEe(T') and R(A;T)'=

R(A; T'). On the other hand, if AE e(T'), then Lemma 2 shows that
(AI- T) has an inverse (U- T)-1 which is closed with (AI- T).
Lemma 1 then shows that D((AI- T)-1) = R (AI- T) is strongly dense
in Y. Hence, by Theorem 1, AE e(T).
Let X be a complex B-space and T a bounded linear operator E L (X, X).
We shall define a function f(T) ofT by Cauchy's type integral
/(A) R(A; T) dA.

c
To this purpose, we denote by F (T) the family of all complex-valued
f(T)
(2ni)-1
functions f (A) which are holomorphic in some neighbourhood of the spectrum a (T) of T; the neighbourhood need not be connected, and can
depend on the function /(A). Let /E F(T), and let an open setU ~ a(T)
of the complex plane be contained in the domain of holomorphy of f,
and suppose further that the boundary au of u consists of a finite number of rectifiable Jordan curves, oriented in positive sense. Then the
bounded linear operator f (T) will be defined by
f(T) = (2ni)- 1
15
Yoslda, Functional Analysis
f /(A) R(A; T) dA,
8U
(1)
226
and the integral on the right may be called a Dunford's integral. By

Cauchy's integral theorem, the value f(T) depends only on the function f
and the operator T, but not on the choice of the domain U.
The following operational calculus holds:
Theorem (N. DuNFORD). Iff and g are in F(T), and x and {J are
complex numbers, then
xf + {Jg is inF(T) and xf(T)
+ {Jg(T) =
(xf + {Jg)(T),
(2)
I g is in F(T) and f(T) g(T) =(/g) (T),

if
f has the Taylor expansion /(A.) =
vergent
00
neighbourhood
(3)
n~o XnAn con-~
U of a (T),
then
(4)
j
1, 2, ... ) be holomorphic in a fixed l
/ (T) = 2: Xn yn (in the operator norm topology),

n~o
let
In E F(T)
(n =
neighbourhood U of a(T). If fn(A.) converges to f(A.)

uniformly on U, then fn(T) converges to f(T) in the
operator norm topology,
(5)
if/EF(T), thenfEF(T') andf(T') =f(T)'.
(6)
Proof. (2) is clear. Proof of (3). Let U1 and U2 be open neighbourhoods

of a(T) whose boundaries 8U1 and 8U2 consist of a finite number of
rectifiable Jordan curves, and assume that U1 + 8U1 ~ U2 and that
u2 + 8U2 is contained in the domain of holomorphy of I and g. Then,
by virtue of the resolvent equation and Cauchy's integral theorem, we
obtain
f f(A.) R(A.; T) d).. f g(ft) R(ft; T) dft

f jf(A.)g(!-l)(ft-A.)- (R(A.;T)-R(f-t;T))dA.df-t
f(T) g(T) = - (4n2)-1
a~
=-(4n2)-1
a~
au, au,
f f (A.) R (A.; T) . 1
f (2ni)-1 f (#- A.)-1 g (#) df-tl.J dA.
(2ni)-1 f g(f-t) R(f-t; T). f(2ni)-1 f (f-t-A.)-1/(A.) d)..~ df-t
= (2ni)-1
a~
= (2ni)-1
a~
a~
a~
f f(A.) g(A.) R(A.; T) dA. =(/g) (T).

au,
Proof of(4). By hypothesis, U must contain a circle {A.; lA. 1 <ra (T) + e}
e > 0, in its interior, where ra(T) is the spectral radius ofT (Theorem 4
00
in Chapter VIII, 2). Hence the power series f(A.) = 2: XnAn converges
uniformly on the circle C ={A.; lA. I < ra (T)
n~o
+ e}, for some e >
0. Hence,
227
by Cauchy's integral theorem and Laurent's expansion R(A.; T)
00
~ A.-nyn- 1 of R(A.; T) (Chapter VIII, 2, (5)),
11=1
(5) is proved by (1), and (6) is also proved by (1) and formula (2) of
the preceding section.
Corollary 1 (Spectral Mapping Theorem). If f is in F(T), then
f(a(T)) = a(f(T)).
Proof. LeU E a (T), and define the function g by g (t-t) = (I (A.) - f (t-t)) /
(A.-t-t) By the Theorem, f(A.)I-f(T)=(AI-T)g(T). Hence, if
(!(A.) I - f(T)) had a bounded inverse B, then g(T) B would be the bounded inverse of (AI- T). Thus A.E a(T) implies that /(A.) E a(f(T)). Let,
conversely, A. E a(/(T)), and assume that A. E f(a(T)). Then the function
g(t-t) = (/(t-t) -A.)-1 must belong to F(T), and so, by the preceding
Theorem, g (T) (I (T) - U) = I which contradicts the assumption
A. E a (I (T)).
Corollary 2. If /E F(T), gE F(t(T)) and h(A.) = g(f(A.)), then h is
in F(T) and h(T) = g(f(T)).
Proof. That hE F (T) follows from Corollary 1. Let U 1 be an open
neighbourhood of a(t(T)) whose boundary U 1 consists of a finite number
of rectifiable Jordan curves such that u1 + 8U1 is contained in the
domain of holomorphy of g. Let U 2 be a neighbourhood of a(T) whose
boundary 8U2 consists of a finite number of rectifiable Jordan curves
such that U 2 + 8U 2 is contained in the domain of holomorphy off and
f(U 2 + 8U 2 ) ~ U 1 Then we have, for A. E 8Uv
f (A.- f(t-t))-1 R(t-t; T) dft,
R(A.; f(T)) = (2ni)-1
i!U,
since the right hand operator S satisfies, by (3), the equation

(AI- f(T)) S = S(U- f(T)) =I. Hence, by Cauchy's integral thearem,
g(A.)R(A.; f(T)) dA.
g(f(T)) = (2ni)-1
i!U,
= (-4n2)-1
f f g(A.) (A.- f(t-t))-1 R(t-t; T) dt-t d}.
i!U1 i!U,
(2ni)-1
f R(t-t; T) g(f(t-t)) dt-t =
h(T).
i!U,
Remark. The introduction of the operational calculus based on a

formula like (1) goes back to the investigations by H. PoiNCARE on
15*
228
continuous groups (1899). The exposition of the operational calculus in

this section is adapted from N. DuNFORD-J. ScHWARTZ [1]. In the next
chapter on semi-groups, we shall frequently make use of Dunford's integral for a closed unbounded operator T.
8. The Isolated Singularities of a Resolvent

Let Ao be an isolated singular point of the resolvent R (A; T) of a closed
linear operator Ton a complex B-space X into X. Then R (A; T) can be
expanded into Laurent series
R(A; T)
00
= .I
(A-Aot An, An= (2ni)- 1
n=-00
(J.-.1.0)-n- 1 R(J.; T) d).,

(1)
where C is a circumference of sufficiently small radius: IA - .1.0 I = e

such that the circle IA - .1.0 I < e does not contain other singularities
than ;. = ).0 , and the integration is performed counter-clockwise. By
virtue of the resolvent equation we obtain
Theorem 1. A's are mutually commutative bounded linear operators
and
T Ak = AkT (k = 0, 1, 2, ... ) ,
AkAm
k > 0, m < - 1,
0 for
(2)
An= (-1t A 0+1 (n > 1),

A-P-Hl = A_pA-q (p, q > 1).
Proof. The boundedness and the mutual commutativity of A's

and the commutativity of A's with T are clear from the integral
representation of A's.
We substitute the expansion of R (J.; T) in the resolvent equation
R (J.; T)- R (p; T) = (p- J.) R (J.; T) R (p; T), and obtain
oo
(A.- Ao)k- (p,- Ao)k-
oo
k=~oo Ak (A.- Ao)- (p,- An) - - k,moo AkAm (A- Ao) (p- Ao) .
The coefficient of An on the left is
+ (A- J.0t- 2 (p- .1.0) + + (p- Aot- 1 , n >

-{(J.- J.0t (p- .1.0)- 1 + (J.- Aot+l (p- .A.0)- 2 +
(J.- Aot- 1
(J.- Ao)- 1 (p- J.0r}, n
<
1,
0.
Hence the terms containing (A- )"')k (p - Ao)m with k > 0 and m < - 1
are missing so that we must have the orthogonality AkAm = 0 (k > 0,
m <- 1). Hence
00
R+ (J.; T) = .I An (J.- Aot

n=O
and
R- (J.; T) =
-1
.I An (A- J.0 )n
n=-00
229
must both satisfy the resolvent equation. Substituting the expansion

of R+ (A.; T) in the resolvent equation
R+ (A.; T) - R+ (A.; T) = (t-t - A.) R+ (A.; T) R+ (t-t; T) ,

we obtain, setting (A.- A.0} = h, (t-t- A.0) = k,
P~ Ap(hP-kP) = (k-h)~~0 AphP)C~Aqkq).

Hence, dividing both sides by (k- h), we obtain
,I A (hP- 1 + hP- 2 k
P=1 p
00
+ + kP- 1) =
00
,I hp kq A A
p,q=O
p q
so that we have -AP+q+ 1 = ApAq (p, q > 0). Thus, in particular,

A 1 = -A6, A 2 = -A 1A 0 = (-1) 2A8, ... , An= (-1t A 0+1 (n > 1).
Similarly, from the resolvent equation for R-(A.; T), we obtain,
setting (A.- A.0 )-1 = h, (t-t- A.0)-1 = k,
00
Y' A
p"':::1
-p
(hP- 1 + hP- 2 k
+ . + kP- 1) = p,q=1
,I hp- 1 kq-l A A
-p -q'
00
so that we have A-p-q+ 1 = A_pA-q (p, q > 1). In particular, we have

A_1
= A:_1, A_2 = A_1A-2, ... , A_n = A_1A-n
(n
> 1).
Theorem 2. We have
An
(T -- Aol) An +I (n > 0) ,
= A-(n+Il = (T- A.0 It A_ 1

(T-A. 0 I) A 0 = A_ 1 -I.
(T- A.ol) A_n
(n 2: 1),
(3)
Proof. By the integral representation of Ak, we see that the range

R (Ak) is in the domain of T, so that we can multiply Ak by Ton the left.
Thus our Theorem is proved by the identity
Theorem 3. If A.0 is a pole of R(A.; T) of order m, then A.o is an eigenvalue ofT. We have
R(A_ 1) =N((A.0 1 -Tt) and R(J -A_1) = R((A.0 1 -Tt) for n> m, (4)
so that, in particular,
230
Proof. Since A_ 1 is a bounded linear operator satisfying A:. 1 = A_ 1,

it is easy to see that
(6)
X2
N(A_ 1)
R(I- A_1), and put also
R(A_ 1 ), N,.
N((J..oi- Tt) and R,.
We put X 1
= R,.(().0 I- Tt).
(7)
Let x EN,. wheren > 1. Then we see, by (T -A.0 It A,._ 1 = (T -A.0 I)A 0
= A_1 - I,
that
0 = A,._ 1(T- A.0 I)" x = (T- A.0 It A,._ 1 x =
(T- Aoi) Aox = A_lx- X so that X= A_lx E x2. Thus N,. with n > 1
belongs to X 2 Let, conversely, x E X 2 Then we have x = A_ 1 y and so
x = A_ 1A_ 1y = A_ 1x by A_ 1 =
1 ; consequently, we have (T- A. 0 I)"
x = A-(n+ 1 ) x by (T- A.0 I)" A_ 1 = A-(n+l) Since A-(n+l) = 0 for n > m
by hypothesis, it follows that X 2 ~ N, for n > m and so
A:.
(8)
Because (T- A.0 I) A_m = A-(m+l) = 0 and A_m =F 0, the number A.0
is an eigenvalue of T.
We see that X 1 = N(A_ 1) = R (I- A_ 1) ~ R,. by (T- A.0 It An_ 1 =
A_ 1 -I. If n > m, then xE R,. (\ N,. implies x = 0. For, if x = (A. 0 I -Tty
and (A.0 I - T)"x = 0, then, by (8), y E N 2 ,. = N,. and therefore x = 0.
Next suppose x E R,. with n > m, and write x = x 1 + x2 where x 1 =
(I- A_1) X E X1, x2 = A_1 X E X2. Then, since x1 ~ R,., x2 = X - x1 E R,..
But x 2 E X 2 = N,. by (8), and so x 2 E R,. f\ N,., x 2 = 0. This proves that
X= x1 E X1. Therefore we have proved that R,. = x1 if n > 111.
Theorem 4. If, in particular, Tis a bounded linear operator such that
X 2 = R (A_ 1 ) is a finite dimensional linear subspace of X, then A.0 is a pole
of R(A.; T).
Proof. Let x 1 , x2 , . . . , xk be a base of the linear space X 2 . Since
x1 , T x1 , T 2 x1 , . . . , Tk x1 are linearly dependent vectors of X 2 , there
exists a non-zero polynomial P 1 (A.) such that P 1 (T) x1 = 0. Similarly
there exist non-zero polynomials P 2 (A.), ... , Pk (A.) such that Pi (T) xi = 0
k
(f = 2, 3, ... , k). Then for the polynomial P(A.) =.II Pj(A.), we must
have P(T)xi
Let
= 0 (f
= 1, ... , k) and hence P(T)x
P (A.) = iX
J=1
= 0 for every xE X 2
.II (A.- A.i)"; (iX =F 0)
J=O
be the factorization of P (A.). Then we can prove that (T- A.0 I)" x = 0
for every X E x2. Assume the contrary, and let Xo E x2 be such that
(T- A.oi)" x 0 =F 0. Then, by P(T) x0 = 0, we see that there exist at
least one Aj (f =F 0) and a polynomial Q(A.) such that
(T -Aji) Q (T) (T -A.oi)" x 0 = 0
231
and y = Q(T) (T- Aol)'" x 0 =1=- 0. Thus y E X 2 is an eigenvector of T

corresponding to the eigenvalue Aj. Hence (..U - T) y = (A -A;) y and
so, multiplying both sides by R(A.; T), we obtain y =(A.-A;) R(A.; T)y
which implies
y = A_ 1 y = (2ni)- 1 R(A.; T) y dA. = (2ni)- 1 f (A.- A.i)- 1 y dA. = 0,
c
c
sufficiently small. This
centre
as
Ao
with
C
circumference
the
by taking
is a contradiction, and so there must exist a positive integer m such
that (T- Aol)m X 2 = 0. Thus, by X 2 = R (A_1) and (T- Aol)" A_1 =
A-(n+l) we see that A-(n+l) = 0 for n > m.
Comments and References
Section 6 is adapted from R. S. PHILLIPS [2]. Section 8 is adapted
from M. NAGUMO [1] and A. TAYLOR [1]. Parts of these sections can
easily be extended to the case of a locally convex linear topological space.
See, e.g., section 13 of the following chapter.
IX. Analytical Theory of Semi-groups

The analytical theory of semi-groups of bounded linear operators in a
B-space deals with the exponential functions in infinite dimensional
function spaces. It is concerned with the problem of determining the
most general bounded linear operator valued function T(t), t > 0, which
satisfies the equations
T(t + s) = T(t) T(s), T(O) =I.
The problem was investigated by E. HILLE [2] and K. YosmA [5] independently of each other around 1948. They introduced the notion of the
infinitesimal generator A of T (t) defined by
A = s-lim t- 1 (T (t) -I),
l.j.O
and discussed the generation of T (t) in terms of A and obtained a characterization of the infinitesimal generator A in terms of the spectral
property of A.
The basic result of the semi-group theory may be considered as a
natural generalization of the theorem of M. H. STONE [2] on one-parameter group of unitary operators in a Hilbert space, which will be explained in a later section. Applications of the theory to stochastic processes
and to the integration of the equations of evolution, which include diffusion
equations, wave equations and Schrodinger equations, will be discussed
in Chapter XIV.
In this chapter, we shall develop the theory of semi-groups of continuous linear operators in locally convex linear topological spaces rather
than in Banach spaces.
232
1. The Semi-group of Class (C0 )

Proposition (E. HILLE). Let X be a B-space, and T 1, t > 0, a oneparameter family of bounded linear operators E L (X, X) satisfying the
semi-group property
TITs= T 1+s for t, s > 0.
(1)
If p(t) =log IIT1 II is bounded from above on the interval (0, a) for each
positive a, then
lim t- 1 log II T 1 ll = inf t- 1 log II T 1ll
(2)
1>0
~~
Proof. We have p(t + s) < p(t) + p(s) from IIT1+sll = IIT1 Tsll ::S
Let f3 = inft-1 P(t). f3 is either finite or - <Xl. Suppose
II T II II Ts II
1
1>0
that f3 is finite. We choose, for any e > 0, a number a> 0 in such a way
that p (a) < ((3 + e) a. Let t > a and n be an integer such that na <
t < (n + 1) a. Then
p(t-na)<nap(a)
p(t-na)
f3 <p(t)<pjna)
=t=t+
t
=ta+
t
<na(f3
+e +
p(t-na)
t
By hypothesis, p (t- na) is bounded from above as t--+ <Xl. Thus,

letting t--+ <Xl in the above inequality, we obtain lim t-1 p (t) = (3. The
l-+00
case f3 = - <Xl may be treated similarly.

Definition 1. If {T1 ; t > 0} ~ L (X, X) satisfy the conditions
T 1 Ts = T 1+s (for t, s
> 0),
(1')
T 0 =I,
(3)
s-lim T 1 x = T 1,x for each t0 > 0 and each x EX,
(4)
t---+t 0
then {T1} is called a semi-group of class (C0 ).

In virtue of the Proposition, we see that a semi-group {T1} of class
(C0) satisfies the condition
IIT1 II <
ef31
(for 0 < t
<
(5)
<Xl),
with constants M > 0 and f3 < <Xl.

The proof is easy. We have only to show that II T 1 II is, for any interval
(0, a) with<Xl >a> 0, bounded on (O,a). Assumethecontraryandletthere
exist a sequence {tn} ~ (0, a) such that II T 1n II > n and lim tn = t0 < a.
By the resonance theorem, II T 1nx II must be unbounded at least for one

xE X, which surely contradicts the strong continuity condition (4).
Remark. By multiplying e-111, we may assume that a semi-group
{T1} of class (C0) is equi-bounded:
n~
IIT1 11 <
(for 0 < t
<
<Xl).
(6)
233
1. The Semi-group of Class {C0 )
If, in particular, M is
< 1, i.e., if
(7)
(for O<t<oo),
then the semi-group {T1} is called a contraction semi-group of class (C0 ).
As for the strong continuity condition (4), we have the following
Theorem. Let a family {T1 ; t 2': 0} of operators E L (X, X) satisfy (1')
and (3). Then condition (4) is equivalent to the condition
I!Tt!l<1
(8)
x EX.
w-lim T 1 x = x for every
qo
Proof. Suppose that (8) is satisfied. Let x 0 be any fixed element of

X. We shall show that s-lim T 1 x 0 = T 1ox0 for each t0 > 0. Consider the
t----+t 0
function x(t) = T 1x0 For each t0 > 0, x(t) is weakly right continuous at
t0 , because w-lim T 1 x = w-lim Th T 1 x0 = T 1 x0 We next prove that
q to
h .j, 0
jjT1II is bounded in a vicinity oft= 0. For, otherwise, there would exist

a sequence {tn} E such that tn t 0 and ~ II T 1nx 0 II = oo, contrary
to the resonance theorem implied by the weak right continuity of x (t) =

T 1xo- Thus by (1'), we see that T 1 x0 = x (t) is bounded on any compact
interval oft. Moreover, x (t) is weakly measurable. For, a right continuous
real-valued function f (t) is Lebesgue measurable, as may be proved from
the fact that, for any 1X, the set {t; f(t) < 1X} is representable as the union
of intervals of positive length. Next let {tn} be the totality of positive
rational numbers, and consider finite linear combinations ~ f3jx(tj)
J
where f3j are rational numbers (if X is a complex linear space, we take
f3i = ai + ibi with rational coefficients ai and bj) These elements form a
countable set M = {xn} such that {x (t); t > 0} is contained in the strong
closure of M. If otherwise, there would exist a number t' such that x (t')
does not belong to Ma. But, being a closed linear subspace of X, Ma is
weakly closed by Theorem 11 in Chapter V, 1; consequently, the condition x(t') E Ma is contradictory to the weak right continuity of x(t), i.e.,
to x (t') = w-lim x (tn).
tn.j.t'
We may thus apply Pettis' theorem in Chapter V, 4 to the effect that

x (t) is strongly measurable, and so by the boundedness of II x (t) II on any
{J
compact interval of t, we may define the Bochner integral

and we have
II/
x(t) dtll
<
llx(t)
II dt for 0 < 1X < f3 < oo.

{J
of the strong continuity in s of the integral
J x (t) dt
"'
By virtue
fJ+s
J x (t + s) dt = J
x+s
x (t) dt,
which is implied by the boundedness of x (t) on any compact interval

of t, N. DuNFORD [3] proved that x (t) is strongly continuous in t > 0.
We shall follow his proof.
234
Let 0 < iX < 1J < fJ < ~- e< ~ with

T'1 T;-'1 x 0 = T'1 x(~- rJ), we have
{J
e>
0. Since x (~)
= T;x 0 =
{J
({3-iX) x(~) = J x(~) drJ = J T'1 x(~-rJ) drJ,

and so, by sup IIT'111 < oo which is implied by (1') and (3), combined
with the boundedness of II T 1 ll near t = 0, we obtain
({)- ~X){x(~ e)- x(~)} = J T 11 {x(~ e- 'f})- x(~ -n)} d1J,
IX
IX
IX:{,t]:{,{J
{J
IX
({3-iX)IIx(~e)-x(~)ll< sup
IX:{,t]:{,{J
e-"'
IIT11 11 J llx(-re)-x(-r)lld-r.
;-{J
The right hand side tends to zero as e t 0, as may be seen by approximating x (r) by finitely-valued functions.
We have thus proved that x(t) is strongly continuous in t > 0. To
prove the strong continuity of x (t) at t = 0, we proceed as follows. For
any positive rational number tn, we have T 1 x(tn) = T 1 T 1nx0 = Tt+tnxo =
x(t + tn) Hence, by the strong continuity of x(t) for t > 0 proved
above, we have s-lim T 1 x (tn) = x (tn). Since each Xm E M is a finite linear
qo
combination of x(tn)'s, we have s-lim T 1xm
qo
Xm (m
1, 2, ... ). On the
other hand, we have, for any t E [0, 1],
< II Ttxm - Xm II + II Xm - Xo II + II Tt (xo- Xm) II

< II Ttxm - Xm II + II Xm - Xo II + o:;:;t;;;;I
sup II Tt II II Xo- Xm II
Hence lim llx(t) -x0 ll < (1 + sup JIT1 11} llxm-xoll, and so s-limx(t)
qo
o,;;e,;;I
qo
= x 0 by inf II x 0 - Xm II = 0. - xmEM
II X (t) -
Xo II
2. The Equi-continuous Semi-group of Class (C0 ) in Locally

Convex Spaces. Examples of Semi-groups
Suggested by the preceding section, we shall pass to a more general
class of semi-groups.
Definition. Let X be a locally convex linear topological space, and
{T1 ; t :;::o: 0} be a one-parameter family of continuous linear operators
E L (X, X) such that
(1)
T 1 Ts = Tt+s To= I,
lim T 1 x = T 1,x for any t0 > 0 and x EX,
(2)
1-->-10
the family of mappings {T1} is equi-continuous in t,

i.e., for any continuous semi-norm p on X, there
(3)
exists a continuous semi-norm q on X such that
p(T 1x) < q (x) for all t > 0 and all x E X.
Such a family {T1} is called an equi-continuous semi-group of class (C 0 }.
235
2. The Equi-continuous Semi-group of Class (C 0 )
The semi-groups {T1} satisfying conditions (1'), (3), (4) and (6) of
Section 1 are example of such equi-continuous semi-groups of class (C0 ).
We shall give concrete examples.
Example 1. Let C [0, = J be the space of bounded uniformly continuous real-valued (or complex-valued) functions on the interval [0, =),
and define T 1, t > 0, on C [0, =] into C [0, =] by
(T 1x) (s) = x(t
+ s).
Condition (1) is trivially satisfied. (2) follows from the uniform continuity of x(s). Finally IITell < 1 and so {T1} is a contraction semigroup of class (C0 ). In this example, we could replace C [0, =J by
C [-=, =J or by LP(-=, =).
Example 2. Consider the space C [- =, =]. Let
Ne(u) = (2nt)- 1 ' 2
e-u'/ 21 ,
-= <
<<X>, t >
which is the Gaussian probability density. Define T 1, t

into C [- =, =] by
JN
00
(T1 x) (s) =
0,
> 0, on C [-=,=]
(s- u) x(u) du, for t > 0,
-00
= x(s), for t = 0.
00
Each T 1 is continuous, since, by
JN
-oo
(s- u) du = 1,
00
IITexll <
llxll J Nz(s-u) du = llxll.

-00
T 0 = I by the definition, and the semi-group property T 1 Ts = Tt+s

is a consequence from the well-known formula concerning the Gaussian
probability distribution:
1
V2:n:(t
+ t'l
e-u'f2(t+t')
1-
00
e-(u-v)'/21 e-v'/21'
dv.
V2:n:e V2:n:e' -oo
This formula may be proved by applying the Fourier transformation

on both sides, remembering (10) and (13) in Chapter VI, 1. To prove the
00
strong continuity in t of T 1, we observe that x (s) =
JN
(s- u) x (s) du.
-00
Thus
00
(T1 x) (s)-x(s) =
JN
(s-u)(x(u)-x(s))du,
-00
which is, by the change of variable (s- u)JVf
z, equal to
00
JN
-00
1 (z)
(x(s- Vtz) -- x(s)) dz.
236
By the uniform continuity of x(s) on (-oo,oo), there exists, for any

> 0, a number 1:5 = 1:5 (s) > 0 such that Ix (s 1) - x (s2 ) I <5
Ivi zl >6
J ( .. )dz
Ivlzl >"
N 1 (z)dz
N 1 (z) dz.
The second term on the right tends to 0 as t-+ 0, since the integral
00
JN
-oo
1 (z)
dz converges. Thus lim sup I (T 1 x) (s)- x(s) I = 0 and hence

i,I.O
s-lim T 1 x = x; consequently, by the Theorem in the preceding section,

qo
we have proved (2).
In the above example, we can replace C [- oo, oo] by LP (- oo, oo).
Consider, for example, L 1 (-oo, oo). In this case, we have
00
JJN
II T 1 x II<
(s- u) lx(u) Ids du < llx II,
-00
by Fubini's theorem. As for the strong continuity, we have, as above,

00
II Ttx-
II
= !If Nl (z)(x(s-
_l
-00
Nl (z)
(_[1
vr.
x(s-
z)- x(s)) dzlds
vr.
z)- x(s) dsJ dz
00
< 2
J Nl(z) dz llxll
-00
Hence, by the Lebesgue-Fatou lemma, we obtain

lim IIT 1 x- xll <
t+O
-oo
N 1 (z) (lim
t.j,O
-oo
lx(s-
Vt z)- x (s) Ids) dz= 0,
because of the continuity in the mean of the Lebesgue integral. which may
be proved by approximating x (s) by finitely-valued functions.
Example 3. Consider C [- oo, oo ]. Let ).
> 0, on C [-oo, oo] into C [-oo, oo] by

(T1 x) (s)
(At)k
>
0, fl
= e-t 2; """"ill x(s- kt-t).
00
k~O
>
0. Define T 1,
237
3. The Infinitesimal Generator of an Equi-continuous Semi-group
We have
(T w (T1x)) (s)
oo (Jew)"' [
= e-.lw ~ m! e-.lt
~ 2_
-.<(w-l-1)
~ pI
P=O
= e-.l(w+t)
&o kT
oo
(M)k
~ 0w)m
[pI
. ~
m=O
(s- kft- mfl) .
(M)P-m ( _ p )]
fl
)I X S
m
(p-
~;! (J.w + J.t)~ x(s -Pfl) =
(Tw+tx) (s).
Thus it is easy to verify that T 1 is a contraction semi-group of class (C0 ).
3. The Infinitesimal Generator of an Equi-continuous

Semi-group of Class (C0 )
Let {T1 ; t > 0} be an equi-continuous semi-group of class (C0 )
defined on a locally convex linear topological space X which we assume
to be sequentially complete. We define the infinitesimal generator A of T 1
by
(1)
Ax= limh-1 (Th-I)x,

h.j,O
i.e., A is the linear operator whose domain is the set D(A) =

{xEX; lim h-1 (Th-I)x exists in X}, and, for xED(A), Ax==
h.j,O
limh-1 (Th-I) x. D(A) is non-empty; it contains at least the vector 0.
h.j,O
Actually D(A) is larger. We can prove

Theorem 1. D (A) is dense in X.
Proof. Let 9ln(s) = ne-ns, n > 0. Consider the linear operator C"'n
which is the Laplace transform of T 1 multiplied by n:
00
f
0
c'fnx =
9ln (s) Tsx ds
for
EX'
(2)
the integral being defined in the sense of Riemann. The ordinary procedure of defining the Riemann integral of numerical functions can be
extended to a function with values in a locally convex, sequentially
complete space X, using continuous semi-norms p on X in place of the
absolute value of a number. The convergence of the improper integral is a
consequence of the equi-continuity of T 1, the inequality
p(q;n(s) T 5 x) = ne-nsp(T5 x)
and the sequential completeness of X.

We see, by
J ne-ns p (Tsx) ds 0,
(3)
and
lim crp X= X for each X EX.
(4)
n-+00
n
00
Then U R(Crpn) will be dense in X, and a fortiori D(A) will be dense in

n~l
X. To prove (3), we start with the formula
h-1 (Th- I) C'Pnx
00
h-1
q;,. (s) Tsxds.
J = 0J
0
00
The change of the order: Th
00
q;,. (s) Th Tsxds- h- 1

00
Th is justified, using the
linearity and the continuity of Th. Thus
00
h-1 (T 4 - I) C'Pnx = h- 1
enh-l
=~h-n
enh
J e-naTaxda-hn J e-nsrsxds
1
oo
f q;,.(s) Tsxds
00
q;,.(s) Ts+hxds- h-1
:=-_! {C'Pnx
-I
ne-naTaxda}-
q;., (s) Tsxds.
By the continuity of q;.,(s) Tsx ins, the second term on the right tends
t 0. Similarly, the first term on the right
tends to nC'Pnx ash t 0. Hence we have
to - IPn (0) T 0 x = - nx as h
A c'l'nx
= n(C'I'n -I)
X,
xE X.
J ne-ns ds =
(5)
00
We next prove (4). We have, by
1,
0
00
C'l'nx-x
= n fe-ns (Tsx-x) ds,

0
<
p(C'I'nx-x)
J e-ns p(Tsx-x) ds = nJ0 + n J = I + I

0
00
00
2,
say,
where {) > 0 is a positive number. For any e > 0, we can choose, by the
continuity in s of Tsx, a {) > 0 such that p (Tsx- x) < e for 0 < s < t>.
Then
J
oo
I 1 < en e-ns ds < en
e-ns ds = e.
For a fixed {)
I2
>
<
0, we have, by the equi-boundedness of {Tsx} in s

00
J e-ns (p(Tsx) + p(x)) ds-+ 0
Hence we have proved (4).
as
nt=
> 0,
3. The Infinitesimal Generator of an Equi-continuous Semi-group
239
Definition. For x EX, we define

D 1 T 1 x =lim h-1 (Tt+h- T 1) x
(6)
h~o
if the right hand limit exists.

Theorem 2. If x E D (A), then x E D (De T 1) and
D 1 T 1 x =A T 1 x
TeAx, t > 0.
(7)
Thus, in particular, the operator A is commutative with Te.

Proof. If xE D(A), then we have, since Te is continuous linear,
TeA X= T 1 lim h-1 (Th-I)x =lim h-1 (T1 Tn-Te)x= lim h-1 (Te+n-T 1)x
h~O
limh-1 (T, -I) Tex
h-J-0
h-J-0
h-J-0
AT1 x.
Thus, if xED (A), then T 1 x ED (A)
and T 1 A x =A T 1 x = lim h-1

h-J-0
(T1+k - T 1)x. We have thus proved that the right derivative of T 1x

exists for each xE D(A). We shall show that the left derivative also
exists and is equal to the right derivative.
For this purpose, take any / 0 E X'. Then, for a fixed xED (A), the
numerical function / 0 (Tex) = <Tex, / 0 ) is continuous in t > 0 and has
right derivative d+ / 0 (T 1 x)fdt which is equal to /0 (A T 1 x) = fo (T1 A x) by
what we have proved above. Hence d+f0 (T1 x)fdt is continuous in t. We
shall prove below a well-known Lemma: if one of the Dini derivatives
/J+j(t), !]_+f(t), TJ-j(t)
and
!]_-j(t)
of a continuous real-valued function f(t) is finite and continuous, then

f(t) is differentiable and the derivative is, of course, continuous and equals
g f(t). Thus / 0 (T1x) is differentiable in t and
t
/0
J d+fo (T.x)fds ds = 0J / (T.A x)ds

0
(T1 x- x) = / 0 (T1 x)- / 0 (T0 x) =
= /0
(j
T.Axds).
Since / 0 EX' was arbitrary, we must have

t
T 1 x-x= jT.Axds
for each
xED(A).
Since T 8 A x is continuous in s, it follows that T 1 x is differentiable in tin

the topology of X and
t+h
D 1 T 1 x =lim h-1 j T.Ax ds = T 1 Ax.
h-+0
Thus we have proved (7).

Proof of the Lemma. We first prove that the condition: TJ+ f (t) > 0
fOr a< t 0. Assume the contrary, and
240
<- s(b- a) with some s > 0. Then, for g(t) = j(t)f(a) + s(t- a), we have lJ+g(a) = 15+j(a) + s > 0, and so, by g(a) = 0,
we must have g(t0 ) > 0 for some t0 >a near a. By the continuity of g(t)
and g(b) < 0, there must exist a t1 with a< t0 < t1 0.
By applying a similar argument to f (t) -ext and to {Jt -I (t), we prove
the following: if one of the Dini derivatives D f (t) satisfies
let f(b) -f(a)
ex
< Df (t) < fJ
on any interval [t1 , t2 ],
then ex < (I (t2 ) -/(t1))/(t2 - t1 ) < {J. Hence, the suprema (and the infima)
on [tv t 2 ] of the four Dini derivatives of a continuous real-valued function f (t) are the same. Thus, in particular, if one of the Dini derivatives of
a continuous real-valued function f (t) is continuous on [tv t 2 ], then the
four Dini derivatives of f (t) must coincide on [tv t 2 ].
4. The Resolvent of the Infinitesimal Generator A

Theorem I. If n > 0, then the operator (ni- A) admits an inverse
R(n; A)= (ni- A)-1 E L(X, X), and
R(n;A) x
J e-ns T xds
00
for
xE X.
(1)
In other words, positive real numbers belong to the resolvent set

of A.
e (A)
Proof. We first show that (ni- A)-1 exists. Suppose that there
exists an x 0 # 0 such that (ni -A) x0 = 0, that is, Ax0 = nx0 Let / 0 be a
continuous linear functional E X' such that / 0 (x0 ) = 1, and set cp (t) =
/ 0 (T1 x 0 ). Since x 0 E D (A), cp (t) is differentiable by Theorem 2 in the preceding section and
If we solve this differential equation under the initial condition cp (0) =

(x0 ) = 1, we get cp(t) = en1 But, cp(t) = /0 (T1x0 ) is bounded in t,
because of the equi-boundedness of T 1 x0 in t > 0 and the continuity of
the functional fo. This is a contradiction, and so the inverse (ni- A)-1
must exist.
Since, by (5) of the preceding section, A Cqnx = n (C'Pn- I) x, we have
(nl- A) C'Pnx = nx fot all x EX. Thus (nl -A) maps R (C'Pn) ~ D(A)
onto X in a one-one way. Hence, a fortiori, (nl- A) must map D (A)
onto X in a one-one way, in virtue of the existence of (nl- A)-1. Therefore, we must have R(C'Pn) = D(A) and (nl- A)-1 = n-1 Ctp,.
/0
241
4. The Resolvent of the Infinitesimal Generator A
Corollary 1. The right half plane of the complex A.-plane is in the

resolvent set e(A) of A, and we have
J e-" T xdt
00
R(A.; A) x =(AI- A)-1 x =
for Re(A.) > 0 and xEX. (2)
Proof. For real fixed r, {e-iTt T 1 ; t > 0}, constitutes an equi-continuous

semi-group of class (C0). It is easy to see that the infinitesimal generator
of this semi-group is equal to (A-ir I). Thus, for any a> 0, the resolvent
R(a + ir; A)= ((a+ ir) I - A)-1 exists and
+ ir) I - A) x = J e-(a+iT)s
00
R((a
for all xE X.
T 8 xds
(2')
Corollary 2.
D(A) = R((AI- A)-1)
when Re(A.)
R(R(A.; A))
>
0,
(3)
AR(A.;A)x=R(A.;A)Ax=(A.R(A.;A)-I)x for xED(A), (4)
for
AR(A.;A)x=(A.R(A.;A)-I)x
lim A.R(J.; A) x
;. too
Proof. Clear from R (A.; A)
(6)
xE X.
for
(5)
xEX,
(AI- A )-1 and (4) of the preceding
section.
Corollary 3. The infinitesimal generator A is a closed operator in the
following sense (Cf. Chapter II, 6):
ifxhED(A)
x E D (A)
and
and
limxh=xEX, limAxh=yEX,then
Ax = y.
Proof. Put (I- A) xh
continuity of (I-
h-->-00
~00
A )-1 ,
zh. Then lim zh

~00
lim xh = lim (I-
~00
~00
that is, x = (I- A)-1 (x- y), (I- A) x
x- y and so, by the
A )-1
zh = (I-- A )-1 (x- y)
= x- y.
This proves that
y=Ax.
Theorem 2. The family of operators
{(A.R(A.; A))}
(7)
is equi-continuous in A.> 0 and inn= 0, 1, 2, ..

Proof. From the resolvent equation (Chapter VIII, 2, (2))
R (f-l; A) - R (A.; A)
(A. - f-l) R (f-l; A) R (A.; A),
we obtain
lim (f-l-A.)-1 (R (f-l; A) -R (A.; a)) x = dR (A.; A) xfdA. = -R (A.; A) 2 x, xE X.
J'-+i.
To derive the above formula, we have to appeal to (2) in order to show

that limR(f-l; A) y = R(A.; A) y, yE X.
J'-+i.
16 Yosida, l'unctional Analysis
242
Therefore, R (A; A) xis infinitely differentiable with respect to A when

Re(A) > 0 and
dnR(A;A)xjdAn=(-l)nn!R(A;A)n+lx
(n=0,1,2, ... ).
(8)
On the other hand, we have, by differentiating (2) n-times with respect

~A,
dnR(A;A) xjdAn =
oo
J e-J.t(-tt T xdt.
(9)
Here the differentiation under the integral sign is justified since {T1 x}
J e-J.t t dt= (n!)/An+l when

00
is equi-bounded in t and
0
).Ml
(AR(A; A))n+l x = -
J e-J.t tn T xdt for

OO
n.1 o
Re(A) > 0. Hence
xE X and Re(A)
and so, for any continuous semi-norm p on X and A> 0, n
p((AR(A; A))+l x)
>
n o
0, (10)
0,
J e-J.t tn dt sup
p(T x) =sup p(T x).
t;;;;o
t;;;;o
_;.n+l 00
< -
>
(11)
This proves Theorem 2 by the equi-continuity of {T1} in t.
5. Examples of Infinitesimal Generators

We first define, for n
so that
>
0,
fn =(I- n-1 A)-1 = nR(n; A),
(1)
(2}
Example 1. (T1 x) (s} = x (t
+ s) on C [0, ex>].
n J e-nt x(t + s} dt =
00
Writing
Yn(s) = Unx) (s) =
we obtain y~ (s) = - n e-n(s-s)x (s)
+ n 2 J e-n(t-s)
J e-n(t-s)x(t) dt,
00
00
x (t) dt = - nx (s)
n Yn(s). Comparing this with the general formula (2):

(A fnx)(s) = n ((]n- I) x)(s),
we obtain Ayn(s) =y~(s). Since R(]n) = R(R(n; A))= D(A}, we have
Ay(s) = y'(s)
for every
yE D(A).
Conversely, let y(s) andy' (s) both belong to C [0, ex>]. We will then show
that yE D(A) and Ay(s) = y'(s). To this purpose, define x(s) by
y'(s) -ny(s) = -nx(s).
Setting Unx}(s) = Yn(s), we obtain, as shown above,
y~(s)-
nyn(s) = - nx(s).
5. Examples of Infinitesimal Generators
243
Hence w (s) = y (s) - y,. (s) satisfies w' (s) = nw (s) and so w (s) = C e"'.
But, w must belong to C [0, ex:>] and this is possible only if C = 0. Hence
y(s) =y,.(s)ED(A) andAy(s) =y'(s).
Therefore, the domain D (A) of A is precisely the set of functions y E
C [0, ex:>] whose first derivative also E C [0, ex:>], and for such a function
y we have A y = y'. We have thus characterized the differential operator
dfds as the infinitesimal generator of the semi-group associated with the
operation of translation by t on the function space C [0, ex:>].
Example 2. We shall give a characterization of the second derivative

d2 fds 2 as the infinitesimal generator of the semi-group associated with
the integral operator by Gaussian kernel. The space is C [-ex:>, ex:>] and
J
00
(T1 x)(s) = (2nt)- 1i 2

We have
(v) r2
n (2nt)- 1i 2 e-nt-(s-v)'/21 dt] dv
-00
x(v) dv if t > 0, = x(s) if t = 0.
j x (v) [ j
y,. (s) = (] ,.x)(s)
e-(s-v)'/21
-oo
Vn j
(2n)-1/2
e-a'-n(s-v)'/2a'
o
(by substituting t = a2fn).
-00
da] dv
Assuming, for the moment, the formula
da
e-(a'+cfa')
V:
e-2c,
c=
Vn Is- v IJV2 > o,
(3)
we get
Yn (s)
J x (v) (n/2)112 e00
V2n Is-vi
dv.
-oo
x (v) being continuous and bounded, we can differentiate twice and obtain
y,.' (s) = n
00
(v) e- y?,;; (v-s) dv- n
y,." (s) =
n{-
x(s) -x(s)
+ V2Vn
+ V2
= - 2n x(s)
J x (v) e- '{2,;
(s-v)
dv,
(s-v)
dv
-oo
x(v) e-V2"
vn _[X
(v) e- V2n (s-v) dv}
+ 2n y,.(s).
Comparing this with the general formula (2):

(Ay,.)(s) = (AJ,.x)(s) = n((J,.-1) x)(s) = n(y,.(s) -x(s)),
we find that Ay,.(s)
16*
y~(s)/2.
Since R(J,.) = R(R(n; A))= D(A), we
244
have proved that
Ay(s)=y"(s)/2
yED(A).
forevery
Conversely, let y (s) andy" (s) both belong to C [-<X>, <X>]. Define x(s) by
y" (s) -- 2ny (s) = - 2nx (s).

Setting Yn(s) = Unx)(s), we obtain, as shown above,
y~ (s)-
2nyn (s) = - 2nx(s).
Hence w(s) = y(s) -Yn(s) satisfies w"(s)- 2nw(s) = 0, and so w(s) =

C1 e'/2n + C2 e-Y2n. This function cannot be bounded unless both C1
and C2 are zero. Thus y (s) = Yn (s), and so y ED (A), (Ay) (s) = y" (s)f2.
Therefore, the differential operator
~ ~22 is characterized as the
infinitesimal generator of the semi-group associated with the integral

transform by the Gaussian kernel on the function space C [-<X>, <X>].
The proof of (3). We start with the well-known formula
00
J e-x' dx = Vn/2.
Setting x
~
____:!_ =
a- cfa, we obtain
00
v-c
e-(a-cfa)'(l
_
2c {
-e
+ cfa2) da =
e2c
00
v<
e-(a'+c'fa') (1
+ cfa2) da
Je-(oo'+c'fa') da + J e-(a'+c'fa') .!:___ da.

}
v<
v<
a
00
00
Setting a= eft in the last integral, we obtain
Vn2n = e2c {
v<
e-<a'+c'fa') da -
v<
e-(c'ft'+J') dt} = e2c
e-(t'+c'ft') dt.
Exercise. Show that the infinitesimal generator A of the semi-group

{T1} on C [-<X>, <X>] given by
(T1 x)(s)
e-J.t
(,1,t)k
2: klx(s--kp)
00~
k=O
(A, fl
>
0),
is the difference operator A :
(Ax)(s)
,.1,
{x(s- p)- x(s)}.
6. The Exponential of a Continuous Linear Operator

whose Powers are Equi-continuous
Proposition. Let X be a locally convex, sequentially complete, linear
topological space. Let B be a continuous linear operator E L (X, X) such
6. The Exponential of a Continuous Linear Operator
245
that {B"; k = 1, 2, ... } is equi-continuous. Then, for each xE X, the

series
00
~ (k!)- 1
k=O
(tB)" x (t > 0)
(1)
converges.
Proof. For any continuous semi-norm p on X, there exists, by the
equi-continuity of {B"}, a continuous semi-norm q on X such that
p(B"x) < q(x) for all k > 0 and xE X. Hence
m
p ( ~ (tB)" xfk!
k=n
Therefore
{.,,#a (tB)"xfk!}
m
.I t"p(B"x)fk! < q(x) .I t"Jk!.
< k=n
k=n
is a Cauchy sequence in the sequentially
complete space X. The limit of this sequence will be denoted by (1).

00
Corollary 1. The mapping x--+ .I (tB)" xfk! defines a continuous

k=O
linear operator which we shall denote by exp(tB).
Proof. By the equi-continuity of {B"}, we can prove that B,. =
.I (tB)"fk! (n = 0, 1, 2, ...) are equi-continuous when t ranges over
k=O
any compact interval. In fact, we have
"
p(B,.x) < .I t"p(B,.x)fk! < q(x) .I t"Jk! < tf. q(x).

k=O
k=O
Hence the limit exp (tB) satisfies
p (exp (tB) x) < exp (t) q (x)
(t
> 0).
(2)
Corollary 2. Let B and C be two continuous linear operators E L (X, X)

such that {B"} and {C"} are equi-continuous. If, moreover, BC = CB,
then we have
(3)
exp(tB) exp(tC) = exp(t(B +C)).
Proof. We have
k
,.c.
q(Cx) < 2" sup q(Cx).

p((B +C)" x) < .I kCs p(B"- CS x) < .I
o~s
s=O
s=O
+C)).
candefineexp(t(B
we
equi-continuous,and
is
Hence{2-" (B +C)"}
series
the
rearrange
we
B,
C
=
C
B
commutativity
the
of
use
By making
00
so that we obtain
00
(i
k=O
.I (t(B
k=O
(tB)"fk!)
+ C)t xfk!
(tC)" xfk!) as
(i
k=O
numericalseries .I(t(b + c))"/k!.

k=O
Corollary 3. For every x E X,
limh-1 (exp(hB) -I) x = Bx,
htO
in the case of
(4)
246
and so, by making use of the semi-group property

exp((t
+ h) B) =
exp (tB) exp (hE)
(5)
proved above, we obtain

D1 exp(tB) x = exp(tB) Ex= B exp(tB) x.
Proof. For any continuous semi-norm

p(h-1(exp(hB)- I) x- Ex)
which surely tends to 0 as h
p on
(6)
X, we obtain, as above
00
00
k=2
k=2
< .I hk- 1 p (Bkx)/k! < q(x) .I hk- 1fk!
t 0.
7. The Representation and the Characterization of Equi-continuous

Semi-groups of Class (C0 ) in Terms of the Corresponding
Infinitesimal Generators
We shall prove the following fundamental
Theorem. Let X be a locally convex, sequentially complete, linear
topological space. Suppose A is a linear operator with domain D (A)
dense in X and range R (A) in X such that the resolvent R (n; A) =
(ni- A)-1 E L (X, X) exists for n = 1, 2, ... Then A is the infinitesimal
generator of a uniquely determined equi-continuous semi-group of class
(C0 ) iff the operators {(I- n-1 A)-m} are equi-continuous inn= 1, 2, ...
and m = 0, 1, ...
Proof. The "only if" part is already proved. We shall prove the "if"
part.
Setting
(1)
we shall prove
lim J,.x = x
n-co
In fact, we have, for xED(A),

x = n-1 J,. (Ax) tends to
boundedness of {],.(Ax)} in n =
{],.} is equi-continuous in n, it
J,.x -
xEX.
Put
Tl"> =
for every xE X.
(2)
A],.x=],.Ax=n(J,.-I)x and so
0 as n-+ oo, in virtue of the equi1, 2, . . . . Since D (A) is dense in X and
follows that lim J,.x = x for every
n->-00
exp(tA],.) = exp(tn(J,.-I))=exp(-nt) exp(nt],.), t>O. (3)
Since {]!} is equi-bounded in n and k, the exp (tn J,.) can be defined, and
we have, as in (2) of the preceding section,
00
p(exp(nt],.)x) <.I (nt)k(k!)-1p(J!x) <exp(nt) q(x).

k=O
247
7. The Representation and the Characterization
Consequently, the operators {T)"l} are equi-continuous in t

n = 1, 2, ... in such a way that
>
0 and
p (T)"l x) < q (x).
(4)
We next remark that fnf m= J mfn for n, m > 0. Thus],. is commutative with Tlml. Hence, by D1 Tl"lx = AJ,.Tl"lx = Tl"lAJ,.x, proved
in the preceding section,
p (Tl"l x- Tlml x)
(j D.(Ti~~T~"lx) ds)
= p ( j Tl~~ T~nl J
J
=p
(A n - A m)
(5)
xds).
Hence, if xED (A), there exists a continuous semi-norm qon X such that
p(Tl"lx- Tlmlx) <
Jq ((AJ,. -Afm) x) ds = tq((J,.A- fmA) x).
Therefore, by (2}, we have proved that
lim
n,.........oo
p(Ti"l x- Tlml x)
= 0
uniformly in twhen t varies on every compact interval. Since D (A) is dense

in the sequentially complete space X, and since the operators {T<"l} are
equi-continuous in t > 0 and in n, we see that lim Tl"l x = T 1 x
n-700
exists for every xE X and t?.: 0 uniformly in t on every compact interval

of t. Thus the operators {T1} are equi-continuous in t > 0, and from the
uniform convergence in t, T 1 x is continuous in t > 0.
We next prove that T 1 satisfies the semi-group property T 1 T 5 = Tt+s
Since r<nl
- T<nl
r<nl
t+st
s ' we have
p (Tt+sx- Tt T,x) < p (Tt+sx- T1'f!,x)
+ p (Tl'f!,x- T~nl Ti"l x)
+ p (Tl"l T~"l x- Tl"l T x) + p (T)"l T,x- T T,x)

< p(Tt+sx- T,l'j!.x) + q(T1"lx- T,x)
+ p ((Tl"l- T1) T,x)- 0 as n - oo.
1
Thus p (Tt+,x - T 1T, x) = 0 for every continuous semi-norm p on X,

which proves that T 1+s = T 1 T,.
Let A be the infinitesimal generator of this equi-continuous semigroup {T1} of class (C0). We have to show that A =A. Let x E D(A).
Then lim Tl"lAJ,.x = T 1 Ax uniformly in ton every compact interval
n-700
oft. For, we have, by (4),
p (T1A X - Tl"l A fnx)
- oo, since lim J nA x = Ax. Hence

~
T 1 x-x= lim (Tin)x-x) =lim JT~n)Afnxds

~
~0
= J (lim T~n) A fnx) ds = J TsA xds

0
'
so that lim t-1 (T1 x- x) =lim t-1 J T 5 A xds exists and equals Ax. We
t.j.O
t.j,O
have thus proved that xED (A) implies xED (A) and Ax= Ax, i.e.
A is an extension of A. A being the infinitesimal generator of the semigroup T 1, we know that, for n> 0, (ni -A) mapsD(A) onto X in
one-one way. But, by hypothesis, (ni- A) also maps D(A) onto X in
one-one way. Therefore the extension A of A must coincide with A.
Finally, the uniqueness of the semi-group T 1 is proved as follows.
Let T 1 be an equi-continuous semi-group of class (C0 ) whose infinitesimal
generator is precisely A. We construct the semi-group T)nl. Since A is
commutative with
we see that Afn and T~n) are commutative with
T 1 Then, for xE D(A), we obtain, as in (5),
A
fe,
p(T)n)X- Ttx) =
(j D5 ('It-.T~n)x)
(j -T1_J~n)
= p
ds)
(6)
(A- Afn) xds).
Thus, by virtue of lim AJ.. x=Ax for all xED(A), we prove lim
n-->00
Tin) x = T 1 x for all x EX similarly to the above proof of the existence of

lim Tin) X, x E X.
n--+OO
Therefore, T 1 x = T 1 x for all x EX, that is, T 1 = T 1

Remark. The above proof shows that, if A is the infinitesimal generator of an equi-continuous semi-group T 1 of class (C0), then
T 1x = lim exp(tA (I- n-1 A)- 1 ) x, x EX,

n-700
(7)
and the convergence in (7) is uniform in t on every compact interval oft.

This is the representation theorem for semi-groups.
Corollary 1. If X is a B-space, then the condition of the Theorem
reads: D (A )a = X and the resolvent (I - n-1 A )-1 exists such that
ii(I -n-1 A)-mll
< C
(n
1, 2, ... ; m
1, 2, ... )
(8)
with a positive constant C which is independent of nand m. In particular,

for the case of a contraction semi-group, the condition reads: D (A)i =X
7. The Representation and the Characterization
249
and the resolvent (I - n-1 A )- 1 exists such that
Ji(I -n-1A)-1 \J < 1 (n = 1, 2, ... ).
(9)
Remark. The above result (9) was obtained independently by E. HILLE [2] and by K. YosmA [5]. The result was extended by W. FELLER
[1], R. S. PHILLIPS [3] and I. MIYADERA [1] and the extension is
given in the form (8). It is to be noted that in condition (8) and (9), we
may replace (n = 1, 2, ... ) by (for all sufficiently large n). The extension of the semi-group theory to locally convex linear topological spaces,
as given in the present book, is suggested by L. SCHWARTZ [3].
Corollary 2. Let X be a B-space, and {T1 ; t > 0} be a family of
bounded linear operators L (X, X) such that
T1 T5
= T 1+5 (t, s >
0), T 0 =I,
(10)
s-lim T 1 x = x for all xE X,
(11)
t.j.O
JJ T 1 \\ < M ef!t for all t > 0, where M
>
0 and {3
> 0 are
independent of t.
(12)
Then (A- {3I) is the infinitesimal generator of the equi-continuous semigroup 51 = e-f!1T 1 of class (C0), where A is the operator defined through
Ax = s-lim t-1 (T1 - I) x. Thus, by Corollary 1, we see that a closed
1.).0
linear operator A with D (A )a = X and R (A) ~ X is the infinitesimal

generator of a semi-group T 1 satisfying (10), (11) and (12) iff the resolvent
(I- n-1(A- {3I))-1 exists such that
\\(I -n-1(A -{3I))-m\\ <M (form= 1, 2, ... and all large n). (13)
This condition may be rewritten as
\\(I -n- 1A)-m\\ < M(1-n- 1{3)-m (form= 1, 2, ... and all large n.
(13')
In particular, for those semi-groups T 1 satisfying (10), (11) and
\\T1 \\
< ef!t for all t > 0,
(14)
condition (13') may be replaced by
\\(I- n-1A)-1 \J < (1- n-1{3)-1 (for all large n).
(13")
An application of the representation theorem to the proof of Weierstrass'

polynomial approximation theorem. Consider the semi-group T 1 defined by
(T1 x) (s) = x(t + s) on C [0, <Xl]. The representation theorem gives
(T1 x) (s)
= x (t +
00 tm
s) = s-lim exp (tA] nx) (s) = s-lim ,21 (A] n)mx (s),
8~
~m=Om.
and the above s-lim is uniform in t on any compact interval of t. From

~
this result we can derive Weierstrass' polynomial approximation theo-
250
rem. Let z(s) be a continuous function on the closed interval [0, 1]. Let
x (s) E C [0, oo] be such that x (s) = z(s) for s E [0, 1]. Put s = 0 in the
above representation of x (t + s). Then we obtain
00
(T1x) (0) = x(t) =s-lim .I tm(Afn)m X (0)/m!

~m=O
C[O, 1],
uniformly in ton [0, 1]. Hence z(t) is a uniform limit on [0, 1] of a sequence
of polynomials in t.
8. Contraction Semi-groups and Dissipative Operators

G. LUMER and R. S. PHILLIPS have discussed contraction semi-groups
by virtue of the notion of semi-scalar product. The infinitesimal generator
of such a semi-group is dissipative in their terminology.
Proposition. (LuMER). To each pair {x, y} of a complex (or real)
normed linear space X, we can associate a complex (or real) number
[x, y] such that
[x
+ y, z]
= [x, z]
+ [y, z],
[Ax, y]
= A[x, y],
[x, x] = [\x\[2,
(1)
\[x,y]\ < \\xii\\YII

[x, y] is called a semi-scalar product of the vectors x andy.
Proof. According to Corollary 2 of Theorem 1 in Chapter IV, 6, there
exists, for each x0 EX, at least one (and let us choose exactly one)
bounded linear functional /..,,EX' such that \\!..,II= llx0 II and
(x0 , /..,,) = llx0 \[ 2 Then clearly
[x, y]
(2)
(x, /y)
defines a semi-scalar product.

Definition. Let a complex (or real) B-space X be endowed with a
semi-scalar product [x, y]. A linear operator A with domain D (A) and
range R(A) both in X is called dissipative (with respect to [x, y]) if
Re [Ax, x] < 0 whenever
xED (A).
(3)
Example. Let X be a Hilbert space. Then a symmetric operator A such

that (Ax, x) < 0 is surely dissipative with respect to the semi-scalar
product [x, y] = (x, y), where (x, y) is the ordinary scalar product.
Theorem (PHILLIPS and LUMER). Let A be a linear operator with domain D(A) and range R(A) both in a complex (or real) B-space X such
that D (A)a =X. Then A generates a contraction semi-group of class (C0 )
in X iff A is dissipative (with respect to any semi-scalar product [x, y])
andR(I-A)=X.
Proof. The "if" part. Let A be dissipative and A. > 0. Then the inverse
(U- A)-1 exists and II (U- A)-1y \\ < _A.-1 IIY II when y E D((U -A)-1).
251
For, if y = A.x- A:;.;, then

.A. llx 11 2 =A. [x, x] < Re(A. [x, x]- [Ax, x]) = Re [y, x] < IIY II llx II, (4)
since A is dissipative. By hypothesis, R (I- A) =X so that A.= 1 is in
the resolvent set e(A) of A, and we have IJR(1; A) II< 1 by (4). If
JA.- 1J < 1, then the resolvent R (A.; A) exists and is given by
R(A.; A)= R(1; A) (I+ (A.-1) R(1; A))-1
= R(1; A)
00
.I ((1-A.) R(1; A))",
n=O
(see Theorem 1 in Chapter VIII, 2). Moreover, (4) implies that

JIR(A.; A) II< A.-1 for).> 0 with lA. -11 < 1. Hence, again by
R (t-t; A) = R (A.; A) (I + (!-'-A.) R (A.; A))-1,
which is valid for!-'> 0 with It-t- A.IJIR (A.; A) Jl < 1, we prove the existence of R(t-t; A) and IIR(t-t; A) IJ < w 1. Repeating the process, we see
that R (A.; A) exists for all A.> 0 and satisfies the estimate II R (A.; A) II < A.-1.
As D (A) is dense by hypothesis, it now follows, from Corollary 1 of the
preceding section, that A generates a contraction semi-group of class (C0).
The "only if" part. Suppose {T1 ; t ~ 0} is a contraction semi-group
of class (C0). Then
Re[T,x-x, x] = Re[T1x,
x] -Jixll2 <
IJT1 xiiJixiJ-JJxll 2 < 0.
Thus, for xED (A). the domain of the infinitesimal generator A ofT, we
have Re [Ax, x] =lim Re{t-1 [T1 x - x, x]} < 0. Hence A is dissipative.
qo
Moreover, we know that R(I -A)= D(R(1; A))= X, since A is the

infinitesimal generator of a contraction semi-group of class (C0).
Corollary. If A is a densely defined closed linear operator such that
D(A) and R(A) are both in a B-space X and if A and its dual operator
A' are both dissipative, then A generates a contraction semi-group of
class (C0 ).
Proof. It sufficies to show that R (I- A) = X. But, since (I- A)-1
is closed with A and continuous, R (I - A) is a closed linear subspace of X.
Thus R (I- A) o:j= X implies the existence of a non-trivial x' E X' such
that
<(x -
Ax), x') = 0
for all x E D (A) .
Hence x'- A' x' = 0, contrary to the dissipativity of A' and x' o:/= 0.
Remark. For further details concerning dissipative operators, see
G. LUMER-R. S. PHILLIPS [1]. See also T. KATO [6].

Definition. An equi-continuous semi-group {T1} of class (C0 ) is called
an equi-continuous group of class (C0). if there exists an equi-continuous
semi-group {T1} of class (C0) with the condition:
252
If we define S, by S 1 = T 1 for t > 0 and S_ 1 = T 1 for t > 0, then

the family of operators st. - 0 0 < t < oo, has the group property
S 1S. = S 1+s (-oo
<
t, s < oo), S0 =I.
(1)
Theorem. Let X be a locally convex, sequentially complete linear

topological space. Suppose A is a linear operator with domain D (A) dense
in X and range R (A) in X. Then A is the infinitesimal generator of an
equi-continuous group of class (C0 ) of operators S 1 E L (X, X) iff the
operators (I -n- 1 A)-m are everywhere defined and equi-continuous in
n = 1, 2, ... and in m = 1, 2, ...
Proof. The "only if" part. Let T 1 = S 1 for t > 0 and T1 = S_ 1 for
t > 0. Let A and A be the infinitesimal generator of T 1 and T 1 , respectively. We have to show that A =-A. If xED (A), then, by putting
x11 = h-1 (T11 - I) x and making use of the equi-continuity of T 11 ,
A
p(T"x"-Ax)
< p(T11 x 11 - T 11 Ax) + p((T11 -I) Ax)

< q(x11 -Ax)+ p ((T11 -I) Ax)
A
where p and q are continuous semi-norms on X such that, for a given p, we

can choose a qsatisfying the above inequality for all h > 0 and all xED (A)
simultaneously. Thus lim T 11 x 11 =Ax, and so xE D(A) implies Ax=
A
h.j,O
limT11 (h- 1 (Th-I))x=limh- 1 (I-T11 )x=-Ax, Hence-A is an

ll.j,O
h.j,O
extension of A~ In the same way, we can prove that

-A. Therefore A =-A.
The "if" part. Define, for t > 0,
T 1x
Tx =
1
lim T)nl x
n->00
lim f)nlx
..-..co
A is an extension of
lim exp(tA (I- n-1 A)-1 ) x,
n->00
= lim
n->00
exp(tA (I- n-1 A)-1) x, where A =-A .
Then T 1 and T 1 are both equi-continuous semi-groups of class (C0). We

have
P (Tt ftx- T)nl f)nl x)
 0. Thus we have

lim T)nlf)nlx = T 1 T1 x. We have incidentally proved the equi-con-
n->00
tinuity of T 1 T 1 by the equi-continuity of T)nlf)nl. On the other hand,

we have T)nl f1ml = f1ml T)nl by the commutativity of (I- n-1 A )-1
and (I - m-1 A)-1 Thus (T(n)
j(n))
(T(n)
j(n))
-- y(n)
j(n)
and so
t
t
s
s
t+s t+so
253
(T1 T 1) with t > 0 enjoys the semi-group property (T1 T1} (T. f.) =
T,+.ft+s Hence {T1T1} is an equi-continuous semi-group of class (C0 ).
If xED(A) = D(A). then
lim h-1(Th ih- I) x = lim Thh-1 (Th- I) x

h}O
11-j.O
+ lim h-1(Th -I) x

h~.O
=Ax+Ax=O,
so that the infinitesimal generator A 1 of {T1 T1} is 0 at every xE D(A).

Since (I- A1} is the inverse of a continuous linear operator (I- A 1)-1 E
L (X, X}, we see that A 1 must be closed. Hence A1 must be= 0, in virtue
of the fact that A1 vanishes on a dense subset D(A) = D(A) of X.
(T1T1) x = lim exp(t. 0 (I- n-1 0)- 1) x = x, that is,
Therefore
.......00
T 1 T 1 =I. We have thus proved the group property of 5 1, - oo < t < oo,
where 5 1 = T 1 and 5_ 1 = f, for t > 0.
Corollary 1. For the case of a B-space X, the condition of the Theorem
reads: D (A )a = X and the resolvent (I- n-1A )-1 exists such that
II(!- n- 1 A)-m II< M (for m = 1, 2, ... and all large In 1. n ;;e: 0). (2)
Forthegroup51 satisfying liSe II< MePitl ({J > 0) forallt,-oo< t < oo,
the condition reads: D(A)a =X and the resolvent (I- n-1A)-1 exists
such that
II (I- n-1A)-m II < M (1-l n-1 1 {3)-m (for m = 1, 2, ...
(3)
and all large In 1. n S 0).
For the particular case IIS1 II < ePitl for all t, -oo < t < oo, the condition reads: D (A)" = X and the resolvent (I- n-1 A )-1 exists such that
II(!- n-1A)-1 11 < (1-l n-1 1{3)-1 (for all large In I, n ;;e: 0}. (4)
Proof. As the proof of Corollary 1 and Corollary 2 in Chapter IX, 7.
Corollary 2 (Stone's theorem). Let U1 , -oo < t < oo, be a group of
unitary operators of class (C0} in a Hilbert space X. Then the infinitesimal
times a self-adjoint operator H.
generator A of U 1 is
V-1
Proof. We have (U1 x, y) = (x, Ut" 1 y) = (x, U_ 1y) and so, by differentiation,
(Ax,y) = (x,-Ay) whenever x and yED(A).
Thus -iA =His symmetric. A being the infinitesimal generator of U1,
(I- n-1 A)-1 = (I- n-1iH)-1 must be a bounded linear operator such
that II (I- n-1iH}-1 11 < 1 for n = 1, 2, ... Hence, taking the case
n = 1, we see that the Cayley transform of His unitary. This proves
that H is self-adjoint.
254
Remark. If A is of the form A=

H, where His a self-adjoint
operator in a Hilbert space X, then condition (4) of Corollary 1 is surely
satisfied, as may be proved by the theory of Cayley's transform. Therefore, A is the infinitesimal generator of a group of contraction operators
U1 in X. It is easy to see that such U1 is unitary. For, a contraction
operator U 1 on a Hilbert space X into X must be unitary, if U11 = U -t
is also a contraction operator on X into X.
10. Holomorphic Semi-groups

We shall introduce an important class of semi-groups, namely, semigroups T 1 which can, as functions of the parameter t, be continued holomorphically into a sector of the complex plane containing the positive
t-axis. We first prove
Lemma. Let X be a locally convex, sequentially complete, linear
topological space. Let {T1 ; t > 0} ~ L (X, X) be an equi-continuous semigroup of class (C0 ). Suppose that, for all t > 0, T 1X ~ D (A), the domain
of the infinitesimal generator A of T 1 Then, for any x E X, T 1 x is infinitely differentiable in t > 0 and we have
Tln) x = (T;1,.)" X
for all
t > 0,
(1)
where T; = D 1 T 1, r;' = D 1 T;, .. . , Tinl = D 1 TlnProof. If t > t0 > 0, then T;x =A T 1 x = T 1_ 1.A T 1.x by the commutativity of A and T 5 , s > 0. Thus
X ~ T 1_ 1 X ~ D (A) when
t > 0, and so r;' x exists for all t > 0 and x E X. Since A is a closed linear
operator, we have
1l.
r;
r;' X =
De (ATe)
= A . n->-oo
lim . n (Tt+1/n- Te)
A Te1zA Te12 x
= A (ATe)
= (T;12) 2 x.
Repeating the same argument, we obtain (1).

Let X be a locally convex, sequentially complete, complex linear
topological space. Let {T1 ; t > 0} ~ L (X, X) be an equi-continuous semigroup of class (C0 ). For such a semi-group, we consider the following three
conditions:
(I) For all t> 0, T 1 xED(A), and there exists a positive constant
C such that the family of operators {(CtT;t} is equi-continuous in
n > 0 and t, 0 < t < 1.
(II) T 1 admits a weakly holomorphic extension T._ given by
and
T._x
00
= .1: (A.- t)" T)nl xjn!

n=O
for
Iarg A. I <
tan- 1 (C e- 1),
(2)
the family of operators {e-'-T;.} is equi-continuous in A. for
Iarg A. I <
tan-1 (2-kC e-1) with some positive constant k.
(3)
255
(III) Let A be the infinitesimal generator of T 1 Then there exists

a positive constant C1 such that the family of operators {(C1 A.R(A.;A))"}
is equi-continuous in n > 0 and in A. with Re (A.) > 1 + s, s > 0.
We can prove
Theorem. The three conditions (I), (II) and (III) are mutually equivalent.
Proof. The implication (I)~ (II). Let p be any continuous seminorm on X. Then, by hypothesis, there exists a continuous semi-norm q
on X such that p((tT;)" x) < c-"q(x), 0 < C for 1 > t > 0, n > 0
and x EX. Hence, by (1), we obtain, for any t > 0,
P((A. - t)"
xjn !)
y<nl
1
IA-t" t I"
<
-
n" _!_ p ((_!___ C T' )"

tfn
n
n ! C"
x)
< ( IA t tL c- 1 e)". q (x)' whenever 0 <
tjn
< 1.
Thus the right side of (2) is convergent for Iarg A. I < tan-1 (C e-1), and
so, by the sequential completeness of the space X, T;.x is well defined
and weakly holomorphic in A. for Iarg A. I < tan-1 (C e-1 ). That is, for any
x EX and IE X', the numerical function I (T1 x) of t, t > 0, admits a
holomorphic extension I(T;.x) for largA.J < tan-1 {Ce-1 ); consequently,
by the Hahn-Banach theorem, we see that T;.x is an extension of T 1x for
largitl<tan-1 {Ce-1). Next put S 1 =e- 1 T 1 Then s;=e- 1 T;-e- 1 T 1
and so, by 0 < te-t < 1 (0 < t) and (I), we easily see that the family
of operators {{2-kCts;r} is equi-continuous in t > 0 and n > 0, in virtue
of the equi-continuity of {T1} in t > 0. The equi-continuous semi-group
S 1 of class {C0 ) satisfies the condition that t > 0 implies S 1 X ~ D (A -I)
= D (A), where (A- I) is the infinitesimal generator of S 1 Therefore,
by the same reasoning applied above to T 1, we can prove that the weakly
holomorphic extension e-J.T;. of S 1 = e- 1 T 1 satisfies the estimate (3).
By the way, we can prove the following
Corollary (due to E. HILLE). If, in particular, X is a complex B-space
and lim lltT;II < e-1 , then X= D(A).
ItO
Proof. For a fixed t
>
0, we have lim II (tjn) T;1,. II < e-1 , and so the

n--->00
series
] (A.- t)" y<nl /
n=O
I -
n. -
n=O
{A-t)" n" (_!___ T' )"

n! n tfn
t"
strongly converges in some sector

{A.; lA.- t 1/t < 1
+b
with some
b > 0}
of the complex A.-plane. This sector surely contains A.
0 in its interior.
256
The implication
(III). We have, by (10) in Chapter IX, 4,
(II)~
;.n+I
(A.R(A.; A))"+l x = -n,
J e-;. t"T xdt

oo
. 0
Re(A.)
for
>
0, xE X.
(4)
Hence, putting 51 = e- 1 T 1, we obtain
((a+ 1 +
--
ir) R(a
(
ir; A))"+l x
J e-(a+iT)tt"S x dt , a >
. )n+l oo
+ +1 H
n.
0.
Let r < 0. Since the integrand is weakly holomorphic, we can deform,

by the estimate (3) and Cauchy's integral theorem, the path of integration: 0 < t < oo to the ray: re;o (0 < r < oo) contained in the sector
0 < arg A.< tan-1(2-kCe-1) of the complex A.-plane. We thus obtain
((a+
ir) R(a
_ (a
+ 1 +1 iT)"+l
n.
ir; A)t+l x
.
r einos,.,ox
eiOdr,
Joo -(a+iT)reiO n
and so, by (3),
p(((a +
ir) R(a
< sup p (S,.wx) I(a + 1nt ir) ln+l

O<r<OO
<q
=
1 (
ir; A)t+l x)
e(-acosO+rsinO)r r"dr
a + 1 + i r ln+l
h
X
Ir Isin
.
6 _a cos Ojn+i , w ere q 1s a contmuous sem1-norm on
1
A similar estimate is also obtained for the case r

with (7) in Section 4, we have proved (III).
>
0. Hence, combined
The implication (III) _. (1). For any continuous semi-norm p on X,

there exists a continuous semi-norm q on X such that
p ((C1 A.R (A.; A)t x) < q(x) whenever Re(A.) > 1 + e, s> 0, and n > 0.
Hence, if Re(A.o)
> 1 + s, we have
P((A.-A.0)" R(A.o; A)" x)
< i~1 ~;.~~;:q(x) (n =
0, 1, 2, ... ).
Thus, if /A.- A.0 //C1 /A.o / < 1, the resolvent R (A., A) exists and is given by
R(A.; A) x =
00
~ (A.0
n=O
A.r R(A.0 ; A)"+l x
such that
p(R(A.; A) x) < (1- C1 1 /A.0 /- 1 /A.-A.0 /)- 1 q(R(A.0 ; A) x).

Therefore, by (III), there exists an angle (}0 with n/2
R (A.; A) exists and satisfies the estimate
p (R(A.; A) x) <
l~l q' (x)
< (}0 <
n such that
with a continuous semi-norm q1 on X
(5)
in the sectors n/2 ::::;; arg A. < 00 and -00

Re (A.) > 0, when lA. I is sufficiently large.
Hence the integral
257
< arg A. < - n/2 and also for
T1 x=(2ni)-1 fe'-'R(A.;A)xdA. (t>O,xEX)

c.
(6)
converges if we take the path of integration C2 = A.(a), -oo <a< oo,

in such a way that lim IA.(a) I = oo and, for some e > 0,
Ia! too
e < arg A. (a) < 00 and -00 < arg A. (a) ::::;; - n/2- e
n/2
at
+oo and ai-oo, respectively; for /a/ not large, A.(a) lies
when
in the right half plane of the complex A.-plane.
We shall show that T 1 coincides with the semi-group T 1 itself. We
first show that lim T1x = x for all x E D (A). Let x0 be any element
1~0
E D(A), and choose any complex number A-a to the right of the contour
c2 of integration, and denote (A.ol- A) Xo = Yo
Then, by the resolvent
equation,
T1x0
T1 R(J.0 ; A) Yo=
(2ni)- 1 f ei.tR(J.; A) R(J.o; A) y0 d).

c.
= (2ni)- 1 f tl.t(Ao -A.)- 1 R(J.; A) y0 d.'.
c,
- (2ni)- 1 f eA1 (.'.0 -J.)- 1 R(J.o; A) Yod.'..
c.
The second integral on the right is equal to zero, as may be seen by
shifting the path of integration to the left. Hence
=
1'1x0 = (2ni)- 1
Je'-'(J.o-.'.)-
c.
R(J.; A) y0 d.'., Yo= (.'. 0 1 -A) x0
Because of the estimate (5), the passage to the limit t t 0 under the
integral sign is justified, and so
J (.'.
0 - A.)- 1 R (A.; A) y 0 dA., Yo= (A.ol- A) x0

c.
To evaluate the right hand integral, we make a closed contour out of the
original path of integration C2 by adjoining the arc of the circle lA. I = r
which is to the right of the path C2 , and throwing away that portion
of the original path C2 which lies outside the circle I.'./ = r. The value of
the integral along the new arc and the discarded arc tends to zero as
r too, in virtue of (5). Hence the value of the integral is equal to the
residue inside the new closed contour, that is, the value R(A.o; A) Yo= x0
We have thus proved lim T1x0 = x0 when x0 E D(A).
lim 1',x0 = (2ni)- 1

t.j,O
t~O
We next show that T' x = A T1x for t > 0 and x EX. We have
R(A.; A) X= D(A) and AR(A.; A)= J.R(J.; A) -1, so that, by the
17 Yoslda, Functional Analysis
258
J eu AR(A.; A)x dA. has a

c,
sense. This integral is equal to A T 1 x, as may be seen by approximating
the integral (6) by Riemann sums and using the fact that A is closed:
lim x.. = x and lim Ax,.= y imply xE D(A) and Ax= y. Therefore
convergence factor
n->00
til, the integral (2ni)- 1
n->00
AT1 x= (2ni)- 1
euAR(A.;A)xdA., t> 0.
c,
On the other hand, by differentiating (6) under the integral sign, we
obtain
= (2ni)- 1 e}JA.R(A.; A) xdA., t > 0.
(7)
c,
t;x
euxdA., and the value

c,
of the last integral is zero, as may be seen by shifting the path of integration to the left.
Thus we have proved that x(t) = Texo, XoED(A), satisfies i)
lim x (t) = x0 , ii) dx (t)jdt = Ax (t) for t > 0, and iii) {.X (t)} is by (6) of
The difference of these two integrals is (2ni)- 1
tj.O
exponential growth when t t oo. On the other hand, since x 0 E D (A)

and since {T1} is equi-continuous in t > 0, we see that x(t) = T 1 x0 also
satisfies limx(t) = x 0 , dx(t)jdt = Ax(t) fort> 0, and {x(t)} is bounded
lj.O
when t > 0. Let us put x(t)- X (t)
= y (t).
Then limy (t)

lj,O
A y (t) fort > 0 and {y (t)} is of exponential growth when t

may consider the Laplace transform
J e-uy(t) dt
0, dy (t)jdt
t oo. Hence we
00
L(A.; y) =
for large positive Re(A.).
We have
p
fJ
{J
J e-uy'(t) dt = J e-AtAy(t) dt =A J e-" y(t) dt,

1
..
"'
"'
0 <X<
fJ<oo,
by approximating the integral by Riemann sums and using the fact that
A is closed. By partial integration, we obtain
,
J e-" y'(t) dt = e-APy({J) -e-.<..y(~X) +A. J e-" y(t) dt,
"'
which tends to A.L(A.;y) as ~XtO, fJtoo. For, y(O)=O and {y({J)} is

of exponential growth as fJ too. Thus again, by using the closure property of A, we obtain
AL(A.; y) = A.L(A.; y)
for large positive
Re(A.).
Since the inverse (H -A)-1 existsfor Re(A.) > 0, we must haveL(A.; y) = 0

when Re (.A.) is large positive. Thus, for any continuous linear functional
259
11. Fractional Powers of Closed Operators
IE X',
we have
00
J e-JJ I (y (t)) dt
0 when Rt (A.) is large positive.
a+ i-c and put

ga(t) = e-atl(y(t)) or = 0 according as
We set A.=
Then,
the
{2n)- 1
J e-itga(t)dt
above
equality
shows
that
t > 0 or t < 0.
the
00
-oo
vanishes identically in
T,
Fourier
transform
-oo<r<oo,
so
that, by Fourier's integral theorem, ga(t) = 0 identically. Thus l(y(t)) = 0

and so we must have y (t) = 0 identically, in virtue of the Hahn-Banach
theorem.
Therefore T1x = T 1 x for all t > 0 and xED (A). D (A) being dense in
X and T1, T 1 both belonging to L (X, X), we easily conclude that T1 x = T 1 x
for all x EX and t > 0. Hence, by defining T0 =I, we have T1 = T 1
for all t > 0. Hence, by (7), r;x = (2ni)- 1 J eJJA.R(A.; A) xdA., t > 0,
c,
and so, by (1) aiid (5), we obtain
Hence
(T;tntx=T/n)x=(2ni)- 1 JeJJA.nR(A.;A)xdA., t>O.

c.
(tT;t x =
{2ni)- 1
Therefore, by (III),
JenJJ(tA.t R(A.; A) xdA.,
c.
t > 0.
p ((tT;t x) < (2n)- 1 q (x) J IenJJ I tn lA. Jn-l d JA. J.

c,
If 0 < t < 1, then the last integral is majorized by c; where C3 is a

certain positive constant. This we see, by splitting the integration path
C2 into the sum of that in the right half-plane Re (A.) > 0 and those in
the left half-plane Re(A.) < 0, and remembering the integral representation of the F-function.
References. The result of the present section is due to K. YosiDA [6].
See also E. HILLE [3] and E. HILLE-R. S. PHILLIPS [1].

Let X be a B-space, and {T1 ; t > 0}
semi-group of class {C0 ). We introduce
1
le,"'(A.) = 2 ni
=
17*
a+ioo
~-ti' dz(a
>
0, t
>
L (X, X) an equi-continuous
0, A.> 0, 0 <ex< 1),
0'-100
{when A.
<
0),
(1)
260
where the branch of z'" is so taken that Re (z"') > 0 for Re (z) > 0. This
branch is a one-valued function in the z-plane cut along the negative
real axis. The convergence of the integral (1) is apparent in virtue of the
convergence factor e- 1"'. FollowingS. BocHNER [2] and R. S. PHILLIPS
[5], we can show that the operators
00
T ,"'x = T x = J / ,"'(s) T.xds

1
(t
>
0),
(2)
(t = 0),
constitute an equi-continuous semi-group of class (C0 ). Moreover, we can
show that {T1} is a holomorphic semi-group (K. YosiDA [8] and V. BALAKRISHNAN [1]). The infinitesimal generator A =A"' of T1 is connected
with the infinitesimal generator A of T 1 by
=X
i"'x =-(-A)"' x
xE D(A),
for
(3)
where the fractional power (-A)" of (-A) is given by

.
(-A)"' x = sm !X.n
n
J A."'o
00
(U- A)- 1 (-Ax) d).
for xE D(A)
(4)
and also by
(-A)"' x
J ;.,-"'00
F(-cx)- 1
(T;.- I) xd).
for
xED (A).
(5)
Formula (4) and (5) were obtained by V. BALAKRISHNAN. For the

resolvent of A"'' we have the following formula due toT. KATO:
(pi-
A"')- 1 =
00
sm !X.n
(r I - A)- 1
n o
p. 2
"'
r
2p.r"' cos !X.7t
+r
2"'
dr.
(6)
In this way, we see the abundance of holomorphic semi-groups among

the class of equi-continuous semi-groups of class (C0 ).
To prove the above result, we have to investigate the properties of
the function / 1,"' (A.) in a series of propositions.
Proposition 1. We have
00
e-ta"'=J e-Aa/1,"'(A.)d). (t>O,a>O).
(7)
Proof. It is easy to see, by the convergence factor e-"'1, that the

function /1,"' (A.) is of exponential growth in A.. By Cauchy's integral theorem, integral (1) is independent of a > 0. Let a > a = Re (z) > 0.
Then, by Cauchy's theorem of residue,
oo
J e-Aa f
1"'
(A.) d).
1
= -.
a+joo [ A(-a}] .<-oo
2n~ a-joo
-1
2ni
J.
a+ioo
a-100
_ez- a
1
A=O
-z"'t
z- a e
e-"'t dz
-ta"'
261
,"'(A.) > 0 for all A.> 0.

Proof. If we set a"'= g(a), e-tx = h(x), then
(8)
/1
(-1t- 1 g<nJ (a)> 0
>
(-1t h<n>(x)
(n
(n
=
=
1, 2, ... ), g(a) > 0
and
0, 1, ... ), when a> 0
and x
>
0.
Hence k(a) = h(g(a)) =e-ta"' satisfies
(-1t k(n) (a)
(-1) h' (x) (-1t- 1 g<nJ (a)
C!P~P.P.> (-1)P h(p,J (x) (-1)P' g<P.+1l (a) ... (-1)Pv g<P.+1l (a)
(C!;l >
O,p0 > 2,p1 > 0, ... ,p. > 0 with Po<.~ Pi= n, and v arbitrary)
>
0, (n
= 0, 1, ... ) .
(9)
That is, the function k (a) =e-ta"' is completely monotone in a> 0.

We next prove the Post-Widder inversion formula
I1,01 (A.)
(n)
(n
(-l)R
-;. )n+1 k(n) -;.
= lim !
nn-+OO
A. > 0
(10)
so that, by (9), f1,01 (A.) > 0. The proof of (10) is obtained as follows. We
find, differentiating (7) n-times,
k(n)(;)=(-1t
sne-sn/Af1,"'(s)ds.
Substituting this into the right side of (10), we get
J[
00
. nn+l 1
~enn!T
)]n / ,"'(s)ds.
-yexp 1--y
Since
lim nnJV2:n:n enn!
11-+00
1 (Stirling's formula),
we have to prove that
,"'(A.o)
/1
= !~ ~
n 312
[Io exp ( 1- ~)T /,"'(s) ds, A > 0.
Let 'YJ be a fixed positive number such that

last integral into three parts,
00
Ao-11
Ao+11
'YJ
< Ao
(11)
We decompose the
00
J = /1 + /2 +fa
J + Ao+'l
J = J + Ao-'1
Since x exp(1- x) increases monotonically in [0, 1] from 0 to 1, we see,
262
11 "'(s)
by the boundedness of
'
in s, that lim ] 1
11->00
0. Next, since
x exp (1- x) decreases monotonically in [1, oo] from 1 to 0, we have

Ao
and so, since
l 1,.,(s}
exp(1- Ao
1J.
1,
is of exponential growth as s too,
J (~t
00
!fa! <
1]) < {3 <
n 312 en pn-n,
exp (- n~s) !lt,a< (s)
Ids~ 0
as n ~ oo.
By the continuity of l 1,"'(s) ins, we have, for any positive number e,
lt,a< (A0)
+e
e < It,"' (s) < lt,a< (Ao)
if we take 1J
>
where
Ao- 1J <
whenever
s<
Ao + 1J,
0 sufficiently small. Thus
Ut, (Ao)- e) fo < !2 < Ut,a< (Ao)

] 0
Ao+'1 3 2 [ s
(
s
n '
1 exp 1- 1
Ao-'7
'"O
'"O
+ e) lo,
(12)
)Jn ds.
(13)
The whole preceding argument is true for the particular case of the
completely monotone function
k(a)
a- 1
J e-;.a,dA.
00
In this case, k(n) (n/Ao) = (-1t n! (Ao/nt+l. Substituting this in (10), we

1. Since (10} and (11) are equivalent,
find that (10) holds for / 1,.. (A.)
(11) must also hold for It" (A.) - 1. Thus, since lim ] 1 = 0 and lim ] 3 = 0
for a general
11,"', we
'
1>-->00
11->00
obtain
1
= n--+00
lim V2Jl 101 ]
Therefore, by (12), we get (11) and the equivalent formula (10) is proved.
Proposition 3.
00
f
0
ft+s,"'(A)
I1,"'(A) dA = 1,
(14)
00
J lt,.,(),-p,) fs,O<(p,) dp,.

0
(15)
Proof. Since the function 11,"' (A) is non-negative, we have, by the

Lebesgue-Fatou lemma and (7),
J lim (e-Aalt,IX(A)) dA <lim e-ta"'= 1.

00
a,j,O
aj-0
Thus lt,a(A) is integrable with respect to A over (0, oo) and so, again by
263
the Lebesgue-Fatou lemma and (7), we obtain (14). We next have, by (7),
=
Je-J.a {fIt,"'(?.- f-l) /s,"' (f-l) df-l} dl.

Je-(A-p)a / (A- p,) d (A- f-l) J e-pa Is,"' (f-l) df-l
e-ta"' e-sa"' -_ e-(t+s)a"' -_ J e-J.a lt+s,O< d, a ---.,.
__ 0 .
1,"'
I
(,)
-_
11.
11.,
Hence, by inverting the Laplace transform as in the preceding section,
we get (15).
00
f
0
8/1,"'(l)f8tdl=0, t>O.
(16)
Proof. By deforming the path of integration in (1) to the union of

two paths re-i 0 (-<Xl <- r < 0) and rei 0 (0 < r < <Xl), where n/2 <
0 < n, we obtain
/ 1,1X(s)
=--;- J exp(sr cos 0- triX cos <XO)

00
X sin (sr sin 0- triX sin rX 0 + 0) dr.

Similarly, by deforming the path of integration in
a+ioo
8ft,O<(J.)j8t = 2 ~i
CA-z1Xt (-z") dz
(17)
o-'oo
to the union of two paths, re-i 0 ( - <Xl

we obtain
/;,IX(s)
= 8/1,1X(s)j8t = ( nl)
<-
X sin (sr sin 0 - triX sin rX 0

then
t;,IX(s)
0 = 00<
< r < <XJ),
exp(sr cos 0- triX cos <XO)
If we take
r < 0) and ri 0 (0
n/(1
+ rX 0 + 0) riX dr.
(18)
<X),
=! j exp((sr+triX)cosOIX)sin((sr-tyli<)sinOIX)riXdr.
(19)
Thus we see, by the factor riX (0 < rX < 1), that t;,IX(s) is integrable with
respect to s over (0, <Xl). Hence, by differentiating (14) with respect to t,
we get (16).
We are now able to prove
Theorem 1. {T1} is a holomorphic semi-group.
Proof. That {T1} enjoys the semi-group property T 1 T 5 = ft+s
(t, s > 0) is clear from (2) and (15). We have, by (2) and (17) with 0 = 0",
A
T 1x = 1"&
J T.xds J exp((sr + triX) cos 00<)

00
00
X sin ((sr- triX) sin 00<
+ 00<) dr,
(20)
264
which gives by the change of variables

S = vt1'1X, r = url/IX
1
(21)
J Tv11/cx xdv 0J exp ((uv + uiX) cos 0,.)

:n 0
00
T 1x = -
00
X sin ((uv- u"') sin (),.
+ ()IX) du.
(20')
The second integral on the right is exactly :n; It,,. (v), and so, by the
equi-boundedness of {IIT1x/j} in t > 0, we see, by (14), that
II T x II <
1
sup II T 1 x 1/
1~0
00
J/
1 ,,.
(v) dv =sup II T 1 x II.
(22)
1~0
By passing to the limit q 0 in (20'), which is justified by the integrability of ft,,.(v) over [0, =),we obtain, by (14),
s-lim T 1 x
tj.O
00
J !Ie<(v)
dv x = x.
o
'
Hence {T1} is an equi-continuous semi-group of class (C0) such that (22)

holds.
By the integrability of t;,,.(s) = 8/1,,.(s)/8t over [0,=) and the equicontinuity of {T1}, we obtain, by differentiating (2) with respect to t
under the integral sign,
T~x
00
J J;,,.(s) T
xds
J T xds J exp ((sr + triX) cos 0,.) sin ((sr- triX) sin ()IX) riX dr,
:n
=-
00
00
(23)
which is, by the change of variables (21),
00
J (Tvtll"' x) /~,IX (v) dv t-
Thus, by the integrability of fi.,IX(v) over [0, =) and the equi-continuity

of {T1} in t > 0, we see that
lim
ItO
llti; II<=.
that is, {T1} is a holomorphic semi-group.

Theorem 2. The infinitesimal generator AIX of T1 is connected with
the infinitesimal generator A ofT by (3), where (-A)"' is defined by (4)
and also by (5). We have, moreover, (6).
265
Proof. By (16) and (23), we obtain

1
T~x = -
J (Ts- I) xdsJ
00
00
:n; 0
exp ((sr
+ tr"') cos 0.,)
X sin ((sr- tr"') sin Oa) r"' dr.
(24)
II (Ts- I) xll is boun-
If x E D(A), then s-lim s-1 (Ts- I) x =Ax and

s~o
ded in s > 0. Thus, we obtain, letting t t 0 in (24),

00
00
J (Ts- J) x ds 0J exp(sr cos()"') sin(sr sin()"') r"' dr

:n; 0
s-lim f;x = _!._

1~0
00
(-F (-~))- 1
s-"'- 1 (Ts- I) xds,
because, by the F-function formulae
J e-cr ~- 1 dr
00
F(z) = cz
(Re(z)
and
>
0, Re(c)
>
(25)
0)
(26)
F(z) F(l- z) = nfsin nz,

we obtain, by
(~
+ 1) ()"' =
n,
-1 J exp(srcosO"') sin(srsinO.,)r"'dr =
00
:n; 0
(ni)- 1 Im((-se'6"')-"'- 1) F(l
+ ~) =
(ni)- 1 Im
J e-<-s' "'lr"'dr
{00
s-"'- 1 n- 1 sin (~n) F(l
+ ~)
_ (-F(- ))-1 -tx-1

__ "'_1 -F(1 + a:)
s
~
r (-1X) r(1 + 1X) - s
Thus, by f~x = AA.,T1x (when t > 0), the continuity at t = 0 of T1 x and

the closure property of the infinitesimal generator A"', we obtain
A"' x =
(-F(-~))- 1
J s-"'- (T -I} xds

00
when
xED(A).
J e-tsT ds, we obtain

00
Hence, by (25), (26) and (tl- A)- 1 =
1{i
~)- 1 F(l + ~)- 1

sm om J t"' ((tl- A)- 1 :n;
A"'x = F(-
00
J t"'0
t- 1 I) xdt
00
=sma::n;
:n;
e- 81 t"'dt} (I- T 8 ) xds
1 (tl-A)- 1 Axdt
for xED(A).
266
Finally we have, by taking(}= n in (17) and (2),

(pJ;rr,- 1
00
00
00
= n- 1
J e-"' T dt
0
J dr J e- 'T ds J exp (-11t- tr'" cos .xn) sin (tr"' sin .xn) dt
00
A,.)- 1 =
(r
I- A)-
1 {[
exp (-11t- tr"' cos .xn) sin (tr"' sin .xn)
dt} dr
Remark. Formula (2) was devised by S. BOCHNER [2] without detailed proof. Cf. R. S. PHILLIPS [5]. That T 1 is a holomorphic semi-group
was proved by K. YOSIDA [8], V. BALAKRISHNAN [1] and T. KATO [2].
Formulae (4) and (5) are due to V. BALAKRISHNAN [1], who, by virtue of
(4), defined the fractional power (-A)"' of a closed linear operator A
satisfying the condition:

the resolvent R (A; A) = (U- A)-1 exists for Re (A)
>
0
(27)
and sup jRe(A)jjjR(A;A)I!<oo.

Re(.I.)>O
He also proved that (-A)" enjoys properties to be demanded for the

fractional power. In fact we have
Theorem 3. Let a closed linear operator A satisfy condition (27). Then,
by (4), a linear operator (-A)"' is defined, and we have .
(-A)"' (-A)II x
with .x
+ {3 <
(-A)"'+II x
if xED (A 2 )
and
0 <<X, {3
(28)
1,
s-lim(-A)"'x=-Ax if
s-lim (-A)"' x = x
a}O
if
(29)
xED(A),
at1
s-lim AR(A; A) x
.1.}0
(30)
0,
and, if A is the infinitesimal generator of an equi-continuous semi-group

T 1 of class (C0),
(A,.) 11 = A,.p, where A,. is the operator
A,. defined through
Kato's formula (6).

Remark. The last formula (31) is due to
llr'"-
J. WATANABE
(31)
[1].
(-Ax) II is, by (27), of order O(r"'- 2 ) when

Proof.
1
r too, and, by (rl -A)- (-Ax)= x-r(rl -A)- 1 x and (27), it is
of orderO(r'"- 1 ) when r t 0. Thus the right side of (4) is convergent.
1 (rl
-A)- 1
267
It is clear that xED (A 2 ) implies (-A)II xED (A). For, by approximat-
ing the integral by Riemann sums and making use of the closure property
of A, we have (-A)IIx ~ D(A). We can thus define (-A)"'(-A)IIx.
(-A)'" (-A)II x = sinncxn
s~n
j j ).11-lp,a.-l R(A.; A) R(p,;A)A 2xdA.dp,

0
can be rewritten, decomposing the domain of integration into those for

which).~ p, and).< p,, as follows.
(aP-1
sin cxn sin {Jn
n o
n
+ a"'-1) da j
).'"+Il-l R(A.a; A) R().; A) A2xd)..
By the resolvent equation R ().; A) - R (p,; A) = (p, - A.) R (A.; A) R (p,; A)

and R(A.; A) (-A)= I -A.R().; A) valid on D(A), we obtain
R(A.a; A) R(A.; A) A 2 x = (1- a)-1{-aR(A.a; A) + R(A.; A)} (-Ax),
and hence
(-A)'" (-A)II X
J
o
I
= sin cxn sin {Jn s-lim

n
ttl
00
A)
).'"+Il-l (-aR(A.a;
= Joo (sin cxn sin {Jn

n
+ oP-1) (1 _
(aP-1
a)-1 da
+ R().; A)) (-Ax) d).
Jl aP- +
a'"- 1 - a-'"- a-ll da)

1-a
X ;.a.+II- 1 R(A.;A)(-Ax)d)..
The coefficient ( ) above is evaluated as n-1 sin n(X + {3), as may be

seen by expanding (1- a)-1 into powers of a. We have thus proved (28).
J ).'"-1 (1 + A.)-1 d). =

00
To prove (29}, we shall make use of

Thus
(-A)"' x - (-A) x = sinncxn
nfsinom.
)."'-l(R().; A)
-A!
1) (-Ax) d)..
We split the integral into two parts, one from 0 to C and the other from
C to oo. For a fixed C, the first part goes to zero as X t 1, since
R (A.; A) (-Ax) = x - ).R ().;A) xis bounded in).> 0. The second part
is, in norm,
~ nst~~:) c"'-1 ~~g I (A.R(A.; A)- A~1) Ax II
Wehaves-limA.R().; A) x = xby x-A.R(A.; A) x = R(A.; A) (-Ax) and

A too
(27). Hence s-lim of the second part is arbitrarily near 0 if we take C
<>~tl
sufficiently large. This proves (29).
268
To prove (30), we split the integral into two parts, one from 0 to C and
the other from C to oo. Because of (27), the second part tends to zero
as IX 0. By R(A.; A) (-Ax)= x -A.R(A.; A) x and the hypothesis
s-limA.R(A.; A) x = 0, the first is, for sufficiently small C, arbitrarily
.qo
near the value (~Xn)- 1 sin ~Xn C"'x, which tends to x when ~X t 0. This
proves (30).
We shall prove (31). By virtue of the representation (6), we obtain
(pi- (A"')p)- 1 =
j j
\.u-J.; e-in{J- p.-J.~ ein{J)
(2 ni)-2 {-
o o
C: e miX-;._ C:. em"') (Cl- A)- dA. dC.

1
X (;. _
This double integral is absolutely convergent in norm, and so we may

interchange the order of integration. Hence we obtain (31), because the
inner integral is
= (2ni)-2
1 . ) dz
J--=-!___
(--1- . - - z-C"'e"""
p.- zP z- C"' e-"'"'
(2ni)- 1 ~
~~-m"'fJ-,.,. _ ~18~niX{J)
where the path of integration C runs from oo e'" to 0 and from 0 to

oo e-m.
An example of the fractional power. If tx = 1/2, then we have, by
taking()= n in (17),
J e-sr sin (tr1' 2) dr =
vn t (
00
!t,1/2 (s) = n-1
n- 1
23
Jlst 3 e-l'/4s.
(32)
Thus, if we take the semi-group {T1} associated with the Gaussian kernel:
(T1 x) (u) = 2 ,~
e-Cu-v)'/4 x(v) dv, x E C [-oo, oo],
y ns -oo
then
(f,,1/2x)
(u) =
j j
-oo
x (v)
4~sa e-C<u-">'+t')/4 ds) dv
00
:n
-oo
tz
+ (~- v)a x (v) dv'
that is, the semi-group {T1} is associated with the Poisson kernel. In this
case, the infinitesimal generator A of T 1 is given by the differential operator d2 fds 2 , while the infinitesimal generator A of T1 is given by the singular
integral operator
J
00
(A 112 x) (s) = s-hmTt.j.O :It
x(s-v) -x(s)
-oo
and not by the differential operator dfds.
+h
dv,
269

We shall denote by exp (t A) the semi-group of class (C0) with the
infinitesimal generator A. Concerning the convergence of semi-groups,
we have
Theorc:m 1. Let X be a locally convex, sequentially complete complex

linear space. Let {exp (tA,.)} ~ L (X, X) be a sequence of equi-continuous
semi-groups of class (C0 ) such that the family of operators {exp(tA,.)} is
equi-continuous in t > 0 and inn= 1, 2, ... Thus we assume that, for
any continuous semi-norm p (x) on X, there exists a continuous semi-norm
q (x) on X such that
p(exp(tA,.) x) < q(x) for all t > 0, xE X and n = 1, 2,...

Suppose that, for some A.o with Re(A.o)
>
(1)
0,
lim R (.1.0 ; A,.) x = J (.1.0) x exists for all x E X in such
'1-+00
a way that the range R (] (J.o)) is dense in X.
(2)
Then ] (.1.0) is the resolvent of the infinitesimal generator A of an equicontinuous semi-group exp(tA) of class (C0) in X and
lim exp(tA,.) x = exp(tA) x for every xE X.
(3)
Moreover, the convergence in (3) is uniform in ton every compact interval

of t.
For the proof, we prepare
Lemma. Let Tt = exp(tA) be an equi-continuous semi-group of class

(C0 ) in X. Then, for any continuous semi-norm p (x) on X, there exists a
continuous semi-norm q (x) on X such that
p(Ttx- (I- tn- 1 A)-n x) < (2n)- 1 t 2 q(A 2 x) (n = 1, 2, ... )
(4)
whenever xED (A 2 ).
Proof. Set T(t, n) =(I -n-1 tA)-". Then we know (Chapter IX,7)
that {T(t, n)} is equi-continuous in t > 0 and n = 1, 2, ... Moreover,
we have (Chapter IX,4), for any xE D(A),
DtT(t, n) =(I- n-1 tA)-n- 1 Ax= A (I- n- 1 tA)-n-l x,

DtTx
TtAx
ATtx.
Thus, by the commutativity of Tt and T(t, n),

t
Ttx- T(t, n) x =
J [DsT(t- s, n) TsxJ ds
=I T(t-s,n)Ts(Ax-(1-t n s
Ar
{5)
1
Ax)ds, xED(A).
270
Hence, if xED(A 2 ), we have, by (I-m- 1A)-1Ax=-m(I-(I- m-1 A)-1)x,

p (T1 x -T(t, n) x) <
J p [ T(t-s, n) T
(I -n-1 (t-s) A)-1 5 n
A 2 x] ds,
and so, by a continuous semi-norm q (x) on X which is independent of x

andn,
p (T1 x- T (t, n) x) < (2n)-1f2q (A 2x).
Corollary. For any x E D(A 2 ), s > 0 and t
p(T1x - (I- sAf[tfsJ x)
<
sq1 (Ax)
> 0,
t;
q(A 2 x), where
q1 (x) is a continuous semi-norm on X which is indepen-
<
dent of x, t and s, and [t/s] is the largest integer
(6)
tjs.
Proof. We have, fort= ns,
p(Tnsx- (I -sA)-n x)
If t
ns
~t
with 0 < u
P(T1 x - Tnsx) = P (
<
s and n
<
[tfs],
f T~xda) < f
ns
2- 1 stq(A~x).
ns
p(TaAx) da
<
sq1 (Ax).
Proof of Theorem 1. By (1) and by (11) in Chapter IX, 4, we see that

{(Re (A.) R (A.; An)m)} is equi-continuous in Re (A.) > 0, in n = 1, 2, ... and
in m = 0, 1, 2, ... From this and by (2), we can prove that J(A.0 ) =
(Aoi- A )-1 with some A and
lim R(A.; An) x
R(A.; A) x whenever Re(A.)
>
0, and}
t;:;onvergence is uniform with respect to A. on every compact subset of the right half-plane Re (A.)
>
(7)
0.
To this purpose, we observe that
and that the series is, by the equi-continuity of {(Re(A.0 ) R(A.~; A;,))m} in
n = 1, 2, ... and m = 0, 1, 2, .. , uniformly convergent for lA.- A.0 I/
Re (Ao) < 1-8 and n = 1, 2, ... Here 8 is a fixed positive number. Therefore, for any b > 0, there exists an m0 and a continuous semi-norm q(x)
on X such that, for lA.- 20 I/ Re (Ao) < 1-8,
m,
P(R(A.; An) X - R(A.; An') x) < m=O

.I JA.o-J.Jm.
271
Hence, by (2), we see that lim R (A.; A,.) x = J (A.) x exists uniformly for
n-->00
(A.- J.o (/Re(A.0 ) < 1- e. In this way, extending the convergence domain
of the sequence {R (A.; A,.)}, we see that
J (A.) x exists
lim R (A.; A) x =
n-->00
and the convergence is uniform
on any compact set of A. in the right half-plane Re (A.)
>
0.
Thus J (A.) is a pseudo-resolvent, because J (A.) satisfies the resolvent

equation with R (A.; A,.) (n = 1, 2, ... ). However, by R (J (A.o))" =X and
the ergodic theorem for pseudo-resolvents in Chapter VIII, 4, we see that
J ().) is the resolvent of a closed linear operator A in such a way that
](A.)= R(A.; A) and D(A) = R (R(A.; A)) is dense in X.
Thus we see that exp(tA) is an equi-continuous semi-group of class
(C0 ) in X. We have to show that (3) is true. But we have, by (6),
p((exp (tA,.)- (I- sA,.)-[tfs]) (I- A,.)- 2 x)

<sq1 (A,.(I-A,.)- 2 x)
+ 2- 1 tsq(A!(I-A,.)- 2 x),
for any x E X, s > 0 and t > 0.

The operators
A,.(I -A,.)-1 =(I -A,.)-1 -I, A,.(I -A)-2 = A,.(I- A,.)-1 (I- A,.)-1
and
are equi-continuous in n
A~ (I- A,.)-2
= (A,. (I- A,.)-1)2
1, 2, ... On the other hand, by (7),
lim (I- sA,.)-[tfs] (I- A,.)- 2 x = (I- sA) -[tis] (I- A)- 2 x
n-->00
uniformly ins and t, if s > 0 is bounded away from 0 and=, and if t

runs over a compact interval of [0, =). Moreover, we have, by (6),
p ((exp (tA)- (I- sA)-[tfsl) (I- A)- 2 x) < sq1 (A (I- A)- 2 x)
+
for every x EX, s
small, we see that
2- 1 tsq(A 2 (I -A)- 2 x)
>
0 and t
lim exp (tA,.) y
n-->00
> 0. Thus, by taking s > 0 sufficiently
exp (tA) y for any y E R (1; A) 2 X,
and the convergence is uniform in t on every compact interval of t.

R(1;A) 2 X being dense in X, we see, by the equi-continuity of exp
(tA) and exp (tA,.) in t > 0 and in n = 1, 2, ... , that (3) holds.
Theorem 2. Let a sequence {exp (tA,.)} of equi-continuous semi-groups

of class (C0 ) in X be such that {exp(tA,.)} is equi-continuous in t > 0 and
inn= 1, 2, ... If, for each xE X,
lim exp(tA,.) x = exp(tA) x
n-->00
272
uniformly with respect to t on every compact interval of t, then

lim R (A.; An) x = R (A.; A) x for each x E X and Re (A.) > 0
n-->00
and the convergence is uniform on every comp:1ct set of A.

in the right half-plane Re (A.) > 0.
Proof. We have
J e-At (exp (tA)- exp (tAn))

00
R (A.; A)
X-
R (A.; An) x =
dt.
Hence, splitting the integral into two parts., one from 0 to C and the
other from C to=, we obtain the result.
Remark. For the case of a Banach space X, Theorem 1 was first
proved by H. F. TROTTER [1]. In his paper, the proof that J (A.) is the
resolvent R (A.; A) is somewhat unclear. This was pointed out by T. KATO.
The proof given above is adapted from Kato's modification of Trotter's
proof of the Banach space case (unpublished).
13. Dual Semi-groups. Phillips' Theorem

Let X be a locally convex sequentially complete linear topological
space, and {T1 ; t > 0} ~ L(X, X) an equi-continuous semi-group of
class {C0 ). Then the family {T;"; t > 0} of operators E L (X', X') where (*}
denotes the dual operators in this section, satisfies the semi-group
property: T;"T; = Ti+s T6 =I*= the identity in X~ (see Theorem 3
in Chapter VII, 1). However, it is not of class (C0 ) in general. For, the
mapping T 1 _,.. Tj does not necessarily conserve the continuity in t (see
Proposition 1 in Chapter VII, 1). But we can show that {T*} is equicontinuous in t > 0. For, we can prove
Proposition 1. If {51 ; t > 0} ~ L (X, X) is equi-continuous in t > 0,
then {5;"; t > 0} ~ L (X', X') is also equi-continuous in t > 0.
Proof. For any bounded set B of X, the set U 5 1 B is by hypothesis
1:?2:0
a bounded set of X. Let U' and V' be the polar sets of B and U 5 1 B:
U' ={x' EX'; sup l<b,x') [ < 1}, V' ={x' EX'; sup
bEB
bEB,t;;;;o
1<5
t;;;;o
b,x') I< 1}.
Then (see Chapter IV, 7), U' and V' are neighbourhoods of 0 of X~. From
I<51 . b' x') I = I< b' 5;" x') I < 1 (when bE B' x' E V')
we see that 5;" V' ~ U' for all t > 0. This proves that {5;"} is equicontinuous in t > 0.
Let A be the infinitesimal generator of the semi-group T 1 Then
D(A)a=X, R(A) ~X, and, for A.> 0, the resolvent (H-A)-1 E
L (X, X) exists in such a way that
{A.m (A.I- A )-m} is equi-continuous in A. > 0 and in m = 0, 1 . . .
(1)
We can prove (cf. Theorem 2 in Chapter VIII, 6)
273
13. Dual Semi-groups. Phillips' Theorem
Proposition 2. For A> 0, the resolvent (AI*- A *)-1 exists and
(AI*- A *)-1 = ((U- A)-1)*.
(2)
Proof. We have (U- A)* = U*- A*. Since (AI- A)-1 E L (X, X),
the operator ((U- A)-1)* E L (X', X') exists. We shall prove that
(U*- A *)-1 exists and equals ((AI- A)-1)*. Suppose there exists an
x' E X' such that (U*- A*) x' = 0. Then 0 = (x, (AI*- A*) x') =
((U- A) x, x') for all xE D(A). But, since R(U- A)= X, we must
have x' = 0. Hence the inverse (AI*- A *)-1 must exist. We have, for
xE X, x' E D(A*),
(x, x') = ((AI- A) (AI- A)-1 x, x') = ((U- A)-1 x, (AI*- A*) x').
and((AI -A)-1)*(AI*-A*)x' =x'

for every x' ED (A*). This proves that ((AI- A)-1)* ~ (AI*- A *)-1.
On the other hand, if xED (A) and x' E D(((U- A)-1)*), then
(x, x') = ((AI- A)-1 (AI- A) x, x') = ((U- A) x, ((AI- A)-1)* x').
This proves that D (A*) = D((U- A)*) ~ R(((U- A)-1)*) and
(U- A)* ((U- A)-1)* x' = x' for every x' E D(((U- A)-1)*), that
is, ((U- A)-1)* ~ (U*- A *)-1. We have thus proved (2).
We are now ready to prove
ThusD(((AI-A)- 1 )*)~R(U*-A*)
Theorem. Let X be a locally convex sequentially complete linear

topological space such that its strong dual space X' is also sequentially
complete. Let {T1} ~ L (X, X), t >- 0, be an equi-continuous semi-group
of class (C0 ) with the infinitesimal generator A. Let us denote by X+ the
closureD (A *t of the domain D (A*) in the strong topology of X'. Let Tt
be the restriction of T~ to x+. Then Tt E L (X+, x+) and {Tt; t > 0}
is an equi-continuous semi-group of class (C0) such that its infinitesimal
generator A+ is the largest restriction of A* with domain and range in X+.
Remark. The above Theorem was proved by R. S. PHILLIPS [2] for
the spepial case of a B-space X. The extension given above is due to
H. KOMATSU [4].
Proof of the Theorem. We have the resolvent equation R(A; A)-R(,u;A)=(.u-A)R (A;A)R(.u;A) and the equi-continuity of
{Am R(A;. A)m} in A.> 0 and in m = 0, 1, 2, ... Thus, by Proposition 1
and 2,
(AI*- A *)-1 - (.ui*- A *)-1 = (.u- A) (U*- A *)-1 (,ui*- A *)-1 (3)
{.Am (H*- A *)-m} is equi-continuous in A > 0 and in m
0, 1, 2, ... (4)
Thus, if we denote by] (A.) the restriction to X+ of (AI*-A *)-1, we have
J (A.) - J (,u) = (,u- A.) J (A.) J (,u)'

{.Am J (A.)m} is equi-continuous in A.> 0 and in m
18 Yoslda, Functional Analysis
(3')
0, 1, .
(4')
274
X. Compact Operators
Since D (A*) is dense in X+ and since (4') holds, we see, as in Chapter IX, 7,
that lim 1..](1..) X= X for xE x+. Thus we have R(J(I..)t = x+ and
J.->eo
so,. by (7') in Chapter VIII, 4, that N(J (I..))= {0}. Thus the pseudoresolvent J (1..) must be the resolvent of a closed linear operator A+ in X+.
Hence, by the sequential completeness of X+ and (4'), A+ is the infinitesimal generator of an equi-continuous semi-group of class (C0 ) of operators
Tt E L(X+, x+). For any xE X andy' Ex+, we have
<(I- m- 1 tA)-m x, y') = <x, (I*- m- 1 tA +rm y'),

and so, by the result of the preceding section, we obtain by letting
m --+ oo the equality <T1 x, y') = <x, Tty'). Hence Tj y' = Tt y', that
is, Tt is the restriction to x+ of Tj.
We finally show that A+ is the largest restriction of A* with domain
and range in X+. It is clear, by the above derivation of the operator A+,
that A+ is a restriction of A*. Suppose that x' ED (A*) and that x' EX+,
A* x' EX+. Then (U*- A*) x' EX+ and hence (U*- A+)-1 (AI*- A*)
x' = x'. Thus, applying (I.. I*- A+) from the left on both sides, we obtain
A*x' = A+x'. This proves that A+ is the largest restriction of A* with
domain as well as range in X+.
Let X andY be complex B-spaces, and let 5 be the unit sphere in X.
An operator TEL (X, Y) is said to be compact or completely continuous
if the image T 5 is relatively compact in Y. For a compact operator T E
L (X, X), the eigenvalue problem can be treated fairly completely, in
the sense that the classical theory of Fredholm concerning linear integral
equations may be extended to the linearfunctional equation Tx- l..x = y
with a complex parameter A. This result is known as the Riesz-5chauder
theory. F. RIESZ [2] and J. SCHAUDER [1].
1. Compact Sets in B-spaces

A compact set in a linear topological space must be bounded. The
converse is, however, not true in general; we know (Chapter III, 2) that
the closed unit sphere of a normed linear space X is strongly compact
iff X is of finite dimension. Let 5 be a compact metric space and C (5)
the B-space of real- or complex-valued continuous functions x(s) on 5,
normed by llxll = sup lx(s) 1- We know (Chapter III, 3) that a subset
sES
{x,.(s)} of C(5) is strongly relatively compact in C(5) iff {x,.(s)} is equibounded and equi-continuous in IX. For the case ofthe space LP (5, 58, m),
1 < p < oo, we have
275
1. Compact Sets in B-spaces
Theorem (FRECHET-KOLMOGOROV). Let S be the real line, 58 the

dx the ordinary Lebesgue
a-ring of Baire subsets B of S and m (B) =
measure of B. Then a subset K of LP(S, 58, m), 1 o:
Ix (s)
IP ds =
(2)
0 uniformly in x E K.
(3)
Proof. Let K be strongly relatively compact. Then K is bounded and

so (1) is true. Let s > 0 be given. Then there exists a finite number
of functions E LP: fi, /2 , . . , / .. such that, for each f E K, there is a f
with 11/-/j II :S::: s. Otherwise, we would have an infinite sequence {lj} c;_ K
with 11/j -/;II> s for f =I= i, contrary to the relative compactness of
K. We then find, by the definition of the Lebesgue integral, finitelyvalued functions gl> g2 , . . , g,. such that 11/j- gj II< s (f = 1, 2, ... , n).
Since each finitely-valued function gj (x) vanishes outside some sufficiently
large interval, we have, for large G,
(Joo + _Z"'
+ (/" +
lt(s)
(Joo + _Z"' lf(s) -- gj(s) lp astp
lp asfp <
_l"'lgj(s)fPasfp < llt-gjll + ( [ + _Z"'Igj(s)!Pas)l/P
IIIi- gi II < 2 s.
This proves (3) by II/- gill < 111-fi II
The proof of (2) is based on the fact that, for the defining function
00
CI (s) of a finite interval I, lim

t~o
Chapter 0, 3).
J
_
ICI (s + t)- CI (s)
IP ds = 0 (see
00
Thus (2) holds for finitely-valued functions
gi (s)
(f = 1, 2, ... , n). Hence we have, for any f E K,
E~ (_[ lf(s +
+ lim(
t~o
_ 00
t) -f(s) IP dsy!P <
~ (_[ if(s + t)-/j(s +
l!j (s + t) -gj (s + t) lp as)l/P + lim(_
+(_[!gj(s)-/j(s)
~o
Igj (s +
t) IP dstp
t)-gj (s) lp ds)lfp
00
IPasfp +(_[ltj(s)-f(s)fPasrp <s+s+ o+s+s.
by taking /j in such a way that 11/-/j II < s. This proves (2).

We next prove the converse part of the Theorem. We define the translation operator T 1 by (Td) (s) = j(t + s). Condition (2) says that
18*
276
s-lim T,f =I uniformly in

1-+0
(Mal) (s) = (2a)-1
IE K.
We next define the mean value
J (T f) (s) dt.
Then, by Holder's inequality and
-a
the Fubini-Tonelli theorem,

IIMal-111<
oo {
< (2a)- 1 (
<
}p )liP
_(2a)- 1 ll(s+t)-:-l(s)ldt ds
_ _ ll(s +
a
00
_ dt_
oo
(2a)- 1
t)
-l(s) IP dt (2a)PIP' ds
ll(s+t)-l(s)IPds
)1/P
)1~
if l<p<oo.
Thus we pave
IIMal-111 0.
We will show that, for a fixed a> 0, the set of functions {(Mal) (s);
IE K} is equi-bounded and equi-continuous. In fact, we have, as above,
a
I(M,.I) (s1) - (M,.f) (s2 ) I < (2a)-1

<
(2a)- 1
J Jl(s1 +
-a
t) -l(s2 + t) I dt
ll(s1 + t) -l(s2 + t) !P dt
)1~
Thus, by (2), we have proved the equi-continuity of the set of functions

{(Mal) (s); IE K} for a fixed a> 0. The equi-boundedness of the set may
be proved similarly. Thus, by the Ascoli-Arzela theorem, there exists, for
any positive a> 0, a finite number of functions Maft, Mal2 , , Main
with /j E K (j = 1, 2, ... , n) such that, for any IE K, there exists some j
for which sup I(M,.I) (s)- (Mal;) (s) l <e. Therefore
lsi:;; ..
II Mal- Mal; W< J llMal) (s) -(Mal;) (s) !P ds

+ J I(Maf)(s)-(M,.I;)(s)IPds.
(4)
lsi>'"
The second term on the right is, by Minkowski's inequality, smaller than
+ ( J ll(s)-/j(s) !Pds)11P+ ( J ll;(s)-(M,.I;) (s) 1Pds)11P)P.

lsi>'"
lsi>'"
The term IIM,.I-1 II is small for sufficiently small a> 0, and, by virtue
( IIM,.I-111
of (3),
lsi>'"
ll(s) -f;(s) IP ds and
lsi>"'
ll;(s)- (M,.I;) (s) IP ds are both
277
2. Compact Operators and Nuclear Operators
small for sufficiently large a> 0, if a> 0 is bounded. Also the first term
on the right of (4) is < 2aeP for an appropriate choice of f. These estimates are valid uniformly with respect to IE K. Thus we have proved
the relative compactness in LP of the set {Maf; IE K} for sufficiently
small a> 0.

Definition 1. Let X andY be B-spaces, and letS be the unit sphere of
X. An operator TEL (X, Y) is 5aid to be compact or completely continuous
if the image T Sis relatively compact in Y.
Example 1. Let K (x, y) be a real- or complex-valued continuous
function defined for- oo <a~ x, y ~ b < oo. Then the integral operator K defined by
b
J K(x, y) l(y) dy
(KI) (x) =
(1)
is compact as an operator E L (C [a, b], C [a, b]).

Proof. Clearly K maps C [a, b] into C[a, b]. Set sup IK(x, y) I= M.
x,y
Then II K I II < (b- a) M II I II so that K S is equi-bounded. By

Schwarz' inequality, we have
I(K I) (xi) -
(K I) (xz) 12 ~
J IK (xi, y)- K (x2, y) 12 dy . aJ /I (y) 12 dy'

a
and hence K Sis equi-continuous, that is,

lim
~tO
sup
lx,-x,l ;;,;;~
I(K I)
(x1 ) - (K I) (x2) I = 0 uniformly in IE S.
Therefore, by the Ascoli-Arzela theorem (Chapter III, 3), the set K Sis
relatively compact inC [a, b].
Example 2. Let K(x, y) be a real- or complex-valued \B-measurable
function on a measure space (5, \B, m) such that
JJ
/K(x, y) 2 m(dx) m(dy) < oo.

ss
Then the integral operator K defined by the kernel K (x, y):
(2)
K(x, y) l(y) m(dy), IE P(S) = P(S, \B, m),

(3)
s
is compact as an operator EL(P(S),P(S)). The kernel K(x,y) satisfying (2) is said to be of the Hilbert-Schmidt type.
Proof. Take any sequence {In} from the unit sphere of P(SJ. We have
to show that the sequence {K In} is relatively compact in L2 (5). Since a
Hilbert space L2 (S) is locally sequentially weakly compact, we may assume
that {In} converges weakly to an element IE L2 (S); otherwise, we choose a
suitable subsequence. By (2) and the Fubini-Tonelli theorem,
(KI) (x)
278
JK(x, y) 12 m(dy) < oo for m-a.e. x. Hence, for such an x,

s
lim (KI,.) (x) = lim
K(x, y) l,(y) m(dy) = lim (!,(.), K(x,. )) =
n--->00 s
n--->00
(!(.), K(x,. )) =
K(x, y) l(y) m(dy). On 'the other hand, we have, by
s
Schwarz' inequality,
1>-->(lO
I (KI,)
(x) 12 <
IK (x, y) \2 m (dy) lln(Y) \2 m (dy) :S I K (x, y) \2 m(dy)

s
s
s
for m-a.e. x.
(4)
I (KI,) (x) 12 m(dx} =

s
I (KI) (x) 12 m(dx). This result, if combined with w-lim K I,= K I,
Hence, by the Lebesgue-Fatou theorem, lim
n--+00
n-->00
implies s-lim K In= K I by Theorem 8 in Chapter V, 1. But, as proved

n--->OO
above in (4), we have
I (Kh) (x) \2 m(dx)

s
and so
<
JJ IK(x, y)
ss
12
m(dy) m(dx)
J jh(y)
12
m(dy),
(5)
Hence, from w-lim

n--->00
1,. =I,
we obtain w-lim K In = K I, because, for

n--->00
any gE P(S), lim (K I,., g)= lim (/,., K*g) = (!, K*g) = (K I. g).
n--->00
n--->00
Theorem. (i) A linear combination of compact operators is compact.

(ii) The product of a compact operator with a bounded linear operator
is compact; thus the set of compact operators E L (X, X) constitutes a
closed two-sided ideal of the algebra L (X, X) of operators. (iii) Let a
sequence {T,} of compact operators E L (X, Y) converge to an operator
Tin the sense of the uniform operator topology, i.e., lim II T- T, II = 0.
n--->00
Then T is also compact.

Proof. (i) and (ii) are clear from the definition of compact operators
The closedness, in the sense of the uniform operator topology, of the ideal
of compact operators in the algebra L (X, X) is implied by (iii).
We shall prove (iii). Let {xh} be a sequence from the closed unit sphere
S of X. By the compact property of each T,., we can choose, by the diagonal method, a subsequence {xh'} such that s-lim T,.xh' exists for every
h--->00
fixed n. We have
1\Txh'- Txk'\\
and so
<
<
\\Txh'- T,xh' I\+ 1\T,.xh'- T,xk' II+ 1\T,.xk'- Txkll
IIT-T,.II + IIT,xh,-T,.xk'll + IIT,.-TI\.
lim IITXh-Txk'JI<2\\T-T,II
h,li-->-oo
Cauchy sequence in the B-space Y.
Hence {Txk'} is a
279
Nuclear Operator. As an application of the Theorem, we shall consider

the nuclear operator introduced by A. GROTHENDIECK [2].
Definition 2. Let X, Y be B-spaces and TEL (X, Y). If there exist a
sequence{/~} s;; X', a sequence {y,.} s;; Yanda sequence {c,.} of numbers
such that
sup Ill~ II< CXJ, sup liY.. II< CXJ, .I !c.. I< CXJ and
n
n
n
(6}
m
T x =s-lim .I c,. <x, /~) y,. in Y for every x EX,
m-->00
n=l
then Tis called a nuclear operator on X into Y.

Remark. The existence of the s-lim in (6} is clear, since
II/~ ci <x, 11> Yili <
1#,.ic1 lllx 111111 IIIIY1 II< constant 1#,.lci I llxll

The nuclear condition says that the s-lim is equal to T x for every
xEX.
Proposition. A nuclear operator Tis compact.
Proof. Define the operata: T,. by
T,.x
..I c1 <x, /1) y1 .

= J=l
(7)
Since the range R(T,.) is of finite dimension, T,. is compact as may be

proved by the Bolzano-Weierstrass theorem. Moreover, by (6} and
Ici I II x II ,
.
II T x - T,. x II = 11.J=n+l ci <x, 11> Yi II < constant J=n+l
we have lim II T- T,. II
.......00
0 and so T must be compact .
An Example of the Nuclear Operator. Let G be a bounded open domain of R", and consider the Hilbert space H~(G). Suppose (k -1) > n.
Then the mapping T
(8)
H~(G) 3q;-+ q; E Hb(G)
is a nuclear operator E L(H~(G), Hb(G)).

Proof. We may assume that the bounded domain G is contained in the
interior of the parallelogram P:
0 < x1 < 2n (i = 1, 2, ... , n).
We recall that H~(G) is the completion of H~(G) = C~(G) with respect

to the norm llq;llk=(.I JID 8 cp(x)i 2 dx) 112 (see Chapterl,lO). We
. isi~kG
extend the functions E H~(G) to be periodic with period 2n in each
variable x 8 by defining the function values as 0 in P- G. The functions
/p(x) = (2n)-"' 2 exp(i/1 x), wherep = (Pv Pz, .. , p,.)

is ann-tuple of integers and
p x =
" p.x.,
.I
S=l
(9)
280
form a complete orthonormal system of L 2 (P) = H8(P). Thus, denoting

by n the 'distributional derivative, we have, for Is I< k, the Fourier
expansion in L 2(P) of functions D 5 tp (x) where tp E if~ (G):
ncp (x) =
~ (D 5 tp, /p)o /p, where (1p, /p)o

{3
f 1p (x) /p (x) dx.
(10)
We have, by
(D 5tp, /p)o = (-1)11 (tp,
and the Parseval relation
n /p)o = m=l
II (if3m)Sm
(tp, /p)o
J ID tp(x) 12 dx < lltp II~ (Is I< k),
~~ (D 5 tp, /p)ol 2 =
{3
the inequality
i(tp, (1 +
lf31 2)k/ 2 /p)o 12 < constant ~
I(D tp, /p)o 12 <

5
ll~k
constant lltp II~.
Therefore the functionalf'p E H~ (G)' defined by

(tp,
<
satisfies sup 11/'pll
f'p>
+ lfJ jZ)k/2 /p)o
= (tp, (1
oo. Moreover,
{3
Yp = (1
[f312pl2 /p
" (if3 1) 51
satisfies sup IIYPIIj < oo by D 5/p =II
l=l
{3
I < oo,
~I cp
{3
where
because, for positive integers {35 ,
L
{3
((Jl
<
+ (12 + ... + (J,.)k-; =
L(
{3
...
fJ1 fJ2
~..
)(k-j)Jn
fln
cp
= (1
t:L
(
p,
+ lf31 2)U-kl/2,
1
((Jl
/p- We also have
+ (12 + ... + (J,.)"
(]_)(k-j)Jn.
fJ1
L
p,
)(k-j)Jn
(]_)(k-j)Jn
fJ2
,-. ( 1 )(k-j)Jn
b (k - j)
1
<oo y~~>.
fJn fln
n
~-
Therefore, we have proved the (Fourier) expansion

tp
= ~
Cp <tp, t'p >Yp
{3
Remark. If there exists a complete orthonormal system {tpj} of eigenfunctions of a given bounded linear operator K E L(L2(S), L2(S)) such
that Kepi= Aj tpj (j = 1, 2, ... ), then, from the Fourier expansion
we obtain
281
3. The Rellich-Garding Theorem
= (Kr:pi, 'Pi), and so, if the eigenvalues Ai are all > 0 and
00
...l: Aj < oo, then the operator K is nuclear. The condition ...l: [(Kr:pi, 'Pi) [
We have Ai
00
;~1
;~1
< oo is surely satisfied, if the operator K is defined by a kernel

{
K 1 (x, y)
an~ K 2 (x, y) are of the Hilbert-Schmidt type.
For,
j~
JK
K (x, y)
(K: K1r:pj. f{Jj) r =
2 (z,
j~
x) K 1 (z, y) m (dz), where the kernels
c~
(Klr:pj. K2r:pj) r <
rr
Klr:pj 11 2 j~ [[K2!f1j rr 2
and, by Parseval's relation, we have

00
00
i~ [[K1 r:pi [[ 2 = i~
j I) K
(z, y) r:pi(Y) m(dy)
2 m(dz)
j i~ I/ K
j fj [K (z,y) [ m(dy)J m(dz) <oo,
00
(z, y) r:pi(Y) m(dy)
m(dz)
00
and similary for ...l: [[ K 2 r:pi [[2. A bounded linear operator K in a sepa;~1
00
rable Hilbert space X is said to be of the trace class if ...l: I(Kr:pi, 'Pi) I< oo
;~1
for arbitrary complete orthonormal systems {r:pi} and {'Pi} of X. For a

general account concerning the trace class and the nuclear operator, see
R. ScHATTEN [1], and I. M. GELFAND-N.Y. VILENKIN [3].
3. The Rellich-Garding Theorem

Theorem (GARDING [1]). Let G be a bounded open domain of Rn. If an
operator TE L(H~(G), H~(G)) satisfies, for j < k,
[[ Tr:p [[,. < C [[r:p IIi for all r:p E H~ (G), where C is a constant,
(1)
then Tis compact as an operator E L(H~(G), H~(G)).

Proof. By the definition of the space H~(G) (see Chapter I, 10), it
would be sufficient to show the following: Let a sequence {r:pv} ~ H~(G) =
C~(G) be such that II'Pvll" < 1 (v = 1, 2, ... ). Then the sequence {Tr:p.}
contains a subsequence strongly convergent in H~ (G). The Fourier
transform cfv(~) = (2n)-nf2 r:p,,(x) exp(-ix~)dx satisfies, by Schwarz'
inequality,
[cr.(~) 12 < (2n)-n
Jdx J lr:p.(x) 2 dx <

1
(2n)-n
Jdx,
and hence {cr~(~)} is equi-bounded in ~ERn and in v. We may assume,

by the boundedness of llr:p. 11 0 , that a subsequence {r:p.,} is weakly conver-
282
gent in L 2 (G) = H8(G). Since, for each ~. the function exp(-ix~)

belongs to L 2 (G), we know that the sequence of bounded functions
tPv' (~) = (cp.,, (2n)-nf2 exp (-ixm0 converges at every ~- Thus, by (1)
and the Parseval relation for the Fourier transform (Chapter VI, 2),
IITcp.,- Tcpl'' II~=
IIT(cp., -cpl'') II~< C2llffJv' -cpl'' 117

./~
= C I.IIDS(cp., -cpl'') II~= c I.IICDS(cp., -cp/''))11~

2
lsi~J
lsi;;;;J
.I.IIn (i~~)St(t/J.,-1/JI',) (~)1112
= c2 Is"[:::;,
t=1
<c21.I.
f /il~:~ct/J.,(~)-q;~',(~))l2d~
sl ~J 1~1 ~
+ c2 c1 f
lEI>
1=1
I~ l2j I1/J., (~l- ~", (~l 12d~,
where c1 is a positive constant.

The first term on the right converges, for fixed r, to 0 asv' and,u' -HxJ.
This we see by the Lebesgue-Fatou Lemma. The second term on the right
is, for r > 1.
< c2 c1 r2j-2k
< c2 cl r2j-2k
!ol>r
Rn
I~ 12k
I~ 12k
lq;., (~l - q;I'' (~l 12 d~
Irrv (~l -
~/!", (~l
12 d~
/~
<
c2c1 c2 r2j-2k !sl;;o;k
2: II (Dsm'Dsm ') 112
TV
Tl'
0
= c2cl c2 r2j- 2k ,s!fk II ns (cp.,- CfJJ<'l II~

= C2C1C2r2i-2k
with a constant c2.
[[cp.,-cp~''[[~
The last term converges, by

lim IITcp,.--Tcp"'lfk=O.
f<
< 4C2C1C2r2i-2k
k, to 0 as r---? oo. Therefore,
v,!-'~
4. Schauder's Theorem
Theorem (ScHAUDER). An operator T E L (X, Y) is compact iff its
dual operator T' is compact.
Proof. Let S, S' be the closed unit sphere in X, Y', respectively. Let
T E L (X, Y) be compact. Let {yj} be an arbitrary sequence in 5'. The
fuuctions Fi (y) = (y, yj) are equi-continuous in the sense that
IFi(y)-Fi(z)[ = l<y-z,yj>l < lly-zll

Moreover, {Fj(y)} is equi-bounded in jon any bounded set of y, since
IFi (y) I< II y II Therefore, by the Ascoli-Arzela theorem, as applied to the
283
5. The Riesz-Schauder Theory
functions {F1 (y)} defined on the compact set (T St, we see that some
subsequence {.z<:; (y)} converges uniformly in y E (T St. Hence
<Tx, yj) = <x, T'yj) converges uniformly in xES, and so {T'. yj}
converges in the strong topology of X'. This proves that T' is compact.
Conversely, letT' be compact. Then, by what we have proved above,
T" is compact. Hence, if S" is the closed unit sphere in X", (T" S") is
relatively compact. We know that Y is isometrically embedded in Y" (see
Theorem 2 in Chapter IV, 8). Hence T S ~ T" S" and soT Sis relatively compact in the strong topology of Y" and so in the strong topology
of Y. Therefore, Tis compact.

We prepare
Lemma (F. RIESZ [2]). Let Vbe a compact operatorE L(X, X), where
X is a B-space. Then, for any complex number Ao# 0, the rangeR (Aoi- V)
is strongly closed.
Proof. We may assume that .Ao = 1. Let {x,.} be a sequence of X such
that y,. =(I-V) x,. converges strongly toy. If {x,.} is bounded, then by,
the compactness of the operator V, there exists a subsequence {x,.} such
that {V x,.} converges strongly. Since x,.. = y,.. + V x,.., {x,.} converges to
some x, and soy= (I-V) x.
We next assume that {l!x,. II} is unbounded. Set T = (I-V) and put
6\:11 = dis(x,., N(T)), where N(T) = {x; Tx = 0}. Take a w,. E N(T) such
that 6\:11 < llx,.- w,. II< (1 + n-1) 6\:11 Then T(x,.- w,.) = Tx,., and so
in the case when {6\:,.} is bounded, we can prove, as above, that y E R (T) =
R(I- V). Suppose that lim 6\:11 = oo. Since z,. = (x,.- w,.)/llx,.- w,.ll
n->00
satisfies liz,. II= 1 and s-lim Tz,. = 0, we can prove, as above, that
n->00
there exists a subsequence {z,..} such that s-lim z,., = w0 , s-lim Tz,.. = 0.
n->00
Hence w0 E N(T). But, if we put z,.- w0 = u,., then, in
n->00
x,.-w,.-w0 llx,.-w,.ll = u,. !lx,.-w,.ll,

the second and the third terms on the left belong to N (T) so that we must
have II u,. II II x,.- w,. II > 6\:11 This is a contradiction, since s-lim u,. = 0,
IJx,.- w,. II< (1
+ n-1) 6\:,. and
n->00
lim
n->00
6\:11
= oo.
We are now able to prove the Riesz-Schauder theory; for convenience

sake, we shall state the theory in a series of three theorems.
Theorem 1. Let V be a compact operator E L (X, X). If Ao # 0 is not
an eigenvalue of V, then A.o is in the resolvent set of V.
Proof. By the preceding Lemma and the hypothesis, the operator
TAo = (J.0 I - V) gives a one-one map of X onto the set R (TA,) which is
strongly closed in X. Hence, by the Corollary of the open mapping theorem in Chapter II, 5, TAo has a continuous inverse. We have to show that
284
= X. If not, the topological image X 1 = T .~,x of X is a proper

closed subspace of X. Hence, if we set x2 = T.t,X1, Xs = TA,x2, .. . ,
then X,.+l is a proper closed subspace of X,. (n = 0, 1, 2, ... ; X 0 = X).
By F. Riesz' theorem in Chapter III, 2, there exists a sequence {y,.} such
that Yn E X,., II Yn II = 1 and dis (y,., x,.+l) > 1/2. Thus, if n > m,
R (T.~,)
/.Q 1 (Vym- Vy,.) = Ym

with some y E Xm+l
+ {-y,.- (TA,Ym- T .t,Yn)/Ao} =
I!VY,.- Vymll ~ 1).0 1/2,
Hence
operator V.
Ym- Y
contrary to the compactness of the
Theorem 2. Let V be a compact operator E L (X, X). Then, (i) its

spectrum consists of an at most countable set of points of the complex plane
which has no point of accumulation except possibly ). = 0; (ii) every
non-zero number in the spectrum of V is an eigenvalue of V of finite
multiplicity; (iii) a non-zero number is an eigenvalue of V iff it is an
eigenvalue of V'.
Proof. By Theorem 1, a non-zero number in the spectrum of Vis an
eigenvalue of V. The same is also true of V', since, by Schauder's theorem, V' is compact when Vis. But the resolvent sets are the same for V
and V' (see Chapter VIII, 6). Hence (iii) is proved. Since the eigenvectors
belonging respectively to different eigenvalues of Vare linearly independent, the proofs of (i) and (ii) are completed if we derive a contradiction
from the following situation:
{ Then~ exists a sequence {x,.} of linearly inde:endent vectors
such that V x,. = ).,.x (n = 1, 2, ... ) and lim ).,. =).=I= 0 .

....-+00
To derive a contradiction, we consider the closed subspace X,. spanned by

xv x 2, . .. , x,.. By F. Riesz' theorem in Chapter III, 2, there exists a
sequence {y,.} such that y,. E X,., II y,. II = 1 and dis (y,., X,._ 1) > 1/2
(n = 2, 3, ... ). If n > m, then
;.;1 Vy,.-
For,
if
;.;;/ V Ym = Yn
y,.
+ (-Ym- ;.;1 TAnYn + ;.;;; 1 T AmYm) =
where zE X,._ 1 .
Yn- z,
n
= ..I {Jj Xj, then we have y,.- ).; 1 Vy,. = ..I {Jj xiJ=1
j~ {JjA; 1A;xjE
J=1
Xn-1
and similarly TAmYmE Xm. Therefore
11).; 1Vy,.-
II > 1/2. This contradicts the compactness of V combined with

the hypothesis lim ).,. -=1= 0.
;.;;; 1 V Ym
....-+00
Ao -=1= 0 be an eigenvalue of a compact operator V E

L (X, X). Then Ao is also an eigenvalue of V' by the preceding theorem.
We can prove: (i) the multiplicities for the eigenvalue Ao are the same for V
Theorem 3. Let
285
and V'. (ii) The equation (A. 0 I - V) x = y admits a solution x iff y E
N (A.ol'- V').L, that is, iff V' I= A.0 1 implies <y, I>= 0. (iii) The equation (A.oi'- V' f) = g admits a solution I iff g EN (Aoi- V).L, that is,
iff V x = A.0 x implies <x, g) = 0.
Proof. Since the eigenvalue A.o #- 0 is an isolated singularity of the

resolvent R(A.; V) =(AI- V)-1, we can expand R(A.; V) in Laurent
series
R(A.; V) =
00
.I (A.-A.ot An.
n=-00
We are particularly interested in the residueA_ 1 = (2ni)-1
R (A.; V) dA..
1.<-;..\=
As was proved in Chapter VIII, 8, A_ 1 is an idempotent, i.e., A:_ 1 = A_ 1

If we set (AI- V)-1 = A.-1 I + V;.,"then from (AI- V) (A.-1 I + V;.) = I
we obtain V;. = V(A.-1 V;. + ).-2 I), and so V;. is compact when V is.
Hence, by
A_ 1
(2ni)-1
+ (2ni)-1
R(A.; V) d).= (2ni)-1
V;. dA. = (2ni)-1
l.l-A.I =
i.<-;..1 =
i.<-;..1 =
i.<-;..1 =
).-ld).. I
V;. d)..
Thus, by the Theorem in Chapter X, 2, A_ 1 is a compact operator.

Therefore, by A_1 X = A_ 1 (A_ 1 X) and the compactness of A_v the
unit sphere of the normed linear space A_ 1 X is relatively compact. Hence,
by F. Riesz' theorem in Chapter III, 2, the range R (A_ 1) is of finite
dimension. On the other hand, V x = A.0 x, x #- 0, implies that
(AI-V)-lx=(A.-A.o)-1 x by (AI-V)x=(A.-A.o)x, and so A_ 1 x=
(2ni)-1
(A.- A.o)-1 dA. x = x. Therefore, the eigenvalue equa-
1.<-A.I=
tion V x = A.ox is equivalent to V x = A.0 x, x E R (A_ 1 ). In the same way,

we can prove that the eigenvalue equation V' I= ).0 1 is equivalent to
V' I= A.0 l, IE R (A~ 1 ). But R (A_ 1 ) and R (A~ 1 ) are of the same dimension. For, A~ 1 1 = g satisfies A~ 1 g = A~ 1 (A~ 1 1) = g, and this is equivalent to <x, g)= <A_ 1 x, g) for all x EX and so the functional g may be
considered as a functional defined on the finite dimensional spaceR (A_ 1).
Now, by the well-known theorem in matrix theory, the eigenvalue
equation V x = A.ox (in R (A_ 1)) and its transposed equation V' I= A.ol
(in R (A~ 1 )) both have the same number of linearly independent solutions. We have thus proved (i). The propositions (ii) and (iii) are already
proved by the Lemma and the closed range theorem (Chapter VII, 5).
Extension ofthe Riesz-Schauder Theory. Let a power vn of VEL (X, X)

be compact for some positive integer n. Then, by the spectral mapping
theorem in Chapter VIII, 7, a(Vn) = a(Vt and, by the compactness of
vn, a(Vn) is either a finite set or else a countable set accumulating only
at 0. Therefore, a(V) is either a finite set or a countable set accumulating
286
only at 0. vn being compact,
(2 ni)-1
R (A; Vn) dJ.
JA-A.J~s
is, for any J.o 7'=- 0 of a(Vn) and sufficiently smalle > 0, of finite-dimensional range. Hence J.0 is a pole of R(J.; Vn) (see Chapter VIII, 8). But
(An I - Vn) = (A- V) W- 1 I + ;.n- 2 V + , + vn- 1 ) and hence
WI- Vn)-1 w-1 I
+ ... + vn-1) =
(AI- V)-1'
which proves that any J.o 7'=- 0 of a ( vn) is a pole of R (A; V) and so is an
eigenvalue of V. These facts enable us to extend the Riesz-Schauder theory
to operators V for which some power vn is sompact. This extension is
highly important in view of its application to concrete problems of
integral equations, such as the Dirichlet problem pertaining to potentials.
See, e.g., 0. D. KELLOGG [1]. It can be proved that the Riesz-Schauder
theory for l 0 = 1 is valid also for an operator VEL (X, X) if there exist a
positive integer m and a compact operator K E L (X, X) such that
IlK- vm II< 1. See K. YoSIDA [9]. It is to be noted here that, if
K 1 (s, t) and K 2 (s, t) are bounded measurable for 0 < s, t < 1, then the
integral operator T defined by
x (s) ~ (T x) (s) = (K1 K 2 x) (s), where (Kjx) (s) =
J Ki (s, t) x (t) dt,
is compact as an operator L(U(O, 1), U(O, 1)). See K. YosmA-Y. MI[10].
MURA-S. KAKUTANI
6. Dirichlet's Problem
Let G be a bounded open domain of Rn, and
L
JsJ,JtJ~m
Dscs1 (x) D 1
a strongly elliptic differential operator with real C00 (Ga) coefficients

C51 (x) = c15 (x). We shall dealonlywithreal-valued functions. Let /E L 2 (G)
and u1 E Hm (G) be given. Consider a distribution solution u 0 E L2 (G) of
Lu
=f
such that (u 0 - u 1) E H 0 (G).
(1}
The latter condition (u 0 - u 1) E H 0 (G) means that each of the distributional derivatives
(2}
is the L2(G)-limit of a sequence {Dicph,j}, where cph,jE CQ"(G) (see Chapter I, 10 ). Thus it gives roughly the boundary conditions:
Di u 0 = Di 'U1 on the boundary 8G of G for
If I <
m.
(3)
In such a sense, (1) will be called a Dirichlet's problem for the operator
287
6. Dirichlet's Problem
L. We shall follow the treatment of the problem as formulated and solved

by L. GARDING [1].
We first solve
(4)
u + cxLu = f, (u-u 1 ) E H0 (G),
where the positive constant ex is so chosen that Garding's inequality
(5)
(rp + cxL *rp, rp) 0 > <5 II'P //! holds whenever rp E ego (G).
Here L*
2: (-1)11+itl D 1 c.1 (x)
IJ,I~m
and b is a positive con-
stant. The existence of such an ex is guaranteed if the coefficients C51 (x)

are continuous on the closure Ga of G. We also have, by m-times partial
differentiation, the inequality
/(rp + cxL*rp, VJ)o/ < y II'PIIm IIVJIIm whenever rp, VJE ego(G),
where y is another positive constant independent of rp and Vi
We have, for u 1 E Hm(G) and rpE ego (G),
(6)
(L*rp, u1 ) 0 =I ((-1)1l+ltl D 1c51 D 5rp, u1 ) 0 =I (-1)11 (c. 1D 5rp, D 1ul)o,

s,t
s,t
by partial differentiation. By Schwarz' inequality, we obtain, remembering that the coefficients c51 are bounded on Ga,
/(L*rp,u1)0 /<'17
lsJ,Itl:;;;m
IID 5 rp// 0 //D 1u 1 // 0
(sup/c.1 (x)/='YJ}
s,t;x
The right hand side is smaller than constant times II'P lim
Thus the linear functional
F (rp) = (rp + cxL *rp, u 1 ) 0 , rp E ego (G),
can be extended to a bounded linear functional defined on H'Q(G) which
is the completion of ego (G) with respect to the norm II'P lim Similarly, we
see, from
/(rp,/)o/ < II'PIIo )/1)/o < JI'PIIm !IIIIo.
that the linear functional (rp, /) 0 of rp E ego (G) can be extended to a
bounded linear functional of rp E H0 (G). Hence, by F. Riesz' representation theorem as applied to the Hilbert space H 0 (G), there exists an
f' = f' (f, u 1 ) E H 0 (G) such that
(rp, /) 0 - (rp
+ cxL*rp, u1 ) 0 =
(rp, f')m whenever rp E ego(G).
Hence, by the Milgram-Lax theorem in Chapter III, 7, applied to the

Hilbert space H 0(G), we have
(rp, /)o- (rp + cxL*rp, u1 )o = (rp, f')m = B(rp, Sf'), Sf' E H 0 (G),
where
B (rp, VJ) = (rp + cxL *rp, VJ)o for rp E ego (G), VJ E H0 (G).
Thus
(rp, /) 0 = (rp + cxL *rp, u 1 + S /') 0 whenever rp E ego (G),
and so u0
u1 +Sf' is the desired solution E L2(G) of (4).
(7)
(8)
288
We shall next discuss the original equation (1). If u0 E L2(G) satisiies

(1); then u2 = u0 - u 1 E H 0 (G) satisfies
(u 0 , L *f}?)o
(uv L *f}?)o
(u2, L *fJ?)o = (I, f}?)o, 9? E COO (G).
We obtain, by partial integration as above,

l(uvL*fJ?)oJ<a constant times ll9?llm
I (1, 9?)o I < Ill llo 119? llo < Ill llo 119? lim
We may thus apply F. Riesz' representation theorem in H 0 (G) to the
linear functional (1, 9?)o- (uv L *f}?)o of 9? E COO (G). Hence there exists a
uniquely determined v E H0 (G) such that
(1, fJ?)o- (uv L *fJ?)o = (v, fJ?)m
whenever 9? E COO (G).
By the Milgram-Lax theorem, applied to (v,fJ?)m, we obtain an 5 1 vEH0 (G)

such that
(v,fJ?)m=B(S 1 v,fJ?)
whenever q?ECOO(G), vEH0 (G).
Thus the Dirichlet problem (1) is equivalent to the problem: For a given
5 1 v E H 0 (G), find a solution u 2 E H 0 (G) of
(u 2 , L *9?)o = B (51 v, 9?), 9? E COO (G).
(1')
Now, for a given u E L 2 (G)= H8(G),

I(u, 9?)o I< llu llo 119? llo < llu llo 119? lim
so that, by F. Riesz' representation theorem in the Hilbert space H 0 (G),
there exists a uniquely determined u' = T u E H 0 (G) such that, whenever
q?E COO(G),
(u,fJ?)o = (u',fJ?)m and llu'llm < llullo
Hence, by the Milgram-Lax theorem, we obtain
(u, f}?)o = (u',fJ?)m =B(S1 u',fJ?) =B(S1 Tu,q?), IIS1 Tu1Jm < b-1 /lullo (9)
Therefore, by (1'), we have, whenever 9? E COO (G),
B (u 2 , 9?)
(u 2 , 9?
+ IXL *9?)o =
(u2 , 9?)o
+ IX (u2 , L *q?) 0
=B(S1 Tu 2 ,q?) +1XB(S1 v,fJ?),

that is,
B(u2 -S 1 Tu 2 -1XS1 v,q?)
Because of the positivity B(fJ?, 9?)
u2-
>
0.
0 of B, we must have
S 1 T u2=a S lv.
(1")
The right hand term aS1 vEH0(G)isaknownfunction. By IIS1 Tullm <

b- 1 llullo, we see that the operator51 T defined on H 0(G) into H 0(G) is
compact (the Rellich-Garding theorem in Chapter X, 3). Therefore, we
Appendix to Chapter X. The Nuclear Space of A.
GROTHENDIECK
289
may apply the Riesz-Schauder theory to the effect that one of the
following alternatives holds:
Either the homogeneous equation u - 5 1 T u = 0 has a non-trivial

solution u E H 0 (G), or the inhomogeneous equation u- S 1 Tu = w has,
for every given wE H 0 (G), a uniquely determined solution u E H'Q(G).
The first alternative corresponds to the case (u, cp + rxL *cp) 0 = (u, cp) 0 ,
that is, to the case Lu = 0. Hence, returning to the original equation (1),
we have
Theorem. One of the following alternatives holds: Either i) the homogeneous equation Lu = 0 has a non-trivial solution u E H 0 (G), or ii) for
any f E L2 (G) and any u 1 E Hm (G), there exists a uniquely determined
solution u 0 E L 2 {G) of Lu = f, u- u 1 E H 0 (G).
Appendix to Chapter X. The Nuclear Space of

A. Grothendieck
The nuclear operator defined in Chapter X, 2 may be extended to
locally convex spaces as follows.
Proposition 1. Let X be a locally convex linear topological space, and
Y a B-space. Suppose that there exist an equi-continuous sequence
{/j} of continuous linear functionals on X, a bounded sequence {Yi} of

n
elements E Yanda sequence ofnon-negativenumbers{c1}with ..I c1 < oo.

J=l
Then
yJ
.I c <x, ()
T x =s-lim
J
n-+oej=l J
(1)
defines a continuous linear operator on X into Y.

Proof. By ~he equi-continuity of {fj}, there exists a continuous seminorm p on X such that sup I(x, /j) I n,
j
11/~.
Cj
(x,
tj>
Yill <
p(x)
~~r IIYill ~~I cj.
This proves that the right hand side of (1) exists and defines a continuous
linear operator Ton X into the B-space Y.
Definition 1. An operator T of the form (1) is said to be a nuclear
operator on X into the B-space Y.
Corollary. A nuclear operator Tis a compact operator in the sense
that it maps a neighbourhood of 0 of X into a relatively compact set
of Y.
290 Appendix to Chapter X. The Nuclear Space of A.
GROTHENDIECK
Proof. We define
n
Tn x =,I cj (x, fj) Yi

;~1
Tn is compact, since the image by Tn of the set V = {x; p (x)
1} of
X is reiatively compact in Y. On the other hand, we have
//Tx-Tnx/J=IJ.i cj(x,fj>y)J<p(x)sup//Yi//.i cj,

I
;~n+l
j-;;;;,1
J~n+l
and so Tnx converges to Tx strongly and uniformly on V. Hence the

operator T is compact.
As was proved in Chapter X, 2, we have a typical example of a nuclear
operator:
Example. Let K be a compact subset of R. Then, for (k- j) > n, the
identity mapping T of H~ (K) into Hb (K) is a nuclear operator.
Proposition 2. Let X be a locally convex linear topological space, and
V a convex balanced neighbourhood of 0 of X. Let Pv (x) =
inf A.
xj.<EV,.<>O
be the Minkowski functional of V. Pv is a continuous semi-norm on X.

Set
Nv={xEX;pv(x)=O}={xEX;A.xEV forall A.>O}.
Then Nv is a closed linear subspace of X, and the quotient space Xv =
X/Nv is a normed linear space by the norm
I/ x
//v
= Pv (x), where xis the residue class mod N v

containing the element x.
(2)
Proof. Let (x- x 1 ) EN v Then Pv (x 1 ) < Pv (x) + Pv (xl- x) = Pv (x),

and similarly Pv(x) < Pv(x1 ). Thus Pv(x) = Pv(x1) if x and x1 are in the
same residue class modNv. We have J/x/lv > 0 and /10/lv = 0. If
//x 1/v = 0, then x Ex implies x E Nv and so x = 0. The triangle inequality is proved by /lx + Yl/v = Pv(x + y) < Pv(x) + Pv(Y) =
/lx//v + Jly /lv We have also JJ,xxiJv = Pv(<Xx) =/<X /Pv(x) =/<X /1/x/lv
Corollary. By the equivalence
(Pv, < Pv,)
++
(V2 s;; V1),
(3)
we can define the canonical mapping
Xv,-+ Xv, (when V 2 s;; V1 )

by associating the residue class xv, (mod N v,) containing x to the residue
class xv, (mod N v,) containing x. The mapping thus obtained is continuous,
since
1/xv,/lv, = Pv,(x) > Pv,(x) = 1/xv,/lv,

We are now ready to give the notion of a nuclear space, introduced in
analysis by A. GROTHENDIECK [2].
GROTHENDIECK
291
Definition 2. A locally convex linear topological space X is said to

be a nuclear space, if, for any convex balanced neighbourhood V of 0,
there exists another convex balanced neighbourhood U ~ V of 0 such
that the canonical mapping
(4)
is nuclear. Here Xv is the completion of the normed linear space Xv.

Example 1. Let RA be the topological product of real number field
R in such a way that RA is the totality of real-valued finite functions x (a)
defined on A and topologized by the system of semi-norms
(5)
Pa(x) = lx(a) I, a EA.
Then RA is a nuclear space.
Proof. Nv is the totality of functions x(a) ERA such that, for some
finite set {ai E A; f = 1, 2, ... , n}, x(aj) = 0 (f = 1, 2, ... , n). Hence
Xv = RAJNv is equivalent to the space of functions xv(a) such that
xv(a) = 0 for a# aj (f = 1, 2, ... , n) and normed by
llxv(a) llv = sup lx(aj)j.

l:;;;j:;;;n
We take N u the totality of functions x (a) ERA such that x (a,.) = 0 for
.x E A' where A' is any finite set of integers containing 1, 2, ... , n.
Thus, for U ~ V, the canonical mapping Xu= RA/Nu~ RA/Nv = Xv
is nuclear. For the mapping is a continuous linear mapping with a finitedimensional range.
Example 2. A nuclear B-space X must be of finite dimension.
Proof. Since X = X v for any convex balanced neighbourhood V of 0
of a B-space, the compactness of the identity mapping X~ X implies
that X is of finite dimension by F. Riesz' theorem in Chapter III, 2.
Example 3. Let K be a compact subset of Rn. Then the space '!JK(Rn)
introduced in Chapter I, 1 is a nuclear space.
Proof. As in Chapter I, 1, let
PK,k(f) =
sup
xEK,Isl:;;;k
ID
/(x) /
be one of the semi-norms which define the topology of '!I K (Rn). Let
Vk = {/E '!IK(Rn); h,k(f) < 1}. Then Nvk is {0}, and Xvk=XfNvt.=
'!IK (Rn)/Nvk is precisely the space '!IK (Rn) normed by PK,k If (k-f) > n,
then it is easy to prove, as in the example following the Corollary of
Definition 1 above, that the canonical mapping of Xvk intoXv1 is a nuclear
transformation. Hence '!IK (Rn) is a nuclear space.
Theorem 1. A locally convex linear topological space X is nuclear,
iff, for any convex balanced neighbourhood V of 0, the canonical mapping
X~ Xv is nuclear.
19*
292
GROTHENDIECK
Proof. Necessity. Let U ~ V be a convex balanced neighbourhood

of 0 of X such that the canonical mapping Xu---+ Xv is nuclear. The
canonical mapping T : X ---+ X v is the product of the canonical mapping
X---+ Xu and the canonical nuclear transformation Xu---+ X v Hence T
must be a nuclear transformation.
Sufficiency. Let the canonical mapping T: X---+ Xv be given by a
nuclear transformation
00
Tx =.I c; <x, fj) Y;

J=l
For any X> 0, the set {x EX; I<x, fj> I <X for j = 1, 2, ... } is a convex
balanced neighbourhood U"' of 0 of X, because of the equi-continuity of
{fj} ~X'. Moreover,
1/ T x 1/v
=I If c; <x, fj) Y; llv
<X
sup 1/Y; 1/v

J
c; whenever x E U"'.
be so small that the right hand side is< 1. Then 1/ T xl/v < 1 and
fj may be considered as belonging to the dual space X~"'
and so
Tx = Tz = ~ c; <x, fj) Y; whenever (x-z)ENu.
0<
Let
U"' ~ V. Each
Thus the canonical mapping Xu"'---+ X vis given by a nuclear transformation

xu"'---+~ c; <xu"', fi> Y;
J
Theorem 2. Let a locally convex linear topological space X be nuclear.

Then, for any convex balanced neighbourhood V of 0 of X, there exists a
convex balanced neighbourhood W ~ V of 0 of X such that Xw is a Hilbert
space.
Proof. The nuclear canonical mapping Xu---+ Xv (U ~ V) defined by
Txu
~ c; (xu,
J
fi> yj
is factored as the product of the two mappings

X: Xu---+ (l 2) and {3: (l2)---+ X V where
is given by
xu---+ {c}'2 <xu, fj)} and {3 is given by {.;;}---+ ~ cJ'2 .;;Y;.

J
The continuity of X is clear from
~ lc] 12 <xu, f1> 12 <(sup

J
j
and that of {3 is proved by
l/f1 1/ //xu l/u) 2 ZJ
c;,
2 ::;; sup 1/ Y; //~ 1/ {M /If. ~ c;

c]12 .;;Y;II~ < ~ c;/1 Y; II~ ~ I.;; 1
II ~
'
J
J
j
J
Let U2 be the inverse image in (l2 ) by {3 oftheunitsphereofXv. Then U2

is a neighbourhood of 0 of (Z2) and so contains a sphere 5 of centre 0 of (l2).
GROTHENDIECK
293
Let W be the inverse image in X of S by the continuous mapping iX

defined as the product of continuous canonical mapping X-+ Xu and the
continuous mapping iX: Xu-+ (l 2 ). Then clearly W ~ V and, for any
xwEXw,
llxwllw = in A=_ inf }. = IIIXxllz (the radius of S).
xf.AEW,.l>O
IXx/.lES,.A>O
Since II liz is the norm in the Hilbert space (l 2 ),Xw isapre-Hilbertspace.

Corollary. Let X be a locally convex nuclear space. Then, for any
convex balanced neighbourhood V of 0 of X, there exist convex balanced
neighbourhoods W1 and W 2 of 0 of X with the properties:
w2 ~ wl ~ V, Xwl and Xw, are Hilbert spaces and the canonical

mappings X-+ Xw,, Xw,-+ Xw Xw -+ Xv are all nuclear.
1,
Therefore, a nuclear space X has a fundamental system {V"'} of neighbourhoods of 0 such that the spaces Xv are Hilbert spaces.
"' It can be proved that:
Further Properties of Nuclear Spaces.
1. A linear subspace and a factor space of a nuclear space are also
nuclear.
2. The topological vector product of a family of nuclear spaces and
the inductive limit of a sequence of nuclear spaces are also nuclear.
3. The strong dual of the inductive limit of a sequence of nuclear
spaces, each of which is an F-space, is also nuclear.
For the proof, see the book by GROTHENDIECK [1] referred to above,
p. 47. As a consequence of 2., the space ~ (Rn), which is the inductive
limit of the sequence {~K,.(Rn); r = 1, 2, ... } (here K, is the sphere
lxl < r of Rn), is nuclear. Hence, by 3., the space ~(Rn)' is also nuclear.
The spaces ~ (Rn), ~ (Rn)', (Rn) and (Rn)' are also nuclear spaces.
The importance of the notion of the nuclear space has recently been
stressed by R. A. MINLOS [1]. He has proved the following generalization
of Kolmogorov' s extension theorem of measures:
Let X be a nuclear space whose topology is defined through a countable
system of convex balanced neighbourhoods of 0. Let X' be the strong dual
space of X. A cylinder set of X' is defined as a set of the form
Z'
{/'EX'; ai
<
<xi,/')
<
bi
(i
1, 2, ... , n)}.
Suppose there is given a set function flo defined and > 0 for all cylinder
sets. Let f1o be a-additive for those cylinder sets Z' with fixed x1 , x 2 , , xnThen, under a compatibility condition and a continuity condition, there
exists a uniquely determined extension of flo which is a-additive and
> 0 for all sets of the smallest a-additive family of sets of X' containing all
the cylinder sets of X'.
For a detailed proof and applications of this result, see I. M. GELFANDN.Y. VILENKIN [3].
294
XL Normed Rings and Spectral Representation

A linear space A over a scalar field (F) is said to be an algebra or
a ring over (F), if to each pair of elements x, yEA a unique product
xy E A is defined with the properties:
(xy) z
x(y
= x (y z)
+ z)
cx{3(xy)
= xy
(associativity),
+ xz
(distributivity),
= (cxx) ({3y).
(1)
If there exists a unit element e such that ex= xe = x for every x E A,

then A is said to be an algebra with a unit. A unit e of A, if it exists, is
uniquely determined. For, if e' be another unit of A, then we must have
ee' = e = e'. If the multiplication is commutative, i.e., xy = yx for
every pa1r x, y E A, then A is called a commutative algebra. Let A be an
algebra with a unit e. If, for an x E A, there exists an x' E A such that
xx' = x' x = e, then x' is called an inverse of x. An inverse x' of x, if it
exists, is uniquely determined. For, if x" be another inverse of x, then
we must have
x" (xx')
= x" e = x" = (x" x) x' = ex' = x'.
Thus we shall denote by x-1 the inverse of x if x has an inverse.

An algebra is called a Banach algebra, or in short a B-algebra if it is a
B-space and satisfies
(2)
IJxyJJ < llxiiiiYII
The inequality
IJx,y,- xy II< Jlxn(Yn- y) II+
<
II (x"- x) Y II
Jlxnllll(y,-y)JJ + JJ(x,-x)JJIIYII
shows that xy is a continuous function of both variables together.

Example 1. Let X be a B-space. Then L (X, X) is a B-algebra with a
unit by the operator sum T + S and operator product T S; the identity
operator I is the unit of this algebra L (X, X), and the operator norm
JJTJI is the norm of the element T of this algebra L(X, X).
Example 2. Let S be a compact topological space. Then C (S) is a
B-algebra by (x 1 + x2) (s) = x1 (s) + x 2 (s), (cxx) (s) = cxx (s), (x1 x2) (s) =
x1 (s) x2 (s) and JJxJJ =sup Jx(s) J.
sES
Example 3. Let B be the totality of continuous functions x (s),

0 < s < 1, which are representable as absolutely convergent Fourier
series:
X
(s) =
00
2: Cne 2"ins with
n=-oo
00
2: lc,l <
n=-00
cx:>.
(3}
1. Maximal Ideals
of a Normed Ring
295
Then it is easy to see that B is a commutative B-algebra with a unit by

the ordinary function sum and function multiplication, normed by
llxll =
00
.I len I
n=-oo
(4)
In the last two examples, the unit is given by the function e (s) = 1
and II e II = 1. In the following sections, we shall be concerned with the
commutative B-algebra with the unit e such that
lie II= 1.
(5)
Such an algebra is called a normed ring.

A Historical Sketch. The notion of Banach algebras was introduced
in analysis by M. NAGUMO [1]. He proved that Cauchy's complex function theory can be extended to functions with values in such an algebra,
and applied it to the investigation of the resolvent of a bounded linear
operator around an isolated singular point. The result is an abstract
treatment of those given in our Chapter VIII, 8. K. YosmA [11] proved
that a connected group embedded in a B-algebra is a Lie group iff the
group is locally compact. This result is an extension of a result due to
] . VON NEUMANN [6] concerning matrix groups. Cf. E. HILLE-R. S. PHILLIPS [1], in which the result of K. YosmA [11] is reproduced.
The ideal theory of normed rings was initiated by I. M. GELFAND [2].
He has shown that such a ring can be represented as the ring of continuous
functions defined on the space of maximal ideals of the ring. By virtue of
this representation, we can give an integration free treatment of the
spectral resolution of bounded normal operators in a Hilbert space; see
K. Y osmA [12]. This result will be exposed in the following sections. The
Gelfand representation may also be applied to a new proof of the Tauberian
theorem of N. WIENER [2]. We shall expose this application in the last
section of this chapter. Forfurtherdetails aboutB-algebras, see N.A.NAIMARK [1], C. E. RICKART [1] and I. M. GELFAND-D. A. RAIKOVG. E. Srwv [5].
1. Maximal Ideals of a Normed Ring

We shall be concerned with a commutative B-algebra B with a unite
such that lie II= 1.
Definition 1. A subset J of B is called an ideal of B if x, y E J implies
that (ax + (Jy) E J and zx E J for every z E B. B itself and {0} are ideals
of B. Ideals other than B and {0} are called non-trivial ideals. A nontrivial ideal J is said to be a maximal ideal if there exists no non-trivial
ideal containing J as a proper subset.
Proposition 1. Each non-trivial ideal ] 0 of B is contained in a maximal
ideal].
296
Proof. Let UoJ be the set of all non-trivial ideals containing ] 0 We

order the ideals of UoJ by inclusion relation; that is, we denote ] 1
2
if ] 1 is a subset of ] 2 . Suppose that{],.} is a linearly ordered subset of
U0 ] and put ]p = U ],.. We shall show that ]pis an upper bound
-< ]
J,.E{l,.}
of {],.}. For, if x, y E ]p, then there exist ideals],., and],., such that
x E ] ,., andy E ] .... Since{],.} is linearly ordered, ] ,.,
(or],.,
and sox andy both belong to],.,; consequently (x- y) E ],., ~ ]p and
zx E fa~,~ ]p for any z E B. This proves that ]pis an ideal. Since the unit
element e is not contained in any]"'' e is not contained in ]p = U ] ...
-< ] ..,
>- ] ,..)
J,.E{J,.}
Thus ]pis a non-trivial ideal containing every ] 01 Therefore, by Zorn's

Lemma, there exists at least one maximal ideal which contains ] 0
Corollary. An element x of B has the inverse .x--1 E B such that
x-1 x = x:x--1 = e iff x is contained in no maximal ideal.
Proof. If .x--1 E B exists, then any ideal ] 3 x must contain e = xx-1
so that ] must coincide with B itself. Let conversely, x be contained in
no maximal ideal. Then the ideal xB = {xb; bE B} =F {0} must coincide
with B itself, since, otherwise, there exists at least one maximal ideal
containing xB 3 x = xe. It follows that xB = B, and so there must exist
an element bE B such that xb =e. By the commutativity of B, we have
xb = bx = e, that is, b = :x--1.
Proposition 2. A maximal ideal ] is a closed linear subspace of B.
Proof. By the continuity of the algebraic operations (addition, multiplication and scalar multiplication) in B, the strong closure
is also an
ideal containing]. Suppose
=F ]. Then = B, because of the maximality of the ideal J. Thus e E
and so there exists an X E J such that
lle-x/1 < 1. x has the inverse .r1 E B which is given by Neumann's
series
e + (e _ x) + (e _ x)2 + ...
r.
For, by II (e-xt II < II e-x lin, the series converges to an element E B

which is the inverse of x, as may be seen by multiplying the series by
x = e- (e-x). Hence e = .x--1 x E ] and so] cannot be a maximal ideal.
Proposition 3. For any ideal] of B, we write
x
y (mod]) or x "'y (mod]) or in short x "'y, if (x- y) E]. (1)
Then x "'y is an equivalence relation, that is, we have
x "' x (reflexivity) ,
x"' y implies y "'x (symmetry),
x "' y and y "'z implies x"' z (transitivity).
We denote by x the set {y; (y - x) E ]} ; it is called the class (mod J)
containing x. Then the classes (x + y), !XX and (xy) are determined independently of the choice of elements x and from the classes and y,
respectively.
1.
297
Maximal Ideals of a Normed Ring
Proof We have to show that x '"" x', y'"" y' implies that (x + y) '""
(x' + y'), !XX '"" iXX and xy '""x' y'. These are clear from the condition
that f is an ideal. For instance, we have xy- x' y' = (x- x') y +
x' (y - y') E f by (x - x') E f and (y - y') E f.
Corollary. The set of classes (mod f) thus constitutes an algebra by
1
(2)
x+y=x+y, iXX=iXx,xy=xy.
Definition 2. The above obtained algebra is called the residue class

algebra of B (mod f) and is denoted by B ff. Thus the mapping x--+
of the algebra B onto B = B!J is a homomorphism, that is, relation (2)
holds.
Proposition 4. Let f be a maximal ideal of B. Then B = B/J is a
field, that is, each non-zero element E B has an inverse :X-1 E B such
that :X-1 = x-1 =
Proof. Suppose the inverse .X-1 does not exist. Then the set xB =
{xb; bE B} is an ideal of B. It is non-trivial since it does not contain e,
but does contain =F 0. The inverse image of an ideal by the homomorphism is an ideal. Therefore, B contains a non-trivial ideal containing f
as a proper subset, contrary to the maximality of the ideal f.
We are now able to prove
Theorem. Let B be a normed ring over the field of complex numbers,
and fa maximal ideal of B. Then the residue class algebra B = B/J is
isomorphic to the complex number field, in the sense that each E B is
represented uniquely as x = fe, where~ is a complex number.
Proof. We shall prove that B = B/J is a normed ring by the norm
x x
e.
llxll =in! llx II
(3)
xEx
B/J is a normed field and so, by the GelfandMazur theorem in Chapter V, 3, B = B!f is isomorphic to the complex
number field.
Now we have lltXxll = ltXIIIxll, and llx +)Ill= ~nf_/lx + Yll ~
If this is proved, then
xEx,yEy
inj IIy II = llx II + II"Y II ; llx y II < IIi lillY II is proved simi+ yEy
If llx II = 0, then there exists a sequence {xn} ~ x such that
inJ II x II
xEx
larly.
s-lim Xn = 0. Hence, for any x Ex, (x- xn) E f and so s-lim (x- xn) = x
n->00
,_co
= ], that is, X= 0. Hence II :X II = 0 is equiwhich proves that X E

valent to x = 0. We have lie II < II e II = 1. If lie 1/ < 1, then there exists
an element x E f such that II e- x II < 1. As in the proof of Proposition 2,
the inverse x-1 exists, which is contradictory to the Corollary of Proposition 1. Thus we must have /le II = 1. Finally, since B is a B-space and f
is a closed linear subspace by Proposition 2, the factor space B = B/J
is complete with respect to the norm (3) (see Chapter I, 11). We have
thus proved the Theorem.
298
Corollary. We shall denote by x (f) the number~ in the representation
x =~e. Thus, for each x E B, we obtain a complex-valued function x(J)
defined on the set{!} of all the maximal ideals of B. Then we have
(x
+ y) (]) =
x(])
+ y(]),
(xy) (]) = x(]) y(J),
(txx) (]) = txx(J),
and
e(J)
(4)
== 1.
We have, moreover,
(5)
sup lx(J) I< llx II'
JE{l}
and
sup lx(J) I= 0 implies
mn
X=
0 iff
n J=
fi{l}
{0}.
(6)
Proof. The mapping X--+ X = X(]) eof the algebra B onto the residue
class algebra B = B/J is a homomorphism, that is, relation (2) holds.
Hence we have (4). Inequality (5) is proved by
I~ I = I~ I lie II = llx II = inf II x II < II x II

xEx
Property (6) is clear, since X(]) = 0 identically on {!} iff X E

Definition 3. The representation
n ].
JE{J}
x--+ x(])
(7)
of the normed ring B, by the ring of functions x (]) defined on the set {]}
of all the maximal ideals J of B, is called the Gel/and-representation of B.
2. The Radical. The Semi-simplicity

Definition 1. Let B be a normed ring over the complex number field,
and {]} the totality of the maximal ideals J of B. Then the ideal
n J is called the radical of the ring B. B is said to be semi-simple if its
JE{J}
radical R =
n J reduces to the zero ideal {0}.
JE{l}
Theorem 1. For any x E B, lim II xn 11 1/n exists and we have

n->00
lim llxnll 11n=suplx(J)I

n->00
JE{J}
(1)
Proof.Settx =sup lx(J) I Then, by llx" II> lxn(]) I= I x(JW, we have

llxnll > txn, and
]E{J}
so
lim llxnll 11n > tx. We have thus to prove

n->00
lim llxn 111/n < tx.

Let IP I> tx. Then, for any] E {!}, x(])- P=I= 0, i.e., (x- {3e) E ].
Hence the inverse (/'1 e - x)-1 exists. Setting p-1 = A., we see that the
inverse (/'1 e- x)-1 =A. (e- A.x)-1 exists whenever lA. I < tX-1 . Moreover,
as in Theorem 1 in Chapter VIII, 2, we see that A. (e- A.x)-1 is, for
n->00
299
lA I<
<X-1 ,
holomorphic in A. Hence we have the Taylor expansion
A(e-Ax)- 1 = A(e
That
+ A.x1 + J.2 x2 + + J."x,. + ).
x,. = x" may be seen by Neumann's series
(e- Ax)- 1 = I
00
n=O
J,."
x"
which is valid for II Ax II < 1. By the convergence of the above Taylor

series, we see that
<lAI-n for large n when lA I< <X- 1 , and so

when IAI- 1 ><X, that is, lim llx"W'" <<X.
Thus llx"ll = IAI-"IIJ."x"ll

lim
n-+00
llx"W'" <
IAI- 1
n-+00
n J
Corollary. The radical R =
JE{J}
of B coincides with the totality
of the generalized nilpotent elements x E B which are defined by
Iimllx"W'"=O.
n-+00
(2)
Definition 2. A complex number J,. is said to belong to the spectrum

of x E B, if the inverse (x- Ae)-1 does not exist in B.
If A. belongs to the spectrum of x, then there exists a maximal ideal
] such that (x- A. e) E]. Conversely, if (x- A. e) belongs to a maximal
ideal ], then the inverse (x- A.e)-1 does not exist. Hence we obtain
Theorem 2. The spectrum of x E B coincides with the totality of
on the space {!} consisting of all the
values taken by the function x
B.
of
J
ideals
maximal
<n
Application of Tychonov's theorem. We define, for any ] 0 E {!},

a fundamental system of neighbourhoods of ] 0 by
(3)
{JE {]}; lxdn -x,(J0 ) I< e, (i = 1, 2, ... , n)},
where e1 > 0, nand x1 E Bare arbitrary. Then{!} becomes a topological

space and each x(J), xE B, becomes a continuous function on{!}. We
have only to verify that if ] 0 = ]p then there exist a neighbourhood
V0 of ] 0 and a neighbourhood V1 of ] 1 with empty intersection. This may
be done as follows. Let x0 E ] 0 and x0 E]1 so thatx0 {]0) = 0 and x0 (] 1 ) =
<X=0. Then V0 ={JE{]};Ix0 (J)I<I~XI/2} and V1 ={!E{J};
x0 (] 1) I < I~X \/2} have an empty intersection.
Ix0
<n -
Theorem 3. The space{]}, topologized as above, is a compact space.

Proof. We attach, to each x E B, the compact set
Kx = {z; izi < llxll}

of the complex z-plane. Then the topological product
K
5 =II
xEB x
300
is a compact space, by virtue of Tychonov's theorem. See Chapter 0. To

any maximal ideal ] 0 E {]},we assign the point
Jlx(]0 ) =s(]0)E5.
xEB
{]} is in one-one correspondence with a subset 5 1 of 5 by the above
correspondence ] 0 --? s (]0 ). Moreover, the topology of{]} is the same as
the relative topology of 5 1 as a subset of 5. Hence, if we can show that 5 1
is a closed subset of the compact space 5, then its topological image{]}
is compact.
To prove that 5 1 is a closed set, we consider an accumulation point
ro = II AxE 5 of the set 5 1 in 5. We shall show that the mapping
xEB
x--? Ax is a homomorphism of the algebra B into the complex number field
(K). Then ] 0 = {x; Ax = 0} becomes, as may be seen from the isomorphism of B/]0 with (K), a maximal ideal of Band (x-A.xe) E ] 0 , that
is, x(]0 ) =Ax This proves that the point ro = II Ax= II x(]0 ) belongs
xEB
xEB
to 5 1 .
We thus have to show that
Ax+:r = Ax + A.,, A.= = ~Ax, A.,., = A.,.A.,, A.. = 1.
We shall, for instance, prove that Ax+:r =A.,.+ A.,... Since ro = 1I Ax is
xEB
an accumulation point of 5v there exists, for any e > 0, a maximal ideal
] such that
lie,.- x(]) I< e, lie,- y(]) I< e, IAx+y- (x + y) (])I< e.
By (x + y) (]) = x(]) + y(]) and the arbitrariness of e > 0, we easily
see that Ax+:Y = Ax + ic,.. is true.
We can now state the fundamental facts about the Gelfand representation x--? x (]) of the normed ring B in the form of
Theorem 4. A normed ring B over the complex number field is represented homomorphically by the ring of functions x (]) on the compact
space {]} of all the maximal ideals J of B. The radical R of B consists of
those and only those elements which are represented by functions identically equal zero on {]}. The representation x--? x (]) is isomorphic iff
the ring B is semi-simple.
Application of the Stone-Weierstrass Theorem (of Chapter 0). The
above obtained ring of functions is dense in the space of all complexvalued continuous functions on {]} with uniform convergence topology
if the ring B is symmetric (or involutive) in the following sense:
For any x E B, there exists an x* E B such that x* (]) = x (]) on {]}. (4)
Examples of Gelfand Representations
]0
Example 1. Let B = C (5) where 5 is a compact topological space, and

a maximal ideal of C(5). Then there exists a point s0 E 5 such that
301
x(s0 ) = 0 for all x E ] 0 . Otherwise, for any s" E 5, there exists an x" E ] 0
such that x" (s") =I= 0. x" (s) being a continuous function, there exists a
neighbourhood V" of s"' such that x"(s) =I= 0 in V". Since 5 is compact,
n
U V".J = 5.
there exists a finite system, say, V", V"', ... , V" n such that j=l
2
Hence the function
n __
E] 0
2: x".(s) x".(s)
x(s) = i=l
~
1
does not vanish over 5, and the inverse x-1, x~ 1 (s) = x (s)~ 1 , of x E ] 0
exists, contrary to the maximality of the ideal ] 0 . Thus we see that ] 0 is
contained in the maximal ideal]'= {xE B; x(s0 ) = 0}. By the maximality of ] 0 , we must have ] 0 = ]'. In this way we see that the space
{]} of the maximal ideals ] of B is in one-one correspondence with the
points s of 5.
Example 2. Let B be the totality of functions x (s), 0 < s
can be represented by absolutely convergent Fourier series:
x(s)
00
= 2:
n=-00
00
Cne2msn,
2:
n=-00
< 1, which
len I < 00.
y(s), (xy) (s) = x(s) y(s)

B is a normed ring by (x + y) (s) = x(s)
and II x II = ~ Icj I . Let ] 0 be a maximal ideal of B. Set e2"is = x1 . Then
. J
x! 1 = e~ 2"'5 , and so, by lx1 Uo) I < llxtll = 1, Jx1 1 {])I = Jxl (])~ 1 1 <
llx1 1 ll = 1, we see that jx1 {]0) I = 1. Hence there exists a point s0 ,
0 < s0 < 1, such that x1 {]0 ) = e2"is,_ Thus Xn = e2"isn = x~ satisfies
00
2: cne2"is,n=x(s 0 ). In this way, we

n=-oo
see that for any maximal ideal ] 0 of B, there exists a point s0 , 0 < s0 < 1,
xnUo) = e2"is,n, and so x(J0 ) =
such that the homomorphism x--+ x (]0 ) is given by x (]0 ) = x (s0), for all
x E B. It is also clear that the rnapping x --+ x (s0) gives a homomorphism
of the algebra B into the complex number field. Therefore we see that
the maximal ideal space of B coincides with {e2"is; 0 ::;;: s < 1}.
Corollary (N. Wiener's theorem). If an absolutely convergent Fourier
series x(s) =
1/x(s)
00
2: cne 2"isn does not vanish on [0, 1], then the function
n=-oo
is also representable as an absolutely convergent Fourier series.

For, x does not belong to any of the maximal ideals of the normed ring
of the above Example 2.
Example 3. We take B 1 = C [0, 1], and define, for x, y E B 1 ,
(x
+ y) (s) =
s
x (s)
+ y (s),
(<Xx) (s)
J x(s- t) y(t) dt and
IJxjj =
<XX (s), (xy) (s)

sup
sE[O,l]
jx(s) J.
302
B 1 is a commutative B-algebra without unit. Adjoining formally a unit e

by the rule ex= xe = x, !!eli= 1, the set B = {z = A.e + x; xE B 1}
becomes a normed ring by the operations
(A.1 e + x1 ) + {A. 2 e + x2) = {A.1 + A. 2) e + (x1 + x2), cx(A.e + x)
1=cde + cxx,
(A.1 e + x1) (A. 2 e + x2) = A.1 A.2e + A.1 x2 + A. 2 x1 + x1 x2
and IIA.e + xll = lA. I+ llxll

We have, by induction,
2
n--I
lx2(s)I<M2s, lx3(s)I<M3s2' ... ,1xn(s)I<M"(ns 1)!'

where M =sup !x(s) I= llxll Thus every xE B 1 is a generalized
sES
nilpotent element of B, due to the fact that lim (n!) 1fn = oo.
n=oo
3. The Spectral Resolution of Bounded Normal Operators

Let X be a Hilbert space, and let a system M of bounded normal
operators E L (X, X) satisfy the conditions:
T, S EM implies T S = S T (commutativity),
(1)
T E M implies T* E M.
(2)
A systemM consisting of a bounded normal operator TEL (X, X) and its

adjoint T* surely satisfies (1) and {2).
Let M' be the totality of operators E L (X, X) which commute with
every TE M, and letB=M" = (M')' be the totality of operatorsEL (X,X)
commutative with every operatorS EM'.
Proposition 1. Every element of B is a normal operator. B is a normed
ring over the complex number field by the operator sum, the operator
product, the unit I (the identity operator) and the operator norm liT II
Proof. M ~ M' by (1), and so M' ~ M". Hence M'" = (M")' ~ M"
and soB= M" is a commutative ring. The identity operator I belongs
to B and is the unit of this algebra B. By (2), we easily see that every
operator E B is normal. Since the multiplication T S and the adjoint
formation T--+ T* in the algebra B are continuous with respect to the
norm of the operator, it is easy to see that the ring B is complete with
respect to the operator norm.
Theorem 1. By the Gelfand representation
B 3 T--+ T(J),
(3)
the ring B is represented isomorphically by the algebra C ({]}) of all

continuous functions T (]) on the compact space {]} of all the maximal
ideals
303
J of B in such a way that
liT II= sup IT(]) I,
(4)
JE{f}
T (]) is real-valued on {]} iff T is self-adjoint,
(5)
T(]) > 0 on{]} iff Tis self-adjoint and positive,

that is, (Tx, x)
> 0 for all xE X.
(6)
Proof. We first show that, for any bounded normal operator T,
(7)
111'211 = IITII 2
By the normality ofT, we see that H = TT* = T*T is self-adjoint.
Hence, by Theorem 3 in Chapter VII, 3,
IITII 2 = &dp (Tx, Tx) =sup I(T*Tx,x) I= IIH II= II T*T 1/ =!ITT* IJ.
llxll~l
li;riJ~l
Since (T*) 2 = (T2)*, T 2 is normal with T. Thus, as above, we obtain

IIT2IJ 2 = IIT* 2 T211, which is, by the commutativity TT* = T*T, equal
to II(T*T) 2 11 = IIH2 11. Since H 2 is self-adjoint, we obtain, again by
Theorem 3 in Chapter VII, 3,
JIHII 2 =sup (Hx, Hx) =sup I(H2 x, x) I= IIH2 11.
llxll~l
IJ;rll~l
= (IITII 2) 2 , that is, 111'211 = IITII 2

11 =
=
Therefore,
because we know already that
We have, by (7), II T II= lim II T"
n--+00
the right hand limit exists (see (3) in Chapter VIII, 2). Hence, by Theorem 4 of the preceding section, the representation (3) is isomorphic and
(4) is true.
Proof of (5). Let a self-adjoint T E B satisfy, for a certain ] 0 E {]},
T(]0) =a +ibwithbo:F 0. Then the self-adjoint operatorS= (T -ai)jbEB
satisfies (I+ 5 2 ) (]0 ) = 1 + i 2 = 0, and so (I+ 5 2) does not have an
inverse in B. But, by Theorem 2 in Chapter VII, 3, (I+ 5 2) has an
inverse which surely belongs to B. Thus, if T E B is self-adjoint,
T(]) must be real-valued. Let TEBbe not self-adjoint, and put
111'211 2
IJHII 2
IIH2
T=
+ T*
2
W'n,
. T- T*
+~~-
Then, since the first term on the right is self-adjoint, the self-adjoint
operator (T- T*)/2i must be =1= 0. Thus, by the isomorphism of
J0 E{]} such that T 2iT* (]0) =I= 0.

T-T*
T+T*
(]0) + i -2-i- {]0) is not real. For, as proved
=
2
representation (3), there must exist a

Hence T(]0 )
above, self-adjoint operators are represented by real-valued functions.

Proof of (6). We first show that
T*(]) = T(]) on{]}.
(8)
304
This is clear, since self-adjoint operators (T + T*)/2 and (T- T*)j2i

are represented by real-valued functions. Therefore, by (4) and the result
of the preceding section, the ring B is represented by the ring of all continuous complex-valued functions on {!} satisfying (5) and (8). Let
T (]) ~ 0 over {J}. Then S (]) = T (]) 1' 2 is a continuous function on
{!}. Hence, by the isomorphism of the representation (3), 5 2 = T. By
(5), we have S = S*. Hence (Tx, x) = (S 2 x, x) = (Sx, Sx) > 0. To
prove, conversely, that the condition (T x, x) > 0 for all x E X implies
T(J) > 0 over{]}, we set T 1 (]) =max(T(J), 0) and T2 (J) = T 1 (J)-T (]).
Then, by what we have proved above, T 1 and T 2 are both E B, self-adjoint and positive: (Tjx, x) > 0 for all x EX (i = 1, 2). Moreover, we
have T 1 T 2 = 0 and T 2 = T 1 - T. The former equality is implied by
T 1 (]) T 2 (]) = 0.
Therefore, we have
0
< (TT2 x, T 2 x) =
(-T~x,
T 2 x) = - (T~x, x) = - (T2 T 2 x, T 2 x)
< 0.
Thus (T~x, x) = 0 and so, by Theorem 3 in Chapter VII, 3, we must

have T~ = 0. Hence, by II T2 ll = lim II~ lll/n, we obtain T 2 = 0. We
n---+00
have thus proved T = T 1 and hence T(J) > 0 on {J}.

We shall thus write T > 0 if Tis self-adjoint and positive. We also
write S > T if (S- T) > 0.
Theorem 2. Let {Tn}
that
B be a sequence of self-adjoint operators such
(9)
Then, for any xE X, s-lim Tnx = Tx exists, i.e., s-lim Tn = Texists and
n-..oo
n---+00
T E B, S > T > Tn (n
1, 2, ...).
Proof. We first remark that, by (6),
E, FEB and E > 0, F > 0 imply E +F:::: 0 and EF > 0.
(10)
Thus 0 < Ti < T~ < < T~ <<52 Hence, for any xE X, a

finite lim (T~x, x) exists. Since, by (6), T~+k > Tn+kTn > T~, we also
n--HlO
have
lim (T~+kx, x) =
n,koo
lim (Tn+kTnx, x)
n,k-+oo
Therefore, lim ((Tn-T m) 2 x, x)

s-lim Tnx
n---+00
n,m--+00
lim
n,m---+oo
Tx exists. That T E B and S
lim (T~x, x).
n-+oo
II Tnx- T mX 11 2 =
0 so that
> T > Tn is clear from the
process of the proof.

Theorem 3. Let a sequence of real-valued functions {Tn (])}, where
T n E B, satisfy the condition
0 < T 1 (]) ~ T 2 (])
< <
T n (])
< <a
finite constant on {J}.
(11)
305
Then, by (6) and Theorem 2, s-lim Tn = T exists. In such a case, we can

n--+oo
prove that
D = {! E {!}; T(J) =I= lim Tn(])}

n--+00
is a set of the first category, and so Dc = {!}-Dis dense in{!}.

Proof. By Theorem 2, T > Tn and so T(J) > lim Tn(]) on {]}.
n--+00
By Baire's theorem in Chapter 0, 2, the set of points of discontinuity of

the function lim Tn (])is of the first category. Hence, if the set Dis not
n--+00
of the first category, then there exists at least one point ] 0 ED at which
lim Tn (]) is continuous. In other words, there exists a positive number()
n--+00
and an open set V (]0 ) 3 ] 0 of {]} such that

T(J) > ()
lim Tn(]) whenever J E V(J0 ).
n--+00
Since the compact space {!} is normal, and since T (]) > lim T n (]) on
n--+00
{]},we may construct, by Urysohn's theorem an open set V1 (]0) 3 ] 0

and a function W(]) E C ({!}) such that 0 < W (f)<() on {!},
V1 Uot ~ v (]0), w (]) = CJ/2 on V1 (]0 ) and w (]) = o on v Uolc.
Hence T(J)- W(]) > lim Tn(]) on{]}, and so, by (6), T- W > T,.
n--+00
(n = 1, 2, ... ). W(]) =!= 0 implies, by the isomorphism (3), that W =F 0,

W > 0. Thus, again by (6), T- W >s-lim Tn, contrary to T =s-lim T,..
n--+00
n--+00
Finally, since{!} is a compact space, the complement Dc = {!}- D

of the set D of the first category must be dense in{!}.
We are now able to prove (K. YosmA [12])
The spectral resolution or the spectral representation of operators E B.
Consider the set C' ({!}) of all complex-valued bounded functions

T' (])on{!} such that T' (])is different from a continuous function T(])
only on a set of the first category. We identify two functions from C' ({!})
if they differ only on a set of the first category. Then C' ({!}) is divided
into classes. Since the complement of a set of the first category is dense
in the compact space {!}, each class T' contains exactly one continuous
function T (]) which corresponds, by the isomorphism B-++ C ({]}), to
an element T E B.
For any T E B and for any complex number z =A + if-t, we put
Ez =the element E B which corresponds to the class E~ containing the
defining function E' (]) of the set {! E {!}; ReT(]) <A, Im T (f) < f-t }.
It is clear that there exists a monotone increasing sequence of continuous
functions In of complex argument such that E~(]) = lim In (T(J)) and
n--+00
20
306
so
E~(])
E C' ({]}).We have then

n
IT(]) - ..I (A;

J=2
+ ip,j) (E~;+iiJ (]) + E~;...+ip;_, (]) - E~;+ip;_, (]))I
A1 = - ()(. -
V2
p,1 = - {J - -
8
-
V2
E~i->+ip; (])
< e on {]} if
< A2 < < An = ()(. = sup IRe T (]) I,

JE{J}
< 1'2 < <
(s~p (Ai- At-1)2
f.tn = {J =sup IImT(])

JE{l}
1,
+ s? (p,i- I'J-l)T'2 < e.
Thus, by the definition of E;(]), we have

n
IT(])- ..I (A;+ ip,1) (EJ..;+ip;(]) + EJ..1_,+ip.;_,(J)

}=2
-EJ..1_,+ip;(]) -EJ..;+ip;_,(J))I < e

on{]}, since the complement of a set of the first category is dense in the
compact space{]}. Therefore, by (4), we have
n
II T- J=2
..I (Aj + ip,1) (EJ..;+ip; + EJ..;_,+ip;_,- E;.1_,+ip 1 -
which we shall write as
T =
E;.;+ip;_,) II < e,
JJzdE
(12)
6 ,
and it is called the spectral resolution of the normal operator T.
4. The Spectral Resolution of a Unitary Operator

If Tis a unitary operator E B, then, by
T(]) T*(J)
T(J) T(])
1,
(1)
we see that the values taken by the function T (]) on {]} are complex
numbers of absolute value 1. From this fact, we can simplify the spectral
resolution
z dE6 ofT.
The defining function E~ (]) of the set {] E {]}; arg(T (])) E (0, 0)},
0 < 0 < 2n, belongs to C' ({]}), and we have, by setting E~(J) = 0,
JJ
E~"(]) =I,
IT (]) -
..It eio; (E~; (]) - E~;_, (])) I < m~x J eio; - eo;_, I
J=
(0 = 00 < 01 < < On = 2n).

Let E 0 (]) be the continuous function on {]} which is different from
E~ (]) only on a set of the first category, and let E 0 be the operator E B
which corresponds toE 6(]) by the isomorphic representationB:;JT-T(J).
Then, as in the preceding section,
II T -
.I ee; (E 6; - E 0;_,) II < max

Ieie; - ee;_, I
j
j=l
307
4. The Spectral Resolution of a Unitary Operator
This we write, in view of the fact e2"i = 1,

2n
T =
J e'
0 dF
(0), where
F(O)
= E0+ 0 -E+o
for 0 < 0 < 2n, F(O)
0, F(2n) =I.
(2)
Here EO+o is defined by E0+ 0 x =s-lim E 0 x, the existence of this limit

OtO
will be proved below.

Theorem 1. The system of operators F(O), 0 < 0 < 2n, satisfies the
conditions:
each F (0) is a projection operator, commutative with
every bounded linear operator commutative with T,
(3)
F(O) F(O') = F(min(O, 0')),
(4)
F(O) = 0, F(2n) =I,
(5)
F(O + 0) = F(O), 0 < 0 < 2n, in the sense that

s-limF(O')x=F(O)x for every xEX.
(6)
O'tO
Proof. It is sufficient to prove that E 0 , 0 < 0 < 2n, satisfies the

conditions:
each E 0 is a projection operator E B,
(3')
Eo Eo- = Emin(O,O'),
E 0 = 0, E 2"' = I,
E 0 x =s-lim Ea-x for every x EX and 0 < 0 < 2n.
(4')
O'tO
(5')
(6')
We have E~ (]) = E 0(]) and E~ (]) 2 = E~ (]). Hence, by the result of

the preceding section, we obtain E 0 = E: and E~ = E 0 This proves (3').
(4') is proved similarly from E~ (]) E~, (]) = E'm.in(O,O') (]), and (5') is
proved similarly. Next let 0,. {- 0. Then E~n (]) > E~n+ (f) > E~ (]),and
E 0n = E exists and
so, by the result in the preceding section, s-lim
~
E (]) = E~ (]) = ~~To E~n (f) on {J} except possibly on a set of the first
category. Thus E = E 0
Example 1. Let a linear operator T defined by
Ty(s) = ei 0 y(s), where y(s) E L 2 (-oo, ooJ.

Tis unitary. We define, when 2nn
F(O) y(s) = y(s)
for
< s<
2 (n
s< 0
2n
J ei
20*
2nn < 2(n
0+2nn<s.
F(O)y(s)=O for
It is easy to see that T =
+ 1) n,
dF(O).
l) n,
308
Example 2. Let a linear operator T 1 be defined by
T 1 x(t)
+ 1)
x(t
in L2(-oo, oo).
T 1 is unitary. By the Fourier transformation
y(s)
Ux(t)
l.i.m. (2n)- 1i 2
11---+00
we obtain Ux(t
T 1 x(t) = x(t
that is, T 1
1)
e'5 Ux(t)
J e-'' x(t) dt,

-n
1
isy(s). Thus
+ 1) = u-1 e' y(s)

5
u- 1 Ty(s) = u- 1 TU x(t),
U-1 TU. Therefore, we have

2n
T1 =
u- 1 F (0)
dF1 (0), where F 1 (0) =
e; 8
U.
The uniqueness of the spectral resolution. Since T-1

y-1(]) = T*(]) = T(J)-1, we easily see that
y-1 =
Let m~x
J
\e'8i- e'8i-'\ <

T
we obtain, by (4),
= T*
and
2n
J e-o dF (0).
(7)
0
8.
Then, from
~ ei81(F(Oi) -F(Oi-1))
J
+ <5,
\\<5\\
< 8,
+ <5', where
IWII < l\(T-b) b\l + l\b(T-b)\1 + 11<52 11
<(/\Til+ 8) 8 + 8(\\TII + 8) + 82.
y2
= ~ e2 i 0i(F (Oj) - F (Oj_ 1))

J
Hence we have T 2 =
yn
2n
Je
2; 8 dF(O},
and, more generally,
0
2n
J ein dF(O)
8
(n
0, 1,
2, ... ) .
(8)
Therefore, if there exists another spectral resolution T
211
J e' dF
8
1 (0)
satisfying (3) to (6), then, for any polynomial p(O) in e; 6 and e-io,
2n
p(O)d((F(O)x,y)-(F1 (0)x,y))=0, (x,yEX).
Thus, by continuity, the above equality holds for any continuous function p (0) with p (0) = p (2n}. Let 0 < 00 < 01 < 2n, and take
Pn(O)
=
=
0 < 00 and for 01
OforO
<
1 for ()0
+n<
1
+n-< 0 <
2n,
0 < Ov
= linear for 00 < 0 < 00
+~
and for 01 <
() < ()1 + ~ .
309
5. The Resolution of the Identity
Then letting n
-'J>-
oo, we obtain, by (6),
2n
J Pn (0) d [(F (8)
X,
y)-(F1 (8) X, y)J= [(F (0) X, y) -(F1 (0) X, y)Jg~= 0,
which is valid for all x, y EX. Hence, letting 80 t 0 and making use of
conditions (5) and (6), we see that F (81) = F 1 (8). Therefore, the spectral
resolution for a unitary operator is uniquely determined.

Definition 1. A family of projections E (A.), - oo < A. < oo, in a Hilbert space X is called a (real) resolution of the identity if it satisfies the
conditions:
E (A.) E (!l) = E(min (A., ft)),
(1)
E (- oo) = 0, E ( + oo) =I, where E (- oo) x =s-lim
E (A.) x and
.<.j,-00
E(+ oo) x =s-lim E(A.) x,

J.too
E(A. + 0) = E(A.), where E(A. + 0) x = s-limE(ft) x.
(2)
(3)
p.j,A
Proposition 1. For any x, y EX, the function (E (A.) x, y) is, as a function of A., of bounded variation.
Proof. Let A.1 < A. 2 < < An- Then, by (1), E (1X, {3] = E ({3)- E (1X)
is a projection. Thus we have, by Schwarz' inequality,
~I (E (A.j-1, Aj] X, y)l =~I (E (A.j-V Aj] X, E (A.j-V Aj] y)
J
II E (A.j-V Aj] X II . II E (A.j-V A.j] y II

< (f II E (A.j-1, Aj] X 11 2) 112 (f II E (A.j-1, A.j] y ll 2 f 12
=(liE (A.1,An]X W) 112 (liE (A.1,An] YW? 12 < llx IIIIY II
<
~
J
For, by the orthogonality
E (A.j-1, A.j] E (A.i-1, A.i] = 0
(i. =F j)
(4)
implied by (1), we have, form> n,
Corollary. For any A., -oo <A.< oo, the operators E(A.
s-limE (A.') and E (A.- 0) = s-limE (A') do exist.
,;.p
J.'t"
Proof. From (5), we see that, if Ant A., then
_lim 1\ E (A.j, A.k] x 112
J,k-+oo
and the same is true for the case An t A..
0,
0) =
310
Proposition 2. Let f (A.) be a complex-valued continuous function

on (- oo, oo), and let x E X. Then we can define, for -oo < ex < {J < oo,
{J
/(A.) dE (A) x
0<
as the s-lim of the Riemann sums
= A.1 < A.2 < < A.,. = {J,
~I (A.j) E (A,;, A.H 1] x, where ex

J
x1 E (A.1, A.1+1],
when the m~x IA.H 1 - A.1 I tends to zero.

J
Proof. /(A.) is uniformly continuous on the compact interval [ex, {J].

Let \f(A.)- /(A.') I< e whenever \A.- A.'! <~.We consider two partitions
of [<X, {J]:
ex= A.1 <<A.,.= {J, m~x \A.H 1 -.Ail<~.
J
ex= P-1
< <
ftm = fJ, max IP-Hl- P-i I < ~

J
and let
ex= v1 < <
Vp
{J, p < m
+ n,
be the superposition of these two partitions. Then, if p,~ E (p,k, ftk+d, we

have
I (A.j) E (A_;, Aj+l] X- I (p,~) E (p,k, P-k+d X
= .J:
8
s
E (v5 ,
V5 +1] X,
with
\8
5 \
< 2e,
and so the square of the norm of the left side is, as in (5),
<
e2
11
E(v5 ,
V5 +1]
w=
e2
I!E(ex, (J] xl! 2 <
e2
l!xl!2.
{J
00
Corollary.WemaydefineJ /(A.}dE(A.)xasthe
00
s-lim
jf(A.)dE(A.)x,
e<.j.-oo,{Jtoo"'
when the right side limit exists.

Theorem 1. For a given x EX, the following three conditions are
mutually equivalent:
J f (A.) dE (A.) x exists,

-oo
00
(6}
00
F (y)
J if(A.) 12 d !IE(A.) x!l2 < oo,

-oo
(7)
J f (A.) d (E (A.) y, x) defines a bounded linear functional.

-oo
(8)
00
Proof. We shall prove the implications (6)--+ (8) ~ (7) ~ (6).
(6)
(8). The scalar product of y with the approximate Riemann
311

00
sum of
J I(.?.)
dE(.?.) x is a bounded linear functional of y. Hence, by
-00
(y, E (.?.) x) = (E (.?.) y, x) and the resonance theorem, we obtain (8).

(8)--+ (7). We apply the operator E (ex, (3] to the approximate Riemann
{J_
J I(J.)dE(J.)x.
sum of y=
Thus,
"'
again by
We then see, by (1),that y=E(a,f3]y.
{1),
00
J I(J.)d(E(J.)x,y)=
-oo
F(y)=
=,
IX
J I(J.)d(E(J.)x,y)
+-oo,{J' too
lim
IX'
IX'
{J'
lim
t-oo,, too
J I(J.)d(E(J.)x,E(cx,f3]y)
IX'
{J'
=,
IX
lim,
.j,-oo,{J too
ji(J.)d(E(cx,f3]E(J.)x,y)
IX'
{J~
JI(J.)d(E(J.)
X,
y)
IIYII 2
IX
Hence IIYII 2
<
IIFIIIIYII, i.e., !IYI[
approximating y
fJ
J I(.?.)
<
IIFII On the other hand, by
dE(.?.) x by Riemann sums, we obtain, by (1),
IX
IIYII 2
IIJt(J.) dE(.?.) xll
fJ
so thatJif(.?.)l 2 diiE(J.)xll 2
J ii(J.)I 2 d liE(.?.) xW,
< IIFII 2 Therefore, by letting
ex-t-oo,
"'
{3 too, we see that (7) is true.

(7)--+ (6). We have, for
II
f(J.) dE(.?.) x-J f(J.) dE(.?.) xll
"'
J if(.?.)
ex'< ex< {3 < fJ',
12
d liE(.?.) xll 2
J If(.?.)
fJ
12
d IIE(.?.)xll 2
as above. Hence (7) implies (6).
Theorem 2. Let f (.A) be a real-valued continuous function. Then,

a self-adjoint operator H with D (H) = D is defined through
00
(Hx, y)
= J f(J.) d(E(J.) x, y),

-co
where
(9)
xED= {x; _[If(.?.) 12 d IIE(.A) xll 2 <=}and any yE X,

and we have HE(.?.)
E (.?.) H, that is, HE (A) is an extension of E (.?.) H.
312
> 0, there exist ex and f3 with

such that II y- E (ex, {3] y II < s. Moreover, we have
Proof. For any y E X and for any s
- = <
ex
< f3 < =
{3
00
J If (A) 2 d II E (A) E (ex, f3]y

1
J If (A) 2 d II E (A) Y
11 2 =
-00
11 2
na =X. His symmetric by
Hence E (ex, {3] y ED and so, by (2),
f(J..) =/(A), (E(A) x, y) = (E(A) y, x).

If y E D (H*) and H* y = y*, then, by E (ex, {3] z E D and (1),
J/(A) d (E (A) z, y).

{3
(z, E (ex, {3] y*) = (E (ex, {3] z, H*y) = (HE (ex, {3] z, y) =
Thus, by the resonance theorem,
J /(A)d(E(A)z,y)=F(z)
-oo
00
lim
<X.j,-oo,fltoo
(z,E(ex,f3]y*)=
is a bounded linear functional. Hence, by the preceding theorem,

00
I!(A)/ 2 diiE(A)yW<=, that is, yED.
-00
Therefore, D = D (H) ~ D (H*). Since H is a symmetric operator, we

have H ~ H* and soH must be self-adjoint, i.e., H = H*.
Finally, let xED (H). Then, by applying E (,u) to the approximate Rie-
J j(A) dE(A) x, we obtain, by (1),

00
mannsumsofHx=
-00
J f (A) d ( E (,u) E (A) x)

00
E (,u) H x =
-00
00
/(A)d(E(A)E(,u)x) =HE(,u)x.
-00
Corollary 1. In the particular case /(A) =A, we have
00
(Hx,y)=
-oo
Ad(E(A)x,y),
xED(H), yEX.
for
(10)
We shall write it symbolically
00
H=
).dE(A),
-00
and call it the spectral resolution or the spectral representation of the selfadjoint operator H.
00
Corollary 2. We have, for H =
J /(A) dE(A) given by (9),
-00
J I/(A)I2di/E(A)xiJ 2
00
I/Hxll2=
-00
whenever
xED(H).
(11)
313
6. The Spectral Resolution of a Self-adjoint Operator
In particular, if His a bounded self-adjoint operator, then
J f(A.td(E(A.)x,y)
00
(Hnx,y)=
x,yEX (n=0,1,2, ... ). (12)
for
-00
Proof. Since E (A.) H x
HE (A.) x for x E D (H), we have, by (1),
f /(A.) d (E (A.) x, H x) = f f(A.) d (HE (A.) x, x)

= f ;(A.) d;. {f f(f-t) df.'(E(f-t) E(A.) x, x)}
f(f-t) d (E (!-') x,x)} = Jf(A.) d liE (A.) x 11
= Jf(A.) d;.
(H x, H x) =
Li
The last part of the Corollary may be proved similarly.
Example. It is easy to see that the multiplication operator
Hx(t)
in L2(-oo,oo)
tx(t)
J A. dE (A.), where
00
admits the spectral resolution H
-00
E (A)
(t) =
for t < ). ,
(t)
0 for
(13)
t >A..
For,
2
d II E (A.)
X l\ 2
f A.
-oo
00
).
00
00
f A.
-oo
d;.
J IX (t) 1 dt = -oof t
-oo
2
2 IX
J A.d(E(A.)x,y)= -ooJ A.d;. -ooJ x(t)y(t}dt= -ooJ t

-oo
00
00
(t)
12
II H X 11 2 '
dt =
00
}.
x(t)y(t)dt=(Hx,y).

Theorem 1. A self-adjoint operator H in a Hilbert space X admits
a uniquely determined spectral resolution.
Proof. The Cayley transform U
= UH = (H- il) (H + iJ)- 1 of the
self-adjoint operator His unitary (see Chapter VII, 4). Let U =

be the spectral resolution of U. Then we have
2n
J ei
dF(O)
F (2:rt- 0) =s-lim F (2:rt- 0) = F (2:rt) =I.

6-!-0
If otherwise, the projectionF(2n)-F(2n-O) would not be equal to the

zero operator. Thus there exists an element y =F 0 such that
(F {2:rt} - F {2:rt- 0)) y = y.
314
Hence, by F(O) F(()') = F(min(O, ()')),

2n
Uy =
e;6 d (F(O)(F(2n)- F(2n- 0))} y
= (F(2n)- F(2n- 0)) y
y.
Thus (y, z) = (Uy, Uz) = (y, Uz) and so (y, z- Uz) = 0 for every
z EX. U being the Cayley transform of a self-adjoint operator H, we
know (see Chapter VII, 4) that the rangeR (I- U) is dense in X. Hence
we must have y = 0, which is a contradiction.
Thus, if we set
A.= -cot(), E(A.) = F(O),
then 0 < () < 2n and- oo < A. < oo are in a topological correspondence.
Hence E(A.) is a resolution of the identity with F(O). We shall show
that the self-adjoint operator
00
H' =
-co
A.dE(A.)
is equal to H. Since H = i(I + U)(I- U)-1 , we have only to show

that
(H'(y-Uy),x)=(i(y+Uy) ,x) forall x,yEX.
But, since D(H')a =X, we may restrict x to be in the domain D(H').
Now, by F(()) F(()') = F(min(O, ()')),
(y- U y, F(()) x) =
2n
J ( l - e;
6')
d 6 (F(()') y, F(()) x)
2n
J (l-e; )du.(F(()) F(()')y,x)
J (l- e'~~') d (F (()') y, x).
Hence
00
(y-Uy,H'x)=
-oo
A.d(y-Uy,E(A.)x)
j (1- e;
J i(l + e' )d(F(O)y,x)=(i(y+ Uy),x).
2n
-cot() d
{ 6
6')
d(F(O') y, x)}
2n
J A. dE (A.)
-co
00
The uniqueness of the spectral representation. SupposeH =
admits another spectral representation H =
J A. dE' (A.)
-co
00
such that
315
E' (A.0 ) #- E (A.0) for some A.0 . Then, by setting
A.=- cot 8, E' (A.)
F' (8),
we have F' (80) #- F(80 ), where A.o =-cot 80 By a similar calculation

00
to the above, we can prove thatthe Cayley transform of

2n
to
J e'
J A. dE' (A.) is equal
-oo
dF' (0). Hence the unitary operator U admits two different
2n
2n
J e'0 dF (0)
J e' 0 dF' (0),
con0
o
trary to what we have proved in Chapter XI, 4.
We have thus proved (see Chapter VII, 3 and 4) the fundamental
result due to J. vox NEUMANN [1]:
spectral representations U =
and U =
Theorem 2. A symmetric operator H has a closed symmetric extension

H**. A closed symmetric operator H admits a uniquely determined spectral representation iff H is self-adjoint. H is self-adjoint iff its Cayley
transform is unitary.
Remark. It sometimes happens, in applications, that H is not selfadjoint but H* is self-adjoint. In such a case, H is said to be essentially
self-adjoint. In this connection, see T. KATO [7] concerning Schrodinger's
operators in quantum mechanics.
The spectral representation of the momentum operator H 1 :
1 d
H 1 x(t)=Tatx(t)
in
P(-oo,oo).
The Fourier transform U defined by

x(t)
= U y(s)
= l.i.m. (2n)- 1' 2

1>-->00
J e'
-n
51
y(s) ds
is unitary and u-1 X (t) = U* X (t) = u X (-t). Hence, denoting by E (A.)

the resolution of the identity given by (13) in Chapter XI, 5, we obtain a
resolution of the identity {E' (A.)}, E' (A.) = U E (A.) U-1 . If both y (s) and
sy(s) belong to L 2 (-oo, oo) n V(-oo, oo), then
1
1 d
dt x (t) = T
T
-1/2 Joo
d (
-oo e"t y (s)
dt (2n)
ds
00
J e' sy(s) ds = U(sy(s)) =
(2nt 112
51
UsU-1 x(t),
-00
or, symbolically,
1
Tat=
u s u-l
(1)
316
Hence, for the self-adjoint operator H = s
J 2 dE(J..), we have
00
-00
U-1 H 1 U y(s)
= s y(s) = Hy(s) whenever y(s),
sy(s) both belong toP(-oo,oo)

For any y(s) E D(H) = D(s ),set
n V(-oo,oo).
Yn(s) = y(s) or= 0 according as JsJ <nor JsJ
>
n.
Then surely Yn(s), syn(s) both E P(-oo, oo) n V(-oo, oo) and, moreover, s-lim Yn = y, s-lim H Yn = H y. Thus, since the self-adjoint operan-+OO
n--KXJ
tors U- 1 H 1 U and Hare closed, we have, by (U- 1 H 1 U) Yn = Hyn,

(U-1 H 1 U) y = Hy whenever y E D(H),
1
that is, U- H 1 U is a self-adjoint extension of the self-adjoint operator
H. Hence, by taking the adjoint, we see that H* =His also an extension of (U-1 H 1 U)* = U-1 H 1 U; consequently U-1 H 1 U =Hand so
H 1 = UHU- 1
00
00
Ad(UE(J.) U-1)=
-00
f ). dE'(J.).
-00
7. Real Operators and Semi-bounded Operators.

Friedrichs' Theorem
Real operators and semi-bounded operators, defined below, have selfadjoint extensions. Thus we can apply von Neumann's theorem to these
extensions to the effect that they admit spectral resolutions.
Definition 1. Let X = L 2 ( S, IE, m) and let H be a symmetric operator
defined in X into X. His said to be a real operator, if i) x(s) ED (H) then
x (s) ED (H), and ii) H maps real-valued functions into real-valued functions.
Example. Let f(s) be a real-valued continuous function in(- oo, oo).
Then, for X= P(-oo,oo), the operator of multiplication by f(s) is a
real operator.
Theorem 1 (J. voN NEUMANN [1]). A real operator H admits a selfadjoint extension.
Proof. Let U = UH be the Cayley transform of H. Then D ( U) =
{(H + iJ) x; xED (H)} consists of the functions obtained by taking the
complex-conjugate of the functions of R(U) ={(H-ii)x;xED(H)}.
Thus, if we define an extension U1 of U through
U1
= U in D(U),
U1 (
c"<p")
=f
c"ip"', where {<p"'} is a complete
orthonormal system of the Hilbert spaceD (U).L,

then U 1 is a unitary extension of U. Therefore the self-adjoint extension
Hl of H exists such that ul = uH, (see Chapter VII, 4).
7. Real Operators and Semi-bounded Operators
317
Definition 2. A symmetric operator H is said to be upper semibounded (or lower semi-bounded) if there exists a real constant IX such that
(Hx, x)
<IX
llxll 2 (or (Hx, x)
>IX
llxll 2) for all xE D(H).
If (H x, x) > 0 for all xED (H), then His said to be a positive operator.
Example. Let q(s) be continuous and non-negative in (-oo, oo).
Consider the operator H defined for C2 functions x (s) with compact
support by
(Hx)(s) = -x"(s)
+ q(s) x(s).
Then His a positive operator in the Hilbert space L2 (- oo, oo), as may be
verified by partial integration.
Theorem 2 (K. FRIEDRICHS [3]). A semi-bounded operator H admits
a self-adjoint extension.
Proof (due to H. FREUDENTHAL [1]). -His lower semi-bounded if H
is upper semi-bounded. If H is lower semi-bounded as above, then
H 1 = H + (1 -IX) I satisfies the condition that (H1 x, x) > II x 11 2 for all
xED (H1). Therefore, since ~XI is self-adjoint, we may assume that the
symmetric operator H satisfies the condition
(Hx,x) > llxll 2 for all xED(H).
(1)
We introduce a new norm II x II' and the associated new scalar product
(x, y)' in D (H) through
(2)
llxii'=(Hx,x), (x,y)'=(Hx,y).
Since His symmetric and satisfies (1), it is easy to see that D (H) becomes
a pre-Hilbert space with respect to llx II' and (x, y)'. We denote by D(H)'
the completion of this pre-Hilbert space.
We shall show that D(H)' is, as an abstract set without topology, a
subset of the set X which is the original Hilbert space. Proof: A Cauchy
sequence {x,.}' of the pre-Hilbert space D(H) satisfies llx,.-xmll' >
llx,.-xmlland lim Jlx,.-xmll' = O;consequently{x,.}isalsoaC auchy
n,fn-1o-OO
sequence of the original Hilbert space X. If we can show, for a Cauchy
sequence {y,.} of D (H)', that
~ IIY.. II' # 0 does not imply
~ JJy,.JI = 0,
(3)
then the correspondence
(4)
is one-one from the Cauchy sequences of D (H) into the Cauchy sequences
of X. Two Cauchy sequences {x,.}', {z,.}' of D(H) (of X) being identified
if~ JJx,.-z,.IJ' = O{if ~ JJx,.-z,.JJ = 0). Since X is complete, we
may thus identify its Cauchy sequence {x,.} with the element x EX such
318
that lim
......oo
llx,.- x II= 0.
Therefore, D (H)' may, as an abstract set with-
out topology, be identified with a subset of X by the correspondence (4).

The proof of (3) is obtained, remembering the continuity of the scalar
product in D (H)' and in X, as follows. lim II x,.- xm II' = 0,
lim
......00
II x,. II' = o.: >

o.:2
n,~oo
0 and lim
......00
\I x,. \I = 0 imply a contradiction
= n,m-+OO
lim (x,., Xm)' = lim (H x,., xm) =
n,m-+00
lim (H x,., 0)
n---+=
= 0.
We next set
D(H*)
n D(H)'.
(5)
Since D(H) ~ D(H*), we must have D(H) ~ D ~ D(H*). Hence we can

define an extension H of H by restricting H* to the domain i5 = D ( H).
We have to show that His self-adjoint.
We first show that His symmetric. Suppose x, y ED; there exist two
sequences {x,.}', {y,.}' of D(H) such that llx-x,.l\'-+ 0, IIY-Ynll'-+ 0
as n-+ = Hence, by the continuity of the scalar product in D (H)',
we see that a finite lim (x,., Ym)' = lim (H x,. Ym), = lim (x,., H Ym)
n,m-+00
exists. This limit is equal to

lim lim (H x,., Ym)
n--+<)0,
m-KXJ
n,tn-J-CO
n,~
= lim (H x,., y) = n--HXl

lim (x,., H y) = (x, H y)
n----KX)
and also to
lim lim (H x,., Ym)
= lim (x, H Ym) = lim (H X, Ym)
Hence His symmetric, that is, H

m-+00 n-+<xJ
~ (H)*.
m-+<X)
= (fiX, y).
Next let xED (H), y EX. Then, by
l(x,y)l < llxii\IYII < llxii'/IYII,

we see that I (x) = (x, y) is a bounded linear functional on the pre-Hilbert
space D(H). Hence l(x) can, by continuity, be extended to a bounded
linear functional on the Hilbert spaceD (H)'. Hence, by F. Riesz' representation theorem as applied to the Hilbert space D (H)', there exists a
uniquely determined y' ED (H)' such that
I (x)
= (x, y) = (x, y')' = (H x, y')
for all xED (H).
This proves that y' E D (H*) and H* y' = y. Hence y' E Dand fi y' = y.
We have thus proved that R(H) =X, and so, by the Corollary of Theorem 1 in Chapter VII, 3, H must be self-adjoint.
319
8. The Spectrum of a Self-adjoint Operator
8. The Spectrum of a Self-adjoint Operator. Rayleigh's

Principle and the Krylov-Weinstein Theorem. The Multiplicity of
the Spectrum
Theorem 1. Let H = il dE (il) be a self-adjoint operator in a Hilbert

space X. Let a(H), Pa(H), Ca(H) and Ra(H) be the spectrum, the point
spectrum, the continuous spectrum and the residual spectrum of H,
respectively. Then (i) a (H) is a set on the real line; (ii) il0 E P a (H) is
equivalent to the condition E (ilo) =I= E (Ao- 0) and the eigenspace of H
corresponding to the eigenvalueil0 is R (E (ilo)- E (ilo- 0)); (iii) il0 E Ca(H)
is equivalent to the condition E (ilo) = E (ilo- 0) in such a way that
E (il1 ) =I= E (/.2 ) whenever il1 < il0 < ~; (iv) Ra (H) is void.
Proof. We know already that, for a self-adjoint operator H, the resolvent set e(H) of H comprises all the complex numbers il with Im (il) =/=: 0
J dE (il) by
-oo
00
(see Chapter VIII, 1). Hence (i) is clear. We have ilol = ilo
the definition of the resolution of the identity {E (il)} and so (H -il0 I) =
J (il- il
-oo
00
0)
dE (il). Hence, as in Corollary 2 of Theorem 2 in Chapter XI, 5,
we obtain
00
jj(H-ilol)xll 2 =
-oo
(il-ilo) 2 diiE(il)xll 2 , xED(H).
(1)
Thus, byE(- CXJ) = 0 and the right continuity of II E (il) x 112 in il, we see
that H x = A.ox iff
{
0) x = E (ilo) x for il
E (il) x = E (Ao
E (A.) X = E (Ao- 0) X = 0 for il < Ao
> ilo and
that is, H x = A.ox iff (E (ilo)- E (ilo- 0)) x = x. This proves (ii). Next
we shall prove (iv). If il0 E Ra(H), then, by (i), il0 is a real number. By
the condition R(H -ilol)" = D((H -iloi)-1)"=1= X, we see that there
exists a y =I= 0 which is orthogonal toR (H- ilol), i.e., ((H- ilol) x, y) =0
for all xED (H). Hence (H x, y) = (ilox, y) = (x, iloY) and soy ED (H*),
H*y = iloY This proves that Hy = A.oy, i.e., A.o is an eigenvalue of H.
Hence we have obtained a contradiction ).0 E Ra(H) n Pa(H), and so
Ra (H) must be a void set.
Let A.o be a real number not belonging to a(H). Then the resolvent
(ilol- H)-1 exists. Hence H;.. = (H- A.0 I) has a continuous inverse
(H- iloi)-1 The last condition is, by (iv), equivalent to il0 E e (H) and
to the condition that there exists a positive number X such that
II (H- ilol) x II >
II x II for all x E D (H).
320
This condition is by (1) equivalent to

00
J (.A-.A
0) 2
dJJE(.A)xii 2 >X 2 /JxJJ 2
forall
xED(H).
(2)
-00
Suppose that we have, for A1 < Ao < A2 with A0 - A1 = A2 - Ao <

E(.A1) = E(.A2). Then, contrary to (2), we obtain
00
J
-oo
X,
00
(A-Ao)2dJJE(.A)xJJ2 <X2
J di!E(.A)xii2=X211xll2
-oo
This proves (iii) by (i), (ii), (iv) and (2).

Remark. The Example in Chapter XI, 5 gives a self-adjoint operator H
for which all real numbers are in the continuous spectrum of H.
Theorem 2. Let H be a bounded self-adjoint operator. Then
sup A= sup (H x, x), inf A= inf (H x, x).

llx/1 :;i;l
lEu(H)
.<Eu (H)
llxll :;i;1
(3)
Proof. Since (H x, x) = (x, H x) = (H x, x) =real number, we can

consider
X1 = inf (H x, x) and X 2 = sup (H x, x).
/lrll:;i;1
llrll:;i;1
Let .A0 E a(H). Then, by Theorem 1, there exists, for every pair
(A1 , A2) of real numbers with A1 < .1.0 < .1. 2, a y = Y.<...:t, '1=- 0 such that
(E (~)- E (.1.1)) y = y. We may assume that IIY II = 1. Hence
J.Ad(E (.A) y,y) = J.Ad IIE(.A) y J/2

= f .Ad //E (.A) (E (.1.2) -E (.A1))y // 2
...
= J Ad II (E (.A) - E (.1.1)) y 11 2.
A,
(Hy, y) =
Thus, by letting A1 t .1.0 , A2 .} .1.0 , we obtain lim (H Y;.,,;.,, YA,,A,) = Ao This

proves that sup A= sup Ao < X2.
AEu(H)
Let us assume that .x2 E a (H). Then, by Theorem 1, there exists a pair
(Av A2) of real numbers such that A1 < X 2 < .1.2 and E (.1.2) = E (.1.1). Hence
I= I -E(.A2) + E(.A1 ), (I -E(.A2)) E(.A1 ) = E(A1 ) (I -E(.A2 )) = 0 and
so either (I-E(~)) orE (.1.1) is not equal to the zero operator. If
(I- E(.A2)) '1=- 0, then there exists a y with IIY II= 1, (I-E (.1.2)) y = y.
In this case, we have
(Hy, y) = Ad IJE(.A) yJ/ 2 = Ad /JE(.A) (I -E(.A2)) Yll 2
J
...f
00
Ad IIE(.A) Yl[ 2 > A2
and in the case E (.1.1) cf=. 0, we obtain

(Hz, z) < .1.1 < X 2 for a z with
> 1X2;
liz II= 1,
E(.A1) z = z.
321
8. The Spectrum of a Self-adjoint Operator
Therefore the assumption <X 2Ea (H) is absurd and so we have proved that
sup A.= sup, H(x, x). Similarly, we can prove that inf A.= inf
AEa(H)
!lx!l~l
l!xll~l
J.Ea(H)
(Hx ,x).
Theorem (KRYLOV-WEINSTEIN). Let H be self-adjoint, and define,

for any x E D(H) with llx II= 1,
<X_.= (Hx,x), f3x= IIHxll
(4)
Then, for any s > 0, we can find a A.0 E a(H) satisfying the inequality
<X_.- ({3;- <X;)l/2_ e < Ao < <Xx

Proof. We have
p; =
<Xx
(Hx, Hx) = (H 2 x, x)
(H X, x)
l!xll 2
JAd II E (A.)
=J d IIE(J.) xl[2,
+ ({3;- <X;)l/2 + e.
JA.
d IIE(A.) x112,
I[2 ,
and so
JJ. d IIE(J.) 11 2<X xf Ad IIE(A.)

= J (A- <Xx)2 d liE (A.) 112.
{3~ -<X~=
(5)
2-
11 2 +<X;
Jd IIE(A.) xll
Therefore, if IIE(A.) xll 2 does not vary in the interval given by (5), we
would obtain a contradiction
p; _<X~> (({3~ _
<X~)l/2
+ e)2 >
{3~
_<X;.
Remark. The so-called Rayleigh principle consists in taking <Xz as

an approximation to the spectrum of the operator H. If we calculate f3x,
then Theorem 3 gives an upper bound of the error when we take
<X_. as an approximation of the spectrum of H. For a concrete application of
such error estimate, we refer the reader to K. YosmA [1].
The Multiplicity of the Spectrum. We shall begin with the cace of a
self-adjoint matrix H = J. dE (J.) in ann-dimensional Hilbert space Xw
Let Av A.2 , . , Ap (p < n) be the eigenvalues of H with the multiplicity
mv m2 , . , mp. respectively
(1 m
j=l
x3,
= n). Let x3,
2
1
x3
m;
be ortho-
normal eigenvectors of H belonging to the eigenvalue J.j(H xis= J.ixi) so

that{xj,; s = 1, 2, ... ,mj} spans the eigenspace E;.1 = R (E (A.j) -E (A.j-0))
of H belonging to the eigenvalue J.i. Then the set {xi,; j = 1, 2, ... , p
and s = 1, 2, ... , mi} is a complete orthonormal system of vectors of
the space Xn, and hence every vector y of Xn is represented uniquely as
the linear combination of xj,'s:
y
m;
_I _I r.Xj Xj
= j=l
s
s=l
8
(6)
322
Thus, denoting by P;.1 the projection (E (A;)- E (Aj- 0)) upon the
eigenspace E;.1 , we have, for any X< {3,
(E ({3)- E (1X)) y
m;
= D<<A;:;;;,p
.I ( _IXi.xi,) = .I P 11 y,
s=l
<X<J.;:;;;,p
(7)
= s~ 1Xj, xj, or =
0 according as X< Aj < f3 or not.

(8)
Hence, for fixed X< f3 and a fixed linear subspace M of X, the set
PJ.; (E ({3)- E (1X)) y
{(E(f3)- E(1X)) y; y EM}

does not contain E;.1 if the dimension of M, dim(M), is< mr Moreover,
for a suitable M with dim (M) = mj, the set {(E ({3) - E (X)) y; y E M}
with X< A1 < f3 contains E 1.r In fact, the statement is true forM containing xj, xj, ... , xjm; In particular, m1 = m2 = = mp = 1 with
p = n iff there exists a fixed vector y E Xn such that the set of vectors
{(E ({3)- E (X)) y; X< {3}
spans the whole space xn"
These considerations lead to the following definitions:
Definition 1. The spectrum of a self-adjoint operator H = JAdE (A.)
in a Hilbert space X is said to be simple, if there exists a fixed vector
yEX such that the set of vectors {(E(f3) -E (1X)) y; X< {3} spans a linear
subspace strongly dense in X.
Definition 2. Let H =
A. dE (A.) be a self-adjoint operator in a Hilbert
space X. For fixed X< {3, consider the linear subspace M of (E ({3)E (1X)) X such that
(E ({3)- E (1X)) M = (E ({3)- E (1X)) X.
(9)
We may take, for example, M = (E(f3) -E(X)) X to meet condition
(9). The minimum of the set of values dim (M), where M satisfies
condition (9), will be called the total multiplicity of the spectrum of H
contained in the interval (X, {3].
Definition 3. The multiplicity of the spectrum of a self-adjoint operator
H =JAdE (A.) at A.= A.o is defined as the limit, as n-+ oo, of the total
multiplicity of the spectrum of H contained in the interval (Ao- n-1 ,
Ao + n-1].
Example. The coordinate operator H, i.e., the operator H defined by
H x(t) = t x(t) in L 2 (-oo, oo), is of the simple spectrum.
Proof. We know that the spectral resolution H =
A. dE (A.) is given by
E (A.) x (t) = x (t) or = 0 according as t < A. or t > A..
Let y (t) be defined by
y(t)
ck
>
0 for k -1 < t
< k (k = 0, 1, 2, ... )
with .Ic~<oo sothat y(t)EL 2 (-oo,oo).

k
9. The General Expansion Theorem
323
Then it is easy to see that the linear combinations of vectors of the form
(E ({3) -E (iX)) y, lX < {3, are strongly dense in the totality of step functions
with compact support and consequently strongly dense in L2(-oo,oo).
The Problem of the Unitary Equivalence. Two self-adjoint operators

H 1 and H 2 in ann-dimensional Hilbert space Xn are said to be unitarily
equivalent to each other if there exists a unitary matrix U in Xn such
that H 1 = U H 2 U-1 . It is well-known that H 1 and H 2 are unitarily
equivalent to each other iff they have the same system of eigenvalues
with respectively the same multiplicities. Thus the eigenvalues together
with the respective multiplicities are the unitary invariants of a self-adjoint matrix.
The investigation of the unitary invariants for a self-adjoint operator in an infinite dimensional Hilbert space goes back to a paper by
E. HELLINGER [1] published in 1909. See, e.g., M. H. STONE [1]. There
the Hilbert space is assumed to be separable. For the non-separable Hilbert
space case, see F. WECKEN [1] and H. NAKANO [1] and also P. R. HALMOS [2]. K. YosmA [13] proved the following theorem:
Let H be a self-adjoint operator in a Hilbert space X, and let us
denote by (H)' the totality of bounded linear operators E L (X, X) which
are commutative with H. Then two self-adjoint operators H 1 and H 2 in
X are unitarily equivalent to each other iff there exists a one-to-one
mapping T of (H1 )' onto {H2)' such that T defines a ring-isomorphism
of the ring {H1)' with the ring (H2 )' in such a way that (T B)* = T B*
for every BE (H1)'.
Thus the algebraic structure of the ring (H1 )' is the unitary invariant
of H 1 .
9. The General Expansion Theorem. A Condition for the Absence

of the Continuous Spectrum
Let H = .A dE (.A} be a self-adjoint operator in a Hilbert space X.

Then, byE(+ oo) =I and E(-oo) = 0, we have the representation
A,
(E(J.2)-E(J.1})xfor everyxEX.
s~lim
J dE(J.) x = A .j. -oo,A,
= A .j. s-lim
too
-oo,A, too A,
1
(1)
We shall call (1) the general expansion theorem associated with the selfadjoint operator H. In concrete cases, it sometimes happens that the
resolvent (.AI- H)- 1 is obtainable more easily than the spectral resoludE (.A). In such a case, the general expansion theorem (1)
tion H =
J;.
21*
324
may be replaced by
x=
s-lim
a.j.-oo,!ltoo
+/
s-lim -2 1 . [
:n~
v,j.O
j ((u- iv) I - Hr
"'
+ iv)I -H)- 1 xdu],
((u
x du
(1')
xEX.
Proof. If v =I= 0, then
J - + ~-,dE (A.) x, for every

00
((u
+ iv) I - H)- 1 x =
ZV-11.
xE X.
-00
For, by approximating the integral by Riemann sums and remembering

the relation E(A.) E(A.') = E(min(A., ).')),we obtain, for Im (t-t) =I= 0,
Ll ,
_[ (!.- p) dE(!.)
j ().-
t-t) d)
-oo
Li
1 p)
dA' (E ().) E (A.')))
l-oo
_[(A- p) d,
_[; 1
j (.?.'
1 dE(!.'))
j 1
dE(!.')!~_[ dE(!.)~ I,
"d, {E(!.) _[ (!.'- p) dE(!.'))
_ [ ' 1 "d,
_ [ ' 1 d,
Ll
Li
(!.'- p) d,. (E (!.) E (!.')))
dE(!.'))~ _[dE(!.) ~ 1.
(!.' - p)
Therefore, we have
J
ll
((u- iv) I - H)- 1 x du
"'
-j
-oo
J ((u + iv)
"'
I - H)- 1 x du
ll
j u-~:-.1.. f u+~:-.?.)
xjj
;v-
dE(A.) XJ
_[dE(!.)
la
ll
d. log (u-
!.) +/d. log (u
Hv - I
A)
9. The General Expansion Theorem
325
which tends, when v,). 0, strongly to

fl-O
<>+0
2nidE (A.) x
+ ni (E ({3)- E ({3- 0)) x + ni (E (ex)- E (ex- 0)) x
=ni(E({J) +E({J-O))x-ni(E(ex) +E(ex-O))x.

This proves formula (1').
Remark. The eigenfunction expansion associated with the second
order differential operator
d2
- dx2
+ q(x)
with real continuous q (x) in an open interval (a, b) was inaugurated by

H. WEYL [2], further developed by M. H. STONE [1], and completed
by E. C. TITCHMARSH [2] and K. KoDAIRA [1] who gave a formula which
determines the expansion explicitly. The expansion is exactly a
concrete application of (1). The crucial point in their theory is to give the
possible boundary conditions at x = a and x = b so that the operator
d2
- dx2
+ q(x)
becomes a self-adjoint operator H in the Hilbert space L2(a, b). Their

theory is very important in that it gives a unified treatment of the
classical expansions in terms of special functions, such as the Fourier
series expansion, the Fourier integral representation, the Hermite polynomials expansion, the Laguerre polynomials expansion and the Bessel
functions expansion. We do not go into details, and refer the reader to
the above cited book by TncHMARSH and the paper by KoDAIRA. Cf.
also N. A. NAIMARK [2], N. DUNFORD- J. SCHWARTZ [5] and K. YoSIDA [1].
The last cited book gives an elementary treatment of the theory.
If the continuous spectrum Ca (H) is absent, then the expansion (1)
will be replaced by a series rather than the integral. We have, for instance, the following
Theorem 1. Let H = A. dE (A.) be a self-adjoint compact operator in
a Hilbert space X. Then (i) Ca (H) contains no real number except possibly
0; (ii) the eigenvalues of H constitute at most a countable system of real
numbers accumulating only at 0; (iii) for any eigenvalue A.0 oF- 0 of H,
the corresponding eigenspace EJ., is of finite dimension.
Proof. Suppose a closed interval [A', A"] on the real line does not
contain the number 0. Then the range R (E (A")- E (A')) is of finite
dimension. If otherwise, there exists, by E. Schmidt's orthogonalization
in Chapter III, 5, a countable orthonormal system {xi} contained in
R(E (A.")- E (A')). We have w-.lim xi= 0, since, by Bessel's inequality,
J->00
326
Hence, by the compactness of the operator H, there exists a subsequence

{xr} such that s-.lim H xr = w-lim H xr = 0. On the other hand, we have
J-..ro
// H xi // 2 =
J-..ro
J.1.2 d // E (I.) xi // 2 = J.1.2 d /IE (I.) (E (A") ,t"
J ).2 d 1/( E (I.) -
E (/.')) xi /12
).'
>
E (A')) xi// 2
II xi /12 . min ( \).' J2' I/." J2)
=min (/A' J2, J.l."/2),

which is a contradiction.
If C" (H) contains a number /.0 =I= 0, then, by Theorem 1 in the preceding section, s-lim (E (1.0 - e) - E {/.0 - e)) x = 0 for any x E X. As proved
e.j-0
above, the range R(E (/.0 + e)- E (Ao- e)) is of finite dimension and
this dimension number is monotone decreasing as e t 0. Hence we see
that (E(/.0 +e) -E(./.0 -e)) = 0 for sufficiently small e > 0, and so,
by Theorem 1 in the preceding section, /.0 cannot be contained in C" (H).
This proves our Theorem.
Corollary 1. Let {.l.i} be the system of all eigenvalues of H different
from 0. Then, for any x EX, we have
,.
x = (E(O)- E(O- 0)) x +s-lim.~ (E(I.j) -E(I.i- 0)) x.

n--..oo ;~1
(2)
Proof. Clear from (1).

Corollary 2 (The Hilbert-Schmidt Expansion Theorem). For any xE X,
we have
n
H x =s-lim.~ l.i(E(I.j)- E(l.i- 0)) x.

11->00 J=l
(3)
Proof. Clear from the continuity of the operator Hand the fact that
H (E (0)- E (0- 0)) = 0 and H (E(I.j)- E (l.i- 0)) x = l.i (E(.I.i)
- E (Aj- 0)) x, the latter being implied by R (E (l.j) -E (.l.j- 0)) = E.t;'
the eigenspace of H belonging to the eigenvalue Aj
Remark. The strong convergence in (3) is uniform in the unit circle
{x; llxll < 1}. For, we have
II Hx-
J /.dE(.I.)x11
l.tl>
=11
J d//E(I.)xll
< e2
J .l.dE(I.)xl1
l.tl;>;
J }.2 dllE(.I.)xll2
< e2 Jd IIE(.I.) xll 2 = e2 llxll 2 < e2.
10. The Peter-Weyl-Neumann Theory

Let G be a totally bounded topological group, metrized by a
distance satisfying the condition (see Chapter VIII, 5)
dis(x,y)=dis(axb,ayb) for every x,y,a and bEG.
(1)
327
10. The Peter-Weyl-Neumann Theorem
Let /(g) 'be a complex-valued bounded uniformly continuous function

defined on G. For any e > 0, set
{2)
V=lyEG;supif(x)-f(y- 1 x)l<s}.
xEG
Then, by the continuity of/, the set Vis an open set containing the unit
e of the group G. Hence the set U = V f\ V-1 , where V-1 = {y-1 ; y E V},
is also open and 3 e. If we put
k(x) = 2-1 (k1 (x)

k1 (X) -_
+ k1 (x-1)),
where
dis(x, U 0 )
dis (x, e)
(dis(x, uc) = inf0 dis(x,y)),
+ dis (x, U a)
(3)
yEU
then we have the results:
k(x) is bounded uniformly continuous on G, and

k(x)
= k(x-1),
k (x)
< k (x) <
1 on G, k(e)
1 and
(4)
0 whenever x E uc.
Hence, for all x, y E G, we obtain
!k(y) (l(x)-f(y- 1 x))! <sk(y).

By taking the mean value (see Chapter VIII, 5) of both sides, we obtain
!My(k(y)) f(x)- My(k(y) f(y- 1 x)) I< e My(k(y)).
We have My(k(y))
!f(x) -My(k0 (y)
>
>
0 by k(y)
f(y- 1 x))l
<
e,
0 and k(y) $ 0. Hence
where
k0 (x) = k(x)fMx(k(x)).
= My(g(y))
Thus, by virtue ofthe invariance
My(g(ya)) of the mean value, we obtain, from (5),
My(g(y-1))
lf(x)-My(k0 (xy- 1 )/(y))i<s forall
My(g(ay))
xEG.
(5)
=
(6)
Proposition I. We shall denote by C (G) the set of all complex-valued

bounded uniformly continuous functions h (g) defined on G. Then
C (G) is a B-space by the norm II h llo = sup Ih (g) I Then, for any b and
hEC(G),
(7)
(bxh) (x) = My(b(xy- 1 )h(y))
also belongs to C (G).

Proof. By dis (x, z) = dis (axe, azc) and the uniform continuity of the
function b(g), there exists, for any tJ > 0, an 'YJ = rJ(tl) > 0 such that
sup Ib (xy-1) - b (x' y-1)
y
I<
tJ
whenever
dis (x, x')
< 'YJ.
Hence, as in the case of Schwarz' inequality, we obtain
IMy(b(xy- 1 ) h(y))- My(b(x'y- 1 ) h(y)) 12
< My((b(xy- 1 ) - b(x'y- 1 )/). My(ih(y) 12) < o2 My(!h(y) !2)

whenever dis (x, x') < 'YJ This proves our Proposition 1.
(8)
328
Proposition 2. C (G) is a pre-Hilbert space by the operation of the

function sum and the scalar product
(b, h)= (bxh*) (e)= My(b(y- 1 ) h(y- 1 )) = My(b(y) h(y)), where
(9}
h*(y) = h(y- 1 ). We shall denote this pre-Hilbert space by C(G).
Proof. Easy.
Proposition 3. We shall denote the completion of the pre-Hilbert space
C(G) by C(G), and the norm in t~e Hilbert ~pace C(G) by II h II =
(h, h) 1/ 2
Then the linear mapping Ton C (G) into C (G) defined by
(Th) (x) = (k0 Xh) (x), xE G,
(10)
can, by the continuity in C (G), be extended to a compact linear operator

Ton C(G) into C(G).
Proof. By virtue of My (1) = 1, we obtain
llhll =
(h,
h) 112 =
My(h(y) h(y)/12 <sup

y
lh(y) I= llhllo
(11)
The continuity of the operator T in C (G) is clear from the Schwarz

inequality for (7), and so, by the denseness of C(G) in C(G), we can
extend T to a bounded linear operator fin C(G). On the other hand, we
see, by (8), that Tis a compact operator on C (G) into C (G). We prove
this bytheAscoli-Arzela theorem. Thus, by (11), we easily see that Tis a
compact operator on C(G) into C(G).
Therefore, by the denseness of C (G) inC (G), we see that the extended
operator Tis also compact as an operator on C (G) into C (G).
We are now ready to prove the Peter-W eyl-Neumann theory on the
representation of almost periodic functions.
It is easy to see that, by k 0 (xy- 1 ) = k0 (yx-1 ), the compact operator Tis
self-adjoint in the Hilbert space C(G). Hence, by the Hilbert-Schmidt
expansion theorem in the preceding section, we obtain
A
llh II< 1, (12)

if we denote by {llm} the system of all eigenvalues of T different from
0, and by P;.m the projection upon the eigenspace of T belonging to the
Th =s-lim ~ AmP;.mh uniformly in h satisfying
~m~l
eigenvalue Am
Since f (g), introduced at the beginning of this section, belongs to
C (G), we have Tf = Tf E C (G). Since the eigenspace R(P;.m) = P;.m C (G)
is of finite dimension, there exists, for every eigenvalue Am, a finite
system {hmj}j~ 1 , ... ,nm of elements E C(G) such that each hER (P;.m) =
329
P;.m C (G) can be represented as a uniquely determined linear combination of hm/s (j = 1, 2, ... , nm) Let
nm
= i~
P;.mh
cihmi where c's are complex numbers.
(13)
Then, since hmjE R(PJ.m), we have Thmj = Amhmi and so, by applying
(8) to the operator T given by (10), we see that hmi = X;/(Ihmj) mu:t
belong to C (G). Hence, by (13), we see that, for each eigenvalue Am ofT,
the eigenspace R (P,.,.) = P;.m C (G) is spanned by the functions hmjE C (G).
Therefore, we have, by (12),
s~ m...; Amlm (x) in the strong topology of C(G),}
(Tf) (x) =
where fm
P;.m f E C (G) for each m.
By applying (8) to the operator T given by (10), we see that
(T21) (x)
= ~ My(ko(xy- 1) m...;1 Amfm(Y)) uniformly in x.
{14)
On the other hand, by (6) and My(k 0 (y)) = 1, we obtain
IM. (k 0 (xz- 1 ) I (z))- M. (k 0 (xz- 1 ) My (k 0 (zy- 1 ) l(y)) I < e,
and so, combined with (6), we have
If (x) - M. (k0 (xz- 1 ) My (k 0 (zy- 1 ) I (y))) I < 2s.
(15)
The left hand side is precisely if(x)- (T2f) (x) I < 2s.
Since T R (P;.m) ~ R (P;.m), we have proved
Theorem 1. The function f (x) can be approximated uniformly on G by

linear combinations of the eigenfunctions of T belonging to the eigenvalues which are different from 0.
We shall take a fixed eigenvalue J.. =!=- 0 of T, and shall denote the
base {hi} ~ C (G) of the corresponding eigenspace P;. C(G) by e1 (x),
e2 (x), .. . , ek(x). Then, by the invariance of the mean value, we obtain,
for any aE G,
My (k 0 (xy- 1 ) ei (ya)) =My (k 0 (xa a- 1 y- 1 ) ej(ya)) = M. (k 0 (xa z- 1 ) ej(z))
(Tej) (xa)
= J..ej(xa).
Since the left hand side is equal to the result of applying the operator T
upon the function ei (ya) of y, we see that, for any given a E G, the function ei (xa) of x must be uniquely represented as a linear combination of
the functions et(x), e2 (x), .. . , ek(x). We have thus
k
ej(xa)=.~dj;(a)e;(x)
=1
(j=1,2, ... ,k),
(16)
330
or, in vectorial notation,
e(xa)=D(a)e(x).
(16')
By e(x ab) = D(ab) e(x), e(xa b)= D(b) e(xa) = D(b) D(a) e(x), we
see, remembering the linear independence of edx), e2 (x), ... , en (x), that
D(ab) = D(b) D(a), D(e) =the unit matrix of degree k.
(17)
By applying E. Schmidt's orthogonaliz~tion, we may assume that {ei (x)}

constitutes an orthonormal system in C(G). Then, by (16),
M_..(ej(xa) e7{x)) = dj;(a)
(18)
and so the elements dii (a) of the matrix D (a) belong to C (G). By the
invariance of the mean value, we see that
M,(ei(ya) e;(ya)) = M,(ej(Y) e;(y)) = b;j

Hence the matrix D (a) gives a linear mapping transforming the orthonormal system {ei(x)} onto the orthonormal system {ej(xa)}. Therefore
D (a) must be a unitary matrix. Hence the tranposed matrix D (a)' of
D (a) is also unitary and we have
D(ab)' = D(a)' D(b)', D(e)' =the unit matrix of degree k. (17')

Hence D (a)' gives a unitary matrix representation of the group G such that
its matrix elements are continuous functions of a. Letting x = e in (16),
we see that each ei (a) is a linear combination of the matrix elements of
the representation D (a)'.
We have thus proved
Theorem 2 (PETER-WEYL-NEUMANN). Let G be a totally bounded
topological group, metrized by a distance satisfying dis (x, y) =
dis(axb, ayb). Let f(g) be any complex-valued bounded uniformly
continuous function defined on G. Then, f (g) can be approximated uniformly on G by linear combinations of the matrix elements of unitary
uniformly continuous matrix representations D (g)' of the group G.
Referring to A. Weil's reduction given in Chapter VIII, 5, we obtain
the following
Corollary. Let G be an abstract group, and /(g) an almost periodic
function on G. Then f (g) can be approximated uniformly on G by linear
combinations of the matrix elements of unitary matrix representations
D (g)' of the group G.
Remark 1. Let the degree of a unitary matrix representation D (g)'
of the group G bed, i.e., the degree of the matrix D (g)' be d. Then each
D (g)' gives a linear mapping of a fixed d-dimensional complex Hilbert
space Xd onto itself. The representation D (g)' is said to be irreducible if
there is no proper linear subspace =I= {0} of xd which is invariant by
applying the mappings D (g)', g E G. Otherwise, the representation
331
D (g)' is called reducible, and there exists a proper linear subspace

Xd,l =F- {0} invariant by every D (g)', g E G. Then the orthogonal com-
plement
xt1
of
xd,1
in
xd
is, by the unitarity of the representation
D(g)', also invariant by every D(g), gE G. If we take for the base of the
linear space Xd the orthonormal system of vectors composed of one

orthonormal system of vectors of Xd, 1 and one of X;};1 , then, by this
choice of an orthonormal base of Xd, the representation D (g)' will be
transformed into the form
UD(g)' u-1
(D cig)' D 2 ~g)'),
1
where U is a fixed unitary matrix.
Hence a reducible unitary representation D (g)' is completely reducible

into the sum of two unitary representations D 1 (g)' and D 2 (g)' of the
group G acting respectively upon Xd, 1 and X;};1 . In this way, we finally
can choose a fixed unitary matrix Ud such that the representation
Ud D (g)' U-;; 1 is the sum of irreducible unitary representations of the
group G. Therefore, in the statements of Theorem 2 and its Corollary,
we can impose the condition that the matrix representations D (g)' are all
irreducible.
Remark 2. In the particular case when G is the additive group of real
numbers, a unitary irreducible representation D (g)' is given by
D (g)'
ei"'g,
where IX is a real number and i =
FJ: .
(19)
For, by the commutativity of the unitary matrices D(g)', gE G, the

representation D (g)' is completely reducible into the sum of one-dimensional unitary representations x(g), that is, complex-valued solutions
X (g) of
X (g1
+ g2) =
X (gl) . X (g2)'
IX (gl) I =
1 (gl, g2 E G)' X (0)
1.
(20)
It is well known that any continuous solution of (20) is of the form

X (g) = eicxg. Hence, any continuous almost periodic function f (g) on the
additive group G of real numbers can be approximated uniformly on G

by linear combinations of e;"'g with real!X's. This constitutes the fundamental theorem in H. Bohr's theory of almost periodic functions.
According to the original definition by Bohr, a continuous function f(x)
(- oo < x < oo) is called almost periodic if for each s > 0 there exists a
positive number p = p (s) such that any interval of the form (t, t + p)
contains at least one r such that
lf(x+r)-f(x)i<s for
-oo<x<oo.
See H. BoHR [1]. S. BocHNER [4] has shown that a continuous function
f(x) (-oo < x < oo) is almost periodic in Bohr's sense iff the following
condition is satisfied: For any sequence of real numbers {an}, the system
of functions {fa. (x); I an (x) :______ f (x + an)} is totally bounded in the topology
332
of the uniform convergence on (- =, =). It was extended by J. VON

NEUMANN [4] to almost periodic functions in a group. Neumann's result
contains as a special case, the Peter-Weyl theory of continuous representation of a compact Lie group (PETER-WEYL [1]). According to our treatment, Bohr's result is easily proved by obsering that lim Is- t I = 0
implies lim [sup lf(asb) --f(atb)l] = 0.
a,b
11. Tannaka's Duality Theorem for Non-commutative Compact

Groups
Let G be a compact (topological) group. This means that G is a compact
topological space as well as a group in such a way that the mapping
(~, y)-+ xy-1
of the product space G X G onto G is continuous.
Proposition 1. A complex-valued continuous function f (g) defined

on a compact group G is uniformly continuous in the following sense:
for any
>
0, there exists a neighbourhood U (e) of
the unit element e of G such that lf(x) -f(y) I< s

whenever xy-1 E U (e) and also whenever x-1y E U (e).
(1}
Proof. Since f(x) is continuous at every point a E G, there exists a

neighbourhood Va of a such that x EVa implies lf(x)- f(a) I< s/2If we denote by Ua the neighbourhood of e defined by Ua = Vaa- 1 =
{va- 1 ; vE V}, then xa- 1 E Ua implies lf(x) -f(a) I< sj2. Let us denote
by Wa the neighbourhood of e such that TV;<;;; Ua, where w; =
{w1 w2 ; wiE Wa (i = 1, 2)}. Obviously, the system of all open sets of
the form Wa a, where a is an arbitrary element of G, covers the whole
space G. G being compact, there exists a finite set {ai; i = 1, 2, ... , n}
such that the system of open sets Wa; ai (i = 1, 2, ... , n) covers G.
We denote by U (e) the intersection of all open sets ofthe system {WaJ
Then U(e) is a neighbourhood of e. We shall show that if xy-1 E U(e),
then lf(x) -f(y) I< s. Since the system Wa; ai covers G, there exists
a number ksuch thatya; 1 E Wak <;;; Uakand therefore lf(y) -f(ak) I< sf2.
Furthermore we have xa; 1 = xy- 1 ya; 1 E U(e) Wak c;;; w;k c;;; UakSO that
lf(x) -f(ak)l < s/2. Combining these two inequalities we get
lf(x) -f(y) I< s.
If we start with a neighbourhood Ua of e such that x- 1 aE Ua implies
lf(x) -f(a) I< s/2, we would obtain a neighbourhood U (e) of e such
that lf(x) -f(y) I< s whenever x-1 y E U(e). Thus taking the intersection of these two U (e)'s as the U (e) in the statement of the Proposition, we complete the proof.
333
11. Tannaka's Duality Theorem
Corollary. A complex-valued continuous function l(x) on a compact

group G is almost periodic on G.
Proof. Let U (e) be the neighbourhood of e given in Proposition 1.
For any a E G, U (e) a is a neighbourhood of a. The compact space G is
covered by the system of open sets U (e) a, a E G, and therefore some
finite subsystem {U(e) a;; i = 1, 2, ... , n} covers G. That is, for any
a E G, there exists some ak with k < n such that aa-,; 1 E U (e). Hence, by
(ax) (akx)- 1 = aa-,;1, we have sup ll(ax)- l(akx) I< s. Similarly, we
can find a finite system {bj, i = "1, 2, ... , m} such that, for any bEG,
there exists some bj with i < m satisfying the inequality
sup \l(a;xb)- l(a;xbi)
i,x
I< s.
Therefore, for any pair a, b of elements E G, we can find ak and bj

(k < n, i < m) such that
sup ll(axb)- l(akxbj) I< 2s.
"
This proves that the system of functions {la,b (x); la,b (x) = l(axb), a and
bEG} is totally bounded by the maximum norm llh II= sup \h(x) I Hence
X
l(x) is almost periodic on G.

We are now able to extend the Peter-Weyl-Neumann Theory of the
preceding section to complex-valued continuous functions l(x) on a
compact group G. For such a function l(x) and s > 0, we set
V =
{Y EG; s~p II (x) -
I (y-1 x) \ < s J.
Then, by the continuity of I, the set V is an open set containing e. By

Urysohn's theorem as applied to the normal space G, we see that there
exists a continuous function k1 (x) defined on G such that
0 < k(x)
< 1 on G, k 1 (e)
= 1
k1 (x) = 0 whenever xE
and
Then the continuous function
k(x) = 2-1 (k1 (x)
+ k1 (x-1))
vc.
(2)
satisfies the condition k (x-1) = k (x) and

0 < k(x) < 1 on G, k(e) = 1 and k(x)
whenever xE uc, where U = VV v- 1 .
0 }
Therefore, if we denote by C (G) the B-space of all complex-valued

continuous functions h(x) defined on G normed by the maximum norm,
we can define a linear operator Ton C (G) into C (G) by
A
(Th) (x) = (k0 Xh) (x), xE G.
(3)
334
Here, as in the preceding section,

(4)
k0 (x) = k(x)fMx(k(x))
and C (G) is the the space C (G) endowed witht the scalar product
(b, h)= My(b(y) h(y)) = (bxh*) (e), h*(y) = h(y-1).
-
~-
(5)
Thus, as in the preceding section, we obtain

Theorem 1. Any complex-valued continuous function I (g) defined on a
compact group G is almost periodic, and I (g) can be approximated uniformly on G by linear combinations of the matrix elements of unitary,
continuous, irreducible matrix representations of G.
We shall say that two matrix representations A 1 (g) and A 2 (g) of a
group G are equivalent if there exists a fixed non-singular matrix B
such that B-1 A 1 (g) B = A 2 (g) for all g E G.
Proposition 2 (I. Schur's lemma). If the representations A 1 (g), A 2 (g)
are irreducible and inequivalent, then there is no matrix B such that
(6)
holds identically in g, except B = 0. In (6), the matrix B is assumed

to be of n 1 rows and n 2 columns, where n 1 , n 2 are the degrees of A 1 (g),
A 2 (g), respectively.
Proof. Let X 1 and X 2 be the linear spaces subject to the linear
transformations A 1 (g) and A 2 (g), respectively. B in (6) can be interpreted as a linear mapping x2 --'>- x1 = B x2 of X 2 onto X 1 . The linear
subspace of X 1 consisting of all vectors x1 of the form B x2 is invariant,
for A dg) x 1 = B x; with x; = A 2 (g) x 2 . By virtue of the irreducibility
of A 1 (g), there are only two possibilities: either B x 2 = 0 for all x2 in X 2 ,
i.e. B = 0, or X 1 = BX2 On the other hand, the set of all vectors x2
in x2 such that B Xz = 0 is an invariant subspace of X2, forB A2 (g) Xz =
Adg) B x 2 = 0. From the irreducibility of A 2 (g) we conclude: either
Bx2 = 0 for all x2 in X 2, i.e. B = 0, or x2 = 0 is the only vector in X 2
such that Bx2 = 0, so that distinct vectors in X 2 go into distinct vectors
in X 1 under the linear mapping B. Hence if B # 0 we conclude that B
defines a one-one linear mapping of X 2 onto X 1 . But this means that B
is a non-singular matrix (n1 = n 2 ) and so A 1 (g) and A 2 (g) would be equivalent.
Proposition 3 (the orthogonality relations). Let A 1 (g) = (afj (g)) and
A 2 (g) = (a~1 (g)) be unitary, continuous, irreducible matrix representations of a compact group G. Then we have the orthogonality relations:
Mg(at(g) ~1 (g))
0, if Adg) is inequivalent to A 2 (g),
(7)
Mg (at (g) a~ 1 (g)) = n1 1 O;k oj1 where n 1 is the degree of AI(G) .
Proof. Let n 1 , n 2 be the degrees of A 1 (g), A 2 (g) respectively. Take any
matrix B of n 1 rows and n 2 columns, and set A(g) =A 1 (g)BA 2 (g-1).
335
Then t11e matrix A = Mg (A (g)) satisfies A 1 (g) A =A A 2 (g). For, we

have, by the invariance of the mean value,
A1 (y) A A 2 (y- 1 ) = Mg (A 1 (y) A 1 (g) B A 2 (g- 1) A 2 (y- 1 ))
Mg (A 1 (yg) BA 2 ((yg)- 1) =A.
'By Schur's lemma, A must be equal to a zero matrix. If we take B = (bj1)

in such a way that only the element biz is not zero, then, by the unitarity
condition A 2 (g-1 ) = A 2 (g)', we obtain
Mg(at(g) a~1 (g)) = 0.

Next we have, as above, A 1 (g) A= A A 1 (g) for A= 211iA 1 (g) B A 1 (g-1 )).
Let ex be any one of the eigenvalues of the matrix A. Then the matrix
(A- exin), where In. denotes the unit matrix of degree n 1 , satisfies
AJ(g) (A --ex In,)
(A -ex I,) AI(g).
Therefore, by Schur's lemma, the matrix (A -ex In) is either non-singular,

or (A- ex In) = 0. The first possibility is excluded since ex is an eigenvalue of A. Thus A = exin. By taking the trace (the sum of the diagonal
elements) of both sides of A = Mg (AJ(g) BAJ(g- 1)), we obtain
n 1 ex = trace(A) = Mg(trace(AI(g) BA 1 (g)- 1) = Mg(trace(B))
trace(B).
Therefore, if we take B = (b11 ) in such a way that b11 = 1 while

the other elements are all zero, then, from Mg (A 1 (g) B Adg- 1 )) = n1 1
trace (B) I 111 , we obtain
Mg (atj (g) a~ 1 (g)) = n1 1 O;k Ojz

Corollary. There exists a set U of mutually inequivalent, continuous,
unitary, irreducible matrix representations U (g) = ( U;j (g)) of G satisfying the following three conditions:
i) for any pair of distinct points gv g2 of G, there exists an U (g) E U

such that U(g1 ) =I= U(g2 ),
ii) if U (g) E U, then the complex confugate representation D (g) of
U (g) also belongs to U,
iii) if U1 (g), U2 (g) are in U, then the product representation Udg) X
U2 (g), explained below, is completely 'reducible to a sum of a
finite number of representations E U.
Proof. The definition of the product representation U 1 (g) X U 2 (g)
is given as follows. Let (ei, et .. . , e~) and (ei, e~, ... , e!) be orthonormal
bases respectively of the finite dimensional complex Hilbert spaces
subject to the linear mappings U1 (g) and U 2 (g), respectively. The product
space of these two Hilbert spaces is spanned by the product base con-
336
sisting of nm vectors e} X e] (i = 1, 2, ... , n; j = 1, 2, ... , m). Referring

to this base, the product representation U1 (g) X U2(g) of U1 (g) =
( u}i (g)) and U 2 (g) = (u~1 (g)) is given by
(Ul (g) X U2 (g)) (e! X ey) =_I u~; (g) u~ (g) (e} X ej).
s,t
Let us take a maximal set U of mutually inequivalent, continuous,

unitary, irreducible matrix representations U (g) satisfying the condition
ii). Then, by Theorem 1, condition i) is satisfied. Condition iii) is also
satisfied since the product representation of two unitary representations
is also unitary and hence is completely reducible.
We are now ready to formulate T. Tannaka's duality theorem. Let 31
be the set of all Fourier polynomials :
.I y}j> ujj> (g) ,

that is, finite linear combinations of ulj> (g), where (ujj> (g)) E U and yjj>
x (g)
denote complex numbers. Then 31 is a ring with the multiplicative unit

u (u (g) _ 1 on G) and with complex multipliers; the sum and the multiplication in the ring 31 being understood as the function sum and the
function multiplication, respectively. Let 'l:: be the set of all linear
homomorphisms T of the ring 31 onto the field of complex numbers such
that
Tu = 1, Tx
Tx, the bar indicating the complex-conjugate.
(8)
'l:: is not void since each g E G induces such a homomorphism Tg:

Tgx = x(g).
(9)
By the condition i) for U, we see that

gl
= g2 implies Tg, =I= Tg,
(10)
Proposition 4. 'l:: may be considered as a group which contains G as

a subgroup.
Proof. We shall define a product T 1 T 2 =Tin 'l:: as follows. Let
u<"'>(g)
(u}j>(g))
(i, j
be members of U. We put, for u~~> (gh) =
I; u}~> (g) u~~> (h),

k
Tu~j> =I; T 1 uh> T 2

k
1, 2, ... , n)
uW.
(11)
By the orthogonal relations (7), we see that the functions u~j> (g)'s, where
(u}j> (g)) E U, are linearly independent on G. Hence, we see that T may
be extended linearly on the whole 31. It is easy to see that the extension T
is also a member of 'l:: and that 'l:: is a group with the unit T. (e = the
unit of G) and r-; 1 = T r' G is isomorphically embedded in this
group 'l:: by the correspondence g--. Tg, as will be seen from (11) and (10).
337

In truth, we have
Theorem 2 (T. TANNAKA). ::t = G, v1z. every T E ::t is equal to a
for all
T x = x (g)
x E ffi .
(12)
Proof. We introduce a weak topology in the group ::t by taking the

sets of the form
{TE %; !Tx,- T.x,l
<
(13)
(i = 1, 2, ... , n)}
as neighbourhoods of the unit T. of the group%. Then %is a compact

space. For, by ,I IT u~i> (g) 12 = (T T) (1) = 1 implied by (8) and (11),
s
%is a closed subset of the topological product of compact spaces:
II fz; lzl <sup 1Txl1,
J
TEX
l
is easily seen that the
It
theorem.
Tychonov's
and so we can apply
one, because a onetopological
a
also
is
Tg
g.embedding
isomorphic
space is a topocompact
a
onto
space
compact
a
of
map
one continuous
logical map. Therefore G may be considered as a closed subgroup of the
compact group %.
By the above weak topology of %, each X (g) E m gives rise to a
continuous function x (T) on the compact group% such that x (Tg) = x (g).
We have only to set x(T) = T x. The set of all these continuous functions x (T) constitutes a ring ffi (%) of complex-valued continuous functions on % satisfying the conditions:
xEIJl
1) 1=u(T)Effi(%),
2) for any pair of distinct points T1 , T 2 of%, there exists an x(T) E
ffi(%) such that x(T1) =I= x(T2 ),
3) for any x(T) E 9!(%), there exists the complex-conjugate function
(T) = x (T) which belongs to ffi (%).
Now let us suppose that %- G is not void. Since a compact space%

is normal, we may apply Urysohn's theorem to the effect that there exist
a point T0 E (%- G) and a continuous function y (T) on ~ such that
y(T) > 0 on %, y(g)
0 on G and y(T0 )
1.
(14)
By the Stone-Weierstrass theorem in Chapter 0, as applied to the ring

ffi(%) satisfying 1), 2) and 3), we see that there exists, for any e > 0, a
function X (g) = .I rl;> uJf> (g) E m such that Iy (T) - I rt'> u~;> (T)) < e
on %, and, in particular, Iy (g) - .I r1.i> u1j> (g) I < e on G. Let ut'l> (g)
= u (g) = 1. Then, by taking the mean value and remembering the orthogonality relation (7), as applied to the compact groups% and G, we obtain
IMT(y(T)) -rii'>l
22
< e,
IMg(y(g)) -yl.i> I< e.
(15)
338
We have thus arrived at a contradiction, since MT(y(T))
M 1 (y(g)) = 0 by (14).
>
0 and
Remark 1. The above proof of Tannaka's theorem is adapted from

K. YosiDA [14]. The original proof is given in T. TANNAKA [1]. Since a
continuous, unitary and irreducible matrix representation of a compact
abelian group G is precisely a continuous function X (g) on G satisfying
x(gt) x(g2) = x(gl g2) and fx(g) I= 1,
Tannaka's theorem contains, as a special case, the duality theorem of
L. PoNTRJAGIN [1]. For further references, seeM. A. NAIMARK [1].
Remark 2. To define mean values of continuous functions /,.(g) (k =
1, 2, ... , m) on G by the method given in Chapter VIII, 5, we have only
to replace the dis (gv g2 ) there by dis (g1 , g2 ) =
sup
I/,. (g g1 h) /,. (g g2 h) I
g,loG;k=1,2, .. ,m
12. Functions of a Self-adjoint Operator
Let H = A. dE (A.) be the spectral resolution of a self-adjoint operator

H in a Hilbert space X. For a complex-valued Baire function /(A.), we
consider the set
D(f(H)) = {xEX; _[it(A.) l2 d IIE(A.) xW
< oo},
(1)
where the integral is taken with respect to the Baire measure m determined by m((A.l> J.2]) = liE(~) x 11 2 -liE(~) x 11 2 As in the case of a
continuous function /(A.), treated in Chapter XI, 5, we see that the integral
<X)
-oo
f(A.) d (E (A.) x, y), xED (t(H)), y EX,
(2)
exists and is finite with respect to the Baire measure m determined by

y)- (E(A.1) x, y). Further we see that (2) gives the
complex conjugate of a bounded linear functional of y. Hence we may
write (2) as (!(H) x, y) by virtue of F. Riesz' representation theorem in
Chapter III, 6. In this way, we can define functions of H:
m ((A.1 , A.2 ]) = (E (~) x,
00
/(H)=
through (1) and (2).
-oo
f(A.)dE(A.)
Example 1. If His bounded self-adjoint and /(A.)

as in Chapter XI, 5,
f(H)
oo
f
-oo
f().) dE().)
= .~ 1XjH1
J=l
(3)
.~ IX1A.1 ,
J=l
then,
339
Example 2. If I (A.) = (A. - i) (A. + i)-1 , then I (H) is equal to the

Cayley transform U H of H. We have, in this case, D(f(H)) =X since
II (A.) I = 1 for real A.. It is easy to see, by E (A.1) E (A.2) = E (min (AI> ~)),
that if we apply the bounded self-adjoint operator (H- il)-1 =
J (A- i)-1 dE (A.) to the operator I (H) = JI (A.) dE (A.) then the result is
(A+ i)-1 dE(A.). Hence I(H) = UH.
equal to (H + iJ)-1 =
Example 3. As in Chapter XI, 5, we have
00
III(H)xll2=
J II(A.)I2diiE(A.)xll2 wh((never
-oo
xED(f(H)).
(4)
Definition. Let A be a not necessarily bounded linear operator, and

B a bounded linear operator in a Hilbert space. If
(5)
xED(A) implies BxED(A) and ABx=BAx,
that is, if AB ~BA, then we shall write BE (A)' and say that B is
commutative with A. We shall thus write the totality of bounded linear
operators B commutative with A by (A)'.
Theorem 1. For a function I(H) of a self-adjoint operator H =
JAdE (A.) in a Hilbert space X, we have
(!(H))'
(6)
~(H)',
that is, f (H) is commutative with every bounded linear operator which
commutes with H. (Since E (A) E (H)' by Theorem 2 in Chapter XI, 5, we
see, in particular, that I(H) is commutative with every E (A).)
Proof. Suppose S E (H)'. Then we can show that S is commutative
with every E (A.). We first show that S is commutative with the Cayley
transform U H of H. For, if x E D (H), then we have, by S E (H)',
+ il) x = (H + il) Sx, (H -if) Sx = S(H -il) x.

By putting (H + il) x = y, we see, from the first of the above relations,
S(H
that
(H+ii)- 1 Sy=S(H+ii)-1 y
Hence
S(H- il) (H
+ ii)-1 =
(H- il) (H
forall
+ ii)-1 5,
Therefore, Sis commutative with (UHY' =
and so
eniiJ
d (SF (0)
X,
y) =
(s j
0
eniiJ
yEX=R(H+ii).
dF (0)
that is, S UH= UHS.
2n
eniiJ
dF (0) (n
X,
y) =
eniiJ
0,
d(F (0)
1, ... )
s x, y).
Hence, as in the proof of the uniqueness of the spectral resolution of a

unitary operator, given in Chapter XI, 4, we obtain SF(O) = F(O) S.
This proves SE(I,) = E (A.) S since E(-cot O)=F(O). Thus, for xED(f(H)),
22*
340
we obtain
(5 J j(A) dE(A)
X,
y)
J/(A) d(5E(A)
X,
y)
J/(A) d(E(A) 5
X,
y)'
that is, 5f(H) ~ f(H) 5.

The above Theorem 1 admits a converse in the form of
Theorem 2 (NEUMANN-Rmsz-MIMURA). Let H be a self-adjoint operator in a separable Hilbert space X. LetT be a closed linear operator in X
such that D (Tt = X. Then a necessary and sufficient condition in order
that T be a function f(H) of H with an everywhere finite Baire function
f (A) is that
(T)' ~ (H)'.
(7)
Proof. We have only to show that condition (7) is sufficient. We may
assume that His a bounded self-adjoint operator. If His not bounded,
then we consider H 1 = tan-I H. Since Itan-I AI < n/2, we easily see that
H 1 is bounded self-adjoint. By Theorem 1, the operator H =tan HI is
commutative with every operator E (HI)'. Thus, by hypothesis, we have
(T)' ~ (H)' ~ (HI)' .
Therefore, if Theorem 2 is proved for bounded Hv then T =/I (HI) =

/I (tan-I H) = / 2 (H), where / 2 (A) = /I (tan-I A).
We may thus assume that His bounded and self-adjoint.
The first step. For any fixed x 0 E D (T), we can find a Baire function
= F (H) x 0 This may be proved as follows. Let
M (x0 ) be the smallest closed linear subspace of X spanned by x 0 , H x 0 ,
H 2 x 0 , , H"x0 , We denote by L the projection upon M(x0 ). Then
(T)' ?JL. For, by HM(x0 ) ~ M(x0 ), we obtain HL = LHL and so
LH = (H L)* = (LH L)* = LH L = H L, that is, L E (H)'. Hence, by
hypothesis, (T)' 3 L.
Therefore T x 0 = T L x 0 = L T x 0 E M (x0 ), and so there exists a
F (A) such that T x 0
sequence {p,. (A)} of polynomials such that

T x0
s-lim
~
p,. (H)
x0
(8)
Hence, by (4), we obtain

00
I!P,.(H)
Xo-
Pm(H)
Xo 11 2
J IP,.(A)- Pm(A) 1 d IIE(A)

2
Xo 11 2
-00
As in the proof of the completeness of L2(5, 58, m), we see that there
exists a Baire function F (A) which is square-integrable with respect to
the measure m determined by m( (A1 , ~]) = II E (~) x0 11 2 - II E (A1) x0 11 2
such that
00
lim
J IF(A)- p,.(A) 1 d IIE(A) x
~-00
0 ll 2
0.
341
Hence, for F(H), we have

lim
fl-->00
00
II(F(H)- p,. (H)) x0 112 =
lim
fl-->00
J IF().) -p,.().) 12 d liE().) x0 112 =

-oo
0.
This proves that T x 0 = F (H) x0 Since F ().) is finite almost everywhere

with respect to the measure m determined by m((~. ~]) =
II E (~) x0 11 2 - II E (~} x0 11 2, we may assume that F (A) is a Baire function
which takes a finite value at every).; we may put F().) = 0 for those;.
for which IF().) I= oo.
The second step. Since X is separable and D (T)a = X, we may choose
a countable sequence {g,.} ~ D (T) such that {g,.} is strongly dense in X.
We set
It
n-1
= g1,/2 = g2- L1 g2, . . ,/,. = g,.- .E Lkg,., where L 11

k=l
(9)
is the projection on the closed linear subspace M (/11).

By the first step, we have (T)' 3 L 11 and so
L 11 g,. ED (T) which implies/,. ED (T)
(n = 1, 2, ... ) .
(10)
We may show that
(11)
and
00
(12)
I= .ELk.
k=l
Suppose that (11) is proved fori, k < n. Then, fori< n,

Ld,. = L;,g,.- L;,
(".i L g,.) = L,g,.- Lig,. =

1
11=1
11
L;,g,.- L;,g,.
0,
L,H11' /,. = H 11' Ld,. = 0.
Thus M (!,.) is orthogonal to M (!,). This proves L;,L,. = L,.L;, = 0.

00
Next we put .E L 11 = P and show that Pg,. = g,. (n = 1, 2, ... ).

11=1
Then, since {g,.} is dense in X, we obtain P =I. But, by (9), we have
n-1
Pg,.=P/,.+ .E PLkg,..
11=1
We also have Pf,. =/,.by /,.E M(f,.). Hence, by PL 11 = L 11 implied by

(11}, we obtain Pg,. = g,. (n = 1, 2, ... ).
The third step. Take a sequence {c,.} of positive numbers such that
k
s-lim .E c,.j,., s-lim .E c,.T/,.

k->oo n=1
k->oo n=1
both exist. For instance, we may take c,. = 2-"

Since Tis a closed operator, we have
{11/,.11 + IIT/,.JI)-1
00
00
n=1
n=1
x0 = .E c,./,.E D(T} and Tx0 =Yo= .E c,.Tj,..
(13)
342
Hence, by the first step,
Tx0
= F(H) x0
(14)
Let BE (H)' be a bounded self-adjoint operator. Then, by hypothesis,

BE (T)'. By Theorem 1, F(H) is commutative with B. Henc-e
F(H) Bx0
= BF(H) x0 = BTx0 = T Bx0
Let en(.A.) be the defining function of the set {.A.; IF(A.)
c;;; 1 PnHk Lm,
where
Pn
(15)
I< n}, and put
= en(H).
Then we can show that

(16)
In fact, we have LmXo
Hence, by (15),
F(H) P,.Hkfm
00
,I c,.Lmfn
n=l
= cmfm, by (11) and fm E M(fm)
= F(H) c;;; 1 PnHk Lmxo = F(H) Bx0 = T Bx0

= Tc;;/P,.HkLmxo = TP,.Hkfm,
that is, for h spanned by Hk fm with fixed m, we have
F(H) P,.h = TP,.h.
(16')
But such h's are dense in M (fm) and so, by (12), we see that if we let m
take all positive integers then the h's are dense in X. Hence we have proved (16') for those h's which are dense in X.
Now P,. is bounded by (4). By the operational calculus given below,
the operator F (H) P,. is equal to the function F., (H) where
F,.(.A.) = F(A.) e,.(.A.) = F(.A.), for IF(.A.)

= 0, for IF(A.)
I> n.
I< n,
Thus F,. (H) = F (H) P n is bounded.

Let h* E X be arbitrary, and let h* = s-.lim hi where hjs are linear
J--+010
combinations of elements of the form Hk fm Such a choice of hi is possible

as proved above. By the continuity of the operator F(H)P,., we have
F(H)Pnh* = sj~ F(H)P,.hi.

Hence, by s-.lim P,.hi
J--+010
= P,.h* and (16'), we see that the closed operator T
satisfies (16).
The fourth step. Let yE D (F (H)) and set y,. = P,.y. Since F (.A.) is finite
everywhere, we must have s-lim P,. =I. Thus s-lim Yn =s-lim PnY = y.
Hence, by (16), we obtain T
n--+ao
n~
F(H). For, s-lim F(H) Pny =s-lim Fn(H)
y = F(H) y whenever yE D(F(H)).
n---+00
n~oo
343
Let y ED (T) and set Yn = P nY Then
Tyn = TPnY = F(H) Pny (by (16)),

TPnY = PnTY (by Pn = en(H) E (H)').
By the operational calculus given below, the function F(H) of H is a
closed operator. Hence, letting n-+ oo in the above relations, we see
that F(H) ~ T. We have thus proved that T = F(H).
An Operational Calculus. We have
Theorem 3. (i) Let /(A.) be the complex-conjugate function of f(A.).
Then DCf(H)) = D(f(H)) and, for x, yE D(f(H)) = D(f(H)), we have
(/(H) x, y)
= (x, !{H) y).
(17)
f(A.)g(A.)d(E(A.)x,y).
(18)
(ii) If xED (/(H)), y ED (g(H)), then

00
(I(H)x,g(H)y)=
-00
(iii) (rxf) (H) x
= rxf(H) x if xED(f(H)).
If xE D(f(H))f\D(g(H)), then
(f +g) (H) X= f(H) X+ g(H) X.
(19)
(iv) If x E D (I (H)), then the condition f (H) x E D (g (H)) is equivalent

to the condition x E D (I g (H)), where f g (A.) = f (A.) g (A.), and we have
g(H) f(H) x = (g f) (H) x.
(20)
(v) If f(A.) is finite everywhere, then f(H) is a normal operator and
f(H)*
= f(H).
(21)
In particular, f(H) is self-adjoint if /(A) is real-valued and finite everywhere.

Proof. (i) D (f(H)) = D (f(H)) is clear and
J f(A.)d(E(A.)x,y)= J f(A.)d(x,E(A.)y)
00
(f(H)x,y)=
00
-00
-00
= (f(H) y, x) = (x,l(H) y).
(ii) We know, by Theorem 1, that E (A.) is commutative with g(H). Thus

00
(f(H)x,g(H)y)=
-oo
00
f(A.)d(E(A.)x,g(H)y)=
j f(A.)d(x,E(A.)g(H)y)
-oo
= _[ f(A.) d (g(H) E (A.) y, x) = _[ f(A.) d (_[ g(p,) d(E (It) E (A.) y, x))
=
_L f(A.) d
oo
_[ g(p,) d(y, E (It) x)
_L f(A.) g(A.) d (E (A.) x, y).

oo
344
XI. N ormed Rings and Spectral Representation

00
(iii) is clear. (iv) Let x satisfy
J /f(A.)/
d JJE(.A.)xJJ 2 <=. Then, by
-00
00
E (A.) E (p,) = E (min (.A.,p,)), the condition
J Jg (A.) J2d JJ E (A.) f (H) x J/ 2< =

-oo
implies, by virtue of the commutativity of E(A.) with f(H),
=>
00
f
-oo
= _[Jg(.A.) /2 d
= _[Jg(.A.) J2 d
00
Jg(.A.)J2dJJE(.A.)/(H)xJJ2=
f
-oo
Jg(.A.)J2dJJ/(H)E(.A.)xJJ2
(_[l!Cu) /2 d //E (p,) E(A.) x 2)

/1
CL
/f(p,) J2 d JJE(p,) xJJ 2) = _[Jg(.A.) f(A.) J2d JJE(.A.) xJJ2.
Since the above calculation may be traced conversely, we see that, under
the hypothesis xE D(f(H)), the two conditions f(H) xE D(g(H)) and
x E D(f g(H)) are equivalent and we have
00
(g(H) f(H)x,y) =
g(A.) d(E(A.) /(H) x,y)
-00
= _[ g(A.) d
(_Lt(p,) d(E(p,)
00
g(A.) /{A.) d(E(A.) x, y)
X,
y))
= ((g f) (H)
:X,
y).
-00
(v) Let us put h(A.) = /f(A.) J +IX, k(A.) = h(.A.)-1, g(.A.) = f(A.) h(.A.)-1,
where IX is any positive integer. Then k (A.) and g (A.) are both bounded
functions. Hence D(k(H)) = D(g(H)) =X. Thus, by (iv),
f(H) = h(H) g(H) = g(H) h(H).
{22)
We have, by (i) and D(k(H)) =X, k(H)* = k(H), i.e., k(H) is selfadjoint. By (iv) we have x = h (H) k (H) x for all x EX and x = k (H) h (H) x
for all xE D(h(H)). Hence h(H) = k(H)-1 . Thus, by Theorem 1 in
Chapter VII, 3, h(H) is self-adjoint. Therefore D(f(H)) = D(h(H)) is
dense in X and so we may define/(H)*. We shall show that f(H)* = l(H).
Let a pair {y, y*} of elements EX be such that (f(H) x, y) = (x, y*) for
all x E D(f(H)). Then, by g(H)* = g(H) (implied by (i)) and (22),
(f(H)
X,
y) = (g(H) h(H)
X,
y)
= (h(H)
X,
g(H) y).
Hence, by xE D(f(H)) = D(h(H)) and the self-adjointness of h(H),

we obtain
g(H) y E D(h(H)) and h(H)g(H) y = y*.
Thus, again by (22), we obtain l(H) y = y* so that f(H)* ={(H). Therefore, by (iv), we see that f(H) is normal, that is,f(H): f(H) = f(H) f(H)*.
345
13. Stone's Theorem and Bochner's Theorem
Corollary. If /(A) is finite everywhere, then f(H) is closed.

Proof. Clear from /(H)**= /(H)*= f(H) = f(H).
A historical note. Theorem 2 was first proved by]. VON NEUMANN
[7] for the case of a bounded self-adjoint operator T. Cf. F. RIEsz [5].
The general case of a closed linear operator T was proved by Y. MIMURA
[2]. The exposition given above is adapted from Y. MIMURA [2] and
B. Sz. NAGY [1].
13. Stone's Theorem and Bochner's Theorem

As an example of functions of a self-adjoint operator, we give
Theorem 1 (M. H. STONE). Let {U1}, -oo < t < oo, be a one-parameter group of class (C0) of unitary operators in a Hilbert space X. Then
U1 = f1 (H), where !t{A.) = exp(it}..) and iH =A, H*

is the the infinitesimal generator of the group U1
= H,
(1)
Conversely, for any self-adjoint operator H in X, U1 = f1 (H) defines a

one-parameter group of class (C0) of unitary operators.
Proof. We have, by the representation theorem of the semi-group
theory,
U1x =s-lim exp (tiH (I- n-1iH)-1) x.
n->00
Since the function g (t) = exp (tiA. (1 - n-1i}..)-1) is smaller than

AdE(A.),
exp ((-ntA. 2 )f(n2 + }..2)) in absolute value, we have, for H =
exp(tiH(I -n-1iH)-1) =
and, moreover,
~-L
00
-oo
exp(1
t~~liJ..)dE(A.)
12 d IIE(A.) xll2
t 'l
Iexp(1 _:-IiJ..)-exp(tiA.)
~_[jexp(n
e:
J-11 2 d liE(}..) xll 2 = 0.
J exp(it}..) dE(A.).
-oo
00
This proves that U1 = / 1 (H) =
For the converse part of Theorem 1, we observe that, by the operational calculus of the preceding section,
fe(H)* = f_,(H)
and /,(H) !.(H)= IHs(H), fo(H) =I.
We also have the strong continuity of / 1 {H} at t = 0 by

l!f,(H)x-xll 2 =
00
J lexp(itA)-1I 2 diiE(A)xJI 2-+0

-oo
as t-+0.
Hence U1 = / 1 (H) is a one-parameter group of class (C0) of unitary

operators.
346
Remark. For the original proof, see M. H. STONE [2], Cf. also J. VON
NEUMANN [8]. Another proof given by E. HoPF [1] is based on a
theorem due to S. BocHNER:
Theorem 2 (BoCHNER). A complex-valued continuous function f(t),
- = < t < =, is representable as
f (t)
Ji
00
1A dv
(.?.) with a non-decreasing,
-00
right-continuous bounded function v (.?.),
(2)
iff I (t) is positive definite in the following sense:

00
00
J J f(t- s) q;(t) q;(s) dt ds >

-oo -oo
for every continuous function q; with compact support.
I (t)
(3)
The proof of Theorem 1 given by E. HoPF starts with the fact that
= (U1 x, x) satisfies condition (3) as may be seen from
J J (U
-oo -oo
00
00
00
1_ 5 X,
x) q; (t) q; (s) dt ds =
= (_[ q; (t)
00
J J (U x, U x) q; (t) q; (s) dt ds
-oo -oo
ut X dt,
_[ q; (s) Usxds) :2: 0.
We shall show that Bochner's theorem is a consequence of Stone's theorem.
Deduction of Theorem 2 from Theorem I. Consider the totality

iJ of complex-valued functions x(t), - = < t <=,such that x(t) = 0
except possibly for a finite set of values oft; the finite set may vary with
x. iJ is a pre-Hilbert space by
(x
+ y) (t) =
(x,y)=
x(t)
~
-oo<t,s<oo
+ y(t),
((xx) (t)
,xx(t) and
f(t-s)x(t)y(s)forx,yEiJ,
(4)
excepting the axiom that (x, x) = 0 implies x = 0. That (x, x) > 0 for
any x E iJ is a simple consequence of the positive definiteness of the function f (t).
Let us set in = {x E iJ; (x, x) = 0}. Then the factor space iJ/in is a
pre-Hilbert space with respect to the scalar product (x, y) = (x, y) where
xis the residue class mod in containing x E iJ. Let X be the completion
of the pre-Hilbert space X= iJfin. The operator Ur defined by
(Urx) (t) = x(t
+ r),
xE iJ,
(5)
surely satisfies the conditions
(Urx, Ury) = (x, y), UrUa = Ur+a and U0 =I.
(6)
347
14. A Canonical Form of a Self-adjoint Operator
Therefore, it is easy to see that {U.,.} naturally defines a unitary operator

U7 in X in such a way that {U,}, -oo < T < oo, constitutes a oneparameter semi-group of class (C0) of unitary operators in X; the strong
continuity in t of U1 follows from the continuity of the function f(t).
00
Hence, by Stone's theorem,
U= f
1
-oo
eiti. dE
by x0 (r) = 1 and x0 (t) = 0 whenever t

/(r) = (U7 x0 , x0 ). Therefore,
/(r)
00
eii.
d IIE(A.)
(A.). Let x0 (t) E g: be defined
T.
Then, by (4) and (5),
:Xoll 2 ,
-00
which proves Bochner's theorem.

Remark. The idea of using a positive definite function to define a preHilbert space as in (4) was systematically applied by B. Sz. NAGY [3] to
various interesting problems concerning the Hilbert space.
14. A Canonical Form of a Self-adjoint Operator

with Simple Spectrum
Let H = A. dE (A.) be a self-adjoint operator in a Hilbert space X

with simple spectrum as defined in Chapter XI, 8. Thus there exists an
element y EX such that the set {(E (!'1)- E (<X)) y; <X< {3} spans a dense
linear subspace of X. We put
a (A.)
(1)
(E (A.) y, y).
Then a(A.) is monotone non-decreasing, right-continuous and bounded.

We shall denote by a(B) the Baire measure determined on Baire sets
on RI from a((a, b]) = a(b)- a(a). We denote by L~(-oo, oo) the totality of complex-valued Baire-measurable functions f (A.), - oo < A. < oo,
such that
11/l\a =
(_[ \f(A.) \2 a(dA.)r
< oo.
ThenL;(-oo,oo) is a Hiloo
bert space by the scalar product (/, g)a
= J f(A.) g(A.) a(dA.)

-oo
f (A.) = g (A.) a-a.e.
with the
convention to consider f = g in L; iff

Theorem. With any /(A.) E L;(-oo, oo) we associate a vector
X defined by
f of
00
i = -oof /(A.) dE(A.) y.
(2)
Then the correspondence f (A.) -+ f is a one-one linear isometric mapping

of L;(-oo, oo) onto X. Let this mapping be denoted by V, i.e.J = Vf.
Then the operator H 1 = v- 1 HV i11 L;(-oo,oo) is precisely the
348
XI. Normed and Rings Spectral Representation
operator of multiplication by A.:

D(H1 ) = D(V- 1 HV) ={/(A.); /(A.) and A./(A.) both E L!(-oo,oo)}
and (H1 f) (A.) =A./(A.) whenever /(A.)ED(H1 ).
(3)
Proof. We have, by E(A.) E(p) = E(min(A.,,u)),

00
(E(A.)y})=
=
and so
-oo
A
00
/(p)d,.(li(A.)y,E(p)y)=
-oo
f(p)d,.(E(p)E(A.)y,y)
J /(,u) d(E(p) y, 'Y) = -ooJ f(p) a(d,u),

-oo
00
(i, g)=
00
J f(A.) d(E (A.) y, i) = J /(A.) g(A.) a(dA.) =
-00
(/, g)a
(4)
-00
Therefore, V maps D(V) = L~(-oo, oo) onto {iJ =

/EL~ (-oo,oo)}
00
J f(A.) dE(A.) y.
-oo
one-to-one, linearly and isometrically. Hence, in particular.

R (V) is a closed linear subspace of X. But R (V) surely contains the
{J
J dE (A.) y =
elements of the form
(E ({J) - E (<X)) y,- oo
< IX< fJ < oo,
"'
and so, by the hypothesis
that the spectrum of H is simple, we see that
R (V) = R (V),. = X. Thus the first half of the Theorem is proved.

Next we have
00
E(A.)i=E(A.)
A
= j
and so, by (4),
-oo
-oo
00
f(p)dE(p)y=
f(p)d,.(E(A.)E(p)y)
-00
/(,u)dE(,u)y,
A
(E(A.)
f. g)= J f(p) g(p) a(d,u).

-oo
(5)
J A. d(E(A.)i.i)<oo, equivalent to {ED(H).

-oo
is equivalent to the condition J A. J/(A.) J a(d.A.) < oo. Moreover, in the
-oo
00
Hence the condition
00
last case, we have, by (20) in Chapter XI, 12,

H VI = H
j=
00
-oo
A. dE (A.)
i,
and hence, by (4) and (5),

(Htf. g)a = (V-1 HVf, g)a = (HV /, V g)= (H j, g)
00
J A.d(E(A.)i.i) = -ooJ A.f(A.)g(A.)a(dA.).

-oo
00
15. The Defect Indices of a Symmetric Operator
349
On the other hand, we have

00
(H1/, g),=
Therefore we must have
-oo
(Hd) (A) g(.i.) a(dA).
(H1 /) (A) = A/ (A) a-ahnost everywhere.
Remark. There is a close connection between self-adjoint operators

with simple spectrum and the Jacobi matrices. See M. H. STONE [1],
p. 275. For a canonical form of a self-adjoint operator with not necessarily
simple spectrum, see J. von Neumann's reduction theory: NEUMANN [9].
15. The Defect Indices of a Symmetric Operator.
The Generalized Resdlution of the Identity
Definition 1. Let U = U H = (H- if) (H + ii)-1 be the Cayley

transform of a closed symmetric operator H in a Hilbert space X. Let
Xfi = D(U8 )l., Xii = R(U8 )l., and let m = dim(Xfi), n = dim{X.H)
be the dimension numbers of Xfi, X.H respectively. Then His said to be
of the defect indices (m, n). H is self-adjoint iff it is of the defect indices
(0, 0) (see Chapter VII, 4).
Proposition 1. The defect indices (m, n) of a closed symmetric operator
H may be defined as follows: m is the dimension number of the linear
subspace {x E X; H* x = i x}; n is the dimension number of the linear
subspace {xE X; H*x = - ix}.
Proof. Clear from Theorem 3 in Chapter VII, 4.
Example 1. Let X= L 2 (0, 1). Let D be the totality of absolutely
continuous functions x(t) E L 2 {0, 1) such that x(O) = x(1) = 0 and
x'(t) E L2{0, 1). Then the operator T 1 defined by T 1 x(t) = i-1 x' (t) on
D = D (T1) is of the defect indices (1, 1).
Proof. As was shown in Example 4 of Chapter VII, 3, T~ = T 2 is
defined by
T 2 x (t) = i-1 x' (t) on D (T2) = {x (t) E L2 (0, 1); x (t) is
absolutely continuous such that x' (t) E L 2 (0, 1)},
Thus the solution y E L 2 (0, 1) of Tty= T 2 y = iy is a distribution solution of the differential equation
y' (t) = - y(t) (y, y' E L2(0, 1)).
{1)
Then z (t) = y (t) exp (t) is a distribution solution of the differential equation
{2)
z'(t) = 0 (z,z'EL2(0, 1)).
We shall show that there exists a constant C such that z(t) = C for a.e.
t E (0, 1). To this purpose, take any function x0 (t) E ego {0, 1) such that
350
J x0 (t) dt =
1 and put
x(t) - x0 (t)
where
by
J x(t) dt = u(t),
w(t)
J u(s) ds,
(t) is an arbitrary function from ego (0, 1). We have wE ego (0, 1)
J u(s) ds =
O.'Therefore, by (2}, we have

1
- J z (t) w' (t) dt =

0
that is,
1
J z (t) u (t) dt =
J z(t) x(t) dt = e J x(t) dt,

0
0
where e
0,
J z(t) x
(t) dt.
This proves, by the arbitrariness of x(t) E ego(O, 1), that z(t) = e for a.e.
tE(0,1).
Hence any solution of T*y = iy is of the form y(t) = e exp(-t).
In the same manner we see that any solution of T*y = - i y is of the
form y(t) = e exp(t). Thus Tis of the defect indices (1, 1).
Definition 2. A symmetric operator H in a Hilbert space X is called
maximal symmetric if there is no proper symmetric extension of H.
Proposition 2. A maximal symmetric operator His closed and H = H**.
A self-adjoint operator His maximal symmetric.
Proof. By Proposition 1 in Chapter VII, 3, H** is a closed symmetric
extension of H. Thus the first half of Proposition 2 is clear. Let H 0 be a
symmetric extension of a self-adjoint operator H. Then, from H ~ H 0 ,
H 0 ~ Hti, we obtain H 0 ~ Hti ~ H* and so, by H =H*, H ~ H 0 ~H.
This proves that a self-adjoint operator His maximal symmetric.
Corollary 1. Every maximal symmetric extension H 0 of a given symmetric operator His also an extension of H**.
Proof. The relation H ~ H0 implies the relation H3 ~ H* andH** ~ H3*.
Since, by the preceding proposition, H 0 = H6*, Corollary 1 is true.
Corollary 2. If H is a symmetric operator such that H* = H**, then
the self-adjoint operator H* is the only maximal symmetric extension
of H.
Proof. Being self-adjoint, H** is maximal symmetric. Thus any maximal symmetric extension H 0 of H, which is also a symmetric extension of
H** by Corollary 1, is identical to H** = H*.
We are thus in a position to state
Definition 3. A symmetric operator H such that H* = H** is said to
be essentially self-adfoint. A self-adjoint operator His said to be hypermaximal. The latter terminology is due to]. VON NEUMANN.
15. The Defect Indices of a Symmetric Opera.tor
351
Example 2. Let X= L 2(--oo, oo), and define an operator H by

ll x(t) = t x (t) for x (t) E cg (- oo, oo). H is surely a symmetric operator
in X. It is easy to see that H* is the coordinate operator defined in
Example 2 in Chapter VII, 3 so that H is essentially self-adjoint. The
operator H defined by H x (t) = i-1 x' (t) for x (t) E Cij (- oo, oo) is also
essentially self-adjoint in X= L2(-oo,oo). For, in this case, H* is the
momentum operator defined in Example 3 in Chapter VII, 3.
Theorem 1. Let the defect indices (m, n) of a closed symmetric operator H satisfy
m = m' + p, n = n' + p (p > 0).
Then there exists a closed symmetric extension H' of H with the defect
indices (m', n').
Proof. Let {IPv !pz, . , IPP IPP+I , IPP+m'}, {tpv tpz, , tpp, tpp+I
... , tpp+n'} be complete orthonormal systems of Xfi = D(UH)l., X}i =
R(UH)l., respectively. Define an isometric extension V of UH by
Vx = UHx
for
p
~(XiiPi
v. i=1
xE D(UH),
p
~(Xitpi
i=1
We have R(I- UH)a =X by Theorem 1 in Chapter VII, 4. Thus, by

R (I - V) ~ R (I- UH) and Theorem 2 in Chapter VII, 4, there exists a
uniquely determined closed symmetric extension H' of H such that
V = (H'- if) (H' + iJ)-1 . The defect indices of H' are (m', n') as may
be seen from dim(D{V).l.) = m', dim(R(V).l.) = n'.
Corollary. A closed symmetric operator H with the defect indices
(m, n) is maximal symmetric iff either m = 0 or n = 0.
Proof. The "only if" part is clear from Theorem 1. If, for instance,
m = 0, then, by D(UH) =X, we must have uH. = UH for a closed
symmetric extension H 0 of H. This proves H 0 = i (I + U H.) (I- UH.)-1
= i(I + UH) (I- UH)-1 =H. Also, in the case n = 0, we see that there
is no proper closed symmetric extension of H.
Example 3. Suppose that {IPv !pz, ... , qJ,., . } is a complete orthonormal system of a (separable) Hilbert space X. Then, by
00
00
00
=1
=1
U ..I (XIP = ..I (XiiPi+I ..I

=1
\()(, \2 < oo
we define a closed isometric operator U such that D (U) =X and

dim(R(U).l.) = 1. If R(I- U)a# X, then there exists an x# 0 such
that xE R(I -U)l.. Consequently ((I- U) x, x) = 0 and so (Ux, x) =
llxll 2 =II Uxll 2 This implies that
I\(I-U)xll2= J\xi\2-(Ux,x)-(x, Ux) + 1\UxJ\ 2
= JJxJJ2-I\xl\2-l\xiJ2
+ llx\12 =
0,
352
that is, U x = x and so, by the above definition of U, x must be zero. This
contradiction shows that R (I- U)a = X. Hence, by Theorem 2 in
Chapter VII, 4, U is the Cayley transform of a closed symmetric operator
H. H is of the defect indices (0, 1) since D ( U) = X and R ( U) .L is spanned
by g;1 . Thus His maximal symmetric without being self-adjoint.
Theorem 2 (M.A. NAIMARK [3]). Let a closed symmetric operator H 1
in a Hilbert space X 1 be of the defect indices (m, n). Then we can construct a Hilbert space X, containing X 1 as a closed linear subspace, and
a closed symmetric operator H in X with the defect indices (m + n,
m + n) such that
H 1 = P(X1) HP(X1 ), where P(X1) is the projection of X onto X 1
Proof. Consider a Hilbert space X 2 of the same dimensionality as X 1 .
We construct a closed symmetric operator H 2 in X 2 with the defect
indices (n, m). For instance, we may take H 2 = - H 1 supposing that X 2
coincides with xl. Then we would have {x E X2; Hi X= ix} = {x E Xl;
Hfx = - ix}, {x E X 2; Hix = - ix} = {xE X 1 ; Hfx = ix} and so the
defect indices of H 2 must be (n, m).
We then consider an operator H defined by
H {x, y} = {H1 x, H 2 y} for {x, y} E X 1 xX2 with xED (H1), y ED (H2 ).
It is easy to see that H is a closed symmetric operator in the product
Hilbert space X= X 1 xX2. The condition
H*{x, y} = i{x, y} (or the condition H*{x, y} = - i{x, y})
means that Hix = ix, Hiy = iy (or Hfx = - ix, Hiy = - iy). Hence
we see that the defect indices of Hare (m + n, m + n).
Corollary. Let, by Theorem 1, fi be a self-adjoint extension of H,
and let
= A. dE (A.) be the spectral resolution of fi. Since H and a
fortiori H are extensions of H 1 when H 1 is considered as an operator in
X, we have the result:
fj J
if
xED(H1)~~1 =P(X1 )~,
then
x=P~X1 )xED(H)
J A. dF(A.) x,
H 1 x = P(X1) H x=P(X1) H P(X1) x =
and'
where
(3)
-00
F(A.) = P(Xl) E(A.) P(Xl).

The system {F(A.); -=<A.<=} surely satisfies the conditions:
F (A.) is a self-adjoint operator in X 1 ,
A.1 < A.2 implies that (F (A.1 ) x, x)
F(A.
+ 0) = F(A.),
(F (A. 2) x, x) for every xE X 1 ,
F(-=) x =s-lim F(A.) x = 0,

.\~-00
F(=)x=s-limF(A.)x=x forall xEX .

.\too
(4)
353
15. The Defect Indices of a Symmetric Operator
Remark 1. For the closed symmetric operator Hv we have
J A.dF(A.)x
00
H 1 x=
for all xED(H1},
-00
where {F(A.);-oo<A.<oo} satisfies (4). In this sense, H 1 admits a

generalized spectral resolution.
Example 4. Let X 1 = L2(-oo, 0), and let D 1 be the totality of
absolutely continuous functions x(t) E L2(-oo, 0) such that x(O) = 0
and x'(t)EL2(-oo,O). Then the operator H 1 defined by H 1 x(t)=
i-1 x' (t) on D 1 = D (H1 ) is a maximal symmetric operator with the defect
indices (0, 1). This we see as in the above Example 1. Let X 2 = L2(0, oo),
and let D 2 be the totality of absolutely continuous functions x (t) E L2 (O,oo)
such that x(O) = 0 and x' (t) E L2(0, oo). Then the operator H 2 defined by
H 2 x (t) = i-1 x' (t) on D 2 = D (H2 ) is a maximal symmetric operator with
the defect indices (1, 0). In this case, X= X 1 xX2 = L2(-oo,oo), and
the operator H in Theorem 2 is given concretely by H x(t) = i-1 x' (t)
for those x(t) E L2(- oo, oo) for which x(O) = 0 and x' (t) E L2(- oo, oo).
Remark 2. Since fi = A. dE(A.) is an extension of H 1 , we have, for
X E D(H1),
00
00
IIH1xii 2 =11Hxll 2 =
J A diiE(A.)xll
2
J A2 d(E(A.)x,x)
-00
-00
00
00
A.2d(E(A.)P(X1)x,P(X1)x)=
A. 2 d(F(A.)x,x).
-00
-00
J A.
00
However, the condition
-oo
d (F (A.) x, x) < oo does not necessarily
imply that xE D(H1). Concerning this point we have

Theorem 3. In the case of a maximal symmetric operator H 1 , we
have, for the corresponding generalized spectral resolution A. dF (A.)
J A.
-oo
00
xED(H1 ) isequivalenttothecondition
d(F(A.)x,x)<oo.
oo
Proof. The reasoning in Remark 2 shows that
J A.2 d II E(A.) x [12<

00
< oo implies that
(5)
J A. d(F(A.)x,x)
2
-oo
oo, i.e., x E D (ii). The operator
-00
when considered as an operator in X 1 , is a symmetric extension of H 1

and D (H') = D (H)(\ X 1 . Thus, by the maximality of Hv we must have
H 1 =H'. Thus D(H1)=D(H')=D(H)f\X1 and so, by H'=
P (X1) if P (X1), the condition
23
J A.2 d II E(A.) x 11
-oo
00
< oo with x E X 1 implies
354
J /,.
00
d(F(A.) x, x) < oo; and conversely, the condition x E D(H1) implies
-00
00
00
J ),2d(F(A.)x,x)= J ),2diiE(A.)xll2<oo.
-oo
-oo
Remark 3. Since, by Theorem 1, any closed symmetric operator can be
extended to a maximal symmetric operator Hv we can apply our Theorem 3 to the effect that (5) is valid. For a detailed exposition of the generalized spectral resolution, see N. I. ACHIESER-l. M. GLASMAN [1] or
B. Sz. NAGY [3]. The spectral representation of a self-adjoint operator
in a Hilbert space can be extended to a certain class of linear operators
in a Banach space with an appropriate modification. This result is due to
N. DuNFORD and may be considered as an "elementary divisor theory" in
infinite dimensional spaces. N. DuNFORD-]. ScHWARTZ [6].
16. The Group-ring V and Wiener's Tauberian Theorem

The Gelfand representation admits another important application in
functional analysis, namely, an operator-theoretical treatment of the
Tauberian theorem of N. WIENER.
The linear space V (- oo, oo) is a ring with respect to the function sum
and the product X defined by
(/X g) (t)
For, by the Fubini-Tonelli theorem,
_[ j_[ f(t-s)
00
J f(t -- s) g(s) ds.
= (I* g) (t) =
(1)
-00
g(s) ds\ dt< _[ lf(t- s) I dt _[ lg(s) Ids

00
= f
-00
00
I!WI&
-00
lgWI&.
We have thus proved
llfxg!l < IIIII llgll, where I! II is the norm in V(-oo,oo).
(2)

Proposition 1. We can introduce formally a multiplicative unit e in
the ring V consisting of all elements given formally by
=Ae + x, xEV(-oo,oo).
In fact V is a normed ring by the following rule:
(A.1 e + x1 ) + (/,.2 e + x2) = (/,.1 + A.2) e + (x1
(A.le
(3)
+ x2) ,
tX(A.e + x) = tXAe + lXX,

xl) X (A.2e + x2) = A.1A.2e + A1X2 + A2X1
I!A.e + xll = lA. I+ l!xll.
+ x1Xx2,
(4)
16. The Group-ring L 1 and Wiener's Tauberian Theorem
355
This normed ring V is called the group-ring of the group R1 of real

numbers written additively. We shall find all maximal ideals of this ring
V. One is I 0 = V(-oo, oo) from which Vis obtained by adjunction
of the unit e. We shall find all maximal ideals I =I= I 0
For any maximal ideal I of the normed ring V, we shall denote by
(z, I) the complex number which corresponds to the element zby the ring
homomorphic mapping V---+ Vjl. Thus (z, I 0) =A. for z = A.e + x,
xE ! 0
Let I be a maximal ideal of U different from I 0 Then there exists a
function x E V(-oo, oo) = ! 0 such that (x, I) =I' 0. We set
x(<X) = (x,., I)j(x, I), where x,.(t) = x(t +<X).
(5)
Then x(O) = 1, and, by \(x,., I) J < JJx,.JJ = J\x\1. we see that \x(<X) J <
\\ x \\/\ (x, I) J. Thus the function X (<X) is bounded in <X. Moreover, by
\x(<X
+ b)- X (<X) I< IJx,.H- x,. II! I(x, I) I,
the function X(<X) is continuous in <X. For, by (2) in Chapter X, 1, we have
J \x (<X +
00
lim
~0-oo
(J
+ t) - x (<X + t) I dt = 0.
On the other hand, we have, by (x,.+ll X x) (t) = (x"' X xp) (t),
(x,.+ll, I) (x, I) = (x,., I) (xp. I)

so that
X(<X +
f3)
= X(<X) x(f3).
(6)
From this we can prove that there exists a uniquely determined real
number ; = ; (I) such that
X(<X)=exp(i;(I)<X).
(7)
In fact, by X (n<X) =X (<Xt and the boundedness of the function x. we

obtain \x(<X)\ < 1. Consequently, by X(<X)X(-<X) =x(O) = 1, we must
have Jx (<X) J = 1. Thus, as a continuous solution of (6) whose absolute
value is one, the function X (<X) must be given in form (7). That the
value~(!) is determined by I independently of the choice ofxEL 1 (-oo,oo)
may be seen from X(<X) (y, I)= (y", I) which is implied by x,.xy =
xxy,..
Every continuous solution of (6) with the absolute value identically
equal to one is called a continuous unitary character of the group R 1 of real
numbers written additively.
We have thus constructed the (continuous unitary) character X (<X)
of the group R 1 with respect to the given maximal ideal I =I= I 0
We next show how to reconstruct the ideal I with respect to this
character or, what amounts to the same thing, to reconstruct the value
(z, I) from X
23*
356
For any y(t) E LI(-oo, oo), we have

00
(xxy) (t) =
J x(t- s) y(s) ds = J x_.(t) y(s) ds.

-oo
-oo
00
Hence, by (5) and the continuity of the ring homomorphism V----? VJI,
we obtain
00
(xxy, I)= (x, I) (y, I)= (x, I)
z(-s) y(s) ds.
-00
Thus, by (x, I) =I= 0, we obtain

00
(y,I) =
z=
Therefore, for any
J y(s) exp(-i ~(I) s) ds.
(8)
-00
.A.e
+ x with xE V(- oo, oo),

00
(z, I)= (I.e, I)+ (x, I)= I.+
J x(s) exp(-i ~(I) s) ds.

-oo
(9)
Conversely, any (continuous, unitary) character z(tX) = exp(i ~ tX)

of the group R 1 defines a ring homomorphism
+ J x(s) exp(-i ~ s) ds
00
z----? .A.
-oo
(z =I.e+ x, x E LI(-oo, oo)) (10)
of V onto the ring of complex numbers. For, we have, by the FubiniTonelli theorem.
00
J (x
-oo
Xx2) (t) exp(-i ~ t) dt
00
J x1 (t)exp(-i~t)dt -ooJ x2 (t)exp(-i~t)dt.

-oo
00
We have thus proved

Theorem 1 (GELFAND [4] and RAIKov [1]). There exists a one-to-one
correspondence between the set of all maximal ideals I =I= V (- oo, oo)
of the group ring L1 of the group R1 and the set of all continuous, unitary
characters z(tX) of this group R 1 This correspondence is defined by formula (9).
We shall show that the normed ring Vis semi-simple or, what amounts
to the same thing (see Chapter XI, 2), that the following theorem is true.
Theorem 2. The normed ring L 1 has no generalized nilpotent elements other than 0.
Proof. Let xE Ll(-oo, oo) and yE P(-oo, oo). Then, by Schwarz'
inequality,
/_[ x(t- s) y(s) dsj
< (_[ !x(t- s) Ids_[
!x(t- s) !!y(s) !2 dsy
12
357
16. The Group-ring L 1 and \Viener's Tauberian Theorem
and so, by the Fubini-Tonelli theorem, we see that the left hand side
belongs to L 2 (-0<.J, =)and
jjxxyjj 2 < llxiiiiYII 2 ,where
Jlfl 2
(11)
isthenorminL2(-=,=).
Hence we can define a bounded linear operator Tx on

LZ(- =, =) by
L 2 (-=,
=)into
00
J x(t-s)y(s)ds
(Txy)(t)=
(12)
whenever xELl(-=,=);
-00
and, moreover, we have
(13)
Tt
= Tx where x* (t) = x (-t).
(14)
Thus, by applying the Fubini-Tonelli theorem again, we obtain

TxT:= Txxx
Txxx
(15)
r:Tx, i.e., Tx is a normal operator.
Hence, by Chapter XI, 3, we have II Tx 11 2 = lim (Jj T;

by
Therefore,
1~00
llxxxxxx
___,
,____,_,__ Xxll2
n factor
1/ 2) 1/n.
JIT;IJz,
we see that, if x is a generalized nilpotent element, then II T,. IJ 2 = 0.

From this fact we easily prove that x must be a zero vector of L 1 (-=,=).
Let now = Ae + x with x E V (- =, =) be a generalized nilpotent
element of the normed ring V. Then, for any maximal ideal I of V, we
must have (z, I) = A + (x, I) = 0. Hence the Fourier transform
00
(2n)- 1' 2
J x(t) exp(-i ~ t) dt
-00
must be identically equal to -(2n)- 1 ' 2 A. From this we prove that A

must be 0. For, we have (the Riemann-Lebesgue theorem)
1-l
x(t) exp(-i
~ t) dtl = 12-
~ 2-1
1 _[
[x(t)-x(t
j \x (t) - x (t + ; ) I dt --+ 0
-oo
+ ;)]exp(-i. ~. t) dtl
as
~ --+ =
Thus a generalized nilpotent element E V must be of the form = x

with x E V (- =, =) and so, by what we have proved above, = x = 0.
We are now in a position to state and prove the Tauberian theorem of
N. Wiener [2]:
Theorem 3. Let x(t) E V(-=, =)be such that its Fourier transform
J x(t)exp(-i~t)dt
00
(2n)- 1i 2
-oo
358
does not vanish for any real f Then, for any y(t)EU(-oo,oo) and
e > 0, we can find the real numbers fl's, the complex numbers x's and a
positive integer N in such a way that
_[I
y(t)- i~ xix(t-{li) dt <e.
Proof. It suffices to find
zE U
in such a way that
i!Y- xxzii < ef2.
We first show that

00
lim
C->OO
(16)
J
-oo
Iy (t) - y<"l (~ I dt
y<"l(t) =..!._ Joo y(t-s)

~
(17)
0, where
(18)
1-c~s.xs ds.
IXS
-00
00
For, by
J (1- cos xs) (xs2)-1 ds = :n (IX> 0), the left side is
-oo
<(:n)-llim
C-->00
Next we have
(2:n)-1/2
-oo
ji!___
-oo
s~oss) ds{-oojly(t)-y(t=)I dt}= 0.

I
(-~)1/21-cos .xu e-i~ du = { 1-l$1/x (l$1

~
(19)
<IX),
(l$1 > x).
.xua
For, we have
"
(2:n)-l/-
.
2 )1/2 "'
~)
J"' (1-l$[/x)e~"d$=(nJ(1-~
cosu$d$
0
-<
(~)1/2 ..!._
~
J(1-1_) d (sin u$) = (~)1/2 Jsin u~ d$
U O
= (~)1/21- COS
:rr:
u=.x
:rr;
IX
UIX
UIX
'
and so we have only to apply Plancherel's theorem in Chapter VI, 2.

Hence, by the Parseval relation of the Fourier transform,
00
y<"'l (t) satisfies
J y<"l (t) exp (-i $ t) dt =
0 if I$I 2
IX.
(20)
-00
Therefore, we may assume that y E U (- oo, oo) in (17) satisfies the

condition
00
J y(t) exp(-i $ t) dt =
-oo
0 whenever l$1 > x.
(21)
Next we choose a positive number {3 and a sufficiently large positive

number y such that [-{3- y, -{3 + y] and [{3 -y, {3 + y] both contain
16. The Group-ring L 1 and wiener's Tauberian Theorem

[-<X,
ex]. Let
and
C 1 (~)
C 2 (~)
359
be the defining functions of the intervals
[-y, y] and [-{3, fl], respectively. Then
u(~) =
(2{3)-1 _ [
to< u(~) <
C 1 (~-17) C
(1J) d1J = 1 for
~E [-ex, ex],
= 0 for sufficiently large I~ I,

(22)
1 for all real~-
By the Parseval relation of the Fourier transform, we have
u(t) =
21{1 (2n) 11:l cl (t) c2 (t).
Thus, by Plancherel's theorem, we see that u(t) belongs to V (- oo,oo) (\

L2 (- oo, oo). Hence we may apply Theorem 1 so that
f(t)
= u(t) =
(2n)- 1' 2 (u, It) with a certain maximal ideal It# I 0 (23)
Moreover, by Plancherel's theorem, the inverse Fourier transform of /(t)

is equal to u (~), i.e., we have
u (~) = (2n)- 112 (/, L~).
We next set
g=
0 < (g, I~)

(g, I;)
e-
(24)
f. Then, by what was proved above,

= 0 whenever ~ E [-ex, ex];
(2n)- 1' 2
< 1; (g, I~)
1 for all sufficiently large \~I
(25)
On the other hand, we have, for x* (t) = x (-t), the relation (x, I~) =
(x*, I~). Thus, by hypothesis, we have (x* X x, I;) = I (x, I~) 12 > 0 for all
real~- Hence the element
g + x*xxEV
satisfies the condition that (g + x* X x, I) > 0 for all maximal ideals I
of V. Consequently, the inverse (g + x*xx)-1 in V does exist.
We define
(26)
z = (g + xxx*)- 1 xx*xy.
Then, by (4), the element xxzbelongs to V(-oo, oo). Moreover, we have

for every real number ~,
(x*,h)(y,IE)
(I)
(I)(-)
-I)
(xXz,
~ = x, E z, I~ = X, ~ ('g, I;) + (x, I;) (x*, I;)
Hence, by (25) and the hypothesis that (y, I~) = 0 whenever\~ I> ex, we
obtain
(xxz, I~)= (y, I~;) for all real~The normed ring Vis semi-simple by Theorem 2. Hence we must have
xxz = y. We have thus proved Theorem 3.
Corollary. Let kdt) belong to V(-oo,oo), and let its Fourier transform vanish for no real argument. Let k 2 (t) belong to V(-oo,oo). Let
360
l(t) be Baire measurable and bounded over (-oo,oo). Let, for a constant C,
00
lim
J k (t- s) l(s) ds =
00
~-00
00
Jk
~-00
2 (t-
s) I (s) ds
(27)
-00
Then
lim
J kdt) dt.
00
Jk
2 (t)
(28)
dt.
-00
Proof. We may plainly suppose that C = 0. And (28} holds for k 2 (t)
of the form k2 (t) = (,xXk1} (t) with x(t} E Ll(-oo, oo). It is easy to
prove that (28) holds for k2 (t) if k 2 (t) =s-lim k(n) (t) in L 1 ( - oo, oo) with
~
k(nl(t)EL1 (-oo,oo) for which (28) holds (n=1,2, ... ). Hence, by

Theorem 3, we see that (28) holds for every k2 (t) ELl(- oo, oo).
Remark. N. WIENER ([1], [2] and [3]) has applied the above Corollary
to a unified treatment of classical results concerning the limit relations
in series and integrals which include a new proof of the prime number
theorem. Cf. also H. R. PITT [1 J. The above proof of Theorem 3 is adapted
from M. FUKAMIYA [1] and I. E. SEGAL [1]. Cf. also M.A. NAIMARK [1]
and C. E. RICKART [1] and the bibliographies cited in these books. For
the sake of comprehension of the scope of the above Corollary, we shall
reproduce Wiener's deduction of the special Tauberian theorem. The
special Tauberian theorem, as formulated by J. E. LITTLEWOOD reads:
00
Theorem 4. Let .I a,.~ converge to s (x) for IxI

n=O
<
1, and let
lim s(x) =C.
(29)
Z--+1-0
Let, moreover,
sup n Jan I = K
n;;;-;1
Then
<
(30)
oc.
00
.Ian=C.
(31)
n=O
Proof. Put
I (x)
[z]
= .I a,.. Then we see, by

n=O
Jl(x}--s(e-1/z}l =
[z]
I2an(1-e-nfx)n=1
<
-
[z]
n=1
K
n
!:': +
X
< 2K + K
that I (x) is bounded.
oo
[z]+1
K e-nfz < 2K
n
-
00
ane-nfzl
[z]+l
I
oo
+K f
J e-uu- 1du =constant,

00
[z]
e-ufzu-1du
361
16. The Group-ring L 1 and Wiener's Tauberian Theorem
Hence, by partial integration,
J e-u"df(u) = 0J xe-u"f(u) du.

-0
n=O
Thus we have
00
00
00
s(e-") = .I ane-n" =
j xe-""f(u) du =lim j e-ee-rr-11t(e"~) e"~drJ.
C =lim
(31')
E-->00 -oo
..,.....o 0
This may be written as
J k1 (t- s) I (e') ds =
-oo
J k1 {t) dt,
-oo
00
00
lim
t->00
where k1 (t) = e- 1e-- 1 , (32)
since
Jk
00
J e- e-- dt = J e-" dx = 1.
00
00
(t) dt
-00
-00
Furthermore, we have
Jk
-oo
J xiue-"dx =
00
00
1 (t) e-iutdt =
+ iu) =F 0.
F(1
Thus we can apply the Corollary and obtain, for

k2 (t)
0 (t
<
0) ; k2 (t)
(t > 0) ,
the limit relation

00
00
J e- dt=C -ooJ k
0
00
C=C
lim x-1
lim
2 (t)dt=
t->00
Jk
-oo
2 (t-s)f(e')ds
J" f(y) dy.
"~
It follows that, if A.> 0,
c=
=
(1
+A.)~- c =
A
~A.x
lim
x-->00
(l+.t)x
f(y) dy
t(y) dy- f t(y) dy}

...!._{(lY)"
A
X
=~
f(x)
+ A.x
(l+.t)x
[f(y)- f(x)] dy (33)
On the other hand, we have, by (30),
(l+.t)x
,1
AX
"
[f(y)-f(x)]dyl
</ J
(l+A)x
"
[(l+A)x]
< ~
[x)+l
[y)
(x)+l
K
[x]
Kdy
n
[A.x]K
< ---cxJ < 2A.K
362
XII. Other Representation Theorems in Linear Spaces
for sufficiently large value of x. Hence, by (33),

lim l/{x)- C I< 2J.K,
%--+00
and since;. is any positive number, we get

lim /(x) =C.
Thus we have proved (31).
%--+00

In this chapter, we shall prove three representation theorems in
linear spaces. The first one, theKrein-Milmantheoremsays thatanon-void
convex compact subset K of a locally convex linear topological space is
equal to the closure of the convex hull of the extremal points of K. The
other two theorems concern the representations of a vector lattice as point
functions and as set functions.
1. Extremal Points. The Krein-Milman Theorem
Definition. Let K be a subset of a real- or complex-linear space X.
A non-void subset M ~ K is said to be an extremal subset of K, if a
proper convex combination cxk1 + (1- ex) k2 , 0 < ex< 1, of two points
k1 and k 2 of K lies in M only if both k1 and k 2 are in M. An extremal set
of K consisting of just one point is called an extremal point of K.
Example. In a three dimensional Euclidean space, the surface of a
solid sphere is an extremal subset of the sphere, and every point of the
surface is an extremal point of the surface.
Theorem (KREIN-MILMAN}. A non-void compact convex subset K of a
locally convex linear topological space X has at least one extremal point.
Proof. The set K is itself an extremal set of K. Let IDl be the totality
of compact extremal subsets M of K. Order IDl by inclusion relation. It
is easy to see that, if IDl1 is a linearly ordered subfamily of Wl, the nonvoid set n M is a compact extremal subset of K which is a lower bound
ME!ln1
for the subfamily IDl1 .

Thus, by Zorn's lemma, IDl contains a minimal element M 0 Suppose
that M 0 contains two distinct points x0 and y0 Then there exists a continuous linear functional I on X such that f(x0 } =I= f(y0 }. We may
assume that Re f(x0} =I= Re /{y0}. M 0 being compact, the set M 1 =
{x E M 0 ; Re f(x) = in Re f(y)} is a proper subset of M 0 On the other
yEM0
hand, if k1 and k2 are points of K such that cxk 1 + (1- ex) k2 E M 1 for
1. Extremal Points. The Krein-Milman Theorem
363
some with 0 < < 1, then, by the extremal property of M 0 , k1 and k 2

both E M 0 It follows from the definition of M 1 that k1 and k2 both E M 1 .
Hence M 1 is a closed extremal subset properly contained in M 0 Since M 0
is a minimal element of W'l, we have arrived at a contradiction. Therefore
M 0 must consist of only one point which is thus an extremal point of K.
Corollary. Let K be a non-void compact convex subset of a locally
convex real linear topological space X. Let E be the totality of the
extremal points of K. Then K coincides with the smallest closed set
lt = 1) of
,e, ( :::::: 0,
containing every convex combination
points e1E E, i.e., K is equal to the closure of the convex hull Conv (E) of E.
Proof. The inclusion E ~ K and the convexity of K imply that
Conv(E)" is contained in the compact set K. Suppose that there exists
a point k0 contained in (K -Conv(E)"). We can then take a point
c E Conv (E)" so that (ko- c) E (Conv (E)"- c). The set (Conv (E)"- c)
being compact convex and 3 0, there exists, by Theorem 3' in Chapter IV, 6,
a continuous real linear functional f on X such that
/(k0 -c)
>
1 and f(k-c) < 1 for (k-c)E(Conv(E)"-c).
Let K 1 =hE K; f(x) = supf(y)} Then, since k0 E K, the set K 1 (\ E
yEK
must be void. Moreover, sinceK is compact, K 1 is a closed extremal subset

of K. On the other hand, any extremal subset of K 1 is also an extremal
subset of K and hence any extremal point of K 1 , which surely exists
by the preceding Theorem, is also an extremal point of K. Since K 1 (\ E
is void, we have arrived at a contradiction.
Remark. The above Theorem and the Corollary were first proved
by M. KREIN-D. MILMAN [1]. The proof given above is adapted from
J. L. KELLEY [2]. It is to be noted that for the unit sphere S = {x EX;
llxll < 1} in a Hilbert space X the extremal points of S are precisely
those on the surface of S, i.e., those of norm 1. This we easily see from
(1) in Chapter I, 5. For applications of the notion of the extremal points
to concrete function spaces, see, e.g., K. HoFFMAN [1].
A simple example. Let C [0, 1] be the space of real-valued continuous
functions x(t) defined on [0, 1] normed by x =max lx(t) I The dual
I
space X= C [0, 1]' is the space of real Baire measures on [0, 1] of bounded total variations. The unit sphere K of X is compact in the weak*
topology of X (see Theorem 1 in the Appendix to Chapter IV). It is
easy to see that extremal points of K are in one-to-one correspondence
with the linear functionals / 1, EX of the form <x. /1,) = x(t0 ), t0 E [0, 1].
The above Corollary says that any linear functional IE X is given as
the weak* limit of the functionals of the form
n
1=1
1=1
..I ;x(t1), where ; > 0, ..I;= 1 and t1 E [0, 1].
364
Recently, G. CHOQUET [1] proved a more precise result: If X is a

metric space, then E is a G6 set and, for every xE K, there exists
a non-negative Baire measure f-lx (B) defined for Baire sets B of X
such that f-lx (X --E) = 0, f-lx (E) = 1 and x =
y f-lx (dy). As for the
uniqueness of
P. A. MEYER [2].
f-lx
and
further
literature,
see G. CHOQUET and
2. Vector Lattices
The notion of "positivity" in concrete function spaces is very important, in theory as well as in application. A systematic abstract treatment of the "positivity" in linear spaces was introduced by F. Rmsz [6],
and further developed by H. FREUDENTHAL [2], G. BIRKHOFF [1]
and many other authors. These results are called the theory of vector
lattice. We shall begin with the definition of the vector lattice.
Definition I. A real linear space X is said to be a vector lattice if X
is a lattice by a partial order relation x < y satisfying the conditions:
< y
implies x
+ z < y + z,
x < y implies <XX< <XY (or <XX> <XY) for every <X
(1)
> 0
(or <X
< 0). (2)
Proposition I. If, in a vector lattice X, we define
x+
= xV
then we have
x V y
= (x-y)+
+ y,
0 and x-
= x 1\
(3)
0,
x 1\ y = - ((-x) V (-y)).
Proof. The one-one mappings x.....,.. x + z and x.....,.. <XX (<X

onto X both preserve the partial order in X.
(4)
>
0) of X
Example. The totality A (5, 58) of real-valued, a-additive set functions

x (B) defined and finite on a a-additive family (5, 58) of sets B ~ 5 is
a vector lattice by
(x
+ y)
(B)
and the partial order x

have, in this case,
x+(B)
= x (B) + y (B), (<Xx) (B) = <XX (B)

< y
defined by x(B)
< y(B)
on 58. In fact, we
supx(N) =the positive variation V(x; B) of x on B.
(5)
N'>,B
Proof. We have to show that V(x; B)= (x V 0) (B). It is clear that

V(x; B)> 0 and x(B) < V(x; B) on 58. If 0 < y(B) and x(B) < y(B)
on 58, then, for any N ~ B, y(B) = y(N) + y(B- N) > x(N) and so
y(B) > V(x; B) on 58.
365
2. Vector Lattices
Proposition 2. In a vector lattice X, we have

(6)
xXy+z=(x+z)X (y+z),
~ > 0,
(7)
Xy) = (~x) 0 (~y) for ~ < 0,
(8)
~ (x Xy) = (~x) ~ (~y)

~ (x
for
xl\y=-(-x)V (-y), x-=-.(-x)+, x+=-(-x)-.
Proof. Clear from (1) and (2).

Corollary.
x + y = x V y + x 1\ y, in particular, x = x+
+ x-.
(9)
(10)
Proof.xVy-x-y =OV (y-x)-y=(-y) V (-x)=-(y/\x).

(11)
x X y = y X x (commutativity),
(12)
. . .
x V (y V z) = (x V y) V z = sup(x, y, z)}
. f (X, y, Z) (associativity), (13)
" y ) 1\ Z_- Ill
_ (X 1\
X 1\ (y 1\ Z) . .
. .
(x 1\ y) V z = (x V z) /\ (y V z) }
(distnbutlVlty).
(x V y) 1\ z = (x 1\ z) V (y 1\ z)
(14)
(15)
Proof. We have only to prove the distributivity. It is clear that

(x 1\ y) V z < x V z, y V z. Let w < x V z, y V z. Then w < x V z =
x + z- x 1\ z and so x + z > w + x 1\ z. Similarly we have y + z 2:
w + y 1\ z. Hence
w+~l\~1\~1\~=~+xl\~1\~+yl\~
and so
<
(x
+ z)
1\ (y
z) = x 1\ y
+ z,
w < (x 1\ y) + z- (x 1\ y) 1\ z = (x 1\ y) V z.
We have thus proved (14). (15) is proved by substituting -x, -y, -z for
x, y, z, respectively in (14).
Remark 1. In a lattice, the distributive identity
(16)
x 1\ (y V z) = (x 1\ y) V (x 1\ z)
does not hold in general. The modular identity:
(17)
x < z implies x V (y 1\ z) = (x V y) 1\ z
the
Then
group.
a
be
G
Let
(15).
is weaker than the distributive identity
that
lattice,
modular
a
constitutes
G
of
totality of invariant subgroups N
is, a lattice satisfying the modular identity (17), if we define N 1 V N 2
and N 1 1\ N 2 as the invariant subgroup generated by N 1 and N 2 and
the invariant subgroup N 1 (\ N 2 , respectively.
Remark 2. A typical example of a distributive lattice is given by the
Boolean algebra; a distributive lattice B is called a Boolean algebra if
it satisfies the conditions: (i) there exist elements I and 0 such that
0 < x < 1 for every x E B, (ii) for any x E B, there exists a uniquely
366
determined complement x' E B which satisfies x V x' =I, x 1\ x' = 0.

The totality B of subsets of a fixed set is a Boolean algebra by defining
the partial order in B by the inclusion relation.
Proposition 4. We define, in a vector lattice X, the absolute value
lxl = x V (-x).
(18)
IxI > 0, and Ix I = 0 iff x = 0,
(19)
lx + Yl < lxJ + JyJ, l~xJ = J~JJxJ.
(20)
Then
Proof. We have
x+ 1\ (-x-) = x+ 1\ (-x)+ = 0.
(21)
For, by 0 = x-x = x V (-x) + x 1\ (-x) > 2(x 1\ (-x)) and the
distributive identity (14), 0 = (x 1\ (-x)) V 0 = x+ 1\ (-x)+. Thus
x+- x- = x+ + (-x)+ = x+ V (-x)+.
On the other hand, from x V (-x) > x 1\ (-x) >- ((-x) V x), we
have x V (-x) ~ 0 and so by (21)
x V (-x) = (x V (-x)) V 0 = x+ V (-x)+ = x+ + (-x)+.
We have thus proved
x+- x- = x+ + (-x)+ = x+ V (-x)+ = x V (-x).
(22)
We now prove (19) and (20). If x = x+ + x- =/= 0, then x+ or (-x-}
is >0 so that lxJ=x+V (-x-)>0. jxl=(x)V(-x)=
11 (x V (-x)) = IJixJ. From Jxl + IYI > x + y, -x-y, we obtain
jxj + jyj > (x + y) V (-x-y) = Jx + yJ.
Remark. The decomposition x = x+ + x-, x+ 1\ (-x-) = 0, is called
the ] ordan decomposition of x; x+, x- and IxI correspond to the positive
variation, the negative variation and the total variation, of a function x(t)
of bounded variation, respectively.
Proposition 5. For any y E X, we have
lx-xiJ = Jx V y-x1 V yJ + lx 1\ y-x1 1\ yJ.
(23)
Proof. We have
Ia- b I =
(a- b)+- (a- b)-
V b - b- (a 1\ b - b)
=aVb-a/\b.
Hence the right side of (23) is, by (10), (14) and (15),
=~V~V~-~V~/\~V~+~/\~V~/\~-~1\~/\~
= (x V x1) V y - (x 1\ x1) V y + (x V x1) 1\ y - (x 1\ x1) 1\ y

= x V x1 + y- (x 1\ x1 + y) = x V x1 - x 1\ x1 = lx- x1 1.
Definition 2. A sequence {x,.} of a vector lattice X is said to 0-converge

to an element x E X, in symbol, 0-li~ x,. = x if there exists a sequence
~
{w,.} such that Jx- x,.j:::;; w,. and w,. t 0. Here w,. t 0 means that
367
2. Vector Lattices
> w2 > and 1\
= 0. The 0-lim x,., if it exists, is uniquely

~
determined. For, let 0-lim x,. = x and 0-lim x,. = y. Then Ix- x,.l : : ;: w,.,
w,. -), 0 and Iy - x,. I < u,., u,. t 0. Thus Ix- y I < Ix - x,. I + Ix,.- y I <
w,. + u,. and (w,. + u,.) t 0 as may be seen from w,. + 1\ u,. >
w1
n61
w,.
1\ (w,.
n;:;;l
+ u,.). This proves that x =
n61
y.
Proposition 6. With respect to the notion of 0-lim, the operations

x + y, x V y and x 1\ yare continuous in x andy.
Proof. Let 0-limx,. = x, 0-limy,. = y. Then lx + y-x,. -y,.l <
~
lx- x,.l
(x,. + y,.) = x +
+ ly-y,. Iimplies 0-lim
......co
IX XY- X,. XYn I < IX XY -
X,.
y. By (23}, we have
XY I + IX,. XY- X,. XYnl
< jx-x,.l + Jy-y,.j,

Xy,.) = (O~x,.) X(O~Y.. )
and so 0-lim (x,.

Remark. We have
But, in general,
0-lim eX x,.
n-+<X>
=eX 0-lim x,..

n"""+OO
0-lim eX,. X =F (lim
......co
eX,.) x.
The former relation is clear from jcXx-cXx,.j = jcXjjx-x,.l. The latter

inequality is proved by the following counter example: We introduce in
the two-dimensional vector space the lexicographic partial order in which
(~ 1 , 171) > {~2 , 1) 2) means that either ~1 > ~2 or ~ 1 = ~2 , 1)1 > 1)2 . It is
easy to see that we obtain a vector lattice. We have, in this lattice,
n-1 (1, 0)
>
(0, 1)
> 0=
(0, 0) (n
1, 2, ... ) .
Hence 0-lim n-1 (1, 0) =F 0. A necessary and sufficient condition for the
~
validity of the equation

0-lim !Xn X =
(X
(24)
tXn-+~X
is the so-called A rchimedean axiom

0-lim n-1 x = 0 for every x > 0.
(25)
This we see by the Jordan decomposition y = y+ + y-.

Definition 3. A subset {x,.} of a vector lattice X is said to be bounded
if there exist y and z such that y < x,. < z for all x,.. X is said to be
complete if, for any bounded set {x,.} of X, sup x,. and inf x,. exist in
"
"
X. Here sup x,. is the least upper bound in the sense of the partial order
" x,. is the largest lower bound in the sense of the partial order
in X, and inf
"
lattice X is said to bP- a-complete, if, for any bounded
in X. A vector
368
sequence {x,.} of X, sup x,. and inf xn exist in X. We define, in a a"~1
n~1
complete vector lattice X,

0-Iim x.. = inf (sup x,.) , 0-lim x,. = sup ( inf x.. ).
~
m,.~m
(26)
m"~m
Proposition 7. 0-lim x,. = x iff 0-lim x,. = 0-lim x,. = x.

n-+oo
n---HX>
Proof. Suppose Ix-= x,. I < w,., w,. t 0. Then x- w,. < x,. < x + w,.
and so we obtain 0-lim (x- w,.) = x < 0-lim x,. < 0-lim (x + w,.) = x,
~
that is, 0-lim x,. = x. Similarly, we obtain 0-lim x .. = x.

n--+-00
We next prove the sufficiency. Put u,. = sup Xm, v,. = inf Xm,
m:?:n
m~n
u,.- v,. = w,.. Then, bv hypothesis, w,. t 0. Also, by x,. < u,. = x +
(u,.- x) < x + (u,.- v,.) = x + w,. and x,. > x- w,. obtained similarly,
we prove Ix- x,. I < w,.. Thus 0-lim x,. = x.
~
Proposition 8. In a a-complete vector lattice X, <XX is continuous in

<X, x with respect to 0-lim.
Proof. We have I<XX-<Xnxnl < I<Xx-<Xx,.l + I<XX,.-<X.. x,.l =
I<X l!x- x,. I + I<X -<Xn llx,. I Hence, if 0-lim x,. = x, lim <X,.= <X, then
~
the 0-lim of the first term on the right is 0. Therefore, by putting

suplx,.l=y,supi<X-<Xm!=.B,., we have to prove 0-lim,B,.y=O.
n~1
m~n
> 0 and ,8,. t 0, we see that 0-lim,B.. y = z exists and

0-lim 2- 1 p,.y = 2- 1 z. Since there exists, for any n, an n 0 such that
,8,., < 2- 1 p,., we must have z = 2- 1 z, that is, z = 0.
But, by y
Proposition 9. In a a-complete vector lattice X, 0-Iim x,. exists iff

n--->OO
0-lim lx,.- Xm I= 0.
(27)
n,m--->00
Proof. The necessity is clear from Ix,.- Xm I < Ix,.- xI + Ix- Xm J.

< Xm + 0-lim Ynm 0-lim x,. >
If we set J x,.- Xm I = Ynm then 0-lim Xn
n--KX:>
n-+OO
n--KX>
Xm - 0-lim Ynm Hence

~
< 0-lim x,.- 0-lim x,. < 0-lim (0-lim Ynm- 0-Iim Ynm) =
~
m--->00
n--->00
n--->00
0.
Thus we have proved the sufficiency.

Proposition 10. A vector lattice X is a-complete iff every monotone
increasing, bounded sequence {x,.} ~X has sup x,. in X.
n~1
Proof. We have only to prove the sufficiency. Let {z,.} be any bounded
sequence in X, and set x,. =sup zm. Then, by hypothesis, sup x,. = z
m~n
n:?:1
exists in X, and z =sup z,.. Similarly, we see that inf z,. = inf( inf zm)
n~1
n~1
n~1 m;;;;n
also exists in X.
369
3. B-lattices and F-lattices
3. B-lattices and F -lattices

Definition. A real B-space (F-space) X is said to be a B-lattice (Flattice) if it is a vector lattice such that
(1)
lx I< IY I implies llx II< IIY II
Examples. C (5), LP (5) are B-lattices by the natural partial order
x < y which means x(s) < y(s) on 5 (x(s) < y(s) a.e. on 5 for the case
LP (5) ). M (5, ~, m) with m (5) < oo is an F -lattice by the natural
partial order as in the case of LP(5). A (5, ~)is a B-lattice by
II x II = IxI (5) = the total variation of x over 5.
We have, by (1) and lx I = I(lx I) I,
In A (5,
llxll
have, moreover,
~),we
x>o,y>o imply
(2)
ll(lxl) II
(3)
llx+yll=llxii+IIYII
S. Kakutani called a B-lattice satisfying (3) an abstract V-space. From

(3) we have
(4)
lxl < IYI implies llxll < IIYII
The norm in A (5,
~)
0-limx,.
n-->00
is continuous in 0-lim, that is
implies
(5)
lim llx,.ll = llxiJ
n-->00
For, inA (5, ~), 0-lim x,. =xis equivalent to the existence of y,.EA (5, ~)
11->00
such that J x- x,. I (5) < y,. (5) with y,. (5) .} 0. It is easy to see that
M(5, ~. m) satisfies (4) and (5).
Proposition 1. A a-complete F-lattice X satisfying (4} and (5) is a
complete lattice. In particular, A (5, ~)and LP(5) are complete lattices.
Proof. Let {x,.} ~X be bounded. We may assume 0 <XIX< x for all
<X, and we shall show that sup x,. exists. Consider the totality {zp} of
"'
Zp obtained as the sup of a finite number of xiX's: Zp
= .V" xiXJ'
3=1
Set
y = s~p II Zp II Then there exists a sequence {Zp;} such that;~ II Zp1 II = y.

If we put z,. =sup Zp; then 0-lim z,. = w exists, and, by (5) and the
j~n
..-.oo
definition of y, we have llw II= y. We shall prove that w =sup XIX.

IX
Suppose that x"' V w > w for a certain XIX. Then, by (4), llxiX V w II>
llw II= y. But, by x,. V w = 0-lim (x"' V z,.), xiX V z,. E {zp} and (5), we
have llxiX V w II
we must have w >
11->00
lim llxiX V z,. II< y, which is a contradiction. Hence

11->00
xiX for all x,.. Let xiX z11 for all
z11 Hence we must have w = sup x,..
Remark.
s-lim x,.
.......00
In C(S),
x. In M (5,
"' x,. = x does not necessarily imply

0-lim
.......00
~.
m), s-lim x,.

11->00
x does not necessarily imply
0-lim x,. = x. To see this, let x1 (s), x2 (s), ... be the defining functions of
.......00
the intervals of [0, 1]:
[o. ~]. [~,
~]. [o. !]. [!,
!]. [!, !J. [!, !]. [o. !]. [!, :] ....
Then the sequence {x,.(s)} ~ M ([0, 1]) converges to 0 asymptotically but

does not converge to 0 a.e., that is, we have s-lim x,. = 0 but 0-lim x, = 0
.......00
does not hold.
.......00
Proposition 2. Let X be an F-lattice. Let a sequences {x,} ~X satisfy

s-lim x,. = x. Then {x,} *-converges to x relative uniformly. This means
.......00
that, from any subsequence {y,.} of {x,.}, we can choose a subsequence

{Yn(kJ} and a z E X such that
IYn(k)
-xj <
k- 1 z
(k = 1, 2, ... ).
(6)
Conversely, if {x,.} *-converges relative uniformly to x, then s-lim x,. = x .

.......00
Proof. We may restrict ourselves to the case x = 0. From lim
II y,.lj =
0,
we see that there exists a sequence {n(k)} of positive integers such that
.......00
II <
00
Then (6) holds by taking z = ~ jky,.(k) j. Conk=1

versely, let condition (6) with x = 0 be satisfied. Then we have
IIYn(kJII < !lk-1 zll from !Yn(kJ! < k- 1 z. Hence S~Yn(k) = 0. Therefore, there cannot exist a subsequence {y,.} of {x,.} such that lim llYn II> 0.
!lkYn(k)
k-2
.......00
Remark. The above proposition is an abstraction of the fact, that

an asymptotically convergent sequence of M(S, ~. m) with m(S) < oo
contains a subsequence which converges m-a.e.
4. A Convergence Theorem of Banach

This theorem is concerned with the almost everywhere convergence
of a sequence of linear operators whose ranges are measurable functions.
See S. BANACH [2]. A lattice-theoretic formulation of the theorem reads
as follows (K. Y OSIDA [15]) :
Theorem. Let X be a real B-space with the norm !Ill and Y a acomplete F -lattice with the quasi-norm 11111 such that
0-limy,. = y
.......00
implies
lim IIYnlh =
.......00
!IYI/1
(1)
4. A Convergence Theorem of
371
BANACH
Let {T..} be a sequence of bounded linear operators E L (X, Y). Suppose

that
0-lim IT.,x I exists for those x' sEX which form a set
n-+00
(2)
G of the second category.
Then, for any x EX, 0-lim T .. x and 0-lim T .. x both exist and the (not
n-+00
necessarily linear) operator

fx =
n-+00
defined by
(0~ T .. x)-(0~ T .. x)
(3)
is continuous as an operator defined on X into Y.

Remark. The space M(S, ~. m) with m(S) < oo satisfies (1) if we
write y1 < y 2 when and only when y 1 (s) < y 2 (s) m-a.e., the quasi-norm
II y 111 being defined by II y 11 1 = Iy (s) I (1 + Iy (s) 1)-1 m (ds). Likewise
s
LP(S, ~. m) with m(S) < oo also satisfies (1) by the same semi-order.
Proof of the Theorem. Set T .. x = y.. , y~ =sup IYm I, y' =sup IYnl
n~l
n~m
and consider the operators V .. x = y~ and V x = y' defined at least on G

into Y. By (23) of the preceding section 2, each V,. is strongly continuous with Tk. Since lim II V,.x- V x 11 1 = 0 by (1), we have
n-+00
lim llk-1 V,.xll = llk-1 Vxll and further lim llk-1 Vxll 1 = 0. These are
~
n-+00
implied by the continuity in .x, y of .xy in the F-space Y. Hence for any
e> O,
s;;
V,.xll
UGk where Gk=fxEX;supjjkn~l
l
1
k=l
<e1.
(4)
By the strong continuity of v.. , each Gk is a strongly closed set of X.

Thus some Gko contains a sphere of X, in virtue of the hypothesis
that G is of the second category. That is, there exist an x 0 EX and a
~ > 0 such that llx0 -xll <~implies sup llk01 V,.xlh <e. Hence, by
putting z = x0 - x, we see that
n~l
supllk01 V,.zlh <supllk01 V,.x0 1l1 +supllk01 V.. xll 1 < 2e,
n~l
n~l
that is, since V,.(k01 z) = k 01 V,.z, we have

sup IIV.. zjh < 2e whenever
n~l
liz II< ~/k0
n~l
This proves that s-lim V,. z = 0 uniformly in n.

1111-+0
G being dense in X, V x is defined for all x E X and V xis strongly
continuous at x = 0 with V 0 = 0. Hence, by
IT xj s 2 V x and i!Tx1 - Tx2 111 sliT(~ -x2) 11 1 ,

we see that f xis strongly continuous at every x EX.
24*
372
Corollary. Under condition (1), the set G
{x EX; 0-lim Tnx exists}

~
either coincides with X or is a set of the first category.
Proof. Suppose the set G be of the second category. Then, by our Theorem, the operator Tis a strongly continuous operator defined on X into Y.
Thus G = {xE X; Ty = 0} is strongly closed in X. Moreover, G is a
linear subspace of X. Hence G must coincide with X. Otherwise, G would
be non dense in X.
5. The Representation of a Vector Lattice as Point Functions
Let a vector lattice X contain a "unit" I with the properties:
I
>
0, and, for any
IE X,
there exists an ex
>
0 such
that -ex!< I< ex I.
(1)
For such a vector lattice X, we can give an analogous representation

similar to the representation of a normed ring as point functions.
An element IE X is called "nilpotent" if n Ill< I (n = 1, 2, ... ).
The set R of all nilpotent elements IE X is called the "radical" of X.
By (20) of Chapter XII, 2, R constitutes a linear subspace of X. Moreover, R is an "ideal" of X in the sense that
IE R and Ig I < II I imply g E R.
(2)
Lemma. Let X 1 and X 2 be vector lattices. A linear operator T defined

on X 1 onto X 2 is called a lattice homomorphism if
X y) =
X (Ty).
(3)
Then T is a lattice homomorphism iff N = {x E X; T x = 0} is an ideal
T(x
(Tx)
of X 1 .
Proof. Let T be a lattice homomorphism. Let x E N and Iy I < IxI

Then, by T(lxl) = T(x V -x) = (Tx) V (T(-x)) = 0, we obtain,
0 < Ty+ = T(y+ 1\ lx/) = Ty+ 1\ T Jxl = 0 and so y+E N. Similarly
we obtain y- E N and thus y = y+ + y- E N.
Next let a linear subspace N = {xE X 1 ; Tx = 0} be an ideal of X 1 .
Then the linear space X 2 is isomorphic to the factor space X 1fN. We
have to show that (x Xy) =
X y, where is the residue class mod N
containing x. But, if y =
i.e., if y- zEN, then, by (23) in Chapter XII, 2,
we obtain
Jx X y-x Xz\ < ly-ziEl\'
so that the residue class (x Xy) is determined independently of the choice

of the representative elements x, y from the residue classes y, respectively.
x,
5. The Representation of a Vector Lattice as Point Functions
373
Remark. The above Lemma may be phrased as follows. Let N be a

linear subspace of a vector lattice. Then the linear-congruence a - b
(mod N) is also a lattice-congruence:
b, a' -
b' (mod N)
implies
X b _ a' X b' (mod N),
iff N is an ideal of X.
Now an ideal N is called "non-trivial" if N =1=- {0}, X. A non-trivial
ideal N is called "maximal" if it is contained in no other ideal =1=- X.
Denote by 9J1 the set of all maximal ideals N of X. The residual
class XfN of X mod any ideal N E 9J1 is "simple", that is, XfN does not
contain non-trivial ideals. It will be proved below that a simple vector
lattice with a unit is linear-lattice isomorphic to the vector lattice of real
numbers, the non-negative elements and the unit I being represented by
non-negative numbers and the number 1. We denote by f(N) the real
number which corresponds to f E X by the linear-lattice-homomorphism
X~ XfN, N E im.
After these preliminaries we may state
Theorem 1. The radical R coincides with the intersection ideal
n N.
NE\m
Proof. The first step. Let X be a simple vector lattice with a unit I.
Then we must have X= {<xi; -oo <ex< oo}.
Proof. X does not contain a nilpotent element f =1=- 0, for otherwise X
would contain a non-trivial ideal N = {g; lg I< r; If I with some r; < oo}.
Hence, by (1), X satisfies the Archimedean axiom:
order-limn-1 lxl
~
for all
xE X.
(4)
Suppose there exists an / 0 E X such that / 0 =1=- y I for any real number y.
Let
ex= inf ex', fJ

fo5.1X'l
sup {J'.
{J'I~/ 0
Then, by (4), {JI < /0 <ex! and{J <ex. Hence (/0 -M)+ =1=- 0, (/0 - M)-=1=- 0
for {J< 15< ex. Thus, by x+ 1\ (-x-) = 0, thesetN0 ={g; lgl < r;(/0 -r'JI)+
with some r; < oo} is a non-trivial ideal, contrary to the hypothesis.
The second step. For any non-trivial ideal N 0 , there exists a maximal
ideal N 1 containing N 0 .
Proof. Let {N0 } be the totality of non-trivial ideals containing N 0 .
We order the ideals of {N0 } by indusion relation, that is, we denote
Na, <NIX, if NIX, is a subset of NIX, Suppose that {NIX} is a linearly ordered subset of {N0 } and set N{J = U NIX. We shall then show that
NIXE{N"'}
Np is an upper bound of {NIX}. For, if x, y E Np, there exist ideals N"' and
NIX, such that x E NIX, andy E NIX,' Since {NIX} is linearly ordered, NIX, ~ N"'
(or N" ~NIX,) and so x and y both belong to N"', This proves that
374
(yx + CJy) EN"' c;;, Np and that lz I< lx I implies z E NIX, c;;, Np. Since
the unit I is contained in no NIX, I is not contained in Np. Therefore Np
is a non-trivial ideal containing every NIX, i.e., Np is an upper bound of
{NIX}. Thus, by Zorn's Lemma, there exists at least one maximal ideal
containing N 0
The third step. R ~ n N. Let f > 0 and nf 0 be not nilpotent. Then we have
Then~
-NEW!
to show that there exists an ideal N E im such that fEN. This may be
proved as follows.
Since f > 0 is not nilpotent, there exists an integer n such that n f $ I.
We may assume that nf ~ I, since otherwise fEN for any N E im and so
we have nothing to prove. Thus suppose p = I - (n f)(\ I> 0. Then,
for any positive integer m, we do not have m p > I. If otherwise, we
would have m-1 I < I - (n f) (\I and hence
(n f) 1\ I
(n f) 1\ (1- m-1) I.
Thus, by (6) in Chapter XII, 2,

(nf-(1-m- 1 )I) 1\ m-1 I=(nf-(1-m- 1 )I) 1\ 0<0,
and so, by the distributivity of the vector lattice,

0 = {(n f- (1-m-1) I) 1\ m-1 I} V 0
= (n f- (1-m-1) I)+ 1\ m-1 I,

that is, (n f- (1- m-1 ) I)+ 1\ I = 0. Put b = (n f- (1- m-1) I)+
and assume that b > 0. By hypothesis (1), we have b < od with some
IX> 1. Then 0 < b = b 1\ tXI and so 0 < (tX-1 b) 1\ I< b 1\ I, contrary
to b 1\ I= 0. Therefore b = 0, i.e., n f < (1- m-1) I. This contradicts
the fact that n f ;;I~ I. Hence the set N 0 = {g; Ig I < rJ IpI for some
rJ < oo} is a non-trivial ideal. N 0 is contained in at least one maximal
ideal N, by the second step. Then 0 = p (N) = 1- (n f(N)) 1\ 1 which
shows that f (N) > 0, that is, fEN.
We have thus proved our Theorem 1.
The vector lattice X= XJR is again a vector lattice with a unit I.
By Theorem 1, the intersection ideal n N of all the maximal ideals N
ii
of X is the zero ideal and X contains no nilpotent element

Hence X satisfies the Archimedean axiom
order-limitn- 1 1~
ntoo
= 0
for all
fE X.
=1=-
0.
(5)
Let N be any maximal ideal of X. Then the factor space XjN is a

simple vector lattice, and so, by the first step in the proof of Theorem 1, XJN is linear-lattice-isomorphic to the vector lattice of real
6. The Representation of a Vector Lattice as Set Functions
375
numbers; the non-negative elements and the unit being represented by

non-negative numbers and 1. We denote by l(N) the real number which
corresponds to 7 by this homomorphism X-+ XJN. We also denote by
IDe the set of all maximal ideals of X. We have thus
Theorem 2. By the correspondence 7-+ l(N), X is linear-lattice-isomorphically mapped on the vector lattice F (Wl) of real-valued bounded
1 on IDe and (iii)
functions on IDe such that (i) 171-+ ll{N) I, (ii) 1 (N)
F(Wl) separates the points of Wl in the sense that
for two different points Nv N 2 of IDe, there exists at

least one
7E X
such that 1(N1) =I=
f (2N) .
(6)
Remark. We introduce a topology in IDe by calling the sets of the form
{N E Wl; ll(N)- ~(N0) I< ei

where -1
(i
1, 2, ... , n),
< 7< 1 for all i}
neighbourhoods of N 0 Then rol is compact since it may be identified

with a closed subset of a topological product (of the same potency as
the cardinal number of the set of elements lEX which satisfy -1 < 7<1)
of the closed intervals [-1, 1]. The proof is entirely similar to the case
of the set of all maximal ideals of a normed ring in Chapter XI, 2.
Moreover, each function J(N) E F(Wl) is continuous on the compact space
IDe topologized in this way. Thus, by the Kakutani-Krein theorem in Chapter 0, 2, we see that F(Wl) is dense in the B-space C(IDC). The above two
theorems are adapted from K. YosmA-M. FuKAMIYA [16]. Cf. also
S. KAKUTANI [4] and M. KREIN-S. KREIN [2].
Let X be a a-complete vector lattice. Choose any positive element
x of X and call it a "unit" of X and write 1 for x; when not ambiguous
we-also write <X for <X 1. A non-negative element eEX is called a "quasiunit" if e 1\ (1- e) = 0. A finite linear combination ,Z
, <X1e1 of quasiunits e1 is called a "step-element", and we call the element y EX "absolutely continuous" (with respect to the unit 1) if y can be expressed as
the 0-lim of a sequence of step-elements. An element z EX is called
"singular" (with respect to the unit 1) if Iz I 1\ 1 = 0.
We shall give an abstract formulation of the Radon-Nikodym theorem
in integration theory.
Theorem. Any element of X is uniquely expressed as the sum of an
absolutely continuous element and a singular element.
Proof. The first step. If f > 0 and f (\ 1 =I= 0, then there exist a
positive number <X and a quasi-unit e!X =I= 0 such that f > <Xeot. We can,
376
in fact, take
e"' =
V {n(rx-1 /-rx-1 / 1\ 1) 1\ 1}.
(1)
n;;;::1
Proof. Put y"' = rx-1 / - rx-1 / 1\ 1. Then we obtain
={..;;;::1
V (2nya
/\ 2)} 1\ 1
and so e"' is a quasi-unit. We obtain f
> rxe"' from
2ea 1\ 1
+ n(rx- 1\ 1)] -n(rx-1 / 1\ 1)

(n + 1) rx-1 / 1\ (n + 1) -n(rx-1 / 1\ 1) < rx-1 /1\
ny"' 1\ 1 = nrx-1 / 1\ [1
<
= e"',
1/
1 < rx-1 j.
If we can show that:y"' 1\ 1 > 0 for some rx > 0, then e"' > 0 for such rx.
Suppose that such positive rx does not exist. Then, for every rx with
0 < rx < 1, we have
rx-1 (rx-1 / - rx-1 / 1\ 1) 1\ rx-1 = 0.

Hence (I -I 1\ rx) 1\ 1 = 0 and, letting rx
contrary to the hypothesis f 1\ 1 =I= 0.
0, we obtain f 1\ 1 = 0,
The second step. Let f > 0, and f > rxe where rx > 0 and e is a
quasi-unit. Then, for 0 < rx' < rx, e"'' > e and f > rx' eO<' where eO<' is
defined by (1).
Proof. For the sake of simplicity, we assumethatrx = 1. ForO<~< 1
1 f 15 + 1 ~ 15
1\ 1 = (1 f 15 + 1
>
(1
f 6
e 6) 1\ ( 1 + 1 f 15)
+ 1e
6) 1\ (1+ 1 e 15)
f
e
=1-<5/\ 1 +1-<5"
Since e is a quasi-unit, we have 2e 1\ 1 = e, e 1\ 1 =e. Hence me 1\ 1
= e when m > 1. Thus, by 1 < (1- ~)-1 , we have (1- ~)-1 e 1\ 1 = e.
Therefore, we obtain, from the above inequality,
<5
e
e
1 - <5 e = 1- <5- 1- <5 1\ 1
<
1 - <5 - 1-----;5 1\ 1 = Y1-"'
and so, by (1)., e < e1_.,.
The third step. The set of all quasi-units constitutes a Boolean

algebra, that is, if e1 and e2 are quasi-units, then e1 V e2 and e1 1\ e2 are
also quasi-units and 0 < e, :::::: 1. The quasi-unit (1 - e) is the complement of e, and 0,1 are the least and the greatest elements, respectively
in the totality of the quasi-units.
377
Proof. The condition e 1\ (1 - e) = 0 is equivalent to 2 e 1\ 1 = e.

Hence, if e1 , e2 are quasi-units, then
2{e1 1\
1\ 1 = (2e1 1\ 1) 1\ (2e 2 1\ 1) = e1 1\ e2 ,
2{e1 V e2 ) 1\ 1 = (2e1 1\ 1) V (2e2 1\ 1) = e1 V e2 ,
so that e1 1\ e2 and e1 V e2 are also quasi-units.
c;;2)
The fourth step. Let f > 0, and set 7 =sup (Jep where sup is taken
for all positive rational numbers (J. Because of the third step, the sup of
a finite number of elements of the form (JiefJ; is a step-element. Hence 1
is absolutely continuous with respect to the unit element 1. We have to
show that g = f -Tis singular with respect to the unit element 1. Suppose
that g is not singular. Then, by the first step, there exist a positive number
ex and a quasi-unite such that g >ex e. Hence f > cxe, and so, by the second
step, there exists, for 0 < cx1 < ex, a quasi-unit ea, > e such that f > cx1 ea,
We may assume that cx1 is a rational number. Thus f > cx1 ea, and so
I = 1+ g > 2cx1 ea, Again, by the second step, there exists, for 0 <
<X~ < <Xt> a quasi-unit e2"''1 > ea, such that I> 2cx~ e2 a'
We may assume
1
that 2cx~ is a rational number so that 1> 2cx~ e2"';. Hence f = 1+ g > 3cx{ e.
Repeating the process, we can prove that, for any rational number <Xn
with 0 < <Xn < <X,
+ 1) <Xne (n = 1, 2, ...)
have (n + 1) <Xne > 2-1 ncxe.
I> (n
> cx/2, we
Hence f > ncxe
(n = 1, 2, ... ) with ex> 0, e > 0. This is a contradiction. For, by the acompleteness of X, the Archimedean axiom holds in X.
The fifth step. Let f = f+ + j- be the Jordan decomposition of a
general element IE X. Applying the fourth step to f+ and j- separately,
we see that f is decomposed as the sum of an absolutely continuous element and a singular element. The uniqueness of the decomposition is
proved if we can show that an element hE X is = 0 if h is absolutely
continuous as well as singular. But, since his absolutely continuous, we
have h = 0-lim hn where hn are step-elements. hn being a step-element,
If we take <Xn
n--->oo
there exists a positive number <Xn such that Ihn I < <Xn 1. Since h is
singular, we have lhl 1\ lhnl = 0. Therefore lhl = lh! 1\ [h[ =
0-lim (lhl 1\ lhnl) = 0.
n---+00
Application to the Radon-Nikodym theorem. Consider the case X =

A (5, m). We know already (Proposition 1 in Chapter XII, 3) that
it is a complete lattice, and that in A (5, m),
x+ (B) = sup x (B') = the positive variation V (x; B) of

B'<;",B
We shall prepare a
x on B.
(2)
378
XII. Other Representation Theorems in Linear
Space~
Proposition. In A (S, 58), let x > 0, z > 0 in such a way that x 1\ z

= 0. Then there exists asetB0 E 58 such thatx(B0 ) = Oandz(S-B0 ) = 0.
+ z, we know, from
Proof. Since x 1\ z = (x- z)(x 1\ z)(B) =
inf (x-z)(B')
F'B
+ z(B) =
inf [x(B')
F'B
(2), that
+ z(B-B')].
(3)
Hence, by the hypothesis x 1\ z = 0, we have

inf [x(B)
BEta
+ z(S-B)] =
(x 1\ z)(S) = 0.
Hence, for any s > 0, there exists a B.E 58 such that x(B.) < s,
z(S- B.) < s. Put B 0 = 1\ ( V B 2
Then, by the a-additivity of
x(B) and z(B),
k~l
n~k
-n).
-n) <lim ..l: 2-n = 0,

V B -n) < lim z(S- B
00
0 < x(B0 ) =lim x ( V B 2

~
n~k
0 < z(S- B 0) = lim z(Sk~
n~k
k~n=k
Corollary. Let e be a quasi-unit with respect to x

Then there exists a set B 1 E 58 such that
e(B) = x(B
n B 1)
2 -k)
>
for all BE 58.
= 0.
0 in A (S, 58).
(4)
Proof. Since (x- e) 1\ e = 0, there exists a set B0 E 58 such that

e(B0 ) = 0, (x- e) (S- B 0 ) = 0. Hence e(S- B 0 ) = x(S- B 0 ) = e(S),
and so e(B) = x(B- B 0 ) = x(B n B 1) where B 1 = S- B 0 .
We are now able to prove the Radon-Nikodym theorem in integration
theory. In virtue of the above Corollary, a quasi-unit e with respect to
x > 0 is, in A (S, 58), a contracted measure e(B) = x(B n B,). Hence a
step-element in A (S, 58) is the integral of the form:
~ Ai
J x(ds),
Br.Bi
that is, indefinite integral of a step function (=finitely-valued function).

Hence an absolutely continuous element in A (S, 58) is an indefinite
integral with respect to the measure x (B). The Proposition above says
that a singular element g (with respect to the unit x) is associated with a
set B 0 E 58 such that x (B 0 ) = 0 and g (B) = g (B n B 0 ) for all BE 58. Such
a measure g(B) is the so-called singular measure (with respect to x(B)).
Thus any element IE A (S, 58) is expressed as the sum of an indefinite
integral (with respect to x(B)) and a measure g(B) which is singular (with
respect to x (B)). This decomposition is unique. The obtained result is
exactly the Radon-Nikodym theorem.
Remark. The above Theorem is adapted from K. YosiDA [2]. Cf.
F. RIESZ [6], H. FREUDENTHAL [2] and S. KAKUTANI [5]. For further
references, see G. BIRKHOFF [1].
379
1. The Markov Process with an Invariant Measure
XIII. Ergodic Theory and Diffusion Theory

These theories constitute fascinating fields of application of the
analytical theory of semi-groups. Mathematically speaking, the ergodic
t
theory is concerned with the "time average" lim t-1

ttoo
JT
0
ds of a semi-
group T 1, and the diffusion theory is concerned with the investigation of a

stochastic process in terms of the infinitesimal generator of the semi-group
intrinsically associated with the stochastic process.
1. The Markov Process with an Invariant Measure

In 1862, an English botanist R. BROWN observed under a microscope
that small particles, pollen of some flower, suspended in a liquid move
chaotically, changing position and direction incessantly. To describe
such a phenomenon, we shall consider the transition probability P(t, x; s, E)
that a particle starting from the position x at time t belongs to the set E
at a later times. The introduction of the transition probability P(t,x;s,E)
is based upon the fundamental hypothesis that the chaotic motion of the
particle after the time moment t is entirely independent of its past
history before the time moment t. That is, the future history of the particle
after the time moment t is entirely determined chaotically if we know
the position x of the particle at time t. The hypothesis that the particle
has no memory of the past implies that the transition probability P satisfies the equation
(1)
P(t, x; u, dy) P(u, y; s, E) for t < u < s,
s
where the integration is performed over the entire space 5 of the chaotic
movement of the particle.
The process of evolution in time governed by a transition probability
satisfying (1) is called a Markov process, and the equation (1) is called
the Chapman-Kolmogorov equation. The Markov process is a natural generalization of the deterministic process for which P(t, x; s, E) = 1 or= 0
according as y E E or y E E; that is, the process in which the particle at
the position x at the time moment t moves to a definite position y =
y (x, t, s) with probability 1 at every fixed later time moment s. The
Markov process P is said to be temporally homogeneous if P(t, x; s, E)
is a function of (s- t) independently of t. In such a case, we shall be
concerned with the transition probability P(t, x, E) that a particle at
the position x is transferred into the set E after the lapse of t units of
time. The equation (1) then becomes
P(t, x; s, E)=
P(t
+ s, x, E)= JP(t, x, dy) P(s, y, E)

s
for t, s > 0.
(2)
380
In a suitable function space X, P (t, x, E) gives rise to a linear transformation Te:

(Ttf)(x) =
P(t, x, dy) f(y), /EX,
(3)
s
such that, by (2), the semi-group property holds:
(4)
A fundamental mathematical question in statistical mechanics is

concerned with the existence of the time average
t
lim t-1
ttoo
J T.fds.
o
(5)
In fact, letS be the phase space of a mechanical system governed by the

classical Hanilltonian equations whose Hamiltonian does not contain the
time variable explicitly. Then a point x of S is moved to the point y1 (x)
of S after the lapse oft units of time in such a way that, by a classical
theorem due to LIOUVILLE, the mapping x~ y1 (x) of S onto S, for each
fixed t, is an equi-measure transformation, that is, the mapping x ~ y1 (x)
leaves the "phase volume" of S invariant. In such a deterministic case,
we have
(Ttl) (x) = f (y1 (x)) ,
(6)
and hence the ergodic hypothesis of BoLTZMANN that
the time average of any physical quantity= the space
average of this physical quantity
is expressed, assuming dx < oo, by
s
lim t-1
f (y. (x)) ds =
f (x) dxj dx for all f EX,
o
s
s
dx denoting the phase volume element of S.
(7)
A natural generalization of the equi-measure transformation
x~ y1 (x) to the case of a Markov process P(t, x, E) is the condition of
the existence of an invariant measure m (dx):
ttoo
Jm(dx) P(t, x, E) =
m(E) for all t
>
0 and all E.
(8}
We are thus lead to the

Definition. Let ~ be a a-additive family of subsets B of a set S such
that S itself E ~. For every t > 0, x E S and E E ~, let there be associated
a function P(t, x, E) such that
P(t, X, E)> 0, P(t, X, S) = 1,
for fixed t and x, P (t, x, E) is a-additive in E E ~,
for fixed t and E, P(t, x, E) is ~-measurable in x,
P(t + s, x, E)= P(t, x, dy) P(s, y, E)
s
(the Chapman-Kolmogorov equation).
(9}
(10)
(11)
(12)
1.
The Markov Process with an Invariant Measure
381
Such a system P(t, x, E) is said to define a Markov process on the phase

space (S, ~). If we further assume that (S, ~. m) is a measure space in
such a way that
(13)
Jm(dx)P(t,x,E)=m(E) forall EE~.
s
then P (t, x, E) is said to be a Markov process with an invariant measure
m(E).
Theorem 1. Let P (t, x, E) be a Markov process with an invariant
measure m such that m(S) <=Let the norm in Xp = LP(S, ~. m) be
denoted by II! liP p > 1. Then, by (3), a bounded linear operator
T 1 E L (Xp. Xp) is defined such that T 1+s = T 1 T 5 (t, s > 0) and
T 1 is positive, i.e., (Ttf) (x)
>
0 on S m-a.e. if f(x)
>
0 on S m-a.e.,
T 1 1 = 1,
(14)
(15)
II Ttl liP< II/IlP for /E Xp = LP(S, ~. m) with p = 1, 2 and=
(16)
Proof. (14) and (15) are clear. Let /E L (S, ~. m). Then, by (9), (10)
and (11), we see that /1 (x) = (Ttf) (x) E L 00 (S, ~. m) is defined and
ll!elloo< ll!lloo Hence, by (9) and (13), we obtain, for fEL 00 (S, ~.m)
with p = 1 or p = 2,
00
life liP={/ m (dx) If P(t,
< {/ m (dx)
X,
dy) f(y)nlfp
[! P(t, x, dy) lf(y) IP ! P(t, x, dy) 1PJYIP
=ffm(dy) lt(Y)IPtP= lltllp

We put, for a non-negative f E LP (S, 58, m) with p = 1 or p = 2,
nf(s) = min(l(s), n), where n is a positive integer.

Then, by the above, we obtain 0 < (nf(s))t < (n+lf(s))e and II (nf)t liP<
lim (nf(s)) 1, then, by the Lebes< II! liP Thus, if we put /1 (s) = 1>-->00
IIJIIP
.
gue-Fatou lemma, life liP< \\1\\p. that is, leE LP(S, 58, m). Again by the
Lebesgue-Fatou lemma, we obtain
f1 (x) =lim
1>-->00
.f P(t, x, dy) (J(y)) > s.f P(t,x,dy) (lim (J(y)))
n->00
P(t, X, dy) f(y).

s
Thus f(y) is integrable with respect to the measure P(t, x0 , dy) for those
x0 's for which / 1 (x0) =I=<=, that is, for m-a.e. x0 Hence, by the Lebesgue-Fatou lemma, we finally obtain
=
.f P(t, x0 , dy) (lim (J(y))) =
1>-->00
lim J P(t, x0 , dy) (J(y)).

1>-->00
382
Therefore, /1 (x0 ) = jP(t,x0 ,dy)f(y) m-a.e. and ll/ellp< II/IlP For a

s
general/ E LP (S, 58, m), we obtain the same result by applying the positive operator T 1 to j+ and j-, separately.
Theorem 2 (K. YosmA). Let P(t, x, E) be a Markoff process with an
invariant measure m such that m(S) < oo. Then, for any /E LP(S, 58, m)
with p = 1 or p = 2, the mean ergodic theorem holds:
n
s-lim n- 1 .I Tkf = /* exists in LP (S, 58, m) and T1 /* = /

n-+00
k=1
whenever f E LP (S, 58, m),
and we have
(17)
(18)
f(s) m(ds) = /* (s) m(ds).
s
s
Proof. In the Hilbert space L2 (S, 58, m), the mean ergodic theorem
(17) holds by (16) and the general mean ergodic theorem in ChapterVIII,3.
Sincem(S) < oo, we see, by Schwarz' inequality, that any/EL2(S, 58, m)
belongs to L1 (S, 58, m) and 11/11 1 < ll/11 2 m(S) 112 . Hence the mean
ergodic theorem lim
J It*
1:
(s)- n-1
(Tk /) (s) [' m (ds) = 0, together
s
k=1
with Ttf* = j*, hold for any /E L2(S, 58, m). Again, by m(S) < oo and
n-+00
0= lim ll/-n/11 1 =lim

n-+00
n-+00
J lf(s)-J(s) lm(ds)
with J(s) =min(/(s),n),
we see that L2(S, 58, m) is V-dense in V(S, 58, m). That is, for any
f E V (S, 58, m) and s > 0, there exists an/, E L2 (S, 58, m) (\ V (S, 58, m)
such that ll/-/.11 1 < s. Hence, by (16), we obtain
:11 n-1
!,
.i Tkj-n- I
1
k=l
k=1
Tkt.[.[' < 111-1.111 < s.

1
The mean ergodic theorem (17) in V(S, 58, m) holds for f. and :so, by
the above inequality, we see that (17) in V (S, 58, m) must hold also
for f.
Since the strong convergence implies the weak convergence, (18) is a
consequence of (17).
Remark I. The above The;>rem 2 is due to K. YosmA [17]. Cf. S. KAKUTANI [6], where, moreover, the m-a.e. convergence
n-1 k~ (Tkf) (x)
is proved for every /E L 00 (S, 58, m). We can prove, when the semi-group T 1
n
is strongly continuous in t, that we may replace s-lim n-1 .I Tk f in (17)

k=1
by s-lim t-1
ttoo
JT
o
fds. We shall not go into the details, since they are
given in the books on ergodic theory due to E. HoPF [1] and

K. JACOBS [1]. We must mention a fine report by S. KAKUTANI [8] on
2. An Individual Ergodic Theorem and Its Applications
383
the development of ergodic theory between HOFF's report in 1937 and

the 1950 International Congress of Mathematicians held at Cambridge.
In order to deal with the ergodic hypothesis (7), we must prove the
individual ergodic theorem to the effect that
n
lim n-1 .L (Tk f) (x) =
n->00
k~l
f* (x)
m-a.e.
In the next section, we shall be concerned with the m-a.e. convergence
,..
of the sequence n-1 .L (Tk f) (x). Our aim is to derive the m-a.e. converk~l
gence from the mean convergence using Banach's convergence theorem

in Chapter XII, 4.

We first prove
Theorem 1 (K. YosmA). Let X 1 be a real, a-complete F-lattice provided with a quasi-norm II x 111 such that
0-lim x,..= x implies lim llxn ll1
~00
~00
= llx ll1
(1)
Let a linear subspace X of X 1 be a real B-space provided with a norm

llx II such that
s-lim x,.. =X in X implies s-lim x,..
-oo
n->eo
=X
in
xl.
(2)
Let {T,..} be a sequence of bounded linear operators defined on X into X

such that
0-lim IT,..x I exists for those x's which form a set S of

n->00
the second category in X.
(3)
Suppose that, to a certain z EX, there corresponds a
lim
n->00
IIT,..z-:ZIJ =
T,..z=z
zEX such that
0,
(n=1,2, ... ),
0-lim (T,..z- T,..Tkz) = 0 for k = 1, 2, ...
(4)
(5)
(6)
n->00
Then
(7)
Proof. Put z
= z + (z- z). We define an operator TonS into X 1 by

Tx =0-lim T,..x-0-lim T,..x .
.......oo
Then, by (5),
0
- -
n->00
< T z < T (z - :Z)
(8)
384
We have T(z-Tkz) = 0 (k = 1, 2, ... ) by (6), and lim
k--+oo
II (z-z)-(z-Tkz) II
= 0 by (4). Hence, by Banach's theorem in Chapter XII, 4, we have

T(z- z) = 0. Hence 0 < Tz < 0, that is, 0-lim Tnz = w exists.
n-->00
We have to showthatw=z. We have, by(4)and(2), lim II Tnz-z lh=O.

Also we have, by {1) and 0-lim Tnz = w, lim
ft.--.+00
H-+00
n-+00
II Tnz- w 11 1 =
0. Hence
W=Z.
Specializing X 1 as the real space M(S, 58, m) and X as the real space
V(S, 58, m), respectively, we can prove the following individual ergodic
theorem.
Theorem 2 (K. Y OSIDA). Let T be a bounded linear operator defined
on V(S, 58, m) into V(S, 58, m) where m(S) <ex>. Suppose that
11rn11 < C <ex> {n = 1, 2, ... ),
(9)
lim In-1
n-->00
m=1
(T"' x) (s) I <
ex>
m-a.e.
(10}
Suppose that, for a certain zE V(S, 58, m),

lim n- 1 (rnz) (s) = 0 m-a.e.,
n-->00
and the sequence
{n- m~ T"'z} contains a subsequence

1
which converges weakly to an element

Then
n
s-lim n- 1 .I Tmz =
and
m=l
n-->00
z,
lim n- 1 .I (Tm z) (s) =

m=l
n-->00
(11)
zE V
(S, 58, m).
(12}
Tz = z,
(13)
z(s)
(14}
m-a.e.
Proof. Consider the space M(S, 58, m), and take the F-lattice X 1 =
M(S, 58, m) with llxll1 =
lx(s) I (1 + lx(s) 1)-1 m(ds), the B-lattice
s
X= L 1 (S, 58, m) with
llxll = sJ lx(s) I m(ds)
and Tn =
n- 1 m=1
.I Tm.
Then the conditions of Theorem 1 are satisfied. We shall, for instance,

verify that (6} is satisfied. We have
+ T 2 + + rn) z
n-1 (Tk+l + yk+2 + ... + yk+n) z
Tnz- TnTkz = n- 1 (T
_
and so, by (11), lim (Tnz- TnTk z) (s) = 0 m-a.e. for k = 1, 2, ...
n-->00
Hence, by taking the arithmetic mean with respect to k, we obtain

lim (Tnz- TnTkz) (s) = 0 m-a.e. for k = 1, 2, ... Conditions (4) and
n-->00
(5), that is condition (13) is a consequence of the mean ergodic theorem

in Chapter VIII, 3.
385
Remark. The above two theorems are adapted from K. YosiDA [15]
and [18]. In these papers there are given other ergodic theorems implied by Theorem 1.
As for condition (11) above, we have the following result due to

E. HoPF [3]:
Theorem 3. LetT be a positive linear operator on the real V (S, >S, m)
into itself with the V-norm [[ T [I< 1. If IEV(S, >S,m) and PEV(S, >S,m)
be such that p (s) > 0 m-a.e., then
~ (T"I) (s))j"g (Tjp) (s) = 0 m-a.e. on the set where p(s) > 0.
and T 1 = 1, then, by taking p (s) = 1, we obtain (11).
If m (S) <
(X)
Proof. It suffices to prove the Theorem for the case

e > 0 arbitrarily and consider the functions
n-l
gn = T" I- ,I y1 p, go
;=0
I > 0. Choose
=I
Let xn(s) be the defining function of the set {s E S; gn(s) > 0}. From
Xngn = g;t = max (g, 0) and gn+l + ep = T gn, we obtain, by the positivity of the operator T and [[T [[ < 1,
Jg;t+l m(ds) + sJXn+lP m(ds) = sJXn+l(gn+l + ep) m(ds)
Xn+lTg,. m(ds) < Xn+lTg;t m(ds) < Tg;t m(ds) < g;t m(ds).
s
s
s
s
Summing up these inequalities from n = 0 on, we obtain
=
J g;t m(ds) + e sJ p i: xk m(ds) < sJg;j m(ds),
s
and hence
k=l
J p . .I xk. m(ds) <

s
00
k=l
00
Jg;j. m(ds).
converges m-a.e. on the set where p (s)
Hence .I
xk (s)
set where
p (s) >
k=l
e-1
0, we must have g,. (s)

0. Thus, on the
0 m-a.e. for all large n.
Therefore, we have proved that (T" I) (s)

all large non the set where
proved our Theorem.
>
n-l
.I
< s k=O
(Tkp) (s) m-a.e. for
0. As e > 0 was arbitrary, we have
As for condition (10), we have the following Theorem due to

R. V. CHACON-D. S. ORNSTEIN [1]:
Theorem 4 (R. V. CHACON-D. S. ORNSTEIN). Let T be a positive
linear operator on the real V(S, >S, m) into itself with the V-norm
25
386
//T// 1
< 1. If/and pare functions in Ll(S, 7B,m) and p(s) > 0, then
~ t~ (Tkf) (s) Jk..to (Tkp)
(s)} is finite m-a.e.
00
.I (T'p) (s) > 0.
on the set where
n~o
If m(S) < oo and T 1 = 1, then, by taking

For the proof, we need a
p (s) = 1, we obtain (10).

-
Lemma(CHACON-0RNSTEIN[1]). If/=f++f-,andif~
1! (Tkf)(s) > 0
n
on a set B, then there exist sequences {dk} and {/k} of non-negative functions such that, for every N,
N
J _Idkm(ds)+JINm(ds)<Jf+m(ds),
s k~O
S
S
00
.I dk(s)
k~O
(15)
j-(s) on B,
=-
(16)
=.I TN-kdk +IN
TNj+
(17)
k~O
Proof. Define inductively
do= 0, /o = f+, f-1 = 0,
/'+ 1 = (T/; +
+ d0 + + d;)+, di+ 1 = T/;-
Note that
r + do +
+ d;
< 0'
/'+ 1 .
(18)
(19)
and that the equality holds on the set where /; (s)
>
0, for
1- +do + + d;- 1)+
/; = (Tfi-1 +
= (T/;-1- /; +
r +do+ +
di-1
+ /;)+
= (d; + j- + d0 + + d;-1 + /;)+.

It follows from (18) that
j
T1 t+ = k~ p-kdk
+ fj
(20)
By the definition, /; is non-negative, and so is d; by the last two equations of (18) and (19). From (20), we have
n
.IPj+
We then prove
j~O
=.I .I p-kd +.If-.

1
j~O k~O
j~O
2: 'Pt+ < 2: dj- 2: 'Pr + 2: tj
j~O
j~o
j~l
j~O
To this purpose, we note that
1: 1 p-k dk = 1: Ti ( Ii dk)
j~Ok~O
j~O
k~O
(21)
(22)
387
and that, by (19) and the positivity of T,
- yi
> yi ( :ij ak) for 1 ::;: i
k=O
;: n.
(23)
Rewriting (22), we have

"
...J: P
J=O
...J: (aj + tj) + r.

u+ + rl < J=O
n
(24)
We will now prove that

00
...J: dj(s)
J=O
+ /-(s) >
0 m-a.e. on B.
(25)
It is clear from the remark after (19) that (25) holds with equality m-a.e.
on the set C = {s E S; fi (s) > 0 for some j > 0}. It remains to show that
(25) holds on the set B -C. This is proved by noting that (24) implies
that on B, and on B- C in particular, we have the inequality
00
...J: (dj
J=O
+ /j)
(s)
+ r(s) >
0.
Now we note that (17) is exactly (20) and that (16) is implied from
(19) and (25). To see that (15) holds, note that we have, by the assumptions on T and (18),
(1ak+T/j)m(ds) = J(1/ak+fi+ 1)m(ds).

J(lak+ti)m(ds)>
k=O
5
k=O
5
s k=O
Hence we obtain (15) by induction on j, since d0 + /0 = j+.

Proof of Theorem 4. It is sufficient to prove the Theorem under the
hypothesis that f (s) 2 0 on S, and only to prove the finiteness of the
indicated supremum when p(s) > 0. The first remark is obvious, and
the second is proved as follows. By the hypothesis that the indicated
supremum is finite at point s where p (s) > 0, we have
lim
{.i ('P+kj) (s)j.i ('P+kp)

J=O
J=O
on the set where (Tkp) (s)

and this implies that lim
>
(s)} is finite m-a.e.
{.i (Ti f) (s)j.i (Pp)

J=O
(26)
0,
J=O
(s)} is finite m-a.e. on
the set where (Tkp) (s) > 0.

Now assume the contrary to the hypothesis. Then
~~ L~ (Ti f) (s) ji~ (Tip) (s)}

is infinite m-a.e. on a set E of positive m-measure, and p (s) > {3 > 0 on
E for some positive constant {3. We have therefore, that for any positive
constant <X,
25*
388
m-a.e. on E. Applying the Lemma with f replaced by(/ -1Xp), we obtain

from (15) and (16) that
j(f-1Xp)+m(ds)
J ,Idkm(ds)>- Ej(f-1Xp)-m(ds).
S k=O
00
>
However, the extreme right term tends to oo as IX too, and the extreme
left term is bounded as IX too. This is a contradiction, and so we have
proved the Theorem.
We have thus proved
Theorem 5. Let T be a positive linear operator on V(5, )8, m) into
itself with the V-norm II T 11 1 < 1. If m (5) < oo and T 1 = 1, then, for
any /E V(5, )8, m), the mean convergence of the sequence {n- 1
.i P t}
J=l
implies its m-a.e. convergence.

As a corollary we obtain
Theorem 6. Let P(t, x, E) be a Markov process with an invariant
measure m on a measure space (5, m, m) such that m(5) < oo. Then,
for the linear operator T 1 defined by (Ttf) (x) =
P(t, x, dy) f(y), we
s
have i) the mean ergodic theorem:
" Tkf=f*
for any /ELP(5, )S,m), s-limn- 1 .I
n->00
k=l
exists in LP(5, m, m) and Tlf* =I* (p = 1, 2),
(27)
and ii) the individual ergodic theorem:

for any /E LP(5, )8, m) with
p=
1 or
p=
2,
finite lim n-1 .I (Tk f) (s) exists m-a.e. and is= f* (s) m-a.e.,
n->00
k=l
and moreover,
f (s) m (ds) =
f* (s) m (ds).
(28)
s
s
Proof. By Theorem 2 in the preceding section, the mean ergodic theorem (27) holds and so we may apply Theorem 5.
Remark. If T 1 is given by an equi-measure transformation x-+ y1 (x)
of 5 onto 5, the result (27) is precisely the mean ergodic theorem of
J. VON NEUMANN [3] and the result (28) is precisely the individual
ergodic theorem of G. D. BrRKHOFF [1] and A. KHINTCHINE [1].
A Historical Sketch. The first operator-theoretical generalization of
the individual ergodic theorem of the Birkhoff-Khintchine type was given
" (Tk f) (x) converges m-a.e.

[1]. He proved that n-1 .I
k=l
when T 1 is defined by a Markov process P(t, x, E) with an invariant
measure mona measure space (5, )8, m) such that m(5) = 1 and f is the
defining function of a set Em. It was remarked by S. KAKUTANI [6] that
by
J. L. DooB
389
3. The Ergodic Hypothesis and the H-theorem
Doob's method is applicable and gives the same result for f merely a
bounded \a-measurable function. E. HoPF [2] then proved the theorem
assuming merely that f is m-integrable. N. DuNFORD-]. ScHWARTZ [4]
extended Hopf's result by proving (28) for a linear operator T 1 which
increases neither the L 1 - nor L 00 -norm, without assuming that T 1 is positive but assuming that Tk = T~ and T1 1 = 1. It is noted that Hopf's and
the Dunford-Schwartz arguments use the idea of our Theorem 1. R. V. CHACON-D. S. ORNSTEIN [1] proved (28) for a positive linear operator T 1 of
L 1-norm < 1, without assuming that T 1 does not increase the L 00 -norm,
and, moreover, without appealing to our Theorem 1. Here, of course, it is
assumed that Tk = T~ and T 1 1 = 1. We will not go into details, since,
for the exposition of this Chapter which deals with Markov processes,
it sufficies to base the ergodic theory upon Theorem 6 which is a consequence of our Theorem 1.

Let P (t, x, E) be a Markov process with an invariant measure m on a
measure space (5, \8, m) such that m(S) = 1. We shall define the ergodicity of the process P(t, x, E) by the condition:
the time average
/* (s} = n-->-00
lim n-1 .E (Tk f) (x)
k~l
= lim
n->00
n-1
i: sJP(k, x, dy) f(y)
k~1
f(x) m(dx) m-a.e.

s
for every f E LP (S, \8, m) (p = 1, 2}.
Since
f* (x) m (dx} =
f (x) m (dx), (1) may be rewritten as
=the space average
(1)
f* (x) = a constant m-a.e. for every f E LP (S, \8, m) (p = 1, 2). (1')

We shall give three different interpretations of the ergodic hypothesis
(1} and (1').
1. Let XB (x) be the defining function of the set BE \8. Then, for
any two sets Bv B 2 E fB, the time average of the probability that the points
of B 1 will be transferred into the set B 2 after k units of time is equal to the
product m(B1) m(B 2}. That is, we have
(x;,, XB) = m (B1) m (B2).
(2)
n
Proof. If/, g E L2(5, \8, m), then the strong convergence n-1 .E Tkf~f*
k=1
in L 2 (S, \8, m) implies, by (1'),

lim n-1
k~1
(Tk f, g)
= (!*, g)
=
I* (x)
Jg (x) m (dx}
Jf (x) m (dx) sJg (x) m (dx)
By taking I (x) = Xn, (x), g (x) = XB, (x), we obtain (2}.
for m-a.e. x.
390
Remark. Since linear combinations of the defining functions XB (x)

are dense in the space LP (S, 58, m) for p = 1 and p = 2, we easily see
that (2) is equivalent to the ergodic hypothesis (1'). (2) says that every
part of Swill be transferred uniformly into every part of Sin the sense
of the time average.
2. The mean ergodic theorem in Chapter VIII, 3 says that the mapping f--+ T* f = f* gives the eigenspace of T 1 belonging to the eigenvalue
1 of T 1 ; this eigenspace is given by the range R(T*). Hence the ergodic
hypothesis (1)' means precisely the hypothesis that R (T*) is of onedimension. Thus the ergodic hypothesis of P (t, x, E) is interpreted in
terms of the spectrum of the operator T 1 .
3. The Markov process P(t, x, E) is said to be metrically transitive
or indecomposable if the following conditions are satisfied:
S cannot be decomposed into the sum of disjoint sets

B 1 , B 2 E 5S such that m(B1) > 0, m(B 2 ) > 0
and P(1, x, E) = 0 for every x E B,, E ~ B; with i =1=- f.
(3)
Proof. Suppose P(t, x, E) is ergodic and that Sis decomposed as in

(3). The defining function Xs, (x) satisfies T 1 XB, = Xs, by the condition
JP (1, x, dy) Xs, (y) =
P (1, x, B 1 ) = 1 = Xs, (x), for x E B 1 ,

= 0 = Xs, (x), for x E B 1 .
But, by m (B1) m (B 2 ) > 0, the function Xs, (x) = x~. (x) cannot be equal
to a constant m-a.e.
Next suppose that the processP(t,x,E) is indecomposable. Let Ttf = f
and we shall show that f* (x) = f (x) equals a constant m-a.e. Since
T 1 maps real-valued functions into real-vaiued functions, we may assume
that the function f (x) is real-valued. If f (x) is not equal to a constant
m-a.e., then there exists a constant <X such that both the sets
B 1 ={sES;f(s)><X} and B 2 ={sES; /(s)<<X}
are of m-measure > 0. Since T 1 (/-<X) = /-<X, the argument under The Angle
Variable of p. 391 implies that T 1 (/-<X)+= (1-<X)+, T 1 (1-<X)- = (1-<X)-.
Hence we would obtain P (1, x, E) = 0 if x E B,, E ~ B; (i =1=- f).
Remark. The notion of the metric transitivity was introduced by
G. D. BIRKHOFF-P. A. SMITH [2] for the case of an equi-measure transformation x--+ y1 (x) of S onto S.
An Example of the Ergodic Equi-measure Transformation. Let S be a
torus. That is, Sis the set of all pairs s = {x, y} of real numbers x,y such
that s = {x, y} and s' = {x', y'} are identified iff x- x' (mod 1) and
y- y' (mod 1); S is topologized by the real number topology of its
coordinates x, y. We consider a mapping
s = {x, y}--+ T 1 s = s1 = {x + t<X, y + t{J}
391
of S onto S which leaves invariant the measure dx dy on S. Here we

assume that the real numbers <X, fJ are integrally linearly independent
0
mod 1 in the following sense: if integers n, k satisfy n<X + k{J
ergodic.
is
s
T
mappings~
(mod 1), then n = k = 0. Then the
1
Proof. Let l(s) E L2(S) be invariant by Tv that is, let l(s) = l(s1 )
dx dy-a.e. on S. We have to show that l(s) =I* (s) =a constant dx dya.e. on S. Let us consider the Fourier coefficients of l(s) and l(s1) =
I(T1 s), respectively.
1 1
JJ l(s) exp(-2:ni(k1 x + k y)) dxdy,

0 0
2
1 1
JJ I(T s)exp(-2:ni(k x+k y))dxdy.
0 0
By T 1 s = {x +<X, y + [J} and the invariance of the measure dx dy by the

T 1 s, the latter integral is equal to
mappings~
1 1
JJ l(s) exp(-2:ni(k x + k y}) exp(2:ni(k + k

1
0 0
1 <X
{J)) dxdy.
Hence, by the uniqueness of the corresponding Fourier coefficients of

l(s) and I(T1 s), we must have
+ k2 [J}) = 1 whenever
JJ l(s) exp(-2:ni(k1 x + k2 y)) dx dy-:/= 0.
exp (2:ni (k1 <X
1 1
0 0
Hence, by the hypothesis concerning <X, {J, we must have
1 1
J l(s} exp(-2:ni(k x + k
0 0
y}) dxdy = 0 unless k1 = k 2 = 0.
Therefore I (s) = I* (s) must reduce to a constant dx dy-a.e.

The Angle Variable. Let T 1 be defined by the Markov process
P(t, x, E) with an invariant measure m on (S, 58, m) such that m (S) = 1.
Let l(s) E L2(S, 58, m) be the eigenvector of T 1 pertaining to the eigenvalue A. of absolute value 1 of T 1 :
T1l = A.l, I-" I= 1.

Then we have T 1 Ill = Ill For, we have, by the positivity of the operator
P(1, x, dy) ll(y) I>
T1 , (T1 Ill) (x) > I(T1 1) (x) I= ll(x) I, that is,
s
ll(x} I In this inequality, we must have the equality for m-a.e. x, as may
be seen by integrating both sides with respect to m (dx) and remembering
the invariance of the measure m. Hence, if P (t, x, E) is ergodic, then
ll(x) I must reduce to a constant m-a.e. Therefore, if we put
l(x) = ll(x) I exp (i8(x)), 0 < 8(x)
<
2:n,
392
we must have
T 1 exp (i
In case A=!= 1, such

process P(t, x, E).
e (x)) =A exp (i e (x)).
e (x) is called an angle variable of the ergodic Markov
The Mixing Hypothesis. Let T 1 be defined by a Markov process

P(t, x, E) with an invariant measure m on (S,
m) such that m(S) = 1.
A stronger condition than the ergodic hypothesis of P (t, x, E) is given by
m,
lim (Tk /,g)=(/*, g)=
f(x) m(dx) g(x) m(dx)

s
s
m).
for every pair {/, g} of vectors E L2 (S,
k-..oo
m,
(4)
This condition is called the mixing hypothesis of the Markov process

P(t, x, E). As in the case of the ergodic hypothesis, it may be interpreted
as follows: every part of S will be transferred into every part of S uniformly
in the long run. As for the examples of the mixing equi-measure transformation x--+ y1 (x), we refer the reader to the above cited book by E. HoPF.
H-theorem. Let P (t, x, E) be a Markov process with an invariant
measure m on (S,
m) such that m (S) = 1. Let us consider the function
m,
H(z) =-zlogz, z> 0.
(5)
We can prove
Theorem (K. YosmA [17]). Let a non-negative function f (x) E

V(S,
m) belong to the Zygmund class, that is, let us assume that
I (x) log+ I (x) m (dx) < oo, where log+ Iz I = log Iz I or = 0 according
s
as Iz I > 1 or Iz I < 1. Then we have
m,
H(t(x)) m(dx) <

H((Ttf) (x)) m(dx).
s
s
Proof. Since H (z) = - z log z satisfies H" (z) = - 1/z < 0 for z
H (z) is a concave function. Hence we obtain
(6)
>
0,
the weighted mean of H (t(x)) ~ H (the weighted mean of f(x)),

and so
P(t, x, dy) H(t(y)) ~ H(f P(t, x, dy) I(Y)).
Integrating with respect to m(dx) and remembering the invariance of the

measure m(E), we obtain (6).
Remark. By virtue of the semi-group property Tt+s

(6), we easily obtain
T 1 T 5 and
H((T1J) (x)) m (dx)< H((Tt)) (x)) m(dx) whenever t 1 < t2

(6')
s
s
This may be considered as an analogue of the classical H -theorem in
statistical mechanics.
393
4. The Ergodic Decomposition of a Markov Process

with a Locally Compact Phase Space
Let S be a separable metric space whose bounded closed sets are
compact. Let ~ be the set of all Baire subsets of S, and consider a Markov
process P(t, x, E) on (S, ~).We assume that
(1)
P(t, x, dy) f(y) E C8(S).
s
The purpose of this section is to give the decomposition of S into ergodic
parts and dissipative part as an extension of the Krylov-Bogolioubov case
of the deterministic, reversible transition process in a compact metric
spaceS (see N. KRYLOV-N. BoGOLIOUBOV [1]). The possibility of such
an extension, to the case where S is a compact metric space and with
condition (1), was observed by K. YosiDA [17] and carried out by
N. BEBOUTOV [1], independently of YosiDA. The extension to the case
of a locally compact space S was given in K. YosmA [19]. We shall
follow the last cited paper.
f(x) E C8(S) implies / 1 (x) =
Lemma 1. Let a linear functional L (f) on the normed linear space

C8(S),normed by the maximalnorm II!//= sup /f(x)J, be non-negative,
~ES
viz., L (f) > 0 when f (x) > 0 on S. Then L (f) is represented by a uniquely
determined regular measure rp (E), which is a-additive and > 0 for Baire
subsets E of S, as follows:
L (f) =
f I (x) rp (dx)
for all
! E cg (S).
Here the regularity of the measure rp means that

rp (E) = inf rp (G) when G ranges over all open sets
(2)
E.
(3)
Proof. We refer the reader to P.R. HALMOS [1].
on
We shall call a non-negative, a-additive regular measure rp (E) defined

.such that rp (S) < 1 an invariant measure for the Markov process if
rp(E) =
f rp(dx) P(t, x, E)
for all t > 0 and E E ~.
(4)
We have then
Lemma 2. For any
lim n-1
~oc,
k=l
/E C8(S}
and for any invariant measure rp(E),
fk(x) = f*(x) (tk(x) = jl(k, x, dy)
/(y)) exists
(5)
rp-a.e. and
Jf* (x) rp (dx) = sJf (x) rp (dx).
(6)
394
This is exactly a corollary of Theorem 6 in the preceding section.

Now, let 1/, 2/, 3/, . . . be dense in the normed linear space C8(5). The
existence of such a sequence is guaranteed by the hypothesis concerning the space 5. Applying Lemma 2 to 1/, J, af, ... and summing up the
exceptional sets of <p-measure zero, we see that there exists a set N of <pmeasure zero with the property:
for any x ~ N and for any f E cg (5), lim n-1 I
k=1
11-+00
fk (x) = f* (x) exists. (7)
A Baire set 5' ~ 5 is said to be of maximal probability if <p (5 - 5') = 0

for every invariant measure <p. Thus the set 5' of all x for which I* (x) =
n
lim n-1 Ilk (x) exists for every IE
11-+00
k=1
cg (5)
is of maximal probability.
Hence if there exists an invariant measure with 
>
0, then there
0.
(8)
For, if otherwise, we would have f* (x) = 0 on 5 for all f E cg (5) and

hence, by (6), I (x) <p (dx) = 0.
s
Let, conversely, (8) be satisfied for a certain g E cg (5) and fm a certain x0 Let a subsequence {n'} of natural numbers be chosen in such a
n'
way that lim (n')-1 .:E gk (x0) = g** (x0). By a diagonal method, we
n'-+oo
k=1
n"
may choose a subsequence {n"} of {n'} such that n~ (n")-1 k~ (jf)k (x0)
exists for j = 1, 2, ... By the denseness of
that
n"
lim (n")-1 I fk (x0 )
n"---+00
If we put l***(x0 )
k=1
U} in
C8(5), we easily see
j*** (x0 ) exists for every f E cg (5).
Lx,(l), then
Lx, (/) = Lx, (/1)

For, by condition (1), we have 11 E cg (5) and so
(9)
By Lemma 1, there exists a regular measure <J?x, (E) such that
Lx, (f)
Surely we have
= f*** (x0 ) =
Jf (x) <J?x, (dx).
0 < <J?x,(E)
< 1,
(10)
(11)
395
and, by (9),
Jl(x) q;" (dx) = J (JP(1, x, dy) I(Y)) q;" (dx).
s s
s
By letting I (x) tend to the defining function of the Baire set E, we see
that q;",(E) is an invariant measure. We have q;",(S) > 0, since, by (8)
and (10), we have L",(g) = g***(x0) = g(x) q;",(dx) > 0. We have
s
thus proved
Theorem 1. A necessary and sufficient condition for the non-existence of non-trivial invariant measures is
IS
lim n-1 .I 1,. (x) =

11-1
IS-+00
I* (x)
= 0 on
s for any IE cg (S).
Definition. We will call the process P(t, x, E) on (S,

if condition (12) is satisfied.
~)
(12)
dissipative
Example. LetS be the half line (0, oo) and ~ the set of all Baire
sets of (0, oo). Then the Markov process P(t, x, E) defined by
P (t, x, E) = 1 if (x + t) E E and = 0 if (x + t) E E
is dissipative.
We shall assume that P (t, x, E) is not dissipative. Let D denote the set
1:
l 11 (x)
n-1
{xES; l*(x) =lim
11=1
~00
Since it is equal to
{xES; (jf)* (x) = 0 for
for all
f=
IE C8(S)}.
1, 2, ...},
Dis a Baire set. We shall call D the dissipative part of S. We can prove
below that q; (D) = 0 for any invariant measure g;, and so, since the
process P(t, x, E) is assumed to be not dissipative, 5 0 = S- D is not
void. We already know that there exists a Baire set 5 1 ~ 5 0 with the
property:
to any x E 5 1 , there corresponds a non-trivial invariant measure
IS
q;" (E) such that I* (x) = lim n-1 I; 1,. (x)

IS-+OO
k=l
f I(z) q;" (dz)
and q;(S0 - 5 1)= 0 for any invariant measure q;.
(13)
For any invariant measure q;, we have (6) and hence, by (13),
Thus we obtain
JI (y) q; (dy) = s,J (sJI (z) f/Jy (dz)) q; (dy).
q; (E) =
and so we have proved
f f/Jy (E) q; (dy),
s,
(14)
396
Theorem 2. Any invariant measure cp may be obtained as a convex

combination of the invariant measure C{Jy (E) with y as a parameter.
We see, from (11) and (14), that the set
52
= {xES; xE Sv CfJx(Sl) = 1}
is of maximal probability. Hence 5 0 is of maximal probability.

For any /E C8(S) and for any invariant measure cp, we have
cp (dx) ( (I* (y) -/* (x)? CfJx (dy))

s,
s,
since the left hand side is equal to
cp (dx) (f!* (y) 2 CfJx (dy))-
(15)
0,
21/* (x) cp (dx) (ft* (y) CfJx (dy))
+ J/* (x) 2 cp (dx) JCfJx (dy)

=
J/* (y)2 cp (dy) -
52
J/* (x)
S1
cp (dx)
J f* (x)
cp (dx) ,
by (13), (14), (6) and the definition of 5 2 . By applying (15) to 1/,

we see that the set
Sa= {x E 5 2 ;
f (!*
J, 3/, . . ,
(y) -I* (x)) 2 CfJx (dy) = 0 for all IE cg (S)}
is of maximal probability.
We shall give the ergodic decomposition of S. For any x E Sa, put
Ex= {y E Sa; I* (y) =I* (x)
for all IE cg (S)}.
(16)
We can then show that each Ex contains a set Ex with the property:
A
CfJx(Ex)=cpx(Ex) and P(1,y,Ex)=1 forany yEEx.
(17)
Proof. By the definition of Sa, we see that/* (y) = f* (x) if the measure
CfJx (E) has variation at the point y. Hence CfJx (Ex) = CfJx (Sa) = 1. Thus, by
the invariance of the measure CfJx, we obtain
1 = C{Jx(Ex) =
JP(1,
S,
Z,
Ex) CfJx(dz) =
JP(1,
Z,
Since 0 < P(1, z, Ex)< 1, there exists a Baire set E1
CfJx (El)
~Ex
CfJx (Ex) and z E E1 implies P (1, z, Ex)
Put
E2 = {zE El, P(1, z, El) = 1}.

Since
Ex) CfJx(dz).
Ex
such that
1.
397
we must have
J (P(1,
Z,
E'
E,J -P(1, z, 1)) ffJx(dz) = 0.
As z E E2 implies P (1, z, 1) = 1 by the definition of E 2 , we obtain
ffJx (El - E2) = 0.

Next put
3 = {zE 2; P(1, z, 2) = 1}.

Then, as above, ffJx(E 3) = ffJx(Ex) In this way, we obtain a sequence {En}
such that
E,.
~ 1 ~
E2 ~ , ffJx(En)
cp,.(Ex) and
P(1,z,En)=1 if zEEn+k (n>O,k>1,E 0 =Ex)

00
We have thus obtained the set Ex=
n En with the demanded property.
n~l
Therefore, we have
Theorem 3. P(t, y, E) defines a Markov process in each E" which is
ergodic, in the sense that
Ex is not decomposable in two parts A and B such

that
cp (A) . cp (B) > 0 for a certain invariant
measure cp and P(1, a, B) = 0 when a E A and
(18)
P(1,b,A)=0 when bEB.
Proof. Let cp be any invariant measure in Ex with cp (Ex) = 1. Then,

by (14), cp (E) =
cp (dz) g;, (E) whenever E ~ Ex. Since cp (E) = ffJx (E)
Ex
for every z E Ex by the definition of Ex we have
cp (E) = f/Jx (E)
j cp (dz) =
Ex
IPx (E) .
Thus there is essentially only one invariant measure IPx (E) in E,..
This proves the ergodicity. For, let Ex be decomposed as in (18). Then,
by the invariance of cp,
cp(C) =
P(1, z, C) cp(dz)
Ex
Thus the measure 1p defined by
whenever
1p(C) = cp(C)jcp(A)
if
C ~Ex.
~A
= 0 if C ~ B
is invariant in Ex and differs from the unique invariant measure cp.
398
o.
The Brownian Motion on a Homogeneous Riemannian Space
There is an interesting interplay between Markov processes and

differential equations. Already in the early thirties, A. KoLMOGOROV
proved, under a certain regularity hypothesis concerning the transition
probability P(t, x, E), that
u(t, x) =
P(t, x, dy) f(y)

s
satisfies the equation of the diffusion type:
au =
at
bij (x)
~'~!:_
ax;8x1
+ ai (x) ax,
au =
Au, t > 0,
(1)
where the differential operator A is elliptic in the local coordinates

(xv x 2 , , xn) of the point x of the phase space 5. Following Einstein's
convention of the tensor notation, it is understood that, e.g., ai (x) 8f8x;
n
means ..I a (x) 8f8xi. For the derivation of equation (1), see A. KoLM0=1
GOROV [1]. The leit-motif in his research was the investigation of local
characteristics of Markov processes.
Following work done by E. HILLE at Scandinavian Congress in 1949
(Cf. E. HILLE [9]) and K. YOSIDA [29] in 1948, W. FELLER [2] began in
1952 a systematic research of this new field in probability theory using
the analytical theory of semi-groups. His investigations were further developed by E. B. DYNKIN [1], [2], K. ITo-H. McKEAN [1], D. RAY, [1]
G. A. HUNT [1], A. A. YusHKEVITCH [1], G. MARUYAMA [1] and many
younger scholars, especially in Japan, USSR and USA. These investigations are called the theory of diffusion. A comprehensive treatise on the
diffusion theory by K. ITo-H. McKEAN [1] is in the course of printing 1 .
We shall sketch some of the salient analytical features of the theory.
Let 5 be a locally compact space such that ~ is the totality of the
Baire sets in 5. To define the spatial homogeneity of a Markov process
P(t, x, E) on the space 5, we suppose that 5 is ann-dimensional, orientable connected C00 Riemannian space such that the full group @ of the
isometries of 5, which is a Lie group, is transitive on 5. That is, for each
pair {x, y} of points E 5, there exists an isometry ME@ such that
M x = y. Then the process P (t, x, E) is spatially homogeneous if
P(t, x, E)= P(t, M x, M E) for each x E 5, E E ~and ME@.
(2)
A temporally and spatially homogeneous Markov process on 5 is

called a Brownian motion on 5, if the following condition, known as the
continuity condition of Lindeberg's type, is satisfied:
limt-1
q
.
dlS(x,y) >
P(t, x, dy) = 0
for every s
Berlin/Gottingen{Heidelberg: Springer.
>
0 and xE 5.
(3}
5. The Brownian Motion on a Homogeneous Riemannian Space
399
Proposition. Let C (5) be the B-space of all bounded uniformly

continuous real-valued functions f(x) on 5 normed by II/ IJ =sup lf(x) I
Define
(Ttf) (x)
=
=
if
j P(t, x, dy) f(y),
f(x), if t
>
0,
(4)
0.
Then {T1} constitutes a contraction semi-group of class (C0) in C (5).

Proof. By P(t, x, E)> 0 and P(t, x, 5) = 1, we obtain
I(Tt f) (x) I <
sup
:Y
II (y) I
and the positivity of the operator T 1 The semi-group property Tt+s = T 1 T 5

is a consequence from the Chapman-Kolmogor ov equation. Ifwedefinea
linear operator M' by (M' f) (x) = f (M x), ME @, we obtain
(5)
For,
(M' T 1 f) (x) = (Ttf) (M x) = j P(t, M x, dy) f(y)
=
=
jP(t,Mx,d(My ))f(My)
s
JP(t, x, dy) f(M y) =
(T1 M'f) (x).
If ME@ be such that M x = x', we have

(T 1 f) (x) - (T1 f) (x')
(T1 f) (x) - (M' T 1 f) (x)
T 1 (f- M' f) (x) .
Hence by the uniform continuity off (x), we see that (Td) (x) is uniformly
continuous and bounded.
To prove the strong continuity in t of T 1, it suffices by the
Theorem in Chapter IX, 1, to verify weak right continuity of T 1 at t = 0.
Therefore, it is surely enough to show that lim (T1 f) (x) = f (x) uniformly
t.j.O
in x. Now
I(Ttf)(x)-f(x)l
 s
P(t,x,dy)[f(y)-f( x)J\+2\Ifll
P(t, x, dy) [f(y)- f(x)] I
d(x,y) > e
P(t,x,dy).
The first term on the right tends to zero uniformly in x as e t 0 and, for fixed
e > 0, the second term on the right tends to zero uniformly in x as t t 0.
The latter fact is implied by (3) and the spatial homogeneity. Thus
lim (T 1 f)(x) = f(x) uniformly in x.
tj,O
400
Theorem. Let x0 be a fixed point of S. Let us assume that the group of

isotropy @0 = {ME G; M x0 = x0} is compact. Being a closed subgroup
of a Lie group@, @0 is a Lie group by a theorem due to E. CARTAN. Let A
be the infinitesimal generator of the semi-group T 1 Then we have the
following results.
[1].
at x0 ,
Iff ED (A) n C2 (5), then, for a coordinate system (xA, x5, ... , x0)
(6)
where
(x0) = lim t-1

q 0
bii (x0 ) =lim t-1

qo
d(....,...
(xi - xb) P (t, x0 , dx),
(7)
(xi- xb) (xi- x6) P (t, x0 , dx),
(8)
d( ....,...t) ;;;;.
>;;>
the limits existing independently of sufficiently small e > 0.
[2]. The set D (A) n C2 (S) is "big" in the sense that, for any coo (S)
function h (x) with compact support there exists an f (x) E D (A) n C2 (S)
such that f(x0 ), 8ff8xb, 82/foxb ox6 are arbitrarily near to h(x0 ), ohfoxb,
82hf8xh 8xb, respectively.
Proof. The first step. Let h (x) be a C00 function with compact support.
If /E D(A), then the "convolution"
(fh)(x)
J f(My x) h(My x0 ) dy
(9)
(!l
(My denotes a generic element of @ and dy a fixed right invariant Haar

measure on @ such that dy = d (y M) for every ME @) is C00 and
belongs to D (A). The above integral exists since the isotropy group @0 is
compact and h has a compact support. By theuniformcontinuityofjand
the compactness of the support of h we can approximate the integral by
k
Riemann sums i~ f(My; x) Ci uniformly in x:

k
(ih)(x) =s-lim .If(My.x) Ci.

k-->00
=1
Since T 1 M' = M'T1, M' commutes with A, i.e., if /E D(A), then

M' f E D (A) and AM' f = M' A f. Putting g (x) = (A f) (x), we obtain
gE C(S) and
A
CJi
f(MYi. x) ci)
i4 (AM~i.
k
f) (x) C;
=.I g(My1 x) C;,

=1
i4 (M~;Af)
(x)
c,
5. The Brownian Motion on a Homogeneous Riemannian Space
401
and the right hand side tends to (g h) (x) = (AI h) (x). Since the
infinitesimal generator A is closed, it follows that I hE D (A) and
A (I h) = A I h. Since S is a homogeneous space of the Lie group &,
i.e., S = &/&0 , we can find a coordinate neighbourhood U of x0 such
that for each x E U there exists an element M = M (x) E & satisfying the
conditions:
i) M (x)
X= X 0 ,
ii) M (x) x0 depends analytically on the coordinates (xi, x2, ... , xn)
of the point x E S.
This is so, since the set {M,. E & ; M,. x = x0 } forms an analytic submanifold of & ; it is one of the cosets of & with respect to the Lie subgroup
& 0 . Hence, by the right invariance of dy, we have
(I h) (x) =
JI (M,. M(x) x) h(M,.M(x) x0) dy

JI(M,.x0 )h(M,.M(x) x0) dy.
The right side is infinitely differentiable in the vicinity of x0 , and
D!(fh) (x) =
JI(M,. x0 )D!h(M,.M(x) x0 )dy.
(10)
The second step. Remembering that D (A) is dense inC (S) and choosing
and h appropriately, we obtain:
there exist C00 functions F 1 (x), F 2 (x), ... , P (x) E D (A) such
that the Jacobian 8 (Fl (x)' .. 'P. (x)) > 0 at x
B(xl, .. . ,xn)
and, moreover
. . . . .F
there exists a C00 function F 0 (x) E D (A) such that

n
(11)
0 > .I (x'- x') 2

(x 3 - x 03 ) - (x- x')
0
0
ax~axt - i=l
(12)
We can use F 1 (x), F 2 (x), ... , P(x) as coordinate functions in a neighbourhood {x; dis(x0 , x) < s} of the point x0 We denote these new local
coordinates by (x1 , x2, ... , xn). Since F (x) E D (A), we have
lim t-1
t.j.O
Jp (t, Xo, dx) (P (x) -
=lim t-1
t .j. 0
dis(x,x0 ) ~
F (xo)) = (A F) (xo)
P (t, x0 , dx) (P (x) - P (x0))
independently of sufficiently small s > 0, by virtue of Lindeberg's

condition (3). Hence there exists, for the coordinate function xl, x2 , , xn
(xi= F), a finite
lim t-1
t .j. 0
26
dis(x,x,) ~
(x 3 - x6) P (t, x0 , dx) = a1(x0)
(13)
402
which is independent of sufficiently small e > 0. Since F 0 E D (A), we

have, again using Linde berg's condition (3),
JP (t, x dx) (F (x) - F (x

=lim t-1
J P (t, x dx) (F (x) -
(A F 0 ) (x) =lim t-1
0,
t.j,O
tt0
dis(x,,x) ~
0,
0 ))
F 0 (x0 ))
li.m [ t-1
xi - x 0i) bFa
ax; P(t , x0 , d x )
tt0
dis(x,,x) ~
0
0 )
+ t-1 dis(x.,x)~
J (xi- x~) (xi-xt) (/::
P(t,x0 ,dx)],
1
X X x=xo+B(x-x,)
where 0 <
() <
1. The first term on the right has a limit ai (x0 )
Hence, by the positivity of P(t, x, E) and (12), we see that

-
lim t-1
ItO
.~(xi- x~) 2 P(t, x 0 , dx)
dis(x,,x)< =1
The third step. Let /E D(A)

we obtain
= t-1
+ t-1
(l(x)- f(x 0 )) P(t, x 0 , dx)
(xi-
dis(x,x0 ) :;>e
f
+ t-1 dis(x,x,)
:;>e
+ t-1
dis(x,x0 ) :;;;.
oo.
(14)
n C2 Then, by expanding f(x)- f(x 0 ),
dis(x,x0)>
<
~F~.
uX 0
x~) 88~- P (t, x 0 , dx)

Xo
. . . . 8
(x'- x ) (x 1 - x6) Ftal P (t, x , dx)
0
2/
Xo Xo
(xi- X~) (xi- x6) Cij (e) p (t, Xo, dx)
= C1 (t, e)+ C2 (t, e)+ C3 (t, e)+ C4 (t, e), say,

where Ci1 (e)-+ 0 as e t 0. We know, by (3), that lim C1 (t, e)= 0 for
ItO
fixed e > 0, and that lim C2 (t, e) = i (x0 ) !1 independently of suffiqo
uXo
ciently small e > 0. By (14) and Schwarz' inequality, we see that
i ,
lim C4 (t, e) = 0, boundedly in t > 0. The left side also has a finite limit
e.j,O
(A f) (x0 ) as t
t 0. So the difference
lim C3 (t, e) -lim C3 (t, e)
qo
ITO
can be made arbitrarily small by taking e > 0 sufficiently small. But,

by (14), Schwarz' inequality and (3), the difference is independent of
6. The Generalized Laplacian of W.
403
FELLER
sufficiently small e > 0. Thus a fmite limit lim C3 (t, e) exists independentqo
ly of sufficiently small e > 0. Since we may choose FE D (A) n C"" (S) in

such a way that 82Ff8x~ 8xj (i, j = 1, 2, ... , n) is arbitrarily near cx;1,
where cxij are arbitrarily given constants, it follows, by an argument
similar to the above, that
a finite lim t-1
t 0
and
(xi- x~) (xixj) P (t, x0 , dx) = bii (x0 ) exists (15)
dis(x,x,) ~
..
()2f
lim c3 (t, e) = b'1 (xo) ~.

uX0
uX0
t,j,O
This completes the proof of our Theorem.

Remark. The above Theorem and proof are adapted from K. YoSIDA
[20]. It is to be noted that bii (x) = bii (x) and
>
bij (x0) ~i~i
because (xi-
0 for every real vector (~v ~2 ,
x~) (x1 -
xj)
~i~j =
.. ,
~n),
(16)
(.i (xi- x~) ~;)2

=1
The Brownian Motion on the Surface of the 3-sphere. In the special

case when S is the surface of the 3-sphere 5 3 and @ is the group of rotations of 5 3 , the infinitesimal generator A of the semi-group induced by a
Brownian motion on S is proved to be of the form
A = CA
1
, where
C is a positive constant and
A is the Laplacian
az on th e surf ace o53 .

a.Oa+
ao sin21 0 otpz
=sin 0 ao sm
(17)
Thus there exists essentially one Brownian motion on the surface of 5 3

For the detail, seeK. YosmA [27].
6. The Generalized Laplacian of W. Feller
Let S be an open interval (r1 , r 2), finite or infinite, on the real line,
be the set of all Baire sets in (rv Yz) Consider a Markov process
and
P(t, x, E) on (5, >8) satisfying Lindeberg's condition
limt-1
qo
lx-yl>
P(t,x,dy)=O forevery xE(rvr2 ) and e>O. (1)
Let C [rv r 2] be the B-space of real-valued, bounded uniformly continuous

functions f(x) defined in (rv r 2) and normed by II! II= sup lf(x) I Then
(Tt f) (x) =
,,
P(t, x, dy) f(y) (t
>
0); = f(x) (t = 0)
(2)
defines a positive contraction semi-group on C [r1 , r 2]. We prove by {1) the

weak right continuity of T 1 at t = 0 so that T 1 is of class (C0 ) by the
Theorem in Chapter IX, 1.
26*
404
Concerning the infinitesimal generator A of the semi-group T 1, we

have the following two theorems due to W. FELLER.
Theorem 1.
A 1 = 0;
(3)
A is of local character, in the sense that, if
IE D (A)
vanishes in a neighbourhood of a point x0 then
(AI) (x0 )
If
IE D (A)
(4)
0;
has a local maximum at x 0 , then
(5)
(A/) (x0 ) < 0.

Proof. (3) is clear from T 1 1
(AI) (x0 ) = limt-1

qo
= 1. We have, by (1),
lx,-yl~
P(t, x0 , dy) (l(y) -l(x0 ))
independently of sufficiently smalls> 0. Hence, by P(t, x, E) > 0, we

easily obtain (4) and (5).
Theorem 2. Let a linear operator A, defined on a linear subspace
D(A) of C [rv r 2 ] into C [r1 , r 2 ], satisfy conditions (3), (4) and (5). Let
us assume that A does not degenerate in the sense that the following
two conditions are satisfied:
there exists at least one lo E D (A) such that (A 10 ) (x)
>
(6)
0 for all xE (rv r 2 ),
there exists a solution v of A v
0 which is linearly
(7)
independent of the function 1 in every subinterval
(xv x 2 ) of the interval (r1 , r 2).

Then there exists a strictly increasing continuous solution s = s (x) of
A s = 0 in (rv r 2 ) such that, if we define a strictly increasing function
m = m(x), not necessarily continuous nor bounded in (r1 , r 2 ), by
m(x) =
J" (Af~
(8)
(t)d(Dtlo(t)),
we obtain the representation:
n:, Dt I (x) in (rv r2 ) for any IE D (A) .

(9)
Here the right derivative Dt with respect to a strictly increasing function
(A I) (x) =
p = p(x)
is defined by
+ ( )
l" g (y + 0) - g (x- 0)
h
.
f
. .
f
D pgx
=Imp( +O)-p(x-O)att epomto contmmtyo
Y
g (x + 0) - g (x- 0)
h
.
f d"
. .
f
p(x + O) -p(x- O) at t e pomt o Iscontmmty o
y.j.x
p,
p.
{10)
FELLER
405
Proof. The first step. Let u E D (A) satisfy A u = 0 and u (x1) = u (x 2 ),

where r 1 < x1 < x 2 < r 2 Then u(x) must reduce to a constant in (xv x 2 ).
If otherwise, u (x) would, for example, assume a maximum at an
interior point x0 of (x1 , x 2 ) in such a way that u (x0 ) - B > u (x1 ) = u (x2 )
with some s > 0. Let / 0 be the function ED (A) given by condition (6).
Then, for a sufficiently large 15 > 0, the function F (x) = /0 (x) + 15u (x)
satisfies F (x0 ) > F (xi) (i = 1, 2). Hence the maximum of F (x) in
[xv x 2 ] is attained at an interiorpointx~although (A F) (x~) = (A/0 ) (x~) >0.
This is a contradiction to condition (5).
The second step. By (7) and the first step, we see that there exists a
strictly increasing continuous solutions (x) of A s = 0. We reparametrize
the interval by s so that we may assume that
the functions 1 and x are both solutions of A f = 0.
(11)
We can then prove that

if (A h) (x) > 0 for all x in a subinterval (x1 , x 2 ) of the
interval (r1 , r 2 ), then h (x) is convex downwards m
(12)
(xl, Xz).
For, the function u (x) = h (x) - ! X X - {J, which satisfies (Au) (x) =
(A h) (x) > 0 for all x in (x1 , x 2 ), can have no local maximum at an interior
point of (x1 , x 2 ).
The third step. Let /E D(A). Then, by (12) and (6), we see that, for
sufficiently large 15 > 0, the two functions / 1 (x) =I (x) + 15/0 (x) and
/ 2 (x) = b/0 (x) are both convex downwards. Being the difference of two
convex functions, f(x) = /1 (x)- / 2 (x) has a right derivative at every
point x of (r1 , r 2 ). Let us put
A I
rp, A / 0
cp0 , Dt / 0 (x)
ft (x).
(13)
For g1 =I -t/0 , we have A g1 =A/ -tA / 0 = rp- trp0 Hence, by

(12),
<
min rp (x) /% (x) implies that g1 (t) is convex downwards
x1 <x<x 2
in (x1 , x2 ) and so Dt g1 (x) is increasing in (xv x2 ).

Applying the same argument fort> max rp(x)frp0 (x), we obtain:
x1 <x<x1
for any sub-interval (x1 , x2) of (a, b), we have

min rp((x)) X (Dt/0 (x 2) - Dt/0 (x1 ))
x,<x<x, 'Po X
< max
tp((x)) X
x, <x<x, 'Po X
<
Dtf(x2) -Dtf(xt)
(Dt/0 (x2 )-Dt/o(xt)).
406
The continuity of the function rp (x)fffJo (x) implies, by the above inequality,
that
J :o(~)d(D"ffo(x)),
x,
D"ff(x2 ) - D"ff(x1 ) =
x,
n:, D"f f.
Remark I. We may consider the operator n:, n: as a generalized
Laplacian in the sense that n: and n:, correspond to the generalized
which is precisely the integrated version of A f =
gradiant and the generalized divergence in one dimension, respectively.

Feller called s = s (x) the canonical scale and m = m (x) the canonical
measure of the Markov process in question. We easily see, from the first
step above, that the function 1 and s constitute a fundamental system of
linearly independent solutions of A v = 0 to the effect that any solution
of A v = 0 can be expressed uniquely as a linear combination of 1 and
s (x). Thus the canonical scale of P (t, x, E) is determined up to a linear
transformation, i.e., another canonical scale s1 must be of the form
s1 = iX s + fJ with <X > 0; hence the canonical measure m1 which corresponds to s1 must be of the form m1 = <X-1 m.
Remark 2. Theorem 2 gives the representation of the infinitesimal
generator A at the interior point x of (r1 , r 2). To determine the operator
A as the infinitesimal generator of a positive contraction semi-group T 1
of class (C0) on C [rv r 2 ] into C [rv r 2 ], we must consider the lateral
condition, that is, the boundary condition of the operator A at both
boundary points r 1 and r 2 in order to describe the domain D (A) of A
concretely and completely. According to Feller [2] and [6] (cf. E. HILLE
[6]), the boundary points r 1 (or r 2) are classified into the regular-boundary,
the exit-boundary, the entrance-boundary and the natural-boundary. To
this purpose, we introduce four quantities:
al =
ff
dm (x) ds (y), Ill=
ff
dm (x) ds (y), ft2 =
~'1<y<x<r~
a2 =
r1 >y>x>r;
The boundary point r; (i
regular in case a;
exit in case
a;
.
entrance 1n case a;
natural in case
1, 2) is called
< <Xl,
< <Xl,
ft;
< <Xl
ft;
<Xl
< <Xl
a, = <Xl, ft; = <Xl
=
<Xl,
ft;
We shall illustrate by simple examples.
ff
r1<y<x<ri
ff
r1 >y>x>r:
ds (x) dm (y) ,
ds (x) dm (y).
(the conditions are

independent of the
. of r 1 an d r 1) .
ch o1ce
2
1
407
FELLER
Example 1. D;;,D; = d 2fdx 2 , S = (-oo,oo). We can takes= x,

m = x. Hence
00
jj
a2 =
dxdy=
oo>:v>x>r~
p,2
JJ
oo>y>x>:
dx dy
J (y-r;) dy =oo,
r;
oo,
and so oo is a natural boundary. Similarly ~oo is also a natural boundary.

Example 2. D;;. Dt = d2fdx2 , S = (-oo, 0). We can take s = x,
m = x. In this case, - oo is a natural boundary and 0 is a regular boundary.
Example 3. D;;. Dt
= x2 d~2 - d~,
= (0, oo). A strictly increasing
J"' e- 11 dt
continuous solutions = s (x) of D;;. Dt s = 0 is given by s (x) =

so that
D+ D+ _ x2e-1fx !:.._ e1fx !:.._
D _ e1fx !:.._
dx"
dx
msdx'
sTherefore ds
a1 =
Ill=
e- 11'" dx, dm
x- 2 e1fx dx. Hence

1
JJ
x- 2 e* dx e-lf:Ydy =
JJ
e-lfx dx y-2 elfy dy =
O<y<x<l
O<y<x<l
J [-e1''"J~ e-l/:Y dy < oc,
J e-lfx [-e1'"Jo dx =
oo.
Thus 0 is an exit boundary. Similarly we see

a2 =
f-l2 =
JJ
elfx x-2 dx e-1/:Y dy =
JJ
e-lfx dx elf:Y y-2 dy =
oo'
J e-lfx [-elf:Y]~ dx =
oo.
00
OO>:Y>X>l
oo>y>x>l
J [- e1fx]i. e-1/:Y dy =
00
That oo is a natural boundary.

d
2d2
+ +
dx, S = (0, 2). As above, we see that
Example 4. Dm D. = x dxz
ds = e1fx dx, dm = x- 2 e-lfx dx, and so we can verify that 0 is an entrance
boundary and 2 is a regular boundary.
Feller's probabilistic interpretation of the above classification is as
follows:
The probability that a particle, located at first in the interior of the
open interval (r1 , r 2), will, after a finite lapse of time, reach a regular
boundary or an exit boundary, is positive; while, the particle can, after a
finite lapse of time, neither reach an entrance boundary nor a natural
boundary.
408
References
FELLER's original paper was published in 1952: W. FELLER [2]. The
proof of Theorem 2 given above is adapted from W. FELLER [3] and [4],
who also delivered an inspiring address [5]. We also refer the readers
to the forthcoming book by K.lTo-H. McKEAN [1] and E. B. DYNKIN [3],
and the references given in these books.
7. An Extension of the Diffusion Operator

Let the possible state of a (not necessarily temporally homogeneous)
Markov process be represented by the point x of an n-dimensional C00
Riemannian space R. We denote by P(s, x, t, E), s < t, the transition
probability that a state x at a time moment sis transferred into a Baire
set E ~ R at a later time moment t.
We shall be concerned with the possible form of the operator As
defined by
(As/) (x) =lim t-1 jP(s,x,s

t'O
+ t, dy) (f(y)-f(x)),
-
/E C8(R),
(1)
without assuming the Linde berg type condition:

lim t-1
qo
d(x,y);;::;e
P(s, x, s
+ t, dy)
= 0 for all positive constants c
(d (x, y) =the geodesic distance between two points x and y).
(2)
We can prove
Theorem. Let there exist an increasing sequence {k} of positive
integers such that, for a fixed pair {s, x},
lim k
P(s, x, s
at 00
d(x,y);;::;a
+ k-1, dy)
J1 ~(;,(:):)"P(s, x, s + k-
1,
= 0 uniformly ink,
dy) is uniformly bounded ink.
(3)
(4)
Suppose that, for a function I (x) E c~ (R)'

a finite lim k
,.__,_00
f JP(s, x, s + k-1 , dy)

LR
f(y) -f(x) ~exists.
We have then, in any fixed local coordinates (x1 , x2 ,

xER,
(As/)
(x)
8/
aj (s, x) 8x
'
J
+ -.
li,mo
,.
+ bij (s, x)
xn) of the point
82
8x. 8x
,
f)/ (y) -I (x) -1 +e(x,y)
d (x y-)2 (yj- xj) 8x-
d(x,y);;::;e l
(5)
8/}
'
'
(6)
7. An Extension of the Diffusion Operator
409
where
G(s, x, E) is non-negative and a-additive for Baire sets
E ~Rand G(s, x, R) < =,
e(x, y) is continuous in
X,
(7)
y such that (I(X, y) = 1 or
0 according as d (x, y) < IJ/2 or > IJ (IJ is a fixed
{8)
constant > 0),

the quadratic form b;j(s, x)
~i~j
(9)
is> 0.
Remark. Formula (6) is given inK. YosmA [26]. It is an extension

of the diffusion operator of the form of the second order elliptic differential operator discussed in the preceding sections. The third term on the
right of (6) is the sum of infinitely many difference operators. The
appearence of such a term is due to the fact that we do not assume the
condition (2) of Lindeberg's type. Formula (6) is an operator-theoretical counterpart of the P. Levy-A. Khintchine-K. Ito infinitely divisible
law in probability theory. About this point, see E. HILLE-PHILLIPS [1],
p. 652.
Proof of Theorem. Consider a sequence of non-negative, a-additive
measures given by
Gk (s, x, E) = k
!(~ (~~~) 2
P (s, x, s
+ k-1 , dy).
(10)
We have, by (3) and (4),
Gk(s, x, E) is uniformly bounded in E and k,
Gk(s, x, dy) = 0 uniformly ink.

lim
at 00 d(x,y)<;;;a
(11)
{12)
Hence, for fixed {s, x} ,the linear functional Lk given by
is non-negative and continuous on the normed linear space cg (R) with

the norm II g II = sup Ig (x) I; and the norm Lk is uniformly bounded in k.
X
Therefore, by the separability of the normed linear space C8(R), we

can choose a subsequence {k'} of {k} such that lim Lk' (g) = L (g) exists
/t-+oo
as a non-negative linear functional on C8(R). By virtue of a lemma in

measure theory (P.R. HALMOS [1 ]) there exists a non-negative, a-additive
measure G(s, x, E) such that G(s, x, R) <=and
lim
k=k'-+oo
JGk(s, x, dy) g(y) = RJG(s, x, dy) g(y)
for all gEC8(R).
(13)
410
~~w
=
[J P(s,
j {[t(y)- f(x) -
e(y, x)
+ f d(y,
x)
~(~(;: x)
x, s
+ k-1 , dy) f(y)- f(x) ]

2
(Yi -xi)
:;J
t(:.(~}:) 2} Gk(s, x, dy)
of
(Yi- xi) oxi Gk (s, x, dy).
(14)
The term {} in the first integral on the right is, for sufficiently small
d(y, x),
of
oaf )1+d(y,x)a
=(yi-xi)ox1 +(y.-x.)(yi-xi) oX,oX1
d(y,x) 2
'
where Xi= Xj + ()(yj- Xj), 0 < () < 1.

Thus {}is bounded and continuous in y. Hence, by (12) and (13), the first
term on the right of (14) tends, as k = k'----). oo, to
G (s, x, dy).
Therefore, by (5),
R
J{}
. .
a f mite
1'tm
k=k'-->00
Je(y,
of Gk (s, x, d y ) = a1 (s, x ) ,-of

d(y x))2 (y1 - x1) ,-' X
<)Xi
<)Xi
exists. Hence, by (3), we obtain (6) by taking

bi1(s, x) =lim
li!fi
s.j.O k=k-->00
d(y,%);:;;;e
(y,- x,) (y1 - x1) P(s, x, s
+ k-1, dy).
8. Markov Processes and Potentials

Let {T1} be an equi-continuous semi-group of class (C0 ) on a function
space X defined by (Ttf) (x) =
P(t, x, dy) f(y) where P(t, x, E) is a
s
Markov process on a measure space (S, ~). Let A be the infinitesimal
generator of T 1 Suggested by the special case where A is the Laplacian,
an element f E X is called harmonic if A f = 0. Then f is harmonic
iff .il(U-A)-1/=/ for every .il> 0. An element /EX is called a
potential iff = lim (.ill - A )-1g with some g. Because, for such an element
.qo
f, there holds by the closedness of A the Poisson equation
A I= limA (.ill- A)-1 g = lim{-g

A.j.O
A.j.O
+ .il(.ill -A)-1g} =-g.
Suppose that X is a vector lattice which is also a locally convex linear

topological space such that every monotone increasing bounded sequence of elements E X converges weakly to an element of X which is
greater than the elements of the sequence. We also assume that the
resolvent J,. = (.ill- A )-1is positive in the sense that I~ 0 implies JA I> 0.
The situation is suggested by the special case where A is the Laplacian
considered in a suitable function space.
8. Markov Processes and Potentials
411
We may thus call subharmonic those elements IE X for which the

inequality A I > 0 holds. By virtue of the positivity of fA, a subharmonic element I satisfies A])J > I for all A> 0. We shall prove an
analogue of a well-known theorem of F. Rmsz concerning ordinary subharmonic functions (see T. RADO [1]):
Theorem. Any subharmonic element xis decomposed as the sum of a
harmonic element x 11 and a potential Xp, where the harmonic part x 11 of
xis given by x11 =lim A].tx and x 11 is the least harmonic majorant of x in
A~O
the sense that any harmonic element xH > x satisfies xH > x11
Proof (K. YosmA [4]). By the resolvent equation
];. - ]p. = (t-t- A)
we obtain
Jd,"
{1)
(I -A];.)= (I+ (t-t-A) ];.) (I -fl,]p.)
(2)
Since x is subharmonic, we see, by the positivity of ];., that
A> fl
implies
A];.x > fl,]p.x > x.
Hence the weak-lim A];.x = xh exists by virtue of the hypothesis conA.J.O
cerning bounded monotone sequences in X. Therefore, by the ergodic

theorem in Chapter VIII, 4, we see that x11 =lim A];.x exists and x11 is
Aj.O
harmonic, i.e., A];.x11 = x11 for all A> 0. We also have Xp = (x- xh) =
lim (I- A]A) x =lim (-A) (AI- A)-1 x =lim (AI- A)-1 (-Ax)
Aj.O
Aj.O
Aj.O
which shows that Xp is a potential.

Let a harmonic element xH satisfy xH > x. Then, by the positivity of
A]A and the harmonic property of xH, we have
xH = A]}.xH > A]}.x and hence

,
xH >lim A];.x = xh.

Aj.O
Researches by Hunt and by Doob. G. A. HuNT [1] gave a condition that a linear operator E L ( cg (Rn), cg (Rn)) becomes one of the resolvents of a semi-group induced by a Markov process P(t, x, E) in R" and
discussed the potential theory associated with this Markov process. We
refer the reader to an excellent exposition of Hunt's theory given in
P. A. MEYER [1]. J. L. DooB [2] discussed the boundary values of harmonic functions f (x) when x tends to the boundary through the path of
the corresponding Markov process. The essential tool in Doob's research
is the limit theorem associated with the martingale theory developed
systematically by him.

The ordinary exponential function solves the initial value problem
dyfdx=IXy, y(O) =y.
412
We consider the diffusion equation

m
8uj8t = Llu, where L1 = ..J: 82 j8x'j is the Laplacian in Rm.

;~1
We wish to find a solution u = u(x, t), t > 0, of this equation satisfying

the initial condition u(x, 0) = f(x), where f(x) = f(xv .. . , xm) is a
given function of x. We shall also study the wave equation
82uj8t2 = Llu, -oo

with the initial data
u(x, 0) = f(x)
and
< t<
(8uj8t) 1 ~o
oo,
= g(x),
and g being given functions. This may be written in vector form as

follows:
~ (:) = (~ ~) (:).
with the initial condition
~~
0)) (f(x))
=
( u(x,
v(x,O)
g(x)
So in a suitable function space, the wave equation is of the same form
as the diffusion (or heat) equation-differentiation with respect to the
time parameter on the left and another operator on the right-or again
similar to the equation dyjdt = !Xy. Since the solution in the last case is
the exponential function, it is suggested that the heat equation and the
wave equation may be solved by properly defining the exponential functions of the operators
L1
and(~ ~)
in suitable function spaces. This is the motivation for the application

of the semi-group theory to Cauchy's problem. It is to be noted that the
Schri:idinger equation
i-1 8uj8t = Hu = (Ll + U(x)) u, where U(x) is a given function,
gives another example of the equation of evolution of the form
8uj8t = Au, t > 0,

(1)
where A is a not necessarily continuous, linear operator in a function
space.
The equation of the form (1) may be called a temporally homogeneous
equation of evolution. We may integrate such an equation by the semigroup theory. In the following three sections, we shall give typical
examples of such an integration. We shall then expound the integration
theory of the temporally inhomogeneous equation of evolution
8uj8t =A (t) u, a< t < b.
(2)
1. Integration of Diffusion Equations in
V (Rm)
413
1. Integration of Diffusion Equations in L2 (Rm)

Consider a diffusion equation
fJujfJt =Au, t
where the differential operator
..
A = a'1 (x) -;;-;.;ux,uxl
>
(1)
0,
+ b' (x) ..,

ux, + c (x)
..
..
( a'1 (x) = a1' (x))
(2)
is strictly elliptic in an m-dimensional euclidean space Rm. We assume that

the real-valued coefficients a, b and c are coo (Rm) functions and that
Ia# (x) 1. s~p W(x) 1. s~p Ic (x) 1. s~p Ia~r.(x) 1.

sup Ib~k (x) I, sup Ia~k'" (x) I) = 'YJ < oo.
max (s~p
(3)
:<
:<
The strictly elliptic hypothesis concerning A means that positive constants A.o and Po exist such that
m
..
Po ..E ~j > a'1 (x) ~i~j 2: Ao ..E ~r on Rm for any real
J=1
(4)
vector ~ = (~v ~2 , ~m)
Let H~ be the space of all real-valued C0 (Rm) functions I (x) normed
=1
by
lllll1 =
Jl
Rm
dx
m
.~ Jl';,dx
+ J-1
Rm
)1/2 ,
(5)
and let H~ be the completion of HA with respect to the norm Ill\ h. Let
similarly H8 be the completion of HA with respect to the norm
(6)
We have thus introduced two real Hilbert spaces HA and H8. and HA
and H~ are 1\ 110-dense in Hg. We know, from the Proposition in Chapter I, 10, that HA coincides with the real Sobolev space W 1 (Rm); we know
also that H8 coincides with the real Hilbert space L 2 (Rm). We shall denote
the scalar product in the Hilbert space HA (or in the Hilbert space H8)
by (/, gh (or by (/, g) 0).
To integrate equation (1) in the complex Hilbert space L 2 (Rm) under
condition (3) and (4), we shall prepare some lemmas. These lemmas will
also play important roles in the following sections.
Lemma 1 (concerning partial integration). Let 1. g E fiA. We have then
(A 1. g) 0 = -
Ja# /,. g,. dx- Ja~1 /x1 gdx

1
Rm
Rm
(7)
414
that is, we may partially integrate, in (A f, g) 0 ,the terms containing the

second order derivatives as if the integrated terms are zero.
Proof. By (3) and the fact that f and g both belong to H6, we see
that aij fx;x;g is integrable over Rm. Hence, by the fact that f and g are
both of compact support,
Jaijlx;x;gdx = - Jaijfx;gx;dx- J a~;/x;gdx.
Rm
Rm
Rm
Remark. The formal adjoint A* of A is defined by

f)2
..
f)
(A*/) (x) = ox; ox; (a'1 (x) f (x)) - ox; W(x) f (x))
+ c (x) f (x) .
(8)
Then, as above, we have the result: If /, g E H6 then we may partially

integrate, in (A* j, g) 0 , the terms containing the first and the second
order derivatives as if the integrated terms are zero. That is, we have
(A*/, g) 0
=-
J aij fx;gx dx- J a~fgx1 dx

1
Rm
Rm
(7')
Corollary. There exist positive constants x, y and

sufficiently small positive constants ()(.,
()(.~
111/li < (/- ()(.Af, f)o <
(1
+ ()(.y) II! IIi
~such
when
that, for all
IE ii6,
(9}
()(.~1/fl/i<(f-()(.A*f,f)o<(l+()(.y)l/fl/i when /EHij,
I(/- ()(.Af, g) 0 / < (1 + ()(.y) /l//1 11/g/11 when /, g EH6,

1
/(1-()(.A*f, g) 0 1< (1 + ()(.y) /lf/l1 Jlg/1 1 when /, gE H 0 ,
I(A/, g)o- (/, Ag)o I< x l/1 /11llg I/o when f,gEH6.
~
(10}
(11}
Proof. (9) and (10) may be proved by (3), (4), (7) and (7') remembering the inequality
2()(.labl
<
()(.(v
which is valid whenever()(. and v are

such as
JaJ 2 + v-1 lbl 2)

>
(12}
0. In fact, we can use an estimate
IR a~;fx;gdxl < i?imn(vllfx;ll~ + V- 1 /lgll~).

We also obtain (11) from
(A/, g)o- (/,A g) 0 = -
J (2a~Jx1 g + a~x1 /g- 2bifx;g- b~jg) dx.
Lemma 2 (concerning the existence of solutions of u- ()(.Au = f).

Let a positive number ()(.0 be so chosen that the above Corollary is valid for
0 < X
<
X0
1. Integration of Diffusion Equations in V (Rm)
415
Then, for any function I (x) E ii5, the equation

u- IXA u =I (0 <X< X0)
(13)
admits a uniquely determined solution u E H5 (\ C (Rm).

Proof. Let us define a bilinear functional B (u, v) = (u- IXA *u, v) 0
defined for functions u, v E ii5. From the above Corollary, we have
00
IB(u, v) I< (1-j- Xy) llull 1 llvllv X<'l.llulli < B(u, u).
(14)
Hence we may extend B (u, v), by continuity, to a bilinear functiona:l
B (u, v) defined for u, v E H5 and satisfying
IB(u, v) I< (1 + Xy) llull 1 llvllv X<'lllulli < B(u, u).

(14')
The linear functional F (u) = (u, 1)0 defined on H5, is a bounded linear
functional by l(u,l)ol < llullo lllllo < llu[l 1 lll[l0 . Hence, by F. Riesz'
representation theorem in the Hilbert space H5, there exists a uniquely
determined v = v (f) E H5 such that (u, 1)0 = (u, v (1)) 1 Thus, by the
Milgram-Lax theorem applied to the Hilbert space H5, we have
(u, 1) 0 = (u, v (1)) 1 = B (u, Sv (!)) for all u E H5, where
S is a bounded linear operator on H5 onto H5.

Let u run over ego (Rm), and let vn E i5 be such that lim
n--+00
(15}
II vn- S v (f) [[ 1
= 0. Then
B(u, Sv(f)) =lim B(u, vn) =lim B(u, vn)
n--+00
n--+00
=lim (u -XA*u, Vn)o = (u-XA*u, Sv(/))0 ,

n--+00
because the norm
II 111 is larger than the norm IIIIo Hence
(u,l) 0 = (u-XA*u,Sv(f)) 0 ,
(15')
that is, Sv (f) E H5 is a distribution solution of the equation (13). Hence,
by the strict ellipticity of (I - X A) and by the fact I E ego (Rm), we see,
from the Corollary of Friedrichs' theorem in Chapter VI, 9, that we may
consider u = Sv(f) E H5 to be a C00 (Rm) solution of (13). Hence u =
Sv (f) E H5 f\ coo (Rm).
The uniqueness of such a solution u of (13) may be proved as follows.
Let a function u E H5 f\ C00 (Rm) satisfy u- X Au = 0. Thus Au E
H5 (\ C00 (Rm) t;;, H8 and so the expression (u- IXA u, u)0 is defined and
= 0. Let Un E ii5 be such that lim llu- un 111 = 0. We obtain, by partial
n--+00
integration as in (9),
0 = (u- X Au, u) 0 = lim (u- iXA u, un)o
n--+00
> X<'l llu IIi.
that is, u = 0.
Corollary 1. Positive constants &0 and 'f/o exist such that, for any
IE i5, the equation
XU- Au =I (0 < &0
+ Ao + 'YJo <
X)
(16}
416
admits a uniquely determined solution u = u1 E Hij (\ C00 (Rm), and we

have the estimate
(17)
Proof. By Schwarz' inequality we have
IJ(1XI-A)ullollullo21((1XI-A)u,u) 0 l for uEifij.
(18)
By partial integration, we have
((!XI -A) u, u) 0 =IX llull~
Jaiiux ux dx + J a% ux udx
1
Rm
- J biux udx- J cuudx.
Rm
Rm
Rm
Hence, we have, by (3), (4) and (12),
llu II~+ Ao (llu IIi -llu II~)

'fJ [v(llulli-llull~) + Y- 1 llull~ + m- 2 lluii~J
(1X-A 0 -m2 'fJ (v- 1 - Y + m- 2 )) llull~ + (.1.0 -m2 rJY) llulli
((!XI- A) u, u)o >IX

-m2
=
Thus by (18) we have, for 'flo = m2 'fJ (v-1 -
II(!XI-A)ullo> (1X-A0 -n0)
11
llullo
+ m-2),
whenever
uEifij,
(17')
by taking v > 0 so small that (./.o- m2 'fJ v) and 'flo are> 0. We then take
&0 so large that for IX> &0
./.o 'flo we can apply Lemma 2
in solving (16).
Since the solution u =UtE H~f\ C00 (Rm) of (16) is approximated by
11111-norm by a sequence of functions E if~, we obtain the estimate (17)
from (18) and (17').
Corollary 2. Consider A as an operator on D (A)= (!XI- A)-1 if~~ H8
into H8. Then the smallest closed extension A in H8 of A admits, for
IX > &0 + ./.o + 'flo, the resolvent (!XI - A)-1 defined on H8 into H8
such that
+ +
(19}
Proof. Clear from Corollary 1 remembering the fact that D (A) and
ifij both are 1111 0-dense in H8.

Thus, by Corollary 2 in Chapter IX, 7, A is the infinitesimal generator
of a semi-group T 1 of class (C0 ) in the B-space H8 such that II T 1 llo < e<l.o+'1o)t
fort> 0.
Actually we can prove
Theorem 1. Let Complex-Hij be the space of all complex-valued
functions
fEC~(Rm)
with
111111 = ( RmJlfl 2 dx+.~

Jllfll2 dx) 112 <<XJ.
J-1 Rm
417
1. Integration of Diffusion Equations in L2 (Rm)
Let fiij and fig be the completions of Complex-Hij with respect to the
norm \\1\\ 1 and \IIIIo, respectively. We know that fiij =(complex) Sobolev space W 1 (Rm) (see Chapter I, 10). It is also clear that fig= (complex)
Hilbert space L 2 (Rm). We consider A as an operator defined on D (A) =
(al-A)- 1 fiij~L 2 (Rm) into L 2 (Rm). Then the smallest closed
extension A in L 2 (Rm) of A is the infinitesimal generator of a holomorphic
semi-group T 1 of class (C0 ) in L 2 (Rm) such that IIT1 1Jo < e(A.+'lolt for
t 2: 0.
Proof. By the preceding Corollary 2 and the reality of the coefficients
in the differential operator A, we see that the range R(exl -A)=
(exl- A) D (A) is IJJI 0-dense in L 2 (Rm) for ex> &0 + Ao + 'YJo Moreover, we have, for (u + iv) E L 2 (Rm) such that (u + iv) E D(A),
Jl(exl -A) (u
+ iv) 1\5 =
ll(exl -A) ul\5
ll(exl -A) vl\5
> (ex- Ao -no) (llu 115 + llv 115)

Hence the inverse (ex!- A)- 1 is a bounded linear operator on L 2 (Rm)
into L 2 (Rm) such that II (exl -A)- 1 1\ 0 < (ex-A-0 -no)- 1 for ex> a0 +
Ao + 'YJo
Therefore, by the Theorem in Chapter IX, 10, we have only to show that
2
(20)
We have, for wED (A),
1\((ex +
i-r) I
wll 0 llwll0 >\(((ex+ i-r) I
-A)
-A) w,
w)ol
(21)
By partial integration, we obtain, as in the proof of (17),

J
Real part (((ex +iT) I - A) w, w) 0 I
=I \ex J\lw \15 +Real part
JbiWXiWdX-
JCWWdX)I
Rm
Rm
> (ex-A-0 -
(i aiiwxJijx;dx + a%;wxJildx
j
'i}o) llw\15 + (A-0 -m2 'i}v) JJwlli
Similarly we have
Imaginary part (((ex +iT) I - A) w, w) 0 I
J
> IJIIw 115-m2 'i}(l\w IIi+ m- 2 1\w 1\5) I> (\r 1-n) llw 115-m2 'iJ llw IIi
Suppose there exists a w0 E D (A), II w0 Jlo =F 0, such that
\Imaginary part (((ex!+ i-r) I - A) w
0, w
0) 0 I< 2-I (J-r 1- 'i}) JJwo 115
for sufficiently large T (or for sufficiently large -T). Then, for such large
r (or -T), we must have
m2
27
1]
JJwo IIi> 2- 1 (J-r 1- 'i}) llwo 115
418
and so
[Real part (((ex+ ir) I- A) w0 , w0 ) 0 [ > (A-0 -m2 rJY)
Hence we have proved Theorem 1 by (21).
([~~ 2111J) [[w0 [15
Theorem 2. For any /E L 2 (Rm), u(t, x) = (Ttf) (x) is infinitely differentiable in t > 0 and in xE Rm, and u (t, x) satisfies the diffusion equation
(1) as well as the initial condition lim II u (t, x) - f (x) [[0 = 0.
qo
Proof. We denote by Tjkl the k-th strong derivative in L 2 (Rm) of
T 1 with respect to t. T 1 being a holomorphic semi-group of class (C0 ) in
L 2 (Rm), we have Tjklf = fi.kTtfE L 2 (Rm) if t > 0 (k = 0, 1, ... ). Since
Ais the smallest closed extensioninL2 (Rm) of A, we see that AkTtfEL2 (Rm)
for fixed t > 0 (k = 0, 1, 2, ... ) if we apply the differential operator Ak
in the distributional sense. Thus, by the Corollary of Friedrichs' theorem
in Chapter VI, 9, u (t, x) is, for fixed t > 0, equal to a C00 (Rm) function
after a correction on a set of measure zero.
Because of the estimate If T 1 flo < e(-'.+>7ol 1, we easily see that, if we
apply, in the distributional sense, the elliptic differential operator
u:2 +A)
any number of times to u (t, x), then the result is locally square integrable
in the product space {t; 0 < t < oo}xRm. Thus again, by the Corollary
of Friedrichs' theorem in Chapter VI, 9, we see that u (t, x) is equal to a
function which is C00 in (t, x), t > 0 and x E Rm, after a correction on a
set of measure zero of the product space. Thus we may consider that
u (t, x) is a genuine solution of (1) satisfying the initial condition
lim llu(t, x)- f(x) llo = 0.
qo
Remark. The above obtained solution u (t, x) satisfies the "forward

and backward unique continuation property":
if, for a fixed t0

then u (t, x)
>
0, u(t0 , x)- 0 on an open set G <;;; Rm,
0 for every t
>
0 and every x E G.
(22)
Proof. Since T 1 is a holomorphic semi-group of class (C 0 ), we have
lim IIT1 +hf- k=O

.i (k!)- 1 hkAkT1 /1! 0 =
for a sufficiently small h. Hence, as in the proof of the completeness of

the space L 2 (Rm), there exists a sequence {n'} of natural numbers such
that
n'
u(t0 +h,x)= lim 2:(k!)- 1 hkAku(t0 ,x) for a.e. xERm.
n'--+00 k=O
By hypothesis given in (22), we have Aku(t0 , x) = 0 in G and so we must

have u (t0 + h, x) = 0 in G for sufficiently small h. Repeating the process
we see that conclusion (22) is true.
419
References
The result of this section is adapted from K. Yosida [21]. As for
the "forward and backward unique continuation property" (22), we have
the more precise result that u (t, x) = 0 for every t > 0 and every x E Rm.
For, S. MIZOHATA [3] has proved a "space-like unique continuation property" of solutions u(t, x) of a diffusion equation to the effect that
u (t, x) = 0 for all t > 0 and all x E Rm if u (t, x) = 0 for all t > 0 and all
x E G. Concerning the holomorphic character of the semi-group T 1 obtained above, see also R. S. PHILLIPS [6]. It is to be noted here that the
unique continuation property of the solutions of a heat equation 8uj8t =L1u
was first proposed and solved by H. YAMABE-S. ITo [1].
There is a fairly complete discussion of parabolic equations from the
view point of the theory of dissipative operators. See R. S. PHILLIPS [7].

in a Compact Riemannian Space
Let R be a connected, orientable m-dimensional C00 Riemannian
space with the metric
(1)
ds 2-- g;j(x) dx i dx j .
Let A be a second order linear partial differential operator in R with
real-valued C00 coefficients:
..
aa
a'l (x) ax ax;
+ b' (x) ax.
(2)
We assume that aij is a symmetric contravariant tensor and that b; (x)

satisfies the transformation rule
Yi =
bk Bxi
axk
+ akj
(3)
Bx .
axk 8x 1
by the coordinate change (x\ x 2 , , xm)--+ (x\ x2 , , .xm) so that the

value (A f) (x) is determined independently of the choice of the local coordinates. We further assume that A is strictly elliptic in the sense that there
exist positive constants Ao and flo such that
m
flo ..I
]=1
;J >
aij
(x) $;$j
>
.1. 0 ..I ~J for every real vector

;=1
(4)
($v ... , $,.) and every x E Rm.
We consider the Cauchy problem in the large on R for the diffusion

equation: to find solution u (t, x) such that
8uj8t = Au, t > 0, u (0, x)
= f (x)
where
function on R.
We shall prove
27*
f (x)
is a given
(5)
420
Theorem. If R is compact so that R is without boundary, then the

equation (5) admits, for any initial function f E C00 (R), a uniquely determined solution u(t, x) which is C00 in (t, x), t > 0 and x E R. This solution
can be represented in the form
u(t, x) =
JP(t, x, dy) f(y)
(6)
where P(t, x, E) is the transition probability of a Markov process on R.

Proof. Let us denote by C (R) the B-space of real-valued continuous
functions f(x) on R normed by II/II= sup
,. lf(x) I We first prove
for any f E COO (R) and any n > 0, we have
max
h
(x)
> f (x) > min
,.
,. h (x) where h (x) = f (x) - n-1 (A f) (x). (7)
Suppose that f (x) attains its maximum at x = x0 We choose a local

coordinate system at x 0 such that aii (x0 ) = ~ij (= 1 or 0 according as
i = j or i =I= 7). Such a choice is possible owing to condition (4). Then
h (x0 ) = f (x0 ) - n-1 (A f) (x0 )
-1 i
=f(x0 )-n
of
-1
b (x0 )"""<i"t-n
uX0
02f
..I"(x 0t) 2 > f(x0 ),
~=1 u
since we have, at the maximum point x0 ,
of
o2f
and o(xA)a < 0.

Thus max
h(x)
>
f(x).
Similarly
we have f(x) >min
,.
,. h(x).
oxt = 0
We shall consider A as an operator on D(A) = C00 (R) ~ C(R) into

C (R). Then, by (7), we see that the inverse (I- n-1 A)-1 exists for n > 0
and I!(I- n-1 A)-1 g II< II gil for g in the range R(I- n-1 A) =
(I- n-1 A) D (A). This range is, for sufficiently large n, strongly dense
in C(R). For, as in the preceding section, we have the result: For any
g E COO (R) and for sufficiently large n > 0, the equation u- n-1 Au = g
admits a uniquely determined solution u E C00 (R). Because, by the compactness of the Riemann spaceR, Lemma 1 concerning the partial integration in the preceding section may be adapted to our Riemannian
space R without boundary. Moreover, C00 (R) is strongly dense in C (R)
as may be seen by the regularization of functions in C (R) (see Proposition 8 in Chapter I, 1).
Hence the smallest closed extension X in C (R) of the operator A
satisfies the conditions:
for sufficiently large n > 0, the resolvent (I- n-1A.)-1 )
exists as a bounded linear operator on C (R) into C (R)
(8)
such that II (I- n-1 ] ) -1 11 < 1,
((I -n-1 A.)-1 h) (x) > OonRwheneverh(x) > OonR,
(9)
(I -n-1 A.)-1 . 1 = 1.
{10)
3. Integration of Wave Equations in a Euclidean Space R"'
421
The positivity of the operator (I- n-1 :A)-1 given in (9) is clear from (7).
The equation (10) follows from A 1 = 0.
Therefore A is the infinitesimal generator of a contraction semigroup T 1 inC (R) of class (C0 ). As in the preceding section, we see, by the
strict ellipticity of A, that, for any /E C00 (R), the function u(t, x) =
(Td) (x) is a C00 function in (t, x) for t > 0 and x E R so that u (t, x) is a
genuine solution of (5).
Since the dual space of the space C (R) is the space of Baire measures
we easily prove the latter part of the Theorem remembering (9)
R,
in
and (10).
3. Integration of Wave Equations in a Euclidean Space Rm

Consider a wave equation
< t < oo,
fJ2uj8t 2 =Au, -oo
(1)
where the differential operator

..
()2
ox,ox; +
A= a'1 (x) - -
ox +
b'(x) -.
(2)
c(x)
is strictly elliptic in an m-dimensional Euclidean space Rm. We assume

that the real-valued C00 coefficients a, b, and c satisfy conditions (3)
and (4) in Chapter XIV, 1. As done there, we shall denote by Hij the
space of all real-valued C3" (Rm) functions I (x) normed by
II I Ill = (RL 12 dx + j~ RL r;j dx) 112

and let Hij (and H8) be the completion of Hij with respect to the norm
f2 dx 2 ) .
II I 111 (and with respect to the norm ll!llo =
(Rl
y1
Lemma. For any pair {I, g} of elements E k/i the equation

(3)
admits, if the integer n be such that In- 1 is sufficiently small, a uniquely

determined solution {u, v}, u and v E Hij f\ C00 (Rm), satisfying the estimate
((u -IX0 Au, u) 0 + iX0 (v, v) 0?/2
1
<
(1 - {J In ~-I)-I ((t- iXoA I, l)o
+ iXo (g, g)o)l/2'
(4)
with positive constants iXo and {J which are independent of n and {I, g}.
Proof. Let u 1 E Hij (\ C00 (Rm) and v1 E Hij (\ C00 (Rm) be respectively the
solutions of
(5)
422
The existence of such solutions for sufficiently small In-11 was proved in
Lemma 2 of Chapter XIV, 1. Then
u
u1
+ n-1 v1 ,
= n-1 Au1 + v1
satisfies (3), i.e., we have u- n- = f, vWe next prove (4). We remark that
1v
Au
= n (v -g) E HA (\ coo (Rm)
H8
n-1 Au=
(6)
g.
and hence, by f, g E ego (Rm)'
n (Au- A f) E HA 1\ C (Rm) ~ H8.

00
We have thus, by (3),
(f- r:x0 A/, /) 0 = (u- n-1 v -CXoA (u- n-1 v), u- n-1 v) 0
= (u- r:x0 A u, u) 0 -
2n-1 (u, v) 0
+ r:x0 n-1 (Au, v) 0 + r:x0 n-1 (A v, u) 0
+ n-2 (v- r:x0 A v, v)0

and
r:x0 (g, g) 0 = r:x0 (v- n-1 Au, v- n-1 A u) 0
= r:x0 (v, v) 0 -
r:x0 n-1 (v, A u) 0 - r:x0 n-1 (Au, v) 0 + r:x0 n-2 (A u, A u) 0
By a limiting process, we can prove that (9), (10) and (11) in Chapter XIV,1
are valid for f = u and g = v. Thus, by (12) in Chapter XIV, 1, there
exists a positive constant {3 satisfying
((t- r:xoA f, f)o
+ IXo (g, g)o? 12 >

-r:x0 ln- 1 11
+ CXo(v, v)o
(Av, u)0 !- 2ln- 1 ll (u, v) 0 1) 1' 2
((u- r:xoA u, u)o
(Au, v) 0 -
> (1- fJ!n- 1 !) ((u -CXoAu, u) 0
+ r:x0 (v, v)0?'2
for sufficiently large In I

The above estimate for the solutions {u, v} belonging to HA 1\ C00 (Rm)
shows that such solutions are uniquely determined by {f, g}.
Corollary. The product space HA 1\ H8 of vectors
,
( u)
v ={u,v},
is a B-space by the norm
II(:)//=
l!{u, v}'
where uEH01 and vEH 0 ,
(7)
II=
+ r:x0 (v, v) 0) 112 ,
(8)
(B (u, u)
where B (1, g) is the extension by continuity with respect to the norm

II 11 1 of the bilinear functional
B(f, g) = (/- r:x 0 Af, g) 0
defined for f, g CHA.
We know that B(u, u) 1' 2 is equivalent to the norm l!ul! 1 (see Chapter XIV, 1):
(9)
r:x0 !5llullr::::.:; B(u, u) < (1 + r:x0 y) llulli
3. Integration of Wave Equations in a Euclidean Space Rm
423
Let the domain D (21:) of the operator
21: =
(~ ~)
(10)
be the vectors {u,v}' EH~xH8 such that u, v E H8 are given by (6). Then
the Lemma shows that the range of the operator
S- n-12{, where S =
(~ ~),
contains all the vectors {/, g}' such that /, g E 1f5, Thus the smallest
closed extension "91 in H~ X H8 of the operator 21: is such that the operator
(S- n-1 ~) with integral parameter n admits, for sufficiently large In I,
an everywhere in H~ X H8 defined inverse (S- n-1 ~j')-1 satisfying
(11)
II(S-n-1 ~)-1 11 < (1-fJin-11)-1.
We are now prepared to prove
Theorem. For any pair{f(x),g(x)} of Cgo(Rm) functions, the equation
(1) admits a CC'" solution u (t, x) satisfying the initial condition
(12)
u(O, x) = f(x), u1 (0, x) = g(x)
and the estimate
(B (u, u)
+ .x0 (u
1,
u1) 0) 112 < exp ({3
It i) (B (/, /) + .x0 (g, g) 0) 112
(13)
Remark. Formula (9) shows that B (u, u) is comparable to the potential energy of the wave (= the solution of (1)) u(t, x), and (u1, u1) 0 is
comparable to the kinetic energy of the wave u (t, x). Thus (13) means
that the total energy of the wave u(t, x) does not increase more rapidly
than exp ({3 It i) when the time t tends to oo. This is a kind of energy
inequality which governs wave equations in general.
Proof of Theorem. Estimate (11) shows that~ is the infinitesimal generator of a group T 1 of class (C0 ) in H~xH8 such that
IIT1 11 <
exp(f31ti), -oo < t < oo.

k = 0, 1, 2, ... ,
for
have,
we
By hypothesis
~k ( ;
$1lk ( : )
Hence, if we put
E ego (Rm) X ego (Rm)
x))
(14)
~ H~ X H8.
(f(x)),
= T
( u(t,
1 g (x)
v (t, x)
then, by the commutativity of ~ with T 1 , we have
~ (u (t, x)) =
fJtk
v (t, x)
akT1
fJtk
= ~k (u (t, x)) E H X Ho
(1g (x))
v (t, x)
(x)
1
0
424
fork= 0, 1, 2, ... Here we denote by 8kTef8i' the k th strong derivative

in H~ X H8. Therefore, by H~ ~ H8 = L 2 (Rm) and the strict ellipticity
of the operator A, we see, as in the proof of Theorem 2 in Chapter XIV, 1,
that u (t, x) is C"" in (t, x) for-= < t < =, x E Rm, and satisfies equation (1) with (12) and estimate (13).
Remark. The result of the present section is adapted from K. YosmA
[22]. Cf. J. L. LIONS [1]. P. D. LAX has kindly communicated to the
present author that the method of integration given in this section is
very similar to that announced by him in Abstract 180, Bull. Amer. Math.
Soc. 58, 192 (1952). It is to be noted here that our method can be modified
to the integration of wave equations in an open domain of a Riemannian
space. There is another approach to the integration of waYe equations
based upon the theory of dissipative semi-groups. See R. S. PHILLIPS [8]
and [9]. The method is closely connected with the theory of symmetric
positive systems by K. FRIEDRICHS [2]. Cf. also P. LAx-R. S. PHILLIPS
[3].
4. Integration of Temporally Inhomogeneous Equations of

Evolution in a Reflexive B-space
We shall be concerned with the integration of the equation
dx(t)fdt =A (t) x(t), a< t <b.
(1)
Here the unknown x (t) is an element of a B-space X, depending on a

real parameter t, while A (t) is a given, in general unbounded, linear
operator with domainD(A (t)) and rangeR(A (t)), both in X, depending
also on t.
T. KATO [3], [4] was the first to make a successful attack on the
problem of integration of (1). He assumed the following four conditions:
(i) The domain D(A (t)) is independent oft and is strongly dense in
X such that, for IX> 0, the resolvent (I -IXA (t))-1 exists as a bounded
linear operator E L (X, X) with the norm < 1.
(ii) The operator B (t, s) =(I- A (t)) (I- A (s))-1 is uniformly bounded in norm for t ~ s.
(iii) B (t, s) is, at least for somes, of bounded variation in tin norm,
that is, for every partition s = t0 < t1 < < tn = t of [a, b],
n-1
.I IIB(ti+ 1 , s) -B(tj, s) II< N(a, b)< oo.

)=0
(iv) B (t, s) is, at least for some s, weakly differentiable in t and
8B (t, s)f8t is strongly continuous in t.
Under these conditions, KATO proved that the limit
U (t, s) x0 =
s-lim
. II exp ((tj+l -tj) A (tj)) x0
maxJti+,-til->-0 J=n-1
4. Integration of Temporally Inhomogeneous Equations
425
exists for every x0 E X and gives the unique solution of (1) with the initial
condition x(s) = x0 at least when x0 E D(A (s)).
Kato's method is thus an abstraction of the classical polygon method
of Cauchy for ordinary differential equations dx (t) fdt = a (t) x (t). Although
the method is very simple and natural in its idea, the proof is somewhat
lengthy because it is concerned with a general B-space X. KATO [3] has
shown that the proof can be simplified if the space X is reflexive.
Another method of integrating equation (1) has been devised by
J. L. LIONS [2]. He assumes the operator A (t) to be an elliptic differential
operator with smooth coefficients depending on t, and seeks distributional solutions x (t) by transforming equation (1) into an integrated form
in concrete function spaces such as the Sobolev spaces Wk 2 (.Q) and their
variants.
We shall not give details of the methods of KATO and LIONS, but
refer the reader toT. KATO [3], [4] and J. L. LIONS [2]. We also refer
the reader to a paper by 0. A. LADYZHENSKAYA-I. M. VISIK [1] which is
motivated by a similar idea as Lions'. It is to be noted here that the
cited book by Lions contains a fairly complete list of references concerning the integration of the equations of evolution up to 1961.
In the present section, we shall give a method of integration of (1)
which is based upon a uniqueness lemma and the local sequential weak
compactness of a reflexive B-space. Our motivation may be described as
follows.
If A (t) = A is independent of t and A is the infinitesimal generator
of a semi-group T 1 of class (C0 ) in X, then the solution x (t) of
dx(t)fdt = Ax(t), a< t < b, x(a) = x0 E D(A) ==the domain of A, (2)
is given by
( )
x(t) = T 1 x 0 = exp((t- a) A) x0 = s~~ exp((t -a) An) x 0 ,
3
where An= A (I- n-1A)-1.

Suggested by this time independent case, let A (t) satisfy two conditions:
A (t) is, for a< t < b, a closed linear operator with
dense domain D(A (t)) ~X and range R(A (t)) ~X
(4)
such that, for A> 0, the resolvent (H -A (t))-1 exists
with the estimate \\(I -A-1 A (t))-1\\ < (1-A-1M)-1
for A> M, where M is independent of A and t;
the strong derivative dA (t)- 1/dt = B (t) exists and is
(5)
strongly continuous in t, a< t :S b.
We then consider the initial value problem:
dxn(t)fdt = An(t) Xn(t), a< t < b, Xn(a) =A (a)-1 y,
(6)
where y is an arbitrary element of X and An (t) =A (t)
1
1
1
1
X (I -n- A (t))- =nUn (t) -I), fn (t) =(I -n- A (t))-
426
Because of (5), ] n (t) is strongly continuously differentiable with respect

to t, and so Jn (t) is uniformly continuous in tin the sense of the norm of
operator. Thus problem (6) has a uniquely determined solution xn (t)
which may be obtained, for instance, by successive approximation starting
with the first approximation exp((t-a) An(a)) A (a)-1y. Hence we are
lead to the problem of finding conditions .for A (t) under which the sequence {xn (t)} converges, strongly or weakly, to the solution x (t) of the
initial value problem
dx(t)jdt =A (t) x(t), a< t::::::: b, x(s) =A (a)-1 y.
(1')
The purpose of the present section is to give a partial answer to this

problem.
Lemma. The solution of (1'), if it exists, must satisfy
!!x(t) II< !!x(a) II exp((t- a) M).
(7)
Proof. We follow an argument essentially given in T. Kato [3]. We

have, for b > 0,
x(t
+b)= x(t) + M (t) x(t) + o(b)

= (I + bA (t)) (I- M (t)) (I- M
{t))-1 X (t) + 0 (b)

1
2
=(I- M (t))- x(t)- b A (t) (I- M (t))-1 A (t) x(t)
+ o(b)
(I- M (t))-1 X (t)- b ((I- M {t))-1 - I) A (t) X (t) + 0 (b).
Because of (4), we have s-lim (I -bA (t))-1 z = z for any z EX (see (2)
=
6j,O
in Chapter IX, 7). Hence, by the estimate in (4), we have
+b) II< (1- bM)-1 iix(t) II+ o(b)

and sod+ llx(t) ilfdt < M llx(t) II This proves the Lemma.
!ix(t
Theorem 1. We assume that X is a reflexive B-space so that any bounded sequence in X contains a subsequence which converges weakly to an
element of X. Besides (4) and (5), we also assume that
A (t) B (t) =A (t) [dA (t)- 1 jdt] is strongly continuous in t.
(8)
Then there exists a uniquely determined solution of (1') which is given by

x (t) = weak- lim Xn (t).
n~
Proof. Since An(t) =n(Jn(t)-I) and I!Jn(t)!! < (1-n-1M)- 1, we

have, for b > 0 and n > M,
!!(I- Mn(t)) x!! =!!(I- bn(]n(t) -I)) xi!> (1 + n b) !!xll

- bn(1-- n-1M)-1 l!x II= (1- bM (1- n- 1M)- 1) l!x II
Hence for sufficiently small b > 0, the bounded inverse (I -bAn(t))-1
exists with the estimate II (I-Mn (t))-1 !1< (1- bM (l-n- 1M)- 1). There-
427
fore, by the Lemma, we see that the solution x,. (t) of (G) satisfies
llx,.(t)ll ~ llx,(a)llexp((t-a)M1 ),M 1 ~2\I(1--n- 1 M)- 1
(U)
Thus x, (t) is uniquely determined by the initial condition Xn (a) =A (a)-1 y

and so we may put
xn(t) = U,.(t, a) x,.(a) = Un(t, a) A (a)- 1 y,

where
II Un(t, a) II~ exp((t- a)M1).
(10}
We next put
y,.(t) = A,(t) Un(t, a) A (a)- 1 y = An(t) x,.(t) = dx,(t)fdt.
(11)
We see, by (5), that A,.(t) is strongly continuously differentiable in t

and
(12)
dA,.(t)fdt = - An(t) B(t) A,.(t),
because we have
o-1 (An (t + o)- A,. (t)) =-A,. (t + o) {[A, (t + o)- 1 - A, (t)-1 ]/o} A, (t),
A, (t)- 1 = A (t)-1 - n- 1 1.
Thus y 11 (t) satisfies the initial value problem
dy,(t)fdt
An(t) y,(t)-A,(t) B(t) y,.(t),
a~t~b,
Yn(a) =],.(a) y. (13}
We see, by (8}, that A,(t) B(t) = ],(t) A (t) B(t) is strongly continuous
in t, and thus uniformly bounded in t and n by (4). Hence
IIAn(t)B(t)II<C for a~tM.

We have, foro> O,y, (t + o) = Yn (t)
(14)
+ OAn (t) y (t)-OAn (t)B (t) Yn (t) +o (o).
Hence, as in the proof of the Lemma, we obtain

and so
[!Yn(t)
II< [ly,(a) llexp((t-a) (M +C)) =llfn(a)y llexp((t-a) (M1 +C))

(15)
~ (1-n-1 M)-1 exp((t-a) (M1 +C)) IIYII
Hence, by dxn(t)fdt = Yn(t), we see that x,(t) is bounded and strongly

continuous in t, uniformly in t and n. Thus, by the reflexivity of X, there
exists a subsequence {n'} of {n} such that weak-lim Xn (t) = x (t) exists
simultaneously for all t, a
We next prove that
:S t ~
n'---700
b.
x (t) ED (A (t)) and A (t) x (t) = weak-lim A,.. (t) x,.. (t)
n'---+oo
is bounded and strongly measurable in t.
(1G)
Indeed, we can prove that A (t) x (t) is strongly continuous in t.

In the proof given below, we shall make use of the fact that, for the dual
428
operators, we have (see Theorem 1 in Chapter VIII, 6)
Jn (t)' = ((I- n-1 A (t))-1)' = (I'- n- 1 A (t)')- 1 ,
(A (t)-1)' = (A (t)')-1. (17)
Since X is reflexive, D (A (t)') is strongly dense in the dual space X' as

will be proved below. Hence, by 1/f,.(t)' II< (1-n-1 M)-1 which
follows from (4) and (17), we see that s-lim fn (t)' f' = f' for every f' E X'.
n->OO
Thus, by weak-lim x,..(t) = x(t), we get

n'~oo
weak-lim
We have
n'-+-oo
Jn' (t)
x,.. (t) = x (t).
<A (t) ],.(t) x,.(t), (A (t)-1)' f'>
(18)
= <J,.(t) x,.(t), f'>
by 17) and so, by (18),

lim <A (t) ],.. (t) x,.. (t), (A (t)-1)'
n~oo
f'>
= <x (t),
f'>.
(19)
Now, by (17), the range R((A(t)-1 )')=D(A(t)') and it is strongly

dense in X'. If otherwise, the reflexivity of X assures the existence of a
w0 EX, w0 =P 0, such that 0 = <w0 , (A (t)-1)' /') = <A (t)-1 w0 , f'). This
proves that A (t)-1 w0 = 0, i.e., w0 = 0 which is a contradiction. Therefore,
by the boundedness in t and n of y,. (t) = A,. (t) x,. (t), we see that
weak-lim A,..(t) x,..(t) = u(t) must exist. We have, by (19), <u(t),
n'->OO
(A (t)-1)' /') = <x(t), f'), that is, A (t)-1u(t) = x(t).

Since the strong continuity in t of A,. (t) x,. (t) implies that the closure
of the set {A,. (t) x,. (t); a< t < b, n = 1, 2, ... } is separable, the weak
measurability of u (t) = weak-lim A,.. (t) x,.. (t) in t implies, by Pettis'
n'->OO
theorem in Chapter V, 4, its strong measurability. To prove that u (t) is

strongly continuous in t, we integrate (13), obtaining
I
Yn (t)
Un (t, a) Yn (a) -
J Un(t, s) A,. (s) B (s) Yn (s) ds.
(20)
The second term on the right is strongly continuous in t, uniformly in

t and n, because of estimate (10), (14) and the boundedness of Yn(t) in t
and n. The first term on the right, U,.(t, a) y,.(a), is strongly continuous
in t, uniformly in t and n, if A,.(a) y,.(a) =A (a) fn(a) fn(a) y is bounded
in n. This we prove as in the above case for the strong continuity in t of
x,.(t) = U,.(t, a) Xn(a). But A,.(a) Yn(a) =A (a) ],.(a) ],.(a) y is bounded
inn if y is in the domain D (A (a)) of A (a), which is strongly dense in X.
Thus, by virtue of the Lemma, we see that, for any y E X, the first term
on the right of (20) is strongly continuous in t, uniformly in t and n. Thus
Yn (t) is strongly continuous in t, uniformly in t and n. This proves that
u (t) = A (t) x (t) = weak-lim y,.. (t) is strongly continuous in t.
~00
429
Therefore, by letting n'--+ =in the relation
Xn (t)
= Xn (a)
we obtain
JAn' (s) Xn (s) ds = A (a)-1y + JAn' (s) Xn (s) ds,

a
x(t) =A(a)-1y
(21)
J A(s)x(s)ds.
a
Hence, by the strong continuity in t of u(t) =A (t) x(t), we see that (1)
is satisfied. Since the solution x (t) is uniquely determined by the initial
value x (a) in virtue of the Lemma, we see that the original sequence
{xn (t)} itself must converge weakly to x (t).
Remark. The above result is adapted from K. YosmA [28]. Our condition (8) implies that the domain D (A (t)) of A (t) is independent oft as
T. KATO did remark to the present author. For, set A (t) (dA (t)- 1/dt) = C (t)
and let W(t) be the solution of the equation dW(t)fdt = - C(t) W(t),
Then d (A (t)-1 W (t))fdt =A (t)-1 C (t) W (t)- A (t)-1 C (t)
W (a) =I.
W(t) = 0, and so
A (a)-1 W (a) = A (a)- 1 = A (t)-1 W (t).
This proves that D (A (t)) = the range R (A (t)-1) is independent of t.

Theorem 2. Let X be a Hilbert space, and let A (t) satisfy, beside (4)
and (5), the further conditions:
A (t) is self-adjoint and (A (t) x, x) < -
II
x 11 2 for
xED(A(t)),
there exists a constant ex with 2-1
:::;:
ex ~ 1 such that
(-A(t))"B(t) is bounded in tin the operator norm.
(22)
(23)
Here the fractional power (-A (t))IX is defined, by virtue of the spectral
resolution A (t) =
-1
J AdE (J.), through

1
-00
-1
(-A(t))"'=
(-J.)"'dE 1 (J.).
(24)
-00
Then the sequence {xn (t)} converges weakly to the uniquely determined
strongly continuous solution of the integral equation (21).
Proof. The proof of Theorem 1 shows that we need only prove that
{Yn (t)} is uniformly bou11ded in t and n.
We have
d
dt
28
II
'
Yn(t)
112
d t , Yn(t) )
= 2 Re (dyn(t)
= 2Re(An(t) y,.(t), Yn(t))- 2Re(An(t) B(t) Yn(t),Yn(t)).
(25)
430
Since -A, (t) = -A (t) (I- n-1A (t))-1 is self-adjoint and non-negative
with -A (t), we have, by Schwarz' inequality,
IRe (An (t) B (t) X, x) I 0. Since An (t) = A (t) ] n (t) =
-1
J A. (1 -oo
spectrum of -An(t) lies in the interval [(1

2-1 < IX < 1, we see that
2-1e-1 II (-An(t)?-"'
112 < II (-An (t))1/2
n-1A.)-1dE1 (A.), the
+ n-1)-1, oo).
X
Hence, by
112 = (-An (t)
X,
x)
for all t and n, by taking e > 0 appropriately.

Therefore we see that the right hand side of (25) is, by (22) and (23),
smaller than
2-1 e II (-An(t))"' B(t) Yn(t) 11 2 = 2-1 e llfn(t)"' (-A (t))" B(t) Yn(t) 11 2
< K IIYn(t)
W, where K
is independent oft and n.

Hence llYn (t) II < llYn (a) II exp((t- a) K 112) and so Yn (t) is bounded in
t and n.
Remark. Let X = L2 (0, 1) and A (t) be the multiplication operator in
P(O, 1):
A (t) x(s) = - (1 + It-s 1-P) x(s).
If f3 > 2, then, by taking IX> 1/2 such that {3(1-IX) -1 > 0, we see
that the conditions of Theorem 2 are satisfied. A similar case was recently
treated by H. TANABE [1]. It seems that his method cannot be applied to
the casef3 = 1. Here we haveB (t) = (1 + It-s IJ- 2 or=- (1 + It-s J)-2
according as t > sort< s. Thus the right side of (25) in this case is < 0
and so 1/Yn(t) II< IIYn(a) JJ. Hence Theorem 2 applies to this case also.
References. The idea of approximating equation (1) by equations (6)
was first devised by the present author. See K. YosmA [23]. In this paper,
the B-space X need not be reflexive. Actually, equation (1) is treated
in the space V under the assumption that A (t) is a second order elliptic
differential operator with very smooth coefficients. It is proved that
the approximate solution Xn (t) E V will tend, as n t oo, to a distributional
solution of the differential equation (1). That x (t) itself is a genuine
solution of the differential equation (1) is proved by using a suitable
parametrix of the adjoint equation to (1):
-dy (s)fds = A* (s) y (s), where A* (s) is the formal dual operator of A (s).
See e.g., K. YosmA [24] and [25]. Cf. also S.lTo [1]. According to
J. KISYNSKI [1], published during the proof reading, the assumption
of the reflexivity of the space X in Theorem 1 can be omitted.
431
5. The Method of Tanabe and Sobolevski

Let X be a complex B-space, and consider the equation of evolution
in X with a given inhomogeneous term f (t) :
dx(t)Jdt = Ax(t)
+ f(t),
(1)
a< t <b.
Then the solution x (t) E X with the initial condition x (a) = x 0 E X is

given formally by the so-called Duhamel Principle from the solution
exp((t- a) A) x of the homogeneous equation dxfdt =Ax:
I
x(t) = exp ((t- a) A) x 0
+ J exp ((t- s) A) f(s) ds.
(2)
This suggests that a temporally inhomogeneous equation in X:

dx(t)Jdt =A (t) x(t), a< t
<
(3)
b,
may be solved formally as follows. We rewrite equation (3) in the form

(4)
dx(t)Jdt =A (a) x(t) + (A (t)- A (a)) x(t).
By virtue of the formalism (2), the solution x(t) of (4) with the initial
condition x(a) = x 0 will be given as the solution of an abstract integral
equation
x(t) = exp ((t- a) A (a)) x 0
(5)
I
+ exp ((t- s) A (s)) (A (s)- A (a)) x(s) ds.
Solving (5) formally by successive approximation, we obtain approximate

solutions:
x1 (t) = exp ((t- a) A (a)) x0 ,
x,.+l (t) = exp((t- a) A (a)) x0

I
+ Jexp ((t- s) A (s)) (A (s) -A (a)) x,. (s) ds.

a
Hence the solution x(t) of (5) would be given formally by

I
x (t) = exp ((t- a) A (a)) x 0
+ J exp ((t- s) A (s)) R (s, a) x0 ds,
(6)
where
R (t, s) =
00
.I Rm (t,s),
m=l
R 1 (t, s) = (A (t) -A (s)) exp ( (t- s) A (s)), s
< t,
=O,s~t.
I
Rm (t, s) =
28*
JR
s
1 (t,
a) Rm-l (a, s) da
(m = 2, 3, . ) .
(7)
432
Justification for the above formal procedure of integration has been

given by H. TANABE [2] using the theory of holomorphic semi-groups
as given in Chapter IX, 10. We shall follow Tanabe's approach, and
assume the following conditions:
For each t E [a, b], A (t) is a closed linear operator
with domain dense in X and range X such that
the resolvent set e (A (t)) of A (t) contains a
fixed angular domain
of the complex }.-plane
(8)
consisting of the origin 0 plus the set {.?.; -{} <
arg A< (} with (} > n/2}. The resolvent (U- A (t))-1
is strongly continuous in t uniformly in }. on any compact set ?;;; e.
There exists positive constants M and N such that,

for A E and t E [a, b], we have IICU- A (t))-1 II< N
(j.?. j- M)-1 whenever j.?. [ > M with N = 1 for real /c.
(9)
The domain D(A (t)) of A (t) is independent of t so

that, by the closed graph theorem in Chapter II, 6, the
operator A (t) A (s)-1 is in L (X, X). It is assumed
(10)
that there exists a positive constant K such that
IIA (t) A (s)-1 - A (r) A (s)-1 [[ < K [t -- r I for s, t and
rE[a,b].
Under these conditions, we can prove
Theorem. For any x 0 EX and s with a< s < b, the equation
dx(t)jdt =A (t) x(t), x(s) = x0 , s < t;;;: b
(3')
admits a uniquely determined solution x (t) EX. This solution is given by
x (t) = U (t, s) x (s) = U (t, s) x0 , where
(11)
U (t, s) = exp ((t- s) A (s)) + W (t, s),

W (t, s)
J exp ((t- a) A (a)) R (a, s) da
with R (t, s) given
(12)
by (7).
For the proof, we prepare three lemmas.
Lemma 1. R (t, s) satisfies, with a constant C,
IIR(t, s) II< KC exp(KC(t-s)),

(13)
and R (t, s) is strongly continuous in a < s < t < b.
Proof. By (8) and (9), we see that each A (s) generates a holomorphic
semi-group which is given by (see Chapter IX, 10)
exp(tA (s))
(2ni)-1
J ei.t (AI- A (s))-
d.?., where C' is
C'
a smooth contour running from
00 e-iO
to
00 eiO
in @.
(14)
433
Hence, by A (s) (U -A (s)) =A (U- A (s))-1 - I, we have, for (b- a)>

t> 0,
JJexp(tA(s))JJ
< C and I!A(s) exp(tA(s))l! < Ct-1 ,
where the positive constant C is independent of t
>
(15)
and sE [a, b].

We have, by (7),
R 1 (t, s) = (A (t)- A (s)) A (s)-1 A (s) exp ((t- s) A (s)), t > s,
and so, by (10) and (15),
(16)
It is also clear, from (8) and (14), that R 1 (t, s) is strongly continuous in
a< s < t < b. Next, by induction, we obtain
t
I!Rm(t, s)
II< J I!RI(t, a) II I!Rm-1(a, s) JJ da

s
<
J (K C)m (a- s)m- 2 ~ m- 2)- 1da

s
-~-
and hence (13) is obtained. In the same way, we see that R (t, s) is strongly
continuous in a < s < t < b.
Lemma 2. For s < -r
<
t, we have
IJR(t,s)-R(-r,s)IJ <
where
c1
c1
G-; +
(t--r)log:
is a positive constant which is independent of s,
:).
i
(17)
and t.
R 1(t, s)- R 1 (-r, s) = (A (t)- A (-r)) exp ((t- s) A (s))
(A (-r)- A (s)) [exp ((t- s) A (s))- exp ((-r- s) A (s))J.
By (10) and (15), the norm of the first term on the right is dominated by
KC (t- -r) (t- s)-1. :rhe second term on the right is
= (A (-r) -A (s))
t-s d
da exp (aA (s)) =(A (-r)- A (s)) A (s)-1
T-S
t-s
J
r-s
A (s) 2 exp (a A (s)) da,
434
and we have, by (15),
IITZS A (s) 2exp(a A (s)) dall < Tz II (A (s) exp (2-1a A (s)) )211 da
t-s
<
(2Cfa) 2 da
= 4C2 [ a1]
T-S
t-s
r-s
= 4C2(t --r) (t- s)-1 (-r- s)-1.
Therefore
IIRdt, s)- R1(-r, s) II < KC (1

On the other hand, by (7),
m=2
m=2
2: Rm(t,s)- 2: Rm(r,s)
JR
t
4C)
tt _
7:
(18)
s.
1 (t,a)R(a,s)da
I
- J R1 (-r,a) R(a,s) da = J R
S
1 (t,
a) R(a, s) da
+ J (R1 (t, a) -R1 {-r, a)) R(a, s) da.

T
The norm of the first term on the right is dominated by

t
J IIR1(t, a) IIIIR(a, s) II da <
K 2 C2 exp (KC (b -a)) (t-r).
The norm of the second term on the right is, by (13) and (18), dominated
by
T
J IIR1(t, a) -R1{-r, a) IIIIR(a, s) II da

s
<K2 C2 (1
+ 4C)
J (t-r)(t-a)T
exp (KC(b -a))
da
t-s
K 1 (t- -r) log -t- -7 :
Therefore we obtain (17).

Lemma 3. For s < t, we have
jjA (t) {exp ((t- s) A (t)) - exp ((t- s) A (s))} II < C2 , where
c2 is a positive constant independent of s and t.
(19)
Proof. We obtain, from (14),

A (t) {exp ((t- s) A (t))- exp ((t- s) A (s))}
(2nif 1
(l(t-s) A (t) (U- A (t))-1 (A (t)- A (s)) (U- A (s))-1dJ..
435
On the other hand, we have A (t) (AI -A (t))-1 =A.(U --A (t))-1 - I,
and so, by (9),
IIA (t)(U- A (t))-1 11 < JA.:A./ M
+ 1 for A.E e and tE
[a, b].
(20)
Hence we obtain (19) by (10) and

II (A (t)- A (s)) (AI- A (s))-111
::=:::
II (A (t)- A (s)) A (s)-1 lfJIA (s) (U- A (s))-111
Proof of the Theorem. We rewrite W (t, s) given in (12) as follows:
W(t,s)
J exp((t-r)A(t))R(t,s)dr
+ J {exp((t- -r) A (-r))- exp ((t- -r) A (t))} R (r, s) dr

s
J exp ((t- -r) A (t)) (R (-r, s)- R (t, s)) dr.
By approximating the integrals by Riemann sums and making use of the

closure property of the operator A (t), we see that we can apply A (t) to
each term of the right side of the above equality. For, by (19), we can
apply A (t) to the second term of the right side; and also to the third term
of the right side by (15) and (17); we also have, by A (t) exp ((t- -r) A (t))
= - d exp ((t- -r) A (t))jdr,
t
J exp((t--r) A (t)) R(t, s) dT =
A (t)
{exp ((t- s) A (t)) -I} R(t, s).
Hence we obtain
A (t) U (t, s) =A (t) exp ((t- s) A (s))
+ {exp ((t- s) A (t))- I} R (t, s)
+ J A (t) {exp ((t- -r) A (-r))- exp ((t- -r) A (t))} R (r, s) dr
(21)
+ J A (t) exp ((t--r) A (t)) (R(-r, s)- R(t, s)) dr.

s
The above proof shows that A (t) U (t, s) is strongly continuous m

a< s < t <band
JJA (t) W(t, s) II< C3 and IIA (t) U(t, s) II< C3 (t- s)-1,
(22)
where C3 is a positive constant independent of sand t.
We next define, for s < (t-c- h)
Uh(t,s) =exp((t-s)A(s))
< t,
t-h
exp((t--r)A(-r))R(-r,s)d-r.
(23)
436
Since a holomorphic semi-group exp (tA (u)) is differentiable in t > 0,

we have
8
8t U,. (t, s) =A (s) exp ((t- s) A (s))
+ exp (hA (t- h)) R (t- h, s)
t-h
+Js
A('r)exp((t-r)A(r))R(r,s)dr.
Hence we have, by (7),

8
8t U,. (t, s)- A (t) U,. (t, s) = exp (hA (t- h)) R (t- h, s)- R 1 (t, s)
(24)
t-h
- sJ R
1 (t,
r)R(r, s) dr.
By (8), (13) and (14), exp(hA(t-h))R(t-h,s) tends strongly to

t 0. Thus we have
R (t, s) as h
s-~\~ ( :t U,. (t, s) -A (t)
U,. (t,
s)) x
(25)
=(R(t,s)-R1 (t,s)-j R 1 (t,a)R(a,s)da)x0 , x0 EX.
The right side must be 0 as may easily be proved by (7). Since we have
s-lim A (t) U,. (t, s) x 0 = A (t) U (t, s) x0
h.J-0
by the reasoning used in proving (21), we obtain from (25)
s-lim~ U,.(t, s) x 0 =A (t) U(t, s) x0 for t > s and x0 E X.

h.J-0 v
(26)
The right side of (26) being strongly continuous in t > s, we see, by

integrating (26) and remembering s-lim U,. (t, s) x0 = U (t, s) x 0 , that
h.J-0
ot
U (t,s) x0 = A (t) U (t, s) x 0 for t
>
s and x0 E X.
(27)
Therefore, x(t) = U(t, s) x0 is the desired solution of (3'). The unique,ness of the solution may be proved as in the preceding section.

The above Theorem and proof are adapted from H. TANABE [2]. To
illustrate his idea we have made Tanabe's conditions somewhat
stronger. Thus, e.g., the condition (9) may be replaced by a weaker
one:
437
For the details, see the above cited paper by H. TANABE which is a refinement of H. TANABE [3] and [4]. It is to be noted that the Russian school
independently hasdevelopedasimilarmethod. See, e.g., P. E. SoBOLEVSKI
[1] and the reference cited in this paper.
Komatsu's work. H. KoMATSU [1] made an important remark regarding Tanabe's result given above. Let L1 be a convex complex neighbourhood of the real segment [a, b] considered as embedded in the complex plane. Suppose that A (t) is defined fortE L1 and satisfies (8) and (9)
in which the phrase "t E [a, b]" is replaced by "t E L1". Assume, moreover, that there exists a bounded linear operator A0 which maps X onto
D, the domain of A (t) which is assumed to be independent of t E L1, in a
one-to-one manner and such that B (t) = A (t) A0 is strongly holomorphic
in t E .d. Under these assumptions, KoMATSU proved that the operator
U (t, s) constructed as above is strongly holomorphic in t E L1 if
larg(t- s) I< 00 with a certain 00 satisfying 0 < 00 < nj2.
The result may be applied to the "forward and backward unique continuation" of solutions of temporally inhomogeneous diffusion equations
as in Chapter XIV, 1. In this connection, we refer to H. KoMATSU [2],
[3] and T. KoTAKE-M. NARASIMHAN [1].
Kato's work. To get rid of the assumption that the domain D (A (t))
is independent of t, T. KATO [6] proved that, in the above Theorem, we
may replace condition (10) by the following:
For a certain positive integer k, the domain D ( (-A (t))ll k)
is independent of t. (Here (-A (t)) 1fk is the fractional
power as defined in Chapter IX, 11.) Further, there
exist constants K 2 > 0 and y with 1- k-1 < y < 1
such that, for s, t E [a, b], II (-A (t)?fk (-A (s)?fk- I II
(10)'
< K21t- slY
Tanabe's and the Kato-Tanabe recent work. With the same motivation
as KATO, H. TANABE [1] devised a method to replace condition (10) by
A (t)-1 is once strongly differentiable in a < t < b

and such that, for positive constants K 3 and ex,
lldAJ:)-1- dAJ:)-111 < Kalt- s I"
Further, there exist positive constants N and

0 < e < 1 such that
II :t (AI- A (t))-111 <
N IJ.Iq-1, !!
<
e with
(10)"
1.
For details, see KATO-TANABE [8]. Their point is to start with the first
approximation exp ((t- a) A (t)) x0 instead of exp ((t- a) A (a))x0
438
Nelson's work on Feynman integrals. The semi-group method of

integration of Schrodinger equations gives an interpretation of
Feynman integrals. See E. NELSON [2].
The Agmon-Nirenberg work. Agmon-Nirenberg [1] discussed the
behaviour as t t oo of solutions of the equation
some B-spaces.
! ~~ --Au= 0 in
Bibliography
AcHIESER, N. I.
[1] (with I. M. GLAZMAN) Theorie der linearen Operatoren im HilbertRaum, Akademie-Verlag 1954.
AGMON, s.
[1] (with L. NIRENBERG) Properties of solutions of ordinary differential
equations in Banach space. Comm. P. and Appl. Math. 16, 121-239
(1963).
AKILOV, G.
(1) See KANTOROVITCH-AKILOV (1).
ALEXANDROV, P.
[1] (with H. HoPF) Topologie, Vol. I, Springer 1935.
ARONSZAJN, N.
[1] Theory of reproducing kernels. Trans. Amer. Math. Soc. 68, 337-404
(1950).
BALAKRISHNAN, V.
[1] Fractional powers of closed operators and the semi-groups generated
by them. Pacific J. Math. 10, 419-437 (1960).
BANACH, S.
[1] Theorie des Operations Lineaires, Warszawa 1932.
[2] Sur la convergence presque partout de fonctionnelles lineaires. Bull.
Sci. Math. France 50, 27-32 and 36-43 (1926).
BEBOUTOV, M.
[1] Markoff chains with a compact state space. Rec. Math. 10, 213---238
(1942).
BERGMAN, S.
[1] The Kernel Function and the Conformal Mapping. Mathematical Surveys, No. 5 (1950).
BERS,L.
[1] Lectures on Elliptic Equations. Summer Seminar in Appl. Math. Univ.
of Colorado, 1957.
BIRKHOFF, G.
[1] Lattice Theory. Colloq. Publ. Amer. Math. Soc., 1940.
BIRKHOFF, G. D.
[1] Proof of the ergodic theorem. Proc. Nat. Acad. Sci. USA 17, 656-660
(1931).
[2] (with P. A. SMITH) Structure analysis of surface transformations. J.
Math. Pures et Appliq. 7, 345-379 (1928).
BocHNER, S.
[1] Integration von Funktionen, deren Wert die Elemente eines Vektorraumes sind. Fund. Math. 20, 262-276 (1933).
[2] Diffusion equations and stochastic processes. Proc. Nat. Acad. Sci.
USA 35, 369-370 (1949).
[3] Vorlesungen iiber Fouriersche Integrale, Akademie-Verlag 1932.
440
Bibliography
[4] Beitrage zur Theorie der fastperiodischen Funktionen. Math. Ann. 96,
119-147 (1927).
BOGOLIOUBOV, N.
[1] See KRYLov-BoGOLIOUBOV [1].
BOHR, H.
[1] Fastperiodische Funktionen, Springer 1932.
BOURBAKI, N.
[1] Topologie Generale. Act. Sci. et Ind., nos. 856, 916, 1029, 1045, 1084,
Hermann 1940-42.
[2] Espaces Vectoriels Topologiques. Act. Sci. et Ind., nos. 1189, 1229,
Hermann 1953-55.
BROWDER, F. E.
[1] Functional analysis and partial differential equations, I-II. Math.
Ann. 188, 55--79 (1959) and 145, 81-226 (1962).
CHACON, R. V.
[1] (with D. S. ORNSTEIN) A general ergodic theorem. Ill. J. Math. 4, 153160 (1960).
CHOQUET, G.
[1] La thoorie des representations integrales dans les ensembles convexes
compacts. Ann. Inst. Fourier 10, 334--344 (1960).
[2] (with P.-A. MEYER) Existence et unicite des representations integrales
dans les convexes compacts quelconques. Ann. Inst. Fourier 18, 139154 (1963).
CLARKSON, J. A.
[1] Uniformly convex spaces. Trans. Amer. Math. Soc. 40, 396-414 (1936).
DIEUDONNE, J.
[1] Recent advances in the theory of locally convex vector spaces. Bull.
Amer. math. Soc. 59, 495--512 (1953).
DooB,
J. L.
[1] Stochastic processes with an integral-valued parameter. Trans. Amer.

Math. Soc. 44, 87-150 (1938).
[2] Probability theory and the first boundary value problem. Ill. J. Math.
2, 19-36 (1958).
[3] Probability methods applied to the first boundary value problem. Proc.
Third Berkeley Symp. on Math. Statist. and Prob. II, 49-80 (1956).
DUNFORD, N.
[1] (with J. ScHWARTz) Linear Operators, Vol. I, Interscience 1958.
[2] Uniformity in linear spaces. Trans. Amer. Math. Soc. 44, 305--356
(1938).
[3) On one-parameter groups of linear transformations. Ann. of Math. 89,
569-573 (1938).
[4] (with J. ScHWARTZ) Convergence almost everywhere of operator averages. J. Rat. Mech. Anal. 5, 129-178 (1956).
[5] (with J. ScHWARTZ) Linear Operators, Vol. II, Interscience 1963.
[6] (with J. ScHWARTZ) Linear Operators, Vol. III, Interscience, to appear.
DYNKIN, E. B.
[1] Markoff processes and semi-groups of operators. Teorya Veroyatn. 1
(1956).
[2] Infinitesimal operators of Markoff processes. Teorya Veroyatn. 1 (1956).
EBERLEIN, w. F.
[1] Weak compactness in Banach spaces. Proc. Nat. Acad. Sci. USA 88,
51-53 (1947).
Bibliography
441
EHRENPREIS, L.
[1] Solutions of some problems of division. Amer. J. Math. 76, 883-903
(1954).
ERDELYI, A.
[1] Operational Calculus and Generalized Functions, Reinhart 1961.
FELLER, W.
[1] On the generation of unbounded semi-groups of bounded linear operators. Ann. of Math. 68, 166-174 (1953).
[2] The parabolic differential equation and the associated semi-group of
transformations. Ann. of Math. 66, 468-519 (1952).
[3] On the intrinsic form for second order differential operators. Ill. J.
Math. 2, No. 1, 1-18 (1958).
[ 4] Generalized second order differential operators and their lateral conditions. Ill. J. Math. 1, No. 4, 459-504 (1957).
[5] Some new connections between probability and classical analysis. Proc.
Internat. Congress of Math. 1958, held at Edinburgh, pp. 69-86.
[6] On differential operators and boundary conditions. Comm. Pure and
Appl. Math. 8, 203-216 (1955).
[7] Boundaries induced by non-negative matrices. Trans. Amer. Math. Soc.
83, 19-54 (1956).
[8] On boundaries and lateral conditions for the Kolmogoroff differential
equations. Ann. of Math. 66, 527-570 (1957).
ForAs, C.
[1] Remarques sur les semi-groupes distributions d'operateurs normaux.
Portugaliae Math. 19, 227-242 (1960).
FREUDENTHAL, H.
[1] Dber die Friedrichssche Fortsetzung halbbeschrankter Hermitescher
Operatoren. Proc. Acad. Amsterdam 39, 832-833 (1936).
[2] Teilweise geordnete Modulen. Proc. Acad. Amsterdam 39, 641-651
(1936).
FRIEDMAN, A.
[1] Generalized Functions and Partial Differential Equations, PrenticeHall 1963.
FRIEDRICHS, K.
[1] Differentiability of solutions of elliptic partial differential equations.
Comm. Pure and Appl. Math. 6, 299-326 (1953).
[2] Symmetric positive systems of differential equations. Comm. Pure and
Appl. Math. 11, 333-418 (1958).
[3] Spektraltheorie halbbeschrankter Operatoren, I-III. Math. Ann. 109,
465-487, 685-713 (1934) and 110, 777-779 (1935).
FUKAMIYA, M.
[1] Topological methods for Tauberian theorem. Tohoku Math. J. 77-87
(1949).
[2] See YosmA-FUKAMIYA [16].
G..i.RDING, L.
[1] Dirichlet's problem for linear elliptic partial differential equations.
Math. Scand. 1, 55-72 (1953).
[2] Some trends and problems in linear partial differential equations. Internat. Congress of Math. 1958, held at Edinburgh, pp. 87-102.
GELFAND, I. M.
[1] (with I. M. SILOV) Generalized Functions, Vol. I-III, Moscow 1958.
[2] Normierte Ringe. Rec. Math. 9, 3-24 (1941).
442
Bibliography
[3] (withN. Y. VILENKIN} SomeApplicationsofHarmonicA nalysis (Vol. IV

of Generalized Functions), Moscow 1961.
[4] (with D. RAIKov) On the theory of characters of commutative topological groups. Doklady Akad. Nauk SSSR 28, 195-198 (1940).
[5] (with D. A. RAIKOV and G. E. ~hLOv) Commutative Normed Rings,
Moscow 1960.
GLAZMAN, I. M.
[1] See AcHIESER-GLAZMAN [1].
GROTHENDIECK, A.
[1] Espaces Vectoriels Topologiques, seconde ed., Sociedade de Mat. de Sao
Paulo 1958.
[2] Produits Tensoriels Topologiques et Espaces Nucleaires. Memoirs of
Amer. Math. Soc. No. 16 (1955).
HAHN, H.
[1] "Ober Folgen linearer Operatoren. Monatsh. fiir Math. und Phys. 32,
3-88 (1922).
[2] "Ober lineare Gleichungssysteme in linearen Raumen. J. reine und angew.
Math. 157, 214-229 (1927).
[3] "Ober die Integrale des Herrn Hellinger und die Orthogonalinvarianten
der quadratischen Formen von unendlich vielen Veranderlichen.
Monatsh. fiir Math. und Phys. 23, 161-224 (1912).
HALMOS, P. R.
[1] Measure Theory, van Nostrand 1950.
[2] Introduction to Hilbert Space and the Theory of Spectral Multiplicity,
Chelsea 1951.
HAUSDORFF, F.
[1] Mengenlehre, W. de Gruyter 1935.
HELLINGER, E.
[1] Neue Begriindung der Theorie quadratischer Formen von unendlichvielen Veranderlichen. J. reine und angew. Math. 136, 210-271 (1909).
HELLY, E.
[1] "Ober Systeme linearer Gleichungen mit unendlich vielen Unbekannten.
Monatsh. fiir Math. und Phys. 31, 60-91 (1921).
HILBERT, D.
[1] Wesen und Ziele einer Analysis der unendlich vielen unabhangigen
Variablen. Rend. Circ. Mat. Palermo 27, 59-74 (1909).
HILLE, E.
[1] (with R. S. PHILLIPS) Functional Analysis and Semi-groups. Colloq.
Publ. Amer. Math. Soc., 1957. It is the second edition of the book
below.
[2] Functional Analysis and Semi-groups. Colloq. Publ. Amer. Math. Soc.,
1948.
[3] On the differentiability of semi-groups of operators. Acta Sci. Math.
Szeged 12B, 19-24 (1950).
[4] On the generation of semi-groups and the theory of conjugate functions.
Proc. R. Physiogr. Soc. Lund 21, 1-13 (1951).
[5] Une generalization du probleme de Cauchy. Ann. Inst. Fourier 4, 31-48
(1952).
[6] The abstract Cauchy problem and Cauchy's problem for parabolic
differential equations. J. d' Analyse Math. 3, 81-196 (1954).
Bibliography
443
[7] Perturbation methods in the study of Kolmogoroff's equations. Proc.

Internat. Congress of Math. 1954, held at Amsterdam, Vol. III,
pp. 365--376.
[8] Linear differential equations in Banach algebras. Proc. Internat.
Symposium on Linear Analysis, held at Jerusalem, 1960, pp. 263-273.
[9] Les probabilites continues en chaine. C. R. Acad. Sci. 230, 34-35 (1950).
HOFFMAN, K.
[1] Banach Spaces of Analytic Functions, Prentice-Hall 1962.
HOFF, E.
[1] Ergodentheorie, Springer 1937.
[2] The general temporally discrete Markoff processes. J. Rat. Mech. and
Anal. 3, 13-45 (1954).
[3] On the ergodic theorem for positive linear operators. J. reine und angew. Math. 205, 101-106 (1961).
HOFF, H.
[1] See ALEXANDROV-HOFF [1].
H6RMANDER, L.
[1] On the theory of general partial differential operators. Acta Math. 94,
161-248 (1955).
[2] Lectures on Linear Partial Differential Equations, Stanford Univ.
1960.
[3] Linear partial differential equations without solutions. Math. Ann. 140,
169-173 (1960).
[4] Local and global properties of fundamental solutions. Math. Scand. o,
27-39 (1957).
[5] On the interior regularity of the solutions of partial differential equations. Comm. Pure and Appl. Math. 9, 197-218 (1958).
[6] Linear Partial Differential Operators, Springer 1963.
HUNT, G. A.
[1] Markoff processes and potentials, I-II. Ill. J. of Math. 1 and 2 (1957
and 1958).
ITo, K.
[1] (with H. P. McKEAN) Diffusion Processes and Their Sample Paths,
Springer, to appear.
IT6, s.
[1] The fundamental solutions of the parabolic differential equations in
differentiable manifold. Osaka Math. J. o, 75--92 (1953).
[2] (with H. YAMABE) A unique continuation theorem for solutions of a
parabolic differential equation. J. Math. Soc. Japan 10, 314-321 (1958).
jACOBS, K.
[1] Neuere Methoden und Ergebnisse der Ergodentheorie, Springer 1960.
JOHN, F.
[1] Plane Waves and Spherical Means Applied to Partial Differential Equations, Interscience 1955.
KAKUTANI, S.
[1] Iteration of linear operations in complex Banach spaces. Proc. Imp.
Acad. Tokyo 14, 295--300 (1938).
[2] See YosmA-MIMURA-KAKUTANI [10].
[3] Weak topology and regularity of Banach spaces. Proc. Imp. Acad.
Tokyo lo, 169-173 (1939).
[4] Concrete representation of abstract (M)-spaces. Ann. of Math. 42, 9941024 (1941).
444
Bibliography
[5) Concrete representation of abstract (L)-spaces and the mean ergodic

theorem. Ann. of Math. 42, 523-537 (1941).
[6] Ergodic theorems and the Markoff processes with a stable distribution.
Proc. Imp. Acad. Tokyo 16, 49-54 (1940).
[7] See YosiDA-KAKUTANI [7].
[8] Ergodic Theory. Proc. Internat. Congress of Math. 1950, held at
Cambridge, Vol. 2, pp. 128-142.
KANTOROVITCH, L.
[1] (with G. AKILOV) Functional Analysis in Normed Spaces, Moscow
1955.
KATO, T.
[1] Remarks on pseudo-resolvents and infinitesimal generators of semigroups. Proc. Japan Acad. 35, 467-468 (1959).
[2] Note on fractional powers of linear operators. Proc. Japan Acad. 36,
94-96 (1960).
[3] Integration of the equation of evolution in a Banach space. J. Math.
Soc. of Japan 5, 208-234 (1953).
[4] On linear differential equations in Banach spaces. Comm. Pure and
Appl. Math. 9, 479-486 (1956).
[5] Fractional powers of dissipative operators. J. Math. Soc. of Japan 13,
246-274 (1961); II, ibid. 14, 242-248 (1962).
[6] Abstract evolution equations of parabolic type in Banach and Hilbert
spaces. Nagoya Math. J. 19, 93-125 (1961).
[7] Fundamental properties of Hamiltonian operators of Schrodinger type.
Trans. Amer. Math. Soc. 70, 195-211 (1950).
[8] (with H. TANABE) On the abstract evolution equation. Osaka Math. J.
14, 107-133 (1962).
KELLEY, J. L.
[1] General Topology, van Nostrand 1955.
[2] Note on a theorem of Krein and Milman. J. Osaka Inst. Sci. Tech.,
Part I, 3, 1-2 (1951).
KELLOGG, 0. D.
[1] Foundations of Potential Theory, Springer 1929.
KHINTCHINE, A.
[1] Zu Birkhoffs Losung des Ergodenproblems. Math. Ann. 107, 485-488
(1933).
KISYNSKI, J.
[1] Sur les operateurs de Green des problemes de Cauchy abstraits. Stud.
Math. 23, 285-328 (1964).
KODAIRA, K.
[1] The eigenvalue problem for ordinary differential equations of the second
order and Heisenberg's theory of 5-matrices. Amer. J. of Math. 71,
921-945 (1949).
KOLMOGOROV, A.
[1] Ober analytische Methoden in der Wahrscheinlichkeitsrechnung. Math.
Ann. 104, 415-458 (1931).
KOMATSU, H.
[1] Abstract analyticity in time and unique continuation property of solutions of a parabolic equation. J. Fac. Sci. Univ. Tokyo, Sect. 1, 9,
Part 1, 1-11 (1961).
[2] A characterization of real analytic functions. Proc. Japan Acad. 36,
90-93 (1960).
Bibliography
445
[3] A proof of Kotake-Narasimhan's theorem. Proc. Japan Acad. 38, 615618 (1962).
[ 4] Semi-groups of operators in locally convex spaces, to appear.
KoTAKE, T.
[1] Sur l'analyticite de la solution du probleme de Cauchy pour certaines
classes d'operateurs paraboliques. C. R. Acad. Sci. Paris 252, 3716-3718
(1961).
[2] (with M.S. NARASIMHAN) Sur la regularite de certains noyaux associes
a un operateur elliptique. C. R. Acad. Sci. Paris 252, 1549-1550 (1961).
KoTHE, G.
[1] Topologische lineare Rii.ume, Vol. I, Springer 1960.
KREIN, M.
[1] (with D. MILMAN) On extreme points of regularly convex sets. Stud.
Math. 9, 133-138 (1940).
[2] (with S. KREIN) On an inner characteristic of the set of all continuous
functions defined on a bicompact Hausdorff space. Doklady Akad.
Nauk SSSR 27, 429-430 (1940).
KREIN, S.
[1] See KREIN-KREIN [2].
KRYLOV, N.
[1] (with N. BoGOLIOUBov) La theorie g{merale de la mesure dans son
application a l'etude des systemes de la mecanique non lineaires. Ann.
of Math. 38, 65-113 (1937).
LADYZHENSKAYA,O.A.
[1] (with I. M. VISIK) Problemes aux limites pour les equations aux derivees partielles et certaines classes d'equations operationnelles. Ousp.
Mat. Nauk 11, 41-97 (1956).
LAx,P.D.
[1] (with A. N. MILGRAM) Parabolic equations, in Contributions to the
Theory of Partial Differential Equations, Princeton 1954.
[2] On Cauchy's problem for hyperbolic equations and the differentiability
of solutions of elliptic equations. Comm. Pure and Appl. Math. 8,
615--633 (1955).
[3] (with R. S. PHILLIPS) Local boundary conditions for dissipative system
of linear partial differential operators. Comm. Pure and Appl. Math.
13, 427-455 (1960).
LERAY, J.
[1] Hyperbolic Differential Equations, Princeton 1952.
LEWY, H.
[1] An example of a smooth linear partial differential equation without
solutions. Ann. of Math. 66, 155--158 (1957).
LIONS, J. L.
[1] Une remarque sur les applications du theoreme de Hille-Yosida. J.
Math. Soc. Japan 9, 62-70 (1957).
[2] Equations Differentielles Operationnelles, Springer 1961.
[3] Espaces d'interpolation et domaines de puissance fractionaires d'operateurs. J. Math. Soc. Japan 14, 233-241 (1962).
[4] Les semi-groups distributions. Portugaliae Math. 19, 141-164 (1960).
LUMER, G.
[1] Semi-inner product spaces. Trans. Amer. Math. Soc. 100, 29-43 (1961).
[2] (with R. S. PHILLIPS) Dissipative operators in a Banach space. Pacific
J. Math. 11, 679-698 (1961).
29
446
Bibliography
MAAK, w.
[1] Fastperiodische Funktionen. Springer 1950.
MALGRANGE, B.
[1] Existence et approximation des solutions des equations aux derivees
partielles et des equations de convolution. Ann. Inst. Fourier 6, 271355 (1955-56).
[2] Sur une classe d'operateurs differentielles hypoelliptiques. Bull. Soc.
Math. France 58, 283-306 (1957).
MARUYAMA, G.
[1] On strong Markoff property. Mem. Kyushu Univ. 13 (1959).
MAZUR, S.
[1] Surles anneaux lineaires. C. R. Acad. Sci. Paris 207, 1025--1027 (1936).
[2] Uber konvexe Mengen in linearen normierten Raumen. Stud. Math. 5,
70-84 (1933).
McKEAN, H.
[1] See ITo-McKEAN [1].
MEYER, P. A.
[1] Seminaire de Theorie du Potentiel, sous Ia direction de M. BRELOTG. CHOQUET-J. DENY, Faculte des Sciences de Paris, 1960-61.
[2] See CHOQUET-MEYER [2].
MIKUSINSKI, J.
[1] Operational Calculus, Pergamon 1959.
MILGRAM, A. N.
[1] See LAX-MILGRAM [1].
MILMAN, D.
[1] On some criteria for the regularity of spaces of the type (B). Doklady
Akad. Nauk SSSR 20, 20 (1938).
[2] See KREIN-MILMAN [1].
MIMURA, Y.
[1] See YosmA-MIMURA-KAKUTANI [10].
[2] Uber Funktionen von Funktionaloperatoren in einem Hilbertschen
Raum. Jap. J. Math. 13, 119-128 (1936).
MINLOS, R. A.
[1] Generalized stochastic processes and the extension of measures. Trudy
Moscow Math. 8, 497-518 (1959).
MIYAD ERA, I.
[1] Generation of a strongly continuous semi-group of operators. Tohoku
Math. J. 4, 109-121 (1952).
MIZOHATA, s.
[1] Hypoellipticite des equations paraboliques. Bull. Soc. Math. France
85, 15--50 (1957).
[2] Analyticite des solutions elementaires des systemes hyperboliques et
paraboliques. Mem. Col!. Sci. Univ. Kyoto 32, 181-212 (1959).
[3] Unicite du prolongement des solutions pour quelques operateurs differentiels paraboliques. Mem. Coli. Sci. Univ. Kyoto, Ser. A 31, 219-239
(1958).
[4] Le probleme de Cauchy pour les equations paraboliques. J. Math. Soc.
Japan 8, 269-299 (1956).
[5] Systemes hyperboliques. J. Math. Soc. Japan 11, 205-233 (1959).
MORREY, c. B.
[1] (with L. NIRENBERG) On the analyticity of the solutions of linear
Bibliography
447
elliptic systems of partial differential equations. Comm. Pure and Appl.

Math. 10, 271-290 (1957).
NAGUMO,M.
[1] Einige analytische Untersuchungen in linearen metrischen Ringen. Jap.
J. Math. 18, 61-80 (1936).
[2] Re-topologization of functional spaces in order that a set of operators
will be continuous. Proc. Japan Acad. 37, 550---552 (1961).
NAGY, B. VON Sz.
[1] Spektraldarstellung linearer Transformationen des Hilbertschen Raumes, Springer 1942.
[2] See RIEsz-NAGY [3].
[3] Prolongements des transformations de l'espace de Hilbert qui sortent
de cet espace. Akad. Kiado, Budapest 1955.
NAIMARK, M. A.
[1] Normed Rings, P. Noordhoff 1959.
[2] Lineare differentiale Operatoren, Akad. Verlag 1960.
[3] tJber Spektralfunktionen eines symmetrischen Operators. Izvestia
Akad. Nauk SSSR 17, 28l"r-296 (1943).
NAKANO, H.
[1] Unitarinvariante hypermaximale normale Operatoren, Ann. of Math.
42, 657-664 (1941).
NARASIMHAN, M. S.
[1] See KoTAKE-NARASIMHAN [2].
NELSON, E.
[1] Analytic vectors. Ann. of Math. 670, 572-615 (1959).
[2] Feynman integrals and the Schrodinger equations. J. of Math. Physics
o, 332-343 (1964).
NEUMANN, J. VON
[1] Allgemeine Eigenwerttheorie Hermitescher Funktionaloperatoren.
Math. Ann. 102, 49-131 (1929).
[2] On rings of operators, III. Ann. of Math. 41, 94-161 (1940).
[3] Zur Operatorenmethode in der klassischen Mechanik. Ann. of Math. 38,
587-643 (1932).
[4] Almost periodic functions in a group. I. Trans. Amer. Math. Soc. 36,
44fr--492 (1934).
[5] tJber adjungierte Funktionaloperatoren. Ann. of Math. 33, 249-310
(1932).
[6] tJber die analytischen Eigenschaften von Gruppen Iinearer Transformationen und ihrer Darstellungen. Math. Z. 80, 3--42 (1929).
[7] Zur Algebra der Funktionaloperatoren und Theorie der normalen
Operatoren. Math. Ann. 102, 370---427 (1929-30).
[8] tJber einen Satz von Herm M. H. Stone. Ann. of Math. 83, 567-573
(1932).
NIRENBERG, L.
[1] Remarks on strongly elliptic partial differential equations. Comm. Pure
and Appl. Math. 8, 643--674 (1955).
[2] On elliptic partial differential equations. Ann. Scuola Norm. Sup. Pisa
18, llfr--162 (1959).
[3] See MoRREY-NIRENBERG [1].
[4] See AGMON-NIRENBERG [1].
ORNSTEIN, D. S.
[1] See CHACON-ORNSTEIN [1].
29*
448
Bibliography
PALEY, R. E. A. c.
[1] (with N. WIENER) Fourier Transforms in the Complex Domain, Colloq.
Publ. Amer. Math. Soc., 1934.
PEETRE, J.
[1] A proof of the hypoellipticity of formally hypoelliptic differential operators. Comm. Pure and Appl. Math. 16, 737-747 (1961).
PETER, F.
[1] (with H. WEYL) Die Vollstandigkeit der primitiven Darstellungen einer
geschlossenen kontinuierlichen Gruppe. Math. Ann. 97, 737-755
(1927).
PETROWSKY, I. G.
[1] Sur l'analyticite des solutions d'equations differentielles. Rec. Math. 47
(1939).
PETTIS, B. J.
[1] On integration in vector spaces. Trans. Amer. Math. Soc. 44, 277-304
(1938).
PHILLIPS, R. s.
[1] See HILLE-PHILLIPS [1].
[2] The adjoint semi-group. Pacific J. Math. o, 269-283 (1955).
[3] An inversion formula for Laplace transform and semi-groups of linear
operators. Ann. of Math. 69, 325---356 (1954).
[4] See LuMER-PHILLIPS [2].
[5] On the generation of semi-groups of linear operators. Pacific J. Math.
2, 343-369 (1952).
[6] On the integration of the diffusion equation with boundaries. Trans.
Amer. Math. Soc. 98, 62-84 (1961).
[7] Dissipative operators and parabolic partial differential operators.
Comm. Pure and Appl. Math. 12, 249-276 (1959).
[8] Dissipative hyperbolic systems. Trans. Amer. Math. Soc. 86, 109-173
(1957).
[9] Dissipative operators and hyperbolic systems of partial differential
equations. Trans. Amer. Math. Soc. 90, 193-254 (1959).
[10] See LAX-PHILLIPS [3].
PITT, H. R.
[1] Tauberian Theorems. Tata Inst. of Fund. Research, 1958.
PONTRJAGIN, L.
[1] Topological Groups, Princeton 1939.
RAno, T.
[1] Subharmonic Functions, Springer 1937.
RAIKOV, D. A.
[1] See GELFAND-RAIKOV [4].
[2] See GELFAND-RAIKov-f:hwv [5].
RAY, D.
[1] Resolvents, transition functions and strongly Markovian processes. Ann.
of Math. 70, 43-72 (1959).
RICKART, c. E.
[1] General Theory of Banach Algebras, van Nostrand 1960.
RIESZ, F.
[1] Zur Theorie des Hilbertschen Raumes. Acta Sci. Math. Szeged 7, 34-38
(1934).
[2] t!ber lineare Funktionalgleichungen. Acta Math. 41, 71-98 (1918).
Bibliography
449
[3] (with B. VON Sz. NAGY) Le<;ons d'Analyse Fonctionelle, Akad. Kiado,
Budapest 1952.
[4] Some mean ergodic theorems. J. London Math. Soc. 13, 274---278
(1938).
[5] Sur les fonctions des transformations hermitiennes dans l'espace de
Hilbert. Acta Sci. Math. Szeged 7, 147-159 (1935).
[6] Sur la Decomposition des Operations Lineaires. Proc. Internat. Congress
of Math. 1928, held at Bologna, Vol. III, 143-148.
RYLL-NARDZEWSKI, C.
[1] See J. MrKusr:NsKI [1].
SAKS, s.
[1] Theory of the Integral, Warszawa 1937.
[2] Addition to the note on some functionals. Trans. Amer. Math. Soc. 3o,
967-974 (1933).
ScHATTEN, R.
[1] A Theory of Cross-spaces, Princeton 1950.
ScHAUDER, J.
[1] Dber lineare, vollstetige Funktionaloperationen. Stud. Math. 2, 1-6
(1930).
ScHWARTZ, J.
[1] See DuNFORD-SCHWARTZ [1].
[3] See DuNFORD-SCHWARTz [5].
SCHWARTZ, L.
[1] TMorie des Distributions, vol. I et II, Hermann 1950, 1951.
[2] Transformation de Laplace des distributions. Comm. Sem. Math. de
l'Univ. de Lund, tome suppl. dedie aM. Riesz, 196-206 (1952).
[3] Lectures on Mixed Problems in Partial Differential Equations and the
Representation of Semi-groups. Tata Inst. Fund. Research, 1958.
[4] Les equations d'evolution liees au produit de compositions. Ann. Inst.
Fourier 2, 16fr-169 (1950-1951).
[5] Expose sur les travaux de Garding, Seminaire Bourbaki, May 1952.
SEGAL, I. E.
[1] The span of the translations of a function in a Lebesgue space. Proc.
Nat. Acad. Sci. USA 30, 165-169 (1944).
SHMULYAN, V. L.
[1] Dber lineare topologische Raume. Math. Sbornik, N. S. 7 (49), 425-448
(1940).
SILOV, G.
[1] See GELFAND-SILOV [1].
[2] See GENFAD-RALKOV-SILOV [5].
SMITH, P. A.
[1] See BrRKHOFF-SMITH [2].
SOBOLEV, S. L.
[1] Sur un tMoreme d'analyse fonctionnelle. Math. Sbornik 4o, 471-496
(1938).
[2] Certaines Applications de !'Analyse Fonctionnelle ala Physique Mathematique, Leningrad 1945.
SOBOLEVSKI, P. E.
[1] Parabolic type equations in Banach spaces. Trudy Moscow Math. 10,
297-350 (1961).
450
Bibliography
STONE, M. H.
[1] Linear Transformations in Hilbert Space and Their Applications to
Analysis. Colloq. Publ. Amer. Math. Soc., 1932.
[2] On one-parameter unitary groups in Hilbert space. Ann. of Math. 33,
643-648 (1932).
SZEGO, G.
[1] Orthogonal Polynomials. Colloq. Pub. Amer. Math. Soc., 1948.
TANABE, H.
[1] Evolution equations of parabolic type. Proc. Japan Acad. 37, 610-613
(1961).
[2] On the equations of evolution in a Banach space. Osaka Math. J. 12,
365--613 (1960).
[3) A class of the equations of evolution in a Banach space. Osaka Math. J.
11, 121-145 (1959).
[4] Remarks on the equations of evolution in a Banach space. Osaka Math.
J. 12, 145--166 (1960).
[5] See KATO-TANABE [8].
TANNAXA, T.
[1] Dualitat der nicht-kommutativen bikompakten Gruppen. Tohoku
Math. J. 53, 1-12 (1938).
TAYLOR, A.
[1] Introduction to Functional Analysis, Wiley 1958.
TITCHMARSH, E. C.
[1] Introduction to the Theory of Fourier Integrals, Oxford 1937.
[2] Eigenfunction Expansion Associated with Second-order Differential
Equations, Vol. I-II, Oxford 1946---1958.
TREJVES, F.
[1] Lectures on Partial Differential Equations with Constant Coefficients.
Notas de Matematica, No. 7, Rio de Janeiro 1961.
TROTTER, H. F.
[1] Approximation of semi-groups of operators. Pacific J. Math. 8, 887-919
(1958).
VILENKIN, N. Y.
[1] See GELFAND-VILENKIN [3].
VISIK, I. M.
[1) See LADYZHENSKAYA-VISIK [1).
VITALI, G.
[1] Sull'integrazioni per serie. Rend. Circ. Mat. di Palermo 23, 137-155
(1907).
WATANABE, J.
[1] On some properties of fractional powers of linear operators. Proc. Japan
Acad. 37, 273-275 (1961).
WECKEN, F. J.
[1] Unitarinvariante selbstadjungierte Operatoren. Math. Ann. 116, 422455 (1939).
WEIL, A.
[1] Sur les fonctions presque periodiques de von Neumann. C. R. Acad.
Sci. Paris 200, 38-40 (1935).
WEYL,H.
[1] The method of orthogonal projection in potential theory. Duke Math. J.
7, 414-444 (1940).
Bibliography
451
[2] "'Ober gewohnliche Differentialgleichungen mit Singularitiiten und die

zugehorigen Entwicklungen willkiirlicher Funktionen. Math. Ann. 68,
220---269 (1910).
[3] See PETER-WEYL [1].
WIENER, N.
[1] See PALEY-WIENER [1].
[2] Tauberian theorems. Ann. of Math. 33, 1-100 (1932).
[3) The Fourier Integral and Certain of Its Applications, Cambridge 1933.
YAMABE, H.
[1] See IT6-YAMABE [2].
YosmA, K.
[1] Lectures on Differential and Integral Equations, Interscience 1960.
[2] Vector lattices and additive set functions. Proc. Imp. Acad. Tokyo 17,
228-232 (1940).
[3] Mean ergodic theorem in Banach spaces. Proc. Imp. Acad. Tokyo 14,
292-294 (1938).
[ 4] Ergodic theorems for pseudo-resolvents. Proc. Japan Acad. 37, 422-425
(1961).
[5] On the differentiability and the representation of one-parameter semigroups of linear operators. J. Math. Soc. Japan 1, 15--21 (1948).
[6] Holomorphic semi-groups in a locally convex linear topological space.
Osaka Math. J., 15, 51-57 (1963).
[7] (with S. KAKUTANI) Operator-theoretical treatment of Markoff process
and mean ergodic theorems. Ann. of Math. 42, 188-228 (1941).
[8] Fractional powers of infinitesimal generators and the analyticity of
the semi-groups generated by them. Proc. Japan Acad. 36, 86-89 (1960).
[9] Quasi-completely continuous linear functional operators. J ap. J. Math.
15, 297-301 (1939).
[10] (withY. MIMURA and S. KAKUTANI) Integral operators with bounded
measurable kernel. Proc. Imp. Acad. Tokyo 14, 359--362 (1938).
[11] On the group embedded in the metrical complete ring. Jap. J. Math. 13,
7-26 (1936).
f12] Normed rings and spectral theorems. Proc. Imp. Acad. Tokyo 19, 356-359 (1943).
[13] On the unitary equivalence in general euclid spaces. Proc. J ap. Acad.
22, 242--245 (1946).
[14] On the duality theorem of non-commutative compact groups. Proc.
Imp. Acad. Tokyo 19, 181-183 (1943).
[15] An abstract treatment of the individual ergodic theorems. Proc. Imp.
Acad. Tokyo 16, 280---284 (1940).
[16] (with M. FUKAMIYA) On vector lattice with a unit, II. Proc. Imp. Acad.
Tokyo 18, 479-482 (1941).
(17] Markoff process with a stable distribution. Proc. Imp. Acad. Tokyo 16,
43-48 (1940).
[18] Ergodic theorems of Birkhoff-Khintchine's type. Jap. J. Math. 17,
31-36 (1940).
[19] Simple Markoff process with a locally compact phase space. Math.
Japonicae 1, 99-103 (1948).
I20] Brownian motion in a homogeneous Riemannian space. Pacific J.
Math. 2, 263--270 (1952).
[21] An abstract analyticity in time for solutions of a diffusion equation.
Proc. Japan Acad. 35, 109-113 (1959).
452
Bibliography
[22] An operator-theoretical integration of the wave equation. J. Math. Soc.

Japan 8, 79-92 (1956).
[23] Semi-group theory and the integration problem of diffusion equations.
Internat. Congress of Math. 1954, held at Amsterdam, Vol. 3, pp. 864-873.
[24] On the integration of diffusion equations in Riemannian spaces. Proc.
Amer. Math. Soc. 3, 864--873 (1952).
[25] On the fundamental solution of the parabolic equations in a Riemannian
space. Proc. Amer. Math. Soc. 3, 864--873 (1952).
[26] An extension of Fokker-Planck's equation. Proc. Japan Acad. 25,
1-3 (1949).
[27] Brownian motion on the surface of 3-sphere. Ann. of Math. Statist. 20,
292-296 (1949).
[28] On the integration of the equation of evolution. J. Fac. Sci. Univ. Tokyo,
Sect. 1, 9, Part 5, 397-402 (1963).
[29] An operator-theoretical treatment of temporally homogeneous Markoff
processes. J. Math. Soc. Japan 1, 244-253 (1949).
YUSHKEVITCH, A. A.
[1] On strong Markoff processes. Teorya Veroyatn. 2 (1957).
Index
absolutely continuous, 17, 375
absorb, 44
absorbing, 24
accumulation point, 3
AcHIESER, N. I., 354, 439
AGMON, S., 438, 439
AKILOV, G., 175, 439
ALEXANDROV, P., 439
almost everywhere, 15
almost periodic function, 218
mean value of, 222
angle variable, 391
Archimedean axiom, 367
ARONSZAJN, N., 96, 439
Ascoli-Arzela theorem, 85
BAIRE, R., function, 18
Baire-Hausdorff theorem, 11
set, 18
theorem, 12
BALAKRISHNAN, V., 260, 439
balanced, 24
BANACH, S., 68, 75, 79, 102, 106, 112,
141, 205, 370, 439
BEBOUTOV, M., 393, 439
BERGMANN, S., 96, 439
BERS, L., 178, 439
Bessel's inequality, 86
BIRKHOFF, G., 364, 439
BIRKHOFF, G. D., 388, 390, 439
BocHNER, S., integral, 132
theorem, 346, 260, 331, 439
BoGOLIOUBov, N., 393, 440
BoHR, H., 331, 440
Boolean algebra, 365
BoREL, E., set, 18
bounded set, 44
BouRBAKI, N., 23, 145, 440
BROWDER, F. E., 440
Brownian motion, 398, 403
carrier(= support), 26
Cartesian product, 3
Cauchy sequence, 11
category, the first, 11
the second, 11
Cayley transform, 202
CHACON, R. V., 385, 389, 440
Chapman-Kolmogoro v equation,
379
CHOQUET, G., 364, 440
CLARKSON, J.A., 127,440
closed graph theorem, 79
closed range theorem, 205
closed set, 3
closure, 3
strong, 125
weak, 125
compact, 4
relatively, 5
complement, 3, 366
complete, 11, 367
orthonormal system, 86
a-complete, 367
weakly, 120
completion, 56
condensation of singularities, 73
conjugate exponent, 33
continuous mapping, 4
convergence, asymptotic, 38
mean, 34
in measure, 38
strong, 32
weak, 120
weak*, 125
convex, 24
countably additive (=a-additive),
15
covering, 4
CRUM, M. M., 166
defect indices, 349
defining function, 17
dense, 11
DiEUDONNE, J., 105, 440
diffusion, 398, 409
454
Index
diffusion, equation, 413, 419

Dirichlet's problem, 286
distribution ( = generalized function), 47
with compact support, 62
direct product of, 65
local structure of a, 100
tempered, 149
tensor product of, 67
distributional derivative, 49
domain, 1
DooB, ] . L., 388, 411, 440
DuFRESNOY, J., 166
DuNFORD, N., integral, 225; 128,
233, 325, 389, 440
DYNKIN, E. B., 398, 408, 440
EBERLEIN, W. F., Eberlein-Shmulyan theorem, 141; 440
Egorov's theorem, 16
EHRENPREIS, L., 183, 441
eigenspace, 209
eigenvalue, 209
eigenvector, 209
elliptic equation, analytically, 193
formally, 193
strongly, 176
equation of evolution, 412
equi-bounded, 85
equi-continuous, 85, 234
ERDELYI, A, 441
ergodic, decomposition, 393
hypothesis, 389
ergodic theorem, of Hille type, 215
individual, 383
mean, 213
essentially bounded, 34
extension, 2
extremal point, 362
FELLER, W., 249, 398, 403, 406, 441
FEYNMAN, R. P., integral, 438
finite intersection property, 6
finitely additive, 71
measure, 118
FOIAS, C., 441
Fourier, coefficient, 86
expansion, 87
integral theorem, 14 7
Fourier-Laplace transform, 161
transform, 146, 151
fractional power of a closed operator,
259
FRECHET, M., 23, 275

space, 52
FREUDENTHAL, H., 317, 364, 441
FRIEDMANN, A., 191, 441
FRIEDRICHS, K., 177,317, 424, 441
FUKAMIYA, M., 360, 375, 441
fundamental solution, 182
fundamental system of neighbourhoods, 28
GARDING, L., inequality, 175, 287, 441
GELFAND, I. M., Gelfand-Mazur
theorem, 128
Gelfand-Raikov theorem, 356
166, 441
general expansion theorem, 323
generalized, derivative, 49
function, 4 7
function with compact support,
62
Laplacian, 403
Leibniz' formula, 51
limit, 103
nilpotent element, 299
spectral resolution, 353
GLAZMAN, I. M., 354, 442
graph, 77
Green's integral theorem, 50
GROTHENDIECK, A., 145, 279, 289,
442
HAHN, H., Hahn-Banach theorem,
102, 105, 106
decomposition, 36
HALMOS, P. R., 119, 393, 409, 442
HAUSDORFF, F., 3, 11, 75, 442
HEAVISIDE, 0., function, 50
operational calculus, 169
HELLINGER, E., 323, 442
HELLY, E., 109, 114, 442
HILBERT, D., Hilbert-Schmidt type
kernel, 277
space, 52
HILLE, E., 215, 231, 232, 249, 255,
295, 398, 442
Hille-Yosida theorem, 248
(The result pertaining to (9) of
Corollary 1 of p. 248 is sometimes
called as the Hille-Yosida theorem)
HOFFMANN, K., 363, 443
Holder's inequality, 33
HOPF, E., 346, 385, 389, 443
Index
HORMANDER, L., 51, 80, 188, 189,
193, 443
H-theorem, 392
HUNT, G. A., 398, 411, 443
hypoellipticity, 80, 189
ideal, 295, 372
maximal, 295
image, 1
(strict) inductive limit, 28
infimum ( = the greatest lower
bount}, 2
infinitesimal generator, 237
integral, 16
Bochner, 132
Dunford, 226
Lebesgue, 19
Radon, 118
invariant measure (with respect a
Markov process}, 379, 393
right invariant Haar measure,
400
inverse, 2
isometric, 113, 202, 203
ITO, K., 398, 408, 443
ITO, S., 419, 443
JACOBS, K., 382, 443
jOHN, F., 443
JoRDAN, C., decomposition, 36, 366
jORDAN, P., 39
KAKUTANI, S., 10, 31, 119, 127, 215,
286, 375, 378, 388, 443
KANTROVITCH, L., 175, 444
KATO, T., 215, 218, 251, 260, 272,
315, 424, 437, 444
KELLEY, J. L., 363, 444
KELLOG, 0. D., 444
KHINTCHINE, A., 388, 444
KISYNSKI, J ., 430, 444
KoDAIRA, K., 325, 444
KoLMOGORov, A., 275, 398, 444
KoMATsu, H., 273, 437, 444
KOTAKE, T., 437, 445
KOTHE, G., 68, 105, 445
KREIN, M., Krein-Milman theorem,
362; 10, 375, 445
KREIN, S., 10, 375, 445
KRYLOV, N., 312, 393, 445
LADYZHENSKAYA, 0. A., 425, 445
Laplace transform, 162, 237
455
lattice, 2
B-, 369
distributive, 365
F-, 369
homomorphism, 372
modular, 365
vector, 364
LEBESGUE, H., Lebesgue-Fatou
lemma, 17
Lebesgue-Nikodym theorem, 93
LERAY, J., 100, 445
LEWY, H., 182, 445
limit point, 3
Lindeberg's condition, 398
linear functional, 21
continuous, 43
linear subspace (or, in short, subspace), 21
linearly, independent, 21
ordered ( = totally ordered), 2
LIONS, J. L., 424, 425, 445
LUMER, G., 250, 445
MAAK, w .. 224, 445
MALGRANGE, B., 182, 193, 446
mapping, 1
MARKOV, A., 119
Markov process, 379
spatially homogeneous, 398
temporally homogeneous, 379
MARUYAMA, G., 398, 446
maximal, 2
ideal, 295
MAZUR, s .. 108, 109, 446
McKEAN, H., 398, 408, 446
measure, 15
Baire, 18
Borel, 18
Lebesgue, 19
regular Borel, 19
topological, 18
measurable, function, 15
set, 15, 20
strongly, 130
weakly, 130
metric, 4
space, 3
metrically transitive, 390
MEYER, P. A., 411, 446
MIKUSU~SKI, J., 166, 169, 446
MILGRAM, A. N., 92, 446
MILMAN, D., 12&, 362, 446
456
Index
MIMURA, Y., 119, 286, 340, 446

MINKOWSKI, H., functional, 24
inequality, 33
MINLOS, R. A., 293, 446
mixing hypothesis, 392
MIYADERA, I., 249, 446
MIZOHATA, S., 419, 446
MoRREY, C. B., 193, 446
multiplicity, of an eigenvalue, 209
of the spectrum at, 321
NAGUMO, M., 231, 295, 447
NAGY, B. Sz.-, 345, 354, 447
NAIMARK, M. A., 295, 325, 352, 360,
447
NAKANO, H., 323, 447
NARASIMHAN, M.S., 437, 447
neighbourhood, 3
NELSON, E., 438, 44 7
NEUMANN, C., series, 69
VON NEUMANN, J., 39, 200, 203, 204,
21~ 22t 31~ 34~ 447
nilpotent, 372
NIRENBERG, L., 178, 193, 438, 44 7
non-dense, 11
norm, 30, 43
of a bounded linear operator, 43
negative, 98, 155
normed, field, 129
linear space, 30
ring, 295
one point compactification, 5
open mapping theorem, 75
open set, 3
operational calculus, 343
Dunford's, 226
Heaviside's, 169
Mikusinski's, 169
operator, adjoint, 196
adjoint differential, 80
bounded linear, 43
closable, 77
closed linear, 77
compact, 274
completely continuous, 274
conjugate, 194
continuous linear, 43
0-, 170
dual, 194
essentially self-adjoint, 350
fractional power of a closed, 259
operator, hypermaximal ( = selfadjoint), 350

identity, 69
inverse, 21
isometric, 202, 203
linear, 21
maximal symmetric, 350
normal, 202
nuclear, 279, 289
pre-closed, 77
projection, 82
real, 316
self-adjoint, 197
symmetric, 197
unitary 202
ORNSTEIN, D. S., 385, 389, 447
orthogonal, 82
base, 86
complement, 82
projection, 82
orthonormal system, 86
PALEY, E. R. A. C., Paley-Wiener
theorem, 161, 162, 163, 448
parabolic equations, 188
PARSEVAL, relation, 86
theorem for the F mrier transform, 154
partial order ( = semi-order), 2
partition of unity, 60
PEETRE, J., 193, 448
Peter-Weyl-Neumann theory, 330
PETROWSKY, I. G., 193, 448
PETTIS, B. J., 131, 448
PHILLIPS, R. S., 224, 231, 250, 260,
272, 295, 419, 424, 448
PITT, H. R., 360, 448
Plancherel's theorem for the Fourier
transform, 153
PoNTRJAGIN, L., 338, 448
PosT, E. L., Post-Widder inversion
formula, 261
projection, 82
pseudo-resolvent, 215
quasi-norm, 31
quasi-normed linear space, 31
quasi-unit, 375
radical, 298, 372
RADO, T., 411, 448
RADON,]., integral, 118
Index
RADON, J., Radon-Nikodym theorem, 377
RAIKOV, D. A., 356, 448
range, 1, 21
rapidly decreasing function, 146
RAY, D., 398, 448
Rayleigh's principle, 321
reflexive, 140
regularization, 29
resolution of the identity, 309
resolvent, 209
equation, 211
set, 209
resonance theorem, 69
restriction, 2
RICKART, C. E., 295, 360, 448
RIEsz, F., Riesz-Markov-Kakutani
theorem, 119
representation theorem, 90
Riesz-Schauder theory, 283; 84,
215, 340, 448
RYLL-NARDZEWSKI, C., 166, 449
SAKS, s .. 70, 449
scalar product, 40
ScHMIDT, E., orthogonalization, 88
SCHWARTZ, J., 325, 389, 448
ScHWARTZ, L., 23, 165, 166, 178, 249,
449
SEGAL, I.E., 360,449
semi-group [of class (C0 )], 232
contraction, 250
convergence of a sequence of semi-groups, 269
differentiability of a, 237
dual, 272
holomorphic, 254
representation of a, 246
semi-norm, 23
semi-reflexive, 139
semi-scalar product, 250
semi-simple, 298
separation (axiom of), 3
SHMULYAN, V. L., 141, 449
SILOV, G., 166, 449
singular, element, 375
measure, 378
slowly increasing function, 150
SMITH, P. A., 390, 449
SoBOLEV, S. L., lemma, 174;
space, 55; 23, 449
SOBOLEVSKI, P. E., 431, 437, 449
space, Banach (or B-), 52
457
space, barrel, 138

bidual, 112
bornologic, 44
discrete topological, 34
dual, 111
eigen, 209
factor, 22, 59
Frckhet (or F-), 52
Hilbert, 52
homogeneous, 401
linear, 20
linear topological, 25
locally convex, 26
locally sequentially weakly compact, 126
metric, 4
normal, 7
normed linear, 30
nuclear, 291
null, 21
pre-Hilbert, 39
product, 6
quasi-normed linear, 31
regular, 7
sequentially complete, 105, 237
Sobolev, 55
topological, 3
uniformly convex, 126
spectral, radius, 211
resolution, 306, 313
spectrum, continuous, 209
point, 209
residual, 209
simple, 322, 34 7
STONE, M. H., theorem, 253, 345
Stone-Weierstrass theorem; 9, 231,
325, 449
subadditive, 23
support ( = carrier), 26
supremum ( = the least upper
bound), 2
SZEGO, G., 90, 450
TANABE, H., 432, 436, 437, 450
TANNAKA, T., duality theorem, 336
TAYLOR, A., 231, 450
TITCHMARSH, E. C., 166, 325, 450
topology, 3
bounded convergence, 110
relative, 3
simple convergence, 110
strong, 111
458
Index
topology, uniform topology of operators, 111

weak, 111
weak*, 111
totally bounded, 13
TREVES, F., 193, 450
TROTTER, H. F., 272, 450
TYCHONOV, A., theorem, 6
uniform boundedness theorem, 68
unitary, equivalence, 323
invariant, 323
operator, 202
URYSOHN, P., theorem, 7
VrLENKIN, N.Y., 293, 450
VrsrK, I. M., 425, 450
Vitali-Hahn-Saks theorem, 70
WATANABE,
J., 266, 450
wave equation, 421

WECKEN, F. ]., 323, 450
WEIERSTRASS, C., polynomial approximation theorem, 8, 249
trigonometric approximation
theorem, 11
WEIL, A., 219, 450
WEYL, H., lemma, 80; 325, 330, 451
WIDDER, D. V., 261
WIENER, N., 161, 165, 357, 451
YAMABE, H., 419, 451
YosmA, K., 90, 95, 178, 215, 231,
249, 259, 260, 286, 295, 325, 338,
370, 375, 378, 383, 384, 392, 393
403, 419, 424, 430, 451
YUSHKEVITCH, A. A., 398, 452
ZoRN, M., lemma, 3

Yosida 1 Functional - Analysis

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Yosida 1 Functional - Analysis

Загружено:

Авторское право:

Доступные форматы

DIE GRUNDLEHREN DER

J.L.DOOB E.HEINZ. F.HIRZEBRUCH

B. ECKMANN UND B.L.VAN DER WAERDEN

SPRINGER-VERLAG BERLIN HEIDELBERG GMBH

SPRINGER-VERLAG BERLIN HEIDELBERG GMBH

Prof. Dr. B. Eckmann

Prof. Dr. B. L. van der Waerden

Springer-Verlag Berlin Heide1berg 1965

Titei Nr. 5106

of the galley proof, correcting errors and improving the presentation.

Semi-norms and Locally Convex Linear Topological Spaces.

TJ. Applications of the Baire-Hausdorff Theorem

IV. The Hahn-Banach Theorems . . . . . . . .

The Hahn-Banach Extension Theorem in Real Linear Spaces

V. Strong Convergence and Weak Convergence

1. The Weak Convergence and The weak* Convergence

VI. Fourier Transform and Differential Equations

The Fourier Transform of Rapidly Decreasing Functions

VII. Dual Operators

VIII. Resolvent and Spectrum. . . .

IX. Analytical Theory of Semi-groups. . . .

The Rellich-Garding Theorem

Appendix to Chapter X. The Nuclear Space of A. GROTHENDIECK 2H!l

Maximal Ideals of a Normed Ring. . .

XII. Other Representation Theorems in Linear Spaces .

Extremal Points. The Krein-Milman Theorem

XIII. Ergodic Theory and Diffusion Theory . . . . . . . .

The Markov Process with an Invariant Measure

1. Integration of Diffusion Equations

ioint if their intersection is void. A family of sets is disjoint if every

of sets is a disjoint family, then the union U Xn may be written in

the form of a sum

Mappings. The term mapping, function and transformation will be

Y1 = j(f-1 (Y1)) for all Y1

Yosida, Functional Analysis

s;; j-1 (t(X1))

x for all xED (f), and /(j-1 (y))

y for ally E R (!).

Then P is said to be partially ordered (or semi-ordered) by the relation

Definition. A partially ordered set Pis said to be linearly ordered (or

Zorn's Lemma. Let P be a non-empty partially ordered set with the

iff is the abbreviation for "if and only if".

(.i (xi <~1

Yi) 2) 112 , where x

Proof. Let there be an accumulation point x0 of a compact set M of a

1, 2, ... , n} which covers M. Then

Gx0' m does not

intersect M. But, since x0 is an accumulation point of M, the open set

GxOl mt 3 Xo must contain a point mE M distinct from Xo This is a

contradiction, and M must be closed.

A consists of integers (1, 2, ... , n), ll Xk is usually

denoted by X 1 xX2 x xXw We introduce a (weak) topology in the

system of compact topological spaces X" is also compact.

of the n-dimensional Euclidean space Rn is compact. From this we see

intersection property. Let {N} be a system of subsets N of X with the

n N~'o any open set G,. o

satisfy the condition of the proposition.

Proof. We assign to each rational number r = k/2n (k = 0, 1, ... , 2n),

= 0 on G(O), and f(t) = supr whenever tE G(Of.

Then, by (i), f(t) = 0 on A and f(t) = 1 on B. We have to prove the

x and multiplying by x, we obtain nx(x

xP yn-p with respect to

Similarly, by differentiating the first expression twice with respect to x and

2c52n_,.. 0 (as n __,. CXJ).