Graduate Texts in Mathematics

68
Editorial Board
F. W. Gehring
P. R. Halmos
Managing Editor

c. C. Moore
 Joachim Weidmann

Linear Operators
in Hilbert Spaces

Translated by
Joseph Szucs

Springer-Verlag
York Heidelberg Berlin
Joachim Weidmann Joseph SzUcs
Mathematisches Seminar der American Mathematical Society
J ohann-Wolfgang-Goethe-Universitat P.O. Box 6248
Institut fur Angewandte Mathematik Providence, RI 02940
Robert-Mayer-Strasse 10 USA
6 Frankfurt a. M.
Federal Republic of Germany

Editorial Board
P. R. HaImos F. W. Gehring c. C. Moore
Managing Editor University of Michigan University of California
Indiana University Department of Mathematics at Berkeley
Department of Mathematics Ann Arbor, Michigan 48104 Department of Mathematics
Bloomington, Indiana 47401 USA Berkeley, California 94720
USA USA

AMS Subject Classifications: 47Axx', 47B05, 47BIO, 47B15, 47B20,

47B25, 47B30, 47E05, 8IA09, 8IAW, 81A45

With I Figure.
Ubrary of Congress Cataloging In PubUcadoD Data
Weidmann, Joachim.
Linear operators in Hilbert spaces.
(Graduate texts in mathematics; 68)
Translation of Lineare Operatoren in Hilbertriiumen.
Bibliography: p.
Includes index.
I. Linear operators. 2. Hilbert space. I. Title. II. Series.
QA329.2.W4413 515'.72 79-12649

Exclusively authorized English translation of the original German edition Lineare

Operatoren in Hilbertriiumen published in the series Mathematische Leitfaden
edited by G. Kothe © by B. G. Teubner, Stuttgart, 1976.

All rights reserved.

No part of this book may be translated or reproduced in any form without written
permission from Springer-Verlag.

Ill> 1980 by Springer-Verlag New York Inc.

9 8 7 6 5 432
ISBN-13: 978-1-4612-6029-5 e-ISBN-13: 978-1-4612-6027-1
DOl: 10.107/978-1-4612-6027-1
To the memory of
Konrad J orgens
 Preface to the English edition

This English edition is almost identical to the German original Lineare

Operatoren in Hilbertriiumen, published by B. G. Teubner, Stuttgart in
1976. A few proofs have been simplified, some additional exercises have
been included, and a small number of new results has been added (e.g.,
Theorem 11.11 and Theorem 11.23). In addition a great number of minor
errors has been corrected.

Frankfurt, January 1980 J. Weidmann

vii
 Preface to the German edition

The purpose of this book is to give an introduction to the theory of linear

operators on Hilbert spaces and then to proceed to the interesting applica-
tions of differential operators to mathematical physics. Besides the usual
introductory courses common to both mathematicians and physicists, only
a fundamental knowledge of complex analysis and of ordinary differential
equations is assumed. The most important results of Lebesgue integration
theory, to the extent that they are used in this book, are compiled with
complete proofs in Appendix A. I hope therefore that students from the
fourth semester on will be able to read this book without major difficulty.
However, it might also be of some interest and use to the teaching and
research mathematician or physicist, since among other things it makes
easily accessible several new results of the spectral theory of differential
operators.
In order to limit the length of the text, I present the results of abstract
functional analysis only insofar as they are significant for this book. I
prove those theorems (for example, the closed graph theorem) that also
hold in more general Banach spaces by Hilbert space methods whenever
this leads to simplification. The typical concepts of Hilbert space theory, .
"orthogonal" and "self-adjoint," stand clearly at the center. The spectral
theorem for self-adjoint operators and its applications are the central
topics of this book. A detailed exposition of the theory of expansions in
terms of generalized eigenfunctions and of the spectral theory of ordinary
differential operators (Weyl-Titchmarsh-Kodaira) was not possible within
the framework of this book.
In the first three chapters pre-Hilbert spaces and Hilbert spaces are
introduced, and their basic geometric and topologic properties are proved.
Chapters 4 and 5 contain the fundamentals of the theory of (not neces-
ix
x Preface to the Gennan edition

sarily bounded) linear operators on Hilbert spaces, including general

spectral theory. Besides the numerous examples scattered throughout the
text, in Chapter 6 certain important classes of linear operators are studied
in detail. Chapter 7 contains the spectral theory of self-adjoint operators
(first for compact operators, and then for the general case), as well as some
important consequences and a detailed characterization of the spectral
points. In Chapter 8 von Neumann's extension theory for symmetric
operators is developed and is applied to, among other things, the Sturm-
Liouville operators. Chapter 9 provides some important results of perturba-
tion theory for self-adjoint operators. Chapter 10 begins with proofs of the
most significant facts about Fourier transforms in L2(R m ), applications to
partial differential operators, in particular to Schrodinger and Dirac opera-
tors, follow. Finally, Chapter 11 gives a short introduction to (time depen-
dent) scattering theory with some typical results; to my regret, I could only
touch upon the far reaching results of recent years.
Exercises are not used later in the text, with a few exceptions. They
mainly serve to deepen understanding of the material and give opportunity
for practice; however, I often use them to formulate further results which I
cannot treat in the text. The level of difficulty of the exercises varies
widely. Because I give many exercises with detailed hints, they can be
solved in general without much difficulty.
Now I want to very heartily thank all those who helped me with the
production of this book. Mrs. Hose turned my notes into an excellent
typed manuscript with infinite diligence. Messrs. R. Hollstein, D. Keirn
and H. Klich spent much time reading the whole manuscript and discuss-
ing with me their suggestions for improvement. Messrs. R. Colgen and W.
Stork helped me with the proofreading. I thank the publisher and the
editors for their pleasant cooperation.
My teacher Konrad Jorgens inspired me to study this subject; he
influenced the present exposition in several ways. I dedicate this volume to
his memory.

Hattersheim am Main, the summer of 1976 Joachim Weidmann

 Contents

Chapter 1
Vector spaces with ~ scalar product, pre-Hilbert spaces
1.1 Sesquilinear forms I
1.2 Scalar products and norms 6

Chapter 2
Hilbert spaces 15
2.1 Convergence and completeness 15
2.2 Topological notions 21

Chapter 3
Orthogonality 29
3.1 The projection theorem 29
3.2 Orthonormal systems and orthonormal bases 34
3.3 Existence of orthonormal bases, dimension of a Hilbert space 42
3.4 Tensor products of Hilbert spaces 47

Chapter 4
Linear operators and their adjoints 50
4.1 Basic notions 50
4.2 Bounded linear operators and functionals 56
4.3 Isomorphisms, completion 63
4.4 Adjoint operator 67
4.5 The theorem of Banach-Steinhaus, strong and weak convergence 74
4.6 Orthogonal projections, isometric and unitary operators 81

xi
xii Contents

Chapter 5
Closed linear operators 88
5.1 Closed and closable operators, the closed graph theorem 88
5.2 The fundamentals of spectral theory 96
5.3 Symmetric and self-adjoint operators 107
5.4 Self-adjoint extensions of symmetric operators 114
5.5 Operators defined by sesquiIinear forms (Friedrichs' extension) 120
5.6 Normal operators 125

Chapter 6
Special classes of linear operators 129
6.1 Finite rank and compact operators 129
6.2 Hilbert-Schmidt operators and Carleman operators 136
6.3 Matrix operators and integral operators 149
6.4 Differential operators on L2(a, b) with constant coefficients 157

Chapter 7
The spectral theory of self-adjoint and normal operators 166
7.1 The spectral theorem for compact operators, the spaces Bp(HI' H~ 166
7.2 Integration with respect to a spectral family 180
7.3 The spectral theorem for self-adjoint operators 191
7.4 Spectra of self-adjoint operators 200
7.5 The spectral theorem for normal operators 210
7.6 One-parameter unitary groups 220

Chapter 8
Self-adjoint extensions of symmetric operators 229
8.1 Defect indices and Cayley transforms 229
8.2 Construction of self-adjoint extensions 237
8.3 Spectra of self-adjoint extensions of a symmetric operator 243
8.4 Second order ordinary differential operators 247
8.5 Analytic vectors and tensor products of self-adjoint operators 259

Chapter 9
Perturbation theory for self-adjoint operators 269
9.1 Relatively bounded perturbations 269
9.2 Relatively compact perturbations and the essential spectrum 273
9.3 Strong resolvent convergence 282

Chapter 10
Differential operators on L2(lRm) 289
10.1 The Fourier transformation on L2(1R"') 289
10.2 Sobolev spaces and differential operators on LilR"')
with constant coefficients 296
10.3 Relatively bounded and relatively compact perturbations 304
Contents xiii

10.4 Essentially self-adjoint Schrodinger operators 313

10.5 Spectra of Schrodinger operators 3.23
10.6 Dirac operators 329

Chapter 11
Scattering theory 337
11.1 Wave operators 337
11.2 The existence and completeness of wave operators 343
11.3 Applications to differential operators on L2(R m ) 354

Appendix A
Lebesgue integration 362
A.IDefinition of the integral 362
A2 Limit theorems 368
A3 Measurable functions and sets 370
A.4The Fubini-Tonelli theorem 374
A5 The Radon-Nikodym theorem 377

Appendix B
A representation theorem for holomorphic
functions with values in a half-plane 381

References 387

Index of symbols 390

Author and subject index 392

Vector spaces with a scalar product,
pre-Hilbert spaces
1

In what follows we consider vector spaces over a field 11(, where II( is either
the field C of complex numbers or the fieldlR of real numbers; accord-
ingly, we speak of a complex or a real vector space. For every cEil( let c*
be the complex conjugate of c; so for c E IR the star has no significance.
As a rule, we assume the most important notions and results of linear
algebra to be known.

1.1 Sesquilinear forms

Let H be a vector space over II<. A mapping s : H X H~ IK is called a
sesquilinear form on H if for all f, g, h E H and a, bEll( we have

s(j, ag + bh) = as(j, g) + bs(j, h), (Ll)

s(af+ bg, h) = a*s(j, h) + b*s(g, h). (1.2)

If (1.2) holds without stars, then s is called a bilinear form on H; in

particular every sesquilinear form on a real vector space is a bilinear form.
Property (1.1) is obviously equivalent to the two properties
sU, g+ h) = sU, g) + sU, h), ( 1.1')
sU, ag) = as(j, g). (1.1")
Similarly, (1.2) is equivalent to
sU+ g, h) = sU, h) + s(g, h), (1.2')
s(af, g) = a*sU, g). (1.2")
2 1 Vector spaces with a scalar product, pre-Hilbert spaces

If s is a sesquilinear form on H, then the mapping q : H-+ II( that is defined

by q(j) .. s(j, f) for each f E H is called the quadratic form on H generated
or induced by s. For each quadratic form q we obviously have

q(aj) = laI 2q(J) for all f E H, a E K; (1.3)

so we have, in particular, q(af) = q(j) for every a E K with lal- 1.
The following theorem shows that in a complex vector space the generat-
ing sesquilinear form is uniquely determined by the quadratic form; for
real vector spaces this is not necessarily true in general; see Exercise 1.2.

Theorem 1.1 (Polarization identity). Let H be a complex vector space, s a

sesquilinear form on H, and q the quadratic form generated by s. Then for a/l
f, g E H we have
s(J, g) - Hq(J+ g) - q(J - g) + iq(J - ig) - iq(J + ig)}. (1.4)

The proof of this identity may be given by calculating the right side of
(1.4) according to the rules (1.1) and (1.2).

Theorem 1.2 (Parallelogram law). Let s be a sesquilinear form on a vector

space H, and let q be the corresponding quadratic form on H. Then for all
f,gEHwe have
q(J+g) + q(J-g) = 2[ q(J)+q(g»). (1.5)

PROOF. For every f, g E H we have

q(J+ g) + q(J - g) = s(J,j) + s(J, g) + s( g,j) + s( g,g)
+ s (j,j) - s (j, g)- s(g,f)+ s (g, g)
- 2q(j)+ 2q(g). 0
A sesquilinear form s on H is said to be Hermitian provided that for every
f, gE H we have
s(J, g) = s(g,j)·. (1.6)
A Hermitian bilinear form on a real vector space is said to be symmetric.
If s is a Hermitian sesquilinear form, and q the quadratic form generated
by s, then we obviously have q(j) E R for allf E H; we say briefly that q is
real. The following theorem shows, among other things, that Hermitian
sesquilinear forms can be characteri}ed by this property of their associated
quadratic forms. We also obtain that symmetric bilinear forms are
uniquely determined by the corresponding quadratic forms.

Theorem 1.3. Let H be a vector space over 11<, s a sesquilinear form on H, and
q the quadratic form generated by s.
1.1 Sesquilinear forms 3

(a) If II( = C, then the following statements are equivalent:

(i) s is symmetric,
(ii) q is real,
(iii) for all f, g E H we have

Res(j,g) - Hq(j+g)-q(j-g)}, (1.7)

(iv) for all f, g E H we have

1m s(j, g) ... Hq(j - ig) - q(j + ig)}. (l.7')

(b) If II( - A, then the following statements are equivalent:
(i) s is symmetric,
(ii) for all f, g E H we have
s(j, g) = Hq(j+ g) - q(j- g)}. (1.8)

PROOF.
(a) (ii) follows from (i): q(f). - s(f, 1). - s(f, 1) = q(f), i.e., q(f) is real.
(iii) follows from (ii): Because q(h) E R for all hE H, it follows from
(1.4) that

Re s(j, g) ... *Re{ q(j+ g) - q(f- g) + iq(j- ig) - iq(j + ig)}

... Hq(j+g)-q(j-g)}·

(iv) follows from (iii): Because of (iii) we have

1m s(j, g) = - Re{ is(j, g)}
... Res(j, -ig) - Hq(j-ig)-q(j+ig)}.
(i) follows from (iv):
s(g,j)· ... Re s(g,f) - i 1m s(g,j) ... 1m s(g, if) - i 1m o$(g, f)
= t {q( g + j) - q( g - f) - iq( g - if) + iq( g + if) }
... Hq(j+ g) - q(j- g) + iq(j- ig) - iq(j+ ig)} - s(j, g);
here we have used (1.3) with a" - I, a" i, and a'" - i.
(b) (ii) follows from (i) by calculating the right side of (ii) while using the
symmetry of s.
(i) follows from (ii):

s(g,1) - Hq(g+f)-q(g-f)}
== Hq(j+g)-q(j-g)} ... s(j,g). 0

A Hermitian sesquilinear form is said to be non~negative when

s(j,f) > 0 for all f E H; (1.9)
 VeclOl: spaces with scalar producl, pre-Hilbert spaces

it is said to be positive when

J) 0 for all f with =I' 10)
Since have s(O, = every positive sesquilinear form non-negative.
We also say that the corresponding quadratic forms are non-negative,
respectively (Because of Theorem 1.3, the word "Hermitian" may
be omitted from this definition in the complex case; this does not hold in
real case, cf. Exercise 1.3.)

Theorem 1,4, If s is a non-negative sesquilinear form on H, and q denotes the

l?entera'ted by s, then lor every E we have Schwarz
inequality

(1.11)
If s is positive, then the equality sign in (1. I) holds if and and are
linearly dependent; the equality s(f, g) = [q(f)q(g)]1/2 holds if andonly if
there exists a c ;;;. such that 1= cg or g cJ.
PROOF. Let gE For all R have
o " q(j + tg) = q(j) + 2t Re (j, g) + t 2q( g).
This second degree polynomial in t has either no root or a double root.
Since this holds a polynomial + + and only if b 2 ac 0, it
follows that

[Re s(j, g) Y" q(j)q( g). (U2)

one chooses a II< that lal I and as(f, g) Is(f, holds, then it
follows from (1.12) with h = ag that
Is(j, g)1 2 = [Re as(j, g)]2 = [Re s(j, h)]2
q(j)q(h) q(j)q(ag) = g);
this is the Schwarz inequality.
Let now be positive and let s(f,g} [q(f)q(g)]1/2 true. If =0,
then the equality g = Of proves the assertion. Consequently, let g =I' O.
Because of equality g)f q(f)q(g) 0, the polynomial consid-
ered above has a double root to; hence we have q(f+ tog)=O i.e., f=
tog. From - toS{g, g) = g) it follows that - to O. we have
Is(f,g)I=[q(f)q(g)]l/2 and choose a and h as above, then s(f,h)=
[q(f)q(h)]'/2 foHows. According to the part proved then have
either g = 0 = Of, or there exists a c > 0 such that f = ch = acg. In both cases
f g are linearly One can verify the converses of the last
two assertions by simple calculation. 0
EXAMPLE 1. For each mE I\J (I\J denotes the set {I, 2, 3, ... } of positive
integers) let be the complex vector space of the
1.1 Sesquilinear forms 5

(f1,f2'" . ,fm), g=(gl' g2" .. , gm)' ... of complex numbers with the
addition
f +g = UI + gl,j2 + g2' ... ,jm + gm)
and multiplication by a E e
af === (afl' af2' ... , afm)'
If (ajk)j, k-I"", m is a complex m x m matrix, then
m
sU, g) = ~ ajJJgk for j, g E em
j,k-I

defines a sesquiIinear form on em. s is Hermitian if and only if the matrix

= I, 2, ... , m we have fl.jk = a~. s is
(ajk ) is Hermitian, i.e., if for every j, k
non-negative (positive) if, for example, (fl.jk) is a diagonal matrix with
non-negative (positive) entries in the diagonal. An important special case
of a positive sesquilinear form on em occurs when (ajk) is the unit matrix.
Then
m
sU, g) = ~ f/8j.
j-I

EXAMPLE 2. On the real vector space IIl m (symmetric, non-negative, posi-

tive) bilinear forms can be given accordingly.

EXAMPLE 3. Let C[O, I] be the complex vector space of complex-valued

continuous functions defined on [0, I] with the addition
(j+g)(x) = f(x) + g(x)
and multiplication by a E e
(aj)(x) = af(x).
If r : [0, I]~e is continuous, then by
sU, g) = fo1f(x)*g(x)r(x) dx f, g E C[O, 1]

°
a sesquilinear form is defined on CrO, I]. It is Hermitian if and only if r is
real-valued; it is non-negative if and only if r(x) :> for all x E [0, 'I]; it is
°
positive if and only if r(x):> for all x E [0, 1] and r does not vanish
identically on any non-trivial interval.

EXAMPLE 4. Let CR[O, I] be the real vector space of real-valued continuous

functions defined on [0, I]. For each continuous function r : [0, 1]~1Il the
bilinear form

s(j, g) = fo1f(x)g(x)r(x) dx f, g E C[O, 1]

6 I Vector spaces with a scalar product, pre-Hilbert spaces

is symmetric. Concerning non-negativity and positivity the same assertions

hold as in Example 3.

EXAMPLE 5. If k:[O, 1] X [0, 1]-).C is continuous, then by

sU, g) = L1L1k(x,Y)J(x)*g(y) dy dx
o 0

a sesquilinear form is defined on C[O, 1]. This is Hermitian if and only if

the kernel k is Hermitian, i.e., if for every x,y E[O, 1] we have k(x,y)=
k(y, x)*.

EXERCISES
1.1. Prove the assertions given in Examples 1-5.
1.2. The matrix

generates a non-zero sesquilinear form on 1Il2 (cf. Example 2), the quadratic
form of which vanishes. Consequently, in a real vector space sesquilinear
forms are not determined uniquely by the corresponding quadratic forms.
1.3. Let s be the sesquilinear form on R2 generated by the matrix

(b ~).
If lal<2 (lal<2), then we have s(f,f) >0 for all fER2 such that f=l'O
(s(f, f) ;;. 0 for all f E 1Il2). If a =F 0, then s is not symmetric.
1.4. Let s be a non-negative sesquilinear form on H, q the quadratic form gener-
ated by s, and N- {fE H: q(f)=O}. Show that
(a) N is a subspace (sub-vectorspace) of H.
(b) If fEN and g E H, then we have s(f, g) = 0 and q(f + g) ~ q(g).
(c) In the Schwarz inequality the equality sign holds if and only if f and g are
linearly dependent modulo N, i.e., if there are numbers a, b E Knot
vanishing simultaneously and such that af + bg E N.
(d) We have s(f, g) = [q(f)q(g»)1/2 if and only if there is a c ;;. 0 such that
f - cg E N or g - cf E N.
1.5. Prove the Cauchy inequality

Ij~lllg{ < j~II.fj12j~l 1gjl2

with the aid of Example 1 and the Schwarz inequality.

1.2 Scalar products and norms

A positive sesquilinear form on H is called a scalar product (or inner
product) on H. In what follows scalar products will be denoted mostly by
<. , .)
and occasionally they will be given an index in order to distinguish
1.2 Scalar products and norms 7

between them. A non-negative sesquilinear form is called a semi-scalar

product. Examples for (semi-) scalar products may be obtained from the
exercises in Section 1.1.
The mapping s : H X H~ II< is a scalar product if and only if for all
f, g, hE H and a E II< we have
(i) s(j, g + h) = s(j, g) + s(j, h),
(ii) s(j, ag) = as(j, g),
(iii) s(j, g)= s(g,f)*, (1.13)
(iv) s(j, f) > 0,
(v) s(j, f) > 0 if f:f= O.
For the proof we only have to observe that the properties (1.1) and (1.2)
follow from (i), (ii) and (iii). Similarly, a mapping s : H X H~II< is a
semi-scalar product if and only if s satisfies properties [(1.13) (i-iv)].
A mapping p : H ~ R is called a norm on H if for all f, g E H and a E II<
we have
(i) p(j) > 0,
(ii) p(af) = lalp(j), (1.14)
(iii) p(j + g) <. p(j) + peg) (triangle inequality),
(iv) p(j) > 0 provided f:f= O.
A mappingp : H~R is called a seminorm on H if it satisfies the properties
[(1.l4) (i-iii)]. In what follows norms will mostly be denoted by II . II and
for more precise distinctions they will occasionally be given different
indices.

REMARK. If P is a seminorm on H, then for all f, g E H we have

p(J± g) ;;;. Ip(J) - p(g)l·

PROOF. The triangle inequality implies

p(J) = p(J- g+ g) <. p(J - g) + peg),
thus
p(J) - peg) <. p(J - g).
Similarly, peg) = peg - f + f) <. peg - f) + p(j); thus
- (p(J) - p( g» <. p(J - g).
From these two inequalities p(j - g) ;;;. Ip(j) - p(g)1 follows. One can show
the inequality p(j + g);;;' Ip(j) - p(g)1 in a similar way. D

EXAMPLE 1. In em (or Rm) let us define two norms by

m
IIflll = ~ I.tl and IIflloo = max{Ltl : j = 1, ... , m}.
j-I

If cj > 0 for j = 1, 2, ... , m, then by

m
p,(J) = ~ cJtl and Poo(J) = max{ cJtl : j = 1, ... , m}
j-I
8 I Vector spaces with a scalar product, pre-Hilbert spaces

two seminorms are defined. These seminorms are norms if all the cJ are
positive.

EXAMPLE 2. If r is a non-negative continuous function on [0, 11, then by

PI(j) = loJ r(x)lf(x)1 dx

and
Poo(j) = max(r(x)lJ(x)l: O~x ~ l}
two seminorms are defined on e[O, I]. These are norms if r does not
vanish identically on any non-trivial interval. For r(x) = I these norms will
be denoted by II . III and II . 1100' respectively:

IIflll = follf(x)1 dx,

IIfll oo = max{lf(x)1 : 0 or>;; x or>;; 1}.

A large number of norms can be generated with the aid of scalar products
because of the following theorem.

Theorem 1.5. If s is a semi-scalar product on H, then p(j) = [s(1, !)1 1/ 2

defines a seminorm on H.
<. , .)
If is a scalar product on H, then IIfll = <f,f)1/2 defines a norm on
/-t.
PROOF. Property [(l.l4) (i)1 follows immediately from [(l.l3) (iv»); [(1.14)
(iv)1 follows from [(1.13) (v)]. It is sufficient to prove the remaining
properties for the first case. Because of [(1.13) (ii)1 and [(1.13) (iii)] we have

pea!) = [s(af, aj) r /2 = [laI2s(j,j) r/2 = lal[ s(j,j) r

/2 = lalp(j),

which is [(1.l4) (ii)]. With the aid of the Schwarz inequality it follows that

p(f+ gi = p(f)2 + 2 Re s(j, g) + p(g)2

<; p(j)2 + 2Is(j, g)1 + p(g)2 ~ p(j)2 + 2p(j)p(g) + p(g)2
= (p(j)+p(g»2,
which is the t-riangle inequality [(1.14) (iii)1. o
From the Schwarz inequality for non-negative sesquilinear forms we
obtain for the norm II . II (seminorm p) induced by a scalar product (. , .)
(semi-scalar product s) that
I<f, g)1 ~ IIflili gil, (U5)
Is(j, g)1 ~ p(j)p( g). (l.lS')
1.2 Scalar products and norms 9

Proposition. If (. , .) is a scalar product on H and II . II denotes the norm

generated by it (cf. Theorem 1.5), then IIf+ gil = I/fl/ + 1/ gl/ if and only if
there exists an a ;> 0 such that f == ag or g = af.
PROOF. If f= ag with a;> 0, then we have
IIf+ gil = 11(1 + a)gll = (1 +a)lIgl1 lIagll + II gil = IIfl/ + I/gl/
=
(this part of the assertion holds for any norm). Conversely, if IIf + gil = IIfl/
+ II gil then
+ 211fllllgll + IIg11 2 = IIf+ gW = IIfll2 + 2 Re <1, g) + IIg11 2,
IIfll2
thus Re (f, g) == IIfllll gil· Using (1.15) this implies <J. g) = /lflill gil. Now
Theorem 1.4 gives the assertion. 0
For a norm 1/ • II (seminormp) induced by a scalar product (semi-scalar
product) the parallelogram identity
(1.16)
respectively
p(j + g)2 + p(j _ g)2 == 2(p(j)2 + p( g)2). (1.16')
follows from Theorem 1.2.
If one considers a (semi-)norm as the length of a vector, then these
equalities have the following geometric meaning: In a parallelogram the
sum of the squares of the diagonals equals the sum of the squares of the
sides. According to (1.4) [respectively (1.8)J the scalar product (.,.)
(respectively semi-scalar product s) which we started with is given by the
polarization identity

= {Hllf+ qIl2_lIf- g1l2+ illf- igll 2 - illf+ igIl 2}, II< =C,
<f,g) Hllf+gIl2-IIf-gIl2}, IK==R,

(1.17)
respectively
~ {p(j + g)2 _ p(j _ gl + ip(j - ;g)2 - ip(j + ig)2}, K == C,
s(j, g) == { I{ 2 2
4 p(j+g) -p(j-g) },IK=R.
(1.17')
The following theorem enables us to decide if a given (semi-)norm is
generated by a (semi-)scalar product.

Theorem 1.6 (Jordan and von Neumann). A norm 1/ • II on a vector space H

is generated by a scalar product (. , .) in the Sense of Theorem 1.5 if and only
if the parallelogram identity (1.16) is satisfied. If this is so then the scalar
product (. , .) is given by (1.17). A corresponding statement holds true for
seminorms and semi-scalar products.
10 I Vector spaces with a scalar product, pre-Hilbert spaces

norm II . II is induced scalar product (. ,

the scalar recaptured from
It remains to that if II . II satisfies
gram and (. , .) is then (. , .) is
and generates the norm II . II. We restrict ourselves to the proof in the
complex case; the real case goes analogously and is even a little simpler.
Let (. , .) be defined by (1.17). We show that (. , .) is a scalar product.
[(1.13) (iv-v)]: For all I E H by virtue of the definition of (. , .) we have
(f, f) - ~ {I/f + flI2 _11011 2+ illf - ifll2 - il/f + if1l2}
= ~ {411/112 2illfll2} ... 11/11 1,
[(1.13) (iv-v)] follow from the corlresplomUng
the norm II . II. time we obtain is
, .).
[(1.13) (iii)]: For all I, g E H we have
(g,/)* = Hllg+ 111 2-l/g- 1112+ ;I/g- i/1l2- il/g+ ifIl2}.
= ~ {III + gll2 -1/1 - gl/2 - illl + igll 2 + illf - igll 2}
= (f,·g)·
For allf, g, hE (1.16) we have
h)
{III + gll2 -III - igl1 2- illl + igW;:
+11/+ h1l2-1I1 - hl1 2 + illl - ihl12 - ill/+ ih112}
(I (!
_ i {II + g; h ) + g; h 112 + II + g; h ) _ g; h 112
-11(/- g; h) + g; h 112-II(f- g; h) _.8;h W

+ill(!_i 8 ;h ill(!-ig;l!
ill(/+ig;h ill(!+ig;~
-=.! {1I/+ g+ h 112+ II g-h 112-11/- g+h 112_11 g- h 112
22222
+ ill/- i g + h 112+ ill g- h 11 2-illf+i g + ~1I2_ ill g- h 112}
2 2 2 2
g+h
= 2<1, -2-)' (1.18)

Since by (1.17) have <1,0) .18) it follows

substituting h
2(f, !g) = <I, (1.19)
1.2 Scalar products and nonns 11

From (US) and (1.l9) it follows that

<f,g)+<f,h)=2{f, g;h) =<f,g+h),

which is the required property.
[(1.13) (ii»): We already know that (f, g) -2(f, g/2). From this and
from property [(1.13) (i)] we obtain by induction that
2- nm(f, g) = <J, 2- llmg) for all n, m E "'0
("'0 is the set of non-negative integers {O, 1, 2, ... }). If a ;> 0, then there
exist numbers ak =2- n(k)m(k) such that ak--'loa as k--'looo. By the proposi-
tion preceding Example 1 we have
IlIf± akgll-lIf± aglll <;; lak - alII gil,
IIf± iakgll-lIf± iaglll <;; lak - alllgll,
therefore because of (1.17)
(f, akg) --'10 <f, ag) as k --'10 00.

From this it follows that

Furthermore, we have
<f, -g) -= Hllf-gIl 2-lIf+gll 2+ illf+;gIl2-lIf-igI1 2 }
=-(f,g);
consequently (f, ag) = a(f, g) for all a E R. As we also have
(f, ig) = ~ {lif + igll 2 -lif - igll 2 + illf + gll2 - illf - g1l2}
= ;(f, g).
The equality (f. ag) = a<f. g) follows for all a E e. The proof for semi-
norms is completely analogous. 0

If H is a (complex or real) vector space and <. , .)

is a scalar product on H,
then we call the pair (H, (. , .»
a vector space with scalar product or a
pre-Hilbert space. If it is clear which scalar product is meant on H, then we
shall briefly write H for the pair mentioned. If II . II is a norm on H. then
we call the pair (H, II . II) a normed space. Here we shall also only write H
in most cases. By Theorem 1.5 the norm 11111 =(f. f>I/2 is defined in a
natural way on every pre-Hilbert space. Therefore in what follows we shall
consider every pre-Hilbert space as a normed space.

EXAMPLE 3. On em respectively Rm by
m
(f, g) = L Jjgj
j-I
12 I Vector spaces with a scalar product, pre-Hilbert spaces

a scalar product is defined. The corresponding norm

m } 1/2
IIfll =
{
j~1 1/;1 2

is the Euclidean length of the vector f, thus II f - gil is the Euclidean distance
of the points f and g.

EXAMPLE 4. On C[O, 1] by

a scalar product and the corresponding norm are defined.

EXAMPLE 5. Let 12 be the Hilbert sequence space, i.e., the set of (real or
complex) sequencesf= Un) = UI,j2' ... ) for which }";~_llfnI2 < co. Then 12
will be a (real or complex) vector space if one defines addition and
multiplication as follows:

f + g = Un + gn)' af = (aj,,) for f, g E 12 and a E II<.

It is clear that this definition of multiplication is meaningful since along

with }";~_ t!.t,,1 2 < co we also have }";~_llafnI2 < co. If f and g are in ' 2 , then
for every N EN we have

consequently we also have

n~llfn + gnl 2 ~ 2L~IIj,,12+ ~llgnI2} < co,

i.e., f +g E 12 , It is easy to see that by

00

<1, g) = ~ .t:gn' f, g E 12
n-I
a scalar product is defined on '2; the series converges, because I.t:gnl <;
(lfnl2 + IgnI 2)/2. The induced norm is

Unless otherwise stated, in what follows '2 will always denote the complex
sequence space.
1.2 Scalar products and norms 13

EXERCISES

1.6. The norms in Examples 1 and 2 are not generated by scalar products.
1.7. The proposition after Theorem 1.5 does not hold true in general for norms
that are not generated by scalar products.
1.8. Let p be a seminorm on H generated by a semi-scalar product and let
N-{fEH :p(/)=O}. We havep(f+g)=p(f)+p(g) if and only if there
exists an a ;;. 0 such that f - ag E N or g - af E N.
1.9. (a) Let A2 be the set of functions f holomorphic on C 1 = {z E C : Izl < I} for
which
llf(x + iy)1 2 dx dy < 00
c,
(the integral can be understood as an improper Riemann integral or as a
Lebesgue integral). A2 is a vector space. By

<f, g)1 - 1c,f(x + (y)*g(x + iy) dx dy, IIflll = llf(x

c,
+ (y)12 dx dy
a scalar product and the corresponding norm are defined on A2.
(b) Let H2 be the set of functions f holomorphic on C 1 for which the limit
lim (2"'IJ(re il Wdt
,-+1 )0

is finite. H2 is a vector space (Hardy-class). By

IIfl12 = {lim 1
,-+1 0
2", .
lJ(re")12 dt
}1/2

a norm is defined on H2. This norm is generated by the scalar product

<f, g)2 = Jim
,-+1
1
0
2",.
f(re'/)*g(re ') dt.
I

(c) If f(z) = 'i.':_ofnzn, g(z) = 'i.':_ognzri are the Taylor series of f and g, then
we have
00 I 00
<1, g)1 "" 7T L n + I.t:gn' <f, g)2 = 27T L .t:gn·

.; ; i f
n-O n-O

(d) H2 is a subspace of A2 and we have "f"~ 11f"~ for E H2.

(e) For allfE H2 we have

IIfll~ = sup {fo 2"'lf(re il )1 2 dt : 0.;;; r < t}.

1.10. Let A be an arbitrary set, let,.,. : A~(O, (0), and let /z(A ; ,.,.) be the set of
functionsf: A~C that vanish outside a countable set (that may vary with/)
and for which 'i.aEA,.,.(a)lf(aW < 00.
(a) (/2A ; ,.,.) is a subspace of the space of all complex valued functions on A.
(b) By

<f, g) = L ,.,.(a)f(a)*g(a), f, g E /2(A ; }!),

aEA

a scalar product is defined on /2(A ; ,.,.).

14 I Vector spaces with a scalar product, pre-Hilbert spaces

1.11. Let A be an arbitrary set; for each IX E A let (Ha' <. , .)a) be a pre-Hilbert
space. Then

fa ~ 0 for at most countably many IX E A, L IIfall! < oo}

aEA

is a vector space (with componentwise addition and multiplication). By

«fa), (ga» = L <fa' ga)a, (fa), (ga) E H,

aEA

a scalar product is defined on H, i.e., (H, (., .» is a pre-Hilbert space.

 2
Hilbert spaces

2.1 Convergence and completeness

Let (H, II . II) be a normed space. A sequence (f,,) in H is said to be
convergent if there exists an f E H such that II f" - fll ~O as n ~ 00. There
exists at most one f E H with "!,, - fll ~O; since from II f" - fll ~O and
II!" - gll~O it follows that Ilf - gil..;; IIf- f,,11 + II!" - gll~O, thusf= g. We
say that the sequence (f,,) tends to f and call f the limit of the sequence U,,).
In symbols we write f= lim"->oof,, or !"~f as n~oo. If no confusion is
possible, we shall occasionally abbreviate these by writing f= lim!", or
f"~l·

Proposition.
(a) From !"~1 it follows that IIl"II~IIfll; the sequence (IIf,,11) is bounded.
(b) If (H, <. , .»
is a pre-Hi/bert space, then we also have that f,,~f and
gn~g imply <1", g,,)~<l, g).

PROOF.
(a) By the proposition preceding Example I of Section 1.2, we have
'"f" II - IIfll''';; II!" - fll; from this the assertion follows.
(b) We have 1<1", g,,) - <1, g)l..;; ,<1", g") - <f", g)1 + 1<1", g) - <1, g)l..;;
II!" II II g" - gil + II!" -/11 II gil ~O, since the sequence (II f" II> is bounded
on account of (a). 0

A sequence (!,,) in H is called a Cauchy. sequence if for each f > 0 there

exists an no EN such that for n, m > no we have Ill" - 1m II ";;f. In what
follows, we shall briefly write for this II!" - lmll~O as n, m~oo. Every
convergent sequence is a Cauchy sequence: if 1 is the limit of the sequence
Un)' then II!" - 1m!! <. !!f" - fll + 111 - lmll~O as n, m~oo. Conversely, in an
15
16 2 Hilbert spaces

arbitrary normed space (or pre-Hilbert space) not every Cauchy sequence
is convergent, as Example 1 below shows.

Proposition.
(a) If (J,,) is a Cauchy sequence, then the sequnce (IIJ"ID is convergent (thus
it is bounded).
(b) If (H, <. , .» is a pre-Hilbert space and (J,,), (gn) are Cauchy sequences,
then the sequence «J", gn» is convergent.
PROOF.
(a) As 11IJ,,1I-lIfmlllo;; IIJ" - fmll, the sequence GlfnlD is a Cauchy sequence
in Ill, thus it is convergent and bounded.
(b) By (a) there exists a c > 0 such that IIfnll" c and II gmll "C for all
n, mEN. Since
I<fn, gn> - <fm, gm>I" I<fn' gn - gm>1 + I<fn - f m, gm>1
< c(1I gn - gmll + IIfn - fml!),
the sequence «J", gn» is also a Cauchy sequence. o
EXAMPLE 1. Let (C[O, I], <. , .»
be the pre-Hilbert space introduced in
Section 1.2, Example 4. We show that not every Cauchy sequence is
convergent. For this let the sequence (J,,) in C[O, IJ be defined in the
following way: fl(x) = 1 for all x E [0, 1], and
1
for 0 <X" '2
I 1 1.
f or -<X<-+--
fn(x) = l-(x-I)n
2 2 n'
o for I+n"x"
I 1
1,
for n = 2, 3, .... This sequence is a Cauchy sequence, since for 2 "n .;;;; m
we have

Ilfn - fmll 2 "jl/2+1/n lJ,,(x) - fm(xW dx " !.

1/2 n
To prove that the sequence (J,,) is not convergent let us assume that there
exists an f E C[O, 1] such that J" ~ f, i.e., "fn - fll ~O. Then for 2 "n
we have
"m
1o1/ If(x) -
2 fl(1/2)+(I/n) If(x)j2 dx
112 dx +

= 1 2If(x) - fm(xW dx + II
1
/ If(x)- fm(xW dx
o (1/2)+(I/n)
.;;;; !ailf(x) - fm(xW dx.
2.1 Convergence and completeness 17

Since the right-hand side tends to 0 as m-"oo, we have

LoI/ 21J(X) - Wdx + fl

(t/2)+(I/n)
lJ(xW dx = 0 for n' > 2.
Since f is continuous, it follows from this that
f (x) = 1 for x E [0, t J,
f(x) = 0 for x E(4, 1].
However, this contradicts the continuity of f. Therefore the sequence Un)
cannot be convergent in C[O, 1].
A normed space (H, II . II) is said to be complete if every Cauchy
sequence is convergent. A complete normed space is called a Banach
space; a complete pre-Hilbert space is called a Hilbert space.

EXAMPLE 2. The space C[O, 1] becomes a Banach space with the norm
(cf. Section 1.2, Example 2)
Ilflloo = max {If(x)1 : O';;;x';;; l}.
(By Exercise 1.6 it is not a Hilbert space.) For suppose Un) is a Cauchy
sequence, i.e., assume that for every € > 0 there exists an no EN such that
for all n, m >no and for all x E [0, 1] we have U;,(x) - fm(x)l .;;; €. Then
Un(x» is convergent for every xE[O, 1]; letf(x)=limfn(x). First we show
that this f is continuous. For t: > 0 let no be chosen as above and for this no
let 8 > 0 be chosen so that for IXI - X2! .;;; 8 we have !/"/x I) - fno(x2)1 .;;; €.
From this it follows for IXI - x 2 1 .;;; 8 that
If(xI) - f(X2)1 .;;; If(x l ) - fno(xl)1 + I/"o(x l ) - fn o(x 2)! + !fno(X2) - f(X2)!
= lim Ifm(x l )
m-i>OO
- fn (xl)1 + Ifn (Xl) - fn (X 2 )!
0 0 0

This proves the continuity of f. Now we show thatfn-"f. For t: >0 let no be
chosen again as above. Then for n > no we have
II/" - fll oo = max {lfn{x) - f{x)1 : O';;;x';;; I}
= max { lim Ifn{x) - fm{x)1 : O';;;x';;;
m~oo
I} .;;; t:,
consequently /" -"f.

EXAMPLE 3. em and Rm are Banach spaces with the norms II . III' II . 1100
and II . " from Section 1.2, Examples 1 and 3. This follows easily from the
fact that a sequence Un) is a Cauchy sequence (convergent sequence) in em
or Rm if and only if it converges componentwise. (The proof can also be
obtained as a special case of Example 4.) em (respectively Rm) is therefore
18 2 Hilbert spaces

a Hilbert space with the scalar product

m
<f. g) = "'2. f!8j.
)-1

EXAMPLE 4. Let the scalar product and the norm in 12 be defined by

~ {~ }1/2
<f. g) == n-1
~ 1:g". IIfll = ~
11-1
f: .
as in Section 1.2. Example 5. We show that 12 is complete. therefore it is a
Hilbert space. Let (f (n» be a Cauchy sequence. f (,,) = (it. n' A n' A n' ... ).
As
I~. n -~. ml oe;; Ilf (n) - f (m)ll,
the sequence (~. n)neN is a Cauchy sequence for eachj EN, i.e., there are
numbers ~ E C such that ~. n ~~ as n ~ 00. It remains to prove that
f=(~) E /2 andf (n)~f as n~oo. For t: >0 let noE N be chosen so that for
n. m ;;'no we have IIf(n)- f(m)1I <t:o Then for all kEN we have
k k
~ I~. n - ~12 = 2~~ ~ I~, n - ~. ml 2 ..; lim sup IIf (n) - f (m)1I 2 < E2,
)-1 )-1 m-->~

therefore also
~

"'2. Iii." - ~12 oe;; t: 2 for n;;' no'

)-1

It follows from this thatf (n) - f E 12 , thus f E 12 , also. and IIf (n) - fll <E for
n ;;. no, i.e .• f (n)~f.

EXAMPLE 5. The Lebesgue space L2(M) for a Lebesgue measurable subset M

of Rm : For the concepts and results of this example a knowledge of
Lebesgue's integration theory is needed (cf. Appendix A). This will be
assumed in what follows. The notions of "measurable," "almost every-
where," and "integrable" refer to Lebesgue measure in Rm.
Let M be a measurable subset of Rm. First we treat the function space

~(M) = {f: fmeasurable complex-valued on M. fMlf(x W dx < oo}.

~(M) is a vector space, since withf, g E ~(M), a E C the functions af and

f + g are also measurable and because of
laf(x)1 = lallf(x)1 and If(x) + g(x)12 .;;; 2lf(xW + 2Ig(x)12
we have
fMlaf(xW dx < 00 and fM1f(x) + g(x)12 dx < 00.
2.1 Convergence and completeness 19

It is obvious that by sU, g) = f"J ·(x)g(x) dx, I, g E f;(M) a semi-scalar

product is defined on f;(M). Let

'.')L( M) = {I: I measurable complex-valued function on M,

I (x) = 0 almost everywhere on M}.
Then '.')L(M) is a subspace of f;(M) and we have
'.')L(M) = {J E f;(M) : sU, f) = OJ.
Now we define
L2 (M) = f;(M)/'.')L(M);

Thus we build equivalence classes in f;(M) by placing two functions in the

same class if they coincide almost everywhere. Addition of these classes
and multiplication by a complex number are defined via representatives: If
j and g are the equivalence classes of I and g and a E C, then
aj= (air, aj+bg = (al+bgr.

The scalar product of two equivalence classes j and g is defined by

<j, g) = sU, g) = fj ·(x)g(x) dx,

where I and g are representatives of j and g. It is evident that this
definition does not depend on the choice of I and g. From <j,j) =0 it
follows that the representatives of j vanish almost everywhere, i.e., j is the
zero element of f;(M)/'.')L(M). Therefore (. , .) is actually a scalar prod-
uct, and L2(M) is thus a pre-Hilbert space. In what follows we shall denote
the functions IE f;(M) and the corresponding equivalence classes j E
L2 (M) by the same symbol f. The function I is then always an arbitrary
representative of the corresponding equivalence class.

Theorem 2.1. (L 2(M), (. , .»

is complete, thus it is a Hilbert space. II In~/,
, 01 Un) such that
then there is a subsequence Un)

,
!,,(x)~/(x) as )-+00, almosteverywherein M

, ,
(here In ( . ) and I( . ) are arbitrary representatives ol!", respectively f).
PROOF. Let Un) be a Cauchy sequence in L2(M). For each) EN there exists
an nJ such that

Without loss of generality we may assume that nJ+ 1 >"l for aU) EN. Then
we have in particular 11/~+1 - In" or;;; rJ. In what follows let In( . ) be an
arbitrary (however, in the rest ot the proof fixed) representative of!".
20 2 Hilbert spaces

For all kEN let gk : M~R be defined by the equality

k
gk(X) = .L linj+,(x) -
J=t
fn/ x ) I·

The sequence (gf( . » is non-decreasing, and

for all kEN. By B. Levi's theorem (Theorem A 7) the sequence gl, and
thus also the sequence (gk)' is convergent almost everywhere. Then the
sequence of the functions

also converges almost everywhere to a measurable functionf( .). We show

thatf( . ) E f;(M) and that in the sense of L2(M) we have fn~f as n~oo.
For each e > 0 let nee) andiCe) be chosen so that for n > nee) and) >)(e) we
have

The functions lin ( . ) - fn( . Ware non-negative, their integrals are

~
bounded by f and for )~oo we have
If" (x)
~
- f,,(xW ~ If(x) - f,,(xW almost everywhere in M.
By Fatou's lemma it follows from this that If( . ) - fn( . Wis integrable and
that we have

f.lf(x) - fn(x)j2 dx ,,;;; ( for n > nee).

M

Therefore f( . ) - fn( . ) E ~(M) and, consequently, f( . ) E ~(M). Besides,

we have II f - fn 112 ,,;;; £ for n > n( f), Le., fn ~ f in the sense of LlC M). The
second part of the assertion is proved by the fact that fn (x)~f(x) almost
~

everywhere. 0

If we look only at real valued functions in this example, then we obtain

the real Hilbert space L2• R( M).

EXAMPLE 6. All the reasoning of Example 5 can be carried out analogously

if p is a measure generated by a regular interval function on Rm (cf. Ap-
pendix A), M is a p-measurable subset of R m and L2(M; p) is the
corresponding space of square integrable functions with respect to p.
Theorem 2.1 holds true for L2(M ; p) also. We omit the details here.
2.2 Topological notions 21

EXERCISES

2.1. Let Un) be a sequence in the normed space (H, II . II) with ~~-dl.{"11 < 00.
(a) In-O and the sequence (~j-I!J) is a Cauchy sequence.
(b) If H is a Banach space, then the sequence (};j-I!J) is convergent; we write
}; j.. I!J for the limit of this sequence.
2.2. (a) In Exercise 1.11 H is a Hilbert space if and only if all Ha are Hilbert
spaces.
(b) The space liA ; fL) of Exercise 1.l0 is a Hilbert space.
(c) The spaces A2 and H2 of Exercise 1.9 are Hilbert spaces.
Hint: This can be proved with the aid of Exercise 1.9(c) or the mean value
property of holomorphic functions.
2.3. (a) Let Ck[O, I] be the vector space of k times continuously differentiable
complex (or real) valued functions defined on [0, 1]. By
k
(i, g)k = ~ f/(j)(x)·g(j)(x) dx
j-O 0

a scalar product is defined on Ck[O, I]. The space (Ck[O, I], (., ')k) is not
complete.
(b) Let W2• k(O, I) be the space of those complex-valued functions on [0, 1)
that are k - 1 times continuously differentiable, whose (k - I)th derivative
is absolutely continuous (cf. Appendix A 5) and whose kth derivate is in
L 2(0, I). By

a scalar product is defined on W2 , k(O, I). The pair (W2, k(O, I), (. , ')k) is a
Hilbert space.
(c) Ck[O, I] is a subspace of W2,k(O, I). For each 1 E W2,k(O, I) there exists a
sequence (j,,) from Ck[O, I] such thatin-i in the sense of W2,k(O, I).

2.2 Topological notions

Let (H, II . II) be a normed space. A subset A of H is said to be open if for
each f E A there exists an t: > 0 such that the ball
K(j,t:) = {gEH: IIg-fll<t:}
lies in A.

EXAMPLE 1. For each r;> 0 and each hE H the ball K(h, r) = {g E H :

II g - hll < r}
is open. It will be called the open ball around h with radius r.
The assertion is obvious for r = 0, as K(h, r) is then empty (the empty
set is open). Now let r> 0, g E K(h, r), then we have
t: = r -II g - hll > 0
22 2 Hilbert spaces

and for eachiE K(g, ()

IIh -111 ;;;; IIh - gil + Ilg - 111 < IIh - gil + ( = r,
i.e., K(g, t':) < K(h, r).
A subset A of H is said to be closed if CA = H\A, the complement of A,
is open.

EXAMPLE 2. For each 1 E H and each r > 0 the ball K(j, r) = {g E H :

is closed, because for g E CK(j, r) we have II g - 111- r > 0
II g - 111 ;;;; r}
and K(g, II g - 111- r) c CK(j, r). The set K(j, r) is called the closed ball
around 1 with radius r.

Closed sets can be characterized in another way. For this we mention

another definition. An element 1 E H is called a contact point of the subset
A of H if for each t': > 0 there exists agE A such that II g - 111 <t:. The set
of all contact points of A will be denoted by A. We obviously have A cA.

Proposition.
(I) A C S implies A C B.
(2) We have 1 E A ii and only ii there exists a sequence (jn) in A such that
I,,~j.
(3) We have A= A.
PROOF. (I) and (2) are clear.
(3) Let f E A, E> O. Then there exists agE A such that II g - 111 < E/2 and
for this g there exists an hE A such that IIh - gil <£/2; consequently
II h - 111 < £. Therefore f E A holds, i.e., Ac A. Since A C A, it follows that
A=A. 0

Theorem 2.2. A is closed. A is closed if and only if A= A. The set A is the

smallest closed subset oj H that contains A.
PROOF. First we show that A is closed, i.e., CA is open. Let 1 E CA. Since
A =:4, then we have 1 E CA, i.e., f is not a contact point of A. Therefore
there is an £ > 0 such that K(j, €) n A = 0 and consequently K(f, t':) C CA.
lf A = A, then A is closed by the first part of our theorem. If A is closed,
then CA is open, i.e., for each 1 E CA there exists an t: > 0 such that
K(j, to) n A = 0. However, this means that no element J of CA is a contact
point of A, therefore A c A and thus A = A.
If B cHis closed and A C S, then it follows that Ac B = S, therefore
Ac& 0

On the basis of Theorem 2.2 it is justified to call A the closure (closed hull)
of A.
2.2 Topological notions 23

EXAMPLE 3. For r > 0 the closed ball K(f, r) is the closure of K(f, r). For if
g E K(f, r), then for all n E N the element g" = ~ -
to K(f, r) and we haveg,,---;.g. Hence K(f, r)CK(j, r). As K(f, r) is closed,
*)( g - f) belongs

we also have K(j, r)c K(f, r).

Theorem 2.3. The closure of a subspace of H is a subspace.

PROOF. Let T be a subspace of H, let f, gET and let a, bE K. Then there

are sequences (J,,) and (gn) in T such thatJ,,---;.J, gn---;.g. It follows that
af + bg = a limJ" + b lim gn = lim (aJ" + bg,,).
As afn + bgn E T, it follows that af + bg E T. o
If (H, /I . IJ) is a normed space and T is a subspace of H, then the
restriction of II . II to T defines a norm on T. Thus T becomes a normed
space (T, II. II) in a natural way. Analogously, if (H, is a pre-Hilbert<. , .»
space, then we can consider T as a pre-Hilbert space (T, <. , .».
Theorem 2.4. A subspace T of a Banach space (H, <. (respectively a»
Hilbert space (H, <. ,
.») is closed if and only if (T, II . II) is a Banach space
(respectively (T, (. , .»
is a Hilbert space).
PROOF. If T is closed, and (J,,) is a Cauchy sequence in T, then there exists
an f E H such that fn ---;. f; therefore f E T = T, i.e., T is complete. If T is
complete and f E T, then there exists a sequence (f,,) from T such that
fn ---;. f; as (fn) is a Cauchy sequence, (fn) is convergent in T, i.e., f E T. 0
Let A and B now be subsets of a normed space H. The set A is said to
be dense relative to B if B c A holds. If, in addition, A C B, then we say
that A is a dense subset of B (or briefly A is dense in B). If A is dense
relative to H, then we say briefly that A is dense.

Proposition. If AI is dense relative to A z and A z is dense relative to A 3 , then

AI is dense relative to A 3 •
-
PROOF. From A3 C A2 and A2 C Al it follows that A3 C AI = AI' o
EXAMPLE 4. A sequence of complex numbers f = (fn) is said to be finitary if
only finitely many members J" are different from zero, i.e., f =
(fl,I2 , ••• ,In' 0, 0, ... ). The set of finitary sequences is a subspace /2,0 of
I . We show that '2 0 is dense (in (2)' Let 1= (fn) be an arbitrary element of
l~. Then for each/E N we have f (j) = (fl' . • . ,Jp 0, 0, ... ) E 12,0 and
00

Ilf - f (J)1I 2 = L I1nl 2 ---;. 0 as j ---;. 00.

n=j+1

Consequen tly I (j) ---;.1, i.e., I E /2, o·

 2 Hilbert spaces

EXAMPLE 5. Let P be the vector space of polynomials in one variable. P

can be considered as a subspace of the pre-Hilbert space (C[O, I], <. , .»
from Section 1.2, Example 4. P is dense in C[O, 1]: By Weierstrass'
approximation theorem (d. Hewitt-Stromberg [18], (7.31» for each con-
tinuous f and for every E > 0 there exists a polynomial p such that
max {If(x)-p(x)1 : x E[O, In
<E. We also have then that IIf-pll <E, i.e.,
f is a contact point of P.

EXAMPLE 6. Let M be a measurable subset of Rm. Let

L2,o(M) = {J E L2(M): there exists a K > 0 such that
If(x)1 <K almost everywhere in M,
andf(x)=O almost everywhere in {xEM: Ixl>K}}.
L2, o( M) is dense in L2 ( M). For let f be an element of L2( M) and for each
n E N let
if Ixl..;n and f(x) <n
otherwise.
Then we have If,,(x)1 < If(x)1 for all n E N and all x E M, andf,,(x)~f(x)
as n~oo. By Lebesgue's dominated convergence theorem it now follows
that

IIf" - fll2 = fM1f,,(x) - f(xW dx ~ 0 as' n ~ 00.

Therefore f" ~ f· Since f" E L2, o( M), the assertion follows.

EXAMPLE 7. A subset J of Rm of the form
J= {x = (Xl' ••• , xm ) E Rm : aj ~ Xj ~bpj = 1, 2, ... , m}
with ap bj E IR, is called an interval in Rm; here any combination of the
signs < and ..; is permitted. A function f: Rm~c is called a step
function if there are finitely many intervals J l , • . . , J" and complex
numbers Cl> ••• , c" such that
,.,
f(x) = ~ cjxAx),
j-I ~

where XA denotes the characteristic function of A, i.e.,

for X EA
otherwise.
m
The set T(Rm) of step functions on R is obviously a vector space (the
linear operations are defined as usual). We show that T(Rm) is a dense
subspace of L2(Rm). To prove this it is enough to show that T(Rm) is dense
in L2 ,o(Rm). It is obvious that T(Rm) c L2,o(Rm ). Let f E L2, o(Rm). Then f is
integrable and there exists (cf. Theorem A6) a sequence (i,,) from T(Rm)
2.2 Topological notions 25

such thatf,,(x)--)of(x) almost everywhere in IR m and

flJ;,(x) - f(x)1 dx --)0 0 as n --)0 00.

(Integrals for which no domain of integration is given are always taken

over the whole space IRm.) If for K;;;. 0 we have If(x)l< K almost every-
where, then we may assume that 1f"(x)l< K for all x E IR m , and for all
n E N. Consequently, we have

IIf" - fll2 < 2K flfn(x) - f(x)1 dx --)0 0 as n --)0 00,

i.e.,f,,~f·

EXAMPLE 8. Let CoOO(lRm) be the space of infinitely many times differentia-

ble complex-valued functions with compact support (i.e., for every f E
Cooo(lRm) there exists a compact subset Kin IR m such that f vanishes outside
K; the smallest set K of this kind is called the support of f. in symbols
supp 1). We show: Cooo(lRm) is a dense subspace of L2(lRm). For the proof it
is enough to show that Cooo(Rm) is dense relative to T(lRm). To prove this it
is enough to show that for every interval J the characteristic function XJ is
a contact point of Cooo(lRm). For this, let us define 8. E Cooo(lRm) by

for Ixl <f,

for Ixl ;;;Of,
and

8. = {fS.(x) dX} -I S•.

The reader can verify himself that 8. E Cooo(lRm) and supp 8. = {x E
Rm : Ixl " f} hold. If J is now an interval in R m and for n E N we define

I
I for with d(x. CJ);;;' -,
f,,(x) = { n
I
o for xER m with d(x. J);;;'-,
n

and 0 <; ],,(x) <; I for all x E R m (here d(x. A) stands for the Euclidean
distance of the point x from the set A). We have ],,(x)~xix) for all x that
do not lie on the boundary of J. Therefore f,,(x)--)oxix) almost every-
where. Thus by the, Lebesgue dominated convergence theorem it follows
26 2 Hilbert spaces

that

i.e., J,,--+XJ in the sense of L2(Rm).

A subset A of a normed space is said to be separable if there exists an at
most countable subset B of A which is dense in A. If T is a subspace of H
and a is a subset of H, then a is said to be total with respect to T if the
linear hull L(a) (the set of finite linear combinations of elements of a, or,
in other words, the smallest subspace of H that contains a) is dense
relative to T. We use the concept total in if acT, and total if T= H.

EXAMPLE 9. The spaces em and Rmare separable, as the set of elements

with rational components (in em this means that the real and imaginary
parts of the components are rational) is enumerable and dense.

Theorem2.5. Let (H, II . 10 be a normed space.

(a) If A is a separable subset of H, then A is separable, also.
(b) If A is separable and Al C A, then Al is separable, too.
(c) A subspace T of H is separable if and only if there exists an at most
countable subset A of H that is total with respect to T.
PROOF.
(a) Let a be at most countable and dense in A. Since A is dense in A, the
set a is also dense in A.
(b) Let a = {J" : n E~} be an at most countable set that is dense relative
to A. Let J be the set of those pairs (n, m) E ~ X ~ for which there
exists an f E Al such that Ilf - f,,11 <;; 11m. For every n, mE J let us
choose a gnm E A\ such that IIgnm - fnll <;; 11m. The set 8\ =
{gnm : (n, m) E J} is then at most countable. We show that it is dense
with respect to AI' Le., A\ is separable. Letf E A\. As a is dense in A,
8 is also dense relative to AI' Therefore for every k E 1\1 there exists an
n(k) such that II flIck) - fll ..;; I I k. Hence (n(k), k) E J and we have
2
II g,,(k). k - fll ..;; II g,,(k), k - fn(k)11 + IIfn(k) - fll ..;; k'
i.e., g,,(k), k --+f as k --+ 00. Hence 8 1 is a dense subset of A\.
(c) If T is separable, then there exists an at most countable subset a which
is dense in T. Since L(a) ~ a, the set L(a) is dense in T, too. Let a
now be at most countable and let T C L(a). Then L(a) is separable, for
the set Lr(B) of finite linear combinations of elements of a with
rational coefficients is dense in L( a) and Lr( a) is countable. Since
T c L(a), the subspace T is separable, also.

EXAMPLE 10. 12 is separable, as the set of unit vectors {en = (81!f)jEN : n E

~) is total in 12 : the linear hull of the unit vectors is 12, O.
2.2 Topological notions 27

L2(R m ) is separable. 7 it is enough

separable. Let So be characteristic functions
end points. The countable and III
the characteristic functions intervals, therefore L( S).
Because L( S) = T(Rm), it follows from this that L( So):J T(Rm), i.e., T(Rm)
is separable.

EXAMPLE 12. For every measurable subset M of Rm the space L2(M) is

separable. The space L2( M) may be considered as a subspace of L2(Rm)
identify each j the element j defined

xEM,
j(x)
xflM.

EXAMPLE 13. Let p be a measure on R m (cf. Appendix A) and let M be a

p-measurable subset of Rm. The HUberl space LiM, p) (d. Section 2.1,
Example 6) is separable. This can be proved for M = R m as in Example II
(in the course of the proof of S C So one has to notice that the boundaries
of intervals in general have measures different from zero). For a general M
assertion by LiM, p) as a subspace , p).
oj step junctions in L2(Rm, p).

2.4. A subset A of a normed space H is separable if and only if its closed linear
hull L(A) is separable.
2.5. Prove that the function 8. from Example 8 is infinitely many times differen-
tiable.
2.6. Let G be an open subset of Rm, let p be a measure on Rm, and let L2(G, p) be
in Section 2.1, ~n.~u,,"V
o(G, p) is the L1(G, p) consisting
'""""'UV1A" with compact then L2• o( G, p) is
is the space of whose supports
are then compact then T(G) is p).
(c) Co""(G), the space of infinitely many times continuously differentiable
functions with compact support in G, is dense in L2(G, p).
2.7. The spaces A2 and Hl of Exercise 1.9 are separable (cc. also Exercise 2.2c).
2.8. (a) Prove the separability of Lia, b) for - 00 <a <b < 00 with the aid of the
Weierstrass approximation theorem (cf. Example 5).
(b) With the prove the separability
(c) Prove of L2(R m ) analogously.
2.9. (a) A subset only if there exists
ana> for all!. g E B.
28 2 Hilbert spaces

(b) If a subspace T is not separable, then for each a > 0 there exists such' a set
B.
(c) /2(A ; p.) is separable if and only if A is at most countable.
2.10. Let H be a pre-Hilbert space and let T be a dense subspace of H.
(a) The closures of (f E T : 11111 .;;; l) and of {f E T: II /11 < I} are equal to
1(0, I).
(b) For every / E H we have II/II = sup {I</, g)1 : gET, II gil ..; I} =
sup W/, g>1 : gE T, II gil < I}.
 3
Orthogonali ty

3.1 The projection theorem

Let (H, <. , .» be a pre-Hilbert space. Two elements /, g E H are said to be
orthogonal (in symbols /1. g) if <1, g) ... O. If /1. g, then we obviously have
1//+ gl/2 = 1//1/ 2+ 1/ g1/2; this formula often is referred to as the Pythagorean
theorem. An element / E H is said to be orthogonal to the subset A of H (in
symbols /1. A), if /1. g for all g E A. Two subsets A and B of H are said to
be orthogonal (in symbols A...L B) if (j, g) = 0 for all / E A, g E B. If A is a
subset of H, then the set A J. = {/ E H : /1- A} is called the orthogonal
complement of A.

Proposition.
(a) We have {O} J. = H, H J.. = {O}, i.e., 0 is the only element orthogonal to
every element.
(b) For every subset A 0/ H the set A J.. is a closed subspace 0/ H.
(c) A C B implies BJ. C A..L.
(d) We have A..L = L(A)..L =L(A)..L.
PROOF.
(a) For every /E H we have <0,/) =0. If /~O, then (j,j) ~O, i.e., / is
not orthogonal to H.
(b) If /, g E A..L and a, bE K, then for all hE A it follows that

<a/ + bg, h) = a*(j, h) + b*<g, h) = 0,

i.e., a/ + bg E A..L. Therefore A..L is a subspace. It remains to prove that

A..L C A..L. Let / E A..L, and let Un) be a sequence from A..L such that

29
30 3 Orthogonality

fn~f. Then we have for all hE A that

<f, h) = lim<fn , h) = 0,

consequently f E A.L .
(c) If fEB.L, then we have <f, h) = 0 for all hE B, therefore also for all
hE A, and thusfE A.L.
(d) Since A C L(A) C L(A), from (c) it follows that
L(A/ cL(A).L cA.L.

It remains to prove that A.L cL(A).L. If f E A.L, then we have (j, h) =

o for all hE A, and therefore for all hE L(A), as well. If hE L(A), then
there exists a sequence (hn) from L(A) such that hn~h. Consequently,
we have
(j, h) = lim<f, hn ) = O.
i.e.,! E L(A).L . o
In order to prove the projection theorem we need an approximation
theorem, which we prove with somewhat more generality than we actually
need. A subset A of a vector space is said to be convex if from x, yEA and
0< a < 1 it follows that ax + (1 - a)y E A. Any subspace is obviously
convex.

Theorem 3.1. Let H be a Hilbert space and let A be a non-empty closed

convex subset of H. Then for each f E H there exists a unique g E A such that
IIf- gil = d(j, A) = inf{llf-hll: hEA}.

PROOF. There always exists a sequence (gn) of elements of A such that

II gn - dU, A). If we replace f by gn - f and g by gm - f in the
fll~d=
parallelogram identity (1.16), then on account of the inequality IIf - hll >d
for all hE A, we have
II gn - gmll 2 = 211 gn - fl12 + 211 gm - fl12 - 411f - i(gn + gm)1I 2
" 211 gn - fll2 + 211 gm - fll2 - 4d 2 ~ 0
as n, m~oo (here we have used the fact that (gn + gm)/2lies in A, since A
is convex). Hence (gn) is a Cauchy sequence. As H is a Hilbert space, there
exists agE H such that gn ~ g. We have g E A, since A is closed. Moreover,
we have
II g - fll = Iimll gn - fll = d.
It remains to prove that g is uniquely defined. If g, hE A are such that
IIf - gil = IIf - hll = d, then for the sequence (gn) = (g, h, g, h, g, ... ) we
obviously have II gn - fll = d. By the above reasoning (gn) is a Cauchy
sequence, i.e. we have g = h. n
3.1 The projection theorem 31

Theorem 3.2 (Projection theorem). Let H be a Hilbert space, and let T be a

closed subspace of H. Then we have T 1. 1. = T. Each f E H can be uniquely
decomposed in the lorm 1 = g + h with gET and h E T 1.. This g is called the
(orthogonal) projection 011 onto T.
PROOF. As T is convex and closed, by Theorem 3.1 there' exists agE T
such that III-gil = dU, T). Let us set h = 1 - g.
hE T1.: We have to prove that for all wET we have (w, h) = O. For
w = 0 this is clear, so let wET, W:f= O. Then for all a E IK the element
g + aw also belongs to T. Therefore
d2 = dU, T)2 < 111 - (g+ aW)1I2 = IIh - awll2
= IIhl1 2 - 2 Re(a(h, w» + lal211wll2
=d2 - 2 Re(a(h, w» + lafllwll 2 •
With a= IIwll-2(w, h) it follows from this that
IIwll-21(w, hW < 0,
so (w, h) =0.
In order to prove the uniqueness of the representation 1= g + h let us
assume that 1 = g + h = g' + h' with g, g' E T, and h, h' E T 1.. Then we have
g - g' E T and h' - h E T1., therefore
g - g' = h' - hE Tn T1. = {o}.
It follows from this that g = g' and h = h'.
lt remains to prove that T = T 1. 1. .
T C T 1. 1.: If 1 E T, then by the definition of T 1. we have (J, g) = 0 for
all gET 1., i.e., 1 is orthogonal to T 1. ,IE T 1. 1. .
T 1. 1. C T: Let 1 E T 1. 1.. On the basis of what we have already proved
the element f may be represented in the form 1 = g + h with gET C T 1. 1.,
h E T 1.. From this it follows that h = 1 - gET 1. n T 1. 1., hence h = 0, i.e.,
l=gE T. 0

Proposition.
(a) Let H be a Hilbert space. For every subset A 01 H we have A1. 1. = L(A),
i. e", A 1. 1. is the smallest closed subspace containing A.
(b) In a Hilbert space H we have A1. = {OJ if and only if L(A)= H holds,
i.e., if A is total.
PROOF.
(a) Since A1. =L(A)1., the projection theorem shows that L(A)=
L(A) 1. 1. = A 1. 1..
(b) If A1.={O}, then we have L(A)=A1.1.={O}1.=H. If A1.1.=H, then
we have A1. = A 1. 1. 1. = H 1. = {OJ, as A 1. is a closed subspace. 0

If T\ and T2 are subspaces of a vector space such that T\ n T2 = {OJ,

then
32 3 Orthogonality

is a direct sum (consequently we write T\ -i- T2), i.e., each element from
T) + T2 has exactly one representation of the form f + g with f E T) and
g E T2 • If T) and T2 are subspaces of a pre-Hilbert space with T)l. T2, then
we have T) n T2 = {O}. In this case we call the direct sum T\ -i- T2 an
orthogonal sum and we denote it by T\ EB T2•

Theorem 3.3.
(a) Let H be a pre-Hilbert space, and let T\ and T2 be orthogonal subspaces.
If T\ EB T2 is closed, then TI and T2 are closed.
(b) If H is a Hilbert space and T\> T2 are closed orthogonal subspaces, then
T\ EB T2 is closed.
(c) If H is a Hilbert space and T and T\ are closed subspaces such that
T\ c T, then there exists exactly one closed subspace T2 such that T2 c T,
T21. T) and T= T\$ T2.

For the subspace T2 , defined uniquely by part (c) of this theorem, we write
briefly T2 = T e T\. The subspace T2 is called the orthogonal complement
()f TI with respect to T. For T= H we obtain that He T) = Til..
PROOF.
(a) We show that T\ is closed (the proof for T2 goes the same way). Let
f E T) and let (fn) be a sequence from T\ such that fn -f. Since
TI C T\ $ T2 , we have TI C T) $ T2 = T) $ T2 • Hence f E TI $ T2 and
thus we have f= g) + g2 with g\ E T\> g2 E T2. On the other hand, it
follows from In E T\ that fn 1. T2 and so f 1. T2, and, consequently,
g2 =f - g \ E T2 n Ti". Therefore g2 = O. From this it follows that f = g
ETI •
(b) We have to prove that T) EB T2 c TI $ T2 • Let f E T\ $ T2 ; then there
exists a sequence (fl, n + An) E TI $ T2 with fl. nET\> An E T2 and
fl, n + f2, n~f. Since

the sequences (fl. n) and (f2, n) are Cauchy sequences. Consequently

fl. n~ fl E T\> A n~ f2 E T2. From this it follows that

f = lim (fl n + f2 n) = f\ + f2 E T\ $ T2·

(c) By Theorem 2.4 T is a Hilbert space, Without loss of generality we
may assume that T = H. In this case let us set T2 = T\.l. Then by the
projection theorem (Theorem 3.2) we have H= TI $ T2 • In order to
prove uniqueness, let us choose an arbitrary subspace T2 such that
H = T\ $ nThen we surely have T2 C T\.l. If f E TiL, then f = f\ + f2
with fl E T"f2 E n Here we must have f, =0, since 0= (fl,f) =
<f\,fl) = IIf11l2. Therefore we have f= f2 E T5., i.e., T\.l C T2 and thus
TI.l= n 0
3.1 The projection theorem 33

EXAMPLE 1. Let - 00 <;; a <;; c <;; b <;; 00. In Lz{ a, b) by

T. = {jEL 2(a,b) :j(x)=O almosteverywherein (a,c)}
T2 = {jEL2(a,b) :j(x)=O almosteverywherein (c,b)}
two subspaces are defined, and we have Lz{a, b) = T. $ T2 : For 8 E T.
and hE T2 we obviously have <g, h> = O. Moreover, for each j E L2(a, b)
we have g=X(a,c)fE T.. h = Xc,,}'E T2 andf= g+ h,

EXAMPLE 2. In L2( - a, a) by
T± = {JEL 2 (-a,a) :j(x)=±j(-x) almosteverywherein (-a,a)}
two subspaces are defined and L2( - a, a) = T + $ T _: For gET + and
hE T_ we have

<8, h> = fa g(x)*h(x) dx =

-0
fO g(x)*h(x) dx - fO g(x)*h(x} dx = O.
-0 -0

Let us setj±(x) = Hj(x} ± j( - x». Then for eachj E L2( - a, a) we have

f ± E T ± and j = j + + j _. The subspace T + is the space of even junctions.
T _ is the space of odd junctions.

If T.. ... , Tn are mutually orthogonal subspaces of H, then we call the

(direct) sum of these spaces an orthogonal sum and we write
n
$ 7j = T. $ T2 ED • . . ED Tn.
j~1

Parts (a) and (b) of Theorem 3.3 can be extended to this case. For
infinitely many subspaces see Exercise 3.3.
If A is an arbitrary set and (Ha, <. , . >
..) is a pre-Hilbert space for each
0: E A, then by Exercise 1.11 the space

H= {j=u',)aEAE II
aEA
Ha:

ja 7'= 0 for at most countably many 0: E A, and ~ Ilfa 112 < 00 }

aEA

is a pre-Hilbert space with the scalar product

<(fa)' (8a »= L
aEA
<ja' ga>a for (fa)' (8a) E H.

By Exercise 2.2a the space H is a Hilbert space if and only if all Ha are
Hilbert spaces. If we identify Ha with the subspace of elements (ffJ)fJ EA
such that jfJ = 0 for p 7'=0:, then the spaces Ha become pairwise orthogonal
subspaces of H. Therefore (H, (. , . >) is called the orthogonal sum of the
spaces (H", (. ">a) in symbols H= $aEAHa (if A is finite, then H=
IIaE Ha)·
34 3 Orthogonality

Theorem 3.4. Let H be a Hi/bert space. If T is a closed subspace and S is a

finite dimensional subspace, then T + S is closed.
PROOF. The problem can be reduced by induction, to the case where Sis
one dimensional; S = L(f). If we write f = 1\ +12 with 11 E T and 12 E T J.,
then we have T + S = T E9 L(f2)' Therefore T + S is closed by Theorem
3.3(b). [J

EXERCISES

3.1. Let H be the pre-Hilbert space {f E C(O, 1) : f(l) = O} with the scalar product
<f,g) = f6f(t)·g(l) dt. The subspace T= {fE H: f6f(t) dt=O} is closed,
T=I' H, and TJ. = (OJ.
3.2. Let H be the pre-Hilbert space C[ -I, I) with the scalar product <1, g) ...
J ~ tf(t)·g(t) dt. The subspaces T\ = {f E H : f(/) = 0 for t.;;; O} and T2 ...
{f E H : f(/)'" 0 for 1 ;;. O} are closed and such that T\.L T2 • The orthogonal
sum T\ EB T2 is not closed (cf. Theorem 3.3(b».
3.3. Let H be a Hilbert space, and let {Ta : a E A} be a family of pairwise
orthogonal subspaces of H.
(a) If (fa)EITaEA To, and fa *0 for at most countably many a, and
l:aEA II fa 112 < 00, thenf= l:aEAfa can be defined. The subspace T of allf
of this form is called the orthogonal sum of To, in symbols T= EBaEA Ta.
Hint: Build a sequence (an) from those a for which fa *0 and define
l:aE,J'a as l:::'-tfa". This definition is independent of the choice of the
sequence (an).
(b) T is closed if and only if all T" are closed.
(c)" If all T" are different from (OJ, then T is separable if and only if A is
countable and all T" are separable.
3.4. Let H be a pre-Hilbert space, let D be a dense subspace of H, and let N be a
finite dimensional subspace of H. Then D n N J. is dense in N J..
Hint: By induction on n =dim N we can reduce the problem to the case n = I,
i.e., N = L( g), g =I' O. Then there exists an h E D such that <g, h) = I. If
fE NJ., then there exists a sequence (fn) from D such that fn~f. For the
sequence f~ = fn - <g, fn) h we then have f~ E D n N J. and f~ -+ f.
3.5. Let H be a pre-Hilbert space and let T\ and T2 be subspaces of H such that
T\.L T2• Then we have T\ EB T2 ::J 1'\ EB 1'2' If H is a Hilbert space, then we have
T\ EB T2- 1'\ EB 1'2'

3.2 Orthonormal systems and orthonormal bases

Let (H, <. ,.»
be a pre-Hilbert space. A family M = {e" : a E A} of
elements from H is called an orthonormal system (ONS) if we have
3.2 Orthonormal systems and orthonormal bases 35

(8ap denotes the Kronecker delta, Le., 8aa = I for all a E A and 8ap = 0 for
a =1= P). An orthonormal system M is called an orthonormal basis (ONB) of
the subspace T if M is total in T (i.e., MeT and L(M):J T). If M is an ONB
of H, then M is called an orthonormal basis.

Proposition.
(a) Each ONS is linearly independent (i.e., every finite subsystem is linearly
independent).
(b) Each ONB M is a maximal ONS (i.e., if M' is an ONS such that
Me M/, then we have M' = M).
(c) If H is a Hilbert space, then each maximal ONS is an ONB.
PROOF.
(a) If {e l , . .. , en} is a finite subsystem of an ONS, then {el" .. , en} is
also an ONS. If l:j_l~~ = 0, then it follows that

Therefore aj = 0 for all j.

(b) Let M be an ONB. If M were not maximal, then there would be an
e E H such that e J.. M; consequently e J.. L(M), this contradicts L(M)
=K .
(c) Let M be a maximal ONS in the Hilbert space H. If M were not total,
i.e., if we had L(M)1. =1= {OJ, then there would be an e E L(M)1. such
that II ell = 1. Hence M' = M U {e} would be a larger ONS which
contradicts the maximality of M. 0

EXAMPLE 1. The set of unit vectors {e 1, ••• , em} is an ONB in II(m (ej is the
vector with I at the jth place and zero otherwise).

EXAMPLE 2. The set of unit vectors {ek = (8kn )neN : kEN) is an ONB in 12,

EXAMPLE 3. An ONB in A2 is {en: nEN o} with en(z)=[(n+I)/'IT]1/2z n.

An ONB of H2 is Un : n E No} with fn(z) = [2'ITr 112 zn. This follows
immediately from Exercise 1.9, in particular part (c).

EXAMPLE 4. In L2(O, I) the set M= {en: n E l} with en(x) = exp (2i'ITnx) is

an ONS, as one can verify by a simple calculation. We show that M is an
ONB, i.e., that M is total. For this let 6[0, IJ = U E C[O, IJ : f(O) =f(l)}.
For each f E 6[0, IJ by Fejer's theorem there exists a sequence (fn) of
trigonometric polynomials Un E L(M» such that fn uniformly tends to f.
We also have then that fn~f in the sense of L:z(O, I), i.e., M is total in
6[0, I]. If we also prove that 6[0, I] is dense in C[O, I], then everything
36 3 Orthogonality

I
will be proved. For f E C[O, I] and n E I\J let us define

f(X) for O<x < I-..!n

fn(x) = I
f(O) + (1- x)n(!( 1- -;;) - f(O») for I -..!n <x < 1.
Then we obviously have fn E C[O, I] and fn ~ j.

EXAMPLE 5. Let Fo = L{ e~ : A E~} with e~ : ~~C, e~(x) = exp (i>..x). On

Fo by
. I fT
<f, g) = hm 2T f(x)*g(x) dx
T->oo - T

a scalar product is defined. For the proof of the existence of this limit it is
sufficient to treat f = e~ and g = e,.. For these we have

<1, g) =
T-HYJ2T -T
f
lim _1_ T el(p.-~)x dx = lim _1_ ~
T->oo2T I1-A
[el(P.-~)T - e -i(P.->.)T]

= 0 for 11 ¥= A.

<1, g) = lim 21TfT dx = I for 11 = A.

T->oo - T

The properties [(1.l3)(i-iv)] of scalar products are obviously satisfied. If

n n
f = ~ ak e"" then <f, f) = ~ lakl 2•
k=l k=l

Therefore we have <f, f) = 0 if and only if f = 0, which is property

[(1.l3)(v)]. By construction, M = {e~ : A E~} is total in Fo, i.e., M is an
ONB in Fo. The space Fo is not a Hilbert space. For if (Ak ) is a sequence of
mutually distinct real numbers and (ak ) is a sequence of complex numbers
such that ak ¥= 0 for all k and ~lakl2 < 00, then the sequence Un) with
fn = ~k_lake,,* is a Cauchy sequence that is not convergent.

The following theorem, known as the Gram-Schmidt orthogonalization

process enables us to generate orthonormal systems and (in separable
spaces) orthonormal bases.

Theorem 3.5. Let H be a pre-Hilbert space. For each finite or countably

infinite set F = {fn} from H there exists a finite or countably infinite
orthonormal system M = {en} such that L( F) = L( M). If F is linearly indepen-
dent, then we can also insure LUI' ... ,In) = L(el' ... , en) for all n. 1 If we
require that in the representation en = ~j_Ia)j the coefficient an is positive,
then M is uniquely determined.

lIn the sequel we write L(e, • ...• en) in place of L({ el' ...• en})'
3.2 Orthonormal systems and orthonormal bases 37

(In what follows we shall always use the Gram-Schmidt orthogonaliza-

tion process with this additional requirement.)
PROOF. It is obviously enough to prove only the last part of the assertion.
Every normed element from LU,) has the form bdll,II-1, with ,bll = 1.
The additional condition a, = bdll,ll-' >0 gives bl = 1. That is, e l =
IIldl-11' So we obviously have L(el)=LUI ), as well. Let us now suppose
that e" ... , en are determined in such a way that L(e" ... , ell)
== LUI' ... , I,,). For every
"
g = ~ bjej + b,,+I!"+1 E L(e" ... , en'!,,+I) =L(j" . .. ,jn+l)
j-I
such that g..L L(e l , ... , en) we then have
0= (e;, g) = b; + b,,+,(e;,ln+,), i = 1, ... , n,
thus b; = - bn + 1< e;, !" + I). Consequently, ell +, necessarily has the form

From the requirement,

,::::,llln+l- ;~I (e;,I,,+I)e;r > 0

l
a,,+1 =

it follows that b,,+ I> 0, therefore bll+ ,Ib,,+ 11-' = 1. Consequently,

By construction, we have e,,+, E L(I" ... ,!,,+ I)' hence L(e" ... , en+,)
C LUI' ... ,!,,+ I)' From the formula for e,,+, it follows that 1,,+ IE
L(e l, ... , en+I), therefore we also have LUI"" ,In+l)cL(e l, ... , e,,+I)'
o
EXAMPLE 6. In L2( - I, I) the set F= Un : n E No} with !,,(x) = xn is a
linearly independent system. The application of Schmidt's process provides
an ONS M= {Pn : n ENol, where Pn(x) = ~j_oanjxj holds with ann> 0; i.e.,
Pn is a polynomial of degree n with a positive leading coefficient. These
polynomials are called the Legendre polynomials. As F is total, the
Legendre polynomials constitute an ONB in L2( - I, I). The polynomial Pn
can be given explicitly:

_ (2n 1)_1(2n+1 )1/2 dn (2 I)" "'.

p" ( x ) - n. -2- dx" x - , n E ''Ilo'

In order to prove this formula it is sufficient to show that the expression

38 3 Orthogonality

for Pn is a degree n whose leading coefficient is

positive and that <Pn, Pm> = 13nm • The first assertion is obvious. For j <m
obtain a (j I)-fold integration parts (the terms
that

d'+
em ( -IY+II _I)m dx =
I
= _ l , _ ( Xj ) , (x 2 O.
-I dx1+ I dX m - 1- 1 ,

implies that >= n =l=m. remains to prove that II 1.

By integration by parts we obtain that

Iff dx

n
=(2n)!n+1 1(- 1+ dx

- I n(n-I)
- (2n). (n + I)(n 2)
II _

= (2n)! n(n-I) ... I II (l+x)2n dx

(n+I)(n+2) ... 2n _I

I
= 2
(n!)--
2n+ I
122n + .

this follows that II Pn = 1.

can in an entirely analogous that generalized Legendre
polynomials
- I [ 2n + 1 ] 1/2 d n n
b, n(X) (b- ] [(x a)(x - n
a

constitute an ONB in L2(a, b).

Theorem Let H pre-Hilbert

(a) If {e p ••• , en} is a (finite) ONS in H, then for eachfE H there exists a
g E L(e l , ••• , en) such that II g - fll = d(j, L(e l , • , • , en»; we have
n
g = ~ <epf)ep
j-I

(b) Let {ea : a be ONS in and let H. at most countably

many of the numbers f> different from and have
Bessel inequality
IIfll2 I<e", 12.
EA
3.2 Orthonormal systems and orthonormal bases 39

(c) An ONS {ea : a E A} is an ONE if and only if for 0/1 f E H the Parseval
equality
IIfll2 = ~ l<ea ,J)1 2
aEA
holds. Then we also have

f = ~ <ea,f)ea for 0/1 f E H.

aEA

PROOF.
(a) For all c" ... , Cn E II< we have

n n
= IIfll2 - ~ I<Ej,J>1 2 + ~ ICj - <ej>f)r
)=1 )-1

Therefore IIf-~j_lcjejll is minimal if and only if cj=<ej>f).

(b) For every finite set {a" ... , an} C A we have

by part (a). Hence

"
~ l<e ,J>1 2 t;; Ilfll2.
aj
)=1

From this the assertion follows because for every f > 0 only finitely
many j E 1\1 exist with the property
. l<ea,JW ;;;. f. ~
(c) Let us assume the Parseval equahty for allfE H. LetfE H, and let (aj )
pe the sequence of those a for which <ea,J) '1=0. Then we have

as n-'HXJ. Consequently, f E L(M), i.e., {ea : a E A} is an ONB. More-

over, it follows that
00

f = L <eal,J>e~ = L <ea,J>ea·
)=1 aEA

Let {ea : aEA} now be an ONB, i.e., let H=L(M). For every fEH
and for every f > 0 there exist n E 1\1, a" ... , an E A and C,' ••• , Cn E
II< such that
40 3 Orthogonality

By parts (a) and (b) it follows from this that

n
o < IIfl12 - L l<e",J)1 2 < IIfll2 - ~ l<e~,f)12
"EA j-I

=lIf-j~1 <e"J'f)e~r <IV-j~1 C)e"r <~.

and consequently that

IIfl12 = ~ I<e",f)(
"EA

Theorem 3.7 (Expansion theorem). Let H be a pre-Hilbert space and let

M- {ea : a EA} be an ONS in H.
(a) If (an) is a sequence of pairwise different elements from A, (cn ) a
sequence from IK, and the series ~cnea" is convergent (i.e.,
lim m _ oo ~~_lCA. exists), then we have (cn) E 12 , If H is a Hi/bert space,
then this series is ~onvergent if and only if (cn) E 12 ,
(b) If g = ~cnea,,' then we have
Cn = <ea,,' g) for all n EN,

II gl12 = ~ Icnl 2
and
<g, f) = ~ <g, ea)<ea,,' f) for all f E H.
(c) The set of all elements from H which can be represented by a convergent
sum ~c"ea" equals L(M).
PROOF.
(a) The sequence (~~-lcnea,,)mEN is a Cauchy sequence if and only if we
have

f c"ea" -
II n=1 f
n= I
cn ea,,112 ~0 for m <k and m, k ~ 00.

This holds true if and only if ~lc,,12 < 00.

(b) We have

II gll2 = lim
m-""oo
f I c"ea,,11
II n-
2
= m~oo
lim f
n= 1
Icnl 2 = ~ Icnl 2

and
3.2 Orthonormal systems and orthonormal bases 41

In the same manner, it follows for all f E H that

(g.J) = lim / f
m ...... oo \ n = ]
c"ea",f) = lim
m-H)IJ
f
n= I
(g, ea")(ea,,.J)

= ~ (g, ea")(ea,,.J).

(c) Each element ~c"ea" is in L{M), by construction. The converse follows

from Theorem 3.6(c), as M is an ONB in L{M). D

Proposition. Let H be a Hilbert space, and let M= {ea : a E A} be an ONS.

For each fEH the vector ~aEA(ea,f)ea is the orthogonal projection of f
onto L(M).
PROOF. If ~j-1Cje"t is an arbitrary element from L(M), then by Theorem
3.6(a) we have

Ilf- L
aEA
(ea,f)eaI12 = IIfll2 - L
aEA
l(ea,f)1 2

~ IIfll2 - L" l(ea/f)1 2

)=\

= Ilf - j~1 (e"t,f)e"tr ~ Ilf - )~I cjeaJr

consequently
lv- a;A (ea.J)ealr = inf {lif-gIl 2 : gEL(M)} = inf {lif-gll z : gEL(M)}.

Theorem 3.8. Let MI and M2 be measurable subsets of ~p and ~q, respec-

tively and let {en : n EN} and {fm : mEN} be orthonormal bases of L2( M\)
and L2(M2 ), respectively. If we define gnm E L2(M. X M2 ) by g"m(x,y)=
en(x)fm(Y) for x EM., Y E M2 , then {gnm : (n, m) E N X N} is an orthonor-
mal basis of Ll M. X M2)'
PROOF. It is obvious that the functions g"m are in Lz(M. X ~) and they
form an orthonormal system. It remains to prove that {gnm : (n, m) E N X
N} is total. Let h be an element of LiM. X Mz) such that hl..gnm for all
n, m. For all x EM. let h/y) = hex, y). By Fubini's theorem we have
hx E L2(M1 ) for all x E M. \ N with some set N of measure zero, and we
have (Parseval's equality)

f.M, IIhxll 1.
00

II h ll 2 = 2 dx = L I(fm' hx)l z dx
M,m=\

= fM'm~llfM2h(X'Y)fm(Y)* dyl Z
dx

= i: f.lf.
m-I M, M2
h(x,y)fm(Y)* dyl2 dx.
42 3 Orthogonality

Let us define k m by

Then k m also belongs to L2(M I ). Another application of Parseval's equality

shows that

I. II. h(x,y)fm(y)* dyl2 dx IIk ll f I(e k m)12

MI M2
= m
2 =
n-I
n,

~ II. km(x)eAx)* dxl2 = f II. I. h(x,y)fm(y)*en(x)* dy dxl2

n-l ~ n-l ~ ~

By summing up, it follows that IIhll = o. o

EXERCISES

3.6. If T= $"EA T" and M" are orthonormal bases in T", then M= U"EAM" is an
ONB in T.
3.7. Let H be an infinite dimensional Hilbert space, and let M be an ONB of H.
The cardinality of every dense subset of H is at least that of M. There exists a
dense subset of H with cardinality equal to that of M.

3.3 Existence of orthonormal bases, dimension of a

Hilbert space
Up to now we have always assumed the existence of orthonormal bases;
only in examples did we see that in certain spaces orthonormal bases exist.
The question is then whether all Hilbert spaces or pre-Hilbert spaces have
orthonormal bases. It is relatively easy to show that each separable
pre-Hilbert space has an ONB. For non-separable spaces it is a little
harder to answer this question.

Theorem 3.9. Let H be a separable pre-Hilbert space.

(a) H possesses an ONB.
(b) If MI is a finite ONS in H, then there exists an ONB in H such that
M-::>M I •
(c) If H is a Hilbert space and MI is an ONS in H, then there exists an
ONB Min H such that M:l MI'
(d) His m-dimensional (m < 00) if and only if there exists an ONB contain-
ing m elements. Then each ONB in H has exactly m elements.
3.3 Existence of orthonormal bases, dimension of a Hilbert space 43

(e) H is infinite dimensional (i.e., not finite-dimensional) if and only if there

exists an ONB containing a countable infinity of elements. Then each
ONB in His enumerably infinite.
PROOF.
(a) follows from (b) if we choose MI =0.
(b) Let H"I'= {OJ and let MI = {e l , ••• , en} be an ONS. As H is separable,
there ¢xists a countable dense subset A = Un : n EN} in H. We define
the elements gl' g2' .••• from A recursively in the following way: let
gl = f", where)1 is the smallest index for which {ej, .•• , en' hi} is
linearly independent. If gl' ... ,gk are defined, then let gk+ 1 = h.d
where jk+1 is the smallest index for which {e l ,···, em' gl' ... ,gk'
j,Jk+1 } is linearly independent. We obviously haveA+I >A. With B=
{el' ... , ell' gl' g2' ... } we then have L(A) c L(B), i.e., B is total. If
we apply Gram-Schmidt's orthogonalization method to B, we obtain
an ONS M, the first n elements of which coincide with e l , ••• , en (as
these are already orthonormal). We have L(M)=L(B)= H, i.e., M is an
ONB in H with MI c M.
(c) Let H be a separable Hilbert space and let MI be an ONS. Then
L(MI)l.. is also separable. Therefore, by part (a), there exists an ONB
Mz of L(MI).l. The set M= MI U M2 is then an ONS, and L(MI U M2 )
= L(M I ) EEl L(M2 ). Consequently, by Exercise 3.5

L(MI U M2 ) = L(MI ) EEl L(M2 ) = L(MI )EElL(M2 ) = I-

hence M is an ONB of H such that MI c M.
(d) Let H be m-dimensional, i.e., assume that the maximal number of
linearly independent elements equals m. As every ONS is linearly
independent, it consists of at most m elements. If M = (e l , ••• , ed is
an ONS with less than m elements, then we have dim L(M) < dim H,
therefore L( M) =1= H. Thus there exists an f E H such that {el>""
ek , f} is linearly independent. The Schmidt orthogonalization process
provides an ONS M' = {e l , ••• , ek' ek + I} such that M C M' , i.e., M is
no ONB. Hence every orthonormal basis has exactly m elements.
(e) If H is infinite dimensional, then every ONB has at least a countable
infinity of elements, for otherwise H would be finite dimensional by
part (d). It remains to prove that each ONB M ... {ea : a E A} is
countable. Let N= {/,. : n EN} be a countable dense subset. For each
a E A there exists an n(a) E N such that IIf,,(a) - e,,11 < Becausei.
lie" - epll"" V2 for a "I'=p, we have
II/"(a) - fn(p)11 ;;. Ilea - epll- IIfn(a) - e,,11 -liep - fn(p)11
;;. v'2 - I > 0 for a "1'= p.
This means that the mapping a~n(a) : A~N is injective. Conse-
quently A is countable. Conversely, if an orthonormal basis in H is
infinite, then H is not finite dimensional. 0
44 3 Orthogonality

Proposition. Every finite dimensional pre-Hilbert space is complete; in partic-

ular, every finite dimensional subspace of a pre-Hilbert space is a closed
subspace.
PROOF. Let H be an m-dimensional pre-Hilbert space. Then there exists an
orthonormal basis {e], ... , em}. Let Un) be a Cauchy sequence in H with
fn = ~j!.]ajnej. Then we have
m
1If.. - .1;,11 2 = ~ lajn - ajpl2.
j=]

thus for each j the sequence (ajn)nEN is a Cauchy sequence as n-:,oo.

Therefore ajn-:,aj as n-:,oo with some aj . Putting
m
f = L aj l1,
j=1
we have

as n-:,oo. Consequently, Un) is convergent in H. D

Proposition. A pre-Hilbert space is separable if and only if it possesses an at
most countable ONB.
PROOF. By Theorem 3.9 each separable pre-Hilbert space possesses an at
most countable ONB. If M is an at most countable ONB in H, then the set
L,(M) of linear combinations of elements of M with rational coefficients is
dense in L(M), and thus it is also dense in H. As L,(M) is at most
countable, H is separable. D
Theorem 3.10. Let H be a Hilbert space.
(a) H possesses an ONB.
(b) If Mo is an ONS. then there exists an ONB Min H such that M:J Mo.
(c) All ONB of H have the same cardinality.
REMARK. Theorem 3.IO(a) and (b) do not hold for (non-separable) pre-Hil-
bert spaces; cf., for example. N. Bourbaki [2]. Chapter 5, §2. Exercise 2.

PROOF. Part (a) follows from part (b) by choosing Mo=0.

(b) Let Wl be the set of all those ONS which contain Mo. un is partially
ordered by the inclusion "c" (i.e., we have M C M for all ME un; from
M] C M2 , M2 C M3 , it follows that M] C M3 ; from MI C M2 , M2 C M] it
follows that M] = M2). If 91 is a linearly ordered subset of un (i.e., for
M], M2 E 91 we have M] C M2 or M2 eM]), then 91 has an upper bound
ME Wl (Le., for every M' E 91 we have M' eM); for the upper bound
M we may take the union of all N E91.
3.3 Existence of orthonormal bases, dimension of a Hilbert space 45

This M is an ONS: If fl' f2 E M, then there exist MI, M2 E 9l such

that fl E MI, f2 E M2. Since MI C M2 or M2 C MI holds, we have I" 12 E
M2 or II' 12 E MI' Therefore 11.1./2'
As M contains all N E 9l, M is an upper bound of 9l. By Zorn's
lemma this implies the existence of at least one maximal element
Mmax E IDl (i.e., for each MEIDl such that Mmax C M we have Mmax
=M).
This Mmax is an ONB: If we had L(Mmaxh'~ H, then, as H is a Hilbert
space, there would be (cf. part (b) of the proposition preceding Theo-
rem 3.3) an I E L(Mmax).l such that IIfll = I, i.e., Mmax U {f} would be
an ONS such that Mmax C Mmax U {f} and Mmax =t= Mmax U {f}; this
would contradict the maximality of Mmax. The requirement Mo C Mmax
is obviously satisfied.
(c) Let M" M2 be ONB of H. If IMII == m < 00 (we write IMI for the
cardinality of M), then by the proposition preceding Theorem 3.10, the
space H is separable and by Theorem 3.9(d) we have dim H= m =
IM2 1·
Now let IMd > INI· For each f E MI let K(f) = {g E M2 : g, < I> :f=
OJ. By Theorem 3.6(b) we have IK(f)1 ~ INI for all I E MI' We have
U {K(f) : I E M.} = M,. since if g E M2 \ U {K(f) : f E Md we would
have g.1. M" therefore g = 0 (as MI is total); however, this is impossible
because all elements of M2 have norm 1. Consequently, it follows that

IM21 ~ ~ IK(f)1 ~ 1M" INI ~ IMII·

JEMI

On the other hand, IM21 > INI is true, for otherwise MI would be finite,
also. We can therefore prove that 1M" ~ IM21 in the same way. 0

The algebraic dimension of a vector space is the cardinality of a maximal

set of linearly independent elements (algebraic basis). In Hilbert spaces it is
useful to introduce another notion of dimension. The dimension (more
precisely, the Hilbert space dimension) of a Hilbert space H is the cardinal-
ity of an ONB of H. By Theorem 3.1O(c) this dimension does not depend
on the choice of the ONB. By Theorem 3.9(d) for finite dimensional
Hilbert spaces the two definitions of dimension coincide; for infinite
dimensional spaces this is not the case, cf. Exercise 3.8.

Proposition. There exist Hilbert spaces 01 arbitrary (Hilbert space) dimen-

sion.
PROOF. Let /( be an arbitrary cardinal num\:Jer, and let A be a set of
cardinality /(. Let liA) be the Hilbert space 12(A; p.) with p.(a) = 1 for all
0: E A (cf. Exercise 1.10 and 2.2(b». The dimension of i 2 (A) equals /( = IAI,
as M = {f" : a E A}, where j,,( {3) = 8,,{3' is an ONB. 0
46 3 Orthogonality

Theorem 3.11.1J H is a Hilbert space and Sand T are closed subspaces oj H

such that S n T.L = {OJ, then we have dim S ~ dim T (dim = Hilbert space
dimension ).
PROOF. Let us distinguish between two different cases.
(a) dim T = k < 00: Assume that dim S > dim T holds. If {e), ... , ek } is
an ONB of T and {fr, ... ,Jk+ d is an ONS in S, then the system of
homogeneous equations (k equations, k + 1 unknowns)
k+1
~ c/em,h) = 0, m = I, ... , k
i-I

has a non-trivial solution. Therefore there exists a non-zero element

J = ~J!: r'cih of S n
T.L, which contradicts the assumption.
(b) dim T ;;.1"'1. Let M) and M2 be orthonormal bases of T and S,
respectively. For each e EM) let K(e) = {f E M2 : (e, f) ~O}. We have
UeEM,K(e)= M2, because forJE M2\ UeEM,K(e) we would haveJ...LM"
thus J...L T; which would contradict the assumption. Since for each
e EM) the set K(e) is at most countable, it follows that IM21 <;
l"'IIMr! = IMr!. 0

EXERCISES

3.8. Let H be a Hilbert space and let A be a countable subset of H such that
L(A) = H. Then H is finite dimensional, i.e., no Hilbert space of algebraic
dimension INI exists.
Hint: Apply the Schmidt orthogonalization process to A; for the resulting
ONB M we have L(M) = H. (It can actually be proved that no infinite
dimensional Hilbert space can have an algebraic dimension smaller than the
cardinality of the continuum; cf. N. Bourbaki [2], Chap. 5, §2, Exercise 1.)
3.9. (a) Let (H, (., .» be a pre-Hilbert space. For any n elementsf" ... ,fn of H
the Gram determinant is defined by DUI,' .. ,jn)"'det «Jj,A». We
have DUI' ... ,fn) ;;. 0; the equality sign holds if and only if the elements
fl' ... ,fn are linearly dependent (in the case n = 2 this is Schwarz'
inequality.)
Hint: Use induction on n. In going from n - I to n use the fact that the
value of the determinant does not change if the first column is replaced
by (Jj,j) - PnUI», where Pig) denotes the orthogonal projection of g
onto LU2, ... , fn)·
(b) Prove the same assertion by using the fact that the matrix
«h, fk) )j. k _ I, ... , n is the product of the matrices
«Jj,e/»j_I, ... ,n and «e/,A»/_I, ... ,m,
I-I, ... ,m k-I, ... ,n

where {el' ... , em} is an ONB of LUI>" . ,fn}'

(c) Prove an analogous statement for semi-scalar products.
3.4 Tensor products of Hilbert spaces 47

3.10. Part (c) of the proposition preceding Example I of Section 3.2 does not hold
in pre-Hilbert spaces.
Hint: In 12 let f= (1/ n), H = 12• 0 , HI = 12•0 n {f) 1.. By Exercise 3.4 we have
HI - {f) J.. If M is an ONB of H, then M is a maximal ONS in 12,0, without
being an ONB in 12 ,°'

3.4 Tensor products of Hilbert spaces

Let HI and H2 be vector spaces over K We denote by F(HI' Hz) the vector
space of formal linear combinations of the pairs (f, g) with f E HI' g E H2•
i.e.,

F(HI' Hz) = { .f cifp~) : cj

)-1
E 'K.,h E HI' gj E H2,} = I, 2, ... ,n; n E r\I}.

The quotient space

HI ® H2 = F(HI' H2)/ N
is called the algebraic tensor product of HI and Hz.
The product HI X H2 can be considered 'as a subset of F( H.. H2 ), if one
identifies (f, g) E HI X H2 with 1(f, g) E F(H .. H2)' The equivalence class
from HI ® Hz defined by (f, g) will be denoted by f® g; these elements are
called simple tensors. Each element of HI ® Hz is representable as a finite
linear combination of simple tensors. Such a linear combination of simple
tensors is equal to zero if and only if it is a finite linear combination of
elements of the form

f f
j-Ik=1
ajb~ ® gk - ( f
j=1
a;t) ® ( f
k~1
bkgk ). (3.2)

In particular, we have

(f
)-1
a)j) ® ( ~
k-I
bkgk ) = f ~
)-1 k-I
ajbk./.i ® gk' (3.3)

If (H .. (. , ')1) and (H2, (. , .)2) are Hilbert spaces over IK, then

s( f
)-1
cih. g), f
k-I
c"Uk' gk») =
)-1
f ctc" (/.i,fDI(gp gk>z
f k-I
defines a sesquilinear form on F(H., H2)' For arbitrary fEN and g E
48 3 Orthogonality

F(H., H2 ) we have s(f, g)= s(g,f)=O, as one can verify by simple calcula-
.tion. Consequently, by

/
\)-1
~ cj/i9 gj' ~ c;J//iJ gk)
)=1
= s( ~
)-1
cifj, g), f
k-I
Ck(j~, gk»)
a sesquilinear form is defined on HI ® H2•
We show that <. , .) is a scalar product on HI ® Hz. In order to prove
this it is enough to show that (f, I) > 0 holds for all I E HI ® Hz, 1"1= O.
Indeed, let I = ~j-l c)fj ® ~ "1= O. If {ed and {ek } are orthonormal bases of
LUI"" ,J,,} and L{gl"'" gn}' respectively, then

1= '2, ck1ek ® e; with Ckl = "2»<ek,fj)<e;, ~), (3.4)

k. I )

and thus
(f,j) = ~ ICk/ 12 > O.
k,l

Therefore (HI ® Hz, (. , .»

is a pre-Hilbert space. The completion of this
pre-Hilbert space (cf. Section 4.3) will be denoted by HI ® H2 and called
the (complete) tensor product of the Hilbert spaces HI and H2.
From (3.4) for each/=~j_lc.IJ®~EH(®H2 it follows by means of
(3.3) that

(3.5)

where {ed and {el} are orthonormal systems in LUI"" ,jn} and
L{ g(, ... ,gn}' respectively; the elements If and gk are contained in
LUI' ... .In} and L{ gl' ... , gn}, respectively.

EXAMPLE I. Let PI and P2 be measures on R, and let HI = L 2(R, PI)'

Hz = L2(JR, pz). By (3.1) an element ~j-Iclfj,
g) from F(H(, Hz) is in N if
and only if the function
n
(x,y) ~ L c)fj(x)~(y), (x,y) E R2
)-1

vanishes almost everywhere with respect to the product measure PI X P2'

The algebraic tensor product HI ® H2 is thus composed of equivalence
classes of functions, square integrable on JR2 with respect to PI X P2' For
I, g E HI ® H2 we have
(I, g) = f I(x,y)*g(x,y) dpI(x) dp2(y)·
As HI ® H2 obviously contains all step functions on R2, the space HI ®H2
is isomorphic to L2(R 2, PI X P2)'
3.4 Tensor products of Hilbert spaces 49

lbeorem 3.11. Let HI and Hz be Hilbert spaces.

(a) If MI and Mz are total subsets of HI and Hz, respectively, then the set
U ® g : f E M» g E Mz} is total in HI ®Hz.
(b) If {ea : a E A} and Ull : {J E B} are orthonormal bases of HI and H2 ,
resp"ectively, then {eer ® f/1 : a E A, fJ E 8) ;s an orthonormal basis of
H,®Hz·
PROOF.
(a) Let ~j-lh®8j E HI ® H2 , E >0. For eachj E {I, 2, ... , n} there exist
elements h' E L(MI ) and gi E L(M2) such that IIjj - 1;'11 IIBjIl <e/2n and
118j - gill IIjjll <E/2n. Then we have
1I..fj ® 8j - jj' ® gjll = lIet - jj') ® gj + ..Ii' ® (8j - ~)II < e/ n
and consequently

Because
n
~..fj' ® g; E L(M,) ®L(M2) =L{J®g: fEMt>gEM2},
j-I

the assertion is proved, since HI ® ~ is dense in HI ®H2 •

(b) By part (a) the set {ea ®fp : a E A, fJ E B} is total in HI ® H2 • More-
over, we have
<e.. ® fll' ea' ® ffJ' >= 8a",,8fJ{J' for all a E A, fJ E B,
i.e., {e.. ® fp : a E A, P E B} is an orthonormal basis. o
EXERCISES

3.11. Two non-zero tensors!1 ®gl andfl®gz are equal to each other if and only if
there exists acE K, c oF 0 that satisfies!2 == ef" g2 =C-Igl •
3.12. Let HI and H2 be Hilbert spaces.
(a) We have dim [HI ~ H2 ] .. (dim HI)(dim H2) (Hilbert space dimensions).
(b) If HI and Hz are different from {O}, then HI ~ H2 is separable if and only
if HI and Hz are both separable.
(c) (HI ® H2 , <. , .»
is complete if and only if HI or H2 is finite dimensional.
4
Linear operators
and their adjoints

4.1 Basic notions

Let HI and H2 be vector spaces over II<. A linear operator T from HI into
H2 is. by definition. a linear mapping of a subspace DCT) of HI into H2•
The subspace DCT) is called the domain of T. The image RCT) = TCDCT»
={Tf : f E DC T)} is called the range of T. Since we only treat linear
operators here. we shall speak only about operators from HI into H2 • If
HI = H2 = H. then T is called an operator on H. A linear operator from H
into II( is called a linear functional. The range of an operator T from HI
into H2 is a subspace of H2• An operator is injective if and only if Tf = 0
implies f = O. In this case the inverse T - I of T is defined by
D(T- I) = R(T). T-Ig = f for g = Tf E R(T).
T- I is a (linear) operator from H2 into HI' For an operator T from HI into
H2 and for a E II( the operator aT is defined by
D( aT) == D( T) and (aT) ... a( Tj) for f E D( aT).
For two operators Sand T from HI into H2 the sum S + T is defined by
D(S + T) = D(S) n D(T). (S + T)f = Sf + Tf for f E D(S + T).
If T is an operator from HI into H2 and S is an operator from H2 into H3•
then the product ST is defined by
D(ST) = {JED(T): TfED(S)},(ST)f= S(Tf) for fED(ST).
If D is a subspace of HI. then the set of those operators from HI into H2
whose domain is D is a vector space over 11(; the zero element is the
operator whose domain is D and which sends all elements of D to O. Let S

50
4.) Basic notions 51

and T be operators from HI into H2 • An operator T is called an extension

of S (or S is a restriction of T) if we have
O(S) cD(T) and Sf = Tf for f E O(S).
For this we write SeT or T:J S.

EXAMPLE l. Let M be a measurable subset of Am and let t : M-+C be a

measurable function on M. The maximal operator of multiplication by t on
L 2( M) is defined by
O(T) = {j E L2(M) : tf E L2(M)}, Tf = if for f E D{T).
The set D(T) is obviously a subspace of L2(M) and T is an operator on
L2(M).

(4.1) OCT) is dense.

PROOF. For each n E '" let M,. = {x EM: !t(x)! <: n}. ThenM,. C M,,+ I and
U :_1 M" .. M. For each f E L2( M) the function f,,'" XM"f belongs to
D( T) and we have f" -+f. 0
(4.2) The following statements are equivalent:
(a) R(T) is dense,
(b) t(x):F 0 almost everywhere in M,
(c) T is injective.
If one of these assumptions is satisfied, then T - I is the multiplication
operator defined by the function

for x EM such that t(x):FO

for x EM such that t(x) .. O.

PROOF. (b) follows from (a): Each f E L2(M), that vanishes outside the set
MI == {x EM: t(x) = OJ, is orthogonal to R(T). Therefore L2(M I ) = {OJ,
i.e., MI has measure O.
(a) follows from (b): Let M" = {x EM: !/(x)l;;. lin}; then M,. C M,.+I
and M \ U :"1 M,. has measure zero. For every g E L2( M) the function
gIl - XM g belongs to R( T) and gIl -+ g holds.
(c) f~l/ows from (b): If Tf= 0, then t(x)f(x) -= 0 almost everywhere in M.
Therefore f(x) ... 0 almost everywhere, too, and thus f= O.
(b) follows from (c): If MI = {x EM: I(x) = OJ, then for all f E L2(M),
vanishing outside MI, we have Tf= 0; therefore f= O. From this it follows
that MI has measure zero.
If one of the above conditions is satisfied, then T is injective and l(x):FO
almost everywhere in M. Hence we have
D(T- I ) ==R(T) = {gEL2(M): there exists anfEL2(M) such thatg= if}
... {gELiM): IlgEL2(M)},
52 4 Linear operators and their adjoinls

and for g E O(T -I) and f E L2(M) such that g = tf we have

T-Ig = f = tlg;
consequently T - I is the multiplication operator induced by t I' 0
=
(4.3) We have OCT) L2(M) if and only if a C exists for which I/(x)1 "C
almost everywhere in M. We have R( T) = L2( M) if and only if a c > 0 exists
for which I/(x)1 ;) c almost everywhere in M.

PROOF. If I/(x)I" C almost everywhere, then If E L2(M) for all f E L2(M).

Therefore OCT) = L2(M). Conversely, let OCT) = L 2(M). Let us assume that
no C exists for which I/(x)l..; C almost everywhere. For n E N set Mn = {x
EM: I/(x)l>n}, Nn=Mn-I\Mn with Mo=M. Then all Mn have positive
measures, and the intersection n:'_ I Mn has measure zero. Therefore there
exists a subsequence (nk ) of N such that all N". have positive measures. We
have I/(x)l;) nk - 1 for x E Nnk . For all kEN let us choose fk E L2(M) in
such a way thatfk vanish outside Nn• and II All = 11k. Since the functions
fk are mutually orthogonal, we have
00

f = ~ AE L2(M)·
k-I

However, If is not in L2(M), i.e., f f/. OCT), this contradicts the fact that
O( T) = L2( M). (A simpler proof of this can be found in Exercise 5.5.)
If I/(x)1 ;) c > 0 almost everywhere, then T - I exists, and for the inducing
function II we have 1/1(x)1 <':c- I almost everywhere. Therefore R(T)
= O(T- I )= L2(M). If R(T) = L2(M), then by (4.2) the operator T is
injective and O(T- I )= L2(M). For the inducing function II we have
1/1(x)l..; C almost everywhere, thus I/(x)1 ;) C - I almost everywhere. 0
If t, s are measurable functions on M and T, S are the multiplication
operators induced by I, s, then T+ S is a restriction of the multiplication
operator induced by t + s, as from tfE L2(M) and sfE L 2(M) it obviously
follows that (/+ s)fE L2(M).

EXAMPLE 2. If '" : R~C is continuous, then by

Tf = J",(x)f(x) dx,f E L2• 0 (R)

a linear functional is defined with O( T) = L2• o(R). If '" E L2(R), then T can
be defined on the whole space L2(R).

The subset N(T) = (fE D(T) : Tf=O} is called the kernel of T. For
every operator T from HI into H2 the set N(T) is a subspace of OCT)
(consequently of HI> as well).
4.1 Basic notions 53

Theorem 4.1. Let be vector over and I_ ••• _ T

linear functionals such that O(T I)= ... = O(Tn) = D(T)= H. If we have
N(T) 7_IN(~), then exist ... , II( such that

PROOF. We prove this by induction on n. be to TI =

then N( T) :J N( T I) = H, therefore T = O. If TI:F 0, then there exists an
fo E H such that TI fo = 1. For every f E H we then have f - T\U)fo N(T 1)
c N(T). Consequently, TUo)TI i.e., TUo)T 1•
Let us now assume that the assertion is true for n - 1 (n > 1). Assume
that N(T):Jn7_IN(1';). If N(Tn):Jn7::\IN(1';), then it follows that
N(T) 7::\1 Therefore by the induction hypothesis we have
T=~7: 1';. So this the assertion holds. If N(Tn);zj i_l N(1';),
then there exists anfoE n7::lN(1';) such that TJo= 1. Let
T(Jo)
For allfE n7::1 N(T;) we have
1;(J - Tf/(J)fo) Ti(J) 1~(J) =0 = 0, 1,2, .. , n-
and

Le.,

f - Tn(J)fo E n N(Ti) cN(T),

i= I
and
To(J) = T(J) - T(JO)Tn(J) = T(J- Tn(J)fo) = O.
Therefore we have

the induction hypothesis it follows from that

n-
T - T(JO)Tn = To = ~ ciTi,
i= \

Le., T linear combination of Tp , .. , T".

Let HI and H2 now be normed spaces. An operator T from HI into Hl is

said to be continuous the f E O( if for every sequence U,,) from
O(T) such that fn~f we have Tf. The operator T is continuous by
definition if T is continuous at each point of OCT). The operator T is said
to bounded if exists C > such that II Tfll Cllfll all
f T). such called bound T.
54 4 Linear operators and their adjoints

Theorem 4.2. Let T be an operator from HI into H2 • Then the following

assertions are equivalent:
(a) T is continuous,
(b) T is continuous at 0,
(c) T is bounded.
PROOF. (b) obviously follows from (a).
(c) follows from (b): Let us assume \hat T is not bounded. Then for every
n E fill there exists an 1" E D(T) such that II Tf,,11 >nll1"lI. From this it
follows that, in particular, fn "I=- 0; without loss of generality we may assume
that IIf,,1I = lin. Consequently we have fn~O and II Tfnll >n(lln)= 1,
which contradicts the continuity of T at zero.
(a) follows from (c): Assume II Tfll -< Cllfll for all f E OCT). If f E OCT)
and Un) is a sequence from OCT) such thatfn~f, then we have
II Tf n - Tfll = IIT(jn-j)1I -< Cllfn-fll~O.

i.e., Tf,,~ Tf, which proves the continuity of T. D

For a bounded operator T from HI into H2 the norm II Til is defined by
IITII = inf {C;>O: IITfil <Cllfll forall fEO(T)} (4.4)
(in Section 4.2 we shall justify the word "norm"). Since for every 10 > 0 we
have
II Tfll -< (II Til + e)IIfil for all f E D(T),
the norm II Til is a bound for T, thus
II Tfll < II TIl IIfil for all f E D(T). (4.5)
EXAMPLE I (Continued). A function s : M~~ is said to be essentially
bounded from above if there exists aCE ~ such that sex) < C almost
everywhere in M. Each C of this kind is called an essential upper bound of
s. The greatest lower bound of all essential upper bounds is called the
essential supremum of s, in symbols ess sup s. It is itself an essential upper
bound for s. Indeed, if Co denotes this greatest lower bound, then for every
n E fill the number Co + (1 In) is an essential upper bound, i.e., sex) - Co -
(II n)" 0 holds almost everywhere. By letting n~oo it follows that s(x)-
Co" 0 almost everywhere. Analogously, we may define the concepts of
essentially bounded from below, essential lower bound, and essential infimum.
A complex-valued function s is said to be essentially bounded, if lsi is
essentially bounded from above.
(4.6) The operator T from Example I is bounded if and only if t is
essentially bounded. We have II Til = ess sup Itl.
PROOF. If t is essentially bounded, and C = ess sup Itl, then we have
I/(x)1 " C almost everywhere. Therefore
4.1 Basic notions ss
i.e., T is bounded and 1/ TI/ <; c. If C = 0, then we have 1/ TI/ = O. If C > 0,
then for every € E (0, C) the set M. = {x EM: II(x)1 ;;. C - €} has a positive
measure and for all f E L2 ( M), that vanish outside M., we have

1/ Tfll2 = f.lt(x)f(xW dx ;;. (C -

M
fl f.
M,
/f(x)J2 dx = (9 - £)2I1fI12.

Therefore II Til ;;. C - E, and thus II Til = c.

If t is not essentially bounded, then for every n E N the set Mn = {x E
M : I(X);;' n} has a positive measure and for every f E OCT), that vanishes
outside Mm we have
IITJII ;;. nllJII·
Therefore T is not bounded. o
EXAMPLE 2 (Continued). The functional T of Example 2 is bounded if
1/1 E Lz(R), since then
ITfl < I/I/II/lifi/·
From Theorem 4.8 (theorem of Riesz) it will follow that T is continuous if
and only if 1/1 E Lz(R).

EXAMPLE 3. Let M, and M2 be measurable subsets of RP and R m, respec-

tively. Then Mz x M, is a measurable subset of Rm+p. The points of Rm+p
can be written in the form (x, y) with x E R m, y E RP. Assume k E L2(M2 x
M,). By Fubini's theorem

f. Ik(x,y)I Z dy < 00 almost everywhere in Mz,

MI

i.e., we have k(x, .) E L2(M.) almost everywhere in M1 • Consequently, for

all f E L2( M,) we can define

(KJ)(x) = f. k(x,y)f{y) dy almost everywhere in Ml ·

MI
Then we have
1/1
I(Kf)(x)1 < IIfll { fMllk(x,Y)ll dy } . (4.7)

For every g E Ll (M2 ) the function h defined by h(x,y) = k(x,y)f(y)g(x) is

integrable on M2 X MI' Therefore by Fubini's theorem the function

g(x)(Kf)(x) = g(x) f.
MI
k(x,y)f(y) dy = f. h(x,y) dy
MI

is a measurable function on M2. If we put M2•n = {x E M2 : Ixl < n} and

6" = XM , then we can see that XM2.n KJ is measurable for every n E N.
2,11
Consequently, Kj is measurable.
56 4 Linear operators and their adjoints

Because of (4.7) we have Kf E L2(M2) and

f. f.
1/2
IIKfll <: {
Ik(x,yW dy dx } Ilfll·
M2 MI

Since the mapping f ~ Kf is obviously linear, we have defined a continu-

ous operator K from L2(M I ) into L2(M2) such that D(K) = L2(M I ) and

f. f. Ik(x,y)1
1/2
IIKII <: { 2 dy dx }
M2 MI

Such an operator is called a Hilbert-Schmidt operator (cf. also Section 6.2).

EXERCISES

4.1. (a) The reasoning of Example 1 can be carried out completely analogously if
M is replaced by an arbitrary a-finite measure space (X, IB, IL). (A measure
space (X, lB, IL) is said to be a-finite if X can be written as the union of
countably many subsets of finite measure.)
(b) In Example 3 replace MI and M2 by two arbitrary a-finite measure spaces
(XI' lBl' ILl) and (X2• lB 2• IL0.
Hint: Observe that the set X2,o={xEX2 : flk(x.y)12dlLl(Y»O} is the
union of the countably many sets X2, n = {x E X2 : f Ik(x. y)1 2 dILI(Y);;'
1/ n} (n E 1\1) with finite measures, and XX2, " E Ll X2 , lB 2 • ILl).
4.2. Let (XI. lBl' MI) and (X2• !H2' IL2) be a-finite measure spaces. Let k : X2 X XI~
C be ILl X MI-measurable and let k(x • .) E LiXI • lB l • ILl) almost everywhere in
X2 • Then for each f E LiXI , !HI' MI) the function Kof defined by

Kof(x) = fXI
k(x,y)f(y) dMI(Y)

is ILl-measurable and we have IKof(x)1 <; IIk(x, .)11 IIfli. By

D( K) = (J E L2(X" !H., MI) : Kof E LiX2, !H2. IL2)}
Kf = Kof for f E D(K)
an operator is defined from L2(X I , !HI. ILl) into L2(X2 , !H2. M2) (such an operator
is called a Carleman operator; cf. also Section 6.2). There are functions k of
this kind for which D(K) = to}.

4.2 Bounded linear operators and functionals

Theorem 4.3. Let HI and H; be normed spaces. Let T be an operator from HI
into H2•
(a) We have
sup {II Tfll : fE D(T), IIfll <: l} = sup {II Till: fE D(T). Ilfll = I}
= sup {IITfll :fED(T), IIfll < l}
4.2 Bounded linear operators and functionals 57

(where the value 00 is allowed). T is bounded if and only if one of these

values is finite; if one is finite then the others are finite, also, and they are
equal to II Til.
(b) If H2 is a pre-Hilbert space and M is a subspace of H2 such that
R( T) c M, then T is bounded if and only if
sup (I<Tf, g)1 : f E D(T), gEM, IIfll = II gil = I}
is finite. This number is then equal to II Til.
(c) If T is an operator on a pre-Hilbert space Hand D(T) is dense, then Tis
bounded if and only if
sup {I<Tf, g)1 : f. g ED(T), IIfll = II gil = l}
is finite. This number is then equal to II Til.
PROOF.
(a) If we define II Til to be equal to 00 if T is unbounded, then we only
have to prove that the three values, which we denote by c l ' C2' C3'
respectively, are all equal to II Til. As II Til is a bound for T, we surely
have II Til> c l • The inequality c i ~ C2 is obvious. If f = 0, then II Tfll =0
EO;c 2 • If 0< IIfll < 1, then with g= Ilfll-~ we have

II gil = 1 and II Tgil = IIfll-IIITfll ~ II Tfll·

Consequently C2 ~ C3 . What remains is to prove that C3 ~ II Til. This is
evident if II Til = O. Therefore suppose II Til> O. For every £ E (0, 1)
there exists an f E D(T) such that II Tfll > (1- ()II Til Ilfll. Hence /76 0
and for g =[(1 + £)llflir ~ we have
1- £
IIgll < 1 and II Tgil ~ -1-II T II·
+£
From this it follows that C3 > (1- £)/(1 + £)11 Til for all £E (0, I) and
thus c3 > II Til·
(b) By (a) we have IITII=sup (IITfll :fED(T), IIfll=l}. On the other
hand,

IITfll = sup {1<Tf, g)1 : gEM, II gil = I}.

Indeed, if Tf= 0, this is obvious. If Tf=l=O, then II Tfll ~ 11< TJ, g)1 for all
gEM such that II gil = 1, and there exists a sequence (gn) from M such
that II gnll = I and gn~1I Tfll-IT!, therefore <Tf, gn)~11 Tfll. These
arguments together give the assertion.
(c) follows from (b) if we choose M= D(T). 0

An operator T is said to be densely defined if OCT) is dense. An operator T,

which is densely defined on a pre-Hilbert space H, is said to be symmetric
if for allj, gE D(T) we have <Tj, g) = <1, Tg).
58 4 Linear operators and their adjoints

Theorem 4.4. Let T be a densely defined operator on a complex pre-Hilbert

space, or a symmetric operator defined on an arbitrary pre-Hilbert space. Tis
bounded if and only if
e = sup {I(f, Tf)1 :fED(T), Ilfll CO; l} < 00.

If T is bounded, then we have

(a) II Til <: 2e, if H is a complex Hilbert space,
(b) II Til = e, if T;s symmetric.
PROOF. Let us set II Til == 00 for an unbounded T. Then we only have to
prove inequality (a) and equality (b).
(a) For all f, g E O(T) such that IIfll CO; 1 and II gil CO; 1, from (1.4) with
sU, g) == (f, Tg) and from (1.16) it follows that

2e ;> te{4I1fI1 2 +41IgI1 2 }

... te {11f+ g112+ Ilf- g112+ IIf+ igll 2 + IIf- igll 2}
;> tl(f+ g, T(j+ g» - (f - g, T(j- g» - i(f + ig, T(J+ ig»
+i(f- ig, T(J- ig»1 = I(f, Tg)l·
The assertion follows from this by Theorem 4.3.
(b) By (1.7) and (1.8) with sU, g) == (f, Tg), it follows for all f, g E O(T)
such that IIfll CO; 1 and II gil CO; 1 that
e ;>4e{lIfIl 2 +lIgI1 2 } =te{III+gIl 2 +llf-gIl 2 }
;> tl(f+ g, T(j+g» - (f- g, T(j-g»1 = IRe (f, Tg)l·
If we choose a E IK so that lal == 1 and a(f, T;g) = 1(1, Tg)1 hold, then
it follows (with h == a*f) that
I(f, Tg)J == (h, Tg) == IRe (h, Tg)1 CO; e.
By Theorem 4.3(c) it follows from this that II Til CO; e. The inequality
e CO; II Til is evident by Theorem 4.3(c). 0
Theorem 4.5. Let T be a bounded operator from a normed space HI into a
Banach space H2• Then there exists a unique bounded extension S of T such
that O(S)==O(T). We have IISII=IITII.
PROOF. Uniqueness: Assume S is a continuous extension of T such that
O(S) == O(T). If 1 E O(S), then there exists a sequence (j,,) from O(T) such
that J" -+ f. As S is continuous, we have Sf = lim Sf" ... lim TI", i.e., S is (if it
exists at all) determined by T uniquely.
Existence: Assume thatfE D(T) and U,,) is a sequence from O(T) such
that f,,-+I. Then (j,,) is a Cauchy sequence. Since T is bounded, the
sequence (TJ,,) is a Cauchy sequence, also, for we have II TIll - Tfmll CO;
II Til Ilf" - fmll· Therefore there exists agE H2 such that TIII-+g. This g is
independent of the choice of, the sequence (j,,) from O(T) with f,,-+f.
4.2 Bounded linear operators and functionals 59

Indeed, if (f~) is another sequence of this kind, then the sequence

(f,. f •• f2' f~ • . . . ) converges to f, also. Hence the sequence
(Tf,. Tfl' Tf2• Tf2 ... ) is also convergent; the limit has to be equal to g.
Consequently, (Tf~) tends to g. as well. Let us define Sf= g.
S is linear: Iff" f2 E D(T) and (f" ,,), (A,,) are sequences in D(T) such
thatf""-"fl,f2,,,-,,f2' then for all a, bEIK it follows that
S(af, + bf2) = lim T(afl, n + bA,,) = lim (aTf,. n + bTA n) = aSfl + bSf2•
S is bounded and II S II = II TIl: For f E D( T) and (],,) from D( T) such
that]"-,,fwe have 1I]"1I-"lIfll and

IISfll = lim II Tfnll ~ lim IIT1l1lfn11 = IITlillfll·

Therefore II S II ~ II TIl. As II S II > II I'll obviously holds, the assertion

~~ 0
The set of those bounded operators from H, into H2 , whose domain is
HI, will be denoted by B(H" H2). By Section 4.1 the set B(HI' H2) is a
vector space.

Theorem 4.6. Let II . II be defined as in (4.4). Then (8(H" H2), II . II) is a

normed space. If H2 ;s a Banach space, then (8(H" H2). II . II) ;s a Banach
space, too.

PROOF. It is clear that II . II is a semi-norm. If II Til = 0, then II Tfll = 0 for all

f E HI such that II fll '" I; therefore Tf = 0 for all f E H .. and thus T = 0,
the zero element in 8(H" H2). Consequently, II . II is a norm. Assume now
that H2 is a Banach space. If (Tn) is a Cauchy sequence in 8( HI> H2 ), then
for every f E HI the sequence (Tn!) is a Cauchy sequence and, conse-
quently, a convergent sequence. Let us define: Tf=lim T"J. Then T is
linear, because for f, g E HI and a, b E IK we have .

T(af+bg) = lim Tn(af+bg) = lim (aT"i+bT"g) = aTf+ bTg.

As (Tn) is a Cauchy sequence, (II Tn I!) is convergent, say II T"II-"C. For all
fE HJ we have
II Tfll = lim II T"ill ~ lim II Tnllllfil = Cllfll,
i.e., T E 8(H" H2)' What remains is to prove that Tn-" T. For every e > 0
there exists an n(t)EN such that IITn- Tmll ~t for n, m >n(e). Therefore,
for n ;;. n( E) and for all f E HI we have

i.e., II Tn - Til '" E for n ;;. n( £). Hence Tn -" T. o

60 4 Linear operators and their adjoints

Theorem 4.7.
(a) Let HI and H2 be normed spaces. If T E B(H., H2), then N(T) is a closed
subspace of HI'
(b) Let T be a linear functional on a Hilbert space H such that OCT) = H.
Then T is continuous if and only if N( T) is closed.
PROOF.
(a) LetfE N(T). Then there exists a sequence Un) from HI such thatfn~f
and TJ,,= 0. Since T is continuous, it follows that Tf= lim Tfn= 0, i.e.,
fE N(T).
(b) If T is continuous, then N(T) is closed by part (a). Let N(T) now be
closed. If N( T) = H, then T = 0; consequently T is continuous. If
*
N(T)* H, then N(T).!. {OJ. Therefore there exists agE N(T).!. such
that g*O. Because <g,j) =0 for allfE N(T), we have N(Tg)~ N(T)
<
for the functional TJ = g, f). By Theorem 4.1 this implies that
T = cTg with some cEil<. Consequently T is continuous. (The proof
can also be carried out analogously to the second part of the proof of
Theorem 4.8, without using Theorem 4.1.) 0

REMARK. If HI and H2 are pre-Hilbert spaces, one may expect that

(B(HI' H2), II . II) is a pre-Hilbert space, also, i.e., the norm is induced by a
scalar product. This holds true if HI = II< or H2 = II<. However, this is not
the case if dim HI ;> 2 and dim H2 ;> 2; cf. Exercise 4.3.

For T E B(HI' H2) and S E B(H2' H3) the product ST is in B(H" H3 ),

since we have
O(ST) = {JEHI : Tf EO(S)=H2} =Hl
and
IISTfll" IISIlIITfll" liS II II TIl IIfll foraH fEH.,
i.e., we have ST E B(H .. H3 ) and

IISTII " IISIlIITII· (4.8)

We write B(H) for B(H, H). For S, T I, T2 E B(H) we have
S(T, + T2) = ST I + ST2, (TI + T 2)S = TIS + T 2S.
The operator I with 0(/) = H and If = f for all f E H obviously belongs to
B(H) and we have
11111 = 1, (4.9)
and IT= TI = T for all T E B(H). The operator I is called the identity
operator on H. The set B(H) is thus an algebra with an identity element. As
the norm on B(H) satisfies relations (4.8) and (4.9), the algebra B(H) is a
normed algebra with an identity element. If H is a Banach space, then B(H)
is complete; we call B(H) a Banach algebra (with an identity element).
4.2 Bounded linear operators and functionals 61

Now we shall study the set of continuous linear functionals defined on a

Hilbert space H, i.e., the set 8(H, K) for a Hilbert space Hover IK. Every
element g E H defines a linear functional Tg such that D(Tg) = H by the
formula
Tg(f) = <g,!).

As ITif)1 " II g/l 11f/l, the functional Tg is bounded, and /I Tg/l " II gil. Since
for!= g we have ITig)1 = /I g1l2, it follows that /I Tgli = /I g/l. Every continu-
ous linear functional defined on H is actually of this form.

Theorem 4.8 (F. Riesz). Let H be a Hilbert space. Every g E H induces a

continuous linear functional on H by Tg(f) = (g,j). We have 1/ Tgil = /I gil.
This mapping of H onto 8(H, IK) is bijective and antilinear, i.e., we have
Tag +bh = a*Tg + b*Th·
PROOF. The first part has already been shown. The antilinearity follows
from
Tag+bh(f) = (ag + bh,J) = a*<g,f) + b*<h,f) = a*Tg(f) + b*Th(f).
As II Tgil = II gil, the mapping g~ Tg is injective. What remains is to prove
that it is also surjective. Let T E 8(H, 11<); we construct agE H such that
T= Tg.
*"
If T = 0, then we can choose g = O. If T 0, then the kernel N( T) = {f E
H : T(f) = O} is a closed subspace of H, different from H, i.e., N(T).L *"
{OJ. Let g E N(T).L such that II gil = 1. Let a = T(g). For every! E H we
obviously have T(f)g - T(g)! E N(T). Therefore T(f)g - T(g)f is or-
thogonal to g, i.e.,
0= <g, T(f)g - T(g)f) = T(f) - a<g,j),
and thus
T(f) = T(g)(g,f) = (a*g,f).
Consequently, we have T= Tao g. ~e can also prove this last part with the
aid of Theorem 4.1, cf. the proof of Theorem 4.7(b).) 0
EXAMPLE I (Continuation of Example 2 from Section 4.1). We can now
show that the continuous function I{I : R~C induces a continuous func-
tional on L2• o(R) by T(f) = Jl{I(x)f (x) dx if and only if t/l E L2(R). If Tis
continuous, then it can be extended uniquely to a continuous functional on
L 2(R) (which we denote by T, as well). By Theorem 4.8 we have T= Tg
with some g E L2(R). Therefore, we have for all f E L2• o(R) that

0= T(f) - Tg(f) = f (l{I(x) - g*(x»f(x) dx.

For an arbitrary n E N let us define

fn{x) = {r(X) - g(x) for

for
Ixl" n,
Ixl >n.
62 4 Linear operators and their adjoints

Then for all n it follows that

hence

Jlt¥(x) - g*(xW dx = O.

From this we can infer that t¥ = g* E LilR).

EXERCISES

4.3. (a) If H is a Hilbert space, then the norm in B(H, 11<) is defined by a scalar
product (i.e., B(H, 11<) is a Hilbert space).
Hint: If T I • T2 are the functionals induced by gl. g2 E H, then let <T I , T 2 )
=<g2' gl)'
(b) If HI and H2 are (pre-) Hilbert spaces, then the norm in B(H I • H2) is
induced by a scalar product if and only if dim HI = 1 or dim H2 = 1.
Hint: If dim HI ;;. 2 and dim Hl ;;. 2. then let fl' f2 E HI' gl' g2 E H2 be such
that <It,b) = <g" g,) = 8ij and T;f=<b.!)gj for fE HI,j= 1, 2. For these
two operators the parallelogram identity does not hold.
4.4. For each Xo E [0, 1) there exists exactly one g E W2• 1(0, 1) such that for all
f E W2• 1(0, 1) we have
f(xo) = 10 1(g*(x)f(x) + g'*(x)f'(x)} dx
(cf. Exercise 2.3(b».
Hint: The functional Tf= f(xo} is continuous on W2• 1(0, 1).
4.5. The set of Hilbert-Schmidt operators on L2( M) (cf. Section 4.1. Example 3) is
a sub-algebra of B(L2(M». It is a Banach algebra with the Hilbert-Schmidt
norm

IIIKIlI = {tfM1k(x,Y)12 dx dY } 1/2.

Hint: If the Hilbert-Schmidt operators K and H are induced by the kernels k

and h, then L= HK is induced by the kernel

l(x.y) = th(x. z)k(z,y) dz.

4.6. (a) Theorem 4.4 does not hold true for non-symmetric operators in real (pre-)
Hilbert spaces. There are non-vanishing operators T such that the
quadratic form qr(f) = <f, Tf) vanishes on D(T).
(b) Show that the constant 2C in Theorem 4.4(a) is optima\.
(c) In a complex (pre-) Hilbert space H the quantity II Tllq = sup {I<i. Tf)1 : f
E D(T). IIfll " I} is a norm on B(H). We have II Til " 211 Tllq " 211 Til.
4.7. Let T be a bounded operator from a Hilbert space HI into a Banach space H2 •
Then there exists an extension S E B(H I , H2) of T such that IISII = II Til.
Hint: Define Sf= 0 for f E D(T).1.
4.3 Isomorphisms, completion 63

4.8. If H is an infinite dimensional Hilbert space, then B(H) is not separable.

Hint: Let {en : n EN} be an ONS in H; for every sequence a - (a,,) from
{O, I} let T"EB(H) be defined by TJ""''2.a"<e,,,f)en ; this is an uncountable
set of operators such that II T" - Til II = I for a 'F p.
4.9. Let H be a Banach space. There are no operators A, B E B(H) such that
AB-BA =/.
Hint: From AB-BA=/ it follows that (n+l)B"=AB,,+l_B"+IA for all
n E 1'1, therefore liB" II ..;: 2/(n + I)IIA II liB II IIB"II, i.e., Bn =0 for large n; this
implies. that 0= B" = B n - I = ... = BO= I.

4.3 Isomorphisms, completion

Let HI and H2 be normed spaces. An operator V from HI into H2 is called
an isometry, if D( V) = HI and II Vflb = IIflll for all f E HI' An isometry V
from HI into H2 is called an isomorphism of HI onto Hz if R( U) = Hz. Every
isomorphism V of HI onto Hz is injective and U - I is an isomorphism of
H2 onto HI'
If HI and Hz are pre-Hilbert spaces and V is an isomorphism of HI onto
<
Hz, then it follows from the polarization identity that Vf, Ug)z = <f, g)1
for all f, g E HI' (The subscripts of the norms and scalar products will be
omitted in the sequel, as it will be always clear from the context, to which
spaces the elements belong.) Two normed spaces HI and Hz are said to be
isomorphic (or equivalent) if there exists an isomorphism of HI onto Hz.

Theorem 4.9. Let HI and H2 be isomorphic normed spaces. HI is a Banach

space (Hilbert space) if and only if Hz is a Banach space (Hilbert space).
PROOF. Let HI be a Banach space and let U be an isomorphism of HI onto
Hz. If (in) is a Cauchy sequence in Hz, then (V - '1,,) is a Cauchy sequence
in HI; hence there exists agE HI such that V - '1,,- g. With f = Vg E Hz we
have fn - f, i.e., Hz is complete. If HI is a Hilbert space, then H2 is complete
and since II V - '111 = II fll for all f E H2 , the parallelogram identity holds in
H2 , i.e., the norm of Hz is defined by a scalar product. As V - I is also an
isomorphism, we can prove analogously the reverse direction. 0
Theorem 4.10. Let H be a Hilbert space, and let A be a set, the cardinality of
which equals the (Hi/bert space) dimension of H. Then H is isomorphic to
Iz(A). In particular, all infinite dimensional separable Hilbert spaces are
isomorphic to '2'
Hilbert spaces having the same dimension are isomorphic to
each other.
PROOF. Let {e" : ex E A} be an ONB of H. For every f = "i.fA. E H let Vf
be the function A_C with (Uf)(a) = fa' It is easy to see that U is an
isomorphism of H onto 'iA). All the other assertions are obvious con-
sequences of this. 0
64 4 Linear operators and their adjoints

It is often useful to know that every normed space (pre-Hilbert space)

can be considered as a dense subspace of a Banach space (Hilbert space).
If H is a normed space (pre-Hilbert space), and A is a Banach space
(Hilbert space), then A is called a completion of H provided that H is
isomorphic to a dense subspace of A.

Theorem 4.11. For each normed space (pre-Hilbert space) H there exists a
completion A. Two arbitrary completions are isomorphic.
PROOF. We construct a completion A. For this let % be the set of all
Cauchy sequences in H. Two Cauchy sequences (f,,) and (g,,) from % are
considered equivalent (in symbols (fn)-(gn»' if IIfn - gnll~O. This obvi-
ously defines an equivalence relation (since we have (f,,)-(fn)' from
(f,,)-(g,,) it follows that (g,,)-(f,,), and (f,,)-(gn) and (g,,)-(h n) imply
(f,,)-(h n». Let A be the set of all equivalence classes. The elements of A
will be denoted by j, g, .... We shall write in particular j=[(!,,») if (f,,)
belongs to the equivalence class j and j = [j] if j is the equivalence class of
the sequences that converge to f E H (notice that fn ~ f and (fn) - ( gn)
imply gIl ~1).
With a[(!,,)] + b[(g,,)] = [(af" + bg,,») the set A becomes a vector space;
the zero element is 0 = [0). We show that a norm can be introduced on A
by putting
II[Un)]1I = lim IIf"lI·
For this, we have to notice that the sequence (jlfnll) is convergent (cf. the
proposition preceding Example I of Section 2.1) and the limit does not
depend on the choice of representatives, since for (f,,)-(g,,) we have
IIIInIl-11 g"III" UIn - g"II~O. The properties of a semi-norm obviously
hold. If 1I[(f,,»)1I = 0, then f,,~O, i.e., we have [Un)] = 0; consequently, we
have defined a norm.
If H is a pre-Hilbert space, then we define a scalar product by

This is obviously a semi-scalar product; since it induces the above norm, it

is a scalar product. Therefore A is a normed space or a pre-Hilbert space,
respectively.
N ow let [H] = ([f) E A : f E H}. The set [H] is obviously a subspace of
A. The space H is isomorphic to [H), since by Uf = [f) an isomorphism of
H onto [H] is defined.
[H] is dense in A: Let j = [(f,,)] E A. For each £ > 0 there exists an
n(£) E N such that Ilf" - fmll ,,£
for n, m;;;' n(£). Therefore for m;;;' n(£) we
have
IIj - [fm] II = ,,1i~ IIf" - fmll " £,

i.e., j is a contact point of [H].

4.3 Isomorphisms, completion 65

It remains to prove that A is complete. Let (k) be a Cauchy sequence in

A. Since [Hl is dense in A, for every n E N there exists a gn E H such that
IIln -[gnlll CO;; lin. Since we have
II gn - gm II = II [ gn] - [ gm] II
..; II[ gn] -],,11 + Ill" - 1m II + 111m - [gm]1I ~ 0
as n, m~oo, the sequence (gn) is a Cauchy sequence. We have

Ilk -[(gn)JII ..; Illk -[gkJII + lI[gd -[(gn)]11

< -kl + "-.00
lim II gk - gnll ~0
as k~oo; consequently, k~[(gn»).
Now let A and R be two completions of H and let 0 and V be the
corresponding isomorphisms of H onto the dense subspaces O(H) and
V(H) of A and R, respectively. Then Vo= (;V- I is an isomorphism of
V'(H) onto (;(H). By Theorem 4.5 Vo can be extended to an element of .v
S(R, A). For every j E ii there exists a sequence (1,,) from V(H) such that
J,,~I We have

II vlII = Jim II V.!,. II = Jim II Vo'!"l1 = Jim 11.1..11 = 11111,

i.e., V is an isometry. In order to prove that V is an isomorphism of R onto
A we have to show that R( V) = A. Let 1E A. Then there exists a sequence
(J,.) from D( H) such that 1" ~ j. If we put = vO - Y 1"
n , then (.1..) is a
Cauchy sequence in R. Therefore there exists an j E R such that.1.. ~I
Then we have

1= lim J" = lim (;V - %= lim v.1.. = vi

Thus 1E R( V), i.e., A and R are isomorphic. o
Proposition. Let HI and H2 be normed spaces (pre-Hilbert spaces) and let HI
be isomorphic to a dense subspace of H2. If AI and A2 are completions of HI
and H2, respectively, then AI and A2 are isomorphic.
PROOF. Let V be an isomorphism of HI onto the dense subspace V(H I ) of
H2, and let V be an isomorphism of H2 onto the dense subspace V(H2 ) of
fl2• Then VV(H I) is a dense subspace of fl2' hence HI is isomorphic to a
dense subspace of A2 , i.e., A2 is a completion of HI and, consequently, it is
isomorphic to AI' 0

Theorem 4.12. Let H be a pre-Hilbert space, and let T I , .•• , Tn be linear

functionals such that O(1j) = Hand L(TI , ••• , Tn) n B(H, K) = {D}. Then
n
M=
)-1
n N{TJ = {JEH: Tj=D for j= 1, ... , n}
66 4 Linear operators and their adjoints

is a dense subspace of H. In particular, the kernel of an unbounded functional

is dense.
PROOF. Without loss of generality we may assume that H is a dense
».
subspace of a Hilbert space (Ho, <- , . (If A is a completion of Hand U
is an isomorphism of H onto a dense subspace of A, then we replace H by
.U( H) and 1j by ~ U - I; the following proof shows that U( M) is dense in
U(H); consequently M is dense in H.)
Since M is surely a subspace, it is enough to show that M is dense in Ho,
i.e., M.l. = {O}. Let g E M.l.. Then for the continuous functional Tg : H-.
11<, Ti=<g,f> we have McN(Tg)' Hence by Theorem 4.1 we have
Tg E L(Tl' ... , Tn)' By assumption Tg = 0; consequently g = O. 0

EXERCISES

4.10. Let H be a pre-Hilbert space with an orthonormal basis {ea : IX E A}. Then
12(A) is a completion of H. (In particular, we obtain that 12 is a completion of
any infinite dimensional separable pre-Hilbert space; in proving this we do
not need Theorem 4.11).
4.11. Let H be a pre-Hilbert space over IK. A mapping A : H-+IK is said to be an
antilinear functional, if for allf, g E H and a, bE IK we have A (af + bg) =- a* Af
+ b* Ag. The functional A is said to be bounded, if there exists a C ;;:. 0 such
that IIAfll..; C IIfll for all f E H. Let H + be the set of bounded antilinear
functionals on H.
(a) An antilinear functional is continuous if and only if it is bounded.
(b) H+ becomes a Banach space with the norm IIAII=sup {IAfl :fEH,
IIfll < l}.
(c) To each gEH there corresponds an AgEH+ defined by Asf=(f,g).
The mapping E : H-+H +, gl-+Ag is isometric.
(d) E(H) is a completion of H.
(e) We have E(H)- H +.
Hint: Use Theorem 4.5 and 4.8. (This exercise provides a completion for
all pre-Hilbert spaces without reference to Theorem 4.11.)
4.12. (a) Let HI and H2 be isomorphic normed spaces. Then HI is separable if and
only if H2 is separable. .
(b) A normed space H is separable if and only if one (and then each) of its
completions is separable.
(c) Every infinite dimensional separable pre-Hilbert space is isomorphic to a
dense subspace of 12 (cf. also Exercise 4.10).
4.13. Let HI be a normed space, H2 a Banach space and fli a completion of HI'
Then 8(H" H2 ) and B(AI , H2) are isomorphic.
4.14. Let HI and H2 be Hilbert spaces. If U is an isomorphism of HI onto H2, then
for every subset M of HI we have U(M.l.)=(UM).l..
4.15. Let G c IR m be open and let L2• o( G) be as in Exercise 2.6(a). Assume that
1/-1' ... ,I/-n : G-+C are locally square integrable (i.e., square integrable on
each compact subset of G) and that L(I/-I"",o/n)nL 2(G)={O}. Then
4.4 Adjoint operator 67

M-{jEL 2 ,o(G): fG'Mx)f(x)dx=O forj=I, ... ,n} is a dense sub-

space of L2( G).
4.16. Let T be a bounded operator from a pre-Hilbert space HI into a Banach
space H2 • Then there exists an extension S E B(HI , H2) of T such that
IISII = II Til·
Hint: Use Exercise 4.7.
4.17. Let Om = {x E R m : Ixl = l} be the unit sphere in Rm and let C(Om) be the
space of continuous functions on Om.
(a) By <1, g) = f rlf(w)g(w) dom(w) a scalar product is defined on C(Om)
(here dom{w) denotes the surface element of Om). Let this pre-Hilbert
space be denoted by C2(Om)'
(b) Let the measure Pm (cf. Appendix A) be defined for every interval J c R m
by Pm(J) = the surface of that part of Om which lies in J. The space
L2(R m, Pm) is a completion of C2(Om); we shall denote it simply by
L 2(Om)'
(c) L2(Om) is separable.
(d) The space of infinitely many times continuously differentiable functions
(i.e., the set of the restrictions of infinitely many times continuously
differentiable functions defined on Rm) is dense in L2(Om)'

4.4 Adjoint operator

Assume that HI and H2 are Hilbert spaces, T is an operator from HI into
H2, and S is an operator from H2 into HI' The operator S is called aformal
adjoint of T if we have
(g, Tf) = (Sg,f) foraB fEO(T),gEO(S).
T is then a formal adjoint of S, also. We say that Sand T are formal
adjoints of each other. The operator So such that O(So}'" {D} is a formal
adjoint of every operator from HI into H2 •
If S is a formal adjoint of T, then for every g E O(S) the linear
functional Lg with
O(Lg) =D(T),Lsf= (g, Tf)

is continuous, since for all f E D(Lg } we have

Lgf = (g, Tf) = (Sg.!),
i.e., Lg is the restriction, to O( T), of the continuous functional TSg induced
by Sg.
If O(T) is dense, and the functional Lg is continuous, then by Theorem
4.5 this functional can be extended to HI = D( T) in a unique way, i.e.,
there exists an element hg E H.. uniquely determined by g and T via
(g, Tf) = Lgf = (hg .!) for all f E OCT).
 68 4 Linear operators and their adjoints

If S is a formal adjoint operator of T, and g E D(S), then we surely have

Sg = hg • Therefore in this case every formally adjoint operator of T is a
. restriction of the adjoint operator T* to be defined below.
Let T be a densely defined operator from HI into H2• and let
D* = {g E H2 : the functional f~( g, Tf) is continuous on D( T) }
= {g E H2 : there exists an hg E HI such that (hg, f> = (g, Tf)
for all f E D( T) } .
The element hg is uniquely determined: If (hi' f) = (g, Tf) = (h 2• f) for all
f E D(T), then hi - h2 E D(T).L = {O}. consequently. hi = h 2•
D* is a subspace of HI and the correspondence D*~Hh g~hg is a linear
transformation. since for gl' g2 E D* and a, bE K we obviously have
hag, +bg, = ahg, + bhg, .
Thus by D( T*) = D*, T* g = hg for g E D( T*) a linear operator T* from HI
into H2 is defined. The operator T* is a formal adjoint of T and is an
extension of all formal adjoints of T.

Theorem 4.13. Let T be a densely defined operator from HI into H2.

(a) If T* is also densely defined, then T** is an extension of T.
(b) We have N(T*) = R(T).L.
PROOF.
(a) As T and T* are formal adjoints of each other, T is a restriction of the
adjoint operator T** of T*.
(b) We have g E N(T*) if and only if g E D(T*) and T*g = 0 hold. Since
D( T) is dense. this is equivalent to the relation
(Tf, g) = (f. T*g) = 0 for all f E D(T).
This holds if and only if g E R(T).L. o
Theorem 4.14. Let T be a densely defined operator from HI into H2 •
(a) T is bounded if and only if T* E B(H2> HI)'
(b) If T is bounded. then II Til = II T*II.
(c) If T is bounded. then T** is the (by Theorem 4.5 uniquely determined)
continuous extension of T to the whole space HI' For T E B(Hh H2) we
have T** = T.

PROOF.
(a) and (b): Let T be bounded. Then for all g E H2 andf E D(T) we have

ILsil = I(g. Tf)1 <; IlglllITlillfll.

i.e .• Lg is continuous for all g E H2. Therefore D(T*) = H2.
4.4 Adjoint operator 69

By Theorem 4.3(b) we have

IIPII = sup {I<T·g,f)1 :fED(T),gEH2, Ilfll=l, IIglI=l}
= sup {I(g, Tf)1 :fED(T),gEH2 , IIfll=l, IIglI=l} = II Til·
If T· E 8(H2' HI), then T" belongs to 8(HI' H2). Hence the restriction
T of T" is also bounded.
(c) By Theorem 4.13 (a) we have T C T". By part (a) we have p. E
8(HI' H2)' As T is densely defined and continuous, the assertion
follows from Theorem 4.5.

EXAMPLE 1. Let MI and M2 be measurable subsets of IR m and IR n , respec-

tively. Let K denote the Hilbert-Schmidt operator from L2(M I ) into L2(M2),
induced by k E L2(M2 x M,). (cf. Section 4.1, Example 3):

(Kf)(x) = f. k(x,y)f(y) dy for f E L2(M I ).

M,

For allf E L2(M,) and g E L2(M2) the function g(x)k(x, y)f(y) is integrable
on M2 X MI' Therefore by Fubini's theorem we have

(g, Kf) = f. g(X).{f. k(x,y)f(y) dY } dx

M2 M,

= f. f(Y){f. k(x,y)·g(x) dX}. dy = <Hg,f),

M, M2

where H is the Hilbert-Schmidt operator induced by the kernel hey, x) =

k(x,y)·. If we define the atijoint kernel k· of the kernel k by
k·(y, x) = k(x,y)·,
then K· is the operator induced by k·.

EXAMPLE 2. Let T be a continuous linear functional on a Hilbert space H,

i.e., a continuous operator from H into II{, We want to compute T·. There
exists a uniquely determined g E H such that
Tf = (g,f) for all f E H.
Hence for all z ElK and all f E H we have
z·Tf = (zg,J) ,
i.e., T· z = zg for all z E II{,

EXAMPLE 3. This example shows that D(T·) = {O} may be true. To prove
this, for every kEN let the sequence (nk,')'EN of positive integers be
70 4 Linear operators and their adjoints

chosen in a way that

{nk , / : I EI\J} n {nj,/: I EI\J} =0 for j:/= k,

U /:I }=
kEN

leave the construction of such sequences to the reader). With these

sequences us define the operator T 12 by
O(T) = /20

Tf = fll•. ,) = ( fill,,' .. ,
1=1 kEN I-I

Let us observe that here all the sums occurring are finite. Moreover, we
Tf E Therefore the operator well-defined.
We show that g D(T*) implies g = Let g = (gn) E O( and (h,,)
= T*g. Then for all f E OCT) = 12,0 we have

Lk Tf) T*g, = L = Lk 1
II~. ,Ink. I
I n

(here one should notice that all sums are finite). If we choosefequal to the
vector (thus =1 fn = 0 n :/=n", then follows
hll." = gk for all / E I\J, k E I\J.
is only possible = 0.. From this follows that g =

Let T be an operator from HI into H2 . The graph of T is the subset

T) {(f, Tf)' E O(
of HI X H2 , where HI X H2 can be considered as a Hilbert space in the
sense of Section 3.1: HI X H2 = HI EEl H2 •

Tbeorem A subset G of X H2 tile graph of an operator HI

into H2 if and only if G is a subspace possessing the following property:
(0, E G g = O. Each subspace of a graph is a graph.
If is an operator HI H2 , G( T) a
subspace, as for (1;, gi) E G(T), a j E II< (i = I, 2) we have
al(f!. + g2) al(fl' a2(f2'
= (adl + a 2f2' T(adl + ad2» E G(T).
If (0, G( T), then it follows that = TO = O.
Let now a of X H2 having the above mentioned
property. We construct an operator T for which G = G(T) holds. For this,
let
T) E HI' there exists g E such that g) E
4.4 Adjoint operator 71

For every f E D(T) there exists exactly one g E Hz such that (1, g) E G, as
(1, g,) E G and (1, gz) E G imply (0, g, - gz) E G (by G being a subspace),
consequently, g, - gz = O. Therefore we can define a mapping T from D(T)
into Hz by
Tf = g for (f, g) E G.

T is linear: If fl.!2 E D(T) and ai' a z ElK, then we have (./;, T./;) E G for
i = I, 2. Hence (as G is a subspace) (adl + a2f2, a l Tf, + a2Tf2) E G. By the
definition of T we have
T(ad, + adz) = at Tf, + a2 Tf2'
By construction, we also have G = G(T). The last assertion can be ob-
tained from this immediately. 0

In the sequel we shall use the mappings

U: HI x Hz ~H2 X HI> U(f'.!2) = (f2' - j,)
V : H, X H2 ~H2 X HI, V(f" f2) = (fz, j,).
U and V are obviously isomorphisms of HI Ee Hz onto H2 Ef) HI' The inverse
operators U -, and V - I are given by

U -) : H2 x H, ~HI X Hz. U -1(f2.!1) = ( - fl.!2) ,

V-I: H2 X H, ~H, X Hz, V-'(jz.!,) = (j"fz)'

Theorem 4.16. Let T be a densely defined operator from HI into Hz. Then we
have
G(T*) = U(G(T)i) = (UG(T»i
(here the symbol ..1 has to be understood in the sense oj HI Ef) H2> respectively
HzEf) HI)'
PROOF. By the definition of T* we have
G(T*) = {(g,h)EH1XH I : (g, Tf)2 = (h.!), forall fED(T)}
= reg, h) EHz X HI : «g, h), (Tj, - j» =0 for all (j, Tj) E G(T)}
= (UG( T»i = U( G( T)i).
The last equality follows simply from the definition of U (cf. Exercise
4.14). 0

Theorem 4.17. Let T be a densely dejined i,yective operator jrom HI into H2 •

(a) We have G(T- I)= VG(T).
(b) Ij R(T) is dense, then T* is also i,yective, and we have T*-I = T- 1*.

PROOF. Part (a) is obvious.

72 4 Linear operators and their adjoints

(b) By Theorem 4.13(b) we have N( T"') = R( T)~ = {O}, i.e., T'" is injective.
As G(T- I)= VG(T), it follows (cf. Exercise 4.14) that

G(T- I",) = U-I(G(T-I)~) = U-IV(G(T)~)

= V-IU(G(T)~) = V-IG(T"') =G(p-I). 0
An operator T on a Hilbert space H is said to be Hermitian, if it is a
formal adjoint of itself, i.e., jf we have
(Tf, g) = (f, Tg) for all f, g E D(T).
An operator Ton H is symmetric (cf. Section 4.2) if it is Hermitian and
densely defined. Since a densely defined operator T is Hermitian jf and
only if it is a restriction of T*, we have: an operator T is symmetric if and
only if T is densely defined and T c P. An operator T on H is said to be
self-adjoint, if T is densely defined and T= T"'.
REMARK. For operators from B(H) the notions of Hermitian, symmetric,
and self-adjoint are equivalent.

Theorem 4.18. An operator T on a complex Hilbert space H is Hermitian if

and only if the quadratic form q(f) = (f, TJ> defined on D( T) is real.
PROOF. By definition, T is Hermitian if and only if the sesquilinear form
s(j, g) = (f, Tg) is Hermitian on D( T). The assertion follows from this by
Theorem 1.3(a). 0
A characterization of symmetric and self-adjoint operators may be
obtained immediately from Theorem 4.16; where U is defined by U(f, g)
=(g, - f) on HEB H.

Proposition. Let T be a densely defined operator on the Hilbert space H.

(a) T is symmetric if and only if
G(T) C U(G(T)~) or UG(T) CG(T)l.
(b) T is self-adjoint if and only if
G(T) = U(G(T)~) or UG(T) =G(T)~,
i.e.,
G(T) 1- UG(T) and G(T) EB UG(T) =HEBH.

Proposition. If T and S are densely defined operators from HI into Hz· and
T C S, then we have S'" C T"'.

Theorem 4.19. Let TI and T2 be densely defined operators from HI into Hz

and from H2 into H3 , respectively.
4.4 Adjoint operator 73

(a) If T 2T. is densely defined, then we have TTTi c(T2 T,)*.

(b) If T2 E B( H2, H3), then we have (T2 T.)* = TT Ti.
PROOF.
(a) We have to show that the operators TTl1 and T2 T, are formal
adjoints of each other. Let f E O(n Tn and g E O(T2 T.). Then f E
O(T!), 11fEO(Tt), gEO(T.), and T 1 gEO(T2). Consequently, from
the definition of the adjoint operator it follows that
(TtTil, g) = (Til, T1g) = <1, T 2T 1g)·
(b) Because of part (a) we only have to prove that O«T2 T.)*) C O(TTTi).
Letf E O«T2 T.)*). As Ti E 8(H3' H2 ), for all g E O(T2 T.) = O(T1), we
have
«T2 T I )*f, g) = (f, T 2 T l g) = <Ti!, Tlg)·
By the definition of the adjoint operator it follows from this that
TilE O(Tn, i.e.,jE O(TTTt). 0
Theorem 4.20. Let Sand T be operators from HI into H2 •
(a) If T is densely defined, then we have (aT)* = a* T* for all a E IK such
that a '1=0.
(b) If T+ S is densely defined, then (T+ S)*::J T* + S*.
(c) If S E 8(HI> H2 ) and T is densely defined, then we have (T+ S)* = T*
+S*.
PROOF.
(a) is evident (it follows from Theorem 4.19).
(b) Letf E O(T* + S*) = O(T*) n O(S*). Then for all g E O(T+ S) = OCT)
n O(S) we have by the definition of the adjoint operator that
«T* + S*)f, g) = (T*f, g) + <S*f, g) = <f, Tg) + (f, Sg)
= <1, (T+ S)g),

i.e.,fE O«T+ S)*) and (T+ S)*f= T*f+ S*j.

(c) Because of part (b) we only have to prove that O«T+ S)*) c O(T* +
S*)= O(T*). Let fE O«T+ S)*). Then for all gE O(T+ S)= OCT)
we have
<[(T+ S)* - S*]f, g) = <1, (T+ S)g) - (f, Sg) = <f, Tg).
From this it follows thatf E O(T*). o
Theorem 4.21. Let T be self-adjoint and i,yective. Then T - I is self-adjoint,
too.
PROOF. R(T) is dense, since we have {O} = N(T) = N(T*) = R(T)J.. Thus
the assertion follows from Theorem 4.17(b). 0
 74 4 Linear operators and their adjoints

EXERCISES

. 4.18. Let Sand T be densely defined operators from HI into H2• and from H2 into
H3• respectively. Assume that TS is densely defined. S is injective and
S - I E B(H2• HI)' Then we have (TS)* = S* T*.
Hint: By Theorem 4.19(a) we only have to show that D«TS)*) c D(S*T*)
holds; for all 1 E D« TS)*) and g E D(TS) = S - I D(T) we have <1, TSg) .,.
«S -I)*(TS)*I. Sg).
4.19. Let T E B(H I• H2 ).
(a) We have T*T E B(H I ). TT* E B(H2) and IIT*TII = IITT*II = IITII2.
(b) T*T and TT* are self-adjoint.

4.5 The theorem of Banach-Steinhaus,

strong and weak convergence
We first prove the theorem of Banach-Steinhaus. which is also known as
the uniform boundedness principle.

Theorem 4.22. Let HI and H2 be Banach spaces. and let M be a subset of

B(H I • H2 ). If M is pointwise bounded (i.e., for each f E HI there exists a
C(f) ~ 0 such that II Tfll ~ C(f) for all T EM), then M is bounded (i.e .•
there exists a C ~ 0 such that II T II ~ C for all T EM).
PROOF. 1. step. It is enough to show that there exist anfo E HI' a p > 0, and
a C' ;> 0 such that II Tfll ~ C' for all f E K(fo. p) and for all T E M. Indeed,
if fo. p, C' have these properties, then for all g E K(O, p) and for all T EM
we have
II Tgil = II T(fo+ g -fo)11 ~ II T(fo+ g)11 + II Tfoll ~ c' + C(fo) = C",
sincefo + g E K(fo, p). Consequently, for all g E K(O, 1) and T E M we have
II Tgil ~ p-IC" = C,
i.e .• II Til "C for all T E M.
2. step. What. remains is to prove the existence of fo' p, and C' with the
above properties. We assume that no such elements exist, i.e., for each
foE HI and for each p>O the set {IITfll: TEM, fEK(fo'p)} is un-
bounded. In particular, the set {II TfI1 : T E M. f E K(O, I)} is unbounded.
Therefore there exist an fl E K(O, 1) and a TI E M such that II Tdlll > 1.
Since TI is continuous, there exists a PI' 0 < PI < 2 -I such that
R(fl' PI) C K(O, 1) and II Tdll > 1 for all f E R(fl' PI)'
Since {I\ Tfll : T E M, f E K(fl' PI)} is unbounded, there exist an f2 E
K(fl' PI) and a T2 EM such that II Tzf211 > 2. As T2 is continuous, there
4.5 The theorem of Banach·Steinhaus, strong and weak convergence 75

exists a P2' 0 < P2 < 2 -2 such that

K(j2' P2) C K(jl' PI) and II T2fll > 2 for all f E K(j2' P2)'
In this way, by induction we obtain a sequence Un) from HI, (Tn)
from M, and (Pn) from (0, I) such thatK(fn+\l Pn+l)cKUn' Pn)' Pn<2- n,
and II Tnill > n for all f E KUn' Pn)' In particular, we have
IIfn - fmll < Pno for n, m > no·
Since Pn < 2- n, it follows from this that (fn) is a Cauchy sequence. Thus
there exists an f E HI such that fn -f. Since for n > m we have fn E
KUm, Pm)' it follows that f E RUm' Pm) for all mEN. Consequently II Tmfll
> m is true for mEN, which contradicts II Tmfll ...: CU)· 0
Let HI and H2 be normed spaces. A sequence (Tn) from B( HI> H2) is said
to be strongly convergent to T E B(H\> H2 ), if for all f E HI we have
Tf= lim Tnf. For this we shall write T= s -lim Tn or Tn ~ T. The operator
T is called the strong limit of the sequence (Tn)' It is obvious that every
sequence (Til) in B(Hh Hz) has at most one strong limit. A sequence (Tn)
from B(Hh H2) is said to be a strong Cauchy sequence, if for every f E HI
the sequence (T,j) is a Cauchy sequence in H2 .Every strongly
convergent sequence is a strong Cauchy sequence.

Theorem 4.23. Let HI and H2 be normed spaces.

(a) If (Tn) is a strongly convergent sequence in B(HI> H2) and T= s-
lim Til' then II Til ...: lim inf II Til II·
(b) If the sequence (Tn) from 8(HI' H2) is bounded and (Tng) is a Cauchy
sequence for every g in a dense subset M of HI> then (Tn) is a strong
Cauchy sequence.
(c) If HI is a Banach space, then every strong Cauchy sequence in 8(HI' H2)
is bounded.
(d) If HI and H2 are Banach spaces and (Tn) is a strong Cauchy sequence in
s
B(Hh H2), then there exists aTE 8(H\I H2) such that Tn _ T.
PROOF.
(a) Let C=lim inf II Tn II· Then there exists a subsequence (Tn) of (Tn)
such that II Tn.II-C as k-HX). Hence for aUfE HI we have

II Tfll = lim II TnJIl ...: lim II Tn.lll1fll = Cllfll,

i.e., II Til ...: C.

(b) Letf E HI> E > O. We have to show that there exists an no E N such that
II T,,j - Tmfll ...: E for all m, n >no' As M is dense, there exists agE M
such that IIf-gll"':E/3C (with C=suP{IITnll :nEN». If we now
choose no in such a way that II Tng - Tmgll ...: E/3 for aU n, m > no, then
76 4 Linear operators and their adjoints

we have
/IT,J- T"JII..; II TnU-g)11 + IITng- Tmgll + II Tm(g-j)1I
..; f for n, m > no'
(c) For every fE HI the sequence (Tn!) is a Cauchy sequence and thus it
is bounded. Consequently, by Theorem 4.22 there exists a C such that
IITnll";C for all nEN.
(d) For every f E HI the sequence (Tn!) is a Cauchy sequence, so it is
convergent in H2. Let us define T by Tf= lim Tnf. Then O(T)= HI'
The operator T is linear, since for all f, g E HI and a, bEll( we have
T(af+bg) = lim TAaf+bg) = Iim(aT,J+bTng) = aTf+ bTg.
By part (c) there exists a C > 0 such that II Tn II ..; C for all n E N.
Consequently, we have
IITfll = lim IIT,JII ..; Cllfll·
i.e., T E 8(HI' H2)' By construction, we obviously have Tn ~ T. 0
EXAMPLE 1. Let the operators Tn E 8(12) be defined by
TnUI,f2,j3' ... ) = Un+I,J,,+2' ... ).
Then for all f E 12 we obviously have Tnf~O, i.e., Tn ~ O. For all f E 12 we
have /I Tnfll ..; IIfll, consequently II Tnll ..; 1. Moreover, for ej = (6j , n) we have
1jej + 1= e l • Therefore II Tn II = 1. From this it follows that strong conver-
gence does not in general imply convergence in the norm of 8(HI> H2)'

Let H be a pre-Hilbert space. A sequence (fn) from H is said to converge

weakly to f E H if for all g E H we have <In' g)~<f, g). For this we write
w
f = w - lim J" or fn ~ f. The element f is called the weak limit of the
sequence (fn)' Every sequence has at most one weak limit. A sequence (fn)
from H is called a weak Cauchy sequence, if for every g E H the sequence
«f", g) ) is a Cauchy sequence in K Every weakly convergent sequence is
a w;;ak Cauchy sequence.

Theorem 4.24. Let H be a pre-Hilbert space.

(a) If (f,,) is a weakly convergent sequence in H, and f = w -limfn, then we
have Hfll..; lim inf IIfnll.
(b) If the sequence (J,,) is bounded in Hand (<I", g» is a Cauchy sequence
for all g from a dense subset M of H, then (f,,) is a weak Cauchy
sequence.
(c) If H is a Hilbert space, then every weak Cauchy sequence is bounded in
H.
(d) If H is a Hilbert space and (fn) is a weak Cauchy sequence in H, then
w
there exists an f E H such that fn ~ J.
4.5 The theorem of Banach-Steinhaus, strong and weak convergence 77

The proof immediately follows from Theorem 4.23, if we notice that the
weak convergence of (fn) is equivalent to the strong convergence of the
sequence (1j,) of the linear functionals induced by fn' In part (d) we have
to use Riesz' theorem (Theorem 4.8). The details are left to the reader.

Theorem 4.25. Let H be a Hilbert space. Every bounded sequence (J,,) in H

contains a weakly convergent subsequence
.
(J"k ).
PROOF. Let M= L{fn : n EN}. Then MfB M1. is dense in H. For every
kEN the sequence «J", fk) )nEN is bounded. Consequently, by induction
we can find, for all j E N, a subsequence (J" )/EN of (J,,) such that
~.,

(f~+.)/EN is a subsequence of (fnj)/EN and «f~."h»)/EN is convergent.

With the diagonal sequence (fn)=(fn,) the sequence (<In,,h))IEN is then
convergent for all j E N. Since for all f E M 1. we have <fn" f) = 0 for all
1 E N, the sequence «fn" f») is convergent for all f from the dense
subspace M fB M 1.. Therefore by Theorem 4.24(b) and (d) (fn) , is weakly
convergent. 0
EXAMPLE 2. Every orthonormal sequence (J,,) weakly converges to zero.
This follows from the Bessel inequality IIf1l2:> ~I<fn' f)12. In particular,
the sequence of unit vectors (e; = (Sjn» in 12 tends to zero weakly. This
example also shows that weak convergence does not imply strong conver-
gence in general.

EXAMPLE 3. For every f = ~ j.. I he; E 12 let the sequence (f(n» be defined by
j<n) = ~j.1 ht.i+n' Then (f(n» converges weakly to zero, since for all g =
~ gjej E 12 we have

12
1<ln), g) IZ = 1j~l.t;gj+" IIfll Zj=~+ II ~l ~0
00 00
<
as n~oo.

EXAMPLE 4. In the pre-Hilbert space Iz,o the sequence (kek ) (with ek = (Snk»
weakly converges to O. However, it is unbounded.

Let HI and Hz be pre-Hilbert spaces. A sequence (Tn) from B(H 1, H2) is

said to converge to T E B(HI' Hz) weakly, if for all f E HI the sequence
<
(Tn!) in Hz weakly converges to Tf, i.e., Tnf, g) ~< Tf, g) for all f E HI
and g E Hz. In this case we shall write T= w -lim Tn or Tn ~ T, and call T
the weak limit of the sequence (Tn)' A sequence (Tn) from B(HJ> H2 ) is said
to be a weak Cauchy sequence if (Tn!) is a weak Cauchy sequence in H2 for
eachfE HI'
Theorem 4.26. Let HI and H2 be pre-Hilbert spaces.
(a) If (Tn) is a weakly convergent sequence in B(H .. H 2 ) and
T= w -lim Tn' then we have II Til < lim inf 1/ Tnl/·
78 4 Linear operators and their adjoints

(b) If the sequence (Tn) from B( HI> H2) is bounded and ( T,J, g» is a
Cauchy sequence for every f E MI and every g E M2, where MI and M2
are dense subsets of HI and H2, respectively, then (T,,) is a weak Cauchy
sequence.
(c) If HI and H2 are Hilbert spaces, then every weak Cauchy sequence (T,,)
from B(HI' H2 ) is bounded.
(d) If HI and H2 are Hi/bert spaces, and (Tn) is a weak Cauchy sequence in
• w
B(H" H2 ), then there eXIsts aTE B(HI' H2) such that Tn ~ T.
PROOF.
(a) Let C = lim inf II T"II and let (T,,) be a subsequence such that II T".II~
C. Then for all f E HI and g E H2 we have
I(Tf, g)1 = lim I(T".1, g)1 " lim II T".11l1fllll gil = Cllflill gil·
The assertion follows from this by Theorem 4.3.
(b) We have to prove that for arbitrary f E HI, g E H2, and € >0 there
exists an no such that
I«T" - T,.,)f, g)1 ,,€ for n, m > no'
Let C = I + (I + 11111 + II gil) sup {II T"II : n E I\I}. Since the sets ~ are
dense in H; (j = I, 2), there exist fo E MI and go E M2 such that
€ €
Ilf - foil " 5C' II g - goll " 5C' IIfoll " IIfll + 1.

If we choose no E 1\1 in such a way that for n, m > no we have I«T,,-

Tm)Jo, go)1 "€/5, then for n, m >no it follows that
1«(Tn - Tm)f, g)1 " I(T"U - fo), g)1
+ I(T,Jo, g - go)1 + 1«(Tn - Tm)fo, go)1
+ I(T"Jo' go- g)1 + I(TmUO - f), g)1
" €.

(c) For every f E HI the sequence (T,,!) is a weak Cauchy sequence in H2 •

Since H2 is a Hilbert space, the sequence (T,.1) is bounded by Theorem
4.24(c). The boundedness of the sequence (II T"ID follows from this via
Theorem 4.22, as HI is also a Hilbert space.
(d) For every fE HI the sequence (Tn!) is a weak Cauchy sequence,
therefore by Theorem 4.24(d) it is weakly convergent in H2 • We define
T by Tf= w -lim T"f for allf E HI' We can prove the linearity of T, as
in the proof of Theorem 4.23(d). By part (c) there exists a C > 0 such
that 1\ T"II "C for all n E 1\1. It follows that
I(Tf, g)1 = lim I(T,J, g)1 " Cllflill gil for all f E H" g E H2•
Consequently T is bounded. By construction, we obviously have
w
~~~ 0
4.5 The theorem of Banach-Steinhaus, strong and weak convergence 79

Proposition.
s s w
(a) T" -+ T implies T" -+ T; T" --+ T implies T" -+ T.
w W
(b) If HI and H2 are Hilbert spaces, then T" -+ T is equivalent to T: -+ T*.

EXAMPLE 5. Let us consider the operators Tn from Example I. The adjoint

operators T: are obviously defined by

T:( i: j;ej ) =
j-I
~ j;e"+j
j= I
for all ~ j;ej E '2'
j-I

By Example 3 we have T: W
--+ O. However, we do not have -+ O. T: s

Therefore, strong convergence does not imply the strong convergence of

the adjoint operators.

Theorem 4.27. Let HI' H2 be Hilbert spaces, and let T be an operator from
HI into H2 such that D( T) = HI' Then the following assertions are equivalent:
(i) T ~ bounded (i.e., /j.-+f implies Tf,,-+ Tj),
(ii) f" -+ f implies Tf" -+ Tf,
(iii) f" -+ f implies Tf" ~ Tj.
w
PROOF. (i) implies (ii): If f" -+ f, then for every g E H2 (notice that T* exists
and T* E B(H2' HI» we have
<g, TJ,,) = <T*g,fn) --+ <T*g,f) = <g, Tf),
Le., T1" ~ Tf·
w
(ii) implies (iii): This is obvious, as fn -+ f implies fn -+ f.
(iii) implies (i): Let us assume that T is not bounded, i.e. there exists a
sequence (1,,) from HI such that IIf,,1I < I and II T1"II "> n 2 • Then we have
(l/n)J,,-+O. Therefore from (iii) it follows that (I/n)Tfn~O. By Theorem
4.24(c) the sequence «I/n}Tf,,) is thus bounded. This contradicts the fact
that
I I
II-Tfnll = -II Tf,,11 "> n.
n n o
Theorem 4.28. Let H be a Hilbert space and let (T,,) be a bounded sequence
of symmetric operators from B( H).
W
(a) If Tn -+ T for some T E B(H), then T is also symmetric.
(b) If the sequence «f, T.J)} is non-decreasing for every f E H, then there
exists aTE B( H) such that Tn ~ T. The same holds true if the sequence
(<1, T.J)} is non-increasing for every f E H.
PROOF.
(a) For all f, g E H we have
<1, Tg) = lim <f, T"g) = lim <T.J, g) = <Tf, g).
Therefore, T is symmetric.
80 4 Linear operators and their adjoints

(b) The sequence «I, T,J» is non-decreasing and bounded for every
1 E H, consequently it is convergent. If C= 2 sup {II T"II : n EN}, then
II T" - Tmll <; C for all n, mEN. The Schwarz inequality applied to the
non-negative sesquilinear form s(g,!)=<g,(T,,-Tm)f) shows that
for all f E H we have

II(T,,- Tm)jll = «T,,- Tm)f, (T,,- Tm)I)'/2 = {s«T,,- Tm)f,J)} 1/2

<; {s« Tn - Tm)l, (Tn - Tm)j)s(f,J)} 1/4
. 2 . } 1/4
= { «Tn - Tm)f, (Tn - Tm) f)<J, (Tn - Tm)f)
<; II(Tn - Tm)III'/411(Tn - Tm)2flll/4<J, (T" - Tm)f)I/4
<; C 3/411 111'/2<1, (Tn - Tm)j)'/4 ---+ 0 as n, m ---+ 00.

So (Tn!) is a Cauchy sequence for every 1 E H, consequently (Tn) is

strongly convergent. 0

EXERCISES

4.20. Let HI' H2 and H3 be normed spaces, and let S", S E B(H2, H3), Tn> T E
s s
B(H .. H2). Sn -'> S, Tn -'> T.
(a) If the sequence (Sn) is bounded, then SIt Tn ~ ST.
(b) If Hz is a Banach space, then Sn Tn ~ ST.
w
4.21. (a) Let H be a Hilbert space. If 1,,-'>1 in Hand 11111 ;;0 lim sup 111,,11, then
1,,-'>/·
Hint: Treat III" - 1112.
(b) Let HI and H2 be Hilbert spaces, and let Tn and T be isometries from HI
. w s
mto H2 such that Tn -'> T. Then we have Tn -'> T.
4.22. Let HI. Hz and H3 be Hilbert spaces, and assume Sm S E B(Hz• H3). Tn' T E
B(HIo Hz).
w s w
(a) Sn -'> S, Tn -'> T imply SIt T" -'> ST.
s w . w
(b) SIt -'> S, Tn -'> T do not Imply Sn Tn -'> ST.
Hint: Let HI = Hz = H3 = Iz, Tn("i:.hej) = "i:.hej+n. Sn = T:. Then we have
s w
. Sn-'>O, Tn-,>O, SnTn-,>I.
w w
(c) Sn -'> S, Tn -+ T imply Sn T" -+ ST.
4.23. Let HI and H2 be Hilbert spaces. and take An' A, from B(HI' Hz} and B
w s •
from B( H2 , HI)' Then An -+ A, A: -+ B imply B ... A ... , therefore that A: -'> A"'.
w
4.24. (a) If H is a finite dimensional Hilbert space, then In -'>1 is equivalent to
1,,-'>1.
(b) If HI and Hz are Hilbert spaces, and Ha. is finite dimensional, then for
Tn' T from B(HI' H2) the statements Tn -'> T and T" ~ T are equivalent;
this holds true in particular for linear functionals on a Hilbert space.
(c) If HI and H2 are finite dimensional, then for Tn> T from B(H I • H2) the
w S • I
statements T" -'> T, T" -'> T, T,,-'> Tare eqUiva ent.
4.6 Orthogonal projections, isometric and unitary operators 81

4.25. Let H be a Hilbert space.

(a) If M is a closed subspace of H and (in) is a sequence in M such that
in ~ i, then we have i E M (we say that M is weakly closed).
(b) If (I.) is a sequence in H such that in ~ i, then there exists a subsequence
Un.) of (j,,) such that (I / m)~'J:_ Jink --'>i as m--'>oo. .
Hint: Treat the case in ~ 0 first and choose the subsequence (in) such
that for k <j we have l<fn•.!,,)1 <F).
(c) A convex subset M of H is cl~sed if and only if it is weakly closed.
4.26. Let H), Hz be Hilbert spaces and let H) be separable. If (Tn) is a bounded
sequence from S(H), Hz), then there exists a weakly convergent subsequence
(Tn)·
4.27. Let H be a complex Hilbert space and assume that (Tn) is a sequence in 8(H)
such that «i, T,,1» is bounded for every i E H. Then the sequence (Tn) is
bounded in 8(H). (For symmetric operators Tn this holds also in real Hilbert
spaces.)
4.28. Let H) and Hz be Hilbert spaces. Assume that (Tn) is a bounded sequence
from S(H), Hz), T E 8(H), Hz), and M) and Mz are dense subsets of HI and
Hz, respectively. [f T"i--,>Tifor aJIiEM) (respectively, <g, Tni)--'><g, TJ>
for all i EM) and g E M2 ), then it follows that Tn':" T (respectively Tn ~ T).

4.6 Orthogonal projections, isometric and unitary

operators
Let H be a Hilbert space and let M be a closed subspace of H. By Theorem
3.2 every f E H can be uniquely represented in the form f = g + h with
gEM and hEM J.; g is called the orthogonal projection of f onto M. If we
define the mapping PM by D(P M) = Hand P,.J = g, then PM is a linear
operator on H such that D( PM) = H, since for fl = g I + hI and f2 = g2 + h2
with lJ.; E M and hj E M J. we have af) + bf2 = (ag l + bg2) + (ahl + bh 2) with
ag l + bg2 E M, ah) + bh z E MJ., therefore PM(afl + bf2) = ag l + bg2 = aP,.J)
+ bP ,.J2' The operator PM is called the orthogonal projection onto M.
Because IIfll2 = II gll2 + IIh1l 2, we have IIP,.JII = II gil' IIfll for all f E H,
i.e., we have lIPMil ,1. If M = {O}, then it is obvious that PM = O. If
M* {O}, andf EM, f*O, then P,.J= f, hence IIPMII = 1. As P,.J= f for all
f E M, it follows that P~ = P MP M= PM' i.e., PM is idempotent. We have
P,.J EM for all f E Hand P,.J= f for all f EM, therefore R(P M) = M. As
P,..,J = 0 if and only if f E M J., we see that N( PM) = M.L. An operator P on
H is called an orthogonal projection if there exists a closed subspace M such
that P= PM'
Theorem 4.29. For an operator P E 8(H) the following statements are
equivalent:
0) P is an orthogonal projection,
82 4 Linear operators and their adjoints

(ii) 1- P is an orthogonal projection,

(iii) P is idempotent and R(P) = N(P) 1.. ,
(iv) P is idempotent and self-adjoint.
We have R(P) .. N(I - P) and N(P) = R(I - P).
PROOF. (i) and (ii) are equivalent: From the definition of orthogonal
projections it follows immediately that P is the orthogonal projection onto
M if and only if 1- P is the orthogonal projection onto M 1... From this it
also follows that R(P) - M == M 1.. 1.. ... N(I - P), N(P) ... M 1.. = R(I - P).
(i) and (iii) are equivalent: Using the above reasoning we only have to
show that (iii) implies (i). As R(P) = N(P) 1.. , the range R(P) is a closed
subspace. For all g E R(P) we have Pg = g, as P is idempotent. If we write
f E H in the form f= g + h, where g E R(P) and hE R(P) 1.. = N(P), then
we have Pf= Pg + Ph == g, Le., P is the orthogonal projection onto R(P).
(iv) follows from (i): P is idempotent, as (iii) follows from (i). For all
fl == gl + hi' f2 = 82 + h2 with gj E R(P), hj E R(P)1.. we have .
<Pfl.J2) = <gl' g2 + h2) = <gl' g2) ... <gl + hI' g2)
= <fl' Pf2)'
i.e., P is self-adjoint.
(iii) follows from (iv): We only have to prove that R(P) = N(P) 1.. • If
f E R(P), f== Pg, then (I - P)f= 0, consequently f E N(I - P). Iff E N(I-
P), thenf- Pf=O, consequently fE R(P). Therefore we have R(P) ... N(I
- P) and thus R(P) is closed. From this it follows that R(P) = R(P) 1.. .L
== N(P*)1.. == N(P).L. 0
Theorem 4.30. Let M and N be closed subspaces of a Hilbert space H, and let
PM and P N be the orthogonal projections onto M and N, respectively.
(a) p ... P MP N is an orthogonal projection if and only if P MP N = P NP M holds;
then we have P=P MnN . We have M.lN if and only if PMPN=O (or
PNPM=O).
(b) Q = PM + PN is an orthogonal projection if and only if M.l N, then we
have Q .. P MtIlN .
(c) R - PM - PN is an orthogonal projection if and only if N eM; we then
have R .. P M9N .
PROOF.
(a) If P - P MP N is an orthogonal projection, then P is self-adjoint. Conse-
quently, PMP N ... P = p* = (PMP N)*';" P~P~ = PNP M. Let PMP N =
PNP M hold. Then it follows that p2=(PMPNi"'PMPNPMPN"'P~P~
= PMP N = P and P* ... (PMP N)* = (PNP M)* ... p~p~ ... PMP N .. P.
Therefore P is an orthogonal projection by Theorem 4.29. Since
P=PMPN=PNP M, we have R(P)cR(PM)nR(PN)-MnN. On the
other hand, iffE M n N, then we have Pf- PMPJ- PMi- f, therefore
M nNe R(P). Consequently, M n N= R(P). It is obvious that PMP N
... 0 if and only if g E R(P M)1.. = M 1.. holds for all g E R(PN) = N, i.e., if
N1- M. The other assertion follows similarly.
4.6 Orthogonal projections, isometric and unitary operators 83

(b) If Q = PM + P N is an orthogonal projection, then 11/112;> II Qfll 2 =

<Qj, I) = <P,.J, I) + <P,J, f) = IIP,.J1I 2 + IIP,Jf For f= PMg it
follows that IIP MgIl 2 ;> IIPMgIl 2 + IIPNPMgII1. Therefore PNPMg=O for
all g E H, i.e., P NP M = O. It follows analogously that P MP N= O. Conse-
quently, by part (a), M1.. N holds. We obviously have R(Q) C R(PM) +
R(PN)=M$N. Conversely, if l=g+hEM$N with gEM, hEN,
then Qj=Qg+Qh=PMg+PNh=g+h=l. Therefore R(Q)=M$N.
If M1.. N, then by part (a), we have P MP N'" P NP M =0, consequently
Q2=(PM+PN)2=p'it+P~=PM+PN"'Q. As the operators PM and
P N are self-adjoint, Q is also self-adjoint, consequently it is an orthogo-
nal projection.
(c) If R ... PM - P N is the orthogonal projection onto the subspace L, then,
because of the equality PM = P L + P N' by part (b) we have L 1.. Nand
M = L $ N ~ N. Therefore L = MeN, i.e., R is the orthogonal projec-
tion onto MeN. Conversely, if N C M and L = MeN, then by part (b)
we have PM = P L+ P N' hence R = PM - P N= P L is an orthogonal .
projection. 0
F or two symmefric operators A, B E B( H) we write A <: B (or B ;> A) if
for all IE H we have <AI, I) 0;;; <BI, I) (by Theorem 4.18 <AI,/) and
<B/, I) are real). A is said to be non-negative if A ;> O.

Theorem 4.31. Let M and N be closed subspaces of the Hilbert space H, and
let PM and PN be the orthogonal projections onto M and N, respectively.
(a) We have 00;;; PM 0;;; f.
(b) The lollowing statements are equivalent:
(i) PM .... P N' (iii) P NP M = PM'
(ii) MeN, (iv) P MP N= PM'
PROOF.
(a) For all IE H we have <01,/) = 00;;; IIP MI1I 2 = <P,.J, I) <: 11/112 =
(II.!)·
(b) (i) implies (ii): If PM <:P N' then IIP"J1I 1 "'<P"J,J)<:.<PJ,J)=
II PN/lil for all IE H. Therefore N(P N) C N(P M) and thus M ... R(P M)
... N(PM).L c N(PN).L ... R(P N) = N.
(ii) implies (iii): By Theorem 4.20 (a) and (b), with L = N e M we have
PNP M=(P M+ PL)P'it= P M= PM'
(iii) implies (iv): As P NP M is an orthogonal projection (namely PM)' by
Theorem 4.30 (a) we have P MP N = P NP M = PM'
(iv) implies (i): Because of the equality P MP N= PM we have, for all
IE H, that
<P"J,J) = II P"J1I 2 = IIP MPJII 2 <: II P JII 2 = <PJ,j). 0

A sequence (Tn) of symmetric operators Tn E 8(H) is said to be mono-

tone (non-decreasing or non-increasing, respectively) if for every f E H the
84 4 Linear operators and their adjoints

sequence « Tnf. f» is monotone (non-decreasing or non-increasing. re-

spectively). Theorem 4.28(b) says that any bounded non-decreasing
sequence of symmetric operators is strongly convergent. For sequences of
orthogonal projections we have

Theorem 4.32.
(a) If (Pn) is a monotone sequence of orthogonal projections acting on the
Hilbert space H, then there exists an orthogonal projection P acting on H
s
such that Pn ~ P.
(b) If (Pn) is non-decreasing (i.e., Pn < Pn+ I)' then P is the orthogonal
projection onto U nENR{Pn).
(c) If (Pn) is non-increasing (i.e., Pn ;;;' Pn+ I)' then P is the orthogonal
projection onto nnENR(Pn).
PROOF.
(a) Because of Theorem 4.28(b) there exists a self-adjoint operator P E
8(H) such that Pn ~ P. As <p 2f, g) = <Pf, Pg) = lim <Pnf, Png) =
lim <Pnf. g) = <Pf. g) for all f, g E H, the operator P is idempotent
and thus it is an orthogonal projection.
(b) Iff..i UnENR(Pn). then Pnf=O for all n EN, consequently Pf= lim Pnf
=0. If fE UnENR(Pn), thenfE R(PnJ for some noEN. Since R(PnJ
C R(Pn} for all n ;;;. no, we have Pnf= PnJ for n;;;' no' Therefore Pf= f,
i.e., UnENR(Pn)cR(P). As R(P) is closed, it follows that UnENR(Pn)
= R(P).
(c) The sequence (Qn) with Qn = 1- Pn is non-decreasing; Q = lim Qn is
therefore the orthogonal projection onto U nENR( Qn)= U nENN(Pn).
Then P = 1- Q is the orthogonal projection onto U nENN(Pn)l. =
n nENN(Pn)l. = n nENR(Pn )· 0
In the calculation of the norm of the difference of two orthogonal
projections the following theorem is often useful.

Theorem 4.33. Let PI and P2 be orthogonal projections acting on the Hilbert

space H. Then we have

where

PROOF.
(a) By the definition of the norm of an operator we have (notice that
R(Pk)l. = N(Pk»
IIP I - P2 11 = sup {II(P I - P2 )fll :fEH,Ilfll< I}
;;;. sup {II(P I - P2 )fll : f E R(PI)l., Ilfll< I}
= sup {IIP2 fll :fER(PI).L, Ilfll< I} = P21'
4.6 Orthogonal projections. isometric and unitary operators 85

The inequality IIP I - P 2 11 ;> PI2 follows in a similar way.

(b) We have PI - P 2 = P,(J - P 2) - (/ - P,)P2• As (J - P 2)f E R(P2)J.. for
all f E H it follows that
IIP,(I - P 2 )fll < P1211 (I - P 2 )fll·
Moreover. since (J - P,)Pd E R(PI)J.. we have

11(1 - P I)PdIl 2 = <(I - P I)P2 f. (I - PI)Pd) = <Pi! - PI)Pd. Pd:

< IIPiI - P I )P2 fll IIPdll < P2111(1 - P I)P2fIlIJPdll.
consequently

This implies for all f E H

II(P I - P 2 )f1l 2 = IIP,(I - P 2 )f - (J - P ,)PdIl 2

= IIP,(I - P 2 )f1l 2 + 11(1 - P ,)PdIl 2
< pM(I - P 2)f1l 2 + p~tl1P2f112
< max {pi2' p~,}(II(J- P2)fIl 2 + II P dIl 2)
= max {P!2' P~I} IIf112.
therefore
o
Let HI and H2 be Hilbert spaces. An operator V from HI into H2 such
that D( V) = HI is called an isometry if II Vfll = IIfll for all f E HI' If V is an
isometry and R( V) = H2• then V is an isomorphism of HI onto H2• In this
case V is called a unitary operator. An operator V from HI into H2 such
that D( V) = HI is called a partial isometry if there exists a closed subspace
M of HI such that
II Vfll = IIfll for f E M, Vf = 0 for f E MJ..

We have R( V) = VM; this shows immediately that R( V) is closed (if

Vf.. -'; g E H2 , then (P ~n) is a Cauchy sequence in M; therefore we have
P r.d" -';f E M, and thus V!,. -'; Vf = g E R( V». The closed subspaces M and
R( V) are called the initial and final domains of V, respectively.

Theorem 4.34. Let HI and H2 be Hilbert spaces and let V be an operator

from HI into H2 such that D( V) = HI'
(a) The following assertions are equivalent:
(i) U is a partial isometry with initial domain M and final domain N,
(ii) R(U)=N and <Vf, Ug)=<PMJ,g) Jorall f,gEH I,
(iii) V" V = PM and VV" = P N'
(iv) V" is a partial isometry with initial domain N and final domain M.
86 4 Linear operators and their adjoints

(b) The following assertions are equivalent:

(i) U is unitary,
<
(ii) R( U) = H2 and Uf, Ug) = <f, g) for all f, g E HI>
(iii) U·U=I H I and UU·=I H 2, i.e., U·= U-I,
(iv) U· is unitary.
PROOF.
(a) The equivalence of (i) and (ii) follows by (1.8) in the real case and by
(1.4) in the complex case:
(i) implies (iv): We have N( U·) = R( U)1. = N 1. and (because (i) implies
(ii» II U· Ufll = II P Mfll = II Ufll for all f E HI> therefore II U· hll = II h II for
all h E R( U) = N. Hence the operator U· is a partial isometry with
initial domain N. If we interchange the roles of U and U· in this
reasoning, then it follows that the final domain of U· is equal to the
initial domain of U·· = U, and, consequently, it is equal to M.
(iv) implies (i) for the same reason.
(i) implies (iii): As (i) implies (ii), it follows that U· U = PM. As (i)
implies (iv), it follows similarly that UU· = P N.
(iii) implies (ii): We have R( U):J R( UU·) = R(P N) = N. Since II U·f11 2
= <UU·f, f) = IIP N fIl 2 , we have N(U·) = N1. and thus R(U)
c N( U·) 1. = N.. Summing up, it follows that R( U) = N. Moreover, we
have <Uf, Ug) = <U· Uf, g) = <PMf, g) for allf, gE H.
(b) This is a special case of part (a). D

Theorem 4.35. If P and Q are orthogonal projections on the Hilbert space H

such that IIP- QII < 1, then we have
(a) dim R(P)=dim R(Q), dim R(l-P)=dim R(/-Q),
(b) P and Q are unitarily equivalent, i.e., there exists a unitary operator U in
H such that Q= UPU- I and P= U-1QU.

PROOF.
(a) We have R(P) n R(Q)1. = {O}, because for f E R(P) n R(Q)1.,f=l= 0 we
would have II(P - Q)fll = IIPfll = IIfll, consequently liP - QII :.> 1
would hold. By Theorem 3.11 it follows from this that dim R(P)";;
dim R(Q). The opposite inequality follows by symmetry. Hence
dim R(P) = dim R(Q). Replacing P and Q by 1- P and [- Q, respec-
tively, we obtain that dim R(l- P) = dim R(l- Q).
(b) By part (a) of Theorem 4.10 there exist unitary operators V and W
from R(P) onto R(Q) and from R(l- P) onto R(/ - Q), respectively.
Then the operator U = VP + W(l- P) is a unitary operator on H such
that U- I = V-1Q + W-I(l- Q). We have
UPU- I = (VP+ W(I - P»p(V-1Q+ W-I(I - Q»)
= VPV-1Q = Q,
since for gER(Q) we have VPV-1g=g. From this P= U-1QU
follows immediately. 0
4.6 Orthogonal projections, isometric and unitary operators 87

EXERCISES

4.29. (a) Let HI and H2 be Hilbert spaces. If there exists a surjective mapping
T E 8(HI' H2 ), then dim H2 .;; dim HI.
Hint: See Exercise 3.7.
(b) Give another proof for Theorem 4.35(a).
Hint: From IIP-QII<1 it follows that R(PQ)=R(P) and R(QP)
=R(Q).
4.30. (a) If (Um) is a sequence of isometric operators and Um ~ U, then U is
isometric.
(b) The strong limit of a sequence of unitary operators is not necessarily
unitary.
Hint: Consider the unitary operators Um on 12 defined by
Um(fn)nEN == (fm,fl,12,fJ, ... ,fm-l,fm+l,fm+2,···) for (fn)nEN E 12•
s
We have Um _ U, where
U(J,.)nEN = (0, fl,f2, 12, ... ) for (J,.)nEN E 12•

4.31. Let M and N be closed subspaces of the Hilbert space H such that dim M <
dim N. Then we have MJ.. n N~ {O}.
Hint: If PN denotes the orthogonal projection onto N, then we have MJ.. n N
=N9PN M.

4.32. Let H be a Hilbert space.

(a) If M and N are closed subspaces of H, then M+ N is not necessarily
closed.
(b) If M is a closed subspace and P is an orthogonal projection, then PM is
not necessarily closed.
Hint: Choose, for example, H= 12 , for M the subspace of the elements
(XI' XI' X2, 2X2' X3. 3X3' ••. ), for N the subspace of the elements (0, Yl' 0,
Y2' 0, Y3, ... ), and for P the projection onto N J...
5
Closed linear operators

5.1 Closed and closable operators,

the closed graph theorem
In what follows H, HI and Hz will always be Hilbert spaces. As long as no
adjoint operators (in particular no symmetric or self-adjoint operators) are
treated, we could also consider Banach spaces; the proofs may be some-
what harder, in this case. An operator T from HI into Hz is said to be
closed if its graph G(T) (cf. Section 4.4) in HI X Hz is closed. An operator
T is said to be closable if G( T) is a graph. From the proof of Theorem 4.15
we know that there exists then a uniquely determined operator T such that
G(T) = G(T); T is closed and is called the closure of T.
Let T be a closed operator. A subspace D of D(T) is called a core of T
provided that for S = TID we have T= S. One should notice that by
Theorem 4.15 the operator S is surely closable, since we have G(S)c G(T).
If T is a closable operator, then D(T) is obviously a core of f.

Proposition.
(a) T is closed if and only if the following holds: If Un) is a sequence in D(T)
that is convergent in HI and the sequence (Tf,.) is convergent in Hz, then
we have lim fn E D(T) and T(lim fn) = lim Tfn'
(b) T is closable if and only if the following holds: If Un) is a sequence in
D(T) such that fn-+O, and the sequence (Tf,.) in Hz is convergent, then we
have lim Tfn = O.
88
S.1 Closed and closable operators, the closed graph theorem 89

(c) If T is closable, then

D( T) = {f E HI: there exists a sequence (fn) from O( T) such that
fn ~ f and for which (Tfn) is also convergent},
Tf = lim Tfn for f E O(T).
(d) If T is closed, then N(T) is closed.
(e) If T is injective, then T is closed if and only if T - I is closed.

Parts (a), (b) and (c) are reformulations of the definitions. Part (d) follows
immediately from Part (a). Part (e) follows from the equality G(T- I )=
VG(T).

Theorem 5.1. Let T be an operator from HI into H2. On OCT) by

a scalar product and the corresponding norm (T-norm or graph norm) are
defined. T is closed if and only if (O(T), <. , ')T) is a Hilbert space.

PROOF. The properties of a semi-scalar product are obviously satisfied.

Because of the inequality <I,J)T). <I, f) the semi-scalar product ')T <. ,
is positive, thus it is a scalar product.
If T is closed and (fn) is a T-Cauchy sequence in O( T) (i.e., a Cauchy
sequence with respect to the T-norm), then (fn) and (Tfn) are Cauchy
sequences in HI and H2 , respectively; therefore there exist f E HI and
g E H2 such that fn~f, Tfn~g. By part (a) of the above proposition we
have f E OCT), Tf= g. We also have

II!" - fll T = {Ilfn - fll2 + II Tf" - g1l2} 1/2 ~ 0

as n~oo, i.e.,fn is convergent in (O(T), ')T)' <.,
Suppose now that (O(T), <. ,
')T) is complete. If (fn) is a sequence in
OCT) and !"~fE HI, Tfn~gE H2, then (fn) and (Tfn) are Cauchy
sequences. Consequently, (!,,) is a T-Cauchy sequence, i.e., there exists an
foED(T) such that IIf,,-foIlT~O. It follows from this that IIfn-foll~O,
liT!" - Tfoll~O, thereforef=foE O(T) and Tf= Tfo=g· 0

Theorem 5.2. Every bounded operator is closable. A bounded operator Tis

closed if and only if O( T) is closed. If T is bounded, then we have
O(T} = OCT); the closure f is the bounded extension of Tonto D(T),
constructed in Theorem 4.5.
PROOF.
1. Let (fn) be a sequence in OCT) such that fn~O and Tfn~g. Then we
have II T!"II " II Til IIfnll ~O. Therefore g = O.
90 5 Closed linear opera tors

2. For aUfE OCT) we have

11111 <; IlfliT ..;; (1 + IITII2)1/2I1fll·

Therefore (1n) is a Cauchy sequence (converging to f E D(T» if and
only if (J,,) is a T-Cauchy sequence (T-converging to 1). Consequently,
(D(T), II . II) is complete if and only if (D(T), II . liT) is complete. From
this the assertion follows via Theorems 2.4 and 5.1.
3. The equality O(f)=O(T) immediately follows from part (c) of the
above proposition. Part (c) also says that f is the extension occurring in
Theorem 4.5. 0
Proposition. Let T be closable and injective. The operator T -I is closable if
and only if f is injective. We then have T -1 ... f -I. If T - I is continuous,
then we have R(f) = R(T).

PROOF. If f.is injective, then f - I is a closed extension of T - I. If T - I is

closable, then VG(f) = VG(T)= G(T -I) is a graph, i.e., T is injective, and
we have f -I = T - I~f - I is continuous, then by Theorem 5.2 we have
R(f) = O(T-I) = O(T-I) == O(T- I)= R(T). 0

Theorem 5.3. Let T be a densely defined operator from HI into H2•

(a) T* is closed. -
(b) T is closable if and only if P is densely defined; we then have T= P*.
(c) If T is closable, then (f)* = T*.

PROOF.
(a) By Theorem 4.16 we have G(T*)=(UG(T».L. Therefore G(T*) is
closed.
(b) Since

G(T)=G(T).L.L = (U-IG(T*».L

= {(J,g)EHI XH2 : <f, T*h)-<g,h)=O forall hED(T*)},

we have (0, g) E G(T) if and only if g E O(T*).L. Therefore

(0, g) E G( T) implies g = 0 if and only if D( T*)= H2 • Consequently,
G( T) is a graph if and only if T* is densely defined. If
O( T*) is dense, then we have G( T**) = U - I( G( T*).L) =
U-IU(G(T).L.L)= G(T)= G(T).
(c) If T is closable, then we have G(T*) = U(G(T).L) = U(G(T).L) =
U( G( f».L = G« T)"'). Therefore T* = (f)*. 0

EXAMPLE 1. The operator T from Section 4.4, Example 3 is not closable, as

O(T*) = {OJ.
5.1 Closed and closable operators, the closed graph theorem 91

Theorem 5.4.
(a) An operator T from HI into H2 is closable if and only if there exists a
closed extension of T.
(b) Every symmetric operator T on the Hilbert space H is closable; T is also
symmetric.
PROOF.
(a) If T is closable, then we have T c T. Therefore T is a closed extension
of T. If S is a closed extension of T, then we have G(T) c G(S)
-G(S), hence G(T)cG(S), and thus G(T) is a graph (cf. Theorem
4.15).
(b) By part (a) the <?'perator T is closable, since T c T* and T* is closed.
For all f, g E D( T) there exist sequences (J,,) and (g,,) from O( T) such
thatf,,-f, g,,-g, Tf,,_Tf and Tg,,_Tg. As T is symmetric, we have

(TI. g) = lim (TI". g,,) - lim (I", Tg,,) == (I, Tg).

Since O( T) is dense, the operator T is symmetric. o
EXAMPLE 2. Let us consider on L2(M) the maximal operator T of multi-
plication by a measurable function t : M_C (cf. Section 4.1, Example 1).
This is the operator defined by
OCT) ... {J E L2(~) : if E LiM)} and Tf == if for f E OCT).

(5.1) T* is the maximal multiplication operator induced by the function t*

(where t*(x) = t(x)*), in particular, we have O(T*) == OCT).
PROOF. It is obvious that D(T) is also the domain of the maximal
multiplication operator induced by t*. Since for all f, g E O( T) we have

<g, Tf) ... fMg(x)*t(X)f(x) dx == fM(t*(x)g(x»*/(x) dx .... <t*g,f),

the maximal operators of multiplication by t and t* are formal adjoints of

each other. What remains is to prove that for g E O(T*) we have t*g E
L2(M). Let g E O(T*). Then for allf E OCT) we have

(T*g,/) =- (g, TI) ... fMg(x)*t(x)f(x) dx,

thus

fM(T*g(x) - g(x)t(x)*)*f(x) dx = 0 for all f E D(T). (5.2)

Let us define, for all n E N, the subsets M" of M by M" = {x EM: !t(x)! "
n}. Then we obviously have M= U:_tM". For every fE L2(M), XM,.!
92 5 Closed linear operators

belongs to D( T), consequently we have

f. (T*g(x) - g(x)t(X)*)*XM (x)f(x) dx = 0

M •
for all f E L2(M).

Since XM (T*g - t*g) E L2 (M), it follows that T*g(x) = t*(x)g(x) almost

everywhe"re in M". Since this holds true for all n EN, we have T*g = t*g,
therefore t* g E L2 ( M). ' 0

(5.3) T is closed.
PROOF. T is the adjoint of the maximal multiplication operator induced by
t*, hence T is closed by Theorem 5.3(a). 0
(5.4) The following assertions are equivalent:
(a) T is self-adjoint,
(b) T is symmetric,
(c) t is real-valued (i.e., t(x) is real almost everywhere in M).
PROOF. (b) obviously follows from (a).
(b) implies (c): If t is not real, then at least one of the sets M\ = {x E
M : 1m t(x) > O} or M2 = {x EM: 1m t(x) < O} has positive measure. If M\
is of positive measure, then for all f E OCT), different from zero and
vanishing outside MI , we have

1m <f, Tf) = f.
MI
If(xW 1m t(x) dx > O.
By Theorem 4.18 the operator T is therefore not symmetric. We can argue
similarly if M2 has positive measure.
(c) implies (a): Since t= t*, we have T= T* by (5.1). 0
(5.5) If M is an open subset of RM and t is locally bounded on M (i.e., t is
essentially bounded on any compact subset of M), then Co""(M) and L2• o(M)
are cores of T.
PROOF. We obviously have CoOO(M) C L2• oeM) C OCT). We prove the asser-
tion for CoOO(M); the other assertion follows from this. We have to prove
that for each f E D( T) and for each ( > 0 there exists an 1. E Co""( M) such
that II!. - fll + II T!. - Tfll .;;; E. If (M,,) is a sequence of open bounded
subsets of M such that M" c M,,+\ and M= U Mno then for all n E N let us
define

We obviously have gn E L2 . oeM), g,,~f, and Tg,,~ Tf. Therefore there

exists an no E N such that II gno - fll + II Tg"o - Tfll .;;; E /2. On the other
hand, there exists a sequence' fm E COOO ( M"J c C?OO( M) such that fm ~ g"o'
Then we also have that Tfm~Tg"o' So there eXIsts an moEN such that
5.1 Closed and closable operators, the closed graph theorem 93

IIfmo - gnoll + II Tfmo - Tgnoll "(/2. The assertion follows from this by taking
1. = f mo' D
Let T and S be operators from HI into H2 and from HI into H3 ,
respectively. The operator S is said to be T-bounded if D( T) c D(S) and
there exists a C;> 0 such that IISfll" CllfllT for allfE D(T), i.e., if S, as
an operator from (D(T), <. , .)r) into H3 , is bounded. Then for all f E
D(T) we have
IISfl1 .;;; COIfIl + II Tfll)·
If S is T-bounded, then the infinum of all numbers b ;> 0 for which an
a ) 0 exists such that
118fll "allfll + bllTfll foraH f E D(T),
is called the T-bound of 8. One should notice that if e is the T-bound of S,
then in general there exists no a;> 0 such that for all fED( T) we have
II Sf II <; allfll + ell Tfll (cf. Exercise 5.4).
Proposition. If T is an arbitrary operator from HI into H2 and 8 E B( HI> H3),
then 8 is T-bounded with T-bound O.

Theorem 5.5. Let T and S be operators from HI into H2 , and let S be

T-bounded with T-bound less than 1. Then T+ 8 is closed (closable) if and
only if T is closed (closable); we have D(T + S) = D(T).
PROOF. As the T-bound of S is less than I, there exist a b < 1 and an a ;> 0
such that 118fll" allfll + bll Tfll for all f E D(T). Consequently, for all
fE D= D(T)= D(T+ 8) we have
- allfll + (1- b)1I TJII ..;; II Tfll - 118fll " II(T+ 8)j1l
..;; IITfll + 118fll " allfll + (1 + b) II Tfll·
From this it follows with a properly chosen C ;> 0 that
II Tfll " c(lIfll + II(T+ S)fll) (5.6)

II(T+ 8)fll " c(lIfll + II Tfll) (5.7)

for all fED. Hence there exists a K ;> 0 such that
IIfliT <; Kllfllr+s and IIfIlT+s" KllfliT'
From this it follows that (D, (. , ')T+S) is complete if and only if (D,
(. , ')r) is. Let T be closable. If (J,,) is a sequence from D(T+ 8) = D(T)
«
such that f" ~O and for which T + S)fn) is convergent in H2 , then by (5.6)
the sequence (Tfn) is a Cauchy sequence. Hence Tfn~O, because T is
closabie. Because of (5.7) from this it follows that (T+ S)fn~O, so T+ S is
also closable. One can show in an analogous way that T is closable
provided that T+ S is closable. By part (c) of the proposition preceding
94 5 Closed linear operators

Theorem 5.1 we have f E D( T + S) if and only if there exists a sequence

(J,,) from D(T+ S) = D(T) for which fn~f and «T+ S)fn) is convergent
Since because of (5.6) and (5.7) «T+ S)f,J is convergent if and only if
(Tfn) is convergent, we have D(T + S) = D(T). 0

Theorem 5.6 (Banach; closed graph theorem). Let HI and H2 be Hilbert

spaces and let T be an operator from HI into H2• Then the following
statements are equivalent:
(a) T is closed and D(T) is closed,
(b) T is bounded and D( T) is closed,
(c) T is bounded and closed.
PROOF. (a) implies (b): We have to show that T is bounded. Without loss
of generality we may assume that D( T) = D( T)= HI (otherwise we could
consider T as an operator from the Hilbert space D(T) into H2)' Conse-
quently, T* is defined. For all g E D( T*) such that II gil < 1 we have
I(T*g,J)1 =I(g, Tf)l< IITfll for all f E HI'
For the linear functionals {Lg : g E D(T*), II gil' I} on HI. where LJ=
(T*g,f), we therefore have
ILiJ)1 ' II Tfll for all f E HI;
consequently they are pointwise bounded. By Theorem 4.22 there exists a
C .., 0 such that
II T*gll = IILgll <C for all g E D(T*) such that II gil , 1.

Therefore T* is bounded and II T*II ,C. As T is closed (consequently

closable), D(T*) is dense and, by Theorem 5.2, closed. We therefore have
D(T*) = H2• i.e., T* E 8(H2' HI)' Since T is closed. this implies that T= T
= T** E B(H" H 2).
The assertions "(b) implies (c)" and "(c) implies (a)" are contained in
Theorem 5.2. 0

Theorem 5.7. Let HI and H2 be Hilbert spaces and let T be an operator from
HI into H2 such that D(T) = HI and D(T*) is dense in H 2• Then T belongs to
B(H .. H2)' In particular, every symmetric operator T on the Hilbert space H
such that D( T) = H is bounded (Hellinger-Toeplitz).
PROOF. By Theorem 5.3 the operator T is closable. Because D(T) = HI' we
have D( T) = D( T), Le., T = f. Therefore T is closed and D( T) = HI' Then
T is bounded by Theorem 5.6. 0
Theorem 5.S. Let HI and H2 be Hilbert spaces and let T be an injective
operator from HI into H2 such that R(T) = H2 • The operator T is closed if
and only if T- I E B(H I , H2)'
5.1 Closed and closable operators, the closed graph theorem 95

PROOF. By Part (e) of the proposition preceding Theorem 5.1 the operator
T is closed if and only if T - I is closed. The assertion follows immediately
from this and from Theorem 5.6. 0

Theorem 5.9. Let H.. H2 and H3 be Hilbert spaces, let T be a closed operator
from HI into H2, and let S be a closable operator from HI into H3 such that
O(S)::> OCT). Then S is T-bounded.

PROOF. On account of Theorem 5.6 it is enough to show that the operator

So from (O(T), (. , '>r) into H3 , defined by O(So) = OCT) and Sof= Sf for
fE OCT), is closed. As O(So) = OCT), it is enough to show that So is
closable. Let Un) be a sequence in OCT) for which IIfnIlT~O and (So'/") is
also convergent. Since II fn II T~O implies Ilfn II ~O and since S is closable,
we obtain from this that So'/" = S./" ~O, thus So is closable. 0

EXERCISES

5.1. (a) Every Carleman operator (cf. Exercise 4.2) is closable.

(b) Any Carleman operator, defined on the whole of L 2(M), is bounded.
5.2. (a) Any densely defined operator T on the Hilbert space H such that
Re (f, Tf) ;;;. 0 for all f E D( T) is closable.
Hint: If (0, g) E G( T), then Re <f, Tf + zg) ;;;. 0 for all fED( T) and
z E K One infers from this that (f, g) = 0 for all fED( T).
(b) The numerical range of an operator T on H is defined by W(T)=
{<J, Tf) : f E D(T), IIfil = I}. The set W(T) is convex (cf. for example,
P. R. Halmos [14], Problem 166). If D(T) is dense and W(Th06. K, then T
is closable.
5.3. (a) Let H be a Hilbert space, and let T be a densely defined operator from H
into em. The operator T is bounded if and only if it is closable (this holds
true for linear functionals in particular).
(b) In the Hilbert space 12 let the functional T be defined by D( T) = 12,0>
Tf = 'i.nf" for f E D( T). Then T is not closable.
5.4. Let T and S be on L2(1R) the maximal operators of multiplication by I(x) =- x 2
and s(x) = x 2 + x, respectively. The operator S is T-bounded with T-bound I.
However, there exists no a;;;. 0 such that IISfll.;;; II Tfll + allfll for aIlfE D(T).
5.5. Prove the results of (4.3) with the aid of the closed graph theorem,
(4.1), (4.2), (4.6), and (5.3).
5.6. Let HI' H2 and H3 be Hilbert spaces, let S E B(HI , H2 ), and let T be a closed
operator from H2 into H) such that R(S) C D(T). Then we have TS E
B(H I , H3)'
Hint: Show that TS is closed, and D(TS) = HI'
5.7. Assume that HI' H2 and H) are Hilbert spaces, T is a closable operator from
HI into H2 , and S is a T-bounded operator from HI into H3 • If the seauence
w w tv
(f,,) from D( T) is such that fn ~ 0 and Tf" ~ 0, then we also have Sfn ~ o.
96 5 Cloned linear operators

5.S. Let H be a Hilbert space. An operator P on H is an orthogonal projection if

and only if P is symmetric, D(P)= H, and p2= P.
Hint: Use Theorem 5.1 (Hellinger-Toeplitz).
5.9. Let H be a vector space, and let (. , .>, and (. , '>2 he scalar products on H
such that (H, (. , .>,) and (H, (. , '>2) are complete. If there exists a C such
that II . lit <; GIl . Ib, then there also exists a C' such that II . 112 <; C'II . II,.
5.10. Let the vector space H be endowed with two scalar products, consequently
with two norms II . II, and II . lb. These norms are said to be coordinated if we
have: (k'2) from IIxnll,-O and IIxn-xlh-O it follows that x=O, or (k 2 ,)
from IIxnlb-O and IIxn - xllt-O it follows that x =0.
(a) The assumptions (k'2) and (k 2 ,) are equivalent.
(b) If the norms II. II, and II· Ib are coordinated, and (H, II . II,) and
(H, II . 112) are complete, then there exist c" C2 > 0 such that II· lit .;;;
ctll·lb<c211·11,·
Hint: The identity map from (H, II . II,) into (H, II . IIv is closed.
5.11. (a) Let H, and H2 be Hilbert spaces. Assume that H, is finite dimensional.
Every operator T from H, into H2 is continuous.
Hint: If {e" ... , em} is an ONB of D(T), then Tf= "£ }(e pf> Te j for all
fE OCT).
(b) If (. , .>, and (. , '>2 are scalar products on a finite dimensional vector
space H, then there exist constants c" C2;;' 0 such that II . II, .;;; ctll . Ih <
c211 . II, (II . IIj denotes the norm defined by (. , .>j)'
5.12. Let T be an operator from H, into H2, and let S be a continuously invertible
operator from H2 into H3 • If Sand T are closed (closable), then ST is also
closed (closable).
5.13. If S is T-bounded with T-bound b, then for any £ > 0 there exists an a ;;. 0
such that II Sfll 2 <; a 2 11 fll2 + (b 2+ £)11 Tfll2 for all fED( T). If we have II Sfll 2 <;
a 2 11f1l2+ b211Tf1l2, then we also have IISfll <allfll + bIITfll··

5.2 The fundamentals of spectral theory

In what follows an operator T from HI into H2 will be said to be bijective if
T is injective and R( T) = H2 •

Theorem 5.10. Let Sand T be bijective operators from HI into H2• If

O(S) c OCT), then

If O(S) = OCT), then

T- I - S-I = T-I(S- T)S-I = S-I(S-T)T-I.

It is enough to prove the first assertion. We shall prove that

PROOF.
T-I=S-I+T-I(S-T)S-I. Since T-'Tf=f for all fEO(T) and
S.2 The fundamentals of spectral theory 97

SS - Ig = g for all g E H2 , it follows that

S-I + T-I(S- T)S-I = T-ITS- I + T-I(S- T)S-I
= T-I(T+ S- T)S-I = T-ISS- I = T- I . 0
Now we show that a closed bijective operator remains bijective even
after the addition of a "not too big" operator.

Theorem 5.11. Assume that HI and H2 are Hilbert spaces, T is a closed

bijective operator Jrom HI into H2, S is an operator Jrom HI into H2 such that
D(S)=> D(T), and liST-III < 1. Then T+ S is also bijective, and we have
00 00

(T+ S)-I = L (-ItT-I(ST- I )" = L (-I)n(T-IS)"T- I ; (5.8)

n=O n=O

the series are convergent in the norm oj B( H2, HI)'

PROOF. For all J E D(T) we have
IISJII = IIST-'TJII ..;; II ST - I IIIITJII,
i.e., S is T-bounded with T-bound less than l. By Theorem 5.5 the
operator T + S is closed, too. Moreover, T + S is injective, because

II(T+ S)JII ;> IITJII -IISJII ;> (1- II ST - I II)II TJII > °
for allJE D(T+ S)= D(T),J*'O.
The two occuring series are obviously identical term by term. Let us
define
p
Ap = L (-ltT-I(ST- I )", pEN.

Then for q >p we have

IIAq - Apll f (- ItT-I(ST-I)nll

=11 n-p+1
q-p-I
"IIT- I IIIIST- l ll p+ 1 ~ IIST- Ill n
n=O
00

" IIT-IIIIIST-IIIP+I L IIST - ' ll n ..;; CIiST-III P + ' ,

where C= IIT-III~::"_oIIST-llin. From this it follows that (Ap) is a Cauchy

sequence in B(H2' Hd, consequently there exists an A E B(H2' HI) such
that Ap~A asp~oo.
We have (T+ S)A = I:
p
{T+ S)Ap = L {-I)n{T+ S)T-I(ST-It = 1+ (-IY(ST-1y+1 ~ I,
n-O
98 5 Closed linear opera tors

as p-+oo. Therefore for all g E H2 we have

Apg -+ Ag and (T+ S)Apg -+ g as p -+ 00.

As (T+ S) is closed, Ag E D(T+ S) and (T+ S)Ag = g, i.e., (T+ S)A = I.

In particular, R(T+ S)= H2, hence T+ S is bijective and we have A ...
(T+ S)-I. 0

Corollary. The statements of Theorem 5.11 hold in particular if S is bounded

and liS II < liT-Ill-I (then we have liST-III" liS II liT-III < 1).

Corollary. Let HI and Hz be Hilbert spaces and let T and Tn (n E N) be

linear operators from HI into Hz such that D(T) C D(Tn) for all n E N.
Assume that T is closed, bijective, and II(T- TII)T-JII-+O as n-+oo. Then
there exists an no E N such that Til is bijective for n >no and II TII- J - T -III
-+0 as n -+ 00. (The assumptions on Til hold in particular if the operators
T- Til are bounded and IIT- Tnll-+O as n-+oo.)
PROOF. With Sn = Tn - T we have IISIIT-III-+O as n-+oo. Therefore there
exists an no E N such that II Sn T - III < 1/2 for n >no' Hence by Theorem
5.11 the operator Til ... T + Sn is bijective for n > no and we have
00 00

IITII- I - T-Ill" ~ IIT-I(SIIT-Itll " IIT-JIlIISnT-JIl ~ 21- m -+0

m-I m-I

as n-+oo. o
In what follows let H be a Hilbert space over K and let T be an operator
on H. The number z E K is called an eigenvalue of T if there exists an
f E D( T), f -:f:: 0 such that Tj = zj, i.e., if the operator z - T = zl - T is not
injective, N(z - T):;I= {O). The subspace N(z - T) is called the eigenspace of
z, the dimension of N(z - T) is called the multiplicity of the eigenvalue.
The element j is called an eigenelement or eigenvector oj T belonging to the
eigenvalue z. If z is not an eigenvalue (i.e., (z - T) is injective), then the
operator
R(z, T) = (z - T)-J
is well-defined. The set
p(T) = {z E K : z - T is injective, and R(z, T) E B(H)} (5.9)
is called the reso'lvent set of T. If T is not closed, then z - T and R(z, T)
are not closed, consequently p( T}= 0. This is why in most of the following
cases we shall assume the closedness of T. For a closed operator T on
H we have by the closed graph theorem that
p(T) = {zEK:z-T is bijective}. (5.10)
S.2 The fundamentals of spectral theory 99

The function
R(., T) : p(T)-+B(H), z R(z, T)
is called the resolvent of T. For any z Ep(T) the operator R(z, T) is called
the resolvent of T at the point z. The set
aCT) == K \ p(T) ... CKP(T) (5.11)
is called the spectrum of T. The set ap(T) of all the eigenvalues of T is
obviously contained in aCT). The set ap(T) is called the point-spectrum of
T.
Theorem 5.12. Let T be a densely defined operator on H. Then a(T')-
a(T)* and p(T*) == p(T)* (here for any subset M of the complex numbers
M*={z*: zEM}).
PROOF. Because of (5.11) it is enough to prove that p(T):ap(T')*. To
prove this it is enough to show that p( T) C p( T*)*, since because of the
equality T**= Twe also have p(T*)Cp(T)*, and thus p(T*)*cp(T).
Let z E p(T). Then z - T is densely defined and bijective. By Theorem
4.17(b) the operator z*- T*==(z- T)* is therefore injective, also, and we
have (z* - T*) -I = «z - T) - 1)* E B( H). Hence z* - T* is bijective, i.e.,
z* Ep(T*), and thus z Ep(T*)*, 0
Theorem 5.13. Let Sand T be closed operators on H.
(a) For all z, z' Ep(T) we have the first resolvent identity
R(z, T) - R(z', T) = (z' - z)R(z, T)R(z', T)

= (z' - z)R(z', T)R(z, T);

in particular, R(z, T) and R(z', T) commute.
(b) If O(S) C OCT), then for all z E peS) n p( T) we have
R(z, T) - R(z, S) = R(z, T)(T- S)R(z, S).
(c) If O(S)=O(T), then for all zEp(S)np(T) we have the second re-
solvent identity
R(z, T) - R(z, S) = R(z, T)(T- S)R(z, S)

= R(z, S)(T- S)R(z, T).

PROOF. The first resolvent identity follows from Theorem 5.10 if in there
we replace T by z - T and S by z - T. The second resolvent identity
follows similarly if T and S are replaced by z - T and z - S, respectively.
o
Theorem 5.14. If T is a closed operator on the Hilbert space H, then p(T) is
open, consequently a( T) is closed. More precisely, if Zo E p( T), then z E p( T)
100 5 Closed linear operators

-I
all z E such that -' zol R(zo, these we have
00

If T E B( H), then we have {z ElK: Izl > II Til} c p( T); the spectrum a( T) is
compact. Furthermore,

R(z, T) = ~ zn-IT n for Izl > II Til; (5.12)

n=O
series called the von Neumann series.
PROOF. Let Zo E peT), and let Iz - zol < IIR(zo, T)II-I. If in Theorem 5.11
we replace T by Zo - T and S by (z - zo)I, then it follows that Z - T= zo-
(z - bijective, consequently z peT). Moreover,. Theorem 5.11
we have

R(z, = «zo T) +

Now let T E B(H) and let Izl;> II Til. If in Theorem 5.11 we replace T by zI
by , then i~ follows - T bijective, therefore Ep(T), and

(z - T)-I = ~ z-n-ITn.
n-O

Hence a( T) C { z ElK: Iz I ..;; 1\ Til}. As a( T) is closed, the compactness of

a( T) follows from this. 0
'l'heorem Let be a operator on Hilbert space The
resolvent R(. , T) : p(T)~B(H) is a continuous function (i.e., for any ZoE
and any sequence (zn)from peT) sitch that zn~zO we have IIR(zll' T)-
T)U If a(T) non-empty, then every Ep(T) have

IIR(z, T)II ;> d(z, a(T»-I.

For every sequence (zn) from peT) such that zn~zO' Zo E a( we therefore
have IIR(zn' T)1I~oo.
Let z E p( such Iz- IIR(zo' T)II-I. Then by Theo-
rem 5.14 we have
00

IIR(z, - R(zo, T)II

As the right side is small when z is close to zo' the continuity at Zo follows
from for any Ep(T). If z E then Theorem 5.14 point z'
S.2 The fundamentals of spectral theory 101

also belongs to p(T) for all z' E K such that Iz' - zl < IIR(z, T)II-I. Conse-
quently, Iz' - zl ;> II R(z, T)1I- 1 for all z' E a(T), and thus
IIR(z, T)II-I < inf (lz'-zl: z'Eo{T)} = d(z, o(T». 0

Let G be an open subset of C, and let X be a Banach space. A function

F: G--,;X is said to be holomorphic if for every Zo E G there exist an r > 0
and a sequence Un) from X such that
00

F(z) = ~ (z - zo)nfn for Iz - zol < r;

n-O
where the convergence has to be understood in the sense of the norm of X.
As in function theory, one can prove that the quantity

ro=[limsup IIfnlll/n]-1 (5.13)

n--.oo

is the radius of convergence of the above series and that the series is
uniformly convergent on each disc around Zo of radius less than ro, while it
is divergent for all z such that Iz - zol >ro (cf., for example, K. Jiir-
gens [191, §4.4). Every holomorphic function is continuous (cf. the proof of
Theorem 5.15).

Theorem 5.16. Let T be a closed operator on the complex Hilbert space H,

and let f, g E H. Then the functions

R(. , T) : p(T)--,;B(H), z~R(z, T)

R{. , T)f: p{T)--,;H, z~R{z, T)f
<g, R(. , T)f> : p(T)--,;C, z~<g, R(z, T)f)

are holomorphic.
PROOF. Let Zo E p(T), and let r = II R(zo, T)II- I • Then for all z E C such
that Iz - zol < r we have
00

R(z, T) = ~ (zo-zfR(zo, Tf+1

n=O

in the sense of the norm convergence in B(H),

00

R(z, T)f = ~ (zo- zfR(zo, Tf+lf ,

n-O

in the sense of the norm convergence in H, and

00

<g, R(z, T)f> = ~ (zo-zf<g, R(zo, Tf+lf ).

n=Q
102 5 Closed linear operators

According to the above definition, these three functions are therefore

holomorphic. 0

Theorem 5.17. Assume that H is a Hilbert space and T E B(H), r(T)=

lim sup IIrlil/".
(a) We have r(T)'" II Tmil i / m for all mEN, and thus r(T)=lim IIT"II I /".
(b) We have r(T) '" II Til and o(T) C {z Ell(: Izl '" r(T)}. For all z E II( such
that Izl >r(T) the operator R(z, T) is given by the Neumann series
(5.12).
(c) If H is complex, then o(T) is non-empty and there exists a z E o(T) such
that Izl=r(T), i.e., we have r(T)=sup {Izl: zEo(T)}.
(d) The statement in part (c) holds for any self-a4Joint T in a real Hilbert
space, as well.
r(T) is called the spectral radius of T.
REMARK. Theorem 5.17(c) does not hold for real Hilbert spaces, as the

example of the operator defined by the matrix (~ 1 ~) on 1R2 shows.

PROOF.
(a) Let mEN. Every n E N can be uniquely represented in the form
n = mPn + qn with Pn' q" E Nand qn < m. If we denote C =
max p, II Til, IIT211, ... , IITm-lll}, then it follows that
IIT nll '" IIT mIl Pn IITQ.1I '" CIITmIlP.,
and thus
r(T) '" lim sup cl/nIlTmU<l/m)-(I/"m)q. = II Tmll i / m.
n-->O()

For every nEN we therefore have r(T)'" II Tnll l/ n, and thusr(T)<;

lim inf IITnlll/n, i.e., lim IIT"II I/ n exists and r(T)=lim II Tnil i / n •
(b) The radius of convergence of the series ~~=oun+ IT n, U E II( is obvi-
ously equal to r( T) - I, because of (5.13). Therefore for all z E II( such
that Izl >r(T) the operator A(z) = ~~_oz -,,-IT" E B(H) is defined; the
series is absolutely convergent in the norm of B(H) (i.e., for every t: > 0
there exists an no E N such that ~':=klz -,,-III/Tn II < f for no <; k <; m).
One can verify easily that
0() 0()

(z - T)A{z) = A(z)(z - T) = L z-nT n - L z-,,-IT n+1 = I,

i.e., A(z) = (z - T)-I and z Ep(T). The Neumann series is therefore

convergent for all z E II( such that Izi >r(T) and it represents the
operator (z - T) -I.
(c) Let us assume that o(T)=0. Then p(T)=C, and by Theorem 5.16 the
function Ffg : z~<J, R(z, T)g> is an entire function for all f, g E H.
Because of the inequality lI(z - T)fll ;;;. (11zll -II TIDllfl1 we have
IIR(z, T)II"'(llzll-IiTID-1 for Izl>IITII, and thus Ffg(z)~O as Izl~
S.2 The fundamentals of spectral theory 103

00. Consequently, Fjg is bounded, and by Liouville's theorem it is

constant. Since Fjiz)~O as Izl~oo, we have Fjg = 0 for all f, g E H,
therefore R(z, T) =O. This is a contradiction. It remains to be proved
that reT) = sup{lzl : z E aCT)} = roo As aCT) is closed and we already
know the inequality ro ~ r(T), it remains to show that reT) ~ roo For
arbitrary f, gE H the function zl-+<f, R(z, T)g> is holomorphic in
{z E C : Izl > ro} and can be uniquely expanded in a Laurent series
there. By (5.12) this Laurent series has the form
00

<J, R(z, T)g> = ~ z-m-I<f, T"'g>.

m-O

Let s >ro. Then the sequence (s-m-I<J, Tmg» is bounded for all
f,g E H. By applications of the Banach-Steinhaus theorem (Theorem
4.22), first to the functionals induced by s - m-ITmg and then to the
operators s-m-1T m, we obtain a C ;;. 0 such that
Is-m-11IITmll '"' C foraH mE~.
This implies that limllTmlll/m ,",s. Since this holds for all s>ro, it
follows that reT) ~ roo
(d) If T is self-adjoint, then by Theorem 4.4(b) there exists a sequence Un)
from H such that II In II = I and <Jn' Tfn>~11 Til or <In, Tfn>~ -II Til·
In the first case it follows that

11(11 TII- T)fnll 2 = II Til 2 II fn 112 - 211 TII<Jn' Tfn> + II Tfnl1 2

~ 1/ TI/ 2 - 21/ TII(t", Tin> + II TII2 ~ 0 as n~ 00.

Analogously, in the second case it follows that

11(11 Til + T)InIi ~ 0 as n~ 00.

Consequently, "Til E aCT) in the first case, and -II Til E oCT) in the
second case. 0
EXAMPLE 1. Let M be a measurable subset of Hm , let I: M~C be
measurable, and let T be the maximal operator of multiplication on LlM)
by t.

(5.14) We have o/T) = {z EC : I-I(Z) has positive measure}, here rl(z)

denotes the sel {x EM: t(x) = z}. (See also Exercise 5.23.)
PROOF. This can be obtained immediately from the results of (4.2) if we
notice that z - T is the multiplication operator generated by z - t. 0
(5.15) We have z E oCT) if and only if for every t: > 0 the sel
1-1({wEC: Iw-zl<t:})={XEM: It(x)-zl<t:} is of positive measure.
In particular, if M is open and t is continuous, then oCT) is the closure of the
range of t. .
104 5 Closed linear operators

. PROOF. Let X = {z E C : {X EM: It(x) - zl < I;} have positive measure for
every I; > O}.
We show that p(T)=CX. We have zEp(T) if and only if z- Tis
bijective, thus if and only if z - T is injective and R(z - T) = L2(M). By
(4.3) this holds if and only if there exists a c > 0 such that Iz - t(x)1 :> c
almost everywhere on M, i.e., if and only if r I( {w E C : Iw - zl <c}) has
measure zero. Consequently, z E p( T) if and only if z fl.1:. 0
The corresponding results hold if we consider the real Hilbert space
L2• R( M) and a real function t.

EXAMPLE 2. If U is a unitary operator on H, then a( U) c {z E C : Izi = I}.

PROOF. By Theorem 5.14 we have {z E C : Izl> I} c p( U). Since U is
bijective and U- I is unitary, we have OEp(U), and thus by Theorem 5.14
{z E C : Izl < I} C p( U). The assertion follows. 0

EXERCISES

5.14. Let HI and H2 be Banach spaces. An operator T from HI into H2 has a

continuous inverse (not necessarily defined on the whole space H2) if and
only if
'( .. inf {IITxll : xED(T).lIxll > I} > O.
We have then that liT-III = ,(-I.
5.15. Let (a,,) be a sequence from C. On 12 by
D(T) = {J- (J;,) E 12 : (a..!,,) E 12 }.
Tf = (a..!,,) for f = (f,,) E D(T)
a closed operator is defined. T is self-adjoint if and only if the sequence (a,,)
is real. We have aiT) = {a" : n EN}. a(T) = ap ( T). Determine ,R(z, T) for
z fI. aiT).
5.16. If Sand T are from B(H). then r(ST) - r(TS).
5.17. Assume that HI and H2 are Hilbert spaces, and T is a bijective operator from
HI onto H2 •
(a) If S is an operator from HI into H2 such that D( S) ::> D( T). ST - I E
B(H2 ). r(ST- I ) < I. and T+ S is closed. then T+ S is also bijective.
(b) If S E B(HIo H2) and r(T-IS) < I. then T+ S is bijective.
In both cases we have
co co
(T+ S)-I = ~ (-l)"T-I(ST- I )" - ~ (-I)"(T-IS)"T- I,
,,-0 ,,-0
where the series converge in the norm of B(H2 • HI)'
Hint: After showing the convergence of the series, denote the sum by A. and
prove that (T+ S)A -A(T+ S)= I. Compare with the proof of Theorem
5.1I, as well.
S.2 The fundamentals of spectral theory 105

5.18. Let HI and H2 be Hilbert spaces, let A, An E B(HI' Hl ), and let An be bijective
for all n E N. If An ~ A and IIAn-11i < C for all n E N, then A is also bijective
and we have An-I~A -I.
Hint: Use Theorem 5.11.
5.19. Assume that HI and Hl are Hilbert spaces; Tn(n EN) and T are closed
bijective operators from HI onto H2 ; the sequence (II Tn-III) is bounded, and
for some core D of T the following holds: For every xED there exists an
n(x) E N such that x E D(Tn) for n;;;. n(x) and also Tnx-) Tx for n~oo. Then
r n- I ~S T- I •
Hint: From Theorem 5.10 it follows that Tn-lg~T-lg for all gET(D);
(II Tn-III) is bounded, it follows that Tn-Ig~ T-Ig
T(D) is dense in Hz; since
for all g E H2 •
5.20. Prove that in Exercise 5.19 no assumptions can be removed.
s •
(a) If T, Tn are bijective and Tn ~ T, then we do not necessarily have
r n-t S
~
T- I
.
Hint: Consider, on 12 , the operators T = I and T,J == (fl' ... ,/n'
(1/ n)J" +l> (1/ n)fn + 2, . . • ); (the sequence II Tn- III is not bounded).
S
(b) If T is bijective and Tn ~ T, then Tn is not necessarily bijective for large
n.
Hint: Consider the operators T = I and T,J ... (fl' ... ,j", 0, 0, ... ) on
12 ,
(c) If the Tn are bijective, (II Tn-III) is bounded and Tn ~ T, then T is not
necessarily bijective.
Hint: Consider the operators T,J-( ... ,/-n-2'/-n-I'/0'/-I'/t'/-2,
fl" .. , f-n>fnJn+l,fn+2' ... ) on /2(1).

5.21. Assume that H is a Hilbert space over IK, U is an open subset of IK, and
R : U~B(H) satisfies the properties (i) R(zt) - R(zz)'" (Z2 - zt)R(zt)R(zz)
for ZI' Zz E U, (ii) N(R(z» = (OJ for all z E U. Then we have
(a) R(zt)R(zz) == R(zz)R(zt) for all Zt> Zz E U,
(b) With K(z)==R(z)-t(zEU) we have K(zt)-zl=K(zz)-zz for all zt>zz
E U; i.e., we can define T= z - K(z).
(c) T is dosed; we have U cp(T) and R(z)= (z - T)-I for z E U.
5.22. Assume that HI and Hz are isomorphic Hilbert spaces, M is a dense subspace
of Hz. and M* H2 • Then B(HI' M)* B(HI' Hz). (From this it follows that
B(HI' H) is in general not a completion of B(HI' H) if H is a completion of
the pre-Hilbert space H.)
Hint: Let U be an isomorphism from Ht onto H2 • Assume that there exists a
sequence (Un) from B(Hl> M) such that Un~ U. Then R( Un) = Hz for large n.
5.23. Let T be the operator of multiplication by a measurable function t on Lz(M)
(cf. Example I). Every eigenvalue of T has infinite multiplicity.
5.24. Let T E B( H), let NT = (/ E H : T"f~O as n -) 00 }, and let BT == (/ E H : the
sequence (T"f) is bounded}.
(a) If r(T) < 1, then NT = H.
(b) If BT = H, then r(T).;; 1.
106 S Closed linear operators

Hint: Use Theorem 4.22.

(c) From NT - H it does not follow in general that r(T) < 1.
Hint: Consider the operator Tf(x) - xf(x) on L 2(O, I).

5.25. If T E 8(H), ~ E II( such that I~I-II Til, and u E H is such that Tu - ~u, then
T·u-~·u.
Hint: Show that II T·u - ~·uIl2 co: 01

5.26. Let A E 8(HI' H2) and B E 8(H2. HI)'

(a) For ~~O we have ~Ep(AB) if and only if ~Ep(BA). In this case we
have (A - BA)-I. *(f
+ B(A-AB)-IA).
(b) Assertion (a) does not hold for ~ - O.
Hint: Let A be an isometric operator such that R(A)=FH2 , and let
B-A·.
(c) Assertion (a) holds for ~ .. 0, as well, provided that at least one of the
operators is bijective.
Hint: AB-A(BA)A- I .

5.27. Let T be an operator on H with non-empty resolvent set, and let ~oEp(T).
We have AEa(T) if and only if (~_~O)-I Ea(R(Ao, T»; for all ~Ep(T) we
have R(A, T) - R(Ao, T)(I - (Ao - A)R(Ao, T»-I - (l - (Ao - A)R(Ao,
T»-IR(Ao, T). In particular, a(R(Ao, T» - {(A -Ao)-I : A E a(T)}.
Hint: Use the first resolvent identity.

5.28. An operator T E 8(H) is said to be nilpotent if there exists an n E N such that

Tn_o. The operator T is said to be quasi-nilpotent if II Tn II I/n ..... o, i.e.,
r(T)-O.
(a) Every nilpotent operator is quasi-nilpotent.
(b) If T is a quasi-nilpotent operator on a complex Hilbert space, then we
have a(T)'" {O}.
(c) If k(x, y) is continuous on 0 co: y co: x co: 1, then the Volterra integral
operator K defined on L2(0, I) by D(K) ... Ll(O, I) and (Kf)(x)-
f 3k(x, y)f(y) dy is quasi-nilpotent.
Hint: Put k(x,y) ... 0 for x <y and M - max {lk(x,y)1 : 0 co:y co: x co: I};
then Ik(II)(X, y)1 co: (Mn/(n - 1)!)lx - yl,,-I for the kernel k(") of K", con-
sequently 11K" II co: (M" /(n - I)!).
5.29. Let H be a complex Hilbert space, let T E 8(H), and let p(t) - Ii_oa }t} be a
polynomial with p(T) ... Ii_oa }T}'" 0 (TO ... f).
(a) If p(z)~O, then zEp(T) and R(z. T)=[p(z)]-lq(Z, 7') with q(z, T) ..
Ij.:Jz}I~_}+ lak Tk-}-I.
(b) If p i~ a polynomial of minimal degree, then a(T) is the set of zeros of p.

5.30. Let T be a closed operator on the Hilbert space H. Then the function
zl-+R(z, T) is holomorphic on p(T) as a function with values in 8(H, D(T»,
where D( T) is equipped with the T-norm.
Hint: Use the inequality II[R(zo, T)]n+ IIIS(H. D(T» .;;; IIR(zo. T)IIB(H. D(T» X
IIR(zo, T)II" and Theorem 5.14.
S.3 Symmotric and solf-adjoint operators 107

5.3 Symmetric and self-adjoint operators

Theorem 5.18. Let T be a Hermitian operator on the pre-Hilbert space H.
Every eigenvalue 0/ T is real; eigenvectors belonging to different eigenvalues
are orthogonal. If H is comple:,c, then for any z E C\R the operator z - Tis
continuously invertible. and we have IIR(z. T)II < 11m zl-I (this holds true in
particular for symmetric and se/f-a4joint operators).
PROOF. Let z be an 'eigenvalue of T and let /E N(z - T). /=FO. Then
z*llIlIl =(T/.f> == (f. T/> = zllflil. thus z == z*. i.e., z E R. If z •• z2 ~re two
distinct eigenvalues and fl' fl are corresponding eigenelements of T, then,
as the Z j are real. we have (z I - zJ(/I' fl> =(T/I.Jl > - (fl' T/l > =O.
Therefore (fl' /l> -= o. With z ... x + Qi (x, y E R) we have for all f E OCT)
that
lI(z - T)flll - lI(x - T)/ + Qifll2

- lI(x - T)f1l2 + IY1211/lll ;> I1m zI211/112.

For z E C' R it follows from this that (z.- T) is i~ective and that for
g - (z - T)f E O(R(z. T» we have
IIR(z. T)gll = II/II < I Imzl-llI(z- T)/II = I Imzl-lligli.
therefore
o
Now we prove a simple criterion for the self-adjointness of a symmetric
operator.

Theorem 5.19. Let T be a symmetric operator on the Hilbert space H. If

H ... N(s - T) + R(s - T) for some s E R. then T is se/f-a4joint and H =
N(s - T) e R(s - T). Special case: If R(s - T) = H, then T is se/f-a4joint
and N(s - T) ... {O}.
PROOF. From T ... T" and T c T* it follows that N(s - T) C N(.s - T)
=R(s - T*).l C R(s - T).L, thus that N(s - T)J. R(s - T) and H= N(s-
T)E9 R(s - T). Therefore we have
N(s - T) = R(s - T).l
and
R(s - T) = N(s - T).l ~ N(s - T).l :J R(s - T*).
We show that D(T·) c OCT). This together with the inclusion T C 1'*
implies T - 1'*. Let fED( T*), g = (s - 1'*)f. Because of the inclusion
R(s- T·)cR(s- T) there exists an fnE D(T)cD(T*) such that (s-
T*)fa'" (s - T)fa = g = (s - T*)f. We therefore have f - /a E N(s - 1'*)
= R(s- T).l ... N(s- T)cO(T) and thusfE D(T). 0
108 5 Closed linear operators

A symmetric operator T on a Hilbert space is said to be essentially

self-adjoint provided that f is self-adjoint.

Theorem 5.20. A symmetric operator T on a Hilbert space is essentially

self-adjoint if and only if T* is symmetric. We then have T= P.

PROOF. If T is essentially self-adjoint, then T* =(f)* = f= T**, conse-

quently T* is self-adjoint (therefore symmetric) and we have f= T*. If T*
is symmetric, then since f is symmetric by Theorem 5.4(b) and (f)* = T*
holds, we have f c(f)* = T* c T** = f, hence f= (f)*. 0
Theorem 5.21. Let T be a symmetric operator on a complex Hilbert space H.
The operator T is self-adjoint (essentially self-adjoint) if and only if R(z ± -
T) = H (R(z ± - T)= H) for some z + with 1m z + > 0 and some z _ with
1m z _ < 0 (this then holds for all z + with 1m z + > 0 and z _ with 1m z _ <
0).
PROOF. Let R(z ± - T) = H for some z+ such that 1m z + > 0 and some z_
such that 1m z_ < O. As T is symmetric, the operators z ± -- T are injective,
therefore bijective, and by Theorem 5.18 we have lI(z+ - T)-III-<
I 1m z ± I-I. By Theorem 5.11 the operator z - T is also bijective for all z
such that Iz - z+1 < 11m z+1 or Iz - z_1 < 11m z_l· Since lI(z - T)-III-<
I 1m Zl-I for all z such that 1m z =1= 0, we can iterate this procedure and·
obtain that (z - T) is bijective for all z E C \~, in partic;ular for z = ± i
(cL also Exercise 5.33).
As T is symmetric, we have T C T*. To prove that T= T*, it is enough
to prove that D( T*) C D( T). To this end, let f E D( T*) and let fo =
(i- T)-I(i- T*)f. Then we havefoED(T)cD(T*), Tfo= T*fo, and (i-
T*)(f - fo) = (i - T*)f - (i - T*)fo = (i - T)fo - (i - T)fo = O. Therefore
f - fo E N(i - T*) = R( - i - T) 1- = H 1- = {O},
thus f = fo E D( T). _
If R(z ± - T)= H, then R(z ± - T) = H by the proposition following
Theorem 5.2. As f is also symmetric by Theorem 5.4(b), the self-adjoint-
ness of f follows from the part already proved.
Let T now be self-adjoint and let z E C\R. Then R(z, T) is closed and
bounded, therefore D(R(z, T» = R(z - T) is closed. Consequently, it is
enough to show that R(z - T)1- = {OJ. To this end, let h.L R(z - T). Then
by Theorem 4.13(b) we have hE N«z - T)*) = N(z· - T). Since T is
self-adjoint, T has no non-real eigenvalue, thus h = O.
If T is essentially self-adjoint, i.e., f is self-adjoint, then by the proposi-
tion following Theorem 5.2 we have

o
Theorem 5.22. If T is a symmetric operator on the complex Hilbert space H
and for some n E N, n ;;;. 2 we have R(i - Tn)= H or R( - i -- Tn)= H (respec-
5.3 Symmetric and self-adjoint operators 109

lively R(i - Tn) = H or R( - i-Tn) = H), then T is essentially self-a4/oint

(respectively self-a4/oint).
PROOF.
--,----,..,-
(a) Let R( i - m)= R( Tn - i)= H. There are numbers y ± E C such that
1m y + > 0, 1m y _ < 0 and y~ = i. We then have
(Tn-i) = (T-y±)(Tn- I +Tn- 2y±+'" +Ty~-2+y~-I),

hence R( T - y ±)~ R( Tn - i)= H. Therefore T is essentially self-

adjoint.
(b) If R(m + i) = H, then we choose y ± E C so that y~ = - i.
The proof of self-adjointness is similar. 0
Theorem 5.23.
(a) The symmetric operator T on the complex Hilbert space H is self-a4/oint
if and only if a( T) C Ill.
(b) If T is self-a4/oint on the (real or complex) Hilbert space H, then
s E a/T) if and only if R(s - T)* H. For z ff. ap(T) we have R(z, T)'" =
R(z"', T).
PROOF.
(a) By Theorems 5.18 and 5.21 the operator T is self-adjoint if and only if
z - T is surjective and continuously invertible for all z E C\ Ill, i.e., if
and only if C\lIlcp(T), or, equivalently, a(T) cIll.
(b) Assume that T is self-adjoint and z ff. ap(T). As a/T) c n, we also have
z'" ff. ap(T). Therefore R(z -:- T)J. = N(z'" - T"') = N(z'" - T) = {O}, i.e.,
R(z - T)= H. Now let R(z - T)= H. Since R(z - T) = H for all z E C\
Ill, we also have R(z'" - T)= H, and thus N(z - T) = N(z - T"') = R(z*
- T)J. = {O}, i.e., z ff. ap(T). If z ff. ap(T), then z - T* = z - T is
densely defined, injective and R(z - T)= H. Therefore R(z, T)* =
«z - T)-I)* = «z - T)",)-I = (z'" - T)-I = R(z"', T) by Theorem
4.17(b). 0
Now we obtain an especially simple characterization of the spectrum (or
the resolvent set) of a self-adjoint operator.

Theorem 5.24. If T is self-adjoint, then the following statements are equiv-

alent:
(i) z Ep(T),
(ii) there exists a c > 0 such that lI(z - T)fll ~ cllfll for all f E OCT) (i.e.,
(z - T) is injective and IIR(z, T)II <c- I ),
.(iii) R(z - T) = H.
(This theorem is in general false for symmetric operators.)

PROOF. If z E p(T), then (z - T) is injective and R(z, T) is continuous. If

(z - T) is injective and R(z, T) is continuous, then z ff. aiT) and thus by
Theorem 5.23(b) the set O(R(z, T» = R(z - T) is dense in H; as R(z, T) is
110 5 Closed linear operators

closed, we have R(z - T) = D(R(z, T» = H. If R(z, T) = Hand z E~, then

N(z - T) = N(z· - T*) = R(z - T)l. = {o}; therefore z - T is bijective, i.e.,
z Ep(T). If 1m z¥oO, then z Ep(T) by Theorem 5.23(a). 0
For the proof of any further criteria for self-adjointness we need some
auxiliary results. In complex Hilbert spaces the theorem of Rellich-Kato
(Theorem 5.28) can be proved directly somewhat more rapidly (cf. Exer-
cise 5.35). Here we obtain it (also for real Hilbert spaces) as a special case
of more general results. The auxiliary results gathered hem will be used at
other places, as well.

Theorem 5.25. Assume that HI and H2 are Hilbert spaces, A and Bare
operators from HI into H2 such that
D(A) C D(B) and IIBfll.;;; CIIAfll for f E D(A)
with some C;;;' 0. For every KE IK let P denote the orthogonal projection
K

(from H2 ) onto R(A + KB). Then IIP PolI~O as K~O.

K -

PROOF. For IKI < (ti2C) and for all f E D(A) we have

II Bfll .;;; CIIAfll .;;; C{II(A + KB)fll + IKIIIBfll} .;;; CII(A + KB)fll + tllBfll,
therefore
IIBfll .;;; 2CII(A + KB)fIl·
For hE R(Po)l. = R(A)l. = R(A)l. we thus have
IIP.hll = sup (I<h, g)l: gER(A + KB), II gil .;;; l}
= sup {I<h, (A + KB)f)1 : f E D(A), II(A + KB)fll .;;; I}
= sup (I<h, KBf) I : fED(A), II(A + KB)fll.;;; 1}
.;;; IKI IIhil sup (IIBfll : f E D(A), II(A + KB)fll .;;; l}
.;;; 2CjKlll h ll·
We can prove in a completely analogous way that for all hE R(P.)l. =
R(A + KB).1
IlPohll .;;; CjKlllhll·
Via Theorem 4.33 we obtain that liP. - Poll .;;; 2CjKI. This proves the
theorem. 0

Theorem 5.26. Assume that HI and H2 are Hilbert spaces, T and S are
operators from HI into H2 ; T is closed, and S is T-bounded. Furthermore,
denote by a the set
a= {z ElK: T+ zS is closed}
and for every z E a let Qz denote the orthogonal projection from HI X H2 onto
G(T + zS). Then a is open and the function z~ Qz is continuous on a (with
respect to the norm topology of B(HI X H2»'
5.3 Symmetric and self-adjoint operators 111

PROOF. If zoE 0, then T+zoS is closed and D(T+zoS)=D(T). The

operator S is therefore (T+ zoS)-bounded by Theorem 5.9. Hence T+ zS
is closed for z sufficiently near zoo Thus 0 is open. Let us define the
operators A and B from HI into HI X H2 by
D(A) = D(T), Af= (j, (T+ zoS)f),
D(B) = D(T), Bf=(O, Sf).
Then the assumptions of Theorem 5.25 are obviously satisfied and the
equalities R(A) = G(T+ zoS) and R(A + ICB)= G(T+(zo+ IC)S) imply the
continuity of z~ Qz at the point zoo 0
With this we can now obtain a general theorem on the perturbation of a
closed operator and its adjoint.

Theorem 5.27. Assume that HI and H2 are Hilbert spaces, T and S are
operators from HI into H2. Let T be densely defined, let S be T-bounded. let
S* be T*-bounded, and let
0= {zEII<: T+zS and T*+z*S* arec/osed},
0 0 = the connected component of 0 that contains zero.
Then (T+zS)*= T*+z*S* for all zEOo'
PROOF. Let Q. be the orthogonal projection (in HI x H2 ) onto G(T + zS)
and let Q; be the orthogonal projection onto U-IG(T*+z*S*). where U
is defined as in Section 4.4. By Theorem 5.26 the operators Qz and Q;
depend continuously on z for z E 0 and Qo+ Qo= I H,xH2 = I (as we have
G(T)teU- I G(T*)=H I XH2). For every zEII< we have T*+z*S*C
(T+ zS)*, i.e.,
U-IG(T*+z*S*) c U-IG(T+zS)* =G(T+zS).l.
By Theorem 4.30(a) we therefore have Q. Q; = Q; Qz = 0, i.e., 1- Q. - Q;
is an orthogonal projection for any z E O. Consequently, III - Qz - Q;II
assumes only the values 0 and 1. Since, on the other hand, 111- Qz - Q;II
depends continuously on z E 0, it follows that III - Qz - Q;II = III - Qo-
QolI=O for all zEOo• therefore G(T*+z*S*)= UG(T+zS).!.=G«T+
zS)*). and thus T* + z*S* = (T+ zS)* for all z E 0 0 , 0

If the relative bounds of S with respect to T and of S* with respect to

T* are less than J, then it follows that (T+ S)* = T* + S* (Hess-
Kato [42]), since in this case {z Ell<: Izl ~ I} c 0 0 by Theorem 5.5. If we
specialize this result for self-adjoint operators T and symmetric operators
S, then we obtain the following important result.

Theorem 5.28 (Rellich-Kato). If Tis self-at/ioint (essentially self-at/ioint) on

the Hilbert space 4, the operator S is symmetric and T-bounded with
112 5 Closed linear operators

T-bound less than 1, then T + S is self-adjoint (essentially self-a4joint with

T + S= T + 8 and DC T + S) = O( T».
PROOF.
(a) If T is self-adjoint (T= T*), then because of the inclusion S C S* the
operator S* is T*-bounded with T*-bound less than 1. In Theorem
5.27 we therefore have {z ElK: Izl" I} cOo' and thus (T + S)* =
T*+ S*= T+S.
(b) Let T now be essentially self-adjoint, i.e., let T be self-adjoint. First we
show that 8 is T-bounded with T-bound less than 1 (more precisely,
equal to the T-bound of S). To this end, letf E O(T). Then there exists
a sequence (f,,) from OCT) such that fn---+f, Tf,,---+ 1:,(. From the T-
boundedness of S it follows that (S/,,) is a Cauchy sequence, therefore
/E 0(8), Sfn---+8/ and
II 8fll = lim II S/"II " lim (a II fn II + bll Tfnl!) = allfll + bll Tfll·

By part (a) the operator T+ S is therefore self-adjoint. From the

inclusion T + SeT + 8 and from the closedness of + 8 it follows 'r
that T+ ScT+8. By Theorem 5.5 and by the equality 0(T+8)
= D(T) it follows that O(T + S) = O(T + S), and therefore
T+s=f+~ 0

Theorem 5.29 (Wtist [58]). Let T be a self-a4joint operator on the Hi/bert

space H, let S be symmetric and T-bounded, and let
n= {zEIK: T+zS and T+z*S are closed},
no = the connected component of n containing zero.
For every z E IR n no the operator T+ zS is self-adjoint. Special case: If
T + t S is closed for all t E [0, 1], then T + S is self-adjoint.
PROOF. We have (T+ zS)* = T+ zS* = T+ zS for every z ElRn no by
Theorem 5.27, 0
Theorem 5.30 (Wtist [57]). Let T be essentially self-adjoint on the complex
Hilbert space H, let S be symmetric with D(T) C O(S) and let IISfll "allfll
+ II Tfll for all f E O( T) with some a ~ O. Then T + S is essentially self-
aqioint.
PROOF. Let A = T + S. We show that R( JL - A).1 = {O} for JL E {i, - i}. For
this, let (In) be a sequence from (0, 1) such that In---+1. By Theorem 5.28 the
operator An = T + In S is essentially self-adjoint for any n E N, and we have
II(A - An)fll = (1- In) II Sfll " allfll + II Tfll - 1.11 Sfll
" allfll + II(T+ tnS)fll = allfll + IIA,JII. (5.16)
5.3 Symmetric and self-adjoint operators 113

Now let hE R( P. - A).L. As An is essentially self-adjoint, there exists an

fn E D(An} = D(A) such that
I
1I(p.-An)Jn - hll < -,
n
n E N.
Therefore
h = lim (p.-An)Jn' (5.17)
Because of the inequality 1I(p.-An)-'II" I we have IIfnll < 1I(p.-An)fnll.
Consequently,
lim sup IIfnll < IIhl!.
By (5.17) it follows from this that
lim sup /lA,Jnll < 2/1hll
and thus because of (5.16)
lim sup II(An - A)J"II < lim sup [all!nll + II Anin II ] <:; ellhll,
with c = a + 2. As D( T) is dense, for every t: > 0 there exists an h. E D( T)
such that IIh - h./I < t:. Because of the relation hE R( P. - A).L, it follows
that

/lh/l2 = lim (h, (p. - An)fn) = lim (h, (A - An)Jn)

= lim [ (h - h., (A - An)J,,)+ (h., (A - An)!n) ]
< Ilh - h./llim sup II(A - An)J,,1I + lim supll(A - An)h.1I IIfnll
+ Jim sup (1- In)I/Sh.IIIIJ,,1/ =
<:; cllhllt: cllhl/t:.
Since this holds true for all t: > 0, we have h = O. o
EXERCISES

5.31. Let T be a symmetric operator on the Hilbert space H.

(a) If His complex, then T is essentially self-adjoint if and only if aiT*) c rio
(b) If H=N(s-T)+R(s-T) for some sER, then T is essentially self-
adjoint and we have N(s - T) = N(s - T) and R(s - T) = R(s - T)
(c) If s - T is continuously invertible and R(s - T)= H for some s E fR, then
T is essentially self-adjoint.
5.32. Let H be a real Hilbert space, and let T be an operator on H.
(a) The space He = H X H, with the addition (f" gIl + (f2g2) =(f, + f2' g, +
g2), multiplication by a scalar (a + ib)(f, g) = (af - bg, ag + bi) and scalar
product «(f" g,), (f2' g2» = <f"h) + i<f" g2) - i(g.,f2) + <g" g2)' is a
complex Hilbert space, the complexijicalion of H.
(b) By setting D(Td={(f,g)EHe :f,gED(T)} and Te(f,,J~=(Tfl' Tf2) a
linear operator Te is defined on He. We have:
(i) Te is bounded (respectively belongs to B(Hd) if and only if T is
bounded (respectively belongs to B(H». We have II Tell-II Til.
114 5 Clo!led linear operators

(ii) O(TC> is dense if and only if OCT) is dense.

(iii) Tc is symmetric if and only if T is symmetric.
(iv) Tc is (essentially) self-adjoint if and only if T is (essentially) self-
adjoint.
Tc is called the complexification of T.
5.33. Let T be a closed symmetric operator on the complex Hilbert space H, and
let p. denote the orthogonal projection onto R(z - T) for 2 E C\R.
(a) The mapping z~p. is continuous on C\R with respect to the norm
convergence in B(H).
(b) If R(zo - T) H for some Zo E C such that 1m Zo > 0 (respectively 1m Zo
=0

< 0), then R(z - T) =- H for all z E C such that 1m z > 0 (respectively
1m z <0).
Hint: Use Theorem 5.25.
5.34. Let T be (essentially) self-adjoint on the Hilbert space H, and let S be
symmetric such that D(T) c D(S) and Re <TI, SI> ;> - (a1l/1l 2 +
bllTIlIlIS/l1) for all/ED(T) with some b<l. Then T+IS is (essentially)
self-adjoint for all t ;;. O.
Hint: For t;;. 0 we have II 1'111 + 11111 "C.(II(T+ IS)f1l + 11/11)" C2<117111 +
11/11>·
5.35. (a) Let T be a self-adjoint operator on a complex Hilbert space, and let S be
T-bounded with T-bound < l. If c > 0 is large enough, then II Sill "
bll( ± ic - T)/II for all IE D(T) with some b < l. Using this, prove the
theorem of RelIich-Kato (Theorem 5.28) for complex Hilbert spaces.
(b) Using part (a) and Exercise 5.32, prove the theorem of RelIich-Kato for
real Hilbert spaces.

5.4 Self-adjoint extensions of symmetric operators

If S is a symmetric operator, then S c S*. For every symmetric extension
T of S we have (cf. the proposition preceding Theorem 4.19) SeT c T* c
S*. It seems plausible that there exists a sufficiently "large" extension T of
S that is self-adjoint. For this we would have
SeT = T* c S*. (5.18)
In this section we begin with some simple investigations concerning the
problem of the existence of a self-adjoint extension of a symmetric opera-
tor (cf. also Sections 5.5 and 8.1 to 8.3).

Theorem 5.31.
(a) If T. C T2 are self-adjoint operators, then T. = T 2.
(b) If S is a symmetric operator and T. and T2 are self-adjoint extensions of
S such that D(T.) C D(T2 ), then T. = T 2 •
(c) If S is essentially self-adjoint, then S is the only self-aa,'ioint extension of
S.
5.4 Self-adjoint extensions of symmetric operators 115

PROOF.
(a) It follows that TI C T2 = T1 c Tr = T I• thus T I '" T 2•
(b) For allf E D(TI) C D(T2) and for all g E D(S) C D(T I) C D(T2) we have
(Td. g) - (f. T 2g) = <1. Sg) == (f. Tlg) == (Td, g).
Since D(S) is dense, it follows that Td= Td for all j E D(T1), i.e.,
TI C T 2• From part (a) it follows that TI == T 2.
(c) If T is a self-adjoint extension of S, then we have SeT, since T is
closed. The equality S == T now follows from part (a). D

The following theorem ensures the existence of self-adjoint extensions for

two large classes of symmetric operators. A symmetric operator S on the
Hilbert space H is said to be bounded from below if there exists a "( E R
such that (J. Sf) ;;. "(IIfll 2 for all f E D(S). Every"( of this kind is called a
lower bound of S. The least upper bound of all lower bounds is also a lower
bound. If 0 is a lower bound of S. then S is said to be non-negative. The
concepts bounded from above, upper bound. and non-positive are defined
similarly. If an operator is bounded from either below or above, then it is
said to be semi-bounded. Besides the semi-bounded operators. the following
simple theorem also treats the continuously invertible symmetric operators.
i.e .• those symmetric operators S. for which II Sf II ;;. yllfll with some y > 0
(cf. Exercise 5.14).

Theorem 5.32. Let S be a symmetric operator on the (real or complex)

Hilbert space H. and assume that (f. Sf) ;;. yllfW with some y E III (respec-
tively II Sfll ;;. y II fll with some "( > 0) for all f E D(S). Then for each "E
( - 00. y) (respectively" E (- y. y» there exists a self-adjoint extension T« of
S such that (f. TJ) ;;. "IIfll 2 (respectively liTJII ;;. 1"llIfll) for all f E
D(TK }. We have N(" - TK ) = N(K - S"') = R(K - T.).l..
PROOF. The operator S is closable and S obviously satisfies the same
assumptions. also. Therefore we can assume without loss of generality that
S is closed. In the first case we have for all "E (- 00. y) and f E D(S),
f"f: 0 that
II(S-,,)fll;;' I«S-K)f.f)llIfll- 1
;0. (Sf. f) Ilfll- 1 - ""fll ;;. (y - ")lIfll·

In the second case we have for all "E (- y. y) and f E D(S) that
II(S - ")fll > IISfll-I"111f1l > (y -1"Dllfll·
Consequently. in both cases S -" is continuously invertible. The range
R(S - ,,) = D«S - K) -1) is therefore closed. From this it follows that

R(S - ,,) + N(S* - ,,) = R(S - ,,) + R(S - ,,).L = H. (5.19)

116 5 Closed linear operators

Because of the equality N(S* - K)n O(S)= N(S- K)= {O}, the sum O(S)
+ N(S* - K) is a direct sum. Hence we can define

O(TK) =O(S) + N(S* - K),

TK(JI + 12) = S11 + K12 for 11 E 0(S),j2 E N(S· - K).

We obviously have N(K - T K ) = N(K - S*). The operator is symmetric, 1:

because D(TK) is dense (since O(TK):J O(S» and for all 11' gl E D(S),
12' g2 E N(S* - K) = R(S - K).l we have (observe that (TK - 1()12 = (TK - K)g2
=0)
<11 + 12' (TK - K)(gl + g2» = <11 + 12' (S - K)gl)
= <11' (S - K)gl) = «S·- K)II' gl)
= «S - K)I gl + g2)
I,

= « TK - K)(JI +12)' gl + g2)'

By Theorem 5.19 the operator T" is self-adjoint, since because of (5.19) we
have H= R(S- K)+ N(S* - K)= R(T,,- K)+ N(T,,- K). Besides, for all 11
E D(S), 12 E N(S* - K) we have

<11 + 12' T,,(JI + 12» = <11> SII) + <S*12,11) + K[ <11,12) + 1112112]

;;;. yllldl 2 + K[ <12,11) + <11,12) + 1112112] ;;;. Kllli + 12112
in the first case, and

II T,,(JI + 12)11 2 = <S11 + 42' S11 + 42)

= IISldl 2 + 1«11' S*12) + K<S*f2 ,j1) + K211f2112
;;;. y2111dl 2 + 1(2[ <11,12) + <12,11) + 1112112] ;;;. K21111 + 12112
in the second case. D

In order to sharpen the results of Theorem 5.32, we: first prove an

extension theorem for bounded Hermitian operators.

Theorem 5.33. If A is a bounded Hermitian operator on the Hilbert space H,

then there exists a self-acfjoint extension B E B( H) 01 A !.uch that II B II =
IIA II. II R(A) is dense, then every self-adjoint extension 01 A is injective.
PROOF. If IIA II = 0, then B = 0 is the required extension. Therefore let
*'
IIA II O. Without loss of generality we may assume that IIA II = I. Since
along with A its closure A is also Hermitian and IIA II = IIXII, we may also
assume that A is closed, i.e., D(A) is closed. Let P be the orthogonal
projection onto O(A). Then we have

A = Al +A2 with Al = PA, A2 = (I-l')A.

5.4 Self-adjoint extensions of symmetric operators 117

We consider AI as an operator from H into the Hilbert space D(A) with

D(AI) = D(A) and A2 as an operator from H into the HilbeEt space_ D(A)i
with D(A 2) = D(A). First we show: There exist extensions A I and A2 of A I
and A 2 , respectively, such that D(AI) = D(A2 ) = H, R(AI) C D(A), R(A0
c D(A)i and
IIA.1112 + 1I12fll2 "lIfll 2 for all f E H.
We define the operator AI by
AI = (AP)".
Then we have IIAIII = IIAPII" IIAII and R(AI)cN(AP)i cD(A). More-
over, for all f E Hand g E D(A) we have

<f,Alg) = <APf, g) = <Pf,Ag) = <1,A l g).

Therefore AI g = A I g, and thus A I c AI' Because of the relation IIA.1I1 "

III III II fll " IIA II Ilfll = II fll the equality
[f, g] = <f, g) - <Ad, Alg)
defines a semiscalar product on H. The set
N= {fEH: [f.J] =o}
is a closed subspace of H. (If f, g E N, a E K, then af E Nand [j + g,
f+g]=2Re[j,g]';;;2{[j.J][g,g]}l j 2=O, thereforef+gEN. If (i,,) is a
sequence from N such that f,,~f E H, then [j.J] = <f,f) - <Ad, 1.1) =
limn-oo{ <1",j,,) - <Ad", Ad,,) =0, therefore fEN). Let Ho= N i and let
Po be the orthogonal projection onto Ho. By construction, we have [j,fJ*
o for all non-zero f E Ho, i.e., [. , .J is a scalar product on Ho. For every
fE H we have

[Pof, Pof] = [f - (/ - Po)!,j - (/ - Po)!]

= [f, f] - 2 Re [f, (I - Po) f] + [ (I - Po) f, (I - Po) f] = [ f, f].

For fE D(A)n N we have

IIA2fll2 = 11(/ - P)AfIl 2 = IIAfll2 - IIPAfll 2

= IIAfll2 -IiA l f1l 2 .;;; IIfll2 -Ii A l f1l 2 = [f,f] = o.

Therefore Ad= A2 g for f - g E D(A) n N. Consequently, the equalities
D(A2 ) = PoD(A),
A2 g = Ad for g = Pof E D(A2 )
118 5 Closed linear operators

define a linear operator from Ho into O(A).L, and for all g = Pof E 0(.42)
we have
11.42 gll2 = II Adll 2= <(I - P)Af, Af) = <Af, Af) - (A d, A I f)
"<f.J)-<A.1,A.1) = [f,f] = [Pof, Poi] =[g,g].
From this it follows that A2 can be extended to an operator C from Ho into
D(A).L such that D(C) = Ho and

II Cfl1 2 " [f, f] for all f E Ho

(cf. also Exercise 4.16), Let us now define A2 by
.42 = CPo'

Then for all f E H

IIAdll 2 = IICPofll 2 "[ Pof, Pof] = [f,f],
consequently
11.4.1112 + 1I.42f1l2 ..;; IIf1l2.
Since for allfE D(A) we have
Ad = CPof = A 2Pof = Ad,
the operator A2 is an extension of A 2. Therefore .4 = Al + A2 is an exten-
sion of A = A 1+ A2 and IIAII ..;; 1. Since for all f E D(A) and all g E H we
have
<A*f, g) = <f, (AI + A2)g) = <f, A,g) + <f, ·,42 g)
.. <1, Alg) = <f, (AP)*g) = <Af, g),
the operator A* is also an extension of A. Hence
B =4(A+A*)
is a self-adjoint extension of A such that liB II = 1.
Now let R(A) be dense and let B be a self-adjoint extension of A. Then
R(B) is also dense, consequently N(B) = N(B*) = R(B).L = {O}. 0

Theorem 5.34. Let S be a symmetric operator on the (real or complex)

Hilbert space H.
(a) If II Sfll ;:. 'YII fll for all f E D(S) with some 'Y > 0, then there exists a
self-a4joint extension T of S such that II Tfll ;:. 'Y II fll for all fED( T).
(b) If S is bounded from below, then there exists a self-adjoint extension T of
S with the same lower bound (cf. also Theorem 5.38, friedrichs' exten-
sion).
PROOF.
(a) The operator A = S - I is Hermitian «ASf, Sg) =<f, Sg) = <Sf, g) =
<Sf, A Sg) for all Sf, Sg E D(A) = R( S»and bounded, IIA II " 'Y - '; A is
S.4 Self-adjoint extensions of symmetric operators 119

injective and R(A) = D(S) is dense. Therefore by Theorem 5.33 there

exists an injective self-adjoint extension B of A such that liB II = IIA II "
1'-1. Then T= B -I is a self-adjoint extension of S and II Tfll ;;. 1'llfll
for all f E D( T).
(b) Without loss of generality we may assume that l' = O. As in the proof of
Theorem 5.32, we can show that 1+ S is continuously invertible. Let us
define A by
A = (/ - S)(/ + S)-I.
Then A is Hermitian, because for aIlf= (I + S)fo' g = (I + S)go E D(A)
= R(I + S) we have
(Af, g> = «(/ - S)fo, (/ + S)go>
= <10' go> - (Sfo, go> + <10' Sgo> - (Sfo, Sgo>
= (fo' go> - <10' Sgo> + (Sfo' g) - (Sfo' Sgo>
= «(/ + S)fo, (/ - S)go> = (f, Ag>.
It follows from the definition of A that

I - A = (/ + S)(/ + S)-I - (I - S)(/ + S)-I = 2S(/+ S)-I,

I +A = 2(/ + S) - I.
Consequently, I + A is injective and
S = (/ - A)(1 + A) - I.

A is bounded with norm IIA II < I, since for all f E D(A) we have
IIfll2 - IIAfll2 = «(I - A)f, (/ + A)f)
= «(I-A)(I+A)-I(I+A)f,(/+A)f)
= (S(I + A)f, (/ + A)f) ;;. O.
Therefore by Theorem 5.33 there exists a self-adjoint extension B of A
such that II B II = IIA II· By the same theorem 1+ B is injective, since
R(I + A) = D(S) is dense. The operator

T = (/-B)(I+B)-I
is therefore an extension of S. For all f=(I+B)fo' g=(I+B)goE
D(T) = R(I + B) we have
(Tf, g) = <(/ - B)fo, (I + B)go> = «(I + B)fo, (I - B)go) = (J, Tg),
<1, Tf> = «(I + B)fo, (/ - B)fo> = Ilfoi/2 - I/ Bf oll 2 > 0,
i.e., T is symmetric and bounded from below with lower bound O. We
have 1+ T=2(J + B)-I and /- T= 2B(I+ B)-I, hence B=(I- T)(1
+T)-I. From this it follows that R(I+T)=D(B)=H, thus Tis
self-adjoint (cf. Theorem 5.19). 0
120 5 Closed linear operators

EXERCISE

5.36. Let A be symmetric and semi-bounded.

(a) If A has only one semi-bounded self-adjoint extension, then A is essen-
tially self-adjoint.
(b) If the lower bound of A is positive and A has only one positive extension,
then A is essentially self-adjoint.

5.5 Operators defined by sesquilinear forms

(Fiedrichs' extension)
In what follows H will always be a Hilbert space. A sesquilinear form s on
H is said to be bounded if there exists a C;;> 0 such that Is(f, g)1 ~
C IIfIl II gil for all f, g E H. The smallest such C is called the norm of s. It
will be denoted by IIsli. If T E 8(H), then the equality t(f, g) = <Tf, g)
defines a bounded sesquilinear form on H. We obviously have II til = II Til·
Conversely, every bounded sesquilinear form induces an operator on 8(H).

Theorem 5.35. If t is a bounded sesquilinear form on H, then there exists

<
exactly one T E 8(H) such that t(f, g) = Tf, g) for all f, g E H. We then
have liT II = II til·
PROOF. For every fE H the function g~t(f, g) is a continuous linear
functional on H, since we havelt(f, g)1 ~ 1I11I11f111i gil. Therefore for each
f E H there exists exactly one j E H such that l(f, g) = <j, g:>. The mapping
f~ j is obviously linear. Let us define T by the equality Tf= j for allf E H.
By Theorem 4.3(b) the operator T is bounded with norm

IITII = sup {1<Tf,g)1 :f,gEH, IIfll=lIglI=l}

= sup {1/(j,g}1 :f,gEH, IIfll=lIgll=l} =: II til·

If TI and T2 are from 8(H) and (Td,g)=t(f,g)=(Td,g) for all

f, g E H, then it follows that TI = T 2, i.e., T is uniquely determined. 0
For unbounded sesquilinear forms the situation is much more com-
plicated. We consider only a special case.

Theorem 5.36. Let (H, (. , -» be a Hi/bert space and let HI be a dense

subspace of H. Assume that a scalar producl <. ,
')1 is defined on HI in such a
way that (HI' <. , ')1) is a Hi/bert space and with some" >0 we have
"lIfI1 2 " IIfll~ for all f E HI' Then there exists exactly one self-adjoint opera-
tor T on H such that

D(T} CHI and (Tf, g) = <1, g)1 for f E D(T},g E HI' (5.20)
5.5 Operators defined by sesquilinear forms (Fiedrichs' extension) 121

T is bounded from below with lower bound K. The operator T can be defined
by the equalities

D( T) = {f E HI : there exists an j E H such that <f, g)1

= <j, g) for all g E H.}, (5.21)

Tf=j
D(T) is dense in HI with respect to the norm 11.111'
PROOF. Existence: The element j in (5.21) is uniquely determined, since HI
is dense. Since the mapping f~ j is also linear, (5.21) defines a linear
operator. We can also consider T as an operator from HI =(H I, (. , ')1)
into H. This operator will be denoted by To. If J denotes the operator from
H into HI defined by
D(J) =HI cH, Jf = f for f E O(J),
then by (5.21) we obviously have
To = J*.

J is closed, since for any sequence (f,,) from D(J) = HI such that./" ~f [in
H] and fn~h [in H.J we have fn~h [in H] because of the inequality
11.11 2 (;,,-III.llf, i.e.,j= h E D(J). Therefore To=J* is densely defined. Thus
D(T) = D(To) is also dense in HI (with respect to 11.111) and, consequently,
in H (with respect to 11.11), as well. By (5.21) we have for al1f, g E D(T) that
<Tf, g) = <f, g)1 = [<g,f)d* = [<Tg,f) J* = <f, Tg),
i.e., T is symmetric. The seif-adjointness of T will fol1ow from Theorem
5.19 if we prove that R(T) = H. For this let f E H be arbitrary. Then
g~<f, g) is a continuous linear functional on HI> since we have

l<f,g)l, Ilfllllgil 'K- I / 2 1Ifllllglll'

Consequently, there exists an J E HI such that
<f,g) = <j,g)1 foraH gEHI·
By (5.21) this means, however, that J E D(T) and f = Tj The semi-
boundedness fol1ows from the inequality

T obviously satisfies (5.20), as well.

Uniqueness: Every operator S that satisfies (5.20) is obviously a restriction
of the operator T defined by (5.21). Since T is self-adjoint, it follows that
SeT c S*. If S is self-adjoint, then we necessarily have S = T. 0
In what follows let D be a dense subspace of H and let s be a
semi-bounded sesquilinear form on D, more precisely, let the inequality
122 5 Closed linear operators

sU, f);;' yllfl1 2 be satisfied by some y E IR for all fED. Then the equality
<f. g). = (1 - y)<f, g) + s(j, g) defines a scalar product on D such that
IIfll.;;' IIIII for all fE D. Moreover, we assume that II. II. is compatible
with II . II in the following sense: If (jll) is a II . II.-Cauchy sequence from D
and IIfllll~O. then we also have IIfnlls~O (cf. Exercise 5.37; in the theory
of sesquilinear forms such a sesquilinear form s is said to be closable). Let
Hs now be a II. II.-completion of D (for example the one that was
constructed in Section 4.3). It follows from the compatibility assumption
that H. may be considered as a subspace of H if the embedding of H. into
H is defined as follows: Let (jll) be a II . II.-Cauchy sequence in D. Then
(In) is a Cauchy sequence in H. Let the element limfll from H correspond
to the element [(In)] of H•. On the basis of the compatibility assumption,
this correspondence is injective and the embedding is continuous with
norm ( 1. The spaces Hand H. are related the same way as H and HI were
in Theorem 5.36 (with K = 1). Let
sU. g) =0 <f, g). - (1- y)<f, g) for f, g E H•.
Therefore s(j. g) = s(j. g) for f, g ED. The sesquilinear form s is called the
closure of s.

Theorem 5.37. Assume that H is a Hilbert space. D is a dense subspace of H

and s ;s a semi-bounded symmetric sesquilinear form on D with lower bound
y. Let II . II. be compatible with II . II. There exists exactly one semi-bounded
self-a4joint operator T with lower bound y such that
D(T) cHs and <Tf, g) = sU, g) for a/l f E Dn D(T), g E D.(5.22)
We have
D( T) = {f E Hs : there exists an i E H such that
sU, g) = <i, g) for all g ED}, (5.23)
Tf = j for f E D( T).

PROOF. If we replace (Hit <. , ')1) by (H., <... ).) in Theorem 5.36, then we
obtain exactly one self-adjoint operator To such that D(To) C H. and
<Tof, g) = <f, g). for all f E D(To), g E H•.
To is semi-bounded with lower bound 1. The operator T= To-(l- y)
obviously possesses the required properties. The uniqueness follows from
the uniqueness of To. Formula (5.22) implies (5.23), since D is dense (in Hs
and in H). []
If S is a semi-bounded symmetric operator with lower bound y, then the
equality
sU, g) = <Sf, g), f, g E D(S)
5.5 Operators defined by sesquilinear forms (Fiedrichs' extension) 123

defines a semi-bounded sesquilinear form s on D(S) with lower bound y.

In this case
(f, g). - (Sf, g) + (1- Y)(f, g) and Ilfll~ - (Sf,!) + (1- y)lIfl1 2
for f, g E D(S). The norm II . II. is
compatible with a II . II: Let (fn) be
II . II.-Cauchy sequence from D(S) such that fn-+-O. Then for all n, mEN
we have
IIfnll; = (fn,fn). a I(fn,fn - f m ). + (f",fm>'1
" IIfnll.llfn - fmll. + II(S+ 1- y)fnllllfmll.
The sequence Olfnll.) is bounded, IIfll -fmll. is small for large nand m and
for any fixed n we have 11(8 + 1- y)fnll'lIfmll-+-O as m-+-oo. Consequently it
follows that IIfllll.-+-O as n-+- 00. This fact makes the following construction
of a self-adjoint extension (Friedrichs' extension) of a semi-bounded sym-
metric operator possible, where the lower (upper) bound remains un-
changed.

Theorem 5.38. Let S be a semi-bounded symmetric operator with lower bound

y. Then there exists a semi-bounded selj-acfjoint extension of S with lower
<
bound y. If we define s(f, g) = Sf. g) for f, g E D( S), and H. as above, then
we have: The operator T defined by
OCT) =D{S·) n H. and Tf = S·f for f E O(T)
is a selj-acfjoint extension of S with lower bound y. The operator T is the only
selj-acfjoint extension of S having the property O(T) C Hs.
PROOF. By Theorem 5.37 there exists exactly one self-adjoint operator T
with O( T) c H. and
(Tf, g) =sU, g) = (Sf, g) for f E D(S) n D(T), g E O(S).
y is a lower bound for T. By (5.23) we have
O( T) = {f E H. : There exists an j E H with sU, g)
=(j,g) forall gED{S)} (5.24)
Tf = j for fED{ T)
We can replace s(f, g) by (f, Sg) in (5.24): If we choose a sequence (f,,)
from O(S) such that 11111 - fll.-+-O, then we obtain
sU, g) = lim sUn' g) = lim (fn. Sg) = (f. Sg).
Consequently, it follows that D(T) = O(S·)n H. and T= S·lo(T)' Because
of the inclusions S c S· and O(S) c H. it follows from this that T is an
extension of S. Let A be an arbitrary self-adjoint extension of S such that
O(A) c H". Then A C S· and O(T) = O(S·) n H. imply that AcT, conse-
quently A = T. 0
124 5 Closed linear operators

Our arguments so far enable us to study the operator product A *A, as

well. If A E 8( H), H2), then we already know from Exercist: 4. I 9 that A *A
is self-adjoint.

Theorem 5.39. Let (HI> <. ,

.»)) and (Hz, <. ,
')2) be Hilbert spaces and let A
be a densely defined closed operator from H) into H2• Then A * A is a
self-adjoint operator on HI with lower bound 0 (A *A is non-negative).
D(A *A) is a core of A. We have N(A *A) = N(A).
PROOF. As A is closed, D(A) is a Hilbert space with the scalar product
<f, g)A = <Af, Ag)z + <f, g)I' and
IlfilA ~ IIflll for allf E D(A). Therefore
by Theorem 5.36 there exists a self-adjoint operator T with lower bound 1
for which
D( T) = {J E D(A) : there exists an j E HI such that
<f,g)A=<j,g)1 forall gED(A)},
Tf=j for fED(T).
On account of the equality <f, g) A = <Af, Ag)2 + <f, g»), this definition
says that f E D(T) if and only if Af E D(A *) (i.e., f E D(A *A» and Tf= j
=A*Af+ f. Hence it follows that T=A*A + 1, A*A = T-l, i.e., A*A is
self-adjoint and non-negative. From Theorem 5.36 it follows tl'\at D(A *A)
is dense in D(A) with respect to II . IIA' i.e., D(A *A) is a core of A. If
f E N(A), then Af= 0 E D(A *) and A * Af= O. Therefore N(A) C N(A * A). If
f E N(A *A), then II Afll2 = <A *Af, f) = O. Therefore N(A *A) c N(A), and
thus N(A*A)= N(A). 0
Theorem 5.40. Let A I and A2 be densely defined closed operators from H into
HI and from H into H2, respectively. Then A r A I = A t A 2 if and only if
D(A 1)= D(A 2} and IIA)fll = IIA2fll for all fE D(A I )= D(A 2}·
PROOF. Assume that D(A)=D(A 2} and IIAtIli = IIAzfl1 for allfED(AI)'
It follows from (1.4) in the complex case and from (1.8) in the real case
that
<A1f,AIg) = <A 2f,A 2 g) forall f,gED(A 1) =D(A 2 ).
Then the construction of Theorem 5.39 provides the same operator for
A =A 1 and A = A 2, consequently ArA 1 = AtA2' If this equality holds, then
for aIlfE D(ArAI)= D(AtA2) we have
IIAlfl12 = <ArAlf,f) = <AtA2f,f) = II A,JII 2
(here we have used the inclusions D(Ar AI)c D(AI) and D(AtA 2)c D(A 2»·
By Theorem 5.39 the subspace D(A r A I) = D(At A 2) is a core of A I and A 2·
As the A I-norm and the A 2-norm coincide on D(A r A I) = D(At A 2 ), it
follows finally that D(A I) = D(Al) and IIA till = IIAzfll for all f E D(A I)
= D(A 2). 0
5.6 Normal operators 125

EXERCISE
5.37. Let H= L2(O, I). Then D= CI[O, I] (the space of continuously differentiable
functions on [0, I]) is a dense subspace of H. The equality sU, g) = <I, g> +
f'G)*g'G) defines a sesquilinear form on D such that 11111. ;;. II/II. The norm
II . II. is not compatible with II . II·

5.6 Normal operators

A densely defined operator T on a Hilbert space H is said to be normal if
OCT) =O(P) and IITfll = IIT*fll forall fED{T).
Every self-adjoint operator is obviously normal.

EXAMPLE I. Let Me Rm be measurable and let t : M~C be a measurable

function. The maximal multiplication operator T induced by t (cf. Section
4.1, Example 1 and Section 5.1, Example 2) is normal. By (5.1) we have
OCT) = O(T*) and for f E OCT) we obviously have

II Tfll2 = jlt(x)f(xW dx = f.lt(x)*f(x)1 2 dx = II T*fIl2.

M M

EXAMPLE 2. According to Section 4.6 an isomorphism U of a Hilbert space

H onto itself is called a unitary operator on H. We have O( U) = O( U*) = H
and II Ufll = II U*fll = IIfll for all f E H, i.e., every unitary operator on a
Hilbert space H is normal.

Proposition.
(I) Every normal operator T is closed and maximal normal (i.e., for every
normal operator N the inclusion TeN implies T';' N).
(2) LeI T be densely defined and closed. Then the following assertions are
equivalent:
(i) T is normal,
(ii) T* is normal,
(iii) T* T = TT*.
(3) If T is normal, then z + T is also normal for every z E K

PROOF.
(I) The T-norm and the T*-norm coincide on D(T). Since T* is closed,
that T is closed follows from Theorem 5.1. If N is normal and TeN,
then OCT) C DeN) = O(N*) c O(T*) = OCT). Therefore D(T) = O(N),
and thus T= N.
(2) The equivalences (i)~(iii) and (ii)~(iii) immediately follow from The-
orem 5.40 with AI = T, A2 = T* and from Proposition (1).
 126 5 Clo!.ed linear operators

(3) We have O«z + T)*) = O(z* + T*) = O(T*) = OCT) = O(z + T) and for
all f E OCT)
lI(z + T)f1l 2 = Izl 211fll 2 + 2 Re <zf, Tf) + II Tflf
= Iz*1211f1l2 + 2 Re <z*f, T*f) + II T*f1l2
= lI(z* + T*)fIl2. 0
Theorem 5.41. Let T be a normal operator.
'(a) For every z E II< we have N(z - T) = N(z* - T*).
(b) If Zl' Z2 are distinct eigenvalues of T and fl,f2 are corresponding eigen-
vectors, then f l 1..f2'
PROOF.
(a) The statement is evident, as z - T is normal.
(b) By part (a) we have
(Zl - z2)<fl,J2) = <Zt!I,J2) - <iI' z 2f 2)
= <T*fl,J2) - <fl' Tf2) == o.
Consequently <fl ,f2) =0. o
Theorem 5.42. Let T be normal and injective. Then we have:
(i) R(T) is dense,
(ii) T* is injective,
(iii) T - I is normal,
(iv) R(T) = R(T*).
PROOF. R(T) is dense because of the equalities R(T).!. = N(T*) = N(T) =
{OJ. Consequently, T* is injective, and we have (T*)-I=(T- I)*. There-
fore it follows that
(T-I)*T- I = (T*)-IT- 1 = (TT*)-I = (T*T)-l
= T-1(T*)-1 = T-1(T- 1)*.

i.e., T - I is normal. It also follows that

Corollary. If T is a normal operator, then R(z, T) = (z - T)-I is normal for

all z t1. alT). This holds in particular for a self-adjoint T.

Theorem 5.43. If T is normal, then

p( T) = {z Ell<: (z - T) is continuously invertible}
= {z Ell<: R( z - T) = H},
(Jp(T) = {ZEII<: R(z- T)~H}.
(Compare with Theorems 5.23 and 5.24 for self-adjoint operators.)
5.6 Normal operators 127

PROOF. If z E peT), then (z - T) -I E B(H) (in particular, z - T is continu-

ously invertible). If z - T is continuously invertible, then (as T is closed)

R( z - T) = D( (z - T) - 1) = D((z - T) - I) = R( z - T),
and
R{z - T)..l = N(z* - T*) = N(z - T) = {O}.
Consequently R(z - T) = H. If R(z - T) = H, then
N(z - T) =N(z* - T*) =R{z - T)..l = {O},
i.e., z - T is injective. As T is closed and R(z - T) = H, it follows that
(z - T)-I E B(H), i.e., z E p(T).
If zEo/T), then R(z-T)..l=N(z*-T*)=N(z-T)*{O}. Therefore
R(z - T)* H. If R(z - T)* H, then N(z - T) = R(z - T)..l =1= (O}. There-
fore z E op(T). 0
Theorem 5.44. If T E B( H) is normal, then the spectral radius r( T) equals
II Til·
PROOF. By Theorem 4.46 we have for all T E B(H)
IIT*TII = sup {1<T*Tf,f) :fEH,·lIfll ~ l}
= sup {II Tfl12 :fEH, IIfll ~ I} = IITlf
In particular, II T211 = II T112 for a self-adjoint T. We obtain by induction
that for a self-adjoint T
II T2"11 = II T1I2" for all n E N.
Let T now be normal. Since reT) ~ II Til always holds, we only have to
prove that reT) > /I T/I. Because of the equality reT) = r(T*) and the
self-adjointness of TT* we have
r(T)2 = r(T)r(T*) = lim
n ..... oo
{IIT 2"IIII(T*tll}I/2"
;;;. lim
n~oo
IIT2"{T*(1I1/2" = lim
n~~
II(TT*)2"U I / 2"
= IITT*U = IITII2. 0

EXERCISES

5.38. If T E B(H), then there exist uniquely determined self-adjoint operators

T., T2 E B(H) such that T= T. + iT2. We have T. = HT+ T*), T2 =(1/2i)
(T - T*). T, is called the real part of T and T2 is called the imaginary part of
T. The operator T is normal if and only if T. and T2 commute, i.e., if and
only if T 1 T2= T2T •.
5.39. Let T be an operator on the Hilbert space H and let M be a closed subspace
of H. The subspace M is said to be invariant under T if T(M n D(T» c M. If
M and M1. are invariant under T and D(T)=[MnD(T»)+[M..lnD(T»),
then M is called a reducing subspace of T.
(a) If M is a reducing subspace, then M 1. is a reducing subspace, also.
128 5 Closed linear operators

(b) If T E B(H), then M is a reducing subspace of T if and only if M is

invariant under T and T*.
(c) M is a reducing subspace of T if and only if TP:) PT for the orthogonal
projection Ponto M.
(d) If T is densely defined, D(T)= O(T*), and M is a reducing subspace of T,
then M is a reducing subspace of P.
5.40. Let T be a normal operator on H.
(a) If M is a reducing subspace of T, then M is also a reducing subspace of
T*.
(b) We have R(T)": R(P). The restriction of Tonto N(T)..L n O(T) is an
injective normal operator on N( T)..L.
Hint: Theorem S.42.
5.41. Let T be a bounded normal operator on H.
(a) If 'A E o( T), then there is a sequence (jn) from H such that IIfnll = I and
('A - T)f,,-+O.
(b) In complex case we have II Til = sup Wf Tf)1 : f E 0(1'), IIfll = I}.
5.42. If T" and T are bounded normal operators such that Tn":~ T, then we also
have r:~ T*.
Hint: T: ~ P and II T:fll-+ II P fll for all f.
5.43. Assume that H= $nENHn> Pn is the orthogonal projection onto H", and Tn is
an operator on Hn. The orthogonal sum T= $"ENTn of the operators T" is
defined by
D(T) = {fEH: P"JED(T,,),
nEN
~ IIT"PJII2< oo},
Tf = ~ TnP"J for f E D(T).
nEN
(a) If all Tn are self-adjoint (normal, closed), then T is also self-adjoint
(normal, closed).
(b) If On is a core of Tn(n E 1\1), then L{ On : n E I\I} is a core of T.
(c) If each Tn is bounded and sup {II Tn II : n E I\I} < 0<:', then T is also
bounded and II Til = sup {II Tn II : nEI\I}.
(d) If all Tn are non-negative, then T is also non-negative. The Friedrichs
extension of T is the orthogonal sum of the Friedrich!: extensions of Tn.
5.44. (a) Let A be a bounded self-adjoint operator and P an orthogonal projection
in a Hilbert space H. Then PAP is self-adjoint. (See also exercise 6.13 for
unbounded A.)
(b) A corresponding result for normal operators does n(H hold in general.

U
Counterexample is

H= C 3 : N = o
o I)
0, P = (00 o
1
I 0 0 o
N is unitary with Nel = e2, Ne2 = e3 and Ne3 = el; P is the projection onto
L(e2' e3). For dim H=2 such an example cannot exist.
(c) If N is normal and P is an orthogonal projection with PN c NP, then
PN P is also normal.
 6
Special classes of linear operators

6.1 Finite rank and compact operators

Let HI and H2 be Hilbert spaces. An operator T from HI into H2 is said to
be of finite rank (of rank m) if R(T) is finite-dimensional (m-dimensional).

Theorem 6.1. Let T be an operator from HI into H2 such that D(T) = HI' The
operator T is a bounded operator of rank m if and only if there are linearly
independent elements fl' ... ,fm from HI and linearly independent elements
g" ... ,gm from H2 such that
m
Tf = ~ <t,j)~ for all f E HI' (6.1)
)-1

Then
m
T*g = '"i. <~, g)t for all g E H2, (6.2)
)-1

and II Til "~j!..,lltll II gjll. The operator T is of rank m if and only if T* is of

rank m. There is no loss of generality in assuming that either {fl' ... , fm} or
{g" ... , gm} is an orthonormal system}
PROOF. If T has the form (6.1), then R(T)c L(gl" .. , gm). For every
Jo E {l, ... , m} there exists an hjo E LUI' ... ,fm) such that "io =I: 0, and
"iol.t forj=l:Jo' Since <to' hjo ) =1:0 and Thjo=<to' "io)8}o' it follows that all
8} are contained in R(T). This implies R(T) = L(g" ... , gm); hence R(T)

IFrom Theorem 7.6 it follows that both {fl' ... ,fm} and {81' ... ,8m} may be chosen to be
orthonormal if suitable scalar factors are added, i.e., Tf = 'i.j_ISj<Jj,f>~.

129
130 Special classes of linear operators

is m-dimensional. Because of the inequality

m
IITfll oro; ~ l<jj,J)llIgjll oro; 11111 ~ Iljjllllgjll,
1 j-1

T is bounded and II Til oro; ~j_ .lljjll II ~II.

Assume T bounded, dim R(T) m, gl"'" gm} is ONB of
R(T), andjj= T"'~ forj= 1, ... , m. Then for allfE HI we have

m m
Tf = ~ gp TJ>~ = ~ <jj, f)gr
j- j-

remains show that the elements fl' ' .. , are linearly independent
Let us assume that this does not hold. There is no loss of generality in
assuming fl Then
m m
f)gj ~<jj,f>(alg,+~), f HI'
j-2

would follow from this that R(T) is at most (m I)-dimensional, which

contradicts the assumption. If T has the form (6.1), then for all f E HI>
H2
m m
<g, TJ) = ~ <Jj,J)<
I
gj) = <~ <~, g)Jj, f).
I

Hence (6.2) ho14s for T"', If is of rank then (6,2) implies that T'" of
rank m, too. The opposite direction follows the same way. Our construc-
tion shows that { . , , ,gm} can be chosen to be an ONS. If the same
reasoning is applied to T", then we obtain that {fl' ... ,fm can be chosen
be ONS. 0
EXAMPLE 1. If Ul' ' .. ,jm} and { .. , , gm} are orthonormal systems in
and H2, respectively, and ttl' . , . , J.tm IK, then

defines an operator T E B(HI , H2) of rank m with TIl max 1 J.tA :

... , m}. We only have to prove the last statement For allJE H
m
II T1112 = L 11LJ121<'~, 1)1 2 OS;; 11111 2max {I ILl :j = 1, ... , m }.
)-1

Therefore II Til max {II;I : j = m}, If jo {I" ., is chosen

such that 11101= max {I J.tjl : j = , ... , m}, then 11 1hoil 11-;01; thus Til
lJ.tjJ
Let HI and H2 be Hilbert spaces. An operator T from HI into H2 is said
to compact if every bounded sequence (fn) from D(T) contains
6.1 Finite rank and compact operators 131

subsequence Unk ) for which (Tfnk ) is convergent (Le., T maps bounded sets
onto relatively compact sets; cf. Exercise 6.1).

Theorem 6.2. Every compact operator is bounded. If T is compact, then f is

also compact.
PROOF. Assume that T is not bounded. Then there exists a sequence (fn)
from D( T) with the properties 1If" II "I and II Tfn II ;;;. n for all n EN.
Therefore, no subsequence Un• ) has the property that (Tfnk ) is convergent;
thus T is not compact. Let T be compact. If (fn) is a bounded sequence
from D( T) =DC T), then there exists a sequence (gn) from D( T) such that
"n -).
II gn - In II Since T is compact, there_exists a subsequence (gn) of
(gn)' for which (Tgn) is convergent. Then (Tfn) is also convergent because
of the inequality II fin. - Tgnk II "n -)11 Til· 0

On the basis of Theorem 6.2, together with Exercise 4.7, there is no loss
of generality in assuming that compact operators from H) into H2 always
belong to B(HIt H2).

Theorem 6.3. Let HI and H2 be Hilbert spaces. An operator T E B(HIt H2) is

compact if and only if Tfn _0 for every weak null-sequence (!,.) from HI'
PROOF. Let T be compact. It is sufficient to prove that every weak
null-sequence (!,.) from HI has a subsequence (fn) such that Tin. _0. Let
(fn) be a weak null-sequence from HI' As T is compact, there exists a
subsequence (fnk ) such that (Tf"k ) is convergent; say g = lim T1.nk . As 1.nIf ~ 0,
by Theorem 4.27 (ii) and Theorem 6.2 it follows that Tfn ~ 0, and thus
Jnk "" w - lim T'
g = lim Tf Jnk = 0. k

Let T now send every weak null-sequence from HI to a null-sequence.

Consider a bounded sequence (fn) from H). By Theorem 4.25 there exists a
weakly convergent subsequence (fn) of (fn); say f n• ~f. Then f n• - f~O
and T(!,.. - f)_0, by assumption. Therefore (Tfn) is convergent. Conse-
quently, T is compact. 0

°
Proposition. Let H be an infinite dimensional Hilbert space. If T E B(H) is
compact, then E aCT).
PROOF. If (en) is an arbitrary orthornormal sequence in H, then Ten_O.
Consequently, T is not continuously invertible. 0

Theorem 6.4. Let H .. H2 and H3 be Hilbert spaces.

(a) If S E B(H2' H3 ) and T E B(HI' H2), and one of these operators is
compact, then ST is compact.
(b) If T), T2 E 8(H), H2) are compact and a, bEll<, then aTI + bT2 is com-
pact.
(c) T E B(H), H2) is compact if and only if T* T is compact.
(d) T E B(HI' H2 ) is compact if and only if T* is compact.
132 6 Special classes of linear operators

sequence of compact "",,"'rf""n' from B( HI> H,) Til

some T E B( HI> H1),

be compact. If weak null-sequence by

Theorem 4.27 (ii) the sequence (Tfn) is also a weak null-sequence. As S
is compact, then STfn ~O; hence ST is compact. Now let T be
compact. Then Tfn~O for every weak null-sequence (jn) from HI' Since
S is continuous, we also have STJ" ~O. Therefore ST is compact in this
case, also.
HI> then TI Hence (aTI
compact, then the is also compact
compact. If (J,,) is nun-sequence in HI,

Consequently, T is compact.
(d) If T is compact, then (T*)* T* = TT* is also compact by part(a). Hence
T* is also compact by part(c).
(e) Let (jn) be a weak null-sequence from HI' The sequence (jn) is then
say Ilf"II"; C for We show that for
there exists an that II Tf,,1i < ~ Let
given. Since II T" exists an mo

II ~ C - I.

Since Tmo is compact, there exists an no E N for which

II TmJnll ..; i~ for all n > no'

It follows from this that for all n >no
o
by Boo(Hh H2 )
will be justified
Boo(H).

Proposition. It follows from Theorem 6.4 that Boo(H) is a closed two-sided

ideal of B( H).
Every rank I bounded operator from HI into H2 is compact (because T
has the form Tf = <g, f)h with g E HI' hE H2 ; for every weak null-
sequence (j,,) have TJ" = <g, Theorem 6.4(b)
then every finite is also compact, Theorem 6.4(e)
is every operator limit, with respect of B(HI> H2 ),
a sequence of operators. Actually, all the compact
operators.
6.1 Finite rank and compact operators 133

Theorem 6.5: An operator T E B(H), H2) is compact if and only if there

exists a sequence (Tn) of finite rank operators from B(HJ, H2) for which
II Tn - TII~O. For every compact operator T the subspaces N(T).!. and R(T)
are separable.
PROOF. One direction has already been proved. Let T now be compact.
First we show that N(T).!. is separable. Let {ea : a E A} be an ONB of
N( T).!.. As T is compact, Tea" ~O for every sequence (a,,) from A such that
a" =foam for n =l=m. It follows from this that for every E > 0 there exist only
finitely many a E A such that II Tea II > E. Consequently, the set A is at most
countable, i.e., N(T).!. is separable.
Let {e" : nEI'\I} be an ONB of N(T).!. (if N(T).!. is finite dimensional,
then the following reasoning is simpler), and let Pm be the orthogonal
projection onto L( e I' . . . ,em)' Then Pm ~ P, the orthogonal projection
onto N(T).!.. The operators Tm = TPm are of rank at most m (Tmf = ~'::_)<
e",J)Te,,); thus they are compact. For every m there exists anfm E H such
that Ilfmll=1 and I/(T-Tm)fmll>IIT-TmIl/2. As «P-Pm)fm,g)=<
fm' (P - Pm)g)~O, it follows that (P - Pm)fm ~O. Therefore (T- Tm)fm =
T(P - Pm)fm~O, since T is compact. Hence 1/ T- Tmll ..; 211(T- Tm)fmll~
0, i.e., Tm~T. If {g,,: nEI'\I} is a countable dense subset of N(T).!., then
{Tg" : n E I'\I} is dense in R( T). Therefore R( T) is also separable. 0
Now we want to study the spectrum of a compact operator. For this we
need the following theorem.
Theorem 6.6. Assume that H is a Hilbert space, T E Beo( H), A ElK, A =1= O.
Then R(A - T) is closed.

PROOF. Let g E R(A - T). Then there exists a sequence (f~) from H such
that gIl = (A - T)f~~g. If /" is the orthogonal projection of f~ onto
N(A - T).!., then gIl = (A - T)f". If the sequence (fn) is not bounded, then
we can assume that 1/ fn II ~ 00 (since this holds for some subsequence of
%
(f,,». For hn = IIfnll- we then have II hn II = I and (A - T)hn~O. As Tis
compact, there exists a subsequence (h"t) of (h,,), for which (Th,,) is
convergent. Then the sequence hn• = A - (I//". II-)g". + Thn) tends to an
element h E H. Because of the relations h" E N(A - T).!. and II h" II = I we
have hE N(A - T).!. and Ilhll = l. On the other hand, (A - T)h = Jimk~oo ('A
- T)h". = 0; thus h E N('A - T). This is a contradiction.
The sequence (f,,) is therefore bounded. Since T is compact, there exists
a subsequence (f,,) of (f,,) for which (T/,,) is convergent. Then the
sequence In. = 'A -I( gn. + Tfn) also tends to an element f E H, and we have

g = lim ('A- T)fn = ('A- T)f E R('A- T). o

k-4OO •

Theorem 6.7. Let H be a Hilbert space over IK, and let T E Boo(H). We have
a(T)n (K\{O}) = a/T) n (K\ {O}). If H is infinite dimensional, then aCT) =
ap(T)u {OJ. The operator T has at most countably many eigenvalues that can
134 6 Special classes of linear operators

cluster only at O. Every non-zero eigenvalue has iinite multiplicity. The

number A =1= 0 is an eigenvalue oi T if and only if A* is an eigenvalue oi T*.
PROOF. 0 E (J(T) by the proposition following Theorem 6.2, if H is infinite
dimensional. To prove (J( T) n (IK \ {O}) = (Jp n (IK \ {O}), it is sufficient to
show that if A=1= 0 is not an eigenvalue of T, then Af£a(T). For this,let us
assume that A=1= 0, N(A - T) = O. Since R(A - T) is closed, A- T is a
bijective mapping from H onto the Hilbert space R(A - T). In what follows
we show that R(A - T) = H. From this we can infer that A E p( T), i.e.,
A f£ (J(T).
Let us now assume that R(A - T) =1= H. Define Ho, and Hn for n ;.. I by
the equalities Ho = Hand HII = R«A - T)"). Then for n E N the subspace
H,,+I is a closed subspace, strictly smaller than H" (the subspace H,,+I is
closed, because (A - T) - I is continuous and H" is closed; from Hllo = HIIO+ I
it would follow that H= (A - T)-"OH"o =(A - T)-"oH"o+1 = HI)' If for every
n E N we choose an element i" E Hn _ I e Hn such that II ill II = I, then (f,,) is
an orthonormal sequence. Since T is compact, Tin ~o. On the other hand,
for all n E N

Tin = Ain - (A - T)f".

Here (A - T)in belongs to H,,; consequently it is orthogonal to in' There-
fore,
II Tf,,11 ;.. IAllli,,1I = IAI·
This contradicts the fact that Ti,,~O. Hence R(A - T) = H.
N ow we show that the multiplicity of every non-zero eigenvalue A is
finite. If we had dim N(A - T) = 00, then there would exist an orthonormal
sequence (f,,) from N(A - T). Since T is compact, we would then have
Tf,,~O, which contradicts the equalities II Ti,,11 = IAlllill1l = 1.\1.
In the next step we show that the eigenvalues can cluster only at O. It
also follows from this that there are at most countably many eigenvalues.
Let us assume that there exists a sequence (>-..,) of pairwise different
eigenvalues of T such that A" ~A =1= O. Then there exists a sequence (1n)
from H such that lif" II = 1 and (An - T)in = O. We know from linear algebra
that the family Un : n E N} is linearly independent. Let Hn = LUI' ...• in)
w
(Ho = {O}). gn E Hn e H,,_I such that II gnli = I for n E N. Then gn ~ 0;
hence Tg,,~O. On the other hand, for all n E N
Tgn = >-..,gn - (An - T)gn.

where (>-.., - T)gn E Hn - I, since it follows from the equality gn = ~j_Ia,,/j

that
11 n-I
(>-.., - T)gn = ~ anj(An - T)h = ~ a,iAn -A)h E Hn- I·
.I-I j~1
6.1 Finite rank and compact operators 135

Because of the relation g" 1- H" _) we therefore have "Tg" II ;> IA"I, which
contradicts the fact that Tgn ~O.
The last assertion follows from the equalities a( 7'*) = a( T)* and (I( 1'*) n
(K\{O}) = (lp(T*)n (K\{O)) (the last equality holds, as T* is also com-
pact). 0

Theorem 6.8. Let H be a Hilbert space and let T E Boo(H). If A~O is an

eigenvalue of T (hence A* is an eigenvalue of T*) then N(A - T) and
N(A* - T*) = R(A - T)l. have the same dimension.
PROOF. We have dim N(A - T)" dim N(A* - T*) or dim N(A* - T*)"
dim N(A - T). We treat the first case. There exists then an isometric
mapping V of N(A - T) into N(A * - T*) = R(A - T)l.. Let P denote the
orthogonal projection onto N(A - T). The operator P is of finite rank,
hence compact. For the compact operator T) = T + VP we then have
N(A - T) = to}, i.e., A is not an eigenvalue for T). By Theorem 6.7, A* is
then not an eigenvalue for Tr, i.e., to} = N(A* - Tn= R(A- T)l. =
(R(A - T) ED R( V» 1. • Consequently, R( V) = R(A - T)l., and thus
dim N(A - T) = dim R( V) = dim R(A - T)l. = dim N(A * - T*). The other
case can be treated similarly. 0

EXERCISES

6.1. Let H be a Hilbert space. A subset A of H is said to be compact provided that

every sequence (In) from A has a subsequence that is convergent in A. The
subset A is said to be relatively compact if A is compact.
(a) A is compact if a!1d only if every open cover of A contains a finite cover of
A.
(b) An operator T from HI into Hz is compact if and only if the set
T A = {Tf : f E A} is relatively compact for every bounded subset A of
D(T).
(c) A subset A of Iz is relatively compact if and only if it is bounded and for
every E > 0 there exists an no E N such that for all f = (fn) E A we have
I~_nolJnlz.;;; (.
(d) Let A be a set of continuous functions defined on IRm, having the following
properties: A is bounded in the sense of L2(lR m ), A is equicontinuous on
every compact subset of IRrn, and for every £ > 0 there exists an r ;;;. 0 such
that f Ixl;;.,I!(x)l z dx .;;; ( for all f E A. Then A is a relatively compact
subset of L2(lRm).
6.2. (a) Let (tn) be a null-sequence from C. The equalities D(T) = 12 and Tf= (tnfn)
for all f = Un) E 12 define a compact operator T E B( 12 ),
(b) Let t : [0, I} X [0, 1)~C be continuous. The equalities D( T) = L2(0, 1) and
Tf(x) = fbt(x,y)f(y)dy for all fELz(O, 1) define a compact operator
T E B(LiO, I».
6.3. If T is a compact operator on the Hilbert space H, then the set T {f E H : Ilfll
.;;; I} is compact.
136 6 Special classes of linear operators

6.4. A sesquilinear form s(. , .) defined on a Hilbert space is said to be compact if

w w
fn ~ 0 and gn ~ 0 imply s(fn' gn)~O.
(a) Every compact sesquilinear form is bounded.
(b) A sesquilinear form is compact if and only if the operator induced by it
(cf. Theorem 5.35) is compact.
6.5. (a) Let T be a compact operator on H. For every XE 11<, X 0 we have the *
Fredholm alternative: Either the equations (X - T)f = g and (X· - T·)h = k
are uniquely solvable for all g, k E H or the homogeneous equations
(X - T)f =0 and (X· - T·)h =0 have nontrivial solutions.
(b) The spaces of solutions of the two homogeneous equations have the same
dimension, and (A - T)f = g is solvable if and only if g is orthogonal to
every solution h of the equation (A· - T·)h = O.
6.6. Let H be a separable Hilbert space.
(a) If (fn) is an increasing sequence of finite rank projections on H sUfh that
Pn ~ I, then Pn TPn ~ T for every compact operator T 011 H.
(b) Boo(H) is a separable subspace of B(H) (cf. Exercise 4.8).
s .
6.6. Let An' A E 8(H2' H3 ) and BE 8(HI' H2)' If An ~A and B IS compact, then
AnB~AB.

6.2 Hilbert-Schmidt operators and Carleman

operators
We begin by studying one of the most important classes of compact
operators, the class of Hilbert-Schmidt operators. Let HI and H2 be Hilbert
spaces. An operator T E B(H .. H2 ) is called a Hilbert-Schmidt operator if
there exists an orthonormal basis {ea : a E A} of HI such that

Theorem 6.9. An operator T E B(H .. H2 ) is a Hilbert-Schmidt operator if

and only if r is a Hilbert-Schmidt operator. Then

1/2 ( )1/2
IITII..;; ( L IITfal1 2 ) = L IIT·ep1l2 < 00 (6.3)
aEA ~EB

for arbitrary orthonormal bases {fa : a E A} of HI and {ep : {J E B} of H2 •

The common value of the square roots in (6.3) is called the Hilbert-
Schmidt norm of T and is denoted by III Till. We have: II Til ..;; III TIll =
IIIT*III, because of (6.3). The set of Hilbert-Schmidt operators is denoted
by Bz(HI , H2) or B2(H) (for the justification of the index 2, cf. Section 7.1).
6.2 Hilbert-Schmidt operators and Carleman operators 137

It is not hard to show that ~(Hh Hz), equipped with the Hilbert-Schmidt
norm, is a Hilbert space (d. Exercise 6.11).
PROOF. Let T E 8(HI' H2) be a Hilbert-Schmidt operator and let {ea :
a E A} be an ONB of HI such that LII Tea 112 < 00. If {eb : f3 E B} is an
arbitrary ONB of Hz, then

~ II T *epW= ~ ~ 1<T*eb, eu )1 2 = ~ ~ I<ep, Tea)l z

PES PES uEA aEA PES

i.e., T* is a Hilbert-Schmidt operator. One can prove the converse in the

same way and obtain the equality sign in (6.3) at the same time. If
{eb : {3 E B} is an ONB of Hz, then for aIlfE HI

II Tfll2 = ~ I<eb, Tf)l z ~ IIfll2~ IIT*ebll 2 •

P p
Consequently II TII2 ~ ~f.lll T*epll2. Hence, (6.3) is completely proved. 0
Theorem 6.10. An operator S from HI into Hz is the restriction of a
Hilbert-Schmidt operator T if and only if S is closable and D(S) contains an
orthonormal basis {ey : y E r} such that LyErilSeyll2 < 00. Every Hilbert-
Schmidt operator is compact.
PROOF. Let T E B(HI' Hz) be a Hilbert-Schmidt operator and let SeT.
Then S is bounded, hence D(S) = D(S). Consequently, D(S) is a Hilbert
space, so it contains an ONB {ey : y E F}. Since {e y": y E r} is contained
in an ONB of H .. it follows by Theorem 6.9 that LIiSeyf = ~II TeyliZ < 00.
Let S now have the given property. We show that S is compact (the
compactness of an arbitrary Hilbert-Schmidt operator follows from this).
- w
Let (J,,) be in D(S) and assume that /" ~O, I': > O. The subset {e y : <ey,fn)
*0 for some n E N} of {ey : y E r} is at most countable; we denote it
simply by {el' e2' ... }. We have

fn = ~ <11,fn)ej and ~ IISI1112 < 00.

j j

Because of the inequality

1 <ep 1,,)IIISeJ II
1 ~ { 71<ep/")121 II S I1W r /2 = /11,,/1 { 1
II Se)1 2r/2,
the series ~/ty,fn)SI1 is convergent. Since S is closed, it follows that

Sfn = ~ <ep fn)SI1'

j
138 6 Special classes of linear operators

It follows from this that for all N E Nand n E N,

IISf,.1I <; ~ l( ep f,.)IIISe;1I

J

<; ~ l(eJ ,f,.)IIISe;1I +{ ~ l( ep f,.)1 2 ~ II seJ11 2} 1/2

J<N J>N j>N

<; ~ l(epf,.)IIISeJII + II f,. II { ~ IIseJ1I2} 1/ 2,

J<N J>N
Because of the inequality ~JII St111 2 < 00 and the boundedness of (1If,.lI)
there exists an No EN such that

Ilf,.ll{ ~ IIseJ 1I2 } 1/2 < 4( forall n E N.

J>No

Since fn ~ 0, there exists an no E N such that

~ l(epfn)IIISeJIl < t( for all n:> no'

j<No

It follows from this that II Sfn II « for n ;;;a no, hence that Sj,1~0, i.e., that S
is compact.
Since the compact operator S is continuous, we have D(S)=D(S). We
define an extension T of S by the equalities D(T) = HI and

T(j + g) = Sf for f E D(S), g E O(S)l..

T is a Hilbert-Schmidt operator, since TEB(HI' H2) with IITII=IISII
(cf. Exercise 4.7) and if {gp : f3 E B} is an ONB of D(S)L, then {ey : 'Y E
T} U {gp : f3 E B} is an ONB of HI such that

~ II Teyl12 + ~ II Tgp ll 2 = ~ II Seyl1 2 < 00. 0

Y fJ Y

Corollary. Let T E B(HI' H2). S E B(H2• H3). and let one of these operators
be a Hilbert-Schmidt operator. Then ST is a Hilbert-Schmidt-operator.
B( H2) is therefore a two-sided ideal of B( H).
PROOF, Let T be a Hilbert-Schmidt operator and let {e" : a E A} be an
ONB of HI' Then

aEA aEA

If S is a Hilbert-Schmidt operator. then T* S* is a Hilbert-Schmidt

operator using Theorem 6.9 and what we have just proved. Hence ST =
(T* S*)* is also a Hilbert-Schmidt operator. 0
Now we return to our earlier definition of a Hilbert-Schmidt operator
(cf. Section 4.1. Example 3).
6.2 Hilbert-Schmidt operators and carleman operators 139

'Theorem 6.11. Let MI and ~ be measurable subsets of RP and IR',

respectively2. The operator T E B(L2(M 1). L2(~» is a Hilbert-Schmidt oper-
ator if and only if there exists a kernel K E L2(M2 X M I ) such that

Tf(x) - J.M.
K(x,y)f(y) dy almost everywhere in M2,f E L2(M1). (6.4)

The a4/o;nt operator T* ;s then induced by the aefjo;nt kernel K+(x,y)-

K(y, x)'.
PROOF. If T is of this form, then T E B(L2(M.), ,L2(M2 »,
by Section 4.1,
Example 3. If {en: n EN} and {1m : mEN} are orthonormal bases of
L2(M1) and ~(~), respectively, then {gnm : (n, m) E N X N} is an OND of
L2(M2 X M 1), where gnm(x, y) - fm(x)enCy)' (ct. Theorem 3.8). Hence

~ II Ten 112 - bnl(Ten,fm>12 - ,tm!fM3fM.K(X,y)·ell(Y)· dyfm(x) dXr

- ~ I(K, gllm>12 ... IIKII2 < 00,
lI,m

i.e., T is a Hilbert-Schmidt operator.

Let T now be a Hilbert-Schmidt operator, and let ell' fm and gllm be
defined as above. Let us define cllm by
C llm - (fm. Tell) for (n, m) E N X N.
Then

so there exists a K E L2( M2 X M1) such that

(gllm' K) - cllm for all (n, m) E N X N.
If To denotes the operator induced by K in the sense of (6.4), then with
h(x,y)-g(x).r(y) we have for allfE L2(M.), gE L2(M2) that

<g, To!) - J. g(x)* J.

M3 M.
K(x,y)!(y) dy dx
'

lI,m lI,m

Hence T- To. o
2 M. and M2 can be replaced by arbitrary measure spaces in most cases. Sometimes we have
to assume that the ~ are a-finite.
140 6 Special classes of linear operators

Theorem 6.12. Let T be an operator from a Hilbert space H into L2(M).

Then the following assertions are equivalent:
(i) T is the restriction of a Hilbert-Schmidt operator,
(ii) there exists a function k : M~ H such that II k( . )11 E L2( M) and
Tf(x) = <k(x),f> almost everywhere in M, f E D(T),
(iii) there exists a function K E L2( M) such that
ITf(x)1 < IIfIlK(x) almost everywhere in M, f E D(T)
(of course, the sets of exceptional points in (ii) and (iii) depend on f and on the
choice of the representative of Tf.)
PROOF. (i) implies (ii): Let T C S and let S E Bz(H, L2(M». We show (ii)
for the operator S. Since S is compact, N(S).L is separable (Theorem 6.5).
Let {e l , e2 , ••• } be an ONS of N(S).L. Then

L !.ISe,,(x)12 dx = L IISe,,1I2 < 00;

" M n
consequently, by B. Levi's theorem, L n ISen (x)1 2 < 00 almost everywhere in
M and f M~nISen(x)12 dx < 00. Hence we can define the function k : M~H
by the formula

k(x) = {o~ (Sen(x»'"en if

otherwise.

With this function k we have for all f E H that

<k(x),f> = L <en,J)Se,,(x) = Sf(x)

"
almost everywhere in M. Because of the equality IIk(x)1I 2 ,= L"ISe,,(xW, the
function IlkOIl belongs to L2(M).
(ii) implies (iii) by taking K(X) = IIk(x)lI.
(Iii) implies (i): It follows from (iii) that

for all f E D( T). So T is bounded. f also satisfies (iii). This can be seen in
the following way. For every f E o(f) = D( T) there exists a sequence (fn)
from D( T) such that fn ~ f; therefore also Tfn ~ ff. By Theorem 2.1 there
exists a subsequence (f,,) such that Tf".(x)~ ff(x) almost everywhere.
Hence

Iff(x)1 = klim ITf" (x)1 < k-+oo·

-+oo·
lim IIf" IIIC(x) = IlfIlIC(x).
6.2 Hilbert-Schmidt operators and Carleman operators 141

almost everywhere. This holds also for the operator S E B(H, L2(M»
defined by

S(f+g) = TJ for JED(T),gED(T).L.

We show that S is a Hilbert-Schmidt operator and IIISII1 2 < JMIK(XW dx. It
is obviously sufficient to show that
n
L II Sej ll 2 < !.IK(X)12 dx for every finite ONS {e" ... , en}
i_I M

in H. Let Sty( •) be an arbitrary (however, in what follows fixed) representa-

tive of Sej . Let us define A : L(e\> ... , en)~L2(M) by the equality

(A(~ ajej))(x) = ~ aiS~(x),

Then A = SIL(el .. "", en) in the sense of L2(M). Since the set Lr(e l , ••• , en) of
linear combinations of el> ... , en with rational coefficients is countable,
there exists a subset N of M of measure 0 for which
/AJ(x)/ < IIJIIK(X) for all x E M\N,j E L,.(e l , , •• , en)'

If J= "2.'J_laj~ E L(e\, ... , en)' and we choose rational sequences (ajk)kEN

so that ajk~aj as k~oo, then withJk =~j-l ajkej it follows that

for all x E M\ N. Consequently,

/AJ(x)1 < IIJIIK(X) for all x E M\N,j E L(el' ... ,en)'
For every x E M\ N the mappingJl-i-AJ(x) is a continuous linear functional
on L(e\, ... , en) whose norm is not greater than K(X). Hence there exists a
k(x)E L(el"'" en)c H such that IIk(x)1I <K(X) and
AJ(x) = <k(x),j) for x E M\N,J E L(el" .. ,en)'
It follows that

f
j-I
II Aej l1 2 = f. f
Mj_1
I<k(x), ~)12 dx = !.lI k (x)1I 2 dx
M
< !.IK(XW dx.
M
0

A linear operator T from a Hilbert space H into L2(M) is called a

Carleman operator if there exists a function k : M~H such that for all
IE D(T)
TJ(x) = <k(x),j) almost everywhere in M. (6.5)
By Theorem 6.12 every Hilbert-Schmidt operator from H into L2 ( M) is a
Carleman operator. "
A function k: M-+H is said to be measurable if the function
<k(·), I) : M~ IK, xl-i-<k(x),j) is measurable for every f E H. If T is a
142 6 Special classes of linear operators

Carleman operator and k is an inducing function of T (in the sense of

<
(6.5», then the function k( . ), f) is measurable for all f E D( T). This
obviously also holds for all f E D( T). If P denotes the onhogonal projec-
tion onto D(T), then <Pk(' ),j) = <k('), Pf) is therefore measurable for
every f E H, i.e., Pk(·) is measurable, and for all f E D(T)
Tf(x) = <Pk(x),j) almost everywhere in M.
Consequently, there is no loss of generality in assuming that k is measur-
able.
If k : M~H is measurable, then the equalities
D(Tk ) = {J E H : <k(· ),f) E L2 (M)},
(6.6)
TJ(x) = <k(x),j) almost everywhere in M,J E D(Tk ),

define an operator Tk from H into L2(M). The operator Tk is called the

maximal Carleman operator induced by k. An operato:r is a Carleman
operator if and only if it is the restriction of a maximal Carleman operator.

Theorem 6.13.
(a) Every Carleman operator is closable. The closure of a Carleman operator
is a Carleman operator. Every maximal Carleman operator is closed.
(b) If Tp T2 are Carleman operators (with inducing functions kl' k;) and
a, bE C, then aTI + bT2 is a Carleman operator (with inducing function
ak l + bk 2 )·
(c) If T is a (maximal) Carleman operator from HI into L 2(M) (induced by
k) and S E B(H2' HI), then TS is a (maximal) Carleman operator from
H2 into L2(M) (induced by S*k).
(d) Let T be an operator from H into L2(M), and let P be the orthogonal
projection onto D( T). The operator T is a Carleman operator if and only
if TP is a Carleman operator.
(e) If H is separable and kl' k2 are inducing functiom of a Carleman
operator T, then Pkl(x) = Pk2(x) almost everywhere in M, where P
denotes the orthogonal projection onto D( T).
PROOF.
(a) Since an operator is a Carleman operator if and only if it is a
restriction of a maximal Carleman operator, it is suffident to show that
every maximal Carleman operator is closed. Let k : M~ H be measur-
able, and let Tk be defined by (6.6). Take a sequence (in) from D(Tk)
for whichfn~fE H, TJn~gE L2(M). Since <k(x),fn>~<k(x),f) (for
all x EM), we have
<k(x),f) = g(x) almost everywhere in M.
HencefED(Tk) and TJ=g, i.e., Tk is closed.
(b) This assertion is obvious.
(c) Again, it is sufficient to show the statement for maximal Carleman
operators. Let Tk be the maximal Carleman operator induced by k.
6.2 Hilbert-Schmidt operators and Carleman operators 143

Then
O(TkS) = {JEH2: SfEO(Tk )} = {fEH2: <k(·), Sf)EL 2(M)}
= {JEH2: <S*k('),J)EL2 (M)},
TkSf(x) = <k(x), Sf) = <S*k(x),f) almost everywhere in M,
i.e., TkS is the maximal Carleman operator induced by S*k.
(d) By part (c), along with T the operator TP is also a Carleman operator.
Since T is a restriction of TP (we have O(TP) = OCT) + O(T)-L and
T= (TP)ID(T»' along with TP the operator T is also a Carleman
operator.
(e) If {e l• e2, ... } is an ONB of OCT) that is contained in D(T), then it
follows that
(kl(x) - k 2(x), ell) = 0 almost everywhere, for all n E N,
and thus
IIP(k,(x) - kix»1/2 = ~ I<k,(x) - k 2 (x), ell )1 2
II

= 0 almost everywhere. o
Theorem 6.14 (Korotkov [46]). An operator T from H into L2(M) is a
Carleman operator if and only if there exists a measurable function K : M~ R
such that for all f E O(T)

ITf(x)1 -< IIfIlK(X) almost everywhere in M. (6.7)

PROOF. If T is a Carleman operator induced by k, then (6.7) holds with

K(X) = I/k(x)ll. Let (6.7) now be satisfied. Then there exists a bounded
function g : M~(O, 00) for which gK E L 2(M); for example we can choose
the function

g(x) = [(I + Ixl)m(l ~ K(X» r \ x EM (for Me Rm).

If G E B(L 2(M» is the operator of multiplication by g, then for all
fE OCT),
IGTf(x)1 -< IIfll g(X)K(X) almost everywhere in M.
By Theorem 6.12 the operator GT is therefore the restriction of a Hilbert-
Schmidt operator and there exists a function k' : M_H such that GTf(x)
= <k'(x), f) almost everywhere in M. With k(x) = g(x) -Ik'(x) we there-
fore have for all f E OCT) that
Tf(x) = g(x)-'(k'(x),f> = (k(x),f> almost everywhere in M. 0
Theorem 6.15. Let H be a separable Hilbert space, and let T be an operator
from H into L2 ( M). The operator T is a Carleman operator if and only if the
series ~"I Te,,(x)1 2 converges almost everywhere for every orthonormal system
.
{ el' e , • •• } In n/7"
2 ..... \~ ).
144 6 Special dassel; of linear operators

PROOF. If T is a Carleman operator, then, as in the proof of Theorem 6.14,

we can find a g : M~(O, 00) for which GT is the restriction of a Hilbert-
Schmidt operator. Then for every ONS {el> e2 , ••• } from OCT)

Consequently, by B. Levi's theorem,

g(X)2L iTen(xW = L iGTen(xW < 00 almost everywhere in M.

n n

Division by g2 gives the assertion.

Now suppose the series ~n i Ten(xW is almost everywher,e convergent for
every ONS {e" e2' ... } in O( T). First we show: If {e" e2' ... } is an
arbitrary ONS in OCT), then TiL(e" e2 •••• ) is a Carleman operator. For this,
let Mo = {x EM: ~ni Ten(x)i 2 = oo}. Mo is of measure zero by our assump-
tion. The function

k(x) = { ~ (TeAx»*en for x E M\Mo,

o otherwise,
induces TiL(e,. e, ... ,), since for f E L(e" e2, ... )

Tf(x) = L <en' f) Ten (x) = <L (Ten(x»*en.!)

n n
= <k(x),f) almost everywhere in M.
Let {e" e2, ... } now be an ONB of OCT), let k be an inducing function of
To = TiL(e,. e, . ... ) and let Tk be the maximal Carleman operator induced by
k. We show that T C Tk . For this, let f E OCT), let {j'.!2' ... } be the
ONS that arises from {j, e., e2 , ••• } by orthogonalization, and let k' be
an inducing function of r;= TiL(f,,j, .... ). Because of the inclusion L(e.,
e2 , ••• ) c L(f., J2' ... ) we have To C T.; hence k and k' are inducing
functions of To. If P is the orthogonal projection onto D( T) =
D( To) = L( e •• e2' ... ), then by Theorem 6.l3(e) we have Pk'(x) = Pk(x)
almost everywhere. Since f E O( T), it follows from this that
<k(x).!) = <Pk(x),f) = <Pk'(x),J) = <k'(x),f) = Td(x);
soJ E O(Tk ) and TJ= Td= TJ. o
Theorem 6.16. An operator T Jrom a separable Hilbert space H into L2(M) is
a Carleman operator if and only if TJ,,(x)~O almost everywhere in M for
every null-sequence Un) Jrom O(T).
PROOF. It is evident from the definition that every Carleman operator has
this property. It remains to prove the reverse direction. By Theorem 6.15 it
is sufficient to show that the series ~ni Ten (x)i 2 is almost everywhere
6.2 Hilbert-Schmidt operators and Carleman operators 145

convergent for every ONS {e l • ez, ... } from O( T). Let {el' ez, ... } be an
ONS in O( T). Assume that there exists a measurable subset N C M such
that A(N»O and ~"ITe,,(x)lz=oo for xEN (A stands for Lebesgue
measure). For all m, lEN let us define Nm • 1 by the equality

Nm ,/ = {XEN: ±
n=1
ITen (XW>m 2 }.

Then N= UIENNm ., for every mEN, and there exists an./(m)EN such
that

Consequently, for No = nmENNm,l(m) we have

A(No) > (1- ~ rm)A(N) > O.

m-I

For all mEN we have

f(m)
~ ITe,,(xW > m2 for x E No'
,,=1

By Exercise 6,9, for every mEN there exist finitely many elements (xm , ) =
(~m,j,I' • . . , ~m.), f(m» E cl(m), j = I, 2, ... ,p(m) for which we have: Ixm,l

=~~:>tI~m,}, ,,1 2 <;; 2m -2 and for every x = (~I' •.• , t'(m» E cl(m) with Ixl 2 >
m 2 there exists ajE {I, ... ,p(m)} for which

Let us set
I(m)

gm,j = ~ Em.}. "e".

n=1

Then for every mEN and for every x E No there exists aj E {I, ... ,p(m)}
such that

Thus, for the sequence

(g,,) = (gl. I' gl, 2' . . . , gl,p(l)' g2, I' . . . , g2,p(2), g), I' ... )

we have: gn -+0 and for every x E No there exists an arbitrarily large n EN

such that Tg,,(x) > 1. This contradicts the assumption. 0
 146 6 Special classes of linear operators

Theorem 6.17. An operator T from Lz(M.) into L2(M2) is a Carleman

. operator if and only if there exists a measurable function K : M2 X M.-loC
such that K(x, .) E L2(M.) almost everywhere in M2 and

Tf(x) == f.
MI
K(x,y)f(y) dy almost everywhere in M1,f E D(T}. (6.8)

Such a kernel K is called a Carleman kernel.

PROOF. If T is induced by a Carleman kernel K in the sense of (6.8), then
the assumption of Theorem 6.14 (Korotkov) is fulfilled with IC(X)'"
IIK(x, .)11: so T is a Carleman operator. If T is a Carleman operator, then
we proceed as in the proof of Theorem 6.14. GT is then a Hilbert-Schmidt
operator from Lz(M1) into Lz(M2); therefore, by Theorem 6.11, it is
induced by a kernel K' E L1(Mz x M1). The kernel K(x, y) == g(x)-IK'(x, y)
is then a Carleman kernel and it induces T. 0
Let K: M-loH be a measurable function, and let Mn == {x EM: IIk(x)1I
..; nand Ixl ..; n}. Let
D(Tk • O) - {g E L2(M): there exists an n E N such that
g(x) = 0 almost everywhere in. M\Mn}.

For every gE D(Tk • O) the equality

<Tk • og, f) = J.Mn g"'(x)<k(x), f) dx for all .f E H

. . of the inequalities
uniquely defines an element Tk og, since because

IfMng"'(X)<k(X),f) dxl" nllfllfMnlg(x)1 dx ..; nX(Mn)I/2I1gllllfll

the function ft-+ f M g"'(x)<k(x), f) dx is a continuous line:ar functional on

H. The mapping g~ Tk. 0 g i~.,obviously linear. Tk,o is therc:fore an operator
from L2(M) into H; D(Tk • O) is dense in L2(M). The operator Tk,o is called
the semi-Carleman operator induced by k.

Theorem 6.18. We have (Tk • 0)* == Tk • (In what follows we write Tt, 0 for
(Tk,o)"'·)
PROOF. By the definition of Tk. O we have for all f E D(Tk) and g E D(Tk. 0)

<g, TJ) - <g, <k('),f» = fMg(x)"'<k(x),f) dx =: <Tk.Og.!),

i.e., the operators Tk and Tk • O are formal adjoints of each other; therefore
Tk c T:'o' It remains to prove that D(T:.o)cD(Tk ). LetfE D(1f:.o). Then
6:2 Hilbert-Schmidt operators and Carleman operators 147

for every g E L2(M,,) and all n E N we have

J. g*(x)(k(x),f) dx =(Tk.Og,j) = (g, Tt,of) = J. g*(x)Tt.of(x) dx.

~ ~
Consequently,

{g*(x){(k(x),f)-Tt,of(x)}dx=O forall gEL 2(M,,).

"
Because of the relation {(k(·), f) - Tt, of(' )}I~ E L2(M,,) it follows from
this that
T: of(x) = (k(x),j) almost everywhere in M".

As this holds for all n, it follows that (k(· ),j) = Tt, of E L2(M), i.e.,
fED(Tk }· 0
If K: M2 X M,-loC is a Carleman kernel and k denotes the mapping
k: M2 -loL2(M,), k(x) = K(x, .), then we write TK = Tk and TK.O= Tk. O' It
follows from the definition of Tk • O (by Fubini's theorem) that for all
fE D(TK • O) we have

TK.of(x) = f M2
K(y, x)*f(y) dy almost everywhere in M,.

Theorem 6.19. Let T be a densely defined Carleman operator from L2(M,)

into L2( M2) that is induced by the Carleman kernel K. The acfjoint T* is a
Carleman operator if and only if K+ is a Carleman kernel (K+(x,y)=
K(y, x)* for (x, y) EM, X M2) and f -::J TK+, o. Then T* is induced by K +.
PROOF. By assumption, T C TK. As T is closable, D(T*) is dense. If T* is
defined by the Carleman kernel H : M, X M2-loC, then T* c TH • Conse-
quently, f. T**-::J Tk= TH,o-::J TH,o' It remains to prove that H(x,y)-
K +(x, y) almost everywhere in MI X M2 • Let
\

MI,,, = {XEMI: fM2IH(x,Y)12dy<.n and 1~I<.n},

Then HIM I.,. XM1: E L2(M I "X M2) and H+IM :il XM I." E L2(M2 X MI ' II)' Hence
I

the function H +, as a kernel on M2 X M" II> is a Carleman kernel. There-

fore, for every n E N, TH. 0IL1(M 1• n } is a Carle man operator induced by H +
and K, Consequently, by Theorem 6.l3(e) we have H+(x,y)-K(x,y)
almost everywhere in M2 X M I ,,, for every n E N. Hence H +(x, y)-
K(x,y) almost everywhere. In particular, K+ is a Carleman kernel and
f:; TK+,o' If K+ is a Carleman kernel and f:; TK+. O' then T*= f*c
(TK +. O)* = TK +; so T* is a Carleman operator induced by K+. 0
148 6 Special classes of linear operators

Corollary. If T is a symmetric Carleman operator on L2(M) with inducing

kernel K, and T* is also a Carleman operator, then the kernel K is
Hermitian, i.e., K(x,y)= K+(x,y) almost everywhere. In particular, the
kernel of a selj-acijoint Carleman operator is Hermitian.

EXERCISES

6.7. Let HI and H2 be Hilbert spaces and let HI be separable. A dosable operator
T from HI into H2 is the restriction of a Hilbert-Schmidt operator if and only
if there exists an orthonormal basis {el' e2, ... } of D(T) such that ~nll Ten 112
< 00 (cf. Theorem 6.10).
6.8. Let HI be a separable Hilbert space and let T be an operator from HI into
L2( M). Assume that
(i) If Un) is a null-sequence from D(T) and (Tfix» is convergent almost
everywhere in M, then Tfn(x)-'>O almost everywhere,
(ii) there exists an orthonormal basis of D(T) such that ~nITen(x)12 < 00
almost everywhere in M.
Then T is a Carleman operator.
Hint: cf. Theorem 6.15. Remark: (i) cannot be replaced by the assumption
of the closability of T.
6.9. For every mEN and C > 0 there are finitely many elements Xj =
(~j. I' . . . , ~. m) E em,j = I, ... , p = p(rn, C), for which Ixjl < 2C - I and for
which we have: for every x = (~I' ... , ~m) E em with Ixl;;.· C there exists a
jE{l, ... ,p} such that

I(x, xj>1 =1 f ~;~j.kl;;.

k-I
1.

6.10. If K is a Hermitian Carleman kernel over M x M and TK. 0 is bounded, then

TK is from B(L2(M» and TK is self-adjoint.

6.11. Let HI and H2 be Hilbert spaces.

(a) The equality (T, S> =~"EA(Te", Sea> defines a scala.r product on the
space B2 ( HI, H2) (here let {e" : a E A} be an arb*ary orthonormal basis
of HI)' The corresponding norm is the Hilbert-Schmidt norm.
(b) Sz(HI , H2) is a Hilbert space with this scalar product.
(c) With this norm, I3z(H) is a Banach algebra (without identity element in
the case dim H is infinite).
(d) We have (S, T> = ('r, S·> for all S, T E B2(H I , H2 ).
6.12. Let T be an operator on L2(M), and let (zo- T)-I be a Carleman operator
for some Zo E p( T).
(a) (z - T) - I is a Carleman operator for all z E p( T).
(b) If T is self-adjoint and kz(x,y) is the kernel of (z - T)-I for some
z E peT), then kz.(x, y) = kz(Y, x)·.
Hint: Notice that [(z - T) ~ II· = (z· - T) ~ 1, and use 'Theorem 6.19.
6.3 Matrix operators and integral operators 149

6.3 Matrix operators and integral operators

Let HI and H2 be infinite dimensional Hilbert spaces over IK, let {en : n E
N} and {e~ : n EN} be orthonormal bases of HI and H2, respectively, and
let (ajk)j, kEN be an (infinite) matrix with ajk E II(. (If HI or H2 is finite
dimensional, then some simplifications arise.) First we show that the
formulae

DCA) = {fEHI: lim

m-+oo
f
k-I
ajk(ek,f> exists for all jEN, and

~ I~
j=1 k-I
ajk(ek,J> 12 < (0), (6.9)

Af = j~1 (k~l ajk<ek,f) )e; for f E DCA)

define a linear operator from HI into H2: Iff, g E D(A) and a, bEll(, then
the limit

lim f
m-+oo k = I
ajk(ek' af + bg> = lim {a
m--+oo
f
k_ I
ajk(ek,J> + b
k
f= ajk(ek, g)}
I
00 00

= a ~ ajk(ek,f) +b ~ ajk(ek, g)
k-I k-l

obviously exists, and we have

. ; 2{ lal2j~11 k~l ajk (ek,f>12

+ Ibl 2 ~ I~
j-I k-I
ajk(ek, g>12} < 00.

Therefore, it follows that af + bg E D(A) and A(af + bg) = aAf + bAg.

Theorem 6.20. Let HI> H2 , {en: n EN}, and {e~ : n EN} be as above. If
(~k) is a matrix such that ~;: tlajkl2 < 00 for all kEN, and A is the operator
from HI into H2 defined by (6.9), then the following holds: D(A) is dense in
HI, A * is a restriction of the operator A + from H2 into HI induced by the
150 Special classes linear operators

a4ioint matrix (a;t) = (a~) analogously to (6.9):

m
: lim L + g> exists for all E and
m .... oo k-I

+g =L (L
j-I k=1
ajt<e~, g»)ej g D(A +

If we also have ~ k _ II ajk 12 < 00 for all j E N, then A is dosed and A * is

densely defined.
PROOF. For all n N we have
00

L ajk(ek, e,,) = ajn

k-I
00

=L )=1
00.

Consequently, L({e,,: n N}) D(A); so is dense. Let

A 11.(( e. : "e Nl)' We show that At = A +; this then implies that A * C A +. It is
easy to see that Ao and A are formal adjoints of each other, i.e., that we
have A C What remains to prove that D(At) C D(A Let g E
D(Ati). Then, because of the relation ek E D(Ao), we have for every kEN
that

C~ ajkej, g)
00

(ek , Atig) = (AOek, g) = ~ aJ(ej, g).

)=

Consequently,

i.e., g E D(A +).

we also have ~k_,lajkI2 < 00, then A = (Ao+)*, where At =
A +IL({e~ : "EN))' Hence A is closed, and A is densely defined. o

Theorem 6,21, Let and be separable Hilbert spaces, and let be a

densely defined operator from HI into H2. The operator T is closable if and
only if there exist orthonormal bases {e,,: E HI and
: E oj and matrix (ajk ) with the properties: ~;:'llajkI2< 00
for all kEN, ~k_llajkll < 00 for all j E N, and T is a restriction of the
6.3 Matrix operators and integral operators 151

operator A deJined by (6.9). The orthonormal bases can be chosen Jrom D(T)
and D( T*), respectively.

PROOF. If T has this form, then it is closable by Theorem 6.20. If T is

closable, then there are orthonormal bases {ell : n EN} of. HI in D( T) and
(since O(T*) is dense) {e~ : n EN} of H2 in O(T*). With

ajk = (ej, Tek) = (T*eJ; ej ), (j, k) E N X N

we then have
00 00

L lajk l2 = II Tek ll 2 < 00, L lajk l2 = II T*e;II 2 < 00,

)=1 k=1

for all kEN and for all j E N, respectively. For every f E OCT)

f If
)=1 k-I
ajk (e k ,f)1 =
2
f If
)-1 k-I
(T*ej, ek )(ek ,f)\2
00 00

L I(T*eJ,f)1 2 = L I(ej, Tf)1 2

)=1 i-I

= II Tfl12 < 00.

Therefore f E D(A). Moreover, we have

Tf = f (ej, Tf)ej = f (T*ej,f)ej = f ( f

)=1 )=1 i-I k-I
(T*e;, ek)(ek,f»)e;

= f ( f ajk(ek,f»)e; = Af·
i=1 k-I o
Proposition. If T is a symmetric operator on a separable Hilbert space H,
then there exists an orthonormal basis {ell: n E N} and a Hermitian matrix
(ajk ) such that ~k- .Iajkl2 < 00 for all j E N, and

PROOF. In the proof of the preceding theorem choose for {ell: n E N} and
{e~ : n EN} the same orthonormal basis in O( T) c D( T*). The series of
equalities

shows this matrix is Hermitian. o

152 6 Special classes of linear operators

Corollary. If T is a symmetric operator on 12 and 12. 0 C D{ T), then there

exists a Hermitian matrix (~k) such that }:f_llajk I2 < 00 for all j E Nand

Tf = TUn) = ( f
k-I
anJk )
nEN
for all f E D( T).

This can be deduced from the previous proposition if we choose the

basis {en: nEN} so that en = (8";)jEN'
Now we prove some simple criteria for the boundedness of the operators
that are induced by matrices in the sense of (6.9).

Theorem 6.22. If }:j, klajkl2 = c 2 < 00, then the operator A defined by (6.9) is
a Hilbert-Schmidt operator, and IliA III = c.
PROOF. By Theorem 6.20, A is densely defined and closed. For the basis
{en: nEN} we have

Therefore, by Theorem 6.10, A is a restriction of a Hilbert-Schmidt

operator. Since A is densely defined and bounded, we have D(A) = D(A)=
HI' and thus A E 8(HI> H2 ). Hence A is a Hilbert-Schmidt operator, and
IIIAIII=C. 0
Theorem 6.23. Let ajk = !J.ikCjk; furthermore, let
L Ibjk l2 0;;;; c~ for a/l j E N,
k
and
L ICjkl 2 <: cf for all kEN,
j

<: L !cjkI 21(ek.J)12

C? j,k
0;;;; C?CfL l(ek.J)1 2
k

= C?CfllfIl 2 •
6.3 Matrix operators and integral operators 153

It follows from this that HI = D(A) and IIAfll .-;;; C I C 2 11fll for allf E D(A) =
HI· 0
Proposition. The assumptions of Theorem 6.23 are satisfied in particular if

L lajkl .-;;; C I2 for all j E N

k
and
L lajkl .-;;; ci for all kEN.
j

For the proof, in Theorem 6.23 let us take bjk = lajk ll / 2 and cjk =
l'7kll/2 sgn '7k' where sgn a =0 for a = 0 and sgn a = lal-Ia for a ~O.

EXAMPLE 1. The matrix

for k '-;;;j,
for k >j
defines a bounded operator A by (6.9) and we have IIA II .-; ; Y6 . For the
proof, let us define bjk = cjk = 0 for j < k and
bjk =j-I/4k- I / 4,Cjk =j-3/4k l / 4 for k '-;;;j.
Then we have

and

~ Icjk l2 = kl/2.~ j-3/2 .-;;; kl/2{ k- 3 / 2 + foo X- dX} 3/ 2

J J=k k

= k- I + 2 .-;;; 3,
for all j E Nand kEN, respectively. The assertion (ol1o'fs from this with
the aid of Theorem 6.23.

Similar arguments can be made for operators T from L 2(M I ) into L 2(M2 ),
that are induced by a measurable function (kernel) K : M2 X MI ~C. Let K
be such a kernel. The equalities

D(TK ) = {1 E L2 (M I ) : K(x, .)1 is integrable over MI almost

everywhere in M2 , and 1. K(. ,y)f(y) dy is in L (M

M,
2 2 )},

TKf(x) = f M,
K(x,y)f(y) dy almost everywhere in M2 (6.10)
154 6 Special classes of linear operators

define a linear operator from L2(M I ) into L2(M2) (notice that the notation
TK is compatible with that of Section 6.2). Special cases of such operators
(Hilbert-Schmidt operators and Carleman operators) have: already been
studied in Section 6.2. Now we want to give a further criterion for the
boundedness of an integral operator (formally this is an exact analogue of
Theorem 6.23 for matrix operators).
Theorem 6.24. Let k : M2 X MI-j>C be measurable, and let KI and K2 be
measurable functions defined on M2 X MI such that K(x, y) =
K I (x,y)K2(x,y) and
J.
M.
IK I (x,y)1 2 dy "ct almost everywhere in M2,

J.
M2
IK2 (x,y)1 2 dx "ct almost everywhere in MI ,

Then the operator TK is in B(L 2(M I ), L2(M2», and II TKII .;;; C I C2 • The adjoint
T; is equal to the operator TK+ induced by the adjoint kernel K+,
PROOF. For every r > 0 let Mir) ... {x E M2 : Ixl "r}. The Lebesque
measure of M2(r) is finite. For every f E L2(M I ) Fubini's theorem implies
that
J.
M2(r) X M.
IK(x,y)f(y)ldx dy " J. M2(r)
{I + [J.
M.
IK(x,y)f(y)1 dy ]2} dx
= X(M2(r» + J. [J.
M2(r) M.
IK,(x,y)IIKix,y)f(y)1 d.v]2 dx

" X(M2(r» + J. [J. IK,(x,y)1 dy J. IK (x,y)f(yW dY] dx

Ml(r) M.
2
M.
2

" X(M (r)) + C J. If(Y)1 J. IK (x, y)12 dX] dy

2 l2 2[ 2
M. M2

" X(M2(r)) + clcillfl1 2 < 00.

By Fubini's theorem again, the integral JM K(x, y)f(y) dy exists almost

everywhere in M2(r) and defines a measurable function there. Since this
holds for all r> 0, the equality
TKf(x) =
,
1. K(x,y)f(y) dy
M.
almost everywhere in M2

defines a measurable function on M2 for every f E L2(M,). For all those

x E M2, for which this integral exists, we have
6.3 Matrix operators and integral operators 155

and thus

Hence, for every f E L1( M1) the function TKf belongs to L1( M1), and
II TKfll "Ct C2 11fll· Since the mapping f~ TKf is obviously linear, we then
»
have TK E B(L1(M t ), L2(M1 with II TKII "C\C1. The corresponding argu-
ments show that K+ induces an operator TK + E B(L1(M2), L2(M t » and
that for all f E L2(Mt> and g E L2(M2 ) the integral

f. M2
Ig(x)lj IK(x,y)f(y)1 dy dx
MI

exists. Therefore, it follows from Fubini's theorem that

<g, TKf) = J. g(x)'" J. K(x,y)f(y) dy dx

M2 MI

= J. f<y){ J. K+(y, x)g(x} dX} '" dy = <TK+g,j)·

MI M2

Hence, TJ: = TK +· o
Corollary. If K : M2 X Mt~C is measurable, and

f. IK(x,y)1 dy "Ct
Ml
almost everywhere in M2,

J. IK(x,y)1 dx "C{
M2
almost everywhere in M t ,

then the operator TK defined by (6.10) is from B(L 2(M), L2(M2 », and
IITKII <;C t C2·
This follows from Theorem 6.24 if we take K\(x, y) = IK(x, Y)lt/2 and
K 2(x,y) = IK(x,y)l tj2 sgn K(x,y).
It is important to observe that the operators occurring in Theorem 6.24
are not necessarily compact (cf. the following example).
EXAMPLE 2. Let M) = M2 = R The kernel K(x, y) = exp (-Ix - yl) satisfies
the assumption of the above corollary. Assume that f E L2( Mt),j '=F 0, and
f(x) > 0 almost everywhere. Then TKf'=FO. Let us setfn{x) = f(x - n). Then
w
fn~O, and

TKfn(x) = f exp ( -Ix - y/)f(y - n) dy

. . f exp (-I(x - n) - yl)f(y) dy = (TKf)(x - n).
156 6 Special classes of linear operators

Consequently, II TKf,,1I = II TKfl1 for all n. So we do not have TKf"--,,,O, i.e.,

TK is not compact.

Theorem 6.25. Let K: M2 X MI--",C be measurable, and let Kp K2 be

measurable such that K(x,y) = K I(x,y)K2(x,y). Let (MI) and (M2) be
increasing sequences of measurable subsets of MI and M2 such that MI =
U}~ I MIj and M2 = U'f'.. I M2), respectively. Assume that for all j E N

J.M2j XM'j
IK(x,y)1 2 dx dy < 00,

and for every ( > °there exists a Jo=joe () EN for which

J.
M,
IK I(x,y)1 2 dy .;,;; ( almost everywhere in M:! \ M2}o'

J.M,\Mljo
IK I(x,y)1 2 dy .;,;; ( almost everywhere in M l ,

f.M 2\M2jo
IK2(x, y)1 2 dx .;,;; ( almost everywhere in MI'

Then TKE B(LiM\), L2 (M2 », and TK is compact.

PROOF. Let us set

H;(x,y) = {K(X,y) for (X,y)EMVXMlj'

° otherwise
and ~. = K - H;. Then THJ is a Hilbert-Schmidt (hence compact) operator,
and II TL II .;,; ( provided that J ;.. Jo( (). The operator TK is therefore the limit
of a seq~ence of Hilbert-Schmidt operators, so that it is compact (Theorem
6.4(e». []

EXAMPLE 3. Let M c IR m be measurable and let

where fl and f2 are bounded measurable functions defined on IR m and

h(x)--",O as Ixl--",oo(j = I, 2); f3 is measurable and fIi3(X)1 dx < 00. With
MI} = Mv = {x EM: Ixl ';';;j} and
K.(x,y) =fl(X)lfix-y)II/2

K2(x,y) = f2(y)lf3(X-y)II/2 sgnfix-y),

the above theorem is applicable; hence TK is compact.
6.4 Differential operators on L1(a, b) with constant coefficients IS7

EXAMPLE 4. Let M be a bounded measurable subset of 81 m , and let

K(x,y) = {lx-Yla-mH(x,y) for (x,y)EMXM,x~y,
o for x=y EM;
where a > 0, and H is a bounded measurable function on M X M. If we set

_ {K(X'Y) for Ix-YI>';,

Kn(x,y) - 1
o for Ix - y I< -n
and Ln = K - Kn , then the Kn induce Hilbert-Schmidt operators, and the
operator induced by L" converges to 0 as n-HlO. Hence TK is compact.

6.4 Differential operators on L 2(a, b) with

constant coefficients
In the following let (a, b) be an arbitrary (non-empty) open interval in R,
i.e., let - 00 "a
< b " 00. Furthermore, let
A,,(a, b) = {J: (a, b)-+C : f,f', ... ,fn - 2) are
continuously differentiable on (a, b) and

fn-I) is absolutely continuous on (a, b)}.

Hence, for f E A,,(a, b) there exists an "nth derivative" fn) for which the
following holds: fn) is integrable over every compact subinterval of (a, b),
and for a <a <P <b and every functbn g, absolutely continuous on (a, b),
we have (cf. Appendix AS)
ffJg(x)fj)(x) dx = g(P)f J - I)({3) - g(a)f(j-I)(a)
a

- fPg'(x)fJ-I)(x) dx
a

forjE{1, ... , n}.

Theorem 6.26. For every n EN, every interval (a, b), and every f > 0 there
> 0 such that lor all j E {O, 1, ... , n - I} and all 1 E An(a, b)
exists a C

fblfJ)(xW dx " £fblfn)(xW dx + C fblf(xW dx.

a a a

Here we have to consider an integral to be eqllal to 00 in case the integrand is

not integrable. The relations f E A,,(a, b) n L2(a, b) and f n ) E L2(a, b) there-
fore imply that fJ) E L2 (a, b) for j E {I, ... , n - I}.
158 6 Special classes of linear operators

PROOF. We prove this by induction on n. For n = I we only have to treat

the case j = O. Then for every f > 0 we may choose C = I. The assertion is
just as obvious for n = 2 and j = O. Therefore, let n = 2, j = I.
Let (a, {J) be an arbitrary bounded subinterval of (a, b), kt L = {J - a>
0, and let J1=(a,a+1L), J2=[a+1L,{J-1L1, J3=({J-1L,{J). For
arbitrary s E J 1 and 1 E J3 , by the mean value theorem there exists an
xo= xo(s, t)E(a, {J) such that
J'(xo) = (/-s)-I[1(/)-I(s)].
With this Xo it follows for all x E(a, {J) that

1J'(x)1 = 1J'(xo) + ~:f"(Y) dyl

.;;; 3L -1(ll(s)1 + 11(t)!) + fiJlf"(Y)1 dy.

IX

This holds for all s E J 1 and t E J3 • The inequality can be itntegrated over
J I with respect to s and over J3 with respect to t. We obtain that

3- 2L 2 1J'(x)1 .;;; f 11(s)1 ds + f 11(t)1 dl + r

~ ~
2L 2jiJ 1f"(Y)1 dy
IX

.;;; jiJ I1(Y)1 dy + 3- 2L 2jiJlf"(Y)1 dy

IX IX

.;;; {L ~iJlJ(Y)12 dy} 1/2 + r2L2{ L ~iJlf"(Y)12 dy} 1/2.

It follows from this that

1J'(x)12 .;;; 2L{ 34L -4~iJll(Y)12 dy + ~iJlf"(Y)12 dy},

and thus by integration over (a, /3) that

jiJlf'(xW dx .;;; 162L -2 jiJ I1(Y)1 2 dy + 2L2 jiJlf"(YW dy.

a a IX

If we divide (a, b) into (finitely or infinitely many) disjoint intervals of

length L, then we obtain

f blJ'(x)1 2 dx .;;; 162L -2 f bll(xW dx + 2L2 fblf"(xW dx.

a a a

As L can be chosen arbitrarily small, the assertion follows for n = 2 and

j= 1.
Let us now assume that the assertion holds for n';;; k (k;;" 2). Let
f E ~+ I(a, b). Since the theorem holds for n = 2 and) = 1, for every 11 > 0
there exists a C 1 ;;.. 0 such that
fblfk)(xW dx .;;; clj\r<k-I)(XW dx + 11jblfk + n (xW dx.
a a a
6.4 Differential operators on L2(a. b) with constant coefficients 159

By the induction hypothesis there exists a C2 > 0 for which

jb1lk-I)(xW dx <; c2 jb 1f(xW dx + 4CI-ljbllk)(XW dx;
a a a
consequently,
jb1lk)(x)1 2 dx <; 4jb1lk)(xW dx + C 1C2 jblJ(XW dx
a a a
+'lJjb1lk+I)(xW dx,
a

and thus
jb1lk)(xW dx < 2C I C2 fblf(xW dx + 2'IJjb1/ k+1)(xW dx.
a a a

This is the assertion for n = k + 1 and j = k. For j <k the assertion now
follows easily with the aid of the induction hypothesis. 0
In what follows we shall use the notation
W2,n(a, b) = {JEAn(a, b)nL2(a, b) :In)EL2(a, b)}.
W2, n(a, b) is called the Sobolev space of order n over (a, b).

Theorem 6.27. Let (a, b) be an arbitrary open interval in R, and let

f E W2 , n( a, b). If - 00 < a, then I;) can be extended continuously to a for all
jE {O, I, ... , n - I}; if a = - 00, then limx->_oo/J)(x)=O. The correspond-
ing assertion holds for the point b.
PROOF. Let c E (a, b). If a> - 00, then

~CIIJ+I)(x)1 dx <; {(c-a) ~cIIJ+I)(xW dX} 1/2 < 00,

because f(J+ I) E L2(a, b) (Theorem 6.26). Hence the limit

lim IJ)(x) = lim {f})(c)-jCIJ+1)(s) dS}

x~a X~Q x

= IJ)(c) - jCIJ+I)(s) ds
a

exists.
Now let a = - 00. For all x E( - 00, c) we have

jCIJ)(s)IJ+I)(s) dx = HI})(c)2- IJ)(xi}.

x

The integral here converges as x -,) - 00, therefore the limit

lim x ->_ 00 IJ)(xi also exists. If this limit were different from zero, then IJ)
could not be in L2(a, c); which would contradict Theorem 6.26. 0
 160 6 Special classes of linear operators

Now we shall study differential operators on Lia, b) that are induced

by the differential form 7'n = ( - in d n/ dx "). The minimal operator Tn. o
induced by 7'n is defined by
D(Tn.O) = Co""(a, b) and Tn. of = T"f for f E D(Tn.o),
The maximal operator Tn induced by 7'n is defined by
D(Tn) = W2 • n(a, b) and Tnf = Tnf for f E D(Tn)'
The maximal operator is obviously defined on the largest possible sub-
space on which Tn can act meaningfully. Below we shall show that
T:' 0 = Tn; if A c Tn. 0 and A*- Tn. O' then A * ~ r:.
0 = Tn' i.e., A * is not a
differential operator induced by ~this is why Tn. 0 is called the minimal
operator (it is also usual to call Tn 0 the minimal operator). For n = 0 we
obviously have To. 0= To = I. .

Theorem 6.28. We have k E R(T". 0) if and only if k E Co""(a, b) and

fa
b .
x1k(x) dx = 0 forall j E {O, I, ... , n-l}.

PROOF. If k =
T".og for some g E D(Tn. 0) = CoOO(a, b), thell we obviously
have kE Co""(a, b), and for j E {O, I, ... , n - I}

fbxJk(x)dx = fbXJ(T"g)(x)dx = (-l)"fb(7'nxJ)g(x)dx = O.

a a a

Conversely, assume now that k E Co""(a, b) has this property, and [a, fl]
is a compact subinterval of (a, b) that contains the support of k. Set

Then we obviously have g E C""(a, b) and g(x) = 0 for x E(a, a]. For
. xE[fl,b) we have

g(x) -= (i)n fX fX ... fX k(x l) dXn ... dx ,

a XI Xn-l

= (W~xk(XI){~~'" ~~_ldXn ••• dx 2 } dx ,

= (i)n jbk(xl)P(x,) dx, = 0,

a

because p(x,) is a polynomial of degree n - 1. Hence, g E Co""(a, b) and

k=Tn.og. 0

Theorem 6.29. We have T:' 0 = Tn'

6.4 Differential operators on L2(a, b) with constant coefficients 161

PROOF. Let i E D( Tn)' g E D( Tn. 0) = COOO( a, g), and let [a, ,8) be a compact
subinterval of (a, b) that contains the support of g. It follows via integra-
tion by parts that
<Tni, g) = f\Tnf)*(x)g(x) dx
a
= f\Tni)*(x)g(x) dx
a

= ff3(T ni)*(x)g(x) dx = j f3f*(X)(Tng)(X) dx

a a

= fbf*(x)(Tng)(X) dx = <I, Tn.og)

a
(the integrated terms vanish, since g, together with all of its derivatives,
vanishes at a and ,8). Consequently, Tn. O and Tn are formal adjoints of
each other. What remains is to prove the inclusion T:' 0 C Tn' Let i E
D(T:'o) and hEA,.(a,b) such that T"h=T:'oi. Such an h exists; for
example we can define h in the following way:
h(x) = (if f Xfx n . . . fX2 (T:'oi)(x l ) dX I ... dXn
c c c

with some cE(a, b). Then it follows by integration by parts (let [a.,8) be
chosen as above) that for all g E D(Tn • 0)
<i. T",og) = <T:'oi. g) = jf3(T:' oi)*(x)g(x) dx
a

= jf3h*(X)Tng(X) dx = jbh*(x)Tn.og(x) dx.

a a
Hence, for all k E R(Tn • o) we have
fb(J(x) - h(x))*k(x) dx = O.
a
The null space of the linear functional F: CoOO(a, b)~C defined as
F(k) = f\i(x) - h(x»*k(x) dx for k E CoOO(a, b)
a
contains therefore R(TI1 • o)' By Theorem 6.28 we have
11-1
R( Tn. 0) =
j~O
n N( Fj);

where the Fj are the linear functionals

Fj : CoOO(a, b) ~ C, Fj(k) = f a
b .
x1k(x) dx.

By Theorem 4.1 there exist complex numbers CO' C I ' • . . , C,,_I such that
n-I
F = L cjFj.
j=O
11-1
Hence, with p(x) = L cl xi we have
i=O
fb(J(x) - h(x) - p(x»*k(x) dx = 0 for all k E CoOO(a. b).
a
 162 6 Special classes of linear operators

For every compact subinterval [a, ,8] we have (f - h - p)l[a.P) E L2(a, ,8)
and (f - h - P)I[a. P)l. CoOO( a, ,8), and thus (f - h - P)I[a. PI = O. Since this
holds for every compact subinterval [a, ,8], it follows that
f(x) = h(x) + p(x) almost everywhere in (a, b).
It follows from this that f E An(a, b) n L2(a, b) and 7'nf= T:' of E L2(a, b).
HencefE O(Tn)· 0

Theorem 6.30. III the case (a, b) = R we have Tn. 0= Tn = T:; thus Tn. O is
essentially self-adjoint and Tn is self-adjoint.
PROOF. The relations T:' 0 = Tn and Tn. 0 C Tn imply that T: C or:. 0 = Tn'
We show that Tn is symmetric; then it will follow that T: C Tn C T:; hence
Til = T: = T:'~ = Tn, o. For f, g E O(Tn} we have

<I, Tng) = )i.~ J:/*(X)(7'ng)(X} dx

= )i.~ {R(c) + J:,,< 7'nf)*(x)g(x) dX}

- }~ f:}7'nf)*(x)g(x} dx = (T,J. g),

because R(e) is a linear combination of terms of the form

. f(f)(± C)*g(k)(± c) withj + k = n - I, and so R(c)~O as C4·00, by Theorem
6E. 0
Theorem 6.31. In case (a, b) '1= R we have

_ { f(a}=f'(a)=·· ·=fn-I)(a)=O ifa>-oo,}

D(Tn• o)'" fE W2• n (a, b}: .
f(b)=f'(b)=·· . =f"-I)(b} =0 if b<oo

In this case Tn, 0'1= Tn (for n > 0), and none of these operators are self-adjoint.
PROOF. We write W£ n(a, b) for the subspace given in the theorem. Let Sn
be the operator induced on Wf lI(a, b) by 7'11' Then one verifies easily that
Sn and Til are formal adjoints ~f each other. Therefore Sn C r: ...
Tn. o. Let
f E D(r:), and let a> - 00. For every j E {O, 1, ... , n - II} there exists a
~ E O(Tn) such that gJk)(a) = 8jk for k E {O, I, .... , n ,- I} and gj(x)
vanishes identically in some neighborhood of b (we choose an arbitrary
smooth function ~ such that g}k)(a) =8jk • and set gj =: cp~l(a. b)' where
cp E CoOO(R) and cp(x)'" I in some neighborhood of a, cp(x) - 0 in some
neighborhood of b). With these ~ we have
II-I
0 ... <f, TII~) - <T:f,~) = (_i)n ~ (-I)"-k-If"-k .. I)(a)*gJk)(a)
k-O
... (- i)n( _ l)n- j - If"- j - I)(a}*.
6.4 Differential operators on Lz(a. b) with constant coefficients 163

As this holds for alljE {O, I, ... , n -l}, it follows that

f(a) = 1'(a) = ... = I,,-I)(a) = O.

It follows analogously that

f(b) = f'(b) = ... = I"-I)(b) == 0,

in case b < 00. Consequently D(T".o) = D(T!) c W£ ,,(a, b), and thus
S" = T" 0= r:.
On the basis of our reasoning so far, it is clear that (for n >0) D(T" 0)
-:1= D(T,,), thus T", 0-:1= T", Because of the equality T", 0*= T", none of th~se
operators are self-adjoint. 0
The question of whether T",o has self-adjoint extensions in the case
(a, b)-:I= IR, will in general be answered in Chapter 8, For even n we can
now give the following theorem.

Theorem 6.32. The operator T 2",o is non-negative. The Friedrichs extension

T2n , F of T2n , 0 is given by the formulae

f(a)=f'(a) = ... =/"-I)(a)=O, if a> -OO'}

D(T2n, F) = { f E D(T2n):
f(b) =1'(b) = , .. =In-I)(b)=O, if b<oo
= W2 ,2n(a, b) n Wf,,,(a, b),
T2".Ff = 'T'2,,! for f E D(T2n , F)'

PROOF. For allfE D(T2n , 0) = CoOO(a, b) we obviously have

(j, T 2n , of) = (_I)n f bf(x)*j<2n)(x) dx = f\rn)(xW dx :> O.

a a

In this case the sesquilinear form s used in Theorem 5,38 is

sU, g) = <In), g<n», f, g E CoOO(a, b).
Therefore the s-norm II . lis is equal to the T",o-norm II . II T•. o
(cr, Theorem 5.1). Hence, the completion of CoOO(a, b) with respect to
II . lis is equal to the completion of D(Tn 0) = CoOO(a, b) with respect to
II . II r...o' and thus it is equal to D(Tn,o)' Con'sequently, by Theorem 6.31
Hs =D(Tn.o) = W£,,(a, b)
f(a) =1'(a) = . . . =/,,-I)(a)=O if a> -OO'}.
= { fE W2.,,(a, b):
f(b) =1'(b) = ... =/,,-I)(b)=O if b<oo
164 6 Special classes of linear operators

The assertion follows from this, because of the equalities D(T2n F) =

D(Tin.o)n Hs= D(T2n )n Hs· 0
. .
In some cases we can explicitly calculate the spectrum of the self-adjoint
operators induced by 'T".

Theorem 6.33.
(a) If (a, b) = IR, then

a(T2n )=[0,00) and a(T2n _ I )=1R forall nEN.

(b) If (a, b) is a half-line «a, 00) or (- 00, b», then for every self-adjoint
extension A of T 2n . O we have
a(A) ~[O, 00).

For the Friedrichs extension T 2n . F of T 2n . 0 we have

a(T2n. F) = [0, 00).

PROOF.
(a) T 2n = T 2n . O is non-negative. Hence for every s < 0 we have

lI(s - T2n )j1l2 = Isl 2 11fll 2 - 2 Re sU, T2nf) + II T2nfl12 > Is1 2 11f1l 2 ,
i.e., s - T 2n is continuously invertible. Therefore, (.- 00,0) Cp(T2n)'
Now we show that every s > 0 lies in a(T2n)' To this end, let cp : IR-')IR
be infinitely often differentiable and let

I for x < 0,
cp(x) = { °<cp(x) < 1 for O<x<l.
o for x>l,
Furthermore, for all mEN let

CPm(x) = cp(lxl- m), x E IR.

Let us set (with s 1/2n > 0)

Then fm E COOO(IR), Ilfmll-') 1 as m-)oo and (s - T 2n )fm = (s - 'T'2n)jm-')0,

as a simple calculation shows. Consequently, s - T 2n is not continu-
ously invertible, and thus [0, 00) C a( T 2n ).
We can show analogously that IR C a(T2n - l ) if for every s E IR we
define the sequence (fm) of functions by the equalities

fm(x) = (2m)-1/2cpm(x) eis l/(2"-I>x, x E IR

6.4 Differential operators on L2(a, b) with constant coefficients 165

(here the root s 1/(211-1) has to be taken in such a way that s 1/(2" - I) < 0
for s < 0 and SI/(211-1);;;. 0 for s;;;' 0).
(b) Without loss of generality we may assume that (a, b) = (0, 00). Then
the inclusion [0, 00) C I7(A) follows as in part (a) if <Pm is replaced by
I/Jm(x) = fPm(x - m - 2)
(notice that I/Jm E CoOO(O, 00». Since the Friedrichs extension T 211• F is
non-negative, it follows that 0(T2", F) C [0, 00) (cf. the beginning of the
proof of part (a»; consequently, 0(T2n, F) = [0, 00). 0
Theorem 6.34. Let T be a self-adjoint operator on L2(a, b) induced by Tn
{i.e., T",oC T= T* C T,,), Let
d1'(x)
(of)(x) = Ln
aix ) - - .
)=0 dx J
be a differential expression such that sup {ia,,(x)i : x E(a, b)} < 1 and
sup {ia;(x)i : x E (a, b), j = 0, I, ... , n -l} < 00. Assume that the operator
S defined by the equalities
D(S) =D(T) and Sf = of for f E D(S)
is symmetric. Then T+ S is self-adjOint.
PROOF. If c = sup {ian(x)i : x E (a, b)}, then by Theorem 6.26, the operator
S is obviously T-bounded with T-bound c < 1. The assertion follows from
this by Theorem 5.28. D

EXERCISE

6.13 Let A be an (unbounded) self-adjoint operator in a Hilbert space Hand P an

orthogonal projection with P D(A) c D(A). Then the operator PAP is in
general not self-adjoint. Counterexamples may be constructed along the
following lines: Let H= Ho$ Ho, B self-adjoint in Ho, C symmetric with
B-bound < I and P(fl,fl) = (fl, 0) for (f.. f2) E H- Ho$ Ho·
(a) The operator Ao( ~ g)
with D(Ao) = D(B)$ D(B), Ao(fl,fv = (Bf2' Bfl)
is self-adjoint.
(b) The operator A = ( ~ ~) = Ao + ( ~ ~) is self-adjoint.
(c) If C is not (essentially) self-adjoint on D(B), then PAP is not (essentially)
self-adjoint. If C has no self-adjoint extension, then PAP has no self-
adjoint extension.
(d) Possible examples for Band C are as follows:
(a) C essentially self-adjoint, but not self-adjoint: B = - d2 / dx 2 , C =
id/dx in L 2(R) with D(B) ... D(C) = W2,2(R),
(fJ)C not essentially self-adjoint: B= -d2 /dx 2 , C=id/dx in L2(0, I)
with D(B) = D( C) = {f E W2,2(0, 1) : 1(0) = j(l) = O},
(y) C has no self-adjoint extension: B= -d 2 /dx 2 , C=id/dx in LiO, 00)
with D(B)=D(C)= {fE W2.iO, (0) : j(O) =O} (see also Section 8.2,
Example 1).
7
The spectral theory of self-adjoint
and normal opera tors

7.1 The spectral theorem for compact operators,

the spaces Bp( HI' H2)
We studied the spectrum of compact operators thoroughly in Section 6.1.
For compact normal operators the results obtained there: may be shar-
pened.

Theorem 7.t. (The spectral theorem for compact normal operators.)

(a) Let T be a compact normal operator on a complex Hilbert space Hand
let {AI' A2, ... } be the non-zero eigenvalues 'of Tj furthermore let
{PI' P2 , ••• } be the orthogonal (finite rank) projections onto the corre-
sponding eigenspaces (cf. Theorems 6.7 and 5.41), then

T = LAjIJj (7.1)
j

this series converges in the norm of 8(H). If T is self-ac:fioint, then this

holds in real Hilbert spaces, as well.
(b) If (~) is a null-sequence (or a finite sequence) from &1:\{O} such that
~ -:l=Ak for j-:l=k, and the IJ are non-zero orthogonal projections of finite
rank such ,that IJPk=Ofor j-:l=k, then the series (7.1),is convergent in
8(H), and T='i.jAjIJ is compact and normal. Furthermore, {AI' A2"" }
is the set of non-zero eigenvalues of T and the R(IJ) are the correspond-
ing eigenspaces. The representation (7.1) is therefore unique in this sense.
If the Aj are real, then Tis self-ac:fioint.
PROOF.
(a) Let M be the closed linear hull of {R(lj) : j = I, 2, ... }, and let P be
the orthogonal projection onto Mol.
166
7.1 The spectral theorem for compact operators, the spaces BP(HI' H2) 167

If MJ... =1= {O}, then for all 1 EM J..., g E H, and j = I, 2, ... we have
<Tl, ~g >= <1, T* Pjg >= >'-/<1, Pjg> = 0,
i.e., TM J... eM J...; we can prove in just the same way that T* M J. eM J....
Hence it follows that the restriction S of T to M J... is a normal operator
on the Hilbert space M J.... Every eigenvalue of S is an eigenvalue of T,
every corresponding eigenvector of S is an eigenvector of T contained
in M J...; therefore it follows that S can only have the eigenvalue 0 (since
M is the closed linear hull of all eigenvectors of T that belong to
non-zero eigenvalues). Therefore (cf. Theorem 6.7), a(S) = {O}, and
thus r(S)=O by Theorem 5.17(c) and (d). Hence S=O by Theorem
5.44, i.e., T vanishes on M J....
Consequently, it follows for 1 E H (whether M J... = {O} or M J. =1= {O})
that

Tl = TPl + T}."; ~1 = }."; TPjl = }.";"A;P;f.

j j j

If the sequence ("A;) is infinite, then for every 1 E Hand mEN

00

}."; IAlIIPj!12.;; sup {lAY :»m+ Qllll1 2•

j=m+1

Since the sequence ("A;) is a null-sequence (cf. Theorem 6.7), this

implies the norm convergence of (7.1).
(b) We can show the convergence of the series just as in part (a). For every
mEN the operator ~;"IAj~ is of finite rank, consequently compact.
The compactness of T follows by Theorem 6.4(e). It is easy to verify
that T is normal. All the Aj are obviously eigenvalues of T, and every
1 E R(~) is an eigenelement of T belonging to the eigenvalue Ai" If
A=I=O is an eigenvalue of T and 1=1=0 is a corresponding eigenvector,
then

0= II(A- T)1112 = 11A - Alll~1112 + IA12111-7 Pjr

i.e., IA-Ajl II~ll1 =0 for aU) and l=~jPj. Since f=l=O, there is a)o
such that ~J=I=O. Hence A = Ajo' Consequently, A#Aj for allj=l=)o, and
thus ~f = 0 for) =1=)0' It follows from this that f E R( Pj ) . 0

Theorem 7.2. (The expansion theorem for compact normal operators). /1 T

is a compact normal operator on a complex Hilbert space, then there exists a
zero-sequence (or a linite sequence) ( IL) from C and an orthonormal sequence
(D from H such that
Tl = }."; JJy<t,J>.t lor all 1 E H. (7.2)
j
168 7 The spectral theory of self-adjoint and normal operators

Conversely, every operator deJined by (7.2) is compact and normal; the

numbers J1.} are eigenvalues oj T; the elements Jj are corresponding eigenvec-
tors. IJ T is selj-a4ioint, then this is true in real Hilbert spaces, as well.
PROOF. Let T="i.j~1j be the representation from Theorem 7.1. For every)
let {gj, I' gj, 2' ••• ,gj,k) be an orthonormal basis of ROj); furthermore,
let Il j, k = 'Aj for k = I, 2, ... , kj' Then by Theorem 7.1 and Theorem 3.7

TJ = ~'}''JP/ = ~'Aj~ <gj,k' P/)gj,k = ~ ~ Ilj,k<gj,k,J)gj,k'

j ) k j k

(7.2) follows from this by changing the indices. We leave: the rest of the
proof to the reader. 0

Theorem 7.3. Let T be a compact normal operator on a complex Hilbert

space, and let ('Jy) be the sequence oj non-zero eigenvalues oj T; every
eigenvalue counted according to its multiplicity, and I~I > I~+d Jor all n.
Then
I'AII = IITII
I~+II = inf sup{1I TJII : J E H,J l.gl' ' .. ,g"" IIJII = I} (7.3)
81"" ,8. EH

Jor n E N (cJ. also Exercise 7.2). IJ Tis selj-a4ioint, then this holds in real
Hilbert spaces, as well.
PROOF. It follows from the relations r(T) = II Til and a(T)\{O} C a/T) that
I'A.I = II Til
(cf. Theorem S.l7(c) and (d». Let (Jj) be .an orthonormal
sequence such that TJj ='AjJj. If we choose gj = Jj for) = I, ... , n, then for
every Jl.gl' ... ,gil we have

II TJII2 = II ~ 'A/Jj,J)JjI1 2 = .~ 1~121<JjJ)12:

J>n J>"

<; 1'A..+d2~ I<Jj,J)1 2 <; 1'A,,+112I1Jf·

j>n

It follows from this that 1~+tI;> inf sup {... }. If gl' ... ,gn are arbitrary,
then there exists an J E LUI' ... ,JII+ I) for which IIJII = 1 and
Jl.gl' ... , gil' We have for this J that

It follows from this that I~+II <; inf sup {... }. Consequently, the theorem
is proved. 0
7.1 The spectral theorem for compact operators, the spaces Bp(H1, H2) 169

If T = ~j'f.ylj is the representation of the compact normal operator T,

given in Theorem 7.1, then for all n EN the equalities

obviously hold. If A)/n is chosen in some way, then the operator

A
n
= ~
~]
A,I/np,
}
j

has the property (An)" = T. The roots 'A)ln can be chosen in a unique way
if we require that, for example, 0.;;; arg A)ln < 2 'IT / n. Therefore, we have
the following theorem.

Theorem 7.4. Let n EN, n >- 2. Every normal compact operator T on a

complex Hilbert space has exactly one normal compact I nth root whose
eigenvalues all lie in {z E C : 0'" arg z < 2'IT / n). Every non-negative self-
acYoint compact operator has exactly one non-negative compact nth root.

PROOF. The operator An = ~j'AFnlj with 0.;;; arg 'A)/n < 2 'IT / n has the re-
quired property. Let B = ~kfJ.kQk be an operator having the same property.
Then we have in particular that

L fJ.:Qk = Bn = T = LAjlj.
k j

The uniqueness statement of Theorem 7.1 assures that J1:;n = Aj and Qj = Pj

(this is true perhaps only after an appropriate reindexing). The inequalities
0<; arg 17 < 2'IT / n imply that J1:; = A)ln; consequently, B = An' If T is non-
negative, then ~ >- 0 for all j; the condition 0", arg 'A)ln < 2'IT / n then
implies that ~I/n >- O. 0

If T is a compact operator from HI into H 2 , then T* T is compact,

self-adjoint, and non-negative. Hence we can define the absolute value of
T by the equality ITI = (T* T)1/2, where (T* T)I/2 is the uniquely de-
termined non-negative square root of T* T (a definition for arbitrary
densely defined and closed operators will be given in Section 7.3). ITI is
obviously compact. The term "absolute value" is justified by the following
theorem.

I Actually, the compactness of the nth root does not have to be assumed. Every norm~l nth
root An of a normal compact operator T is compact: If (An)n = T, then (A: An)n = 1'* T;
consequently, (A: An)n is compact. The compactness of A: An follows from this by Theorem
7.20 (for n =2 this follows from Theorem 6.4(c) because of the equality (A!A2)2 =
(A!A~·(A!A~). The compactness of An follows from Theorem 6.4(c).
170 7 The spectral theory of self-adjoint al~d normal operators

Theorem 7.5. Let T be compact. Then III Tlfll = II Tfll for all f E H. There
exists an isometric operator U from R(I TI) onto R( T) such that T = UI TI,
and ITI = U-IT. The representation T= UITI is called the polar decom-
position of T.
PROOF. For allfE H we have
IIITlfli z = (ITlf, ITlf) == (ITI1",J) = (T*IJ,f)
= (Tf, Tf) = II Tf1l2.
If for every f E H we set
V(I Tlf) = TJ,
then V is obviously a linear isometric mapping of R(I TI) onto R( T). Then
U= V is an isometry from R(ITI) onto R(T) (cf. the proof of Theorem
4.11), and we have T= UITI· 0

If T is a compact operator, then the non-zero eigenvalues of ITI are

called the singular numbers or singular values or s-numbers of T. In the
following let (~(T» denote the (possibly finite) non-increa.sing sequence of
the singular numbers of T; every number counted according to its multipl-
icity as an eigenvalue of ITI. For 0 <p < 00 we denote by B,(H I , Hz) the
set of all compact" operators T out of B(HI' Hz) for which
~ [siT)Y < 00.
j

We write B..:.(HI> H2) for the set of compact operators belonging to

B(HJ> H2) (cr. Section 6.1). For BP(H, H) we briefly write Bp(H).

Theorem 7.6. Let T E B..:.(Hh Hz), s) = siT). Then there exist orthonormal
sequences (h) from HI and (g) from H2 (these sequences can be finite) for
which
Tf = ~ sj(h,f)8j for all f E HI>
j

T*g = ~ ~(8j, g)h for all g E H2 ,

j

ITlf = ~ ~(h, f)h for all f E HI>

)

1T*lg ... ~ sl8j, g)8j for all g E Hz.

j

The elements f; and 8j are eigenelements of ITI and IT·I, respectively. In

particular, T, I TI, T* and IT·I have the same singular values, and the
following assertions are equivalent:
T E B,(HI' Hz), ITI E BiHI)' T* E S,(Hz, HI), IT·I E S,(Hz)'

PROOF. By Theorem 7.2 the compact operator ITI has a :representation of

the above form. Since ITI is non-negative, all the ~ are positive. Conse-
7.1 The spectral theorem for compact operators, the spaces Bp( H" Hz) 171

quently, with the operator V from Theorem 7.5 it follows for allf E HI that

Tf = VITlf = V( ~ S;<~,f)~) = ~ S;<~,f)Uij'

J J

Along with (1.J), the sequence (Vi.J) is also an orthonormal sequence;

therefore, we have the required representation of T. For all f E HI and
gE H2 we have

<i, T*g) = <Ti, g) = ( 7Sj<1.J,j)~, g)

= (i' ~:S/~, g)jj );

the required representation of T* follows from this. We can deduce from
these identities that

for all g E H2 • Consequently, we have the required representation of

IT*I r= (T** T*)1/2. The remaining assertions are now clear. 0

With the aid of Theorem 7.3 we have the opportunity of determining the
singular numbers.

Theorem 7.7. Let Sand T be from B.,.(H, HI)' Then

sl(T) - II TIl,
3;+ I(T) ... inf
8,,··· .8J EH
sup {II Tfll : f E H,f 1.g l , .•. ,g1' IIfll = I} (7.4)
for all j E N, and

SJ+k+I(S+T) <;sJ+I(S)+Sk+I(T) forall j,kENo' (7.5)

(7.6)

-!I(ST) ... siT*S*) <; IISlisi T ) ... IIS*lIsiT*). (7.7)

172 1 The spectral theory of self-adjoint and normal operators

PROOF. The formulae (7.4) follow from (7.3), since the ~(T) are the
. eigenvalues of ITI and since II Tfll = II ITlfll for all f E H. For S, T E
Boo(H, H,) and}, kENo
Sj+k+ ,(S + T)
= gil' ..
inf
,8j+k
sup {II(S + T)fll : f E H,f J..g, •... , gj+k' llill = I}
<:
8h'"
inf
,gj+k
sup {IISfll + IITfll : f E H,f J..g" ... , g)+k' IIfll = I}
<: inf [sup{IISfll:fEH,fJ..g" .•. ,gpllfll=I}
gl' ... 18j+k

+sup {II Till :fEH,fJ..gj +, •... , gj+k, IIfll = I}]

= Bit···
inf sup {IISfll :fEH.fJ..g,,···
18j
,~, Ilfll=I}
+ inf
8j+.' ... ,8j+k
sup {II Till : fEH,fJ..g j +" .. . , gj+k' IIfll = I}
= S j+'(S) + sk+,(T).
For T E Boo(H, H,), S E Boo(H" H2 ) and }. kENo we have (set
II Tfll-'IISTfll =0 if Tf=O)
Sj+k+,(ST) = inf sup {IiSTfll :fEH,iJ..g,,··· .gj+k' IIfll=I}
8 ..... ,8j+k EH

<: inf sup {IiSTfll :fEH,fJ..g,.··· ,gk' T*gk+"

8., ... ,8k EH

... , T*gk+P IIfll = I}

= inf sup { IIISI:llIII1I1Tfll :IEH,fJ..g,,···, gk:;

8 ..... ,8k EH .

Tf J..gk +" ... , gk+j; 11111 = I}

inf [sup {IIShll : hE H" hJ..gk+,,·· ., gk+j' IIhll = 1}
!:i"" ,8k EH
8k+ ..... ,8k+jEH.

x sup {II Tfll : f E H,I J..g" ... , gk' IIfll = l}]

= ~+,(S)sk+,(T).
If T E Boo(H, H,) and S E B(H" H2 ), then for allj E No
sj+,(ST) = inf sup {IISTIII : f EH,f J..g" ... , gj' IIfll = I}
8p •• . 8j EH

<: IISII inf sup {II TIll : ... } = IISlIsj+,(T).

8., ... ,8jEH

The remaining equalities follow from the equalities siA) = ~~(A *). D
7.1 The spectral theorem for compact operators. the spaces Bp(H., H2) 173

In what follows we write

IITllp = { L Is)T}IP riP

J
for T E Bp(H, HI) (0 <p < 00).
Theorem 7.8.
(a) If S, T E BiH, HI) (0 <p < 00), then S + T also belongs to Bp(H, HI),
and
IIS+ Tllp + IITllp) for p
~ 2 11p (1I S lip ~ 1,2
liS + TII~ ~ 2(IISII~ + II TII~) for p ~ 1.
The sets Bp(H, HI) are therefore vector spaces.
(b) If TEBp(H, HI), SEBq(H I, H2 ) (O<p,q<oo) and (l/r)=(I/p)+
(I/ q), then ST E Br(H, H2 ) and
IISTlir ~ 211rllSllqll Tllp·2
(c) If T E Bp(H, HI) and S E B(Hl' H2), then ST E Bp(H, H2), and we have
IISTll p ~ IISIIIITllp'
The corresponding assertions hold for T E B( H, HI) and S E Bp( H.. H2)'
REMARK. Theorem 7.S(a) and (c) imply that the sets Bp(H) are two-sided
ideals of B(H) and we have IISTll p ~ IISlllITllp and II TSli p ~ IISllliTilp
for S E B(H) and T E BiH).
PROOF.
(a) By virtue of (7.5) we have

L siS + TY = L {S2j_I(S + TY + S2j(S + TY}

j j

~ L {(Sj(S) + s)T)y + (Sj(S) + Sj+ I(T)Y}.

j

If p ~ I, then it follows by the Minkowski inequality for the ~-norm

that -

IIS+ TiI~ ~[(1SiSYfIP +(1S)TyfIPr

+ [( 1 s)Syr
lP
+ (~:Sj+I(TYrlpr
~ 2[ IISllp+ II TllpY·
2More accurate studies show that the II . lip are norms for p ;;. I (cf. Theorem 7.12 for p = I).
and that IISTI!." IISlIqllTllp for (l/p)+(I/q)= I; d. (5). Lemma XI. 9.14 or (10). Theorem
111.7.1.
174 7 The spectral theory of self-adjoint and normal operators

If p" I, then we can use the elementary inequality lal P + I/lIP >
la + I1I P (proof:
it is sufficient to prove the case lal + 1111 = 1). With the
aid of this inequality we obtain that

liS + Til; " ~ {2~(SY+sj(TY+sj+I(TY}

j

" 2~ (Sj(sY + ~(TY) = 2(IISII; + II Til;).

j

(b) As (r/p) + (r/q) = I, it follows from (7.6) with the aid of HOlder's
inequality that
1/,
IISTII, = ( ~ sj(ST)')
J

" {~SiS)'Si T)' + ~ siS)'Sj+ I( T)' } 1/,

." {(~Sj(str/q(~Sj(TYr/p +(~SiS)qr/q(7Sj+I(TYr/p}I/'

" 2 1/'II S lI qIlTli p'
Assertion (c) follows from (7.7). D

Theorem 7.9. Let p,q,r>O with (I/p)+ (l/q) ={I/r}. We have TE

B,{ H, HI) if and only if there exist operators TI E Bp( H, H2) and T2 E
BiH2' HI) (with an arbitrary Hilbert space H2) for which T= T 2T I; the
operators TI and T2 can be chosen such that II Til, = II Tilip II T211 q'
PROOF. By Theorem 7.8(b) we have T2TI E B,(H, HI) for TI E Bp(H, H2)
and T2 E Bq(H2' HI)' Now let T E B,(H, HI) and let (cf. 1beorem 7.6)

Tf = ~ ~(T)(Jj, f)gj, f E H,
j

where (Jj) and (~) are orthonormal sequences in H and HI> respectively. If
{hi' h2' ... } is an ONS in a Hilbert space H2 and we define TI and T2 by
the equalities

Td = ~ siTY/P(Jj,f)hj> f E H,
j

T2h = ~ siTy/q(hj, h)gj> hE H2,

j

then obviously T= T 2 T\. The numbers siT),/p and siTy/q are the singu-
lar numbers of TI and T 2, respectively. Therefore, T) E Bp(H, H2 ) and
7.1 The spectral theorem for compact operators, the spaces Bp(H1, H2) 175

t/P
II Ttllp = ( ~ s;(T)('/P)P ) = II TII'/P
J

consequently, II Til, = II Tilipil T 2 11 q • D

Theorem 7.10.
(a) The set ~(H, HI) coincides with the set of Hilbert-Schmidt operators.
For T E ~(H, HI) we have II Till = III Till·
(b) We have T E BI(H, HI) if and only if there exist operators TI E
B2( H, H2) and T2 E B2( H2, HI) (with an arbitrary Hilbert space H2) such
that T = T2 T I; the operators TI and T2 can be chosen such that II Till =
II T.lbll T 2 1b·
PROOF.
(a) Let T E ~(H, HI)' If fl' f 2, . .. are the orthonormalized eigenelements
of ITI that belong to the non-zero eigenvalues sj(T) and if {ga : a E
A} is an ONB of N(lT!) = N(T), then UI,J2' ... } U ega : a E A} is an
ONB of H. and we have

~ II Tijl12 + ~ II Tgall 2 = ~ III Tlijll2 = ~ Sj(T)2 = II TII~ < 00.

j a j j

Consequently. T is a Hilbert-Schmidt operator with III Till = II Tlb· If T

is a Hilbert-Schmidt (therefore compact) operator from H into HI and
fl' f2' . .. are chosen as above, then

~ Sj( T)2 = ~ II Tijll2 ..;; III TIW < 00.

j j

i.e., T E B2(H, HI)'

(b) This follows from Theorem 7.9 for r = I and p = q = 2. D
The set BI(H, HI) is also called the trace class of operators from H into
HI' This term originates from the fact that for T E Bt(H) a trace can be
defined by
tr (T) = ~ <ea , Tea>, (7.8)
a

where {ea : a E A} is an ONB of H. This is so, because then T = T2 TI with

appropriately chosen Hilbert-Schmidt operators T I , T2 and
1/2
~ I<ea• Tea> I = ~ I<T!ea, Tlea>l..;;
0: a
(
~ II T!ea ll 2 ~ II T lea ll 2)
a a
< 00.
For matrices the trace does not depend on the choice of the basis with
respect to which the matrix is determined. Analogously, the definition of
the trace does not depend on the choice of the ONB in the above case.
176 7 The spectral theory of self-adjoint and normal operators

Theorem 7.n.
(a) Definition (7.8) of the trace of an operator T E BI(H) is independent of
the choice of the orthonormal basis.
(b) If TI E BiHI> H2) and T2 E B2(H 2, HI) or TI E BI(H I , H2) and T2 E
B(H2' HI)' then tr (TI T 2 ) = tr (T2T I). (This also holds for TI E
Bp(HI' H2) and T 2 EBq (H 2, HI) with (l/p)+(I/q)=:I; cf.Exercise
7.6(c).)
PROOF.
(a) By Theorem 7.IO(b) there exist operators TI E B2(h', H2 ) and T2 E
Bz.(H2,H) such that T=T2 T I. If {e,,:aEA} and {fp:f3EB} are
arbitrary orthonormal bases of Hand H2 , respectively, then

=L Up, TITdp ), (7.9)

p

where all sums have at most count ably many summands and the sums
are absolutely convergent. If we choose another ONB {e~ : a E A} of
H, then it follows from this that

This is the required independence.

(b) The first assertion has been proved in part (a). If TI E: BI(HI> H2 ) and
T2 E B(H2' HI), then there exist operators BE B2(H I, H3) and A E
Bz.(H3 , H2) such that TI = AB. Thus it follows from (7.9) that

tr (T2T I) = tr «T2A)B) = tr (B(T2A»

= tr «BT2)A) = tr (A(BT2» = tr {T I T 2),
since T 2 A, B, BT2 , and A are Hilbert-Schmidt operators. o
Theorem 7.12. An operator T from HI into H2 such that D(T) = HI is in
BI(H I, H2) if and only if there are sequences (CPn) from HI and (tPm) from H2
such that IICPnll = IItPnll = I, and there is a sequence (zn) from II{ for which
~Iznl < 00 and
(7.10)
The norm II Till is the infimum of those sums ~Iznl for which there are
normed sequences (cpn)from HI and (tPn) from H2 such that (7.10) holds. II . III
is a norm on BI(H I , H2), the so-called trace norm.
7.1 The spectral theorem for compact operators, the spaces ~(Hh Hz) 177

BI (HI> H2), then

moreover, it follows the
lUU",1\U ... If (7.10) holds, the
finite operators Tm, T,J= Iz,,(rpn,J)1fn in the norm
of B(HI' H2 ) (because II T- Tmll ..; ~n>mlznD. Hence, we obtain with the
notation of Theorem 7.6 that

thus, T E BI(HI> H2)' As this holds for every representation of the form
(7.10), the equality given for II Till follows. The fact that II . III is a norm
follows immediately from this. 0

exactly

7.2. In (7.3) and (7.4) we can replace "inf sup" by "min max".
7.3. Theorems 7.1 and 7.2 do not hold for normal operators on real Hilbert
spaces. As an example, one can consider the operator induced by the matrix
(_~ ~) on H2; this operator has no eigenvalue.
normal operator on exists a Zo E p(T) T)
",Ullll"""'.then R(z, T) every Z EP(T),
seq,uellce (Aj) from K such asj--?oo and a sequence
projections that ~Pk = 8jk~'
< oo}, and D(T).

7.5. Let {A,.: aEA} be a family in IK\{O} for which J..,..,.Ap for a.,.p, and let
{Pc< : a E A} be a family of orthogonal projections on the Hilbert space H
such that PaPp = 8ap Pa for a, PEA.
(a) The equalities

D(T) = {fEH: ~ IAaI2I1Pa1112<OO} and

aEA

~ J..,.Pa1 for
aEA

define operator on H •
(b) Every .,. 0 is an eigenvalue N(J..,. - T) = R(Pa).
178 7 The spectral theory of self-adjoint and normal operators

(c) T is compact if and only if we have:

(i) dim R(P",) < 00 for all a E A.
(ii) for every t: > 0 there are only finitely many a E A such that I~I ;;. t:
and P",'f=O.
7.6. (a) Let 0 <p < 00. For every T E BiHI' H2) there exists a sequence (T,,) of
finite rank operators such that II T"lI p -< II Tllp and II T- T"lIp~O as n~
00.
Hint: Theorem 7.6.
(b) For every TE BI(H) we have Itr (T)I -< II TIll'
Hint: If Tf= ~l\}<Ii,j)$i. then define the trace of l' by means of an
OND that contains {fl. f2 • ... }.
(c) For TEB/H I • H2). SEBq(H2 • HI). where (l/p)+(I/q)=l. we have
tr (ST) - tr (TS).
Hint: Without loss of generality we may assume that q < 2. Let (T,,) be
chosen as in part (a). Then T" E B2(H I • H~. S E B2(H2• HI); conse-
quently. by Theorem 7.1l(b) we have tr (ST,,) = tr {T"S). Moreover.
Itr (ST) - tr (ST,,)I -< 211SIIqll T - T"lI p and Itr (TS) - tr (T"S)I -<
2I1SlIqIlT- T"lIp~O.
7.7. Prove Theorem 7.2 without reference to Theorem 6.7:
(a) There is an eigenvalue tJ.1 of T such that ItJ.11 == II Til.
Hint: By Theorem 5.17(c) and (d) and Theorem 5.43 there exist a tJ.1 E K
and a sequence (g,,) from H for which I tJ.11 == II Til. II g"II-I. and (tJ.I-
T)g,,~O. The sequence (g,,) has a convergent subsequence (g"k); the
element fl -lim g"k is an eigenvector of T belonging to the eigenvalue tJ.1'
(b) Let the eigenvalues tJ. ...... J.In (I tJ.11 ;;. ItJ.21 ;;. •.• ;;. I J.lnD and the eigenele-
mentsf..... •f" be determined. The restriction T" of T to LU.. ... • f,,)J.
is a normal operator on LUI' ... • f,,)J.. J.In+1 and f,,+1 are obtained by
using (a) for T".
(c) Prove (7.2).
7.8. Let T and S be operators on H such that p(T)np(S),,=0, and let 10.0 Ep(T)
n peS). 0 <p -< 00. If R(1o.o. T) - R(7\o. S) E ~(H). then R(A. T) - R(A. S) E
Bp(H) for all AEp(A)np(B).
Hint: We have
(A- TI)-I - (A- T 2)-1
- (AO- TI)(A- TI)-I[(1o.o- T1)-I-(>.o- T 2 )-IJO.o- T2 )(A- T2 )-I.

7.9. Give a proof. independent of Theorem 7.12. that II . III is a norm on

BI(H I • H~.
Hint: Let Af- ~s;<Jj. f)g;. Bf- ~S1'<Ii". f)g1'. and (A + B)f- ~s)
<Ii. f)g} as in Theorem 7.6. Then

IIA + BII! - ~«A +B)!J.g})

- ~ ~ <Ii. A*gD<gk' $i) + ~ ~ <h. B*gk><g;:. gl)
j k j k

-< IIAIII + IIBII 1•

7.1 The spectral theorem for compact operators. the spaces 8,,( HI> H 2) 179

7.10. (a) If Tn E B",,(HI • H2), TE B(H I, H2), and II Tn - TII~O as n~oo. then for
every kEN we have sk(Tn)~Sk(T) as n~oo.
Hint: Use (7.5).
(b) If Tn E ~(Hl> H2). T E B(H I • H2). IIT- Tnll~O as n~oo, and
lim inf n ..... oo 1/ Tnl/p < 00, then T E Bp(HI' H 2 ), and II Tllp .;;
lim infn..... oo II Tnllr
(c) If (Tn) is a sequence from Bp(HI' H2) such that 1/ Tn - Tml/p~O as n, m~
00, then there exists aTE Bp(HI' H~ such that 1/ T- Tnl/p~O as n~oo.
(Consequently, BI(H I , H2 ) is a Banach space.)
7.11. Let S, T E B(H I , H2) be bijective. Then S - T is in Bp(HI' H2) if and only if
S-I- T- I is in Bp(H2, HI)'
Hint: S-I- T- I == S -I(T_ S)T- I .
7.12. Let S, TE B(H), S- TE BP(H).
(a) We have S" - Tn E 8p(H) for all n E N.
Hint: S"- T"="ij:JTj(S- T)S"-j-I.
(b) We have peS) - peT) E BP(H) for every polynomial p.
7.13. If TEBp(H.,H2) for somep<oo, then T*TEBp/2(H I) and (T*T)"E
~/2iH)c BI(H) for n >p/2. We have II Til =lim" ..... oo [tr (T·T)"]1/2n.
Hint: {"ij_dajl"}I/"~max {laA :j= 1, ... , k} as n~oo.
7.14. Let HI and H2 be Hilbert spaces, and let T E B",,(HI , H2). For every (> 0
there exists a finite-dimensional subspace M. of HI such that 1/ TJII .;; (I/JII for
allJE M.l.. .
Hint: Use the representation in Theorem 7.6.
7.15. Let HI and H2 be Hilbert spaces. Let Hi = B(H I , IK) be the Hilbert space of
continuous linear functionals on HI (cf. Exercise 4.3(a».
(a) HI ® H2 is isomorphic to the space of bounded finite rank operators from
HI into H2; we can make the element "ij_ICj~®~ from H,®H2 corre-
spond to the operator T for which D(T) = HI and Tf= "ij.lcjLJ(f)~.
(b) We havell"ij.lcjLj®gjll""IITlh; the space Hl~H2 can be identified with
the space ~(HI' H2) of Hilbert-Schmidt operators from HI into H2.
(c) A norm II . liT on H{®H2 is called a "cross"-norm if IIL®gIlT-
IIL1I1I811. The completion H;~TH2 of H;®Hl with respect to 11.11. can
be identified with a subspace B.(HI , H2) of B",,(H I , H2). The spaces
B.(HI , H2) are two-sided ideals of B(HI, H2 ).
(d) For every "cross"-norm II . liT we have BI(HI• H~ C B.(H I • H2). Further-
more. B.(HI' H2) - BI(HI• H2), if we choose
180 7 The spectral theory of sell-adjoint and normal operators

Integration with respect a spectral family

1) A Hi/bert space H function E: R_B(H)
Ihe following nrnnprllp.<
(a) E(t) is an orthogonal projection for every IE R,
(b) E(s) <; E(/) for s "t (monotonicity),
E(I ~ E(t) all as + (continuity the
(d) E(/)_O as 1_ - 00, E(/) _ I as 1_00.
REMARK. Property (7,1l(c» is not essential, it is needed only m the
uniqueness statement of Theorem 7.17, Right continuity be replaced
by left continuity; then we have to replace limB->o+ by lima...o_ in Theorem
7.17.

EXAMPLE Let M measurable of and let M_R be a

measurable function. For every t E R let
M{/)
M(t) is obviously a measurable subset of M.
{XE
"
The equality XM{I)f Jor f L 2(M) ER a
on
PROOF. Properties (7.11(a» and (7.1l(b» are evidently satisfied. We show
l(c»: E R, E and (En) be null-sequence of nn"iti.rp
numbers. Then
II(E(t + £,,) - E(/»JII2 = f. (XM(t+<,,)(x) - XM(/)(x»If(xW dx.
M
Because the relations 0 XM(I +<,,)(x) XM(t)(x) " 1 XM(/+<,,)(X)-
XM(t)(x)_O as n_ 00 for all x E M, it follows by Lebesgue's theorem that
+ £,,) E(t»J as 00.

As this holds for every zero-sequence (£,,), property (7.11{c» follows. We

can show (7.11(d» similarly, since we have
XM(/)(X) - 1 all M, as 00.

XM(t)(x) - 0 for all x EM, as t - - 00. 0

EXAMPLE Let {Pa E A} family of right continuous non-decreas-
ing functions defined on 1Ft

(7.13) The equalities

E(t)(fa} X(- IlUa)
= (X( - 00, td.. ) for (fa) E $
.. EA
L2(R, Pa), I E R

define spectral family on EDaEA P.. ).

7.2 Integration with respect to a spectral family 181

PROOF. It is clear that the E(t) are orthogonal projections. The increasing
character follows from the equality

f (s, I)
If.. {xW dp,,{x) ~ 0 for all a E A.

Because

1(s, I)
/f.. (xW dp,,(x) .;; 1/f,,(xW dp,,(x)
R
for all s, t and all a

and

it follows from this that

II(E(t) - E(s»(!,.)11 2 = L
a
1
(s,l)
11a(x)1 2 dp,,(x) ~ 0, t ~ s+,

i.e., E is right continuous. The remaining assertions E(t)~l as t~ + 00

and E(t)-+O as t~ - 00 are clear. 0

EXAMPLE 3. Let (iy) be a sequence of pairwise different real numbers, and

let (P) be a sequence of orthogonal projections on H such that ljPk =0 for
j=l=k and ffijR(lj) = H.

(7.14) The equalities

E{t)f = L ljf for f E Hand t E ~
{J: Aj';l}

define a spectral family on H.

PROOF. Properties (7. Il(a», (7.11(b» and (7.II(d» are clear; we leave their
proof to the reader, and only prove the right continuity here. For every
tE~ and t:>0

\I(E(t + t:) - E(t»fIl2 = ~ Illjfll 2•

U: I<Aj';I+<)

The sum converges to zero as t:~O+ (since the series ~)lljfIl2 is conver-
gent and for every noE N there exists an t:>0 such that }y~(t,
t+e]forj<.n o)· 0
Let E be a spectral family on the Hilbert space H. For every f E H
define
pit) = <f, E(t)f> = IIE(t)fI12, t E~. (7.15)
182 7 The spectral theory of self-adjoint and normal operators

The function PI : R~R is obviously bounded, non-decreasing, and right

continuous; limt ->_ 00 pit) = 0, lim t .... oo pit) = Ilfll2.
A function u : R~K is said to be E-measurable if it is PI -measurable for
every f E H (cf. Appendix A). Non-trivial examples of E-measurable func-
tions (for every spectral family E) are all continuous functions, all step
functions, and all functions that are pointwise limits of step functions; all
Borel measurable functions are E-measurable. The function U o with uo(t) =
1 for all t E R is in L2(R, Pj) for every f E H. Consequently, by Theorem
A 14, every bounded E-measurable function u : R~K belongs to L2(R, PI)
for allfE H.
For a step function u we can define the integral fu(t) dE(t) by the
equality

f L cjxAt) dE(t) = L cyE(0),

n n

)=1 J j=1

where
E(a, b]) = E(b) - E(a), E«a, b» = E(b -) - 1\(a),
E([a,b]) = E(b)- E(a-), E([a,b)) = E(b-)-- E(a-)
(here let E(t - ) = s -lim......o+ E(t - t:); this limit exists by Theorem 4.32
and is an orthogonal projection). For any step function U : R~K and for
every f E H we obviously have

Ilf u{t) dE{t)~1

2

= fl u (t)1 2 dpjt), (7.16)

as a simple calculation shows. If f E Hand U E LlR, Pj)' then by Section

2.2, Examples 7 and 13, there exists a sequence (un) of step functions for
which un~u in L2(R, PI).3 Then

Ilf ull{t) dE(t)! - f um{t) dE(t)~1

2

= f IUII{t) - U m(t)12 dpjt)

~ () as n, m ~ 00,

i.e., the sequence (fun dE(t)f) is a Cauchy sequence in H. Therefore, we

can make the definition

lim fUII(t) dE{t)f·

f u(t) dE(t)f = n->oo (7.17)

This definition is obviously independent of the choice of the sequence (un)'

and we have

II! u(t) dE(t)~1

2
= !lu{tW dpjt). (7.18)

] Here and in the sequel Ll(R, Pf} is meant to be the real or the complex Lrspace according
as H is real or complex. .
7.2 Integration with respect to a spectral family 183

For u, v E L2(R, PI) and a, bEll( it follows from (7.17) immediately that

f (au(t) + OO(t» dE(t)f af u(t) dE(t)f + bf vet) dE(t)f· (7.19)

=

The integral just defined is therefore linear. 0

In our further studies we shall use the following auxiliary theorem; for
functions that are pointwise limits of step functions (all functions explicitly
occurring in the following are of this kind) this auxiliary theorem is not
needed.

Auxiliary Theorem 7.13. Let E be a spectral family on H and let u : R~IK

be an E-measurable function. If {f" ... ,,,fp} is a finite set in H, then there
exists a sequence (un) of step junctions that converges to u almost everywhere
with respect to Ph for j = I, ... ,p. If u is bounded, then the sequence can be
chosen to be bounded.
PROOF. It is enough to show that there exists an h E H for which every
Ph-null set is also a PJi-null set forj= 1, ... ,p (then we choose a sequence
of step functions th~t converges to u Ph-almost everywhere). In order to
prove this, it is enough to find, for any two elements f, = f and f2 = g, an
element h E H for which every set of Ph -measure zero is of PI - and
Pg-measure zero; the rest is simple induction.
For this, set M = L{ E( t)f: t E R}. Let us introduce th~ notations: P is
the orthogonal projection onto M, g, = Pg, g2 = (I - P)g = g - g" and
h = f + g2' Then for arbitrary intervals II and '2
in R we obviously have
E(1I)f1. E(l2)g2 and E(lI)g, 1. E(l2)g2 (since E(lI)f EM, E(1,)gl EM, and
E(12)g2 E M.l).
Let N be a set of Ph-measure zero. Then there exists a sequence (8n ) for
which N C 8m and 8n = U m Jnm , where the Jnm are at most countably many
mutually disjoint intervals for fixed n E N, and

~ Ph(Jnm ) ~ 0 as n~ 00.
m

Because of the equalities

for all n, m, it follows from this that

~ PiJnm) ~ 0
m
and Lm Pg,(Jnm ) ~ 0 as n~ 00.

Consequently, N is a set of Prmeasure zero.

184 7 The spectral theory of self-adjoint and normal operators

For all t E IR we have

II ~ E(Jnm)E(t)~r = IIE(t) ~ E(Jnm)~12 < II ~ E(Jnm)~(

=LIIE(Jnm)!1I2=LPiJnm)~0 as n~oo.
m m

Since the set {E(t)f: t E IR} is total in M and the norms of the operators
k~};mE(Jnm)k are less than or equal to 1, Theorem 4.23 implies that
ImE(Jnm)k~O for all k E M. As gl E M, it follows from th:IS that

~ Pg,(Jnm ) = ~ IIE(Jnm )gtlI 2 = II ~ E(Jnm)gl\r -~ 0,

consequently,

as n~oo. Therefore, the Pg-measure of N equals zero. o

For f, g E H and a bounded E-measurable function u : IR~IK, according
to the polarization identities (1.4) and (1.8), respectively, we define
fu(t) d<g, E(t)f)
= ~ {fU(/) dpg+j(/) - fu(t) dpg-it) + ifu(/) dPg_ij(/)
-if u( I) dpg+ii t)} if IK = C,
= {{fu(t) dpg+it)- fu(t) dpg_it)} if IK = IR.

With this definition we obtain for any bounded E-measurable functions u

and v that
fv(t)*u(t) d<g, E(t)f) = <fv(t) dE(t)g, fu(t) dE(t)!). (7.20)
For step functions u and v this is evident. In the general (;ase this follows
by Lebesgue's theorem if, according to Auxiliary theorem 7.13, we choose
bounded sequences (un) and (vn ) of step functions that converge
Pj-' pg-, P;+!' Pg-_j' Pg-+ij' Pg----ij almost everywhere to u and v, respectively.
Now we are in a position to generate linear operators by means of
integrals with respect to a spectral family.

Theorem 7.14. Let E be a spectral family on the Hilbert space H, and let
u: IR~IK be an E-measurable function. Then the formulae
D(E(u» = {JEH: UEL 2 (R Pj}}
£(u)f = fu(t) dE(t)f for f E D(£(u») (7.21)
define a normal operator £(u) on H. For (7.21) we briefly write
feu) = fu(t) dE(t).
7.2 Integration with respect to a spectral family 185

If u, V : IR~IK are arbitrary E-measurable functions, a, bE IK, and

cp,,{x) = {I if lu(x)1 ""n, I/;,,(x) = {I if Iv{x)1 <n,

o otherwise, 0 otherwise,
then it follows:
(a) For all f E D(£(u» and g E D(£(v» we have

(£(v)g, £(u)f) = lim Nn(t)v(t)*cp,,(t)u(t) d(g, E(t)!);

"--+00.

for the latter we briefly write fv(t)*u(t) d(g, E(t)f).

(b) For all f E D(£(u» we have

1I£(u)fIl2 = Jlu(tW dpit).

(c) If u is bounded, then feu) E B(H) and

1I£(u)1I < sup{lu(t)1 : t E IR}.

(d) If u(t) = I for all t E IR, then feu) = I.
(e) For every f E D(£(u» and all g E H we have

(g, £(u)f) = fU(/) d(g, E(/)f).

(f) If u(t) ;> c for all t E IR, then

(f, £(u)f) ;> cllfll2 for all f E O(£(u».

(g) Beau + bV):J a£(u) + bE(v), D(£(u) + E(v» = D(£(jul +'Ivl».

(h) ~(uv):J ~(u)£(v), P(£(u)£(o») = D(£(v» n D(£(uv».
(i) E(u*) = E(u)*, D(E(u*» = D(E(u».
PROOF. The mapping £(u): D(£(u»~H is well-defined, because of
(7. 1])-(7.2 I). We show that this mapping is lin!ar. It is cl~ar that f E
D(E(u» and a E K imply that af E D(E(u» and E(u)(a1) = aE(u)f.
First assume that u is bounded. Then D(£(u» = H. By Auxiliary theo-
rem 7.13, for arbitrary f, gE H there is a bounded sequence (u,,) of step
functions that converges to u almost everywhere with respect to Pi' PI and
Pi+1f Then~un~u in~ L2(1R, Pi)' L~(IR, Pg) and L2~(IR, Pi+g); therefore~ E(u,,)f
~E(u)f, E(un)g~E(u)g and E(u,,)(f + g)--.E(u)(f+ g). Since E(u,,) is
obviously linear, it follows that

B(u)(f + g) = lim E(u,,)(J + g)

"->00

= lim (E(u,,)! + B(u,,)g) = £(u)f + £(u)g.

"-->00

Now let u be an arbitrary E-measurable function, and let f, g E D(B(u».

Then (IIP,,(t)U(/)j) converges to lu(t)1 monotonically for all t E IR. Using the
186 7 The spectral theory of self-adjoint and normal operators

identity we have just proved and on the basis of (7.18) we obtain that

{J1'P,,(t)u(t)1 2 dpj+g(t)} 1/2

= 1I£(IPn u)(f+g)1I
« 1I£('P"u)fll + 1I£('P"u)gll
= (JIIP,,(t)u(t)12 dpj(t)} 1/2 + (JIIP,,(t)u(tW d,Pg(t)} 1/2
« (Jlu(t)12 dpit)} 1/2 + (Jlu(tW dpg(t)} 1/2
= 1I£(u)fll + 1I£(u)gll < 00.

It follows from thls by B. Levi's theorem that u E LllFt, Pj+g); conse-

quently, f + g E D(E(u» and
£(u)(f+g) = lim £('P"u)(f+g)
" ..... 00

= lim (£(IP"u)f+£('P"u)g) = E(u)f+ £(u)g.

" .... 00

So B(u) is linear. We obtain from the proof of (i) that B(u) is normal.
(a) This equality is clear for step functions u and v. For bounded E-
measurable functions u and v the equality follows by means of Aux-
iliary theorem 7.13. In both cases the passage to the limit does not
actually take place, as '11" = til" = I for large n. If u and v are arbitrary
E-measurable functions, then we have for f E D(B(u» and g E D(B(v»
that
<B(v)g, B(u)f) = lim <B(tiI"v)g, E(IP"u)f)
" ..... 00

= J~~ JtiI,,(t)v(t)*'P,,(t)u(t) d< g, E(t)f).

(b) follows from (7.18) with v = u, g =f.
(c) Since u is bounded, we have u E L2(1R, Pj) for all f E hi. Consequently,
D(B(u»'" H. The estimate of the norm immediately follows from (b).
(d) By (c) we have B(u) E B(H). Furthermore, X<_II.III(t)-.u(t) - 1 for all
t E IR; therefore,
B(u)f ... n-+oo lim (E(n)f - E( - n)f) - f.
lim B(X<_" ' "I)f'" n-+oo
(e) follows from (a) with v'" 1, by taking (d) into account.
(f) immediately follows from (e).
(g)4 Iff E D(B(u) + B(v»..., D(B(u» n D(B(v», then u, v E L2(1R, PI); conse-
'luently. au + b£ E L2(1R, PI). i.e., f E D(B(au + bv». The ~quality
E(au + bv)f- aE(u)f + bE(v)f therefore follows from (7.19). Smce for
E-measurable functions u, v we have u, v E L2,(IR, Pf) if and only if
lui + Ivl E L2(1R, PI)' we have D(E(u) + E(v» - D(B(lul + Ivl».

4 Properties (g), (h). and (i) follow more easily from Theorem 7.16.
7.2 Integration with respect to a spectral family 187

(h) By (a) and (d) we have for all bounded E-measurable functions cp, If;
and all f, E H that
(g, E(cp)f) = (E(l)g, E(cp)f) = (E(cp*)g, EO)!) = (E(cp*)g,f);
conseq uen tly,

(g, E(cp)E(If;)f) = (E(cp*)g, E(If;)f) f

= cp(t)If;(t) d(g, E(t)f)
= (g, E(W)!)'
For bounded E-measurable If; therefore have E(cp)E(If;)
= E(W). " " A A A

Let fE D(E(u)E(v», i.e., let f E D(E(v» and E(v)f E D(E(u». As

function cp"U is bounded for fixed n E it follows cp"ulj;mv--+
cp"uv in L2(IR, Pj) as m--+oo. Consequently,
E(u)E(v)f = n-+oo
E(cp"u){ limm~oo

= lim E(cp!lu)E(If;mv)f
1'1-+00

The existence this means the sequence (cp"uv) is a Cauchy

sequence in L2(1R, Pj)' Since, moreover, cp,,(t)u(t)v(t)--+u(t)v(t) for all
~, follo~s th,!it uv ~elongs to Pj); '" con~equently, J E
D(E(uv» and E(u)E(v)f= E(uv)f. Therefore, D(E(u)E(v» c D(E(v»
n D(E(uv» and E(u)E(v) C E(uv). Iff E D(E(v»n D(E(uv», then
E(uv)f- lim n~oo
lim E(cp"uIf;mv)]
m~oo
= n-+oo
lim lim E(cp"u)E(If;mv)f
m~oo

= E(cp" u) E(v)f.
1he existen£e of this limit means" tha~ u E L2(1R, Pi(I))j); consequently,
E(v)f D(E(u», and thus f D(E(u)E(v».
(i) We first show that D(E(u» is dense. For this we prove that for every
fE H and every mEN we have E(CPm)fE D(E(u»; because of the limit
relation]- lim E(CPm)fit follows from this that D(E(u» dense. Let
m-tooo A
f E H, mEN and g = E( CPm)!' Then by (h) we have for all n ;> m that

!lcp/l(t)u(t)j2 dpg(t) = IIE(cp"u)E(CPm)!1I 1 ... IIE(CPmu)fIl2 < 00.

Therefore, E L2(R, Pg)' i.~, E(CPm)f- JI E D(E(u» all mEN.

We obviously have D(E(u*» = D(E(u». By (e) we have for f, g E
D(E(u» == D(E(u*» that

(g, E(u)f) = f u(t) d(g, E(t)f) = {f u(t)* d(f, E(t)g)}*

= (J, E(u*)g)* = <E(u*)g,j),
188 7 The spectral theory of self-adjoint and normal operators

i.e., E(u) and E(u*) are formal adjoints of each other. It remains to
prove that D(E(u)*) c D(E(u*». Let g E D(E(u)*). Then for all f E
D(E(u»
(E(u)*g,f) = (g, E(u)f) = lim (g, E(fPnu)f)
n---+oo
= n-+oo
lim (E(fPnu*)g,f).

In particular, for every f E Hand mEN

(E(fPm)E(u)*g,f) = (E(u)*g, E(fPm)J)
= n-+oo
lim (E( fPnu*)g, E( fPm)J) = n--+oo
lim (E( fPmfPnu*)g, f)
= <E(fPmu*)g,f).

Consequently, for all g E D(E(u)*) and all mEN

E(fPm)E(u)*g = E(fPmu*)g,
and thus

E(u)*g = m--+oc
lim E(fPmu*)g = mlim
....... oo
!<Pm(t)u(t)" dE(t)g.

The existence of this limit means that the sequence (fPmu*) converges in
L 2(R, Pg)' i.e., u* E L 2(R, Pg), llnd thus g ~ D(E(u*». ..
We have in particular D(E(u)*) = D(E(u», and (by (b» IIE(u)fll =
IIE(u)*fll for alifE D(E(u», i.e., E(u) is normal. 0

EXAMPLE 1 (continued). Let E be the spectral family of Example 1. If

Lj_l CjX~(t), then
u : R~C is a step function, u(t) =

(f u(t) dE(t)f)(x) = .± cJx~(g(x»f(x)

J=l

= u(g(x»)J(x), f E L2(M).
Therefore, it follows for every E-measurable function u that
D(E(u» = {J E L2 (M) : (u 0 g)f E L2(M)}
and
(E(u)J)(x) = u(g(x»f(x) for f E D(E(u»,
i.e., E(u) is the maximal operator of multiplication by u 0 g. For u = id we
obtain the operator of multiplication by g.

EXAMPLE 2 (continued). If E is the spectral family of Example 2 on

EEl aEA L2(R, Pa) and u : R~C is an E-measurable function (cf. Exercise
7.18), then

D(.E(u» = {(fa) E EEl L2 (R, Pa) : (uJa) E EEl L2(R, Pa)}

aEA aEA
and
E(u)(fa) = (uJrJ f~r f E D(E(u».
7.2 Integration with respect to a spectral family 189

We call this operator the maximal multiplication operator induced by u on

EB a EA L2(1R, Pa ). The proof is along the same lines as in Example I.

EXAMPLE 3 (continued). Let E be the spectral family of Example 3. Then

every function u: IR~IK is E-measurable, since for all tE{A,- :j=
1, 2, ... }

as R\{Aj :j= I, 2, ... } is a set of Prmeasure zero for allfEH, we have

found a sequence of step functions that converges to u Pralmost every-
where for all f E H. It is now easy to see that for all u : 1R41K

D(E(u» = {f E H: ~ IU(A)I ZllljfIl 2 < 00 },

J

i(u)J = ~ u(A,-)PjJ for J E D(E(u».

j

For u = id we obtain in particular that

and
i(id)J = ~ AjljJ for J E D(i(id».
j

Theorem 7.15. Let HI and H2 be Hilbert spaces, let U be a unitary operator

Jrom HI onto Hz, and let E be a spectral family on HI' Then by the formula
F(t) = UE(t)U- I , t E III
a spectral Jamily is deJined on Hz. A Junction u : 1Il~1K is F-measurable if
and only if it is E-measurable. If ft is defined analogously to i, and u is
E-measurable, then

PROOF. It is clear that F is a spectral family on H2 • If pfit) = IIE(t)JII 2 and

O'it) = II F(t)gIl2, then pIt) = O'ult) obviously holds. Consequently, u is
E-measurable if and only if it is F-measurable, and L2(1R, Pj) = L2(R, O'Uj)'
The equality F(u) = UE(u)U- 1 is evident for any step function u. The
assertion follows from this fact immediately. 0
Theorem 7.16. Let E be a spectral family on the Hilbert space H. Then there
exists a Jamily {Pa : a E A} of right continuous non-decreasing Junctions (the
cardinality of A is at most the dimension of H) and a unitary operator
190 7 The spectral theory of self-adjoint and normal operators

V: H~e"EAL2(R, p,,)jor which

E(t) = V-1F(t)V for a/l t E R
with the spectral family F from Example 2. For every E-measurable function
u
E(u) = V-1F(u)V,
where F(u) is the maximal operator of multiplication by u on
ffiaEALiR, Pa)'
PROOF. For any fE H, f=l=O let H.J=L{E(t)j: t ER} and let pJ.t)=
IIE(I)fI1 2• Then the formula

vj • O( ~
j-I
Cj E(0)J) = ~ cjX(-oo.t,l
j=1

defines an isometric mapping of L{ E(t)f : t E R} into L2(R, Pj)' as can be

verified easily. For all gEL{ E(t)f: t E IR}
Vj,o(E(t)g) = X(-oo.I)u"og·

The range of u"o contains the space of left continuous step functions, thus
it is dense in L2(R, Pj) (observe that the left continuous ~ep functions are
dense in the space of step functions). The closure 0= ~1.0 is therefore a
unitary operator from H.r onto L2(R, Pj)' and for all g E Ii-
Vj(E( t)g) = X( - 00, II ll.rg.

With the aid of Zorn's lemma we see immediately that there exists a
maximal system {H.t. : IX E A} such that H.ra.i H.r" for IX =F f3 (partial ordering
= inclusion, upper bound = union). We write Ha for Hr., and show that
H= eaEAHa • If we had H=I= ffiaEAHa , then there would be agE H, g=l=O
such that g.i H" for all IX E A. Then we would also have E( t)g..L Ha for all
IX E A, and thus Hg.i eaEAHa ; this contradicts maximality.
Let Pa = PIa' let Va : H" ~ L2(R, p,,) be the corresponding unitary opera-
tors, and let P" be the orthogonal projections onto H", then

is a unitary operator from H onto ffi aEA L2(1R, p,,). For all g E H we have

VE(t)g = (VaPaE(t)g) = (V"E(t)Pag)

= (X(-OO,I)V"Pag) = F(t)Vg.

E(u) = V-1F(u)V follows by Theorem 7.13. The rest of the assertion

follows from Example 2. 0
7.3 The spectral theorem for self-adjoint operators 191

EXERCISES

7.16. Let E be a spectral family, and let u : R~D< be a continuous function. For
- 00 < a < b < 00 we can define the integral f ~ u(t) dE(t) as a Riemann-
StieJtjes integral, i.e., the integral is the limit in B(H) of the sums
n
~ u(l;) =F (E(I;)-E(I;_I»
)-1
with a = 10 < II < . . . < In = b, provided that the maximal length of the
intervals tends to O.

7.17. Let E be a spectral family on H.

(a) For every sequence (!,.) from H there is an hE H for which we have:
Every set of Ph-measure zero is of p~-measure zero for allj E N.
(b) If H is separable, then there exists an hE H for which we have: Every set
of Ph-measure zero is of Prmeasure zero for all f E H.
(c) If H is separable, and u : R~IK is an E-measurable function, then there
exists a sequence (un) of step functions that converges to u Pralmost
everywhere for all f E H.
7.18. Let E be the spectral family of Example 2. A function u : R~C is E-measur-
able if and only if it is p,,-measurable for every a E A.

7.3 The spectral theorem for self-adjoint operators

If u is a real-valued E-measurable function on R, then the operator E(u) is
self-adjoint by Theorem 7.14(i). We show in this section that every self-
adjoint operator can be represented in this way and there exists exactly one
such representation with u = id.
Theorem 7.17 (Spectral theorem). For every selj-atijoint operator T on the
Hilbert space H there exists exactly one spectral family E for which T=
E(id), or in another notation, T= Jt dE(t) (cf. Theorem 7.14). In the com-
plex case the spectral family E is given by
(g, (E(b)-E(a»f> = lim lim -2.
I fb+8 (g, (R(t-iE, T)
8 .... 0+ <....0+ 'IT I 0+6
- R(t + iE, T»f> dt . (7.22)
for all f, g E Hand - 00 < a . ;;;; b < 00. We stry that E is the spectral family of
T.
PROOF. First we assume that H is complex.
Uniqueness: If T= E(id), then z - T= E(z - id) by Theorem 7.14(g).
Then for all z E C such that 1m z =1=0 we have by Theorem 7.14(h), with the
notation uz(t) = (z - I)-I, that
(z - T)E( uz)f = E«z - id)uz)f = f for all f E H,
E(uz)(z - T)J = E(uAz - id»)J = f for all f E D(T).
192 7 The spectral theory of self-adjoint and normal operators

. Consequently, E(uz ) = R(z, T) for all z E C such that 1m z .pO. This im-
plies via Theorem 7.14(e) that

(f, R(z, T)f> f

= (z - I) -I dPj(/) for all f e:: H.
The functions PIt) = II E(/)fIl2 and F(z) = (f, R(z, T)f> therefore satisfy
the assumption of Theorem B I of the Appendix, and thus for all t E R

IIE(t)fIl2 = (f, E(t)f> = lim lim --I fl+8 Im(f, R(s + if, T)f> ds
8--+0+ .--+0+ 'TT - 00

= lim lim 21. f'+B(f, (R(s-if, T)-R(s+if, T»f> ds.

8 ....0+ .....0+ 'TTl - 00

(7.22) follows from this with the aid of the polarization formula (1.4). Since
(7.22) holds for all f, g E H, the uniqueness has been proven.
Existence: If there exists a spectral family E such that T = E(id), then
(7.22) must hold. Therefore we study whether (7.22) defines a spectral
family E with the property E(id) = T. -For every f E H Ithe function Fj
defined by the equality Flz) = (f, R(z, T)f> satisfies the assumptions of
Theorem B3 (Appendix), since Fj is holomorphic for 1m z > 0 by Theorem
5.16 and we have
1m Fj(z) = Im(f, R(z, T)f> ... Im«z - T)R(z, T)f, R(z, T)f>
= IIR(z, T)f112 1m z* < 0 for 1m z > 0
and (by Theorem 5.18)
IFj(z) 1m zl <; 11m zl-llIfll 21lm zl = IIf1l2.
Consequently,
<1, R(z, T)f> = Flz) = f (z - 1)-1 dw(j, t), (7.23)
where
w(j, I)
-
= 11m .
11m 1 fl+8
-2' - .
(f, (R(s - I f , T) - R(s + If, T»f> ds.
8..... 0+ ......0+ 'TT I -00

w(j, t) is a non-decreasing and right continuous function of t, and w(j, I)

--+0 as t--+ - 00, w(j, t) <; IIfll2 for all t E R. Equation (7.23) holds for all
z E C\R since <1, R(z*, T)f> = (f, R(z, T)f>*. Furthermore, we define

w(g,j,t) = lim 1
lim -2' . .
f'+8 (g,(R(s-u,T)-R(s+lf,T»f>ds;
8-+0+ ...... 0+ 'TT 1 - 00

the existence of this limit follows by means of the polarization identity for
the sesquilinear form (g, j)"""'( g, (R(s - if, T) - R(s + if, T»f>.
The mapping (g,j)....."w(g,f, I) is a bounded non-negative sesquilinear
form on H for every t E R. The sesquilinearity is clear from the definition;
moreover, w(j, f, t) = w(j, t) > 0 for all 1 E R. The Schwarz inequality and
7.3 The spectral theorem for self-adjoint operators 193

the inequality w(f, I) " 111112 imply for all 1, g E Hand t E R that
Iw(g,l, tW " w(g, l)w(J, I) " IIg1l2111112.
Therefore, by Theorem 5.35 there exists, for every t E IR, an operator
E(t) E 8(H) for which IIE(I)II " I and
(g, E(t)l> = w(g,l, t) for all 1, g E H.
It is obvious that E(t) is self-adjoint and E(/) > O.
Now we show that E is a spectral family. For this we first show that
E(s)E(t) = E(min(s, I» for all s, 1 E R. For all z E C\R and for all 1 E H

(g, R(z, T)J> = f (z - 1)-1 dw(g,l, I) = ! (z - 1)-1 d(g, E(t)I>.

(7.24)
This follows from (7.23) using the polarization identity. Consequently, the
first resolvent identity implies for all z, z' E C \ R with z =l=z'

! (z - 1)-1 d(R(z'·, T)g, E(t)l) = (R(z'·, T)g, R(z, T)J)

= (g, R(z', T)R(z, T)J)

= (z'-z)-t{(g, R(z, T)I> -(g, R(z', T)I)}

= (Z'-Z)-I![(Z-/)-I_(Z'-t)-I] d(g, E(t)l)

= !(z-t)-I(z'-t)-I d(g,E(t)f)

= ! (z - 1)-1 d,f~}z' - S)-I d(g, E(s)I).

It follows from this by Theorem B2 (Appendix) that

!(Z'-S)-I ds(g, E(s)E(t)l) = (g, R(z', T)E(/)I>

= (R(z'·, T)g, E(t)l)

= f'
-co
(z' - s) - I d(g, E(s)I>.

Therefore it follows for all 1, g E Hand s, 1 E R again by Theorem

B2 (Appendix) that

for s "t,
(g, E(s)E(/)I) = {(g, E(s)1>
(g, E(t)1) for 1 "S.
This means that for all s, 1E R
E(s)E(/) = E(min(s, t».
194 7 The spectral theory of self-adjoint and nonnal operators

E(ti E(t). E( t) orthogonal projections

for all t E IR, and E(s) "- E(t) for s "- t (cf. Theorems 4.29 and 4.31). Thus
(7.1 I(a» and (7. 11 (b») are satisfied. The right continuity (7.11(c» follows
the formula

t + () w(J, o as ...... 0

since w(j, . ) is right continuous. Moreover, IIE(I)111 2 = wU, t) ...... O as t ......

Therefore, ~0 t ...... - It only remains prove that
~I 00. E( .) monotone, E(/) strongly wnverges an
orthogonal projection E(oo) as t ...... oo (cf. Theorem 4.32). We have

<j, E( 00 )1) = <1, E(t)j>.

Consequently, E( 00) > E(t) for all t E IR. Let F- 1- E( 00), Then

- E(t) = 0, R

It follows from this for all j, g E H, 1m z =F 0 that

g, R(z, T)Fj) J(z - d<g, E(t)Fj)

Hence R(z, T)Fj= for all H, and thus F= i.e., = I. Conse-

have that is a spectral
R(z, T) = E(uz ) by (7.24) and Theorem 7.14. This implit:s that E(z - id)
= z - T and E(id) = T (Theorem 7.14(h) and (g) respectiv,ely). Hence, the
theorem is proved in the complex case.
T is a self-adjoint operator on real Hilbert spnce H, we
consider the self-adjoint operator T c on the complex Hilbert space Hc
Exercise 5.32). what have Tc exactly one
Ec which Edid) = The restriction Ec to is a
spectral family on H such that E(id) = T. If F were another spectral family
on H with the property that F(id) = T, then the compl(:xification of F
would be another Fe such that Fdid) = T c' There-
fore, Fc = E c ' and thus F= E. The details are left to the reader (cf. Exer-
cise 7.25). 0

EXAMPLE 1. If T is the operator of multiplication by a real function g on

E9 aEA L2 (1R, Pa) (cf. Section 7.2, Example 2 (continued», then E(t) is the
operator of multiplication by the characteristic function of {s E IR : g(s) "-
t}. We also write briefly E(t) X{SER : g($) <I}' The corresponding result
holds for multiplication operators on L2(M). The proofs are contained in
Examples I and 2 of Section 7.2.
7.3 The spectral theorem for self-adjoint operators 195

EXAMPLE 2. Let P be an orthogonal projection on H. Then

o for 1<0,
Ep(/) = { [-P for 0<1<1,
[ for I> 1
is the spectral family of P, since for this spectral family we have

f u(t) dE(t) = u(O)(I - P) + u(l)P;

consequently,

f t dE(t) = P.
In particular, the spectral family of the zero operator is given by

Eo(t) = {O[ for t < 0,

for t > 0,

and that of the identity operator by

EAt) = {OJ for t<l,

for t> 1.

EXAMPLE 3. Assume that T is a compact self-adjoint operator on H, (A.i) is

the sequence of non-zero eigenvalues of T, (lj) is the sequence of the
orthogonal projections onto the eigenspaces N(A.i - T) and Po is the or-
thogonal projection onto N(T). Then the equality

~ Pj for t<O,
( ) {J : ".I "',}
Etf= {
~ Pj+Pof for t>O
{J: Aj""}

defines the spectral family of T, since

E(id)f = ~ AjP) + OPoJ = TJ for all f E H

j

by Theorem 7.1.

Theorem 7.18 (Spectral representation theorem). Let T be a se/j-ac(joint

operator on H. Then there exist a family {Pot : a E A} of right continuous
non-decreasing Junctions and a unitary operator U from H onto
EB .. eA L2(R, Pot) for which
T= U-1TidU,
where Tid denotes the maximal operator of multiplication by the function id
196 7 The spectral theory of self-adjoint and normal operators

on fD et E A L2(R, Pet)' For the spectral family E of T we have E(t) =

U - I X<_ 00. tJU.
REMARK. In place of the assertion of Theorem 7.18 we can briefly say the
following: Every self-adjoint operator is unitary equivalent to a multiplica-
tion operator by the function u = id; these are operators that we already
know quite well. The cardinality of the set A is at most equal to the
dimension of H. The spaces L2(R, Pet) have to be chosen to be real when H
is real. One disadvantage of this theorem compared to Theorem 7.17 is that
this representation is not unique.
PROOF. If E is the spectral family of T, and fD etEA L2 (R, p,,), U, and Fare
constructed as in Theorem 7.16, then by Theorem 7.16 and Theorem 7.17
T = E(id) = U - Ift(id) U = U - ITid U,
E(t) = U-IF(t)U = U-IX(-oo.tlU, o
If E is the spectral family of the self-adjoint operator T, and u is an
E-measurable function, then we write u(T) for E(u). We already know that
uz(T) = (z - T) - I for uz(t) = (z - t) - I. The following theort!m gives further
justification for this notation.

Theorem 7.19. If u(t) = "i}_Ocjt), then u(T) = "i}=oc) Tj, where we set TO = I.
PROOF. The assertion obviously holds for n = 0 (cf. Theorem 7.14(d». Let
us assume that it holds for polynomials of degree <n - l. Then veT) =
"i} _ Ic) T) - I for v( t) = "i.j. Iclj - I. Because of the equality u = v' id + CO' it
follows from Theorem 7.14(h) that
u(T) :::) v(T)T+ col.
Since for n;;;' 1 we moreover have OCT) = O(E(id»:::) O(u(T», it follows
from Theorem 7.14(h) that

O(u(T» = D(T) n O(u(T» = D(v(T)T),

and thus

u(T) = v(T)T+ Col = (f)=1

CjTj-I)T+ Col =
.
:~
)·-0
ciT). 0

Let E be a spectral family on H. A subset M of Ill! is said to be

E-measurable if its characteristic function XM is E-measurable. We write
E(M) for E(XM)' We have E(R\M) = E(l- XM) = I - E(XM) = 1- E(M). If
E is the spectral family of T, then E(M) = XM(T). The operators E(M) are
orthogonal projections.

Proposition. Let T be a self-adjoint operator on H, and let E be its spectral

family.
7.3 The spectral theorem for self-adjoint operators 197

(1) For any E-measurable subset M oj R the subspace R(E(M» is a reducing

subspace oj T, i.e., E(M)T C TE(M).
(2) Let OeM, T)= E(M)O(T). We have <J, TJ) <y1lJ1I 2 Jor all JE
0« - 00, y), T), J:I=O, and <J, TJ) < yllJII 2 Jor all J E 0( - 00, y], T).
Similar statements holdJor (y, 00) and [y, 00).
(3) We have <1, Tf) < yllf1l 2 Jor all J E OCT) if and only if E(t) = I Jor all
t .,. y. We have <1, TJ) .,. yllJII 2 Jor all J E OCT) if and only if E(t) = 0
Jor a/l t <y.
(4) For every bounded interval J the subspace R(E(J» is contained in OCT)
and TE( J) = E(id XI) E B( H).
(5) Assume that / E Rand s > O. We have J E OCT) and II(T- t)JII<sIlJII
Jor every J E R(E(t + s) - E(t - s».
(6) IJ u is a real-valued E-measurable Junction, then we have Jor the spectral
Jamily F oj u(T) that F(t) = E({s E R : u(s) < t}) Jor every t E Rand
F(M) = E({ s E R : u(s) EM}) Jor every Borel set M.

The prooJs are obvious when T is the operator Tid on $aEAL2(R, Pa). In
the general case we use the spectral representation theorem 7.18. As to Part
6, observe that F(M) = X",(u(T» = (X", 0 u)(T), X", 0 u = X{.EN: u(.)E"'} and
use Section 7.2, Example 2.
Now we are in a position to define the nth root of an arbitrary
non-negative self-adjoint operator and give the polar decomposition of
unbounded operators.

Theorem 7.20.
(a) Every non-negative self-adjoint operator T possesses exactly one non-
negative self-adjoint nth root TI/n. IJ E is the spectral Jamily oj T, then
Tl/n = Jtlln dE(t) (here t lln ., 0 Jor t" 0; Jor t < 0 the value oj t lln is
immaterial, as the ~-measure oj ( - 00, 0) vanishes Jor every J E H). IJ T
is compact, then T In is also compact.
(b) Let T be a densely deJined closed operator Jrom HI into H2. The operator
T can be uniquely represented in the Jorm T = US, where S is a
non-negative self-adjoint operator on HI and U is a partial isometry with
initial domain R{S) andJinal domain R{T). We have S=(T*T)1/2; we
again write 1TI Jor (T* T)1/2.
PROOF.
(a) By Theorem 7.14(f), (h) and (i) the given operator T lin is a non-nega-
tive nth root of T. By Part 6 of the last proposition the spectral family
of Tl/n is

If S is an arbitrary non-negative nth. root of T with spectral family F,

198 7 The spectral theory of self-adjoint and normal operators

then the spectral family of T= sn is

F ( ) {O for t < 0,
n t = F(t l / n ) for t > O.
The equality E = Fn follows because of the uniqueness of the spectral
family of T. Consequently, F= E 1/ n, and thus s= TI/n.
If T is compact, then there exists a compact non-negative nth root of
T by Theorem 7.4. The operator TI/n constructed here is then com-
pact, because of the uniqueness of the nth root.
(b) If T= US is such a representation, then T*T=S*U*US=SU*US=
S2, since U* U is the orthogonal projection onto R(S). The equalities
S = (T* T)I/2 = ITI follow from part (a). This proves the uniqueness,
since U is uniquely determined by the equality UITll=,Tl. It remains
to prove the existence of such a representation. By 1beorem 5.40 we
have 0(1 TI) = OCT) and II ITllll = II Tll1 for all 1 E OCT). The mapping
V: R(ITD~R(T), ITll~Tlis therefore isometric, and T= VITI. The
operator V can be uniquely extended to an isometric mapping V acting
from R(I TI) onto R( T). The equality U(J + g) = Vl for 1 E R(I TI) and
g E R(I TI).1 proves the assertion. D

The boundedness and the norm of self-adjoint operators can be seen

from their spectral family.

Theorem 7.21. A self-acijoint T on H is bounded if and only if there exist real

numbers YI and Y2 lor which
E(t) = {O lor t <y.,
I lor t >h
We can then choose
YI = m = inf{ <1, Tl) : 1 E O(T), 11111 = l},.
Y2 = M = sup{ <1, Tl) : lEO(T), 11111 = I}.

For m <t <M we have E(t)=FO and E(t)=FI.

PROOF. By Theorem 4.4 the operator T is bounded if and only if m and M
are finite. By Part 3 of the last proposition this is equivalent to the
assertions that E( t) = 0 for t < m and E(I) = I for t > M. If we had E(lo) = 0
for some to>m, then we would have <1, Tl) > toll 1112 for alllE D(T), on
account of Part 3 of the last proposition. This contradicts the definition of
m. The relation E(I)9I= I for t < M follows similarly. D

EXERCISES

7.19. Assume that T is a self-adjoint operator, E is its spectral family, and M is a

closed subspace of H. We have M= R(E(t» if and only if M is a reducing
subspace of T and <f. Tf) <; tllfll 2 for f E O(T) n M and <I, Tf) > tllfll 2 for
fE O(T)n M.1.
7.3 The spectral theorem for self-adjoint operators 199

7.20. Suppose that T is self-adjoint with spectral family E, a, bE p(T) n IR, and r
is a positively oriented Jordan curve for which (a, b) n o(T) lies inside rand
all other points of the spectrum of T lie outside r.
Then E(b) - E(o) =
(2'1T i)-I f rR(z, T) dz in the sense of the Riemann integral.

7.21. Assume that T is self-adjoint, 0 < p < I, and 1 E D(T), TP = f IP dE(t) with
an arbitrary choice of the branch of the function IR-+C, I~ IP. Then
II TPlll <IITII1 P llll1 l - p •
Hint: Holder's inequality for II TP111 2 = f 1/12p dIlE(/)111 2 •
7.22. Let T be a non-negative self-adjoint operator, and let>.. > O. Then (>.. + T)-l
= (1/>") f O'e-' cos (>.. -1/2STI/2) ds, in the sense of the improper Riemann-
Stieltjes integral.
7.23. (a) Let TI and T2 be densely defined and closed. Assume that D(TI ) = D(T~
and II Ttfll = II Tdll for all 1 E O( TI)' Then ITil = IT 2 1· If TI and T2 are
self-adjoint and non-negative, then TI = T 2 •
(b) A densely defined closed operator is normal if and only if D(1'*T)
=D( TT*) and II 1'* TIll = II TT*111 for 1 E D( 1'* T).
7.24. If T is self-adjoint and u, v: IR-+IR are Borel functions, then u(v(T»=
(u 0 v)(T).
Hint: Use the spectral representation theorem 7.18.
7.25. Assume that H is a real Hilbert space, T is a self-adjoint operator on H; He
and Te are the complexifications of Hand T, respectively (cf. Exercise 5.32).
(a) The mapping K : He-+ He, (j, g)~(j, - g) for 1, g E H has the properties
KK = I and K(ahl + bh 2 ) = a* Khl + b* Kh2 for all hi' h2 E He (K is called
a conjugalion; cf. also Section 8.1). We have H={hEHc: Kh=h}.
(b) We have KTe = TeK and K(z - Td- I = (z* - Td-IK for all z Ep(Td.
(c) If Ec is the spectral family of T e, then KEd/) = Ed/)K for all IE IR.
Hint: Use (7.22).
(d) The formula E(I) = Edt)IH' 1 E IR defines a spectral family on H such that
T= £(id).
(e) E is the only spectral family for which T= E'(id). (This proves Theorem
7.17 for real Hilbert spaces.)
7.26. Let T be a densely defined closed operator on H and let T= UITI be its
polar decomposition.
(a) We have N(JTJ) = N(T) and R(jTj)=R(1'*).
(b) We have 1'* = ITIU* and 11'*1 = UITIU*.
Hint: T1'* ... (VI TIU*)( VI TI V*), and UI TI V* is non-negative and self-
adjoint.
(c) Prove, furthermore, that T= VITI ... IT*IV= V1'*V, 1'* = V*I1'*1 =
ITIU* = V*TV·, ITI = V·T= 1'*V= V*I1'*IU, 11'*1 = UT* = TV· =
UITIV·.
(d) If T is normal, then ITI=I1'*I, UITI=ITIU, V·ITI=ITIV·, and
R(T)=R(T*)= ROT!).
(e) If T is normal, then the operators T}(T*)k are normal for all j, kENo;
furthermore, O(T}(T*)k) = D(T}+k), (Ti(T*)k)* = Tk(T*Y and
II T}(T*i'fll = II Ti+ kll1 for 1 E D(T}+k).
7.27. Let S be closed and symmetric but not self-adjoint, and let T-ISI. Then·
200 7 The spectral theory of self-adjoint and nonnal operators

T + AS is self-adjoint for ,\ E( -I, I), closed and not self-adjoint for IAI> I,
and not closed for IAI = 1. (S is T-bounded with T-bounded 1.)

7.28. Let T be a self-adjoint operator on L2(M) and let (zo- T)--n be a Carleman
operator for some Zo E p( T).
(a) (z - T)-n is a Carleman operator for every z Ep(T).
(b) E(b) - E(a) is a Carleman operator for all a, bE IR.
Hint: (z - Tnzo - T)-" and (z - Tt(E(b) - E(a» are bounded.
7.29. (a) If T is a non-negative self-adjoint operator, then D(T) is a core of TI/2.
(b) If A is symmetric and non-negative, T is the Friedrichs extension of A,
and S is an arbitrary non-negative self-adjoint extension of A, then
D(TI/2) C D(S 1/2).
Hint: D(TI/2) is the completion of D(A) with respect to the norm
{lIfll 2 + <f, Af)} 1/2.
7.30. Let Sand T be non-negative self-adjoint operators. We write T, S if
0(SI/2)C 0(TI/2) and IIT'/~II' IIS'/~II for aBfE 0(SI/2).
(a) If OEp(T), then T,S if and only if S-I,T- 1 (i.e., <f,S-'J)<-
<f, T - 'J) for all f E H).
Hint: Show that T,S~IIT'/2S-'/211' 1~{IIS-I/2Tl/~1I <- 11111 for all
fED(TI/2)}~S-1 ,T- I .
(b) If A is symmetric and non-negative, T is the Friedrichs extension of A,
and S is an arbitrary non-negative self-adjoint extension of A, then S, T.
7.30' Let T be self-adjoint on a complex Hilbert space, and let A be T-bounded.
(a) The relative bound of A equals lim,-+_ooIlA(it - T)-'II.
(b) If T is bounded from below, then the relative bound of A equals
limt-+oollA(t + T)-'II·

7.4 Spectra of self-adjoint operators

We know from Section 5.3 that the spectrum of a self-adjoint operator is a
closed subset of the real axis (of course, this is a non-trivial statement only
in the complex case). In this section we want to study how the spectral
points of a self-adjoint operator may be characterized by means of its
spectral family. .
Let T he a self-adjoint operator on H. The spectrum aCT) and the point
spectrum a/T) are defined as in Section 5.2.

Theorem 7.22. Let T be a selj-amoint operator on H, let E be the spectral

family of T, and let To be a restriction of T for which 1~ = T. Then the
following statements are equivalent:
(i) s E aCT);
(ii) there exists a sequence (i,,) from D( T) for which lim inf II J" II > 0 and
(s - T)fn~O;
(iii) there exists a sequence (gn) from D( To) for which lim inq gn II
(s - TO)gn~O;
> °and

(iv) E(s+f.)-E(s-f.)'i=Ofor every f.>O.

7.4 Spectra of self-adjoint operators 201

PROOF. The equivalence of (i) and (ii) immediately follows from Theorem
5.24.
(ii) implies (iii): Since OCTo) is a core of T, for every n EN there exists a

°
gnEO(TO) for which IIgn-fnll<n-: 1 and IITogn-Tfnll<n- l • Hence,
lim inf II gnll > and (s - TO)gn- O.
(iii) implies (ii): This is clear because of the inclusion OCTo) C OCT).
(ii) implies (iv): Assume that (iv) does not hold, i.e., that there exists an
e > 0 such that E(s + e) - E(s - e) = 0. If (J,,) is the sequence from (ii), then

lI(s - T)fnll 2 = fls - 112 dIlE(t)fnIl 2 ;;. (2 f dIlE(t)J,.1I2

= e2 11fnll 2
(here we have used the fact that Is - tl ;;. f almost everywhere relative to the
measure induced by p/" = II E( . )fn 11 2). This is a contradiction because
(s - T)fn -0 and lim inf II fn II > 0.
°
(iv) implies (ii): We have E(s + n -I) - E(s - n -I) ~ for every n E N,
i.e., there exists an fn E R(E(s + n -1) - E(s - n -I» such that IIfnll = l. For
this sequence we have lim inf IIfnll = 1 and

lI(s - T)fnll 2 = fls - 112 dIlE(/)jnIl 2 < n- 2 f dIlE(t)jnIl 2

= n- 2 I1fnIl 2 _ O. o
Corollary 1. Let a <b. If E(b) - E(a) ~O then (a, b] n aCT) ~0. We have
(a, b) n a(T)~0 if and only if E(b -) - E(a)~O.
PROOF.
(a) Assume that (a, b] c p( T). Then by Theorem 7.22 the spectral family E
is constant in some neighborhood of s for every s E (a, b]. Conse-
quently, E is constant in (a, b], and thus E(b) - E(a) = s -
lim .....o+(E(b) - E(a + f» = 0.
(b) If (a, b)n a(T)=0, then we can prove, as in part (a), that E is
constant in (a, b), i.e., that E(b - ) - E(a) = s -lim<->o+(E(b - f) - E(a
+e»=O. If AE(a, b)na(T), and £>0 is so small that (A-e, A+e]c
(a, b), then E(A + f) - E(A - e) ~ 0, so that E(b - ) - E(a) #0. 0

Corollary 2. A self-acijoint operator T is bounded from below if and only if its

spectrum is bounded from below. The greatest lower bound of T is equal to
min aCT).

PROOF. By Part 3 of the proposition preceding Theorem 7.20, we have

(f, Tf);;' yllfll 2 for allf E OCT) if and only if E(t) = 0 for t <Yo If E(t) = 0
for I <y, then by Theorem 7.22 no spectraipoint of T can lie in (- 00, y).
Therefore, min aCT) ;;'y. If aCT) is bounded from below, then E is con-
stant in (- 00, min aCT»~. Consequently, E(/) = 0 for t < min aCT), and
thus y ;;. min a( T). 0
 202 7 The spectral theory of self-adjoint and normal operators

Theorem 7.23. Let T, To and E be as in Theorem 7.22. Then the lollowing

. assertions are equivalent:
(i) s E ap(T); .
(ii) there exists a Cauchy sequence (f,,) Irom D(T) lor which lim II In II >0
and (s - T)ln .....O;
(iii) there exists a Cauchy sequence (gn) Irom D(To) lor which limll gnll > 0
and (s - To)g" .....O;
(iv) E(s) - E(s - ) =1= O.
We have N(s - T) = R(E(s) - E(s - ».
PROOF. (i) implies (ii): If I is an eigenelement of T belonging to the
eigenvalue s, then we can choose the constant sequence In = I.
(ii) implies (iii): Since D(To) is a core of T, for every n E: N there exists a
gn E D(To) for which IIgn - 1,,11 <:n- I and II Togn - Tlnll <:n- I • Everything
follows from this.
(iii) implies (i): Let/= lim gn' Then (s - T)/= lim(s - 10)gn =0.
(i) implies (iv): Let I be an eigenelement of T belonging to the eigen-
value s. Then

i.e., Is - tl = 0 almost everywhere with respect to the measure induced by

p,=IIE(.)/1I 2 • Therefore, E(t)/is constant in (-oo,s) and in (s, 00), and
thus E(s - )/=0, E(s)/- E(s + )/= I. Hence, E(s) - E(s - )=1=0.
(iv) implies (i): For every IE R(E(s) - E(s -»
we obviously have
lI(s - T)fl1 2 = Jls - tl 2 dll E(t)/1I 2 = O.
It follows from the last two steps that N(s - T) = R(E(s)- E(s - 0 ».
Proposition. Any isolated point A 01 the spectrum 01 a self-a4ioint operator T
is an eigenvalue 01 T.
PROOF. There is an f:>0 such that [A-f:, A+f:]n a(T)=' {A}. Hence, by
Corollary I to Theorem 7.22, E is constant in [X - e:, X) and in (X, X+ f:].
. Since XE a( T), by Theorem 7.22 we have

E(X) - E(X -) = E(X + e:) - E(X - f:) =1= 01,

i.e., X is an eigenvalue of T. o
The essential spectrum ae(T) of a self-adjoint operator T is the set of
those points (of a(T» that are either accumulation points of aCT) or
isolated eigenvalues of infinite multiplicity. The set aiT)- a(T)\ae(T) is
called the discrete spectrum of T. By the last proposition ad( T) is the set of
those eigenvalues of finite multiplicity that are isolated points of a(T). We
say that T has a pure discrete spectrum if a.( T) is empty.
7.4 Spectra of self-adjoint operators 203

Theorem 7.24. Let T, To and E be as in Theorem 7.22. Then the following

statements are equivalent:
(i) sEae(T);
w
(ii) there exists a sequence (f,,) from D(T) for which fll ~ 0, lim inf Ilfllll > 0
and (s- T)fll~O;
W
(iii) there exists a sequence (gil) from D( To) for which gil ~ 0, lim inf II gil II
> 0 and (s - T O>gll~O;
(iv) for every t: > 0 we have dim R(E(s + t:) - E(s - t:» = 00.
PROOF. (i) implies (ii): If s is an eigenvalue of infinite multiplicity, then
there exists an orthonormal sequence (f,,) in N(s - T); this sequence has
the properties required in (ii). If s is an accumulation point of aCT), then
there exists a sequence (Sll) from aCT) such that Sll +s, Sll +sm for n +m,
and Sn~S as n-+oo. Let us now choose t:1I > 0 so small that the intervals
(SII - €", Sll + t:1I ) be mutually disjoint. Since Sll E aCT), we have E(sll + £n)-
E(sll - (11) + O. Let us choose a normed element.l,. from R(E(s" + (11 ) -
E(sll - (11»' Then we obviously have <fll,fm) = 811m and (s - T).I,.~O.
(ii) implies (iii): For every n E N there exists a g" E D(To) for which
II gil - .1,.11 "n - 1 and II To gn - TfnII "n - I. All the properties required in (iii)
follow from this.
(iii) implies (iv): Assume that we have dim R(E(s + £) - E(s - f»~ < 00
for some t: > 0, i.e., that the projection E(s + t:) - E(s - £) is compact. For
the sequence (gn) from (iii) we then have (E(s + £) - E(s - £»gll~O. Con-
sequently,

lI(s - T)g1l1l2 = jls - tl 2 dIlE(t)gnIl 2

:> t;2[ j dll E(t)gnIl 2 - j X(s-<, S+<j(t) dIlE(t)gnIl 2 ]

= t: 2 [ II gnll 2 -1I(E(s + £) - E(s - £»g,,1I 2 ],

and thus

This is in contradiction with (iii).

(iv) implies (i): If dim R(E(s) - E(s - » = 00, then s is an eigenvalue of
infinite multiplicity (Theorem 7.23). Therefore, s E a.(T). Let R(E(s)-
E(s -» be finite-dimensional, but let dim R(E(s + £) - E(s - t:» ... 00 for
all t: > O. Then the set (3 - (, 3) U (3, 3 + (1 contains at least one spectral
point for every t: > 0 by Corollary 1 to Theorem 7.22; hence s is an
accumulation point of aCT). 0

Theorems 7.22 and 7.24 provide natural characterizations of aiT). We

will not give these explicitly here. They are partly contained in the
following propositions.
204 7 The spectral theory of self-adjoint and normal operators

Proposition. If a < b and dim R(E(b - ) - E(a» = m (m E I~), then aCT) n

(a, b) consists of only isolated eigenvalues of finite multiplicity. The sum of
the multiplicities of these eigenvalues equals m.
PROOF. By Theorem 7.24 we have (a, b)n ae(T) =0, i.e., (a, b)n a(T)c
ad(T). Let AI' A2 ,
••• be the eigenvalues of T in (a, b) (there are at most
countably many of these, since they cannot accumulate ill the interior of
(a, b». Then
n
E(b-)-E(a) > ~(E(Aj)-E(Aj-)) forall nEN.
j-I

Therefore, only finitely many eigenvalues AI' A2' ... , Ak <:an lie in (a, b),
and
k
dim R(E(b -) - E(a» = ~ dim (E(Aj) - E(~i - )).
j=1

The right-hand side equals the sum of the multiplicities of the eigenvalues
AI"" ,Ak • 0
Proposition. If dim R(E(b» = m < 00 for some bE IR, then T is bounded
from below.
PROOF. By the previous proposition the interval (a, b) contains at most m
spectral points for any a < b; hence (- 00, b) contains at most m spectral
points. The smallest of these finitely many eigenvalues is a lower bound for
T by Corollary 2 to Theorem 7.22.

Proposition. If dim R(E(b) - E(a» = 00, then ae(T) n [a, b] *0.

PROOF. If we had 0e(T)n[a, b]=0, then for every sEra, b] there would
be an € > 0 such that dim R(E(s + to) - E(s - f»~ < 00. The interval [a, b]
could be covered by finitely many intervals of this kind. This implies that
dim R(E(b) - E(a» < 00, which is a contradiction. 0

Theorem 7.25. Let T be a self-adjoint operator on H, and let H= HI ED H2ED

H3 with dim H3 = m < 00. Assume that the orthogonal projE'ction Pj onto H.J
maps OCT) into itself for j = I, 2.5 If

then (a, b) n aCT) consists of only isolated eigenvalues; the sum of the
mUltiplicities of these eigenvalues is at most m.

5Then P3 also maps D(T) into itself.

7.4 Spectra of self-adjoint operators 205

PROOF. Let us assume that dim R(E(b -) - E(a»;;' m + l. Then there

exists anf E R(E(b - ) - E(a»n H/- such thatf*O. Hence,f= Pd + Pd,
and. putting c = (a + b)/2, we have

II(T-c)fI1 2 = !X<Q.b)(t)(t- cf d Il E (t)fI1 2 < (b;afllfI12.

With properly chosen ~E 11<. lajl = I it follows from this that

I«T- c)f, Pd)1 + I<P2 f. (T- c)f)1
= ai<Pd. (T- c) f) + a!<Pd. (T- c)f)
= <aIPlf+ a2 P2f. (T-c)f) ~ lIalPd+ a2Pdllll(T-c)fll

= {II P dIl 2 + IIP2 fln 1/2 11 (T_ c)fll = Ilfllll(T- c)f\l < b; a IIf1l2.
Consequently.
bllPdll 2 ~ <Pd. TP2 f) = <Pd, Tf) - <Tf, Pd) + <Pd, TPd)
~ c(IIP2 fIl 2 -\l P dIl 2 ) + <Pd, (T- c)f)
-«T- c)f. Pd) + allPdll 2

~ bllPdll 2- b; a IIfll2 + I<P2f, (T- c)f)1

+I«T- c)f. Pd)1

< bllPdll 2 •
This is a contradiction. Therefore, dim R(E(b -) - E(a» <m. and the
assertion follows from the first proposition after Theorem 7.24. 0

A corresponding theorem holds in the case when (a. b) is a half-line.

Theorem 7.26.
(a) Let T be a self-adjoint operator on H, and let HI be a closed subspace of
H such that dim H).L = m < 00. Assume that PI D(T) C D(T) for the
orthogonal projection PI onto HI> and that
<f. Tf) ;;. bllfl1 2 for f E P)D(T).
Then (- 00, b) n oCT) consists of only isolated eigenvalues; the sum of
the multiplicities of these eigenvalues is at most m. The operator T is
bounded from below.
(b) Let T be self-adjoint on H, and let HI be an m-dimensional subspace of
D(T). Assume that <f, Tf) -< allfll2 for all f E H). Then dim R(E(a»:.>
m.
PROOF.
(a) With H2 = {OJ and H3 = HI.L the assumptions of Theorem 7.25 are
fulfilled for every a <b. Consequently, (- 00, b)n oCT) contains only
206 7 The spectral theory of self-adjoint aE,d normal operators

isolated eigenvalues of finite multiplicity, with total multiplicity less

than or equal to m. In particular, aCT) n (- 00, b) is a finite set;
therefore, So = min a( T) exists. Hence, T is bounded from below by
Corollary 2 to Theorem 7.22.
(b) If we had dim R(E(a» <m, then there would be an f E HI for which
f;j:.O andf.lR(E(a». For thisf

(f, Tf> " allfl1 2 < <f, Tf>,

by assumption and by part (2) of the proposition preceding Theorem
7.20. This is a contradiction. 0

In some investigations another partition of the spectrum is useful. For

that we first need some definitions. Let T again be a self-adjoint operator
on H with spectral family E. Let fi, denote the closed linear hull of all
eigenelements of T. We call fi, = fi,(T) the discontinuous subspace of H
with respect to T. The orthogonal complement of Hp is called the continu-
ous subspace of H with respect to T. This is denoted by He = He(T). The
singular continuous subspace Hse = H.,e( T) of H with respect to T is the set of
those f E He for which there exists a Borel set N C II;t of Lebesgue measure
zero (briefly: a Borel null set) such that E(N)f= f. The subspace Hse is
closed. This can be seen in the following way. If (jn) is a sequence from
Hse, ]" ~ f, and the Nn are Borel null sets such that E( Nn)!n = fn' then
N= U nNn is also a null set and E(N)f= lim E(N)]" = lim E(Nn)!n = limfn
= f. Since He is closed, f lies in He' hence in Hse. The orthogonal comple-
ment of Hse relative to He (i.e., He e Hse) is called the absolutely continuous
subspace of H relative to T. This is denoted by Hac = HacCT). The singular
subspace Hs of H with respect to T is defined by the equality Hs = Hs(T)
= Hp $ Hsc. Let Pp' Pc' Psc ' Pac' and Ps denote the orthogonal projections
onto these subspaces.

Theorem 7.27. Let T be a self-acfjoint operator on H with spectral family E.

Denote, for every f E H, by Pi the measure induced on IR by means of
IIE( . )f112.
(a) fi, equals the set of those f E H for which there exists an at most
countable set A c II;t such that pllI;t\ A) = 0, i.e.,for which the measure Pi
is concentrated on (at most) countably many points.
(b) He is the set of those f E H for which p/{ t}) =0 for eve~y t E 1I;t, i.e., for
which the function t~ II E(1)fIl 2 is continuous. (For f E He we obviously
have pIA) = 0 for every at most countable set A c R)
(c) Hs equals the set of those f E H for which there exists a Borel null set N
such that pllI;t\ N) = 0, i.e., for which Pi is singular with respect to
Lebesgue measure.
(d) Hac equals the set of those f E H for which pi N) = 0 for every Borel null
set N, i.e., for which Pi is absolutely continuous with respect to Lebesgue
measure.
7.4 Spectra of self-adjoint operators 207

PROOF.
(a) If h is an eigenelement for the eigenvalue Aj> andf= ~~ IC~, then we
obviously have E({Aj :jEp"J})f=j. If A={Aj :jEp"J} and E(A)f=f,
then
00 00

f= L E({Aj})f= L (E(Aj)-E(Aj-»f.
)=1 )-1

Since (E(A) - E(A) - »f (when it is different from zero) is an eigenele-

ment belonging to the eigenvalue Ai' the element f lies in the closed hull
of the eigenvectors.
(b) We have E({t})fE Ii, and <1, E({t})f) = IIE({t})fIl 2 for every fE H
and every fER. If fEHc=Ii,1., then we have pJ{t})=IIE({t})fIl2=
(f, E({t})f) = 0 for every t E R. Let pJ{t}) = 0 for every t E R. Then
pJA)=O for every at most countable set AcR; hence IIE(A)fIl2=
pJA) = O. If g E Ii. and A is an at most countable set such that
E(A)g=g, then (f,g)=(f,E(A)g)=(E(A)f,g)=(O,g)=O, i.e.,
fEH/=Hc·
(c) We have Hs = Ji, + H$C" If f= ~ + isc with ~ E Ii, and fs< E Hsc ' then
there exist an at most countable set A c R and a Borel null set N for
which E(A)~ =~, E(N)ise = ise' and thus E(A U N)f= f; the set AU N
is a Borel null set. Conversely, let E(N)f= f for some Borel null set N.
The set A of jump points of the non-decreasing function tH-IIE(t)fIl 2 is
at most countable, and E(A)f Eli,. Let g E Ji, be arbitrary. Then there
exists an at most countable set A' such that E(A') = g. Since E({ I})(f
-E(A)j)=E({t})f-E({t}nA)f=O for every IER (as {t}={t}nA
for lEA and E({/})f=E({t}nA)f=O for t=t=A), it follows that
E(A')(f - E(A)j) = 0, and thus
(g,f- E(A)f) = (E(A')g,f- E(A)f)
= (g, E(A')(j- E(A)j» = 0,
i.e., f - E(A)f Eli,1. = He. Because of the equality E(N)(f - E(A)j) =
f- E(A)f we have f- E(A)fE H.sc; therefore, f= E(A)f+(f- E(A)j)
E Hp + Hsc = Hs.
(d) We have Hac = He e H.c = H.1.. Assume that f E Hac = Hs1.. Then
II E( N)f1l 2 = (f, E( N)f) = 0 for every Borel null set N, since E( N)f E
Hs. Let us now assume that E(N)f= 0 for every Borel null set N. If
g E Hs. then by part c) there exists a Borel null set N such that
g = E(N)g. It follows from this that
(f. g) = <1. E(N)g) = (E(N)f. g) = O.
and thus f E H/ = Hac. o
Let M be a closed subspace of H. and let P be the orthogonal projection
onto M. We say that M reduces the operator T if PT c TP (this obviously
implies POeT) c 0('1'). as well as (I - P)O(T) C D(T) and (J - P)T C.
208 7 The spectral theory of self-adjoint and normal operators

T(J - P); d. also Exercise 5.39 and Section. 7.3, Proposition 1). The
formulae O(TM) = M n OCT) and T ~= Tf for f E O(TM) define an opera-
tor on M. The subspace M is a reducing subspace of T if and only if M.L is
a reducing subspace of T. Then O(T) = O(TM) + O(TMJ.)·

Theorem 7.28. Let T be a self-adjoint operator on H, with the spectral family

E, and let M be a reducing subspace of T. Then T M and T M" are self-adjoint
on M and MJ., respectively. We have o(T) = o(TM) U o(TMC). The subspace
M reduces T if and only if PE(t) = E(t)P for every t E R, where P denotes
the orthogonal projection onto M.
PROOF. O(TM) is dense in M, since Me D(TM)= O(T)J. n M= {O} because
of the equality O(T) = O(TM ) + D(TMJ.). As a restriction of a self-adjoint
operator, T M is surely Hermitian. Therefore, T M is symmetric on M. It
remains to prove that O«TM)*) c O(TM)' Let g E D(TM)*)' Then
«TM)*g,f) = «TM)*g,fl) = <g, T ~I) = <g, T ~I + = <g, Tf)
TMd2)
for all f= fl + f2 E O(TM) + O(TMJ.) = OCT); i.e., gEM n D(T*) = M n
OCT) = O(TM)' We can show the self-adjointness of T MJ. analogously. We
have
lI(z- T)f11 2 = lI(z- T M)J11I 2 + lI{z- T MJ.)f211 2
for every z E IK and for f= fl + f2 E O(TM) + O(TMJ.) = OCT). It follows
from this by Theorem 5.24 that z E peT) if and only if z E peTM) n peTMJ.),
i.e., oCT) = oCTM) U oCTMJ.)·
If M reduces the operator T, then R(z, T)P= R(z, T)P(z - T)R(z, T)
c R(z, T)(z - T)PR(z, T) = PR(z, T). Therefore, R(z, T)P = PR(z, T),
because O(R(z, T)P) = H. It follows, by formula (7.22) for the spectral
family E, that E(t)P = PE(t) for all t E R Now let the equalities E(t)P =
PE(t) (t E IR) hold true. Then

Jltl2 dIlE{t)PfIl2 = JI/1 2 dIlPE{t)fIl 2 ~ JI/1 2 dIlE{/)fW < co

for every f E OCT). Consequently, Pf E OCT). This implies that O(TP):J

O(T) = O(PT). If f E OCT) and (un) is a sequence of step functions such
that Un tends to id in L2(1R, Pj)' then un tends to id in ~2(1R, Ppj) also, and
PTf = PE(id)f = P lim E(un)J = lim PE(un)J = lim E(un)Pf
= E(id)Pf = TPj. 0

Theorem 7.29. Let T be a self-adjoint operator on H. The subspaces

Hp> Hc' Hsc, Hac' and Hs reduce the operator T.
PROOF. It is obviously sufficient to show that Hp and Hs reduce the
operator T. The remaining assertions follow from this, because He = H/,
Hsc = Hse Hp' and Hae = H/.
For any f E Hp there is an at most countable subset A of IR such that
E(A)f= j. Consequently, E(t)f= E(t)E(A)f= E(A)E(/)f E Hp for all 1 E IR,
7.4 Spectra of self-adjoint operators 209

i.e., E(t)Pp = PpE(t)Pp. It follows from this that

PpE(t) = (E(t)Pp)* = (PpE(t)Pp)* = PpE(t)Pp = E(t)Pp.
We can show in an entirely analogous way that PsE(t) = E(t)Ps if we
replace A by the null set N for which E(N)f= f. 0

We denote by Tp' Tc' Tsc ' Tac ' and Ts the restrictions of T to
Ji" Hc' Hsc' Hac, and Ha· These operators are called the (spectral) discon-
tinuous, continuous, singular continuous, absolutely continuous, and singular
parts of T.
The continuous spectrum CJc(T), singular continuous spectrum CJac(T), ab-
solutely continuous spectrum CJac(T), and the singular spectrum CJs(T) of T
are defined as the spectrum of Tc' Tsc ' Tac ' and Ta, respectively. In contrast
with this, the point spectrum CJ/T) is defined as the set of eigenvalues of T
(these are also the eigenvalues of Tp; however, in general we only have
CJ(Tp) = CJp( T); cf. Exercise 7.33). The sets CJe<T), CJsc(T), CJac(T), and CJ.(T)
are closed (as they are spectra). We obviously have CJ(T)=CJp(T)UCJsc(T)
U CJae< T) = CJs( T) U CJac ( T) = CJ/ T)u CJc( T).
We say that T has a pure point spectrum, pure continuous spectrum, pure
singular continuous spectrum, pure absolutely continuous spectrum and pure
singular spectrum if Hp = H, Hc = H, Hac = H, Hac = H, and Hs = H, respec-
tively. We then have CJ(T)=CJp(T), CJ(T)=CJc(T), CJ(T)=CJsc(T),CJ(T)=
CJaiT), and CJ(T)= CJs(T), respectively.

EXAMPLE 1. Let Pp' Psc ' and Pac be measures on IR. Let Pp be concentrated
on a countable set (i.e., there exists a countable set A such that pilR \ A) =
0), let Psc be singular continuous (i.e., there exists a Borel null set N such
that Psc(1R \ N) = 0 and Psc( {t}) = 0 for every t E IR), let Pac be absolutely
continuous (i.e., Pac ( N) = 0 for every Borel null set N). Let T be the
operator of multiplication by the function id on L2(R, pp) Ee LilR, Psc ) Ee
LilR, Pac)' Then fi, = L2(R, pp)' Hc = L2(1R, PsC> Ee L2 (1R, PaC>, Hsc = L2 (1R, Psc)'
Hac = L2(R PaJ, and Hs = L2(1R, pp) Ee L2(R, PsJ. The proof will be left to
the reader. We shall show in Exercise 7.34 that every self-adjoint operator
is the orthogonal sum of operators of this type.

EXERCISES
7.31. If T is a self-adjoint operator on H with pure point spectrum, and {ea : a E
A} is an orthonormal basis of eigenelements with corresponding eigenvalues
{Aa : a E A}, then D(T) is equal to the set of those f E H for which
~aIAa<ea,jW < 00; we have Tf= '2. aAa<ea,j)ea for f E D(T).
7.32. Let T be a self-adjoint operator on the infinite-dimensional Hilbert space H.
(a) If T is bounded, then ae(T) 7'= 0.
(b) T is compact if and only if T is bounded and a.(T) ... {OJ.
(c) If H is separable, then B<XJ(H) is the only closed ideal in B{H).
210 7 The spectral theory of self-adjoint alld normal operators

7.33. For any self-adjoint operator T we have the following:

(a) ap(T) ... a(J;,): however, ap(T) is in general not closed, and thus ap(Th~'
a( Tp) in general.
(b) aCT) - ap(T) U aAT) ... ap(T) U a.c(T) U aac(T) = a.(T) U aac(T).
(c) aiT)- ac(T)u(ap(T) \a.t<T».

7.34. Every self-adjoint operator is unitarily equivalent to a maximal operator of

multiplication by the function id on ($" EA L2 (IIl, p,,»EB(EB pEs L2(IIl, ap»EB
(EB yEr L2(R, 'Ty»: where the measures p", ap, 'Ty have the following properties:
for every p" there exists a countable set A" such that
p,,(R \.4,,) = 0: ap( {t}) ... 0 for every t E IR, and there exists a Borel null set Np
such that ap (Ill \ Np) = 0 (f3 E B); all 'Ty are absolutely continuous with
respect to Lebesgue measure.
Hint: Apply Theorem 7.18 to Tp, Tsc , and Tac.
7.35. Let T be a self-adjoint with spectral family E, and let u : IR-~IR be E-measur-
able. Then we have the following:
(a) a(u(T» c u(a(T»; if u is continuous on a(T), then o(u{T»'" u(a(T»: if
T is bounded and u is continuous, then a(u(T» = u(a(T»:
(b) fi,(T) c Hp(u(T», u(op(T» c ap(u(T»:
(c) Hac(u(T» c Hac(T), oac(u(T» c u( oac(T».
7.36. Let T be self-adjoint with spectral family E, and let u,v : IIl~C be E-
measurable.
(a) If u(t) = v(t) for all t E a(T), then u(T) - v(T).
(b) If T has a pure point spectrum, then it is sufficient to assume that
u(t) ... vet) for all t E ap(T) in (a).
7.37. Let T be a self-adjoint operator with spectral family E, and let M be a
subspace of D( T) such that II(;\. - T)/II "ell/il for all I EM.
(a) dim R(E(;\' + c) - E(;\' - e - » ;> dim M.
(b) If dim M ... 00, then o.(T) n [;\. - e, ;\. + e] ,,=0.
7.38. Let T be a self-adjoint operator on H. The operator T hilS a pure discrete
spectrum if and only if (;\. - T) -\ is compact for every ;\. E, p( T). If H is not
separable, then a,(T),,= 0.
7.39. Let A be a symmetric semi-bounded operator, let T be the Friedrichs
extension of A, and let S be an arbitrary semi-bounded self-adjoint extension
of A.
(a) dim ET(t) " dim Es(t) for every t E Ill.
Hint: Exercise 7.30c and Theorem 7.26b.
(b) If S has a pure discrete spectrum, then T also has a pure discrete
spectrum.

7.5 The spectral theorem for normal operators

We have shown in Section 7.3 that every self-adjoint operator can be
represented in the form f Rt dE(t), where E is a (real) spectral family
defined on iii. Here we show that every normal operator can be repre-
7.5 The spectral theorem for normal operators 211

sented in a corresponding way as an "integral" fez dG(z), where G is a

spectral family defined on C.
Let H be a Hilbert space. A function G : C_B(H) is called a complex
spectral family if there are real spectral families E and F such that
G(t + is) = E(t)F(s) = F(s)E(t) for all s, t E ~.

Theorem 7.30. Let G be a complex spectral family on H with G(t + is) =

E(t)F(s).
(a) G(t + i s)G(t' + is') = G(min{t, t'} + i min{ s, s'D for all s, s', t, t' E R;
in particular, G(z).;;; G(z') for all z, z' E C such that Re Z .;;; Re z' and
1m z.;;; 1m z'.
(b) zn-z, Re Zn >z, and 1m Zn >z imply G(zn)~ G(z).
(c) If (zn) is a sequence from C for which Re Zn- - 00 or 1m zn- - 00, then
s
G(zn) - 0; if (zn) is a sequence from C for which Re zn-oo and
s
1m zn-oo, then G(zn)-I.
(d) The spectral families E and F are uniquely determined by G; we have
E(t) = s -Iims-."" G(t + is) and F(s) = s -Iim t .... "" G(t + is).
(e) If we set G(J) ... E(J 1)F(J2) for an arbitrary interval J= J 1 X J 2 = {z E
C : Re Z E J 1, 1m Z E J2}, then G(J)G(J') = 0 for J n J' = 0.
(f) The equality "II. J) = II G( J)fll z for all intervals J in C defines a regular
interval function on C for every f E H (cf. Appendix AI).
PROOF.
(a) For all s, s', I, I' E ~

G(t + is)G(t' + is') = E(t)F(s)E(t')F(s')

= E(t)E(t')F(s)F(s') = E(min{t, t'})F(min{s, s'};
= G(min{ I, I'} + i min{ s, s'}).
(b) Let (En) and ('I1n) be null-sequences from [0, 00). Then
lim II( G(z + En + i'l1n) - G(z»)!11 2
n->oo
= lim (E(t + En)F(s + 'I1n)f - E(t)F(s)f,j)
n .... ""
= lim {(F(s
n->oo
+ 'I1n)f, E(t + En)!) - (F(s)f, E(t)!)} = 0
for all Z = t + is and f E H.
(c) Assume that Zn = tn + isn, and In- - 00 or sn- - 00. Then for allf E H

since in (F(sn)f, E(tn)!) at least one factor tends to zero, while the
other remains bounded. If tn_oo and sn-oo, then it follows for all
fE H that
)i~ II! - G(zn)fIl2 = Ji~ {lIfll 2 - (F(sn)!, E(tn)!)} ... O.
212 7 The spectral theory of self-adjoint and nonnal operators

s
(d) It follows from the formula F(s)--,>! as s--,>oo that
E(t)J= lim F(s)E(t)J= lim G(t+is)f
s~oo $--+00

for all J E H. The assertion for F follows in a similar way.

(e) If J=J\XJ2> J'=J;xJ2 and JnJ'=0, then J\nJ;=00r J 2 nJ2
=0, and thus
G(J)G(J') = E(J\)E(J{)F(J2 )F(J2) = 0,
(f) If J) is an arbitrary interval in ~, then we can show easily that there
exists a s sequence' (J\ II) of open intervals for which J\ c J\ , " and
E(J1, II) --'> E(J1) (the sequence (J 1.,,) has to be chosen so that we have
XJ,)t)--'>XJ,(t) fOI all t E ~). Let J2." be a similar sequence with the
property F( J2• ,,) --'> F( J2 ). Then
y/(J1X J2 ) = <F(J2 )J, E(J1)J) = }i~ <F(J2.,,)J, E(J\, n)!)

= }i~ IIG{J1,n X J2.,,)JI1 2 = }i~ yiJl.n xJ2.,,)

for all J E H. Since the intervals J 1." X J 2." are open and J\ X J 2 C J 1."
X J 2• '" this proves the regularity of Y/' 0

If for the step function u : C--,>IK,

u(z} = .L" cj X1;(z)

)-1

we define the integral with respect to the complex spectral family G by the
equality

Ju(z) dG{z) = L"

)=1
cj G(0),

we can use the same arguments as we did in Section 7.2 whcm we discussed
integration with respect to a real spectral family. Let us also denote by Y/
the measure that is induced by the interval function y,.
We say that a
function u : C--,>IK is G-measurable if it is Yrmeasurable for all J E H. If
u E L 2(C, y,), then we can define the integral

Ieu(z} dG(z)J
just as in Section 7.2. For every G-measurable function u: C--,>IK the
formulae
O( G(u» = {J E H : u E L2 (C, Y,)},

G(u)J = L u(z) dG(z)J for JED( G(u»

7.5 The spectral theorem for normal operators 213

define a normal operator on H. We also write

G(u) = Ie u(z) dG(z).

The assertions of Theorem 7.14 (and their proofs) remain valid and will be
used in the following without further explanation.

Tbeorem 7.31 (The spectral theorem for bounded normal operators). Let H
be a complex Hilbert space, and let T E 8(H) be a normal operator. Then
there exists exactly one complex spectral family G for which

T = Lz dG(z).
We have G(t + is) = E(t)F(s) = F(s)E(t) for s, t E R; where E and F are the
spectral families of the self-adjoint operators A =(T+ 1'*)/2 and B=(T-
T*)/2i,. respectively. G(z) = I for z E C such that Re z:> II Til and 1m z:>
II Til, and G(z) = 0 for z E C such that Re z < -II Til or 1m z < -II Til.
Moreover, T = A + iB and A B = BA. The operators A and B are called the
real part and the imaginary part of T.
PROOF. If A and B are defined as in the theorem, then it is obvious that
A* =A, B* = B, T=A +iB, and

AB = ~i (T2- TT*+ T*T- T*2) = ~i (T2_ r*2) = BA.

Then we also have

R(z,A)R(z',B) =[(z'-B)(z-A)r l =[(z-A)(z'-B)r l

= R(z', B)R(z, A)
for all z, z' E C\R For the spectral families E and F of A and B,
respectively, it now follows with the aid of formula (7.22) that
E(t)F(s) = F(s)E(t) for all s, t E R.

Now we define the complex spectral family G by the formula

G(t+is)=E(t)F(s) forall s,tER.

Then
G(t + is) = E(t)F(s) = I for t > II Til and s > II Til,
since IIA II .;;; II Til and II B II .;;; II Til. Similarly,
G(t+is) = E(t)F(s) = 0 for t < -IiTii or s < -IiTii.
If for (un) we choose a bounded sequence of step functions defined on R
214 7 The spectral theory of self-adjoint and normal operators

which converges to id uniformly for It I ~ II TIl, then

A = f. t dE(t) = lim f. Un(t) dE(t) = lim Jrc un(Re z) dG{z)
R n-HlO R n ..... oo

= fc Re z dG(z),

where we have used the fact that G({zEC: RezEJ})='E(J). We can

obtain similarly that

B = fclm z dG(z).

Consequently, it follows that

T = A + iB = 1c Re z dG(z) + i JrIm
c
z dG(z)

= fc(Rez+i Imz) dG(z) = fcz dG(z).

It only remains to prove the uniqueness of G. Let G' be a complex spectral

family for which

T = Lz dG'(z) and G'(t + is) = E'(t)F'(s) = F'(s)E'(t).

Then

A = HT+ T*) = if.c (z + z*) dG'(z) = f. Re z dG'(z) = 1t dE'(t).

C R

The unicity of the spectral family of self-adjoint operators gives that

E = E'. We can show similarly that F= F'. Therefore, G = (i'. 0
If G is the spectral family of the operator T in the sense of Theorem
7.31, then we also write u(T) for G(u). Theorem 7.19 also holds for normal
operators.

EXAMPLE 1. Assume T is a compact normal operator on the complex

Hilbert space H, P\I' A2, ••• } are its non-zero eigenvalues, P j is the
orthogonal projection onto N(A j - T), AO = 0, and Po is the orthogonal
projection onto N(T). Then the formula
G(z)=~{Pj:jENo,ReAj~Rez,ImAj~Imz} for zEC
defines a complex spectral family, and we have

T = fc z dG(z).

The proof goes as in the self-adjoint case.

7.5 The spectral theorem for normal operators 215

Theorem 7.32 (The spectral theorem for normal operators). Let T be a

normal operator on the complex Hilbert space H. Then there exists a unique
complex spectral family G for which

T= kzdG{z).

The operators A = (T + T*) /2 and B = (T - T*) /2i are self-adjoint. For the
spectral families E and F of A and B, respectively, we have
G{t + is) = E{t)F{s) = F{s)E(t).
We have T = A + iB and T* = A - iB. The operators A and B are called the
real part and the imaginary part of T (cf. Exercise 7.49).
PROOF. Since this theorem has little significance in the applications, we
shall not work out its proof in detail. The operator S = T* T is self-adjoint
by Theorem 5.39. Let Eo denote its spectral family. TS .. ST holds,
because T is normal. Hence R(z,S)TcTR(z,S) for all zEC\R. It
follows from this by (7.22) that Eo(t)T C TEo(t). This implies that the
subspaces
Hn = R( Eo« - n, - n + 1] U (n - I, n J»), n E N

reduce the operator T (d. Section 7.4). The same holds for T*. Let
Tn = TIIi,' We have Hn C O(T* T) c OCT) by Section 7.3, Proposition 4.
Therefore, O(Tn} = H". For all f E Hn we have
IITJII2 = <T*Tf,f) ~ nllfll 2 ,
i.e., Tn E B(Hn). Similarly, (T,,)* E B(Hn), (T,,)* = T*11i, and II(Tn)*fll =
II T*fll = II Tfll = II TJII for all f E Hn, i.e., Tn is a bounded normal operator
on H". Consequently, by Theorem 7.31 there is a complex spectral family
Gn such that

Tn = J> dG..{z),

where Gn(t + is) = E,,(t)Fn(s) = F,,(s)Eit) for s, t E R with the spectral

families En and Fn of

An = t(Tn + T,n = 4(T+ T*)11i. and Bn = ;i (T" - T:) = ;i (T- T*)IIi.'

respectively. In what follows we consider the operators G,,(z), En(t), and
Fn(s) to be defined on H (more precisely, we should write InGn(z)Pn, etc.,
where In is the embedding of Hn into Hand Pn is the orthogonal projection
onto H,,). Since H= fBnEI'IlH", the sums

"EN

E(t) = ~ E..{t) and F(s) = ~ Fn(s)

nEI'Il "EN
216 7 The spectral theory of self-adjoint and normal operators

exist for all z E C and s. t E ~ in the strong sense. It is easy to see that E
and F are (real) spectral families and G(t + is) = E(t)F(s) =: F(s)E(t).
Now let Do = L{ Hn : n EN}. Then Do is a core of T. For every j E Do
there exists an N E N such that j E $~_ I Hn- Then

Lz dGn(z)J
N N
TJ = L TnP,j = L
n=1 n-I C

= Lz d( ~ Gn(z»)J Lz dG(z)j
C n=1
=
C
= G(id)j.

The restriction To of T to Do is therefore contained in the normal operator

G(id). Then we also have T= To e G(id). and it follows by Section 5.6,
Proposition 1 that T = G(id).
Similarly,

AJ = HT+ T*)J = J/ dE(t)j, BJ = ;i (T- T*)J = Ls dF(s)j

for JE Do. As the restrictions of (T+ T*)/2 and (T- P)/2i to Do are
essentially self-adjoint by Exercise 5.43, it follows from this that

HT+ T*) = A = Jt dE(t), ;i (T- T*) = B = Is dF(s).

We have TeA +iB, by construction. In order to prove that T=A +iB,

we therefore have to prove that D(A +iB) = D(A)n D(B) e D(T). If JE
D(A)n D(B), then (observe that G({z EC : Re z E J}):= E(J), G({z E
C: ImzEJ})=F(J»

{lzl2 dIlG(z)JII 2 = {(IRe Z12+ 11m Z12) dIlG(z)JI1 2

consequently,] E D(T). D
EXAMPLE 2. Let 'I be a measure on C( = ~2) and let T be th,e maximal
operator of multiplication by the function id, i.e., let
D(T) = {jEL 2 (C, 'I): idJEL 2(C, 'I)},
TJ = idJ for J E D(T).
Then T is normal, and the spectral family G of T is given by
G(z)J = XzJ for J E L2(C. 'I),
where Xz is the characteristic function of the set {w E C : Re w " Re z and
1m w (; 1m z}.
7.5 The spectral theorem for normal operators 217

We can show in an entirely analogous way as for self-adjoint operators

that every normal operator is unitarily equivalent to an orthogonal sum of
operators of the kind considered in Example 2. We give the theorem
without proof.

Theorem 7.33 (The spectral representation theorem for normal operators).

If T is a normal operator on a complex Hilbert space, then there exist a
family {Y.. : a E A} oj measures on C and a unitary operator U: H~
EB .. EA L2(C, Ya ) Jar which
T = U-1TidU;
where Tid is the maximal operator of multiplication by the Junction id on
EB aEA L2(C, Y.. ), i.e.,
D(Tid ) = {Ua)aEA E EB L2 (C, Ya) : (idJ.. )aEA E ED L2(C, Ya)},
aEA aEA
TidUa) .. EA= (i<JJ",)aEA for Ua)",EAED(Tid )·

The spectral points of a normal operator can be classified similarly to

those of a self-adjoint operator, and they can be characterized by means of
the spectral family. We do not go into details and mention only the
following result.

Theorem 7.34. Let T be a normal operator on a complex Hilbert space, and

let G be the spectral family of T.
(a) z E o(T) if and only if
G(z + f + if) + G(z - f - if) - G(z + f - if) - G(z - f + if) *- 0
for every f > O.
(b) IIR(z, T)II = d(z, O(T»-I Jar z Ep(T), where d(z,o(T» denotes the
distance of the point z from o( T).

For the proof: (a) The proof is analogous to that of Theorem 7.22.
Observe that G(XM) = G(b l + ib2 ) + G(a J + ia2 ) - G(a l + ib2) - G(b l + ia2)
for M={zEC: a\<Rez.;;b l ,a2 <Imz.;;b2 }.
(b) /lR(z, T)/I >d(z, o(T»-J by Theorem 5.15. Since we have R(z, T)=
uz(T) with uz{w) = (z - w)- t, and as luz(w)1 .;; d(z, o(T»-1 G-almost every-
where, it follows that /lR(z, T)II ';;d(z, O(T»-I.

Theorem 7.35. Assume that T" (n EN) and T are bounded normal operators
on the complex Hilbert space Hand "T - T"II ~O as n ~ 00. Then
o( T) = lim o( T,,)
n->oo
= {z E C: there exists a sequence (z,,) from C Jar which
z" E oCT,,) and z,,~z}.
218 7 The spectral theory of self-adjoint and normal operators

(This is an assertion about the continuity of the set-valued function T~

oCT) defined on the set of bounded normal operators. We I;annot allow all
bounded operators here, and II T - Tnll ~O cannot be replaced by
Tn ~ T; cr. Exerctses 7.41 and 7.42.)
s •

PROOF. If zf1.o(T), i.e., zEp(T), then (second Corollary to Theorem 5.11)

zEp(Tn ) for sufficiently large nand lI(z- T,,)-I_(Z- T)-III~O. Hence,
lI(z- T,,)-III~II(z- T)-III as n~oo. Since d(z, o(T,,»=IIR(z, T,,)II- I, we
have lim..... oo d(z, o(T,,»=IIR(z, T)II- I >O. Consequently, the point z is
not contained in lim oCT,,).
Assume z E o( T). Then by Theorem 5.43 there is a sequence (i,,) from H
for which IIfnll = I and (z - T)f,,-+O. Then
(z - T")f,, = (z - T)f" - (T" - T)f" ~ 0;
hence d(z,o(T,,»=IIR(z, Tn)II-I<II(z-TII)fnll-+O. It follows from this
that z E lim oCT,,). 0
Theorem 7.36. If U is a unitary operator on a complex Hilbert space, then
there exists a real spectral family E for which E( t) == 0 for t <: 0, E( t) == I for
t;> 2'1T and U= fe il dE(t) (cj. also Exercise 7.46).

PROOF. By Section 5.2, Example 2 the spectrum of U is contained in

{zEe: Izl=l}={eil : 0<t<2'1T}, i.e., G({eil : 0<t<2'1T}=I, where G
denotes the complex spectral family of U. Then the formu!lae
for t < 0,
for 0<t<2~T,
for t ~ 2'1T
define a real spectral family. We can verify easily that U= Ieil dE(t). 0
EXERCISES

7.40. A function G : C~8(H) is a complex spectral family if and only if G(z) is an

orthogonal projection for every z E C and properties (a), (b) and (c) of
Theorem 7.30 are satisfied.
7.41. We cannot replace norm convergence by strong convergence in Theorem
7.35.
Hint: Let the operators TT E 8(12) be defined by Tm(/,,)neN=(/),h,·· .,
1m' 0, 0, 0, ... ). Then Tm ~ I, o(Tm) = {O, I} and o(I) = {I}.
7.42. Let the operators Tm , mEN, and T from B(l2(l» be defined by the formulae
/,,+) for n,.,-I,
Tm(f"),,ez = (g")"EZ with gIl =
{ !/o for n = -1,

T(f")"EZ = (g")"EZ with gIl = { 1.0,,+1 for

for
Ii'''' -I,
n=-1.
7.S The spectral theorem for nonnal operators 219

(a) We have II Til-II Tmll = 1. Therefore, a(T) C {z EC : Izl< I} and o(Tm)

c{zEC:lzl<I}.
(b) For every z E C such that Izl < 1, the vector (fn)nEZ defined by in = 0 for
n < 0 and in = zn for n ;;. 0 is an eigenelement of T belonging to the
eigenvalue z; hence a(T) = {z E C : Izl< I}.
(c) T'; 1 E 8(12(1» for all mEN, and r(T'; I) = 1. Consequently, {z E C : Izl
> I} cp(Tm- I ), and thus {z EC : Izl < I} cP(Tm ) (cf. Exercise 5.27).
(d) Theorem 7.35 does not hold in general if the operators Tm and T are not
normal.
7.43. Assume that T is a bounded normal operator on the complex Hilbert space
H, r> II Til, and u is a function, holomorphic in {z E C : Izl < r} with
u(z) = ~j..ou jzj for Izl <r. Then u(T) = ~j..ou jT), where the series con-
verges in B(H). For self-adjoint operators this statement holds in real Hilbert
spaces, as well.
7.44. Let TI and T2 be similar bounded normal operators on the complex Hilbert
space H, i.e., let there exist a bijective operator S E B(H) such that ST1 ==
T 2S. Prove that TI and T2 are unitarily equivalent.
(a) S = ezT,S e - zT, for all z E C.
Hint: Exercise 7.43.
(b) eizTiSe-izTf=ei(z1J+zOT,lS e-i(zTf+zoT,). Consequently, lIeiz1JS e-izTfil <
IISII for all z E C.
(c) S_eiz1JSe-izTfforallzEC.
Hint: The Liouville theorem and part (b).
(d) S11 = 11 S.
Hint: Differentiate in (c) with respect to z and substitute z = O.
(e) T1S·S"",S·ST1, TdSl=ISITI and T.JSI-1=ISI-IT I ·
(f) If S- UISI is the polar decomposition of S, then U is unitary, lSI is
bijective, and U - SIS I.r
(g) We have TI == U - IT2 U with the operator U from part (f). Hence TI and
T2 are unitarily equivalent.
7.45. Assume that H is a complex Hilbert space, S, T E 8(H), ST= TS, and Tis
normal.
(a) Then sr ... T*S.
Hint: Use the reasoning of Exercise 7.44(b), (c) and (d) for T, - T2 - T.
(b) If G is the spectral family of T, then SG(z)'" G(z)S for all z E C.
7.46. Prove the uniqueness of the spectral family E in Theorem 7.36.
7.47. If T is a normal operator on a complex Hilbert space and n E N, then there
exists exactly one normal operator A for which A n = T and G( {z - rei'" : r ;;.
0, 0 < cp < 2'17/ n}) == I for the spectral family G of A.
7.48. If A and B are self-adjoint (not necessarily bounded) operators on a complex
Hilbert space with spectral families E and F, and E(t)F(s) == F(s)E(t) for all
s, t E iii, then T = A + iB is normal.
7.49. Decomposition in real- and imaginary parts for arbitrary operators (cf. Exer-
cise 5.38 and Theorem 1.32);
(a) If T is a densely defined operator on H such that D(T) C D(r), then the
220 7 The spectral theory of self-adjoint and normal operators

operators A=(T+T*)j2 and B=(T-T*)j2i are symmetric, D(A)

= D(B), and T=A +iB.
(b) If A and B are symmetric, D(A)=D(B), and T=A+iB, then A=
(T+ T*)j2 and B=(T- T*)j2i.
(c) Even if T is closed and D(T) = D(T*), we cannot expec1; that A and B in
(a) are essentially self-adjoint.
Hint: Choose T=ICI+(iCj2) or T=iICl+(Cj2), where C is a closed
symmetric but not self-adjoint operator.
(d) Even if A and B are self-adjoint, we cannot expect that T in (b) is closed.
Hint: Take an unbounded self-adjoint operator C. Define A and B by
the formulae D(A)= D(B) = D(C)ffi D(C), AU, g) = (Cg, Cf) and BU, g)
= i( Cg, Cf) for j, g E D( C).

7.6 One-parameter unitary groups

One example of the significance of the theory of self-adjoint operators is
shown by Stone's theorem (cf. Theorem 7.37 and Theorem 7.38). We shall
learn more about this later in this section.
Let H be a Hilbert space. A family {B(t) : t E iii} of operators out of
8( H) is called a one-parameter group if
B(O) = I and B(s)B(t) = B(s + t) for all s, t E iii.
(This is then a "representation" of the additive group iii by operators on
H.) The one-parameter group {B( t) : t E R} is said to be strongly continu-
ous if the function

is continuous for every f E H.

Let {B(t) : t E iii} be a one-parameter group of operators on H. The
operator A defined by the formulae

D(A) = {f E H: lim.!. (B(t) - I)f exists},

1--.0 t

Af=lim.!.(B(t)-I)f for fED(A)

1-->0 t
is called the infinitesimal generator of {B(t) : t E iii}.
It can be proved that every strongly continuous one-parameter group
possesses a densely defined infinitesimal generator. In the following we
only consider (one-parameter) unitary groups (i.e., one-parameter groups of
unitary operators). The situation is somewhat simpler in that case.

Theorem 737. Let T be a self-adjoint operator on the complex Hilbert space

H, let E be the spectral family of T, and let

U(t) = eitT = fe its dE(s) for t E iii.

7.6 One-parameter unitary groups 221

Then {U(t) : t E IR} is a strongly continuous (one-parameter) unitary group.

The infinitesimal generator is iT. We have U(t)fE D(T) for all fE D(T)
and t E IR.
PROOF. By Theorem 7.14 we have
U(t) E 8(H) and U(t)* = U( - t) = U(t)-I for all t E IR,
i.e., U(t) is unitary for all t E IR (cL Theorem 4.34). For all x, y E IR

leix - elY I = le i (x- y )/2 _ e- i (x- y )/21 = 21sin x~ y I.

It follows from this that

II V(t)f - V(t')fI1 2 = II j (e i/' - eit'S) dE(s)fI12

= jleilS ei/ 'sl 2 dIlE(s)fIl2

r
_

= 4jlsin (1 -2 t ')s dIIE(s)fIl2

for all f E Hand t, t' E IR. Because of the relations

(t-t')S\ (t-I')s
\sin 2 .;;; 1 and sin 2 ~0 as I' ~ I,

it follows by Lebesgue's theorem that

II V(/)f - V(t')fll ~ 0 as t' ~ I.

This proves the strong continuity of Vet). For all f E H and IE IR we have

1
I(V(/)-1)1= If'(ellS-I) dE(s)1·
1
Because of the relations
1 .
- (e'lS - I) ~ is as I ~ 0, s E IR,
1
and

1+(e -1)1.;;;
i/s lsi for s, I E R, t "1= 0,

the right-hand side converges, as t~O, if and only if the function u(s) = lsi
belongs to L2(IR, PI) (with pfis) = IIE(s)fIl 2), i.e., if and only if 1 E D(T). The
limit equals j Tf. Consequently, iT is the infinitesimal generator of { V(t) : t
E IR}.
If 1 E D( T), then for every t E IR

JIsl2 dsllE(s)U(t)fI12 = jlsl2 dsil U(t)E(s)1112 = jlsl2 dIlE(s)1112 < 00,

i.e., V(t)f E D( T). 0

222 7 The spectral theory of self-adjoint and nonnal operators

Actually, every strongly continuous (one-parameter) unitary group can

be represented in this form.

Theorem 7.38 (Stone). Let {U(t): t E JR} be a strongly continuous (one-

parameter) unitary group on the complex Hilbert space H. Then there exists a
uniquely determined self-a4ioint operator T on H for which
U(t) = eitT for all t E JR.

If H is separable, then strong continuity can be replaced by weak measura-

bility, i.e., it is sufficient to require that the function
<f, U(· )g) : JR --'» C, I ~ <f, U(t)g)
is measurable (with respect to Lebesgue measure on JR) for all j, g E H.
PROOF. The equality U( t) = eitT implies that iT is the infinitesimal genera-
tor of {U(/) : t E JR}. This proves the unicity of T and, at the same time,
provides an opportunity to construct T. In what follows let A be the
infinitesimal generator of { U( t) : I E JR 1 and let T = - iA. We show that T
is essentially self-adjoint and U(t) = eitT• Since iT is then the infinitesimal
generator of {U(/) : t E JR}, it follows that T= T. First, let us assume that
the group is strongly continuous.
D(A) is dense: For every ({I E COOO(JR) and every f E H the equality

f<p = J ({I(s) U(s)f ds

defines an frp E H (the integral is extended only to the support of ({I and can
be understood as a Riemann integral). We have

I(
I U(t) - J)f<p = I1J((I(s)( U(s + t) - U(s»f ds
= +f [((I(S) U(s + I) - ({I(S) U(s)]f ds
= +J [ ({I(S - t) U(s) - ({I ( s) U( s) ] f ds

= J +«({I(s-t)-({I(s»U(s)f ds --'» - J~D/(s)U(s)fds

as I --,»0. The set Do = U<p : f E H, ({I E COOO(JR)} is therefore contained in
D(A). If «((In) is a sequence from COOO(JR) such that
I
({In(s) = 0 for Isl;;.-,
n
({In(s) ;;. 0 for all s E JR,

J ({In(s) ds = I for all n E ~,

7.6 One-parameter unitary groups 223

thenf<p. ~f as n~oo, since

IIf"", - fll = II! 'Pn(s)( U(s) - 1)f dsll

..;; SUP{II(U(S)-I)fll: -;<s<;}.

Hence Do, and thus D(A), too, is dense in H.

T= - iA is symmetric: If f, g E D(T) = D(A), then

<g, Tf) = -i<g,Af) = -lim

1--+0
i(g, !(U(t)-1)f)
t
= -lim i(!(U(-/)-I)g,f)
1--+0 1

= lim i( _1 (U( - t) - 1)g, f) = i<Ag, f)

1-+0 - t

= <Tg,f)·
R( ± i - T) is dense in H: Assume g E R(i - T).l = N(i + r). Since for
every f<p E Do and all t E ~

U(/)f", = U(t) J 'P(s) U(s)f ds = J 'P(s - t) U(s)f ds E Do,

it follows that
d
dt<g, U(/)f",) = <g, AU(t)f",) = <A*g, U(I)j",)
= <- iT*g, U(t)f",) = -<g, U(t)f",).

The function h(t) = <g, U(t)f<p) is therefore a solution of the differential

equation h' = - h, i.e., we have h(t) = e-'h(O). Since U(t) is unitary, h is
bounded; however this is possible only if <g,j",) = <g, U(O)f'9) = h(O) = O.
Since this holds for all f", E Do, we have g = O. Consequently, R(i - T)= H.
We can show similarly that R( -i- T)= H. Hence, T is essentially self-
adjoint.
We have U(t) = eilT : Let V(t) = eilT and f E Do. Because of the relation
fE D(f) we also have V(t)fE D(i) by Theorem 7.37 and d/dt V(t)f=
iTV(t)f. Since, moreover, U(t)f E Do C D(T) for all t E R it follows that

:t (U( t)f - V( I)f) = i TU(t)f - i TV(t)! = i T( U( t)f - Vet)!).

Consequently, because T is self-adjoint,

d 2 -
dl IIU(/)f- V(t)fll = 2 Re<U(t)f- V(t)f, iT(U(t)f- V(t)f» = O.

It follows from this that U(t)f = V(t)f for all t E ~ and all f E Do, because
U(O)f = V(O)f. Since Do is dense, this implies that U(t) = V( t) = eilT•
224 7 The spectral theory of self-adjoint and normal operators

It remains to prove that weak measurability implies strong continuity in

·the separable case. Let IE H. As t~<U(t)/. g) is bounded (with bound
11/11 II gil) and measurable for all g E H. the function
g ~ loa <U(t)/. g) dt
is a continuous linear functional with norm" all III for every a> O. By the
Riesz representation theorem (Theorem 4.8) there exists an la E H for
which

<fa. g) = foa<U(I)/.g) dt.

Therefore.
<U(s)/a. g) = <Ia. U( - s)g) = loa<U(t)/. U( - s)g) dt

= r <U(t + s)/. g) dt = f
J,
a a
<
+ s U(t)/. g) d/.
o s
and thus

!< U(s)/a• g) - <la' g)1 "Ifo s<U(t)/. g) dtl + I£a+s <U(t)j. g) dtl

" 21s1 11/11 II gil·

Hence.
<U(s)fa, g) -+ <la' g) as s -+ 0,
i.e., U(· )/a is weakly continuous at the origin. Because of the equality
"U(s)fall = llfall. the continuity of U(· )/a at the origin follows from this,
since
II U(s)fa - fa 112 = II U(s)lall 2 - 2 Re< U(s)fa, la) + II/all 2 -+ 0 as s -+ O.

We show, in addition, that the set of elements fa is dense in H. It follows

from this that U(s) is strongly continuous at the origin, and thus every-
where. Let h be orthogonal to all 101.' and let {e" : n EN} be an orthonor-
mal basis of H. Then

ra<U(t)e", h) dt = <ell . a' h) = 0

)0

for all n E N and all a > O. It follows from this that for all n EN we have
<U(t)e", h) =0 almost everywhere in (0,00) (cf. Theorem AI6(c». Conse-
quently, there is a to> 0 such that
<U(to)e", h) = 0 for all n E N.
As U(to) is unitary, {U(to)e" : n E N} is an orthonormal basis. Hence, we
must have h = O. 0
7.6 One-parameter unitary groups 225

Corollary. Let { U(t) : t E R} be a strongly continuous unitary group, and let

iT be its infinitesimal generator. Then the initial value problem
I d
T dt u(t) == Tu(t), u(O) = f

is uniquely solvable for every f E O( T) and the solution is u(t) == U( t)f. (A

solution is a continuously differentiable function defined on R with values in
OCT) that satisfies the differential equation.)
PROOF. As U(t)fEO(T) for all fEO(T) and all tER, the function
u(t) == U(t)f is a solution of the initial value problem. If u and v are
solutions, then u(O)-v(O)=O, and dldt lIu(t)-v(t)1I 2 =2 Re<u(t)-v(t),
>
i T( u( t) - v(t» = 0, i.e., u(t) = v(t) for all t E R. 0
A few more words about the significance of Stone's theorem: In quan-
tum mechanics the states of a system are described by the normalized
elements of a Hilbert space. If u(t) is the state of the system at time t, then
we write u( t) = U( t)u(O); for reasons derived from physics, U(t) has to be a
linear operator. Of course, every state has to be possible at any time; since,
moreover, U(t) has to preserve the norm of the states, it follows that U(t) is
unitary. If, in addition, we require strong continuity (weak measurability in
the separable case), which is quite plausible on a physical basis, then it
follows that there exists a self-adjoint operator T such that U(t) = eilT; in
particular,

+ :t U(t)f = TU(t)f

for all f E OCT). Since the time dependent Schrodinger equation is of this
form, this implies that the Schrodinger operators must be self-adjoint.

EXAMPLE l. Let H = L2(R). The formula

U(t)f(x) = f(x + t), f E L 2(R)

defines a strongly continuous unitary group. The infinitesimal generator is

A =iT, where
OCT) = W2 , J(R} and Tf = -if'

(for W2,I(R) see Section 6.4). It is obvious that T contains the operator
TJ,o defined by the equalities

O(TJ,o) =Co""(R) and T1,of= -if'.

As by Theorem 6.30 T"o is essentially self-adjoint, the assertion now
follows.
226 7 The spectral theory of self-adjoint and normal operators

EXAMPLE 2. Assume that H= L2(M) and g : M~R is a measurable func-

tion. The equality
U(t)f(x) = exp{itg(x))f(x), f E L 2(M)

defines a strongly continuous unitary group. It is easy to see that the

infinitesimal generator is A = iT, where T is the maximal operator of
multiplication by the function g.

Theorem 7.39. Let T be a self-adjoint operator on the complex Hilbert space

H, and let M be a closed subspace of H. If eisTf E M for all f E M and s E R,
then M reduces T and eisT M = M, eisT M 1. = M 1. for all s E R.
PROOF. We have eisT M = M for all s E R, since

eisTMcM for all sE R

by assumption, and because every f E M can be written in the form

for every s E R
For all f E M, gEM i and s E R we have

<eiSTg,f) = <g, e-i.Tf) = o.

Consequently, eisT M i C M i , and thus eisT M i = M 1.. If P is the orthogo-
nal projection onto M, then

P eisT = eisTp for all s E R,

because of what we have just shown. We have to show that PT c TP. Let
fE D(T). Then
-i . -i .
PTf = P lim - (e'ITf - f) = lim - (ellTPf - Pf).
1-.0 t 1-+0 t

Hence, Pf E D( T) and TPf = PTf. o

We next prove the following special case of a theorem of H. Trotter for
unitary groups that are generated by the sum of two self-adjoint operators.

Theorem 7.40. Let T, S, and T+ S be self-adjoint operators on the complex

Hi/bert space H. Then

eil(T+S) =s- lim [ei(I/")Tei(I/")S]"

" ..... 00

for all t E R.
7.6 One-parameter unitary groups 227

PROOF. For every fE D(T+ S)= D(T)n D(S)

r l(e ilT ei/S - ei/(T+S»f
= t -Ie eilT - 1)f + r 'eiIT( ei/S - 1)f - t -I( ei/(T+ S) - 1)f

~iTf+iSf-i(T+S)f= 0 as I~O.
In particular, for every f E D( T + S) there exists a C(f) ;;;. 0 for which
IIt-'(eiITei/S-ei/(T+S»fll"; C(j) foraH IER\{O}.
Since the space D(T+ S) is a Hilbert space with the (T+ S)-scalar
product <. ,
'>T+S' by the uniform boundedness principle (Theorem 4.22)
there exists a C ;;;. 0 such that
1I1-'(eiITei/S-eit(T+S»)jIl"; CllfllT+s for fED(T+S),tER\{O}.
If E is the spectral family of T+ S, then for fE D(T+ S)

lIeis(T+S)f - eis'(T+S)fll~+s = f (t2 + 1)leiSI - eis'/1 2 dIlE(t)fW

..; 4f(t2+1)lsinC~SI t)12dIlE(t)fIl2~O as S'~S,
since the integrand is bounded by (t 2 + I) (because of fED( T + S) this
function is integrable) and tends to 0 pointwise. Hence, for fixed f E D(T
+ S) and r> 0 the family of functions
<P,: [-r, r] ~H, IE R\{O},
<p,es) = t-l(e itT ei/S_ei/(T+S»eis(T+S)f

is equicontinuous. Since eis(T+ slj E D( T + S), we moreover have q:>/(S)~O

as t~O for an arbitrary S E [ - r, r). Consequently, <P, uniformly converges
to 0 on [- r, r) as t~O. Let us now make the following estimates
II[ (ei(l/n) Tei(l/n)S)n - ei/(T+S)]~1

= I :~: (e i(I/II)T ei(I/II)S)k[ ei(//II)T ei (I/II)S - ei (I/")(T+S)](ei (l/I)(T+S»"-I-k 1

1
'" n max II[ e (I/II)T ei(I/II)S -
i ei(l/n)(T+S)]e is(T+S)J11
lsi c;; III

= III max II(.!..) -I [ei(I/II)T ei(I/II)S - ei(//n)(T+ S) ]eis(T+ S)~I

lsi c;; III n J I
'" It I max{ !P1/n(S): SE [ -III, It I]}.
For fixed t the last expression tends to zero as n~oo, as we have already
proved (choose [-r, r)=[-III, ItlD. This proves the convergence for fE
D(T+ S). Since D(T+ S) is dense and all operators have norm 1, the
required strong convergence follows (cf. Theorem 4.23). 0
228 7 The spectral theory of self-adjoint and normal operators

If the operators Sand T are bounded from below, then an analogous

result holds for e- 1T, e- 's and e-/(T+S) with t > O.

Theorem 7.41. Let T, S, and T + S be selj-a4ioint operators on the Hilbert

space H. Assume that these operators are bounded from below. Then
e-/(T+S) = s - lim [e-(I/n)T e-(I/n)S]n
n~oo

for al/ t;;' O.

The proof follows that of Theorem 7.40. We consider only nonnegative t

and s; the details can be left to the reader.

EXERCISES

7.50. Let U be a unitary operator.

(a) There exists a strongly continuous unitary group (U(t) : 1 E R} for which
U(l) om U and whose infinitesimal generator has norm c;; 2'17".
Hint: If G is the (complex) spectral family of U and we define z t = r' eiltp
for z=rti'P (Oc;;r, 0c;;'P<2'17", and tER), then we can choose U(t)-
f zt dG(z).
(b) Prove Theorem 7.36 with the aid of part (a) and Stone's theorem.
7.51. If T is seH-adjoint and V is a T-bounded operator, then 11-+ Ve 1t7J is
continuous for every f E D(T).
7.52. Let T J and T2 be seH-adjoint operators on H. and assume that the strong
limit w- s -lim t ... oo eitT2 e- ilTI exists (W is called a wave operalor; cf. Sec-
tion ll.l). We have:
(a) W is isometric.
(b) R(W) reduces euT2 •
(c) R(W) reduces T2 •
(d) TI and T 2 1R(w) are unitarily equivalent; TI = W- 1T 2 1R(w) W.
 Self -adjoint extensions
of symmetric operators
8

In Sections 5.4 and 5.5 we have already learned that certain symmetric
operators (the semi-bounded and continuously invertible ones) possess
self-adjoint extensions. The question of whether all (or which) symmetric
operators have self-adjoint extensions could not be answered there. The
key to our studies was the fact that A- T was continuously invertible for
some AE iii; however, this is not always the case. In this chapter we
develop the von Neumann extension theory, which completely answers this
question. Moreover, we shall prove certain theorems about the spectra of
all self-adjoint extensions of a symmetric operator.

8.1 Defect indices and Cayley transforms

First let T be an arbitrary linear operator on a Hilbert space H. The set

1"( T) = {z Ell{: there exists a k( z) > 0 such that

II(z-T)fll>k(z)IIfil forall fED(T)}
is called the regularity domain of T.

Proposition.
1. We have z E I'(T) if and only if (z - T) is continuously invertible. Then
II(z- T)-lll <k(Z)-1 (observe that (z- T)-l need not belong to B(H».
2. If H is complex and T is Hermitian, then C\ R c I'(T).
3. If T is isometric, then II( \ {z Ell{: z = I} c I'( T).
4. I'(T) is open.
229
230 8 Self-adjoint extensions of symmetric operators

from the inequality ;;;. k(z)lIfll for that

injective and II(z I. If (z- T) ,n,p'r.n.!p and
bounded, then
lI(z- T)fll ;;;. lI(z- T)-III-III(z- T)-I(Z- T)fll
= lI(z- T)-IIl-llIfli.

Consequently, z E nT) and we can choose k(z)= II(z - T)-Ill-i.

2. Forz=a+ib andfED(T)

lI(z- T)f1l2 = b2 11fl12 ;;;. b21lf1rz.

r(T) and we can =Ibl·
then
lI(z - T)fll ;;;. III Tfll -IzlIIfllI = 11 -Izlillfll·
Consequently, z E nT) and we can choose k(z) = II-izil.
4. Let Zo E nT). If z E K such that Iz - zol <k(zo), then
II(z - T)fll ;;;. lI(zo - T)fll -Iz - zoillfil ;;;. (k(zo) -Iz - zoDllfll
T). Therefore, nT) is open. o
R(z - T)l. defect space of
I.".uu.a. number f3( T, z) = dim is called the

Theorem 8.1. The defect index f3(T, z) is constant on each connected subset
of r(T). If T is Hermitian, then the defect index is constant in the upper and
lower half-planes.
sufficient to show that is locally constant
Zo E r( T) there o such that f3(T,
with the nrnnPT!hl t:. Replace A
the orthogonal prOI!ecl:lOn
Pz is the orthogonal

IIPz - Pzoll = IIQz - QZoll ~ 0 as z ~ zoo

If we choose t: > 0 such that II Pz - PZoll <I for Iz - zol < t:, then it follows
from Theorem 4.35 that
f3( T, zo) for
If T is Hermitian, upper and lower U""'-V""""_" are connected
subsets of I'(T) t'ropmiUl()ll 2); therefore, index is constant
there.
8.1 Defect indices and Cayley transforms 231

1fT Hermitian on complex Hilbert then define

y +(T) = peT, i), y -(T) = peT, -i)
(of course, here can be Z E for which 1m Z >
and - i by an arbitrary z E C for which 1m z < 0). The pair
(y +(T), y _(T» = (y +, Y _) is called the defect indices of T.
REMARK. We can reformulate the result of Theorem in of this
definition: A symmetric operator is essentially self-adjoint if and only if its
defect indices are equal to (0, 0). A closed symmetric operator is self-adjoint if
and its indices are (0,

Let T be a symmetric operator on a complex Hilbert space. The Cayley

transform of T defined the
v = (i - T)( - i - T) - = - (i - T)(i + T) - I.

V is therefore a linear operator from R( - i - T) onto R(i - T).

Theorem 8.2. Let T be a symmetric operator on the complex Hilbert space H.

The Cayley transform of T is an isometric mapping of R( - i - T) onto
R(i - The R(I is dense in H, T= V)(I V) - I.
um,quelY determined V.
PROOF. For every g=(-i- T)fER(-i- T)=D(V) we have
II = 11(i 1')(-i -l g11 2 1I(i-
= IIfl12 + /I Tfll2 = lIe - i- T)f11 2 = II g1l2.
Consequently, is isometric. It clear that R( = R(i since
(-i-T)-l maps O(V)=R(-i-T) onto OCT) and (i-T) maps OCT)
onto R(i - T). We have

- V + (i T)(i + I = [ (i + (i ] (i
-I
+ T)- ,
J+ V = J - (i- T)(i+ T)-I = 2T(i+ T)-I.
In particular, V)= is J- injective, and
T = i(1 + V)(I - V) - I. o
REMARK. We could define a Cayley transform
Vz = (z - T)(z* - T)-I
for every such that 1m O. Then Vz is isometric mapping of
R(z* - T) onto R(z - T), and

T= z*
232 8 Self-adjoint extensions of symmetric operators

In what follows V can be replaced by Vz • We use z = i, as this saves us

from using unnecessary indices.

Theorem 8.3. An operator Von the complex Hi/bert space H is the Cayley
transform of a symmetric operator T if and only if V has the following
properties:
(i) V is an isometric mapping of D( V) onto R( V),
(ii) R(I - V) is dense in H.
The symmetric operator T is given by the equality T = i( I + V)( I - V) - I.
PROOF. If V is the Cayley transform of T, then V has properties (i) and (ii)
by Theorem 8.2. We also have then that T = i(l + V)(I - V) -I. Let V now
be an operator with properties (i) and (ii). Then I - V is injective, since the
equality Vg = g implies that
<g,f - Vf) = (g,!) - <g, Vf) = <g,f) - Vg, V!) <
=<g,f)-<g,f)=O forall jED(V),
i.e., that g E R(I - V).l, and thus g = O. Therefore, we can define an
operator T by the equality
T = i(J + V)(J - V)-I.
By assumption, D( T) = R(I - V) is dense. For all f = (I - V)jl and g =
(I - V)gl from D(T) = R(l- V) we have

<Tf, g) = -i<(J + V)(J - V)-Ij , g) = -i<(J + V)fl' (J - V)gl)

= -i{<JI,gl)+<Vjl,gl)-<fl' Vg 1 )-<Vjl' Vg 1) }
= -i{<Vfl' Vg1)+<Vfl,gl)-<fl' Vg1)-<J.. gl)}
= i<(J- V)jl' (J+ V)gl) = i<J, (J+ V)(J- V)-Ig)
= <f, Tg).
Consequently, T is symmetric.
It remains to prove that V is the Cayley transform of T. This im-
mediately follows from
(i- T) = i - i(J+ V)(J- V)-I = i[(J- V)-(J+ V)](J- V)-I
= -2i V(J - V)-l,
( - i - T).= -i[(J- V)+(J+ V)](J- V)-l = -2i(J- V)-l. 0

Theorem 8.4. Let T be a symmetric operator on a complex Hi/bert space, and

let V denote its Cayley transform.
(a) The following statements are equivalent:
(i) T is closed,
(ii) V is closed,
8.1 Defect indices and Cayley transforms 233

(iii) D( V) = R(i + T) is closed,

(iv) R( V) = R(i - T) is closed.
(b) T is self-adjoint if and only if V is unitary.
PROOF.
(a) (i) is equivalent to (iii) and to (iv): T is closed if and only if (± i-
T) - I is closed. The bounded operator (± i - T) - I is closed if and
only if its domain D«(I - T)-I) = R(i - T) = R( V) or D« - i-
T) -I) = R(i + T) = D( V) is closed (Theorem 5.6).
(ii) is equivalent to (iii): The bounded operator V is closed if and only
if its domain is closed.
(b) T is self-adjoint if and only if R(i - T) = R( - i - T) = H, i.e., if and
only if D( V) = R( V) = H. This is equivalent to the statement that V is
unitary. 0
For the construction of symmetric or self-adjoint extensions of a sym-
metric operator the following theorem is essential. The proof of this
theorem is obvious.

Theorem 8.5. Let TI and T2 be symmetric operators on a complex Hilbert

space, and let VI and V2 denote their Cayley transforms. Then T, C T2 if and
only if VI C V2•

Consequently, we can obtain all symmetric extensions of a symmetric

operator T if we determine all those extensions V' of the Cayley transform
Vof T which possess property (I) of Theorem 8.3 (since V has property (ii),
V' automatically does, too) and we calculate the corresponding symmetric
operators T' = i(I + V')(I - V') - I (Theorems 8.5 and 8.3). We can obtain
all self-adjoint extensions (provided that such exist) if we determine all
unitary extensions V' (Theorems 8.5 and 8.4b). In particular, T has
self-adjoint extensions if and only if V has unitary extensions. The follow-
ing theorem makes it possible to explicitly construct the extensions V' of
V. Here we assume, without loss of generality, that T is closed.

Theorem 8.6. Let T be a closed symmetric operator on a complex Hilbert

space, and let V denote its Cayley transform.
(a) V'is the Cayley transform of a closed symmetric extension T' of T if and
only if the following holds: There exist closed subspaces F _ of R(i - T)l.
and F+ of R( - i - T)l. and an isometric mapping Vof F+ onto F _ for
which
D(V') =R(-i- T') =R(-i- T)$F+.
V'(f + g) = Vf + Vg for f E R( - j - T), g E F+,
R( V') = R(i - T') = R(i - T) $ F _.
The spaces F _ and F+ have the same dimension.
234 SeU-adjoinl extensions of symmetric operators

(b) The operator V' in part (a) is unitary (i.e., T' is self-adjoint) if and only
if = F + R( i -
(c) T possesses self-adjoint extensions if and only if its defect indices are
equal.
PROOF.
If has the given form, then is an isometric of
R( - i - T) EEl F + onto R(i - T) EEl F _. Consequently, V' satisfies
assumption of Theorem 8.3, Since R(I V) dense, is
also dense, so that V' also satisfies (ii) of Theorem 8.3. Therefore, V'is
the transform of symmetric extension of Since V an
isomorphism of F + onto F _, we have dim F + = dim F _. If V'is the
Cayley transform a extension T' of T, then put =
R(i - T') e R(i - T), F+ = R( - i - T') e R( - i - T), and V= V'k.
Viis unitary if and only if V') H = V'), i.e., if only if
F+ = R(-i- T).l. and F_ = R(i- T).l..
By (a) (b) the V possesses a extension and only
if there exists an isometric mapping V of R( - i - T).l.. onto R(i - T).l...
This happens and only dim - - T).l.. = dim R(i T).l.. 0
Theorem 8.7. Let be operator on complex Hilbert space. The
operator T is essentially self-adjoint if and only T has one
self-adjoint extension.
(For the real case, compare Exercise 8.4.)
PROOF. If is essentially then f is the only self-adjoint
extension of T by Theorem 5.31(c). We show: If T is not essentially
self-adjoint, if f is not self-adjoint, then has either no infinitely
many self-adjoint extensions. If the defect indices of f are different, then f
thus T) has self-adjoint extension. If the defect are equal
(> 0, as f is not self-adjoint), then there are infinitely many unitary
Ii . R( - T).l.. ~R(i - T).l.. (proof!), and therefore infinitely
many self-adjoint extensions. 0
Now we are in a position to define certain classes of symmetric opera-
tors that have self-adjoint extensions,

Theorem S.S. Let T be a symmetric operator on a complex Hi/bert space.

If (T) III T has self-adjoint extensions.
(b) If T is semibounded, then T has self-adjoint extensions.
(The statements of this theorem have been proved in another way
in Sections 5.4 and 5.5.)
PROOF.
*'
(a) I'( T) is connected, since r( T) n III 0. Then y + (T) = y _ (T) by Theo-
rem
(b) Let T be bounded, for example, from below, and let c be a lower
8.1 Defect indices and Cayley transfonns 235

bound of T. Then
II(A- T)fll ;> (f, (T-A)f)lIfll- 1 ;> (c-A)lIfll
for A<C and all fE D(T), f~O. Consequently, in this case we also
have ItT)n R~0. 0

Let H be a complex Hilbert space. A mapping K of H onto itself is

called a conjugation if
(a) K(af + bg) = a* Kf + b* Kg for f, g E H, a, bE C,
(b) K2=I. (S.I)
(c) (Kf, Kg)=(g,f) for f,gEH.
An operator Ton H is said to be K-real if
(a) KO(T) c OCT),
(b) TKf= KTf for f E D(T). (S.2)

Theorem 8.9. Let H be a complex Hilbert space, and let K be a conjugation

on H. If T is a K-real symmetric operator on H, then T possesses self-adjoint
extensions.
PROOF. It follows from (S.l(b» and (S.2(a» that D(T):J KD(T):J K 20(T)
= D(T). Consequently, KD(T) = OCT). Iff E R(i - T)l., then
(Kf, (-i- T)Kg) = (KJ, K(i- T)g) = «i- T)g,j) = 0
for all g E D(T). Therefore, Kf E R( - i - T).L. We can show similarly that
if f E R( - i - T)l., then Kf E R(i - T).L. As K2 = I, we have R( - i - T)l.
= K R(i - T).L. Since {ea : 0: E A} is an orthonormal basis of R(i - T)l. if
and only if {Kea : 0: E A} is an orthonormal basis of R( - i - T)l., it
follows that dim R( - i - T).L = dim R(i - T)l.. 0

EXAMPLE 1. The formula

Kf(x) = f(x)*, f E L2(M)
defines a conjugation (the natural conjugation) on L2(M). The conditions
given by (8.1) are obviously satisfied.

(8.3) Let GeRm be open and let T be defined on L2( G) by the equalities
D( T) = COOO ( G), and Tf = - Af + qf for fED( T)
(here A ... ~j_ I(a 2/ ax}) denotes the Laplace differential form). Assume that
the function q is real-valued and measurable on G and belongs to L2, lac( G)
(i.e., it is square integrable over every compact subset of G). Then T is
obviously symmetric and K-real. Consequently, T has self-adjoint extensions
by Theorem S.9.
(8.4) Let K(. , .) : M X M_R, be a Hermitian Carleman kernel. Then the
operator TK,o from Section 6.2 is symmetric and K-real. TK,o therefore
possesses self-adjoint extensions.
'236 8 Self-adjoint extensions of symmetric operators

EXAMPLE 2. The formula

Kf(x) = f( - x)*
defines a conjugation on L2(lll m ) and on L 2 { x E Ill m : - a j <Xj <aj}.

(8.5) Let T be defined 011 L2(1ll) or on L2( - a, a) by the formulae

O( T) = CoOO(Ol) or O( T) = COOO( - a, a)
I
Tf = -:-1' for f E OCT).
1

Then we obviously have Kf E O( T) and

TKf(x) = +ddxf( - x)* = ~i 1'( - x)* = ~i (K1')(x) = KTf(x)

for all J E OCT). Hence Tis K-real, and thus possesses selj-atijoint exten-
sions.
(In Section 6.4 we could prove this only for the case of LiOl).)

EXAMPLE 3. The formula

K(f.. J2) = (ft,N) for (f\,J2) E L2(M) E9 L2(M)
defines a conjugation on LiM)E9 L2(M). The operator T defined by
OCT) = Cooo(M) E9 CoOO(M),

T(f\,J2) = (f2' - Jt) for (f1,J2) E OCT)

is symmetric, since

<T(f.. J2)' (gp g2» = ! J2*g\ dx - ! J;*g2 dx

= - !Ng; dx+ !ft*gzdx = <(f},J~), T(g},g2»

for all (f\, J2), (gl' g2) E OCT). Moreover, T is obviously K-real. Therefore,
T possesses self-adjoint extensions.

Theorem 8.10. Let T be a closed symmetric operator on the complex Hilbert

space H with equal Jinite deject indices (m, m). IJ T} and T2 are selj-atijoint
extensions oj T, then (z - T1)-1 - (z - T 2 ) - , is oj rank at most m Jor every
z E peT}) n p(T2). (ThereJore, it is in Bp(H) Jor all p > 0.)

PROOF. Every z E peT,) n P(T2) obviously belongs to I'(T); consequently,

R(z - T)1- is m-dimensional. Since (z - T})- '1= (z - T 2)- '1= (z - T)- '1
for J E R(z - T), we have
(z- T,)-I - (z- T 2)-1 = (z- T,)-'-(z- T 2)-I)p,
8.2 Construction of self-adjoint extensions 237

where P denotes the projection onto R(z - T)l... Hence, dim R«z - T1)-1
-(z-T2)-I)<m. 0

EXERCISES

8.1. For every self-adjoint operator T on a complex Hilbert space there exists a
conjugation K for which T is K-real.
Hint: Use a spectral representation of T (Theorem 7.18) and the natural
conjugation on Ell" L2(1R, p",).
8.2. Let K denote the natural conjugation on L2{M). If T is a K-real self-adjoint
operator on L2(M) and (z- T)-l is a Carleman operator for all zEp(T)
(cf. Exercise 6.12) with kernel kAx, y), then kz(x, y) = k.(y, x) almost every-
where in M X M.

8.2 Construction of self-adjoint extensions

In this section we wish to give the explicit construction of the self-adjoint
extensions of a symmetric operator with equal defect indices. For a closed
symmetric operator T on the complex Hilbert space H we set
N+ =N(i- T*) =R(-i- T)l..,
N_ =N(-i- T*) =R(i- T)l...

Theorem 8.11 (The first formula of von Neumann) .. Let T be a closed

symmetric operator on a complex Hilbert space. Then
O(T*) =O(T)';" N+ .;.. N_ (direct sum),
T*Uo+g++g-) = Tfo+ig+ -ig_ for foEO(T),g+ EN+,g_ EN_.

PROOF. Since N+ C O(T*) and N_ c O(T*), we obviously have OCT) +

N+ + N_ c O(T*). We show that we have equality here, i.e., every
f E O( T*) can be written in the form f = fo + g + + g _ with 10 E O( T),
g + E N+, and g _ E N _. To this end, let f E O( T*). Then by the projection
theorem we can decompose (-i - T*)f into its components in N+ and in
N:/: = R( - i - T),
(-i- T*)f = (-i- T)fo + g, (-i- T)fo E R( - i - T),g E N+.
Since T*fo = Tfo and T*g = ig, we then have (with g + = ig/2)

T*U - fo - g +) = T* f - Tfo +~g = - if + ifo - ~ g

= -iU-fo)+ig+ = -iU-fo-g+).

If we set g _ = f - fo - g +, then g _ E N_ and f = fo + g + + g _.

238 8 Self-adjoint extensions of symmetric operators

It remains to prove that the sum is direct, i.e., that the relations
0= fo+ g+ + g-,fo E D(T), g+ E N+, and g_ E N_ imply J;)=g+ =g_ =0.
It follows from the equality 0 = fo + g + + g _ that
o= T*(Jo + g + + g _) = Tfo + ig + - ig _.
We obtain from this that

and
( - i - T)Jo = 2ig + ;
consequently, g _ E N_ n R(i - T) = {OJ, g + E N+ n R( - i - T) = {OJ.
Therefore, g _ = g + = 0, and thus fo = 0, also. 0

Theorem 8.12 (The second formula of von Neumann). U?t T be a closed

symmetric operator on a complex Hilbert space.
(a) T' is a closed symmetric extension of T if and only if the following holds:
There are closed subspaces F + of N + and F _ of N _ and an isometric
mapping V of F + onto F _ such that
D(T') =D(T) + {g+ Vg : gEF+}

and
T'(jo+ g+ Vg) = Tfo + ig - iVg
= T*(jo+g+Vg) for foED(T),gEF+.
(b) T' is self-adjoint if and only if F + = N+ and F _ = N_.

PROOF. This theorem immediately follows from Theorem 8.6 if we show

that the operator T' of Theorem 8.6 can be represented in the above form.
We have (with V as in Theorem 8.6)
D( T') = R(I - V') = (I - V')D( V') = (I - V')(D( V) + F+)
= (I - V)D( V) + (I - V)F +
=D(T) + {g- Vg: gEF+}.
The sum is direct, as {g - Vg : g E F + } c F + + F_ c N+ + N _. Since T' c
T*, we have moreover that

for all fo E D(T) and g E F +. The assertion follows by taking V= - V. 0

As long as the subspaces N+ and N_ are known, this theorem enables

us to determine all closed symmetric extensions (in particular, all self-
adjoint extensions) of a symmetric operator.
8.2 Construction of self-adjoint extensions 239

Let T and T' be linear operators such that T C T'; then we say that T' is
a finite-dimensional (m-dimensional) extension of T if the quotient space
D(T')/ D(T) is finite-dimensional (m-dimensional). We also say that T is a
finite-dimensional (m-dimensional) restriction of T'.

Theorem 8.13. Let T be a closed symmetric operator on a complex Hilbert

space, and let T' be a symmetric extension of T.
(a) T' is an m-dimensional extension if and only if F+ (defined in Theorem
8.12) is m-dimensional.
(b) If T has defect indices (m, m), then a symmetric extension T' of T is
self-acfioint if and only if T' is an m-dimensional extension of T.
PROOF.
(a) As D(T') = D(T) + (J + V)F+ is a direct sum, we have
dim D(T')/D(T) = dim(l+ V)F+. Since VF+=F_cN_, F+cN+
and N_ n N+ = {O}, we obviously have dim(I + V)F + = dim F +.
(b) T' is an m-dimensional extension if and only if dim F + = dim F _ = m.
Since F+ C N+ and F_ C N_, this holds if and only if F+ = N+ and
F_=N_. 0
An operator Ton H is said to be maximal symmetric if we have T= A
for every symmetric operator A such that T CA. As the closure of a
symmetric operator is symmetric, every maximal symmetric operator is
closed.

Theorem 8.14.
(a) A closed symmetric operator T is maximal symmetric if and only if at
least one of its defect indices is equal to O.
(b) Every self-acfioint operator is maximal symmetric.
(c) Let T be a closed symmetric operator with equal finite defect indices.
Then every maximal symmetric extension of T is self-adjoint.
PROOF.
(a) By Theorem 8.12(a) we can construct proper symmetric extensions if
and only if both defect indices are different from zero.
(b) This follows from (a), since for every self-adjoint operator both defect
indices are equal to 0.
(c) An extension T' is maximal symmetric if and only if it has the form
given in Theorem 8.12(a) with F+ =N+ or F_ =N_. Dimensionality
arguments then show that F+ = N+ and F_ = N_, so that T' is
self-adjoint. 0
EXAMPLE 1. Consider the operator T 1• O of Section 6.4 (cf. Theorems 6.29
and 6.31) defined on LiO, 00) by
- - I -
D(Tl.O) = {jEW2. 1(O, 00) :f(O)=O} and T1.of= Tf' for fED{T 1. o)·
~40 8 Self-adjoint extensions of symmetric operators

We have Tto= Tto= T t , where

O(Tt) = W2. 1(0, 00) and Ttl = -;-11 J' for f E O(T I).

Then N+ = N(i - T t ) is the set of those solutions of the differential equa-

tion if - (f'/i) = 0, hence of the differential equation f + J' = 0, that lie in
L2(0, 00). As the solutions of this differential equation are given by
f(x) = ce- x , we have N+"" {O}. The subspace N_ = N( -1 - T I ) is the set
of those solutions of the differential equation - f + J' = 0 that lie in
L 2(O, 00). Consequently, N_ = {O}. Therefore, the defect indices of T t • o are
different, and thus T t • o possesses no self-adjoint extension.

EXAMPLE 2. Consider the operator T,.o from Section 6.4 (cf. Theorem 6.31)
defined on L2(a, b), - 00 <a <b < 00 by the formulae

~) = {J E W2• I(a, b) : f(a) = feb) =O}

and
-
TI.of = T1J' for f E
0(-
T 1• )· O

We have Tt. 0= Tt 0 = T t , where

1
O(T,) = W2 , t(a, b) and Ttl .. -;-J' for f E D(TI).
1
We also have
N+ = N(i - T t ) = L(e+) with e+(x) = exp(b - x),
and
N_ = N( -i - T t ) = L(e_) with e_(x) = exp(x - a).
The defect indices are therefore equal, and thus T I • o has self-adjoint
extensions. We want to construct these extensions. It is obvious that
lIe+ II = lie_II, so that all unitary mappings of N+ onto N_ are given by
formula
V,,(ce+) = cei"e_ forall cEC (0 <,'} < 271').
Consequently, all self-adjoint extensions S" of T I • o are given by

It is usual and convenient to describe the domains of differential

operators as the restrictions of the maximal operators (here T\) with the
aid of boundary conditions.
8.2 Construction of self-adjoint extensions 241

(8.6) We have

D(SiJ) = {JED(TI) :f(a)=8(1'Jo)f(b)},

where 8(1'Jo)=(1 + eiiJeb-o)-I(eb-o + e iiJ ). The mapping 8: [0, 2'11)~C is
bijective as a map from [0,2'11) onto the unit circle.
PROOF. For every f= fo+ ce+ + ceiiJe_ E D(SiJ) withfoE D(T1. a) we have

f(a) ce + (a) + ceiiJe _ (a)

e b - o + eiiJ
f ( b) = ce + ( b) + ce e _ ( b) = I + e iiJ eb- 0
iiJ •

Consequently,J(a)=8(1'Jo)f(b). Now letf(a)=8(1'Jo)f(b); then we have

(J-ce+ -ceiiJe_)(a) = f(a) - c(eb-o+e iiJ ) = 0,
(J- ce+ - ceiiJL)(b) = 8(1')-lf (a) - e(l +eiiJeb- o)
= 8(1')-lf (a) - f(a)8(1'Jo)-1 = °
with c = f(a)(e b - o + eiiJ)-I. Hence f - ce + - eeiiJe_ E D(T 1. a), and thus f E
D(S,,). It is clear that 8 is a bijective map of [0, 2'11) onto the unit circle. 0

(8.7) The eigenvalues i\" and the normalized eigenelements J.. of S" are
given by the formulae

i\" = (a-b)-I(a+2'11n),
fn{x) = Cn exp(i,\x), (n EZ)
where a is chosen so that e ia = 8( 1') and the cn are normalizing factors.
PROOF. i\ and f are an eigenvalue and a corresponding eigenelement of S"
if and only if

i\f= ~f'.
1
and f(a) = 8(1')f(b).

As aU solutions of the equation Aj= l' Ii have the form f(x) = ceiAx , we
obtain from the boundary condition that eW, = 8(1'Jo )ei.\b. Therefore, eiA(o-b)
= 8(1') = e ia . It follows from this that
i\(a - b) = a modulo 2'11;
consequently,

and
J..(x) = cn exp(ii\"x) for n E Z. o
In the reasoning of Section 8.1 K-real symmetric operators played an
important role; they possess self-adjoint extensions. Now we can show that
they also possess K-real self-adjoint extensions.
242 8 Self-adjoint extensions of symmetric operators

Theorem 8.1S. Let T be a K-real symmetric operator on the complex Hilbert

space H. For every K-real self-adjoint extension T' of T there exists an
orthonormal basis {e", : a E A} of N + such that

V'( ~ c",e",) = ~ c",Ke", for ~ !c",1 2 < O:l

holds for the Cayley transform v' of T. If {e", : a E A]- is an arbitrary
orthonormal basis of N +, then the unitary operator

V: N+ ~N_. V(~ c",e",) = ~ c",Ke", for ~ Ic",1 2 < 00

induces a K-real self-adjoint extension of T in the sense of Theorem 8.12.

PROOF. If {e", : a E A} is an orthonormal basis of N+, V is defined as in
the theorem, and T' denotes the self-adjoint extension of T defined by V,
then

K(Jo + ~ c"'( e", + /(e",») = Kfo + ~ c;( e", + Ke",) E O( T')

and
T' K(Jo + ~ c"'( e", + Ke",») = TKfo + ~ c;(ie", - iKe",)

= K(Tfo+ ~ c"'( -iKe.. +ie",»)

= KT'(Jo+ ~ ca(ea + Kea»)

for all fo + ~ ca(ea + Ke",) E O(T). Consequently, T' is K-real.

Let T' now be a K-real self-adjoint extension of T, and let V' be the
Cayley transform of T.l With the aid of Zorn's lemma we can show the
existence of a maximal orthonormal system {e", : a E A} in N+ with the
property
V'( ~ cae",) = ~ caKe", for ~ !c",1 2 < 00.
Then the formulae
D(S) =O(T) + {~c",(e., + Kea ) : ~ !c.l<:lO},
Sf = T'f for f E D(S)
define a K-real symmetric extension of T (this can be proved as above). If
we assume that {e", : a E A} is not an orthonormal basis of N+, then there
is a non-zero element f of R( -i - S).L. Then V'f E R(i - S).L, Kf E
R(i - S).L, and KV'f E R( - i - S).L; consequently,J + KV'j E R( -i - S)J..
If f + KV'f = 0, then the orthonormal system {e", : a E A} can be enlarged
by taking the element e = illfll-Y, since then Kf= - V'f and thus Ke = V'e.
If f+'KV'f*,O, then we can choose e= IIf+ KV'fll-I(j+ KV'!). since
V'(j + KV'f) = V'f + V' KV'f = V'f + Kf = K(j + KV'f).
II am indebted to Dr. Jiirgen Voigt for the following proof.
8.3 Spectra of self-adjoint extensions of a symmetric operator 243

Here we have used the fact that V' = (i - T')( - i - T') - I, K(i - T') r=
(-i - T')K, and K(-i - T,)-I = (i - T,)-I(i - T')K(-i - T,)-I =
(i - T,)-IK( -i - T')( - i - T,)-I = (i - T,)-IK, so that V' KV' = K. There-
fore, in both cases we obtain a contradiction to the maximality of the
system {eo : a E A}. 0

EXERCISES

8.3. Assume that T is a symmetric operator on a real Hilbert space H, the space He
is the complexification of H, and K is the conjugation defined in He as in
Exercise 7.25.
(a) The complexification Te of T is symmetric and K-real; Te therefore
possesses K-real self-adjoint extensions.
(b) The K-real self-adjoint extensions S of Te have the form S-(T')c, where
the T' are self-adjoint extensions of T.
(c) Every symmetric operator on a real Hilbert space has self-adjoint exten-
sions.
8.4. A symmetric operator (on a real or complex Hilbert space) is essentially
self-adjoint if and only if it has a unique self-adjoint extension. (The complex
case was considered in Theorem 8.7.)
8.5. Let T be a symmetric operator. If Tn is maximal symmetric for some n E 1'1,
n> I, then T m is essentially self-adjoint for mEN, m 0;; n, and T m .. fm.
Hint: First consider the complex case. f is self-adjoint by Theorem 5.22; the
assumption and the inclusion Tnc f" imply that T"- fn; D(T") is a core of
fm for m 0;; n; therefore, D(Tm) is a core of f m, too.
8.6 (a) If T is a K-real operator, then T$ is also K-rea!.
(b) If T' is a symmetric extension of a K-real operator T, then T' is K-real if
and only if K D( T') c D( T').
(c) Let T be a K-real symmetric operator, and let {e" : a E A}. {fa: a E A} be
orthonormal bases of N+. The operators i\ ~2 : N+~N_ defined by the
formulae

~I( ~ c"e,,) - ~ c"Ke", ~2{ ~ c.,ja) - ~ c"Kja for ~ Ical2 < 00

are equal (i.e., the K-real self-adjoint extensions induced by VI and V2 are
equal) if and only if <e",jfJ> is real for all a, fJ EA.
8.7. Let T be a symmetric K-real operator with defect (I, I). Then every self-
adjoint extension of T is also K-rea!.
Hint: cf. Theorem 8.15.

8.3 Spectra of self-adjoint extensions of

a symmetric operator
In this section we study what can be said about the spectra of the
self-adjoint extensions of a given symmetric operator (with equal defect
indices).
244 8 Self-adjoint extensions of symmetric operators

In what follows T will always be a closed symmetric operator. Let us set

n(T, ;\.) = dim N(;\. - T). If ;\. is an eigenvalue of T, then n(T, A) is the
multiplicity of ;\.. If ;\. is not an eigenvalue, then n( T, ;\.) = O.

Theorem 8.16. If T' is an m-dimensional extension of T, then

dim(N(;\.- T')8N(;\.- T» <; m.
If, in addition, n( T, ;\.) < 00, then n( T', ;\.) - n( T, ;\.) <; m.
PROOF. It is obvious that N(;\. - T) c N(;\. - T'). The fonnula
(N(;\.- T')8N(;\.- T» n OCT) = {O}
implies
(N(;\.- T')8N(;\.- T» + OCT) cD(T').
Therefore,
dim(N(;\. - T')8N(;\. - T» <; dim O(T')/O(T) = m.
If n(T, ;\.) < 00, then the second assertion follows from this. o
Let Hz = N(z - T).l. for every z E K. The operator T obviously maps
N(z-T) into itself. We also have T(HznO(T»cHz, since for allfE
Hzn O(T) and gE N(z- T)
<g, Tf) = <Tg,j) = z*<g,f) = 0,
i.e., Tf E N(z - T).l. = Hz. Consequently, Hz is a reducing subspace of T.
We denote by Tz the restriction of T to Hz, i.e.,
O(Tz ) =Hz n OCT) and TJ = Tf for f E O(Tz).
It is obvious that z - Tz is injective (as we have excluded exactly the null
space). Being a restriction of a symmetric operator, Tz is Hermitian. D(Tz ).
is dense in Hz, since any f E Hz that is orthogonal to D(Tz ) is also
orthogonal to O(T) = O(Tz) + N(z - T). Therefore, Tz is a symmetric oper-
ator on Hz. The operator Tz is closed.
In the following we call S( T) = D< \ 1'( T) the spectral kernel of T. The set
SeCT) = {ZED<: (z-Tz)-I is unbounded or n(T,z)=oo}
is called the essential spectral kernel of T.

Theorem 8.17. Lei T be a closed symmetric operator.

(a) We have S..(T) C S(T) c iii and S(T) C u(T).
(b) If T' is a closed symmetric extension of T, then S(T) C S(T') and
S,,(T) c SiT').
(c) If T' is a finite-dimensional symmetric extension of T, then SeCT')
= SiT).
(d) If T is self-adjoint, then S(T) = u(T) and Se(T) = ue(T).
8.3 Spectra of seIr-adjoint extensions of a symmetric operator 245

PROOF.
(a) If AE SAT), then dim N(A - T) = 00 or (A - T,>.T I is unbounded. It is
clear that in both cases A does not lie in I'(T), i.e., AE SeT). It is also
evident that SeT) c IR, since I'(T) contains the upper and lower
half-planes. (A - T) is not continuously invertible for AE SeT); there-
fore, SeT) c aCT).
(b) The inclusion reT') c I'(T) is evident because of the definition of
reT). Hence, SeT) c SeT'). We show that SeCT) C SeCT'). To this end,
let A E SeCT). If neT', A.) = 00, then A. E SAT'). Consequently, we can
assume without loss of generality that neT, A)" neT', A) < 00, i.e., that
(A. - T>.) - I is unbounded. Then there exists a sequence Un) in D( T>.) for
which 1/ J" 1/ = I and (A - T>.)fn ~o. The sequence (J,,) contains no con-
vergent subsequence, since J". ~f would imply II fll = I and T In. ~Af,
i.e.,jE D(T>.) (as T>. is closed) and (A.- T>.)J=O; this would contradict
the injectivity of A - T>.. Let· P now denote the orthogonal projection
onto the finite-dimensional subspace N(A - T'), and let gn = (1- P)J".
Since P is compact, there exist a subsequence Ut') of Un) and an h E H
for which PJ". ~h. The sequence (gn) with gn. = (l- P)J". is therefore
not convergent, and

(A - T~)gn. = (A. - T')fn• - (A - T')PJ... = (A - T)J... ~ O.

Hence, (A. - T~) - I is unbounded, and thus AE See T').
(c) It is sufficient to prove that SiT') C SeCT), since we already have
SeCT) C SeCT'). Assume that A fI. SAT). Then we have to prove that
A fI. SeCT'). It follows from n(T, A) < 00 by Theorem 8.16 that n(T', A)
<; neT, A) + m < 00. The operator (A - T>.)-I is continuous and closed;
therefore, R(A - T) = R(A - T>.) = D«A - T>.) -I) is closed. Since T' is a
finite-dimensional extension of T, there are finitely many elements
JI> ... ,iJ such that
R(A- T') =R(A- T) + LUI"" ,iJ}.
Therefore, D(A - TD - I = RCA - TD = R(A - T') is closed by Theorem
3.4, and thus the closed operator (A - TD- 1 is continuous. Hence,
A fI. See T').
(d) The equality SeT) = aCT) follows from the characterization of the
spectral points of a self-adjoint operator given in Theorem 5.24. We
show that See T) = oe( T). Let AE See T). If n( T, A) = 00, then AE ae( T).
If (A - T,.,) - 1 is not continuous, then we show that A is an accumulation
point of the spectrum of T. If this were not the case, then there would
be an f > 0 for which (A - f, A+ f) n a( T) c {A}, and with the spectral
family E of T we would then have

II (A - T>.)fIl 2 = II (A. - T)f1l 2 = 1It-AI", IA. '--- tl 2 dIlE(t)fIl2 ;> f211f1l2

246 8 Self-adjoint extensions of symmetric operators

for all fE N(>"- T)J. n D(T)= R(E({>"}»J. n D(T), which contradicts

the discontinuity of (>.. - T,..) -I. Hence we have>.. E a,,( T) in this case,
also, and thus SiT) C ae(T).
Assume that>.. E a,,(T). If >.. is an eigenvalue of infinite multiplicity,
then>.. E S~(T). If >.. is an accumulation point of aCT), then there is a
sequence Un) for which

f" ER(E(>..+;;)-E(>..)+E(>..-)-E(>..-*)), IIfnll = 1.

We have fn E R(E({>"}))J. n OCT) = Hh n D(T) and (>.. - Th)fn =

(>.. - T)fn~O; consequently, (>.. - Th)-I is not continuous, so that>.. E
~(T). 0

Theorem 8.18. Let T be a closed symmetric operator on a complex Hilbert

space with equal finite defect indices. Then all self-a4ioint extensions of T
have the same essential spectrum. If some self-a4ioint extension of T has a
pure discrete spectrum, then all self-a4ioint extensions of T do, too.
PROOF. The first assertion immediately follows from Theorem 8.17(c) and
(d). The second assertion follows from the fact that the spectrum is discrete
if and only if the essential spectrum is empty. 0

Theorem 8.19. Let T be a closed symmetric operator on a complex Hilbert

space with equal finite defect indices (m, m) and assume that
II(>..-T)fll;;. cllfll forall fED(T)
with some>.. E Rand c > 0. Then every self-a4ioint extension T' of T has the
follOWing property: a( T') n (>.. - c, >.. + c) contains only isolated eigenvalues
with total multiplicity ..; m.
PROOF. By the first proposition after Theorem 7.24 we only have to prove
that dim R(E(>.. + c -) - E(>.. - c»"; m for the spectral family E of T'.
Assume that dim R(E(>.. + c -) - E(>.. - c» >m. Since

dim D( T') / O( T) = m and R( E(A + c - ) - E(A - c» c O( T'),

there exists an f E R(E(A + c -) - E(>.. - c» n O(T), fol=O. For this f we
have

fIX-tl<cIA- t I2 dIlE(t)fIl 2
1/2
{ }
cllfll"; II(A-T)fll = <cllfll,

which is a contradiction. o
Corollary 1. Let T be a closed symmetric operator on a complex Hilbert
space with finite defect indices (m, m), and let TI and T2 be self-adjoint
extensions of T. If a(TI) n (a, b) = 0, then a(T2) n (a, b) consists of only
isolated eigenvalues of total multiplicity ..; m.
8.4 Second order ordinary differential operators 247

PROOF. If - 00 < a < b < 00, then T satisfies the assumptions of Theorem
8.19 with A= (a + b)j2 and c = (b - a)/2, since for all f E O(T)

II(A.- T)fll2 = II(A.- T I )f1l2 = i

lA-II ;.c
IA.- tl 2 dllE 1(t)fIl 2 ~ c 211fIl2,

where EI is the spectral family of T I . The assertion therefore follows by

taking T' = T2 • If (a, b) is unbounded, then Theorem 8.19 can be applied
to every bounded subinterval (a', b') c (a, b). 0
Corollary 2. If T is a closed symmetric operator on a complex Hilbert space,
bounded from below with lower bound y and finite defect indices (m, m), and
T' is a self-acfjoinf extension of T, then oCT') n (- 00, y) consists of only
isolated eigenvalues with total multiplicity <m.
PROOF. Theorem 8.19 can be applied with any A< Y and c = y - A., since
II(A-T)fll ~ <f,(T-A)f)lIfll- 1 ~ (y-A)lIfll
for all f E OCT), f =1= O. o
EXERCISE

8.8. Let T be a closed symmetric operator with equal finite defect indices (m, m),
and let T. and T2 be self-adjoint extensions of T with spectral families EI and
E 2• Then
dim R(E2(b -) - Eia» < m + dim R(E.(b -) - E.(a».
Hint: Use Exercise 7.37.

8.4 Second order ordinary differential operators

In this section we would like to apply the results of'Sections 8.1 to 8.3 to
second order ordinary differential operators. This way we obtain part of
the theory of Sturm-LioviIle operators developed by Weyl, Titchmarsh,
and Kodaira. For further results and examples we refer the reader to
Hel/wig [15] and Jorgens-Rellich [20].
Let (a, b) be a bounded or unbounded interval in IR, and let r : (a, b)_
IR be a measurable and almost everywhere positive locally integrable
function (i.e., let it be integrable over every compact subinterval of (a, b».
In the following we consider the Hilbert space L2(a, b, r). This is the space
of (equivalence classes of) measurable functions f defined on (a, b) for
which J:lf(x)1 2r(x) dx < 00. The scalar product on L2(a, b, r) is
<f, g) = fbf(x)*g(x)r(x) dx.
a

We denote the corresponding norm by II . II. The formula Ur : L2(a, b, r)

248 8 Self-adjoint extensions of symmetric operators

~L2(a, b), U,f= rl/'f defines an isomorphism of L 2(a, b, r) onto L 2(a, b);
this shows, in particular, that L2(a, b, r) is a Hilbert space.
First we consider differential forms L of the type

Lf = 1. ( - (pf'), + is!, + i(sf)' + qJ}, (8.8)

r
where the coefficients p, q, r, s satisfy the following assumptions:

(8.9) (a) p, q, r, and s are real-valued continuous functions defined on

(a, b); moreover, p and s are continuously differentiable.
(b) p(x) > 0 and rex) > 0 for all x E(a, b).

L is said to be regular at a if a> - 00 and the coefficients p, q, rand s

can be continuously extended to [a, b) withp(a) >0 and rea) >0. Regular-
ity at b (b< 00) can be defined in a corresponding way. L is said to be
regular if L is regular at a and b. L is said to be singular at a (singular at b,
singular) if L is not regular at a (at b, at a or b).
We now define operators on L2(a, b, r) with the aid of a differential
form L such as the one given by (8.8). The maximal operator T induced by
L is defined by the formulae
O( T) = {J E L2 (a, b, r): f is continuously differentiable,
!' is absolutely continuous on (a, b), and Lf E L 2(a, b, r)}2
and
Tf = Lf for f E D(T).
The minimal operator To induced by L is defined by the formulae
O( To) = {f E O( T): the support of f is compact and contained in (a, b)},
and
Tof = Tf for fED( To)·

Theorem 8.20. Let L be as in (8.8). The operator To is symmetric. If s = 0,

then To has equal defect indices, i.e., To has self-adjoint extensions.
PROOF. The Hermitian character of To follows by integration by parts.
OCTo) is dense, because CoOO(a, b) c OCTo). Therefore, To is symmetric. If
s = 0, then To is K-real for the natural conjugation on L2(a, b, r) (Kf= 1*).
The assertion follows from this by Theorem 8.9. 0
If z E C and g : (a, b)~C is a locally integrable function, then we say
thatf: (a, b)~C is a solution of the equation (L - z)f= g if f is continu-
ously differentiable, !' is absolutely continuous, and (L - z)f(x) = g(x)
2 Since l' is absolutely continuous, P1' is also absolutely continuous. Let (P1')' be the
derivative of P1' in the sense of Appendix AS.
8.4 Second order ordinary differential operators 249

almost everywhere in (a, b). Every solution! of the homogeneous equation

(L - z)f= 0 is obviously twice continuously differentiable, since (pf')' is
continuous in this case.
The solutions of the homogeneous equation (L - z)f = 0 constitute a
two-dimensional (complex) vector space. 3 Two solutions U 1' u2 constitute a
fundamental system (i.e., they are linearly independent) if the modified
Wronskian determinant

"
W(u p U2' x) = p(x) det ( UI(X)
u,(x)
UiX»)
u;(x) = p(x)(U I(X)U2(x) - ul(x)uix»
does not vanish for some (and then for all) x E (a, b).
If g : (a, b)~C is locally integrable and u l , U2 is a fundamental system
for the equation (L - z)u = 0, then the solutions h of the equation (L - z)h
= g are given by the formula

hex) = CIU1(X) + C2U2(X) + ul(x) fXW(UI, uz,y)-I U2 (y)g(y)r(y) dy

c

- U2(X) fX W(UI' u2,y)-l ul (y)g(y)r(y) dy, (8.10)

c

where C E (a, b) and C 1, C2 E C.

For continuously differentiable functions f, g : (a, b)~C we define
[f, g Jx = p(x)(J'(x)*g(x) - f(x)*g'(x» + 2is(x)f(x)*g(x)
for x E (a, b). If, in addition, l' and g' are absolutely continuous, then

fP {f(x)* Lg(x) - (Lf(x»*g(x) lr(x) dx = [f, g] f3 - [f, g]a (8.11)

a

for [a,,B]c(a,b). It follows from this that for f,gED(T) the limits
[J. g]Q = limx--+a+[J, g]x and [j, gh = limx--.b_[j, g]x exist. We have
<1. Tg) - <Tf, g) = [f. g]b - [f, gJa for all f, g E D(T). (8.12)

Theorem 8.21. Let L2,0(a, b, r) be the subspace of those functions in

L2(a, b, r) that vanish almost everywhere near a and b. Then

R(To) = {kEL 2 ,0(a, b, r): fbu(x)*k(x)r(x) dx=O

a

for every solution u of the equation Lu = O}.

PROOF. We denote the subspace on the right hand side by R. For f E D(To)
and for every solution u of the equation Lu = 0 we obtain via integration

3 Concerning the results mentioned here about ordinary differential equations we refer to the
textbooks on this subject, for example, Knobloch-Kappel [23], Chapter I.
250 8 Self-adjoint extensions of synunetric operators

by parts that

fb u(x)*(Tof)(x)r(x) dx = fb(Lu)(x)*f(x)r(x) dx = O.
a a

Therefore, R(To) cR. Now let k E R, and let [a, ,8] be a compact subinter-
val of (a, b) with the property that k vanishes outside [a, ,8]. For c E (a, a)
and c. = C2 = 0 let h be the solution of the equation Lh = k given by (8.10)
for z = O. Then hi is absolutely continuous, and h(x) = 0 for x E(a, a). For
every solution u of the equation Lu = 0 and for every Xo E (a, a), x E (,8, b)

[u, h Jx == [ u, h t - [u, h Jxo

== fX {u(y)*k(y) - (Lu(y»*h(y)}r(y) dy = O.
Xo

As this holds for every solution u of the equation Lu = 0, it follows that

°
h(x) = for all x E (,8, b) (if we choose the solution u for which u(x) = 0
and u'(x) = 1). Therefore, h E D(To) and Toh = k. 0
Theorem 8.22. We have Tti = T. The operator To is essentially self-adjoint if
and only if T is symmetric. Then To = T.
PROOF. Integration by parts shows that To and T are formal adjoints of
each other. To prove that Tti = T, it remains to prove that D(Tti) C D(T).
Letf E D(Tti). Then g = T6f is locally integrable. Let h be a solution of the
equation Lh == g. Then

f\f(x) - h(x»*(Tok)(x)r(x) dy
a

= f b«T6f)(X) - (Lh)(x»*k(x)r(x) dx == 0
a

for all k E D(To). Hence, N(F)::J R( To) for the functional

F: L2,o(a, b,r)~C, l~ fbU(x)-h(x»*/(x)r(x)dx.

a

Consequently, by Theorem 8.21 and Theorem 4.1 we have F- ciF. + C2F2

with appropriate c.' C1 E C and
Fj: L2,o(a, b, r) ~ C, I ~ fb uix)*l(x)r(x) dx for j == 1,2,
a

where u.' U2 is a fundamental system of the differential equation Lu == O.

This implies (compare with the proof of Theorem 6.29) that
f(x) - h(x) = + C2U1(X) almost everywhere in (a, b).
c.u.(x)
Hence,f is a solution of the equation Lf= g. Sincef E L2(a, b, r), it follows
that! E D(T). The rest follows from Theorem 5.20. 0
8.4 Second order ordinary differential operators 251

Theorem 8.23. The deject index 'Y + = 'Y +(To) (y _ = 'Y _(To» is equal to the
number oj linearly independent solutions oj the equation (L + i)u
= 0 «L - i)u = 0) that lie in L2(a, b, r). IJ L is regular, then the deject
indices oj To are equal to (2, 2).
PROOF. We have R(i - To).l = N(i + T) and R( -i - To).l = N( -i + T).
Furthermore, N( ± i + T) is equal to the set of those solutions of the
equation (L ± i)u = 0 that lie in L2 (a, b, r). If L is regular, then every
solution of the equation (L ± i)u = 0 is in L2(a, b, r). Consequently,
dim N( ± i + T) = 2. 0
Hence for the defect indices y + and y _ of To there are only three
possible values: 0, I, and 2. If the defect indices are (0, 0), then To = T is
self-adjoint. If the defect indices are equal and different from zero, then
(y +' Y_) = (y, y) = (I, 1) or (2,2). Consequently, by Theorem 8.l3(b) every
y-dimensional symmetric extension of To is self-adjoint.
Theorem 8.18 and Corollary 2 to Theorem 8.19 immediately imply

Theorem 8.24. Let L be a difJerential Jorm such as in (8.8). All self-adjoint

extensions oj To have the same essential spectrum. IJ To is semibounded, then
all self-adjoint extensions oj To are semibounded.

Theorem 8.25. Let L be a regular difJerential Jorm oj the kind (8.8). Then we
have the Jollowing:
(a) For every J E O(T) the Junctions J and l' are continuously extendible to
[a, b). For J, gE O(T) we have
[J, gL =
p(x)(f'(x)*g(x) - J(x)*g'(x)) + 2isJ(x)*g(x) Jor all x E [a, b]

(b) We have 0(1'0) = {f E O(T) : J(a) = 1'(a) = J(b) = 1'(b) = OJ.

PROOF.
(a) If J E O(T) and g = TJ, then J can be represented in the form (8.10)
with a fundamental system u1' U2 of the equation Lu = O. As the
functions uj and u; are continuously extendible to [a, b), the same
follows also for J from this representation. The rest can be obtained
from the definitions of [. , .la and [. , .h.
(b) There exists a cp E O(T) such that cp(a) = 0, cp'(a) = I and cp(x) = 0 for x
near b (it is enough to choose cp twice continuously differentiable).
Then for every J E O( To) it follows from part (a) that

0= <cp, ToJ> - <Tcp,J> = [CP,J]b - [cp,JL = -p(a)J(a),

and thus J(a) ... O. If we now choose cp such that cp(a) = I and cp'(a) - 0,
252 8 Self-adjoint extensions of symmetric operators

then we find thatf'(a)=O, too. We can show similarly thatf(b)==f'(b)

= O. Since To C T, we therefore obtain that

o(To) cOo = (JEO(T) : f(a)=1'(a)=f(b)=1'(b)=O}.

Let To be defined by the formulae O( To) = Do and Tof = Tf for
f E OCTo). Then by part (a)
<f, Tg) - <Tof, g) = [I, g]b - [f, gt = 0
for all f E O( To) and g E O( T), i.e., To and T are formal adjoints of
each other. Hence, To C T* = To, and thus Do c o(ro)' 0

We perform the construction of self-adjoint extensions only for the case

s = O. Hence, in the following we only consider Sturm-Liouville differential
forms

Lf = rI {- (p1')' + qJ}, (S.13)

where p, q, and r satisfy the assumptions (S.9(a» and (S.9(b». In this case
To always possesses self-adjoint extensions by Theorem S.20.
We have (djdx)W(u p u2 ,x)=0 for any two solutions u"u 2 of the
equation (L - z)u = 0, as can be easily verified. Therefore, W(u" u 2 , x) is
constant in (a, b). We briefly write W(u" u 2) for this value.
For any two continuously differentiable functions f, g : (a, b)~C and
x E (a, b) we have now

[f, g t = p(x)(J'(x)*g(x) - f(x)*g'(x» = - W(J*, g, x). (S.14)

Corresponding assertions hold in the regular case for x = a and x = b,

respectively.

Theorem 8.26. Let L be a regular Sturm- Liouville differential form as in

(S.13).
(a) The formulae

f( a) cos a - l' (a) sin (( = 0 )

O( Ta. (J) = { f E OCT): feb) cos f3 - f'(b) sin I~ = 0 '

Ta. tJ = Tf for f E 0( Ta. (J)

define a self-adjoint extension of To for arbitrary a, f3 E [0, 'IT).4

4The boundary conditions occurring here are called "separated boundary conditions", since
every boundary condition affects only one boundary point. There are also "mixed boundary
conditions" that define self-adjoint extensions of To (cr. Exercises 8.10 and 8.11).
8.4 Second order ordinary differential operators 253

(b) For every Z E p(Ta, p) the resolvent R z = (z - Ta, p)-I has the form
Rzg(x) = W(ua, ub)-J{ ub(x) ~xua<y)g(y)r(y) dy

+ ua(x) ~b Ub(y)g(y)r(y) dy},

where Ua and ub are non-trivial solutions of the equation (L - z)u = 0 that

satisfy the boundary condition at a and b, respectively (hence, for
example, ua(a) = sin a, u~(a) = cos a, ub(b) = sin /3, and u;'(b) = cos /3).
(c) The operators Ta. p have pure discrete spectrum. Every eigenvalue is
simp/e.
PROOF.
(a) We can verify easily that Ta,p is symmetric. Hence, it is su.!.ficient to
show that Ta , p is an (at least) two-dimensional extension of To' To this
end, let us choose fila and fIIb from O( T) in such a way that we have
fIIa(a) = sin a, fII~(a) = cos a, fIIa(X) = 0 near b,
fIIb(b)=sin/3,fII;'(b)=cos/3, fIIb(X) =0 near a.
These elem~nts obviously lie in- O( Ta , p) and are linearly indep'~ndent
modulo OCTo)' i.e., no non-trivial linear combination lies in OCTo).
(b) The functions ua and ub are linearly independent, since otherwise ua
would fulfill the boundary conditions at a and b, and z would be an
eigenvalue of Ta , p, which would contradict the relation z E p( Ta , p)'
Therefore, W( ua ' Ub) ~ O. Let K be the integral operator given in the
theorem. If we define Ur as at the beginning of this section, then

( )!
U,KUr - 1 is an integral operator on L2( a , b) with kernel

k x, y =
W(u", ub)-lrl/2(x)Ub(X)u,,(y)rl/2(y)
I
W(u a, ub ) - rl/2(x)ua(X)Ub(y)rl/2(y)
for
for
x >y,
x <yo

The function k obviously belongs to L2 [(a, b) X (a, b)], i.e., K is a

Hilbert-Schmidt operator. Hence, K belongs to B(L 2(a, b, r». For all
g E Lia, b, r) we have
Kg(x) = cu,,(x) + W(u", Ub)-I { ub(x) ~x u,,(y)g(y)r(y) dy

- ua(x) ~x ub(y)g(y)r(y) dy} (8.15)

with c = W(ua, Ub)-IJ:Ub(y)g(y)~(y) dy. Therefore, by (8.10) Kg is a

solution of the equation (z - L)u = g. We can infer from (8.15) that for
g E LZ,o(a, b, r) the function Kg is a multiple of ua in a neighborhood
of a and a multiple of Ub in a neighborhood of b. Consequently,
Kg E O(Ta, p). The operator K therefore coincides with R z on
Lz.o(a, b, r). Since L2 • 0(a, b, r) is dense and since the operators K and
Rz are continuous, it follows that K = Rz •
254 8 Self-adjoint extensions of symmetric operators

(c) R z is a normal and injective Hilbert-Schmidt operator (cf. the proof of

part (b». Consequently, there exist an orthonormal basis {in : n E N}
in Lla, b, r) and a null-sequence (zn) for which z" *0 (n EN) and

RJ = L z"<J.. ,J)J,, for all f E Lia, b, r).

"EN

It follows from this that

Ta.tJ= L (z-z,,-l)<f.. ,f)f.. forall fED(Ta.p)·

.. EN

Consequently, o( Ta. p) = {z - zn- 1 : n EN}. As every solution of the

equation (A - Ta. p)u = 0 is determined by the boundary condition at
one boundary point up to a constant factor, every eigenvalue is simple.
Hence, o( Ta. p) = 0d( Ta. p). 0

Let us now turn to the singular case (more precisely, the not necessarily
regular case).

Theorem 8.27 (The Weyl alternative). Let L be a Sturm-Liouville differen-

tial form defined on (a, b), and let c E (a, b). Either every solution u of the
equation (L - z)u = 0 lies in L2 ( c, b, r) for every z E C or for every z E C
there exists at least one solution u of the equation (L - z) u = 0 for which
u f/. L2(c, b, r). In the second case, for every z E C\ IR there exists (up to a
factor) exactly one solution u of the equation (L - z)u = 0 for which u E
L2(C, b, r).

According to H. Weyl we say in the first case that we have the limit
circle case (LCC) at b; in the second case we say that we have the limit
point case (LPC) at b. The terminology can be explained from the original
construction of Weyl (cf. H. Weyl [56]; cf. also Hellwig [15] and Jorgens-
Rellich (20)). A corresponding theorem holds for the boundary point a.
The limit circle case at a and limit point case at a are defined similarly.

PROOF. In order to prove the alternative, it is sufficient to show the

following: If there exists a Zo E C such that u E L2(C, b, r) for every solution
u of the equation (L - zo)u = 0, then this holds for all z E C. Let VI' V 2 be a
fundamental system of the equation (L - zo)v = O. (We can assume, with-
out loss of generality, that W( VI' V2) = I.) We have (L - zo)u = (z - zo)u
for every solution u of the equation (L - z)u = O. It follows from this by
(8.10) that

u(x) = c\v\(x) + C2V2(X)

+ (z - Zo) IX {V\(X)V2(y) - V2(X)V\(y)} u(y)r(y) dy.
e
8.4 Second order ordinary differential operators 255

As V E L2 ( c, b, r), there exists a dE (c, b) such that

~b v(x)2r(x) dx < (2M)-I.
Consequently, for all XI E(d, b)
f dX, lu(xWr(x) dx .;;; 2c 2 fX' 2
d v(x) r(x) dx

+ M ~Xlv(x)2r(x){ ~Xllu(YWr(y) dy} dx

.;;; 2c 2fb v(x)2r(x) dx + 4f X' lu(yWr(y) dy,
d c
and thus
fX'lu(xWr(x) dx < 4c 2fb v(x)2r(x) dx + fdlu(yWr(y) dy.
d d c
This implies that u E L2(d, b, r), and thus u E L2(c, b, r), as well, since u is
continuous on (a, b).
Let us now assume that we have the limit point case at b, i.e., that for
every z E C there exists (up to a constant factor) at most one solution u of
the equation (L - z)u = 0 such that u E L2(c, b, r). It remains to prove that
for every z E C\ IR there exists at least one solution with this property. For
this we consider the differential form L on the interval (c, b). The form L
is obviously regular at c. Let To and T be the minimal and the maximal
operators on L2 (c, b, r) induced by L. It is sufficient to show that T is not
symmetric, since this implies that To has positive defect indices, i.e., that
N(z - T) =F {OJ for every z E C \ R In order to prove this we use two twice
continuously differentiable functions CPI' CP2 : [c, b)~C for which
CPI(C) = 1, cpi(c) = 0, CPI(X) = 0 near b,
cpic) = 0, cp;(c) = 1, cpix) = 0 near b,
Then CPI' CP2 E OCT) and
<CPI' TCP2) - <TCPI' CP2) = - [CPI' CP2],. = p(c) =1= 0,
i.e., T is not symmetric. o
Auxiliary theorem 8.28. Let L be a Sturm-Liouville differential form on
(a, b).
(a) [j, g]" =Ofor fE OCTo) and gE OCT).
256 8 Self-adjoint extensions of symmetric operators

(b) Let us have the limit circle case at a. If u is a solution of the equation
(L - z)u =0 for some z E C, the function Uo is twice continuously dif-
ferentiable on (a, b), a~d we have uo(x) = u(x) near a and uo(x) = 0 near
b, then uoE D(T)\D(To)'
(c) If we have the limit point case at a, then [J, g]a'" 0 for all f, g E D(T).
Corresponding assertions hold for b.
PROOF.
(a) Assume f E D(To) and g E D(T). Then there is a go E D(T) such that
go(x) = g(x) near a and go(x)'" 0 near b (proofl). Therefore,

[f, g t ... [1, got ... - {<f, Tg o) - <Tof, go)} = O.

(b) Uo obviously lies in D(T). If v is a further solution of the equation
(L - z)u ... 0 such that W( u, v) + 0, and Vo is defined analogously to uo,
then we also have v~ E D(T) and
[uo, v3t ... [u, v*t ... - W(u*, v*) ... - W(u, v)* + O.
e
Thus, Uo D(To) by part (a).
(c) We may assume without loss of generality that L is regular at b
(otherwise we consider L on (a, c) with some cE(a, b». Then the
defect indices of To are (I, I) by Theorems 8.27 and 8.23. Let u l ' u2 be
linearly independent solutions of the equation Lu = 0, and let VI' V2 be
twice continuously differentiable functions for which tJ(x)", uix) near
band tJ(x)'" 0 near a. By part (b) the elements VI' V2 belong to D(T)
and are linearly independent modulo D(To). Consequently, D(T)=
D(To) + L{ VI' V2}' This implies that for arbitrary f, g E D(T) there are
elements f o, go E D(To) that coincide with f and g, respectively, in a
neighborhood of a. It follows by part (a) that [J, g]a ... [fo, gO]a ... O. 0

We are now in a position to give self-adjoint extensions of To in the

singular case, as well.

Theorem 8.29. Let L be a Sturm-Liouville differential form (8.13). Moreover,

let XE R, and let v and w be real solutions of the equation (L - X)u ... O.
(a) The operator Tv. w defined by the formulae

. [v, 1] a = 0 if we have the LCC at a }

D(T w) = { fED(T). ,
v. [w,f]b'" 0 if we have the LCC at b
Tv . ..J... Tf for f E D( To. w)s
define a self-adjoint extension of To.

5 If we have the limit point case at a and/or b, then the index v and/or)l, has no significance.
cr. also footnote 4.
8.4 Second order ordinary differential operators 257

(b) For z E C\~ the resolvent R z = (z - T". w)-I is of the form

Rzg(x) = W(ua, Ub )-l{ Ub(X)~xUa(y)g(y)r(y) dy

+ ua(x) fb Ub(y)g(y)r(y) dy},
x

where ua and ub are the solutions of the equation (L - z)u =0, uniquely
determined up to a factor by the conditions

[v, uaL = 0 if we have the LCC at a, respectively

ua E L2 (a, c, r) if we have the LPC at a,
[w, ub ] b = 0 if we have the LCC at b, respectively
ub E L 2 (c, b, r) if we have the LPC at b.

(c) If we have the limit circle case at both a and b, then Tv. w has a pure
discrete spectrum.
(d) All eigenvalues of Tv. ware simple.

PROOF.
(a) If we have the limit point case at both boundary points, then

<f, Tg) - <Tf, g) = [1, g]b - [f, gL = 0 for all j, g E D(T)

by Auxiliary theorem 8.28. Consequently, T is symmetric and thus

self-adjoint by Theorem 8.22. This is the assertion in this case.
If we have the limit circle case at a and the limit point case at b, then
the defect indices are (1, I), as immediately follows from Theorem 8.27
(the Weyl alternative). If Vo is a twice continuously differentiable
function for which vo(x) = v(x) near a and vo(x) = 0 near b, then
Vo E D(T)\ D(!sJ) by Auxiliary theorem 8.28(b). Moreover, we obvi-
ously have D(To) + L{ vol c D(Tv , w)' If U is a solution of the equation
(L - A)U = 0 that is linearly independent of v and if Uo is defined
analogously to v o, then Uo fl D( Tv,~. Therefore Tv. w is a proper restric-
tion of T, and thus-E( Tv, w) = D( To) + L{ vol (since T is a two-dimen-
sional extension of To). With the aid of this representation of D(Tv . w)
we can immediately see that T". w is symmetric. Hence, T". w is a
one-dimensional symmetric extension of To, and thus it is self-adjoint.
If we have the limit circle case at both boundary points, then the
defect indices are (2, 2). We can show in an entirely analogous way as
in the case just treated that Tv. w is a two-dimensional symmetric
extension of To. Consequently, Tv. w is self-adjoint.
258 8 Self-adjoint extensions of symmetric operators

(b) If we have the limit point case at a (respectively at b), then ua

(respectively Ub) is determined up to a factor. If we have! the limit circle
case at a and U I , u2 is a fundamental system of the equation (L - z)u =
0, then because of the equality [v, CUI + dU 2 ]a = c[v, utla + d[v, u2 ]a'
there is at least one non-trivial linear combination ua = CUI + dU 2 for
which [v, ua]a = O. On the other hand, we do not have [:0, uda = [v, u2]a
= 0, since otherwise there would exist at least one non-trivial solution
of the equation (Tv,w-z)f=O, which contradicts the relation zEC\!R
C p(Tv, w)' We can handle the boundary point b in a similar way.
Let Ko be the restriction, to L2 ,o(a, b, r), of the integral operator
given in the formulation of our theorem. As in Theorem 8.26(b), we
can show that Ko coincides with the restriction of Rz to L2 ,o(a, b, r). If
K is the maximal integral operator defined by the formula in our
theorem, then K is a closed operator (as UrKUr- 1 is a Carleman
operator on L2(a, b». Since Ko is continuous and densely defined, K is
also continuous, and K= R z •
(c) This can be proved in exactly the same way as in Theorem 8.26(c),
because Ua ' Ub E Lia, b, r).
(d) If we have the limit point case at at least one boundary point, then the
assertion is clear, since the space of solutions of the equation (X - T)u
=0 is at most one-dimensional (Theorem 8.27). Let us now assume
that we have the limit circle case at both boundary points. Assume that
X is an eigenvalue of multiplicity 2 (the mUltiplicity cannot be greater),
and let Up U2 be linearly independent eigenelements. Let uj , a and uj , b
be twice continuously differentiable functions for which
uj,a(x) = u/x) near a,uj,a(x) = 0 near b,
Uj,b(X) = 0 near a, Uj,b(X) = uj(x) near b
for j = 1, 2. These are four elements from D(Tv w) that are linearly
independent modulo D(To). This is a contradicti~n to

o
EXERCISES

8.9. Let L be a regular differential form of the form (8.8). The formulae D(TI ) -
{f E D(T) : f(a)-f(b) ... O} and Ttf ... Tf for f E D(T1) define a self-adjoint
extension of To.
8.10. Let Lf(x) - f"(x) + q(x) for x E (0, I). Assume that q is continuously extend-
ible to [0, I].
(a) The operator Ta defined by the formulae
D(T.) = (f E D(T) : f(O) ... ,'}f(I), 1'(0) -= 151'(I)},
T.f - Tf for f E D(T.)
is a self-adjoint extension of To for every ,., E C such that I,'} 1- 1 (the
boundary conditions are mixed).
 8.S Analytic vectors and tensor products of self-adjoint operators 259

(b) Prove, with the aid of (a) (for the case q - 0) that the eigenvalues are in
general not simple in the case of mixed boundary conditions.
. 8.11, Let L be a Sturm-Liouville differential form with limit circle case at a and b.
(a) If A is real, and v, ware real linearly independent solutions of the
equation (L - y)u - 0, then the formulae D( To, w) .. D( To) + L {v, w},
T(o, w)f- Tf for f E D(T(u, w) define a self-adjoint extension of To.
(b) We have D(T(u, w» = (f E D(T) : [V,JJb - [v,JJa" [W,fJb -[w,na =O}
(these are mixed boundary conditions),
8.12. The representation of R z given in Theorem 8.29(b) also holds for z ERn
p(To, w)'
8.13. (a) Let L be a Sturm-Liouville differential form on (a, b) such that
(q(x)/ r(x»;;. y for x;;. xo. If g f£. L1,(xo, b, r) for g(x) ... f~op(S)-1 ds,
then we have the limit point case at b.
Hint: Consider that solution u of the equation (L - y)u = 0 for which
u(xo)'" u'(xo)'" l.
(b) Let L be a Sturm-Liouville differential form on (0, I) such that p(x) =
rex) == I and q(x);;. cx- 2 with c;;. 3/4. Then we have the limit point case
at O.
8.14. Let L be a differential form of the form (8.8) with p == l.
(a) Consider the unitary operator U defined on L2(a, b, r) by the formula
(Uf)(x)=exp(-if~s(t) dt)f(x). Then

ULU-1g(x) - r(~) (-g"(x)+(q(x)-s2(x»g(x)}.

(b) To has equal defect indices.

8.15. (a) The self-adjoint extensions of To given in Theorem 8.26 and Theorem
8.29 are K-real (Kf- r).
(b) The operators Tf from Exercise 8.10 are not K-real for f f£. R.
(c) If we have the limit point case at at least one boundary point, then
Theorem 8.29 provides all K-real self-adjoint extensions of To. (These are
all self·adjoint extensions of To by Exercise 8.7.)

8.5 Analytic vectors and tensor products of

self-adjoint operators
With the aid of the results of Section 8.1 we can also prove the criterion of
Nelson for the essential self.adjointness of symmetric operators. For this
we need the notion of analytic vectors.
Let S be a symmetric operator on the Hilbert space H. Introduce the
notation Coo(S)= n:.oD(S"). An elementfE Coo(S) is called an analytic
vector of S if there exists a t(f) > 0 such that

f
11-0
~
n.
IIS'111 < 00 for It I < tU)
260 8 Self-adjoint extensions of ,:ymmetric operators

Theorem 8.30. If S is a self-adjoint operator and f is an analytic vector of S,

then

for every z E II( such that Izl < t(f).

PROOF. Let E denote the spectral family of S. Then

{f:lezSI2 dIlE(s)fIl 2 } 1/2 = Ilf:ezs dE(s)~1

=llf:n~o (~r dE(S)~1

..; ~ Izl,n IlfM sn dE(s) III
n.
n=O -M JI

< i: Izl,n IIS'111

n- O
n.

for every M >0 and z E II( such that Izl <t(f). Letting M tend to 00, we
obtain that f E D( eZs ). Furthermore,

eZSf = f: ezs dE(s)f = f: n~o (~r dE(s)f

= n~o :~ f:sn dE(s)f+ f:n=~+1 (~r dE(s)f

m nOOn
=
,
I1m ~ z sn'f
"-' --, ~ Z S"j
= "-' --, ,
m ... oo n-O n. n-O n.
since

Ilf OO

-oon-m+1
~ (zs( dE(S)fll =
n.
lim IlfM
M ... oo -Mn-m+1
~ (zs( dE(S)fll
n.

= lim
M-+oo
~ I
II n-m+ z; fM sn dE(S)fll
n. -M

= i:
n-m+1
Iz~nIlS'1I1~O
n.
as m~oo. 0
8.5 Analytic vectors and tensor products of self-adjoint operators 261

Theorem 8.31 (Nelson). Let T be a symmetric operator on the Hilbert space

H. Assume that the set of analytic vectors of T is dense. Then T is essentially
self-adjoint.
PROOF.
(a) First we consider the complex case. Introduce the notations R = H EB H
and T = T EB ( - T) (i.e., D( T) = D( T) EB D( T) and T(j, g) = (Tf, - Tg)
for (j, g) E D( T». If ('Y +, 'Y _) denote the defect indices of T, then the
defect indices of Tare ('Y + + 'Y _, 'Y + + 'Y _). Therefore, f possesses
self-adjoint extensions. We can also see from the defect indices of f
that T is essentially self-adjoint if and only if f is essentially self-
adjoint. Hence, it is enough to prove that T is essentially self-adjoint.
The set of analytic vectors of T is dense: If f and g are analytic
vectors of T, then (j, g) is an analytic vector of f. Consequently, it is
sufficient to show the following (where we write T instead of T): If T
is a symmetric operator with equal defect indices and a dense set of
analytic vectors, then T is essentially self-adjoint.
Let S be a self-adjoint extension of T. Iff is an analytic vector of T,
then f is also an analytic vector of S. Because the formula II S" eissfll =
lIe issS111 = II S111 = II T111, the element eis".! is also an analytic vector of
S for every s E ~, and t(e i' S!) = t(j). Then for every s E ~, z E C such
that Iz - sl < t(j) and for every g E H we have by Theorem 8.30 that
f E D(ei(z-s)S) =D(eizS )
and

Consequently, F is holomorphic in {z E C : 111m zl < t(f)}.

If g E R(i - T).l. = N( - i - T*), then (T*yg = ( - ilg for all n E t\lQ.
Therefore, (T1, g) = (-il(f, g). This implies for all n E t\l o that
F(")(O) = (g, (is)''f) = (g, (iT)"!) = (-l)"(g,j) = (-I)"F(O).
Hence, for all s E ~

(g, eissf) = F(s) = e-'F(O) = e-S(g,f).

Since e- isS is unitary, the function F is bounded. Consequently,
(g, f) = 0 for every analytic vector f of T. Hence, R(i - T).l. = {OJ.
The equality R( - i - T).l. = {OJ follows similarly.
(b) Let H now be a real Hilbert space. Then the set of analytic vectors of
the complexification Tc is dense, namely it is equal to the complex
linear hull of the analytic vectors of T (observe that the set of analytic
vectors is a vector space). Tc is therefore essentially self-adjoint. Then
T is also essentially self-adjoint by Exercise 5.32(b). 0
262 8 Self-adjoint extensions of symmetric operators

Corollary. A closed symmetric operator Tis self-acijoint if and only if the set
of analytic vectors of T is dense.
PROOF. If the set of analytic vectors of T is dense, then T is essentially
self-adjoint by Theorem 8.31. As T is closed, it is self-adjoint.
Let T now be self-adjoint, and let E denote the spectral family of T.
Then all elements of R(E(t) - E( - t» are analytic vectors of T for every
t> 0, since II T"l'1I "tNllfll for f E R(E(t) - E( - t». As U r>oR(E(t)-
E( - t» is dense in H, the assertion follows. 0

The above results enable us to prove the essential self-adjointness of

tensor products of operators. Let HI and H2 be Hilbert spaces, and let us
consider the space H= HI ®H2 (cf. Section 3.4). If TI and T2 are operators
on HI and H2, respectively, then we define the operator TI (8) T2 on HI ®H2
by the formulae

and

(TI ® T2)C~) c/./g;~) = j~l cjTtfj ® T2~'

In order to prove that this is a linear operator, it is sufficient to prove that
this definition is independent of the representation of the: elements from
D(T)® D(T2) as linear combinations of simple tensors; the linearity then
follows directly from the definition. In order to prove this, we have to show
that };j_tc)j®gj=O implies };'J_tcjTtfj®T21J.i=O. By (3.3) we have
};j_ I Cij ® IJ.i = 0 if and only if this sum can be written as a finite linear
combination of elements of the form

Then };j_tcjTtfj® T 2gj is also a linear combination of I~lements of the

same form, and thus it is equal to zero.
In what follows we study, for two given operators T t and T 2 , the
operators
A = TI ® T2 and B = TI ® 12 + II ® T 2 •
We have D(A) = D(B) = D(T)® D(T2).

Theorem 8.32. Let HI> H2 , T t , T 2 , A and B be as above.

(a) A is different from zero (i.e., there exists an f E D(A) such that Af=l=O) if
and only if T) and T2 are different from zero. If A is different from zero,
then A is bounded if and only if T t and T2 are bounded. Then IIA" =
II TtIlIl T211·
(b) If D(B) =1= {OJ, then B is bounded if and only if T t and T2 are bounded.
Then liB II "IITd' + IIT2 11.
8.S Analytic vectors and tensor products of self-adjoint operators 263

(c) If TI and T2 are densely defined, then A and B are also densely defined,
and A '" ~ Tt ® T! and B'" ~ Tt ® 12 + II ® Tr If TI and T2 are symmet-
ric, then A and B are symmetric.
PROOF.
(a) If TI and T2 are different from zero, then there are elements 11 E D( T I)
and f2 E D( T2) for which T.1I:I= 0 and Tzf2:1= O. Hence A (f, ® f0 0, *
i.e., A =1= O. If one of the operators T, and T2 is zero, then A U I ® f0 =
T.1I ® Tzf2 = 0 for all fl E D(T,) and f2 E D(T2). Because of the equal-
ity D(A) = LU, 0f2 : f, E D(T,), f2 E D(T2)}' it follows from this that
Af=O for alllE D(A), i.e., that A =0.
Assume now that A:I= 0 is bounded. Then for all 11 E D( T I ), f2 E
D(T2 ) such that IIf.! I = IIf211 = 1 we have
II T.1.!1 II Tzf211 = II T.1, ® T2f211
IIA (f, ® f 2)11 " IIA II. (8.16)
=

As T2:1= 0, there is an f2 E D( T2) such that II f211 = 1 and T212:1= O.

Therefore, it follows that

Consequently, T. is bounded. We can prove in just the same way that

T2 is bounded. The left side of (8.16) assumes values arbitrarily close to
II Till II T211; thus II T,IIII T211 " IIA II.
Now let TI and T2 be bounded. We show that A is bounded and
IIAII" II TtlIIIT211. Because of the formula A =(T.®I2)(II®T2) it is
sufficient to prove that the operators TI 0 I z and II ® T2 are bounded
and II TI ® 1211 " II T.!I and 1111 ® T211 " II T211· We show the first inequal-
ity. To do this, we use the fact that according to (3.5) every element
fED(T 1)®D(/2)=D(T1)®H2 can be written in the formf="i.'J_lh®
tj with orthonormal elements fJ, j = 1, ... , n. For such an f we have

II(T.®I2)1l1 z =11 t
J=I
T.Jj®fJI1
2
~ IITI/i1l 2
=J-I
n
" II TtlI2 ~ IIhll 2 = II Td1 211f1l2.
j-I

(b) If T. and T2 are bounded, then the bounded ness of T. ® 12 and I. ® T2

follows as in (a). Therefore, liB II "II T .II+II TzlI· If TI is unbounded
and D(T2 ) is different from zero, then there exist a sequence Un) from
D(TI) for which IIJ"II = I and II Tlfnll~oo and agE D(Tz) for which
II gil = l. Then
IIB(J,,®g)1I ;;. IIT.1n ® gll-Ilfn ® Tzgll = IIT.J"II- IIT2 gli ~ 00

as n~oo. Hence, B is unbounded. The same follows if T2 is un-

bounded and D(TI ) is different from zero.
264 8 Self-adjoint extensions of symmetric operators

(c) We can easily verify that A and 'Ii ® 11 (respectively B and 'Ii ® 12 +
I, ® TV are formal adjoints of each other. This implies that A *::>
Tr ® 11 and B*::> 'Ii ® 12 + I, ® 11. The last assertion immediately
follows from this. 0

Theorem 8.33. Let T, and T2 be essentially self-acijoint operators on H, and

H2, respectively. Then the operators A = T, ® T2 and B = T, ®I2 + I, ® T2
are essentially self-acijoint on H = H, ®H2.
PROOF.
(a) First we assume that T, and T2 are self-adjoint. Then A and Bare
symmetric according to Theorem 8.32(c). We construct total sets of
analytic vectors for A and B. As the linear combinations of analytic
vectors are again analytic vectors, the assertion follows from this by
Theorem 8.31.
First we consider the operator A. Let M, and M2 be the sets of
analytic vectors of T't and Ti, respectively. We show that all f,®f2
such that Ii E "1 are analytic vectors of A. If
t"
I ,1I1J2%1I < 00
00
for j = 1, 2 and 0.; t < to'
"-0 n.
then
00 tn
"~O n! IIA "U, ®f2) II
t"
= I ,II Trf, ® TU211
00

n-O n.
00 tn
= n~o n! IITrJ,III1 T2'f2 11
00 In 00 In } '/2
.;
{
n~o n! II TrJdl 2n~o n! IIT2'i2111
00 In 00 In } II;!
= {
I ,(Tr"f,,J,> n-O
n-O n.
I -n•l (Ti"f2,f2>

.; IIfdl'/2I1f211 1/ 1{ ~o :~ IIT't"fdl n;o :~ II Ti"f211) '/2 < 00

for 0 .; I < to' Consequently, f, ® f2 is an analytic vector of A. As the "1

are dense subsets of H.J (j = I, 2), the set of these analytic vectors is
total in H, ®H2 •
Now we consider B. Let M, and M2 now be the sets of analytic
vectors of T, and T2 , respectively. Assume that for Ii E: "1 we have
00 tn
I ,1I1J%1I < 00 for j = 1,2 and 0.; t <to'
n-O n.
8.S Analytic vectors and tensor products of self-adjoint operators 265

<00
for 0 .;;; t < to' Consequently, fl 0 f2 is an analytic vector of B. The set of
these analytic vectors is total in HI ®H2.
(b) Let TI and T2 now be essentially self-adjoint, i.e., let i l and i2 be
self-adjoint. The operators A and B are symmetric, and thus closable.
We can verify easily that

A= T I 0 T2 -:J r; 0 r;,
jj = T I®I2 +I10T2 -:J fl 012 + 1 10 f 2.
The essential self-adjointness of A and B now follows from part (a) of
the proof. 0

Theorem 8.34. Let TI and T2 be self-adjoint on HI and H2 with spectral

f:
families Eland E 2, respectively. Then

<11 0f2' E(t)(gI0g2» = <11, EI(t-s)gl) d s <f2' Eis)g2)

for the spectral family E of jj = TI 012 + 11 0 T2 and for all fl' gl E

HI' f2' g2 E H2•

PROOF. The formulae FI(t) = EI(t) 012 and F 2(t) = 1 1 0 Ell), t E IR define
spectral families on H= HI ®H2. We show this for F I :
(a) Because of the formulae Ft = (EI 012)* -:J Et 012=EI 012 we have
Ft = Fl' It is obvious that FI(t) FI (s)f = FI(s)FI(t)f = FI(t)f for f E
HI 0 H2 and t ';;s. Due to continuity arguments (cf. Theorem 8.32(a»
this holds for all f E HI ®H2. In particular, FI(t? = FI(f), i.e., all FI(f)
are orthogonal projections.
266 8 Self-adjoint extensions of lIymmetric operators

(b) We have FI (t)F\ (s) = F\(s)F\(t) = F\(t) for I <;; s by part (a); this proves
that F\ is increasing.
(c) For f= f\ ®f2 we have
IIFI(I + E)f - F\(/)fll = II(EI(t + E) - EI (t»flllllf211 --l> 0 as E -+ 0 + .
This proves the right continuity of FI(/)f for .all f E HI ® H2
(cf. Theorem 4.23(b».
(d) For f= fl ®f2 we have F\(t)f= E\(t)fl ®f2-+0 as t-+ - 00 and
F\ (t) f --l>f as t -+ 00. This implies the corresponding assertion for all
f E HI ®H2 (cf. Theorem 4.23(b».
The spectral families F\ and F2 commute, since
FI(t)Fis)f = EI(t)fl ® E2(S)f2 = F2(s)F I(I')f
for all f= fl ®f2. This then holds for all f E HI ®H2• Consequently, the
equality
G(t+is) = F\(t)F2(s), s, t E IR
defines a complex spectral family on HI ®H2• We obviously have G(t + is)
= E\(t) ® E2(S). We show that

B= G(u) = f (Re Z + 1m z) dG(z),

where u is the function defined by the formulae u : C-+Dl, u(z) .... Re z +

1m z. As Band G(u) are self-adjoint (u is real-valued), it is enough to
prove that Be G(u). It will follow from this that Be G(u), and thus
B = G(u). For f= fl ®f2 E D(Tt)® O(T,.) we have

jlu(z)1 2 dIlG(z)fI12 <;; 2f {(Re z)'- + (1m Z)2} dIlG(z)fIl2

= 2{f(Rez)2 dIlG(z)fIl 2+ f(Imz)2 d IlG(Z)fIl 2}

= 2{ f t 2 dIlE (t)flI1 Ilf211 2+ f

t 1 Sl dIlE2(s)f2111IlfdI 2}
... 2{1I Tdtll211f2112 + IIfd1 211T2f2112} <: 00.

Consequently,! ED( G(u» and

G(u)f = f (Re z + 1m z) dG(z)f = f Re z dG(z)f + JIm z dG(z)f

= (ftdEI(t)ft)®f2+fl®(fsdE,.{S)fl) = Bf·
If we consider linear combinations of such elements, then it follows that
Be G(u).
Now we can give the spectral family E of B ... G(u). We have
E(/) = G( (z E C : Re z + 1m z <;; I» ... G(X{zec: Ru+lmz <I})'
8.S Analytic vectors and tensor products of self-adjoint operators 267

E( . ) is obviously a spectral family. For every f E HI ® H2 we have

ft2dIlE(t)fIl2 == f(Rez+lmz)2 dIlG(z)fIl 2 •

Consequently, D(E(id» =0 D( G(u» = D(ii). For all f E D(ii)

Jt dE(t)f = f u(z) dG(z)f = Bf.

Hence, E(id) = ii, i.e., E is the spectral family of Ii.

By Fubini's theorem we have for all fl ® f2 E HJ ® H2 that

IIE(t)(fJ®f2)1I 2 == iRez+lm z<t

dll G(z)(fJ ®f2)112
= fOOoof~:S d"IIE J(u)fdI 2 d.IIEis)f211 2
= fOOoo"EJ(t - s)fdl 2 d.IIE2(s)f2112.

The assertion follows from this by means of the polarization identity. 0

Theorem 8.35. Let T J and T2 be self-a4joint operators on complex Hilbert
spaces, and let B = T J® 12 + II ® T2. Then

exp(itii) = exp(itT.)®exp(itT2 ).
PROOF. For all simple tensors f®g such that f E D(T.) and g E D(T2) we
havef®gED(B), and thus

:t [exp(itii)(f®g)] = iii exp(itii)(f®g).

Hence, u(t) = exp(itii)(f® g) is a solution of the initial value problem
d -
dt u(t) = iBu(t), u(O) = f ® g.

On the other hand, it is easy to prove that

u(t) = [exp(itTI)f] ® g + f ® [exp(itT2 )g]
is also a solution of this initial value problem. The solution is uniquely
determined by the Corollary to Theorem 7.38. This proves the assertion. 0

EXERCISES

8.16. Let T be a self-adjoint multiplication operator on L2(R, p) (respectively ori

EB" eA L2(R, Pa». Give a dense set of analytic vectors of T.
8.17. There are essentially self-adjoint operators whose sets of analytic vectors are
not dense.
Hint: Let T be defined by the equalities O(T)-CoOO(R), Tf-(I/i)f'+qf,
268 8 Self-adjoint extensions of symmetric operators

where q(x) = 0 for x .;; 0, and q(x) = I for x > O. If f is an analytic vector of T,
then f (0) = O. If the set of analytic vectors of T were dense, then the same
would hold for the operator Toe Twith D(To) = (IE CoOO(R) :f(O)=O}; the
defect indices of To are equal to (I, I).
8.18. Let Tl and T2 be the operators of multiplication by the variables on
EB aEA L2(1R, Pal and on EB IlEs L2(R, all)' respectively.
(a) TI ® T2 is the operator of multiplication by X I X2 on EBaEA,IlESL2(1R2,
Pa X on)·
(b) TI ® 12 + II ® T2 is the operator of multiplication by Xl + X2 on
EBaEA.IlESLiIR2, p" X all)'
(c) Prove Theorem 8.34 with the aid of (b) and the spectral representation
theorem.
8.19. (a) If Tl ® T2 is different from zero, then Tl ® T2 is symmetric if and only if
there exists a e E K, e -=1= 0 for which eTI and e - IT2 are symmetric.
(b) Tl ® 12 + II ® T2 is symmetric if and only if there exists a e E IR for which
TI - icJ l and T2 + ieh are' symmetric; we have TI ® 12 + II ® T2 =
(TI - ieII)® 12 + II ®(T2 + ieI2). (If HI and H2 are real Hilbert spaces,
then Tl ® 12 + II ® T2 is symmetric if and only if TI and T2 are symmet-
ric.)
8.20. Let TI and T2 be self-adjoint. Then for B = TI ® h + II ® T2 we have
o(B) = {AEIR: A=AI+A2,Aj Eo(1)},
op(B) = {AEIR: A=A. +A2, AjEOp(1)}
The multiplicity n(A) of an eigenvalue A of B is equal to ~.\, +.\2_.\n.(AI)n2(A~,
where ~(Aj) is the multiplicity of the eigenvalue Aj of 1).
8.21. Assume that T 1, T2 and B are as in Exercise 8.20, T2 has a pure point
spectrum, and Pa denotes the projection onto N( T2 - s). Then
E(t) = ~ E.(t - s)®Ps ,
sEop (T2)

where E and EI denote the spectral families of Band T., respectively.

8.22. Let T. and T2 be self-adjoint operators with spectral families E. and E 2. If
l!.·.(t)E 2(s)'" E 2(s)E.(t) for all t, s E IR, then the operators T. + T2 and T. T2
are essentially self-adjoint.
Hint: For all bounded intervals J. and J 2 the set R(E.(J.)E2(J2» consists of
analytic vectors of TI + T2 and TI T 2·
 9
Perturbation theory for
self-adjoint operators

Here we will deal almost exclusively with the perturbation theory for
self-adjoint and essentially self-adjoint operators. Essentially two questions
arise:

(9.1) Let T be a self-adjoint or essentially self-adjoint operator on the

Hilbert space H. Let V be a symmetric operator, a perturbation. Is T + V
also self-adjoint or essentially self-adjoint?
(9.2) Let T be a self-adjoint operator and assume that we know certain
properties of its spectrum. Can we say anything about the spectral properties
ofT+ V (or T+ V)?

We already answered question (9.1) in Section 5.3. Now we tum our

attention to question (9.2). We shall study the question of whether T and
T + V have the same essential spectrum and whether the semi-bounded-
ness of T implies that of T + V. Moreover, we obtain results concerning
the perturbation of the discrete spectrum and concerning the continuous
dependence, on the given operator, of the spectrum and the spectral
family. Further results of this kind are also included in Section 7.5. For the
absolutely continuous spectrum, see also Chapter 11.

9.1 Relatively bounded perturbations

First we consider the case where the unperturbed operator T is bounded
from below.
Theorem 9.1. Let T be self-adjoint and bounded from below with lower bound
'IT' Let V be symmetric and T-bounded with T-bound < 1. Then T + V is
269
270 9 Perturbation theory for self-adjoint operators-

self-adjoint and bounded from below. If

II Vfll .;; allfll + bllTfIl for all f E D(T)
with some b < I, then

'I = Yr - max{ I ~ b' a + blYrl}

is a lower bound of T + V.
PROOF. By Corollary 2 to Theorem 7.22 it is sufficient to show that
(-00, 'I) is contained in p(T+ V), i.e., that the operator T+ V-A=
(T - A) + V is bijective for every A < y. By Theorem 5.11, this is surely the
case if II VR(A, T)II < 1. We obtain from the spectral theorem that
II VR(A, T)II .;; aIlR(A, T)II + bll TR(A, T)II
<: a(Yr-A)-1 + b sup{ltl(t-A)-1 : D'Yr}
= a(Yr-A)-1 + b max{l, IYrl(Yr-A)-I}

= max{ a(Yr-A)-1 + b, a(Yr-A)-1

+ bIYrl(Yr- A)-I}.
The last expression is obviously less than 1 for A <'I. o
Theorem 9.2. Let T be self-adjoint and bounded from below, and let V be
symmetric and T-bounded. If T + p. V is closed for all p. E [0, 11, then T + V is
self-adjoint and bounded from below_
PROOF. The operator T+ V is self-adjoint by Theorem 5.27. For every
11
the operator V is relatively bounded with respect to T + p. V, i.e.,
° °
p. E [0,
there exist al-' > and bl-' > for which
II Vfll .;; al'llfll + bl'lI(T+ p.V)fll·
Consequently, for I p. - p./1 < (2b,,)-1
IlVfll .;; al-'llfll + bl'(II(T+/LIV)JII+I/L'-p.IIIVfll)
< al'lIfll + bl-'II(T+ p.'V)fll + 411 Vfll,
and thus
IlVfll < 2al-' II fll + 2b"II(T+ /L' V)fll-
The segment [0, I] is covered by the open intervals (p. - (2b,,) - I, P. +
(2bl')-I), p.E[O, I]. Consequently, there are finitely many 1-'-1"'" P.n for
which the corresponding intervals cover the whole interval [0, I]. The
operator V is therefore (T + P. V)-bounded for all /L E [0, I] with relative
bound b=max{2bllj :j= I,. _., n}. If we choose mEN such that b/m<
I holds, then by successive applications of Theorem 9.1 we obtain the
semi-bounded ness of T+(1/m)V, T+2(1/m)V, . .. , T+ V. 0
9.1 Relatively bounded perturbations 271

Theorem 9.1 also enables us to prove the following useful inequalities.

Theorem 9.3 (Heinz). Assume that T is self-atijoint, non-negative, S is

symmetric, D(T) C D(S), and IISfl1 ..;; II Tfll for all f E D(T). Then
1<1, Sf) 1.;,; <1, Tf) for all fED{ T).

PROOF. Theorem 9.1 applies with V = KS for every K E ( - 1, 1) if we take

a= 0, b = IKI, and 'IT = O. Then 'I = O. Consequently, T + KS is self-adjoint
and non-negative for every K E (-I, 1). For K~ ± 1 we get

<f, (T+ S)f) ;;;. O} for all f ED{T),

<1, (T- S)f) ;;;. 0
and thus
<f, Tf) ;;;. 1<1, Sf)1 for all f E D{T). o
Theorem 9.4. Let Sand T be self-atijoint and non-negative.
(a) D(T) C D(S) and II Sfll .so; II Tfll for all f E D(T) imply D(TI/2)
C D(SI/2) and IISI/111 ..;; II T I/111for a/l f E D(T I/ 2).
(b) D(T) c D(S) implies D(TI/2) C D(S 1/2). The equality D(T) = D(S) im-
plies D(TI/2) = D(S 1/2).

PROOF.
(a) It follows from Theorem 9.3 that
liS 1/1112 = <J,Sf)';'; <f, Tf) = IITI/1112 forall fED{T).

Let f E D(T 1/2). Since D(T) is a core of TI/2, there is a sequence (f,,)
from D(T) for whichf,,~f and Tl/1,,~Tl/1. Then (SI/1n) is also a
Cauchy sequence. Therefore, f E D(S 1/2) and S 1/1= limn-+""S 1/1".
Consequen tly,
IIsl/111 = n-.oo
lim IISI/:t.1I
n
.so; lim
n~oo
II Tl/:t.n II = IIT I/111.

(b) Because of the inclusion D(T) C D(S), the operator S is T-bounded by

Theorem 5.9, i.e., there exists a c ;;;. 0 such that

IISfll .;,; c(lIfll + II Tfll) .;,; Y2 c(lIfll 2 + II Tf1l2) 1/2

< V2 c(lIfll 2 + 2<f, Tf) + II Tf1l2) 1/2 = V2 cll(I + T)fll.
By part (a), D«J + T)I/2) is therefore contained in D(S 1/2). We can
immediately infer from the spectral theorem that D(TI/2) = D((l +
T)1/2). Consequently, D(TI/2) C D(S 1/2). If D(T) = D(S), then it also
follows that D(S 1/2) C D(Tl/2), and thus that D(SI/2) = D(TI/2). 0

Now we prove a result concerning the continuous dependence, on the

given operator, of !!:~ spectrum and the essential spectrum.
.272 9 Perturbation theory for self-adjoint operators

Theorem 9.5. Let T and Tn (n E N) be self-adjoint, and assume that

OCT) = O(Tn)' Assume, furthermore, that there are null-sequences (an) and
(b n) from R for which
II(T- Tn)!11 <; anllfll + bnll Tfll for all f E ,D(T).

Then u(T) = limn ... ",,u(Tn) and aiT) = limn-+ 00 ui Tn)'

PROOF. We have to prove that A E ae(T) (respectively A E u(T» if and only
if there is a sequence (~) for which ~ E ae(Tn) (respectively ~ E u(Tn»
and An-A. There is no loss of generality in considering only the point

°
A = 0. Let E and En denote the spectral families of T and. Tn'
(a) If E a(T), then there is a sequence (J,,) from D(T) for which IIf,,1I = 1
and Tfn _0. Then
d(O, a(T,,» <; IIT"!nll <; II Tf,,11 + II(T- T,,)f,,11
<; (I + b,,)11 Tf,,11 + a"lIfnll _ ° as n_ 00.

° °
Therefore, d(O, a(T,,»_O, i.e., E limn ... oou(Tn).
If fI. u( T), then T is bijective. Since

II(T- Tn+A)T-'1I <; a"IIT-'1I + b" + IAIIIT-'II,

by Theorem 5.11 T" - A is also bijective for IAI < (211 T-'ID-'
°
and
sufficiently large n E N. Hence, f/.limn-+ooo(Tn)'
(b) Assume that OEae(T). Then dim R(E(£)-E(-t:»=oo for every £>
0, by Theorem 7.24. If no E N is such that an + bn £ <£ for all n ;;> no,
then it follows for all n ;;>no andfE R(E(£)- E(-£)}.j=l=O that
IIT"!II <; IITfll + II(T- Tn)fll <; £lIfll + a,,11111 + b"IITfll
< 2£lIfll·
It follows from this that dim R(E,,(2t:) - En( - 2£» = 0(>, since otherwise
there would exist an

f E R(E,,(2t:) - En( -2t:».L n R(E(t:) - E( -f»~, f "i= 0;

for this f we would have
II T"!II < 2t:llfll·
2t:llfll <;
By the proposition preceding Theorem 7.25 we have
"i= 0 for every n;;> no' As £ > 0 was arbitrary, it follows that E
o~( T,,)
°
n [ - 2£, 2£]

limn-+ooaiTn)'
°
Let us assume that fI. oe(T), i.e., that there exists an £ > 0 such that
dim R(E(£) - E( - (0» < 00. If no EN is such that a" <,,/3 and bn < 1/3
for all n ;;> no, then
IIT,JII ;;> IITfll-II(T- Tn)fll > IITfll- jt:llfll- jllTfli
= ~II Tfll - jt:llfll ;;> j£llfll
9.2 Relatively compact perturbations and the essential spectrum 273

for all n ;> no and f E R(E( f) - E( - f»1. ,f=1= O. It follows from this that
dim R(En{ e/3) - En{ - f/3» < 00, since otherwise there would exist an

f E R( EnOf) - En( - if)) n R(E(f) - E( - f»1., f =1= 0;

for this f we would have

~fllfll;> IIT,j1l >~fllfll.

Consequently, (- f/3, f/3) n oe(Tn) = 0 for all n;> no, and thus 0 t;l:
limn-+oooiTn)' 0
Corollary. Assume that T is self-adjoint, V is symmetric and T-bounded, and
a denotes the set of those real /L for which T+ /LV is self-adjoint. Then the
set-valued functions /L~o(T+ /LV) and /L~oe(T+ /LV) are continuous on a
(i.e., if /Lo,p."EO, and /Ln~/Lo, then o(T+/LoV)=limn-+ooo(T+/LnV) and
0e(T+ /Lo V) = limn-+oo0eC T + /Ln V».
PROOF. V is (T + Po V)-bounded for every 1'0 Ea. Therefore,

1I[(T+/LoV)-(T+/LnV)]fU = l/Lo-ltnlllVfll
< IPo - p."lallfll + I/Lo - IlnlbU(T+ JLoV)fll·
Consequently, the operators T + Ito V and T + Itn V satisfy the assumptions
of Theorem 9.5. 0

EXERCISES

9.1. The converse of Theorem 9.4(a) does not hold. If we consider the operators
induced by the matrices

S = 3U D and T = 4(6 ~)
on C2 (with the usual scalar product), then S..; T, but not S2..; T2.
9.2. Let the assumptions of Theorem 9.5 be satisfied. Assume that a, P E iii n p(T)
and a <po Then a, p EP(Tn) for large nand II(En(P)- En(a»-(E(P)-
E(o:»II_O as n_oo.
Hint: The second resolvent identity, Exercise 7.20, and Theorem 9.5.

9.2 Relatively compact perturbations and

the essential spectrum
By Theorem 7.24 the number A E R belongs to the essential spectrum of a
self-adjoint operator T if and only if there exists a sequence Un) from D(T)
for which
J" ~ 0, lim inf II fn II =1= 0, (A - T)J" ~ O.
n-+oo
(9.3)
274 9 Perturbation theory for scM-adjoint operators

Such a sequence is called a singular sequence for T and 'A. With the aid of
this characterization, we can prove the following theorem.

Theorem 9.6. Let TI and T2 be self-adjoint operators with the spectral

families E, and E 2•
(a) IJ R(EI(J» c D(T2) and (T, - T2)E 1(J) is compact for every bounded
interval J, then (JiT,) C (Je(T2 ).
(b) If the assumptions oj (a) are satisfied and D(T,) C D(T2), then every
sequence that is singular for TI and 'A is also singular for T2 and A.
PROOF.
(a) Assume that 'AE (Je(T,). As in the proof of Theorem 7.24 (part (i)
implies (ii» we can show that there exists a singular sequence Un) for
T, and 'A that is contained in R(E,('A + I) - E,('A - 1» (cf. also part (b)
of this proof). As Un) tends to 0 weakly and as (T, - T 2 ) (E,(A + 1)-
E,('A-I» is compact, we have
(T, - T 2 )Jn = (T, - T2)( E) ('A + 1) - E \('A - 1) )f" ~ 0
as n~oo (cf. Theorem 6.3). We obtain from this that
II(T2 - 'A)J"II <; II(T2 - T\)f" II + II(T) - 'A)fnll ~ 0
as n~oo. Therefore, Un) is a singular sequence for T2 and 'A, and thus
'A E (JiT2)'
(b) Due to the inclusion D(T,)c D(T2 ), the operator T2 is relatively
bounded with respect to T\ (cf. Theorem 5.9). Let (j~) be a singular
sequence for T, and 'A. Then

11(1 - E\('A+ 1) + E)('A-I»f,,11 2 = f (-00,.\-']

+ f('\+',0<:')
1 dIlE,(t)f,,1I 2

<; foo
-00
I'll. - tl 2 dIlE\(t)/nIl 2

(9.4)
and

II T,(1 - E\('A + 1) + E\('A - 1»1,,11 2

= 1(-oo,i\-I]
+ f
(i\+I,oo)
Itl 2 dIlE\(t)1,,1I 2

<; 21'A1211(1 - E1('A + I) + E,('A --1»1,,11 2

+ 211 ('A - T\)1" 112 ~ 0 as n ~ 00. (9.5)

9.2 Relatively compact perturbations and the essential spectrum 275

Hence, «E,(;\+ 1)- E,(;\-l»fn) is a singular sequence for T, and ;\

that is contained in R(E,(;\ + I) - E,(;\ - I». As in part (a), it follows
that

Since T2 is bounded with respect to T" we can derive from (9.4) and
(9.5) that

and thus
(;\ - TZ)jn -l> 0 as n -l> 00.

Consequently, (J,,) is a singular sequence for T2 and ;\. o

If in Theorem 9.6 we make the assumptions symmetrical with respect to
T, and T 2 , then we obtain a criterion for the coincidence of the essential
spectra and for the coincidence of the singular sequences (in the following
we say that T, and T2 have the same singular sequences if (J,,) is a singular
sequence for T, and ;\ if and only if (J,,) is a singular sequence for T2 and
;\). However, the result is not very useful in this form since too many
properties of T, and T2 are explicitly assumed; usually only one of these
operators (the unperturbed one) is known accurately. In what follows we
give conditions that imply the assumptions of Theorem 9.6. First we need
some preparation.
Let H" Hz, and H3 be Hilbert spaces. Let A be an operator from H, into
H2• An operator B from HI into H3 is said to be A-compact if D(A) c D(B)
and B, as a mapping from (D(A), II . IIA) into H3t is compact. If A is
bounded and D(A) = Hit then it is obvious that B is A-compact if and only
if B is compact (since the norms II . 1/ and II . IIA are then equivalent).

Proposition. If A is an AI-bounded operatorand B is A-compact, then B is

also A ,-compact. If A is densely defined and closed, then B is A-compact if
and only if B is IA I-compact.
PROOF. Any II .IIA,-bounded set is also II. II A-bounded. Consequently,
every II . IIA ,-bounded set is mapped by B onto a relatively compact set.
The second assertion follows from the equalities D(A) = D(IAI) and II . IIA
=II·IIIAI· 0

If B is an A -compact operator, then B, as an operator from

(D(A), II . IIA) into H3 , is bounded, i.e., B is A-bounded. In fact, much
more is true.

Theorem 9.7. Let A be an operator from H, into H2, and let B be an

A-compact operator from H, into H3 • If A or B is closable, then B is
A-bounded with A-bound zero.
"276 9 Perturbation theory for self-adjoint operators

PROOF. Let assume that the A -bound B is there exists

an f > 0 with the property that for every n E N there is an fn E D(A) such
IIBJ,.II nil fn II t:IIAfnll can choose any positive number f that is
than A-bound of B). we put liB!" then Bgnll and
f(1I gnll + IIAgnll) < nil gnll + t:IIAgnll < II Bgn II == 1
all n E such that ;> f. follows from this that gn~O and II <
gn +IIAgnll I/f.
First let B be closable. As B is A -compact, there exists a subsequence
of for which Bgn• ~h E H3 • The formula = 1
l. This is a contradiction because gn~O and the c10sability
Now let A be closable. Without loss of generality we can assume that
D(B) = D(A). By Theorem 6.2 the operator B can be extended to an
A-compact operator on D(A). Consequently, wc can assume without loss
of generality that A is closed. Since IIAgnll < liE, there exists sub-
w
sequence (gnt) of (gn) such that Agn• ~h E H2 (cf. Theorem 4.25). Hence,
= w Agn) G(A )= and thus h It follows by
Theorem that Bgnk -'i'0, which contradicts the equality II II = I, 0
Corollary. Let B be closable. B -Nl'mnnrt if and only if it is (A + B)-

B be A -compact. Since D(A) c D( B), we have D(A + B)

By Theorem 9,7 there a;> 0 such that Bfll <allfll +
fE Consequently, all f D(A)
IIAfll < 2(IIAfll-IIBfll + allfll) < 2(II(A + B)fll + allfll),
is B)-bounded. The (A + of follows the
above proposition. can prove similarly the other direction. 0

Let A closed operator HI into H2, and let B be an

HI into Then following assertions are equivalent,
(i) B is A-compact.
(ii) J,. E D(A),J,. ~O and Af" ~O imply that Bfn~O.
If A self-adjoint on HI E denotes its spectral then these
assertions are equivalent
(iii) BE(J) is compact for every bounded interval J, and B is A-bounded with
-bound zero.
w
PROOF. (i) implies (ii): If fn ~0 and ~o, then
(J,., g)A = (fn' g) + (Af", Ag) ~ 0
for all E D(A i.e., weakly tends zero in the: Hilbert space
(D(A), (. , '),4). It follows from this by Theorem 6.3 that Bfn-O.
(ii) implies (i): Let (J,.) be a weak null-sequence in (D(A), (. , ,),4)' i.e.,
assume that (fn' + (At" _0 all (g, E G(A). Theorem 6.3
9.2 Relatively compact perturbations and the essential spectrum 277

is sufficient to show that (Bin) tends to zero as n-4oo. Since (in' g) +

(Aln' h) =0 for all (g, h)E G(A).L, we have
Un' g) + (Aln' h) -40 for all (g, h) E HI EB H2 •
It follows from this that J" ~ 0 and AJ" ~ 0, and thus BJ" -40 (because of
(ii».
(i) implies (iii): By Theorem 9.7 the operator B is A-bounded with
A-bound zero. Let (J,,) be a bounded sequence in HI' Then the sequence
(AE(J)J,,) is also bounded, because

IIAE(J)J,,1I2 = ~t2 dIIE(t)inI1 2 < 11J,,1I 2 sup { t 2 : t E J).

.
Consequently, (E(J)J,,) is a bounded sequence in (D(A), (. , .)A)' and by
(i) there exists a subsequence (f,n) for which (BE(J)ln) is convergent.
Hence, BE(J) is compact.
~

(iii) implies (i): Let (J,,) be a weak null-sequence from (D(A), (., ')A)'
i.e., assume that (In' g) + (Ain' h) -40 for all (g, h) E HI EB HI (cr. the "(ii)
II'
implies (i)" part of the proof). Then in -40, and thus B(E(N) - E( - N»in
-40 for all N > O. Since 'the A -bound of B is equal to 0, for every f > 0
there is a C > 0 such that

II BIll < ~ IIA/II + C II ill for all I E D(A).

Therefore, for all n E N and sufficiently large N

IIB(I - E(N) + E( - N»J"II < ~ IIAJ"II + CII(I - E(N) + E( - N»)Jnll

€ C
< illAJ,,1I + N IIAJ"II < fIlAJ"II,
and thus
lim sup IIBJ"II " f lim sup IIAJ"II.
n-+oo n-.oo

Since the sequence (AJ,,) is bounded and since f > 0 was arbitrary, it
follows that Bln~O. Consequently, B is A-compact. 0
Now we can prove an old result that is essentially due to H. Weyl.

Theorem 9.9. Let T be a self-alijoint operator on the Hilbert space H, and let
V be a symmetric T-compact operator. Then T+ V is self-alijoint, T and
T+ V have the same singular sequences, and t:1e (T) = aeCT+ V).
PROOF. By Theorem 9.7 the operator V is T-bounded with T-bound O.
Therefore, T+ V is self-adjoint by Theorem 5.28. V is also (T+ V)-
compact by the corollary to Theorem 9.7. Now it follows from Theorem
9.8 that TI = T and T2 = T+ V satisfy the assumptions of Theorem 9.6(b),
278 9 Perturbation theory for seH-adjoint operators

just as T. = T+ V and T2 = T do. Consequently, T and T+ V have the

same singular sequences, and 0e( T + V) = 0e(T). 0

In applications (particularly to differential operators) it is important that

the assumptions of Theorem 9.9 can be somewhat weakened. First we
prove some preparatory theorems.

Theorem 9.10. Let T be a selj-acijoint operator on H such that p(T) =1:0, and
let p > O. An operator V is TP-compact (respectively TP-bounded) if and only
if V(z - T)-P is compact (respectively bounded) for some (and then for all)
z Ep(T).

(The operators TP and (z - T)-P are defined with the aid of the spectral
theorem by the formulae TP = ft P dE(t) and (z - T)-P = f(z - t)-P dE(t),
where t~tP and t~(z - t)-P are chosen to be continuous on o(T).)
PROOF. We obviously have D(TP)=D«z- T)P); the TP-norm and the
(z - T)p-norm are equivalent. Consequently, V is TP-compact (TP-
bounded) if and only if it is (z - T)p-compact «z - T)p-bounded). Since
(z - T)P : D(TP)~H is continuous and continuously invertible, V: D(TP)
~H is compact (bounded) if and only' if V(z- T)'-P is compact
(bounded). 0
Theorem 9.11. Let T be a selj-acijoint operator with spectral family E, and let
V be a T-bounded operator. Then
(a) V is TP-bounded with TP-bound zero for all p > 1.
(b) If V is TP-compact for some p > 0, then VE(J) is compact for every
bounded interval J.
(c) If VE(J) ;s compact for every bounded interval J, then V is TP-compact
for every p > I.
(d) V ;s T-compact if and only if it is T 2-compact and T-bounded with
T-bound zero.
PROOF.
(a) There are numbers a, b > b such that II Vfll "allfll + bl! Tfll for all
f E D(T). We have D(TP) C D(T) for p > I, and thus
II Vfll II V(E(N) - E( - N»1I1 + II V(I - E(N) + E( - N»fll
<;
<; all1l1 + bN IIfll + allfll + bll T(I - E(N) + E( - N»fll

<; (2a + bN)llfll + bN I-p II TPfll

for all f E D(TP) and N > O. Since N can be chosen arbitrarily large,
the assertion follows from this.
(b) Let Un) be a bounded sequence. Then (E(J)fn) is a bounded sequence
in (D(TP), (. , ')T')' Since V is TP-compact, there exists a subsequence
(E(J)J,,) for which (VE(J)fn.) is convergent. Hence, VE(J) is compact.
9.2 Relatively compact perturbations and the essential spectrum 279

(c) The operator

V(I+ITl)-P(E(N)-E(-N» = V(E(N)-E(-N»(l+ITI)-P
is compact for every N ;.. 0, and

+ ITI)-P(I - E(N) + E( - N»II .;;; II V(I + ITI)-'II(I + N)I-p.

II V(I
Consequently, Vel + IT\) -P = limN .... co V(l + ITj)-P(E(N) - E( - N»,
and thus Vel + IT\)-P is compact. V is therefore TP-compact by
Theorem 9.10 and the proposition preceding Theorem 9.7.
(d) This follows from Theorem 9.8 and parts (b) and (c) of this theorem. 0

Theorem 9.12. Let TI and T2 be self-ac!Joint operators and assume that

D(TI)=D(T2). Put V= T 2 - T 1•
(a) V is T[-compact if and only if it is Ti-compact.
(b) If V IS T[-compact (or Ti-compact), then (z - T I)-2 - (z - T 2)-2 is
compact for every zEp(T I)np(T2).
(c) If V is T I2-compact, then for every TI-bounded operator W we have the
following: W is T l2-compact if and only if it is T;-compact.
PROOF. Write Rj = (z _1j)-1 for z E p(T I) n p(T2 ). Then the operators VRj
are bounded for j = I, 2. (If H is real and o( T 1) U o( T 2) = IR, then H must
be complexified in order that we may have p( T I) n P(T2) =1= 0.)
(a) If V is T?-compact, then VR I2 is compact by Theorem 9.10. It follows
from the resolvent identity R2 - RI = RI VR 2 = R 2VR I that
Ri = (RI + R 2 VR I)(R I + RI VR 2) = (I + R 2 V)R?(I + VR 2 ), (9.6)
and thus
(9.7)
It follows from this that VRi is compact. Therefore, V is Ti-compact.
We can prove similarly the reverse direction.
(b) Let V be T?-compact. (9.6) implies that

Ri - R? = R[VR2 + R 2VRf(I + VR 2 )
=[(z*- T 2)-I V(Z*- T I)-2]* + R2VR[(I+ VR 2);

here we have used the equality

<f, (z*- T 2 )-I V(Z*- T 1)-2g ) = <RtVR2f,g)
for all J, g E H. As V(z* - T I)-2 and VR? are compact, the compact-
ness of Ri - RI2 follows.
(c) Let W be T I2-compact. Then WRf is compact. It follows from (9.6)
that
280 9 Perturbation theory for self-adjoint operators

As WRt and VR 12 are compact, WRi is also compact, i.e., W is

Tf-compact. We can prove similarly the other direction. 0
Theorem 9.13. Let T be a self-adjoint operator on H, and denote its spectral
family by E. Assume that V is symmetric, O(T) C O( V), and T+ V is
self-adjoint. Assume, furthermore, that VE(J) is compact for every bounded
interval J (this condition can be rep/aced by the following: V is TP-compact
for some p > 0). Then T and T + V have the same singular sequences. In
particular, (J.(T) = (Je(T+ V).
PROOF. It follows from the assumptions that V is T 2 -compact, and thus
also (T+ V)2-compact by Theorem 9.12. Therefore VE'(J) is also compact
for every bounded interval J, where E' denotes the spectral family of
T+ V. The assumptions of Theorem 9.6(b) are therefore satisfied for
T( = T, T z = T+ V and for T( = T+ V, T z = T. If V is TP-compact for
some p > 0, then the compactness of VE( J) follows from Theorem 9.1 I(b)
for every bounded interval J. 0

The assumptions of Theorem 9.13 do not guarantee that T + V is

semi-bounded in case T is semi-bounded.

EXAMPLE 1. Let T be a semi-bounded self-adjoint operator with discrete

spectrum, i.e., assume that there exist an orthonormal basis {en: n E N}
and a sequence (An) for which An ---) 00 and

OCT) = {fEH: L 1~121<en,J)12<00},

nEI'IJ

Tf = L An<en,J)en for f E D(T).

nEI'IJ

Furthermore, write V = - 2 T. Then V is T-bounded and - T = T + V is

self-adjoint. Since the space R(E(J» is finite-dimensional for every
bounded interval J, the operator VE(J) is compact. Consequently, all
assumptions of Theorem 9.13 are satisfied, T is bounded from below, and
T + V is not bounded from below.

The following theorem studies the behavior of gaps in the essential

spectrum of T in the case of a non-negative T2-compact perturbation.

Theorem 9.14. Let T be a self-adjoint operator on H such that Ge(T) n (a, b)

= 0. Assume that the point b is not an accumulation point of those eigenval-
ues of T that belong to (a, b). Assume, furthermore, that V is symmetric,
non-negative, TZ-compact and T-bounded with T-bound < 1. Then Ge( T + V)
n (a, b) = 0, and b is not an accumulation point of those eigenvalues of
T+ V that belong to (a, b). (If V <; 0, then a similar result holds for the point
a.)
9.2 Relatively compact perturbations and the essential spectrum 281

PROOF. We only have to prove the second assertion. We can assume,

without loss of generality, that (a, b) = (-1, 1). We use Theorem 7.25.
Since, by assumption, V is T-bounded with T-bound < 1, there exists a
c ;> 0 for which
IlVfll " c(lIfll + II(T+ sV)fll) for all f E OCT) and s E [0, 1].
Let So E [0, I] be chosen such that I is not an accumulation point of
eigenvalues of T + So V belonging to ( - I, I). By assumption, this holds for
So = 0 in any event. Let Eo denote the spectral family of T + So V. Since
ae< T + So V) n ( - 1, 1) = 0 (cf. Theorem 9.13), Eo(l - ) - Eo(O) is of finite
rank.
For all f E OCT) we have
liVfll" c(lIfll + II(T+soV)fll) " 2c ll(I+IT+soVi)fll·
It follows from this by Theorem 9.3 that
<J, Vf) " 2e<J, (1 + IT+ soVDf) for f E OCT),
and thus
<f, Vf) " 2c<J, (1- T- soV)f) for f E R(Eo(O» n OCT).
If s>so and s-so"(l/4e), then T+sV=(T+soV)+(s-so)V is self-
adjoint and
<J, (T+ sV)f) = <J, (T+ soV)f) + (s - so)(-f, Vf)
" <J, (T + So V)J) + (l / 4c )<J, VJ)

" <J, (T+soV)J) +i<J, (1- T-soV)J)

= ~<J, (T+soV)J) +~llfIl2 "!llfIl 2

for J E R(Eo(O» n O( T). It is obvious that
<I, (T+ sV)f) ;;. <f, (T+ soV)f) ;;. IIfll2
for f E R(/ - Eo(1- »n
OCT).
Since R(Eo(l-) - Eo(O» is finite-dimensional, Theorem 7.25 implies the
following: The interval (~, I) contains at most finitely many points of the
spectrum of T + s V, and thus 1 is not an accumulation point of the
eigenvalues of T + s V from ( - I, I).
If we choose mEN such that m ;;. 4c, and p. = 1/ m, then in this way we
can prove the assertion step by step for T+ p.V, T+ 2p.V, ... , T+ mp.V=
T + V, starting with So = o. 0
REMARK. In Theorem 9.14 the T-boundedness of V with T-bound < 1 is
not necessary. Instead, it is enough to assume that V is T-bounded and
T + s V is self-adjoint for all s E [0, I]. For the proof see the technique used
in Theorem 9.2.
282 9 Perturbation theory for sdf-adjoint operators

EXERCISES

9.3. Let TI and T2 be self-adjoint such that D(TI):J D(T2).

(a) If TI has a pure discrete spectrum, then so does T 2•
(b) If {"1 1)} and {Al>} are the eigenvalues of TI and T 2, respectively (each
eigenvalue counted according to its multiplicity), then IAP>I ;;.aN'>1 + b for
allj, with appropriate numbers a> 0 and b ;;. o.
Hint: Use the equality (i - T2) -I ... (i - T2) -I(i - T1)(i - T I) -I, the
boundedness of (i - T2)-I(i - T I ), and (7.3).
9.4. Let T be self-adjoint, and denote by E its spectral family. If V is T-bounded,
and VE(J) is compact for every bounded interval J, then V is I(T)-compact
for every E-measurable function 1 such that limltl ... oolt- ~(t)l- 00.
9.5. For a normal operator A let us define the essential spectrum c'e(A) to be the set
of accumulation points of a(A) plus the set of eigenvalues of infinite multiplic-
ity. (This definition extends the definition for self-adjoint operators.)
(a) AE ae(A) if and only if there exists a sequence Un) from D(A) such that
lim infn.... oo II In II > 0, In ~ 0 and (A - A)In~O (Un) is a singular sequence for
A and A).
(b) If AI and A2 are normal, D(A 1)=D(A 2), and p(A 1)n,l(A21*0, and if
A 1- A2 is Af-compact for some p > 0, then aiA I) C a,(A21.
(c) Let TI and T2 be normal. If a.«z - TI)-I) = a,«z - T 2)-I) for some
z E p(TI) n p(T2), then a.(TI) = a.(T2). If (z - TI)-I - (,; - T2l- 1 is com-
pact for some z E p( T I) n p( T2), then ae( T I) - ai T2l.
(d) With the aid of (c) and Theorem 8.10 prove that all self-adjoint extensions
of a symmetric operator with finite defect indices have the same essential
spectrum.
9.6. Let Tbe a self-adjoint operator such that (a, b)n a(T)-0 or (a, b)n ae(T)-
O. If V;;. 0 is symmetric, D(T) C D(V), T+ V is self-adjoint, and <I. VI> <:
- <1, TI> + (a + 'IJ)IIIII 2 for all 1 E E(a)D(T) with some 'IJ < b - a, then (a +
'IJ, b)n a(T+ V)-0 or (a+ 'IJ, b)n a.(T+ V)-0, respectively. This holds in
particular if V is bounded, symmetric, and 0 <: V <: 'IJ.
Hint: Use Theorem 7.25.
9.7. Let H be a Hilbert space, and let A be an unbounded linear functional on H.
(A is a non-closable operator from H into K.) A is A-compact; however, the
A -bound of A equals 1. (In Theorem 9.7 the assumption that A or Bare
closable cannot be dropped.)

9.3 Strong resolvent convergence

If Tn{n E 1\1) and T are self-adjoint operators on the complex Hilbert space
H, then we say that the sequence (Tn) converges to T in the sense of the
strong resolvent convergence if (z- Tn)-I~(z- T)-J for some zEC\!R.
Then this holds for all z E C \ R on the basis of the following theorem.
9.3 Strong resolvent convergence 283

Theorem 9.15. Let Tn (n E 1\1) and T be self-a4ioint operators on the complex

Hilbert space H. /f (zo - Tn) - 1 ~ (zo - T) -I for some Zo E C \ R, then
(z - Til) - 1 ~ (z - T) - I for all z E C \ R.
PROOF. If z E C and Iz - zol < 11m zol, then by Theorem 5.14 we have
00
(z- TII)-If- (z- T)-If= ~ (ZO-Z)k[(ZO- Tn)-k-I_(zo- T)-k-I]f
k-O
for all f E H. Therefore,
N
lI(z-Tn)-'f-(z-T)-'fll';;; ~ IZo-zlkll(zo-TII)-k-lf

- (zo- T)-k-'fll +2 ~ Izo - zlklIm zol-k-II/fll

k>N
for every N E 1\1. The second sum will be arbitrarily small if N is chosen
large enough. The first sum tends to 0 for fixed N as n- 00, since
(zo - Tn)-k-I ~ (zo - T)-k-I. Consequently, the assertion follows for all
z E C such that Iz - zol < 11m zol. An iterative application of this step
provides the assertion for all z in the half-plane where Zo lies. The limit
relation (z- TII)-I~(z- T)-I implies that (z*- TII)-I~(z*- T)-I and
(as (z - TII)-I and (z - T)-I are normal) that
lI(z* - T,,)-'fll = lI(z - TII )-lfll - lI (z - T)-'fll = I/(z* - T)-Ifll·
Therefore, for all f E H
lI(z* - Til) -If - (z* - T) - lfll 2
- lI(z* - Til) - lfll 2 - 2 Re«z* - Til) -If, (z* - T) -If) + lI(z* - T) -lfll 2
-1I(z* - T)-lfI/2 - 2 Re«z* - T)-'f, (z* - T)-If) + lI(z* - T)- 'fIl 2
- O.
Hence, the assertion holds for all z E C \ R. o
Now we prove a few sufficient conditions for strong resolvent conver-
gence.

Theorem 9.16. Let TII(n E 1\1) and T be self-a4ioint operators on the complex
Hilbert space H. The sequence (Til) converges to T in the sense of the strong
resolvent convergence if one of the following assumptions is satisfied:
(i) There is a core Do of T such that for every f E Do there exists an no E 1\1
with the properties that f E D( Tn) for n ;> no and TJ - Tf.
(ii) The operators Til and T are bounded and Tn ~ T.
(iii) D(Tn} = D(T} for all n E 1\1 and there are null sequences (all) and (bn)
such that
II(T- T,,)fll " a"lIfll + bllil Tfli for all f E D(T).
284 9 Perturbation theory for se:lf-adjoint operators

(iv) G(T)=limn ..... ooG(Tn), i.e., G(T) is the set of those elements (f, g)from
H EB H for which there exists a sequence (in) such that fn E D(Tn) and
(fn, TJn)~(f, g) (graph convergence).
PROOF.
(i) We have
(i- T,,)-'f- (i- T)-'f
= (i-Tn)-'(Tn-T)(i-T)-'f--l'O as n~oo
for all f E H such that (i - T) - ':! E Do. As Do is a core of T, the set of
these f is dense in H. Therefore, (i - Tn) - , ~ (i - T) -, by Theorem
4.23.
(ii) or (iii) implies (i); these cases are therefore also proved.
(iv) It obviously follows from the formula G(T)=limn ..... ooG(Tn) that
G(i - T) = lim n ..... oo G(i - T,,). Let g E H be arbitrary. Then there is an
f E D( T) such that g = (i - T)f. Furthermore, there is a sequence (fn)
for which /n E D(T,,), f,,~f, and (i - T,,)fn~(i - T)f== g. Due to the
inequality 1I(i - Tn) -III (; 1 it follows from this that

lI(i- T,,)-'g - (i- T)-I gil (; 1I(i- Tn)-Ig -f,,11 + 111. - (i- T)-'gll
== 1I(i - Tn) -I( g - (i - T,,)/n)1I + IIf" - fll
~o.

o
Theorem 9.17. Let T" (n EN) and T be self-adjoint operators on the complex
Hilbert space H, and assume that (i - Tn)-' ~ (i - T)-I. Then u(Tn) ~ u(T)
for every continuous bounded function defined on R.
PROOF. First we assume that the limits lim,±oou(t) exist, and liml-+_oou(t)
= lim, ..... oou(t). These are the functions that can be considered as continu-
ous functions defined on the Alexandroff-compactification IR of R. We
consider the space C(R) with the maximum norm. The polynomials in
(i - t) - I and ( - i - t) -, can be considered as elements of C(R). The set P
of these polynomials has the following properties: (i) the constant func-
tions lie in P, (ii) the elements of P separate the points of R (i.e., for
x, y E Rsuch that x =FY there exists a u E P for which u(x) =Fu(y», and (iii)
if u E P, then u* E P. By the complex form of the Stone-Weierstrass
theorem (cf. Hewitt-Stromberg [18], Theorem (7.34» P is therefore dense in
C(R). Consequently, for every u E C(R) there exists a sequence (um ) from
P such that max'€Rlu(t)- u",(t)I~O as rn~oo.
Now let f E Hand £ > 0 be given. Then there is an rno E N such that
lu(t) - u",(t)1 (; (E/31Ifll) for all t E IR and m;;;' mo. As U",(/) is a polynomial
in (i- t)-I and (-i- t)-I, we have
um(Tn) ~ um(T) as n~ 00
9.3 Strong resolvent convergence 285

for all mEN. Consequently, there exists an no EN such that

lIumo (Tn)] - umJT)jll "i for Ii > no·

Hence, it follows for n > no that
II u( Tn)] - u( T)fll " II u( Tn)] - umo( Tn)]11 + II um/ Tn)] - umo( T)fll
+ lIumo(T)f - u(T)fll " £,
since lIu(Tn) - u'.l'o(Tn)1I ,,(£/3I1fll) and lIu(T) - umo(T)1I " (£/3I1fll).
Therefore, u(Tn) ~ u(T).
Now let u be an arbitrary continuous and bounded function defined on
R Let (IPm) be a sequence of continuous functions with compact supports
that tends to I non-decreasing. Then
IIIPm(T)f - fll2 = f IIPm(t) - WdIlE(t)fIl2 ~ 0 as m~ 00

for every f E H. Because of uIPm E C(R) we have

u(Tn)IPm(Tn) ~ u(T)IPm(T) as n~ 00

for every mEN, by the first part of the proof. For all n, mEN
II u( Tn)! - u( T)fll
" lIu(Tn)] - u(Tn)IPm(T)fll + lIu(Tn)IPm(T)f - u(Tn)IPm(Tn)fll
+ lIu(Tn)IPm(Tn)! - u(T)IPm(T)fll + lIu(T)IPm(T)f - u(T)fll
.;;; lIu(T,,)1I11f - IPm(T)fll + lIu(T)lllIf - IPm(T)fll
+ lIu(T,,)IIIIIPm(T)f - IPm(Tn)]1I + lIu(T,,)IPm(T,,)f - u(T)IPm(T)fll·
The first two terms on the right side will be small for sufficiently large m
(observe that Ilu(Tn)II" sup{lu(t)1 : t E R}). The last two terms will be
small for fixed m if n is chosen sufficiently large. Consequently, the
assertion is proved. D

Now we can prove, in particular, that the unitary group induced by a

self-adjoint operator depends on this operator continuously in the strong
sense.

Theorem 9.18. Let Tn (n E N) and T be selj-acijoint operators on the camp/ex

Hilbert space H. Assume that (i - Tn) - I ~ (i - T) - 1.
(a) eilT• ~ eilT for all t E R
(b) If Til > Y and T > Y for some y E R, then we a/so have e - IT. ~ e - IT for
all t > O.
PROOF.
(a) The function s~eils is continuous and bounded on R. The assertion
therefore follows from Theorem 9.17.
286 9 Perturbation theory for self-adjoint operators

(b) With u(s) = e- Is for s ~ y and u(s) = e- /y for s <y we have u(T) = e- IT
and u(Tn)=e- IT•• As, on the other hand, u is continuous and bounded,
Theorem 9.17 can be applied again. 0

Now we shall investigate the influence of the strong resolvent conver-

gence on the spectral family.

Theorem 9.19. Let Tn(n E N) and T be self-adjoint operators on the complex

Hilbert space H, and assume that (i'- Tn)-l ~ (i - T)-l. If En and E are the
spectral families of Tn and T, respectively, then, as n~oo,

En(t)~E(t)
s for all t E R such that E(t) = E(t - ).
En(t -) ~ E(t)

PROOF. Assume that E(t) = E( t -). Let (<<Pm) respectively (o/m) be non-
decreasing respectively non-increasing sequences of continuous functions
such that «Pm(s)~X( _ 00, I)(S), o/m(s)~X( _ 00, 'Ies), l«Pm(s)l..;; t, and 100m(s)1 ..;; t
for all s E R. Then for all f E H
II( «Pm{ Tn) - En{ t - ) )J1I 2

= fl«Pm(s)-X(-00,/)(sWdIlEn{s)fIl2~O as m~oo
(Lebesgue's theorem). Therefore,
«pm( T,,) ~ En( t - ).
It follows similarly that
o/m(T,,) ~ E,,(t),
and (because E(t) = E(t - »
«Pm(T) ~ E(t) and o/m(T) ~ E(t).

Hence, for every f E H and every t: > 0 there are continuous functions
00, I) and 0/ ~ X{ - 00, I) for which
«P " X( -

1I00(T)f- «p(T)fll < ~;

we can choose «P = «Pmo and 0/ = o/mo with a sufficiently large mo' By
Theorem 9.17 there is an no E N such that
E
II «p( T)f - «p( T,,)fll < "5
for all n ~ no'
1I00(T)f-0/(T,,)fll <~
We therefore have
t:
II 0/( Tn)f - <pC Tn)! II < 3"5 for all n;;> nC).
9.3 Strong resolvent convergence 287

Since
1/2
{
IIE(t)j - cp(T)]1I = jlx<-oo./j(S) - cp(sW dIlE(s)]1I2 }
1/2
..; { jll/l(S)-cp(sW d IlE(S)]1I 2 }
f:
= 1Il/I(T)] - cp(T)]1I ..; '5
and
f:
II En( 1 - )] - cp( Tn)jll ..; II En( I)] - cp( Tn)jll ..; 11"'( T,,)] - cp( T,,)jll < 3'5'
we obtain that
IIE(I)] - En(t)]11
< IIE(t)j - cp(T)j1l + IIcp(T)] - cp(Tn)]1I + IIcp(Tn)] - En(I)]1I
< II l/I( T)] - cp( T)]II + II cp( T)j - cp( Tn)]11 + II cp( T,,)] - l/I( Tn)]11
< f: for all n;> no'
It follows similarly that
IIE(I)] - E,,(t -)jll ..; f: for all n;;;' no.

o
It is worth noting that the results of Theorems 9.5 and 9.19 are not
comparable. It is clear that o(T) = limHOOo(T,,) does not imply E(t) =
S -lim,,->ooEn(t).
Conversely, from En(t) ~ E(t) (for all 1 E R) we cannot infer o(T,,)~
o( T), as the following example shows.

EXAMPLE 1. Let H be a separable infinite dimensional Hilbert space, and

let {em : mEN} be an orthonormal basis of H. For every n E N let Tn be
the orthogonal projection onto L{em : 1 <m <n}. Then Tn~I as n~oo.
The spectral family E" of T" is given by the equality
o for t<O,
En(l) = { 1- Tn for 0<1< I,
I for t;> l.

Therefore, En( t) ~ E(t) for all t E R, where E denotes the spectral family
of I, i.e.,

E(t)={O fort<l,
I fori;> 1.
On the other hand, o(T,,) = {O, I} for all n EN, while 0(/) = {I} (cf. Exer-
cise 7.41).
288 9 Perturbation theory for self-adjoint operators

EXERCISE

9,8, real Hilbert T and Tn(n E 1'1)

(i)-(iv) of
IISSllmtlU0I1S

(a) For the spectral families En and E of Tn and T, respectively, we have

En( t) ~ E( t) for every t E R such that E is continuous at t,
(b) If there exists ayE Ii such that T;;. y and Tn;;' y for all n E 1'1, then
e-'T·~e-IT for all t;;. O.
Hint: Complexify.
 10
Differential operators on L2(IRm)

10.l The Fourier transformation on L2(lR m )

In what follows we shall use so-called multi indices. A multiindex (of m
components) is an m-tuple a = (ai' ... , am) of non-negative integers a) E
l\Io,j= 1, 2, ... ,m. The absolute value of a is defined by the formula

We set, for every x E R m , m

xa = II X/'i.
j=1
Correspondingly, we write
a IT (~~)aJ
D =
j=1 lOX)
= (_i)la l IT ~.
j-Iax?
The space of rapidly decreasing functions (the Schwartz space) S(Rm) is
the vector space of arbitrarily many times continuously differentiable
functions f: Rm_c for which we have the following: For every multiin-
dex a and for every p E 1\1 0 there exists a cap ;;;. 0 such that

IxlPI D'Y'(x) I <;; cap for all x E Rm.

It is obvious that this assumption can also be formulated in the following

way: For every muItiindex a and for arbitrary p, q E 1\10 there exists a
capq ;;;. 0 such that

(I + IxlYI D'Y'(x) I <;; capq(l + Ixl)-q for all x E Rm.

289
290 10 Differential operators on L 2(Rm)

This formulation shows that S{lRm) c t.,,(lRm) for all p E [1, 00]. In particu-
lar, for all / E S(lRm) we can define the Fourier trans/ormation Fo by the
integral

where

m
xy = L :>c.JYj for X,Y E IRm.
)-1

Theorem 10.1. We have FoS(lRm) c S(lRm). For every / E S(lRm) and every
multiindex a
D"Fo! = (-I)'a'FoM,,/, M"Fo! = Fo D"'j,
where (M"f)(x) = x "'j(x).
PROOF. It is easy to see that the function

(Fo!)(x) = (2'1T)-m/2 f e-ixy/(y) dy

is arbitrarily many times continuously differentiable. The differentiation

can be done under the integral sign, i.e.,

D"(Fo!)(x) ... (2'1T)-m/2( -I)''''f e-iXY[y"!(y)] dy.

It follows from this for any multiindex {3 that

x.8(D"Fof)(x) - (2'1T)-m/2( -l)''''+'.8'fDff(e-iX)')[y"'j(y)] dy

= (2'1T)-m/2( -l)''''f e- ixy D.8[y"'j(y)] dy.

Since D.8M,,! is in S(lRm), too, x.8(D"Fof)(x) is bounded for all a and {3.
Therefore, Fo! E S(Rm ), and
MpD"Fo/ = (-l)''''FoD.8MJ.
Both formulae follow from this. D
Theorem 10.2. The function {to : IRm_IR defined by the equality
{to(x) = exp( -4IxI2) for all x E IRm
is in S(lRm). We have Fo{to = {to.

1 If no domain of integration is indicated, then the integral always is to be taken over Rm.
10.1 The Fourier transformation on L 2(R m ) 291

PROOF. The reader can verify I} E To prove that Fol} =

we first consider the case m = 1. In this case we obviously have the first
order differential equation
t'if(x) xt'i(x)
It follows from this by Theorem 10.1 that
+
i.e., Fol} satisfies the same differential equation as I}. Due to the equalities
(Fot'i)(O) (27T)-1/2! I}(x) = 1=
we obtain that = I}. The result can be derived for an arbitrary m by
taking products. o
Theorem 10.3. Fourier transformation Fo bijective linear mapping
S(lRm) onto itself. We have
(FO-lg)(X) = -m I eiXYg(y) 2! g S(lRm).

Moreover, (Fof)(x) = (Fo 1){ - x) for every fE S{lRm), and we have F~-= [.
PROOF. For all i, g E S(lRm)
! g(Y)(Fof)(y)e ixy = (2'1T) m/2 Jg(y)e Je-iy'j(z)
iXY

J J
= (2'1T)-m I 2 J(z) e-i(Z-x)Yg(y) dz

= Ji(z)(Fog)(z - x) dz
= f (Fog)(z)f(z x) dz.

With g.(x)'" g{ t:x) we have for g E S(lRm) and t: > 0 that

(Fog.)(x) J
(2'77)-m/2 e-iXYg(~) dy

J
_ (2'77)- mI2 t:- m e-ixy/cg(y) dy

t:-m(Fog)( ~);
therefore,
f g(9'){F J)(y)e
o ixy dy =

=J(Fogc)(z)f(z + x) dz
= t:- m f (Fog)( -; )i{z + x) dz
f (Fog)(z)f(u x) dz.
292 10 Differential operators on Lz(Rm)

If we replace here g by the function f) from Theorem 10.2, then it follows

for ~~O (as Fo1t=f) and f)(O) = I) that

(27T) - m/2 f eiXY(FoJ)(y) dy = f)(O)(2'IT) -m/2 f eiXY(FoJ)(y) dy

= lim (27T)-m/2feiXYf)(EY)(FoJ)(y) dy
<-->0

= lim (27T)-m/2ff)(y)f(ty + x) dy
<-->0

= (27T)-m/2f(x) f f)(y) dy = f(x).

It follows from this that Fo is injective and Fo- 1 has the given form.
Moreover, for all f E S(lRm)

(FJf)(x) = (27T)-m/2 f e-iXY(Fof)(y) dy = f( - x),

and thus F04f(x) = f(x), i.e., F~ = I. Since R(Fo):J R(F~) = R(I) = S(lR m),
the mapping Fo is surjective. 0
In what follows we consider Fo as an operator on L:!(lRm) such that
O(Fo) = S(lRm).

Theorem 10.4. We have II Fofll = II fll and II Fo- ~II = II fll for all f E S(lRm)
(here II. II denotes the norm in L2(lRm». Fo and FO- 1 possess uniquely
determined extensions F and F belonging to 8(L2(lRm». Th~ operators F and
F are unitary, and F= F* = F- 1. We have F 4 = I. The operator F is called
the Fourier transformation on L2(lRm).
PROOF. For f, g E S(lRm) we have

(f, g) = ff(x)*(F;IFog)(x) dx

dZ] dY} dx
r
= f f(x)*(27T)-m {f eiXY [ f e-iyzg(z)

= f (2'IT)-m {f e-iy-'!(x) dx {f e-V-Zg(z) dZ} dy = (Fof, Fog)·

In particular, II Fofll = Ilfll, and thus IlFo-111 = IIfll· Since S(lRm) is dense
»,
in L2 (lR m ) (as Cooo(lRm) c S(lR m there exist uniquely determined exten-
sions F and F of Fo and F O- 1 from 8(L2(lRm». We obviously have
IIFfll = IIFfll = IIfll for all f E L2(lRm). If f, g E L2(lR m) and Un)' (gn) are
sequences from S(lRm) such thatfn~f and g,,~g, then
(Ff, g) = lim (Fof", g,,)
n---+oo
= lim (fn. Fo-1g,,)
n~oo
= <f, fig),
10.1 The Fourier transformation on L2(R m ) 293

i.e., F* = F. Moreover,
lim FFo- '1" = lim FoFo- '1" = n---+oo
FFJ = n--+oo n~oo
lim fn = j,

i.e., FF= I. We can prove similarly that FF= I. Therefore, F= F-·, and
thus F is unitary. From F~ = I1S(Rm) it follows that F4 = I. 0

The mappings Fo and F.9-· can be extended to L.(lR m) in a natural way.

These extensions F. and FI are defined for f E LI(lR m) by the formulae,

J
(Ftf)(x) = (2'17') -m12 e-ixYf(y) dy

(Ftf)(x) = (2'lT)-m/2 JeixYf(y) dy.

It is easy to see that FI f and PI f are continuous functions such that
I(Ftf)(x)I"; (2'lT)-m/211flll and I(Ptf)(x)I"; (2'lT)- mI2 1Ifll l for all x E IRm.

Theorem 10.5. The mappings FI and P, of L.(lRm) into the space Coo(lRm) of
continuous bounded functions defined on IRm are injective. For fE L.(lRm)n
L2(lR m) we have
(Ftf)(x) = (FJ)(x) and (F.f)(x) = (F-1)(x)
almost everywhere in IRm.
PROOF. Take anJfrom L,(lR m ) for which FIJ=O (i.e., (Ftf)(x) = 0 for all
x E IRm). We have to prove that J = O. It follows from the equality FI J = 0
that

JJ(x)(Fog)(x) dx = (2'lT)-m I 2 f f e-i"Yg(y)J(x) dy dx

= Jg(y)(Ftf)(y) dy = 0
for all g E S(lRm). Since Fog runs over the whole space S(lR m),

f f(x)h(x) dx = 0

for all h E S(lRm). Then this holds also for all continuous functions h
defined on IR m and having compact support. Define K(n) = {x E IR m : Ixl
..; n} for n EN, let x" be the characteristic function of K(n), and define

sgnJ(x) = {~(X)IJ(X)I-I for f(x) =1=0,

for J(x)=O.
Then Xn sgnJ is measurable, Ix,,(x) sgnJ(x)1 ..; 1, and there exists a
sequence (h) of continuous functions with supports in K(n) such that
294 10 Differential operators on L2(R m )

~(x)-sgnf(x) almost everywhere in K(n). We can assume without loss of

generality that Ihix)1 <; I for all x E K(n) (since we can replace h.i(x) by
h.i(x)Ih.i(x)I-1 for those x E K(n) for which Ih/x)1 > I). Consequently,

1
K(n)
If(x)1 dx = .lim f.
J-+OO K(n)
f(x)hJ"(x) dx = O.

Therefore,f(x) =0 almost everywhere in K(n). As this hoMs for all nEN,

we have f(x) = 0 almost everywhere. FI is then injective. The injectivity of
i l can be proved similarly.
If f E L I(lIl m) n L2(R m) and!" = xJ, thenfn-f in L I(lIl m) and in L2(lIlm).
For every n EN there exists a IPn E Cooo(K(n» such that

£(n)lfn(x) - IPn(xW dx <; nv(~(n»'

and thus

1 Ifn(x)-IPn(x)ldx<; {v(K(n»l Ifn(X)-IPn(.XWdX}1/2

K(n) K(n)

<; (* r/2,
where V(K(n» denotes the volume of K(n). Consequently, lPn-fin LI(Rm)
and in L2(R m ). Therefore,

and
(FJ)(x) = l.i.m.
n ..... oo
(21T)-m/2je- iX)'IPn(Y) dy.

Here "l.i.m. = limit in mean" stands for the limit in L2(lIlm). Hence, (Ff)(x)
= (Ftf)(x) almost everywhere. We can prove in a similar way that
(F-1)(x)== (itf)(x) almost everywhere. 0

(Ff)(x) = 1.i.m.(21T)-m/21 e-iX)'f(y) dy.

n-+oo K(n)
A similar formula holds for F - I.

PROOF. The functions xJ belong to LI(lRm)n L2(Rm), and xJ-f in

L2(Rm). Therefore, F(xJ)- Ff in L2(R m). Since by Theor,em 10.5

(F(xJ»(x) = (21T)-m/21 e-iXYf(y) dy, almost eveIywhere in Rm ,

K(n)
the assertion follows. D
10.1 The Fourier transformation on L2(RM) 295

For f. g E L2(~m) the convolution f *g is defined by the integral

(f * g)(x) = (2'17)-m/2 J f(x - y)g(y) dy.

The integral exists for all x E ~m. because f(x -. ) E L2(~m) and g( . ) E
L2(~m) imply f(x - . )g( . ) E LI(~m). (For the convolution of LI-functions,
see Exercise 10.1.) Moreover,
(f*g)(x) = (2'17)-m/2Jf(x-y)g(y)dy

= (2'17) -m/2 J g(x - y)f(y) dy = (g * f)(x).

Theorem 10.7 (The convolution theorem). For f, g E L2(~m)

(a) FI(Ff· Fg)=FI(F-Y' F-Ig)=f*g.
(b) The following assertions are equivalent:
(i) Ff . Fg E L2(~m),
(ii) F-Y' F-IgEL2(~m),
(iii) f * g E L2(~m).
In this case

PROOF.
(a) FI(Ff· Fg)(x) = «Ff)*, hex, . » with
h(x,y) = (2'17)-m/2e iXY (Fg)(y)

= J.
(2'17)-ml.i.m.
n->oo K(n)
e-iy(z-x)g(z) dz

= (2'17)-ml.i.m. J. e-iyzg(z + x) dz = (2'17)-m/2(Fg)(y),

n->oo K(n)

where gAz) = g(z + x). Moreover, let us set f _(x) = f( - x). Then it is
obvious that (Ff)* = F1*-, and thus
FI(Ff· Fg)(x) = (2'17)-m/2<F1*-, Fgx) = (2'17)-m/\1*-, gx)

= (27r)-m/2!f(-z)g(z+x)dz = (f*g)(x).

We can prove analogously that FI(F-~ . F-Ig) = f * g.

(b) If Ff . Fg E Li~m). then in the formula f * g = FI(Ff . Fg) we can
replace the operator FI by F - I and obtain that f * g E L2(~m) and
f * g -= F-I(Ff . Fg). If f * g E L2(~m), then with hi = Ff . Fg E
LI(~m) and h2 = F(f * g) E L2(~m) we have

(Flhl)(x) = (f*g)(x) = (F- 1h2)(x) almost everywhere.

296 10 Differential operators on L2(R"')

We show that hl(x) = hz(x) almost everywhere, and thus Ff . Fg = hI

E Lz(R m ). To prove this it is sufficient to prove that f(hl(x)-
hz(x»tp(x) dx - 0 for all tp E S(Rm) (cf. the proof of Theorem 10.5).
This follows from the equalities

fhl(X)tp(X) dx = fhl(X)(Fo-1Fotp)(x) dx

= f (Flh1)(x)(Fotp)(x) dx

= f(F-1hz)(x)(Fotp)(x) dx

= «F-1h z)"', Frp) -= <Fh~, Frp) = <h~ q»

= f h2(x)q>(x) dx.

The equivalence of (i) and (ii) and the second equality follow from the
formula F(h)=(F-1h)_. 0

EXERCISES

10.1. The convolution U. g)(x) = (2'IT)-m/z f j(x- y)g(y) dy is defined almost

everywhere for all j, g E LI(Rm ), and is a function from Ll(R"'). We have
FlU. g) = (Fd) . (F. g).
10.2. If f and Fd are from L1(Rm), thenj .. P1(Fd).
10.3. (a) For f E Lz(R m) and a E R m let fa be defined by the equality ja(x) =
J(x + a). Then (FJa)(x)" eiJCtl(Ff)(x).
(b) LetjE Lz(Rm) and let (Ff)(x) +0 almost everywhere in IRm. Then the set
{fa : a E iii"'} is total in L2(R m ) (Wiener's theorem).
Hint: If g..L {f" : a E Rm}. then F.(Ff . Fg) .. O.
(c) Let {} be as in Theorem 10.2. Then the set {{}" : a E IRm} is total in
Lz(lRm ).

10.4. We have (Fd)(x)-,>O as Ixl-'>oo for every j E L.(lRm ) (Riemann-Lebesque).

Hint: S(IR"') is dense in L.(lRm ), and /,,-'>j in L](IR"') implies Fdn-,>Fd
uniformly in IR"'.

10.2 Sobolev spaces and differential operators

on L2(lRm) with constant coefficients
In what follows, for all s :> 0 define
k.(x) = (1 + Ixlzy/2 for x E Rm
and
10.2 Sobolev spaces and differential operators on L 2(Rm) with constant coefficients 297

L2• s(IIl'") is obviously a dense subspace of L2(lIl m ). The equality

defines a scalar product on L2 • • (lIlm). We denote the corresponding norm

by II • 11(.). The space (L 2• • (IIl'"), (. , . )(.» is a separable Hilbert space, since
U. : L2•• (lIl m) -7L 2(1Il'"), UJ = kJ
is an isomorphism of L2. •(lIlm) onto L2(lIim).
The Soholev space of order s is defined by the equality
W2.• (R m ) = {fEL2(lIlm): FfEL2 .• (lIlm)} = F- 1L2 •• (R m ).
W2• • (IIl'") is therefore a dense subspace of L 2(R m ), and the equality

<f, g). = <Ff, Fg)(.) for f, g E W2•• (lIi m )

defines a scalar product of W2•• (Rm). We denote the corresponding norm
by 1/ • " •. Since F is an isomorphism of W2 •• (lIim) onto L2 , .(lIim), the space
W2, .(Rm) is also a separable Hilbert space.
First we show that the functions from W2 , .(lIlm) are differentiable in a
certain weak sense.

Theorem 10.S.
(a) Let s;;;' 1, ~ = (8jl , ~'2' • • . ,8jm ), (Mjg)(x) = xjg(x) and h. ,(x) =
f(x + fW) for j = 1, 2, ... ,m. Then for all f E W2, sCllim) and j =
1,2, ... , m

in L2(R m ). If f E S(lIim), then this limit equals Dj with a =

(8jl , 8j2 , ... '~'m)' We write Dj for this limit in case f E W2,s(Rm), as
well.
(b) If a is a multiindex and lal ~s, then the derivative Dj can be computed
by iteration. The order of differentiation is irrelevant.
(c) If s E No. then II . II. is equivalent to the norms

PROOF.
(a) For all x E iii"'

F(~(Jj,
If •
-f»)(x) = ~(ei-'j'
If
-1)(Ff)(x),
298 10 Differential operators on Lz(Rm)

and

It follows from this that F(I/itXb.• - f)~M.JFf. and thus

(I/i€)(h.• - f) ~ F-1M.JFf as € ~ O.
By Theorem 10.1 this limit is equal to D,,! for all f E S{Rm).
(b) If lal = I, then it is obvious that F Daf= xaPf E L2• • _I(Rm). Conse-
quently. D,,! E W2• s_I(R m). If s > 2. then by part (a) we can therefore
differentiate further. The commutativity of the order oj[ differentiation
follows from the equality D<y= p-1MaFj.
(c) We have IIfll; = IIk.PfIl 2•

IIfll;.o = ~ II MaFfll2 and IIfll;.1 = IIFfll2 + ~ II Ma Ff1l2.

lal<s lal-s
Since
I + ~ Ix al2 " C1(I + Ixl 2Y " C2 ~ Ix al2
lal-s lal<'

"C3(I+j_r.~~.mxls)" C3(I+I<t~slxaI2)
with appropriate constants C 1• C 2 • and C3 (that depend only on m and
s), all three norms are equivalent. 0

Theorem 10.9. SupposefE W2. s(R m ) and a is a multiindex such that lal "s.
(a) We have <D,,!. g) = <f. Dag ) for all g E W2.lal(Rm).
(b) The element Daf E L2(R m ) is uniquely determined by the equality
<Daf. g) = <f. Dag) for all g E COCO(IR:"').

PROOF.
(a) If g E W2.lal(Rm). then Fg E L2•lal(R"'), and thus Fg belongs to the
domain of the operator of multiplication by xa .. There:fore.
<D"!. g) = <F Daf. Fg) = <MaFf, Fg) = <Fl. MaFg)
= <Ff. F Dag) = <I. Dag ).
(b) By part (a) we have <Daf. g) == <I. Dag ) for evc~ry g E CoCO(R"')
C W2.lal(Rm). If fa is a further element from L2(Rm) su(:h that <fa. g> ==
<I. D"'g). then <la - D,,!. g> = 0 for all g E Coco(Rm). Therefore, fa =
Daf. since Coco(jRm) is dense in L2(Rm). 0

Theorem 10.10.
(a) For every s > 0 the set CoCO(R"') is dense in S(R"') with respect to the
norm II . II.·
10.2 Sobolev spaces and differential operators on Ll(Rnt) with constant coefficients 299

(b) For every s > 0 the set Cooo(lJilm) is dense in W2, .(lJilm) with respect to the
norm II· lis'

PROOF.
(a) Suppose rE No and r >s. We show that Cooo(Rm) is dense in S(lJil m)
with respect to II .lIr,o; since /I. II. ",CII· IIr,o' the assertion follows
from this. In order to prove the former statement, let t't E CoOO(R) be
such that t't(t) = 1 for t", 1, t't(t) =0 for t> 2 and 0", t't(t)", 1 for all
t E IR. For every n E N let t'tn E Cooo(Rm) be defined by the equality
t'tix) = t't(n-tlxj) for all x E Rm. If fE S(Rm) and we define/" = t't,jE
Cooo(lRm) for all n EN, then /lO"fn - 0'111 ...... 0 as n ...... oo for every multi-
index a. It follows from this that II!" - fllr, 0 ...... 0 as n ...... oo.
(b) Because of part (a) we only have to show that S(Rm) is dense in
W2 , .(lRm). Since F is an isomorphism of W2 , .(lRm) onto L2 , .(lRm) that
maps S(Rm) onto itself, it is sufficient to prove that S(Rm) is dense in
L2, .(Rm). This is surely true, as Cooo(Rm) is dense in L2 , .(Rm) (we can
prove this the same way as we did the corresponding assertion for
L2(lJilm), cf. Section 2.2, Example 8). 0

An m-variable polynomial P of degree r has the form

P(x) = ~ cax",
lal<'
where ca E C and Ca ~O for at least one a such that lal = r. If P is a
polynomial of degree r, then the formula
m a"l
P(O) = ~ c" 0" == ~ ca( _i)la l IT ~
1"1<' /a/<, j-1 ax)

defines a differential form of order r. We always assume that r=l'O, i.e., we

only consider non-trivial differential operators. In what follows let P be a
(fixed) polynomial. The equalities
D(To) = Cooo(lRm), To! = P(O)f for f E CoOO(lJilm)
define a differential operator on L2(lJil m) with constant coefficients. If we
denote by Mg the operator of multiplication by the function g, then the
following theorem holds.

Theorem 10.11.
(a) To is closable. For T=ro we have T=F-IMpF. T is called the
maximal differential operator with constant coefficients induced by P. The
operator T3 is equal to the maximal differential operator induced by the
conjugate polynomial p •.
:r::,...:::.-..,~--=~
(b) We have a(T)={P(x): xElRm} and (A- T)-I=F-1M('A_p)-.F for
A E p( T). The operator T has no eigenvalues.
300 10 Differential operators on L2(Rm)

If (1 + IE then - T)- is a Carleman operator for every

;\Ep(T), and

(;\-
with
'(;\- ') E

PROOF.
First define by the equalities
fE

Then To C TI and TI = F - IMp 1F, where denotes the restriction

of Mp s(~m) Theorem 'to.1), Since (as restriction a
closed operator) is closable, the operator TI that is unitarily equivalent
to M P, 1 is also closable. Therefore, To is closable, also. Now we show
that To and 1= It will follow from this

T = To= TI = F-1Mp, IF= F-1Mp,IF = F--]MpF.

To = : Since C T I , we have To C Com;equently, is

sufficient to prove that O( = S(Rm) octo)' For every E let
us construct, as in the proof of Theorem 10.to(a), a sequence Un) from
Cooo(~m) such that IID'Jn oafll ~O for all a, It follows from this that
II fn - -;,0 and Tofn - -;,0. f E and thus
O(T]) C OCTo). __
M p, 1= Mp: For this it is sufficient to prove that NIp, 0= M p , where
o denotes the restriction of Mp Cooo(~rn), In order to prove this
we have to the following: For every f D(Mp) E
L2(~m) : Pf E Li~m)} there exists a sequence Un) from Cooo(~m) such
that and ~Mpf. 'rhis can also be tht: same way in
Section Example 8 (cf. also the of Theorem lO(b».
Finally, Tti = Tti = T* = F-1(Mp)* F= F-1Mp*F. (cf. Section 5.1,
Example I, (5.1».
Since and Mp unitarily equivalent,

For;\ peT)

(;\- T)-I = (;\- F-IMpFr l

= 1(;\- Mp)Ffl F- 1(;\ Mp)-

Because of the assumption r>O, the set {xE~m: P(x)=s} is a null

set for every sEC (proof!). By Section 5.2, Example I, (5.14) the
M p therefore has no eigenvalue. Then the same for,
10.2 Sobolev spaces and differential operators on L 2(Rm) with constant coefficients 301

(c) Assume that (1 + IPI)-I E L2(R m ) and AEp(T), i.e., there is an 1J >0
such that IA - P(x)1 >1J for all x E Rm. For all x E Rm such that
IP(x)1 > 21AI + 1 we have the estimate

IA - P(x)1 > IP(x)I-IAI = (tIP(x)I-IAI) + 4IP(x)1

> 4+ 4I P (x)1 = HI + IP(x)l).
Consequently,
IA - p(X)I-1 <:; 2(1 + IP(x)l)-1
for x E Rm such that (1 + IP(x)l)-1 <:; t(l + IAI)-I.
Since, besides, IA - p(X)I-1 <:; 1J -I for all x E Rm, it follows that (A-
p)-I E L2(Rm). The convolution theorem (Theorem 10.7) therefore
implies that
(A - T) - If = F - I[ (A - P) - IFf] = h" * f,
where h" = F-I«A - p)-I). Since h" E L2(R m ), the operator (A - T)-I
is a Carleman operator. 0
REMARK. In Theorem 1O.I1(b) the closure is superfluous for m = I, as can
be easily seen. This is not true for m> I, as the example of P(xp x 2 ) =
(l-X 1x 2)2+ X; shows; in this case we have {P(x): xER2 }=(0, 00).

Theorem 10.12. The following assertions are equivalent:

(i) A II coefficients c" of P are real.
(ii) To is symmetric.
(iii) To is essentially self-adjoint.
(iv) T is self-adjoint.

The proof immediately follows from Theorem 1O.lI(a).

Theorem 10.13. Let T be a self-adjoint differential operator with constant

coefficients induced by P, and let E denote the spectral family of T.
(a) For all sER

E(s) = F-IMX(XERm : P(x)~s}F.

(b) If IP(x)l~oo as Ixl~oo, then E(t) - E(s)is a Carleman operator for all
s, t E R such that s <:; t, and

(E(t) - E(s»f(x) = (277) -m12 J es,l(x - y)f(y) dy,

where es,l = F-1X(XERm: s<P(x)<;t} E L2(Rm).

(c) If P(X)-HIO as Ixl~oo, then E(s) is a Carleman operator for al/ s E R.
The same holds for 1- E(s) provided that P(x)~ - 00 as Ixl~oo.
 302 IO Differential operators on L2(R"')

PROOF .
. (a) The first assertion is clear, since
F(s) = MX{XERm: P(x)..;s) for s E IR

is the spectral family of M p.

(b) We have
E(t) - E(s) = F-1MX{XER":s<P(x)..;t)F.

Since IP(x)I-HO as Ixl~oo, the set {x E IR m : s <P(x)·' t} is bounded.

Therefore, X{xERm: s<P(X)</} belongs to L2(lRm). The assertion then
follows from Theorem 10.7.
(c) If P(x)~oo as Ixl~oo, then {xElRm: P(x)";;;;s} is compact, and thus
X(xERm: P(x)<.s}E L2(lRm). If P(x)~ - 00 as Ixl~oo, then X{XER": P(x»s)
E L2(lRm). The assertion follows from Theorem 10.7 in both cases. 0

A polynomial P of degree r and the operators To and :r induced by P

are said to be elliptic if there exists a C ;;. 0 such that

1+ IP(x)1 ;;. cO + IxI2)'/2 = Ck,(x) for all x E IRm.

(Observe that we always have 1+ IP(x)I";;;; C' k,(x) with an appropriate
choice of C' ;;. 0.) The principal pari of P is given by

P,(x) = ~ c"x".
1,,1=,
Correspondingly, the principal pari of P(O) is given by

P,(O) = ~ c" 0".

1,,1=,

Theorem 10.14. Let P be a polynomial of degree r, and let T be the maximal

differential operator induced by P. Then the following statE'ments are equi-
valent:
(i) P is elliptic.
(ii) The principal part of P vanishes only for x = O.
(iii) O(T) = W2• ,(lRm).
In this case the norms II . II T and II . II, are equivalent.
Let us assume that there is an Xo E IR m, x o :;6 0 such
PROOF. (i) implies (ii):
that P,(xo) = O. Then we also have P,(sxo) = 0 for all s E R. Therefore,

IP(sxo)1 = I ~ c"sla1xol";;;; C(I+lsxoI2)('-I)/2,

1,,1<'
in contradiction with the ellipticity of P.
10.2 Sobolev spaces and differential operators on L2(R m ) with constant coefficients 303

(ii) implies (i): Let 1J = min{IPAx)1 : x E Rm , Ixl = I}. Then for all x
ERm

IP(x)1 = Ipr(x) + ~ Caxal;> Ixl r1J - C(I + IxI 2)(r-l)/2.

lal<r
The ellipticity of P follows from this.
(i) implies (iii): We have
D(T) = F-I{JEL2(Rm): PfEL 2(R m)}
= Foo.l {J E L2(R m) : krI E L2(R m)} = W2, r(Rm).
Since
IIfll2 + II Tfll2 = II Ffll2 + IIMpFfIl2,
Ilfll~ =
the equivalence of the norms II . II T and II . II r follows.
(iii) implies (i): We obtain from the equality D(T) = W2, r(Rm) that
D(Mp) = FD(T) =L2• r(R m ) =D(MkJ
By Theorem 5.9 Mk is relatively bounded with respect to M p , therefore
with respect-to MI~I' as well (since D(Mp)=D(M1PI ) and IIM1P1fll=
IIMpfll). The bounded ness of Mk,(l + M1Pr l, i.e., the boundedness of the
function kr(l + IPI)-I, follows from this by Theorem 9.9. Consequently, P
is elliptic. 0

Theorem 10.15. Let T be a self-adjoint elliptic differential operator with

constant coefficients on L2(Rm).
(a) If m > I, then Tis semibounded.
(b) If T is bounded from below, then E(t) is a Carleman operator for every
tER.
PROOF.
(a) Since T is self-adjoint, P is real-valued. As T is elliptic, IP(x)I-HX) as
Ixl~oo. Consequently, IP(x)I>O for alllxl;>co. Because of the con-
tinuity of P it follows (due to the assumption m> I) that P(x) > 0 for
alllxl ;>co or P(x) < 0 for alllxl ;;;'co; hence P(x)~oo or P(x)~- 00
as Ixl~oo. The boundedness from below or from above follows from
this.
(b) If T is elliptic and bounded from below, then we obtain (as in the
proof of part (a» that P(X)-Hx) as IxI--HlO. The assertion follows by
Theorem 10.13. 0

Corollary. Let T be a self-adjoint differential operator on L2(R m) with

constant coefficients. Then
I. O'(T) = R or O'(T) = [y, 00) or O'(T) = (-- 00, y).
2. If T is elliptic and m> 1, then O'(T) = [y, 00) or O'(T) = (- 00, y].
304 10 Differential operators on L2(Rm )

3. We always have Hp(T) = {O}, Hc(T) = L2(lR m), and o(TJ = oiT) = 0c(T)
(cf, also Exercise 10.7).

EXAMPLE l. m = I, P(x) = x, P(D) = (I/i)(d/dx). Then Tis elliptic and not

semibounded; a(T) = IR.

EXAMPLE 2. m~l, P(x)=~;"lx}=lxI2, P(D)=-~j_l(a2/ax])=-d.

Then T is elliptic and bounded from below; o( T) = [0, (0).

EXAMPLE 3. m = 2, P(x) = x; + x1, P(D) = - (a 2lax?) + (a 4/ax1). Then T

is not elliptic but is bounded from below; a( T) = [0, 00].

EXERCISES

10.5. Let f belong to W2•• (lIl m ) with s >mI2. Then f is Lipschitz with exponent
8 E (0, I] n (0, s - (mI2», i.e., there exists a C ;;. 0 such that If(x) - f(y)l..;;
Cjx-yl6 for all x,yER m •
10.6. For rENo such that r < m the set Cooo(lIlm \ (O}) is dense ill W2• ,(Rm).
Hint: If {} E Cooo(ll~m) is such that {}(x) = I for Ixl.;; 112 and (}(x) = 0 for
. w w
Ixl ;;. I and Dn(x) = D(nx), then D,J ~ 0; therefore, (I - Dn)f·~ f in the sense of
W2• ,(lIlm) (cf. Exercise 4.25).
10.7. If P is a non-constant polynomial, then Mp and F-1MpF have a pure
absolutely continuous spectrum.

10.3 Relatively bounded and relatively compact

perturbations
In this section we first give conditions in order that an operator
Mq : W2.,(lRm)~L2(lRm), fH>qf be bounded or compact. For the sake of
simplicity, we only consider integers r; this is sufficient in most applica-
tions to differential operators. (Corresponding results for arbitrary r can be
found in M. Schechter [34], Chapter 6.)

Theorem 10.16. Let 0, s < r (not necessarily integral). Then for every '11 >0
there exists a C'I ;;. 0 such that

11111. ~ 'I111f1l, + C'IlIlll for all f E W2 .,(lR m ).

PROOF. The assertion is equivalent to the inequality

10.3 Relatively bounded and relatively compact perturbations 305

For all N > 0 we have

IIfllt.) = Jlf(x)I2(1 + IxI2y dx

0;;; (I + N 2 )' IIxl";NIf(xW dx + (I + N y-rIIxl>NIf(xW(l + Ix12)' dx
2

+ N 2Yllfl1 2 + (I + N2y-rllfll~r)'
0;;; (I
Due to the inequality s - r < 0, the assertion follows from this if N is
chosen large enough. 0
A measurable function q : Rm~c is said to be locally square integrable
if f)q E L2(R m ) for every f) E Cooo(Rm). This holds if and only if Iql2 is
integrable over every compact subset of Rm. The set of locally square
integrable functions obviously constitutes a (complex) vector space. This
space will be denoted by L2,loc(Rm). For every q E L2 , loc(R m) let us define

Nq(x) = {I Ix-yl..;'
Iq(y)lldy }I/2 forall xERm.

Nq is obviously locally bounded, i.e., it is bounded on every compact

subset. Nq is even continuous (proof!).
For every measurable function q : Rm~c and every pER let us define
(the value 00 is allowed)

for p<m,2

for p>m.

We denote by Mp,loc(R m) the vector space of measurable functions q : R m

~c for which Mq, / . ) is a locally bounded function. Mp(Rm) denotes the
subspace of those q E Mp. loc(Rm) for which Mq, p( .) is bounded. For
q E Mp(Rm) we set
Mq,p = sup{Mq,p(x): XERm}.
For all PI' P2 E R such that PI 0;;; P2 we have
Mp"loc(Rm) C M p" loc(Rm) e L2, loc(Rm),
Mp) (Rm) eMP (Rm).
2'
(10.1)
if q E Mp,(R m ), then we obviously have
M q , P2 ..;; M q , p,' (10.2)

2 This definition goes back to F. Stummel [52] for p < 4.

~06 \0 Differential operators on L2(R m )

EXAMPLE 1. Every bounded measurable function belongs to Mp(lRm) for

every p>O.

EXAMPLE 2. Assume a E IRm , c ;> 0, and 0 < 8 < m /2. Assume, furthermore,
that the function q : IRm ~C is measurable and Iq(x)1 ;;;; clx - al- 8 almost
everywhere in IRm. Then q E Mp(lRm) for all p > 28. This is obvious for
p > m, since q is locally square integrable and q(x)~O as Ixl~ 00. Now
suppose that 28 < p < m. Then

M;' p(x) "' c2 {1 Ix - yl < I

+ 1Ix - yl < I
Iy - al- 26 lx - ylP-m dY }
ly-al<lx-YI ly-al>lx-YI

"' c2 { 1Iy-al< 1
Iy - a!p-26-m dy +
Ix-YI< I
1
Ix - yIP-28-m dy }

= 2c 21 lyIP-28-m dy = C < 00.

Iyl < I
Of course, sums of such functions also belong to Mp(lRm).

Theorem 10.17. Assume that r E Nand p < 2r.

(a) There is a constant C > 0 such that
IIqfll"' CMq,pllfll, forall qEMp(lRm) andall jEW2.,(lRm ).
(b) For every q E Mp(lRm) and every '11 > 0 there is a C'1 such that
IIqfll "' 1I11fll, + C'1l1fll for all f E W2.,(lRm).
PROOF. Take a i} E Cooo(lRm) such that i}(0) == 1 and i}(x) = 0 for Ixl ;> 1, and
let i}. E Cooo(lRm) be defined by the equality
i}s(x) = i}(S-IX) for x E IRm and s E(O, 1].

Let Om denote the unit sphere in IRm. Then for all f E CoOO(lRm), s E (0, 1)
and wEO m

f(O) = i}s(O)f(O) = - los :t (i}s(tw)f(tw» dt == •••

== (-1)' (SfS ... fS aa', (i}s(t\w)f(t\w» dt\ ... dt,_\ dt,

Jo I, 12 t1

== (-I)'Ia S :t~(i}s(t\w)f(t\W»{Ia'l . .. Ia'r-1 dt,.... dt 2 } dt l

(-I)'
= (r-I)! 10 sa'
Fr(i}s(tw)f(tw»t,-I dt.
10.3 Relatively bounded and relatively compact perturbations 307

With
at
g(y) == g(tw) == a? (D-.(Iw)f(tw» for y co tw,O < I <s
we therefore obtain by integration over the unit sphere Om that

= cll'tr-11Ig(,w)1 dw dl == eli lylr-mlg(y)1 dy,

o Om Iyl<s
where Vm denotes the area of Om and dw the area element on Om' Since

I g(y)1 < C2( ~ ID'1(y)1 + s-t ~ ID'1(Y)I) = : hey),

lal<r lal<r
we obtain for all E > 0 that
If(0)1 2 < e 3 i
. Iyl <,
Iyj<-m dy fIyl <.Ih(y)J2lyI2r-<-m dy
< C4s<i lyI2r-<-m { ~ 10'1(y)12 + s-2r ~ 10'1(Y)1 2 } dy.
Iyl < 1 lal <r lal <r

It follows similarly that for all s E Rm

If(xW <: C4 s< {f Ix-yl<1

Ix - yI2r-<-m ~
lal<r
10'1(y)1 2 dy

+s- 2r fIx - yl < Ix-yI2r-<-m lal~<t 10'1(y)l2dY}.

1

In order to estimate II qfll we distinguish between two different cases.

First assume that 2r > m. In this case, without loss of generality we can
assume on the basis of (10.1) and (10.2) that p >m. Furthermore, choose
E=2r- m. Then

jlq(x)f(xW dx < C4 S 2r - mjlq(xW

x {fIx-yl< llal<r
~ 10'1(y)1 2 dy + s-2r fIx-yl< Ilal<r
~ 10'1(y)1 dY } dx 2

< c4s 2r - mM;'pj{ L 10'1(y)j2+s-2r L 10'1(Y)12 } dy

lal<r lal<r
< esM;'p{ s2r- m llfl/; + s-mllfll;_d·
308

Now let 2r " m. Then p <m. Choose ~ = 2r - p. We have

flq(x)f(xW dx " C4S 2r -P f,q(x W{ f.Ix-yl<; ylP-m 2:

lal<;r
ID1(y)12 dy

+s-2r I
Ix- <; I
Ix - ylP-m L
l"l<r
ID1(yW dY} dx

"C6Mq~ {s2r- P llfll;+s-Pllfll;_I}'

In both cases we obtain assertion (a) for f E Cooo(~m) provided that we
take, for example, s =
Since by Theorem 10.16 for every J.L > 0 there exists a CI" ;;;. 0 such that

IIfll;-1 " J.Lllfll; + C"lIfll 2 ,

assertion (b) follows for all f E Cooo(~m) in both cases provided that first s
and then J.L chosen small enough.
f W2, r(~m), then there exists a sequence Un) from Cooo(~m) for
which J" ~f in the sense of W2.l~m). Since the sequence Un) then
converges also in L2(~m), there a Un) that converges to
f almost everywhere. Then this holds for (qfn.) and qf, as well. Since the
sequences (111".11,) and 011".11> are bounded, (1lqfn.lD is also bounded and
Fatou's lemma implies that qf E L2(~m) and

lim inf
k-,>oo
{
f dx }
1/2

lim infllqfnk
k-,>oo
"lim inf('ljll!"
k-,>oo'
+ C"IIfn,lI)
A

'I} C"lIfll·
We can prove assertion (a) for f o
Using this we can now give conditions for the relative bounded-
ness and relative compactness of the perturbations of a closed operator T
such that D( T) W2• ,(~m).

Theorem 10.lS. Assume r E ~ and T is a closed operator on Lz(W) such that

D( T) c Let be an operator L2(n m ) such that
Vf = ~ q"D"j for f E Wz,,(R"').
lal<r

Let the functions qa be measurable, and assume that qa bounded for = r,

l':1"1_,sup{lq,,(x)l: E ~m} = c, q" M".<~m) with Pa < 2(r -Ial)for la <r.
Then V is T-bounded. If IIfll, <dllTfil + ellfll for a/l ff=' D(T), then the
T-bound V less or equal to dc. If = for r, then
T-bound of V is equal to O.
10.3 Relatively bounded and relatively compact perturbations 309

PROOF. If Mk is the operator of multiplication by kr = (1 + I . 12)'/2 on

L2 CR m ) and T; = F -IMk F, then 1', is a self-adjoint operator on L2CRm ) and
D(Tr) = W2• rCRm). By Theorem 5.9 the operator Tr is T-bounded, i.e., there
exist contants d and e such that II Trfll ,dll Tfll + ellfll for allfE D(T). Let

v, = L qa O '" and Vo = L qa 0"'.

1"'1-' lal<r
We obviously have
IIVJII ,cllfll, = cllTJII ,cdllTfll + cellfll·
It is therefore sufficient to show that for every f >0 there exists a C. ;;. 0
such that
II Vofll
fllfll, + C.llill for all f E W2• r (R m ).
,
Since oaf E W2 • r_lal(R m) (because MaFf E L2 . '_lal(Rm», by Theorem 10.17
for every '1/ > 0 there exists a C1j;;;' 0 such that for all lal <r and f E
W2 .,(R m )

Ilqa D"'fll ' 'l/IIO"'1I1'-lal + C1jIlD"'fll ' 'l/lIfll, + C1j II filial'

This gives, together with Theorem 10.16, the assertion. o
In order to prove the relative compactness of perturbations, we need the
following auxiliary results.

Theorem 10.19. Let 0 ,s <r (not necessarily integral). The mapping

4> : W2• r(R m) ~W2.s(Rm), f~ ((Jf
is compact for every q> E Cooo(Rm).
PROOF. The operator 4> is compact if and only if the operator K defined on
L 2(R m ) by the formula

is compact, since the four operators at the ends are unitary (cf. Section
10.2 for Us and Ur). By Theorem 10.7 we obviously have for f E L2(R m)
that

(Kf)(x) = J(l + IxI y/21f!(x - y)(1 + lyI2)-r 2f (y) dy

2 I

with 1f! = (27T)-m /2 Fq> E S(Rm). The operator K is therefore an integral

operator with kernel
310 to Differential operators on L1(Rm )

Since 1/1 E S(R m), for every I E ~ there is a BI such that

1I/I(x)1 < BI(I + Ixl}-I for all x E Rm.

For n E ~ let us set
for Iyl<n,
Iyl>n, hll(x,y) = k(x,y) - k,,(x,y).
for
We shall prove the assertion by showing that for the operators K" and H"
induced by kll and h" the following holds: K" is a Hilbert-Schmidt operator
and IIH"II-+0 as n-HtJ.
For Iyl <.n and Ixl > 2n we have

II/I(x - y)1 < BI(l + Ix - yl)-I < BI(I +4I x lr"·

It follows from this that for 1= m + I + s and an appropriate constant C

1I/I(x-y)I<.C(I+lxl}-m-l-s for Iyl<n andall xERm.

We therefore have

Iyl <n,
Iyl>n,
and thus k" E L2(R'" x Rm), i.e., K" is a Hilbert-Schmidt opl~rator.
To estimate the norm of H", we use the corollary to The:orem 6.24. We
have

1 Iyl ;."
1I/I(x-y)I(l+lyI2)-r/2 dy < C.(l+n)-rjll/l(x-y)ldy

< Cil+n)-r < Cil+lxl}-s(l+ny-r

for Ixl <. n. Since 1/1 E S(Rm), we have

1 Iyl ;."
1I/I(x-y)I(i+lyI2)-r/2 dy < c 3 j(l+lx-yl}-m-r(l+IYI}-rdy

< c3 { { (I + Ix - yl)-m-'(1 + Iyl)-r dy

Ix - yl <: Ixl/2

+{ (l+lx-yl)-m- rdY }
Ix - yl > Ixl/2

< Cil + Ixl)-r j (1 + Ix - yl)-m-r dy

+C3 1lyl>lxl/2
(l+lyl)-m-rdy
10.3 Relatively bounded and relatively compact perturbations 311

for Ixl >n. Consequently, it follows for all x E ~m that

Furthermore,

J(1 + Ixl)sl1j;(x - y)ldx " J(l + IxlY(l + yl) -m-s-I

C7 Ix - dx

= c {1 7 +1 +lxI)S(1+Ix-yl)-m-s-I dX}
(1
Ixl <; 21yl Ixl> 21yl

" C (I + 21ylY J(l +

7{ yl)-m-s-I Ix - dx

+ J(I + IxI)S(1 +4I x lr m

-
s- 1 dX)

" Cg(I + Iyl)s.

Hence, for all y E ~m such that Iyl >n (and thus for all y E ~m)

It follows therefore from the Corollary to Theorem 6.24 that IIBnll <:
C9 (1 + ny-r _0 as n_oo. 0

Theorem 10.20. Assume that sEN, q E Mp(~m)for some p < 2s and Nix)-
o as Ixl- 00. Then the mapping
Q : W2.s(~m) _Li~m), f~ qf

is compact.
PROOF. Let Mk be the operator of multiplication by the function ks' and let
~=F-IMkF:We have to prove the ~-compactness of Q. By Theorem
1O.17(b) th~ operator Q is obviously ~-bounded with T,.-bound zero.
Therefore, by Theorem 9.11 it is enough to prove the T,.2-compactness of
Q. This, in turn, is equivalent to the compactness of

We shall prove this in what follows.

If p :> m, then set 'T = p. Then

If p <m, then we choose a 'T for which p <'T < min(m, 2s). Then Holder's
inequality with the exponents p = (m - p)/( 'T - p) and p' = (m - p)/(m - 'T)
312 \0 Differential operators on L2(R m )

gives

Consequently, in each case we have found a T for which

T < 2s and Mq,~(x) ~ 0 as Ixl ~ 00.

Now let cp E Co""(lRm) be such that cp(x) = 1 for Ixl '" 1, cp(x) = 0 for Ixl ~ 2
and 0", cp(x) '" 1 for all x E IRm. Let the functions CPn E Co""(lRm) be defined
by the equality CPn(x) = cp(n -IX) for n E N and x E IRm. Then by Theorem
10.19 the operator
<l>n : W2,2r(lRm ) ~W2,s(lRm), ff-4 cP,J
is compact for all n E N. Since the mapping Q : W2,.(lRm)~L2(lRm) is
bounded by Theorem 10.17, the compactness of
Q<I>n : W2. 2r(lRm ) ~L2(Rm), ff-4 qcp,J
follows. As Mq,1'(x)~O when Ixl~oo, we obviously havc~ M(I-IJI.)q,1'~O
when n~oo. It follows from this for the operators
An : W2, 2r(lRm ) ~L2(lRm), ff-4 q(l - .pn)f
that IIAnll~O (cf. Theorem 10.17). This implies that Q=limn--.""Q<I>n'
Hence, Q is compact. 0

Theorem 10.21. Let r, T, and V be defined as in Theorem lO.IS.

o
(a) If qa = for lal = rand
NqJx) ~ 0 as Ixl ~ 00 for lal < r,
then V is T-compact.
(b) If T;s selj-a4ioint, D(TP) C W2,s(lRm) for some p > 1 and s >r, and
qa(x)~O for and lal = r,
N q• (x)~O for and lal <t,
then V is T'-compact for every t> l.

PROOF. The mappings

W2,,(lRm) ~W2.'_lal(lRm), ff-4 D~
10.4 Essentially self-adjoint Schrodinger operators 313

are hounded, and by Theorem 10.20 the mappings

W2,'_lal(~m) ~L2(~m), g ~ qag
are compact for lal < r. The compactness of the operators
Va : W2,,(~m) ~L2(~m), fH qaDi
follows from this for all ex such that lal <r. This gives the T-compactness
of V for part (a) of the assertion.
We can show the compactness of
Va : W2.s(~m) ~L2(~m), fH qaDi
for lal = r in an analogous way. Since D(TP) C W2, s(~m), this gives the
TP-compactness of V. As V is T-bounded by Theorem 10.18, we obtain the
T'-compactness of V for every t> I by Theorem 9.11. This proves part (b).
o
EXERCISES

10.8. (a) Assume that q E 4. loc(R m) for some p ;;. 2 (i.e., q is measurable and IqlP
is integrable over every compact subset of Rm). If p> 2mlp for p > 2
and p ;;. m for p = 2, then q E Mp,lo.,(lRm).
(b) Prove a corresponding result for Mp(Rm).
10.9. Let T be defined by the equalities D(T)= W2,2(1R), Tf= -Aj+qf, where q
is a continuous real-valued function with compact support. If fR q(/) dl < 0,
then T has at least one negative eigenvalue.
Hint: Theorem 6.33, 7.26(b) and 10.2 I(a).
10.10. Let T be a self-adjoint operator on L2(Rm).
(a) If D( T) C W2 , ,(Rm) for some r > m 12, then (A - T) - 1 is a Carleman
operator for every AEp(T). The operator E(b)- E(a) is a Carleman
operator for all a, b E R.
(b) If D(Tn) c W2, ,(Rm) for some r >m/2, then E(b)- E(a) is a Carleman
operator for all a, bE IR.

10.4 Essentially self-adjoint Schrodinger operators

In this section we consider operators on L2(lijm) that are induced by
differential forms
m
7f = L (Dj - bif + qf,
j=1

with Dj = (lji)(ajax.;) and with real-valued functions h.; E cl(lijm) and

q E L2 ,loc(Rm) (we have used the notation (Dj - b)'2j= (Dj - bj)[(Dj -
bj)J]). The operator T defined on L2(lijm) by
D(T) = Cooo(lijm) and Tf = 7f for f E D(T) (10,3)
314 10 Differential c'perators on L2(R'")

is obviously symmetric. For f E D( T)

m
Tf - -df - 2 ~ bjDJ+ (bl+i divb+ q)f (10.4)
j-I

with b-(b l • bl •...• bm ) and bl-Ij_Ib}.

In non-relativistic quantum mechanics operators of this form occur as
Schrodinger operators of systems consisting of finitely many charged par-
ticles in an electromagnetic field. For m'" 3 we encounter the Schrodinger
operator of one particle in an electric field with potential q and a magnetic
field with the vector potential b-(b l • bl • b3).

(lO.S)

For m - 3N the Schrodinger operator of a system of N charged particles has

this form: If in this case we write

R3N 3 x = (XI. I' XI. 2' XI. 3' X2. I' .••• XN_I. 3' XN.I' XN.2' XN.3)'
where JC.J - (JC.J. I' JC.J. 2' JC.J. 3) are the coordinates of the jth particle, then
N 3 2
Tf(x) = ~ ~ (Djk - ~k(Xj» f(x) + q(x)f(x) (10.6)
j - I k-l

with
N I-I N
q(x) ... ~ ~ qj,(Xj - x,) + ~ qix) (10.7)
1-2j-1 j-I
and
bjk ... ~bk for j = 1.2, ... ,N and k ... 1,.2.3, '
where the factor ~ depends on the charge of the jth particle. (Here we have
replaced by I all physical quantities that are irrelevant for the properties
studied here.)

Theorem 10.22. Let the operator T be defined as in (10.3). Assume that

b l •••• , bm E C l(Rm) are bounded with bounded derivatiV!s, and q E Mp(R'")
for some p < 4. Then T is essentially self-a4joint and D(T) .. W2 2(R m ). The
operator T (and thus also T) is bounded from below. •
The proof can be obtained immediately from Theorems S.28, 10.18. and
9.1 if we consider - fl as the unperturbed operator (cf. the representation
(10.4) of T).
The operators under (10.S) and (10.6) satisfy the assumptions of Theo-
rem 10.22 in almost all physically realizable cases.
In this section we actually prove that under much more general assump-
tions these operators still remain essentially self-adjoint, while D(T) is then
generally no longer equal to W2• 2(Rm). The following theorem is due to
T. Ikebe and T. Kato [44] in a somewhat more general form. Our presenta-
10.4 Essentially self-adjoint Schrodinger operators 315

tion relies on a proof of e.G. Simader [51] for a somewhat more general
result.
For r E N let W2, "loc(RnI ) denote the space of (equivalence classes of)
the functionsj: RnI_C for which iijE W2,,(RnI ) for all ii E Cooo(Rm), For
every j E W2, "loc(Rm ) and j E {I, 2, .. , , m} there is exactly one i.J E
W2, ,_I, loc(Rm ) such that for all k > 0

'XJci.J limXk~(i.J.-j)
= .... 0 It: '
in the sense of L2(RnI),

where .ti,. is defined as in Theorem 10,8 and Xk denotes the characteristic

function of Mk = {x E Rm : Ixl <; k } (on Mk the function /i can be defined
by the equality M/i = XkOi iij) where I} E CoOO(RnI) and I}(x) == 1 for x E
Mk ; this definition of M/i is independent of ii). For j E W2 , ,(Rm) we have
.ti == OJ. Consequently, for j E W1, "loc(lRnI ) we define OJ= /i. If I} E
Co""(RnI) is such that ii(x) = I for x E Mk , then
OJ(iij)(x) == O),(x) almost everywhere in Mk •
Since OJ E W2, ,_I,loc(Rm), we can define successively all derivatives Oa.j
for lal <;r; we have Oa.jE W2,r_lal,loc(Rm). The function hj belongs to
W2, I, Joc(Rm) provided that h E C I(Rm) and j E W2, J,loc(R m). In particular,
the expression ~j-l(OJ - bit is meaningful for bj E C I(R"') and j E
W2, 2, Joc(R"').

Theorem 10.23. Let the operator T be dejined as in (10.3). Let bl' b2 , ••• , bm
E C l(Rm) and let q belong to Mp , Joc(Rm) jor some P < 4. Moreover, assume
that q = ql + q2 with
ql E Mp(Rm) jor some p< 4,
q2(X) >- Clxl2 jor some C >0 and all x E Rm.
Then T is essentially self-atfjoint. We have

m
fj = ~ (OJ - b)2j + qf jor j E D(f).
)-1

For the proof we need the following auxiliary theorems.

Auxiliary theorem 10.24. Ij q EMiR"'), then Iqll/2 E Mer(R"') jor every

a> p/2. (A similar result holds jor Mp,loc(Rm) and Mer,loc(R m ».
PROOF. If p> 2m, then it follows by the Schwarz inequality that

1 Ix-yl<1
Iq(y)ldy <; {c1 Ix-yl<1
Iq(y)1 2 dY } 1/2 " C 1/2Mq, p

for all x E Rm, i.e., !qll/2 E Minm) for" >m, and thus for all " > p/2.
316 10 Differential operators on Lz(Rm)

If m <. p < 2m and > pl2 ~ m/2, then

f.
(J

Iq(y)"x - yla-m dy
Ix-yl<J

<. {f.Ix-yl< Iq(y)1 dy f.Ix-yl

I
2
<I
Ix - ylla-2," dY } 1/2 <. CMq , p'

because 20' - 2m > - m,

If p <m and a> p/2, then it follows for E- 0' - (pI2) > (I that
f. Ix-yl<1
Iq(y)lIx - yla-," dy

== f.
Ix-yl<1
Ix - yl,-(,"/2)[ Iq(y )Hx - yla-'-(,"/2)] dy

Consequently, the assertion is proved in each case. o

Auxiliary theorem 10.25. Let the functions bii = 1,2, ... , m) and qi be the
same as in Theorem 10.23. For every TJ > 0 there exists a C1J ~ 0 such that
m
<lqllf, f) <: TJ ~ II(Oj - b)fll 2 + C"lIfll 2
i-I
for al/ f E Co""(IR'").
PROOF. Iqd I / 2 belongs to M..(R'") for some (J < 2 by Auxiliary theorem
10.24. Take a If E CoOO(R'") such that If(X) == 1 for x E supp f. and set
h=lgradlfl+llfl. Set, furthermore, cp,(x)=(lf(xW+E)I/2 for every £>0.
Then by Theorem 10.17
<Iqllf,f) = IlIqlj1/~12 <: Illqtl)/~.lfI12 <. t IIcp.lfll~ + C11ICP.lfli l
m
<. TJ ~ !IIf(x)OjCP.(xW dx + C2111grad Iflcp. + 'If'cpi

*
i-I
m
= .~ f CP.(x)-21 lf (xWlf(x)* OJ(x) - f(x)(Dj(x»*,2 dx
J-l

*
+ C2 11 hcp.1I 2
m
= ~ f cp.-2Ir(Oj - bj) - f(Oj - bj)*r! dx + C211 hCP.1I 2
J-I
m
<. TJ ~ f cp.-2IfI 2IDj - bjl2 dx + C2 11hCP.1I 2
i-I
m
<. TJ ~ fIDj(x) - bix)f(xW dx + C2 I1hCP.11 2.
j-I

The assertion follows by letting £ tend to 0, because h(x) = 1 for x E supp f.

r1
lOA Essentially self-adjoint Schrodinger operators 317

Auxiliary theorem 10.26. Let T be defined as in Theorem 10.23. For n E '"

let

R" = {x E R m : n ..; Ixl..;; 2n} and Q" = {x E R m : I or;;; Ixl or;;; 3n }.

There exists a C ;;.. 0 such that for a/l f E Co""(Rm) and a/l n E '"

and
m
~
j_ 1
1
Ixl.:"
I(Dj - bj)fl 2 dx ..; 21 Ixl.: 2"
ITfl2 dx + Cn21
Ixl.: 2n
Ifl2 dx.

PROOF. Let 'II E Co""(Rm) be such that

I for I" Ixl" 2,
o " 'II(x) " I for all x E Rm, 'II(x) = { 0 for
Ixl " i and Ixl ;;.. 3.
Moreover, for all n E '" set 'IIix) = 'II(n -IX). Since'll" E Co""(Rm) and
supp 'II" C Qn> supp(grad 11,,) C Q" "R",
by Auxiliary theorem 10.25 we have for all f E Co""(Rm) that

1171" Tfll2 + 111I,J11 2 ;;.. 2Re<lInTf,'I/,J) = 2Re<Tf,'I/;.t)

= 2ReL~1 «D) - bj)f, (D) - b)( 'I/;.t) + <qq,J, 'I/,J) }

m m
= 2Re { ~ II'I/,,{Dj - bj~fll2 + 2 L <'1/,,(0) - b)f, (OJ'll,,)f)
)-1 )=1

+ <qt'l/,J, 1I,J) + <q211,J, 'I/,J) }

m 1 m
;;.. 2 { )~tll'IJ,,(Oj-b)fIl2_4 ~ll1l1n(D)-b)fIl2

- 4 ~ II (OJ'll,,)Jll z - ~~ 11(0) - b)'II,JII Z

)=1 )=1

- Ctll'I/,Jll z - Czn21111JII2}.
Since
II(Oj - b)l1Jll z = II {D)'II,,)f + 'I/,,{Oj - b)fll 2
..;; 2{ II {OJ'll,,)fIl 2 + II '1/" (D) - b)fIl 2 },
and
318 10 Differential operators on L2(Rm)

we can further estimate:

11'11.. Tfll2
1 m
+ 11'I1..f1l2 > 2" ~ II '11.. (OJ - b)f11 2 -
)=1
C4 n2 i lJ(xW
On
dx
I
> 2" ~
m

j-I
i I(Oj - bj(x»)f(xW
R.
dx - C 4,,2 i On
If(x)j2 dx.

The first assertion follows from this immediately. The second assertion
can be proved analogously if we replace 11.. by a function K.. E Cooo(lRm)
such that K.. (x) = 1 for Ixl "n and KII(x) = 0 for Ixl > 2n. 0
Auxlllary theorem 10.27. Let T be defined as in Theorem 110.23. For every
f E Cooo(Rm) there is a C > 0 such that
II~( fJ)11 " C {II Tfll + IIfll} for all f E Cooo(lRm).

PROOF. By (10.4) and Auxiliary theorem 10.26 we have for f E Cooo(lRm)

that

1I~(fJ)1I = II T(fJ) + 2)~l bj O/fJ) - (b 2+ i div b + q)fJ11

= ~1'}Tf- (~f)f +2)~1 (bjf + Djf) OJ- (b 2+:, div b + q)f~1

" IIfTfll + 1I(~f)fll + 2L~1 {bjf + Ojf) 0~1
+ lI(b 2 + i div b)ffll + IIq1'}fll
"CI{IITfll+llfll} + IIq1'}fll,
where C1 depends on * (more precisely, on the supremum of 1*1, Igrad *1.
*).
*
I~f I and on the supremum of b2 and Idiv bl on supp If we set
qo(x) = q(x) for x E supp and qo(x) = 0 for x f1. supp f, then qo obviously
belongs to Mp(Rm) for some p < 4, and by Theorem 10.17 and by the
inequality II gll2 " lI~gll + II gil
II q(*J) II = IIqo(*f)1I ..;;; tll~(1'}J)1I + C 2111'}fll,
and thus

The assertion follows from this. o

PROOF OF THEOREM 10.23.
(a) First we prove the essential self-adjointness of T. Take a cp E Cooo(lRm)
such that .
I for Ixl.;~ 1,·
o " cp(x) ..;;; 1, cp(x) = { 0 for Ixl ;.2.
10.4 Essentially self-adjoint Schrodinger operators 319

For every n E N let <pix) -= <p(n -IX), and let T" be defined similarly to
T provided that we replace the functions bj and q, by the functions
bj,,,(x) = <P3"(X)~(x), q/,,,(x) = <P3,,(X)q,(x),
Then it is obvious that for all n E Nand f E CoOO(Rm)
<p" Tf = <p" T"J and T( <p,J) = T,,( <p,J),
Now assume that g E R(i - T).l, We have to show that g ==0. Since the
operators T" are essentially self-adjoint by Theorem 10.22, for every
n E N there exists an f" E CoOO(lRm) for which

1I<p"g - (i - T,,)f,,11 <; !.

n
Then we have in particular that (d. Theorem 5.1S)

IIf,,1I <; lI(i - T,,)f,,11 <; Ii<p"gll +!.

n
Therefore, the sequence (in) is bounded. Auxiliary theorem 10.26
implies that

f f f
m
~ I(Dj - b) !"12 dx <; 2 IT!n1 2 dx + C In 2 11,,1 2dx
j-1 R. Q. Q.

= 2f IT,J,,1 2 dx + C,n2f 1!"12 dx

Q. Q.

= 2 fQ.
l(i- T")!,, - <p"g + (<p"g-if,,)jZ dx

<; 4n- 2 + 4f l<p"g - if,,12 dx+C,n 2

Q.
fQ.
1!,,12 dx
<; 4n -2 + SII<p"gIl2 + C2n2 11!,,112
<; C3{ 1 + n211<PngIl 2) <; C3 (1 + nll<p"gll)2,
and thus (because IDj<p,,1 <; Kn-' with K= sup (Igrad <p(x)1 : x E Rm})

IIj~1 {Dj<p,,)(Dj - bj, n)!"ll = L~, (Dj<p,,)(Dj - bj)!,,11

= {f.I.f
R. } - I
(Dj<p,,)(Dj - b)J,,12 dX} 1/2
320 \0 Differential operators on L2(R m )

Consequently, because of the equality <g, (i - T,,)( f/J,J,,) >= <g, (i - T)

(f/J,J,,» - 0 we obtain the estimate

*
1I'P"gIl2 = <'Png, (i- Tn)Jn +['Png-(i- Tn)!,,]>

<;; I<g, 'Pn(i- T,,)!,,>I + 1I'P"gll

= I<g, (i - T,,)('P"!,,) + 2 ~ (O)'P,,)(O) -

}-I
b), ")!,, - (tlf/J,,)J ,,>1

I
+ -;;1I'Pngll

<; {~.I gl2 dX} 1/2 { 2t~1 (O)'Pn)(O) - bj • n)j"I1+ II (tlqJn)!n II }

I
+ -;; lIf/Jngll

<;; C6 { ~.I gl2 dx }

1/2 { I I } + -;;IlI'Pngll·
-;; + lI'Pngll + n211!nll

Since the left side tends to II gll2 and the right side tend:i to 0 as n~oo,
it follows from this that g =: O. Hence, R(i - T).L = {O}. We can prove
similarly that R( - i - T).L = {O}. The operator T is therefore essen-
tially self-adjoint.
(b) If JEL 2(R m)n W2• 2. loc(R m) and ~j_I(Oj-b):r+qfEL2(Rm), then
! E O(T·) = O(T) and
m
T· J = TJ = L (OJ - b/! + qf,
)-1

because for every g E Cooo(Rm) = O( T) with supp g c {x E Rm : Ixl <; k}

and for any {to E CoOO(Rm) with {J(x) = I for Ixl <;; k we have
m
<1, Tg> = <{toJ, Tg> = <L (0) - b))2( {Jj) + q{toJ, g>
)-1
m
= <L (0)- bjfJ+ qf, g>.
)-1

Now letJ E O(T). There exists a sequence (fn) from C(r"(R m) for which
In ~J and TJn-+ TJ. In particular, (fn) and (Tf,,) are Cauchy sequences.
Then (tl( {toJ,,» is also a Cauchy sequence for every {J E Cooo(Rm) by
Auxiliary theorem 10.27. Since IIhlb <; IItlhll + IIhll for hE C.f(Rm), the
sequence ({JJ,,) is also a Cauchy sequence in W2• 2(R"'). Let g be the
limit of the sequence ({JJ,,) in W2,2(Rm ). Since {JJ,,-+l"if in L2(R m ), we
then have {JJ= g E W2• 2(R m ).
10.4 Essentially self-adjoint SchrOdinger operators 321

For an arbitrary r > 0 let ~ E Co""(/R m ) be such that ~(x) = 1 for

Ixl < r. Then for all g E Co""({x E /Rm : Ixl <r})

(ij, g) = (T*j, g) = (f, Tg) = (~j, Tg)

m
= ( ~ (Dj - b)2(~j) + qi}j, g)
j-I

As the set of these g is dense in L2({xE/Rm : Ixl <r}), it follows that

m
fj(x) = ~ (Dj - bix»2j(x) + q(x)j(x)
j-I

almost everywhere in {XE/Rffl : Ixl <r}. Since r was arbitrary, this

equality holds almost everywhere in /R m • In particular, ~j_I(Dj - b}1
+ qf belongs to L2(/R m ). 0
REMARK. The operator T of Theorem 10.23 is still essentially self-adjoint if
q= ql + q2 with

ql E Mp(/R m ) for some p < 4, q2 E L2, loc(/Rm),

q2(X);> - Clxl 2 for some C ;> 0 and aU x E /Rm .

The above proof can then be employed without change if we first show
that the corresponding operators Tn are essentially self-adjoint. A proof is
given, for example, by CG. Simader (51). The reader can find further
references concerning this circle of problems there.
In order to apply the results of Section 9.2, it is useful to have criteria for
the T-compactness and the T 2 -compactness of a perturbation of T. We
shall prove such a criterion now. If A and B are operators on L2(/Rm), then
we say that A is B-smal/ at injinity if A is B-bounded and for every t: > 0
there exists an r(t:);> 0 such that

IIAjll .;;; t:(11 Bfll + Ilfll)

for alljE D(B) such thatj(x)=O for Ixl <r(t:}.
Theorem 10.28. Let T be dejined as in Theorem 10.23. Ij V is closed and
T-smal/ at injini!l, then V is f 2-compact. If, in addition, the T-bound oj V is
zero, then V is T-compact.
PROOF. If V is T-bounded with T-bound zero, then it is also f-bounded
with f-bound zero. Consequently, the second statement follows from the
first with the aid of Theorem 9.lld. It remains to prove the first statement.
322 10 Differential operators on L2(Rm)

First step: The operator A defined by the formulae

OCA) -oCr) and Af = ~f for f E OCA)

is T-compact for every ~ E CoOO(Rm).
Proof of the first step. It is sufficient to prove that A is T-compact. Let tf
be an element of CoOO(Rm) such that tf(x) = 1 for x E supp i/o. The operator
B defined by the equalities
DeB) =O(T) and Bf = ~(tff) for f E D(B)
is T -bounded by Auxiliary theorem 10.27. Since Af = A (tff) and since A is
~-compact by Theorem 10.21, the operator A is B-compact, and thus also
T-compact (cf. the proposition preceding Theorem 9.7).
Second step: The operator C defined by the equalities

O(C)=O(T) and Cf=~Tf-T(~f) for fED(T)

is T-compact for every 1') E Cooo(Rm).
Proof of the second step. Again, it is sufficient to show that C is
T-compact. Since C is a differential operator of the first order having
continuous coefficients with compact support, the B-compactness and thus
also the T-compactness of C follow as in the first step.
Third step: V is T2-compact.
Proof of the third step. It is clear that V is also T-small at infinity. Let
t: > 0 be given, and let r( t:) be chosen according to the definition of
T-smallness at infinity. Moreover, let ~ be an element of Cooo(Rm) such
that ~(x)-l for Ixl<;r(t:) and O<;~(x)" I for all xERm. Since V is
T-bounded by assumption, we have for all f E o(f) with appropriate
a, b;;. 0 that
IIVfl1 <; IIV( ~f)lI + IIV[ (I - 1')f] II
<; all1')fll + bIlT(1')f)1I + t:{II(l-1')fll+IIT[(l-1')f]lI}
" all,'}fll + bll~Tfll + (b + €)II1')Tf - T(~f)1I
+t:{II(I -,'})ill + 11(1- 1')Till}.
If we replace here i by a sequence Un) from D( T2) for which in ~ 0 and
T'fn ~ 0, then we also have Tin ~ 0, and by steps 1 and 2 the first three
terms of the ri~t side converge to zero. The fourth term is bounded by
t: sup (lIfnll + II TfniD. Since € > 0 was arbitrary, it follows that II Vfnll~O,
and thus V is T 2-compact. 0

EXERCISE

10.11. Let T be the self-adjoint Schrodinger operator from Theorem 10.23 for
m<:3. Then (z-T)-I is a Carleman operator for every zEp(T). The
operator E(b) - E(a) is a Carleman operator for all a, bE IR.
Hint: M,(z - T)-I is a Carleman operator for i} E Co"'(Rm ).
10.S Spectra of Schrodinger operators 323

10.5 Spectra of Schrodinger operators

In this section we prove some properties of the spectra of self-adjoint
operators of the form considered in the previous section. The results can be
applied equally well to one particle and several particle Schrodinger
operators (Theorems 10. 29{a), 10.30 and 10.33).
In what follows let T always have the form
m
Tf = ~ (OJ - bYf + qf for f E D( T),
j-I
(10.8)
where the coefficients satisfy the assumptions of Theorem 10.23. Therefore,
T is essentially self-adjoint. We denote by S the closure of T (that is, the
uniquely determined self-adjoint extension of T). Moreover, let the self-
adjoint operator So be defined by (cf. Section 10.2)
D(So) -=W2. 2(Rm) and Sof= -aj for fED(So), (10.9)
and let V be defined by the formulae
D( V) ... D(S) n W2• 2(R m),
m
(10.10)
Vf= Sf-Sof= -2~ bj OJ+ (b 2 +idivb+q)f.
j-I

Theorem 10.29.
(a) If q _ E Mp(Rm) for some p < 4, then S is bounded from below (here
q _(x) = max (- q(x), 0». If q > 0, then S ;s non-negative.
(b) If b2 E Mp(R m), div bE Mp(Rm), q E Mp(Rm) for some p < 4, and

1 (b 4(Y)+ldivb(Y)12+lq(Y)12)dy-+O as
Ix->,1.,; 1
Ixl-oo,
then V is relatively compact with respect to So, and Pe(S) =[0, 00).
PROOF.
(a) By Auxiliary theorem 10.25 with "1 = 1 we have for allf E Cooo(Rm) that
m
<f, Sf) = ~ /I (OJ - b)fll 2 + <f, qf)
j-I
m
> ~ II(Oj-b)fIl 2 -<q_J,f) > -CdlfIl2.
j-I

This then holds for all J E D(S), since Cooo(Rm) is a core of S. If q ;;. 0,
then q _ ... O. The above estimate then gives that S ;;. O.
(b) We have bj E Ma(Rm) for some C1 < 2 (j = 1, 2, ... , m) by Auxiliary
theorem 10.24. Consequently, the assertion follows from Theorem
10.21 together with (10.10) and Theorem 10.11. 0
 ,
324 10 Differential operators on L 2(Rm)

The spectrum of the operator So is [0, (0); the operator So has no

. eigenvalue. We can actually prove for a large class of operators S of the
form given above (without magnetic field) that no eigenvalues lie in [0, (0).

Theorem 10.30. Assume that 0 = 0 fQr j = 1, 2, ... , m, q E M peRm) fQr SQme

p<4, and
q(ax) = a-1'q(x) fQr x E Rm,,{O}
with SQme y E (0, 2). If f is an eigenelement .of S belQnging to the eigenvalue
X, then
2(f, - 6.f) + y(f, qf) = 0 (virial theQrem).
The interval [0, (0) cQntains nQ eigenvalue .of S. (FQr a sQmewhat mQre
general result we refer tQ J. Weidmann [55].)

PROOF. If f is an eigenelement of S belonging to the eigenvalue X, then

- 6.f(x) = Xf(x) - q(x)f(x). (10.11)
It follows from this for every a> 0 using the notation f(J(x) = f(ax) that
- 6.fa(x) = - a2(~f)(ax) = a 2(Xj(ax) - q(ax)f(ax»
= a 2N'a(x) - a Z -1'q(x)fa(x). (10.12)
It follows from (10.11) and (10.12) that
0= ( - - (f, -6.fa)
~f,Ja)

= X(f,fa) - (qf,Ja) - a 2X(f,Ja) + a Z -1'(qf,Ja)

= (1- a 2 )X<J,fa) + (a 2-1' -1)(qf,Ja)'

For a =1= 1 we obtain by dividing by (a - 1) that

a 2 -1' -1
(a+ I)X(f,fa) = a-I (qf,Ja)'
By letting a tend to 1, it follows from this that
2Xllfl1 2 = (2- y)(f, qf).
Since - ~f + qf = Xj, this implies
2<J, -6.f) = -2<J, qf) + 2Xllfll 2 = - y<J, qf)
(this is the virial theorem) and
yXIl fll2 = (2 - y)<J, qf - N') = (2 - y)(f, 6.f).
It follows from the last equality that A < 0 provided that f =1= O. 0
REMARK. The assumptions of Theorem 10.30 are satisfied in particular by
every Schrodinger operator without magnetic field with pure Coulomb
interaction (this also holds for many-particle operators).
10.S Spectra of Schrodinger operators 325

The operators of Theorem 10. 29(b) also satisfy the assumptions of part
(a); the negative part of the spectrum therefore consists of at most
countably many eigenvalues of finite multiplicity that are bounded from
below and can only cluster at O. The following theorem shows that in
many interesting cases there actually exist infinitely many negative eigen-
values.

Theorem 10.31. Let S be as in Theorem 1O.29(b). Assume further that there

exist constants C ;> 0, e> 0 and r ;> 0 such that

q(x) < - Cixl-2+· for all x E Rm with Ixl;> r,

and

Then S has infinitely many negative eigenvalues accumulating at zero.

PROOF. We only have to prove that S has infinitely many negative

eigenvalues. According to Theorem 7.26(b) it is sufficient to find an
infinite-dimensional subspace M of D(S) with <f, Sf) < 0 for every non-
vanishing f EM. Let * E COO(Rm ) such that Ii* Ii = I and supp {} c
{XERm: I <lxl<2}. Then for the function {}lx)=t m / 2{}(t- 1X) we have
11*,11 = I and supp *, C {x E Rm : t < Ixl < 2t}. Therefore it follows for t >r
that
m
<{}" S{},) = <{}" q{}, - Ll*, + (b 2 + i div b ){}, - 2 :L bjDi},)
j=1

< -0-2+<11*,11 2 + t- 2<{}, -Ll{}) + t- 2 +p(t)II{},W

m
+2t- I+«/2)p(t)I/2 L <{}" Di},)
j=1

<- 0- 2 +< + t- 2 <{}, -Ll{}) + t-2+'p(t)

m
+2t- 2+«/2)p(t)I/2 L <{}, Dj *).
j=1

(Here we have used that Iblx)l< t - I +«/2p(t)1/2 for Ixl ;> t, which follows
from the definition of p.) Therefore there exists a to;> 0 such that
<*" S{},) <0 for t;> to'
Let now In = {}2"t o' n E No. These functions have mutually disjoint supports.
Therefore M = LUll: n E No} is infinite-dimensional and <f, Sf) < 0 for
all f E M, f =1= O. This implies the theorem. 0

Now we want to show that the smallest eigenvalue of a Schrodinger

operator (without magnetic field) is always simple, i.e., that the system has
a uniquely determined ground state. The proof is essentially taken from a
work of W. Faris [41]. We need some preparation.
326 10 Differential operators on L2(Rm)

following for an Lz(Rm) we write

f is real and > 0) almost .."","",:r"",r..
"""AU"""" are said to be (positive). A
said to be if f ;;;. 0 implies
said to be positivity improving if f ;> 0 and f =F 0 imply Af > O.

Theorem 10.32. Let A E 8(Ll(Rm» be a real 3 positivity improving self-

adjoint operator. Assume that IIA II is an eigenvalue of A. Then the multiplic-
ity of the eigenvalue IIA II equals 1 and there is an f> 0 that spans the
II-A).

"""'YU'''' that 1=F 0 and Since A is real,

(otherwise we 1 by Re f or
(j+Kf)/2= = (IIAlll+ =
IIA A(Imf)= IIA From the inequality it
follows that ±Al<Alfl. Therefore, IAfl<AIfI, and thus
(f, AI) < <111, IAfl) < <IfI, Alfl)·
This implies that
IIAlillfill = (f,Af) < <lfl,AIfI) < IIAlllllfll12 = II A llllfll 2,
<f, Ifl)·
Let f+ and f _ by the
f+(x) =

Then If I= 1+ + f _. Consequently,

(f+, Af- >= H<111, Alii) - <f, Af)} = O.

we have f+ =0 or 1- =0, since f + =1= 0 and f _=F 0 imply

thus that <1+, Af_ Consequently, we have
We can assume, or generality,
and 1=F 0, it then we even havef
"""nnln\! improving.

will be proved that f spans the

A). For every element g of NOlA 11- A) the functions Re g and 1m g do not
change sign. Such an element can only be orthogonal to the positive
elementf if g = O. Therefore, N(IIA 11- A) = L(f). 0

Theorem 10.33. Let S be defined as above with =0 (j= 1, 2, ... , m),

q E Mp , loc(R m) for some is bounded from
below. If the lowest o(S) is an eigenvalue, is simple.

3 Here "real" refers conjugation K on =J (x)* (cf. Section

8.1, Example I),
10.5 Spectra of Schrodinger operators 327

PROOF. By Theorem 1O.29(a) the operators S and So - q _ are bounded

from below. The lower bound of So - q _ is, at the same time, a lower
bound of the operators Sn and S - Qn used in steps 2 and 3. These
operators therefore have a common lower bound, so that Theorem 9.18(b)
can be applied.
First step: Let So be defined as above. Then the operator exp( - tSO> is
order improving for all t > o.
Proof of the first step. With -t1"t(x) = exp( - tlxl 2 ) for x E ~m we have
exp( - ISo) = F - 1M" F, I

where M;;I denotes the operator of multiplication by lit' Hence, by the

convolution theorem (Theorem 10.7) the operator exp( - tSo) is equal to
the operator of convolution by the function F -I{},. With {}(x) =
exp( -lxI 2 /2) we obtain from Theorem 10.2 that

( F - l{}J( x) J
= (2'7T) - m/2 eiX)'{},(y) dy

= (2 'IT ) JeiX)'.n-( v'2i y) dy

- m/2

= (2/)-m/2(2'11rm/2 Jexp(i vit z ){}(z) dz

= (2t)-m/2(F-I{})(_X_) = (2/)-m/2{}(_X_) >0

v'2i V2i
for all x E ~m and t > O. Since the operator of convolution by a positive
function is obviously positivity improving, the assertion follows.

Second slep: exp( - tS) is positivity preserving for all t;> O.

Proof of the second step. For every n E N let qn be defined by the
equality
if Iq(x)l.;;; n
if Iq(x)1 >n.

r
Let S" be the operator defined by q" instead of q. By Theorem 7.41

exp( - tS") = s - 2~n;, [exp ( - I so) exp ( - -f Qn)

for all I > 0, where Qn denotes the operator of multiplication by qn' Since
every term on the right side is order preserving, it follows from this that
exp( - tS") is also positivity preserving. Since by Theorem 9.16(i) (with
Do=Coao(~m» we have (i-S")-I~(i-S)-l, it follows by Theorem
9.l8(b) that
exp( - tS") ~ exp( - IS) for all t > O.
Therefore, exp( - IS) is also positivity preserving for all I;;' O.
328 10 Differential operators on L2(R"')

°
Third slep: If f > 0, f 7'= 0, g > and g 7'= 0, then there exists a I > Osuch
that <f, exp( - tS)g) > 0.
Proof of the third slep. It is sufficient to prove that if f > and f 7'= 0,
then
°
KU) = {g EL2(~m) : g> 0, <g, exp( -IS)f) =0 for all t > O}

itself by exp( - sS) for s> 0, since s, I> °

contains only the zero element. The set K(f) is closed. It is mapped into
and g E K(f) imply that
<exp( - sS)g, exp( - IS)f) = < g, exp[ - (s + t)S)f) = 0. This then holds
for exp(sQn)' as well: It follows from f > 0, g > 0, < g, exp( - tS)f) = and
°
exp( - tS)f > (d. step 2) that g(x)[exp( - tS)J](x) = almost every- ° °

r.
where; we also have then that <exp(sQn)g, exp( - tS)f) = 0. Since

exp( - t(S - Qn» =S- 1~~ [ex p ( - f s) exp ( f Q,,)

the operator exp( - t(S - Q,,» also maps the closed set K(f) into itself.
Since, moreover,

exp( - t( S - Q,,» ~ exp( - tSo) for all t > 0,

this follows also for exp( - ISo). If g E K(f), then we therefore have
°
<g, exp( - ISo)!) = for all I> 0. Because exp( - ISo) is positivity improv-
ing, it follows from this that g = 0.
Fourlh step: If X E ~ is smaller than the lower bound of S, then
(S - X)-I is positivity improving.
Proof of Ihe fourth step. If y is the lower bound of S, then
lIe- I (S-X)1I '" e-I(Y-A) for t > 0.
Consequently, the following integrals exist. For all f, g E L2(~m)

foooeAt<g, e-,sf> dt = 10 00
<g, e-I(S-X)f) dt

= Lo 1IA.oo)
oo
ds<g, E(s)f) dt
e-/(S-X)

= 1 LOOe-I(S-A) dt ds<g,E(s)f)
[A. 00) 0

= 1IA,oo)(S-X)-I ds<g, E(s)f) <g, (S-X)-I f ). =

° °
If f > 0, f 7'= 0, g > 0 and g 7'= 0, then the second and third steps imply that
<g, e-ISf) > for all I> and <g, e-'oSf) > 0 for some to> O. Since the
function t~<g, e-'Sf) is continuous on [0, 00), it follows from this that
< g, (S - X) -1) > O. Therefore, (S - X) - I is positivity improving (if we had
(S-X)-1}0 for somef>O such thatf7'=O, then there would be a g>O
such that g7'=O and g(x)=O for all x with (S-X)-1(x)7'=0; we would
therefore have <g, (S-X)-~> =0).
10.6 Dirac operators 329

Fifth step: The smallest eigenvalue of S is simple.

Proof of the fifth step. Since the lowest point y of o(S) is an eigenvalue,
II(S-X)-lll =(y-X)-l is an eigenvalue of the positive operator (S-X)-·.
This eigenvalue is simple by Theorem 10.32 and the fourth step. Conse-
quently, the assertion follows for S. 0

EXERCISE

10.12. Let T be defined as in Theorem 10.23, and assume that q(X)-HX> as Ixl~oo.
Then T has a pure discrete spectrum.
Hint: The identity operator is T-small at infinity and its T-bound is equal
to O. The compactness of (i - fr-·
follows from this by Theorems 10.28,
9.1 I(d) and 9.10.

10.6 Dirac operators

In this section we consider the Hilbert space

L2(R 3)4 = L2(R 3) EB L2(R 3) EB LiR3) EB L2(R 3).

The elements of this space are the 4-tuples f = (f.,J2,J3' f4) of elements
Jj E L2(R 3 ), and the scalar product is defined by the equality
4
<f, g> = ~
j=l
j /i(x)*8j(x) dx.
t
The elements of L2 (R 3 may also be considered as equivalence classes of
functionsf: R3~C4; then

<f, g) = jU(x), g(x» dx,

where (. , .) is ihe usual scalar product in C 4 «~, 11) = ~j-.~/1/.; for ~, 11 E
C 4 ; the corresponding norm in C4 will be denoted by 1. I). For any
continuously differentiable function f: R3 ~ C 4 , f(x) = (f. (x), fix),
fix), fix» let us define

(Dj)(x) =i I( a a a
3x f .(x), ax f2 (x), ax/ix ), 3x/4(x)
a ) j = 1,2,3.
J J :J :J

For an arbitrary function f: R3 ~C4 and a 4 X 4 matrix-valued function y

we define yf by the equality

(yf)(x) = y(x)f(x) = (
k-l
± Ylk(x)fk(x), ... , ±
k=l
Y4k(X)fk (X»);

we use a similar definition if y is a constant matrix. The operator norm of

4 X 4 matrices corresponding to the norm 1. 1on C 4 will be denoted by 1. I,
as well.
330 10 Differential operators on Ll(Rm)

The free Dirac operator (which describes the free electron in relativistic
quantum mechanics) is defined on L2(R 3)4 by the differential form
3
Tf(x) = ~ (Xj Dj(x) + f3f(X)
j-I
with the matrices

~) for j = 1, 2, 3, fJ = [6 ~ g g]
gg -6 -~
6)' °2 = (~ -~), °3 = (6 _~) (the Pauli matrices).

Here we have made all physical constants 1 again. Since the matrices (Xj
and fJ are Hermitian, the operator To defined by

D(To) =CO (R3)"

OO and Tof = Tf for f E D(To)
is symmetric on L2(R 3)4, as a simple computation shows.

Theorem 10.34. The operator To defined above is essentially self-adjoint. For

T= To we have:
(i) D(T) = W2• I(R 3)4 and II TJII = IIflll Jor J E D(T),4
(ii) o(T) = (- 00, -1] u (1, 00),
(iii) T has no eigenvalue.

PROOF.
(i) Let TI be defined by

D(TI ) == S(R3)" and Td"" TJ for f E D(TI ).

It is clear that TI c To (cf. the proof of Theorem 10. 11 (a», and thus
T - To =- T I • On the other hand,
FTIF- I == M p • I ,
where

and
0 X3 XI - i X2
0 XI +i X2 -X3
P(x) = -1 0
X3 XI - i x2
XI + i x2 -X3 0 -1

4 In what follows we also write II • 111' and II • 11(1') for the norms in W2.p(Rm)4 and L1•p (R"')4,
respectively.
10.6 Dirac operators 331

(Here the Fourier transformation is applied to the elements of L l(1R3t

component-wise.) Mp. 1- Mp, where Mp denotes the maximal operator of
multiplication by the matrix-valued function P (cf. the proof of Theorem
1O.1l(a»:
D(Mp) - {iEL2(R3t: PjEL2(R3t}, Mpj- Pj.
The reader can easily verify that the eigenvalues pix) and the correspond-
ing orthogonal eigenelements ~(x) of the matrix P(x) are given by the
formulae (observe that OJ Ok + Ok OJ = 2 8jk ( ~) 6
PI(X) - P2(X) = - P3(X) == - P4(X) - (1 + IxI2)1/2,

(6) (~)
et(x) - O'tXI + 0'2 X2 + 0'3 X3 ( 1) , ez{x)- O'IX, + 0'2 X2 + 0'3 X3 (0) ,
1+ (1 + IxI2)1/2 0 1 + (I + IxI2)t/2 I
_ 0IXI + 0'2 X2+ 03 X3 ( I) _ 0IX, + 02 X2 + 0'3 X3 (0)
lix)-
1+ (I + IxI2)'/2 , e4(x)= 0 1+ (1 + IxI2)1/2 I

(~) (~)
The corresponding normalized eigenelements are

e/x) - [2 + 21xI2 + 2(1 + IxI2)1/2]'/2[ 2 + IxI 2 + 2(1 + IxI2)'/2r '/2~(x).

Consequently,
4
(MpJ)(x) == ~ p/x)(f1(x),j(x»eix) for all j D(Mp).

r
E
}-1
It follows that

D( M,) = {f E L,(R')' : [j;/;(x)'I( ~(x), f(x»)I' E L,(R')}

= {j E L2(Rl)4 : (I + I . 12)1/2f E L2(R3)4} = L 2• 1(R 3 )4,

D(T) ... F-'D(Mp) == F- 1L2• I (R l )4 = W2• I (R l t,

Tf .. F-'MpFj for f E D(T),
and
IITfll = IIPFfl1 = 11(1 +1·1 2)'/2 Ffll == IIFfll(l) = IIflll'
(ii) If A> I, then there is an Xo E Rl for which p,(xo) == A. Since PI is
continuous, for every £ > 0 there is a ball K. around Xo such that
Ip,(x) - AI <; t: for all x E I{.. If ." E L2(R 3) is such that 1/ "" 0 and
332 10 Differential operators on L2(R m )

lJ(X) = 0 for x fI. K., then for f = lJel E L2{1R 3 )4

II(X- Mp)fll2 = fiX - PI(xWllJ(xW dx <;; f211lJII2 = f21if1l2.

Since f > 0 was arbitrary, X- Mp is not continuously invertible. If
X <;; - 1, then we can prove this analogously if we use P3 and e3 instead
of PI and e l • Therefore, (- 00, -l]u [1,00) c o(Mp) = oCT).
If now X E ( - I, I), then
4
II(X- Mp)fli 2 = L fiX - Pj(xWI(ei x ),f(x»)j2 dx
j-I
4
~ (t-IXI)2 L fl(ei x ),f(x»1 2 dx = (1-IAlillfIl 2
j=1
for all j E LiIR 3)4. Consequently, X- M p is continuously invertible.
Hence, (- 1, I) C peT) = p(Mp), and thus oCT) = (- 00, -1] U [I, 00).
(iii) If (X - Mp)f= 0, then we must have
(X-pix»(ej(x),f(x» = 0 almost everywhere (j= 1, 2, 3, 4).
As the set of zeros of X- Jlj(x) is a null set (it is either empty, consists
of the origin, or is a sphere), we must have
(eix),j(x» = 0 almost everywhere (j=1,2,3,4).
It follows from this (as e/x) is a basis in C 4) that f(x) = 0 almost
everywhere; therefore, f= O. Consequently, no X can be an eigenvalue
of M p (and hence of T). 0
The Dirac operator of an electron in the electric field with potential q
has the form
S = T+ Q,
where Q is the operator of multiplication by a function q : 1R3 --+ IR and
Qf"" qj= (qjl' qj2' qj3' qj4)· Because of physical reasons, the Coulomb
potential q(x) = c(l/Ix!) is particularly interesting. As this q does not
belong to Mp(1R3) for p < 2, the results of Section 10.3 cannot be applied to
it. The following auxiliary theorem enables us to treat such potentials. We
prove it with somewhat more generality than we need.

Auxiliary theorem 10.35. For f E Cooo(lRm) (m ~ 3) and for f E CoOO(R"- {O})

(m= 1)

fl xl- 21J(x)1 2 dx <;; 4 2 f

(m-2)
f
Iff(XW dx.
j=1 Xj
10.6 Dirac operators 333

For f E CO""(R2" (O})

fl x l- 2 1f(x)1 2 dx <;; 4fllnlxW.f If,f(xW dx.

}-I x}

PROOF. Let Um denote the unit sphere in Rm and let dWm be the area
element on Um' For m > 3 we have

= lim foo r
.~o •
m- 3 (
Jam
If(rwW dWm dr

- _2_
m-2 (
rm -foo 2 Re ( f(rw)*
Jam
~
vr
f(rw) dWm dr}.

Here the first term vanishes as £-+0, since the integral tends to f(O)! dWm
as £-+0. Therefore,

The assertion follows from this for m;;> 3. The case m = 1 can be proved
similarly though with certain simplifications. In the case m = 2 a logarith-
mic term arises after the integration by parts; everything else goes as
above. [J

Theorem 10.36. Assume that q = ql + q2' where ql and q2 are measurable

Hermitian 4 X 4 matrix-valued functions such that
1
Iql(x)1 <;; C IxT and qi.) E Mp(R3) for some p < 2.
Then Q is T-bounded with T-bound less than or equal to 2C. If C < 1/2,
t
then S = T + Q is essentially self-a4Joint on CO""(R3 and self-a4Joint on
W2• JCR3)4.
334 10 Differential operators on L2(R m )

PROOF. By Auxiliary theorem 10.35

3 4
= 4C 2 ~ ~ flxk{F,~)(XW dx
k-I j - I

for aUfE Cooo (1R 3)4. The operator QI induced by ql is therefore T-bounded
with T-bound" 2C. The operator Q2 induced by q2 has T-bound 0, since
D(T)= W2 , 1(1R3)4 (cf. Theorem 10.18). Consequently, the T-bound of Q=
QI + Q2 is not greater than 2C. The remaining assertions follow from
Theorem 5.28. 0

Theorem 10.37. Let q be as in Theorem 10.36. Assume, moreover, that

Nlql(x)~O as Ixl~oo. Then Q is T 2 -compact. If C< 1/2, then O'e(S)=
O'e(T) = (- 00, - I]u [I, 00). If, in addition, q" 0, then the eigenvalues of S
in (- I, 1) can only accumulate at I; if q ;) 0, then the eigenvalues of S in
( - I, 1) can only accumulate at - 1.

PROOF. We have T2= P-1M;P. Therefore,

Since Iql = Iql + q21 E Mp(1R3) for some p < 4, the operator Q is T2-compact
(cf. Theorem 10.21). (We can show in the same way that the operator Q2
induced by q2 is T-compact.) The remaining assertions follow from Theo-
rems 9.13, 9.14 and 10.36. 0
In analogy with the viria) theorem (Theorem 10.30), we can now prove
yet the following result.

Theorem 10.38. Let q be as in Theorem 10.36 with some C < 1/2. Moreover,
assume that q(ax) = q(x)/ a Jor all a> 0 and x E 1R3 " , {O}. Then S = T + Q
has no eigenvalue in ( - 00, - 1) U (1, 00). IJ q( x) is diagonal, then - I and 1
are not eigenvalues oj S, either.
10.6 Dirac operators 335

PROOF. If (;\- S)f=O, then the functionfa(x)= f(ax) obviously belongs to

D(S) and
3
(Sfo)(X) = :L ajDl(ax) + (13 + q(x»f(ax)
j=1
3
= a:L aAD))(ax) + (13 + q(x»f(ax)
j=1

= aL~l aj(Dl)(ax) +(13+ q(ax»f(ax) } + f3(l-a)f(ax)

= a(Sf)(ax) + 13(1- a)fo(x)
= aAf(ax) + f3(I - a)fa(x) = aAfaCx) + f3{l- a)1a(x).

It follows from this that

0= <Sf,Jo> - <1, Sfa> = ;\<f,fo) - a;\<f,fo> - (1- a)<f, f3fo>

= (1- a)<1, (;\ - f3)fa>·
For a =1= 1 we can divide by (I - a). Taking the limit as a~ 1 gives

o= <f, (;\ - f3)f>.

For ;\E(-oo, -I)u(l, 00) the matrix ;\-13 is strictly positive or

strictly negative; this equality is therefore possible only if f = O.
If ;\ = 1 and q is diagonal, then it follows first that i3 = f4 = O. The
eigenvalue equality (l - S)f = 0 then gives

This implies that il = f2 = 0, and thus f = O. We can argue similarly if

;\= -I. 0
To conclude this section, we give a general criterion for the essential
self-adjointness of Dirac operators (cr. also Exercise 10.13).

Theorem 10.39. Assume that q = ql + q2' where ql and q2 are measurable

Hermitian 4 X 4 matrix-valued junctions such that
I
Iql(X)1 C;;; Cr;r jor some

Iq2( . )1 E Mp , loc(R3) jor some

336 10 Differential operators on L2(R"')

Then S = T +Q is essentially self-adjoint on Cooo (1R3t. The closure S is given

by the formulae

D(S) = {f E L2 (1R 3)4 n W2• 1.loc(1R3)4 : 7'f + qf E L2(1R3)4} ,

Sf = 7'f + qf for f E D(S).

PROOF. Let epn be as in the proof of Theorem 10.23, let lin = CPJ"q, and let
S" = T + Q", where Q" is the operator of multiplication by li". The function
li" satisfies the assumption of Theorem 10.36 for every n E N; the operator
S" is therefore essentially self-adjoint on Cooo(1R 3)4.
The rest of the proof follows the proof of Theorem 10.23; however, it is
essentially simpler: Let g E R(i - S)1.. Since S" is essentially self-adjoint,
for every n E N there exists an f" such that

Hence, IIf,,11 .;;; lI(i - S,,)j,,11 ,; ; lIep"gll + (lin). It follows from this that

lIep"gl12 = <ep" g, (i - S")f,, + (ep"g - (i - S,,)f..] >

.;;; I<g, ep,,(i - S,,)f,,>1 + ~ lIep"gll
3 I
.;;; I<g, (i - S)(ep"!,,»1 + CIII gil L Ilf" DJ'Pnll + -;; IIcp"gll
j=l

I
.;;; C2 -(11
n
gil II f" II + Ilep"gll) ~ 0 as n~ 00.

This implies that g = O. Consequently, R(i - S)= LiIR 3)4. The equality
R( - i - S)= L2(1R 3)4 follows analogously. The rest of the proof is analogous
to part (b) of the proof of Theorem 10.23. 0

EXERCISE

10.13. (a) In the first part of Theorem 10.36 the assumption on q. can be replaced
by the assumption Iq.(x)l.;; ~7'_.cjlx - ajl-· with N different points
aJ EA 3 •
(b) In the second part of Theorem 10.36, in Theorem lO.37, and in Theorem
10.39 the assumption on q. can be replaced by the assumption Iq.(x)1 <;
~J_.c)x-ajl-' with N different points aj ER 3 and O<;Cj<t for j ....
I, ... , N.
(c) In Theorem 10.39 we can allow an infinite sum of the above form for
the bound of q. provided that the aj have no accumulation point.
Hint: If " E Co""(R"'), "(x) = 1 for Ixl.;; s /2, "(x) = 0 for Ixl;;. s, and
0.;; "(x) .;; 1 for all x E R 3 , then there is a C ;;. 0 such that

jlxl- 2 1"(x)!(xW dx .;; 41 L

3

Ixl<s j_1
ID./(xW dx + C 1
Ixl<s
1!(x)j2 dx.
 11
Scattering theory

11.1 Wave operators

The theory of wave operators provides a useful means of studying the
absolutely continuous spectrum. We wish to present this theory briefly in
what follows. For any two self-adjoint operators TI and T2 on the complex
Hilbert space H we define O±(T2, T I) by the equalities

D(O±(T2' T I» = {fEH: lim eiIT'e-iIT'f exists},

I-+± 00

It is obvious that O±(T2• T I) are linear operators. Since eilT'e-ilT, is

unitary for all t E Ill. the operators O±(Tz' T.) are obviously isometric (as
»
mappings from D(O±(T2' T I into H).

Theorem n.t. The subspaces D(O±(T2' T I» are closed and reduce T I. The
subspaces R(O±(T2 • T.» are closed and reduce T z. We have R(O±(Tz' T I »
= D(O±(TI• T2» and O±(TI• T z)'" O±(T2• T.)-I. Moreover,for all s E III

and

where E. and E z are the spectral families of T. and T z• respectively. We have

O±(Tz• TI)u(T I) = u(Tz}O±(T2• T I) for every bounded continuous function
u : III ~ III (" intertwining" property).
337
338 11 Scattering theory

PROOF. First we show that D(Sl+(T2, T I» is closed. LetfE: D(Sl+(T2, T.».

We have to show that for every E > 0 there exists a 10 E R such that

for all s, t >10 , In order to prove this, we take an fo E D(Sl+(T2, T I» such

that IIf - foil <: E/3. Since lim,.... oo ei ITl e- ItT fo exists, there is a to E R such
that

for all s, I> to' Consequently, for all s, I> to

lIeilTl e-iITi- eiaTl e-i.Till <: lIeilT2 e-iIT'(j_ fo)1I
+lIeisT2 e-isT'(jo- f)1I + lIe ilT2 e-itTio - eisT2 e-iSTioll <: E.

We can show in a similar way that D(Sl_(T2' T.» is closed. As the

operators Sl%(T2, T.) are isometric, the ranges R(Sl%(T2, T.» are also
closed.
If g == Sl+(T2, TI)f E R(Sl+(T2, T.» with some f E D(Sl+(T2, T I then »,
lIeitT, e-itT'g-fll"'lIg-eiIT2e-iITill~O

as t~oo, i.e., g E D(Sl+(T., T2» and Sl+(T" T2) g- f. If g E D(Sl+(T., T 0)

and f- Sl+(T., T 2)g, then we can show similarly that f E D(Sl+(T2, T I»
and g - Sl+(T2, T.)f. Therefore, R(Sl+(T2' T.» == D(Sl+(T., T0) and
Sl+(TI, T2)-Sl+(T2, T I)-·. We can show analogously that R(Sl_(T2' T,»
-D(Sl_(T., T2» and Sl_(T., T 2)=Sl_(T2, T I)-·.
If f E D(Sl+(T2, T I» and g = Sl+(T2, T.)f, then for every s E R

»
Consequently, ei.Ti E D(Sl+(T2' T I for all s E R. By Theorem 7.39 the
operator T. is therefore reduced by D(Sl+(T2, T I». We can prove this
assertion for D(~L (T2' T,» similarly. Since R(Sl % (T2' T.» ==
D(Sl%(T" T 2», the operator T2 is reduced by R(Sl%(T2' T,).
For every s E R the operator eisT, maps the space D(Sl+(T2, T.» onto
itself (cf. Theorem 7.39). Since for all f E D(Sl + (T2' T.» and all s E R

Sl+(T2, T,) eisT'J = lim eilTl e-i(l-s)T'f

, .... 00

= eisTl lim ei(I-S)Tl e-i(l-s)T'f = eisT1Sl+(T2' T,)f,

, .... 00

it follows that Sl + (T2' T,) ei sT, = ei ST1Sl + (T2• Ti)' A corresponding ~quality
holds for Sl_(T2' T,).
11.1 Wave operators 339

For j = I, 2, 1m z > 0, and all], g E H

- ifooo ei ZI<1, e- i l7Jg ) dt = -i fooo ei zi {j e- i Is d<1, Ej(s)g) } dt

= -ij{foooei,(Z-S)dt} d<],E/s)g)

= j(Z-S)-l d<], ~(s)g)

= <1, (Z-1j)-l g ).

Similarly, for 1m z < 0

i{Ooo eiZl <1, e- il7Jg) dt = <1, (z _1j)-J g ).

Together with the representation (7.22) of the spectral family, it follows

therefore that O±(T2, T J)E J(s)=E2(s)O±(T2, TJ) for all sEIR. The last
assertion follows from this, because U(1j) = fU(/) d~(t). 0
REMARK. If ] is an eigenelement of T J for the eigenvalue A. then] E
D(O+(T2, T J» if and only if] is also an eigenelement of T2 for the same
eigenvalue. The same holds for O_(T2• T I ).
PROOF. If] E D(T2) and Td= Ai. then
ei IT, e - i IT'i = ei IT, e - i I"J = ei IA e - i I"J = ]
for all I E R Consequently,] E D(O ± (T2• TJ ». If ] ».
E D(O +(T2, T J then
for every s E R
lIeis(Tl-A)]_]II"'lIei(l+s)Tle-i('+s)A]_ei'Tle-i'''JII~O as t~oo

(since ei('+s)T2e-i(,+s)~and eiIT2e-il~converge to O+(T2, T J)!). Hence,

ei s(T.-A) ] =] for all s E IR.
This implies that] E D(T2 - X) = D(T2) and

o
»
In what follows we write M C T ,D(O±(T2, TJ if M is a closed subspace
»
of D(O:!:(T2 , TJ that reduces T J. If M c T ,D(O:!:(T2• T J» and P denotes
the orthogonal projection onto M, then we define
W±(T2• T J, P) = O±(T2 , TJ)P.
These operators are called generalized wave operators. In particular, we
write W:!:(T2 , T J} for W:!:(T2 , T J, I) provided that D(O:!:(T2 , T J}} = H. The
340 II Scattering theory

operators W ±(T2 , T,) are called wave operators. In the sequel we say that
W±(T2 , T I, P) exists if R(P)c T,D(Q.±(T2 , T I». (Some of the following
statements still hold when TI is not reduced by R(P).) The operators
W ±(T2, T" P) are obviously partially isometric with initial domain R(P)
and final domain R(W ±(T2 , T I, P». Now we prove a few simple proper-
ties of the generalized wave operators.

Theorem 11.2. Assume that T, and T2 are self-a4ioint operators on the

complex Hilbert space H, M, c T,D(Q.+(T2 , T,», P, is the orthogonal projec-
tion onto M,. and M2 = R(W+(T2 , T" PI»' Then M2 cTP«(!+(T" T 2», and
with the notation W + = W +(T2, T" PI) we have
eisT,W + = W + eisT , for all s E R
and
T2W+ = W+T,P,:> W+T,.
If P 2 is the orthogonal projection onto M2, then
T, W! = W! T 2 P 2 :> W! T 2 •
The operators T,IM, and T21M, are unitarily equivalent. If M, C Hc(TI) or
M, C Hac(T,), then M2 C Hc (T2) or M2 C Hac(T2)' respectively. A similar result
holds for W _(T2• T" PI)'
PROOF. Since M, is closed and is contained in D(Q.+(T2' T , », the subspace
M2 = Q.+(T2 , T,)M, is also closed. By Theorem 11.1

eisT,W+ = eisT'Q.+(T2' T,)P I = Q.+(T2 , T,) eL,T,p,

= Q. + (T2' T)P
, , e
isT , = WeisT,
+

for all s ERIn particular, ei ST>g E M2 for all g = W + f E M2 and all s E R,

i.e., M2 reduces Ticf. Theorem 7.39). The equality ei sT,W + = W + ei sT,
implies that

The restrictions of ei sT, and ei sT, to M, and M2, respectively, hence are
unitarily equivalent. This then follows also for their infintesimal genera-
tors, i.e., T 2 P2 = W + T,P, W!, and thus
T2W+ = T 2P 2W+ = W+T\P,W!W+ = W+T\P,:> W+TI'
By taking adjoints, we obtain via Theorem 4.19 that
W! T2 = Wt. Ti C (T2W +)* C (W + T 1)* = Tr W! = T\ W'!.
On the basis of what we have proved so far, T21M, and TdM, are unitarily
equivalent. If M\ c Hc(T,) or M, C Hac(T\). then T1IM, has a pure continu-
ous or a pure absolutely continuous spectrum, correspondingly. Then this
11.1 Wave operators 341

also holds for the operator T2 1M2 ' unitarily equivalent to TIIM" i.e., M2
C HAT2) or M2 C Hac (T2), respectively. 0

Theorem 11.3. Assume that TI and T2 are self-adjoint operators on the

»,
complex Hilbert space H, MI C T,D(Q+(T2, T I PI is the orthogonal projec-
tion onto M» M2 = R( W +(T2, Tp PI»' and P2 is the orthogonal projection
onto M2. Then with the notation W + = W +(T2, T I, PI) we have, as t~oo,
that

e- i tT2W + _ e- i tT'P I ~O, (11.1)

ei tT, e- i tT2W + ~ PI' ( 11.2)
eitT , e-itT2P2~ W!, ( 11.3)
( W + - I) e - i tT,p 1 ~ 0, ( 11.4)
( W! - I) e - i tT,p 1 ~ 0, ( 11.5)
eilT,W+ e-itT'~PI' (11.6)
eitT,W* e-itT,~p (11.7)
+ I'

(1 - P 2 ) e- i tT'P I ~O.
.
( 11.8)

We have W +(T I , T 2 , P2) = W +(T2, T!, PI)*. Similar results hold for W _ =
W _(T2 , T I , PI) as t~ - 00.

PROOF. (Il.l) follows from the definition of W + by multiplication by

e -i tT2. (11.2) follows from (11.1) by multiplication by ei tT,. If we multiply
(11.2) by W! from the right, then we obtain (11.3). Relation (11.4) follows
from (ILl), because e- i tT.W + = W + e- i tT, = W + PI e- i tT, =
W+ e-itTIPI' If we multiply (11.4) by W,!, then (11.5) follows, because
W!W+ e-itTIPI =P I e-itT'P I =e-itT,p l • Relations (11.6) and (11.7)
follow from (11.4) and (11.5), respectively, by multiplication by eitTI .
Finally, (11.8) follows from the equalities
11(1 - P2) e- i tT'Pdll = lIei tT'(I - P2) e- i tT'P.!1I
= 11(1 - P 2 ) ei tT. e -i tT'Pdll ~ 11(1 - P2 ) W + fll 0= o.

The following chain rule is useful in many investigations.

Theorem 11.4. Assume that Tp T 2, and T3 are self-adjoint operators on the

»,
complex Hilbert space H, MI C T,D(Q+(T2, T I M2 c T ,D(Q+(T3• TJ),1j is
the orthogonal projection onto '1 (j= I, 2), and R(W+(T2' T I• PI» C M2 •
342 1:1 Scattering theory

Then M, C D(0+(T3• T,» and

W +(T3 • T,. PI) = W +(T3 • T 2 • P 2 ) W +(T2 • T,. 1'\).
A similar result holds for W _ .
PROOF. For f E M,
eitT, e- itT'.! ... (eitT, e-UTl)(eitTl e-itT')f
= eilT, e-itT'O+(T2' T,)f
+ eitT) e- i tT, [ ei tT, e- i tT'.! - 0+(T2• T,)f]
--+ 0+(T3 • T 2)0+(T2• T,)f as t --+ ex:l.
since 0+(T2• T,)f= W +(T2• T,. P,)f E M2 and
ei tT, e -i tT'.! - 0+ (T2• T,)f --+ o.
Consequently. M, C D(0+(T3• T,» and
0+(T3• T,)f = 0+(T3• T 2)0+(T2• T,)f for f EM,.
Because of the inclusion 0+(T2• T,)M, e M2 we have
W +(T3 • T,. PI) = 0+(T3• T,)P, = (0+(T3 • T 2 )U+(T2 • T,)P,
=0+(T3• T 2)P20+(T2• T,)P,
= W+(T3• T 2• P 2 )W+(T2 • T,. PI). o
Theorem 11.5. Assume that TI and T2 are self-adjoint operators on the
complex Hilbert space H. M, and M2 are closed subspaces. J\ and P 2 are the
orthogonal projections onto MI and M2• respectively. If MI e rp(0+(T2• T I ».
M2 cT,D(O+(TI• T 2». R(W +(T2• T I, P,» e M2 and R(W +(TI, T 2 , PJ)
eM .. then

and
R(W+(T2• T I, PI» =M2' R(W+(T I• T 2• P2» =MI •
PROOF. We obviously have W+(T I, T I, PI)=P 1 and W+(T2 • T 2, P 2)=P2 •
It therefore follows from Theorem 11.4 that
W+(T I, T 2 , P 2 )W+(T2 , T I, PI) = PI'
W+(T2 , T"P,)W+(T" T 2.P2) = P2 -
Because of the formulae R( W +(T I, T 2 , P 2» e MI and PI =
W +(T2, T" P I )* W +(T2• T,. PI) it follows that
W +(T" T 2 , P 2 ) = P, W +(T" T 2 , P2 )
W +(T2• T" P,)* W +(T2 , T" PI) W +(T,. T 2, P 2)
= W +(T2• T" P,)* P2 •
11.2 The existence and completeness of wave operators 343

Because W +(T2 , Tp PI)· is a partial isometry with initial domain

R(W+(T2' Tp P I »cM2, we have
W +(TI , T 2, P2) = W +(T2• T I • PI)· P2 = W +(T2, T I, P I )*.

The equality R( W +(T2 , T I , P 2» = M2 follows from the fact that

W +(T2, T I , PI) = W +(TI , T 2, P 2)* is a partial isometry with final domain
M2. The equality R( W +(TI , T 2, P~) = M) follows from W +(TI> T 2• P~ =
W +(T2, T I , PI)· correspondingly. 0
On the basis of the remark after the end of Theorem 11.1, an eigenele-
ment f of T) belongs to D(~+(T2' T I » only if f is also an eigenelement of
T2 for the same eigenvalue. Consequently, it is unrealistic to expect that an
eigenelement belongs to D(~+ (T2' T\». We shall therefore always assume
in the sequel that M\ C Hc(T\). If M\ c Hc(T) n D(~+(T2' T I » and P J is the
orthogonal projection onto M\> then R(W+(T2, Tp PI»~ c Hc(T~ by Theo-
rem 11.2. Actually, for technical reasons we shall consider only subspaces
M\ of Hac( T I). Particularly, we obtain from Theorems 11.2 and 11.5:

Theorem 11.6. Let T\ and T2 be self-adjoint operators on the complex

Hilbert space H, and let lJ, ac be the orthogonal projections onto Hac(T.J) for
j= I. 2. If Hac(T\)CD(~+(T2' T I» and Hac(T~CD(~+(TJ' T~). then

R(W+(T2, T I, PI,ac» = Hac(T2), R(W+(TJ' T 2, P 2,ac» =Hac(TJ)

and
W+(T I• T 2, P 2,ac) = W+(T2, T I, PI,oJ*;
the absolutely continuous parts of T J and T2 are unitarily equivalent. The
corresponding results hold for W _. (In the assertions of this theorem lJ, ac
and Hac(T.J) can be replaced by Pi, c and Hc(1j).)

Corollary. Assume that W +(T2, T" PI,ac) exists. Then R(W+(T2• T I •

PI, ac» = R(P2, aJ if and only if W +(T I• T 2, P 2, ac) also exists.

11.2 The existence and completeness of

wave operators
Useful existence results for W ±(T2 , T I , P) are known only in the case
where P = PI, ac (or P" PI, QC)' Consequently, we shall consider only this
case in what follows. If, for example, TI is a (non-trivial) differential
operator on L2(R m ) with constant coefficients, then PI, ac = I. Therefore in
many cases there is no loss of generality in assuming that P "PI, ac'
344 11 Scattering theory

The wave operator W + (Tz' T I, PI, ac) is said to be complete if

R( W +(Tz, T I, PI, aJ) = Hac(TZ}. By the Corollary to Theorem 11.6 this is
the case if and only if W +(T I, T z, P z ac) also exists; this wave operator is
then also complete, and the absolutely continuous parts T I. <Ie and T2, ae are
unitarily equivalent. A similar result holds for W _(Tz• T I• PI, ae) and
W_(TI' T z• P2• ae ). If R(W_(Tz• T I• P I,ae»=R(W+(T2• T I , P I. ac then ».
the scattering operator S(Tz' T I)= W+(Tz• T I• P I. ac )*W_(T2 , T I• PI,ac) is
obviously a unitary operator on Hac(T I); this holds in particular if
W +(Tz, T I• P I. ac ) and W _(Tz' T I , PI • ac) are complete.
The purpose of this section is to prove a few abstract criteria for the
existence and completeness of W ± (Tz• T I, PI. ac)' These will then be ap-
plied to differential operators in the next section. The r,eader can find
further references, for example, in T. Kato [45].
Assertions similar to those occuring in the following theorem are known
in the literature as Cook's lemma.

Theorem 11.7. Let TI and Tz be selj-ar:fjoint operators on the complex

Hilbert space H. If e - i IT'f E D( T I) n D( T z) for all t E R. the function
R -'J>H, t ~ (Tz - T I) e- i IT'f

is continuous, and

L:"(T T Z- I) e-itT'fil dt < 00,

then f E D(U±(Tz' T I »·
PROOF. Since e- i IT'f E D(T2)n D(TI). the function U(t)I= ei IT, e- i IT'f is
differentiable for all tEn. and its derivative is

:t(U(t)f) = i eiIT2 (T2- T I) e-iIT'f.

This derivative is continuous by assumption. Therefore,

O(t)f- O(s)f = if'e
s
iUT2 (Tz- TI)e-iuT'fdu,

IIU(t)f - O(s)fll ..;; f'II(T2 - T I) e- i uT'f1l duo

$

Since the integral over (- 00, 00) is bounded, the limit:; U±(T2• TI)f-
limH ± ooU(t)f exist. 0

Theorem 11.S. Let TI and Tz be selj-ar:fjoint operators on the complex

Hilbert space H, and let E be the spectral family of T I. If fE Hac(T I),
H.,= L{ E(t)f : t E R}, PI is the orthogonal projection onto HI' H.,c D(TI) n
D(Tz). and (Tz - TI}PJE BI(H), then H.,c D(U±(Tz. T I »·
11.2 The existence and completeness of wave operators 345

PROOF. The proof is in three steps.

First step: Without loss of generality we can assume that Hr= L2(J) with
J a bounded interval and that the restriction T1,f of TI to the reducing
subspace Hf = LlJ) is equal to the operator of multiplication by the
variable.

Proof of the first step. The subspace Hf reduces T I , because it obviously

reduces E(t) for all 1 E R (cf. Theorem 7.28). The operator TI,f is self-
adjoint on Hr and is defined on the whole of Hr. Therefore, Tl,f is bounded
(Theorem 5.6). Hence there exists an interval (II' I~ such that IIE(t)fll =0
for I";; II and IIE(I)fll = IIfl! for t > 12 ,
If we set Pjt)=IIE(/)fll for IER, then there exists (cf. the proof of
Theorem 7.16) a unitary mapping Vf : Hr--~L2(R, Pf} for which VfTI,fVf - 1
is the multiplication operator on L2(R, Pf):
(J-jTI,fJ-j-lg)(X) = xg(x) for g E LiR, Pj)'
Because f E Hac(T1), the function Pf is absolutely continuous, i.e., there
exists a a E LI(R) such that

pit) = fl o(s) ds = [Ia(s) ds for t E R.

- 00 11

Denote 5.t = (t E R : a(/) ~ O}, (Since a is uniquely determined only up to

a Lebesgue null set, Sf is uniquely determined only up to a null set; the
reader can verify that this plays no role in what follows.) The set 5.t is
measurable and can obviously be chosen to be bounded. The formula

w.r: L2(R, Pf) ~LiSf)' g ~ a l / 2g

defines a unitary mapping. With the notation Uf = WjJ-j the operator
UfTI,f~ - I is the multiplication operator on L2( Sf). Let us extend the
operator Uj to the operator
U : H~Hr.l. ffi L2 (Sj), g~ «(I - Pf)g, UfPfg)·
U is obviously unitary; U maps Hr onto L2( 5.t) and transforms T\,j to the
multiplication operator on Ll Sf}. The subspace Hr.l. and the restriction of
TI to Hr.l. remain invariant under U. Consequently, we may assume
without loss of generality that Hr = L2( S) with some bounded measureable
subset S of R and that Tl,f is the multiplication operator on L2 ( S).
Let J be a bounded open interval with S c J, and denote S= J \ Sand
PI = H ffi L2( S) = Hi" ffi L2( J). The ",operators 1) on H (j = I, 2) will be
extended to self-adjoint operators 1) on PI if we define
O(i;) =0(1) $ L2(S),
i;(g+ h) = 1)g + Mh for g E 0(1), hE L2(S),
346 1:1 Scattering theory

M denotes the multiplication on L2( S). The restriction of

to L2(J) is then equal to the multiplication operator on L2(J). We
therefore have L2(J) Hae(T I ), If we choose anj E L2(J) such thatf<x) = I
all x E then] Hae( TI )

Hj: = L{ F(t)]: tE IR} = L2(J),

where F denotes the spectral of Since T\ vanishes on L1( S),
the operator (T2 - TI)Pj belongs to 8 1(A), where Pi denotes the orthogo-
projection onto If we prove that HjC D(n + CT2 , TI », then it follows
H.r C (T2' Since tTl e -- i tr'g = ei tT, tTIg gE we
also have then that H.rcD(n±(T2, T I ».

Second Let T2 - . The assertion of theof,em holds VPj

has the form
n
L
j-I
g)0 for all EH,

where ~ E COOO ( J), '1 E H and II bjll = II hjll = I. Then for all g E COOO( J) (with
=ei IT, tTl) -

1.9)

F is Fourier transformation L2(1R) and g/g taken be

to outside J). A similar estimate holds for the integrals over
( - 00, t) and (- 00, s).

Proof of second step. e' tT'g E H.r D(T2 ) all g

t E H. The function
~(

t- c
is therefore continuous for all g E H.r. Moreover, for g E H.r = L2(J)

II VPr-ilTlgll = j £gj"(x)e- i dx~11

n
~ ~1I(2'11)1/2IF( ~·g)(t)l.
j-I

If g E COOO(J), then the (extended) function ~.g belongs to CoOO(R), and

11.2 The existence and completeness of wave operators 347

thus F( ~* g) E S(R). Hence, in this case

f:IIVPje-iIT,glI dt < 00 forall gECooo(J).

»
Therefore, CoOO(J) c D(Q±(T2' T I by Theorem 11.7. Since the subspaces
D(Q:!:(T2' T I »
are closed by Theorem 11.1, it follows that H.t= LiJ)
C D(Q:!:(T2' T\».
With Q+ =Q+(T2' T I ) we have for gE COoo(J) that
IIQ+ g - Q(t)gIl2 = IIQ+ gll2 - 2 Re<Q+ g, Q(t)g) + IIQ(t)gIl2
= 211Q+ gll2 - 2 Re<Q+ g, Q(t)g)
= 2 Re<Q+ g, Q+ g - Q(t)g)
= 2 Re i foo <Q+ g, eiST'VPj e-isT1g) ds
I

= -21m foo<O+e-iST'g, Vp~-iST'g) ds (cf. Theorem ILl)

I
n 00
= -2 1m ~ ejf <Q+e-isT'g, hj)<~' e-isT'g) ds
j-I I

. ; 2.f Icjl{ {"'1<0+e-iST1g, 0)1 2 dsfool<~, e-iSTlg)12 dS}

J-1 I I
1/2.

Since e - i sT lg E H.t, we have

fool<Q+ e-isT1g, 0W ds = fool<Q+Pj e-isT1g, 0)1 2 ds

I I

= fool<e-iST,g, (Q+Pj )*0)1 2 ds

I

= fOOl! ei SXg(x)* [(Q+Pj )*hj](x) dX( ds

..; 21TIIF- I (g*(Q+Pj )*hj }1I 2 = 21TlIg*(Q+Pj )*011 2
..; 21TlIglI!'II(Q+Pj )*011 2 ..; 21TlIglI!,
and
fool<g1' e- isT'g)1 2 ds = 21T fooIF(g/g)(s)12 ds.
I I

Consequently, we obtain that

110+ g - 0(t)gIl2 < 41T1I gil 00 ~

n
lsi {fOO IF(~*g)(sW ds }1/2 ,
J=1 I

and hence the assertion follows from the inequalities

1I0(t)g - 0(s)gIl2 < {IIO(t)g - 0+ gil + IIQ+ g - 0(s)gll)2
< 2(110+ g-0(t)gIl2+ 110+ g-0(S)gIl2}.
348 II Scattering theory

Third step: The assertion of the theorem holds in general.

Proof of the third step. Let the operator Vo be defined by the equality

Vo = (VPf )* + VPf - PfVPf ·

Vo is obviously self-adjoint and belongs to BI(H). For all g E ~ and hE H

<h, Vog) = <h, (VPf)*g + Vg - PfVg)
= <VPfh - VPfh, g) + <h, Vg) = <h, Vg),
Le., VOPf = VPf .
The operator Vo E BI(H) has the form
00

Vog = ~ cj<gp g)gj for g E H,

)-1

where ~ E H, h)1I = 1 and ~j_llcjl < 00. For allj E N we choose sequences
(gj, "),, E N from H with the properties Pfgj . n E COoo( J). II gj. "i I .;;; 1 and gj,,, ~
gj as n~oo. With

"
V"g = ~ c/gj .", g)gj,n for g E H, n E r~
)=1

it follows that II V,. - Voll ~O, since for every N E Nand n ~ N

N 00

II V" - Voll .;;; 2 ~ Icjlll gj,,. - gjl! +2 ~ IcJ

)=1 )=N+I

If we now set T2,,. = T2 + (V,. - Vo), then by Theorem 9.16 and Theorem
9.18.

If we set ~,.(t) = ei IT2. ne -i IT" then

~,,(t)~~(t) as n~ 00, tE R
The second step can be applied to T2,,, in place of T2, since
(T2, II - T1)Pf = (V + V,. - VO)Pf = V,.Pf ,

and V,.Pj has the form required in the second step:

n

V"Pfg =L cj<Pfgj ,,,. g)~,n for g E H.

)=1

Therefore, (11.9) holds for all g E COOO ( J) with ~n( . ) in place of ~( . ) and
with Pj~,,, in place of ~. We obtain from this for n~oo that

1I~(t)g - n(s)gIl2 .;;; 8'17'11 gil 00 )~IIcA {J,ooIF«Pf g)*g)(u)1 du 2 f/2

+ {~ooIF«Pfg)*g)(uW du f /2]; (1l.l0)

11.2 The existence and completeness of wave operators 349

for the proof we notice that F{(Pj~. n)*g) converges to F«Pjg)*g) in L2(R)
as n~ 00. Since the right side of (l I. 10) tends to zero as t, s~oo, we have
COCO(J) C o(fl+(T2' T I »,and thus Hf =L 2(J)cO(fl+(T2 , T I». We can
prove similarly that H.J C O(fl_ (T2• T I ».
This completes the proof of Theorem 11.8. 0

Theorem 11.9. Let TI and T2 be self-adjoint operators on the complex

Hilbert space H, and let E be the spectral family of T I.
(a) If J is a bounded interval, R(E(J)P •. ac ) C O(T2),and (T2 -
TI)E(J)P I, ac E BI(H), then the wave operators W ±(T2, T •• E(J)P •. ac)
exist.
(b) If the assumption of (a) is satisfied for every bounded interval J, then the
wave operators W ±(T2, T., P I. ac ) exist.

PROOF.
(a) If fER(E(J)PI,ac) then ~cR(E(J)P •. ac)' and thus (T2 -T I)Pf E
B.(H). By Theorem 11.8 Hj C O(fl±(T2' T I »; therefore, f E
O(fl±(T2' T I». As this holds for all fE R(E(J)P •. ac )' we have
»,
R(E(J)P I. ac )cO(fl±(T2, T I i.e., W±(T2' T I, E(J)P I. ac ) exist.
(b) R(E(J)P •. ac) C O(fl±(T2' T I» for every bounded interval J, by part
(a). Since the linear hull of these spaces (for all bounded intervals J) is
dense in R( P I, ac)' the assertion follows. 0

Theorem 11.10. Let TI and T2 be self-adjoint operators on the complex

Hilbert space H. The wave operators W ±(T2• Tp P I. ac ) and
W ±(T\> T 2 , P2. ac) exist and are complete provided that R(EI(J» C O(T~,
R{ElJ» C O{T I) and (T2 - TI)E;(J) E BI(H) for j= 1,2 and for every
bounded interval J. This holds in particular if one of the following assump-
tions is satisfied:
(i) There is an operator V E BI{H) such that T2 = T. + V (Kato-Rosen-
blum).
(ii) O(T I)= O(T2) and (T2 - T,)(z - T.)-2 E BI{H)for some z Ep(T I).
(iii) O(T])= O(T2) and (T2 - TI)(z - T 2)-2 E B](H) for some z Ep(T~.
(iv) O(T I) = D(T2), and there exists an n E N such that (T2 - TI)(z - TI)-n
E BI(H) for some z E p(T I) and (T2 - TI)(z - T 2)-n E B](H) for some
z Ep(T2).
(v) There exists an nEN such that D(T2)-:JO(Tn-:JR(EiJ»for every
bounded interval J and (T2 - T,)(z - TI)-n E B.(H)for some z Ep(T,).
PROOF. If (T2 - TI)E;(J) E BI(H) for j = I, 2 and for every bounded inter-
val J, then the wave operators W ±(T2, T I, P I. ac ) and W ±(TI, T 2, P 2. ac )
exist by Theorem 11.9(b). These wave operators are complete by the
Corollary to Theorem 11.6.
(i) obviously implies (iv) for all n E N. The assumptions (ii) and (iii) are
equivalent according to (9.7) (observe that the assumptions of (9.7) are
350 11 Scattering theory

fulfilled. (iv) follows from (ii) and (iii). In order to prove that (i), (ii), (iii),
or (iv) implies the assertion, it is sufficient to prove that (iv) implies the
assertion: For every bounded interval J
(T2 - T.)E.(J) = (T2 - T.)(z - T.)-n(z - T.tE.(J) ,= B.(H).
The formula (T2 - T.)E2(J) E B.(H) follows similarly.
It remains to prove that (v) implies the assertion: For this, we prove
(T2 - T.)E.(J) E B.(H) as above. Since R(E2(J» c O«z - 1'.)"), the opera-
tor (z - T.tE2(J) is bounded, and thus
(T2 - T.)E2(J) = (T2 - T.)(z - T.)-n(z - T.)"E2(J) E B.(H). 0

In the case 0(T1) = 0(T2) Theorem 11.10 can be essentially sharpened.

Theorem 11.11. Let T, and T2 be self-adjoint operators on the complex

Hilbert space H. Assume that O(T.) = 0(T2) and (T2 - T,)E,(J) E B,(H)for
every bounded interval J. Then the wave operators W ±(T2, T I) exist and are
complete.

I am indebted to Dr. R. Colgen for the proof given here. The case where
0(T,,/2) = 0(TlI2) can be also treated without many changes. First we
prove the following auxiliary theorem.

Auxiliary theorem 11.12. Let T, and T2 be self-adjoint Gperators on the

complex Hilbert space H. Assum~ that O(T,) = 0(T2).
(a) 11(1 - E,« - n, n»)E2(J)II-+O as n-+oo for every bounded interval J.
(b) If the limits s-liml-+±OOE1«-n, n»e iIT , e-iIT2E2(J)P2.ac exist for all
n E N, then the limits
s- lim eiIT'e-iIT2E(J)P
2 2, ac
t ........ ± 00

also exist.
PROOF.
(a) It is obvious that (I - E I « - n, n»)E2(J) = (I - E 1 «-
n, n)))(TI + i)-I
X (T 1 + i)(T2 + i)-I(T2 + i)E2(J) for all n E N. Since (1; + i)E2(J) and
(T I + i)(T2 + i)-I are bounded and since 11(1 - E I «-
n, n»)(TI + i)-III
<; (1/ n), the assertion follows.
(b) This assertion immediately follows from
lIei IT, e- i IT' E 2(J)P 2, ac - E,« - n, n»e i IT, e- i IT2E 2(J)P 2,acl!
= lIei IT'(I - E,« - n, n»)E2(J) e- i IT'P2. aell
;;; 1\(1 - E,« - n, n)))EiJ)II-'> 0 as n -+ 00.

D
11.2 The existence and completeness of wave operators 351

PROOF OF THEOREM 11.11. The existence of W ± (T2• T,) follows from

Theorem 11.9(b), because R(E,(J» c D(T.) = D(T2 ) for every bounded
interval J. It therefore remains to prove the existence of
W ±(T•• T 2 • E 2 (J)P2 aJ =s- lim ei/T , e-i/T'E2(J)P2 ac
I I~±OO •

for every bounded interval J.

Let J be a bounded interval. As (T2 - T.)E.(J) E B.(H). the operator
Vo; = (T2 - T.)E.(J) + «T2 - T.)E.(J»* - E,(J)(T2 - T.)EJ(J)
is a self-adjoint element of B.(H) and
VoE.(J) = (T2 - T.)E,(J).
E,(J) Vo = (VoE,(J»* :> E,(J)T2 - T,E,(J).
Define T3 = T2 - Vo. Since R(EiJ» c O(T) = O(T2). we then have

E.(J) VOE3(J) = (E.(J)T2- T.EJ(J»EiJ).

and thus
T.E.(J)E3(J) - E J(J)T3E3(J) = (T,E,(J) - E,(J)T2 + E,(J) VO)E3(J) = O.
If we define.A : ~~B(H) by the equality
A(t) = E • (J)e i/T, e-i/T'E) (J)P 3. ac.
then

:t (u. A(t)v) = i(u. ei ,T'(T,E.(J)EiJ) - E,(J)T3E3(J» e- i/T,p3• acv) = 0

for all u. v E H. A is therefore constant, i.e., A(t) = A (0) = E.(J)E3(J)P3, ac
for all t E R.
Since T3 = T2 - Vo with Vo E B.(H). the operators
W±(T3• T2 • E 2 (J)P2. a J = s -lim ei/T, e-i/T2EiJ)P2,ac
exist by Theorem 11.9(a). and

W±(T3• T 2 • E 2(J)P2,ac) = E 3(J)P3 ,ac W ±(T3 • T 2• E 2(J)P2,aJ

by Theorem 11.2. This implies
s- lim00 (1- E 3 (J)P3, ac ) ei/T, e- i1T2E 2 (J)P2. ac
t ----. ±
= 0,
and thus the existence of
s - lim E,(J) eitT, e-iIT1EiJ)P2.ac
t-+± 00

lim E (J) eilT, e- itT, eilT, e- i1T2E (J)P

= s - I-+±oo J 2 2.ac
= S -
t_ ±
lim00 E J(J) eilT, e-iIT'E3(J)P3, ac eilTJ e- i1T2E 2(J)P2. ac

= s- lim A(O) e itTJ e-iIT'E (J)P .

t~±oo 2 2.QC
352 II Scattering theory

As this holds for all bounded intervals J (in particular, for J = ( - n, n), as
well), the existence of
s- lim E«-n,n»eitTte-itT2E(J)P
t .... ±oo I 2 2,ac

also follows for all n E N. Consequently, the assertion of the theorem is

obtained by Auxiliary theorem 1l.12(b). 0

In what follows we would like to prove a version of the so-called

in variance principle for wave operators (d. also T. Kato [22), X.4).

Theorem 11.13 (lnvariance principle). Let the assumptions of Theorem ll.S

be satisfied (let H.r and PI be defined as there). Assume that the function
-& : IR~IR is twice continuously differentiable and -&'(x) > 0 (respectively
-&' (x) < 0) for all x E R Then HI C O(Q ± (-&( T 2), -&( T I))), and Q ± (-&( T 2),
-&(TI»g = Q±(T2' TI)g (respectively Q±(-&(T2)' -&(TI»g = Q:;:(T2' T1)g) for
all g E HI'

For the proof we need an auxiliary theorem.

Auxiliary theorem 11.14. Let -& : R~IR be twice continuously differentiable.

Assume that -&'(x) > 0 (respectively -&'(x) < 0) for all x E IR. Then for all
g E L2(1R)
IIgll 2 ;;> {,olF(e-iS"'(.)g)(u Wdu ~O

as s~oo (respectively s~ - 00). (A similar result holds for f~oo if we

exchange "s~oo" with "s~ - 00".)
PROOF. The inequality is clear since F: L2(1R)~L2(1R) is unitary and
lIe-i~(.)gll = II gil for all s E R It is therefore enough to prove the conver-
gence relation for all g from a total subset of L2(1R), for example, for
characteristic functions of bounded intervals. We consider the case -&' > O.
Then for g = X(a. b), S > 0, and u > 0
F( e- i s",(. )g)( u) = (217) -1/2 fb e - i ux e- i s-'(x) dx
a

= i(27T)-1/2f\u+s-&'(x»-I~e-i\lX-is-'(X) dx
a aX
e - i ux-i s-'(x) ]b
- '(2 )-1/2 { [
- I 17 U + si}'(X)
a

+ fb s-&"(X) e- i ux-is",(x) dxl.

a (U+s~'(x)f f
There exist C. >0 and C 2 ;;' 0 such that -&'(x);;. C I and I~ "(x)l< C 2 for all
11.2 The existence and completeness of wave operators 353

x E [a, b). Consequently, for all u > 0 and s >0

IF(e-isD(.)g)(u)1 ..; CiU+S)-I,
and thus
l oOOIF(e-isD(·)g)(uWdu..; C32S- 1 forall s > o.
.
The case {f' < 0 can be treated in a similar way if we consider u > 0 and
s<O. 0
PROOF OF THEOREM 11.13. As in the proof of Theorem 11.8, we can assume
without loss of generality that H.r =- L2(J) with some bounded interval J
and that the restriction of TI to L2( J) is equal to the operator of multi-
plication by the variable.
Let {f' > O. We consider n+: First we remark that (11.10) obviously
holds for gE C(J), as well. If we replace g by e-i",<T1)g=e-isD(')g(.),
then we obtain for s=-O and t ~ 00 that
IIn+(T2 , T I ) e-isD(T1)g - e- isD(T1)gIl2

..; 8'1TlIg1i00 ,f ICjl{lOOIF(pjg )* e- isD(.)g)(u)j2 dU}

,-I 0
j 1/2.

By Auxiliary theorem 11.14, the expressions in the braces are bounded by

II gll~1I ~1I2 = II gll~ and tend to 0 as s~oo. Consequently,
n+(T2 , T.) e-isD(T')g - e-isD(T')g ~ 0 as s ~ 00

for all g E COoo( J). Then this holds for all g E L2( J), as well. Hence, for all
g E L2(J)
e- isD (T2 )n+(T2 , TI)g - e-isD(T')g ~ 0 as s ~ 00

(cf. Theorem 11.1). Multiplication by ei "'{T,) gives assertion in this case.

The other cases can be handled similarly. 0

Theorem H.tS. Let T. and T2 be selj-at/ioint operators on the complex

Hi/bert space H, and let E denote the spectral family of T I . Assume that
R(E(J)P •. ac) C D(T2) and (T2 - T1)E(J)P I. ac E BI(H) for every bounded
interval J. Assume, furthermore, that the function {f : R~R is twice continu-
ously differentiable and {f'(x) > 0 (or {f'(x) < 0) for all x E R Then
Hac({f(TI»=- Hac(T I), the wave operators W ±({f(T2), {f(T I), PI,ac) exist, and
W:t({f(T2), {f(T I), PI. ac) = W:!:(T2 , T I , PI. ac} (or W ±({f(T2 ), {f(T I), PI. oC>
= W +(TZ' T I , P I. ac ), respectively).
PROOF. A subset N of IR is a null set if and only if {feN) is a null set. If F is
the spectral family of {f( T I), then F( S) == E( {f -ie
S» for every Borel set S
(cf. Section 7.3, Proposition 6). It follows that Hac({f(TI = Hac(TI)' The »
rest of the assertion follows from Theorem 11.13 (cf. the proof of Theorem
11.9). 0
354 11 Scattering theory

Theorem 11.16. Let TI and T2 be self-adjoint operators on the complex

Hitbertspace H.Assumethat TI~Y' T2~y,and(TI-'A)-P·-(T2-'A)-PE
8 1(H) for some 'A E iii such that 'A < Y and some pEN. Then the wave
operators W±(T2' T l , Pl. at) exist and are complete.
PROOF. Let 1'} : 1Ii~1Ii be a twice continuously differentiable function such
that 1'}(t) = 'A + t - lip for t ~ (y - 'A) -p and 1'}'(t) < 0 for all t E iii (hence,
1'}«1j - 'A)-P) = 1j for j = I, 2). The wave operators

W±(T2, T l , Pl,aJ = W+(T2 -'A)-P, (Tl-'A)-P, Pl,ac)

and
W ±(Tl , T 2, P 2,ac) = W:;:(TI-A)-P, (T2 -'A)-P,P2,aJ
therefore exist according to Theorems 11.10 and 11.15. The assertion thus
follows. 0

11.3 Applications to differential operators on L2(~m)

If Tl is a self-adjoint differential operator on L2(lIim) with constant coef-

ficients, then in many cases the existence of W ±(T2 , T 1) ,;;an be proved
without using Theorem lI.S or any of its consequences. We only have to
apply Theorem 11.7 and appropriate estimates of exp( - i tTI)f that can be
proved with the aid of the Fourier transformation. Here we prove a special
case of a result of K. Veselic and J. Weidmann [53], [54] (compare also with
L. Hormander [43]).

Auxiliary theorem 11.17. Let A C llim be a closed set. Assume that the
function h : llim~ iii is infinitely many times continuously differentiable in
Rm"A and grad h(x)~O for x E llim"A. Then for eveIJ' function gE
Cooo(lIim" A) and for every p ~ 0 there is a constant C such that

jor a/l x E R m and t E R" {OJ (here g is considered as a junction defined on

Rm with g(x) .... 0 jor x E A).

PROOF, For every y E Rm"A there is aj E {l, 2, ... , m} for which <ljh(y)
- (3 /3Yj)h(y) ~ O. Since <ljh is continuous, there is a neighborhood Uy of y
in Rm"A such that ajh(z)~O for all z E Uy • The compact set supp g can
be covered by finitely many such neighborhoods UN ... , Uyn • Then there
are functions 1'}j E CoOO(Rm) such that supp 1'}j c UYj and ~j-I1'}iY) = I for all
y E supp g (partition of unity; cf. W. Rudin [33], Theorem 6.20). It is obvi-
ously sufficient to prove the assertion for the functions 1'}jg in place of g,
11.3 Applications to differential operators on L 2(Rm) 355

i.e., we can assume that ajE {t, 2, ... , m} exists for which ajh(y)~O for
all y E supp g. (In many concrete cases there is a j E {I, 2, ... , m} such
that ajh(y)~O for all y E IRm"A; then the foregoing step of the proof is
superfluous. )
Hence, let ajh(y)~O for ally Esuppg. Then by k-fold partial integra-
tion with respect to Yj we obtain for all kEN and t =1= 0 that

j eiXY-ilh(Y)g(y) dy = je-i1h(Y)a.h(y) eiXYg(y) dy

J ajh(y)

= (i t)-lje-i1h(y)a.[ eiXYg(y) 1dy

II
J Ojh(y)

= . . . = ( 1' t )-kj e -ilh(Y)aj [ Ojh(y)

I 0j [ ,., Ojh(y) 1. , ,
aj [eixyg(y) dy

= (i t)-k j eiXy-ilh(Y)X(Xj>y) dy,

where
k
X(~'Y) = L fJ/y)x!
p-o
and the supports of the functions fJp E CoOO(lRm\A) are contained in supp g.
Therefore,

We consequently obtain for every kEN that

1j ei xy-i Ih(Y)g(y) dYI ..; Ckltrk(1 + Ixl)k.

This estimate evidently holds for k = 0, as well.
Now let kENo be such that k..; p <k + l. Then

Itl-P(I + IxlY > Itl-k(l + Ixl)k for Itl-I(I + Ixl) > 1,

Itl-P(I + Ixl)p > Itl-(k+l)(I + Ixl)k+1 for Itr 1(1 + Ixl) < 1.
If we apply the above result with k and k - I, then we obtain the assertion.
o
In what follows, for a real-valued measurable function h defined on IRm
let Mh denote the maximal operator of multiplication by h on L2(R m ), and
let Th = F -IMhF. If h is a polynomial, then Th is a differential operator
with constant coefficients.
356 II Scattering theory

Theorem 11.18. Let us assume that with a closed set A c Rm of measure zero
we have
hE C""(Rm\A) and grad h(x) =1= 0 for x f£. A.
Let V be a symmetric operator on L2(R m) such that S(IR/ m) c D( V), and
assume that there exist apE No, a e> 1, and a C :;;. 0 with the property that
for all r:;;' 0

IWfll < C(l + r)-sIIfIV for all f E S(Rm) such that f(x} = 0 for Ixl < r.
Then for every self-adjoint extension T of Th +V (provided that any exist)
the wave operators W ±(T, T h) exist.

In the following we shall not prove this theorem but a somewhat more
general one that also considers operators on L2(Rm)M (for example, Dirac
operators). For this let H be a measurable function defined on R m whose
values are M X M Hermitian matrices (such a function is said to be
measurable if the entries of the matrix are measurable functions). Let MN
again be the maximal "operator of multiplication" by H on L2(Rm)M and
let TN = F-JMNF. The operator TN is evidently self-adjoint. We denote by
hJ(x), ... , hM(x) the M (not necessarily different) eigenvalues of H(x) and
by eJ(x), ... ,eM(x) the corresponding normalized eigenelements. (There
is a great deal of freedom [especially if multiple eigenvalues occur) in the
choice of these functions; in what follows it will be possible to choose them
in such a way that the functions hj and ej are infinitely many times
differentiable.) The operator TN is a differential operator if all entries of
the matrix function H are polynomials; however, the functions hj and l'l
are in general then not polynomials (cf., for example, the Dirac operator,
Section 10.6).
A function H defined on Rm whose values are Hermitian M x M
matrices is said to be permissible if the functions ~ and Ej can be chosen
such that there exists a closed set A C Rm of measure zero for which

hj E C""(Rm\A), ej E C""(Rm\A)M,
grad ~(x) =1= 0 for x fI. A (j= 1, ... , M).

Theorem 11.19. Let H be a permissible function. Let V be a symmetric

operator on L2(Rm)M with the properties: S(Rm)M c D( V), and there exists a
p E No, a e> 1, and a C :;;. 0 such that for all r:;;' 0

II Vfll <: C(l + r)-8 I1 fll/ for all f E S(Rm)M such that
f(x) = 0 for Ixl < r.
111 • lip is the norm of W2,p(R m ).
2Here II . lip denotes the norm in W2•p(R",)M.
11.3 Applications to differential operators on L 2(R m ) 357

Then the wave operators W:!:(T, TH) exist for every self-adjoint extension T
of TH + V.

PROOF. We prove that if Ff E Cooo(lRm \ A)M, then f satisfies the assumptions

of Theorem 11.7. Since Cooo(lRm\A)M is dense in LllRm)M=LllRm\A)M,
this will prove the assertion.
Let FfE CoOO(lRm\A)M. Then, in particular,jE S(lRm)M. The assumption
on V (for r = 0) implies that
IIV(e-i.TH-e-itTH)fll';; CII(e-i.TH_e-itTH)fllp
= CII(e- isH(, l _ e- i tH(, ')Ffll(P'

= C {jleiSH(Xl _ e- i tH( Xl I21(Ff)(xW{I + Ixl 2Y dX} 1/2

~ 0 as s ~ t.

The function t~ Ve- i tTHf is therefore continuous on R

Now let {f E COO(IR) be such that 0.;; {f(s) .;; I, D(s) = I for s .;; 0 and
{f(s) = 0 for s> 1. With some J.L E(l/e, 1) let {ft E Co""(lRm) be defined by
the equality {flx) = {f(lxl-ltll') for all x E IR m and t E R
The assumption on V (for r = 0) implies (cf. Theorem 1O.8(c» that

= C 1 max
lal';;p
{fIxl,;;W+l
(2'1T)-m/2IjeiXYe-i IH(Yly"'(Ff)(y) dyl2 dX} 1/ 2,

The operator induced on eM by the matrix e- i tH(Yl can be written in the

form
M
e-itH(Yl~ = L e-it~(Yl(eiY), ~)eiY) for ~ E eM,
j=1

where (, , ,) is the scalar product in eM, Consequently,

Jei "Ye- i tH(Yly"'(Ff)(y) dy Jei"Ye- i t~(Yl(eiy),Y"'(Ff)(y»eiY) dy,

M'
= ~
j-I

Since FfE Co""(lRm\A)M, we also havey"'FfE Cooo(lRm\A)M, Therefore, the

functions y~(eiY), y"'(Ff)(y»eiY) also belong to COOO(lRm\ A)M, Hence
we can apply Auxiliary theorem 11.17 to the last integral. We thus obtain
for all p > 0 and It I ;;;. 1 that

IIV{fte-itT'i1i .;; c2 {I Ixl<ltIP+ I

(l +lx/)2Pltr2p dX}1/2

.;; C 3 1tI P(I'-I)+(mI'12l.

·358 1 I Scattering theory

We obtain similarly from the assumption on V (for r = Itl":> that

II V(I-1'Jo,)e- i tTHf11 .r;;; C4(1 + ItI")-e max lIe- i tTHD'111
lal<p
.r;;; C4maxIlD'1l1ltl-"e = Csl'I-"e.
a<p
If we choose p so large that p( p. - I) + (mp./2).r;;; - p.9 < - I, then
IIVe-itTHfll .r;;; IIV1'Jote-i1THfll + IIV(I-1'Jo,)e-itTHfll .r;;; C 6ltr"e
for It I > 1. Since t~ Ve- i ITHf is continuous, the integrability of t~
II V e-ilTHfll follows from this. 0

Theorem 11.20.
(a) Let T J be the self-adjOint operator defined by the formulae D(T J)
= W2, 2(lJim), T Jf = - Af for f E D( T J). Let V be defined by the equality
VJ(x) = ~ c.. (x) D" f(x) for f E S(lJim),
lal<k
°
and assume that Ic.. (x)l.r;;; C(l + Ixl)-'-' for some C :> and some £ > 0.
Then the wave operators W ± (T2, T J) exist for every se(f-adjoint exten-
sion T2 of T J+ Y.
(b) A corresponding result holds if TI is the free Dirac operator on L2(1It 3 t
and the c..(x) are 4 X 4 matrices.
PROOF.
(a) We have TJ=Th with h(x)=lxI 2• Since gradh(x)*O for x*O, the
assumption of Theorem 11.18 is obviously satisfied with p = k.
(b) The functions hj and ej are known from Section 10.6 and satisfy the
assumptions of Theorem 11.l9 with A = {OJ. Theorem 11.19 is there-
fore applicable if we choose p = k. 0

The assumptions of Theorem 11.20 on the coefficients (~ can be essen-

tially weakened (cf. Exercise 11.1).
In Theorems 11.18 to 11.20 we showed the existence of W ± (T2 , T J) in
many cases where T J is a· differential operator with constant coefficients.
Since in these cases Hac(T J) = H (cf. Exercise 10.7; this is clear for - A and
the unperturbed Dirac operator), this is equivalent to the existence of
W ±(T2, T J, PJ,ac)' We cannot expect that the completeness (i.e., the equal-
» »
ity R( W ± (T2 , T J = Hac( T2 can be proved under such general assump-
tions. A few simple assertions can be proved with the aid of the results of
Section 11.2. In order to be able to apply them we need the following
auxiliary theorems.

Auxiliary theorem 11.21. If r, s :> 0, r - s >m/2 and qJ E ~r(lItm), then the

operator
11.3 Applications to differential operators on L2(Rm ) 359

belongs to B 2( W 2 , r(Rm), W 2, s(R m», If r - s > m, then 4> E

B I ( Ww, lRm), W2, iRm».
PROOF. We recall the proof of Theorem 10.19. 4> is a Hilbert-Schmidt
operator if and only if the integral ope~ator K on L2(R m ) defined there is a
Hilbert-Schmidt operator. The kernel of K is given by the fonnula

k(x,y) = (I + IxI 2y/21/1(X -y){l + lyI2)-r/2 for X,y E Rm

with 1/1 = Fcp E S(Rm). Because of the inequality 11/I(x)l,.;; C(I + IxI 2)-1 we
have

1
CI{I + IxI2)-I+S/2{1 + lyI2)-r/2 for Iyl ~ -&Ix/,
,.;; c1{I + Ix - YI2)-I{1 + lyI2)(s/2)-(r/2) for Iyl > ilx/.
If we choose I > (m/4) + (s/2), then kEL 2(R m xRm); the operator K is
therefore a Hilbert-Schmidt operator.
Now assume that r - s >m, Sl = (r + s)/2 and CPI E Cooo(Rm) with cp\(x) =
1 for x E supp cpo Then 4> can be considered the product of the mappings
4>' : W2,r(Rm) ~W2,s,(Rm), ff-'; cpd,
4>" : W2 ,s,(Rm) ~W2,s(Rm), ff-'; cpf.
As both of these operators are Hilbert-Schmidt, <I> belongs to
B,(W2,r(Rm ), W2,s(Rm » (cf. Theorem 7.9). 0

Auxiliary theorem lI.22. Let T be a self-adjoint operator on L2(Rm). Assume

that D(T") C W2, tCRm) for some n E ~ and some t >m. Let V be a symmet-
ric operator on L2(R m) such that D( V)::J W2 , s(Rm) for some s E [0,
t - m). Assume that there exist a C ;;;. 0 and a e >m such that for all r;;;' 0

for all fE W2,s(R m) such that f(x) =0 for Ixl <r. Then

V(i- T)-" E B\(LiRm».

(An analogous result holds in LiRm)M.)
PROOF. Let 6 = 6 1 E Cooo(Rm) be defined as in Section 2.2, Example 8 (i.e.,
0", 6(x), 6(x) = 0 for Ixl;;;' 1 and f 6(x) dx = I). For all m-tuples Y =
(YI"'" ym )E71.m let
Qy = {x E Rm : Yj <:. Xj < Yj + 1, j = I, 2, ... , m },

CPy(x) = f 6(x - Y)Xa,(Y) dy.

360 11 Scattering theory

Then we obviously have

~ qiy(X) = 1 for all x E IIl m ,

yEl'"

where the sum is finite for every x E Rm (with at most 3m summands). By

Auxiliary theorem 11.21 the operator
<l>y : W2. I(Rm) - W2 jlllm), f 1--4 qiJ
belongs to BI ( W2• ,(R m), W2. s(lIlm» for every 'Y E zm, and the trace norm
lI<I>ylls,(w, t' W 2,,) does not depend on 'Y (we have <fly = 7'y<flo7' _y' where 7'y
denotes the operator of translation by y : (7'J)(x) = f(x - y».
The operator (i - T)-" is a continuous mapping from L2(lIl m) into
W2,I(lRm ): As in the proof of Theorem 10.18 we can prove that II . II Tn>
Cdl . III for some C I > 0, and thus
lI(i- T)-"fIl7 ~ C I- 2 {IIT"(i- T)-'Y1I 2 + 1I(i- T)-'Y1I 2 } ~ C 2 11f112.

Now let Oy={XElIl m :,(;-I<:'xj <'Yj+2 for j=I,2, ... ,m}, CPy=
f8(x - Y)XQy(Y) dy, and
4)y : W2.s(ll~m) _W2,s(Rm), fl--4 ipJ.
Then 4)y<fly = <fly, and V4)y, as an operator from W2,iRm) into L2(Rm), is,
by assumption, bounded by
-
IIV<fI y llB(w2",L 2) <:. C3(1+lyl)- .
e

Observe now that (BI' II . III) is a Banach space by Exercise 7.10 and
that by Theorem 7.8(c) IIABIII <:.IIAlldIBIl and IICAIII·~ IICIlIIAIiI for
bounded Band C. It therefore follows that

11V(i - T)-nll l = II L V4)y<fl)i - T)-nIl 1 <:. L IIV~y<fl)i - T)-nIl 1

yEZ'" yEZ'"

o
From Auxiliary theorem 11.22 we can immediately derive criteria for the
existence and completeness of wave operators with the aid of Theorem
11.10. (These existence statements are weaker than those contained in
Theorem 11.20). Here we only give a typical result. Actually, for m > 2
much better results can be proved using entirely different methods. We
shall not consider them here (compare with, for example, S. T.
Kuroda [47, 48] and M. Schechter [49, SO] for further references).
J 1.3 Applications to differential operators on L2(R m ) 361

TIleorem 11.23. Let T\ be equal to - A, with O( T\) = W2.2(~m). Let V be

symmetric such that O( V)::J O(T\), and let T2 = T\ + V be self-adjoint.
e
Assume that there exist an s > 0, a > m and a C > 0 such that for all r ;> 0

UVfll " C(1 + r)-eUflls

for a/l fE W2,s<~m)n O(V) such that f(x)=O for Ixl <r. Then the wave
operators W ±(T2, T\) exist and are complete.
PROOF. Since O(T\)'! = W 2 , 2n(~m), by Auxiliary theorem 11.22 V(i - T\)-n
E B\(L2(~m» for every n E N such that 2n - s >m, and thus
VE\(J) E B\(L2(~m» for every bounded interval J.
The assertion therefore follows from Theorem 11.11. o
REMARK. The assumptions of Theorem 11.23 hold in particular if V is a
differential operator of order ,,2 whose coefficients decrease as Ixl- e for
some e > m. An analogous result can be proved for Dirac operators.

EXERCISES

11.1. The assertion of Theorem 11.20 holds also if the functions ca are locally
square integrable and Nc (x).;;; C(l + Ixl)-\-< for some C ;;. 0 and some (> O.
Hint: Choose p >k + (';;/2) in the proof.
11.2. Assume m .;;; 3, T\ == - A, D(TI ) = W2 , iRm) and q E L2(Rm) n LI(R m). If V is
the operator of multiplication by q, then V is TI-bounded with TI-bound
zero; consequently, T2 = TI + V is self-adjoint and D(T2) = D(TI). The wave
operators W;!:(T2' T I ) exist and are complete.
Hint: (T I + S)-I - (T2 + S)-I = (TI + s)-IV(T2 + S)-I E B I(L 2(R m » for
sufficiently large s, since
IVII/2(TI +8)-1 and IVII/2(T2+s)-1 = IVII/2(TI +8)-I(TI +8)(T2 +s)-1
are Hilbert-Schmidt operators (Theorem 11.16).
Appendix A

Lebesgue integration

In this appendix we shall compile and prove a few results of Lebesgue

integration theory that are used in several places in this book. We essen-
tially follow the presentation of F. Riesz and B. Sz.-Nagy [31]; but notice
that only the measure induced by the volume function is :;tudied in detail
there. For complete presentations of the theory of measure and integration
we refer the reader to, for example, E. Hewitt and K. Stromberg [18] or W.
Rudin [32].

A.I Definition of the integral

Let t = t(Rm)1 be the set of bounded intervals J = J1 X ')2 X ... X Jm in
Rm, where the ~ ar~ arl2itrary open, half-open, closed, OM-point or empty
intervals in R. Let t = ~(Rm) be the set of finite unions of intervals from
t. It is obvious that every M E ican be written as a union of finitely many
mutually disjoint intervals from t.
A mapping p : ~-+R is called a function of an inte1'1'Jal or a interval
function on Rm if we have:
(AI) Monotonicity: J(, J 2 E t and J 1 C J 2 imply p(J1) ..; p(J-z).
(A2) Additivity: J" J 2 E~, J, n J 2 = 0 and J, u J-z E t imply p(J1 U J 2)
= p( J,) + p( J2).
The mapping p can be extended to i by the formulae
n n
p(M) == L p(~) for M= U ~ with ~ n Jk = (2:: for j oF k.
j-I j-I

tIn the following we omit Rm if no confusion is possible.

362
A.) Definition of the integral 363

This definition is obviously independent of the choice of the representation

M = U 1-10. This extended mapping is also monotone and additive.
The monotonicity and additivity immediately imply that
p(M) ;> 0 for all MEi; p(0) = O. (A3)
An interval function p is said to be regular if we have:
(A4) For every J E ~ and every E > 0 there exists an open interval j
such that J c j and p(j) ( p(J) + E.
It is easy to see that p is regular if and only if we have:
(AS) For every J E ~ and every E > 0 there exists a closed interval J
such that J c J and p(J);> p(J) - E.
A regular interval function on IRm will be called a measure in the sequel.

EXAMPLE I. The volume function

m
A(J) = II (b)-a j ) for J= {XElRm: aj~xj~0}
j-I

is a measure. A is called the Lebesgue measure.

EXAMPLE 2. Iff: IR~IR is a right continuous non-decreasing function, then

the formula

f(b) - f(a) for J=(a,b],

pi J) = ! f(b)-f(a-)
feb -) - f(a)
for
for
J= [a, b],
J= (a, b),
feb -) - f(a -) for J= [a, b)
defines a measure.

EXAMPLE 3. If PI is a measure on IRP and P2 is a measure on IRq, then the

equality
p(J. x J2) = PI(J 1)P2(J2) for J. E HiRP), J2 E HIR'l)
defines a measure on IRp+q.

In what follows let p always be a measure on IRm. A set N C IRm is called

a p-null set if for every E > 0 there exists a sequence (J,,) from ~ such that

Nc U J" and ~ p(J,,) < E.

"EN "EN

Since p is regular, these intervals can always be chosen to be open.

EXAMPLE 4. All finite and countable subsets of IRm are A-null sets.
364 Appendix A: Lebesgue integration

EXAMPLE 5. Every subset of a p-null set is a p-null set.

EXAMPLE 6. Countable unions of p-null sets are p-null sets: If the Nk(k E
> 0 there are sequences (Jk.n)nEN from
N) are p-null sets, then for every t:
~ for which Nk C U nENJk. n and ~nENp(Jk, n) <t:2- n . This implies that
U kENNk C Un, kENJk,n and ~n, kENp(Jk,n) <t:o

In what follows we say that a certain assertion holds p-almost everywhere

(or for p-almost all x) if there exists a. p-null set N such that the given
assertion holds for all x E Rm \ N. In particular, we write f = g or J" -4 g if
p p
there exists a p-null set N such that f(x) = g(x) or fn(x)-4g(x) for all
x E Rm \ N, respectively.
A function f: Rm -4 C is called a step function if
n
f = L cjXJ. with J.J E ~ for j = 1, ... , n
j-l ~

(here XM denotes the characteristic function of M). Of course, the intervals

J.J are not uniquely determined by f. However, we can choose the intervals
J.J to be mutually disjoint. We denote the set of step functions defined on
Rm by T = T(Rm). The set T(Rm) is a complex vector spac,e.
For an f E T we define the p-integral by the equality

f f dp = f f(x) dp(x) = L cjp(J.J)

n n
for f = ~: cjXJ.·
j-l j=l ~

This definition is independent of the choice of the representation of f as a

linear combination of characteristic functions.
The following properties of the p-integral of step functions are obvious:
(A6) If f, gET are real-valued and f:r;;; g, then ff dp:r;;; Jg dp.
(A7) IJjdpl:r;;; flfl dp for every fE T.
(A~) The mapping T-4C, f~ If dp is linear.
Concerning the extension of this notion of an integral to a wider class of
functions we need a few preliminary remarks. First we consider only
real-valued functions.

i
Auxiliary theorem At. If ME and (In) is a sequence from ~ such that
Me U nEN/m then p(M):r;;; ~nENp(/n)'
PROOF. Let M = U ;-1 J.J with pairwise disjoint interval:; J.J. Since p is
regular, for every t: > 0 there exist closed intervals 0
for which C J.J and 0
p(0);;;' p(J.J) - t:/(2k) and open intervals Tn for which In C Tn and ~nENp(ln)
:r;;;~nENp(ln)+t:/2. The set 1Cf= uj=10 is a compact subset of M and
p(ICf);;;'p(M)-t:/2. Consequently, there is an NEN such that ICfc
A.I Definition of the integral 365

U ~_I in> and thus

As this holds for all f > 0, the assertion follows. o

Auxiliary theorem A2. If (J,,) is a non-increasing sequence from T and J" ;;. 0,
~ 0, then ffn dp~O.
fn
p
PROOF. Let No be the p-nuli set on which the sequence Un) does not
converge to zero, let K E ~ be chosen such that we have K ;;.f, (then we
also have K ;;. fn for all n EN), and let 10 be an interval such that f,(x) = 0
for x E ~m \10 (then J,,(x) = 0 for x E ~m\ 10 and all n EN).
For a given f > 0 let f' = f( K + p(lo» -I. Let ~n be the set of mutually
disjoint intervals where J" assumes constant values not smaller than f'.
Then with the notation Mn = U JE§J we have
MI ::JM2 ~ .. , and n Mn eNo,
nEN

because of the mono tonicity of the sequence Un)' This implies that

n
00 00

Mk = U (Mn \Mn+ l ) U Mn e U (Mn \Mn+d UNo. (A9)

nEN

For every n E N let '5'n be a finite set of mutually disjoint intervals for
which Mn \Mn+1 = U JE'5: J. Since MI ~ U~::(Mn \Mn+1) for kEN and
(Mn \ Mn+ J) n (Mm \ Mm+ J)"= 0 for n =i=m, we have

and therefore
00

L L p(J) , p(MJ) < 00.

n-I JE~n

Consequently, there exists a ko E N such that

00 f'
n~k J~~np{J) < '2 for k;;. k o·

Since the p-null set No can be covered by countably many intervals of total
measure <f' /2, by (A9) the set Mk can be covered by countably many
intervals of total measure <f'. By Auxiliary theorem Al we therefore have
p( Mk ) < f' for k ;;. ko' and thus

fA dp '" Kp{Md + f'p{lO) '" f'{K+p{lo» = f for k;;. ko' 0

.366 Appendix A: Lebesgue integration

Auxiliary tbeorem A3. If Un) is a non-decreasing sequence from T and the

sequence of integrals (fin dp) is bounded, then there exists a function
f: Rm - R for which fn -p f·

PROOF. Without loss of generality we may assume that the: functionsfn are
non-negative (otherwise we would consider the sequence Un - fl»' Let No
be the set of those x E Rm for which the sequence Un(x» diverges. We have
to show that No is a p-null set.
Let £ > 0 be given. Let C ;;. 0 be such that ffn dp .;;; C for all n E "'. Let
Cj"n be the set of disjoint intervals on which fn assumes constant values not
smaller than C / £, and define

Nn = U J= {X E R
m : fn(x);;, C) for n E "'.
JE~n £

If N= {x E Rm : fn(x);;' C/ €. for some n E "'}, then No C /1/= UnENNn and

Nn C Nn+ I' We can therefore choose a sequence (Jk ) of disjoint intervals so
that
k(n) 00

Nn =U Jk for all n E '" and thus N= U Jk ~No.

k-I k-I

It therefore follows for all n E '" that

~n)

k-I
C
~ p(Jk ) -
£
=
JE~n
C
~ p(J)- .;;;
€.
JIn dp .;;; C,

and hence

i.e., No is a p-null set. o

Auxiliary theorem A4. Let Un) and (gn) be non-decreasing sequences from T
such that fn - f. gn - g and f < g. Assume that the sequences of integrals
pep
(ffn dp) and (f gn dp) are bounded. Then

lim Jfn dp
n-+oo
< n--+-OlO
lim Jgn dp.

PROOF. For every mE'" the sequence Um - gn)nEN is non-increasing and

p - lim
n-+oo
(1.m - gn ) =p 1.m - g .;;;p f- g .;;;p 0 .

Consequently, for every mE'" the sequence «fm - gn)+)nEN is non-in-

creasing and p-converges to 0 (here h + = max{O, h}). By Auxiliary theorem
A.I Definition of the integral 367

A2 we therefore have fUm - gn)+ dp~O as n~oo, and thus

.s;;: nlim IUm - gn)+ dp = 0 for all mE 1\1.

..... oo

The assertion follows from this if we let m tend to 00. o

In what follows let TI = TI(lRm, p) be the set of the functions f: Rm~R
for which there exists a non-decreasing sequence (J,,) from T for which
J" -+ f and the sequence of integrals is bounded. Since the sequence of
p

integrals (I
fn d P) is non-decreasing and bounded, it is convergent and we
can define .

I f dp = nlim
..... oo
If" dp.

By Auxiliary theorem A4 this definition is independent of the choice of the

sequence (J,,) (having the required properties). If f, g E TI and a, b > 0,
then af + bg obviously also belongs to TI and

I (af + bg) dp = a I f dp + b f g dp.

Now let T2 = T2(R m , p) be the real vector space that is spanned by T I, i.e.,
T2 = {f = fl - f2 : fl'/2 E Td· On T2 let us define the p-integral by the
equality

I f dp = I fl dp - f f2 dp for f = fl - f2 with 1j E TI·

This definition is independent of the choice of the functions fl and f2' since
f= fl - f2 = gl - g2 with fl' f2' g" g2 E TI implies fl + g2 = gl + f2' ff, dp +
fg2 dp= fg, dp+ Jf2 dp, and thus

If, dp - I f2 dp = fg I dp - I g2 dp.

The elements of T2(R m, p) are called p-integrable functions (observe that

only real functions have been considered so far).

Theorem AS.
(a) The mapping T2-+R,f~Jfdp is linear.
(b) f, g E T2 and f <, g imply ff dp.s;;: Jg dp.
p
(c) If fE T2, then Ifl'/+'/- E T2.
(d) Iff dpl <, flfl dp for every f E T2.
368 Appendix A: Lebesgue integration

PROOF.
(a) This assertion is evident.
(b) It is obviously sufficient to consider the case f = 0 (g ,:::an be replaced
by g - f). Let g = gl - g2 with gl' g2 E T •. Then g2 <I; gl' and hence
p
Ig2 dp" IgI dp by Auxiliary theorem A4, i.e., Ig dp;;;' O.
(c) Letf=fl- f2 withfl.J2E T •. Then max{f•.J2} E T. and min{fl,f2} E
T. (proof!). The assertion then follows from the formulae
If I = max{J.,f2} - min{JI,f2}'
f+ = max{J•.J2} - f2' f - = max{J.,f2} -- fl'
(d) We have

Iff dpi = Iff + dp - f f - dpl

"ff+dp+ jf_dp= jU++f_)dp==jlfldp. 0

Theorem A6. For every function f E T2 there exists a sequ,~nce (f,,) from T
-)0 f and Iii" - fl dp-)oO. In particular, ff" dp-)o If dp.
such that f"
p

PROOF. Let f= fl - f2 with fl> f2 E T •. Then there exist non-decreasing

sequences (h, "),, E N from T such that Ii." -: hand n." dp-)o Iii dp for
j= 1, 2. Forf,,=fl • "-f2,,
•
we havef,,-)ofand
p
Ilf" dp- Ifdpi "/lf,,- fl dp
= flf•. " - A" - (f. - f2)1 dp " f(fl - fl.,,) dp + f(f2 - f2, ,,) dp~O as n-)o 00.
o

A.2 Limit theorems

The following theorem asserts that the extension proceS!i of the previous
section (which took us from T over TI to T2) does not lead from T2 to any
wider class of functions. In the rest of this section we shall prove theorems
which show, under which assumptions the passage to the limit and the
integration are exchangeable. These theorems show the eS8ential advantage
of the Lebesgue integral when compared with the RiemalID integral.

Theorem A7 (B. Levi). Let (f,,) be a monotone sequence (non-decreasing or

non-increasing) from T2 for which the sequence of int/?grals (ffn dp) is
bounded. Then there is an f E Tz for which
A.2 Limit theorems 369

PROOF. We may assume without loss of generality that (!,,) is non-decreas-

ing (otherwise we consider (- fn»'
First step: There are non-decreasing sequences (gn) and (hn) from T.
with bounded sequences of integrals for which !" = gn - hn (i.e., it is
sufficient to prove the theorem for a sequence Un) from T.).
Proof of the first step. Define k. = f. and's = fj - fj_. for j ;;;. 2. Then
fn = ~'i_.kj and k n ;;;. 0 for all n EN. We have kj = kj, • - kj ,2 with's, .' kj ,2
E TI • For ~very j EN tpere exists an ~ E ! such that ~ <'s, 2 l!nd I(kj , 2 -
~) dp ~ 2-:.If ~e set kj, 2 = kj, 2 - ~ and's, I = kj, • -~, then kj, •• kj, 2 E TI
and 's =, 's, • - kj ,2 for all j E N. The series ~;'.J IS, 2 dp is convergent
(since I's, 2 dp < r j ). If C is a bound of the sequence (ffn dp), then
n
.~
}=I
J
kj dp = J fn dp ~ C for all n E N.

Then the series ~j:. II kj, • dp is also convergent. The functions

n n
gn = ~ ~.. and hn = ~ ~. 2
j-I j=1

have the required properties.

Second step: The theorem holds for any sequence (!,,) from T•.
Proof of the second step. For every n EN there is a non-decreasing
sequence (gn, m)mEN from T with a bounded sequence of integrals such
that gn , m ~
p fn' Define gm = max {gn , m : n ~ m}. Then (gm) is a non-decreas-
ing sequence from T. Since gn • m "fn
p
"fm
p
for n "m, we have gm "p fm and
Igm dp ~ Ifm dp" C for all mEN. By Auxiliary Theorem A3 there exists
an f E T. for which gn ~ f and I gn dp~ If dp. From the inequality
p
gn • m ~p gm for all n <m we obtain, by letting m~oo, thatfn <f, p
and thus
gn ~p fn "p f· Consequently,Jn ~ f and If n dp~ Jf dp. 0
p

Theorem AS (Lebesgue's dominated convergence theorem). Let Un) be a

sequence from T2 for which fn ~ f. Assume that there exists agE T2 such that
p
IJ"I or.; g for
p
all n E N. Then f also belongs to T2 and Ifn dp~ If dp.

PROOF. For all nE'" let gn=sup{J",fn+.""}' Then gn belongs to T2 :

Since max {fn,Jn + d = (J" - fn+ .)+ + fn+ IE T2• it follows by induction that
max{fn,fn+ I , • •• ,fn+d E T2 • Moreover, since

Jmax{jn,fn+.' ... • fn+k} dp " Jg dp for all k E '"

and max{J",J,,+I" .. ,fn+k}~gn as k~oo, the function gn indeed belongs

to T2 by B. Levi's theorem. The sequence (gn) is non-increasing and
370 Appendix A: Lebesgue integration

I gn dp;> - I g dp. Therefore, f = p - lim gn E T2 and I gn dp~ If dp by B.

Levi's theorem.
If we define hn = inf{j", j,,+., ... }, then we can show analogously that
hn ~ f and I hn dp~ If dp. The assertion then follows, because hn "j" " gn'
p
o
Theorem A9. Let Un) be a sequence from T2 and assume that fn ~ f. If there
p
exists agE T2 such that If I " g, then f also belongs to T2•
p

PROOF. Let.in = mint g, max{j", - g}} for all n E N. Then.in E T2 • l.inl " g
for all n E N, and.in ~ j. The assertion therefore follows from Lebesgue's
p
theorem. 0

Theorem AIO (Fatou's lemma). Let Un) be a non-negative sequence from T2

for which the sequence of the integrals is bounded and fn ~ f. Then f E T2 and
p

J f dp " lim inf Jfn dp.

n->oo

PROOF. Let hn = inf{fn.Jn+.' ... } for all n E N. The sequence (h n) is

non-decreasing and hn ~ f. The inequalities hn <: fn+k for all n, kEN imply
p

J hn dp <: lim inf Jfn+k dp = lim inf ffk dp, n EN.

k->oo k->oo

Consequently, B. Levi's theorem implies that f E T2 and

J f dp = n-+oo
lim J hn dp " lim inC Jfn dp.
n---+oo
o

A.3 Measurable functions and sets

In the sequel let p be a measure on Rm. A function f: R'" ~ C is said to be

p-measurable if there exists a sequence Un) from T for which fn ~ j. It is
obvious that every continuous function defined on Rm is p-me:surable.
The sum, product, and quotient (if the denominator #: 0) of two p-measur-
p
able functions are p-measurable. Along with f, the function g 0 f is also
p-measurable for every continuous function g : C~C. In particular, If I is
p-measurable if f is p-measurable. Every f E T2(R m , p) is p-measurable
(cf. Theorem A6).
A.3 Measurable functions and sets 371

Theorem All.
(a) Iff: Rm~R is p-measurable and there exists agE T2 such that Ifl '"' g.
then f also belongs to Tz.
(b) If (J,,) is a sequence of p-measurable functions such that fn ~ f, then f is
p
p-measurable.
PROOF.
(a) This immediately follows from Theorem A9.
(b) Let h E Tz be such that hex) > 0 for all x E R m (the reader is advised to
prove the existence of such a function). With gn = (h + Iinj)-tJifn for all
n E N we have gn ~ g = (h + Ifl) -IJif. Since all gn are measurable and
p
since Ignl.;; h, we obtain that all gn belong to T2. Consequently, it
follows from Lebesgue's theorem that g E Tz. The function g is there-
fore p-measurable. This then holds for f= (h -I gj)-thg, as well. 0
A subset M of R m is said to be p-measurable if its characteristic function
XM is p-measurable. If XM is p-integrable. then the measure p(M) of M is
defined by the equality p(M) = fXM dp. If M is p-mea~urable and XM is not
p-integrable, then we set p(M) = 00. All sets M E ~(Rm) are obviously
p-measurable and have the finite measure p( M); for these M the definition
coincides with the earlier one. Every p-null set N is p-measurable with
p(N)=O.

Theorem A12. Countable unions and intersections of p-measurable sets, as

well as the complement of a p-measurable set are p-measurable. If
M I, M20 .. , are disjoint p-measurable sets, then

If Ml ~ M2 ~ ... are p-measurable sets and p(M t ) < 00, then

The proof of this theorem can be left to the reader. One has to consider
the characteristic functions XM and apply the previous theorems. The cases
where infinite measures occur"have to be treated carefully.

Theorem A13. A function f: Rm ~ R is p-measurable if and only if the set

M. = {x E Rm : f(x) ;;. s} is p-measurable for every s E R. The same holds for
the sets {xER m :f(x),",s}, {xERm :f(x»s} and {xERm : f(x)<s}.
372 Appendix A: Lebesgue integration

PROOF. Let f be p-measurable. We can assume without loss of generality

that s = 1 (otherwise we replace f by f - s + 1). Th~: function g =
min{ 1, max{O, f}} is p-measurable and g"-X/4 as n-H.('. The function X/4
is therefore p-measurable.
Let II.(. now be p-measurable for all s E R. Then the sets

{ x E IR m : f( x ) < I} = R m \ M, and M( s, I) = {x E IR"' : s .;;; j( x) < I }

are also p-measurable. The functions

f,,(x) = {kn-.for~EM(k-l)n-.'kn-.)'k=-n2+1, -n 2+2, ... ,n 2 ,

o otherwIse
are therefore p-measurable for all n E~. The assertion follows, because
L~ 0
The family of Borel sets in Rm is the smallest family of subsets of Rm
that contains all intervals and is closed with respect to taking complements
and countable unions and intersections. It is obvious that all open and all
closed subsets of Rm are Borel sets. A function f : Rm.....,. R is said to be
Borel measurable if the set {x E Rm : f( x) > s} is a Borel sell for every s E R
(A functionf: Rm_c is said to be Borel measurable if Ref and Imf are
Borel measurable.)
If p is a measure on IR m , then every Borel set is p-measurable and every
Borel-measurable function defined on Rm is p-measurabl(l. (The measures
considered here are hence called Borel measures, as well.)
Now we extend the notion of the integral to complex valued functions.
A function f: Rm_c is said to be p-integrable if Ref, ImfE T2 • We
define

Jf dp JRe f dp + i J
= 1m f dp.

The set of p-integrable functions is a complex vector space. It will be

denoted by E.(R m , p). The mapping E.(lRrn , p)-C,f~ffdp is obviously
linear.

Theorem A14.
(a) Iff: Rm_c is p-measurable and there exists a p-integrable function g
such thai If I .;;; g, then f is p-integrable and Iff dpl.;;; fg dp.
p
(b) Iff dpl.;;; flfl dp for every f E E.(R m , p).
(c) For every f E E.(R m , p) there exists a sequence U,,) from T for which
fif,,-fldp>O.
(d) For any p-measurable function f: Rm_c we have f ,= 0 if and only if
p
f E E.(lRm , p) and Jlfl dp = O.
A.3 Measurable functions and sets 373

PROOF.
(a) The functions Re f and 1m fare p-measurable, and IRe fl "p g and
IImfl " g. The functions Ref and Imf are therefore p-integrable by
p
Theorem All(a), i.e.,fE e.(R m , p). The inequality follows from (b).
(b) Let a = sgn(ff dp)*. Then by Theorem AS(d)

Iff dpl = af f dp = f Re( af) dp " f lafl dp = f Ifl dp.

(c) This follows by applying Theorem A6 to Ref and Imj.
(d) The equality f= 0 implies If I = 0, fE e.(R m, p) (since the sequence
p p
whose members are all zero p-converges to 1), and fIJI dp = O. If
flfl dp=O, then B. Levi's theorem can be applied to the sequence (g,,)
with gIl = nlfl. Hence there exists agE T2 for which gIl = nlfl--+ g. This
p
is possible only if f= O. 0
Theorem AIS (Lebesgue). Let (J,.) be a sequence from e.(R m, p) such that
e,
f" --+ f. Assume that there exists agE (Rm, p) such that If"I" g for all
p
n E 1\1. Then f E e.(R m , p) and ff" dp--+ ff dp.

PROOF. We can obviously apply Theorem A8 to the sequences (Re f,,) and
(lmf,,). This gives the assertion. 0

e
Corollary. If (f,,) is a sequence from 1(Rm, p) such that l:::'_ .Ilf,,1 dp < 00,
then there exists an fEe.(Rm,p) such that l:j-l./i--+f and ffdp=
p
l:::'_.ff" dp.
PROOF. We can apply B. Levi's theorem to the sequence (l:j-,I./iI)"EN and
obtain a gET2 ce.(R m,p) for which l:j-tl./il--+g. Consequently, there
exists a functionf: Rm--+C such that '2.'j-l./i--+J as n--+oo. Since l'2.j-,./i1
p
" g for all n E 1\1, the assertion follows from Lebesgue's theorem. 0

Let M be a p-measurable subset of Rm. A functionf: M--+C is said to be

p-measurable (p-;ntegrable) if the function
for xEM
for x f1. M
is p-measurable (p-integrable). If f: M--+C is p-integrable, then we define

J.fdp .. fjdp.
M

If f: Rm--+C is p-measurable (p-integrable),2en the restrictionflM of f to

M is p-measurable (p-integrable), since flM= X,,/. If f: Rm--+c is p-
374 Appendix A: Lebesgue integration

measurable and flM is p-integrable, then we define

fj dp = fjlM dp = f x"J dp.

We denote the vector space of p-integrable functions f: M-+C by
el(M, p).

Theorem A16.
(a) If f: M-+C is p-integrable, (M,,) is a sequence of mutually disjoint
p-measurable subsets of R"', and M = U "ENM", then f Mf dp =
"2"ENJ M"f dp.
(b) If f: M-+R is p-integrable and J,J dp <.;ap(K) (J Kf dp ;>ap(K» for
every p-measurable subset K of M, then f <.; a (f;> a).
p p
(c) If f: IR"'-+C is p-measurable, f E el(J, p), and JJf dp= 0 for all
bounded intervals J, then f = O.
p

PROOF.
(a) Apply the above corollary to the sequence (xM"j).
(b) Let K" = {x EM: f(x»a}. Then x~(f- a);> 0 and JMX~(f- a) dp=
O. By Theorem A14(d) we therefore have x~(f - a) = O. This gives the
• P
assertion.
(c) It is sufficient to consider real f. It is obvious that f gf dp = 0 for every
gET. If I is a bounded interval and Mel is p-measurable, then there
exists a sequence (g") from T such that 0 <.; g" <; I, g"( x) = 0 for x f/. I,
and gIl -+ XM' Then it follows from Lebesgue's theorem for the sequence
p
(g"f) that J"J dp = JX"J dp == O. By part (b) we obtain that f = O. 0
P

A.4 The Fubini-Tonelli theorem

In what follows let PI and P2 be measures on RP and IRq, respectively. Let p
denote the product measure on Rp+q (cf. AI, Example 3).

Auxiliary Theorem A17. If N is a p-null set in W+ q , then for PI-a/most al/

x E RP the set {y E IRq : (x, y) EN} is a Prnull set, i.e.,
NI = {xERP: {yElRq: (X,Y)EN} isnota p'}.-nullset)
is a PI-null set.
PROOF. Since N is a p-null set, there is a sequence (Jk ) of intervals for
which Jk = Jl,k X J 2, k' N C U kENJk' "2 kEN P(Jk) < 00, and each zEN is
covered by infinitely many Jk (we choose the union of infinitely many
A.4 The Fubini-Tonelli theorem 375

covers by intervals with total measures 2 - I, 2 - 2, 2 - 3, • • . ). We have

We can therefore apply B. Levi's theorem to the sequence of"step functions

Consequently, there is a PI-null set FI such that

~
keN
XJ,,.{x) f XJ". dp2 < 00 for all x E RP \ Fl'

It remains to prove that NI c Fl' Let Xo E RP \ Fl' Then

For every y E Rq such that (xo, y) E N the element (xo, y) belongs to

infinitely many Jk = J I, k X J2, k' The non-decreasing sequence

is therefore divergent. Since the corresponding sequence of integrals with

respect to y is bounded, it follows from B. Levi's theorem (or from
Auxiliary theorem A3) that {y E Rq : (xo, y) EN} is a P2-null set. There-
fore, Xo fl N .. and thus NI c Fl' 0
Theorem AI8 (Fubini). Let PI' P2 and P be as above, and let f E el(RPTq, p).
Then we have: For PI-almost a/l x E IRP the function f(x, .) belongs to
el (IRq, P2)' The function F defined by the equality

F(x) _ {!.. flX,Y) dp,(y)

otherwise

A similar result holds if we exchange the roles of x and y. To express the

content of this theorem, we briefly write

I f dp = ~p {I,/(x, y) dp2(y) } dpl(x)

= ~q{~!(x,y) dPI(X)} dP2(y), (AID)
376 Appendix A: LClbesgue integration

PROOF. Since f= fl - f2 + if3 - if4 with t

E h it is enough to study the
case f E TI. Then there exists a non-decreasing sequence (f,,) from T and a
p-null set N such that fll(z)~f(z) for all z E IRp+q\ Nand Ifll dp~ If dp.
Formula (A 10) is evident for the step functions fll' Define gll(x) =
ffll(x,y) dpiY). Then (gil) is a non-decreasing sequence from T(W) and
IglI dpi "If dp. By B. Levi's theorem there hence exists agE el(W, PI) for
which gil ~ g and I gil dpl~ I g dpl' Consequently, by AUKiliary theorem
PI
Al7 the set
No = {XEW: {yElRq: (x,y)EN} is not ap2-null set}
U {x E W : (gn(x» is not convergent}

is a PI-null set. For x fl No the non-decreasing sequence Un(x, . ) has the

properties

By B. Levi's theorem f(x, . ) therefore belongs to el(lRq, P2) for all x fl No

and

Since g is PI-integrable and g = F, the function F is also PI-integrable and

PI

J F dpi = J g dpi = J~~ J gn dpi

= "->00
lim f. {f. f,,(x,y) dP2(y)} dpl(x) = lim Jj" dp = Jfdp.
RP R9 "->00

o
Theorem A19 (Tonelli). Let PI' P2 and P be as above, and letf: IRp+q~c be
p-measurable. Assume that f(x, . ) E el(lRq, P2) for PI-almost all x E IRP and
that the function F defined by the formula

F(x) _ {t.. If(x.Y)1 dp,(y).

otherwise
belongs to el(RP, PI)' Then f E el(IRP+q, p).
PROOF. For every n E N let
M" = {(x,y)ElRp+q: If(x,y)1 <n, l(x,y)1 <n}, In = XM"f.
Since f is p-measurable, every fn is p-integrable by Tht:orerns A13 and
A.S The Radon-Nikodym theorem 377

A14(a). The sequence (Ifni) is non-decreasing and Ifnl~lfl. An application

of Fubini's theorem to IJ"I gives

fiJ"I dp = Lp {L)fn(x,Y)1 dP2(Y)} dpt(x)

<; Lp{~)f(x,Y)1 dP2(Y)} dPt(x) < 00.

Consequently, the p-integrability of Ifl follows from B. Levi's theorem.

Sincefis p-measurable,jE et(RP+q, p) by Theorem AI4(a). 0

A.S The Radon-Nikodym theorem

Let p and p, be two measures on Rm. The measure p, is said to be absolutely
continuous with respect to p (in symbols: p,«.p) if every p-null set is also a
p,-null set. (Then every p-measurable set is p,-measurable, as well).

Theorem A20 (Radon-Nikodym). Let p and p, be two measures on Hm. We

have p,«.p if and only if there exists a p-measurable non-negative function
h : Rm-+R such that XJh E et(R m , p) for every bounded interval J and
p,(M) = JXMh dp for every p-measurable set M (here weconsider the integral
to be equal to 00 if XMh is not p-integrable). Every p-measurable function is
also p,-measurable. Iff: Rm~c is p-measurable and p,-integrable, then

f f dp, f jh dp.
=

PROOF. If IL has the above form, then p,«.p obviously holds. Now let p,«.p
and let J be an arbitrary bounded interval in Rm. Let us consider the
Hilbert space L2(J, 'T) with the measure 'T = P + p,. The mapping

L2 (J, 'T) -+ c, f~ Jf dlL

J

is a continuous linear functional, since 1f.J dp,1 <; JJlfl d'T <; 'T(J)t/2I1fll.
By the Riesz representation theorem (Theorem 4.8) there exists agE
L 2(J, '1") (more precisely, a gE ~(J, 'T» such that

Jf dlL = f gf d'T
J J
for f E L (J, 'T).
2 (All)

If here we replace f by XM' where M is an arbitrary p-measurable subset of

J, then we obtain

p,(M) =f XM dp,
J
= fgXM d'T =
J
f. g d'T.
M
378 Appendix A: Lebesgue integration

Since p.( M) ..;; '7"( M), it follows that

o . ;; fMg d'7" ..;; '7"(M).

It follows from this by Theorem AI6(b) that

o . ;; g(x) ..;; I for p-almost all x E J.
We obtain from (A II) that

f (1- g)i dp. f gf dp

J
=
J
for i E LiJ, '7"). (Al2)

Now set N= {x E J: g(x) = I} and L= J\N. Then (AU) implies for

i=XN that

p(N) = f XN dp = f gXN dp = f (1- g)XN dp. = 0,

and thus p.(N) = 0, as well, because p.«p. If in (AI2) we set i= (1 + g + g2
+ . . . + gn)XN' then it follows for all n E N that

f. (I-gn+l)dp.= f.g(l+g+'" +gn)d/l.

M M
The integrands of both sides constitute non-decreasing sequences the
integrals of which are bounded by J-I.( M). The left integrand converges to
XL' and hence the left side tends to p,(M n L). By B. Levi's theorem there
exists an hEL,(J,p) such that g(l+g+'" +gn)~h as n~oo and
p
f Mg(1 + g + ... + gn) dp~ f Mh dp. Consequently, it follows that

p.(M) = p,(Ln M) + p,(Nn M) = p.(Ln M) = {h dp.

Since g ;) 0, we also have h ;;. o. Without loss of generality, we can choose
p p
h ;;.0.
Now let (In ) be a sequence of disjoint intervals for which IRm = U nENJn'
Let hn be functions such that

/L(M) = f. hn dp for every p-measurable set MCJn •

M

Let h : IRm~R be defined by the equalities h(x) = hn(x) for x E I n • Then h

has all properties required, since for every p-measurable subset M of IRm

p.(M) = ~ p.(Mn I n ) =
nEN
~ f.
nEN Mn~
h dp = f. h dp.
M
If i is p-measurable, then there is a sequence (j,,) from T such that in ~ i.
p
Consequently, we also have in ~ i, and thus i is p.-measur,able.
fl
A.S The Radon-Nikodym theorem 379

It remains to prove the last assertion for non-negative functions f. For

every n E N let
M",k = {x E R m ; (k - I)r" ';;;f(x) <kr"} for k = 1,2, ... , 22"
and
22 •
fro = ~ (k - 1)2-"Xfo(,.k·
k-)

J.
22• 22•
Jfn dJ-L = ~ (k-I)2-"J-L(M",k) = ~ (k-l)2-" h dp = ff"h dp.
k= I k= I M•. k

It follows from this by B. Levi's theorem that If dJ-L = ffh dp. D

A function F : R~C is said to be absolutely continuous if there exists a
A-measurable (i.e., Lebesgue measurable) function f: IR~C that is A-inte-
grable (Lebesgue integrable) over every bounded interval and

F(x) = F(O) + foxf(t) dA(t).2

This function f is called the derivative of F. (It is possible to show that F is

A-almost everywhere differentiable and F'(x) = f(x).) The derivative f is
A
uniquely determined by F. (This follows from Theorem A16(c).)
If p is a measure on IR, then P«A if and only if the function

f: R ~ IR, f(x) = { p«O, x» for x>O

-p(x,O]) for x.;;; 0
is absolutely continuous, i.e., if p is induced by an absolutely continuous
function f in the sense of Section AI, Example 2.
Let F and G be absolutely continuous functions on R, and denote by f
and g their respective derivatives. If F(O) == G(O) == 0, then we obtain by
Fubini's theorem that

fo\Fg + fG) dA = fox { get) folf(s) dA(s) + f(t) fo'g(s) dA(s)} dA(t)

= fox {f(s) ~Xg(t) dA(t) + g(s) ~xf(t) dA(t) } dA(s)

== 2F(x)G(x) - foxUG+gF) dA,

and thus
F(x)G(x) = fo\Fg+ fG) dA.
lHere we set H- - J~ for x<O.
380 Appendix A: u,besgue integration

If we do not necessarily have F(O) = G(O) = 0, then

F(x)G(x) - F(O)G(O) = fox(Fg+ JG) <lA.

Consequently, for - 00 <a <b < 00 we have the formula of integration by

parts

fba Fg dX = F(b)G(b) - F(a)G(a) - fba fG dX. (Al3)

 Appendix B

A representation theorem for

holomorphic functions with
values in a half-plane

A function h : R~C is said to be of bounded variation when it can be

written in the form h = hI - h2 + ih) - ih4' where the functions ~ : R~ R
are non-decreasing bounded functions. (We can show that h is of this form
if and only if there is a C ;> 0 such that ~"lh(bll) - h(all )1 "C for every
sequence «a", b"]) of disjoint intervals. The smallest C of this kind is called
the variation of h. We do not need this result here.) If h is a right
continuous function of bounded variation, then the integral

roo (z-t)-I dh(t)

-00
for z EC\R

can be considered as a Riemann-Stieltjes integral. We will retain this

notation in the sequel. We can also view this integral as a linear combina-
tion of the corresponding integrals with respect to the measures p~
(cf. Section AI, Example 2). Consequently, the theorems of Appendix A
are at our disposal.

Theorem Bl. Assume that w : R~R is right continuous and of bounded

variation, liml-+ _ 00 w( t) = 0, and

f(z) = f
-00
OO I
(z - t)- dw(t) for z E G= {z EC : 1m z >O}.

(a) For a/l t E R we have the Stieltjes inversion formula

wet) = lim lim -

-1 fl+8 Imf(s+i t:) ds.
8-->0+ <-->0+ 'TT - 00

(b) If fez) = 0 for all z E G, then w(t) = 0 for all t E R

381·
382 Appendix B: Holomorphic functions with values in a half-plane

PROOF. Part (b) follows from part (a). Consequently, it is sufficient to

prove (a). Since w is real-valued,

Imf(s+if) = L:Im[(s+it:-U)-'] dW(u)

= -fL:[(s-u)2+ f 2r' dw(u)

for every t: > O. It follows from this by Fubini's theorem that

f' ImJ(s + if) ds = -foo f' t: 2 ds dw(u)

-00 -co (s- u) + t: 2
-00

= - f OO[ r-u 'IT]

-00 arctan-f-+I <ilw(u).
Since
r-u 'IT
larctan-t:- + II <; 'IT for all r E!R
and
'IT for r >u,
r- U 'IT 'IT
arctan--+-.....:.. { - for r= u, as f:""':"O,
t: 2 2
o for r <u,
Lebesgue's theorem implies that

lim
( ..... 0.
f'
- 00
Imf(s + it:) ds

= - 1(-00,,)
'IT dw( u) - f i
{,}
d w( u) - 1(r,oo)0 d w( u)
'IT 7T
= -'lTw(r-)-I[w(r)-w(r-)] = -I[w(r)+w(r-)].

(In order to be able to apply Lebesgue's theorem, we write J ... dw =

J . .. dw, - J ... dW2' where WI and w2 are non-decreasing right continu-
ous functions, w = w, - W2' and lim....... _""w,(u)= lim....... _oc: w2(u)= 0.) If we
set r = t + 6 with 6 > 0 and let 8 tend to zero, then the assertion follows. 0

Theorem B2. Assume that w : IR.....:..C is right continuous, of bounded varia-

tion, and lim, __ ""w(t) = O. If f~",,(z - 1)-1 dw(t) = 0 for all z EC\!R then
w( I) - 0 for al/ 1 E R.

PROOF. For z E G= {z EC : 1m z >O} we have

f oo (z - t)-
-00
I
dw(t) =0
and
Appendix B: Holomorphic functions with values in a half-plane 383

Therefore,
J 00
-00
(z - t) - 1 d [Re w( t)] = J 00
-00
(z - t) - ~ d [1m w( t)] =0
for all z E G. It then follows from Theorem B 1(b) that Re w(t) = 1m w(t) =
0, and thus w(t) = 0 for all t E R. 0

lbeorem B3 (Herglotz). Let G .. {z E C : 1m z > O}, and let J: G-C be

holomorphic such that ImJ(z) ~ 0 and IJ(z)lm zl ~ M Jor al/ z E G. Then
there exists a unique right continuous non-decreasing Junction w : iii_iii Jor
which w( 1)-0 as t _ - 00 and

J(z) = foo (z -
-00
t)-I dw(t) Jor all z E G.

For all t E iii we have w(t) ~ M and

wet) = lim lim -

-I fl+6 ImJ(s + it:) ds.
6-->0+ <--->0+ 'IT - 00

PROOF. The last equality will follow from Theorem BI(a) if we prove the
existence of a function w having the remaining properties.

y-Im z
r,-II+N
r~-N
r .. -J+II+m

f .z*+2ie' x-Re z
o r

For O<t: <r let the paths 1',., r:, and r«J be defined as the above figure
shows. For z = x + iy E G such that 1m z =y >t: and for r > Izl the point
z* + 2i t: lies outside 1',.. Therefore, by the Cauchy integral formula

J(z) = - 12.
'IT 1
f.r,(!-z)-'J(r) d!

== _1_.
2'IT 1
f.r,(z - z* - 2 i t:)[(! - z)(r - z* -2 i t:)rIJ(n d!
= .!.. f.r,(y -
'IT
t:)[(r - z)(! - z* -2 i t:)] -'J(r) dC·
384 Appendix B: Holomorphic functions with values in a haH-plane

For t E r; we have for fixed z that

ij{nl" £-IM and l{y-£)I[(t-z)(t-z*-2i£)r l " Cr- 2•
The integral over r; therefore tends to °as r~oo, and there remains

j{z) ="!'f.roo (y-£)[(t-z)(t-z*-2i£)rlf(nd~

7T

= "!'foo (y-£)[(t+i£-z)(t-i£-z*)rlj(t+i£) dt
7T - 00

=..!. Joo (y-£)[(x-t)2+(y-£)2r lj{t+i£) dt.

7T -00

If we set v(z) = Imj(z), then it follows for 0 <£ < 1m z = y that

v{z) = .; foooo (y - £)[ (x - t)2 + (y - £)zrlv(t + i £) dt.

The inequalities lyv(z)1 " If(z) 1m zl <; M imply for 0<£ <y that

!..!. foo (y - £)2[ (x - 1)2+ (y - £)zrlv(t +i £) d/l = I(y -- £)v(z)1 " M.

." -00

By 10ttingy~oo, we obtain from Fatou's lemma (observe that v <;0) that

v( . +i £) is integrable over Rand

o " --If
7T -
00

00
vet + it:) dt <; M for all t: > 0.
Since

for 0 < t: <y,

it follows that
foo {{y - t:)[(x - 1)2 + (y - £irl-y[ (x - 1)2+ yZrl}t>(1 +i £) dl ~ 0
-00

as t:~O + . Therefore, for all z E G

v{z) = lim "!'fOO y[(x - 1)2+ y2rlv(/+i t:} dt .

• ->0+ 7T - 00

In what follows let

".{t) = -}
7T
ft -00
v(s+i t:) ds for 1 E R, t: > O.
The functions ". are all non-decreasing and bounded, 0 "".( t) "M for all
t E R.I Let us construct, with the aid of the diagonal process, a positive null
IThe following steps can be much shortened if we make use of the fact that the family of
measures induced by {". : 0 < ~ < I} is compact in the "vague" topology.
Appendix 8: Holomorphic functions with values in a half-plane 385

sequence (£11) in such a way that (,'Jo.,,(t» is convergent for all rational
numbers t. If we set
t'Jo(t) = lim ,'Jo.. (t) for rational t,
11--+ 00

then ,'Jo(s) <. ,'Jo(t) for all rational sand t such that s <. t. If we extend ,'Jo to a
function t'Jo : III ~ III by defining
,'Jo(t) = inf{ t'Jo(s) : s >1, s rational} for irrational I,
then t'Jo is obviously non-decreasing, and limHOO(t'Jo(t) - t'Jo( - t» <. M.
We show that in the sense of the Riemann-Stieltjes integral

v(z) = - f OO

-00
Y[ (x - t)
2 I
+ y2r dt'Jo(t) for z E G.
Since

= - lim
€--+o+
foo
_ 00
y[(x-t)2+ y 2r 1 d
d t t'Jo.(t)dt

- lim foo y[(x-t)2+ y 2r l d,'Jo.(t),

€--+o+ - 00

this assertion is equivalent to the equality

For the proof of this equality we notice that if we wish to approximate this
Riemann-Stieltjes integral (with a continuous integrand) by Riemann
sums, then it is enough to consider only partitions of (- 00, 00) with
rational division points. For every such rational partition P and for fixed
z = x + i y let Up, Lp, Up, n' and L p, n be the upper and lower sums of the
integrals

respectively

that correspond to P. For every rational partition P we obviously have

UP,n~ Up and Lp,II~Lp, For every 8> 0 there exists a rational partition P
for which Up - Lp <.8/2. For this P there is an no E 1\1 such that IUp, n -
Upl <.8/2 and ILp , l I - Lpl <.8/2 for all n > no· Since Lp,lI <'JII <. Up ,lI and
Lp <.J <. Up, it follows that IJ -Jnl <.8 for n >no' Therefore, JII~J as
n~oo.
386 Appendix B: Holomorphic functions with values in a half-plane

Consequently, we have shown that for z E G

ImJ(z) = v(z) = - j'JO y[(x-t)2+ y 2r l dt'J(t)

= 1m f_: (z -
-00

t)-I dt'J(t).

Since J and z~ f~ ",,(z -- t) - I dt'J(t) are holomorphic in G and have the

same imaginary part, it follows that

J(z) = foo (z -
-00
t)-I dt'J(t) + C with some C 1= IR.

Because IJ(z) 1m zl .;;; M and

1(lm z) foo (z -
-00
t)-I dt'J(t)1 .;;; foo I dt'J(t) .;;; M
-00
for z E G,

we must also have IC 1m zl .;;; 2M, and thus C = o.

If we now define
5(t) = 8lim
...... 0+
t'J(t+8) for tE R

and
w(t) = 5(t) - lim 5(x)
$-+- 00
for t E R,

then w has the required properties: The passage from t'J to 5 does not
change anything in the integral formula we have just proved, since 5 has at
most countably many points of discontinuity and they c:an be avoided
during the formation of the partitions. The passage from {. to w does not
influence the integral formula. 0
 References

(A) Textbooks and monographs

[1] Achieser, N. I.; Glasman, I. M.: Theorie der linearen Operatoren im Hi-
Ibertraum. 5. Aufl. Berlin 1968
[2] Bourbaki, N.: Elements de mathematique, Livre V; Espace vectoriels topolo-
giques. Paris 1967
[3] Berberian, S. K.: Lectures in functional analysis and operator theory. New
York-Heidelberg-Berlin 1974
[4] Dunford, N.; Schwartz, J. T.: Linear operators, Part I: General theory. New
York-London-Sydney 1958/1964
[5] Dunford, N.; Schwartz, J. T.: Linear operators, Part II: Spectral theory, self
adjoint operators in Hilbert space. New York-London-Sydney 1963
[6] Dunford, N.; Schwartz, J. T.: Linear operators, Part III: Spectral operators.
New York-London-Sydney-Toronto 1971
[7] Faris, W. G.: Self-adjoint operators. Berlin-Heidelberg-New York 1975
[8] Foias, C.; Sz.-Nagy, B.: Harmonic analysis of operators on Hilbert space.
Budapest-Amsterdam-London 1970
[9] Friedrichs, K. 0.: Spectral theory of operators in Hilbert space. Berlin-Heidel-
berg-New York 1973
[10] Gohberg, I. c.; Krein, M. G.: Introduction to the theory of non-self-adjoint
operators in Hilbert space. Providence, R. I. 1969
(IIJ Goldberg, S.: Unbounded linear operators, theory and applications. New
York-St. Louis-San Francisco-Toronto-London-Sydney 1966
(12) Halmos, P. R.: Introduction to Hilbert space and the theory of spectral
multiplicity. New York 1951
[13] Halmos, P. R.: Finite dimensional vector spaces. Princeton, N. J.-Toronto-
New York-London 1958
[14] Halmos, P. R.: A Hilbert space problem book. New York-Heidelberg 1974
(15) Hellwig, G.: Differentialoperatoren der mathematischen Physik. Berlin-Got-
tingen-Heidelberg 1964
387
388 References

(16] Helmberg, G.: Introduction to spectral theory in Hilbert spa,ce. Amsterdam-

London 1969
(17] Heuser, H.: Funktionalanalysis. Stuttgart 1975
(l8] Hewitt, E.; Stromberg, K.: Real and abstract analysis. Berlin-Heidelberg-New
York 1965
(19] Jorgens, K.: Lineare Integraloperatoren. Stuttgart 1970
(20] Jorgens, K.; Rellich, F.: Eigenwerttheorie gewohnlicher Differentialgleichun-
gen. Berlin-Heidelberg-New York 1976
(21) Jorgens, K.; Weidmann, J.: Spectral properties of Hamiltonian operators.
Berlin-Heidelberg-New York 1973
(22] Kato, T.: Perturbation theory for linear operators. Berlin-Heidelberg-New
York. Second Edition 1976 .
(23) Knobloch, H. W.; Kappel, F.: Gewohnliche Differentialgleichungen. Stuttgart
1974
(24] Maurin, K.: Methods of Hilbert spaces. Warschau 1967
(25] Maurin, K.: General eigenfunction expansions l\nd unitary mpresentations of
topological groups. Warschau 1968
(26] Nelson, E.: Topics in dynamics I: Flows. Princeton, N. J. 1969
(27) Neumark, M. A.: Lineare Differentialoperatoren. Berlin 1960
(28] Putnam, C. R.: Commutation properties of Hilbert space operators. Berlin-
Heidelberg-New York 1967
(29] Reed, M.; Simon, B.: Methods of modem mathematical physics I: Functional
analysis. New York-London 1972
(30] Reed, M.; Simon, B.: Methods of modem mathematical physics II: Fourier
analysis, self-adjointness. New York-London 1975
(31) Riesz, F.; Sz.-Nagy, B.: Vorlesungen tiber Funktionalanalysi:;. Berlin 1956
(32) Rudin, W.: Real and complex analysis. New York 1966
(33) Rudin, W.: Functional analysis. New York 1973
[34] Schechter, M.: Spectra of partial differential operators. Amsterdam-London
1971
[35] Stone, M. H.: Linear transformations in Hilbert space and their applications
to analysis. New York 1932
(36) Sz.-Nagy, B.: Spektraldarstellung linearer Transformationen des Hilbertschen
Raumes. Berichtigter Nachdruck. Berlin-Heidelberg-New York 1967
(37] Triebel, H.: Hohere Analysis. Berlin 1972
(38] Voigt, A.; Wloka, J.: Hilbertraume und elliptische Differentialoperatoren.
Mannheim-Wien-Ztirich 1975
[39] Wloka, J.: Funktionalanalysis und Anwendungen. Berlin-New York 1971
[40] Yosida, K.: Functional analysis. Berlin-Heidelberg-New York 1965

(B) Articles referred to in the book

[41] Faris, W. G.: Invariant cones and uniqueness of the ground state for fermion
systems. J. Mathematical Phys. 13 (1972) 1285-1290
[42] Hess, P.; Kato, T.: Perturbation of closed operators and their adjoints.
Comment. Math. Helv. 45 (1970) 524-529
[43] Hormander, L.: The existence of wave operators in scattering theory. Math. Z.
146 (1976) 69-91
References 389

(44) Ikebe, T.j Kato, T.: Uniqueness of the self-adjoint extension of singular
elliptic differential operators. Arch. Rational Mech. Anal. 9 (1962) 77-92
(45) Kato, T.: Scattering theory and perturbation of continuous spectra. Actes,
Congres intern. math., 1970. Band I, 135-140
(46) Korotkov, V. B.: Integral operators with Carleman kernels. Dokl. Akad. Nauk
SSSR 165 (1965) 1496-1499 (russisch)
(47) Kuroda, S. T.: Scattering theory for differential operators, I, operator theory.
J. Math. Soc. Japan 25 (1973) 75-104
(48) Kuroda, S. T.: Scattering theory for differential operators, II, self adjoint
elliptic operators. J. Math. Soc. Japan 25 (1973) 222-234
[49) Schechter, M.: Scattering theory for elliptic operators of arbitrary order.
Comment. Math. Helv. 49 (1974) 84-113
[50) Schechter, M.: A unified approach to scattering. J. Math. Pures Appl. 53
(1974) 373-396
[51) Simader, C. G.: Bemerkung tiber Schrodinger-Operatoren mit stark singularen
Potentialen. Math. Z. 138 (1974) 53-70
[52) Stummel, P.: Singulare elliptische Differentialoperatoren in Hilbertschen Rau-
men. Math. Ann. 132 (1956) 150-176
[53) Veselic, K.j Weidmann, J.: Existenz der Wellenoperatoren fUr eine allgemeine
Klasse von Operatoren. Math. Z. 134 (1973) 255-274
[54) Veselic, K.j Weidmann, J.: Asymptotic estimates of wave functions and the
existence of wave operators. J. Functional Analysis 17 (1974) 61-77
[55) Weidmann, J.: The vi rial theorem and its application to spectral theory of
Schrodinger operators. Bull. Amer. Math. Soc. 73 (1967) 452-456
[56) Weyl, H.: Uber gewohnliche Differentialgleichungen mit Singuiaritiiten und
die zugehOrigen Entwicklungen willktirlicher Funktionen. Math. Ann. 68
(1909) 220-269
[57] Wtist, R.: Generalisations of Rellich's theorem on perturbations of (essen-
tially) selfadjoint operators. Math. Z. 119 (1971) 276-280
[58] Wtist, R.: Holomorphic operator families and stability of selfadjointness.
Math. Z. 125 (1972) 349-358
'Index of symbols

A2 13 12, 12,0 12,23

X 22 12(A, p.) 13
B(HI' H 2) 59 L(B), L,.(B) 26
Bee(HIo H2), Bee(H) 132 E:z(M) 18
~(HI' H2), B2(H) 136 (L2(M), <',' », L (M) 2 19
Bp(HI , H2), Bp(H) 170 L2(M; p) 20
C,Cm 1,4 L2• 0(M) 24
CI 13 L2• R (M) 20
C[O, Ij, CR[O, 1J 5 L 2 ,o(G; p) 27
Ck[O, Ij 21 L2• • (Rm), <." >(.) 296,297
Co""(Rm), Co""( G) 25,27 L2(Rm)M 329,356
dim 46 el(R m, p) 372
D(T) 50 IMI 45
p(u) 184 M q• p (·) 305
F o, F, F, F I , FI 289-292 M/Rm), Mp,lok(Rm) 305
GKF,GPF 254 N, No 4, II
G(T) 70 neT, A) 244
G(u) 212 m(M) 19
H2 13 N+,N_ 237
(H, II' I\}, (H,<' ,.» II N(T) 52,98
Hac(T), Hc(T), HiT), N q (·) 305
H.( T), H.c( T) . 206 OND,ONS 34, 35
He, Te 199 PM 81
~=HRm) 362 R,Rm 1,5
i ... i(Rm) 362 R(T) 50
K I s-lim,~ 75
K(j, e) 21 supp 25
lim,~ 15 SeT), SeCT) 244
l.i.m. 294 S=S(Rm) 289
390
Index of symbols 391

T- I 50 p(T) 98
1'* 68 pIt) 181
f 88 11( T), I1p( T) 99
ITI 170, 197 l1e( T), I1d( T) 202
T(R m), T(G) 24,27,364 l1ae(T), l1e (T), l1iT),
T\ = TI(R m) 367 I1sc(T) 209
Tl - Tl(Rm) 367 O±(Tl , T I ) 337
Tk , Tx 142, 147, 153
Tk.O' Tx.o 146, 147
Tn' Tn. 0 160
C 22
TDC' Tc' 7;" T" Tsc 209 1.. 29
tr(T) 175 + 32
w-lim, ~ 76, 77 EB,e 32
Wl.n(a, b) 21, 160 ®,* 47, 48, 262
Wl •• (lRm) 297 11'11 7
Wl • r. iok(R m) 315 11'111' 11'1100 7
W ±(Tl , T\, P), <',' > 6
W±(Tl' T I ) 339 <',' >k 21
111'111 136
<.,. >r, II . II r 89
peT, z) 230 <." >(s), <." >s 297
y+=y+(T), C 339
y_ = y_(T) 231 r
l'(T) 229 ==,
p
~
p
364
Author and subject index

absolutely continuous antilinear functional 66

function 157, 379 approximation theorem 30
measure 377
part of an operator 209
spectrum 209 ball 21
subspace 206 ,closed 22
absolute value ,open 21
of a multi index 289 Banach
of an operator 169 algebra 60
additivity of an interval function space 17
362 of operators 59
adjoint theorem 94
,formal 67 Banach-Steinhaus theorem 74
kernel 69, 139 basis
of a linear functional 69 ,algebraic 45
operator 50, 67, 68 , existence of orthonormal 42,
algebra 44
,Banach 60 of a tensor product ~·9
,normed 60 ,orthonormal 34, 35
algebraic Bessel inequality 38
basis 45 bijective operator %
dimension 45 bilinear form I
tensor product 47 ,symmetric 2
almost all 363 Borel
almost everywhere 18, 363 measurable 182, 312
alternative measure 372
,Fredholm 136 set 372
,Weyl 254 bound
analytic vector 259 ,essential 54

392
Author and subject index 393

bound (cont.) ,weakly 81

,lower 115 closure 22
of an operator 53 of an operator 88
relative 93, 200 of a sesquilinear form 122
upper 44, 115 Colgen, R. 350
boundary conditions 240 commutator equation 63
,mixed 253, 258, 259 commute 127
,separated 253 compact 135
bounded normal operators
, essentiall y 54 , expansion theorem 167
from above 1I5, 118, 269 , spectral theorem 166
linear functional 56 operator 129, 130
operator 53, 56 ,spectrum 133
,pointwise 74 , relati vel y 135
,relatively 93, 200 sesquilinear form 136
sesquilinear form 120 ,vague 384
variation 381 compatible norms 122
Bourbaki, N. 46 complement 22
,orthogonal 29, 32
complete 17
Carleman kernel 146, 235 tensor product 48
Carleman operator 56,95, 136, 141, wave operators 344, 349, 350, 354,
200,300,301,303 361
characterization 143, 144 completion 63, 64, 66
inducing function 142 of a separable Hilbert space 66
,maximal 142 complexification 113, 114, 199,
Cauchy 243
inequality 6 complex numbers
sequence 15 complex spectral family 211
,strong 75 conjugation 199, 235, 243
,weak 76,77 ,natural 235
Cayley transform 231 contact point 22
chain rule 341 continuity from the right 180
characteristic function 24 continuous dependence of the
characterization of spectrum 218, 271
Carleman operators 143, 144 invertibility, stability 97
Hilbert-Schmidt operators 139, linear functional 60
140 operator 53
orthogonal projections 81 part of an operator 209
the spectrum 200 spectrum 209
closable strongly 220
operator 88 subspace 206
sesquilinear form 122 convergence
closed ,dominated 369, 373
ball 22 , graph - 284
graph theorem 88, 94 radius !OI
hull 22 ,strong 74
operator 88 , strong resolvent 282
set 22 ,weak 74
394 Author and subject index

convergent Dirac operator 329, 330

,absolutely 102 ,eigenvalues 334
sequence 15 , essentially self-adjoint 333, 336
,strongly 75 ,free 330
,weakly 76, 77 ,spectrum 334
convex subset 30 with electric field 332
convolution 295 direct sum 32
Cook's lemma 344 discontinuous
coordinated norms 96 part 209
core of an operator 88 subspace 206
Coulomb interaction 324 discrete spectrum 202
cross-norm 179 characterization 203
distance, Euclidean 12, 25
domain
decomposi lion, polar 197, 199 ,final 85
defect ,initial 85
index 230,231 of an operator 50
space 230 dominated convergence theorem 369,
degree of a polynomial 299 373
delta, Kronecker 35 Dunford, N. 173
dense 23
in 23
, relative to a subset 23 eigenelement 98
subset 23 eigenspace 98
subspaces of L, 24, 25 eigenvalue 98, 107
densely defined operator 57 ,isolated 204
derivative of an absolutely continuous ,smallest 325, 326
function 379 eigenvector 98, 107
determinant element, maximal 45
,Gram 46 elliptic
, Wronskian 249 operator 302
diagonal polynomial 302
matrix 5 E-measurable 182
sequence 77 equality, Parseval 39
differential form 299 equivalence relation 64
,Laplace 235 equivalent
,Sturm-Liouville 252 Cauchy sequences 64
differential operator spaces 63
,elliptic 302 ,unitary 86
,maximal 160,299 essential
,minimal 160 bound 54
with constant coefficients 157, infimum 54
299 spectral kernel 243
with constant spectral family 301 spectrum 202, 273 .. 282
with constant spectrum 299 characterization 203
dimension 42, 45 invariance 280
,algebraic 45 supremum 54
, Hilbert space 45 essentially
of a Hilbert space 42 bounded 54
Author and subject index 395

essentially (cont.) , rapidly decreasing 289

self-adjoint 108,231,234,243 , step- 24, 363
Euclidean functional
distance 12, 25 ,antilinear 66
length 12 , bounded linear 56
even function 33 ,linear 50, 53
expansion theorem 40
for compact normal operators 167
extension gap in the essential spectrum 280
,finite-dimensional 239
generation of norms 8
,Friedrichs' 118,120,123,200
generator, infinitesimal 220
of an operator 51, 58 G-measurable 212
, real self-adjoint 241, 242
Gohberg, I.e. 173
, self-adjoint 114, 229, 237, 243
Gram determinant 46
theory, von Neumann 229
Gram-Schmidt orthonormalization 36
graph 70
convergence 284
Faris, W. 325
norm 89
Fatou's lemma 370
of the adjoi nt operator 71
Fejer's theorem 35
of the inverse operator 71
final doinain of a partial isometry 85
ground state 325
finitary sequence 23
group
finite-dimensional 43
, one-parameter 220
extension 239
,unitary 220
operator 129
restriction 239
finite rank operator 129
formal adjoint 67 Halmos, P.R. 95
form Hardy-class 13
, bilinear Heinz, E. 271
,quadratic 2 Hellinger-Toeplitz theorem 94
, sesquilinear Hellwig, G. 247,254
Fourier transformation 289, 290, 292 Herglotz theorem 383
Fredholm alternative 136 Hermitian
Friedrichs' extension 118,120,123, operator 72, 107
200 sesquilinear form 2
Fubini theorem 375 Hess, P. 111
function Hewitt, E. 24, 284, 362
, absolutely continuous 157, 379 Hilbert
, Borel measurable 372 sequence space 12
, characteristic 24 space 17
, even 33 , dimension 42, 45
, integrable 367, 372, 373 Hilbert-Schmidt
, measurable 141, 356, 370, norm 62, 136
373 operator 56, 62, 69, 136
,odd 33 characterization 139, 140
of an interval 362 holomorphic function 101
of bounded variation 381 Hormander, L. 354
, permissible 356 hull, closed 22
396 Author and subject index

ideal, two-sided 132, 138, 173 isometric operator 81, 85

identity isometry 63, 85
element 60 , partial 85
operator 60 isomorphic spaces 63
, polarization 2, 9 isomorphism 63, 85
Ikebe, T. 315
imaginary part 127,213,215,219
independent, linearly 35 Jordan-von Neumann theorem 9
inequality JOrgens, K. WI, 247, 254
,Bessel 38
,Cauchy 6
, Schwartz 4, 46 Kappel, F. 249
, triangle 7. Kato, T. 111,114,315,344,349,352
infimum, essential 54 kernel 6, 52, 69, 153
infinite-dimensional 43 ,adjoint 69, 139
infinitesimal generator 220 ,Carleman 146
initial domain of a partial isometry 85 ,Hilbert-Schmidt 139
injective operator 50 of an integral operator 153
inner product 6 ,spectral 243
integers Knobloch, H.W. 249
,non-negative II Kodaira, K. 247
, positive 4 Korotkov, V.B. 143
integrable 18 K-real 235
,function 367, 372, 373 Krein, M.G. 173
integral 182, 363 Kronecker delta 35
,Lebesgue 368 Kurado, S.T. 360
operator 55, 149, 153
, Volterra \06
, Riemann 368 Laplace differential form 235
, Riemann-Stieltjes 381 la w, parallelogram 2, <I
with respect to a spectral family 184, LCC 254
212 Lebesgue
integration integral 368
by parts 380 integration 18, 362
,Lebesgue 18 measure 18, 363
with rf~spect to a spectral family 180, space 18
212 theorem 369, 373
interaction, Coulomb 324 Legendre polynomials 37
intertwining property 337 ,generalized 38
interval 24 lemma, Zorn's 45, 190
function 362 length
, regular 363 of a vector 9
invariance , Euclidean 12
of the essential spectrum 280 Levi theorem 368
principle 352 lim 15
invariant subspace 128 limit 15
inverse operator 50 circle case 254
inversion formula, Stieltjes 381 in mean, l.i.m. 294
isolated point of the spectrum 204 point case 254
Author and subject index 397

limit (cont.) mixed boundary conditions 253,258,

,strong 75 259
theorems 368 monotone operator sequence 79, 83
,weak 76,77 monotonicity 180
linear of an interval function 362
functional 50, 53 multi index 289
,bounded 56 ,length 289
,continuous 61 , absolute value 289
hull 22 multiplication operator 51, 91, 125
operator 50 ,maximal 189
,bounded 56 ,spectrum 103
linearly dependent modulo a subspace 6 , spectral family 216
linearly independent 35 multiplicity of an eigenvalue 98
modulo a subspace 253
linearly ordered 44
locally natural conjugation 235
square integrable 305 Nelson, E. 261
constant 230 Neumann, J. von 9, 229
local Sobolev space 315 first formula 237
lower bound 115 second formula 238
LPC 254 Neumann series 100, 102
nilpotent 106
,quasi- 106
matrix 5 non-decreasing operator sequence 83
,diagonal 5 non-increasing operator sequence 83
function, permissible 356 non-negative
operator 149 elements 326
,Pauli 330 integers 11
maximal operators 83, 115
Carleman operator 142 quadratic form 4
differential operator 160,248 sesquilinear form 3
element 45 non-positive operator 115
multiplication operator 51, 189 norm 6,7
normal 125 ,coordinated 96
operator 51, 160, 189,248 ,cross- 179
orthonormal system 35 , generated by a scalar product 8
symmetric operator 239 ,graph- 89
measurable 18, 182, 196 ,Hilbert-Schmidt 62, 136
,Borel- 182, 372 not generated by a scalar product 13
function 141, 356, 370, 373 of an operator 54
with values in a Hilbert space 142 of a sesquilinear form 120
set 370, 371 of a symmetric operator 58
measure 363, 371 ,trace- 176
, absolutely continuous 377 normal operator 125, 210
,Borel 372 , essential spectrum 282
,Lebesgue 18, 363 normed
,product 374 algebra 60
m-dimensional 42 ,complete 17
minimal operator 160, 248 space 11
398 Author and subject index

nth root 196, 197 product 50

null set 362 , real symmetric 241
numbers ,normal 125, 199
, real , not closable 90
, complex ,self-adjoint 72, 107
numerical range 95 ,semibounded 234
,semi-Carleman 146
sequence
odd function 33 ,monotone 79, 83
ONB 34, 35, 42, 44 ,non-decreasing 83
one-parameter group 220 ,non-increasing 83
ONS 34,35 ,Sturm-Liouville 24'1, 252
open sum 50
ball 21 ,symmetric 72, 107
set 21 ,unitary 81, 85, 125
operator 50 , Volterra integral 106
, absolute value 169 ordered
,adjoint 50, 67, 68 , linearly 44
algebra 60 ,partially 44
,bijective 96 order of a differential form 299
, bounded linear 53, 56 orthogonal 29
,Carleman 56,95, 136, 141,200, complement 29, 32
300, 301, 303 projection 31,41,81
,closable 88 characterization 81.
,closed 88 sum 32,33,34, 128
,closure 88 orthogonalization, Gram··Schmidt 36
,compact 129, 130 orthonormal
,continuous 53 basis 34, 35
,core 88 construction 36
defined by a sesquilinear form 120 existence 42, 44
,differential 157 system 34
,Dirac 329, 330 ,maximal 35
, densely defined 57
,elliptic 302
, finite rank 129 parallelogram law 2, 9
,Hermitian 72, 107 Parseval equality 39
,Hilbert-Schmidt 56, 62, 69, 136 partial
,identity 60 integration 380
,injective 50 isometry 85
,integral- 55, 149, 153 partially ordered 44
,inverse 50 Pauli matrices 330
, isometric 81, 85 permissible function 356
,linear 50 perturbation 269
,maximal 160, 248 , relatively bounded 269, 304
,matrix- 149 , relatively compact 273, 311
, minimal 160 point
, multiplication 51,91,125 , contact- 22
, non-negative 83, 115 characterization 202
, non-positive 115 spectrum 99, 200
Author and subject index 399

pointwise bounded 74 radius

polar decomposition 197, 199 of a ball 21, 22
polarization identity 2, 9 of convergence 10 1
polynomial 299 ,spectral 102
polynomials Radon-Nikodym theorem, 377
, generalized Legendre 38 range
,Legendre 37 ,numerical 95
,trigonometric 35 . of an operator 50
positive rapidly decreasing function 289
elements 326 real
integers 4 extension 241, 242, 243, 259
quadratic form 4 numbers 1
sesquilinear form 4 operator 235
positively part 127,213,215,219
improving 326 quadratic form 2
preserving operator 326 self-adjoint extension 241, 242
pre-Hilbert space 11 symmetric operator 241
principle with respect to a conjugation
,invariance 352 235
, uniform boundedness 74 reducing subspace 128, 207
principle part 302 regular
product boundary point 248
,inner 6 interval function 363
measure 374 regularity domain 229
of operators 50 relative bound 93, 200
,scalar 6 relatively
,semi-scalar 7 bounded 93, 200, 269, 304
,tensor- 47 compact 135, 273, 275, 304,
projection 309
,orthogonal 31, 41, 81 Rellich, F. 247, 254
theorem 29, 31 Rellich-Kato theorem III, 114
pure representation 220
absolutely continuous spectrum resolvent 99
209 identity, first 99
continuous spectrum 209 identity, second 99
point spectrum 209 of a differential operator with constant
singular continuous spectrum 209 coefficients 300
singular spectrum 209 set 98
Pythagorean theorem 29 restriction
,finite-dimensional 239
of an operator 51
quadratic form 2 Riemann integral 368
,non-negative 4 Riemann-Lebesgue theorem 296'
,positive 4 Riemann-Stieltjes integral 381
,real 2 Riesz, F. 61, 362
quantum mechanics 225 theorem 61
,non-relativistic 314 root, nth 169, 197
,relativistic 330 Rosenblum, M. 349
quasi-nilpotent 106 Rudin, W. 354,362
400 Author and subject index

scalar product 6 space, Hilbert- 12

, vector space with 11 , weakly convergent 74
scattering series, Neumann 100, 102
operator 344 sesquilinear form 1
theory 337 ,bounded 120
Schechter, M. 304, 360 ,closable 122
Schmidt orthonormalization 36 ,closure 122
Schrodinger operator 313,314 ,compact 136
eigenvalues . 324 ,Hermitian 2
, essentially self-adjoint 313 ,non-negative 3
of N particles 314 ,positive 4
of one particle 314 ,semi-bounded 121
spectrum 323 set
Schwartz, J.T. 173 ,Borel 372
Schwartz space 289 ,closed 22
Schwarz inequality 4, 46 convex 30
self-adjoint 231 ,measurable 370, 371
,essentially 108,231,234,243 ,null- 363
extension 114 ,open 21
construction 237 sigma-finite 56
of a bounded Hermitian Simader, C.O. 315,321
operator 116 similar operators 219
of continuously invertible simple tensor 47
operators 115, 118 singular
of ordinary differential boundary point 248
operators 240, 251, 256 continuous part of an operator 209
of semi-bounded operators 11, 118, continuous spectrum 209
123 continuous subspace 206
spectrum 243 numbers 170
operator 72, 107 part of an operator 209
criteria 108, 109 sequence 274,280,282
spectral theorem 191 spectrum 209
semi-bounded subspace 206
operator 234 values 170
sesquilinear form 121 small at infinity 321
semi-Carleman operator 146 s-numbers 170
seminorm 7 Sobolev space 159, 297
semi-scalar product 7 space
separable 26 ,Banach 17
separability , complete normed 17
of I, 26 ,Hilbert 17
of L,(M) 27 sequence 12
separated boundary conditions 253 ,Lebesgue 18
sequence ,normed 11
,Cauchy 15 , not separable 27
,convergent 15 ,pre-Hilbert 11
,diagonal 77 ,Schwartz 289
, fini tary 23 ,separable 27
,singular 274, 280, 282 ,Sobolev 159, 297
Author and subject index 401

spectral family 180 strongly

,complex 211 continuous 220
examples 180, 181 convergent 75
of a compact normal operator 214 Stummel, F. 305
of a compact self-adjoint Sturm-Liouville operator 247, 252
operator 195 subset
of a differential operator with constant ,compact 135
coefficients 301 ,closed 22
of an orthogonal projection 195 ,convex 30
of a self-adjoint operator 191 ,dense 23
of the identity operator 195 ,open 21
of the multiplication operator 194 , relatively compact 135
of the zero operator 195 ,separable 26
spectral subspace 6, 23
kernel 243 , absolutely continuous 206
,essential 243 ,continuous 206
radius 102 ,discontinuous 206
representation theorem 195,217 ,invariant 128
theory 96 ,reducing 128, 207
theorem for compact normal ,singular 206
operators 166 continuous 206
theorem for normal operators 213, sum
215 ,direct 32
theorem for self-adjoint operators of operators 50
191 ,orthogonal 32, 33, 34, 128
spectrum 99, 200 support 25
, absolutely continuous 209 supremum, essential 54
characterization 200 symmetric
,continuous 209 bilinear form 2
,discrete 202, 203 operator 72, 107
,essential 202, 273, 282 ,maximal 239
of a compact operator 133 ,norm 58
of a multiplication operator 103 system, orthonormal 34
of a Schrodinger operator 323 Sz.-Nagy, B. 362
of a unitary operator 104
,point 99, 200
, singular continuous 209 T-bounded 93
square integrable, locally 305 T-compact 275
stability of continuous invertibility TP-compact 278
97 TZ-compact 279
step function 24, 363 tensor product 47
Stieltjes inversion formula 381 ,algebraic 47
Stone theorem 222 basic 49
Stromberg, K. 24, 284, 362 ,complete 48
strong of L2 -spaces 48
Cauchy sequence 75 of self-adjoint operators
convergence 74 259
limit 75 tensor, simple 47
resolvent convergence 282 Titshmarsh, E.C. 247
402 Author and subject inde)(

theorem uniform boundedness principle 74

,approximation 30 unitary
,Banach 94 equivalent 86
,Banach-Steinhaus 74 group 220
, closed graph 88, 94 operator 81,85, 125
,convolution 295 spectrum 104
,expansion 40 unit Vf'~tnr
,Fatou 370 in IKm 35
,Fejer's 35 in I, 35
,Fubini-Tonelli 374, 375, 376 upper bound 44, 115
,Heinz 271
, HelIinger-Toeplitz 94
,Herglotz 383 vague compact 384
, Jordan-v. Neumann 9 variation 381
,Kato-Rosenblum 349 vector
,Korotkov 143 ,analytic 259
, Lebesgue's dominated space with scalar proCiuct II
convergence 369, 373 Veselic, K. 354
,Levi 368 vi rial theorem 324, 334
,Nelson 261 Voigt, J. 242
,projection 29, 31 Volterra integral operator 106
,Pythagorean 29 volume function 363
,Radon-Nikodym 377
,Rellich-Kato III, 114 wave operator 228, 337, 340
,Riemann-Lebesgue 296 completeness 343, 344
, Riesz, F. 61 existence 343
,spectral 166, 191,213,215 ,generalized 339
,Stone 222, 225 weak
,Trotter 226 Cauchy sequence 71>, 77
,Wiener 296 convergence 74
,virial 324, 334 limit 76, 77
,Wiist 112 weakly
T-norm 89 closed 81
Tonelli theorem 376 convergent 76, 77
topological notions 21 Weidmann, J. 324, 3:;4
total 26 Wiener, N. 296
trace 175 Weyl, H. 247,254
class 175 Wey1, H., alternative 254
norm 176 Wronskian determinant 249
triangle inequality 7 Wiist, R. 112
trigonometric polynomials 35
Trotter, H. 226
transformation, Fourier 289, 290, 292 Zorn's lemma 45, 190