P. 1
A Hilbert Space Problem Book

A Hilbert Space Problem Book

|Views: 1,078|Likes:
Published by Michael Raba

More info:

Published by: Michael Raba on Apr 09, 2012
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

11/19/2013

pdf

text

original

Graduate Texts in Mathematics

Springer Jl&tl!Yfff

Paul R. Hatmos

A Hilbert Space

Probt~fti~(liok

Second~ l:dition

......._____

. . .UI

naer-¥erlag York Heidelberg Berlin

Paul R. Halmos

,A Hilbert~Space~ Problem Book
,---.__

Second Edition, Revised and Enlarged

Springer-Verlag New York Heidelberg Berlin

Editorial Board

P.R. Halmos
Managing Editor

F. W. Gehring
University of Michigan Department of Mathematics Ann Arbor, Michigan 48104 U.S.A.

C. C. Moore
University of California at Berkeley Department of Mathematics Berkeley, California 94720 U.S.A.

Indiana University Department of Mathematics Bloomington, Indiana 47405 U.S.A.

AMS Subject Classifications (1980): 46-01, OOA07, 46CXX

Library of Congress Cataloging in Publication Data Halmos, Paul R. (Paul Richard), 1916A Hilbert space problem book. (Graduate texts in mathematics; 19) Bibliography: p. Includes index. I. Hilbert spaces-Problems, exercises, etc. I. Title. II. Series. QA322.4.H34 1982 515.7'33 82-763 AACR2

© 1974, 1982 by Springer-Verlag New York Inc. All rights reserved. No part of this book may be translated or reproduced in any form without written permission from Springer-Verlag, 175 Fifth Avenue, New York, New York 10010, U.S.A.
Typeset by Composition House Ltd., Salisbury, England. Printed and bound by R. R. Donnelley & Sons, Harrisonburg, VA. Printed in the United States of America.

9 8 7 6 54 3 2 1

ISBN 0-387-90685-1 Springer-Verlag New York Heidelberg Berlin ISBN 3-540-90685-1 Springer-Verlag Berlin Heidelberg New York

To J. U. M.

Preface

The only way to learn mathematics is to do mathematics. That tenet is the foundation of the do-it-yourself, Socratic, or Texas method, the method in which the teacher plays the role of an omniscient but largely uncommunicative referee between the learner and the facts. Although that method is usually and perhaps necessarily oral, this book tries to use the same method to give a written exposition of certain topics in Hilbert space theory. The right way to read mathematics is first to read the definitions of the concepts and the statements of the theorems, and then, putting the book aside, to try to discover the appropriate proofs. If the theorems are not trivial, the attempt might fail, but it is likely to be instructive just the same. To the passive reader a routine computation and a miracle of ingenuity come with equal ease, and later, when he must depend on himself, he will find that they went as easily as they came. The active reader, who has found out what does not work, is in a much better position to understand the reason for the success of the author's method, and, later, to find answers that are not in books. This book was written for the active reader. The first part consists of problems, frequently preceded by definitions and motivation, and sometimes followed by corollaries and historical remarks. Most of the problems are statements to be proved, but some are questions (is it?, what is?), and some are challenges (construct, determine). The second part, a very short one, consists of hints. A hint is a word, or a paragraph, usually intended to help the reader find a solution. The hint itself is not necessarily a condensed solution of the problem; it may just point to what I regard as the heart of the matter. Sometimes a problem contains a trap, and the hint may serve to chide the reader for rushing in too recklessly. The third part, the vii

and the hint did not help. and the convergence of nets. and real and complex analysis. in mathematics. It can perfectly well happen that a relatively unsophisticated fact about operators is the best tool for the solution of an elementary-sounding problem about the geometry of vectors. (Reference: [61]. however. measure theory. such concepts as subbase for a topology. discuss some of the history of the subject. and of special cases (what happens in the finite-dimensional case?). To get maximum profit from the book the reader should know the elementary techniques and results of general topology.) From real analysis I need. the best thing to do at first is to go on to another problem. on the other hand. answers. you solved a problem. I use concepts such as a-fields and LP spaces. and specializations that you did not think of. of course. I did not try to achieve maximal generality in all the directions that the problems have contact with. not legal technicalities. precompact metric spaces. almost as bad: it tends to conceal the beauty of the subject and to delay or even to make impossible an understanding of the full truth. What makes the assertion true? What would make it false? Problems in life. do not necessarily arise in increasing order of depth and difficulty. An insistently linear view of the intricate architecture of mathematics is. do not hesitate to use it later. but try to think of related questions. depending on the nature of the problem. I made an attempt to exclude dull problems with routine answers. and results such as that LP convergent sequences have almost everywhere convergent subsequences. The topics treated range from fairly standard textbook material to the boundary of what is known.PREFACE longest. of generalizations (what if the operator is not normal?). or constructions. and even in this book. with no apology and no reference. at least. every problem in the book puzzled me once. Lindelof spaces. I tried to communicate ideas and techniques and to let the reader generalize for himself. The solution may introduce some standard nomenclature. the facts about the derivatives of absolutely continuous functions. I use. look at the hint.) From measure theory. and such results as the metrizability of compact spaces with a countable base. may throw valuable light on the solution. and the Lebesgue dominated convergence theorem. If you cannot solve a problem. If the problem was a statement. its use. and mention some pertinent references. Do not be discouraged if the solution of an early problem borrows from the future and uses the results of a later discussion. Do not just answer the question. If. and the compactness of the Cartesian product of compact spaces. and here is how it goes) will miss a lot of the point. generalizations. (Reference: [87]. consists of solutions: proofs. The logical error of circular reasoning must be avoided. connectedness. anyway. The problems are intended to be challenges to thought. and the Weierstrass polyviii . A reader who offers solutions in the strict sense only (this is what was asked. You may find modifications. and then the solution. or possible misuse. and he will miss a lot of fun.

equivalently. I now use f and g for vectors. pedantic {0}. Thus.PREFACE nomial approximation theorem. and the maximum modulus principle. As far as Hilbert space is concerned. is denoted by sgn z. as usual.) Incidentally kernel and range are abbreviated as ker and ran. at the very least. except in a few small details.L and ran. and operator means bounded linear transformation. Some knowledge of that subject is a prerequisite. (Reference: [26]. and determinant and trace are abbreviated as det and tr. (2) the continuous analogue of a matrix. (Reference: [120]. the barred arrow in x ~ x 2 denotes the function q> defined by V V '/i ix . by Reand Im. is regrettable but unavoidable. instead of the correct. i. To avoid even a momentary misunderstanding.. the dimension of the quotient space it defines). I cannot imagine any circumstances where it could lead to serious error. to denote (1) null-space. and an operator. The notation and terminology are mostly standard and used with no explanation. When I ascribe a result to someone without an accompanying bracketed reference number. then M v N is the smallest closed linear manifold that includes both M and N.) This is not an introduction to Hilbert space theory. by the way. The symbols (as a prefix) and v (as an infix) are used to denote spans. z/izl orO according as z i= 0 or z = 0.) From complex analysis I need such things as Taylor and Laurent series. if M and N are sets of vectors. a function. instead of x andy (the latter are too useful for points in measure spaces and such). more than likely the result is a folk theorem. so that if M is an arbitrary set of vectors. dimension is abbreviated as dim. subuniform convergence. Ideally the reader should know something like the first two chapters of [50]. (The simpler notation is obviously more convenient. I follow [50]. and it is not a whit more illogical than the simultaneous use of the symbol "0" for a number. then M is the smallest closed linear manifold that includes M. The zero subspace of a Hilbert space is denoted by 0. in conformity with current fashion. however. and (3) the reproducing function associated with a functional Hilbert space. a study of the elements of Hilbert space theory should proceed concurrently with the reading of this book. it does not seem to lead to any confusion. their orthogonal complements are abbreviated as ker. for instance. I am referring to an oral communication or an unpublished preprint. Real and imaginary parts are denoted. a vector. but in many cases I could find no reference. (The triple use of the word. Subspace. I use "kernel" instead of "nullspace". and.L. then M i is the smallest closed linear manifold that includes each Mi. When I make no ascription. means closed linear manifold. The "signum" of a complex number z.) The co-dimension of a subspace is the dimension of its orthogonal complement (or. I am not claiming originality. I tried to indicate where I learned the problems and the solutions and where further information about them is available. and if {Mi} is a family of sets of vectors. The arrow in a symbol such asfn --+/indicates that a sequence ifn} tends to the limit/. I write {0} for the set of complex numbers consisting of 0 alone.e.

I am truly grateful to them all just the same. not everything could be fitted in. something like ten of the problems (or their solutions) were substantially revised. The journal literature needs time. and Toeplitz and Hankel operators.. g). My special thanks are due to Ronald Douglas. Since I have been teaching Hilbert space by the problem method for many years. n = 1. ! 2 . and the weak and strong operator topologies. and much of it is not yet ready for that. just as integrals in calculus and quantifiers in logic bind theirs. and it has typographical advantages. cyclic operators. (Whether the actual number is 8 or 9 or ll or 12 depends on how a "substantial" revision is defined.} is the sequence whose n-th term is x. however. (The image of a set M of complex numbers under the mapping z ~ z* is M*.) Since the inner product of two vectors f and g is always denoted by (f. Some of that progress is visible in the difference between the two editions. and {x. but it is widely used by physicists. and Donald Hadwin and David Schwab for the second. I hope that the number of incorrect or awkward statements and proofs is smaller in this edition. In principle equations such as (x ~ x 2 )(y) = y 2 make sense.H. the three most frequent ones are total sets of vectors. · · ·. ••• ). I use z*. to become understood and simplified enough for expository presentation in a book of this sort.R. thus {x} is the set whose only element is x. and I am sorry about that. This tends to make mathematicians nervous. The first edition had 199 problems. this one has 199 . Without them this book could not exist. the Calkin algebra. I use braces to denote both sets and sequences. Eric Nordgren. another symbol is needed for their ordered pair. X . 2. it is in harmony with the standard notation for the adjoints of operators. I had to choose. it is not the sort of book that could have been written in isolation from the mathematical community. Maybe next time. I use (f. to ripen. For the complex conjugate of a complex number z. In accordance with inconsistent but widely accepted practice. almost the whole manuscript) and stopped me from making many foolish mistakes.) Operator theory has made much progress since the first edition of this book appeared in 1967. This leads to the systematic use of the angular bracket to enclose the coordinates of a vector. 3. Indiana University P. This could lead to confusion. Each of them read the whole manuscript (well. but in context it does not seem to do so. and Carl Pearcy for the first edition. the symbol M suggests topological closure. Even in the part that is reaqy.9 + 60. I omitted beautiful and useful facts about essential spectra. I owe thanks for their help to more friends among students and colleagues than I could possibly name here. as in (j0 . g). j 1 .PREFACE q>(x) = x 2 • (Note that barred arrows" bind" their variables. In any event.) The new problems have to do with several subjects.

SPACES 13. Vector sums and the modular law 16. Weak continuity of norm and inner product 21. Local compactness and dimension 17. Strict convexity 5. Weak metrizability of the unit ball xi . Infinite Vandermondes II. Approximate bases 2. T -total sets 12. Total sets 9. Weak closure of subspaces 20. Vector sums 14.Contents 1. Measure in Hilbert space 3. Limits of quadratic forms 2. Semicontinuity of norm 22. Representation of linear functionals 4. Schwarz inequality 3. WEAK TOPOLOGY 19. Linear dimension 8. Separability and dimension 18. Weak compactness of the unit ball 24. Uniqueness of crinkled arcs 7. VECTORS 1. Weak separability 23. Infinitely total sets 10. Lattice of subspaces 15. Continuous curves 6.

30. Unbounded symmetric transformations 7. Dimension in inner-product spaces 55. Closed graph theorem 59. Real functions in H 2 34. Dirichlet problem 5. Uniform boundedness of linear transformations 52. Column-finite matrices Schur test Hilbert matrix Exponential Hilbert matrix Positivity of the Hilbert matrix Series of vectors 6. Weak closure of the unit sphere Weak metrizability and separability Uniform boundedness Weak metrizability of Hilbert space 2 Linear functionals on 1 Weak completeness 4. Analytic Hilbert spaces 32. Analytic characterization of H 2 36. Diagonal operators Multiplications on /2 Spectrum of a diagonal operator Norm of a multiplication xii . 63. Kernel functions 38. Boundedness on bases 51. 46.CONTENTS 25. Basis for A 2 33. Total orthonormal sets 56. Radial limits 41. 45. 29. 28. 49. INFINITE MATRICES 44. 27. 64. Functional Hilbert spaces 37. Range inclusion and factorization 60. Invertible transformations 53. 26. Products in H 2 35. Preservation of dimension 57. Continuity of extension 40. Projections of equal rank 58. Conjugation in functional Hilbert spaces 39. MULTIPLICATION OPERATORS 61. BOUNDEDNESS AND INVERTIBILITY 50. Bounded approximation 42. Diminishable complements 54. 62. ANALYTIC FUNCTIONS 31. 47. 48. Multiplicativity of extension 43.

Closure of approximate point spectrum 78. 85. 84. Weighted sequence spaces 96. Similarity of weighted shifts 91. OPERATOR MATRICES 70. Spectrum of a multiplication 68. Analyticity of resolvents 87. Analytic quasinilpotents 98. 8-1. Spectral radius 89. Non-emptiness of spectra 88.CONTENTS 65. Continuity of inversion 101. Operator determinants with a finite entry 9. Norm and spectral radius of a weighted shift 92. Residual spectrum of a normal operator Spectral parts of a diagonal operator Spectral parts of a multiplication Unilateral shift Structure of the set of eigenvectors Bilateral shift Spectrum of a functional multiplication 11. EXAMPLES OF SPECTRA 79. 83. Boundary of spectrum 10. Multipliers of functional Hilbert spaces 8. Roundedness of multipliers 66. 80. Spectrum of a product 77. Spectral mapping theorem 75. Roundedness of multiplications 67. SPECTRAL RADIUS 86. Operator determinants 72. Weighted shifts 90. NORM TOPOLOGY 99. Spectra and conjugation 74. Commutative operator determinants 71. 82. PROPERTIES OF SPECTRA 73. Power norms 93. One-point spectrum 97. Interior of conjugate class xiii . Metric space of operators 100. Multiplications on functional Hilbert spaces 69. Spectrum of a direct sum 12. Approximate point spectrum of a weighted shift 95. Eigenvalues of weighted shifts 94. Similarity and spectrum 76.

Continuity of adjoint 111. Maximal partial isometries 129. Normal continuity of spectrum 106. 121. 139. 141. Continuity of spectral radius 105. Partial isometries 128. 137. Unitary equivalence for partial isometries 133. Strong normal continuity of adjoint Strong bounded continuity of multiplication Strong operator versus weak vector convergence Strong semicontinuity of spectrum Increasing sequences of Hermitian operators Square roots Infimum of two projections 15. Continuity of norm 109. Separate continuity of multiplication 113. Continuity of multiplication 112. 122. STRONG OPERATOR TOPOLOGY 116. Sequential continuity of multiplication 114. Semicontinuity of operator norm 110. OPERATOR TOPOLOGIES 107. Topologies for operators 108. 118. Weak sequential continuity of squaring 115. PARTIAL ISOMETRIES 123.CONTENTS 102. 136. Components of the space of partial isometries 132. and nullity 131. 138. 117. Polynomially diagonal operators 126. Closure and connectedness of partial isometries 130. Continuity of spectrum 103. Polar decomposition Maximal polar representation Extreme points Quasinormal operators Mixed Schwarz inequality Quasinormal weighted shifts Density of invertible operators Connectedness of invertible operators xiv . Quasinilpotent perturbations of spectra 13. Weak convergence of projections 14. 140. POLAR DECOMPOSITION 134. Spectral mapping theorem for normal operators 124. 135. Rank. Semicontinuity of spectrum 104. 120. Decreasing squares 125. co-rank. Spectrum of a partial isometry 16. Continuity of the functional calculus 127. 119.

185. UNILATERAL SHIFT 142. 184. Compact operators 171. 143. 177. 155. 153. 152. Normal compact operators Hilbert-Schmidt operators Compact versus Hilbert-Schmidt Limits of operators of finite rank Ideals of operators Compactness on bases Square root of a compact operator Fredholm alternative Range of a compact operator Atkinson's theorem Weyl's theorem Perturbed spectrum Shift modulo compact operators Distance from shift to compact operators XV . 163. PROPERTIES OF COMPACTNESS 169. CYCLIC VECTORS 160. 150. Diagonal compact operators 173. 164. 146. 157. Reducing subspaces of normal operators Products of symmetries Unilateral shift versus normal operators Square root of shift Commutant of the bilateral shift Commutant of the unilateral shift Commutant of the unilateral shift as limit Characterization of isometries Distance from shift to unitary operators Square roots of shifts Shifts as universal operators Similarity to parts of shifts Similarity to contractions Wandering subspaces Special invariant subspaces of the shift Invariant subspaces of the shift F. 166. 148.CONTENTS 17. 154. 176. 167. 180. 159. 161. 162. 156. 175. 168. 144. 178. Mixed continuity 170. 145. 179. 151. 174. 183. Cyclic vectors Density of cyclic operators Density of non-cyclic operators Cyclicity of a direct sum Cyclic vectors of adjoints Cyclic vectors of a position operator Totality of cyclic vectors Cyclic operators and matrices Dense orbits 19. 172. 149. i47. Riesz theorem Reducible weighted shifts 18. and M. 158. 165. 181. 182.

Quasinormal invariants 197. Hyponormal idempotents 209. Spectrum and numerical range 215. Filling in holes 202. Algebras of normal operators 194. Normaloid. Numerical range of a compact operator 214. and spectraloid operators 220. Closure of numerical range 213. Similarity of subnormal operators 200. Hyponormal operators 204. NUMERICAL RANGE 210. Spectral measure of the unit disc 195. Unitary dilations Images of subspaces Weak closures and dilations Strong closures and extensions Strong limits of hyponormal operators Unitary power dilations xvi . Unbounded Volterra kernels 188. Volterra integration operator 189. 227. Norm 1. 224. Normal and subnormal partial isometries 205. EXAMPLES OF COMPACTNESS 186. Norm powers and power norms 206. Bounded Volterra kernels 187. Higher-dimensional numerical range 212. Power inequality 23. Numerical radius 219. Skew-symmetric Volterra operator 190. 226. 225. Minimal normal extensions 198. Compact hyponormal operators 207. Powers of hyponormal operators 22. Donoghue lattice 21. UNITARY DILATIONS 222. SUBNORMAL OPERA TORS 192. 223. Toeplitz-Hausdorff theorem 211. compact imaginary part 208. Hyponormal. Putnam-Fuglede theorem 193.CONTENTS 20. convexoid. Quasinilpotence and numerical range 216. spectrum {1} 191. Spectral inclusion theorem 201. Extensions of finite co-dimension 203. Subnormal operators 196. Normality and numerical range 217. Polynomials in the shift 199. Subnormality and numerical range 218. Continuity of numerical range 221.

von Neumann's inequality 24. Laurent operators and matrices 242. Eigenvalues of Hermitian Toeplitz operators 249. Kleinecke-Shirokov theorem 233. Spectral inclusion theorem for Toeplitz operators 246. Projections as self-commutators 238. Compact Toeplitz products 245. Analytic Toeplitz operators 248. TOEPLITZ OPERATORS 241. Spectrum of a Hermitian Toeplitz operator REFERENCES LIST OF SYMBOLS INDEX xvii . Commutator subgroup 25. Positive self-commutators 237. Direct sums as commutators 236. Multiplicative commutators 239. Toeplitz operators and matrices 243.CONTENTS 228. Unitary multiplicative commutators 240. Commutators 231. Continuous Toeplitz products 247. Limits of commutators 232. Operators with large kernels 235. Ergodic theorem 229. COMMUTATORS 230. Distance from a commutator to the identity 234. Zero-divisors 250. Toeplitz products 244.

.

PROBLEMS .

.

CHAPTER 1 Vectors 1. Most people who say they study the theory of Hilbert spaces in fact study operator theory.(f) = qJ(j. Since "sesqui" means "one and a half" in Latin. and. f).) More honestly put. cf. The reason is that the algebra and geometry of vectors. quadratic forms. Limits of quadratic forms. Such an existential definition makes it awkward to answer even the simplest algebraic questions. Recall to begin with that a bilinear functional on a complex vector space H is sometimes defined as a complex-valued function on the Cartesian product of H with itself that is linear in its first argument and conjugate linear in the second. in this context and in other more general ones. f). use "semilinear" instead of "conjugate linear". [50. such as whether the sum of two quadratic forms is a quadratic form (yes). Some mathematicians. linear functionals. (The symbol ijJ is used there instead of qJ. incidentally. Is the limit of a sequence of quadratic forms a quadratic form? 3 . but the operators on it. a quadratic form is a function l/1 for which there exists a sesquilinear form qJ such that l/f(f) = qJ(j. "form" instead of "functional". A quadratic form is defined in [50] as a function qJ. it has been suggested that a bilinear functional is more accurately described as a sesquilinear form. perhaps some are both. 12].. The objects of chief interest in the study of a 1 Hilbert space are not the vectors in the space. subspaces.associated with a sesquilinear form qJ via the equation qJ. and the like are easier than operator theory and are pretty well worked out. p. Some of these easy and known things are useful and some are amusing. and whether the product of two quadratic forms is a quadratic form (no). Problem 1.

f)· q. gW and 2 (f. by definition.(f. !) 2 • This proof of the Schwarz inequality does not work for sesquilinear forms unless they are strictly positive. One proof of the Schwarz inequality consists of the verification of one line. A subset of a real vector space is convex if. In a real vector space (and hence. What are the facts? Is strict positiveness necessary? Problem 2. and therefore. Representation of linear functionals. The Riesz representation theorem says that to each bounded linear functional ~ on a Hilbert space H there corresponds a vector g in H such that ~(f) = (f. more powerful. This one line proves also that if the inequality degenerates to an equality. Find a coordinatized proof of the Riesz representation theorem. 32]) are so elegant that they conceal what is really going on. it is mandatory that the proof be such. where 0 ~ t ~ 1.(f. Strict convexity. Schwarz inequality. g) are equal to I a 1 ([. that the role of convexity is sometimes not so clearly visible in it as in other vector spaces. in particular. symmetric. g) for all f and g? 3 3. structure. by definition. say g = af. g) 12 ~ q. the set of all vectors f with II !II < 1. An easy example of a convex set in a Hilbert space is the unit ball. for each pair of vectors that it contains. in a complex vector space) the segment joining two distinct vectors f and g is. Problem 3. (The adjective "closed" can be used to distinguish the unit ball from its 4 . the set of all vectors of the form tf + (1 . f) · (g.(g. Convexity plays an increasingly important role in modern vector space theory. then f and g are linearly dependent. the set of all vectors f with 11!11 ~ 1. namely: : 2iillgll 2f . according to current mathematical ethics. gW perspicuous in the form given.ICJ. 4 4. then one of them is a scalar multiple of the other. g)gji 2 • 1 1 . Another example is the open unit ball. The trouble is that most coordinate-free proofs (such as the one in [50.t)g. g) for all f. It might perhaps be more elegant to multiply through by llgll 2 so that the result should hold for g = 0 also. If q> is a positive. Hilbert space is so rich in other. and then both I(f. but the identity seems to be more = llfll 2 ·llgll 2 .CJ. sesquilinear form. The statement is "invariant" or "coordinate-free". is it necessarily true that I q. which is. p. The converse is trivial: iff and g are linearly dependent.PROBLEMS 2 2. it contains all the vectors of the segment joining them.

determined by the intervals [a. a simple continuous curve with the property that every two non-overlapping chords of it are orthogonal. Continuous curves. Unlike convexity. i. the concept of strict convexity is not purely algebraic. An infinite-dimensional Hilbert space is even roomier than it looks. it is very easy. 5 5 . . b] is the vector f(b) . . but in order for it to make sense the space must have a topology. and if 0 < t < 1 (the emphasis is that t # 0 and t # 1). this convex set has just four extreme points (namely 1.VECTORS open version. The open unit disc in the complex plane has no extreme points. in the complex plane. If a curve could do so for every pair of non-overlapping chords. i.i). then the curve makes a right-angle turn during the passage between their farthest end-points.f(a). The unit ball of every Hilbert space is strictly convex. The problem is stated here to call attention to a circle of ideas and to prepare the ground for some later work.1. The chord of the curve f determined by the parameter interval [a. Problem 5. and . A continuous curve in a Hilbert space H is a continuous function from the closed unit interval into H. then it would seem to be making a sudden right-angle turn at each point.1. and hence. The set of all those complex numbers z for which IRe z I + IIm z I :. a striking way to demonstrate its spaciousness is to study continuous curves in it. in particular. d].) These examples are of geometric interest even in the extreme case of a (complex) Hilbert space of dimension 1. It makes sense in many spaces other than Hilbert spaces. and . then it is called an extreme point of the set. No great intrinsic interest is claimed for it. respectively.:. The expression "boundary point" is used here in its ordinary topological sense. 5. d] have at most an end-point in common. Problem 4. A closed convex set in a Hilbert space is called strictly convex if all its boundary points are extreme points. If a point of a convex set does not belong to the interior of any segment in the set. If two non-overlapping chords are orthogonal. The extreme points of the closed unit disc in the complex plane are just the points on its perimeter (the unit circle). 1 is convex (it consists of the interior and boundary of the square whose vertices are 1.i). b] and [c.t)g is a point of the segment joining two distinct vectors f and g. The closed unit disc in the complex plane is strictly convex. preferably one that is properly related to the linear structure. the curve is simple if the function is one-to-one. Construct. are non-overlapping if the intervals [a. it could not have a tangent at any point. If h = tf + (1 . Two chords. they reduce then to the closed and the open unit disc. for every infinite-dimensional Hilbert space. then h is called an interior point of that segment. but is in fact used only when unusual emphasis is necessary. b] and [c.

There are two other useful ways in which one crinkled arc can be changed into another. Why is that? Surely there must be a reason. G. since H has. It is true. it has an orthogonal dimension. Linear dimension. A unified way to approach the two concepts is first to prove that all bases of H have the same cardinal number. see [81] and [146]. The concept of dimension can mean two different things for a Hilbert space H. Does there exist a Hilbert space whose linear dimension is ~o? 8 8. but for present purposes irrelevant. that every vector is a finite linear combination of the vectors in any Hamel basis. and. for different concrete representations. it turns out.) An orthonormal basis for H is a maximal orthonormal subset of H. Total sets. (Recall that an infinite set is called linearly independent if each finite subset of it is linearly independent. a simple continuous curve with every two non-overlapping chords orthogonal-cf. Since H is a vector space. an inner product structure. Problem 6. In what follows a crinkled arc will be called normalized in case all three of these normalizations have been applied to it. 7 7. A subset of a Hilbert space is total if its span is the entire space. and. Any two normalized crinkled arcs are unitarily equivalent to reparametrizations of one another. more generally. and then to define the dimension of H as the common cardinal number of all bases. to topological vector 6 . Solution 5) is psychologically unique. A Hamel basis for H (also called a linear basis) is a maximal linearly independent subset of H. The other is unitary equivalence: fix a unitary operator U on H and replace f by Uf. the difference between the two concepts is in the definition of basis. Miracle: that's all. Remedy: normalize so that the span of the range off is H. there is a good one. (3) Span: fix a Hilbert space H 0 and replace H (the range space of f) by H EB H 0 . There are three trivial senses in which crinkled arcs are not unique. discovered by G. everyone who tries it seems to come up with the same answer. it has a linear dimension. (Generalizations to Banach spaces. One is reparametrization: fix an increasing homeomorphism cp of [0. It is an interesting empirical fact that the example of a "crinkled arc" (that is. Uniqueness of crinkled arcs. (The analogues of the finite expansions appropriate to the linear theory are the Fourier expansions always used in Hilbert space.) Problem 7. (1) Translation: fix a vector fo and replace the arc f by f + f 0 • Remedy: normalize so that f(O) = 0. The reason is the existence of a pleasant and strong uniqueness theorem. (2) Scale: fix a positive number a and replace the arc f by af.PROBLEMS 6 6. l] onto itself and replace f by f o cp. Johnson [80]. in addition. Remedy: normalize so that llf(l)ll = 1.

···). f 2 . ••• } be a total set. fn}· Another example of an "infinitely total" set. It is trivial to construct an infinite sequence such that its terms form a total set and such that this remains true no matter how many terms are omitted from the beginning.j~. choose a non-zero vector f 1 in E 1 . in £ 2 (0. 2. and choose fn+ 1 so that it is in En+ 1 but not in VUt. consider a countable base {E 1 . ··-). is the set of all powers J. The generalization (unpublished) was discovered and proved by R. To get an inductive construction started. There exists a total set in a Hilbert space that continues to be total when any one element is omitted but ceases to be total when any two elements are omitted.fo. not sequences. 1) (i. the statement is the one in Problem 8. Infinite Vandermondes. Infinitely total sets.. The Hilbert space P consists. note that En+ 1 is not included in V {!1 .See Solution 11. n. ~b ~ 2 .)A sharper way to formulate what is wanted is to ask whether there exists a linearly independent total set that remains total after the omission of each finite subset. 9 10. · · ·. ···)of complex numbers such that 10 7 . one way to see it is to construct a linearly independent dense set. For n = I. Can a set be such that the removal of every finite subset always leaves it total? (Note: the question is about sets..e. there exists a total set in Hilbert space that continues to be total when any n of its elements are omitted but ceases to be total when any n + 1 elements are omitted. · · ·. Problem 9. for instance. F. the open balls with centers at a countable dense set and rational radii).) Can a set be so imperturbably total that the removal of any single element always leaves it total? The answer is obviously yes: any dense set is an example. This is not surprising but some of the behavior of total sets is..VECTORS spaces are immediate.finite subset. Indeed: let {f0 . given jj in Ei. but needing more analytic machinery. j = 1. and form the sequence <fo. the span is nowhere dense). by definition. fix)= x".} (because.f3 .J1 . fn} is linearly independent.f2 .fo. then it has at least one infinite subset whose omission leaves it total also. 9. 1.. Wiser in 1974.f1. J.!o. The answer is yes. · · ·.g. · · ·. To do that. in some respects simpler. E 2 . f 1. n = 0. If a set remains total after the omission of each . so that Ut.f1. The statement of Problem 8 has a natural broad generalization: for each non-negative integer n. For the induction step. · · ·} for the norm topology of a separable infinite-dimensional Hilbert space (e. Problem 8. The result is obvious for n = 0: any orthonormal basis is an example. of all infinite sequences (~ 0 .

and specialize (to finite-dimensional spaces). • • ·} is an orthonormal set in H such that 00 L llej j= 1 /. Approximate bases. This is a hard one.*· n=O Jal < 1. ···}is an orthonormal basis for a Hilbert space H. <IJo. see [114. f 2 .(. 86].11 2 < oo. 2. Does there exist an infinite total set such that every infinite subset ofit is total? 12 12. 3. 8 . IJ1.PROBLEMS The vector operations are coordinatewise and the inner product is defined by 00 (<~o' ~1. Problem 12. T-total sets. then the vectors . For a related exposition. There are many problems of this type. e 3 . . and detailed references. If {e 1 . IJz. · · -)) = Problem 10. and if {!1 . Problem 11.. ' 2 determine the span of the set ofall fk's in 1 • Generalize (to other collections of vectors). ~z. · · -)./fO < L ~niJ. f 3 . No.. and if k = 1. span H (and hence form an orthonormal basis for H). e 2 . 11 11. The version above is discussed in [14]. the first one is apparently due to Paley and Wiener.

just put N = 0. This means that the collection is partially ordered (by inclusion). M. then M + N need not be closed (see [50. Here is the conclusion under another very strong but frequently usable additional assumption. p. if no additional assumptions are made. 14. Problem 13. Orthogonality may be too strong an assumption. 28]. and if N is a subspace (a closed linear manifold) in H.finite-dimensional linear manifold in a Hilbert space H. Vector sums. as written above. It is known that something is necessary. and N.CHAPTER 2 Spaces 13. There is a weakening of this distributivity condition. If M is a . and Problems 52-55 below). called modularity. If M and N are orthogonal subspaces of a Hilbert space. a lattice is called modular if the distributive law. then M + N is closed (and therefore M + N = M v N). holds at 9 . but it is sufficient to ensure the conclusion. Lattice of subspaces. and that any two elements M and N of it have a least upper bound or supremum (namely the span M v N) and a greatest lower bound or infimum (namely the intersection M 1"'1 N). then the vector sum M + N is necessarily closed (and is therefore equal to the span M v N). A lattice is called distributive if 14 (in the notation appropriate to subspaces) L n (M v N) = (L n M) v (L 1"'1 N) identically in L. 13 The result has the corollary (which it is also easy to prove directly) that every finite-dimensional linear manifold is closed. The collection of all subspaces of a Hilbert space is a lattice.

PROBLEMS least when N c: L. and the identity (with the proviso N c: L still in force). Vector sums and the modular law. Two possible kinds of misbehavior for su bspaces are connected with each other. The answers either are a simple yes or no. Problem 17. Finite-dimensional Euclidean spaces have. the vector sum M + N is closed if and only if the modular equation Ln(M v N) =(LnM) v N is true whenever N c: L. Thus. For which cardinal numbers m is the lattice of sub spaces of a Hilbert space of dimension m modular? distributive? 15 15. Problem 15. if one of them is ruled out. that it is finite). Local compactness and dimension.. Many global topological questions are easy to answer for Hilbert space. Since a Hilbert space is geometrically indistinguishable from any other Hilbert space of the same dimension. if and only if it is finite- 17 17. The same sort of problem could be posed backwards: given some information about the dimension of a Hilbert space (e. Problem 14. For subspaces M and N of a Hilbert space. but a Hilbert space is compact if and only if it is the trivial space with dimension 0. L n N becomes Ln(MvN)=(LnM)vN N. it is clear that the modularity or distributivity of its lattice of subspaces can depend on its dimension only. In that case.g. it might be useful to generalize that too to infinite dimensions. Such problems sometimes have useful and elegant solutions. Various attempts have been made to do so (see [92] and [132]). every Hilbert space is connected. of course. Infinite-dimensional Hilbert spaces are properly regarded as the most successful infinite-dimensional generalizations of finite-dimensional Euclidean spaces. or depend on the dimension. for instance. A Hilbert space is locally compact dimensional. A Hilbert space His separable if and only if dim H ~ ~0 • 18 18. The un10 . Problem 16. Measure in Hilbert space. then the other one cannot happen either. Separability and dimension. 16 16. a measure. find topological properties that distinguish such a space from Hilbert spaces of all other dimensions. in addition to their algebraic and topological structure.

then the following problem indicates that the unsophisticated approach is doomed to fail. oo for all Borel sets M. so that 0 . Problem 18. 11 . the word "measure" is used to describe a set function satisfying just these conditions. JJ.) If. defined on (at least) the collection of all Borel sets (the cr-field generated by the open sets). For each measure in an infinite-dimensional Hilbert space.{j + M) = JJ.::..) In order that !J. the measure of every non-empty ball is infinite.(M) . (Warning: the parenthetical definition of Borel sets in the preceding sentence is not the same as the one in [61]..(M) for every vector f and for every Borel set M. for now... (The second condition means that JJ. it makes sense to require that every non-empty open set have positive measure and that measure be invariant under translation. be suitably related to the other structure of the space.::.SPACES sophisticated approach is to seek a countably additive set function !J.

it is a topological space.fo. go) I < e}. i = 1. many technical questions automatically demand attention. a function may be described as weakly continuous. k}. for instance. sets of the form {f: llf. The use of a topological word without a modifier always refers to the strong topology. gk are vectors.e. Whenever a set is endowed with a topology. i. · · ·. this convention has already been observed in the preceding problems. Facts about these topologies are described by the grammatically appropriate use of "weak" and "strong". g 1 . where fo (the center) is a vector and e (the radius) is a positive number. and e is a positive number. Another topology. A base for the strong topology is the collection of open balls. • · ·. A Hilbert space is a metric space. called the weak topology. as such. It follows that a base for the weak topology is the collection of all sets of the form {f: l(f. A subbase (not a base) for the weak topology is the collection of all sets of the form {f: l(f. or a sequence as strongly convergent. The metric topology (or norm topology) of a Hilbert space is often called the strong topology.)l < e. plays an important role in the theory of Hilbert spaces. where k is a positive integer. and. Weak closure of subspaces.CHAPTER 3 Weak Topology 19 19. the meanings of such phrases should be obvious. Thus.fo. (Which separation axioms does the space satisfy? Is it compact? Is it connected?) If a large class of sets is in sight (for 12 .foil < e}. g. f 0 ..

(xn) ~ 21 q. g) is separately weakly continuous in each of its two variables. (A sequence. in fact. Nevertheless every subspace of a Hilbert space (i. be said that the possible discontinuity of the norm is the only difference between weak convergence and strong convergence: a weakly convergent sequence (or net) on which the norm behaves itself is automatically strongly convergent. en) = 1 for all n. but (en.llfll. Semicontinuity of norm. the function f ~--> (f.f (weak) and 20 llfnll . Problem 20.) This.e. similarly q. or a net. (Is a weakly compact set strongly closed?) Most such questions. g) to their inner product (f. g) is weakly continuous. implies that.. the function g ~--> (f. A counterexample has already been seen. so to speak: norm fails to be upper semicontinuous. It is natural to ask whether the mapping is weakly continuous jointly in its two variables. but it is easy to see that the answer is no. g) is weakly continuous. (Which ones are locally compact? Which ones are separable?) If the set (or sets) already had some structure. Every weakly closed set is strongly closed. then the relations of the topologies to one another must be clarified. or an important application.. then fn. though natural. The questions that do appear justify their appearance by some (perhaps subjective) test. Problem 19. and. such as a surprising answer. unavoidable. g) for each g. for each fixed vector f.. These two assertions between them say that the mapping from ordered pairs <f. Un} is weakly convergent to f if and only if (!. the class of all Hilbert spaces). it was used there for a slightly different purpose. in fact. e 3 . If!. this is practically the definition of the weak topology. 21. a tricky proof. e 2 .(f. together with the (Hermitian) symmetry of the inner product. the connection between the old structure and the new topology should be investigated. finally. (Is the closed unit ball compact? Are inner products continuous?) If. every strongly closed linear manifold) is weakly closed.WEAK TOPOLOGY example. then classification problems arise. If {e 1 . more than one topology is considered.g).0 (weak). For each fixed vector g. The misbehavior of the example that shows the weak discontinuity of norm (Problem 20) is at the top.x (sequence or net). This example shows at the same time that the norm is not weakly continuous. then en. It could. Weak continuity of norm and inner product. for that reason most such questions do not appear below. is lower semicontinuous if n 13 . but the converse is not true.f (strong). 20. in Solution 19. are not likely to be inspiring. Definition: a real-valued function on a topological space is upper semicontinuous if limsup q.(x) whenever xn. • • ·} is an orthonormal sequence.

.. It is very important. limsupn t:p(xn). must H be separable? 27 27. so that if t:p is both lower and upper semicontinuous.. Is every weakly separable Hilbert space separable? 23 23. but even compact sets are better if they are metric. Explicitly: if fn-+ f (weak). of the set of all unit vectors)? If a set is weakly dense in a Hilbert space. it follows that if a Hilbert space is separable (that is. Weak metrizability and separability. What is the weak closure of the unit sphere (i.. Problem 23.PROBLEMS whenever xn -+ x. Is the weak topology of the unit ball in a separable Hilbert space metrizable? 25 25. llfnll + e whenever n ~ n0 • 22 22. then llfll .i. it is natural to ask if it is weakly metrizable also. Weak separability. then it is weakly separable.e.) Misbehavior at the bottom cannot occur. Weak closure of the unit sphere. Problem 21. does it follow that its intersection with the unit ball is weakly dense in the unit ball? 26 26. The celebrated "principle of uniform boundedness" (true for all Banach spaces) is the assertion that a pointwise bounded 14 . What about the converse? Problem 22. which is a characteristic property of continuity. The result is sometimes known as the Tychonoff-Alaoglu theorem. Uniform boundedness. It is as hard as it is important..:. Problem 24. Norm is weakly lower semicontinuous. Since the strong closure of every set is included in its weak closure (see Solution 19). The closed unit ball in a Hilbert space is weakly compact. Weak metrizability of the unit ball. liminfn llfnll· Equivalently: for every e > 0. Once the unit ball is known to be weakly compact.:. 24 24. strongly separable). there exists an n0 such that llfll . Problem 26. then liminf and limsup are forced to be equal. If the weak topology of the unit ball in a Hilbert space His metrizable. Weak compactness of the unit ball.. (Here is how to remember which way the inequalities must point: always liminfn t:p(xn) . Problem 25. Compactness is good.

(f) such that I(f. the map g f-+ (f. g) sends the g's in a weakly compact set onto a compact and therefore bounded set of numbers. Reason: for each f. it means that for each fin H there exists a positive constant cx. The weak topology ofan infinite-dimensional Hit bert space is not metrizable. For unbounded sets it is not. cx. then 15 . so that a weakly compact set is weakly bounded.2. show that for bounded sets the weak topology is well behaved. 29 2 1 and cx. (In this context a proof is "elementary" if it does not use the Baire category theorem. Some of the preceding results. g) I ~ cx. p. The proof is completely elementary: since convergent sequences of numbers are bounded. /3 3 .(f) for all g in T. g)l ~ /311!11 for all fin Hand all gin T. [3 2. The assumption of pointwise boundedness for a subset T of H could also be called weak boundedness. it follows that a weakly convergent sequence of vectors is weakly bounded. Problem 27.WEAK TOPOLOGY collection of bounded linear functionals is bounded. Linear functionals on /2 • If (cx. 29. Nothing like this is true for nets. The usual proof of the general principle is a mildly involved category argument. One easy generalization of the sequence result that is available is that every weakly compact set is bounded. 49]. · • ·) E (/3 1. of course. 28. Problem 28. 28 notably the weak compactness of the unit ball and the principle of uniform boundedness. The desired conclusion means that there exists a positive constant f3 such that l(f. The shortest proof of this is tricky.3. 2 · ·-) E 1 .l. A standard reference for a general treatment of the principle of uniform boundedness is [39. Find an elementary proof of the principle of uniform bounded ness for Hilbert space. this conclusion is equivalent to llgll ~ f3 for all g in T. Weak metrizability of Hilbert space. It is clear that every bounded subset of a Hilbert space is weakly bounded.) A frequently used corollary of the principle of uniform boundedness is the assertion that a weakly convergent sequence must be bounded. The assumption and the conclusion can be expressed in the terminology appropriate to a Hilbert space H as follows. The principle of uniform boundedness (for vectors in a Hilbert space) is the converse: every weakly bounded set is bounded.

To say of a Hilbert space. it says that 1 sequences are the 2 1 only ones whose product with every 1 sequence is in 1 • Problem29. A sequence {gn} of vectors in a Hilbert space is a weak Cauchy sequence if (surely this definition is guessable) the numerical sequence {(f. If the conclusion is known to hold for sequences only. (b) Which Hilbert spaces are sequentially weakly complete? 16 . that it is weakly complete means that every weak Cauchy net has a weak limit (in the set under consideration). Weak completeness. IfL:nlanPnl < oo whenever 2 Ln1anl 2 < 00. or a subset of one. 30.PROBLEMS 2 The following assertion is a kind of converse. gn)} is a Cauchy sequence for each fin the space. (a) No infinite-dimensional Hilbert space is weakly complete. then L 1Pn1 n 30 < 00. the space is called sequentially weakly complete. Problem 30. Weak Cauchy nets are defined exactly the same way: just replace "sequence" by "net" throughout.

f----1)/n · z" for lzl < 1 and n = 0. it is necessary to enrich it by the addition of some external structure.CHAPTER 4 Analytic Functions 31. 31 Problem 31. then the en's form an orthonormal basis or A 2 . I. The typical way to construct these examples is to consider a region D ("region" means a non-empty open connected subset of the complex plane). with Taylor (n + I)/n(f. then rxn = 33. Real functions in H 2 • Except for size (dimension) one Hilbert space is very like another. the spaces A 2 (D) are of interest because of 17 32 33 . Analytic Hilbert spaces. g) = Lf(z)g(z)* dJ1(z). series L:'=o rxnz". 1. · · ·. It is also an innerproduct space with respect to the inner product defined by (f. the corresponding function space will be denoted simply by A2 • No matter what Dis. Basis for A2 • Problem 32. Analytic functions enter Hilbert space theory in several ways. To make a Hilbert space more interesting than its neighbors. Thus.. en) for n = 0. one of their roles is to provide illuminating examples. If fE A 2 . D = { z: Iz I < 1}. If en(z) = j"(~ . 2. or does it have to be completed before it becomes one? 32. let 11 be planar Lebesgue measure in D. · · ·. Is the space A2 (D) of square-integrable analytic functions on a region D a Hilbert space. for instance. and let A2 (D) be the set of all complex-valued functions that are analytic throughout D and square-integrable with respect to 11· The most important special case is the one in which D is the open unit disc. the set A2 (D) is a vector space with respect to pointwise addition and scalar multiplication. 2.

A related space. as long as the qualification "almost everywhere" is kept in mind at all times. ± 1. 18 . equivalently. Another important Hilbert space. Let C be the unit circle (that means circumference) in the complex plane. then. and this custom was followed in the preceding paragraph. and.g.. endowed with some structure not usually found in a Hilbert space. The custom is not likely to lead its user astray.l. it will be denoted by H 2 *. Similarly H 1 is the set of all those elements f of L 1 for which these same equations hold. they appear in questions (which must be asked and answered) about the limiting behavior of the functions on the boundary.(C) = 2n). Fourier expansions with respect to the orthonormal basis {en: n = 0..1. known as H 2 (H is for Hardy this time). is defined as follows.( C) = I (instead of J. but they are pushed elsewhere. equivalently H 2 is the orthogonal complement in L 2 of {e _ ~> e _ 2 . normalized so that J. C = {z: lzl = 1}.l. H 2 . Problem 33. If en(z) = z" for lzl = 1 (n = 0.. ± 1. The following statement is evidence in that direction. the functions in H 2 are anxious to behave like analytic functions. 0. H 00 is the set of all those functions f in L 00 for which Jfen* dJ. = 0 (n = -1. H 2 .1.. when asserted. e _ 3 . What gives Hi.. · · ·} are formally similar to the Laurent expansions that occur in analytic function theory. In that approach (see Problem 35) the almost everywhere difficulties are still present. the Weierstrass theorem on approximation by polynomials) that the en's form an orthonormal basis for L 2 • (Finite linear combinations of the en's are called trigonometric polynomials. ±2. Iff is a real function in H 2 . -3. Thus "bounded" means "essentially bounded". ± 2. the elements of H 2 * are called co-analytic. A subset ofH 2 (a linear manifold but not a subspace) of considerable technical significance is the set H 00 of bounded functions in H 2 . by definition. be Lebesgue measure (the extension of arc length) on the Borel sets of C.PROBLEMS the analytic properties of their elements.. is the span of the en's with n . as holding almost everywhere.l. Independently of the approach used to study them. and H 00 their special flavor is the structure of the semigroup of non-negative integers within the additive group of all integers. and let J. · · ·). the subspace of L 2 spanned by the en's with n .. and 00 H as functions. all statements such as "f = 0" or "f is real" or " I f I = I " are to be interpreted. -2. then f is a constant.. Some authors define the Hardy spaces so as to make them honest function spaces (consisting of functions analytic on the unit disc). it is an easy consequence of standard approximation theorems (e. by elementary calculus. · · ·). the functions en form an orthonormal set in L 2 (J.) The space H 2 is.). similarly. playing a role dual to that of H 2 . 0. It is customary to speak of the elements of spaces such as H 1 . The analogy motivates calling the functions in H 2 the analytic elements of L 2 . • • ·}.

(z) = L:':o f3nzn. Analytic characterization of H 2 • If f E H 2 . The mapping f ~ j (obviously linear) establishes a one-to-one correspondence between H 2 and the set H2 of those functions analytic in D whose series of Taylor coefficients is sq uare-summable. p. Iff E H 2 . 35 I:'= Ln Problem 35. they will be called functional Hilbert spaces. and if <p. If a measure space has a non-empty set of measure zero (and this is usually the case). 35. Many of the popular examples of Hilbert spaces are called function spaces. then L:'=o ian 12 < oo. it all comes to the same 2 thing.) The direct statement is more useful in Hilbert space theory than the converse. The following one is an easily accessible sample. t/1) is defined to be I:':o anf3n*· In view of the one-toone correspondence f ~ J between H 2 and H2 . The product of two functions in H 2 is in H 1 . and therefore the radius of convergence of the power series 0 a/zn is greater than or equal to 1. but they are not. its image j in H may be spoken of as the extension off into the interior (cf. and there is no natural way to identify such equivalence classes with representative elements. Solution 40). It follows from the usual expression for the radius of convergence in terms of the coefficients that the power series 00=o anzn defines an analytic function j in the open unit disc D. Products in H 2 • The deepest statements about the Hardy spaces have to do with their multiplicative structure. for them.nzn. with Fourier expansion f = L:'=o an en. then the L 2 space over it consists not of functions. 34 A kind of converse of this statement is true: it says that every function in H 1 is the product of two functions in H 2 • (See [75. A functional 19 36 . however. <p(z) = L:'=o tY. and the techniques used in the proof of the direct statement are nearer to the ones appropriate to this book. If <p is an analytic function in the open unit disc. with bounded concentric L 2 norms. equivalently. the series are bounded. L:'=o lanl 2 converges if and only if the norms ll<fJrll 2 Many authors define H 2 to be H . with <p(z) = L:':o rt. then the inner product (<p. 52]. If <p and t/1 are two such functions. H 2 consists of analytic functions in the unit disc with square-summable Taylor series. that is. Problem 34. then <fJr E H 2 for each r. Functional Hilbert spaces. this concept makes sense for elements of H 00 also.(z) = <p(rz) for 0 < r < 1 and lzl = 1. but of equivalence classes of functions modulo sets of measure zero.ANALYTIC FUNCTIONS 34. Since H 00 is included in H 2 . the set of all their extensions will be denoted by H00 • 36.nzn and tf. a class of examples of Hilbert spaces whose elements are bona fide functions. There is. or.

· · -). The usual sequence spaces are trivial examples of functional Hilbert spaces (whether the length of the sequences is finite or infinite). g) depends linearly on/. and let fl. Give an example of a Hilbert space offunctions such that the vector operations are pointwise. Given H. the role of X is played by the index set. An early and still useful reference for functional Hilbert spaces is [5]. then (af + [Jg)(x) = rxf(x) + [Jg(x) for each x in X. over X say. the evaluation functionals on H are linear. the Hilbert space structure of His related to X in two ways (the only two natural ways it could be).g)! ~ IIIII · llgll = IIlii · l!gl!. the correspondence is linear. then the linear functional f ~----+ f(y) on His bounded for each yin X. g). Ky) for all f. Problem 36. consider the standard orthonormal basis {el.. what is its kernel function? For a convenient notation to express the answer in. Write. ' en} in en (where e/i) = bij• the :L 20 . is called the kernel function or the reproducing kernel of H. This definition yields a bona fide inner product if and only if A is positive and invertible. Since if(q)l = !([. write X = H. by definition. it follows that fl. but not all the evaluationfunctionals are bounded. such that !f(x)! ~ Yxllfll for all fin H. There is a trivial way of representing every Hilbert space as a functional one..e. The function K on X x X. consequently. g) (whence. since (f. and. If H is a functional Hilbert space. and (2) to each x in X there corresponds a positive constant Yx. is a functional Hilbert space. it follows that fl.. to complex-valued functions f and g on X by {f. in particular. 3. i. the evaluation functionals on H are bounded. There is a natural correspondence f ~----+ l from H to H. y) = Ky(x). Kernel functions. By the Riesz representation theorem the correspondence is one-to-one. If HA is the vector space en with inner product defined by the positive linear transformation A. there exists.PROBLEMS Hilbert space is a Hilbert space H of complex-valued functions on a (nonempty) set X. g)A to be (Af. · · ·. In other words. It is required that (1) if/and g are in H and if a and f3 are scalars.. for each yin X. . The most trivial examples of functional Hilbert spaces are obtained by modifying the standard inner product in en (n = 1. g)= (f. g) = "modify" it. defined by l(g) = (f. is a Hilbert space isomorphism. is a Hilbert space. The correspondence f ~----+ J between H and fl. start with X= {1. 37 37. and define(/. n}. IIlii = 11/11). g) for all gin X. defined by K(x.. 2. More typical examples of functional Hilbert spaces are the spaces A2 and H2 of analytic functions. an element KY of H such that f(y) = (f. i. then HA is a functional Hilbert space. be the set of all those functions f on X (=H) that are bounded conjugate-linear functionals. (l. and define the "standard" inner product of two fWgU)*. consider a linear transformation A on en.e.

and it follows that K(i. If {ei} is an orthonormal basis for afunctional Hilbert space H. j What are the kernel functions of A 2 and ofH 2 ? The kernel functions of A 2 and of Bergman kernel and the Szego kernel. respectively. the function K is the matrix of A. Ki) = (f. does itfollow that llf*ll = llfll? 38 Whenever the answer is yes for all fin the space. the unit disc. AKi) whenever f E HA and j = 1. as the 38. so that Ki = A -lei. Ki)A = (Af. the complex conjugate f* may fail to belong to H. 21 . n. eJ In other words. Call a functional Hilbert space self-conjugate if it is closed under the formation of complex conjugates. then the kernel function K ofH is given by K(x. Note that the Hermitian character of the function K persists in the general case. A more sophisticated example is the set of all square-integrable complex harmonic functions in.1 with respect to the standard basis. Is the postulated connection strong enough to extend to complex conjugation? What does the question mean? Possible interpretation: is conjugation isometric? Problem 38. g*) = (f.) Consequence: AKi = ei. then fU) = (f. alternatively. y) = L e/x)e/y)*. an example is the sequence space P. j) = K/i) = (A -tei. If f is an element of a functional Hilbert space H. H2 are known. in this sense : Problem 37. (The quickest way to describe complex harmonic functions is to say that they are the functions of the form u + iv. the spaces H 2 and H2 yield examples. say. (Since A is positive. to the mean value property. or. ei) = (f. g)* for all f and g. Iff is an element ofa self-conjugate functional Hilbert space. Other classical definitions refer to the solutions of Laplace's equation. · · ·. If the kernel function of HA is K.ANALYTIC FUNCTIONS Kronecker delta). Conjugation in functional Hilbert spaces. then a routine polarization argument shows that (f*. where each of u and v is the real part of some analytic function. it is Hermitian.) The definition of functional Hilbert spaces requires a strong connection between the unitary geometry of the space and the values of the functions the space consists of.

Iff E H 00 .(z) . Equivalently.t . the convergence is uniform on each disc {z: Iz I ~ r}. e0 ) = 0 and such that u + iv E H 2 . That is: if fn . 40 40. then . then f is bounded. Problem 40. Radial limits. Bounded approximation.fmultiplic'ative? 43 43. Multiplicativity of extension. does it follow that . Dirichlet problem. (i. If an element f of H 2 is such that the corresponding analytic function . Problem 41. but also in the sense appropriate to analytic functions.f(z) for Iz I < 1. in fact. alternatively.fis bounded? 42 42. to each real u in L 2 there corresponds a unique fin H 2 such that (f. To each real function u in L 2 there corresponds a unique real . and. 2 Problem 39. .function v in L 2 such that (v.e. Is the mapping f 1-+ . v is the Hilbert transform of u.PROBLEMS 39 39.1 O<r<l. f E H 00 ).f. Continuity of extension. 22 . The extension mapping f 1-+ (from H 2 to H ) is continuous not only in the Hilbert space sense.. The relation between u and vis expressed by saying that they are conjugate functions. e 0 ) is real and such that Ref = u. 41 41.t f in H 2 . Problem 42.fin H2 is bounded. Problem 43.

Suppose that {ei} is an orthonormal basis for a Hilbert space H. but it sometimes helps. and what does is not so useful. matrices reduce qualitative geometric statements to explicit algebraic computations. and operator multiplication corresponds to the matrix product defined by the familiar formula Yii = I IY.k = (eb A*e. the linear operations on matrices are the obvious ones. Many problems about operators on finitedimensional spaces can be solved with the aid of matrices. Not much of matrix theory carries over to infinite-dimensional spaces. column. The correspondence from operators to matrices (induced by a fixed basis) has the usual algebraic properties. diagonal) can nevertheless be used in their familiar senses. If A is an operator on H. e. The index set is arbitrary here. it does not necessarily consist of positive integers. the entries of the matrix that arises this way are given by rxii = (Aei. as usual..k/3kj• k There are several ways of showing that these sums do not run into convergence trouble. here is one. adjoint corresponds to conjugate transpose. then each Aei has a Fourier expansion..). Familiar words (such as row.). the first index indicates rows and the second one columns. The zero matrix and the unit matrix are what they ought to be. Column-finite matrices. 44 Aei = I i rxiie. then the matrix is formed by writing the coefficients in the expansion of Aei as the j column. Since rx. it follows that for each 23 . Note that if.CHAPTER 5 Infinite Matrices 44.

. L:j !aij/ 2 ·llfl/ 2 .PROBLEMS fixed i the family {a. . they are not sufficient. If r:xij ~ 0.// ~ L. f3kj = (Bej. ej))e. This is the first significant way in which infinite matrix theory differs from the finite version: every operator corresponds to a matrix. if A is an operator on a Hilbert space H. Every operator has a column-finite matrix. since. ej)/ 2 ~ L:j la.). Conclusion (via the Schwarz inequality): for fixed i and j the family {r:x. ek). striking but less useful than it looks.j/ 2 ·llfl/ 2 for each i 2 and each f. that is. . to write down in matricial terms the condition that a linear transformation is everywhere defined and bounded. similarly. ·).k} is square-summable. While the algebra of infinite matrices is more or less reasonable.) There are no elegant and usable necessary and sufficient conditions. Li L:j /a. As long as there is a fixed basis in the background. the analysis is not. of course.) A sufficient condition of the same kind is that the family of all entries be squaresummable. L r:xijqj ~ YPi j (i = 0. one operator may be assigned many matrices. but the result is neither elegant nor usable. if. Problem 44. 2. It is perfectly possible.. More precisely. the correspondence from operators to matrices is one-to-one.k/3kJ is (absolutely) summable. Schur test. if P. The following result (due in substance to Schur [129]) is an example.j/ 2 < oo. 1. as soon as the basis is allowed to vary. An enticing game is to choose the basis so as to make the matrix as simple as possible.jPi ~ f3qj i u 24 = 0. (Proof: since !L:j a. the matrix entry (Aej. j = 0.if. Questions about norms and spectra are likely to be recalcitrant. then there exists an orthonormal basis {ej} for H such that . it follows that for each fixed} the family {f3kj} is square-summable. and if f3 andy are positive numbers such that L OC. (Example: the unit matrix. e. · · ·). Here is a sample theorem. 1.) This condition is not necessary. and it is hard to say which ones do. ·). then there exists an operator A such that aij = (Aej. 1. (Lj au(f. Each of the few answers that is known is considered a respectable mathematical accomplishment. it follows that IlL. (Example: the diagonal matrix whose n-th diagonal term is n. It follows from the preceding paragraph that each row and each column of the matrix of each operator is square-summable. 2. . but not every matrix corresponds to an operator. These are necessary conditions on a matrix in order that it arise from an operator. 2.) vanishes for all but finitely many i's. > 0 and qj > 0 (i. Problem 45. e. 45 45.for each j. Reference : [ 141].

Problem 47. can be asked for other functions cp. Exponential Hilbert matrix. Hilbert matrix. consequence: the corresponding operator is positive. . the main assertion of Problem 46 is that the matrix is "bounded" (meaning that it is the matrix of some operator).e. Positivity of the Hilbert matrix. for instance. that a be in F (Problem 29). {f. i. p. Series of vectors. aii = cp(i + j). of course) with IIAII 2 ~ f3y and with matrix <aii) (with respect to a suitable orthonormal basis).j = 0. For an interesting concrete question not covered by either of these examples consider this one: if the functions fn in L 2 (0. · · · . The same question.(x) = xn. and other sharper ones. 47. What is its norm? 48. The classical Hilbert matrix (Problem 46) is also Hermitian. If. n = 1. 2. the norm of the matrix is in fact equal to 46 n ([67. Thus. then the series Ln an fn sometimes converges in Hand sometimes does not. The exponential Hilbert matrix (Problem 47) is Hermitian and its spectrum is positive (Solution 47). the Hilbert matrix (see Problem 46) is the Hankel matrix that corresponds to the function cp defined by cp(x) = 1/(x + 1) (i.INFINITE MATRICES then there exists an operator A (on a separable infinite-dimensional Hilbert space. and if Ln I an 12 < oo. For a related result.. is not quite so easily visible. then a necessary and sufficient condition that Ln lanfnl < oo for every sequence Un} in F is.. If {an} is a sequence of complex numbers and lfn} is a sequence of vectors in a Hilbert space H. Problem 46. His the !dimensional vector space of complex numbers. A matrix whose <i..j) entry is a function of i + j only is called a Hankel matrix. see Problem 173. I) are defined by f.. however. There exists an operator A (on a separable infinitedimensional Hilbert space) with IIAII ~ nandwithmatrix<l!(i + j + 1)) (i. A pleasant function is given by cp(x) = 2-<x+ll.j = 0. 1. then a necessary and sufficient condition for the convergence of 2 Ln anf~ is that the sequence a be in 1 • If. Is the Hilbert matrix positive? 49. The matrix <2-(i+j+ll) 47 is bounded. Problem 48. for another example. 46. again. does it follow that the 48 49 25 . 2. The matrix is named after Hilbert. · · ·. In that case all questions have a simple answer. 1. 226]). ·).} is an orthonormal sequence. for instance. and a pertinent reference. its spectrum. 2.

Problem 49.PROBLEMS series Ln anfn converges in L 2 ? It turns out that a general question along these lines has an elegant and usable answer.. Under what conditions on a sequence {f.} of vectors in a Hilbert space H does the series Ln anfn converge in Hfor every sequence a in / 2 ? 26 .

Boundedness on bases. Roundedness is a useful and natural condition. · · ·}. of course. The obvious mistakes just mentioned are based on the assumption that the converses of these assertions are true. Much of the two-space theory is an easy adaptation of the one-space theory. (b) Is there such an example that annihilates a basis? (c) Is there such an example that is bounded on each basis? (d) Give examples of operators of arbitrarily large norms that are bounded by 1 on a basis. e2 . (a) Give an example of an unbounded linear transformation that is bounded on a basis. If A is an operator on a Hilbert space H with an orthonormal basis {e 1 . e 3 . if A = 0. it is important to know that boundedness is more than just the conjunction of an infinite number of conditions.CHAPTER 6 Roundedness and Invertibility 50.forinstance. and. Uniform boundedness of linear transformations. the norm of A. in symbols IIAII. the inner product is formed in K.if. IIAII ~ 1. one for each element of a basis. If H and K are Hilbert spaces. the inner product (Af. from its mildest algebraic aspects to its most complicated topological ones. 51 27 .thenthenumbers IIAenll arebounded. Sometimes linear transformations between two Hilbert spaces play a role even when the center of the stage is occupied by operators on one Hilbert space. a linear transformation A from H into K is bounded if there exists a positive number a such that IIAfll ~ allfll for allf in H. Given a bounded linear transformation A. but it is a very strong condition on a linear transformation. To avoid certain obvious mistakes. Problem 50. then A en = 0 for all n. is the infimum of all such values of a. (e) Could all the operators in such an example be normal? 50 51. g) makes sense whenever f is in H and g is in K. then II A en II ~ 1 for all n. The condition has a profound effect throughout operator theory.

.£ a(f. g) I . then there exists a unique mapping A.1 must be bounded.1 is linear. it is identically equal to (j.PROBLEMS For fixed g the inner product defines a bounded linear functional of f. In the sense of pure set theory the converse is true: if A maps H one-to-one onto K. The principle of uniform boundedness for linear transformations is the best possible converse. All these algebraic statements are trivialities. g) such that I(Af. this assertion remains true with A and A* interchanged. however. Problem 51. and. It is not obvious. If A is invertible. The algebraic properties of this kind of adjoint can be stated and proved the same way as for the classical kind. the mapping A. it is conceivable that A could be invertible as a set-theoretic mapping but not invertible as an operator. By definition (Af. that there exist a positive number <5 such that I Af I ~ <5 llf I for every fin H. the generalization of the principle of uniform boundedness from linear functionals to linear transformations is somewhat subtler. here the left inner product is formed inK and the right one in H.£ {J(f) for all A in Q. It is clear that every bounded set is strongly bounded and every strongly bounded set is weakly bounded. that the linear transformation A. Every weakly bounded set ofbounded linear transformations is bounded. g) = (f. g) for some g in H. consequently. it is a bounded linear transformation A* from K into H. To guarantee that A -J is bounded it is customary to strengthen the condition that A be one-to-one. The proper strengthening is to require that A be bounded from below. The set Q is strongly bounded if for each f in H there exists a positive constant {J(f) such that IIA/11 . The mapping from g to g is the adjoint of A. (It is trivial to verify that if A is bounded from below. A *g) whenever f E Hand g E K. A bounded linear transformation A from a Hilbert space H to a Hilbert space K is invertible if there exists a bounded linear transformation B (from K into H) such that AB = 1 (=the identity operator on K) and BA = 1 (=the identity operator on H). then A is a one-to-one mapping ofH onto K. An especially important (but no less easily proved) connection between A and A* is that the orthogonal complement of the range of A is equal to the kernel of A*. i. The generalization can be formulated almost exactly the same way as the special case: a pointwise bounded collection of bounded linear transformations is uniformly bounded. Invertible transformations. A set Q of linear transformations (from H into K) is weakly bounded if for each f in H and each g in K there exists a positive constant a(f.e. then the other usual condition (onto) can be weakened: the requirement that the range of A be equal to K can be 28 .1 from K to H such that AA 1 = 1 and A 1 A = 1. then A is indeed one-to-one. since A** = A.) If that strengthened condition is satisfied. 52 52. The assumption of pointwise boundedness can be formulated in a "weak" manner and a "strong" one. g) for all A in Q.

if they are invertible.ROUNDEDNESS AND INVERTIBILITY replaced by the requirement that the range of A be dense in K. and its consequences. A is 2 defined on 1 by A <~ 1 .. 53 29 . it does not contain the sequence <1. it is one of the pleasantest and most useful facts about operator theory that they are the same after all. M + N = H EB R. · · ·). A complement of a subspace M in a Hilbert space H is a subspace N such that M n N = 0 and M v N = H. it is dense in 1 . [50. since. The theorems and the examples seem to indicate that set-theoretic invertibility and opera to rial invertibility are indeed distinct. • · ·) = <~h -!~ 2 . Bn>--+ <J. see Problem 58. and does it follow that g E R? (possibly no). for instance. the set of all vectors of the form <J~ Af). !~ 3 .1 and A* -• is the adjoint of the other. ~ 2 . 110]. Let A be an operator on a Hilbert space H.) IsM+ N closed? This means: if <fn. !. does it follow that f E H? (trivially yes). then A is invertible. Diminishable complements. ~ 3 . g) belong toM + N? The answer is if and only if it has the form <u. IfH and K are Hilbert spaces. (In other words. Problem 52. cf. Observe that the linear transformations A and A* are invertible together. '12. g). and let N be the "graph" of A. where fn E H and 9n E R. 2 Since this range contains all finitely non-zero sequences. '73. then each of A . If. the construction itself takes place in the direct sum H EB H. i. not bounded from below. their ranges would be closed. It is perhaps worth a short digression to discuss the possibility of the range of an operator not being closed. · · ·) with L n 2 1'1nl 2 < oo. 38]). Another example: for fin L 2 (0. i.e. and if A is a bounded linear transformation that maps Hone-to-one onto K. if they were. ···).e. it is not closed. Operators with non-closed ranges can be used to give a very simple example of two subspaces whose vector sum is not closed. the vector sum of two subspaces need not be closed either.then the range of A consists of all vectors <111. 0) + <v. a vector in H EB H has that form if and only if its second coordinate belongs to the range R of the operator A. Av) = <u + v. of course. The result is a special case of the socalled closed graph theorem. the set of all vectors (in H EB H) of the form <J. Since A can be chosen so that R is not closed. These operators are. 0). p. Conclusion: M + N is closed in H EB H if and only if R is closed in H. however. Let M be the "x-axis ". 53. It is trivial to verify that both M and N are subspaces of H EB H. !. When does <J. since u and v are arbitrary. 1). Av). In sum: A is invertible if and only if it is bounded from below and has a dense range (see [50. p. define (Af)(x) = xf(x). The corresponding statement about Banach spaces is usually proved by means of the Baire category theorem..

An important question about operators is what do they do to the geometry of the underlying space. 55 55.f. ej)ej. Assume that {ej} is a maximal orthonormal set. (a) It is possible for a subspace M to have a diminishable complement N. Solution 32. i. Problem 54. Orthogonal dimension can therefore be defined for every inner product space. The reason is that the condition M n N = 0 implies the uniqueness of the representation of a vector in the form f + g (f E M. Is every maximal orthonormal set in an inner-product space total? 56 56. Total orthonormal sets. for an extreme example. form the Fourier expansion Lj (f.PROBLEMS Problem 53. There exists a linear manifold G dense in a Hilbert space H such that dim G -. If. dim H. N. the standard proof makes no use of completeness. given an arbitrary vector f. e)ej is orthogonal to each ej. Indeed. then M v N cannot possibly be equal to M + N. A total orthonormal set in an inner-product space (not necessarily a Hilbert space) is maximal. g E N). but completeness is needed to guarantee that the sum exists in the space. but although it makes sense for "bad" inner-product spaces. Bessel's inequality guarantees that the terms are not too large to be added.f. collapses every space to a 0-dimensional one. The crucial point in this proof is the formation of Lj (f. in the sense that there exists a complement N 0 with N 0 c N and dim(N n N~) infinite? The statement (a) is a generalization of the existence of non-closed vector sums.f. e)ej. Preservation of dimension. however. since the difference f . M + N. a linear transformation on a finite-dimensional vector space is one-to-one 30 . with N 0 c N. in the sense that there exists another complement N 0 . its properties are likely to be bad too.Lj (f. N 0 -. word for word as for Hilbert spaces (the common cardinal number of all maximal orthonormal sets). (b) Can a complement N be infinitely diminishable. use maximality to infer the desired conclusion.. if any part of N is discarded. the rest cannot form a conspiracy with M to recapture it. It is familiar from the study of finite-dimensional vector spaces that a linear transformation can lower dimension: the transformation 0. A possible proof of the converse goes as follows. (Cf. or is it in the nature of things? Problem 55. Dimension in inner-product spaces. 54 54.e. ifN is a diminishable complement ofM. the existence of subspaces M and N such that M v N -. The statement that any two maximal orthonormal sets have the same cardinal number is true for every inner product space.) Is this use of completeness an awkwardness in the proof.

that is. The assertion is known as the closed graph theorem for Hilbert spaces. Closed graph theorem. then dim H ~dim K. Problem 57.ROUNDEDNESS AND INVERTIBILITY (i. g) (elements of H Ef> K) for which Af = g. at the very least it serves as a warning that the same object is about to be viewed from a different angle. and it is hard to see what is accomplished by giving it another name. If an operator A is "left divisible" by an operator B. then it cannot lower dimension. If there exists a one-to-one bounded linear transformation from a Hilbert space H into a Hilbert space K. (The terminology is standard. Range inclusion and factorization. 57. what the closed graph theorem says is that they can occur on incomplete inner-product spaces only (or nonclosed linear manifolds in Hilbert spaces). 57]). Problem 58. Is the converse true? 59 31 . 58. since the same can be said about the inverse transformation (from the range back to the domain). It is curious that it should be so.e. the graph of A is A. its kernel is 0). its proof for Banach spaces is usually based on a category argument ([39. then equality holds. The theorem does not make the subject of closed but unbounded linear transformations trivial. the generalization of this finite-dimensional result to arbitrary Hilbert spaces. in a sense. it follows that dimension is preserved.. This is a special case of Problem 130. Nevertheless most mathematicians cheerfully accept the unnecessary word. a function. A linear transformation from a Hilbert space into a Hilbert space is closed 57 liP . then ran A c ran B. 59. p. According to a widely adopted approach to the foundations of mathematics. The following assertion is. Problem 56.) A linear transformation is called closed if its graph is a closed set. if there exists an operator X such that A = BX. According to that approach. by definition. Such transformations occur frequently in the applications of functional analysis. If the image ofH is dense inK. but it is. If P and Q are projections such that P and Q have the same rank. The graph of a linear transformation A (not necessarily bounded) between inner product spaces H and K (not necessarily complete) is the set of all those ordered pairs (f. Projections of equal rank. then 58 if and only if it is bounded.Qll < l. is a set of ordered pairs satisfying a certain univalence condition.

that is. !fran A c ran B. Unbounded symmetric transformations. then A is "majorized" by B. g) = (f.) The converse is true.PROBLEMS Problem 59. does it follow that A is left divisible by B? The corresponding question about right divisibility is different. there exists a constant (J( such that IIAfll ~ (J(IIBfll for all f. then II XII will do for()(. It is advisable to use this neutral term (rather than" Hermitian" or" self-adjoint"). because in the customary approach to Hermitian operators (A = A*) boundedness is an assumption necessary for the very formulation of the definition. (a) Is a symmetric linear transformation on an inner-product space H necessarily bounded? (b) What (fH is a Hilbert space? 32 . the proof is a straightforward geometric construction. (If A = XB. and. easier to answer. as it happens. Reference: [38]. Is there really a distinction here? Problem 60. 60 60. A linear transformation A (not necessarily bounded) on an inner-product space H (not necessarily complete) is called symmetric if (Af. If A is right divisible by B. Ag) for all f and gin H.

g.. and then never mentioned again. a 1.. if it is bounded. Problem 61. the emphasis is on both the cardinal number (~ 0 ) and the order (w) of the underlying index set. Operator theory. for eachj. like every other part of mathe- 61 matics. Diagonal operators. constructions by induction). but in most discussions of diagonal operators a basis is (perhaps tacitly) fixed in advance. An operator A is called a diagonal operator if Ae1 is a scalar multiple of e1. Suppose. then the equations Ae1 = a1e1 uniquely determine an operator A. That special case makes possible the use of some convenient language (e. and spectra. cannot be properly studied without a large stock of concrete examples.g. inverses. that His a Hilbert space and that {eJ is a family of vectors that constitute an orthonormal basis for H. diagonal operators can be characterized in invariant terms as normal operators whose eigenvectors span the space. The definition of a diagonal operator depends. say Ae1 = a1e1. and IIAII = supJiaJI· 33 . on the basis {ei}. a2 . A necessary and sufficient condition that a family {a1} be the diagonal of a diagonal operator is that it be bounded. of course.) Usually diagonal operators are associated with an orthonormal sequence. • • ·) to denote the diagonal operator with the indicated diagonal terms. which can then be examined for the behavior of their norms. for a modest first step. "the first element of the diagonal") and the use of some convenient techniques (e. It is sometimes convenient to use the symbol diag(a 0 . The purpose of several of the problems that follow is to build a stock of concrete operators. Alternatively. the family {a1} may properly be called the diagonal of A. (The proof of the characterization is an easy exercise.CHAPTER 7 Multiplication Operators 61.

in fact. ~ 2 .. i. Each sequence {ocn} of complex scalars induces a 2 linear transformation A that maps 1 into the vector space of all (not necessarily square-summable) sequences. of some diagonal operator). oc 3 ~ 3 . and with a norm (II {ocn} 1/ = supn Iocn 1). Consequence: the spectrum of a diagonal operator is the closure of the set of its diagonal terms. oc 2 ~ 2 . then <oc 1~ 1 . The correspondence preserves the norm also (see Problem 61).. A necessary and sufficient condition for this to happen is that {ocn} be bounded away from 0. Half of Problem 61 implies that if A is an operator (i. • · · ).e. OC3~3. The result has the following useful corollary: every non-empty compact subset of the complex plane is the spectrum of some operator (and. That problem considered sequences that 2 multiply z2 into 11 (and concluded that they must belong to 1 ). What happens if the boundedness assumption on A is dropped? Problem 62. • ·-) E 1 . ~3. The correspondence preserves not only the familiar algebraic operations.e.. that is. that 2 if <~ 1 . See Problem 66 for the generalization to L 2 • 63 63. i. then it is easy to verify that the correspondence {ocn} ~ A. . where A is the operator on H such that Aen = ocnen for all n.. Problem 63. if {ocn} ~ A. A diagonal operator with diagonal {ocn} is an invertible operator if and only if the sequence {ocn} is an invertible sequence. and form a diagonal operator with that sequence as its diagonal. is an isomorphism (an embedding) of the sequence algebra into the algebra of operators on H. Multiplications on P. a bounded linear transformation of P into itself). ~ 3 . The set of all bounded sequences {ocn} of complex numbers is an algebra (pointwise operations). if there exists a bounded sequence {/Jn} such that ocnPn = 1 for all n.. · · ·) = <oc1~1' OC2~2. Can an unbounded sequence of scalars induce a (possibly 2 unbounded) transformation of 1 into itself? The emphasis is that all z2 is in the domain of the transformation.:. with a conjugation ({ocn} ~ {ocn*}). by definition A<~t' ~2. then the sequence {ocn} is bounded. A bounded sequence {ocn} will be called invertible if it has an inverse in this algebra.PROBLEMS 62 62. Spectrum of a diagonal operator. The question should be compared with Problem 29. then {ocn*} ~A*. that there exist a positive number fJ such that locnl . Proof: find a sequence of complex numbers dense in the prescribed compact set. this one considers sequences that multiply [2 into [2 (and asks whether they must belong to lCL))..: fJ for all n. If His a Hilbert space with an orthonormal basis {en}. -) E P.e. with unit (ocn = 1 for all n). 34 .e. but also conjugation. i.

then the function cp is bounded. What. is the norm of the multiplication induced by cp? 65.e. (Here. Diagonal operators are special cases of a 64 general measure-theoretic construction. then multiplication operators reduce to diagonal operators. Can an unbounded function induce a (possibly unbounded) transformation ofL 2 (for a a-finite measure) into itself? 66 This is the generalization to measures of Problem 62. If an operator A on L 2 (for a a-finite measure) is such that Af = cp · f for all fin L 2 (for some function ([) ). Problem 64. a bounded linear transformation of L 2 into itself). then ([) is measurable and bounded. A 67 35 .1 is the counting measure (the measure of every set is the number of elements in it). two functions are identified if they differ on a set of measure zero only. for short) induced by cp is the operator A on L 2 (J. as elsewhere in measure theory.1) defined by (Af)(x) = cp(x)f(x) for all x in X. Spectrum of a multiplication. The set of all bounded measurable functions (identified modulo sets of measure zero) is an algebra (pointwise operations). Half of Problem 65 implies that if A is an operator (i. but the details become a little fussy at times. Suppose that X is a measure space with measure f.l· If cp is a complex-valued bounded (i. and with a norm (llcplloo).. Norm of a multiplication. What happens if the boundedness assumption on A is dropped? Problem 66. if there exists a bounded measurable function 1/J such that cp(x)!/J(x) = 1 for almost every x.. Roundedness of multiplications. A bounded measurable function is invertible if it has an inverse in this algebra. A sample is the generalization of the assertion that if a sequence is the diagonal of a diagonal operator. with unit (cp(x) = 1 for all x).) If X is the set of all positive integers and f. then the multiplication operator (or just multiplication. with a conjugation (cp~--+cp*). Some parts of the theory of diagonal operators extend to multiplication operators almost verbatim.MULTIPLICATION OPERA TORS 64. Problem 65. Each complex-valued measurable function cp induces a linear transformation A that maps L 2 into the vector space of all (not necessarily square-integrable) measurable functions.e. Much of the theory of diagonal operators 65 extends to multiplication operators on measure spaces. 67. then it is bounded. by definition (Af)(x) = cp(x)f(x).. Roundedness of multipliers. 66. i. as follows. This applies to the bounded cp's as well as to the square-integrable f's.e. in terms of the multiplier cp. essentiallyoounded) measurable function on X.

Multiplications on functional Hilbert spaces. In a certain sense the space 1 has too many multipliers: most of them do not belong to the space. Suppose that His a functional Hilbert space over a set X.1(N) has positive measure. Can this happen for infinite-dimensional spaces? 36 . In a certain sense the space has too few multipliers: most of the functions in the space are not among them. If X is finite and ifH consists of all functions on X. then qJ* ~A*. the easiest infinite-dimensional space.PROBLEMS necessary and sufficient condition for this to happen is that qJ be bounded away from 0 almost everywhere. is an isomorphism (an embedding) of the function algebra into the algebra of operators on L 2 • The correspondence preserves not only the familiar algebraic operations. The multiplication operator on L2 (for a a-finite measure) induced by qJ is an invertible operator if and only if qJ is an invertible function. where A is the multiplication operator defined by (Af)(x) = qJ(x)f(x). and therefore multiplication by qJ is necessarily an operator on V. (a) If Af = qJ · f. then the set of multipliers ofH is neither too large nor too small: it consists exactly of the elements of H. The role played by the range of a sequence is played. the correspondence preserves the norm also (see Solution 64).. and suppose that qJ is a complex-valued function on X such that qJ · f E H whenever f E H. that there exist a positive number b such that IqJ(x) I . It is frequently interesting and important to determine all multipliers of a functional Hilbert space. Suppose that His afunctional Hilbert space. 68 V into itself.?. Consequence: the spectrum of a multiplication is the essential range of the multiplier. then 68. is the linear transformation A bounded? (b) If Af = qJ · f and if A is bounded. by the essential range of a function qJ. Solution 68 says that every multiplier is bounded. 2 For 1 . a sequence) be a 2 multiplier is that it be bounded. but also the conjugation. Multipliers of functional Hilbert spaces.e. Are the analogues of these assertions true for functional Hilbert spaces? Problem 68. the set qJ . that is. If a function qJ multiplies qJ is necessarily bounded (Solution 66). The space A2 behaves differently: for it a necessary and sufficient condition that a function be a multiplier is that it be bounded and belong to the space. that is the set of all complex numbers A. b for almost every x. If the measure is a-finite. it is easy to prove that a necessary and sufficient condition that a function (i. i.e. Problem 67. in the general case. is the function qJ bounded? 69 69. such that for each neighborhood N of A. The correspondence qJ ~A. by definition. A function qJ on X is a multiplier of H if qJ · f E H for every fin H.. if qJ ~A. over a set X say.

37 . but almost.. that His an algebra. what the problem asks for. Construct an infinite-dimensional functional Hilbert space H such that the multipliers ofH are exactly the elements ofH. If 1 E H. the construction of an infinite-dimensional functional Hilbert space that is an algebra with unit (under pointwise functional multiplication) is not quite.e. then. of course. i. but trivial examples show that the converse is not true.MULTIPLICATION OPERATORS Problem 69. The constant function 1 is a multiplier of every H. Thus. To say that every element ofH is a multiplier is the same as to say that His closed under multiplication. the algebra H has a unit. to say that every multiplier ofH belongs to H is the same as to say that 1 E H. hence.

's are subs paces of H. that H = H 1 EB H 2 EB H 3 EB · · ·. has a decomposition: Ajj = gti + g2i + g31 + · · ·. The gii's depend. Some of the matrix theory associated with orthonormal bases deserves to be extended to more general direct sums. whose entry in row i and column j is the projection onto the i component of the restriction of A to H 1 . Each Afj. being an element ofH. then Af=Aft+Af2+Af3+ . Commutative operator determinants. If A = 0.1).i = 0 when i =.) If the direct sum is viewed as an "internal" one. then A.i = Aiif1.. The linear operations on operator matrices are the obvious ones. Suppose. on jj.i is a bounded linear transformation from H 1 to H.CHAPTER 8 Operator Matrices 70 70. ·.). of course. The construction is finished: corresponding to each A on H there is a matrix (A. to be specific. with!. then Ail = 0 for all i andj..6 j and Aii = 1 (on H. if A= 1 (on H). with gii in H. If A is an operator on H. It follows that g. then the elements f of H are sums f=ft+f2+f3+"·. The matrix 38 . so that the H. and the dependence is linear and continuous. where A. An orthonormal basis serves to express a Hilbert space as the direct sum of one-dimensional subspaces. (Uncountable direct sums work just as well.. The correspondence from operators to matrices (induced by a fixed direct decomposition) has all the right algebraic properties.. and finite ones even better. in H.

lfC and D commute. If A. Problem 71. Problem 70. that is.BC be invertible.BC be invertible. the order of the factors must be watched with care. C. The multiplication of operators corresponds to the matrix product defined by Lk A. 39 . the matrix of A* has the entry Aj.OPERATOR MATRICES of A* is the adjoint transpose of the matrix of A. and and sufficient condition that if Dis invertible.kBkj· There is no convergence trouble here. There are many situations in which the invertibility of an operator matrix 71 plays a central role but in which the entries are not commutative. and operators on that direct sum are expressed as two-by-two matrices whose entries are operators on H. then a necessary be invertible is that AD . 71. For finite matrices more is known (cf. The theory of operator matrices does not become trivial even if the number of direct summands is small (say two) and even if all the direct summands are identical. then the condition becomes unnecessary and insufficient. any special case is worth knowing. the role of what wasH in the preceding paragraph is played now by the direct sum H $ H.BC have the same determinant. Operator determinants. then a necessary and sufficient condition that the operator matrix be invertible is that the formal determinant AD . The following situation is the one that occurs most frequently: a Hilbert space His given. and D are pairwise commutative operators on a Hilbert space. but there may be commutativity trouble. The proof for the general case can be arranged so as to yield this strengthened version for the finite-dimensional case. [130]): if C and D commute.* in row i and column j. Construct examples to show that if the assumption that D is invertible is dropped. then and AD . B.

PROBLEMS 72 72. Problem 72. B. Problem 71) if both A and Dare invertible. The converse (if M is invertible. Note that A operates on H. Operator determinants with a finite entry. 40 . D operates on K. In at least one special case of interest the converse that was false in the preceding paragraph becomes true. and (cf. then both A and D are invertible. Operator matrices define operators on direct sums of Hilbert spaces whether the direct summands are identical or not. If A. with dim H < oo. and if M = (~ ~) is an invertible operator on H $ K. then M is invertible. then the operator matrix M = (~ ~) induces (is) an operator on H $ H. and B maps K into H. then A and D are) is not true (see Problem 71 again). IfH and K are Hilbert spaces. Consequence: the spectrum of M is the union of the spectra of A and D. and Dare operators on a Hilbert space H.

The question reduces therefore to this: why is a non-invertible operator not invertible? There are several possible ways of answering the question. The result is a partition of the spectrum into five parts. a number A belongs to 11(A) if and only if there exists a sequence Un} of unit vectors such that II(A .A is a proper subspace of H (i.e.11 0 ) by a diameter perpendicular to the overlap. Schematically: think of the spectrum as the union of two overlapping discs (11 and r). then 73 spec A = 11(A) u r(A). To say that A is in the spectrum of A means that A .. The set 11(A) is called the approximate point spectrum of A. Consequence: if spec A is the spectrum of A. and if r(A) is the set of complex numbers A such that the closure of the range of A . It is often useful to ask of a point in the spectrum of an operator how it got there.e. The set r(A) is called the compression spectrum of A. An important subset of the approximate point spectrum is the point spectrum I1 0 (A).A is not bounded from below. distinct from H).A)fn I ~ 0. but he would be well advised to postpone the attempt until he has 41 / . 11 0 (A) is the set of all eigenvalues of A). they have led to several (confusingly overlapping) classifications of spectra. The born taxonomist may amuse himself by trying to see which one of the 25 a priori possibilities is realizable. a number A belongs to it if and only if there exists a unit vector f such that Af = )/(i. Spectra and conjugation. if 11(A) is the set of complex numbers A such that A . each one of which may be sometimes present and sometimes absent. Perhaps the simplest approach to the subject is to recall that if an operator is bounded from below and has a dense range. then it is invertible.A is not invertible.CHAPTER 9 Properties of Spectra 73. one of which (11) is divided into two parts (11 0 and 11 ..

(The sets n and r may overlap. through illuminating special examples. 53]) is called a spectral mapping theorem. The most useful things to know are the relations of spectra to the algebra and topology of the complex plane. such as inversion. Two operators A and B are similar if there exists an invertible operator P such that p-t AP = B. orr in place ofspec? What about the spectral mapping theorem for inversion (p(z) = 1/z when z # 0). Similarity and spectrum. Similar operators have the same spectrum. and the approximate point spectrum when an operator is replaced by its adjoint? 74 74. Problem 74. or n. conjugation. There is no consensus on which concepts and symbols are most convenient in this part of operator theory. What happens to the point spectrum. and the present one is no exception. and the continuous spectrum n . but. The study ofthe latter might be called spectral theory. Problem 75. 569]). p. This is a good opportunity to comment on a sometimes confusing aspect of the nomenclature of operator theory. p. and there are things called spectra for all operators. A once popular approach was to divide the spectrum into three disjoint sets. In the normal case the spectral theorem gives information about spectral theory.(r u ll 0 ). and the same compression spectrum. Spectral theory in the present sense of the phrase is one of the easiest aspects of operator theory.PROBLEMS seen several more eJ<amples of operators than have appeared in this book so far. of course. Perhaps the easiest algebraic questions concern conjugation. There is something called the spectral theorem for normal operators (see Problem 123). Spectral mapping theorem. the same point spectrum. and sometimes it is. Apparently every book introduces its own terminology. The best way to master these concepts is. the compression spectrum. the same approximate point spectrum. Is the spectral mapping theorem for polynomials true with ll 0 . Problem 73. they help in the study of the examples. then spec p(A) = p(spec A) (see [50.ll 0 . namely. other instances of it have to do with functions other than polynomials.) As for symbols: the spectrum is often a (or :E) instead of spec. that information can be bought cheaper elsewhere. the residual spectrum r . and wide classes of analytic functions ([39. or r? 75 75. but a few general facts should come first. An assertion such as that if A is an operator and pis a polynomial. 42 . usually. applied to invertible operators. with llo' or n. the point spectrum ll 0 . examples will be easy to construct a little later.

) This implies (Problem 75) that if at least one of A and B is invertible. Problem 76. If neither A nor B is invertible. then. In the finite-dimensional case more is known: with no invertibility assumptions. Closure of approximate point spectrum. (For the proof. Problem 77. the two products need not have the same spectrum (many examples occur below). The non-zero elements of spec AB and spec BA are the same. then AB and BA have the same spectrum. 76 77 78 43 . AB and BA always have the same characteristic polynomial. in the infinite-dimensional case. then AB and BA are similar. Boundary of spectrum. Spectrum of a product. but their spectra cannot differ by much.1 (BA)B in case B is. Here is the precise assertion. Is the approximate point spectrum always closed? 78.1(AB)A in case A is invertible or AB = B. apply BA = A . The boundary of the spectrum of an operator is included in the approximate point spectrum. and if at least one of them is invertible. If A and B are operators. Problem 78. 77.PROPERTIES OF SPECTRA 76.

Problem 80. then r(A) = IT 0 (A) (and therefore spec A = IT(A )). compression spectrum. ~z. Problem 79. If A is normal. · · -). Spectral parts of a multiplication. 80 80.IT 0 (A). the determination of the fine structure of the spectrum requires another look. 44 ~o' ~1. 82 82. the unilateral shift is the operator U on ZZ defined by U<~o' ~1. Alternativeformulation: the residual spectrum ofa normal operator is always empty. the most amenable large class known. For each multiplication. Residual spectrum of a normal operator. Unilateral shift. 81 81. Problem 81. Spectral parts of a diagonal operator. Recall that the residual spectrum of A is r(A) . find its point spectrum. the worst spectral pathology cannot occur. The spectrum of a diagonal operator was determined (Problem 63) as the closure of its diagonal. Perhaps the simplest way to define it is to consider the Hilbert space ZZ of squaresummable sequences. For each diagonal operator. The time has come to consider special cases. is called the unilateral shift. which plays a vital role in all parts of Hilbert space theory. The most important single operator. com- pression spectrum. .CHAPTER 10 Examples of Spectra 79 79. find its point spectrum. ~z. The first result is that for normal operators. and approximate point spectrum. · · ·) = <O. and approximate point spectrum.

In other words. Is there a non-scalar operator whose set ofeigenvectors has a non-empty interior? 84. 1. Properly speaking the unilateral shift is a unitary equivalence class of operators. A familiar space that comes equipped with an orthonormal basis indexed by non-negative integers is H 2 (see Problem 33). To define it. but no confusion will result from regarding it as one operator with many different manifestations. ~ 1 . Norms are not only kept within reasonable bounds. ~ 2 .EXAMPLES OF SPECTRA (The unilateral shift has already occurred in this book. Is there an interesting answer? Problem 83. Bilateral shift. and in most of the study of U they may be taken as its definition. Problem 82. The unilateral shift shows that the 83 set of eigenvalues of an operator can be much richer than finite-dimensional experience suggests. To say that it is the "same". the unilateral shift is the same as the multiplication operator on H 2 defined by (Uf)(z) = zf(z). If en is the vector <~ 0 . compression spectrum. the effect of shifting forward by one index is the same as the effect of multiplication by e 1 . . and what are its parts (point spectrum. ···)for which ~n = 1 and~. en(z) = zn. depending on how the definition of eigenvector was formulated. ·). A close relative of the unilateral shift is the bilateral shift. see Solution 71. every vector is an eigenvector. and approximate point spectrum)? What are the answers to the same questions for the adjoint of the unilateral shift? 83. in fact. the unilateral shift is an isometry. The existence of an isometry whose range is not the whole space is characteristic of infinite-dimensional spaces. although it was not named until now. then the en's form an orthonormal basis for 1 • The effect of U on this basis is described by (n = 0. 1. it is true with room to spare. or every non-zero vector is one. in that space. but they are preserved exactly. What is the spectrum of the unilateral shift. Structure of the set of eigenvectors.= 0 whenever 2 i =f n (n = 0. 2. These equations uniquely determine U.) Linearity is obvious. let H be the Hilbert space of all two-way (bilateral) square- 84 45 . Since.. · · -). a convenient one that will be maintained throughout the sequel. The range of U is not F but a proper subspace ofF. As for boundedness. the subspace of vectors with vanishing first coordinate. 2. and. What about the set of eigenvectors? A dull answer to the question is given by scalars. to speak of" the" unilateral shift is a slight abuse oflanguage.

~2. ~1. ± 2. Caution: as the facts for multiplications on L 2 spaces show (cf. Linearity is obvious. The latter kind is usually harder to study. the bilateral shift. ~-2• ~-1• (~o). (n = 0. ~1. and what are its parts (point spectrum. Since the range of the bilateral shift is the entire space H. 46 . is an isometry. Halmos and A. The effect of W on this basis is described by <· ··. The elements ofH are most conveniently written in the form <· ··' ~-2• ~-1• (~o). Spectrum of a functional multiplication. the term in parentheses indicates the one corresponding to the index 0. it does. · · ·). A necessary and sufficient condition that an operator A on a Hilbert space H be representable as a multiplication on a functional Hilbert space is that the eigenvectors of A* span H. like the unilateral one. What is the spectrum of the bilateral shift. ~o' ~b · · ). compression spectrum. Shields. ~-3• ~-2. If en is the vector ~-b (~ 0 ). The bilateral shift is the operator W on H defined by W(.· ·. Problem 84.n (n = 0. however. either in the legitimate sense (on an L 2 ) or in the extended sense (on a functional Hilbert space). The result seems to be due to P. ±1. then the en's form an orthonormal basis for H.PROBLEMS summable sequences. R. ± 1. Every operator studied so far has been a multiplication. have the advantage of having a satisfactory characterization in terms of its spectrum.= 0 whenever i #. it is even unitary. ±2. L. Problem 85. · · ·) = <· ··.(~-1). ~2• · · ) . ~ 1 . and boundedness is true with room to spare. ···)for which ~n = 1 and~. and approximate point spectrum)? lfhat are the answers to the same questions for the adjoint of the bilateral shift? 85 85. ···). Solution 81) this characterization is applicable to functional Hilbert spaces only.

this makes it seem plausible that the resolvent is a well-behaved function.. To say that every finite matrix has an eigenvalue is the same as to say that the characteristic 47 . ~----? (cp(A. Does every operator have a non-empty 87 spectrum? The question was bound to arise sooner or later. (To distinguish this concept from other closely related ones. and at oo. quite generally.) = cp(l/.) can be assigned a value at the origin so that it becomes analytic there.. its values are operators on H. then (just as for numerical functions) cp will be called analytic at oo. The domain of pis the complement of the spectrum of A.) =(A. Non-emptiness of spectra. 86 The definition of the resolvent is very explicit.A..1 . does not belong to the spectrum of A. 3]. To formulate this precisely.1. VII. write p(A. Problem 86. 87. If . consider.CHAPTER 11 Spectral Radius 86. its value at oo is (the operator) 0.. 1.}. The resolvent of every operator is analytic at each point of its domain. and cp is assigned at oo the value of t/1 at 0.)f. the numerical function . is invertible. (When it is necessary to indicate the dependence of the function p on the operator A.1. for each f and g in H. Analyticity of resolvents. it is sometimes called weak analyticity.) In case the function t/1 defined by t/1(. write p = p A-) The function p is called the resolvent of A.1. Suppose that A is an operator on a Hilbert space H. functions cp whose domains are open sets in the complex plane and whose values are operators on H. Even the finitedimensional case shows that the question is non-trivial. Such a function cp will be called analytic if. For a detailed study of resolvents. 1. then the operator A . g) (with the same domain as cp) is analytic in the usual sense. see [39.

the finite-dimensional case of the general question about spectra is as deep as the fundamental theorem of algebra. Every operator has a non-empty spectrum. Weighted shifts.£ IIAII. · · ·. then two-dimensional examples. More explicitly. and that is no more and no less general than to say that every polynomial equation (with complex coefficients) has at least one (complex) zero. · · -). suppose that {en} is an orthonormal basis (n = 0. Clearly 0 .£ r(A) .PROBLEMS polynomial of every finite matrix has at least one zero. 88 88. whose proof is usually based on the theory of complex analytic functions. For each operator A. It is a somewhat less easy consequence. then r(AB) . It is an easy consequence of this result that if A and B are commutative operators. that if A and B commute. The spectral radius of an operator A.i:A. Problem 88.EspecA}. ± 1. It should not be too surprising now that the theory of such functions enters the study of operators in every case (whether the dimension is finite or infinite). A weighted shift is the product of a shift (one-sided or two) and a compatible diagonal operator. the tool that makes it easy is the following assertion. is defined by r(A) = sup{IA. 89 89. in the sense that the indicated limit always exists and has the indicated value. Problem 87. but still a matter of no more than a little fussy analysis with inequalities. 2.£ r(A) + r(B). r(A) =limn 11Anll 11n. in symbols r(A). Spectral radius. In other words. 1. the spectral mapping theorem implies also that r(An) = (r(A)t for every positive integer n. such as A=(o 0 ' o) 1 show that neither the submultiplicative nor the subadditive property persists. and suppose that {an} is a bounded sequence of complex numbers (the set of n's 48 . If no commutativity assumptions are made. ±2. It frequently turns out that the spectral radius of an operator is easy to compute even when it is hard to find the spectrum. then r(A + B) .£ r(A)r(B). or else n = 0.

then rxA. is in the spectrum and if 1rx I = 1.then A and B are unitarily equivalent. Similarity of weighted shifts. the reason is that. Is the converse of Problem 89 true? Suppose. then there exists a unitary operator U such that Uen = fn for all n. but even the little that is makes them almost indispensable in the construction of examples and counterexamples. · · ·).} are orthonormal bases. a translation of the weights produces a unitarily equivalent shift. If some of the weights are allowed to be zero. then there is at least one n such that lrxnl =1. with diagonals {an} and Wn}. If {en} and {f. That is: if Aen=rxnen+ 1 and Ben=rxn+ 1 en+ 1 (n=O. ± 1. where Sis a shift (Sen = en+ 1) and Pis a diagonal operator with diagonal {an} (Pen = an en). in fact. does it follow that lrxnl = IPnl for all n? The answer can be quite elusive. that the spectrum of a weighted shift has circular symmetry: if A. by rights. ··-). if. whose weights have the same moduli as the corresponding weights of A. Since unitarily equivalent operators are "abstractly identical".lrxn+ 1 1. It is sometimes convenient to use the symbol shift<a0 . A weighted shift is an operator of the form SP. A discussion of two weighted shifts should.SPECTRAL RADIUS being the same as before). if A and Bare unitarily equivalent. First. If. • • ·) to denote the weighted shift with the indicated weights. then the weighted shifts A= SP and B = SQ are unitarily equivalent. Problem 89. the sequence {I rxn I} is not constant. If P and Q are diagonal operators. with weights {rxn} and Wn}. then. the situation is in part annoying and in part trivial. The answer is no. and U can be carried along gratis with any unitary equivalence proof. in other words. but with the right approach it is easy. n = 0. refer to two orthonormal bases. the weighted shift with weights an is unitarily equivalent to the weighted shift with weights Irxn 1. a weighted shift is not altered by multiplication by a number of modulus 1. it follows that A and rxA are unitarily equivalent. if A is a weighted shift and if rx is a complex number of modulus 1. In other words. that A and B are weighted shifts. 90. for bilateral shifts. ± 2. a 1 . Not everything about weighted shifts is known. but the generality gained that way is shallow. In the good case (no zero weights). ±1. a2 . and if !ani= IPnl for all n. is in the spectrum. the kernel of the adjoint A* of a unilateral weighted shift is 90 49 . W is the bilateral shift (Wen = en+ 1 . this is what really justifies the use of the word "weight". This implies. however. there is never any loss of generality in restricting attention to weighted shifts whose weights are non-negative. then W* A W = B. Second. ±2. Unilateral shifts behave differently. since rxA is a weighted shift. to within unitary equivalence. The result about the unitary equivalence of weighted shifts has two useful consequences. for instance.

a modification whose difficulties are more notational than conceptual. in general. • • • . Problem 93. Similarity is a less severe restriction than unitary equivalence. The power norms of an operator A. and from this. it follows that U is a diagonal operator. 2. then A*" and B*" are unitarily equivalent. If A and B are unilateral weighted shifts. the kernel of A* 2 is spanned by e0 and e 1. Problem 91. it follows that. L. it is possible to get a satisfactory condition. Power norms. for the similarity of two-sided shifts. 2. · · ·? 93 93. 50 . questions about similarity are usually easier to answer. that is the numbers IIAkll. this was done by R. 1. If A and Bare unitarily equivalent weighted shifts. 91 91. · · ·). for each n. A = U*BU. 3. The exact determination of the spectrum and its parts for arbitrary weighted shifts is a non-trivial problem. constitute an interesting sequence. Express the norm and the spectral radius of a weighted shift in terms of its weights. en_ 1 } is invariant under U. then U must send ker A*" onto kerB*". For which sequences {pk} of positive numbers does there exist an operator A such that Pk = IIAkll. the kernel of A*" is spanned by e0 . k = 0. This settles the unitary equivalence theory for weighted shifts with nonzero weights. what about similarity? Problem 90. Eigenvalues of weighted shifts. • • ·. if. with non-zero weights {an} and {f3n}.PROBLEMS spanned by e0 . Since the diagonal entries of a unitary diagonal matrix have modulus 1. say. like that in Problem 90. in turn. 92 92. en-l (n = 1. Kelley. then a necessary and sufficient condition that A and B be similar is that the sequence of quotients be bounded away from 0 and from oo. By a modification of the argument for one-sided shifts. and. Problem 92. Norm and spectral radius of a weighted shift. the effect of A on en can differ from that of B by a factor of modulus 1 only. This implies that the span of {e0 . Find all the eigenvalues of all unilateral weighted shifts (with non-zero weights) and of their adjoints. Here is a useful fragment.

it may be called a weighted sequence space.SPECTRAL RADIUS The possible presence of 0 among the weights is not a genuine difficulty but a nuisance. ) E l 2 (p). IIS/11 is bounded by a constant multiple of II /II? The answer is easy. <1'fo. The expression "weighted shift" means one thing. Problem 98).) When is the shift an operator on this space? When. what it could have meant is to modify by attaching weights to the space. · . then Sf = ~ 0 . of course. 2 for all n. What it does mean is to modify the ordinary shift on the ordinary sequence space l2 by attaching weights to the transformation. To get an explicit description of the alternative. and. is it true that if f = <~o. ~ 1 . but the good problems about shifts have to do with non-zero weights. IIS/11 = 2 n=l IPnl~n-tl 2 =n=l ~Pn-tl~n-tl 2 I Pn-l 00 ~ l:t.. If A is a weighted shift. in other words. ~ 2 . p 2 . then <O. ~t> ~ 2 . · · ·)) = n=O L Pn~n1'fn*• the set l (p) is a Hilbert space. With respect to the coordinatewise linear operations and the inner product defined by 00 94 95 «~o' ~1• ~2• 2 • · ·). 11/ll 2 2 • 51 . then the circular symmetry of each part of the spectrum (Problem 89) implies that every complex number of modulus r(A) belongs to the approximate point spectrum of A. A unilateral weighted shift. one of whose weights vanishes. • • ·} be a sequence of strictly positive numbers. the bilateral case can be treated similarly. but it could just as well have meant something else. 2 L Pnl~n1 n=O 2 = r:t. Question: does it ever include more? Problem 94. is the vector whose coordinate with index n is 1 and all other coordinates are 0. The presence of an infinite number of zero weights can cause some interesting trouble (cf. as f varies 2 over l (p). (All this is unilateral. · . ) E F(p). where en. Is there a weighted shift whose approximate point spectrum has interior points? 95. Weighted sequence spaces. 94. For each operator A there exists a complex number of modulus r(A) in the approximate point spectrum of A (Problem 78). If Pn+ t!Pn ~ r:t. p 1.llenll.. · ·-) with L:'=o Pnl~nl 2 < oo. ~ 2 . It is almost obvious that this necessary condition is also sufficient. becomes thereby the direct sum of a finite-dimensional operator and another weighted shift. Since llenll 2 = Pn• this condition says that the sequence {Pn+ t!Pn} is bounded. let p = {p 0 . and let l 2 (p) be the set of complex sequences <~ 0 . ~ 1 . Consequence: the approximate point spectrum of a weighted shift always includes a circle (possibly degenerate). such that lien+ 1 11 ~ r:t. Approximate point spectrum of a weighted shift. 1'/t> 1Jz. An obviously necessary condition is that there exist a positive constant r:t.

A nilpotent operator always has a non-trivial kernel. (The proof is of no relevance to Hilbert space theory. if cp(z) = z for z in D. Similar techniques show the existence of operators A with IT 0 (A) = 0 and spec A = [0. (To say that M is the support of J1 means that if N is an open set with Jl(M n N) = 0. The spectral mapping theorem implies that if A is nilpotent.) It follows that every such set is the spectrum of an operator with no eigenvalues. Problem 96. an operator A is quasinilpotent if limn IIA"II 11" = 0. say. If p = {Pn} is a sequence of positive numbers such that {Pn + tlPn} is bounded. moreover. Whenever a compact set M in the plane is the support of a measure (on the Borel sets) that gives zero weight to each single point. The proof in Problem 63 (every non-empty compact subset of the plane is the spectrum of some operator) is not sufficiently elastic to yield examples of all the different ways spectral parts can behave. then the spectrum of A is the closure 15. Construct a quasinilpotent operator whose point spectrum is empty. One-point spectrum. If Dis a bounded region. The expression for the spectral radius in terms of norms implies that if A is quasinilpotent. in terms of {Pn}. It is obvious that nilpotence implies quasinilpotence. compact. 1]. and hence a non-empty point spectrum. and if A is the multiplication operator induced by qJ on the L 2 space of planar Lebesgue measure in D. Observe that on finite-dimensional spaces such a construction is clearly impossible. the converse is true.PROBLEMS Every question about weighted shifts on the ordinary sequence space can be re-asked about the ordinary shift on weighted sequence spaces. That proof used diagonal operators. but the point spectrum of A is empty. for quasinilpotent operators that is not so. from that proof alone it is not possible to infer the existence of operators whose point spectrum is empty. just use linear Lebesgue measure in [0. which always have eigenvalues. then spec A = {0}. here is a sample. is the spectral radius of the shift on l2 (p)? 96 96. and that. An operator is nilpotent if some positive integral power of it is zero (and the least such power is the index of nil potence).) It is a routine exercise in topological measure theory to prove that every non-empty. what. What about sets that are not perfect? A very satisfactory answer can be given in terms of the appropriate analytic generalization of the algebraic concept of nil potence. then spec A = {0}. Problem 95. 52 . then M is the spectrum of an operator with no eigenvalues. perfect set (no isolated points) in the plane is the support of a measure (on the Borel sets) that gives zero weight to each single point. Multiplication operators come to the rescue. 1]. then M n N = 0.

Analytic quasinilpotents. more precisely. and its spectrum is the closure of the union of their spectra (Problem 63). Polynomials in an operator make sense. The correspondence f ~ f(A) between functions and operators respects the algebraic structures involved. what about the obvious extension to "infinite polynomials". Reason: since 11Anll 11n < e for n sufficiently large. The extension of this result from two direct summands to any finite number is a trivial induction. or. it is an extension of the trivial functional calculus for polynomials. Such an operator is an infinite direct sum. f(z) = L:'=o anzn. The simplest of them has to do with operators that have the smallest possible spectrum. the homomorphism from functions f analytic at 0 to operators f(A) (where A is a fixed quasinilpotent operator) is called a functional calculus. If A is such an operator. that is. What happens if the number of summands is infinite? A possible clue to the answer is the behavior of diagonal operators on infinite-dimensional spaces. by a natural extension oflanguage it seems proper to say that A is analytic if there exists a non-zero function f analytic at 0 such that f(A) = 0. This operator series converges (in the norm).SPECTRAL RADIUS 97. and the same is true of the point spectrum. Is every quasinilpotent operator analytic? 98. convergent when. quasinilpotent operators. each summand of which is an operator on a one-dimensional space. Spectrum of a direct sum. and the compression spectrum. An operator A is called algebraic if there exists a non-zero polynomial p such that p(A) = 0. more precisely. where f is an analytic function and A is an operator. it follows that IIAnll <en for n sufficiently large. power series? There are several possible approaches to defining expressions such as f(A ). Is the spectrum ofa direct sum ofoperators always the closure ofthe union of their spectra? 97 98 53 . and iff is a function that is analytic in a neighborhood of 0. lzl < e. The spectrum of the direct sum of two operators is the union of their spectra. then f has a power series expansion. say. in that case write f(A) = L:'=o anAn. This phenomenon. Problem 98. is the generalization to "infinite polynomials" true? Problem 97. Every nilpotent operator is algebraic. or. the approximate point spectrum. sums and products of functions correspond to sums and products of operators.

Continuity ofinversion. incidentally. the convergence of the series L:'=o A" was inferred from the assumption IIA I < 1.CHAPTER 12 Norm Topology 99 99. does it follow that the metric space B(H) of operators on it is separable? 100 100. Metric space of operators. it is customary and necessary to ask about the continuity of the pertinent algebraic operations. the set B(H) of all operators on a Hilbert space H becomes a metric space. it is not to be expected that the operator space is. such as the space of operators on a Hilbert space. indeed. In the present case it turns out that all the elementary algebraic operations (linear combination. it is easy to prove that it is not. In Solution 86. The proof is the kind of routine analysis every mathematician has to work through at least once in his life. and. Thus. multiplication) are continuous in all their variables simultaneously. The answer is yes. Soon after the introduction of a topology on an algebraic structure. If the distance between two operators A and B is defined to be IIA . it is not more than minimum courtesy to ask whether or not the space is complete. Some of the standard metric and topological questions about that space have more interesting answers than others.Bll.) So much for completeness. for instance. has been tacitly used already. it offers no surprises. what about separability? If the underlying Hilbert space is not separable. If a Hilbert space H is separable. The alert reader should have noted that the justification of this inference is in the completeness result just mentioned. That leaves one more natural question along these lines. (It takes less alertness to notice that the very concept of convergence refers to some topology. conjugation. The result. and the norm of an 54 . Problem 99.

for instance. One of them is about conjugate classes. In more technical language: are there any isolated singular operators? The answer is no.) 101. quite possible for a conjugate class to be closed. It would be quite reasonable to try to define what it means for a function of this 55 . Is the mapping A~ A. (They are the equivalence classes of the relation of similarity. Is there an open conjugate class? 102. (Example: the standard finite-dimensional similarity theory implies that all the matrices with a =!= 0. Is the open set of invertible operators connected also? That question is much harder. so is tA for all scalars t. consider A = 0. whether or not invertible operators can completely surround a singular (=non-invertible) one. (A trivial example is given by A = B = 1. The main algebraic operation not mentioned above is inversion. Continuity of spectrum. the mapping t ~ tA is a continuous curve that joins the operator 0 to the operator A. for an example. Interior of conjugate class. What are the topological properties of 101 similarity? The question includes several precise subquestions. their closure contains (~ ~). Since not every operator is invertible.NORM TOPOLOGY operator is also a continuous function of its argument. Problem 100. the question of the continuity of inversion makes sense on only a subset of the space of operators.) Conjugate classes do not have to be closed sets. The proofs are boring. are similar. The spectrum (restricted for a moment to 102 operators on just one fixed Hilbert space) is a function whose domain consists of operators and whose range consists of compact sets of complex numbers. Here is an amusing puzzle about exponentiation and invertibility with a perhaps unexpected solution: determine all operators A and B such that B is invertible and An -+ B as n -+ oo. Reason: if A is singular. (~ ~) . In more detail: the conjugate class of an operator A is the set of all operators similar to A. It does not say. can they be? Problem 101.1 of that set onto itself continuous? The statement that the set of invertible operators is open does not answer all questions about the geometry of that set.) It is. the set of singular operators is (arcwise) connected. Conjugate classes do not have to be open (consider A = 0 again). see Problem 141. The set of invertible operators is open. however.

167]. (The proof is immediate from 56 . cannot belong to limsupn spec An. Problem 102. it follows that limsup spec An c spec A. p. another is [ 112. Problem 103. then limsupn An is the set of all limit points of those sequences {A. Problem 21. so is the spectral radius.B I < 8. ¢: spec A. The proofs needed for this assertion are straightforward. · · · and if k = oo. Since the spectrum is upper semicontinuous (Problem 103). 35]. the answer is always no. and if An ~ A and A. 2.0 or n = 0. in outline form they go like this. · ··. Continuity of spectral radius.) What are the spectra of the operators Ak (k = 1. If spec is known to be metrically upper semicontinuous. Is the spectrum continuous? The following example is designed to prove that however the question is interpreted. n The two definitions are equivalent. Semicontinuity of spectrum. 2. conversely. 104 104. then spec B c A0 • The sequential definition has to be preceded by an auxiliary concept: if {An} is a sequence of sets of complex numbers. p. Metric definition: to each open set A0 that includes spec A there corresponds a positive number 8 such that if II A . That is: to each operator A and to each positive number b there corresponds a positive number 8 such that if IIA . If k = 1. Could it happen the other way? Is there a small spectrum with arbitrarily near large spectra? The answer turns out to be no. then separate A. Spectrum is upper semicontinuous. This is a standard result. then the negation of metric upper semicontinuity leads to a sequence {An} with An ~ A that contradicts the assumption. Sequential definition of upper semicontinuity: whenever An ~ A. (Cf.n} for which An E An for each n.) Upper semicontinuity for a set-valued function such as spec has at least two definitions. let Ak be the two-sided weighted shift such that Aken is en+ 1 or (1/k)en+ 1 according as n #. and spec A by disjoint open sets. and infer that A. then r(B) < r(A) + b. a metric one and a sequential one. One standard reference is [74. the relevant concept is that of an upper semicontinuous function. (Put 1/oo = 0.Bll < 8.PROBLEMS kind to be continuous. The semicontinuity of related functions is discussed in [63]. 3. If. spec is known to be sequentially upper semicontinuous and A0 is an open set that includes spec A. 3. The example of Problem 102 shows that there exists an operator with a large spectrum in every neighborhood of which there are operators with relatively small spectra. oo)? 103 103.

then lr(A) .NORM TOPOLOGY Problem 103. define liminfn An to be the set of all limits of those convergent sequences v. (Recall that limsupn An was defined as the set of all limit points of not necessarily convergent sequences of this kind. 105. Normal continuity of spectrum.n} for which An E An for each n. what about the spectral radius? Problem 104. Note that the example in Problem 102 gives no information.) Normal operators frequently behave not only algebraically but also topologically better than others.) A general characterization is given in [27]. a necessary and sufficient condition for continuity at A is that spec A c liminf spec An. does it follow that r(An) --+ r(A)? This is hard. (Example: if spec A is totally disconnected. To say that spec is continuous at a particular operator A means that An --+ A implies spec An--+ spec A. then A is a point of continuity of spec.r(B)I < f>? Equivalently: if An--+ A. in that case the spectral radius is equal to 1 for each term of the sequence and also for the limit. see Problem 103. more generally. The determination of the set of all points of continuity of spec is a nontrivial problem. Is it true that to each operator A and to each positive number f> there corresponds a positive number e such that if IIA .Bll < e. (Sample result: the restriction of spec to every commutative set is continuous. Solution 106. Special case: if A is of finite rank.) The spectrum is not continuous (Problem 102). if A is compact. n n in that case the common value deserves to be called limn An. Cf. or. Another useful question asks whether the restriction of spec to various sets of operators is continuous. Does it make sense to speak of the continuity (not just semicontinuity) of a set-valued function such as spec? The most convenient answer is in sequential terms. Is the restriction of spec to the set of normal operators continuous? 57 . Since upper semicontinuity is true for all A. Some interesting subsets of it were studied by Newburgh in his seminal paper [I 02]. what happens if spec is restricted to them? Problem 105. then A is a point of continuity of spec. If {An} is a sequence of sets of complex numbers.) The sequence {An} is called convergent (An --+ A) if and only if 105 liminf An = limsup An.

p}. If A and Bare operators such that spec(A for n + nB) = spec A = 0. the answer is easily seen to be no. If. and if spec( A + nB) = spec A for just one non-zero value of n. There is a sense in which the operator Bin the preceding example is small: its spectrum is small. then spec(A + B) = spec A = {a. 58 . Does the converse behave any better? That is: does a perturbation that produces only a small change in the spectrum have to be small? If small is interpreted in the sense of norm. Is that the clue? Problem 106. 1. for instance. · · · . the spectrum can undergo a large change. The discontinuity of the spectrum shows that if an operator is perturbed by a "small" summand. Quasinilpotent perturbations of spectra. 2. A = (~ ~) and B = (~ ~). The operator A can in fact be perturbed by arbitrarily large summands without changing its spectrum at all: spec(A + nB) = spec A for all n.PROBLEMS 106 106. then spec B = 0. If. but B is not small. A is a scalar. for instance. does it follow that B is quasinilpotent? The answer is clearly yes in some degenerate cases.

i = 1. and the strong topology is smaller than the norm topology. every weak open 59 . (Anf. The easiest questions to settle are the ones about comparison. The corresponding concepts of convergence (for sequences and nets) are easy to describe: An--+ A strongly if and only if An f --+ Af strongly for each f (i. Here k is a positive integer. • • ·.ll < e. to distinguish it from the others. where f and g are vectors and e > 0.A0 )/.. A subbase for the weak operator topology is the collection of all sets of the form '{A: I((A. The next two are natural outgrowths for operators of the strong and weak topologies for vectors. it is usually called the norm topology or the uniform topology. k}.Ao)fll < e}.. j. g) for eachf and g). and e is a positive number.CHAPTER 13 Operator Topologies 107. A Hilbert space has two useful topologies (weak and strong).A 0 )j. The weak topology is smaller (weaker) than the strong topology. g)--+ (Aj. the space of operators on a Hilbert space has several. as above (as always) a base is the collection of all finite intersections of such sets. correspondingly a base is the collection of all sets of the form {A: II(A. A subbase for the strong operator topology is the collection of all sets of the form 107 {A: II(A . The metric topology induced by the norm is one of them. g)i < e}. · · ·.e. Topologies for operators. In other words. For a slightly different and often very efficient definition of the strong and weak operator topologies see Problems 224 and 225. / 1. are vectors. II(An . and An--+ A weakly if and only if Anf--+ Ajweakly for eachj(i.e.A)f II --+ 0 for each f).

cf. Here are a few sample questions. a small additional effort must be made to avoid confusion. None of these acronyms will ever be seen below. it has a conjugate space. but even those few are more like routine operator extensions of the corresponding vector questions and answers (in Chapters 2 and 3) than inspiring novelties. In most cases the operator proofs are not ingenious corollaries of the vector facts but repetitions of them. Associated with these questions there is a small verbal misfortune.des a strong neighborhood of that operator. always refers to the norm. see. namely the ones given by inner products (A~ (Af. and. and if IIAnf -A/II IIAn. Problem 16. imitate Solution 23. People who do that use SOT also (strong). 60 . There is a big difference between the two. (1) Is B 1 (H) (the closed unit ball in B(H)) compact? (A topological term such as "compact". In still other words: every weak neighborhood of each operator inclu. when it is. regrettably. as such. the implications are reversible. and every strong neighborhood includes a metric neighborhood. strongly separable?. Solution 115. then The plethora of topologies yields many questions. (The answer is a special case of the corresponding theorem about arbitrary Banach spaces.All-+ 0. strongly metrizable? (Yes. This weak topology is defined in terms of the set of all bounded linear functionals on B(H). Problem 107. cf. with curt answers and no proofs. g)-+ (Af. the weak topology. but they stop short of NOT (norm). one school of thought abandons the grammatically appropriate form of "weak" in favor of a parenthetical WOT. is B 1(H) separable? (No. Again: norm convergence implies strong convergence. If (Anf.Afll -+ 0.) Is it weakly separable?. then 0 uniformly for 11/11 = 1. for instance. to all parts. with some extra notational complications. The definition of the weak topology for operators above uses only some linear functionals on B(H). with no indication of topology. g) IIAnf . "weak" for operators will always refer to the topology defined by inner products. The Banach space weak topology is rarely used in the operator context.PROBLEMS set is a strong open set.) (2) If H is separable. -+ uniformly for llgll = 1. and every strong open set is norm ope~. Problems 17 and 24. A few of those questions are interesting and have useful answers. g)). which induces a topology in B(H) called. weakly metrizable?.) (3) Is B 1 (H) weakly compact? Yes. see Solution 99. In the presence of uniformity on the unit sphere. and strong convergence implies weak convergence. In an attempt to do so. These facts are immediate from the definitions. The space B(H) is a Banach space.) Answer: yes if and only if the underlying Hilbert space H is finitedimensional. standing for weak operator topology. Is B 1 (H) strongly compact? No.

and the proof is a matter of routine. The proof is easy.IIA. and for operators in Solution 108 (where it was shown that the norm is not even strongly sequentially continuous). The most useful. Continuity of multiplication. Continuity of adjoint. i. endowed with one of the appropriate operator topologies) the proof that something is continuous is usually dull.e.. The norm is not weakly continuous. the function A r-+ I A II) continuous? 109. Multiplication is continuous with respect to the uniform topology and discontinuous with respect to the strong and weak topologies.. Equivalently: for every a > 0 there exists an n 0 such that IIAII . Which of the three topologies (uniform. Problem 108. Problem 110. it is true that the linear operations on operators (aA + {JB) are continuous in all variables simultaneously.II.e. available. The assertion means that if {An} is a net converging weakly to A.OPERA TOR TOPOLOGIES 108. strong. The norm (of an operator) is weakly lower semicontinuous (and a fortiori strongly lower semicontinuous).II ~ B whenever n ~ n 0 • IIAII 110. It is usual. to interpret them as referring to the mapping that sends an ordered pair (A. B) onto the product AB. then ~ liminf. weak) . strong. for instance. Continuity of norm.e. 110 111 questions concern products. Which of the three topologies (uniform. a continuity statement about products has a "joint" and a "separate" interpretation. (Readers who have never been through this routine are urged to check it before proceeding. Semicontinuity of operator norm.) Here is a related question that is easy but not quite so mechanical. but the counterexamples are hard.. when nothing is said to the contrary. 61 . Problem 111. the mapping A r-+ A*) continuous? 111. Half of continuity is. and most recalcitrant. In the study of topological algebraic structures 108 (such as the algebra of operators on a Hilbert space. to interpret such statements in the "joint" sense. IIA. Problem 109. makes the norm (i. however. weak) makes the adjoint (i. Thus. the interesting problems arise in proving that something is not continuous. Since a product (unlike the norm and the adjoint) is a function of two variables. 109 for vectors this was discussed in Problem 20. the quickest ones depend on unfair trickery.

Problem 112. ·)? 115 115.--. Separate continuity of multiplication. 3. it is separately continuous in each of its arguments with respect to both topologies. Weak convergence of projections. Is every weakly convergent sequence of projections strongly convergent? 62 . then it converges strongly to P. it can be used only when multiplication is commutative. Separate continuity (strong and weak) of multiplication is a feeble substitute for joint continuity. . Weak sequential continuity of squaring. then A. How much does the restriction to projections simplify matters? Problem 115. For which operators A (on an infinite-dimensional Hilbert space) is squaring weakly sequentially continuous? In other words. at least for some operators? Problem 114. Problem 113. Proof: for each f. A weakly. Sequential continuity of multiplication. A 2 weakly (n = 1.. 113 113.PROBLEMS 112 112. A slightly more precise formulation runs as follows. for instance. Each of the mappings A~----+. and therefore Problem 20 is applicable. AB strongly. In one respect. note. another feeble (but sometimes usable) substitute is joint continuity in the sequential sense. does that mean that there is a hope for the weak sequential continuity of squaring.b) 2 ).AB (for fixed B) and B ~----+. then multiplication is continuous also.. respectively.. however. (b) Does the assertion remain true if "strongly" is replaced by "weakly" in both hypothesis and conclusion? 114 114. then AnBn --.AB (for fixed A) is both strongly and weakly continuous. (a) If {An} and {Bn} are sequences of operators that strongly converge to A and B. Although multiplication is not jointly continuous with respect to either the strong topology or the weak. 2. for which A is it true that if An--. and proofs of that answer can be deduced from much of what precedes. Operator multiplication is not commutative.. Are the weak and the strong operator topologies the same? The answer is no. Since the argument depends on the identity ab = i ((a + b) 2 -(a . the different ways that sequential continuity of multiplication behaves. There is a familiar argument that can be used frequently to show that if the linear operations and squaring are continuous. weak and strong are alike: if a net {Pn} of projections converges weakly to a projection P.

CHAPTER 14 Strong Operator Topology 116. Strong bounded continuity of multiplication. Even though the adjoint is not strongly continuous. Strong semicontinuity of spectrum. and if {IIBnll} is bounded. The restriction of the adjoint to the set of normal operators is strongly continuous. 117. does it follow that Anfn---+ Af weakly? 119. The spectrum of an operator varies upper semicontinuously (Problem 103). then the spectrum can increase only a little. and if {II An II} is bounded. If {An} and {Bn} are nets that converge strongly to 0. then {AnBn} converges strongly to AB. Strong normal continuity of adjoint. it has an important continuous part. That is: if {An} and {Bn} are nets that converge strongly to A and B. Problem 118. does it follow that {AnBn} converges strongly to 0? 118. If {JJ is a sequence of vectors and {An} is a sequence of operators such that fn ---+ f weakly and An ---+ A strongly. What if the strong topology is used in place of the norm topology? 116 117 118 119 63 . The crux of the proof that multiplication is strongly sequentially continuous (Solution 113) is boundedness. Is this result symmetric with respect to the interchange of right and left? Problem 117. Problem 116. an operator is replaced by one that is near it in the norm topology. respectively. that is. Strong operator versus weak vector convergence. If.

The assumptions imply that the sequence {!/! n} is convergent and hence (Solution 1) that the limit !/! is a quadratic form. If E and Fare projections with ranges M and N. The assertion that a positive operator has a unique positive square root is an easy consequence of the spectral theorem.PROBLEMS Problem 119. Is the spectrum strongly upper semicontinuous? What can be said about the spectral radius? 120 120. It follows that !/!(/) = (Af.A)+ B/). In some approaches to spectral theory. then each !/In is a quadratic form. f) for all n and all f). then it is easy to find the projection with range N n M\ and if M j_ N. Thus. A bounded increasing sequence of Hermitian operators is weakly convergent (to a necessarily Hermitian operator). · · ·.find the projection E A F with range M 1"'1 N. f) for some (necessarily Hermitian) operator A. then the sequence {Bn} is strongly convergent.. 64 . the existence of square roots is proved first. 122 122. then it is sometimes easy and sometimes hard to find. f). To see this. (Anf. n = 0. 1.B) 2 =A. Increasing sequences of Hermitian operators. (AJ. The following assertion shows how to get square roots without the spectral theorem. If A is an operator such that 0 {Bn} is defined recursively by the equations Bo ~ A ~ 1. then (1 . in terms of E and F. if M c N. then it is easy to find the projection with range M v N. bounded by a (i. polarization justifies the conclusion that An~ A (weakly). and the spectral theorem is based on that result. Is a bounded increasing sequence of Hermitian operators necessarily strongly convergent? uniformly convergent? 121 121. Problem 122. the projections onto various geometric constructs formed with M and N. however. for instance. In the absence of such special assumptions. Does the same conclusion follow with respect to the strong and the uniform topologies? Problem 120. If limn Bn = B.e.e. Problem 121. 2. the problems become more interesting. f) ~ allfll 2 for all n and all f). and if a sequence =0 and Bn+ 1 = -!{(1 . Things are likely to be easy if E and F commute. Square roots. If !/!if) = (Anf. Infimum of two projections. If E and Fare projections with ranges M and N.. suppose that {An} is an increasing sequence of Hermitian operators (i. f) ~ (An+ J.

and closed under multiplication by arbitrary scalars). containing 1. then.e.e. a collection closed under addition and multiplication. 65 . closed under adjunction). A von Neumann algebra is an algebra of operators (i. The most obvious way to make it precise is to describe certain classes of operators by the requirement that they be closed under some familiar algebraic and topological operations. and strongly closed (i. then it contains E A F. self-adjoint (i. p. Although most mathematicians would read the statement of such a problem with sympathetic understanding.. the problem is this: prove that if a von Neumann algebra contains two projections E and F. vol. 2.. and then try to prove that whenever E and F belong to such a class. it must be admitted that rigorously speaking it does not really mean anything. 55]. The most famous and useful classes pertinent here are the von Neumann algebras (called "rings of operators" by von Neumann). Reference: [150.STRONG OPERA TOR TOPOLOGY The problem is to find an "expression" for the projection described. closed with respect to the strong operator topology).e. then so does E A F. For von Neumann algebras..

the reader may choose one of 66 . the assumption of a-finiteness in the treatment of multiplications is a matter of convenience. Sometimes bounded operator representations of function algebras play the central role.) The general version. and powerful statement that does not attain maximal generality (it applies to only one operator at a time. Spectral mapping theorem for normal operators. (Every finite normal matrix is unitarily equivalent to a diagonal one. not to algebras of operators). does not have a universally accepted formulation. [40. see [56]. In the second place. There is a short. applicable to infinite-dimensional spaces. pp. That statement will be called the spectral theorem in what follows. 911-912]. and that has the advantage of being a straightforward generalization of the familiar statement about diagonal forms. for two reasons.CHAPTER 15 Partial Isometries 123 123. The multiplication version of the spectral theorem has a technical drawback: the measures that it uses may fail to be a-finite. This is not a tragedy. In the sequel when reference is made to the spectral theorem. non-a-finite measures need to be considered only when the underlying Hilbert space is not separable. but that does have all classical formulations of the spectral theorem as easy corollaries. In the first place. it says that every normal operator is unitarily equivalent to a multiplication. the most important statement about them is the spectral theorem. the pathology of measures accompanies the pathology of operators. and sometimes Stieltjes integrals with unorthodox multiplicative properties. Normal operators constitute the most important tractable class of operators known. not of necessity (see [131]). The statement can be proved by exactly the same techniques as are usually needed for the spectral theorem. Students of operator theory generally agree that the finite-dimensional version of the spectral theorem has to do with diagonal forms. simple.

In order to be sure that this makes sense. Often a long but "elementary" proof gives more insight. then cp 2 = cp. By definition. it is the fate of many big theorems to be more honored in evasion than in use. A non-spectral treatment of a related property of square roots is in Problem 124. for example.e.PARTIAL ISOMETRIES two courses: treat the general case and proceed with the caution it requires. it follows that the complement of spec A is covered by a countable collection of neighborhoods with that property. every point in the complement of spec A has a neighborhood whose inverse image under cp has measure zero. and write B =A . the operator F(A) is then the multiplication induced by the composite function Fa cp. then the range of cp comes to be included in its essential range. The spectral theorem makes possible a clear and efficient description of the so-called functional calculus. Powerful general theorems exist to be used. This implies that IIBII = r(B) = 0 and hence that B = 0.) Compare this with the proof via the spectral theorem: if cp is a function whose range consists of the two numbers 0 and 1. i. p. with multiplier cp. or restrict attention to the separable case and proceed with the ease that the loss of generality permits. and their willful avoidance can lose insight at least as often as gain it. the approximation trickery of Problem 121 is fun. on a measure space X. and hence that the inverse image of the entire complement of spec A has measure zero. 55]. without using the spectral theorem. and has its uses. For another example. The reason is not just mathematical mischievousness. with spectrum included in the real line) is Hermitian. If A is a normal operator and ifF is a bounded Borel measurable function on spec A. the spectral theorem yields an immediate and perspicuous proof that every positive operator has a positive square root (because every positive measurable function has one). see [50. at worst. try to prove. For a final example. This sort of ritual circumlocution is common to many parts of mathematics. Thus. To define F(A) represent A as a multiplication.. spec B = {0}. Since the plane is a Lindelof space. a set of measure zero. and.e. that if the domain of cp is altered by. This is not to say that use of the spectral theorem is to be avoided at all costs.. than a short proof whose brevity is made possible by a powerful but overly specialized tool. consider the assertion that a Hermitian operator whose spectrum consists of the two numbers 0 and 1 is a projection. but for Hermitian operators it is completely elementary. The proof goes as follows. To prove it. In some contexts some authors choose to avoid a proof that uses the spectral theorem even if the alternative is longer and more involved.A 2 • Clearly B is Hermitian. (It is true for all normal operators that the norm is equal to the spectral radius. say. it is necessary to know that cp maps almost every point of X into spec A. then the functional calculus yields an operator F(A). let A be the operator. by the spectral mapping theorem. 67 . and leads to more fruitful generalizations. but it is not nearly so transparent. that every normal operator with a real spectrum (i.

for a counterexample. Decreasing squares. Problem 124. but the result is not likely to remain true. but continuous functions are sometimes easier to handle. does it follow that A is diagonal? 126 126. for instance.g. and if 1 + A + · · · + An is diagonal for some positive integer n.e. then A 2 ~ A. Polynomially diagonal operators. There is no hope for the statement spec F(A) = F(spec A). This is a typical use of the spectral theorem to prove an algebraic fact about Hermitian operators. Suppose. Continuity of the functional calculus. every function of it is normal. <p > I on a set of positive measure. if. it is very special in that it refers to normal operators only. it has good properties as a function of the first argument (the function).e. see. some people have found the possibility of getting along without the spectral theorem a little more than commonly elusive in this case. Problem 123. it follows. Proof: if <p is a measurable function and 0 ~ <p ~ I. for Hermitian operators. If A is a normal operator and ifF is a continuous function on spec A. The problem that follows is a spectral mapping theorem. Problem 123. but it is very general in that it allows all continuous functions. that F(A) = A* A ( = 1A12). and. 0 ~ A ~ I. if. Something like F(A) might seem to make sense sometimes even for nonnormal A's. If an operator A is diagonal (see Problem 61 ). then it is false that <p 2 ~ <p.. Is the converse true? (Note that if A is diagonal. then spec F(A) = F(spec A). A functional calculus is a mapping from functions and operators to operators. For a fixed value of the second argument (the operator). then F(A) = A*A. for every operator A. F*(A) = (F(A))*). for instance. and define F(A). then 0 ~ A ~ I. What properties can it have as a function of the second argument? 68 . then <p 2 ~ <p. If A is a normal operator on a separable Hilbert space. consequently. characteristic functions of Borel sets are of importance). that if F(A) = IAI 2. for instance.. then it is normal.. i. as A* A. Its principal property is that it is an algebraic homomorphism that preserves conjugation also (i. then all the powers of A are diagonal.) Problem 125. 125 125. 0 ~A ~ 1 if and only if A 2 ~A. conversely. The functions F that occur in the applications of the functional calculus are not always continuous (e. and so is any sum of such powers. 124 124.PROBLEMS The mapping F ~ F(A) has many pleasant properties. If A is a positive contraction. contemplate the unilateral shift. Prove without using the spectral theorem that. A is a Hermitian operator such that A 2 ~ A. conversely.

IfF is a continuous function on the complex plane. since 11/11 2 = 11911 2 + llhll 2 . defined for all normal operators. the range is sometimes called the final space. g). An isometry is a distance-preserving transformation: II Uf . ker U. An isometry is a linear transformation U (from a 127 Hilbert space into itself. It is sometimes convenient to consider linear transformations U that act isometrically on a subset (usually a linear manifold. Assertion: the mapping is norm continuous. Partial isometries. it is easy to verify that a partial isometry U is bounded.) For partial isometries. A partial isometry is a linear transformation that is isometric on the orthogonal complement of its kernel. (3) (U*Uf. in fact if U is not 0. then 11/11 = II Ufll = II Ug + Uhll = IIUhll = llhll. (To pass from (2) to (3). fix a polynomial p and consider the mapping p(X). continuous? 127. and if An is the scalar 1/n. The definition of partial isometries is deceptively simple. and these examples continue the deception.f) = (f. (What needs proof is that if II Ufll = llfll. The initial space of a partial isometry U turns out to be equal to the set of all those vectors f for which II Ufll = 11!11. Write f = g + h. but F(An) ( = 1) does not converge to F(O) ( =0).) The range of a partial isometry is equal to the image of the initial space and is necessarily closed.PARTIAL ISOMETRIES For a special example. A necessary and sufficient condition that a linear transformation U be an isometry is that U*U = 1. it follows that g = 0. they are the isometries (and. Proof: obvious. the image is a complete metric space. g)= (f. but not necessarily a subspace) of a Hilbert space. then An -+ 0.) Caution: the conditions U* U = 1 and UU* = 1 are not equivalent.f). with g E ker U and h l. The orthogonal complement of the kernel of a partial isometry is frequently called its initial space. or from one Hilbert space into another) such that II Ufll = 11!11 for all f. There are two large classes of examples of partial isometries that are in a sense opposite extreme cases. in particular.Ugll = II! -gil for all f and g. and the projections. then II U II = 1. and (4) U*U = 1 are mutually equivalent. for instance. the unitary operators). polarize. Does the functional calculus for normal operators remain continuous in this sense if polynomials are replaced by more general functions? The answer is sometimes no. What is the source of the trouble? Is it just that the function F in this example is not continuous? X~---+ Problem 126. The latter condition is satisfied in case U* is an isometry. (Since U is isometric on the initial space. 69 . F is the characteristic function of (0. this just means that I!Ufll = II! II for all fin that subset. Indeed: the conditions (1) IIUfll 2 = 11!11 2 . (2) (U*Uj. If. since addition and multiplication are norm continuous. In any case. is the mapping A~---+ F(A). defined for all operators X. oo) (considered as a subset of the complex plane). the structure of partial isometries can be quite complicated. however. ker U. in that case U is called a co-isometry. then fl.

It is natural to define a (partial) order for partial isometries as follows: if U and V are partial isometries. is the one according to which A ~ B if and only if (Af. but it cannot be concluded that U and V necessarily agree on the smaller initial space. This implies that if UU* = I. again. then a necessary and sufficient condition that M reduce U is that UM = M. The easy way to be a maximal partial isometry is to be unitary..) It follows that if U ~ V with respect to the present order. If U is unitary on H and if M is a subspace of H. then.e. and it can happen that M reduces U but UM ¥= M.PROBLEMS Problem 127. f) ~ (Bf. usually considered for Hermitian operators only. if U is an isometry. if all that is known about the partial isometries U and Vis that U* U ~ V* V. What if U is a maximal partial isometry? 70 . moreover. A bounded linear transformation U is a partial isometry if and only if U = UU*U. then U* U ~ V* V with respect to the usual order for operators. Are there any other maximal partial isometries? One way to get the answer is to observe that if U ~ V. then the initial space of U is the range of U* U. A bounded linear transformation U is a partial isometry if and only if U*U is a projection. This implies that the initial space of U is included in the initial space of V. V* agrees with U* on the initial space of U*. the initial space of U is included in the initial space of V. i. to be sure. i.) The converse is not true. UU* ~ VV*. if U is a co-isometry. of course. the initial space of U*) is included in the final space of V (the initial space of V*). (Cf. and. Corollary 2. Corollary 1. in particular. then the only partial isometry that can dominate U is U itself: an isometry is a maximal partial isometry.. and hence. If U*U = 1. the characterization of initial spaces given in Problem 127. Conclusion: a partial isometry is maximal if and only if either it or its adjoint is an isometry. 128 128.e.. In other words. In that case it is easy to enlarge U to a partial isometry that maps a prescribed unit vector in ker U onto a prescribed unit vector in ker U* (and. Corollary 3. then the final space of U (i.e. The adjoint of a partial isometry is a partial isometry. then U* ~ V*. If U is merely a partial isometry. Maximal partial isometries. Note that U* U ~ V* V in this sense is equivalent to II Ufll ~ IIVfll for all f. then it can happen that UM = M but M does not reduce U. f) for all f. (The "usual" order. write U ~ V in case V agrees with U on the initial space of U.L U). agrees with U on ker. then. If U is a partial isometry. U is maximal. with initial space and final space interchanged. then both U and U* have non-zero kernels. If a partial isometry U is neither an isometry nor a co-isometry. if U ~ V.

p'(U) = dim ran. 129 130. Some statements about partial isometries are slightly awkward just because 0 must be counted as one of them. and the nullity of U. It follows that v(U*) = p'(U) and p'(U*) = v(U). called the rank. p'. then there exist partial isometries with rank p. and v let P(p. and vice versa. 129. then so is U*. the co-rank. the same co-rank. 130 For each fixed p. and v(U) = v(V). and the same nullity.PARTIAL ISOMETRIES Problem 128. they are such that both p + p' and p + v are equal to the dimension of the underlying Hilbert space. and v are useful is that they are continuous.L U. For this reason. v) constitute a partition of the space P of all partial isometries. With this explanation the meaning of the continuity assertion becomes unambiguous: if U is sufficiently near to V. To interpret this statement. it is a consequence of the statement of Problem 130 that each set 71 .L U. One reason that the functions p. v) be the set of partial isometries (on a fixed Hilbert space) with rank p. Clearly the sets of the form P(p. (Symmetry demands the consideration ofv'(U) =dim ker. Closure and connectedness of partial isometries. If U and V are partial isometries such that II U .) Recall that if U is a partial isometry. since U is isometric on ker.) It is easy to see that if p. the co-nullity of U. respectively. p'. p'. for instance.) These three cardinal numbers. are not completely independent of one another. then U and V have the same rank. If U is a partial isometry. and nullity v. and v are any three cardinal numbers such that p + p' = p + v. co-rank p'. and nullity v. and v(U) = dim ker U. it is the only partial isometry in the interior of the unit ball. but there is no point in it. (That U is a partial isometry is not really important in these definitions. (Caution: subtraction of infinite cardinal numbers is slippery. p'. Problem 130. the set of all partial isometries is obviously not connected.VII < 1. The following assertion is a precise quantitative formulation of the result. p'(U) = p'(V). The operator 0 is an isolated point of the set of partial isometries. co-rank. co-rank p'.L U it follows that v' = p. What about the partial isometries on the boundary of the unit ball? Problem 129. The set of all non-zero partial isometries is closed but not connected (with respect to the norm topology of operators). use the norm topology for the space P of partial isometries (on a fixed Hilbert space). write p(U) = dim ran U. the initial space of U* is the final space of U. Discover the implication relations between the statements "UM = M" and "M reduces U" when U is a maximal partial isometry. and nullity. Rank. and use the discrete topology for cardinal numbers. then p(U) = p(V). p'. it does not follow that p' = v. similar definitions can be made for arbitrary operators.

it is a component of the set P of all partial isometries. Question: what are the other components? Answer: the sets of the form P(p. If U1 = eitA.AA * is positive. then M(A) and M(B) are unitarily equivalent. then there exists a measurable real-valued function () on that space such that <p = e. 0) (on a Hilbert space of dimension p) is connected. then there exists a Hermitian operator A such that U = eiA. 0 ~ t ~ 1. Consequence: every contraction can be extended to a partial isometry. It follows that there exists a unique positive operator whose square is 1 . The reason is that the exponential function has a Borel measurable inverse (in fact many of them) on the complement of the origin in the complex plane. co-rank.AA *. If A and B are unitarily equivalent contractions. If cp is a measurable function on a measure space. v) is both open and closed. the open-closed set P(p. p'. 0. This is easy to prove.PROBLEMS P(p. the result of the preceding paragraph can be expressed as follows: if U is a unitary operator. Problem 131 Each pair of partial isometries (on the same Hilbert space) with the same rank.) In the language of the functional calculus. and so is the set of all unitary operators = v = 0). Proof (via Problem 127): check that MM*M = M. Components of the space of partial isometries. call it A'. co-rank.o almost everywhere. and extend it by requiring one-sided continuity on. There are many ways that a possibly "bad" operator A can be used to manufacture a "good" one. and conversely. (Choose a continuous logarithm on the complement of the negative real axis. Unitary equivalence for partial isometries. 131 131. Samples: A+ A* and 72 . and nullity. and nullity. If A is a contraction (that means IIA II ~ 1). say. It follows that the set of all isometries (p' (v = 0) is both open and closed. In the notation of Problem 130. What it essentially says is that a measurable function always has a measurable logarithm. Conclusion: the set of all unitary operators is arcwise connected. Problem 132. 132 132. then t ~ V 1 is a continuous curve of unitary operators joining 1 ( = U 0 ) to U ( = U1). the upper half plane. can be joined by a continuous curve of partial isometries with the same rank. such that Icp I = 1 almost everywhere. p'. Assertion: the operator matrix M(A) = (~ ~') is a partial isometry. then 1 . v).

e. and. The chief interest of the assertion of Problem 132 is that. Problem 63). usually false that if the constructs are unitarily equivalent. for the special partial isometry construct it deals with. What conditions must a set of complex numbers satisfy in order that it be the spectrum of some non-unitary partial isometry? 73 .PARTIAL ISO ME TRIES None ofthese ways yields sufficiently many usable unitary invariants for A. If the spectrum of a partial isometry does not contain the origin. however. It is usually easy to prove that if A and Bare unitarily equivalent. Since every non-empty compact subset of the unit circle is the spectrum of some unitary operator (cf. and translations and changes of scale do not affect unitary equivalence. therefore. its spectrum is necessarily a subset of the closed unit disc. its spectrum is a subset of the unit circle (perimeter). What about the non-invertible ones? Problem 133. the converse happens to be true. then the original operators themselves are. The unitary equivalence problem for invertible contractions is. What conditions must a set of complex 133 numbers satisfy in order that it be the spectrum of some partial isometry? Since a partial isometry is a contraction.. It is. The end product of all this is a reduction of the general unitary equivalence problem to the special case of partial isometries. then so are the various constructs in which they appear. the problem of characterizing the spectra of invertible partial isometries is solved. in turn. 133. trivially equivalent to the unitary equivalence problem for arbitrary operators. Spectrum of a partial isometry. then it is unitary. The reason is that by a translation (A t-t A + IX) and a change of scale (A t-t (JA) every operator becomes an invertible contraction. The result is that the unitary equivalence problem for an apparently very small class of operators (partial isometries) is equivalent to the problem for the much larger class of invertible contractions. i. if a partial isometry is invertible.

Every complex number is the product of a nonnegative number and a number of modulus 1. If the given matrix is invertible. It is possible to get a satisfactory uniqueness theorem for every matrix. The corresponding left-handed theory (A = PU) follows by a systematic exploitation of adjoints. In the infinite-dimensional case. this polar decomposition is unique. then. more accurately. and this additional condition uniquely determines them.) The correct statements are just as easy for transformations between different spaces as for operators on one space. then there exists a partial isometry U (from H to K) and there exists a positive operator P (on H) such that A = UP The transformations U and P can be found so that ker U = ker P. 74 . and if the order of the factors is specified (UP or PU). once again. this polar decomposition is unique. The generalization to finite matrices says that every complex matrix is the product of a positive matrix and a unitary one. and it does not become true even if U is required to be merely isometric. this is a point at which partial isometries can profitably enter the study of finite-dimensional vector spaces. with U unitary and P positive. Polar decomposition. Problem 134.CHAPTER 16 Polar Decomposition 134 134. (The construction of concrete counterexamples may not be obvious now. but it will soon be an easy by-product of the general theory. except for the number 0. or. The representation of A as the product of the unique U and P satisfying the stated conditions is called the polar decomposition of A. the right-handed polar decomposition of A. !fA is a bounded linear transformation from a Hilbert space H to a Hilbert space K. but only at the expense of changing the kind of factors admitted. partial isometries are unavoidable. It is not true that every operator on a Hilbert space is equal to a product UP.

g) defines an inner product (not necessarily strictly positive). 136. if any. It is clear that every normal operator is quasinormal. and a necessary and sufficient condition that U be a co-isometry is that the range of A be dense.g).JA*A. A possible guess is to replace A on the right (majorant) side by . but since the equation (. Mixed Schwarz inequality. Extreme points. Quasinormal operators. = UP is the polar decomposition of A. it is of interest to determine the extreme points. If. 138. An operator A is called quasinormal if A commutes with A* A. probably no). The converse is obviously false.POLAR DECOMPOSITION Corollary 1. Problem 135. it is not clear what. 138 (To feel comfortable about this relation it helps to recall that the positiveness of A implies (jA) 2 = A.) If A is not positive. positive and symmetric.) Problem 137.jA) 2 =A is not available.g)i 2 ~ (Af. then A* A = 1 and therefore A commutes with A* A. An operator with polar decomposition UP is quasinormal 137 if and only if UP= PU. A is an isometry. If A is a positive operator. then U* A = P. then the inner product expression (Af. g) still defines something (sesquilinear. the most elementary of these leads to the concept of quasinormality. What are the extreme points of the closed unit ball in the space of operators on a Hilbert space? 136 137. Quasinormal operators (under another name) were first introduced and studied in [18]. Maximal polar representation. Every bounded linear transformation is the product of a maximal partial isometry and a positive operator. 75 .f)·(Ag. for instance. The closed unit ball in the space of operators is convex. yes. For every interesting convex set. If A = UP is the polar decomposition of A. it follows that A satisfies the Schwarz-like inequality i(Af. If A Corollary 2. then A is not normal. Problem 136. then a necessary and sufficient condition that U be an isometry is that A be oneto-one. then (Af. The condition of normality can be weakened in various ways. but if A is not unitary. 135 135. (For a concrete example consider the unilateral shift. Schwarz-like inequality prevails.

Quasinormal weighted shifts. the space is infinite-dimensional and the operator is not normal. It sometimes happens that a theorem is easy to prove for invertible operators but elusive in the general case. g)? What if jAA* is used in place of fl*"A? if they are both used. Problem 139. There is a tempting generalization nearby that is frequently a rich source of illuminating examples. then there is trouble. Density of invertible operators. and. The set of all invertible operators is connected. Problem 141.e. What about the set of left-invertible operators? 141 141. In the infinite-dimensional case the approximation technique works.PROBLEMS JA*A? Why not JAA*? What (Reason: if z is a complex number. Which weighted shifts are quasinormal? 140 140. f)· (jA*Ag. The unilateral shift is an example of a quasinormal operator that is not normal. The set of all operators that have either a left or a right inverse is dense. for normal operators. This makes it useful to know that every finite (square) matrix is the limit of invertible matrices. however. by changing the small values of the multiplier. 76 . one in each factor? 139 139. but the set ofall operators that have both a left and a right inverse (i. Is it true for every operator A that i(Af. gW ~ (JA*Af. Problem 140. Connectedness of invertible operators. with no difficulty. the set of all invertible operators) is not.) If. approximate it by operators that are bounded from below. (Invoke the spectral theorem to represent the given operator as a multiplication. then ~ = izi.) Objection: why Problem 138..

The unilateral shift." involutory". Unilateral shift versus normal operators.. Reducing subspaces of normal operators. Products of symmetries. Problem 143. The main point of Problem 142 143 144 142 is to help solve Problem 143 (and. 77 . The main point of Problem 143 is to emphasize the role of certain shift operators. to provide a non-trivial application of the spectral theorem). A symmetry is a unitary involution. 143. Problem 142. both algebraic and analytic. then it possesses the third. and "Hermitian". The following assertion is one way to say that the spectral theorem provides many reducing subspaces. The techniques for discovering and proving these properties are frequently valuable even when the properties themselves have no visible immediate application. Discuss the assertion: every unitary operator is the product of a finite number of symmetries. Here are three sample questions. It may be pertinent to recall that if an operator possesses any two of the properties" unitary". One of the principal achievements of the spectral theorem is to reduce the study of a normal operator to subspaces with various desirable properties.e. If A is a normal operator on an infinite-dimensional Hilbert space H. an operator Q such that Q*Q = QQ* = Q 2 = 1.CHAPTER 17 Unilateral Shift 142. the proof is completely elementary algebraic manipulation. i. all with the same infinite dimension. has many curious properties. then H is the direct sum of a countably infinite collection of subs paces that reduce A. 144. incidentally. Shifts (including the simple unilateral and bilateral ones introduced before) are a basic tool in operator theory. in particular.

and L 2 is formed with normalized Lebesgue measure.finite number of normal operators? (b) What is the norm of the real part of the unilateral shift? (c) How far is the unilateral shift from the set of normal operators? The last question takes seriously the informal question:" How far does the unilateral shift miss being normal?" The question can be asked for every operator and the answer is a unitary invariant that may occasionally be useful. In some special cases the determination of the commutant is accessible by relatively elementary methods. (a) Is the unilateral shift the product of a . if U is the unilateral shift. 147 147. Here en(z) = zn (n = 0. Problem 145. The bilateral shift W can be viewed as multiplication bye 1 on L 2 of the unit circle (cf. If the bilateral shift is regarded as a multiplication. The commutant of an operator (or of a set of operators) is the set of all operators that commute with it (or with each operator in the set). One of the most important purposes of the so-called multiplicity theory is to discuss the commutants of normal operators. The commutant is one of the most useful things to know about an operator. On second thought 78 . The generalizations are obtained by replacing the unit circle by an arbitrary bounded Borel set X in the complex plane and replacing Lebesgue measure by an arbitrary finite Borel measure in X. ±2. Does the unilateral shift have a square root? In other words. Problem 146. The wording suggests a superficially plausible conjecture: perhaps the commutant of the unilateral shift consists of the restrictions to H 2 of all multiplications. The generalization of the bilateral shift is the multiplication induced by e 1 (where e 1(z) = z for all z in X). Problem 84). Square root of shift. 145 145. The unilateral shift is the restriction of the bilateral shift to H 2 . Corollary. a case in point is the bilateral shift. ± 1. Commutant of the unilateral shift. The commutant of the bilateral shift is the set of all multiplications. then its commutant can be described as the set of all multiplications on the same L 2 (Problem 146). Commutant of the bilateral shift.PROBLEMS Problem 144. · · ·) whenever lzl = 1. does there exist an operator V such that V 2 = U? 146 146. Both the main statement and the corollary have natural generalizations that can be bought at the same price. Each reducing subspace of the bilateral shift is determined by a Borel subset M of the circle as the set of all functions (in L 2 ) that vanish outside M.

Commutant of the unilateral shift as limit.. Problem 34). 2. and corresponding insight into. the contradiction has arrived. and.e. the multiplier itself is in H 2 (and hence in H 00 ). where en(z) = zn for all z in X. 148. n = 0. however. the main statement has a natural generalization. a mixture of the two kinds is possible. if V 2 = U. then H 2 is invariant under the induced multiplication (cf. sometimes denoted by H 2 (J1). The unilateral shift is irreducible. 148 149. an example of the latter kind is the unilateral shift. The trouble is that the structure ofH 2 (J1) within L 2 (J1) depends strongly on X and 11. Every operator that commutes with the unilateral shift is the limit (strong operator topology) of a sequence of polynomials in the unilateral shift. It follows 149 79 . and the conjecture makes sense. Indeed. Characterization of isometries. and infer that (<fJ(z)) 2 = z almost everywhere. The characterization of the commutant of the unilateral shift yields a curious alternative proof of. More precisely. the restriction of a multiplication to H 2 is not necessarily an operator on H 2 . i. Just as for the bilateral shift. · · · . 1. The generalization of the unilateral shift is the restriction to H 2 (J1) of the multiplication induced by e 1 • The corollary does not generalize so smoothly as the main statement. Problem 147. If. happen that H 2 (J1) = L 2 (J1). The commutant of the unilateral shift is the set of all restrictions to H 2 of multiplications by multipliers in H 00 • Corollary. What can an isometry look like? Some isometries are unitary. and therefore Vis the restriction to H 2 of the multiplication induced by a function <fJ in H 00 • Apply V 2 to e0 . for instance. is the span in L 2 (J1) of the functions en. that the function e1 has an analytic square root. (There is no point in forming direct sums of unitary operators-they are no more unitary than the summands. and replace Lebesgue measure by an arbitrary finite Borel measure J1 in X. Since a direct sum (finite or infinite) of isometries is an isometry.) The useful theorem along these lines is that that is the only way to get isometries. it can. The generalization of H 2 . Replace the unit circle by an arbitrary bounded Borel subset X of the complex plane. Problem 148. consequently. apply U to e0 .UNILATERAL SHIFT this is absurd: H 2 need not be invariant under a multiplication. then V commutes with U. This implies that (<P(z)) 2 = z in the unit disc (see Solution 42). and some are not. the direct sum of a unitary operator and a number of copies (finite or infinite) of the unilateral shift is an isometry. in the sense that its only reducing subspaces are 0 and H 2 . the assertion that U has no square root (Solution 145).

but make no demands on the value of the co-rank. They play a vital role in the study of all operators. or. U 2 N. U 2 e0 . Which unilateral shifts have square roots? 80 . Consider an isometry U on a Hilbert space H such that the subspaces (UH)\ U(UH). Unilateral shifts of higher multiplicities are just as important as the simple one. Square roots of shifts. . In view of this comment another slight reformulation is possible: the unilateral shift is an isometry U of co-rank I on a Hilbert space H such that the subspaces ( UH). and if there exists a unit vector e0 in H such that the vectors e0 . or a direct sum of one or more copies of the unilateral shift. ···are pairwise orthogonal and span H. Problem 149 shows that they are exactly the pure isometries.) Most of these remarks are implicit in Solution 149.L. Ue 0 . Distance from shift to unitary operators. then it must be equal to the co-range (UH). The co-rank of a shift (also called its multiplicity) constitutes a complete set of unitary invariants for it. 150 150. is a puzzle that makes contact with Problem 145 and with a curious infinite-dimensional manifestation of Sylvester's law of nullity. · · · form an orthonormal basis for H. An isometry for which the unitary direct summand is absent is called pure. A generalization lies near at hand. U is the unilateral shift. the original unilateral shift is determined (to within unitary equivalence) as the shift of multiplicity 1 (the simple unilateral shift).. If there is such a subspace N. This characterization of the unilateral shift can be reformulated as follows: U is an isometry on a Hilbert space H for which there exists a onedimensional subspace N such that the subspaces N. U 2 (UH). Here. (Since U is an isometry. Problem 150. How far is the unilateral shift from the set of unitary operators? 151 151. Problem 149. Every such isometry rna y be called a shift (a unilateral shift). or a direct sum of a unitary operator and some copies of the unilateral shift. it is in fact one of the fundamental building blocks out of which all isometries are constructed. Every isometry is either unitary..PROBLEMS that the unilateral shift is more than just an example of an isometry. with interesting and peculiar properties.L. by a slight abuse of language.L. it follows that they must be pairwise orthogonal.L. U( UH). If U is an isometry on a Hilbert space H. U 2 (UH). Problem 151. to begin their study.L.L. UN. ···span H. then (obviously) U is unitarily equivalent to the unilateral shift. are pairwise orthogonal and span H. not only isometries.

a measure of size that similarity does preserve.1 AS. There is a condition on the spectral radius alone that is sufficient for similarity to a part of the adjoint of a shift. 81 152 153 . Corollary 2. moreover. turn out to be universal operators. r(A) < 1. then sn ---t 0 strongly. it follows that r(A) ~ 1. can be obtained by forming the n-th power of U. where n is a positive integer. (That is: if An ~ 0 strongly. and it follows that if A is a part of U*. Problem 152) is necessary for similarity also. cf. A part of an operator is a restriction of it to an invariant subspace. shifts of multiplicity n. Similarity to parts of shifts. Corollary 1. (More generally. What about parts of the adjoints of unilateral shifts? If U is a unilateral shift. it follows that An ~ 0 (strong). 153. Every operator whose spectrum is included in the interior of the unit disc is similar to a contraction whose powers tend strongly to 0. When is an operator A similar to a part of the adjoint of a shift U? Since similarity need not preserve norm. Solution 110). The study of shifts of infinite multiplicity is much harder and much more important than that of the finite ones.) Since there are many operators A such that r(A) ~ I but An does not tend to 0 in any sense (example: 1). namely the spectral radius. in fact. Every contraction whose powers tend strongly to 0 is unitarily equivalent to a part of the adjoint of a unilateral shift. then IIA II ~ 1. however.) This remark shows (see Problem 145) that the answer to Problem 151 is neither "all" nor "none". or rather their adjoints. since r(U*) = 1. 152. Every operator whose spectrum is included in the interior of the unit disc is similar to a strict contraction. [ 43] and [32. Every operator whose spectrum is included in the interior of the unit disc is similar to a part of the adjoint of a unilateral shift. The miraculous and useful fact is that these two obviously necessary conditions are also sufficient. it is. Problem 153. they cannot be obtained from U in such a simple manner. Since. There is. the study of the parts of unilateral shifts does not promise anything new. but it is quite a bit stronger than r(A) ~ 1 . Problem 152. then II U II = II U* II = 1. then it is quite easy to see that U 2 is a shift of multiplicity 2. The most important aspect of shifts is that they. the condition on the spectral radius is obviously not sufficient. For many purposes similarity is just as good as unitary equivalence. It is easy to see that this necessary condition is not sufficient. cf. Shifts as universal operators. Each part of an isometry is an isometry. 33].UNILATERAL SHIFT If U is the simple unilateral shift. and if B = s. there is no obvious condition that II A II must satisfy. U*n ~ 0 in the strong topology (cf. The reason is that one of the necessary conditions for unitary equivalence (An ~ 0 strongly.

but a useful and non-trivial one..PROBLEMS (A strict contraction is an operator A with II All < 1. namely the unilateral shift.1 AS I for all invertible operators S. in both hypothesis and conclusion. then there exists a (necessarily unique) onedimensional wandering subspace N such that M = V':= 0 U"N.) Corollary 3. the following concise statement is just a reformulation of 82 . In other words. iff and g are inN.) For at least one operator. The spectral radius of every operator A is the infimum of the numbers IIS. then umf j_ U"g. If A is an operator on a Hilbert space H. positive. The equation connecting M and N can be expressed by saying that every non-zero part of the simple unilateral shift is a shift. observe that UM = V:. (To prove that this correspondence is one-to-one. 154 154. even more is true: in that case umN j_ U"N whenever m and n are any two distinct integers. Wandering subspaces. negative. Problem 155. in this sense: if U is an isometry. This concept is especially useful in the study of isometries. so that N = M n (UM). U"g) = (U*"Umf. g). un+kg).) Wandering subspaces are important because they are connected with invariant subspaces. Similarity to contractions. If r(A) ~ 1. One of the most difficult open problems of operator theory is to determine exactly which operators are similar to contractions [59].) If U is unitary. are replaced by the corresponding weak inequalities? Problem 154. These simple but beautiful and general results are in [119]. If U is an isomt:try and N is a wandering subspace for U.L.o U"N. and note that (Umf. To add that dim N = 1 is perhaps an unsurprising sharpening. Every quasinilpotent operator is similar to operators with arbitrarily small norms. If U is the (simple) unilateral shift and ifM is a non-zero subspace invariant under U. a subspace N of H is called wandering for A if it is orthogonal to all its images under the positive powers of A. then umN j_ U"N whenever m and n are distinct positive integers. or zero. (Proof: reduce to the case m > n. Corollary 4. is A similar to a contraction? 155 155. the correspondence is invertible. g) = (um-nf.0= 1 U"N. What happens to that result when the strict inequalities in it. In view of these comments. U"g) = (Um+kf. The correspondence is given by setting M = V:. then there is a natural one-to-one correspondence between all wandering subspaces N and some invariant subspaces M. Corollary 2 of Problem 153 gives a sufficient condition for similarity to a strict contraction. (Proof: find k so that m + k and n + k are positive and note that (Umf.

157 83 . The lattice operations (intersection and span) applied to them yield some not particularly startling new examples. · · ·. A non-zero subspace M of H 2 is invariant under U if and only if there exists a function cp in H 00 .L(A. 3. and consider the span Mk of the en's with n . and V'f:= 1 M/(A. · · ·). A good concrete special case to look at is the unilateral shift. and the other kind. After this elementary observation most students of the subject must stop and think.) is invariant under U. Note that the spaces Mk considered above are the same as the spaces Mk(O)..) = k. k. such that M is the range of the restriction to H 2 of the multiplication induced by cp. With almost no additional effort.1f.)is theorthogonalcomplementof {/. interesting. it follows that the orthogonal complement of the singleton {f. 156 157.. = I:'=o A. The easiest way to obtain an invariant subspace of the unilateral shift U is to fix a positive integer k. since each complex number A. of modulus less than I is a simple eigenvalue of U* (Solution 82). uk. and then the well seems to run dry.. 156. 2. regard U as the restriction to H 2 of the multiplication induced by e1 • Problem 157.. The unilateral shift has no reducing subspaces (Problem 147). Problem 156. then Mk(A. the question remains as to how many of the other kind it has and what they look like. it is not at all obvious that any other invariant subspaces exist. with corresponding eigenvector f. of constant modulus I almost everywhere. lfMk(A.) (see Problem 156) are examples..) = H 2 . Invariant subspaces of the shift. and profitable thing to do is to accumulate experimental evidence by examining concrete special cases and seeing what their invariant subspaces look like. and only the obviously appropriate changes in the statement.nen.} is a non-trivial subspace invariant under U.. dim Mk . New inspiration can be obtained by abandoning the sequential point of view and embracing the functional one. Special invariant subspaces of the shift.. Indeed. What are the invariant subspaces of the unilateral shift? The spaces Mk and their generalizations Mk(A. A recollection of the spectral behavior of U is helpful here. One of the most recalcitrant unsolved problems of Hilbert space theory is whether or not every operator has a non-trivial invariant subspace.UNILATERAL SHIFT the problem: every non-zero part of the simple unilateral shift is (unitarily equivalent to) the simple unilateral shift. all these considerations extend to shifts of higher multiplicities.. A promising. There are two kinds of invariant subspaces: the kind whose orthogonal complement is also invariant (the reducing subspaces).} (k = 1.

cf. Concisely: there are no zero-divisors in H 2 . If M and N are non-zero subspaces invariant under the unilateral shift. the functions such as q> (functions in H 00 . For no very compelling reason. see [75. then M n N i= 0. p. then f = 0 almost everywhere or g = 0 almost everywhere. In more informal language. The elements of H 2 are related to certain analytic functions on the unit disc (Problem 35). the characterization of the invariant subspaces of the unilateral shift does that. then <p and t/1 are constant multiples of one another. For a more general discussion of the F. 71]. It is always a pleasure to see a piece of current (soft) mathematics reach into the past to illuminate and simplify some of the work of the founding fathers on (hard) analysis. and M. then q> is divisible by t/J. of constant modulus 1) are called inner functions. If q> and t/1 are inner functions such that q> • H 2 c t/1 · H 2 .PROBLEMS This basic result is due to Beurling [13]. they tend to imitate the behavior of analytic functions. Corollary. that is). by constants of modulus 1. Corollary 2 says that the lattice of invariant subspaces of the unilateral shift is about as far as can be from being complemented. An important theorem of F.47]. 54. F. A function in H 2 vanishes either almost everywhere or almost nowhere. in the sense that there exists an inner function() such that <p = t/J. Problem 158. Correspondingly it is suggestive to write M = q>. M can be described as the set of all multiples of q> (multiples by functions in H 2 . A crucial property of an analytic function is that it cannot vanish very often without vanishing everywhere. and only almost everywhere at that. Here is a sample. Corollary 1. It has received considerable attention since then. 84 . here is one possible formulation. [89. Iff and g are in H 2 and if fg = 0 almost everywhere. but it does solve some. Riesz asserts that the elements ofH 2 exhibit the same kind of behavior. although they themselves are defined on the unit circle only. If <p • H 2 = t/J · H 2 . H 2 . Riesz theorem. 158 158. e. Corollary 2. and M. Riesz theorem. The characterization in terms of inner functions does not solve all problems about invariant subspaces of the shift. and. and M.

L. It is due to R. If A is a bilateral weighted shift with strictly positive weights an. Kelley. ± 1. Problem 159. however. 159 85 . There is. then a necessary and sufficient condition that A be reducible is that the sequence {an} be periodic. it has to do with the reducibility of two-sided weighted shifts. n = 0. · · ·. Reducible weighted shifts. ±2. Very little of the theory of reducing and invariant subspaces of the bilateral and the unilateral shift is known for weighted shifts. one striking fact that deserves mention.UNILATERAL SHIFT 159.

is dense in H. Some trace of the relation between the existence of cyclic vectors and multiplicity 1 is visible even for non-diagonal matrices. A. If. a trivial example is e0 • On finite-dimensional spaces the existence of a cyclic vector indicates something like multiplicity 1. A 2 . An operator A on a Hilbert space H has a cyclic vector f if the vectors f. at most n of the matrices 1. If. · · ·. 1). If the A.. then <~ 2 *. on the other hand.n)~n) for every polynomial p. If. For f = <~ 1 . 86 . · • ·. otherwise the i coordinate of p(A)f is 0 for all p. then p(A. to be precise. • · · span H. that the eigenvalues be simple). 2 . ~n) to be cyclic. then a necessary and sufficient condition that A have a cyclic vector is that the diagonal entries be distinct (i. . A 2{. then the p(A)j's exhaust the whole space.~ 1 *. for instance.b ···. ·' p(A. it is clearly necessary that ~i =I= 0 for each i.An. Equivalently. p(A. · · ·. the A's are distinct. Cyclic vectors. A is a finite diagonal matrix. if A is a finite matrix. Reason: by virtue of the Hamilton-Cayley equation. at most n of the vectors Aij EB Aig are linearly independent. so that if none of the ~·s vanishes.'s are not distinct. A. Af. The simple unilateral shift has many cyclic vectors.n) can be prescribed arbitrarily. and consequently.e. • • • are linearly independent (where n is the size of A). if the diagonal entries are A.. Thus. 1 = A.CHAPTER 18 Cyclic Vectors 160 160. then p(A)(~b · · ·' ~n) = <P(Al)~b . 0. for instance. it follows that their span can never be 2n-dimensi onal. f is a cyclic vector for A in case the set of all vectors of the form p(A)f. no matter what f and g are. then the direct sum A EB A cannot have a cyclic vector. where p varies over all polynomials. nothing is sufficient to make f cyclic. 0) is orthogonal to p(A)f for all p. Indeed.

then its co-rank is at most 1. then U* EB U* (which is the adjoint of a unilateral shift of multiplicity 2) has a cyclic vector. These facts were first noticed by D. that is one of the achievements of the multiplicity theory of normal operators ([50. then U* has a cyclic vector. then so does A*. «ex. For non-normal operators.) 87 . There is. 2. consider a candidate for a cyclic vector of U EB U. the close connection between multiplicity 1 and the existence of cyclic vectors persists in infinitedimensional spaces. surely. it is reasonable to expect that A* also has. The methods of the preceding paragraphs are very parochially finitedimensional. · · ·). It is possible for a direct sum A EB A to have a cyclic vector. in other words. For normal operators. in non-separable spaces there are no cyclic vectors. If (ex. If U is a unilateral shift of multiplicity not greater than t{ 0 . and recall that every matrix is similar to its transpose. this motivates the conjecture that if A has a cyclic vector. · · ·)) for all n ( = 0. E. if a matrix has a cyclic vector. · · ·). the co-rank of U EB U is 2. ({3. (IJo. 1. and therefore the space is separable. · · ·)) is orthogonal to (U EB Vt«eo. 0. that indicates that the theory of cyclic vectors in infinitedimensional spaces is likely to be refractory. e1. then a countable set spans the space.CYCLIC VECTORS If A has multiplicity 1 in any sense. (Here is a slightly more sophisticated way of expressing the same proof: if an operator has a cyclic vector. but it is not at all obvious that its adjoint has one. 0. suitably reinterpreted. and. The first strange consequence of the present assertion is that if U is the simple unilateral shift.and that proves that the cyclic candidate fails. Problem 160. and it is possible for A to have a cyclic vector when A* does not. but that by itself does not imply anything shocking. then so does its complex conjugate. ···). things are peculiar. note that. f/2. It does. It is obvious that the simple unilateral shift has a cyclic vector. If there is a cyclic vector. For finite matrices this is true. 0. the same is true for direct sums with more direct summands). To prove the negative assertion. {3) is an arbitrary non-zero vector orthogonal to IJo) in the usual two-dimensional complex inner product space. 0. 1J1. The promised strange behavior becomes completely exposed with the remark that U EB U cannot have a cyclic vector (and. Sarason. to begin with. and it is. even in spaces of uncountable dimension. then the vector <eo. all the more. This difficulty can be got around. a trivial difficulty with cardinal numbers. For a proof. III]). e2.

If dim H = ~0 . and hence is a "large" set in the sense of Baire category. Is U EB U* cyclic? 164 164. what about its complement? Even special cases of the question are not completely trivial. the dimension of the corresponding eigenspace. does it follow that f is cyclic for A*? 165 165. !. It is always good to know all the cyclic vectors of an operator. then the operators with n distinct eigenvalues constitute a dense set. . Problem 162. If the dimension is n. Density of non-cyclic operators.) Does normality improve matters? Problem 164. Iff is a cyclic vector for a normal operator A. (Example: C ~) and (1. it follows also that the set of cyclic 1 ~ 0 0) as operators is not closed. 0. (The relevant multiplicity is the geometric one. a necessary and sufficient condition for an operator to be cyclic is that every eigenvalue be of multiplicity 1.PROBLEMS 161 161. t. Is the unilateral shift a limit of non-cyclic operators? 163 163. it follows that the set of cyclic operators is dense. 0. Cyclic vectors of adjoints. Cyclic vectors of a position operator. 1/n. Density of cyclic operators.) If the underlying space is finite-dimensional. and U* EB U* are cyclic. and even if both an operator and its adjoint are cyclic. but U EB U is not.) It follows that in the finite-dimensional case the set of cyclic operators is open. ···. · · · ). but to find them all is rarely easy.) In the separable infinite-dimensional case the set of cyclic operators is not open. °) ( is the set of cyclic operators on H dense? 162 162. then it is a cyclic operator. U*. (Since not every operator is cyclic. Concrete example: the cyclic operator diag(l. Operators constructed out of the unilateral shift exhibit various cyclicity properties: U. Even if an operator is cyclic. Concrete example: ( /n 0 -1/n 0 0 n ~ oo. Problem 161. Problem 163. t. they don't necessarily have the same cyclic vectors. its adjoint does not have to be (remember U EB U). 0). At least one question remains. If dim H = ~ 0 . Cyclicity of a direct sum. ·) is the limit of the non-cyclic operators diag< 1. Even for 88 .. How large is the set of cyclic operators? (It is convenient to say that if an operator has a cyclic vector. 0. the set of cyclic operators is not dense.

for some operators.CYCLIC VECTORS one of the most natural operators on an infinite-dimensional space the problem leads to some interesting analysis. Is there an operator for which the span of the set of cyclic vectors is a non-trivial subspace (that is. 2. What are all the cyclic vectors of the position operator on L 2 (0. Problem 167.) is defined by (Af)(z) = zf(z). then it is cyclic. with no scalar multiples and no sums allowed. then the position operator A on L 2 (p. that are not cyclic are the scalars. rxij = 0 whenever i > J). Solution 165 describes examples for which the set of cyclic vectors is. To say that an operator A has a cyclic vector fis to say that finite linear combinations of the vectors A"f are dense. every non-zero vector is cyclic. A much stronger property is conceivable. upper triangular) if every entry below the main diagonal is zero (that is. 1. Totality of cyclic vectors. Extensions of this language (triangular + k) are obviously possible. 166. What's special about a cyclic matrix (meaning the matrix. it is not known whether that is actually possible. is a Borel measure with compact support in the complex plane. but are not needed now. Another conceivable answer is that. n = 0. Problem 165. 166 167 168 89 . Despite this bad news. quite large. Dense orbits. of a cyclic operator)? Nothing much: the only matrices of size 2. Is there an operator A on a Hilbert space Hand a vector fin H such that the orbit off under A is dense in H? The orbit off is the set of all vectors of the form A"f. The pertinent concept is a slight generalization of triangularity. (a) If an operator has a matrix that is triangular + 1 and is such that none of the entries on the diagonal just below the main one is zero (rxij ¥= 0 when i = j + 1). How many cyclic vectors does an operator have? One possible answer is none at all. Cyclic operators and matrices. matrices can be helpful in the study of cyclic operators. · · ·. for instance. A matrix <rxij) is triangular (specifically. 1)? Recall that if p. rxij = 0 whenever i > j + 1). different from both 0 and the whole space)? 167. (b) Is the converse true? 168. is it possible? Problem 168. Can it be medium-sized? Problem 166. with respect to some orthonormal basis. A matrix is triangular + 1 if every entry more than one step below the main diagonal is zero (that is. nevertheless.

(w --+ w). A linear transformation on a Hilbert space is called compact (also completely continuous) if its restriction to the unit ball is (w --+ s) continuous (see Problem 169). Corresponding to the strong (s) and weak (w) topologies for a Hilbert space H. there are four possible interpretations of continuity for a transformation from H into H: they are the ones suggested by the symbols (s --+ s). Thus. a linear transformation is compact if it maps each bounded weakly convergent net onto a strongly convergent net. it is fortunate that three of them collapse into one. equivalently it means that the direct image under A of a net s-convergent to f is a net w-convergent to Af Four different kinds of continuity would be too much of a good thing. Mixed continuity. The image of the closed unit ball under an operator on a Hilbert space is always strongly closed. and continuity (w --+ s) implies that A has finite rank. Equivalently. (s --+ w). Problem 169. 170 170. it follows 90 . Corollary. Compact operators. to say that A is continuous (s --+ w) means that the inverse image under A of each w-open set is s-open. For a linear transformation A the three kinds of continuity (s--+ s). and (s--+ w) are equivalent (and hence each is equivalent to boundedness). It is perhaps worth observing that for linear transformations of finite rank all four kinds of continuity are equivalent. this is a trivial finite-dimensional assertion. (w--+ w). Since weakly convergent sequences are bounded.CHAPTER 19 Properties of Compactness 169 169. and (w --+ s).

PROPERTIES OF COMPACTNESS that a compact linear transformation maps every weakly convergent sequence onto a strongly convergent one. 171. Reason: since f =I= 0 and 1/11 f I ~ 1. The set K of all compact operators on a Hilbert space is a closed self-adjoint (two-sided) ideal. here treated as consequences of the continuity conditions used above to define compact linear transformations. together with the assumption that that image is precompact. If A = 0.e. and "ideal" means that linear combinations of operators in K are in K and that products with at least one factor in K are in K. "self-adjoint" means that if A E K. the answer is yes for them. Since the unit ball is weakly compact.) The compactness conditions.. the reason is that the image of the unit ball is the unit ball. that every compact linear transformation is bounded. (The copverse just proved is not universally true for Banach spaces. this function attains its maximum.. For infinite-dimensional spaces. (This implies. (Proof: a bounded set is included in some closed ball. Diagonal compact operators. incidentally. then A* E K. then f can be chosen to have norm 1. and in an infinite91 171 . and the mapping g f--+ llg I is strongly continuous. p. so that I Af II = I A I for some f with I f I . this. then it maps the closed unit ball onto a compact set. if A =I= 0.) This implies that the image of each bounded set is precompact (i. IIAII11/11 - Problem 170.) The converse implication is also true: if a linear transformation maps bounded sets onto precompact sets. To prove this. Is the identity operator compact? Since in finite-dimensional spaces the strong and the weak topologies coincide. then f necessarily has norm 1. can in fact be shown to be equivalent to those continuity conditions and are frequently used to define compact linear transformations.) An occasionally useful property of compact operators is that they "attain their norm".::. observe first that compact (and precompact) sets are bounded. 484].) It follows from the corollary to Problem 169 that the image of the closed unit ball is strongly closed. has a strongly compact closure). 1. Precisely said: if A is compact. then there exists a unit vector f such that I A/II = I All· The reason is that the mapping f f--+ Af is (w--+ s) continuous on the unit ball. The image of the closed unit ball under a compact linear transformation is strongly compact. Here "closed" refers to the norm topology. IIAII -11/11 = IIA/11:::. it follows that IIAII:::. the answer is no. implies that that image is actually compact. (Proof: the closed unit ball is weakly compact. it follows that/ f--+ I A/II is weakly continuous on the unit ball. and that therefore a linear transformation that maps bounded sets onto precompact sets is necessarily bounded itself. (See [39.

PROBLEMS

dimensional space the unit ball cannot be strongly compact (Problem 16). The indistinguishability of the strong and the weak topologies in finitedimensional spaces yields a large class of examples of compact operators, namely all operators of finite rank. Examples of a slightly more complicated structure can be obtained by exploiting the fact that the set of compact operator is closed.
Problem 171. A diagonal operator with diagonal {an} is compact i{ and only if an --. 0 as n --. oo. Corollary. A weighted shift with weights {an: n = 0, 1, 2, ···}is compact if and only if an --. 0 as n --. oo.

172

172. Normal compact operators. It is easy to see that if a normal operator has the property that every non-zero element in its spectrum is isolated (i.e., is not a cluster point of the spectrum), then it is a diagonal operator. (For each non-zero eigenvalue A. of A, choose an orthonormal basis for the subspace {f: Af = A/}; the union of all these little bases, together with a basis for the kernel of A, is a basis for the whole space.) If, moreover, each non-zero eigenvalue has finite multiplicity, then the operator is compact. (Compare Problem 171; note that under the assumed conditions the set of eigenvalues is necessarily countable.) The remarkable and useful fact along these lines goes in the converse direction. Problem 172. The spectrum of a compact normal operator is countable; all its non-zero elements are eigenvalues offinite multiplicity. Corollary. Every compact normal operator is the direct sum of the operator 0 (on a space that can be anything from absent to non-separable) and a diagonal operator (on a separable space).

A less sharp but shorter formulation of the corollary is this: every compact normal operator is diagonal.

173

173. Hilbert-Schmidt operators. Matrices have valuable "continuous" generalizations. The idea is to replace sums by integrals, and it works-up to a point. To see where it goes wrong, consider a measure space X with measure J1 (a-finite as usual), and consider a measurable function K on the product space X x X. A function of two variables, such as K, is what a generalized matrix can be expected to be. Suppose that A is an operator on L 2 (J1) whose relation to K is similar to the usual relation of an operator to its matrix. In precise terms this means that if jE L 2 (J1), then (Af)(x)

=

f K(x, y)f(y)dJl(Y)
92

for almost every x. Under these conditions A is called an integral operator and K is called its kernel.

PROPERTIES OF COMPACTNESS

In the study of a Hilbert space H, to say "select an orthonormal basis" is a special case of saying "select a particular way of representing H as L 2 ". Many phenomena in L 2 spaces are the natural "continuous" generalizations of more familiar phenomena in sequence spaces. One simple fact about sequence spaces is that every operator on them has a matrix, and this is true whether the sequences (families) that enter are finite or infinite. (It is the reverse procedure that goes wrong in the infinite case. From operators to matrices all is well; it is from matrices to operators that there is trouble.) On this evidence it is reasonable to guess that every operator on L 2 has a kernel, i.e., that every operator is an integral operator. This guess is wrong, hopelessly wrong. The trouble is not with wild operators, and it is not with wild measures; it arises already if the operator is the identity and if the measure is Lebesgue measure (in the line or in any interval). In fact, if J1 is Lebesgue measure, then the identity is not an integral operator. The proof is accessible, but it reveals more about integrals than about operators; see, for instance, [66, p. 41]. What about the reverse problem? Under what conditions does a kernel induce an operator? Since the question includes the corresponding question for matrices, it is not reasonable to look for necessary and sufficient conditions. A somewhat special sufficient condition, which is nevertheless both natural and useful, is that the kernel be square integrable. Suppose, to be quite precise, that X is a measure space with a-finite measure Jl, and suppose that K is a complex-valued measurable function on X x X such that 1Kl 2 is integrable with respect to the product measure J1 x fl. It follows that, for almost every x, the function y ~ K(x, y) is in L 2 (J1), and hence that the product function y ~ K(x, y)f(y) is integrable whenever jE L 2(J1). Since, moreover,

II

2

f K(x, y)f(y)dfl(Y)

1

dfl(X) dJ1(y)· f1J(y)l dJ1(Y))dJ1(X) = IIKII
2

~ J( JIK(x, y)l

2

2

·11!11 2

(where IlK II is the norm of Kin L 2 (J1 x /1)), it follows that the equation

(Af)(x) = f K(x, y)f(y)dfl(Y)
defines an operator (with kernel K) on L 2 (J1). The inequality implies also that IIAII ~ IIKII· Integral operators with kernels of this type (i.e., kernels in L 2 (J1 x /1)) are called Hilbert-Schmidt operators. A good reference for their properties is [127]. The correspondence K ~ A is a one-to-one linear mapping from L 2 (J1 x Jl) to operators on V(Jl). If A has a kernel K (in L 2 (J1 x Jl)), then A* has the kernel K defined by
K(x, y)

= (K(y, x))*.
93

PROBLEMS

If A and B have kernels Hand K (in L 2 (J1 x Jl)), then AB has the kernel HK defined by
(HK)(x, y) =

f

H(x, z)K(z, y)dJl(z).

The proofs of all these algebraic assertions are straightforward computations with integrals. On the analytic side, the situation is just as pleasant. If {Kn} is a sequence of kernels in V(Jl x Jl) such that Kn ~ K (in the norm of L 2 (J1 x Jl)), and if the corresponding operators are An (for Kn) and A (for K), then IIAn - All ~ 0. The proof is immediate from the inequality between the norm of an integral operator and the norm of its kernel.
Problem 173. Every Hilbert-Schmidt operator is compact.

These considerations apply, in particular, when the space is the set of positive integers with the counting measure. It follows that if the entries of a matrix are square-summable, then it is bounded (in the sense that it defines an operator) and compact (in view of the assertion of Problem 173). It should also be remarked that the Schur test (Problem 45) for the boundedness of a matrix has a straightforward generalization to a theorem about kernels; see [21].

174

174. Compact versus Hilbert-Schmidt. Problem 174. Is every compact operator a Hilbert-Schmidt operator?

175

175. Limits of operators of finite rank. Every example of a compact operator seen so far (diagonal operators, weighted shifts, integral operators) was proved to be compact by showing it to be a limit of operators of finite rank. That is no accident. Problem 175. Every compact operator is the limit (in the norm) of operators offinite rank.

The generalization of the assertion to arbitrary Banach spaces was an unsolved problem for a long time. It is now known to be false; see [29] and [41].

176

176. Ideals of operators. An ideal of operators is proper if it does not contain every operator. An easy example of an ideal of operators on a Hilbert space is the set of all operators of finite rank on that space; if the space is infinitedimensional, that ideal is proper. Another example is the set of all compact operators; again, if the space is infinite-dimensional, that ideal is proper. The second of these examples is closed; in the infinite-dimensional case the first one is not.

94

PROPERTIES OF COMPACTNESS

Problem 176. If His a separable Hilbert space, then the collection ofcompact operators is the only non-zero closed proper ideal of operators on H.

Similar results hold for non-separable spaces, but the formulations and proufs are fussier and much less interesting.
177. Compactness on bases. An orthonormal sequence converges weakly to 0, and a compact operator is (w-+ s) continuous. It follows that if an operator A is compact, then A maps each orthonormal sequence onto a strong null sequence. To what extent is the converse true? It is plain enough what the converse means, but there are some closely related questions that deserve a look. Could it be, for instance, that if an operator A maps some orthonormal sequence onto a strong null sequence, then A is compact? Certainly not-that's absurd. Reason: the orthonormal sequence in question could be just half of an orthonormal basis on the other half of which A is large. (Example: a projection with infinite rank and nullity.) Very well: try again. Could it be that (1) if A maps some orthonormal basis onto a strong null sequence, then A is compact? (To avoid the irrelevant distractions of large cardinal numbers, it is best to assume here that the underlying Hilbert space is separable, and that the orthonormal bases to be studied come presented as sequences.) It is conceivable that the answer to question (1) is yes. The plain converse question originally asked is this: (2) if A maps every orthonormal basis onto a strong null sequence, then is A compact? It is conceivable that the answer to question (2) is no. The implications between the possible answers to (1) and (2) are clear: if (1) is yes, then so is (2); if (2) is no, then so is (1); in the other two cases the answer to one question leaves the other one open. What are the facts? Problem 177. If an operator A (on a Hilbert space of dimension ~ 0 ) maps an orthonormal basis onto a sequence that converges strongly to 0, is A compact? What if A maps every orthonormal basis onto a strong null sequence? 178. Square root of a compact operator. It is easy to construct non-compact operators whose square is compact; in fact, it is easy to construct noncompact operators that are nilpotent of index 2. (Cf. Problem 96.) What about the normal case? Problem 178. Do there exist non-compact normal operators whose square is compact? 179. Fredholm alternative. The principal spectral fact about a compact operator (normal or no) on a Hilbert space is that a non-zero number can get into the spectrum via the point spectrum only. 95

177

178

179

PROBLEMS

Problem 179. IfC is compact, then spec C- {0} c:: IT 0 (C).

Equivalently: if Cis compact, and if A is a non-zero complex number, then either A is an eigenvalue of C or C - A is invertible. In this form the statement is frequently called the Fredholm alternative. It has a facetious but substantially accurate formulation in terms of the equation (C- A)f = g, in which g is regarded as given and f as unknown; according to that formulation, if the solution is unique, then it exists.
Corollary. A compact operator whose point spectrum is empty is quasinilpotent.

180

180. Range of a compact operator. Problem 180. Every (closed) subspace included in the range of a compact operator is finite-dimensional. Corollary. Every non-zero eigenvalue of a compact operator has finite multiplicity.

181

181. Atkinson's theorem. An operator A is called a Fredholm operator if (1) ran A is closed and both ker A and ran-L A are finite-dimensional. (The last two conditions can be expressed by saying that the nullity and the co-rank of A are finite.) An operator A is invertible modulo the ideal of operators of finite rank if (2) there exists an operator B such that both 1 - AB and 1 - BA have finite rank. An operator A is invertible modulo the ideal of compact operators if (3) there exists an operator B such that both 1 - AB and 1 - BA are compact. Problem 181. An operator A is (1) a Fredholm operator if and only if it is (2) invertible modulo the ideal of operators offinite rank, or, alternatively, if and only if it is (3) invertible modulo the ideal of compact operators.

The result is due to Atkinson [7].

182

182. Weyl's theorem. The process of adding a compact operator to a given one is sometimes known as perturbation. The accepted attitude toward perturbation is that compact operators are "small"; the addition of a compact operator cannot (or should not) make for radical changes. Problem 182. If the difference between two operators is compact, then their spectra are the same except for eigenvalues. More explicitly: if A - B is compact, and if A E spec A - IT 0 (A), then A E spec B.

Note that for B = 0 the statement follows from Problem 179.

96

PROPERTIES OF COMPACTNESS

183. Perturbed spectrum. The spectrum of an operator changes, of course, when a compact operator is added to it, but in some sense not very much. Eigenvalues may come and go, but otherwise the spectrum remains invariant. In another sense, however, the spectrum can be profoundly affected by the addition of a compact operator. Problem 183. There exists a unitary operator U and there exists a compact operator C such that the spectrum of U

183

+ C is the entire unit disc.
184

184. Shift modulo compact operators. Weyl's theorem (Problem 182) implies that if U is the unilateral shift and if C is compact, then the spectrum of U + C includes the unit disc. (Here is a small curiosity. The reason the spectrum of U + C includes the unit disc is that U has no eigenvalues. The adjoint U* has many eigenvalues, so that this reasoning does not apply to it, but the conclusion does. Reason: the spectrum of U* + C is obtained from the spectrum of U + C* by reflection through the real axis, and C* is just as compact as C.) More is true [137]: every point of the open unit disc is an eigenvalue of ( U + C)*. It follows from the preceding paragraph that U + ·C can never be invertible (the spectrum cannot avoid 0), and it follows also that U + C can never be quasinilpotent (the spectrum cannot consist of 0 alone). Briefly: if invertibility and quasinilpotence are regarded as good properties, then not only is U bad, but it cannot be improved by a perturbation. Perhaps the best property an operator can have (and U does not have) is normality; can a perturbation improve U in this respect? Problem 184. If U is the unilateral shift, does there exist a compact operator C such that U + C is normal?

Freeman [ 44] has a result that is pertinent to this circle of ideas; he proves that, for a large class of compact operators C, the perturbed shift U + C is similar to the unperturbed shift U.
185. Distance from shift to compact operators. An increasingly valuable part of the study of operators is a subject called non-commutative approximation theory. The following question is a small sample of it. Problem 185. What is the distance from the unilateral shift to the set of all compact operators?

185

97

CHAPTER 20

Examples of Compactness

186

186. Bounded Volterra kernels. Integral operators are generalized matrices. Experience with matrices shows that the more zeros they have, the easier they are to compute with; triangular matrices, in particular, are usually quite tractable. Which integral operators are the right generalizations of triangular matrices? For the answer it is convenient to specialize drastically the measure spaces considered; in what follows the only X will be the unit interval, and the only f..l will be Lebesgue measure. (The theory can be treated somewhat more generally; see [115].) A Volterra kernel is a kernel Kin L 2 (f..l x f..l) such that K(x, y) = 0 when x < y. Equivalently: a Volterra kernel is a Hilbert-Schmidt kernel that is triangular in the sense that it vanishes above the diagonal (x = y) of the unit square. In view of this definition, the effect of the integral operator A (Volterra operator) induced by a Volterra kernel K can be described by the equation
(Af)(x) =

J:

K(x, y)f(y)dy.

If the diagonal terms of a finite triangular matrix vanish, then the matrix is nilpotent. Since the diagonal of the unit square has measure 0, and since from the point of view of Hilbert space sets of measure 0 are negligible, the condition of vanishing on the diagonal does not have an obvious continuous analogue. It turns out nevertheless that the zero values of a Volterra kernel above the diagonal win out over the non-zero values below.

Problem 186. A Volterra operator with a bounded kernel is quasinilpotent.

98

187. the assumption is that the kernel is bounded almost everywhere in the unit square. some are easy to answer and some are not. V is indefinite integration.) This is a pleasantly simple integral operator. If J~ f(y)dy = 0 for almost every x. Better: every vector in the range of V. It follows that Re V has rank l . y): 0 ~ y ~ x ~ 1}. it follows that V + V* is the integral operator whose kernel is equal to the constant function l almost everywhere. in fact. it follows that V is a perturbation (by an operator of rank 1 at that) of a skew Hermitian operator. considered as an equivalence class of functions modulo sets of measure 0. How important is the boundedness 187 assumption in Problem 186? Problem 187. Thus. so that the kernel of V* is the characteristic function of the triangle {(x. not to the operator. 99 . Since the functions in the range of V are not only continuous but. (The operators V* and V + V* are of course not Volterra operators. a moment's reflection should serve to show that it is the projection whose range is the (one-dimensional) space of constants. There are many other natural questions about V. contains a unique continuous function. here is a simple sample. In still other words.) Since V* is the integral operator whose kernel is the "conjugate transpose" of the kernel of V.EXAMPLES OF COMPACTNESS Caution: "bounded" here refers to the kernel. the equation holds for every x. y): 0 ~ x ~ y ~ l}. Is every Volterra operator quasinilpotent? 188. The simplest non-trivial Volterra operator is the one whose kernel is the characteristic function of the triangle {(x. Volterra integration operator. the result is that f(x) = 0 for almost every x. with the constant of integration adjusted so that every function in the range of V vanishes at 0.) The answer is no. As for natural questions that are not so easily disposeQ of. Unbounded Volterra kernels. (Note that every function in the range of Vis continuous. for instance. differentiable almost everywhere. The theory of Hilbert-Schmidt operators in general and Volterra operators in particular answers many questions about V. 1) by (Vf)(x) 188 = S: f(y)dy. then. Here is an easy one: does V annihilate any non-zero vectors? (Equivalently: "does V have a non-trivial kernel?". Explicitly this is the Volterra operator V defined on L 2 (0. and it is quasinilpotent (because it is a Volterra operator). Vis compact (because it is a Hilbert-Schmidt operator). by continuity. but that way terminological confusion lies. since V = Re V + i Im V. the equation can be differentiated.

There is an operator V0 on L 2 ( -1. If a triangular matrix has only 1's on the main diagonal. The question is simple to state: does every operator on an infinite-dimensional Hilbert space have a non-trivial invariant subspace? "Non-trivial" means different from both 0 and the whole space.PROBLEMS Problem 188. and most exasperating unsolved problems of operator theory is the problem of invariant subspaces. According to a dictum of P6lya's. in an attempt to get a positive result. of course. The reasoning that led to this conclusion was very finite-dimensional. the chances of finding a counterexample are greater. for instance. other than 1. many weakenings of the invariant subspace problem are either trivial or as difficult as the fullstrength problem. 100 . Norm 1.£ Iy I :. can it be patched up to yield the same conclusion for infinite-dimensional spaces? Problem 190. and the scholar's first task is to find the latter. Every finite matrix is unitarily equivalent to a triangular matrix. 1). but. then the answer is yes. such that spec A = {1} and IIAII = 1? 191 191. spectrum {1}. Skew-symmetric Volterra operator. Donoghue lattice. (For an elegant discussion. on the other hand. "subspace" is replaced by "linear manifold" (not necessarily closed). For finite-dimensional spaces there is. no problem.) If. no construction has yet been verified. for each unanswered question there is an easier unanswered question. despite periodically circulating rumors. "invariant" means that the operator maps it into itself. One of the most important. "Hilbert space" is replaced by "Banach space". y): 0 :.£ IxI :. Find the spectrum and the norm of the skew-symmetric Volterra operator V0 • 190 190. see [124]. The conclusion is that on a finite-dimensional Hilbert space the identity is the only contraction with spectrum {1}. as long as the complex field is used. + 1) (Lebesgue measure) that bears a faint formal resemblance to the operator V on L2 (0. What is the norm of V? 189 189. the fundamental theorem of algebra implies the existence of eigenvectors. most difficult. If. and easy. by definition Note that V0 is the integral operator induced by the kernel that is the characteristic function of the butterfly {(x. Is there an operator A. then its norm is at least 1. Even that dictum is hard to apply here.£ 1}. Problem 189. the norm can be equal to 1 only in case the matrix is the identity.

e. a wider class of them is described in [1 03]. The second operator whose invariant subspaces have received detailed study is the Volterra integration operator ([17. Ma is the subspace of those functions that vanish almost everywhere on [0. The collection of all subspaces invariant under some particular operator is a lattice (closed under the formation of intersections and spans). Is there an operator whose lattice of invariant subspaces is isomorphic to the positive integers? The question must be formulated with a little more care: every invariant subspace lattice has a largest element. The cheapest way to get one is to invoke the spectral theorem and to conclude that normal operators always have non-trivial invariant subspaces.n} is a monotone sequence (r:t. n = 0. n = 0. the first significant step in this direction is the work of Beurling [13] (Problem 157). · · · is invariant under such a shift. If A is a contraction such that neither of the sequences {An} and {A *n} tends strongly to 0.n ~ r:t. 58. but the generalizations are still closely tied to compactness. 1. en+l' en+z. An elegant way of obtaining these results is to reduce the study of the Volterra integration operator (as far as invariant subspaces are concerned) to that of the unilateral shift.) The lattice of invariant subspaces of V* is in an obvious way isomorphic to the closed unit interval. grows Ma shrinks. It is helpful to approach the subject from a different direction: instead of searching for counterexamples. · · · . 1). If (Vf)(x) = J~ f(y)dy for fin L 2 (0. ("Anti-" because as r:t. 2. ···)of positive numbers (r:t. The span of the basis vectors en. The first such operator was discovered by Donoghue [37]. between the indicated integers (including oo) and all invariant subspaces? The answer is yes. One way to state the result about Vis to say that its lattice of invariant subspaces is anti-isomorphic to the closed unit interval. in [0. The exact formulation is easy: is there an operator for which there is a one-to-one and order-preserving correspondence n ~ Mn. here is a sample. 93]. · · ·.n} will be called a monotone f2 shift. One way to do this is to fix attention on a particular operator and to characterize all its invariant subspaces. a more extensive bibliography is in [ 40]. i. and if. 1.n+h n = 0. but harder to prove. 2. That result has been generalized [12. 3. 91. then Ma is invariant under V. Non-compact results are few. 37. for each r:t. this was done in [123]. 122]).]. and a detailed treatment in [111]. oo. Suppose that {r:t.n > 0) such that l:::o r:t. 1]. 2./ < oo. the span Mn of 101 . A bird's eye view of the subject is in [71]. 84. The results for it are easier to describe than for the shift. r:t. then A has a non-trivial invariant subspace [100].. Nothing along these lines is easy. 1. The unilateral weighted shift with the weight sequence {r:t. the principal result is that every invariant subspace of V is one of the Ma 's. The orthogonal complement. study the structure of some non-counterexamples.EXAMPLES OF COMPACTNESS Positive results are known for some special classes of operators. The earliest non-trivial result along these lines is the assertion that compact operators always have non-trivial invariant subspaces [6].

en-l• is invariant under the adjoint. then there exists an integer n ( = 1. ···. 2. Problem 191.the principal result is that every invariant subspace of that adjoint is one of these orthogonal complements. · · -)such that M = Mn. If A is the adjoint of a monotone 12 shift. 3. 2. · · ·. n = 1. and if M is a non-trivial subspace invariant under A. 3. 102 .PROBLEMS e0 .

but whose proof is not. to be sure. however. in the sense that for finite-dimensional spaces they are either meaningless or trivial. Some of the natural questions about normal operators have the same answers for finite-dimensional spaces as for infinite-dimensional ones. but the techniques are not. A familiar and typical example of a theorem whose statement is easily generalizable from the finite to the infinite. For finite-dimensional spaces the statement was known to be true and trivial. more generally. that sometimes the finite and the infinite proofs are intrinsically different. It is more striking because it was for many years an unsolved problem. a case in point is the statement that any two bases have the same cardinal number. to say that the techniques are different means that the natural finite-dimensional techniques are not generalizable to infinite-dimensional spaces. It should be added. Sometimes. so that neither can be adapted to yield the result of the other.CHAPTER 21 Subnormal Operators 192. on the other hand. or. are properly infinite-dimensional. questions about shifts. for infinite-dimensional spaces it was unknown. Putnam-Fuglede theorem. either the question or the answer must be reformulated in order to bring the finite and the infinite into harmony. The algebraically simplest formulation is that if A is a normal operator and if B is an operator that commutes with A. As for the technique. The Fuglede theorem (cf. A more striking example is the Fuglede commutativity theorem. questions about subnormal operators are likely to belong to this category (see Problem 195). experience shows that an infinite-dimensional proof can usually be adapted to the finite-dimensional case. Between these two extremes there are the questions for which the answers are invariant under change of dimension. is the spectral theorem. Some questions. Solution 146) can be formulated in several ways. then B commutes 103 192 . and the techniques used to prove the answers are the same.

then they are unitarily equivalent. If. then A* commutes with B. obtained by splitting the two roles of A between two normal operators. B. It remains true for compact operators [157]. The properties of an operator are intimately connected with how it enters various algebraic structures. The Putnam generalization (i. Away from finite-dimensional spaces even this result becomes recalcitrant. A + iD. is true and useful. One of the techniques that can be used to prove the Fuglede theorem is to characterize in terms of the geometry of 104 . just take the adjoint of the assumed equation AB = BA. and therefore so are A + iC. with A. then BA also is normal. Here is a precise formulation. the adjoint of A 1 B = BA 2 is. A pertinent positive result was obtained by Wiegmann [156]. If two normal operators are similar. then A. the modified assertion. the proof seems to need the Fuglede theorem. then that algebra is commutative. the assumed normality implies that everything commutes. In the latter form the assertion is that in a certain special situation commutativity is transitive. and B + iD. In this connection it should be mentioned that the product of not necessarily commutative normal operators is very reluctant to be normal. and A commutes with B.) The operator A plays a double role in the Fuglede theorem. then A 1 *B = BA 2 *. (Proof: if A + iB and C + iD are in the algebra. conversely. however.PROBLEMS with A* also.. Problem 192. Algebras of normal operators. Equivalently: if A* commutes with A. Corollary. and D are in the algebra. Spectral measure of the unit disc. it is trivial that if an operator belongs to a commutative algebra that is closed under the formation of adjoints. and D Hermitian. C. Is an algebra ofnormal operators necessarily commutative? 194 194. B. Is the product of two commutative normal operators normal? The answer is yes. it says that if H is a finite-dimensional Hilbert space and if A and B are normal operators on H such that AB is normal. 193 193. Thus. then that operator is normal. not obvious even if B is Hermitian. C. If A 1 and A 2 are normal operators and such that A 1 B = BA 2 . which is not what is wanted. but it is false in the general case [85]. in that case.) Question: what if the condition on adjoints is dropped? Problem 193. an algebra is closed under the formation of adjoints and consists of normal operators only. because adjoints are formable. Problem 192) is.e. if B is an operator Observe that the Fuglede theorem is trivial in case B is Hermitian (even if A is not necessarily normal). for instance. BA 1 * = A 2 *B. (In general it is not. B + iC. and the proof is the same for spaces of all dimensions.

That technique is useful in other contexts too. The theory of normal operators is so successful 195 that much of the theory of non-normal operators is modeled after it. The operatorial analogue of the preceding numerical observations should be something like this: if A is a normal operator on a Hilbert space H.SUBNORMAL OPERATORS Hilbert space the spectral subspaces associated with a normal operator. in particular. there are no f's in L 2 (other than 0) such that IIAnfll ~ 8nllfll for all n. An operator is subnormal if it has a normal extension. The same result can be formulated in a more elementary manner in the language of multiplication operators. and consequently the operator A is invertible. x denotes the characteristic function of the set indicated by its subscript. Conclusion: under these circumstances 1 cp(x) 1 > 8 almost everywhere. One way this result can sometimes be put to good use is this: if. in some sense. By translations and changes of scale the spectral subspaces associated with all discs can be characterized similarly. then the set E of those vectors fin H for which IIAnfll ~ llfll. If A is the multiplication operator induced by a bounded measurable function cp on a measure space. as usual. 3. There is a close connection between complex-valued functions and normal operators. One weakening of normality is quasinormality (see Problem 137). which goes in an altogether different direction. n = 1. 195. then a necessary and sufficient condition that an element fin L 2 be such that X"'-'<D)f = f is that IIAnfll ~ llfll for every positive integer n. A natural way to extend a successful theory is to weaken some of its hypotheses slightly and hope that the results are weakened only slightly. the precise formulation is that E is a subspace of H and the projection on E is the value of the spectral measure associated with A on the closed unit disc in the complex plane. and if D = {z: lzl ~ 1}. n = 1. This trivial observation extends to complex-valued functions: {x: lcp(x)l ~ l} = {x: lcp(x)ln ~ 1. Subnormal operators constitute a considerably more useful and deeper generalization. then the subspace off's that vanish on the complement of the set {x: lcp(x)l ~ 8} is 0. Subnormal operators. Problem 194. an operator A on a Hilbert space H is subnormal if there exists a normal operator B on a Hilbert space K such that H is a 105 . 3. More precisely. should be. · · · . a necessary and sufficient condition that a vector f be invariant under multiplication by the characteristic function of {x: lcp(x)l ~ 8} (8 > 0) is that IIAnfll ~ 8nllfll for all n. the part of H on which A is below 1. 2. for some positive number 8. 2. · · ·}. Here. This is true. and therefore the set {x: l<t>(x)l ~ 8} is (almost) empty. A necessary and sufficient condition that a complex number have modulus less than or equal to I is that all its powers have the same property.

it follows that M = K. only to a subtle analytic approach [24]. what is its minimal normal extension? 106 . The result is just as subnormal as the unilateral shift. and everyone does. again. but. Normality implies quasinormality. Problem 195. The present assertion is that quasinormality implies subnormality. but that takes a little proving. then the unilateral shift would be normal. A normal extension B (on K) of a subnormal operator A (on H) is minimal if there is no reducing subspace of B between H and K. What is the right article for minimal normal extensions: "a" or "the"? Problem 197. The explicit determination of the minimal normal extension of a subnormal operator can be a non-trivial problem. Solution 202 or Problem 203. but a straightforward computation shows that if it were also quasinormal. Quasinormal invariants. IfU is the unilateral shift and pis a polynomial. To get a counterexample. as in all others. Typical example: the minimal normal extension of the unilateral shift is the bilateral shift. the converse is false. then p(U) is a subnormal operator. In view of this result. Every normal operator is trivially subnormal. In other words. after many years. UB 1 = B 2 U) and is equal to the identity on H.PROBLEMS subspace of K. for subnormal operators it was very hard and yielded. A more interesting and typical example of a subnormal operator is the unilateral shift.. The invariant subspace problem is easy for normal operators. Problem 196. cf. Every quasinormal operator on a space of dimension greater than 1 has a non-trivial invariant subspace. add a non-zero scalar to the unilateral shift. then there exists an isometry U from K1 onto K2 that carries B 1 onto B 2 (i. Problem 198. In this respect. and the restriction of B to H coincides with A. B is minimal over A if whenever M reduces B and H c M.e. If B 1 and B 2 (on K 1 and K2 ) are minimal normal extensions of the subnormal operator A on H. the subspace H is invariant under the operator B. any special case where that extension is accessible is worth looking at. Polynomials in the shift. quasinormal operators are between the two. 198 198. Every quasinormal operator is subnormal. On finite-dimensional spaces every subnormal operator is normal. 197 197. the bilateral shift is a normal extension. it is permissible to speak of" the" minimal normal extension of a subnormal operator. Minimal normal extensions. but not conversely (witness the unilateral shift). 196 196.

if A is subnormal and B is its minimal normal extension.e. A hole in a compact subset of the complex plane is a bounded component of its complement. and if ~ is a hole of spec B. as an operator grows. 202. Problem 201. Differently expressed: if A c: B. then spec A c: spec B. in particular. then ~ is either included in or disjoint from spec A. Spectral inclusion theorem. 201.. Can a subnormal but non-normal operator on a Hilbert space H have a normal extension to a Hilbert space K when dim(K f'\ H. Problem 202. This informal expression can be given a precise technical meaning. then spec A is the unit disc. and hence that. If an operator A is a restnctwn of an operator B to an invariant subspace H of B. If A is subnormal and then spec B c: spec A. or. is this one of the respects in which they succeed? Problem 199.f for some scalar A. Similarity of subnormal operators. Extensions of finite co-dimension. Subnormal operators are designed to imitate the properties of normal ones. Are two similar subnormal operators necessarily unitarily equivalent? 200. if B is its minimal normal extension. Filling in holes. An equally easy verification shows that as an operator grows.L) is finite? 202 107 . then f is an eigenvector of B.). its point spectrum grows. its spectrum grows. f E H and Af = A.SUBNORMAL OPERA TORS 199. The first non-trivial example of a subnormal operator shows that this conjecture is false: if A is the unilateral shift and B is the bilateral shift. its approximate point spectrum grows. The spectral inclusion theorem (Problem 200) for 201 subnormal operators can be sharpened in an interesting and surprising manner. Problem 200. 199 200 if B is its minimal normal extension. It turns out that this counterexample illustrates the general case better than do the plausibility arguments based on eigenvalues. whereas spec B is only the perimeter of the unit disc. For normal operators similarity implies unitary equivalence (Problem 192). then ll 0 (A) c: ll 0 (B). it is tempting to conjecture that as an operator grows. and iff is an eigenvector of A (i. exact or approximate. In view of these very natural observations. If A is subnormal. The result is that the spectrum of a subnormal operator is always obtained from the spectrum of its minimal normal extension by "filling in some of the holes". Reference: [51].

its restriction to His an operator on H. It is easy to express the relation (A c B) between A and Bin terms of the matrix of B. it is equivalent to the operator inequality AA* .£A* A. it follows that the matrix A* A B*B. that restriction is equal to A*.f) and IIA/11 2 = (A*Af. Since the operator PB* on K leaves H invariant. Iff and g are in H.AA* 0. and. Hyponormal operators. Corresponding to the decomposition K = H EB Hj_. This relation between A* and B* has a curious consequence. that is true for B. and (1) says that the restriction of B to His A. according to the preceding chain of equations. then IIA*/11 = IIPB*/11 .PROBLEMS 203 203. Iff E H. Ag) = (f. The condition (2) says that His invariant under B. Pg) = (PB*f. If A (on H) is subnormal. in particular. g).£ IIA/11) can be reformulated in another useful way. then (A*f. 108 . That is the answer: A*f = PB*f for every fin H.£ liB*! II = IlB/II (by normality)= IIA/11· The result (IIA*/11 . g)= (f. Thus 2 so that B* = (~: . what is the relation between A* and B*? The answer is best expressed in terms of the projection P from K onto H. This implies that A*A.BB* = ( R*A A*R ) (AA* + RR* R*R + S*S SR* must vanish. Indeed: IIA*/11 = (AA*f.AA* = RR*.f). g)= (B*f.*} RS*) SS* Since B is normal. and (2) the one below it (southwest) is 0. and hence that ~ A*A. The curious inequality that subnormal operators always satisfy can also be obtained from an illuminating matrix calculation. and. with normal extension B (on K). a necessary and sufficient condition for it is that (1) the principal (northwest) entry is A. Bg) = (B*f. every operator on K can be expressed as an operator matrix.

as follows. Since tr(AB) is always equal to tr(BA). it is subnormal if and only if it is the direct sum of an isometry and zero. Bishop's result is easy to state: the set of all subnormal operators is exactly the strong closure of the set of all normal operators.AA* is a positive operator with trace 0.) The result of the preceding paragraphs is that every subnormal operator is hyponormal. but this is how it grew. U*U.. whereas AA *is a non-trivial projection. and then U would be normal. clearly A* A .AA *)is always 0. On a finite-dimensional space every hyponormal operator is normal. the unilateral shift. The dull dual result is. The characterization is of"finite character". Normal and subnormal partial isometrics. 204 In both cases.?. Problem 195). if A* A. then A is not subnormal. i. and A* A = 1. then A is subnormal. The techniques used are almost sufficient to yield an intrinsic characterization of subnormality ([49. A partial isometry is normal if and only if it is the direct sum of a unitary operator and zero. AA *. An operator A such that A* A .If A = U*. Problem 204. AA*. A* A. The most efficient proof of this assertion is a trace argument. then it would satisfy the inequality A* A. that every cosubnormal operator is co-hyponormal. AA* has been called hyponormal. symmetry could be restored to the universe by the introduction of the dual concept of co-subnormality.) If A is co-subnormal in this sense. and therefore A* A . 16]).?. (Proposed definition: the adjoint is subnormal. This is not easy. (See Problem 225. What was thus proved is a generalization of the statement that on a finite-dimensional space every subnormal operator is normal (cf. then A* A .) 204. "Intrinsic" means that the characterization is expressed in terms of the action of the operator on the vectors in its domain. Problem 203. With still more work of the same kind an elegant topological characterization of subnormality can be obtained.?. it follows that tr(A *A . and it seems to be here to stay.) If it were deemed absolutely essential.?. 109 . Give an example of a hyponormal operator that is not subnormal. in the sense that it depends on the behavior of the operator on all possible finite sets of vectors. (Reason: if it were. If A = U. of course. The nomenclature is not especially suggestive.e. and not in terms of the existence of something outside that domain.SUBNORMAL OPERATORS There is a curious lack of symmetry here: why should A* A play a role so significantly different from that of AA *? A little meditation on the unilateral shift may help. AA*. one or the other of the direct summands may be absent.?.?. this was first done by Bishop [15]. (The dual kind might be called co-hyponormal.AA* = 0. then AA* . UU*. Note that "hypo" is in Greek what "sub" is in Latin.

a natural question remains unanswered.PROBLEMS 205 205. but A is certainly not normal. then IIAnll ~ IIAIIn = (r(A)t = r(An) ~ IIAnll. What if only one of Re A and Im A is compact? Problem 207. the converse becomes false. 207 207. The only hyponormal quasinilpotent operator is 0. so that equality holds all the way through. Compact hyponormal operators. As soon as the dimension becomes greater than 2. then A is normal. The set T of those operators A such that II An II = IIAIIn for every positive integer n has at the very least. then A E T) is to refer to the spectral theorem. r(A) = IIAII. It follows from the discussion of hyponormal operators on finite-dimensional spaces (Problem 203) that a hyponormal operator of finite rank (on a possibly infinite-dimensional space) is always normal. Does there exist a hyponormal operator that is not normal but has compact imaginary part? 110 . a certain curiosity value. 136]. The answer turns out to be affirmative. then IIAnll = IIAIIn for every positive integer n. Since neither the assertion nor its proof have any merit. Reference: [4. Hyponormal. The quickest (but not the most elementary) proof of the direct assertion (if A is normal. and therefore r(A) = IIAII. If A is hyponormal. Corollary. conversely. the latter is omitted. 206 206. If AET. Every compact hyponormal operator is normal. if. then A is normal. Since for subnormal and hyponormal operators that theorem is not available. If. For two-by-two matrices an unpleasant computation proves a strong converse: if IIA 2 11 = IIAII 2 . A=(~ ~1 0 ~). Norm powers and power norms. The result of Problem 206 can therefore be stated in this form: if A is hyponormal and both Re A and Im A are compact. 10. then 11Anll 11n = IIAII. compact imaginary part. 0 then IIAnll~ 1 for all n. An operator A is compact if and only if both Re A and Im A are compact. Problem 205. Conclusion: A ET if and only if r(A) = IIAII· The definition ofT implies that every normal operator belongs toT (and so does the conclusion ofthe preceding paragraph). What about limits of operators of finite rank? Problem 206. for example.

Hyponormal idempotents.SUBNORMAL OPERATORS It should be kept in mind that even if A is hyponormal. is that iA is hyponormal at the same time as A is and Im(iA) = ReA. the hyponormal operator used in Solution 203 has the property that all its powers are hyponormal also. A normal operator with one good property 208 usually has several good properties. however. A typical example of this kind of statement is that an idempotent normal operator must be Hermitian (and hence a projection). bound to be at least as difficult as the construction of a hyponormal operator that is not subnormal (Problem 203). 209 This is not easy. The converse is not true. in fact. It is. 111 . Give an example of a hyponormal operator whose square is not hyponormal. To what extent do the useful generalizations of normality behave normally? Problem 208. Every power of a normal operator is normal. This trivial observation has as an almost equally trivial consequence the statement that every power of a subnormal operator is subnormal. since any solution of Problem 209 is automatically a solution of Problem 203. A* need not be. What is trivially true. For hyponormal operators the facts are different. Problem 209. 208. Powers of hyponormal operators. Is every quasinormal idempotent a projection? What about subnormal idempotents? What about hyponormal ones? 209.

apparently as an afterthought. Nowadays they play a secondary role. The numerical range of an operator A is the set W(A) of all complex numbers of the form (Af. but not vice versa. Toeplitz.) The numerical range of A is the range of the restriction to the unit sphere of the quadratic form associated with A. Toeplitz-Hausdorff theorem. This is not to say that the quadratic point of view is dead. Most quadratic questions about an operator are questions about its numerical range. In early studies of Hilbert space (by Hilbert. First comes an operator A on a Hilbert space H. (The image of the unit ball is the union of all the closed segments that join the origin to points of the numerical range. 112 . f). where f varies over all vectors on the unit sphere. and then. One reason for the emphasis on the image of the unit sphere is that the image ofthe unit ball. it still suggests questions that are interesting with answers that can be useful. f) on H. (Important: 11/11 = 1. Hellinger. if A= G~). comes the numerical-valued function/~--> (Af. not 11/11 . sometimes called its field of values.:£ 1. and others) the objects of chief interest were quadratic forms.CHAPTER 22 Numerical Range 210 210. and also the entire range. then W(A) is the closed unit interval (easy).) The determination of the numerical range of an operator is sometimes easy. Here are some sample results. If A=(~ ~). are easily described in terms of it. the entire range is the union of all the closed rays from the origin through points of the numerical range.

The result is known as the Toeplitz-Hausdorfftheorem. then the set of all n-tuples of the form <(Ad. so normalized that llfll = liB II = I. minor axis I and major axis ./31/o and the major axis is Ia . ( 0 I 0 where A is a complex number of modulus I. If A= 0 0 . then W(A) is the closed elliptical disc with foci at 0 and I. then W(A) is the (circular) disc with center a and radius !IIA . (Anf. but more interesting). · · · . If A is a two-by-two matrix with distinct eigenvalues a and /3. There is a theorem that covers all these cases. The following assertion describes the most important common property of all these examples.sin t 0 ~ t ~ 2n (where the value at 0 is taken to be the limit from the right).) If A= 0) 0 0 0. If A is the Volterra integration operator (see Problem 188). then W(A) is a closed elliptical disc with foci at a and f3./31/o. if A= (0 0) 1 1 . (Cf.}2 (analytic geometry at its worst). if y = l(f. (Cf.cos t t2 + i --=-- t . and corresponding eigenvectors f and g.all. A couple of three-dimensional examples will demonstrate that the twodimensional case is not typical. Problem 216. If A has only one eigenvalue a. then the minor axis is yla. Problem 216. Consideration of real and imaginary parts shows that it is a special case (n = 2) of the following general assertion: if A 1 .) The higher the dimension. where 113 . The numerical range of an operator is always convex. I 0 (0 0 I then W(A) is the union of all the closed segments that join the point 1 to points of the closed disc with center 0 and radius!. • • • . f). f)). then W(A) is the set lying between the curves t f--+ I .NUMERICAL RANGE then W(A) is the closed disc with center 0 and radius ! (easy. the stranger the numerical range can be.. Problem 210.1:) I 0 0. An are Hermitian operators. then W(A) is the equilateral triangle (interior and boundary) whose vertices are the three cube roots of A. g)l and o = ~.

in symbols Wk(A): it is the set of all complex numbers of the form tr PAP.PROBLEMS 11/11 = 1. True or false. The numerical range of an operator isn't always closed. It is false for n = 3 in dimension 2. but left open the possibility that it had interior holes. where P varies over all projections of rank k. it is a pity that it is so very false. Replace 1 by an arbitrary positive integer k. where P varies over all projections of rank 1. that is. the assertion seems to be a natural generalization of the ToeplitzHausdorff theorem. Closure of numerical range. Problem 211. The number 0 plays a 114 . Solution 212 exhibited a compact operator A such that W(A) = (0. The trace of PAP can be computed by finding the (one-by-one) matrix of the restriction of PAP to the range of P. with respect to the (one-element) basis {f}. 211 211. Problem 212. Is the k-numerical range of an operator always convex? 212 212. is a convex subset of n-dimensional real Euclidean space. The first paper on the subject was by Toeplitz [142]. 0 is in the closure of W(A) but not in W(A) itself. The numerical range can be regarded as the one-dimensional case of a multi-dimensional concept. 213 213. The result about the Volterra integration operator is due to A. then PAP is an operator of rank 1. and therefore a finite-dimensional concept such as trace makes sense for it. If A is an arbitrary operator. Observe that in the finite-dimensional case the numerical range of an operator is a continuous image of a compact set. not even if the operator is compact. 1]. Donoghue [36] reexamined the facts and presented some pertinent computations. and obtain the k-numerical range of A. since Pf = f. Give examples of operators whose numerical range is not closed.f) = (APf.f)f for all g. To see how that goes.f). counterexamples are easy to come by. These remarks can be summarized as follows: W(A) is equal to the set of all complex numbers of the form tr PAP. Pf) = (Af. and hence necessarily compact. Numerical range of a compact operator. the value of that trace is (PAPf. who proved that the boundary of W(A) is a convex curve. Brown. The ordinary numerical range is the k-numerical range with k = 1. Hausdorff [69] proved that it did not. Higher-dimensional numerical range. recall the expression of a projection P of rank 1 in terms of a unit vector fin its range: Pg = (g.

then. the set W(A) may fail to be closed. Is there a compact operator A such that 0 E W(A) but W(A) is not closed? 214. 215 216. Perhaps the most useful formulation of this fact for the plane goes as follows: the convex hull of a compact set is the intersection of all the closed half planes that include it. To formulate the result precisely. 115 .NUMERICAL RANGE special role with respect to the spectrum of a compact operator. conv M is the intersection of all the convex sets that include M. but the numerical range is large ({z: lzl ~ 1}). conv M is the smallest convex set that includes M. in other words. does it play an equally special role with respect to the numerical range? Problem 213. 0 E W(A). It is a non-trivial fact of finite-dimensional Euclidean geometry that the convex hull of a compact set is closed. in symbols conv M. then spec A c: W(B) + iW(C) is the Bendixson-Hirsch theorem. so that from 0 E W(A) it does not follow that 0 E W(A). By definition. Spectrum and numerical range. with B and C Hermitian. A discouraging example is 216 the spectrum is small ({0}). Is it true just the same? Problem 215. by Problem 214. Can the closure of the numerical range be very much larger than the spectrum? The answer is yes. Observe that any such example is a solution of Problem 212. Problem 214. A useful reference for all this is [144]. it is necessary to introduce the concept of the convex hull of a set M. So much for making convex sets out of closed sets. Among normal operators such extreme examples do not exist. for them the closure of the numerical range is as small as the universal properties of spectra and numerical ranges permit. If A is a quasinilpotent operator. 214 215. The trivial corollary that asserts that if A = B + iC. Quasinilpotence and numerical range. Give an example of a quasinilpotent operator A such that 0 ¢ W(A). The closure of the numerical range includes the spectrum. By Solution 212. it is true and easy to prove that the closure of a convex set is convex. Normality and numerical range. The reverse process of making closed sets out of convex sets is much simpler to deal with.

As an application consider the matrix ( 0 0 A. like the spectrum. the result just proved implies that its numerical range is the convex hull of its eigenvalues.).£ 116 . It is easy to prove that w is a norm.) 1 0 0 ' 0 1 0 where IA.. and w(A) = 0 if and only if A= 0. called the numerical radius.) Some of the properties of the numerical radius lie near the surface. A different generalization of this special assertion is that the numerical range of a direct sum is the convex hull of the numerical ranges of its summands. and recapture the assertion (made in passing in Problem 210) about the domed-cone shape of the numerical range of 0 0 0) 1 0 0 . and therefore normal. (Cf.e W(A)}. Problem 217. The closure of the numerical range of a normal operator is the convex hull of its spectrum. There is a closely related numerical function w. the definition of spectral radius. w(ocA) = loci· w(A) for each scalar oc. The numerical range. Since this matrix is unitary. and w(A +B). For an example. others are quite deep. Does the conclusion of Problem 216 remain true hypothesis "normal" is replaced by "subnormal"? if in the 218 218. 0 0 1 ( 217 217.PROBLEMS Problem 216. consider the direct sum of and (1). associates a set with each operator. The general result includes the special assertion that the numerical range of every finite diagonal matrix is the convex hull of its diagonal entries. it is a set-valued function of operators. I = 1. Problem 88. and this proves the assertion (made in passing in Problem 210) that the numerical range of this particular matrix is the triangle whose vertices are the cube roots of A. The eigenvalues are easy to compute (they are the cube roots of A. Numerical radius. That is: w(A) ~ 0. defined by w(A) =sup{ IA. I: A. Subnormality and numerical range. The proof of the generalization is straightforward.

w(A* A) = IIA 11 2 . write M 220 + (s) = {z +a: z EM. convexoid. nilpotent) operators shows that nothing like the reverse inequality could be true. Since the Hausdorff metric is defined for compact sets. N) between them is the infimum of all positive numbers s such that both M c: N + (s) and N c: M + (s). As for the continuity question. Problems 102 and 103. then that disc includes W(A). there is an easy inequality between the spectral radius and the numerical radius: r(A) ~ w(A).NUMERICAL RANGE w(A) + w(B). in the sense that w(U* AU) = w(A) whenever U is unitary. for instance. Wintner called an operator A with w(A) = IIAII normaloid.) The norm w has many other pleasant properties. and spectraloid operators. then A is invertible. since the closed disc with center 0 and radius r(A) includes spec A and is convex. It is a consequence of Problem 218 that every normaloid operator is spectraloid. for that matter. call them convexoid. In this notation. p. if M and N are compact sets. then w(A) = IIAII. not W.) The best way to ask the question is in terms of the Hausdorff metric for compact subsets of the plane. This implies th~t w(A) ~ r(A). and w is unitarily invariant. It is also true that every convexoid operator is spectraloid. Still another (nameless) property of a normal operator A is that r(A) = w(A). 33]. it follows that if A is convexoid. and hence that A is spectraloid. = 219. This norm is equivalent to the ordinary operator norm. call an operator with this property spectra/aid. In what sense is the numerical range a continuous function of its argument? (Cf. w(A*) = w(A). Problem 218. Problem 219.A)< 1. (b) Ifw(A) II All. Since spec A c: W(A) (Problem 214). Normaloid. Discuss the implication relations between the properties of being convexoid and normaloid. the Hausdorff distance d(M. Indeed. Another useful (but nameless) property of a normal operator A (Problem 216) is that W(A) is the convex hull of spec A. To define that metric. To have a temporary label for (not necessarily normal) operators with this property. (a) Ifw(l. the appropriate function to discuss is W. then r(A) = IIAII. Continuity of numerical range. If A is normal. 219 220. it still has as 117 . in the sense that each is bounded by a constant multiple of the other: -!IIAII ~ w(A) ~ IIAII· (See [50. The existence of quasinilpotent (or. thus. Ia I < s} for each set M of complex numbers and each positive number s.

then so is W. Indeed: w(AB) = w(A 3 ) = ! > w(A) ·! = w(A)w(B). The only shred of multiplicative behavior that has not yet been ruled out is the power inequality 118 ! . it follows that for normal operators w is submultiplicative (because if A and Bare normal. 221 221. for instance. then IIA II = w(A) and liB II = w(B)). if w is discontinuous. Is W weakly continuous? strongly? uniformly? And what about w? The only thing that is immediately obvious is that if W is continuous with respect to any topology. i.:£ w(A)w(B)). w(AB) is not always equal to w(A)w(B). The example used to show that w is not submultiplicative shows also that the constant 4 is best possible here. Discuss the continuity of Wand w in the weak.:£ IIAII · IIBII. and consequently. strong. here it does not.. The value of w(A) is slightly harder to compute.) The next best thing would be for w to be submultiplicative (w(AB) . Power inequality. but that is false too. but it is not needed.:£ 4w(A)w(B) (because IIA II . Thus. but they exist. (Example with commutative normal operators: if A = (~ ~) and B = (~ ~). Here is one: A~(~ ~ ~) and B = A 2 • It is easy to see that w(A 2 ) = w(A 3 ) = t. the almost obvious relation w(A) < 1 will do. Examples of commutative operators A and B for which w(AB) > w(A)w(B) are a little harder to come by.e. w is certainly not multiplicative. The good properties of the numerical range and the numerical radius have to do with convexity and linearity.PROBLEMS many interpretations as there are topologies for operators. then so is w. then w(A) = w(B) = 1 and w(AB) = 0. Problem 220. Commutativity sometimes helps. then w(A) = w(B) =!and w(AB) = 1. the relations between the numerical range and the multiplicative properties of operators are less smooth.:£ IIABII . (Example: if A= (~ ~) and B = (~ ~). and for operators in general w(AB) . and uniform operator topologies.) Since w(AB) ..:£ 2w(A) and liB II ~ 2w(B)).

NUMERICAL RANGE

This turns out to be true, but remarkably tricky. Even for two-by-two matrices there is no simple computation that yields the result. If not the dimension but the exponent is specialized, if, say, n = 2, then relatively easy proofs exist, but even they require surprisingly delicate handling. The general case requires either brute force or ingenuity.

Problem 221. If A is an operator such that w(A) for every positive integer n.

~

1, then w(An)

~

1

The statement is obviously a consequence of the power inequality. To show that it also implies the power inequality, reason as follows. If w(A) = 0, then A = 0, and everything is trivial. If w(A) ¥= 0, then write B = Ajw(A), note that w(B) ~ 1, use the statement of Problem 221 to infer that w(Bn) ~ 1, and conclude that w(An) ~ (w(A))n. Generalizations of the theorem are known. Here is a nice one: if p is a polynomial such that p(O) = 0 and lp(z)l ~ 1 whenever lzl ~ 1, and if A is an operator such that w(A) ~ 1, then w(p(A)) ~ I. With a little care, polynomials can be replaced by analytic functions, and, with a lot of care, the unit disc (which enters by the emphasis on the inequality Iz I ~ 1) can be replaced by other compact convex sets. The first proof of the power inequality is due to C. A. Berger; the first generalizations along the lines mentioned in the preceding paragraph were derived by J. G. Stampfli. The first published version, in a quite general form, appears in [86]. An interesting generalization along completely different lines appears in [ 101].

119

CHAPTER 23

Unitary Dilations

222

222. Unitary dilations. Suppose that H is a subspace of a Hilbert space K, and let P be the (orthogonal) projection from K onto H. Each operator Bon K induces in a natural way an operator A on H defined for each f in H by

Af= PB{. The relation between A and B can also be expressed by AP = PBP. Under these conditions the operator A is called the compression of B to H and B is called a dilation of A to K. This geometric definition of compression and dilation is to be contrasted with the customary concepts of restriction and extension: if it happens that H is invariant under B, then it is not necessary to project Bf back into H (it is already there), and, in that case, A is the restriction of B to H and B is an extension of A to K. Restrictionextension is a special case of compression-dilation, the special case in which the operator on the larger space leaves the smaller space invariant. There are algebraic roads that lead to compressions and dilations, as well as geometric ones. One such road goes via quadratic forms. It makes sense to consider the quadratic form associated with Band to consider it for vectors of H only (i.e., to restrict it to H). This restriction is a quadratic form on H, and, therefore, it is induced by an operator on H; that operator is the compression A. In other words, compression and dilation for operators are not only analogous to (and generalizations of) restriction and extension, but, in the framework of quadratic forms, they are restriction and extension: the quadratic form of A is the restriction of the quadratic form of B to H, and the quadratic form of B is an extension of the quadratic form of A to K. Still another manifestation of compressions and dilations in Hilbert space theory is in connection with operator matrices. If K is decomposed into H
120

UNITARY DILATIONS

and H.l, and, correspondingly, operators on K are written in terms of matrices (whose entries are operators on Hand H.l and linear transformations between Hand H.L), then a necessary and sufficient condition that B be a dilation of A is that the matrix of B have the form

Problem 222. (a) If IIAII ~ 1, then A has a unitary dilation. (b) If 0 ~ A ~ 1, then A has a dilation that is a projection. Note that in both cases the assumptions are clearly necessary. If A has a dilationBthatisacontraction,then IIAfll = IIPBfll ~ IIBfll ~ llfll forallf in H, and if A has a positive dilation B, then (Af, f) = (Bf, f) ~ 0 for all fin H. Corollary. Every operator has a normal dilation. 223. Images of subspaces. The range of an operator is not necessarily closed, and, all the more, the image of a subspace under an operator is not necessarily closed. If the operator is very well behaved (bounded from below), all is well. Just how well does it have to behave? Problem 223. Is the image of a subspace under a projection closed? 224. Weak closures and dilations. Dilations have easy but sometimes surprising applications to the weak operator topology. Basic weak neighborhoods were defined (Problem 107) in terms of finite sets of vectors. An equivalent definition is in terms of projections of finite rank. Assertion: a base for the weak operator topology is the collection of all sets of the form {A: IIF(A - A0 )FII < e}, where F is a projection of finite rank and e is a positive number. An efficient way to prove the assertion is to compare the pseudonorms IIAIIJ,g defined by vectors with the ones

223

224

= i(Af, g) I

IIAIIF = IIFAFII defined by projections of finite rank. Given vectors f 1, • • ·, fn, g 1, • • ·, gn, let F be the projection whose range is the span of all the f's and g's, and note that II All fHJ = i(FAFfj, g)l ~ liF AFII. M 2 , where M = max{llf1 ll, ... , 11.1;.11, llg 1 ll, ... , llgnll}. In the reverse direction, given a projection F of finite rank, let {e 1, ···,en} be an orthonormal basis for ran F and note that IIAII/ ~

L L I(FAFej, eiW ~ L L IIAIIeJoe/·
i

j

i

j

121

PROBLEMS

It follows from the definition of the weak topology via projections that if A 0 is an operator on an infinite-dimensional Hilbert space Hand ifS is a set of operators on H such that every compression of A 0 to a finite-dimensional subspace has a dilation in S, then A 0 itself belongs to the weak closure of S. Indeed: the hypothesis means that for every projection F of finite rank there exists an operator S0 in S such that F A 0 F = FS0 F. The hypothesis implies

(with room to spare) that for each 8 the set
{A: I!F(A - A0 )FII <
8}

meets S, and, consequently, that A 0 belongs to the weak closure of S.

Problem 224. On an infinite-dimensional Hilbert space, what are the weak closures of the sets U (unitary operators), N (normal operators), and P (projections)?
Sometimes it is good to know the weak sequential closures of these sets; for some pertinent information see [98].

225

225. Strong closures and extensions. The strong operator topology stands in the same relation to extension as the weak operator topology to dilation (see Problem 224). Assertion: a base for the strong operator topology is the collection of all sets of the form
{A: II(A- Ao)FII < 8},

where F is a projection of finite rank and 8 is a positive number. An efficient way to prove the assertion is to compare the pseudonorms IIA11 1 defined by vectors with the ones IIAIIF = IIAFII defined by projections of finite rank. Given vectors f 1, • • ·, fn, let F be the projection whose range is the span of all the f's and note that IIAIIJj = IIAFJJII ~ IIAFII·M, where M = max{llf1 11, · .. , llfnll }. In the reverse direction, given a projection F of finite rank, let {e 1, ···,en} be an orthonormal basis for ran F, and note that, for all f, IIAFfll

= IIAfll

= IIA
=

t

(Ff, ei)eill

lit

(f, ei)Aeill

~ llfll· t IIAeill,

so that

122

UNITARY DILATIONS

One application of this approach to the strong topology is a simple proof that the set of nilpotent operators of index 2 is strongly dense; see Solution 111. It is sufficient to prove that every operator A of finite rank belongs to the strong closure of the set of nilpotents of index 2. Let H 0 be a finite-dimensional subspace that includes ran A, let H 1 be a subspace of the same dimension as H 0 and orthogonal to it, and consider the matrix corresponding to A with respect to the decomposition H = H 0 Ei3 H 1 Ei3 (H 0 Ei3 H 1 ).L:

(

Ao 0 0. 0 0 0 0
211AoiiAo
B

* *)

If B is the operator whose matrix with respect to the same decomposition is

0 0 0

-A 0

0

then B is nilpotent of index 2; if F is the projection

(~ ~ ~).
0 0 0
then II(A - B)FII <e. That is: B belongs to the strong neighborhood of A determined bye and F. The technique is based on the nilpotence of all matrices of the form

it is related to the alternative proof of the weak density of the set of normal operators (Solution 224), based on the normality of all matrices of the form

(:* ~*).
Another application of the same technique, based on the idempotence of all matrices of the form

~1 ~A) 1~: A)
(with IX = e/2111 - A II) shows that the set of all idempotent operators is strongly dense. 123

PROBLEMS

The strong density of nilpotents and idempotents is part of the unbounded pathology of operator theory; in bounded sets it cannot happen. That is: the set of all nilpotent contractions of index 2 is strongly closed, and so is the set of all idempotent contractions. Proof: Solution 113. It follows from the definition of the strong topology via projections that if A 0 is an operator on an infinite-dimensional Hilbert space H and S is a set of operators on H such that every restriction of A 0 to a finite-dimensional subspace has an extension in S, then A 0 itself belongs to the strong closure of S. (The restriction of A 0 to a subspace makes sense whether the subspace is invariant under A 0 or not; in any event the restriction is a bounded linear transformation from the subspace into H.) Indeed: the hypothesis means that for every projection F of finite rank there exists an operator S0 in S such that A 0 F = S 0 F. The hypothesis implies, with room to spare, that for each e the set
{A:

II(A - A 0 )FII < e}

meets S, and, consequently, that A 0 belongs to the strong closure of S. As an example, consider the set S of subnormal operators on H. If A 0 is subnormal, then, by definition, A 0 has a normal extension to a larger space K. The same is true, therefore, of every restriction of A 0 to a finitedimensional subspace M of H. Since the pairs (M, K) are obviously isomorphic to the corresponding pairs (M, H), it follows that S is a subset of the strong closure of the set of normal operators on H. The fact is that S is strongly closed (Problem 203), but that's harder to prove.
Problem 225. On an infinite-dimensional Hilbert space, what are the strong closures of the sets U (unitary operators) and P (projections)? What about the set of co-isometries?

226

226. Strong limits of hyponormal operators. Problem 226. What is the strong closure of the set of hyponormal operators?

227

227. Unitary power dilations. The least unitary looking contraction is 0, but even it has a unitary dilation. The construction of Solution 222 exhibits it as

The construction is canonical, in a sense, but it does not have many useful algebraic properties. It is not necessarily true, for instance, that the square of a dilation is a dilation of the square; indeed, the square of the dilation of 0 exhibited above is

124

UNITARY DILATIONS

which is not a dilation of the square of 0. Is there a unitary dilation of 0 that is fair to squares? The answer is yes:

GH)
is an example. The square of this dilation is

(~

~ ~),

1 0 0

which is a dilation of the square of 0. Unfortunately, however, this dilation is not perfect either; its cube is

1 0 0) (
0 1 0 ( 0 0 0 1 0

0 1 0 ' 0 0 1

which is not a dilation of the cube of 0. The cube injustice can be remedied by passage to
0 I) 0 0 0 0 ' 1 0

but then fourth powers fail. There is no end to inductive greed; the clearly suggested final demand is for a unitary dilation of 0 with the property that all its powers are dilations of 0. In matrix language the demand is for a unitary matrix with the property that one of its diagonal entries is 0 and that, moreover, the corresponding entry in all its powers is also 0. Brief meditation on the preceding finite examples, or just inspired guessing, might suggest the answer; the bilateral shift will work, with the (0, 0) entry playing the distinguished role. (Caution: the unilateral shift is not unitary.) The general definition suggested by the preceding considerations is this: an operator B is a power dilation (sometimes called a strong dilation) of an operator A if B" is a dilation of A" for n = I, 2, 3, · · · . Problem 227. Every contraction has a unitary power dilation. In all fairness to dilations, it should be mentioned that they all have at least one useful algebraic property: if B is a dilation of A, then B* is a dilation of A*. The quickest proof is via quadratic forms: if (Af, f) = (Bf, f) for each fin the domain of A, then, for the samej's, (A*f, f)= (f, Af) = (Af, f)* = (Bf, f)* = (f, Bf) = (B*f, f). One consequence of this is that if B is a power dilation of A, then B* is a power dilation of A*.
125

e.PROBLEMS The power dilation theorem was first proved by Nagy [95]. Their definition is similar to that of minimal normal extensions (Problem 197).. it asserts that if U is a unitary operator on a Hilbert space. whose generalizations are widely applicable. and they too are uniquely determined by the given operator (to within unitary equivalence). -l:uJ= . If A is a contraction on a Hilbert space H. Ergodic theorem. in fact. then the averages 1 n-1 -Lui n i=O from a convergent sequence. the projection whose range is the subspace {f: Uf = f}. consider separately the cases u = 1 and u # 1. i. This is an amusing and simple piece of classical analysis. 126 . then is a strongly convergent sequence of operators on H. The subject has received quite a lot of attention since then. To prove the statement. then each average is equal to 1. The most plausible operatorial generalization of the result of the preceding paragraph is known as the mean ergodic theorem for unitary operators. If u is a complex number of modulus I. If u = 1. good summaries of results are in [96] and [94]. then 1" ·1 11 1 1 u" :s. Schreiber [128] proved that all strict contractions (see Problem 153) on separable Hilbert spaces have the same minimal unitary power dilation. An especially interesting aspect of the theory concerns minimal unitary power dilations. then the averages form a strongly convergent sequence. A more informative statement of the ergodic theorem might go on to describe the limit. The curious fact is that knowledge of the minimal unitary power dilation of an operator is not so helpful as one might think. Nagy [97] extended the result to non-separable spaces. If u # 1. It is less obvious that a similar ergodic theorem is true not only for unitary operators but for all contractions. it is. namely a bilateral shift. Problem 228. ni=O n(1-u) -nl1-ui I 1 and the limit is 0. the subspace of fixed points of u. and the limit is 1. 228 228.

The general context to which the theorem belongs is the theory of spectral sets.UNITARY DILATIONS 229. then they can be applied to every operator.£ IIPIIv for every polynomial p. when it does. If II All .£ 1 and if Dis the closed unit disc. If.EM}.) In view of the last remark. the result can be false. or even compact. There is an easy way around both obstacles: consider only such functions F for which F(A) does make sense. however.) It is not obvious how this inequality can be generalized to non-normal operators. a spectral set for A is a set M such that spec A c: M and such that if F is a bounded rational function on M (i. in the role of A. IfF is a bounded complex-valued function defined on a set M. 127 . A viable theory can be built on these special considerations. the inequality between norms will fail. If A is a normal operator with spectrum A. a spectral set for an operator is a set such that the appropriate norm inequality holds for all rational functions on the set. a rational function with no poles in the closure of M). If the only functions considered are polynomials. and that is usually done. von Neumann's inequality. the inequality llp(A)II .)I:A. seriously change the definition. To demand the norm inequality for polynomials only does. then llp(A)II . write IIFIIM = sup{IF(A. for which the inequality does hold.£ IIFIIM· (Note that the condition on the poles of the admissible F's implies that F(A) makes sense for each such F. [70]. and. A moderately sophisticated complex function argument (cf.e. the spectrum of the operator is too small. then llp(A)II = IIAII and IIPIIspecA = 0. There are two obstacles: in general. (Reference: [151].£ liP II spec A holds only if A = 0. A is quasinilpotent and p(z) = z.) 229 Problem 229. however. and consider only such sets.. That theory is concerned with rational functions instead of just polynomials. Roughly speaking. (For this purpose a set is sufficiently simple if it is compact and its complement is connected. is sometimes known as the von Neumann inequality. then IIF(A)II .£ IIFIIA· (Equality does not hold in general. Precisely. F(A) does not make sense. the von Neumann inequality is frequently stated as follows: the closed unit disc is a spectral set for every contraction. then IIF(A)II . and if F is a bounded Borel measurable function on A. [90]) can be used to show that the polynomial definition and the rational function definition are the same in case the set in question is sufficiently simple. F may take a few large values that have no measure-theoretically detectable influence on F(A). lf. The earliest positive result. which is still the most incisive and informative statement along these lines. for instance.) It turns out that the theory loses no generality if the definition of spectral set demands that the set be closed.

commutators on Banach spaces). and the trace of a product of two factors is independent of their order. For the general (not necessarily finite-dimensional) case. + oo) and let P and Q be the differentiation transformation and the position transformation. A mathematical formulation of the famous Heisenberg uncertainty principle is that a certain pair of linear transformations P and Q satisfies. they are due to Wintner 128 .. after suitable normalizations. their domains are far from being the whole space. the equation PQ . quite different from one another.oo. Commutators. It is easy enough to produce a concrete example of this behavior.CHAPTER 24 Commutators 230 230. The finite-dimensional case is easy to settle. the only scalar with trace 0 is 0 itself. and they misbehave in many other ways. These are not bounded linear transformations. consider L 2 ( . [2]).g. The question of the preceding paragraph can be phrased this way: "Is 1 a commutator?" The answer is no. The only scalar commutator is 0. define a commutator as an operator of the form PQ . More is known: in fact a finite square matrix is a commutator if and only if it has trace 0 ([135]. Can this misbehavior be avoided? To phrase the question precisely. (Pf)(x) = f'(x) and (Qf)(x) = xf(x)).QP = I. of course. It follows that the trace of a commutator is always zero. More general uses of the word can be found in the literature (e. Problem 230. Trace is linear. respectively (that is. two beautiful proofs are known. That settles the negative statement. the main thing that it is intended to exclude is the unbounded case. where P and Q are operators on a Hilbert space. and most of them do not conflict with the present definition. The reason is that in that case the concept of trace is available.QP.

and.) What about a projection (on an infinite-dimensional Hilbert space) with nullity 1. the most surprising fact in this subject is that on separable spaces the condition is necessary also [22]. such that liPnI ~ 1:1. it is the unique projection with nullity 0.QnPn)ll --+ 0 as n--+ oo? The Brown-Pearcy characterization of commutators (see Problem 230) implies that the answer is yes. The corollary gives a sufficient condition that an operator be a noncommutator. The quotient algebra is a normed algebra.QnPn} converges in the norm to an operator C. Granted that the identity is not a commutator. Corollary. as follows. and II Qn I ~ 1:1. Consider the normed algebra of all operators and in it the ideal of compact operators. Since a projection with nullity 1 is a very special example of such a sum. an element e such that ef = fe = f for all f. with no change. this means that the identity cannot be equal to the sum of a commutator and a compact operator. In that algebra the unit element is not a commutator (by Wintner and Wielandt). If {Pn} and {Qn} are bounded sequences of operators (i.e. (Recall that the nullity of an operator is the dimension of its kernel. if there exists a positive number 1:1. The sum of a compact operator and a non-zero scalar is not a commutator. The proof is not short. The algebraic character of the Wintner and Wielandt proofs can be used to get more information about commutators.COMMUTATORS [160] and Wielandt [158].. then C :f:. for all n). 231. is it at least a limit of commutators? Do there. moreover. The following statement summarizes what the proof proves. A unit in a normed algebra is. the proof is complete. intuition is right [55]. Problem 231. In other words: the identity cannot be the limit of commutators formed from bounded sequences. and ifthe sequence {PnQn. to arbitrary complex normed algebras with unit. exist sequences {Pn} and {Qn} of operators such that Ill . of course. 1. Reference: [20]. In other words: on a separable space every operator that is not the sum of a non-zero scalar and a compact operator is a commutator.) A more modest result is more easily accessible. The identity is a projection. A normed algebra is a normed vector space that is at the same time an algebra such that llfgll ~ llfll·llgll for all f and g. that II ell = 1. translated back to operators. Both apply. can it be a commutator? Intuition cries out for a negative answer. it is customary to require. 129 231 . in other words. Limits of commutators.(PnQn . for once. (See also Problem 235.

who proved. it follows that pn = 0 for some n. then 111 .QP and C commutes with P. then an easy inductive argument proves that P"Q . 211Pnii·IIQII for every positive integer n. Operators with large kernels. The first general theorem of this sort is due to Jacobson [79]. then C is quasinilpotent. (b) If C commutes with P. or if P is normal). however. Problem 232. commutators can come arbitrarily near to 1.QP) is nilpotent. Commutators with this sort of commutativity property have received some attention.. 233 233.q ~ 1.. under suitable finiteness assumptions.e. all the results of the preceding problems are negative. Problem 233. that if C = PQ . 234 234. Example: p = (~ ~). then C = 0.. Kaplansky conjectured that if nilpotence is replaced by its appropriate generalization.QP and if P is hyponormal (hence. commutators cannot be equal to 1. In infinitedimensional Hilbert spaces finiteness conditions are not likely to be satisfied.QP = 1?) fits into the context of their theory. By Wintner and Wielandt.q ~ 1. then the Jacobson theorem will extend to operators. Since it is impossible that n . in particular. This is a not unreasonable generalization of the theorem about scalars. An easy way for PQ . and he turned out to be right. Distance from a commutator to the identity. they all say that something is not a commutator. ifC = PQ. and ifC commutes with P. 211QII for all n. How does the proof use the assumption that Cis a scalar? An examination of Wielandt's proof suggests at least part of the answer: it is important that C commutes with P. Kleinecke-Shirokov theorem. that P ( = PQ . i. The result of Problem 230 says that if C = PQ .QP.q ~ 1 for all commutators C. Usually. 130 . after all the only nilpotent scalar is 0. If P and Q are operators.. the original question (PQ . then Ill . independently. a commutator is anxious to stay far from 1. If the underlying Hilbert space is finite-dimensional.QP to commute with Pis for it to be equal to P. by Kleinecke [88] and Shirokov [134]...QP and if C is a scalar. then C is nilpotent. quasinilpotence.QPn = nPn. If that happens. if P is an isometry. The proof was discovered. (a) If C = PQ . and this implies that niiPnll . Q= (~ ~). then it is an easy exercise in linear algebra to prove that Ill .. by Brown-Pearcy.PROBLEMS 232 232. As far as the construction of com- mutators is concerned.

) In this language. If. conversely. 53]. If. even if the word has not. A is an arbitrary operator on H (it could even be the identity). If the matrix of an operator on H E9 H is then that operator has a large kernel.J defines an operator.QP ~~~ 0 0 0 0 ~J Since the direct sum of infinitely many copies of H is the direct sum of the first copy and the others. suppose that His an infinite-dimensional Hilbert space and consider the infinite direct sum H E9 H E9 H E9 · · · . if His infinite-dimensional. in particular. Operators on this large space can be represented as infinite matrices whose entries are operators on H. and. moreover. if 0 1 Q= 0 0 0 0 0 0 1 0 0 1 ~J 0 0 0 0 then it can be painlessly verified that PQ. then H (regarded as one of the axes of the direct sum H E9 H) is a large subspace of H E9 H. that kernel reduces A. cf. an operator on an infinite-dimensional Hilbert space has a 131 . then the matrix p ~(~ A 0 0 0 0 A 0 0 . it follows that every two-by-two operator matrix of the form is a commutator [52. It is worth while reformulating the result without matrices. Problem 142. and since the direct sum of the others is isomorphic (unitarily equivalent) to H.COMMUTATORS To get a positive result. Call a subspace M of a Hilbert space H large if dim M = dim H. (The idea has appeared before.

Thus. then the infinite direct sum A EB A EB A EB · · · is a commutator. each non-empty compact subset of the plane (i. The reason is that a linear functional that deserves the name "trace" must vanish on all commutators. Corollary 1. Another immediate corollary is that the identity is the limit (in the norm) of commutators. Direct sums as commutators.) In view of these remarks the matrix result of the preceding paragraph can be formulated as follows: every operator with a large reducing kernel is a commutator.The theory of self-commutators has some interest.AA *. it answers many of the obvious questions about them.. This result can be improved [106]. for instance. Corollary 2. The techniques needed for the proof contain the germ (a very rudimentary germ. compare Problems 231 and 233. identically. and hence. any set that can be a spectrum at all) is the spectrum of some commutator. The self-commutator of an operator A is the operator A* A . Corollary 2 shows that nothing like a trace can exist on the algebra of all operators on an infinite-dimensional Hilbert space. 236 236. Ifan operator A on a separable Hilbert space is not a scalar. Every operator with a large kernel is a commutator. Problem 234. Every operator on an infinite-dimensional Hilbert space is the sum of two commutators. If the dimension of that orthogonal complement is too small. Even though this result is far from a complete characterization of commutators. then that operator can be represented by a matrix of the form (Represent the space as the direct sum of the kernel and its orthogonal complement. by Corollary 2.PROBLEMS large reducing kernel. to be sure) of what is needed for the general characterization of commutators [22]. more precisely. enlarge it by adjoining "half" the kernel. Positive self-commutators. It is known that a finite square matrix is a self-commutator if and only if it is 132 . Problem 235.e. it is an immediate corollary that the spectrum of a commutator is quite arbitrary. On an infinite-dimensional Hilbert space commutators are strongly dense. 235 235.

In ring theory it means PQ . A positive self-commutator cannot be invertible. by Problem 236. Can C be a projection. in that case C = 0. It is natural to ask just how positive a self-commutator can be. Problem 236.1 (multiplicative commutators). If A is a nonnormal isometry (i. The interesting question here concerns what may be called abnormal operators. the most familar ones are the projections.COMMUTATORS Hermitian and has trace 0 ([140]). A little judicious guessing about trace versus determinant.AA * = 1 . is strictly infinite-dimensional). and. the direct sum of a unilateral shift of non-zero multiplicity and a unitary operator-see Problem 149). operators that have no normal direct summands. It is not difficult to produce an example of an abnormal operator whose self-commutator is a projection: the unilateral shift will do. Among the positive operators with non-trivial kernels. and. and the answer is not very. then I A I = 1 and C = A* A . Some of those analogues are true. What is interesting is that in the presence of the norm condition (II All = 1) this is the only way to produce examples. such that A* A . then A is an isometry. The word "commutator" occurs in two distinct mathematical contexts. Reference: [108]. if so. how? The most obvious way for C to be a projection is for A to be normal. 237. however. in group theory it means PQP. is likely to lead to the formulation of multiplicative analogues of the results about additive commutators.e. Whatever other ways there might be. Problem 237. (a) If A is an abnormal operator of norm 1.AA * is a projection.AA * reduces A. Otherwise said. they can always be combined with a normal operator (direct sum) to yield still another way.. (b) Does the statement remain true if the norm condition is not assumed? 238. The easiest way for C to be not invertible is to have a non-trivial kernel. i. Cis not invertible. Multiplicative commutators.AA * is the projection onto the kernel of A*.AA * 237 is positive. is only trivially different. then. What about the analogue of the additive theorem according to which the only scalar that is an additive commutator is 0? 238 133 . by the way.QP (additive commutators). If a self-commutator C = A* A .e. A is abnormal if no non-zero subspace of the kernel of A* A . more generally. An obvious place where self-commutators could enter is in the theory of hyponormal operators. about logarithm versus exponential. Projections as self-commutators. which. a necessary and sufficient condition that A be hyponormal is that the self-commutator of A be positive.1 Q. That self-commutators can be non-trivially positive is a relatively rare phenomenon (which..

in other words. The positive assertion of Problem 238 can be greatly strengthened. Unitary multiplicative commutators. The commutator subgroup of a group is the smallest subgroup that contains all elements of the form PQP. If H is an infinite-dimensional Hilbert space. The determinant of a multiplicative commutator is 1. The set of all invertible operators on a Hilbert space is a multiplicative group. It turns out that the necessity proof is algebraic. in analogy with standard finite-dimensional_terminology. in the sense that it yields the same necessary condition for an arbitrary complex normed algebra with unit. 240 240. it follows. For finite-dimensional spaces determinants can be brought into play. it is the subgroup generated by all commutators (multiplicative ones. One of the biggest steps toward the strengthened theory is the following assertion. it may be called the full linear group of the space. Problem 240. of course).1 Q. From this. that if a commutator is congruent to a scalar modulo the ideal of compact operators. a modification of the argument that works for infinite-dimensional spaces shows that the condition is sufficient as well.1 . then a necessary and sufficient condition that a scalar a acting on H be a multiplicative commutator is that Ia I = 1. just as in the additive theory. and the only scalars whose determinants are 1 are the roots of unity of order equal to the dimension of the space. then that scalar must have modulus 1. What is the commutator subgroup ofthefulllinear group of an infinite-dimensional Hilbert space? 134 . This proves that on an n-dimensional space a necessary condition for a scalar a to be a multiplicative commutator is that a" = 1. Problem 239. 239 239.PROBLEMS Problem 238. On an infinite-dimensional Hilbert space every unitary operator is a multiplicative commutator. just as in the additive theory. in turn. Commutator subgroup.

j) such that A. ···). A necessary and sufficient condition that an operator on L 2 be a Laurent operator L"' is that its matrix (A.CHAPTER 25 Toeplitz Operators 241.In words: a Laurent matrix is one all of whose diagonals (parallel to the main diagonal) are constants. · · ·) has a simple form. Problem 241. where <p = Ln anen is the Fourier expansion of (p. Laurent operators and matrices. If <pis a bounded measurable function on the unit circle. define a Laurent matrix as a (bilaterally) infinite matrix 0-. 242. · · ·} be a Laurent matrix. if that condition is satisfied.) Multiplications are. To describe the relation.ii = IX.+ t. and H 2 is a distinguished 135 .ii for all i and} (=0. those about numerical range. moreover. then A.g. ± 1. Laurent operators (multiplications) 242 are distinguished operators on L2 (of the unit circle). in symbols L"'. n = 0. and aside from the pathology of the uncountable. Multiplications are the prototypes of 241 normal operators.-i. and spectrum) have obvious answers. and most of the obvious questions about them (e. norm.. not too sensitive to a change of space. ± 2. ± 2.j+ t = A. ±2. (This is not to say that every question about them has been answered. ± 1. then the multiplication induced by <p on L 2 (with respect to normalized Lebesgue measure Jl) is sometimes called the Laurent operator induced by <p.ii) with respect to the basis {en: n = 0.. The matrix of L"' with respect to the familiar standard orthonormal basis in L 2 (en(z) = z". ± 1. elegantly related to <p. aside from the slightly fussy combinatorics of atoms. what happens on the unit interval or the unit circle is typical of what can happen anywhere. Toeplitz operators and matrices.

The description of what happens is called the theory of Toeplitz operators. A necessary and sufficient condition that an operator A on H 2 be a Toeplitz operator is that U* AU = A. A necessary and sufficient condition that an operator on H 2 be a Toeplitz operator T'P is that its matrix O. The structural differences between the Laurent theory and the Toeplitz theory are profound. but for Toeplitz matrices they go forward only. · · ·). i. the matrix of a Laurent operator with respect to that basis has an especially simple form.e. characterizes Toeplitz operators. Problem 242. if that condition is satisfied. it is just that the compressed operator has the compressed matrix.i+ l. 2. then A. then the Toeplitz operator T'P induced by <p is defined by T'Pf = P(q. In words: a Toeplitz matrix is one all of whose diagonals (parallel to the main diagonal) are constants. note that the statement of Problem 242 implies that the correspondence <p f-+ T'P is one-to-one. 1.ii = ai _i.PROBLEMS subspace of L2 . Since W is unitary. The first. The second. Corollary 1. the simplest non-trivial example of a Toeplitz operator is the unilateral shift U ( = I. The corresponding statements are true about H 2 and Toeplitz operators. U* AU = A. something interesting is bound to happen if Laurent operators are compressed to H 2 .J There is a natural basis in L 2 . The corresponding equations for U say quite different things. in case q... ·f) for all fin H 2 . for Laurent matrices both indices go both ways from 0. is analytic (see Problem 33). and if <pis a bounded measurable function. correspondingly. but the difference between the two kinds of matrices is superficial and easy to describe. is called analytic in case q.j+ 1 = A. The unilateral shift U does for Toeplitz operators what the bilateral shift W does for Laurent operators-but does it differently. AU = U A. · · ·} be a Toeplitz matrix. The simplest non-trivial example of a Laurent operator is the bilateral shift W ( = Le).ii> with respect to the basis {en: n = 0. = Ln an en is the Fourier expansion of q. The necessity of the condition should not be surprising: in terms of an undefined but self-explanatory phrase. where q. The Toeplitz operator T'P induced by q. To state them. 1.) Observe that an analytic Toeplitz operator 136 .ii) such that A.. there is no difference between W* A W = A and A W = W A. define a Toeplitz matrix as a (unilaterally) infinite matrix (A.ij for all i and j ( = 0. characterizes analytic Toeplitz operators (see Problem 147). Explicitly: if P is the projection from L 2 onto H 2 . 2. is not only in L 00 but in H 00 • (To justify the definition.

Toeplitz products. A counterexample for both assertions is given by the unilateral shift and its adjoint. indeed T<P = T<P * if and only if q> = cp*. it follows that T<P is positive if and only if (L<Pf. The algebraic structure of the set of all Laurent 243 operators holds no surprises: everything is true and everything is easy. with fin H 2 . but some are hard. and since the product w. The only compact Toeplitz operator is 0. Since W commutes with L<P. and an isometry (supremum norm to operator norm).) The argument that proved that the set of all Laurent operators is weakly closed works for Toeplitz operators too. linear operations. linear operations. everything is settled. The mapping q> ~---+ L<P from bounded measurable functions to operators is an algebraic homomorphism (it preserves unit. it follows that the set of all Laurent operators is weakly (and hence strongly) closed. and conjugation). Some of the corresponding Toeplitz statements are true and easy. Indeed. wnf) ~ 0 whenever f E Hz. wnf) ~ 0 whenever f E H 2 (and n is an arbitrary integer). which is q> ~---+ T<P. or unknown. Reference: [19]. The easiest statements about the multiplicative properties of Toeplitz operators are negative: the set of all Toeplitz operators is certainly not commutative and certainly not closed under multiplication. is dense in L 2 . Since the set of all W)'s. 243. since (T<Pf. The easiest statements concern unit. and hence to q> ~ 0. Both U and U* are Toeplitz operators.1 A W by U* AU (cf. It is also true that T<P is positive if and only if q> is positive. f) ~ 0 for all fin Hz. Corollary 1 of Problem 242). multiplication. the condition can also be expressed in this form: (L<P W). Since the Laurent operators constitute the commutant of W (Problem 146). Alternatively. it is not only a compression but a restriction of the corresponding Laurent operator. f) whenever f E Hz. Corollary 2.1 A W is weakly continuous in its middle factor. but the product U*U (which is equal to the Toeplitz operator 1) is not the same as the product UU* (which is not a Toeplitz operator). and conjugation: since both the mappings q> ~---+ L<P and 2 L<P~---+(PL<P)IH 2 (=the restriction of PL<P to H = T<P) preserve the algebraic structures named. One way to prove that UU* is not a Toeplitz operator is to use Corollary 1 of Problem 242: since U*(UU*)U = (U*U)(U*U) = 1 ( =/: UU*). It is a trivial consequence of the preceding paragraph that a Toeplitz operator T<P is Hermitian if and only if q> is real. (The preservation of adjunction is true for compressions in general. 137 . just replace w. f) = (L<Pf. the condition is equivalent to L<P ~ 0. When is the product of two Toeplitz operators a Toeplitz operator? The answer is: rarely. the spectrum of L<P is the essential range of q>. the same is true of their composite. see Problem 227.TOEPLITZ OPERATORS is subnormal. The latter condition is equivalent to this one: (WnL<Pf. or false. this negative result could have been obtained via Problem 242 by a direct look at the matrix of UU*.

unfortunately the converse is false. This is equivalent to Lip being bounded from below and hence to cp being bounded away from 0. The result raises a hope that it is necessary to nip in the bud. so that 0 f/: ll(Tip). but Tip e0 = Pe _ 1 = 0. 244 244.Tipi/J is compact? 245 245. Problem 245 can be used to show that in some respects Toeplitz operators behave as if they were normal. This is a spectral inclusion theorem. does it follow that TipTI/J. The Toeplitz mapping cp H Tip from functions to operators is not multiplicative. for instance. cp = e _ t. Although the spectral behavior of Toeplitz operators is relatively bad. If Tip is bounded from below. Corollary. if the condition is satisfied. then the spectral structure of Tip would be much more easily predictable from cp than in fact it is. here. by Problem 245. the "larger" operator has the smaller spectrum. Corollary 1. but there is a relatively easy inequality ([68]) that answers some of the natural questions. then ll(L) c ll(T). Concisely: among the Toeplitz operators there are no zero-divisors. If cp and t/J are in L 00 . Here are some samples. indeed. As for the norm of Tip. Questions about the norms and the spectra of Toeplitz operators are considerably more difficult than those for Laurent operators. About the spectrum ofT nothing is obvious. too. then r(Tip) = IITipll = llcplloo· 138 . then Tip T"' = Tipi/J. is it at least multiplicative modulo compact operators? Problem 244. Problem 243 says that the product of two Toeplitz operators is a Toeplitz operator if and only if the first factor is co-analytic or the second one is analytic. Problem 245. Spectral inclusion theorem for Toeplitz operators. If L and Tare the Laurent and the Toeplitz operators induced by a bounded measurable function. The Toeplitz operator Tip induced by cp is called co-analytic in case cp is co-analytic (see Problem 33). A necessary and sufficient condition that the product of two Toeplitz operators be zero is that at least one factor be zero.PROBLEMS Problem 243. In this language. so that Tip has a non-trivial kernel. If. then cp is bounded away from 0. 0 f/: ll(Lip). then. A necessary and sufficient condition that the product Tip T"' of two Toeplitz operators be a Toeplitz operator is that either cp* or t/J be analytic. that much is obvious because Tis a compression of L. If the converse were true. If cp is a bounded measurable function. formally similar to Problem 200. all that is obvious at first glance is that II Tip II ~ II Lip II ( = llcpll 00 ). Compact Toeplitz products.

TOEPLITZ OPERATORS Corollary 1 says. Problem 242) that that correspondence is one-to-one. 248. then the spectrum of L"' is the essential range of <p.. Problem 247. 246. determine the point spectrum of the Hermitian Toeplitz operator T"'. Corollary 3. The multiplicative properties ofT. The spectral behavior ofT"' is influenced by the complex analytic behavior of (prather than by the merely set-theoretic behavior of <p. Corollary 4. Riesz theorem (Problem 158) implies that either <p = J. For Hermitian Toeplitz operators this reasoning does not apply: there is nothing to stop a non-constant real-valued function from being constant on a set of positive measure. then the spectrum of T"' is what may be called the essential range of{(>. Continuous Toeplitz products. that the correspondence <p f---+ T. are bad both absolutely and relatively (relative to compact operators. then the F. 246 247 Here is still another way to express the result. if <p E H 00 . Eigenvalues of Hermitian Toeplitz operators. There are no quasinilpotent Toeplitz operators other than 0. does it follow that T"' Tt/1 . If <p E H 00 . among other things. that is). Analytic Toeplitz operators. If <pEL 00 .f for some fin H 2 . The operative word is "analytic". The example in Solution 244 used discontinuous functions. 248 139 .T"'"' is compact? 247. then the spectrum ofT"' is the closure of the image of the open unit disc D under the corresponding element{(> ofH 00 . and M. Roughly speaking.. in other words spec T"' = {(>(D). is norm-preserving. The closure of the numerical range of a Toeplitz operator is the convex hull of its spectrum. the reason is that an analytic function cannot take a constant value on a set of positive measure without being a constant. The easiest Toeplitz operators are the analytic ones.. or f = 0. Can an analytic Toeplitz operator have an eigenvalue? Except in the trivial case of scalar operators. Corollary 2. this recaptures the result (cf. Given a real-valued function <p in L oo. Reference: [159]. the answer is no. Problem 248. Recall that associated with each <pin H 00 there is a function {(>analytic in the open unit disc D (see Problem 35). The reason is that if qJ is analytic and qJ • f = J. If <p and ljJ are continuous. is that what made it work? Problem 246. Every Toeplitz operator with a real spectrum is Hermitian. but even for them much more care is needed than for multiplications.

Zero-divisors. Example: if A = 1 (on an infinite-dimensional space) and B = diag(l. ! . that with this interpretation the answer is no for Toeplitz operators. to rule out this sort of construction. At least one question pertinent to the present context still remains: what about Toeplitz operators? Problem 249. then A EB B and B EB A are non-compact Hermitian operators with kernel 0. The answer. is still the same. the approach is right. · · ·) (on the same space). it is still yes.. rule out eigenvalues. Another way to interpret "small" is as "compact". however. Given a real-valued function cp in L oo. 140 . Unless. Indeed: there are noncompact zero-divisors. but the answer for operators in general doesn't change. and still easy. t. however.g. whose product is compact. however. One way to make the question more challenging might seem to be to insist that the factors be" good" operators. and in that case the answer is trivially yes: there are operators that are zero-divisors. Spectrum of a Hermitian Toeplitz operator. Very well. Hermitian operators with kernel 0. but it turns out that it is still yes. the construction can be quite laborious. Do there exist two Hermitian operators with no eigenvalues whose product is compact? Now at last the answer is not quite so near the surface. e.PROBLEMS 249 249. determine the spectrum of the Hermitian Toeplitz operator T"'. It is relevant to recall. see Widom [154. A trivial way is to interpret "small" to mean "zero". 155]. For more recent and more general studies of the spectra of Toeplitz operators. Do there exist non-zero Toeplitz operators whose product is compact? Equivalently: are there Toeplitz zero-divisors modulo the compact operators? 250 250. Problem 250. Can operators that are not small have small products? There are many ways to interpret the question.

HINTS .

.

is to contemplate spectral measures. Determine the orthogonal complement of the span. Chapter 2. Problem 3. If g +hE L. Problem 14. Consider characteristic functions in L 2 (0. Problem 7. l(ei. For the converse: use Solution 14. A countably infinite set has an uncountable collection of infinite subsets such that the intersection of two distinct ones among them is always finite. Polarize. then gEL n M. Problem 2.. Consider the differences normal basis{···. Problem 12. Problem 5. Iff and g are normalized crinkled arcs. Problem 4. Prove that M + N is complete.)l = l(e. e_ 1 . Problem 11. Use uniqueness: iff = Li rxiei. In an infinite-dimensional space there always exist two subspaces whose vector sum is different from their span. An alternative hint. Use inner products to reduce the problem to the strict convexity of the unit disc. · · ·}. Omit an infinite subset by omitting one element at a time. jj)l. then ~(f)= Li rxi~(e). Spaces Problem 13. Problem 8. Use Solution 10. e0 . Consider a perturbation of the given form by a small multiple of a strictly positive one. Problem 10. en en-l formed from an ortho- Problem 9.jj. There is no loss of generality in assuming that dim M = 1. e 1 . 1). define cp on [0. 1] so that llg(t)ll = II f(cp(t))ll.Chapter 1. Vectors Problem 1. Problem 6.. Problem 15. 143 . with gin M and h inN. for those who know about spectral measures./. e.

given an orthonormal basis {e 1 . Expand I fn .define a metric by d(f. Chaper 3. Problem 26. ···}. Consider the set of all complex-valued functions ~ on H such that I~(f) I ~ I f I for aUf. Problem 24. How many basis elements can an open ball of diameter Problem 17. Problem 18. add suitable multiples of orthonormal vectors to convert it to a unit vector. The span of a weakly dense set is the whole space. g)=~ J 1 2 i l(f. Given a countable dense set. e) I· Problem 25. then it is weakly separable. and show that the linear functionals of norm less than or equal to 1 form a closed subset. Given a countable dense set. Problem 23. use rational coefficients. If the unit ball is weakly metrizable. with g =f. e3 .g. Use the definition of weak convergence. g). show that the resulting collection of neighborhoods is a base for the weak topology. Un. Given a vector in the open unit ball.HINTS J2 contain? Problem 16. endowed with the product topology. Weak Topology Problem 19. Caution: is weak closure the same as weak sequential closure? Problem 20. using finite subsets of itself for the vector parameters and reciprocals of positive integers for the numerical parameters of the neighborhoods. Alternatively. Consider orthonormal sets.f 1 2 • Problem 21. Problem 22. define all possible basic weak neighborhoods of each of its elements. approximate each element of a basis close enough to exclude all other basis elements. Given a countable basis. g)-+ (f. e 2 . 144 . Fit infinitely many balls of the same radius inside any given ball of positive radius.

Is conjugation continuous? Problem 34. Construct a sequence that has a weak cluster point but whose norms tend to oo. The value of an analytic function at the center of a disc is equal to its average over the disc. Consider partial sums and use the principle of uniform boundedness. What is the connection between the concepts of convergence appropriate to power series and Fourier series? Problem 33. use a Hamel basis to construct a Cauchy net {g 1 } such that (f. Problem 28. This implies that evaluation at a point of D is a bounded linear functional on A2 (D). Chapter 4. Problem 29. and hence that Cauchy sequences in the norm are Cauchy sequences in the sense of uniform convergence on compact sets. If the conclusion is false. 145 . gn) defines a bounded linear functional. an orthonormal sequence such that the inner product of each term with a suitable element of the given weakly bounded set is very large. g1 )--+ ~(f) for each f. Problem 32. Analytic Functions Problem 31. then form a suitable (infinite) linear combination of the terms of that orthonormal sequence.ANALYTIC FUNCTIONS Problem 27. (a) Given an unbounded linear functional ~. A necessary and sufficient condition that is that the numbers (0 < r < 1) be bounded. Use continuity of the partial sums at r = 1. then ~(f) = limn (f. inductively. Problem 30. Is the Fourier series of a product the same as the formal product of the Fourier series? Problem 35. (b) If {gn} is a weak Cauchy sequence. then construct.

Problem 38. Problem 45. Look at the sum of the squares of the matrix entries. Examine the finite-dimensional case to see whether or not the isometry of conjugation implies any restriction on the kernel function. Problem 37. = q. and then try to invert the process. f. Problem 43. Chapter 5. Approximate panding concentric circles. by the values of J on ex- Problem 41. Write L aiiei as c r:: ~ Ii <vv. in the norm. Start with a well-behaved functional Hilbert space and adjoin a point to its domain. Use the maximum modulus principle and Fejer's theorem about the Cesaro convergence of Fourier series. Apply Problem 45 with P. Problem 42. Problem 40. Problem 46. Given the Fourier expansion of an element ofH 2 . first find the Fourier expansion of its real part. Assume that one factor is bounded and use Problem 34. and apply the Schwarz inequality. = 1/Ji+l. Is it a Gramian? 146 . Treat the case of dimension ~ 0 only. ensure that it is a basis by choosing every other element of it so that the span of that element and its predecessors includes the successive terms of a prescribed basis.HINTS Problem 36. Problem 39. use the general expression of a kernel function as a series. To evaluate the Bergman and the Szeg6 kernels.ij vq) (~'j) . Problem 47. Note that the operator has rank 1. Infinite Matrices Problem 44. Construct the desired orthonormal set inductively. Problem 48. Use the kernel function of H2 .

(a) Use the x-axis and the graph of an operator. then there must be a row consisting of nothing but O's. Problem 59. if ~ 0 ~dim K < dim H. Prove that A* is bounded from below by proving that the in verse image under A* of the unit sphere in H is bounded. to prove boundedness. For (a) and (b). For a counterexample. for (c) imitate Solution 27. use the closed graph theorem. Problem 54. Problem 56. For a proof. Problem 60. for (d) use a matrix with a large but finite first row. Write the given linear transformation as a matrix (with respect to orthonormal bases of Hand K). Use Problem 54. apply either the closed graph theorem or the principle of uniform boundedness.ROUNDEDNESS AND INVERTIBILITY Problem 49. (b) Form infinite direct sums. Consider the Gramian matrix ((jj. diminish the graph by restricting the operator to a subspace. The equation Af = BXf uniquely determines an Xf in ker. extend an orthonormal basis to a Hamel basis. Apply Problem 52 to the mapping that projects the graph onto the domain. Problem 55. Problem 57. Use the principle of uniform boundedness. 147 . look at unbounded diagonal matrices. Problem 58.LB. J. Roundedness and Invertibility Problem 50. Chapter 6. Problem 51. Problem 52. Apply the principle of uniform boundedness for linear functionals twice.)). Consider the graph of a linear transformation that maps a Hamel basis of a separable Hilbert space onto an orthonormal basis of a non-separable one. Use Problem 56. Problem 53.

HINTS Chapter 7. Multiplication Operators Problem 61. For the boundedness of the multiplier. Necessity: if an operator commutes with each entry of the matrix. this implies that Icp(x) I ~ 1 whenever j(x) :j:. then IIA/11 ~ (llcplloo. If e > 0 and if f is the characteristic function of a set of positive finite measure on which Icp(x) I > I cp I 00 . then I cpn · f I ~ II fll for every positive integer nand for every fin L 2 . Chapter 8. Problem 71. Problem 66. Problem 67. Problem 64. Imitate the discrete case (Solution 63). lcxil = IIAeill and 2 · ~ lcxi~il 2 ~ (s~p lcxi1) ~ l~ii 2 J J J Problem 62.e)·llfll. Operator Matrices Problem 70. The inverse operator must send en onto (1/ctn)en. Problem 68. Multiply on the right by 148 . 1] whose derivatives belong to L2 . 0. Consider the set of all those absolutely continuous functions on [0. Iflcxnl ~ n. Sufficiency: use Cramer's rule. imitate the "slick" proof in Solution 65.e. then it commutes with each entry of the inverse. then there exists an fin H such that f(x) :j:. Imitate the discrete case (Solution 62). or prove that a multiplication is necessarily closed and apply the closed graph theorem. use the closed graph theorem. Problem 69. For the boundedness of the multiplication. then the sequence {1/cxn} belongs to 12 • Problem 63. Problem 65. If II A II = 1. assume that if x EX. 0.

).AB) .EXAMPLES OF SPECTRA with T chosen so as to annihilate the lower left entry of the product. Problem 76. Use the same technique for IT.An) .1f. and its adjoint. · ·) ~ <O. Problem 72. To prove IT 0 (p(A)) c: p(IT 0 (A)). For r: the range of p. apply the result with A* in place of A.ll are bounded from below by 1/11 Pll. Problem 74. -+ A. ~2.A. Examples of Spectra Problem 79.. then (A . A. Problem 77. Problem 80.1 into a geometric series. given a in IT 0 (p(A)). Suppose that An ¢ spec A. E spec A. · · ·).A.1 AP is included in the image under p. The kernel of an operator is the orthogonal complement of the range of its adjoint. f =I= Problem 78. Chapter 9. 149 . then <p = A. Problem75.a.n) 1/11 . ~o' ~1.)(A .1 of the range of A. factor p(A. Use Problem 79. If <p · f = A. Problem 81.) . for r.A. then it is invariant under the inverse. whenever f =I= 0.f almost everywhere. Properties of Spectra Problem 73. and An 0 andf l_ ran(A . Pretend that it is legitimate to expand (1 . If A is normal.1/-+ O II(A . If a finite-dimensional subspace is invariant under an invertible operator. Prove that the complement is open. ~2. For IT: if llfnll=l. then the numbers IIP. and. then IT 0 (A) = (IT 0 (A *))*. If Chapter 10. Look 2 for counterexamples formed out of the operator on 1 defined by <~o' ~1.

Compute that TI 0 (U) is empty and TI 0 (U*) is the open unit disc.HINTS Problem 82. Spectral Radius Problem 86. Imitate the coordinate technique used for the unweighted unilateral shift. define the multiplier as the conjugate of the corresponding eigenvalue.A is bounded from below. Problem 88. for each fin that domain.1 BS. Problem 90. Problem 93. then they all belong to the same eigenvalue of A. then the matrix of S must be lower triangular. Apply Liouville's theorem on bounded entire functions to the resolvent. For the norm: S is an isometry. Write r(A) = (A. find the matrix entries in row n + 1. ~ 2 . If A = s.o)r. 2.1 0 I 00 ((A. • ·-) = (~ 1 . Look for a diagonal operator D such that AD= DB. · · · . and then use the principle of uniform boundedness. ~ 1 . Problem 91. Use the analyticity of the resolvent to conclude that r is analytic for IAI < 1/r(A). If all the vectors in a convex subset of a Hilbert space are eigenvectors of an operator A. Problem 83."I1) -1 . Problem85. Verify that U*(~ 0 . 1. 150 .1(A0 A. Problem 89. then U .A0 I is sufficiently small. n=O Problem 87. If A0 is not in the spectrum of A and if IA . Look at weighted shifts. Use a spanning set of eigenvectors of A* for the domain. Problem 92. If IAI < 1.A ). ~ 2 . ~3• • • ·). column n. Represent W as a multiplication. Chapter 11. n = 0. Problem 84. and therefore liP I = IISPII· For the spectral radius: use Problem 88.A ). then PA(A) =(A.

Uf = • . Approximate a weighted unilateral shift with positive spectral radius by weighted shifts with enough zero weights to make them nilpotent. In the infinite-dimensional case. 2 and prove that U is an isometry from l 2 (p) onto 1 that transforms the shift 2 2 on 1 (p) onto a weighted shift on 1 • Problem 96. Alternatively.. 1). Consider a sequence dense in an interval and let the weights contain arbitrarily long blocks of each term of the sequence. Find the spectral radius of both Ak and Ak. apply Solution 91. Try unilateral weighted shifts with infinitely many zero weights. Iff = (~ 0 . Problem 98. to the set of singular operators is positive on the complement of spec A. write <$o~o' JP. Problem 105. Problem 106. ~ 2 . Problem lOt. factor out the largest possible power of z fromf(z).AA 0 .t = (A 0 . Think of projections on L 2 (0. for normal operators. Problem 100. If A 0 is invertible.A)A 0 .1 . use the geometric series trick to prove that A is invertible and to obtain a bound on IIA . apply Solution 91. The distance from A . ¢ liminfn spec An? Recall that. If f(A) = 0. ) E l 2 (p). Problem 95. Norm Topology Problem 99. Try unilateral weighted shifts. use the fact that the set of singular operators is closed. the reciprocal of a bound is a suitable e. What does it mean that A. ~ 1 . Problem 104. Add a small scalar.1 11.~t• -JP. spectral radius is equal to norm. the norm of the resolvent is bounded on the complement of A0 .A. modify Solution 104. then 1 . Problem 97. For a sequential proof. Problem 103.NORM TOPOLOGY Problem 94.1 . Problem 102.~2• ·· ·). 151 . Chapter 12. In the finite-dimensional case spectrum is continuous.

Use Problem 20. assume that I(Anf. Problem 111. Let U be the unilateral shift. Chapter 14. Problem 112. Problem 117. consider the projections onto a decreasing sequence of subspaces. If {An} is increasing and converges to A weakly. The set of all nilpotent operators of index 2 is strongly dense. (a) Use the principle of uniform boundedness. Consider an increasing sequence of projections. g) I < s for all unit vectors g. Problem 119. imitate the first part. Use nets. Use the slip-away technique of Solution 114. use suitable multiples of powers of U* for the A's. then the positive square root of A . For the second part. For the first part. Problem 109. Problem 120. Problem 114. For a counterexample with respect to the strong topology. (~ ~) with one axis nailed Problem 115. Problem 118. Problem 110. For a counterexample with respect to the strong topology. Problem 113. use suitable multiples of powers of U for the B's.An converges to 0 strongly. and.HINTS Chapter 13. Strong Operator Topology Problem 116. consider the powers of the adjoint of the unilateral shift. and replace g by An f/11 An f 11. Operator Topologies Problem 107. Consider square roots of down and the other slipping out far away. Imitate Solution 21. (b) Look at powers of the unilateral shift. consider powers of the adjoint of the unilateral shift. Consider the adjoint of the unilateral shift. similarly. Problem 108. For a counter152 .

evaluate (U*Uf. For rank: the restriction of U to the initial space of Vis one-to-one.. For connectedness: if U is a partial isometry. M. for each eigenvalue A. if pis a non-zero polynomial such that p(A) is diagonal. If U is a partial isometry with initial space M. do the same thing. Study the sequence of powers of EFE.).}. If N is a neighborhood of F(A. of p(A). Problem 126. The Bn's form a bounded increasing sequence. if U*U is a projection with range M. ker U. Problem 128. For nullity: iff E ker V andf J.Vfll = llfll· Problem 131. then F. if Vis an isometry. say.1 {z: p(z) =A.VII < 1. Use the Weierstrass polynomial approximation theorem in the plane. Chapter 15. Represent A as multiplication operator by a function <p.1(N) is a neighborhood of A. is disjoint from F(spec A). for this let U be the adjoint of the unilateral shift and let M be the (one-dimensional) subspace of eigenvectors belonging to a non-zero eigenvalue. f) when f EM and when f J. The only troublesome part is to find a co-isometry U and a non-reducing subspace M such that UM = M. then IIUf. then consider. Compare 11Afll 2 with IIJA/11 2 • Problem 125. For closure: A is a partial isometry if and only if A = AA* A. Find a unitary operator that matches up initial spaces. then some neighborhood of A. then U is an isometry. the set <p.¢ F(spec A). Problem 129. and if II U .PARTIAL ISOMETRIES example with respect to the uniform topology. Problem 130. Problem 124. consider sequences of projections. Partial Isometries Problem 123. Problem 122. Problem 127. and another that matches up final spaces. and find continuous curves that join each of them to the identity. Problem 121. 153 . If A.

To prove that maximal partial isometries are extreme points. The question and the answer are mildly interesting but the method is not: if A is a weighted shift. Problem 139. then it transforms M(A)M(A)* onto M(B)M(B)*. Every partial isometry has a maximal enlargement. find a contraction A with spectrum A. If UP commutes with P 2 . Problem 137. then it commutes with P. one of which shifts the resulting two-way sequence of subspaces forward and the other backward. use the polar decomposition of A to prove the fourth. For the negative one: a left-invertible operator that is not right-invertible cannot be the limit of right-invertible operators. just compute (A* A)A and A(A* A). If U is a unitary operator matrix of size 2 that transforms M(A) onto M(B). and extend A to a partial isometry. use Problem 135. Look for 2-dimensional counterexamples to three of the four possible questions. Problem 133. Chapter 16. so that UP . and argue that it is enough to prove the existence of two orthogonal reducing subspaces of infinite dimension. Apply Problem 142. show that every contraction is the average of two maximal partial isometries. Chapter 17.HINTS Problem 132. Problem 143. Problem 136. apply Problem 135. For the positive result. Unilateral Shift Problem 142. If a compact subset A of the closed unit disc contains 0. Prove it by the consideration of spectral measures. Problem 138. Problem 141.PU annihilates ran P. and define U by U Pf = Af on ran P and Problem 135. Assume that His separable. Polar Decomposition Problem 134. = A* A. 154 . and factor the given unitary operator into two operators. Put P 2 by Uf = 0 on ker P. Consider polar decompositions UP and join both U and P to 1. To prove the converse. Problem 140. use Problem 4.

Problem 151. (a) If a normal operator has a one-sided inverse. . then the reverse inequality holds. complement of the range of V. Put Ae0 = cp. Problem 145. Problem 152. If V 2 = U*. then L:'=o IIA"II 2 z" converges at z = 1. by the Fuglede commutativity theorem. ·). and use the technique of Solution 65. (b) Since 1 is an approximate eigenvalue of the unilateral shift. Problem 149.L and apply the results of Solution 149. Prove that Mk . If U has multiplicity m and V is a square root of U*. (c) There is no invertible operator within 1 of the unilateral shift. = J1 - A* A Problem 153. If A= s-tcs. Problem 147.L(A. and consequently an equivalent norm is defined by 11!11 0 2 = L:'=o IIA"fll 2 • Problem 154. and if t/J is a bounded measurable function on the circle. Write N = M n (UM). Problem 156. by Sylvester's law of nullity.. Every function in H 00 is the limit almost everywhere of a bounded sequence of polynomials. T Af. then. then n:=o n:=o Problem 150. If r(A) < 1. If V is an isometry on H. If W commutes with an operator A. If m is finite. 155 . t/J. Problem 146. then it is invertible. write T and assign to each vector f the sequence (Tf. cf. and recall that -1 belongs to the spectrum of the unilateral shift. Use Problem 149. and if N is the orthogonal V"H = (V"Nl. then. the same is true of the real part. To prove dim N = 1. prove that At/1 = cp.) is invariant under U*. use Solution 65. Solution 41. m ~ 2 null V. Begin as for Solution 146. If II A II ~ 1 and A" -+ 0 strongly.1 II·IIC"II·IISII· Problem 155. imitate Solution 66.UNILATERAL SHIFT Problem 144. Problem 148. t/J(W) commutes with A. then IIA"II ~ IIS. T A 2f. assume the existence of two orthogonal unit vectors/and ginN and use Parseval's equation to compute 11!11 2 + llull 2 • It is helpful to regard U as the restriction of the bilateral shift. then dim ker V ~ 1 and V maps the underlying Hilbert space onto itself.

~ 1 . for each measurable subset E of the circle. and apply Problem 157 toM. Iff is a cyclic vector of A. consider (g*. (a) Consider the first basis vector. Use Problem 155 to express Min terms of a wandering subspace N. There exists an orthonormal sequence Un} such that 11(1 . Problem 164.112 k '>k (~ 0 . consider a vector such that lim 1. 156 . with p a polynomial. 0 and hence there exist projections Pn of rank 2 such that (1 .U*)f. form vectors whose components are subsequences of this sequence {~n}· Problem 161.Pn)(1 .HINTS Problem 157. let M be the least subspace of H 2 that contains f and is invariant under U. (b) Gram-Schmidt. let Mj be the span of the e/s with n j (mod p). and examine the Fourier expansion of a unit vector inN. Problem 162. Problem 159. ~ 2 . Problem 165. Cyclic Vectors Problem 160. Problem 158. Sufficiency: assume {ocn} periodic of period p.. and examine its matrix. Approximate A*"! by p(A)f. Problem 166. Problem 163. Consider shifts of multiplicity greater than 1.U. For shifts of higher multiplicity.ocA)Pf is a cyclic vector of A for each positive integer p.U) ___. use Fejer's theorem. 1 . observe that each vector has a unique representation in the form fo + · · · + JP. then <pK c: K.ll --. · ·-) n= 1 I l~n+k12 = 0. Given fin H 2 . Problem 167. Necessity: consider a Hermitian operator that commutes with A (and hence with A* and with A*A). If K is the cyclic subspace spanned by a function that is never 0. then (1 . and if <p is a bounded measurable function. with jj in Mj. For each f and g in /2 . = Chapter 18.where the stars indicate coordinatewise complex conjugation. For the simple shift. consider the set of all those f's for which jj(z) = 0 for all j and for all z in the complement of E. -!*). and if 0 < llocAII < 1.

{0} c: IT 0 (C). Problem 176. Problem 172. Infer. s) continuity recall that a basic weak neighborhood depends on a finite set of vectors. Problem 175. that the part of the spectrum of a normal compact operator that lies outside a closed disc with center at the origin consists of a finite number of eigenvalues with finite multiplicities. To prove self-adjointness. Problem 170. Problem 179. Approximate by simple functions. via the spectral theorem. use the polar decomposition. and consider the orthogonal complement of their span. Properties of Compactness Problem 169. Every non-compact Hermitian operator is bounded from below on some infinite-dimensional invariant subspace. Consider the direct sum of a sequence of projections of rank 1. In the discussion of (w __. then the subspace is finite-dimensional.PROPERTIES OF COMPACTNESS Problem 168. Chapter 19. Problem 178. Use the polar decomposition and Problem 172. Problem 171. Approximate by diagonal operators of finite rank. Consider the collection of all operators that map orthonormal sequences to strong null sequences. then IT(C). Problem 173. separated by suitably long sequences of zeroes. Problem 177. Consider 2U* (where U is the unilateral shift acting on /2 ). its restriction to such a subspace is invertible. Use nets. and IT 0 (C) is countable. Construct a vector f by stringing together suitable multiples of the initial segments of the vectors in a countable dense set. If the restriction of a compact operator to an invariant subspace is invertible. then the sum of the eigenvalues of A* A is finite. If A is a Hilbert-Schmidt operator. Use the spectral theorem. Every operator of rank 1 belongs to every non-zero ideal. 157 . If C is compact. Problem 174.

that ker A = 0. = 0. Examples of Compactness Problem 186. in that case prove that there exist scalars A. From (3) to (1): if 1 . Given the shift U and a compact operator C.1(A Problem 183. Problem 184. Can a Volterra operator have a non-zero eigenvalue? Problem 188.1.HINTS Problem 180. estimate the spectral radius of U .BA.L A is invertible. where Vis the Volterra integration Problem 191. apply Solution 179 to 1 . then An is a Volterra operator with kernel bounded by cn/(n. Problem 181. B(1 + B. + V). but the spectrum of (U + C)*(U + C) is small. Problem 187. From (1) to (2): the restriction of A to ker. so that M contains e0 . the restriction of A to the inverse image of M is invertible. Perturb the bilateral shift by an operator of rank 1.C. and solve it.B)). In that case. with no loss. Assume. if M is a subspace included in ran A. 1).nAnf-+ e0 . use induction to conclude that M contains ek for every positive integer k. then the spectrum + Cis large. + 1) with L 2 (0. Caution: there is more than one interesting way of making the identification. Identify L2 ( -1.BA is compact.1)!.n such that A. Assume A.f into a differential equation. then A = . 158 . If C is compact and U + C is normal. Reduce to the case where M contains a vector f with infinitely many non-zero Fourier coefficients. If A is a Volterra operator with kernel bounded by c. Problem 190. of U Problem 185. note that if B is invertible. Problem 182. Chapter 20. By differentiation convert the equation V* Vf = A. and determine the two-by-two operator matrix corresponding · to such an identification. Put A= (1 operator. Express V*V as an integral operator. 1) EB V(O. Problem 189.

. If A commutes with a Hermitian B.. IX 1 . r > 1}. 11/11 for all n. then the matrix ((A1j. consider A Problem 194. and do all this so that the two extensions commute. ···}is hyponormal if and only ifi1Xnl 2 ~ lan+tl 2 for all n. • · ·. Apply Fuglede's theorem to two-by-two operator matrices made out of A 1. then it commutes with the spectral measure of B. Problem 201. Problem 197.. J. A 2 . then U ('f.} Problem 198. Problem 196. Problem 203. Use Problem Problem 202.)) is positive definite. 78. then 11/11 2 . and consider the powers of the adjoint of the operator p(W). Form a subnormal operator by restricting a suitable multiplication on L 2 to the closure of the set of all polynomials.. Problem 199.i B 1 *if) = Li B 2 *ijj. Let W be the bilateral shift. The desired isometry U must be such that if {!1 . JMr r2 nlfl 2 df. and if / 0 . with measure defined so as to be normalized Lebesgue measure in the circle and a unit mass at the center. If A (on H) is subnormal. Once ker A = 0. Problem 200. and if M. IX 2 . fn are vectors in H. is a finite subset of H. consider the polar decomposition of A.. extend the isometric factor to a unitary two-by-two matrix.. and B.. then so is B. It is sufficient to prove that if A is invertible. • • • .. extend the positive factor to a positive two-by-two matrix. 159 . Show that ker A reduces A and throw it away. Problem 195. Use Problem 194. Subnormal Operators Problem 192. Consider the measure space consisting of the unit circle together with its center. A weighted shift with weights {1X 0 . If IIAn/11 ~ + Band A + iB.l. Problem 193. Every finite-dimensional subspace invariant under a normal operator B reduces B. Both~- spec A and~ n spec A are open..SUBNORMAL OPERATORS Chapter 21. A1. = {x: lcp(x)l. Given A and B.

Problem 212. The closure of the numerical range includes both the compression spectrum (the complex conjugate of the point spectrum of the adjoint) and the approximate point spectrum. Problem 211. If A is compact also. Try a diagonal operator. P. The quadratic form associated with a compact operator is weakly continuous on the unit ball. Problem 208. Problem 207. then the span of the eigenvectors of A reduces A.···. Problem 209. k. If A is hyponormal. Given a hyponormal idempotent P./. Try the unilateral shift.J forM. i = 1. and apply Problem 179 and Problem 205. • • ·. If P and Q are projections of rank k. Problem 213. then 11Anfll 2 for every vector f.···. · · ·. cxk such that Tfi = cx. and apply the Toeplitz-Hausdorff theorem k times. If M and N are k-dimensional Hilbert spaces and if T is a linear transformation from M to N. and {g 1. Use Problem 149. decompose the space into ran P and ranJ. Consider the eigenvectors of the imaginary part.g. ~ 11An+ 1 11·11An. then consider the restriction of A to the orthogonal complement of that span.. Chapter 22. Note incidentally that the general Toeplitz-Hausdorff theorem is equivalent to its 2-dimensional special case. Problem 215.1 ll·llfll 2 Problem 206.1 . If A is hyponormal. Problem 214.HINTS Problem 204. Let V be the Volterra integration operator and consider 1 . Numerical Range Problem 210. Try a linear combination of the unilateral shift and its adjoint. A set in the plane is convex if and only if its intersection with every straight line is connected. then there exist orthonormal bases {!1.gk} for N. apply this statement to the restriction of QP to the range of P. 160 .(1 + v). Problem 205. and there exist positive scalars cx 1 .

Bll < a and II! I = 1.:! 1 is that Re(1 .zA).:. reduce the thing to be proved to the statement that if the values of a function are in the right half plane. with the projection characterization of the weak topology. and consider A ( JA(1. (a) Prove the contrapositive. then Afn . Problem 223. 3. 1. Write down the partial fraction expansion of 1/(1 .. then so is the value of its integral with respect to a positive measure. and 21'6.fn ~ 0. Problem 219. Are the suggested closures indeed closed? 161 . 2. and let the resulting formulas suggest the solution in the general case.214. For U use Problem 222(a).UNITARY DILATIONS Problem 216. Problem 220. A necessary and sufficient condition that w(A) . and consider (~ ~) and (~ ~).A Problem 224. with non-closed range. together. · · · . If IIA . in both cases. (b) Imitate (a). Problem 221. 0 . n = 1. Chapter 23. Write M IIA I = 1 and = (~ ~). fn) ~ 1. let N be a normal operator whose spectrum is the closed disc with center 0 and radius t. then (Af. Use Problems 200. Unitary Dilations Problem 222.z") and replace z by zA. (b) If (Afn. use analytic geometry to prove it. Find an operator A. (a) Suppose that the given Hilbert space is one-dimensional real Euclidean space and the dilation space is a plane.1 ~ 0 for every z in the open unit disc. A . Problem 218.:.. + (a). Use the spectral theorem.A) jA(l- A)) 1. Examine the meaning of the assertion in this case. Problem 217. f) E W(B) Let U be the unilateral shift and consider U*n.. and for P Problem 222(b).

then so is B $ C. Find a unitary power dilation of A.QP. for P note that every idempotent contraction is a projection.1 AT has a non-zero kernel.1 AT with itself countably many times. Commutators Problem 230. Look for a bilaterally infinite matrix that does the job. for co-isometries use Problem 152. (a) Generalize the formula for the "derivative" of a power to the non-commutative case. where P is the pertinent unilateral shift. and use that evaluation to estimate its norm. Consider the Banach space of all bounded sequences of vectors. apply Problem 234 to the direct sum of A + T. Examine the corresponding matrix representation of the given operator. Fix P and consider determine ~"Q". Chapter 24. Problem 226. Wielandt: assume PQ . Wintner: assume that P is invertible and examine the spectral implications of PQ = QP + 1. Problem 231. and try to represent it as PQ . Use the spectral theorem to prove the assertion for unitary operators. Prove and use the lemma that if B + C is a commutator. If an operator B is such that to every vector g there corresponds a hyponormg'l operator that agrees with B on both g and B*g. Problem 229. and observe that each bounded sequence of operators induces an operator on that space. Find an invertible operator T such that A + T. evaluate P"Q . use the techniques and results of Solution 222. (b) Use the Kleinecke-Shirokov theorem. Problem 233. 162 .QP = 1. Problem 228.QP". then B itself is hyponormal. For U use Problem 149. Problem 234. Represent the space as an infinite direct sum in such a way that all summands after the first are in the kernel. and then use the existence of unitary power dilations to infer it for all contractions. modulo null sequences.HINTS Problem 225. Problem 227.QP as a function of Q. Problem 235. Problem 232. ~Q = PQ . and imitate Wielandt's proof.

Use Problem 142.. and use that to prove that if 0 e ll(L). For sufficiency: use Problem 146. Problem 242. Prove that W*"TPW"-+ L strongly. and. For necessity: compute. Problem 238. and prove that Cfn-+ 0.L(l . n = 0. then Yi+t. for a fixed y in D and a fixed fin H2 . together with a multiplicative adaptation of the introduction to Problem 234. If <Yii> is the matrix ofT"' T"'. · · ·. If C = A* A . adapt the Wintner argument from the additive theory. (2) ker(l -A* A) reduces A. Consider the characteristic functions of a set and its complement. in ll(A). For sufficiency.AA *). (a) Prove that (1) A is quasinormal. try a (bilateral) diagonal operator and a bilateral shift. choose A. Chapter 25. (b) Consider a weighted bilateral shift. Since g(y) = 0. Problem 237.i+t = Yii + ('J'i+tf3-j-t• where Problem 244. it 163 . find {!. for necessity.~(y))/(z).-+ 0. For sufficiency: write Anf = W*" APW"f for all f in L2 .)f. Let K be the kernel function of H2 . and prove that the sequence {An} is weakly convergent. For necessity: compute. Use Problem 142.A.AA * ~ 0. and (3) ker. Problem 243.TOEPLITZ OPERATORS Problem 236. and then try a diagonal operator matrix and a bilateral shift. Problem 245.f. Problem 240. II = 1 and (A . 2. with all the weights equal to either 1 or J2.A* A) c:: ker(A *A . to prove that every invertible normal operator is the product of two commutators. in an operator matrix imitation of the technique that worked in Problem 238. write g(z) = (~(z) . Toeplitz Operators Problem 241. Problem 239. 1. Problem 246.} so that II . What if q> is a trigonometric polynomial? Problem 247. then 0 e ll(T).

Ky and hence that cp(y) is in the (compression) spectrum Problem 248. then <p · f* implies that sgn <p is constant. and this 164 .HINTS follows that ofT"'. then <p · f* · f is real and belongs Problem 249. Modify Solution 244 so as to be able to apply Solution 246. If <p is real and T"' is invertible. g l. = 0. E H2 . Problem 250. If <p is real and T"'f to H 1.

SOLUTIONS .

.

. more generally.<f + g)) .+ . Indeed. and. Since each l/ln is a quadratic form. the Schwarz inequality is true for not necessarily strictly positive forms. sesquilinear form on the same space. moreover. is strictly positive. Yes.it/I(J. defined by 1 t/J+(f. Associated with each function cp of two variables there is a function cp. g) = t/I(J. Solution 2. the set of natural numbers) has nothing to do with the facts here. t/J(f) for each vector f). The limit ofa sequence ofquadratic forms is a quadratic form. and write..g)) + il/f(t(f + ig)) .l/1+-.of one variable. and one way to prove it is to reduce the general case to the strictly positive case.t/I(J._. given cp. As for finding a strictly positive form cp+ (on every real or complex 167 . for nets of arbitrary structure. and hence that t/1 is a quadratic form. Since. apply the Schwarz inequality to it and let e tend to 0. 2 cp.. then t/J = t/J+-. for each positive number e.<f - ig)). t/1 (that is. The form cp. if t/1 is a quadratic form.e. defined by cp-(f) = cp(f. let cp + be an arbitrary strictly positive. PROOF. f).CHAPTER 1 Vectors Solution 1. then cp = cp..<f ... symmetric. If { t/Jn} is a sequence of quadratic forms and if l/ln __. associated with each function t/J of one variable there is a function l/1+ of two variables. = cp + ecp+. The index set for sequences (i. then t/ln + _..it follows that t/1 = t/1+-. t/Jn(f) __. If cp is a sesquilinear form.l/1+ and l/ln +.. the proof is just as valid for ordered sequences of arbitrary length. it follows that each t/Jn + is a sesquilinear form and hence that l/1+ is one too. t/Jn = l/Jn +-.

g).SOLUTIONS vector space): just use Hamel bases.: {3iei) = 2. Begin by observing that 1 = (f. II gil ~ 1. write q>+ (2.j~(ej) = 2: ff. 4 Solution 4. The boundary points of the closed unit ball are the vectors on the unit sphere (that is. and llhll ~ 1.iei. For each finite set J of indices. and hence that This result justifies writing g = ~U) = j L {3iei. To motivate the approach. the vectors f with I f II = 1). write g1 = Lj eJ {3iei.i/3/. The thing to prove therefore is that iff= tg + (1 . h).iei. 11!11 = 1.t)(f. Since L {3i = (g. e) = (ei. is to prove the convergence of the series Li {3iei. g)* = ~(ei)*. 3 Solution 3.: ff. g) 168 + (1 .t)h. assume for a moment that it is already known that ~(f) = (f. . Iff = Li ff. J J J The sums are formally infinite but only finitely non-zero. where Pi = ~(e)*.j/3/ = u. 2. j If the existence of the Riesz representation is known. the unit vectors. Choose an arbitrary but fixed orthonormal basis {e} and expand g accordingly: g = {3iei. from the point of view of the present approach to the existence proof. where 0 ~ t ~ 1. If {ej} is one. this reasoning proves uniqueness and exhibits the coordinates of the representing vector. f) = (f.) = 2: 1Pjl jeJ 2 • and therefore This implies that jeJ 2: lf3jl 2 ~ WI. and the proof is complete. g) for some g. then f = g =h. The main problem.: ff. Then ~(g.t)h) = t(f. tg + (1 . the vector g could be captured by writing g = L ~(e)*ei. then j 2: ff.

but that does not imply that it is unavoidable. Since every infinite-dimensional Hilbert space has a subspace isomorphic to L 2 (0. If E is the spectral measure on the Borel sets of [0. Is it? An examination of the finite-dimensional situation i~. Solution 5. 6 169 . f(t)) = llf(s)ll 2 • Proof: (f(s). g) = (f. It is convenient to begin with a few small auxiliary statements true about every normalized crinkled arc f. It is not absolutely necessary to consider only continuous spectral measures whose support is the entire interval. Write g = af and h = f3f. f(t)) = (f(s) . t]. it follows that (f.f(O). Constructions similar to the one given above are familiar in the theory of spectral measures (cf. g) I ~ 1 and I(f. h) 1~ 1. and. The description is easy. and this implies that both g and hare multiples of f. but it is wise. (f(t))(s) = 1 or 0 according as 0 ~ s ~ t or t < s ~ 1. p. 1] such that E(M) is. those assumptions guarantee that every non-zero vector will work in the role of e. 58]). If 0 ~ a ~ b ~ 1. similarly. As for the existence of tangents: it is easy to see that the difference quotients do not tend to a limit at any point. which shows quite explicitly that f is not differentiable anywhere. quite instructive. 1 = /3*. t])e. Since 1 = (f. f(t) . it is sufficient to describe the construction for that special space. (1) If 0 ~ s ~ t ~ 1. and if e is the function constantly equal to 1.f(a)ll 2 = Jl(f(b))(s).f(s) + f(s)). then the curve f above is given by f(t) = E([O. then llf(b). this implies that f is continuous. [50. let f(t) be the characteristic function of the interval [0. The result says that the Schwarz inequality degenerates. 1).VECTORS Since I(f. the proof is complete.a. h) = 1. multiplication by the characteristic function of M. this step uses the strict convexity of the closed unit disc. for each Borel set M. then (f(s).(f(a))(sW ds = 5 fds = b. Solution 6. The verifications of simplicity and of the orthogonality conditions are obvious. in other words. Indeed. If 0 ~ t ~ 1. This remark shows how to construct many examples of suddenly turning continuous curves: use different spectral measures and apply them to different vectors. both for f and g and for f and h. g) = (f. Infinite-dimensionality was explicitly used in the particular proof given above. af) = a*.

and hence they too must be linearly independent. in other words. hence linearly independent.g(s). • • ·. first on the range off only. Suppose that f and g are normalized crinkled arcs. is isometric. then its linear dimension is greater than or equal to 2~ 0 • (Recall that if either the linear dimension or the orthogonal dimension of a Hilbert space is finite. the mapping U is isometric.f(<p(s)). then llf(s). Now define a mapping U. (3) The mapping t ~---+ llf(t)ll is strictly monotone and continuous. Proof: immediate from (1). f(tn) span the same n-dimensional space. by the definition of <p.f(<p(s)). Proof: if 0 < t 1 < t 2 < · · · < tn ~ 1. note that. by Uf(<p(t)) = g(t).f(tn_ 1 ) are pairwise orthogonal. The next thing to prove is that the extended U is still isometric. f(tn) .llf(s)ll 2 . and then the inverse of the mapping t ~---+ I f(t)ll.f(<p(s)))ll 2 = llg(t)- = = g(s)ll 2 llg(t)ll 2 .SOLUTIONS (2) If 0 ~ s ~ t ~ 1. 1] to [0. Proof: immediate from (2). The second and last step is to observe that each vector in the linear span of ran f is a finite linear combination of orthogonal differences such as f(<p(t)) . The first step in this direction is to consider two numbers s and t. Define the mapping <p from [0. · · ·. and observe that II U(f(<p(t)). If the orthogonal dimension ofa Hilbert space is infinite.) 170 . 1] to be the following composition: first the mapping t ~---+ llg(t)ll. it follows that U. f(t 2 ) . then the vectors f(t 1 ). then so is the other. Since the nonzero elements of ran f are linearly independent. Since the square norm is additive for orthogonal summands. it follows that both the domain and the range of the extended U are dense in H. Since both ran f and ran g span H.llg(s)ll 2 llf(<p(t))ll 2 . The rest is trivial: extend by continuity and get a unitary operator U on H such that g(t) = Uf(<p(t)). the vectors f(t 1 ). as defined so far.f(<p(s))ll 2 . and the two are equal.llf(<p(s))ll 2 = llf(<p(t)). The ground is now prepared for the uniqueness proof. 7 Solution 7. (4) The non-zero values of f are linearly independent. 1] such that llg(t)ll = llf(<p(t))ll for all t.f(t)ll 2 = llf(t)ll 2 . and its image under U is the same linear combination of the corresponding orthogonal differences g(t) . 0 ~ s < t ~ 1. The result is a reparametrization <p of [0.f(t 1). U can be extended to the linear (not necessarily closed) span of ran f so as to become linear. U is isometric on differences such as f(<p(t)) .

enen. • · ·} is a linearly independent sequence. For each positive real number t. For each real number t. The same argument proves. Proof: the Gram-Schmidt orthogonalization process implies 171 . . since. 0 0 ·). let J 1 be an infinite set of rational numbers that has t as its only cluster point. Each L 1 is infinite. Suppose. write J. f 3 . for instance. that rx 1 = 0. 0 < t < 1. constitute a linearly independent set with cardinal number 2~ 0 • The problem as originally stated has at least one solution that is both short and elementary. 1). Here is a quick outline of a possible construction. however. consisting of infinite sets of positive integers. and the vectors en Ln en en LneJ. say Lf= 1 r:xifrt = 0. that a finite linear combination of the f's vanishes. then there exists a vector that is not a (finite) linear combination of the f's. such that J. let L1 be the part of L in the open band between the line of slope t through the origin and the parallel line at distance 2 above it. en en g(t) = (1. Since there is a one-to-one correspondence between the positive integers and the rational numbers. constitute a linearly independent set with cardinal number 2~0 • Alternatively. . it follows that J 11 contains at least one integer. for each i =I= 1.VECTORS PROOF. k. Consider the (countable) set L of all those points in the plane whose coordinates are integers (i. t. f 2 . and let f = (Fourier expansion) be an arbitrary vector such that =I= 0 for all n. [An alternative construction of uncountably many "almost disjoint" subsets of a countable set is based on an elegant geometric observation of J. t 2 . 0 < t ~ 1. Assertion: if {!1 . and hence. that r:xi = 0 for each i = 1. 1). the lattice points in the plane). This result is the main reason why the concept of linear dimension is of no interest in Hilbert space theory. every infinite-dimensional Hilbert space may be assumed to include L 2 (0. e3 ..e. and the vectors f(t). f'l L 1 is bounded whenever s =I= t. exhibited in Solution 4.. of course. ···}is acountablyinfinite orthonormal set in a Hilbert space H. of cardinal number 2~ 0 . every infinite-dimensional Hilbert space may be assumed to include /2 . In a Hilbert space context "dimension" always means" orthogonal dimension". It follows that r:x 1 = 0. that does not belong to any J 11 (i =f. L. There are shorter solutions of the problem. which has several applications: there exists a collection {J1}. but the preceding argument has the virtue of being elementary in a sense in which they are not. since =I= 0. R. The main tool is the following curious piece of set theory. it follows that the intersection of two distinct L/s is always finite. If {J!} is a collection of sets of positive integers of the kind described above. say n. Buddenhagen [25]. e2 .] Suppose now that {e 1 . f'l J 1 is finite whenever s =f.} of vectors is linearly independent. t. Since. Thus. indeed. the set ] 11 contains infinitely many integers that do not belong to J 11 .. = Assertion: the collection {J. it is sufficient to prove the existence of sets of rational numbers with the stated property.

n = 0. .1. In more detail: if {!1.. 2. ek+z) = ···. • • ·}. for instance. The proof is due to L.. · · ·} be an orthonormal basis. then (g. the conditions amount to the equality of a finite number of Fourier coefficients. ftc} is not total. for each n. in particular. ' f. to be sure. An even shorter (but less elementary) solution consists of two words: Baire category. = en . ek-3) = · · · · Since infinitely many Fourier coefficients can be equal only if they vanish.. 2. ±2. then T . every coefficient is different from 0. ···}be an orthonormal basis. The argument of the preceding paragraph still applies. and letT be the set {f.} for any n.} for n = 1. . e 3 . This proves that T .. f 3 . ek+ z) = . a proper subspace and. and (g. · • ·} such that V{e1. · · ·. ± 1. if the Fourier expansion of a vector g with respect to the e's has infinitely many non-zero coefficients. ···to get (g.{ed contains the vector f . ·. e2 ... How can a vector g be orthogonal to T .{jj. ···}is an arbitrary sequence. it 172 . k + 2. e1_ 1) = (g.. ek+ 1) = (g. ek-1) = (g. 3. ek-1). g = e1 + · · · + ek_ 1. (g. An alternative proof (with different merits) goes as follows.E.{fi. ek) = (g. f. which is harmless. These conditions imply that (g. If. ek+l) = (g. k . e1_ 2 ) = (g.. e1 _ 3 ) = · · ·.en_ 1. If. Let {en: n = 0. ' en} = vu1.. . e2 . and letT be the set of all vectors of the form f. put n = k .l T . ftc}. ek) = (g.{fk} is always total (and implies. Let {en: n = 1. en_ 1 ) whenever n =F k. then (g. e1. ek-z) = (g. e3 ... J. · · ·. 3. it follows that the Hn's cannot fill up a complete metric space. on the other hand. then Hn = VU1. fk} is not total. f 2 . Put n = k + 1. Since the span of the set T .# rxnen.j < k. 8 Solution 8. it follows that g = 0. e1) = · · · = (g.2. that T itself is total). · · · to get (g. then all the conditions are satisfied. f. . ±2.{ftc}? If that happens. a nowhere dense set. eJ = 0 when n ~ k and when n < j.{jj.SOLUTIONS the existence of an orthonormal sequence {e 1. and that implies that T . therefore. when j ~ n < k. but the result this time is that if g . Wallen. then g does not belong to V {!1. ± 1. en) = (g.. • • ·.} is. let f be a vector in whose Fourier expansion.

the 173 .:-J ~nrxk*n = 0. + 1 . i. the same method applies if the powers rxk are replaced by arbitrary numbers rxk (and. ·. · • ·} (there is no loss of generality in the assumption of countability). Conclusion: T. + 1. rxk*ei.}. · · · .rxi *ek is orthogonal toT.1 ) is orthogonal to each fk> then 2::. 2 + 1 and approximate them to within ! by vectors in VUn 2 + 2 . then the vector g = rxk *ei . · · ·). Reason: (f. (Note that if L~=o l~n1 2 < oo. · · ·}. m) are defined by i'k = <1. .) The result is a shallow generalization of the well known facts about Vandermonde matrices. rxnk is replaced by rxkn). j < k.rxkrxi = 0. • • · .ork=123··· ' J' ' ' ' PROOF.. fn. ·.. • · ·. • • ·. then the power series Lnco=o ~nzn has radius of convergence greater than or equal to 1. and consequently it vanishes identically.fn2 + 1 . The phrasing of the problem is deceptive. on the other hand.. 2 \ 3 9 10 \ • • ·) rx t. Iff= <~ 0 . 3. the power series L ~nZn n=O vanishes for z = rx*k (k = 1.e.. · .. etc. · . / 3 . Conclusion: ~n = 0 for all n. then omit f 1 and approximate it to within t by a vector in V {/2 .. and therefore f = 0.{ed is total for each k.fn. f 2 . If the given set is {!1 . 2. The solution has nothing to do with the arithmetic structure of the powers rxk. 3.then 0 = (f. and if the vectors ~ i'k (k = 1. then the span of {/1 . and the proof suggested above is adaptable to the finite-dimensional case. correspondingly. ~ 2 . n=O co In other words. ~ 1 . fn./n. lfO < lrxl < 1 and fk = < rx\ rx 1. ). ed is not total.VECTORS follows that T . + 2 . · · ·). then omit / 1 . If lm 2 is the m-dimensional Hilbert space of all sequences < 0 . Solution 10. ~m _ 1 ) of length m ( = 1. /n 3 ). 2 then the fk's span 1 . provided only that the numbers rxk cluster somewhere in the interior of the unit disc. rxkm. Perhaps the quickest approach is to look for a vector f orthogonal to all the}. g)= (rxiei + rxkek. rxb ..s. then omit / 1 . Solution 9. iff = <~ 0 . 2. ek}. + 1 and approximate them to within t by vectors in V Un. ~m.fz. i'k) = co L ~nrx*nk. If. fm} is lm 2 • Indeed.{ei. fnJ.{ei. .. • · ·} that is included in the span of its complement in {f1 . f. . The result is an infinite set {f1. .1 ) (where 0 ~ lrxkl < 1 and the rxk's are distinct).rx/ek) = rxirxk.

p./'.W 1.. with ak -+ 0. ak 3.then the set {k k = 1. If fk = (1. A set such that every infinite subset of it is total deserves to be called totally total. 11 Solution 11.£ L lie. A more direct way is to define the dimension of Has the minimal cardinal number of a total set in H. For instance. It is common mathematical experience that a proof of the existence of minimal sets of a certain kind is likely to prove a stronger statement. the collection of total sets is only partially ordered. Caution: this does not say that dimension is the cardinal number of a minimal total set. It seems that the more you know the harder this is. One way to define dimension was mentioned in Problem 7: prove that all bases have the same cardinal number. . and hence. · · ·} is total in F. e.SOLUTIONS polynomial _L. This is so trivial that it is hard to see how it could have any usable consequences. To say that the stronger statement is false is to say that there exists a total set such that no total subset of it is minimal.) I = I(e. yield L lle1 - . The basic observation is that 12 I(e1 .. JJ)I 2 2 (by Parseval) . The following ingenious proof was offered by a student who at the time was completely innocent of the techniques of Hilbert space [143]. see [47.J ~nzn of degree m .fj.fk's is total also. and then define dimension as the common cardinal number of all bases. ···). a2 . The example described in Problem 10 works. Solution 12. ak. and so do many of the generalizations mentioned in Solution 10. 3. since every infinite subsequence of the ak's has the same properties as the original sequence.. 2. The collection of cardinal numbers under consideration is well ordered. namely that every one of the sets of that kind includes a minimal one.fj) I for all i and j. .f}ll 2 = = L L l(e1 j i I I l(e. it follows that every infinite subset of the . (For further comments on dimension see Problems 54 and 55. For total sets in Hilbert space the mere existence of minimal ones is obvious (orthonormal bases). It does. 174 . A final comment along these lines has to do with the concept of dimension. and hence identically.fj.:r.. however. or. in particular. 214].1 (at most) vanishes at m distinct points. For sufficiently unpleasant topological vector spaces the answer is no. the order of the words is important. let <a 1. • • ·) be a sequence of distinct complex numbers in the unit disc.) Does every separable Banach space have a minimal total set? The question seems to be open. t-total. but the stronger statement is false..11 i (by Bessel). has a least element./'. abbreviated. i j .. and the existence of such a set is obvious from Solution 11. e. ak 2.

en.. first on the linear combinations of the e/s only. The invertibility of A implies that the vectors et> .11 2 • I. JJW lie.VECTORS The assumed finiteness condition implies that the Bessel inequality must have been an equality for each i: I j l(e.JJ)r 2 2 • j>n . .Jn+ 2 . fn + 1 . 52]) that A has a (unique) extension to an invertible operator on H (which may as well be denoted by A again). ... = b.. belongs to the span of the ... • . a.fj's span H. Consequence: each e. which is a shade less elementary. .:. and then define a linear transformation A.···) span H.:.Jn+ 1.J. for each i./.l = n. b.. . I I'jl j>n llej .. It follows that if M is the span of fn+ 1.en+t>en+ 2 .f}ll j>n 2 • It follows that 1 . but whose technique is more often usable. which is strictly less than 1. the . fn+z.. so that I j (e.:.) It follows that I. (The elementary lemma that is being used is that ifO. p.A is bounded (as far as it is defined) by j>n I llej.:.. for each i and a. . < oo.fj's. 11/11 2 • I llej. Iff = Ij 'jej (finite sum). There is an alternative proof.f}ll 2 < 1. }j)}j = e.f}ll 2 . then 111 /ifll = 2 IIJn'iej. J. • • • . indeed. Begin by choosing a positive integer n so that I j>n llej... for each i...J. fn + 2 . This implies ([50. · • • span H. ·. by writing and Aej = }j if j > n. en. then dim M. and therefore.}j)}j = e..··· (the images under A of et>···. = b. Conclusion: the vectors f 1 ..JJII I . then a. I j (e. 175 .

go)+ (rxgo + g)ll 2 ~ lrxl 2 ·lifo . so that M + N consists of all the vectors of the form rxf0 + g. It follows that rxn -+ rx and gn -+ g. Indeed. say. If foE N. and consequently h = limn (rxnfo + gn) = rxfo + g. then Solution 13. that is.g0 l_ rxg 0 + g). where rx is a scalar and gEN. g0 is the unique vector in N for which fo-go j_ N. N.goll and therefore llgll = ll(rxfo +g). therefore. -lifo.CHAPTER 2 Spaces 13 + N is closed. Suppose. if rxnfo + gn-+ h. 176 . Observe now that if g is a vector in N. It is sufficient to prove that if dim M = 1. the general case is obtained by induction on the dimension. let g0 be the projection of fo in N. then M + N = N. with g in N of course. or lrxl s llrxfo +gil .goll 2 (since fo . If fo ¢.rxfoll ~ llrxfo +gil + ::1° ~ ~~:~·llfoll· These inequalities imply that M + N (the set of all rxf0 + g's) is closed. so that {rxnfo + gn} is a Cauchy sequence. in this case there is nothing to prove. that M is spanned by a single vector fo. then the inequalities imply that both {rxn} and {gn} are Cauchy sequences. then M l!rxfo + gl! 2 = l!rxUo.

hE L. so that each open ball of diameter fi (or less) can contain at most one element of each orthonormal basis. and let L be the span of Nand f 0 . find a vector fo in M v N that does not belong toM + N. and hence that H is finite-dimensional. and therefore gEL n M. All the remaining parts depend on easy finitedimensional geometry only. i. They should be supplied by the reader. every vector in L is of the form r:xf0 + g. To infer finite-dimensionality. L n M = 0.SPACES Solution 14. IfH is infinite-dimensional. Conclusion: (L n M) v N = N. then the modular law fails) is exactly what the proof of Solution 14 proves. so does their intersection. Suppose now that M v N = M + N. 177 15 16 . (This is in essence the usual proof that the normal subgroups of a group constitute a modular lattice. PROOF. then (by assumption) f = g + h. Lis equal to the vector sum of Nand the onedimensional space spanned by fo. which does not contain fo.. in particular. In a Hilbert space of dimension n ( < ~ 0 ) the (closed) unit ball is a closed and bounded subset of 2n-dimensional real Euclidean space. then it has subspaces M and N such that = 0 and M + N ¥: M v N (cf. then r:xj~ EM + N. that H is a locally compact Hilbert space.. conversely. and. this implies that r:x = 0 and hence that g = 0. so that gEL. Both L and M v N contain fo. The lattice of subspaces of a Hilbert space His modular if and only if dim H < ~ 0 (i. Given M and N. The preceding argument is the only part of the proof in which infinitedimensionality plays any role. so is the closed unit ball. By Problem 13. On the other hand. His finite-dimensional). and therefore the closed unit ball is compact. Suppose. since the open balls constitute a base for the topology. Solution 16. it follows that the space is locally compact. since translations and changes of scale are homeomorphisms. Since -hEN c:: L and f E L. with g in M and h in N. The inclusion :::> in the modular law is true for all lattices. The argument in the preceding paragraph reverses to this extent: the assumption of local compactness implies that each closed ball is compact. If f E L n (M v N). who is urged to be sure he can do so before he abandons the subject. therefore. Conclusion: f E (L n M) + N. just observe that both L n M and N are included in both L and M v N. it follows that f . that every closed ball is compact.e. Reason: if r:xf0 + gEM (with g in N). the compactness of the latter implies that every orthonormal basis is finite. where r:x is a scalar and g is in N. 14 M n N Solution 15. and. Problem 52). it is distributive if and only if dim H ~ 1. It follows.) The reverse implication (ifM + N ¥: M v N for some M and N. recall that the distance between two orthogonal unit vectors is fi. The collection of all open balls of diameter fi is an open cover of the closed unit ball.e.

apply Gram-Schmidt to orthonormalize. since its span is the same as that of the original sequence {!. begin by discarding from the sequence Un} all terms that are linear combinations of earlier ones. Iff E En. consider the open balls En with center at (r/2)en and radius r/4. that is. f 3 . Suppose. Since that process is frequently described for linearly independent sequences only. Once that is done. 178 . then This implies that if n ¥= m.}. · · ·} is an infinite orthonormal set in the space. Since a measure is. The Gram-Schmidt process yields an alternative approach to the converse. and consequently H is separable.. that {!1 .. conversely. If dim H ~ ~ 0 . Solution 18. and hence that the measure of E must be infinite. • • ·} is a countable set dense in H. rfi r r and hence that if n ¥= m. If E is such a ball.gil ~ . the mapping j ~ ni is one-to-one. e2 . then 18 so that En c E.gill < fi/2.. all the En's have the same measure. then En and Em are disjoint. it follows that E includes infinitely many disjoint Borel sets of the same positive measure. by definition. If {gi} is an orthonormal basis for H. Since. Since two open balls of radius fi/2 whose centers are distinct g/s are disjoint. there is no loss of generality in considering balls with center at 0 only. The set of all such rational linear combinations is countable. and if {e 1 .4 > 0. e 3 . / 2 . then II f .2 . it is a basis. En = {f: II!.. this implies that the cardinal number of the set of indices j is not greater than ~ 0 .4 .(r/2)enll < r/4}. then H has a countable orthonormal basis. Iff E En and g E Em. by in variance. invariant under translation. Since every vector in H is the limit of finite linear combinations of basis vectors. The resulting orthonormal set is surely countable.SOLUTIONS 17 Solution 17. it follows that every vector in H is the limit of such linear combinations with coefficients whose real and imaginary parts are rational. with radius r (>0). then for each index j there exists an index ni such that II!.

(f. If S is a weakly closed set in H and if {f. and therefore/ eS. The remedy. Reason: for each vector j. g). en) are the Fourier coefficients of j. observe that in an infinite-dimensional space weakly open sets are unbounded. however..--.. To prove what needs proof. g) for every g. It remains to prove that subspaces are weakly closed..e. Another way of settling the converse is to exhibit a strongly open set that is not weakly open. II!. At this point in this book the weak topology is not known to be metrizable.(f. so thatfn--. then.L and hence that f EM.. then en--+ 0 (weak). just observe that the sequential argument works without the change of a single symbol if the word "sequence" is replaced by "net". but that does not yet justify the conclusion that M is weakly closed. the inner products (f. This proves that weakly closed sets are strongly closed. by definition. in fact.CHAPTER 3 Weak Topology Solution 19.} is a sequence in a subspace M. e3 .. If {f. It follows that the set of all en's is not closed in the weak topology. g)--+ (f. and iff.../II· llgll --. This argument shows that M contains the limits of all weakly convergent sequences in M. Since each f. sequential closure may not be the same as closure. one such set is the open unit ball. The falsity of the converse (i. the proof shows that the strong closure of each set is included in its weak closure. .. that a strongly closed set need not be weakly closed) can be deduced from the curious observation that if {et> e 2 . 179 . is orthogonal to M\ it follows that f j_ M. therefore. g) I . then I(J. they are the terms of an absolutely squareconvergent series. f (strong). ···}is an orthonormal sequence. /(weak).. is easy. and net closure is always the same as closure. and. f (weak).} is a sequence of 19 vectors in S withfn --. o. in the strong topology it is discrete and therefore closed.

For each gin the unit ball. g) I is small when n is large. it is therefore sufficient to prove that if a set S is weakly dense in a Hilbert space H. it must be equal to H. in D.. Conclusion: I f. As for continuity : gj-+ g (weak) if and only if (f. If fn-+ f (weak). it is weakly closed. happens if and only if b(gi)-+ b(g) in D.. . by Problem 19. Every weakly separable Hilbert space is separable. g 1 ) = (f. Solution 22. 422]. Indeed. with the customary product topology. being at the same time weakly dense in H.SOLUTIONS With a different proof a stronger theorem can be proved. gj)-+ (f. Consequence: 11!11 2 ~ 11/nll ·11!11 + e for large n. and. in turn. p. and that. l(f~ g). There is a one-sentence proofthat the span of a weakly dense set is the whole space. Caution: it is not only more elegant but it is also safer to argue without sequences. the terms with minus signs tend to 11!11 2 . by assumption. 22 PROOF. if (f. Given the Hilbert space H. Looked at from the right point of view this is obvious. as follows. the mapping f ~--+ (f. if b(g 1 ) = b(g 2 ). The Riesz theorem on the representation of linear functionals on H implies that 180 . in particular. that is. then clearly g 1 = g 2 . it follows that the only vector orthogonal to a weakly dense set isO. The span of S is.f)l + e for each e > 0 when n is sufficiently large. by definition. as asserted. and hence. g) is a point. it turns out that the crucial concept is convexity. 23 Solution 23. The mapping b thus defined is a homeomorphism from the unit ball (with the weak topology) into D (with the product topology). then.(f. The proof depends on a familiar trivial computation: Since f.f 11 2 -+ 0. Reference: [39. 21 Solution 21. then the span of S is equal to H. a (strongly closed) subspace.-+ f (weak). say b(g). the theorem is that every strongly closed convex set is weakly closed. To be precise. so does the first term. that l(f. and let D be the Cartesian product of all the D1 's. so that b is one-to-one. It is not a priori obvious that iff is in the weak closure of S.f)l ~ l(fn. and therefore 11/11 2 ~ (liminfn 11/nll) ·11/11 +e. The proof is completed by letting e tend to 0. 20 Solution 20. The span of a countable set is always a (strongly) separable subspace. Since a vector orthogonal to a set is orthogonal to the weak closure of the set. for each fin H let D be the closed 1 disc {z: Iz I ~ I f II} in the complex plane.. g) for each fin H.. for each g. it follows. g 2 ) for all f. then f is the limit of a sequence in S.

it is sufficient to prove the existence of a countable base for the weak topology of H 1 • For this purpose. In a separable Hilbert space the weak topology of the unit ball is metrizable. and e H e<a tit + <Xz fz) is continuous on D. · · ·. Since the unit ball H 1 of His weakly compact (Problem 23). · · ·. In more detail. f 2 ) ofD defined by e e E(al. r) c: U. let {hi:j = 1.WEAK TOPOLOGY the range of b consists exactly of those elements of D (complex-valued functions on H) that are in fact linear functionals of norm less than or equal to 1 on H. Since the definition of product topology implies that each of the functions e H eul ).fz) = {e ED: e(aJl + <Xzfz) = ale(fl) + <XzeUz)}. as soon as that is done. 3. a 2 . The property of being a linear functional is a property of" finite character". q. To prove: if foE Ht. 3. PROOF 1. and r such that foE U(p. That is: is a linear functional if and only if it satisfies equations (infinitely many of them) each of which involves only a finite number of elements of H. f 1 . and form the subset E(ap a2 . and e is a positive number. k is a positive integer. <Xz. this implies that the set of all linear functionals is closed in D.f1 . e H eUz). g 1. and hence that the range of b is compact. it follows that each set E(a 1. and consider the basic weak neighborhoods (in H 1) defined by 24 where p. r = 1. The argument so far succeeded in constructing a homeomorphism b from the unit ball into the compact Hausdorff spaceD. f 2 ) is closed. 2. The remainder of the argument will show that that range is closed (and therefore compact) in D. ···}be a countable set dense in the space. gk are arbitrary vectors. ! 1. consider fixed pairs of scalars a 1 and a 2 and vectors f 1 and f 2 . q. Solution 24. 2. f 2 ). the weak compactness of the unit ball will follow. 181 . and it succeeded in identifying the range of b. ft. q. The assertion about properties of finite character amounts to this: the set of all linear functionals in D (the range of b) is the intersection of all the sets of the form E(a 1 . and if then there exist integers p. a 2 . The proof above differs from the proof of a more general TychonoffAlaoglu theorem (the unit ball of the conjugate space of a Banach space is weak * compact) in notation only.

hi)i ~lifo. Specifically: choose q so that 1/q < e/3. hi) I + l(f .hhll < lj2q.fo..hhil is small. if merit it be. The tail of the series for lin I is uniformly small for all n (in fact. choose r so that j.h) I ~ l(f. and the proof is PROOF 2.fo I is very small. ej)l ~ 11/11. e3 . q. e)i. There is an alternative procedure that sheds some light on the problem and has the merit. g)= If .g I whenever f and g are in Ht> then dis a metric for H 1. r}. Un. ~ 1 3' E e 1 and II j~ I ~ 1).foli·iihiil + II/. ei)-+ 0 as n-+ oo for each j. so that. k.SOLUTIONS The proof is based on the usual inequality device: l(f.foll·ilg. If d(f. To show that d metrizes the weak topology of H 1. (There is no loss of generality in assuming that the basis is infinite.. · · ·. k) choose j.hp. that it exhibits a concrete metric for the weak topology ofH 1• Let {e 1.l\ < qm · m < 3' e and II!. J since l(f.hP. choose p so that llhP -foil < 1/qm. g) I ~ l(f.) For each vector f write 1/1 = ~ 2i l(f.) Assume that fn-+ 0 (weak).hhll < 2· 2q < (recall that I f I complete. hi) I+ llhP. qm q so that / 0 E U(p. then IUo. ~ r for i = 1. the series converges and defines a norm. · · ·.hPII·IIhill < -·m = -. ···}be an orthonormal basis for H. and choose p so that II hP .hP. hi) I+ i(hp. · · ·.!oll·llg. (Caution: the metric d is defined for all H but its relation to the weak topology of H is not the same as its relation to the weak topology of H 1. It follows that f U.foil ·llhi. Iff E 1 1 U(p. it is sufficient to prove that fn-+ 0 (weak) if and only if Ifni-+ 0. · · ·. e2 . . so that llg.fo.fo.hP. · · ·. and.hill· Argue as follows: for each i ( = 1. r) and i = 1. If j = 1. r). where m = max{llhill: j = 1. then 1 l(f.. choose j. r. in particular. The uniform boundedness of the elements of H 1 is what is needed in the argument below. k. g. q.. so that llg. finally. the 1 182 . hh)l < q < 3' l e llhP. in the finite-dimensional case all these topological questions become trivial.

but. n = 1.. Solution 25. Choose g so as to make llh -gil small (and therefore l(fn. and Um mH 1 =H. h)l . The answer to the second question is no. then IUn. In view of Solution 24. 3. h .) To do that. This implies that if g is a finite linear combination of e/s. 2. Since en ~ 0 weakly. 2. and is therefore not dense in the unit ball. then f is the weak limit of a sequence of vectors f. then it is weakly separable (since it is weakly compact). 2. 26 183 .llfll 2)en. e)~ 0 as n ~ oo for eachj. it follows that fn ~ f weakly. with llfnll = 1. h. In an infinite-dimensional Hilbert space the weak sequential closure of the unit sphere is the unit ball. If h E H. g)~ 0.) The proof is completed by recalling (Solution 22) that weakly separable Hilbert spaces are separable.WEAK TOPOLOGY tail of the series for If I is uniformly small for all fin H 1). PROOF. 3. therefore. g) I small. Such linear combinations are dense. the unit ball.. Assume that IfnI ~ 0. Solution 26.. If Sr and Hr are the sphere and the ball of radius r (>0) and center 0. In the present case the assumed weak convergence implies that each particular partial sum of the series for IfnI becomes small as n becomes large. the intersection of that exterior with the unit ball is empty. (Recall that the unit ball is weakly closed. then H is separable. s. ffthe weak topology of the unit ball in a Hilbert space His metrizable. then Un. it follows that(!. 3. · ··}be a countable set weakly dense in H 1 . that the weak sequential closure of Sr includes Hs whenever r .) Conclusion: fn ~ 0 (weak). the weak closure of the unit sphere is the same as its weak sequential closure. all the more. and it follows that IfnI ~ 0.. the result is an easy corollary of what was just proved. the mfn's are weakly dense in mH 1 . obviously. and then choose n so as to make IUn. · · ·.g)l will be just as small). Since the sum of the series for IfnI dominates each term. (Reason: for fixed m. · · ·. is weakly metrizable. Let Un: n = 1. is that if 11!11 < 1. Consequence: the exterior of the unit ball is weakly dense (because it is the union of the S/s with r > 1)...g) I + I(J. What is to be proved. then what is already known is that the weak sequential closure of sr always includes "" and hence. The set of all vectors of the form mfn. l(f. PROOF. let {en} be an orthonormal sequence of vectors orthogonal to f and write 25 fn = f + (jl . g) I.:. If H 1.:. (This is a standard argument that is sometimes isolated as a lemma: a bounded sequence that satisfies the condition for weak convergence on a dense set is weakly convergent. n = I. m. is weakly dense in H.

e. 59]. after n steps.9z)l ~ 2(a(e1) + 2). This proof is due to D. Induction yields. Sarason. such that l(en+l.-a(e. Suppose that Tis a weakly bounded set in Hand that. if {e 1.. as before. contrary to the present assumption. almost the general case is in [1.. i= 1 and all is well. If His finite-dimensional.gn+1)1 ~ (n + o(t1! a(e. then it cannot be weakly bounded either. g 3 )1 ~ 3(a(e 1) + !a(e 2 ) + 3). as before.en and g 1.en} is an orthonormal basis for H. Special cases of it occur in [148. an element gn+ 1 ofT and a unit vector en+ 1 orthogonal to e 1. gn+ 1) + -+ (en+ b gn+ 1) I n 1 I•= z n 1 ~- i= 1 n 1 1 L -. · · ·.. gn+ 1)1 = 1 _L -1: (e. then they are bounded on H.)(e. arrive. A consequence ofthis assumption is the existence of an element g 1 ofT and a unit vector e 1 such that l(e 1. g) I ~ a(f) for all gin T. and assume that the conclusion is false. gn.(1n + 1)· (n 1 a(e. then j(f. a couple of elegant examples are [76] and [72]. E. the proof is easy. 45]. ···. ···. and such that l(e 3 . Now put f = L~ 1 (1/i)e... Argue.. Since l(f. such that l(ez. g 1)1 ~ 1. that the linear functionals induced by T cannot be bounded on the orthogonal complement ofthe at most four-dimensional space spanned by eb e2 . A consequence of this argument is the existence of an element 9 2 ofT and a unit vector e2 orthogonal to e 1 and 9 1. footnote 32] and [138. it follows that if T is not bounded. 184 . g) I= (tl(f. e. g) I i~l(e. p. Are the linear functionals induced by T (i.e. e. specifically.) + . gb g 2 . p. g) for gin T) bounded on the orthogonal complement of the at most two-dimensional space spanned by e1 and g 1? If so.. to the existence of an element g 3 ofT and a unit vector e 3 orthogonal to eb e 2 and gb g 2 .)) 2 .) + n + 1) Lz . Indeed. the mappings f ~----> (f. Assume now that H is infinite-dimensional.SOLUTIONS 27 Solution 27. Continue in the same vein. The production of non-category proofs (for Banach spaces) became a cottage industry for a while.)e. z n + i= 1 = n + 1.) + n + 1). gW ~ Ji~l(f.)l ~ 11!11 · 2 • I n (a(e. j(f. g) I I= Iit1(f.

Since I:'= 1 I(g. • ··}be an infinite orthonormal set in Hand let E be the set of all vectors of the form en.. of the set S1 = {f: I f/1 = 1}) contains 0. Since no infinite subset of E is bounded. (otherwise square both sides and contemplate the harmonic series). 2. The one above is simpler. (The sum of a finite number of square-summable sequences is square-summable. k} is a basic weak neighborhood of 0. This leads to a contradiction: the unbounded sequence {f~} IS weakly convergent (to 0). and does not construct something as explicitly as the preceding argument. If there were a metric d for the weak topology. p. cx 3 . and. e2 . therefore. that l/gkll 2 ~ (3 for all k.) It follows that there is at least one value of n for which L~=t /(g. in particular. The underlying directed set is the set D of all basic weak neighborhoods of 0 (ordered by reverse inclusion). /{g. Suppose indeed that Jn · 28 U: I(f. there would exist a vector/. gi)/ < e for each i (=I. 2 Solution 29. . choose ( !) an element fD of E in D. 2.. Conclusion (from the principle of uniform boundedness): there exists a positive constant (3 such 2 ~ (3. that the sequence {gd of vectors in 1 is weakly bounded. n = 1. The construction of such a net is a standard technique of general topology.. 3. It goes as follows. Adams. The weak closure of the unit sphere (i. en)/) 2 < oo. 1 /f3i/ 29 L"'= Lf= 185 ... The weak non-metrizability of H can be established by proving that no sequence in E converges weakly to 0. 0) < 1/n. the desired result is an immediate consequence of the principle of uniform boundedness. 2 k = 1. gk)) is bounded.) I < e. 0. The first construction of this kind is due to von Neumann [148. Assertion: the origin belongs to the weak closure of E. ··-)is in 1 . Shields. for each/in 1 . in the case at hand it can be described as follows. Conclusion: there is no such metric. . ···. Let {e1 . the sequence <(f. gk) = L~=t cxif3i-+ 2 1 cxif3i· It follows that.) The net Dr--+ fD converges weakly to 0. en)/ < ejJn. for each D there is a positive intege~: n such that Jn en ED. k). so that clearly gk E 1 . cx 2 .. Write gk = <f3 1*.e. Obvious extension: for each positive integer n. en)/ < ejJn for each i. It is sometimes good to remember that every cluster point of a set is the limit of a net whose terms are in the set. in Sn such that d(fn. (3k *. 3. it follows that L:'= 1 (L~= 1 / (g. it was discovered by A. 380]. e3 . · · ·. then. i = 1. . T. · · · .WEAK TOPOLOGY Solution 28. g. Proof: if D 0 is a (basic) weak neighborhood of 0. en) / 2 < oo for each i. see Problem 20. then. If n is chosen so that this inequality is satisfied.. the weak closure of the sphere Sn = {f: II f I = n} contains 0. in other words. If DE D. .. L. It proceeds by contradiction. 0. and therefore /(J/i ·en. 0.. (The first paragraph above shows that such an element always exists. 2 i. then (f. then D 0 ED and fD E D 0 whenever D c D0 • An alternative elegant proof was suggested by G. · · ·)..Iff= <cx 1 .e. for each n. but it involves no computation.

to something not in the space. then (f. To make this precise. Suppose that X is a measure space with a-finite measure Jl. sufficiently general version of the closed graph theorem. gK) = 0. k = 1. It follows from the principle of uniform boundedness that {gn} is bounded. Since ~ is unbounded. there exists a net J 1-t g 1 of vectors such that ~if) = ([. so that ~ifo)-+ (/0 . apply it to the linear transformation f 1-t fg* from V into V. including all a-finite ones. Since limn (f. (b) Every Hilbert space is sequentially weakly complete. on an infinite-dimensional Hilbert space such things always exist. g) for all f. 30 Solution 30. so that {gn} is weakly bounded. For those who know about the closed graph theorem. g 1 ) for each f and for each J. g) for each fo. and. that is impossible. and suppose that g is a measurable function on X with the property that its product with every function in L 2 (J1) belongs to L 1 (J1). Let {Ek} be an increasing sequence of sets of finite measure such that Uk Ek = X and such that g is bounded on each Ek. it provides an alternative approach.. This means that gn--+ g (weak). Suppose indeed that g1 -+ g weakly. but not quite. This Cauchy net cannot possibly converge weakly to anything. gn) = (f. Consider the linear functional ~ 1 that is equal to ~ on M 1 and equal to 0 on M/. then {(f. see Problem 58. The rest of the proof is the obvious modification of the preceding discrete proof. let M1 be the (finite-dimensional) subspace spanned by the e/s with j in J. (The finite sets J are ordered by inclusion. gn)} is a Cauchy sequence. then ~ 10 (/0 ) = ~(/0 ). and since the boundedness of {gn} implies that the linear functional ~ is bounded. corresponding to each finite subset J of the index set. for each f in H. fixed throughout what follows. it follows that there exists a vector g in H such that limn (f. it follows that ~(/0 ) = (/0 . let~ be an unbounded linear functional.) Given/0 . it follows that {g 1} is a weak Cauchy net. · · ·. If both J and K include 1 0 . just replace sums by integrals. and therefore bounded.) Write gk = XEkg* (where XEk is the characteristic function of Ek).SOLUTIONS The method generalizes to many measure spaces. so that {gn} does indeed have a weak limit. the conclusion is that g belongs to L 2 (Jl). Since the ~/s are bounded (finite-dimensionality). (a) The idea is that a sufficiently "large" Cauchy net can turn out to be anxious to converge to an "unbounded vector". For a discussion of an almost. 3. let ] 0 be a finite set such that foE M10 . (Here is where u-finiteness comes in.) Then let {ei} be a Hamel basis. 186 . (Use a Hamel basis to make one. i. 2. PROOF. gn) = ~(f) exists for each f in H. g1 ) (f. of course.e. If {gn} is a weak Cauchy sequence in H. g) for each fixed/0 • As soon as 1 0 is so large that / 0 E M10 .

for each fixed A. and. the functional v. the Taylor series off converges uniformly. It is 31 convenient to present the proof in three steps.f dJ. then f(A_) = rcr JD There is no loss of generality in restricting attention to the unit disc D 1 in the role of D. (1) If Dis open disc with center A. the inner-product space A2 (D) is complete. accordingly.L(Z). that f E A 2 ( =A 2 (D 1 )) with Taylor series L:'=o cxnzn. the general case reduces to this special case by an appropriate translation and change of scale. then 1 Iv. D 1 = {z: lzl < 1}. ED and f E A2 (D). and radius r.L(Z) = ~ j f(z)dJ. 11!11· v rcr 187 . and let D. (2) If v. is linear. consequently.. 0 < r < 1.) whenever A. it is term-byterm integrable. and iff E A2 (D)..) is the radius of the largest open disc with center A.. PROOF.L(Z) 2 • = IXo • rcr Since If I is integrable over D 1 . that is entirely included in D. For each region D. If r = r(A. This implies that L/(z)dJ. be the disc {z: lzl < r}.CHAPTER 4 Analytic Functions Solution 31.L as r.L.JD. In each D..(f) = f(A. n~O 1Xn Lrzn dJ.1. 0 < r < 1. Return now to the case of a general region D. then.::.(f)l ~ r. the proof of (1) is complete. it follows that SDJ dJ. since IX 0 = f(O). Suppose.

so that r(A) is bounded away from 0 when A E K. then the sequence {J. and it follows also (put n = m) that llenll 2 = 1.SOLUTIONS Let D0 be the largest open disc with center A that is entirely included in D. 188 . It follows that a subsequence of Un} converges to g almost everywhere. and consequently. from the point of view of the standard theory of Hilbert spaces. 1 = nr 2 1 f(AW the proof of (2) is complete. in fact.. If.(z) = 12" rei<n-mwPn+mP dp de Jo Jo r"+m+2 = 2nbnm n+m+ 2. 24].-r(-A) llfn. the proof above is in [64]. then JDr I z"z*m dp. The latter makes explicit use of the Riesz-Fischer theorem (the completeness of L 2). em) is routine calculus. This implies that there exists an analytic function f on D such that fn(A)--+ /(A) for all A in D.fmll for every A in D.} of functions is uniformly convergent on K. 1 32 Solution 32.(z) ~ I lf(z)l 2 dp. and hence that f = g almost everywhere. as before. and hence that A2 (D) is complete. but not necessarily analytic function g on D such that fn --+ g in the mean of order 2.) implies the existence of a complex-valued. At the same time the completeness of the Hilbert space V(p. p.(z) (by (1)).e. then (en. em) = 0.} is a Cauchy sequence in A2 (D). instead of proving it in the particular case at hand. This proves orthonormality. It follows (put r = 1) that if n =1= m. Since 11/11 2 = I lf(zW dp.(z) I = nr 2 -\I 1 2 2 (by the Schwarz inequality) l-\ I nr JDo f(z)dp. Suppose that {!. it is simpler than the analytic argument given by Bergman. r(A) is the radius of the largest open disc with center at A that is entirely included in D. This implies that f is square-integrable. that f E A2 (D). Dr = {z: lzl < r}. It follows that if K is a compact subset of D. here. (3) The proof of the main assertion is now within reach. It follows from (2) that lfn(A). i.(z) JD JDo ~ nr JDo f(z)dp. square-integrable. The evaluation of the inner products (en. These facts were first discussed by Bergman [11.fm(A)I ~ -=-.

The trouble is that the kind of convergence it talks about is wrong. em) as r ~ 1. The point is that in the absence of completeness the convergence of Fourier expansions cannot be guaranteed. q. it is term-by-term integrable. then f(z) = l:. is this. which uses completeness in a less underhanded manner. e 2 . these facts by themselves do not imply that the series converges in the metric (norm) of A2 • There is a simple way around the difficulty: prove something else. r2m+2 =rxm m+1. en) = J nj(n + 1) · rxn. the series J<z)z*m = I 00 rxnznz*m n=O converges uniformly. Specifically. The argument proves that the orthonormal set {e 0 . ···}is maximal.) An alternative proof that the en's form a basis. (Cf. If f E A2 . this shows that each fin A2 is a linear combination of the en's. i Since If· em* I is integrable over D 1. it follows that f ·em* df.e. The argument is almost right. then (f. that Svr I nlrxnl n 2 n=O +1 converges. This implies that 00 rn+m+2 f(z)z*m df. this conclusion squarely meets and overcomes the obstacle that stops the naive argument via power series expansions. and.l ~ Jv 1 f ·em* dJi = (f. consequently. In each D. · · ·. and the proof is complete. 1. Note that the argument above makes tacit use of the completeness of A2 . It follows that the series whose n-th term is Jnj(n + 1) · rxne. with Taylor series I:'=o rxnzn. Problem 55. 189 .l(Z) = L CXn · 2nbnm Dr n=O n +m +2 n. and this is an immediate consequence of the second statement in Problem 32 (the statement about the relation between the Taylor and Fourier coefficients). then f = 0. 2.%o rxnF/{!i + 1) · en(z). via the Bessel inequality. Although L:'=o rxnzn converges to f(z) at each z. e 1 . This implies.d. it is sufficient to prove that if f E A2 and f j_ en for n = 0. Iff E A2 .ANALYTIC FUNCTIONS To prove that the e. 0 < r < 1. a maximal orthonormal set deserves to be called a basis only if the space is complete.s form a complete orthonormal set. and even uniformly in each compact subset of the disc. The proof of the special case can be adapted to the general case.(z) converges in the mean of order 2. That statement is a straightforward generalization of (I) in Solution 31 (which is concerned with e0 only). as follows. with Taylor series L:'= 0 rxn zn. it is tempting to argue as follows.

e. n (J 34 it follows that iff = f*. it follows that the constants are the only real functions in H 1 . It is sometimes useful to know that this last argument applies to L 1 as well as to L 2 . Like the assertion (Problem 33) about real functions in H 2 . it follows that a subsequence of them converges to f almost everywhere. If. Reason: for each term 1Xn/3k-n' either n < 0. then 1Xn = IX_/ for all n. Iff and g are in L 2 .. in which case k . For any f in L 2 (not only H 2 ). n=O I--< oo. or n ~ 0. n + 1 I~. Formally the assertion is almost obvious. 1Xn =IX_/. IX 2 . /3 1 . it follows that IX_/ = fe-n* dfl)* = f*en* dfl. The trouble with this argument is its assumption that complex conjugation distributes over Fourier expansion. moreover.l. n=O oo I1Xnl2 with the inner product of <IX 0 . then. The assumption can be avoided this way: since 1Xn = J fen* df. f E H 2 . IX 1 . i. with Fourier expansion f = 1Xnen. and the desired result follows from the continuity of conjugation. with Fourier expansions n m then If. then it follows that 1Xn = 0 whenever n =I= 0.SOLUTIONS The result establishes a natural isomorphism between A2 and the Hilbert space of all those sequences < IX 1. indeed. f and g are in H 2 . ···)and <f30 . then 1Xn/3k-n = 0 whenever k < 0. so that iff= f*. the assertion is formally obvious. in which case 1Xn = 0./32 . IX 2 . In J Solution 34. ···)given by oo IX/3* n+1 33 Solution 33. so that 1Xn = 0 whenever n < 0. in the mean of order 2. complex conjugation In = yields f* = IIXn*e_* = IIXn*e-n n IIX-n*en. that assumption must be justified or avoided. It can be justified this way: the finite subsums of 1Xnen converge to fin the sense of the norm of L 2 . This assumption can be justified by appeal to the subsequence technique In 190 . The trouble with this argument is the assumption that the Fourier series of a product is equal to the formal product of the Fourier series of the factors. moreover. and hence that f = IX 0 . so that IXn = f3n = 0 whenever n < 0. • • -) for which IXo.n < 0 and therefore f3k-n = 0.

Since.r2 ") 2 + f3r + y.l 2 (1 n=O k Fork fixed. Since the Fourier coefficients Yn of gek * are given by Yn = it follows that Jgek*en* djl = Jgek+n * djl = fgek * dJ1 = n n f3k+n• f L r:xn Y-n = L r:xnf3k-n• It is an immediate corollary of this result that iff E H 00 and g E H 2 . the 0-th Fourier coefficient of fg is given by Apply this result with g replaced by the product gek *. it converges in every other useful sense. It is true also that iff E H 00 and g E H 1 . for each fixed r. since f3r ~ f3 for all r. this can be done because the finite sum is a polynomial in r (and hence continuous) that vanishes when r = 1.r ") . then the f3. Solution 35. then fg E H 1 .ANALYTIC FUNCTIONS used in Solution 33.'s are bounded. g*) is equal to L r:xnf3-n (by Parseval. then fg E H 2 . chooser so that I:=o fan/ 2(1. and this 2 implies that L:'= 0 I r:xn 1 ~ 1 + y. it follows. in particular. conversely. in other words.r 2 ") < 1. Suppose now. and by the results of Solution 33 on the Fourier coefficients of complex conjugates). 191 . that f3r ~ y for all r. then ({Jr(z) = L:'=o r:xnr"z" for 0 < r < 1 and Iz I = 1.'s (0 < r < 1) be bounded. the latter series converges uniformly on the unit circle. Since II({Jrll 2 = L:'=o lr:xnl 2 r 2 ". If qJ(z) = L:'=o r:xnz" for lzl < 1. Conclusion: L~=o lr:xnl 2 ~ 1 + y for all k. Alternatively the assumption can be avoided. it follows that if f3 < oo. L lrx. and hence that (/Jr E H 2 . In one direction the result is trivial. It follows that for each positive integer k. then a necessary and sufficient condition that f3 < oo is that the f3. 35 ~ ~ n=O I k 1r:xnl 2 (1 . but the proof requires one additional bit of analytic complication. the second (and principal) assertion reduces to this: if f3 = L:'=o lr:xnl 2 and f3r = L:'=o lr:xnl 2 r 2 ". The inner product (f. that the Fourier series expansion of ({Jr in L 2 is L:'=o r:xnr"en. as follows.

Let H+ be the set (pointwise vector space) of all those functions f + on X+ whose restriction to X. y) = (Ky.SOLUTIONS 36 Solution 36. and. it follows (by the result just obtained) that the kernel function K of A 2 is given by 1 00 K(x. ·) form an orthonormal basis (see Problem 32).z) 2 when lzl < 1 (discover this by integrating the left side. ej)ei = L ej(y)*ei.g. with orthonormal basis {ei} and kernel function K. and therefore that If+ ( <p) I can be large for unit vectors f+.. it follows that 1 1 K(x. is in H.(1 . 37 Solution 37. 2. Iff= L~o r:xnen and if IYI < 1. with restrictions jand g. H+ is a Hilbert space of functions. then 1(Y) = L:'=o r:xnyn and consequently 1(y) = (f. where Ky = L:'=o y*nen. and since <p is not bounded. If His a functional Hilbert space. say f.. y) = . for each yin X. g+) = (f. 1. consequently. g). zn for lzl < 1 (n = 0. write Ky(x) = K(x. write (f+. Ky). and whose value at <p is equal to <p(f). and adjoin a point that acts like an unbounded linear functional.) Since L:'=o (n + 1)zn = 1/(1 . consider the Fourier expansion of Ky: Ky = L (Ky. it follows that I<p(f) I can be large for unit vectors f. and write x+ =Xu {<p}.) The vector space H+ with this inner product is isomorphic to H with its original inner product (e. y). (Equivalently: define the inner product of the extensions off and g to be equal to the inner product off and g. y) = . (Equivalently: extend each fin H to a function f + on x+ by writingf+(<p) = <p(f). j j Parseval's identity implies that K(x. To be specific: let <p be an unbounded linear functional on H (such a thing exists because H is infinite-dimensional). or verify it by expanding the right). and. n As for H2 : by definition it consists of the functions 1 on the unit disc that correspond to the elements f of H 2 . over a set X say. over a set X say.) Iff+ and g+ are in H+. . (Note that x and y here are complex numbers in the open unit disc.xy *)2. j In A the functions en defined by en(z) 2 = j(n + 1)/n. Start with an arbitrary infinite-dimensional functional Hilbert space H. via the restriction mapping). This proves two things at once: it proves that 1 H 1(y) is 192 . Kx) = L ej(x)ej(y)*. Since <p Ex+ andj+(<p) = <p(f) for all fin H.L (n n n=O + 1)xny*n.

y) = 1/(1 - xy*). so that f = 0. Ky) 38 = f(y) = (f*(y))* = (f*. n=O lxl < 1. It follows that the norm of every vector in H can be calculated from its inner products with the values of K. IYI < 1. if. K/). and hence K is real. Yi)) 1/2. Ky)* = (f. that the vectors Ky (y E X) span H. y) = L xny*n. of course. Li Ct:JKy)l' IILJ Ct:}Ky)l where the supremum is extended over all non-zero finite linear combinations of the form Li Ct:1Ky1 • The numerator is equal to It Ct:/f(yj) I· and the denominator is equal to (~~a. Precisely: I !II =sup l(f. That settles necessity. conjugation is isometric. The norm off* is. PROOF. a/ K(y. a similar supremum. and therefore (f. Ky *). then (f*. Reason: if f . Ky)* for all f and all y.. then f(y) = (f. Consequence: Ky = K/. with numerator which is equal to 193 . Ky) = (f. Ky) = 0. Observe. In closed form: K(x. A necessary and sufficient condition that conjugation in a self-conjugate functional Hilbert space be isometric is that the kernel function be real. in addition.l Ky for all y. Solution 38. sufficiency lies a little deeper.ANALYTIC FUNCTIONS a bounded linear functional (so that H2 is a functional Hilbert space). to begin with. and it proves that the kernel function of f(Z is given by <Xl K(x. It is always true that (f.

In view of this solution. it follows that if K is assumed to be real. The most useful way to approach the problem is to prove that the values off (on the unit circle) are.(z) = ](rz). lzl = 1.. 1). an inner product is defined in cz by the positive matrix A 11!11 = = ( ~i :).£ llfn .(y). Ky)l .£ r. which is exactly the same as the original denominator (for f). y) is always equal to (K(y. are in H (see Problem 35).(y). the assertion is that f. and if IYI < 1.*rxjK(yj. 2 0 < r < 1. the preceding paragraph implies therefore that the suprema defining I f I and I f* I are the same.*rxjK(y. the result is (~ ~ IX. then vrs and 11/*11 = = 1. in some sense.z · Solution 40.].. then IJ.. Let K be the kernel function of H2 (see Solution 37)..](y)l Since I(J..](y)l 40 ..!II· IIKyll· it follows that if Iy I . 1 Since K(x. x))*.£ 11/. Iff.!11· 1 1 _ . and iff is the vector (i.~ in Hz.) 194 . it is easy to construct finite-dimensional examples in which conjugation is not an isometry. The proof is complete. write f.SOLUTIONS and denominator same as for f The new numerator (for f*) is obtained from the old one by replacing each rx j by its conjugate.. ~ f (as r ~ 1) in the sense of convergence in the norm of Hz.)) 1/Z.. y)r z. 39 Solution 39. y. (The boundedness of] is not relevant yet. for instance. The functions f. The function J determines f-but how? Taylor and Fourier expansions do not reveal much about such structural properties as boundedness. then f IJ. If that replacement is made in the denominator. For this purpose. limits of the values of J (on the unit disc).. then the last displayed expression is equal to (f ~ rx. The set of all finite linear combinations of the form Lj rxjKyj and the set of all those of the form Lj rx/ KYj are the same set. If.

41 J on the disc and the essential Consider the following two assertions. can be replaced by non-tangential convergence. it is true that f. then J is bounded. the assertion about the boundedness off is an immediate consequence.. that if a point z in the disc tends radially to a boundary point z 0 . moreover. (2) will be used below without any further apology. Both these assertions are known parts of analysis: (I) is a consequence of Fejer's theorem about the Cesaro convergence of Fourier series.. I ~ 1 and such that f. say.ll 2 = n=O L lanl 2 (1 2 00 . find trigonometric polynomials f. for almost every z 0 • The result can be strengthened. using (1).. and.r") 2 • It follows that for each positive integer k llf.) PROOF. These analytic delicacies are at the center of the stage for some parts of mathematics. The assertion f. such that If. but after that it will be buttressed by the outline of a proof. it follows that f. then f. 1!1 ~ I.f. ... --+ f is true in a sense different from (better than?) convergence in the mean of order 2.r") + L n=k+l lanl 2 · The desired result (II! .ll 2 ~ n=O L lanl 2 (1 k 00 . and. consequently.t fin the norm of L 2 • 195 . In any case. assume (this is just a matter of normalization) that If I ~ 1. llf. norm convergence is enough. then the function value ](z) tends to f(z 0 ). f is in H 00 . then so are the fn's. Iff E H 00 . II J II 00 = II f II oo. (2) seems to be far better known. Given fin H 00 . in the context of Hilbert space. then there exists a sequence Un} of trigonometric polynomials converging to f in the norm of L 2 . (The norms are the supremum of supremum off on the boundary.t f almost everywhere for a suitable subsequence {rn}. n .f. Since convergence in L 2 implies the existence of subsequences converging almost everywhere. (I) If /E L 00 . This says. then lp(z)l ~ I whenever lzl < I. if.. in fact. and. Of the two assertions. and (2) is the maximum modulus principle for polynomials. and that. It is easy to derive the boundedness conclusion about J from the two assertions of the preceding paragraph. Solution 41.ANALYTIC FUNCTIONS To prove the assertion. (I) will be used also. in other words. for instance. radial convergence..t f almost everywhere. in fact.= L:'=o anr"en. recall that iff = L:'=o an en. such that lfn I ~ 1 for all n.f. rn . II is small when r is near to 1) is now easy: choose k large enough to make the second summand small (this is independent of r). (2) If pis a polynomial and if lp(z)l ~ I whenever lzl = I. .. and then choose r near enough to 1 to make the first summand small.t I.

it is obvious. write Dk = Li+=k-k ei (k = 0. Iliz) I ~ 1 for all n and all z.. 2 1 (Computational trick: note that if lzl = 1. it follows from Problem 39 that /.iei )= -k Clearly sk-> fin L . Since. I z 1 = 1) write J Dk(z) = 1 + 2 Re L zi (k = 1.zl 2 = 2 Re(1 . 0.. thefn's themselves are (can be chosen to be) in H 00 .__according to (1). and.z).. 2. The inequality I J I oo ~ II f I oo is implicit in the proof above.z" . ·). it follows that Kn dJ! = 1.. It remains to look at a proof of (1). 2.(z)--.nen. it does not yield the necessary boundedness results. For the proof. Conclusion: 1/(z)l ~ 1 for all z. for z # 1 (but. it follows that Kn(z) . write +k (k = 0. ](z) whenever Iz I < 1. 3. j= 1 k and apply the formula for the sum of a geometric series to get Db)= 2 Reek1~z:. respectively. n k=O the sequences of functions Dk and Kn are known as the Dirichlet and the Fejer kernels. 3. moreover. as asserted.e. 0. · · ·). This is proved by computation. then 11 .) This turns out to be good enough: if If I ~ 1. I1 . 2.and 1 n-1 Kn = . but the proof is non-trivial. To get the reverse inequality. Since JDk dfl = Je0 dfl = 1.. observe that the sum telescopes.. fortunately. into the interior are polynomials. then so are the tn's. Re zn ~ 1 (recall that Izn I = Iz In = 1). 1 .. ··-). ·).. it does not follow that lskl ~ 1. Since fn-> f in the norm of H 2 . i. 196 . use Solution 40 (f...Re zn . (Note that iff E H 2 .) The remedy is to consider the averages 2 1 n-1 tn =- n k=o 2: sk (n = 1. 1.. (If 1!1 ~ 1. but this is not good enough.L Dk (n = 1. {In fact it is known that tn -> f almost everywhere.2 ). sk = L IY. the fact is not needed here. and get Kiz) =- n Re 1.) Substitute into the expression for Kn. 2.. then it does follow that Itn I ~ 1. .-> jas r-> 1). it follows that their extensions J.z This makes it obvious that Kn is real. For z = 1. and in fact positive. . · · ·). Iff= Ln IY.) Clearly tn-> fin L 2 . of course. 2. 3. By {2).SOLUTIONS Since. 1. The principal property of the Kn's is that their values are real.

g E H 2 . This way leads to some not overwhelming but extraneous analytic difficulties... no such thing ash is defined. namely to note that h E H 1. The trouble with the question as phrased is that it is easier to answer than to interpret. then. consequently. In other words: formal multiplication applies to both Fourier and Taylor series. JKn(y)f(y*z)df. (There is another way out. and 197 . this implies that if If I . 1. by the Lebesgue bounded convergence theorem. Reason: if it is in the norm. Here is one more technical remark that is sometimes useful: under the assumptions of (1) it makes no difference whether the convergence in the conclusion is in the norm or almost everywhere... say. then a subsequence converges almost everywhere.) Once the question makes sense. Solution 42. and.. f from u.. then ltn(z)l. and h = fg. by Problem 34. then h is not necessarily in H 2 .-l(Y) JDk(y)f(y*z)d{t(y). f Kn(Y)If(y*z)ldf..-l(Y). In order to motivate the construction of. Iff and g are in H 2 and h = fg. if it is almost everywhere. Iff E H 00 . Solution 43. Suppose therefore thatf E H 2 with Fourier expansion! = L::'=o ()(nen.-l(y). the result there is that the Fourier coefficients of hare expressed in terms of those of f and g in exactly the same way as the Taylor coefficients of J · g are expressed in terms of those of J and g. and to extend the process of passage into the interior to H 1 . it is also in the norm. f Kn(y)df. it is a 43 good idea to turn the process around and to study the way u is obtained from f. The proof is over... The simplest way out is to assume that one factor (say f) is bounded. then h = ]g. the answer is automatic from Solution 34. 42 PRooF.-l(Y) = 1. and hence the definition given in Problem 35 does not apply to h.ANALYTIC FUNCTIONS To apply these results to j~ note that sk(z) = jxk f f(y)y*izi d{t(y) = = and hence that tn(z) = JDk(y*z)f(y)df. and the question makes sense. in that case Solution 34 shows that hE H 2 . the mapping from one to the other is multiplicative.

then Re Du = u for every real u in L2 . n<O !rxn (n > 0). It is important to note that the Hilbert transform of a bounded function need not be bounded... As for the formulation in terms of v: given u. then V = Imf = ~(f*. If I. then ~ ' 2~n = = l~n - i~n (n > 0). Given u = Ln ~nen. the function u is in L 2 • If the Fourier expansion of u is u = Ln ~nen. put v = Im Du.O l:rxn*e-n. Since the sequence of rx's is square-summable.. it is easy to get explicit expressions for the Fourier coefficients of v.O + L rxn*e-n) ni. these algebraic trivialities are all there is to the Dirichlet problem on the unit disc. or. as above.. Since Iu I . the result can be concisely expressed: 1Jn = ( -i sgn n)~n for all n.SOLUTIONS write u = Re f. (n If Im ~ 0 = 0. the difference f .D Ref is a purely imaginary constant that can be prescribed arbitrarily. If. an element f of H 2 is defined by f = LnsO rxnen. The formal expression for v in terms of u makes sense even when u is not necessarily real. and therefore ~ 0 = Re rx 0 and ~n = { 1 * z!X-n ~n It is now clear how to go in the other direction. It is not quite true that D Ref = f for every f in H 2 . u = Ln ~nen and f = Du = Lni.O rxnen.f)=~ ( ni. with ~ 0 real. and the terminology (conjugate function.l:rxnen)· n<:o - If V = L 1Jnen. in other words (consider extensions to the interior) that unbounded analytic functions can have bounded real parts. then (see Solution 33) u = !(f = Re + /*) rxo = !( L rxnen ni. Since Im Du = .O + L !rxnen n>O + L !rx-n *en. write rxo = ~o and rxn = 2~n = 2~-n* = ~n + ~-/ (n > 0).Re(iDu). 198 .. and. with = ~ -n *. Hilbert transform) remains the same. in particular.. but it is almost true. Write f = Du (D for Dirichlet). (n < 0). Standard example: f(z) = i log(l . < 0). As far as L 2 functions on the unit circle are concerned. 1Jo = Im ~ 0 and 1Jn = { i l · 2~n .z)..

the new matrix elements do not interfere with the old columns. Since e 1 and Ae 1 span a subspace of dimension at most 2. There is a difficulty. e 2 }. it follows that. Since e 1 . e 2 . If they do not. en+ t}. while only feebly used in the proof.en. reduces A. there is no loss of generality in assuming that His separable in the first place. eliminates the discomfort of having to worry about the pathology of the uncountable. however. (A Jacobi matrix is a Hermitian matrix all whose non-zero entries are on 199 44 . it follows that when the en's are embedded into an orthonormal basis. In the infinite-dimensional case (Aei. invariant under A. it follows that. and Ae 2 span a subspace of dimension at most 3. unless dim H = 1. The finite-dimensional case is transparent and. This comment. then no such troubles can arise. e 3 }. Since His the direct sum of separable subspaces that reduce A. from the present point of view. and hence. e. ···. e. This proof.) are different from 0. there is no reason to suppose that the en's form a basis. there exists a unit vector e3 orthogonal to e 1 and e2 such that Ae 2 E V {e 1 . there exists a unit vector e2 orthogonal to e 1 such that Ae 1 E V {e 1 . An inductive repetition of this argument yields an orthonormal sequence {e 1. That is. in the Hermitian case. There is a tempting attack on the proof that is doomed to failure but is illuminating just the same. then the process of embedding them into an orthonormal basis may ruin the column-finiteness of the matrix. e2 . unless dim H = 2. If A happens to be Hermitian. for Hermitian A. e 3 . in any case. it could happen that for some e orthogonal to all the en's infinitely many of the Fourier coefficients (Ae.CHAPTER 5 Infinite Matrices Solution 44. shows more than was promised: it shows that every Hermitian operator has a Jacobi matrix. Let e 1 be an arbitrary unit vector. The span of the en's is. • • ·} (which is finite only in case dim H < oo) such that Aen E V {e 1. uninteresting. e2 .) = 0 when i > j + 1. and everything seems to be settled.

and then find a unit vector e 4 orthogonal to e 1.. e 4 }. e 2 . then = I~ · YL-qj1 -La .) Indeed: if (Aei. ei) = 0 whenever i > ii. e 2 . . eiJ· This last condition implies that (Aei. (Jqj = 2 fJ. the argument is completed by inductively enlarging the en's to an orthonormal basis selected the same way as the en's were. .. and e 3 such that Ae 2 E V {e 1. Let {fhf2.. e 3 . Conclusion: the sequence {e 1 .SOLUTIONS either the main diagonal or its two neighboring diagonals. j 2 1 i !J ' ~ y L 1~·1 j qj 2 _J_. In the non-Hermitian case the argument has to be refined (and the conclusion weakened to the form originally given) as follows. that none of the elements on the diagonals next to the main one vanish. Some authors require also that the matrix be irreducible. Continue in this way.. catching alternately one of the f. ~ 2 . then (ei. If <~ 0 . ) is a finitely non-zero sequence of complex numbers (i. 45 Solution 45. p.) Next find a unit vector e 3 orthogonal to e 1 and e 2 such thatf2 E V {e 1 . · · ·} is orthonormal by construction. there exists a unit vector e orthogonal to M such that g E M v { e}. Aei) = 0 when i > j + 1. • • ·) = / \ ~ IXoj~i' ~ IX1j~j' ~ IX2j~j' · · ·) J J J 200 . ~ 1 . e3 }. and the proof is complete. y L I~jl 2 • j These inequalities imply that the operator A on / defined by A<~o' ~1. ~2. The selection of the needed new e is always possible. e 2 .e. satisfies the conditions. The general lemma is this: for each finite-dimensional subspace M and for each vector g.e. · · ·} be an orthonormal basis for H. ~n = 0 for n sufficiently large).. (Once again restrict attention to the infinite-dimensional case. e 2 . and it has the property that for each n there is an in (calculable in case of need) such that Aen E V {et. · .'s and the next as yet uncaught A en. ei) = 0 when i > j + 1.. Put e 1 = k Find a unit vector e 2 orthogonal to e 1 such that Ae 1 E V {eh e 2 }. i.f3 . it forms a basis because its span contains eachfn. e 3 .

every vector in the range of the operator A under study is a multiple of fo. If fo is the first column.. 1.. The spectrum of A consists oftwo eigenvalues: 0 with infinite multiplicity.} of vectors such that (f. in fact. be defined by f. 1). a sequence {!... j = 0. in some Hilbert space. Solution 49. consequently. (Note that A is Hermitian. and. 2. it is therefore sufficient to exhibit. For the proof. Joo du 1C = vJ+l o u2 + 1= Solution 47. in other words. let the Hilbert space be L 2 (0. and with multiplicity 1. apply Problem 45 with Pi= qi = 1/Ji -l Since the Hilbert matrix is symmetric. 1 1 f o = <1 4• 8• . The verification depends on elementary calculus.(x) = xn. The assertion is easy to prove: since (2-<i+i+ 1 >) 2 < oo. To do that. J. and let the vectors f. j) entry is the inner product u. 201 48 49 . Af = 2(/. In other words. To prove that the Hilbert matrix is positive. The result is a coro~of Problem 45. The answer is that the Gramian matrix ((jj. The determination of the norm is made easy by the special arithmetic structure of the entries of the matrix. that there exist an operator with that matrix).INFINITE MATRICES Solution 46.} of vectors is the matrix whose (i.) L:'= i Solution 48. Li Li 47 . It follows on general grounds that the Hilbert matrix has a unique positive square root. it is completely trivial to prove that every Gramian is positive. it is sufficient to verify one of the two inequalities (with f3 = y = n). The Gramian of a finite or infinite sequence{!.. z. The rest is elementary calculus.jj). · · ·). · . / 0 ) /0 for every vector f It follows that IIA I = 211/o 1 2 = 2 1 l/4n = 2/3. jj) = 1/(i + j + 1) (i. > then all other columns are scalar multiples of fo.)) be bounded (or. as follows: 46 < L'' 2 (x + /: =2 f'-=----. It is not difficult to prove that every positive matrix is a Gramian. A has rank 1. What is it? No explicit description of it seems to be known. boundedness follows from the sufficient condition mentioned in the preliminary discussion of Problem 44.

and that implies (in view of the identity involving I T.h)ajai* = (Ga. Note that the result is a generalization of the numerical fact (Problem 29): the "Gramian" <Pj{J.ll} is bounded. in particular.SOLUTIONS In one direction the proof is straightforward.. this answers the question about The quickest proof of the converse is via uniform boundedness. is bounded.a} is bounded. therefore. r where G is the Gramian (considered as an operator on /2 ). k=l Clearly each T. and. 1). For each n = 1. a1 2 and the Gramian) that the Gramian is bounded. If the Gramian is bounded.ak fk.a L. a). For a fixed a. Conclusion (see Problem 51): the sequence {II T. 3..*) of a sequence of scalars is bounded if and only if the sequence is in /2 • 202 . and that implies convergence. the sequence {T. Note. 2. the sequence {T./. then 11 ~ anfn =(t ajij. and if a is a finitely non-zero sequence.)=~ t (Jj. Consequence: II~ anfn ~ I Gil ·llall L 2 (0. r = 2 .a} consists of the partial sums of a series that is convergent by assumption. ···.let Tn be the linear transformation from P to H defined by n T.. ~<'X. that since the Hilbert matrix is a bounded Gramian (Problem 46).

) For part (d). · · · . by induction. and hence that IIAenll ~ 1 for all n. e1 )(e 1 + · · · + ek) forallf. e2 . and find a Hamel basis for H that contains each en. e 2 . + ek)e1. • • ·} for H such that supn IIAenll < oo. then there exists a unit vector en+ 1 orthogonal to {e 1. Let {e 1 .CHAPTER 6 Roundedness and Invertibility Solution 50. The answer to (c) is no: if a linear transformation A is bounded on each orthonormal basis.sothat. If A were bounded. in fact. Reference: [121]. ···. there exists an orthonormal basis {e 1.. then A is bounded. One way to see that is to imitate the easy beginning part of Solution 27.. It is interesting to observe that unbounded examples for part (a) are not only easy to come by. . It follows that Aen = e 1 or 0 according as n ~ k or n > k. then its vanishing on each en would imply that A = 0. ···}be an orthonormal basis for a Hilbert space H.A*e 1 = e1 + · · · + ek. en} such that II A en+ 1 11 ~ n + 1. then there exists a unit vector e 1 such that 11Ae 1 l ~ 1. whenever e1 . A unique linear transformation A is defined on H by the requirement that Af0 = fo and Af = 0 for all other elements of the selected basis: in particular. .inparticular. n. That is: for every linear transformation A on H. j = 1. (If A is not bounded..en are orthonormal vectors with IIAeill ~ j. el + . Aen = 0 for all n. it is impossible to avoid them. e 3 . and. Let fo be an arbitrary but fixed element of that Hamel basis distinct from each en (see Solution 7). This solves parts (a) and (b) of the problem. e3 .itfollowsthat IIAII = IIA*II ~ IIA*e1ll = lle1 + · · · + ekll = 203 Jk. Since (easy computation) A*f = (f. . choose an arbitrary but fixed positive integer k and define an operator A (depending on k) by 50 Af = (f..

SOLUTIONS A simple alternative way to say all this is to describe the matrix of A with respect to the basis {e 1 .. Fix an arbitrary vector g0 and write T 0 = {A*g 0 : A E Q}. Since j{f. It is sufficient to prove that A* is invertible. 52 Solution 52. j(Af. Since l(g. which are 1's. then llgll ~ 1/b for some £5. As for (e). that have the same unboundedness properties as the ones in (d). Af)l = I(A*g. To prove it.j ~ k and rxij = 0 otherwise. e 3 . Suppose that Q is a weakly bounded set of bounded linear transformations from H to K. consequently. This implies that IIAII ~ y. it is now a projection. consider the transformation 0 in the role of A*: for it the implication from IIA*gll = I to 11911 ~ 1/<5 is vacuously valid. The conclusion can be obtained from two successive applications of the principle of uniform boundedness for vectors (Problem 27). bllgll for some £5 (and all gin K). and. e2 . Summary: it is sufficient to prove that if S = {h: IIA*hll = 1}. specifically. and that. 51 Solution 51. which is true because the range of A is dense in K. and the proof is complete. Next. it is sufficient to prove that A* is bounded from below. the set Tis weakly bounded in K. g) for all A in Q. and therefore there exists a constant y such that IIA/11 ~ Y whenever A E Q and/ E H 1 . go) I ~ rx(f. go). it is sufficient to prove that if IIA*gll = 1. ···}:all the entries are 0 except the first k entries in the first row. The range of A* is dense in H (because the kernel of A is trivial). If the matrix is divided so that each non-zero entry becomes 1/k. Since it is Hermitian. where H 1 is the unit ball of H. Example: rxii = 1/fl if 1 ~ i. then the matrix becomes by idempotent. Jk:. write T = {Af: A E Q. yes. and hence has norm I. (The full force of the assumption that A maps H onto K will be used in a moment. and even Hermitian ones. Caution: the last reduction uses the assumption that the kernel of A* is trivial.f)l ~ {J(g) · 11/11 ~ {J(g).:. g) I ~ rx(f. Conclusion: the original matrix has norm Jk:. This means that IIA*gll . 204 . there are normal matrices. A*go)l = I(Aj. and therefore there exists a constant {J(g 0 ) such that IIA*g 0 ll ~ {J(g 0 ) for all A in Q.) To see the difficulty. the set T 0 is weakly bounded in H.J E Hd..

This cannot be done if H is finite-dimensional. . Assume that these conditions are satisfied. the construction of a diminishable complement can be accomplished by constructing an operator A on H and a proper subspace H0 of H such that ker A = 0 and AH 0 is dense in H. define g 0 so that its coordinates with indices n + 1. indeed. The. Consider. h) (=In (1/n)~n) = 0. ~ 2 . In view of all this. take g in K. and that can be done by proving that it is weakly bounded. p. Since f is finitely non-zero. ···))for which (g..to define H 0 . find fin H so that Af = g. More precisely. All known constructions can be generalized so as to solve Problem 53 also. the plan is to find a vector g 0 such that (g 0 . h)= -rx (so that g + g 0 E H 0 ) and such that 11Ag0 l is small (so that A(g + g0 ) is near to f). . ~ 3 . Solution 53. Question: under what conditions on A is N a complement of M? Answer: if and only if ker A = 0 and ran A = H. · ·-) = <~1> 1~ 2 .. in the role of N 0 . that part of the graph of A that lies over H 0 . p. and let n be a positive integer such that all the coordinates of g after the n-th are equal to 0. only thing that needs proof is that AH 0 is dense in f2. ~ 3 . but by a suitable perturbation it can be put there. The only mild challenge is in the case at hand. To do that. 1. !. Ag) for some g in H 0 . ···). (a) The construction of non-closed vector sums is either laborious ([138. and for that purpose it is sufficient to prove that each finitely non-zero vector fin f2 can be approximated arbitrarily closely by vectors in AH 0 . for example. h) I= l(f. n + p are equal to -rx(n 53 + p 1) -rx(n + 2) p -rx(n + p) p 205 . -i-~ 3 . 28]) or mildly sophisticated (Problem 52). n + 2. it is very easy to do. consider a subspace H 0 of H and. h) = rx. there exists a (unique and necessarily finitely non-zero) vector g such that Ag = f. Question: under what conditions on H 0 is N 0 still a complement of M? Answer: if and only if AH0 is dense in H. consider. For this purpose let p be a large positive integer. That is: H 0 is the set of all those vectors g (=<~ 1 . and observe that l(g. ···)and let H 0 be its orthogonal complement. and let M be the x-axis and N the graph of A in H EB H. an operator A on a Hilbert space H. 21] and [50. The proof is complete. Suppose. ~ 2 . let N 0 be the set of all vectors in H EB H that are of the form <o. in an attempt to diminish N. that (g. let h be the vector (1.ROUNDEDNESS AND INVERTIBILITY then the set S is bounded. A*h)l ~ 11/11 for all h in S. Let H be 12 and define A (as in Problem 52) by A<~ 1 . h)l = I(Af. There is no reason why the vector g should be in H 0 . if H is infinite-dimensional and the kernel requirement is omitted.

and let F be a subset of K such that E u F is a Hamel basis forK. 0) E G (the closure) for each f in K. Solution 55. is a maximal orthonormal set in G. and IIA liz go lct. Tf) with fin K. The cardinal number ofF is c (the power of the continuum). A. Let G be the graph ofT. No. The proof goes as follows. withe in E. and H0 by H EB H EB H EB · · ·.l <e. 0) E G for each e in E. just form the direct sum of infinitely many copies of that example. then f .ct. Note that ran T is dense in L. This. and hence G = H. maximal orthonormal sets need not be total. Let K be a Hilbert space of dimension ~ 0 . Suppose. From this. for g in L.SOLUTIONS and all other coordinates are 0. implies that <f. A EB A EB A EB · · · .l E. To get an infinitely multiple example of the (lattice) algebraic phenomenon observed in the preceding example. and 55 are concerned with the pathology of linear manifolds. 54 Solution 54. that G is a dense linear manifold in H with dim G =1= dim H. and H0 EB H0 EB H0 EB · · · . That is: replace H. One way to get an example is to use Problem 54.lct. hence the span of E 0 in His H. Since ran Tis dense in L. and let T be a linear transformation from K into L defined so that T maps E into 0 and maps F one-to-one onto an orthonormal basis for L. 0) = <O. it follows that (f. g). whence f = 0. in turn. but algebraic ones not. i. Since <e. see Solution 55. therefore G contains every <O. Tf) .<f.e. indeed. For a different approach. the linear manifold G is dense in H. 54. The verification that nothing goes wrong is straightforward. in each case a transformation with dense range can be used to 206 55 . and therefore Tf = 0. · · ·. If E 0 is such a set. then the intersection of its span in H with G is a closed subspace of G that includes E 0 • Hence that intersection is G. (b) Analytic phenomena (having to do with convergence) are likely to misbehave when infinitely repeated.lz -Ppz-p' which is indeed small if p is large.jp. 0) for all e in E. The coordinates of Ag 0 with indices n + 1. 0). Indeed.l2 . hence E 0 is an orthonormal basis for H. Tf) E G for eachfin K. and let H be the direct sum K EB L. let E be an orthonormal basis for K (of cardinal number ~ 0 ). h) = -ct. and that contradicts the first consequence mentioned above. The dimension of H is dim K + dim L = c. in turn. n + 2. An immediate consequence is that no orthonormal basis for H is included in G. it follows that (g 0 .. Pertinent reference: [ 48]. if <f. Let L be a Hilbert space of dimension c. n + pare all equal to . Tf) . the set of all <f. what is the dimension of G? Answer: ~ 0 • Reason: the set of all vectors of the form <e. Problems 53. it follows that there is no orthonormal set that is total for G.

since N 1"1 N 0 . however. is orthogonal to each gj and therefore to K. it follows that A/. Indeed: iff E Hz 1"1 N. Riesz theorem implies that the only function in both L z(E) and Hz is 0. where h E Hz*. Since. ···}is. Assertion: N is a complement of Hz. and its techniques are frequently applicable.l = V(E follows that the complement His infinitely diminishable. accordingly. and hence that f* = h* . and M. therefore. but. if p E P. therefore f = 0 and therefore p 0 (since p E Hz and pj_ Hz). If dim K < dim H. Since Jl(E) > 0. Suppose.l c: Hz*) and f vanishes on a set of positive measure (namely E). by the F. and p + f j_ e. at this point.p. Solution 56. and M. it words Lz(E0 ) is a complement of Hz.E 0 have positive measure. 2. It is short. Riesz theorem implies that f 0. and. consider the Hilbert space Lz and the subspace Hz. 1. For a similar alternative approach to 55. however. the F. The result is that f*en E Hz. the F. and M.p*. there exist non-zero functions in Lz(E). it borrows from the future: it needs the analytic power of the F. the assumed in- 56 equality between the dimensions of H and K implies the existence of an i such that (fi. Riesz theorem implies that f = 0. the range of A is included in K. Riesz theorem. Each A*gj can be expanded in terms of countably many f's.E) > 0.J E V(E). 207 . it does not need much machinery. then there is no loss of generality in assuming that K c: H. Let E be a Borel subset of C (the unit circle) such that both E and C-E have positive measure. In the notation of Problem 33. Let {f. it follows that A/. then f vanishes on a set of positive measure. This implies that f = h . ez.. ez. let P ( c: Hz) be the set of all polynomials. If E 0 is a Borel subset of E such that both E 0 and E . Multiply the latter equation by e. and M. and let N be Lz(E) (=the set of those elements ofLz that vanish outside E). The reason is that its span (in Lz) is Hz. Indeed. ···. Riesz theorem about the Hardy space Hz (Problem 158). then f E Hz* (because (Hz). Since (fi. f*e. So much for Problem 53. maximal in P + Lz(E). and hence. ···}(notation as in Problem 33) is not total in P + Lz(E). A*g) = (A/. If f j_ Hz v N. then the result just proved applies to E 0 as well as to E. = 0. e 1 . this assumption excludes trivial cases only. since Jl( C . Its disadvantage is that.. respectively. = 0. A*gj) = 0 for all j. In other E 0 ). There is a much more sophisticated construction that deserves to be widely known. the F. n = 0. The orthonormal set {e0 . g). e 1 . that A is an operator on H with range included in K. The orthonormal set {e0 . it is to be proved that ker A is not trivial.then p + f j_ Hz.} and {gj} be orthonormal bases of H and K. and consider the vector sumP + L z(E) as an inner product space in its own right. and M. where n is sufficiently large to make p*en belong to Hz. where h* E Hz. Assume that dim K is infinite.BOUNDEDNESS AND INVERTIBILITY construct a large subspace with surprising properties..

Observe first that no non-zero vector in the range of P is annihilated by Q. it follows that Af. 58 Solution 58. it follows that g = Af. then a set of cardinal number dim H is dense in H.) It follows that a set of cardinal number dim H is dense in K. (Use rational linear combinations. [63].. 57 Solution 57.. if Pf = f and Qf = 0.0 would imply liP . and consider the linear transformation B from G to H defined by B(f. is a bounded linear transformation. incidentally.. As for the case of equality: if ran A is dense in K. and therefore/ #. of course. If A is a one-to-one linear transformation from H into K. all is trivial. then. Af). Af) = f. that A is bounded. is based on a nontrivial theory (polar decomposition). The proof of the converse is less trivial. it is a trick based on Problem 52.. A/)11 2 . (It is due. the assertion there is a special case of the closed graph theorem. since IIB<f. Solution 17. If dim H is finite.1 from H into G. say <fn. A/)11 2 = 11/11 2 . but. Afn> ~ (f. the conclusion is that B is invertible.. cf.Qll ~ 1. L. cf.1 / = (f. with polar decomposition UP (see Problem 134). If dim H is infinite.. Afn>} be a sequence of vectors in the graph of A converging to something. to G. that 11/11 2 + IIA/11 2 . Clearly B is a one-to-one mapping from G onto H... From this it follows that the restriction of Q to the range of P is a one-to-one bounded linear transformation from that range into the range of Q. This.. Let {<J~. Since fn ~ f and A is continuous. first. which. all/11 2 for some a (and all fin H). is not an especially helpful comment for someone who wants to know how to prove the closed graph theorem. Indeed. for a statement that is completely natural. it follows that U is an isometry. Since G is a closed subset of the complete space H EB K. and this implies that dim K . since ker A is 0. Suppose that A is a linear transformation from H to K. Equivalently the conclusion says that the mapping B. This means. and therefore that the rank of P is less than or equal to the rank of Q (Problem 56). however..) There is a quick proof. It goes as follows. Let G be the graph of A. by definition. The proof above is elementary.Qfll = 11/11. and hence that (f. and all is ready for an application of Problem 52. since at the same time Afn ~ g. it follows that B is bounded. It is worth remarking that the derivation of the result from Problem 52 is reversible. dim H. then ran U is equal to K. g). the boundedness of A is an immediate consequence. ~ Af. Weiss. g) is in the graph of A. 11/11 2 + IIA/11 2 = II(/.SOLUTIONS Consider next the statement about equality. and not just how to bounce back and forth between it and a reformulation. it is not at all completely obvious. 208 . The conclusion follows by symmetry. then liP/ . defined by B. it is complete. and suppose.

59 < moreover. n n and therefore f' = Af. n = 1. g). 3~ 3 . · • and g = <11 1 . (b) This is an easy consequence of the closed graph theorem. f'. ~ 2 . That is.. in an obvious manner A is determined by the diagonal matrix whose sequence of diagonal terms is (1.with ~n = 0 for all sufficiently large n. since Afn = Bgn for each n. ···). this proves that A is closed. Since. 1] 2 . and therefore bounded. use the principle of uniform boundedness directly. Af and Bgn-. ~ 3 . obviously. g) = lim (Aj.) The verification that X is linear is routine.. (There is no ambiguity here. for all g. if A is symmetric. then g = 0. Ag) are equal to 1 n~n11n*· The linear transformation A is not bounded.-. Ag) = (Af.. the image under A of the unit ball is weakly bounded. ~ 3 . <f. indeed if {/. ~ 3 .LB is closed. an element ofHisa sequence <~ 1 . then. 3.-. the" sufficiently large" may vary with the sequence. Reason: if Af = 0. and. 1] 3 . g) and (f. f and Aj. The linear transformation A is symmetric. Indeed. g)= L:'=t ~n17n*· Let A be the linear transformation that maps each sequence <~ 1 . (a) On incomplete inner-product spaces unbounded symmetric transformations do exist. so that A = BX. 2~ 2 . · · ·. g) belongs to the graph of X. Bg.. gn) -. and iffn-. ···). so that g belongs to ker. Af = BXf for all f.. ~ 2 . and hence that A is bounded. · · -). ~ 2 . it follows that Af = Bg. If llgll ~ 1. The answer is yes. then l(f. g) = lim(/. Solution 60.write (f.} is the sequence whose n-th term is I and all other terms are 0. PROOF. write Xf =g. Alternatively. 209 . Define inner product in H the natural way: iff = <~ 1 . Ag) = (f.LB).ROUNDEDNESS AND INVERTIBILITY Solution 59. Indeed.LB. it follows that <J. ·> L:'= (f'.. indeed both (Af. ker. gn) is in the graph of X. and fn. then A/.. then llfnll = I and IIA/. Note first that ran B = BH = B(ker. Ag)l = I(Af. if <fn. every symmetric linear transformation is bounded. ··-)onto <~ 1 .LB with Af = Bg. 60 (a) Let H be the complex vector space of all finitely non-zero infinite sequences. g). g) I~ IIA/11 for all f. in other words.11 = n. • ·-). (b) On a Hilbert space. 2. 3. It follows that for each fin H there is a unique g in ker. all that needs proof is that X is bounded.. 2. That comes quickly from the closed graph theorem..

The construction is reasonably straightforward.l:J ~ 1 e1 = LJ a1 ~ 1 e1 . then there exists a sequence {~n} such that Ln I ~n 12 < oo but Ln lan~nl 2 = oo. the assertion is this: if {an} is not bounded. with Ae1 = a1e1. then so that {a1} is bounded and sup1 la1 1 ~ IIAII· The reverse inequality follows from the relations IIA ~ ~jejr II~ aj~jejr ~ laj~j~2 = = ~ (s~p 1a11Y. If {an} is not bounded. 62 Solution 62. all it takes is a slight change of PROOF. Clearly A is a diagonal operator. then Ian I takes arbitrarily large values. the preceding computations imply that A is an operator. If {an} is a sequence of complex scalars. ~ ~~jl 2 = (s~p 1aj1Y ·I/~ ~jejr Given a bounded family {a1}. such that 2 2 Ian ~n 1 < 00 whenever Ln I~n 1 < oo. 210 . then {an} is bounded. If A is a diagonal operator. Ln Expressed contrapositively. define A by A . The proof of uniqueness is implicit in the construction: via Fourier expansions the behavior of an operator on a basis determines its behavior everywhere. and the diagonal of A is exactly the sequence {a1}. There is no loss of generality in assuming that Ian I ~ n.CHAPTER 7 Multiplication Operators 61 Solution 61.

of positive measure. ···. not just a set of complex numbers. and hence that the sequence {o:n} is invertible.< Ln Solution 63.en.11. "the closure of the diagonal" does not make rigorous sense. then II A II = II q> II oo (=the 1 .£ = fM 1 · dfl = JL(M). the omission of some o:'s. If A is the multiplication induced by a bounded measurable function q> on a a-finite measure space. and. If.} is bounded away from 0 if and only if A. is invertible if and only if A. IIAfll 2 = J1q> · fl dfl 2 . say M. then llfll and 2 .A. it would be a pity to let the purist have his way. therefore. Indeed. if {fJn} is a bounded sequence such that o:nf3n = 1 for all n. 3. As for the spectrum: the assertion here is that A . Let J1 be the underlying measure.£ llq>lloo 2 · J1fl dfl 2 = llq>lloo ·llfll 2 2 . (The purist has a small right to object. then the diagonal operator B with diagonal {fJn} acts as the inverse of A. as a cluster point. the proof is complete.) The assertion is equivalent to this: {o:n -A. and.MULTIPLICATION OPERA TORS notation. ~n = 1/o:n. n 2 but lo:n~nl diverges. Since PROOF.enll IIA . then Iq>(x)l > llq>ll 00 . it follows that II All llq>lloo· In the proof of the reverse inequality a pathological snag is possible. this implies that the sequence {1/o:n} is bounded.A. Solution 64. The usage is an instance of a deservedly popular kind of abuse of language. then A and {o:n} are invertible together. The diagonal is a sequence of complex numbers. Since this is obvious. so that 1 A len=.} has 0 as a limit point if and only if the set {o:n} has A. unambiguous and concise. is not in the closure of {o:n}. A sensible way to begin the proof is to note that if e > 0.e on a set.then l~nl 2 n n oo. The assertion is that if A is a diagonal operator with diagonal {o:n}. I . Contrapositively: the sequence {o:n .1 (ctnen) = en. possibly. 63 since IIA .£I_. until further notice that assumption will not be used. It is instructive to see how far the proof can get without the assumption that J1 is a-finite. then A . 2. Iff is the characteristic function of M. Conversely: if A is invertible. 211 . n = 1. does not belong to the closure of the diagonal {o:n}.£ 1Xn 1 64 essential supremum of Iq> I). in that case.

otherwise the best that can be asserted is the inequality IIAII ~ llq>lloo· Every finite or a-finite measure is locally finite. The objection may not seem very strong. Conclusion: if the measure is locally finite (meaning that every measurable set of positive measure has a measurable subset of finite positive measure). so is its quotient by f To prove boundedness. I q> I 00 . is that the measure space may be pathological enough to admit measurable sets of positive measure (in fact infinite measure) with the property that the measure of each of their measurable subsets is either 0 or oo.e)llfll. since q> • f is in L 2 . Iq> I ~ IIA II) almost everywhere on that set. After all. Measurability is easy. for instance. of course. consequently. then q> = 0.u is. even if the measurable set {x: Iq>(x) I f.u({xd) = l and . observe that llq>" ·!II = IIA"fll ~ IIAII" ·IIIII for every positive integer n.SOLUTIONS It follows that IIAJII f.) From this it follows that if f =F 0 on some set of positive measure. then llq>lloo = l.e} has infinite measure. This proof is quick.u({x2 }) = oo. This is true.e. it is not the one that would suggest itself immediately. and prove that M must have 212 . the solution (as stated above) is complete. and hence that IIAII f. (llq>lloo . A more natural (and equally quick) approach is this: to prove that Iq> 1~ IIA I almost everywhere. otherwise write l/J = q>/IIAII. the given a-finite measure. and rewrite the preceding inequality in the form (Here . then 11/J I ~ 1 (i. If A = 0. and there is nothing to prove. then the norm of each multiplication is the essential supremum of the multiplier. but a little too slick. then the conclusion is that Iq> I ~ IIA I almost everywhere. however. llq>lloo .U) is the one-dimensional space consisting of all those functions on X that vanish at x 2 • If q> is the characteristic function of the singleton {x 2 }. The most practical way to avoid excessive pathology with (usually) hardly any loss iri generality is to assume a-finiteness. Iff is chosen (as above) so that f =F 0 almost everywhere.. let M be a measurable set of finite measure on which lq>l > IIAII. If. since this is true for all e. it follows that IIA I f. then L 2 (. The difficulty. What is wrong is that M may have infinite measure. If that is done. and. llq> I 00 • The proof is over. the reasoning above works perfectly well if M is taken to be a measurable subset of finite positive measure. There is no way out of this difficulty. Since the measure is a-finite. but the norm of the induced multiplication operator is 0. it is measurable.e. X consists of two points x 1 and x 2 and if . but it is wrong. 65 Solution 65. there exists an element f of L 2 that does not vanish anywhere.

then cp is essentially bounded. j. if f is the characteristic function of M. as follows. If cp is a complex-valued function such that cp · f E L 2 (for a a-finite measure) whenever f E L 2 .MULTIPLICATION OPERATORS f measure 0. Since 1 J/ZMJ. then either = 0 almost everywhere. since the function cp · f is not. -+ f almost everywhere and gn -+ g almost everywhere. an advantage other than artificial polish: unlike the more natural proof. cf. however. /1 dJ1 = Llcpl dJ1 > IIAII J1(M) = IIAII ·11!11 2 2 2 2 2 . let A be the linear transformation that multiplies each element of L 2 by cp.. The proof in the preceding paragraph has. Proof: as above. and prove that A is closed. The situation is this: suppose that H is a subspace of L 2 . Conclusion.. the latter possibility is contradictory. PROOF. as before: cp is measurable and bounded (by IIA II).. and suppose that (f. g) (i. Solution 66. gn = cp · j.e. it. cp(x)' · the function f is in L 2 . 213 .. Solution 65. If cp is not bounded. Suppose that (f. or IIA/11 2 = Jlcp.). then there exists a disjoint sequence {Mn} of measurable sets of positive finite measure such that cp(x) ~ n whenever x E Mn. For another proof. suppose that an operator A on H is such that Af = cp · f for all fin H.) Define a function f as follows: if x E Mn for some n. (There is no trouble in proving that cp is measurable. One way to proceed is to generalize the discrete (diagonal) construction (Solution 62).e. gn)-+ (f.-+ f and gn-+ g).. There is no loss of generality in assuming that j. gn) belongs to the graph of A (i. works in a certain curious but useful situation.. Indeed. then 66 f(x) = otherwise f(x) = 0. and suppose that H contains a nowhere vanishing function.

68 Solution 68.~ g). 3. = cp ·f.) Since cp · A . then the multiplication operator induced by tf.. The second proof is worth a second glance. (a) A multiplication transformation on a functional Hilbert space is necessarily bounded. it is true for a suitable subsequence. 67 Solution 67. Suppose. that <fn. The assertion about the spectrum reduces to the one about invertibility.e. measuretheoretic principle. every function that can be written down is measurable. (This sort of thing is the operator analogue of a vague. it follows that A.(x) ~ g(x) for all x. The beginner is advised to examine the reduction in complete detail. The trouble is that there is nothing in the definition of a functional Hilbert space to prevent the existence of points x in 214 . Since fn ~ f almost everywhere. indeed. and therefore lcpl ~ 1/IIA. then fn ~ f weakly). it follows that cp · fn ~ cp · f almost everywhere. and suppose that <fn. In measure theory. that <f. For invertibility: if cp · t/1 = 1. PROOF.1f= (1/cp) ·f whenever jEL 2 • Conclusion (from Solution 65): 11/cpl ~ IIA. conversely. but on first acquaintance it frequently appears to be. g) (i. let M be the set (linear manifold) of all those fin L 2 for which cp · f E V. cp · fn ~ g almost everywhere. then it must be bounded. · · ·.) Briefly: multiplications (bounded or not) are closed. The concept of essential range is no more slippery than other measure-theoretic concepts in which alterations on null sets are gratis. = Af.f. every transformation that can be written down is closed. This implies that cp can vanish on a set of measure 0 at most. that A is invertible. since. it follows that g = cp · f almost everywhere. The second proof above proves that the linear transformation (from Minto L 2 ) that maps each f in M onto cp · f is a closed transformation.1 11 almost everywhere. i. in addition.e. g") is in the graph of A.1 11. Since convergence in H implies pointwise convergence (if fn ~ f strongly. If cp is an arbitrary (not necessarily bounded) measurable function. at the same time. (Otherwise take for f the characteristic function of a set of positive finite measure on which cp vanishes. in operator theory. ~ f and g. A proof can be based on the closed graph theorem. and therefore (by Problem 65) cp is bounded. n = 1. acts as the inverse of A. Conclusion (via the closed graph theorem): A is bounded.. The closed graph theorem can then be invoked to prove that if..(x) ~ cp(x)f(x) for all x.SOLUTIONS if this is not true for the sequence lfn}. gn) ~ <f.1f = f. The answer to (b) is not quite yes. 2. g) is in the graph of A. Since g. Suppose. a multiplication has all V for its domain. but well-known and correct. it follows that fn(x) ~ f(x) and g. The concept of multiplication operator can be profitably generalized to unbo)lnded multipliers. it follows that g = Af Conclusion: A is closed and therefore bounded. and since cp(x)f.

Kx = 0. As long as infinitely many null-points are present. The relation K(x. Kv) = Kx(y).. it follows that = I(AKx. x)l 11Kxll 2 = K(x. then every (necessarily bounded) multiplication on H is induced by a bounded multiplier. Let K be the kernel function of the space (cf. y) = Ky(x) = (Kv. at the same time. x) = 0 for some x.e. Reason: ( t/J" · f)(x) is bounded by some multiple of II t/J" · f II· Since for each x there is anfsuch thatj(x) # 0. Null-points play the same role for functional Hilbert spaces as atoms of infinite measure play for L 2 spaces (cf. (AKx)(y) = (AKx." null-points" are as easy to eliminate as they are to produce. Shields. The situation can be produced at will. then <p = 0. x). KJ implies that IK(x. Solution 64). Solution 69. Ky). l<p(x)l ~ IIAII). then lt/J(x)l ~ 1 (i. and hence f(x) = (f. enlarge X arbitrarily. Since AKx = <p · Kx for each x. At the same time. (b) If H is a functional Hilbert space with no null-points. L. 1] whose derivatives belong to L 2 . and there is nothing to prove. define inner 215 69 . it is due to A. Solution 65). Note that ll<p" ·!II = IIA"fll ~ IIAII" ·11!11 whenever n is a positive integer and f is in H (cf. and since always (Kx. The assumption that there are no null-points guarantees that this does not happen. Problem 37). y)l ~ IIKxll · IIKyll = jK(x. Let H be the set of all those absolutely continuous (complexvalued) functions on [0. however. PROOF. x)jK(y. Conclusion: j<p(x)l ~ IIAII. x)j. y) = 0 for all y. If A = 0. and rewrite the preceding inequality in the form llt/1"· !II~ 11!11· From this it follows that if f(x) # 0. the answer to (b) must be no.. without changing the effect that multiplication by it has on the elements of H. x)l ~ IIAII·IK(x. and extend each function in H so as to be 0 at the new points. Kx). given Hand X. i.e. Reason: any function on X can be redefined at the null-points so as to become unbounded. y). omit them all from X and restrict each function in H to the remaining set. KJI ~ IIAII·I1Kxll 2 • Since 11Kxll 2 = (Kx. then K(x. and since. it follows that l<p(x)K(x.MULTIPLICATION OPERATORS X such that f(x) = 0 for all fin H. it follows that I<pi~ IIAII everywhere. so that l<p(x)K(x. It follows that if K(x. otherwise write t/1 = <p/IIAII. Kx) = 0 for all f. Here is an alternative proof that is more in the usual spirit of functional Hilbert spaces.

Shields. Since 1 obviously belongs to H. so that f'(x) = 0 almost everywhere. this proves that evaluations are bounded and hence that H is a functional Hilbert space. all the requirements are satisfied. then lf(x)l 2 g =I f(O) + f f'(t)dt 1.. This proves that the inner product is strictly positive.SOLUTIONS product in H by (f.£ x . then {!nCO)} is a numerical Cauchy sequence and Un'} is a Cauchy sequence in L 2 • It follows that fnCO) ~ rx and fn' ~ g. This example is due to A. L. since. 216 . and therefore f is a constant. ~ fin H. for some complex number rx and for some g in L 2 . and thus obtain an f such that f.£ 2(1/(0)1 2 2 + f1f'(t)l dt) 2 =211/11 2 . This proves that H is complete. If Un} is a Cauchy sequence in H. it follows that (fg)' ( = fg' + f'g) belongs to L 2 and hence that fg E H. put f(x) = rx + J~ g(t)dt. then f and g are bounded. Iff and g are in H. If 11/11 = 0. it follows that f = 0.£ 1. however. then J~ I f'(xW dx = 0. f(O) = 0. g) = f(O)g(O)* + f'(x)g'(x)* dx. If 0 .

If the matrix S is such that det S is an invertible element of M. then det S makes sense. IfM is a commutative subring of a ring N with unit. as an element of M: just apply the usual definition. 70 but. where Tis a matrix over N and 1 is the identity matrix of appropriate size. Sufficiency was discussed above. it turns out. suppose that ST = TS = 1. then it is profitable to consider the (commutative) ring with unit generated by those entries. It may not be obvious. The assertion of Problem 70 is often useful in operator theory. in that case. This remark settles the sufficiency part of Problem 70. if S is a finite square matrix over a commutative ring M. according to which the determinant of a matrix of size n is a sum of n! terms with appropriate signs. If an operator is represented as a matrix whose entries are commutative operators. then all is well. It is conceivable that a matrix S over the small ring has an inverse whose entries are in the large ring but not in the small one. Commutative rings have a determinant theory that is not much more frightening than in the numerical case. then a necessary and sufficient condition that a finite square matrix S over M be invertible over N is that det S be invertible in N. but all is well. it is not at all obvious that anything at all follows about the element det S of the small ring. namely the ring of all operators.CHAPTER 8 Operator Matrices Solution 70. and. Indeed. The difficulty of the necessity part is the presence in the background of a large non-commutative ring. the classical reasoning of Cramer's rule shows that S has an inverse. To prove necessity. 217 . the context it properly belongs to is a much more general part of algebra.

McLaughlin. since each entry of S is fit to play the role of a. Since is always invertible. each entry of Tis fit to play the role of a. everything else is almost automatic. A. The statement of the theorem so proved is due to J. it follows that all the entries of T commute with each other. That is: the ring generated by the entries of S and T together is commutative. Suppose now that a is an element of N that commutes with each entry of S.1 C and conclude that is invertible if and only if is invertible. The equation ST = TS = 1 (over that ring) implies that det S · det T = det T · det S = 1. as well as with the entries of S.e. it follows that each of those entries commutes with each entry of T. then RT = TR.d. set T = . the proof is due to M. with inverse it follows that (~ ~) and (~ ~) (~ ~) are invertible together. and if RS = SR. of the same size as S. This is the heart of the argument. Proof: multiply the assumption fore and aft by T. Indeed. and let R be the "scalar" matrix a · 1. The product works out to A+ BT B) ( C + DT D . therefore. Zorn. 218 . Since the assumed commutativity implies that R and S commute. 71 Solution 71. Since.SOLUTIONS An easy lemma is needed: if R is a matrix over M. the lemma implies that R and T commute. E.D. and that in turn implies that a commutes with each entry of T. q.

since PE = 1 and EP + BR = 1.1 C is invertible.1 C and proceed to consider the invertibility of B) 0 (E D. it follows that is invertible if and only if AD .BD. Suppose that is the inverse of then PE ( RE PB + QD) = (EP + BR RB + SD DR EQ DS + BS) = (1 0) 0 1 .BD.OPERATOR MATRICES Introduce the temporary abbreviation E =A . it follows thai E is invertible (and. The assumption that Dis invertible is still in force. Since D is invertible. multiplication by D does not affect any statement of invertibility.BD.l = P).1 CD is invertible. it comes in now and 219 . The proof is an easy computation. Up to this point the assumed commutativity of C and D was not needed. If E also is invertible. then so is E. Since DR = 0 and D is invertible. Now unabbreviate and conclude that is invertible if and only if A . in fact. then so is with inverse The converse is also true: if is invertible. E . it follows that R = 0.

BC = 1. The unsymmetry of the hypothesis (why C and D? and why D. or. it follows that BD . Since C and D commute. As for the counterexamples.BC is invertible. or (3) DA . ···). ez. assume that A is invertible. what the proof proves is that det( ~ ~) = det(AD - BC). or (4) A and C. e2. but even the numerical value of the determinant. Define A and D by and D<eo. What rights does (1) AD .BC be invertible. This remark (together with the result proved above) implies that in the finite-dimensional case the invertibility assumption is superfluous: if C and D commute. Conclusion: is invertible if and only if AD . then a necessary and sufficient condition that be invertible is that AD . ···) = <O. alternatively. and put B = C = 0. or (2) B and D.1 CD = BC. assume that D is invertible and make the commutativity hypothesis about (I) C and D.BC. It is well known and obvious that if the underlying Hilbert space is finitedimensional.CB do not have? Symmetry is restored not by changing the statement but by enlarging the context. but 220 . It follows that AD . e1. then the invertible operators are dense in the metric space of all operators.SOLUTIONS serves to make the statement more palatable. and make the commutativity hypothesis about (3) A and B.CB. To get a conclusion about all possible versions of the formal determinant. Actually the proof proves more than this: since multiplication by leaves unchanged not only the property ofinvertibility. an efficient place to find them is F. The theorem is one of four. The point is that the conclusion is equally unsymmetric. or (4) AD . eo.1 ?) is not so ugly as first it may seem. el.BC have that (2) DA .

1H = H. ~~' ~ 2 . regardless of the sizes of the entries. · · -). the rest of the argument is universally valid. and the proof of the lemma is complete. 0. and its inverse is the corresponding entry of the inverse matrix. B is defined by f = <1. Proof: multiply the two matrices in both possible orders and look. (Easy examples show that the assumption of finite-dimensionality is indispensable here.) Since MH c: H.) If.. . 0. This implies that M. g). it follows that dim MH = dim H. Solution 72. with inverse but the formal determinant DA has a non-trivial kernel. then. 0. The lemma applies to the case at hand. it follows from the lemma that if M is invertible.1 has the form B') (A' D'. If 72 M=(~ ~). · · -). ·) = <~ 0 . H is not really a subspace of H EB K. each diagonal entry in the matrix is invertible. then it is invariant under the inverse too. and since (by invertibility) M preserves linear independence. 221 .) To avoid the introduction of extra notation. but it becomes one by an obvious identification. 0. then M. 0. then is invertible. 0 Finite-dimensionality has served its purpose by now. where g = 0. (To be sure.OPERATOR MATRICES has a non-trivial kernel. on the other hand. then His invariant under M. let H EB K be the space. and B<~ 0 . and M the operator. Once it is known that a triangular matrix has a triangular inverse. H the subspace. and hence (by finite-dimensionality) that MH = H. (Look at <J. It is convenient to begin with a lemma of some independent interest: if a finite-dimensional subspace is invariant under an invertible operator.

The assertion is that if A* . then IT 0 (A*) = r(A)* PROOF.* is not bounded from below. together with the assumption that A* is bounded from below. If A is an operator and pis a polynomial. If the product of a finite number of operators (1) has a non-zero kernel. Corollary.* has a non-zero orthogonal complement. The second equation is the best that can be said about the relation between IT and conjugation. IT 0 (A) = r(A*)* and IT(A) u IT(A*)* =spec A. or (2) is not bounded from below. then IT 0 (p(A)) = p(IT 0 (A)). has a non-zero kernel. then at least one factor must have the same property.A. the same equations are true if A is an invertible operator and p(z) = ljz for z "I= 0. If A is an operator. this. PROOF. then A* . If A. then its kernel is trivial. and A . IT(p(A)) = p(IT(A)). then one of A* . The proof is trivial: if A is bounded from below. or (3) has a range that is not dense. It is convenient to make three elementary observations before the proof really begins.CHAPTER 9 Properties of Spectra 73 Solution 73. 74 Solution 74. Equivalently (with an obvious change of notation) it is to be proved that if both A* and A are bounded from below. and therefore the range of A* is dense. both these implications are reversible. is not invertible. implies that A* is invertible.A. then A* is invertible.A. if the factors 222 . and r(p(A)) = p(r(A)). and IT(A*) u IT(A)* = spec A*.A. E IT 0 (A *). and therefore the range of A -A. Replace A by A*. PROOF.

)--+ 0. except that (2). and therefore saves time later. 11/nll = 1. division by IIP.1/=~f.1/nll are bounded from below by 1/IIPII. The arguments for n. and apply (1). the conclusion is that a = p(A0 ) for some number )" 0 in ll 0 (A). (3) If Afn. Use the same method. a E ll 0 (p(A)). An alternative method is available for r: apply the result for ll 0 to A*.1!). and hence that p(ll 0 (A)) c: ll 0 (p(A)).1(A . and obtain A.1 of the range of A . where 11/nll = 1. conjugate. Solution 75. then p.AP ( liP 1 fnll 75 . to get the inversion spectral mapping theorem for ll.AP. then p(A) f = p(A 0 ) f The longer sentence is adaptable to the other cases.p(A 0 ) is divisible by A . but starting with Afn . or r. If A is invertible and Af = Aj with f =F 0.A is invertible.A. with no loss of generality. Conclusion: llo(A) c: llo(A 1 -1 ).1 AP(P. then the range of B cannot be. then A =F 0. liP 1 fnll 1 ( p-1/.1 to both sides of the equation.1AP(P. . Derive the result for r by applying the result for ll 0 to the adjoint. or (3). or (2) a sequence of unit vectors onto a null sequence.1 and form reciprocals to get the reverse inclusion. and hence that ll 0 (p(A)) c: p(ll 0 (A)). ) --+ O.) If.A ( =P. and apply Solution 73. ) 1 p. divide by A. This implies that p-1/. are exactly the same. on the other hand. then express p(A). it follows.A)P = p. (1) If A . then p(A 0 ) E ll 0 (p(A)).a as a product of factors such as A. then argue from the right: if B does not already do so. then so is P. Apply A.PROPERTIES OF SPECTRA commute. consequently. (2) If Af = Aj.(P. are used instead of (1). that if Ao E ll 0 (A). 0. argue from the left: if the range of A is dense. Since p(A) . This means that a E p(ll 0 (A)). Turn now to inversion.)/.A)P).1 AP..1fn). by (1 ). and. this implies that 223 .A0 .1(A .A.1(Af.A0 . Now for the proofs of the spectral mapping theorems. Replace A by A .1/n) = P. The idea of the proofs is perhaps best suggested by the following sentences.A. Assume. The norms IIP. (4) If g belongs to the range of p.1!) = A.A(P. (This part of the statement can be proved much more simply: if Af = Ao f. then p. then A must. that the polynomial p has positive degree and leading coefficient 1. If AB sends (1) a non-zero vector onto 0.Afn _.1/nll does not affect convergence to 0. (3) If the range of AB is not dense. then the converse of each of these statements is true. then g belongs to the image under p.Afn--+ 0.

. there is nothing to prove.ofll ~ IIAf.A.A.A. say C. it follows that + IIA. A convenient attack is to prove that the complement of TI(A) is open.A. then so is 1 . is bounded from below. of I . then A . reduces the theorem to the general ring-theoretic assertion: if 1 .A. # 0.AB can be written in the form I + AB + ABAB + · · ·.1(A . then 0 E TI(A). The proof itself consists of verifying that if C(1 . If A.f. 78 Solution 78.AB is invertible. Division by -A.A I ~ 0 as n ~ oo.BA is I + BA + BABA + · · · = 1 + B(l + AB + ABAB + · · ·)A = 1 + BCA. It is to be proved that if A.1 AP. If 0 E TI(A). the inverse of 1 . then AB . 77 Solution 77.1. the approximate point spectrum TI(A) is closed.AB) = (1 . are invertible or non-invertible together. Since A is not invertible.1 (in place of A and P) implies its own converse. This assertion applied to p.BA)(1 + BCA) = 1.0 is bounded from below. 76 Solution 76. PRooF.SOLUTIONS if the closure of the range of A .ofll ~ c511fll for all f. either 0 E TI(A) or 0 E r(A).A.A. For each operator A. The motivation for the proof of this assertion (but not the proof itself) comes from pretending that the inverse. + BCA)(l . The four proofs just given show that each named part of the spectrum of A is included in the corresponding part for p. similarly.BA) = (1 . The verification is straightforward.A.BA.0 I is sufficiently small. say IIAf .AB)C then (1 = 1.fll for all A. then the closure of the range of p.lA. and that. and BA ..)P is not H either. then A .A. is not H. It is convenient (but not compulsory) to prove the following slightly more general assertion: if {An} is a sequence of invertible operators and if A is a non-invertible operator such that II An .1 AP and p.A-oDII/11 ~ IIAf- A/II· This implies that if IA. It is therefore sufficient to prove that A is not 224 .. It is a little easier to see if the assumption on C is rewritten in the form CAB = ABC = C .0 is not in TI(A).ofll (c5.A.A. Since 11-Af.

= I An. it follows from the preceding paragraph that A. 225 .PROPERTIES OF SPECTRA bounded from below (i. and hence that I Afn I --> 0. The operators A .A)f.II Since llfnll = I.ll ~ IIAn.A.All however.(A - A. It follows that there exist numbers A. it follows that II(An .n). I-> 0.)ll = IA. is on the boundary of spec A. and write An -lj f. To derive the original spectral assertion.A. = IIAnfnll 2 + 11Afnll 2 ~ 11Afnll 2 . suppose that A. since II(A.e. it follows that IIAnf~. Since.A.Afnll 2 --> 0. Afn E ran A and An fn _L ran A.An are invertible and A . Suppose then that f is a non-zero vector orthogonal to ran A.n. E IT( A).. that 0 E Il(A)) under the assumption that ran A is not dense.n not in spec A such that An --> A. is not.

It follows that II(A . The first assertion is that a number is an eigenvelue of A if and only if it is equal to one of the a/s. If A is the multiplication induced by a multiplier <p (over a a-finite measure space). 80 PROOF. then both TI 0 (A) and f(A) are equal to the diagonal. then <p(x) =A whenever f(x) =1= 0.A*)fll for every A. By an obvious subtraction. "If" is trivial: each ai is an eigenvalue of A. Since a diagonal operator is normal. <p(x) = A on a set M of positive measure. then at least one of the a/s vanishes. every ~i must. then Af = Li ai~iei. 81 Solution 81. Now that TI 0 (A) is known. since no ai vanishes. PROOF. The conclusion follows from Solution 73. Normality says that IIAfll = IIA*fll for every vector f. Contrapositively: if ai =!= 0 for all j. and that the approximate point spectrum is the same as the entire spectrum. the result of Problem 79 applies. the function <p must take the value A on a set of positive measure. and TI(A) (=spec A) is the closure of the diagonal. and iff is the charac- 226 . If. then Af = 0 implies f = 0. and TI(A) (=spec A) is the essential range of <p.CHAPTER 10 Examples of Spectra 79 Solution 79. it follows that f(A) also is equal to the diagonal. so that Af = 0 is equivalent to ai~i = 0 for allj. Indeed: iff = Li ~iei. This implies that in order for A to be an eigenvalue of A. If A is a diagonal operator. and hence that TI 0 (A) = (TI 0 (A*))*. the" only if" is equivalent to this: if A has a non-zero kernel. Suppose that {ej} is an orthonormal basis such that Aei = aiei. then both TI 0 (A) and f(A) are equal to the set of those complex numbers A for which <p. Iff E L 2 and <p(x)f(x) = Af(x) almost everywhere. Solution 80.1( {A}) has positive measure. conversely.A)fll = II(A* .

then f E L 2 .. · · ·) it follows that U*e 0 = 0.j = (ei-1• e). 2. it is nonsense. ·). 82 It is wise to treat U and U* together.(f.then (0. and r(U*) = 0. If Uf = Af. This is not only false. 227 . its representation as a multiplication on H 2 ) is deceptive. and therefore U*e. PROOF. . Now for the spectrum and its parts. ~t> ~2• · · ·) = (~1• ~2• ~3• · · ·). H 2 is not invariant under multiplication by e _ 1 .. Consequence: r(U*) = 0. but there is no reason to assume that it will work for a subspace of L 2 • The correct expression for U* on H 2 is given by (U*f)(z) = z*(f(z) . ~o. then spec U = D ( = the closed unit disc). ll(U*) = D. where/= (~ 0 .EXAMPLES OF SPECTRA teristic function of a measurable subset of M of positive finite measure. A~t• A~2• · · ·). = e. 2. ~1• ~2• · ·-) = (A~o. and hence the same is true of U*.j+1 = b._ 1 (i = 1. and ~n = A~n+ 1 for all n.-t. and hence that ll 0 (U) = 0.. 1. so that I VII = 1. then b. 3. The remaining assertions are proved just as in Solution 80. it is tempting to think that if f E H 2 . ll 0 (U*) = D .j = 0. then (U*f)(z) = z*f(z). and r(U) = D . whether together with U or separately.C ( = the interior of the unit disc). Since U is an isometry. it is advisable to know what it is. f ¥= 0. it follows that the spectrum of U is included in the closed unit disc. ~ 1 • ~ 2 . For the adjoint: spec U* = D. The functional representation of U (i. To treat U*. Since (for i.e. ll(U) = C ( = the unit circle). if i > 0.C. each gives information about the other. The correspondence between adjunction and conjugation works for L 2 . so that 0 = A~ 0 . If U is the unilateral shift. e0 )). since the adjoint of a multiplication operator is multiplication by the complex conjugate function. ···). so that A is an eigenvalue of A. ll 0 (U) = 0. and Af = )/. This implies that ~n = 0 for all n (look separately at the cases A= 0 and A¥= 0). Solution 82. In terms of coordinates the result is that U*(~o.

then II Uf . Since the set of points of equality of two operators is a 228 .t3. a= y and {J = y. is bounded from below.A. that iff is an eigenvector belonging to a non-zero eigenvalue. Since spectra are closed. for all n. and.. then f = (1/A. For U* the situation is different: since TI 0 is always included in TI. and. and let g be an element of E (g ::/= 0). Conclusion: TI 0 (U*) is the open unit disc (and consequently r(U) is the open unit disc). that they be square-summable) is that IA. · · ·) = (A. it follows that TI( U*) is the closed unit disc. consequently. 83 Solution 83. If they are linearly independent."~ 0 .) The range of un consists of all vectors orthogonal to all the e/s with 0 ~ j < n.1..)An+ 1g. If U*f = A. it follows that both TI(U) and TI(U*) include the unit circle.. which has some geometric merit. If IA. If the set of eigenvectors of an operator A has a non-empty interior. then o = II U/11 = llfll. ·).)Af = (1/A. . Trivial for n = 0. because of linear independence. and hence that they are equal to it. then af + {Jg = Af + Ag = A(f +g) = yf + yg. in that disc is a simple eigenvalue of U* (i. and since TI 0 ( U*) is the open unit disc. = (1.II·IIfll for all f. this proves that TI(U) is exactly the unit circle. All that remains is to find TI(U) and TI(U*). then/= 0.IIA. . otherwise a necessary and sufficient condition that the resulting ~'s be the coordinates of a vector (i. ~2.~n' or ~n+ 1 = ). (t ran un consists of 0 alone. A.. . (normalized so that (/. the corresponding eigenvector /. I < 1. so that ~n+ 1 = A.~o' . Since the boundary of the spectrum of every operator is included in the approximate point spectrum. (Proof by induction. it has multiplicity 1). This is obvious iff and g are linearly dependent.SOLUTIONS Here is an alternative proof that U has no eigenvalues. This proves that U has no eigenvalues different from 0.fll ~ Ill Ufll .~2. then <~1. Each A. Ag = {Jg.f. then f and g belong to the same eigenvalue. .. then A is a scalar.e.. so that U -A. true for every operator A. I < I. Suppose indeed that Eisa non-empty open ball consisting of eigenvectors of an operator A.~1. e0 ) = 1) is given by f. ~3.1. and A(!(f + g)) = y(!(f + g)). · · -). The eigenvalue 0 is excluded by the isometric property of U: if Uf = 0. and if Af = af. Assertion: iff E E. If ~ 0 = 0. It is a trivial fact.flll = 11 - IA. Consequence: throughout the open set E the operator A agrees with the (scalar) operator a.e.. iff = Ang.. it follows that both spec U and spec U* include the closed unit disc. then f belongs to ran An for every positive integer n.t2.

· · ·) form an orthonormal basis for L 2 . and such that the Kx's span H. when f = 0 also. g) = (f.1 shifts backward the same way as W shifts forward. As for W*. then f = 0 (since the Kx's span H). This justifies the definition (. it is just a notational convenience. clearly w.f is linear. ± 2. and since the effect on them of shifting forward by one index is the same as the effect of multiplying by el> it follows that the bilateral shift is the same as the multiplication operator on L 2 defined by (Wf)(z) = zf(z). 84 PRooF.. the bilateral shift W has a natural functional representation on L 2 (p) (where p is normalized Lebesgue measure on the unit circle. Since both sides of the equation depend continuously on f. the unitary operator R determined by the conditions Ren = e_n (n = 0. The calculation of W. R. it follows that the set of eigenvectors (0 included) is always a closed set. ± 2.e. the set H of all functions of the form 85 229 . Suppose first that the eigenvectors of A* span H. but there is also another way to do it. Solution 85. part for part. the equation holds. · · ·) transforms W onto W* (i. so that A. With this definition of inner product. ± 1. Indeed. everything follows from Solution 81.= (Af.. Corresponding to the functional representation of U on H 2 . f)/11/11 2 • The eigenvalue equation can therefore be rewritten in the form I f 1 2 Af = (Af. Since the functions en defined by en(z) = zn (n = 0. g). and since a subspace with a non-empty interior is the whole space. and r(W) = 0. The same equations are true. since W is unitary. Kx). if (f. Let X be an index set such that corresponding to each x in X there is an eigenvector Kx of A*. Solution 84. The determination of the spectrum of W. ± 1.fbe the function on X defined by .f. then (Af. f) = A.f = 0. its adjoint is the same as its inverse. Observe that if Af = A. in particular.e. Kx) = 0 for all x.) It follows that A* Kx = q>(x)* Kx. If W is the bilateral shift. can follow the pattern indicated in the study of the unilateral shift U (Solution 82). denote the eigenvalue corresponding to Kx by q>(x)*. to obtain one from the other. TI(W) = C.f with f ¥= 0. i.(f. Conclusion: the spectrum of W* is equal to the spectrum of W. f)f. For eachfin H. see Problem 33). and. TI 0 (W) = 0. of course. its study can be reduced to that of W.for the adjoint W*. There is a thoroughgoing symmetry between Wand W*. let. and of the fine structure of that spectrum. This settles everything for W. it follows that A = a. f).n. a better way.1 takes no effort at all.f(x) = (f. and the same is true. then spec W = C ( = the unit circle). If . for each of the usual parts of the spectrum. (The conjugation has no profound significance here. The correspondence fr--+ .EXAMPLES OF SPECTRA subspace. In more pedantic language: W and W* are unitarily equivalent. just replace n by .1 W R = W*).

then J (with fin l(f. KJI ~ 11/II·IIKxll (A])(x) = (Af)-(x) = (Af. since in a functional Hilbert space the set of Kx's always spans the space. so that (Af)(x) = cp(x)f(x).<P(x)*KJ = 0 for aUf It follows that A*Kx = qJ(x)*Kx.Kx) = cp(x)(f. 230 . Kx) = cp(x)](x).Kx) (where Kx(y) = K(y.e. A*Kx) = (f. and therefore (f. cp(x)* Kx) = cp(x)(f.. if A is a multiplication (with multiplier cp. then (Af. KJ = (f.SOLUTIONS H) becomes a functional Hilbert space. [Note: l](x)l = = 11/II·IIKxii. The converse is proved by retracing the steps of the last computation. Compare the construction with what is known about the unilateral shift (Solution 82). In detail. say) on a functional Hilbert space H with domain X and kernel function K.x)).J Let A be the image of A under the isomorphism f f-+ J (i. the proof is complete. Af = (Af)-). A*Kx. so that A is a multiplication.

in terms of A. then A is invertible. 1. The conclusion is that if IA.1 = 1 +A+ A 2 + · . 52].A.CHAPTER 11 Spectral Radius Solution 86.A..A.. note that 1 1 A . express A. As for A.) . To prove that (A . )11 < 1.0 .A. If lA.o)-1(.. and (A. for A. and the series trick can be applied. 1 is sufficiently small.A) A. p.1(. n=O 00 It follows that iff and g are in H. A. and obvious algebraic manipulations prove that its sum indeed acts as the inverse of 1 . ).) Suppose now that .. and hence that p is analytic at ..A..0 is not in the spectrum of A..A). 1.1 is analytic in A...A. then (p(A.A. 0 ) .o))n..1. Cf.A. see also Problem 99.0 is invertible. A standard trick for proving operator functions analytic is the 86 identity (1.0 . then A ... [50..(A.0 ) . then II(A.A. 1. 1. 0 ) = (A . (Replace A by 1 . = (A . then the series converges (with respect to the operator norm). 1.. .A.1(.A and recapture the assertion that if Ill . 231 . 1..0 )(1 .A. . 0 : A . 1.0 )).(A .. g) = L ((A n=O 00 .)-1 =(A.1.A II < 1. 0 I is sufficiently small. near .o)n in a neighborhood of .)f.A.A.. is invertible... = oo.(1 . .o)-n-1/.= .. If IIAII < 1. 1. 1.1. 1.o)-1 0 0 0 L ((A. 1.0 . 1. 1.A. so that A . g)(A. .

Since (r(A)Y it follows that = r(A") ~ IIA"II. The principle of uniform boundedness yields the 232 . Since this is absurd (replace f and g by (A . Thus.. If then r(A) = -A(1 . g) (i. This implies that the sequence {((AA)"f. g) is bounded in a neighbor- hood of oo and therefore in the whole plane. If the spectrum of A is empty. so that r(A) ~ IIA"II 11" for all n. JAI < 1/r(A)). and hence that A .1/A is invertible whenever IAI is sufficiently small (but different from 0). the assumption of empty spectrum is untenable. 88 Solution 88. note that 1 . Analyticity doesn't have to be treated via power series. Problem 100) that p(l/A) ~ 0 as A--. the numerical function ('r:f.AA ~ 1 as A~ 0. Proceed by contradiction.J.1/. the function A~--+((A. r(A) ~ liminf IIA"II 11". the function (p Af. it follows that ((A. Since PG) =(A-~) the series trick applies again: 1 = -A(1. ·). r(A) = p(~) = -A(1 + AA + A2 A 2 + . then (p. g) is a constant.e.. The definition by differentiability can be used too. g) n=O converges whenever 1AI < 1/r(A). g)) is an entire function for each f and g.AA)-1. g)= 0 identically in f. Liouville's theorem implies that (p Af. and the factor -A in front guarantees that r(O) = 0. Since. 0. The parenthetical series converges for small A. n The reverse inequality leans on the analytic character of the resolvent (Problem 86). it follows that its Taylor series -A I 00 A"(A"f.e. g.. and A. since PA( oo) = 0.1 whenever A =1= 0 and 1/A is not in the spectrum of A. 87 Solution 87.1/.A). for instance.A)f and f). for each f and g. but it requires a slight loan from the future. to prove that p is analytic at oo.SOLUTIONS whenever A =1= 0.AA).A). since p A( oo) = 0. and hence (from the continuity of inversion. g) is analytic as long as 11/AI > r(A) (i. g)} is bounded for each such A.

(J 2 . 2. the operator D transforms A onto B. if an # 0. let it induce a diagonal operator D. if an = 0.1. then an = 0 (since. Put o0 = 1. column n (n = 0.. It follows (apply both sides to en) that anon = f3non+ 1 for each n. and determine the other o's by recursion. note that since Ion I = 1 for all n. 89 s.o I(Jo. Solution 89. the boundedness conditions are satisfied. the operator D is unitary. I < 1/r(A). in that case put on+ 1 = 1. The result is that (Jn+ l.1 (instead of S) and with the equation AS. The steps leading to this sequence are reversible. To get boundedness (away from oo ). n ~a Since this is true whenever IA. The asserted unitary equivalence can be implemented by a diagonal operator.I"·IIA"II} is bounded. · · ·) for both sides. then put On = 1. and.n+ 1 an = f3n(Jn. and evaluate the matrix entries in row n + 1. (J 1 . Suppose first that S is an invertible operator such that A = 90 f3o · . (Jn+t = f3o · · · Pnlao ···a". It follows that A*= S*B*S*.ol Consequence: {lao·· ·an//3 0 · · · /3nl} is bounded away from 0. and assume that AD = DB. This implies that the matrix <(Jii> of Sis lower triangular.SPECTRAL RADIUS conclusion that the sequence {IA.I·limsup 11A"II 11" ~ 1. Use the argument that worked for unitary equivalence to infer that S* sends kerB*" onto ker A*". it follows that limsup IIA"II 11" ~ r(A).n• and hence that Solution 90. put on+ 1 = (anlf3n)on. note that since ADen = DBen for all n. en+ 1) I ~ _£j]_. finally. To see which diagonal operator to use.1BS. let s be the (invertible) diagonal operator with diagonal sequence {(J 0 .1. put On = (f3n/an)On+ 1 . a" I = I(Jn+ 1.. (Jo. · · ·}. Consider first the positive n's. work with s. by assumption. then and therefore IA.o (Jo. Iani= 1/3nl). Given the sequence. If f3n # 0. and verify that SA= BS.n+ 1 I = I(Sen+ 1. conversely. f3n Iao · . n The proof is complete.1 B (instead of SA = BS). and hence that A*"= S*B*"S*. That is. If IA.I"·IIA"II for all n. work backwards. For negative n's (if there are any) apply the same process in the other direction. The result is a sequence {on} of complex numbers of modulus 1. If f3n = 0. then write (Jo = 1. Assume that Dis a diagonal operator with diagonal {on}. 1.1 = s. Consider the equation SA = BS. 233 . If.

then it is the sequence of power norms of some operator A. What this shows is that Ak is the product of an isometry (namely the k-th power of the associated shift) and a diagonal operator (namely the one whose n-th diagonal term is the product of k consecutive IX's starting with 1Xn). n = 1. if A is a weighted shift.. If Pk = IIAkll. 2.then Pi+k ~ PiPk. IX 1. · • ·}. SO that 0 = Aeo and 234 ..1. then Af = (0. IXo eo. be chosen as a weighted shift. 2. then ll 0 (A) = 0.1). Since an~ Pn-lPdPn. where f = <eo. If A is a weighted shift with weights 1Xn.. Given the p's define the IX'S: put IX 0 =Po and 1Xn = PniPn. PROOF.f. The expression for the spectral radius follows immediately. IX 2 . the IX'S are bounded. and since a shift is an isometry. as they must be to define a weighted shift. ···. If A is the unilateral weighted shift with weights {1X 0 . In sequential (coordinate) notation. etc. 1. If Aen = 1Xnen+ 1. 1. it follows that the norm of A is equal to the norm of the associated diagonal operator. A 3 en = 1Xn1Xn+l1Xn+zen+ 3 . in fact.I 11". ···}. The rest is a natural computation: IIAkll = sup (1Xn1Xn+ 1 . but there are cases when it can be used to compute something. 3. then IIAII = SUPni1Xnl. explicitly. · · ·. · · ·.SOLUTIONS 91 Solution 91. IX1 . The operator A can. Indeed. . or. and ll 0 (A*) is a disc with center 0 and radius liminfn lfli'~J 1X. J. The disc may be open or closed. IXz ez. and r(A) = limk supn lfl~~J 1Xn+ii 11k. Since A is the product of a shift and the diagonal operator with diagonal {1Xn}. 2.1 . · · ·). If some of them are 0. To prove the assertion about the spectral radius. 93 Solution 93.1 . and if 1Xn =I 0 for n = 0. 2. if a sequence {pk} satisfies these inequalities.then IIAkll = supn{1Xn1Xn+ 1 · · ·IXn+k. -). conversely. e2. PROOF. The proof for A is the same as for the unweighted unilateral shift (Solution 82). evaluate the powers 2 of A. IXt. IX 2 . with (positive) weight sequence {1X 0 .) The result and the proof are due to L. 92 Solution 92. and it may degenerate to the origin only. then A en = 1Xn1Xn+ 1en+z. Conclusion: the norm of Ak is the supremum of the moduli of the "sliding products" of length k. Wallen. the proof can be modified in an obvious way. IIAkll = s~p I k-1 J11Xn+i I (k = 1. k = 0. e1. IXn+k-1) n (The reasoning assumed that none of the p's is 0. · · ·). and. The expression for r looks mildly complicated. if Af = A. 3.

oc 2 *~ 3 . · · ·).: n = 0. 2. 1. Sequentially: iff= (~ 0 . = 1 (n 0. *· Since a sequence of numbers defines a vector if and only if it is squaresummable. Solution 94. and therefore the power series converges in the open unit disc. The radius of the disc can be obtained from the formula for the radius of convergence of a power series. 3. This implies that ~. ··-). in arbitrarily long blocks. The answer is yes: the approximate point spectrum of a weighted shift can fill an annulus.: n = 0. then there exists a unit vector f such that IIA/. If a. that A is a weighted shift with weight sequence {oc. If a. ~ 2 . and write f = (1/jk. = A~n + 1 for all n. Solution 91._ 1 if n > 0. If = = = a. that proves that the A's satisfying it form a disc. 1.~. and therefore the power series converges at the origin only. I A. 2. then flj. to be precise. cf.:-J oc. Suppose. · · ·}.) 1 em+ j· It follows that if a weight A. ···}that is being shifted. then flj.A/II ~ 2/fi. · · ·).= 1/(n + 1). = (n + 1) . This implies that if n > 1. = 0 for all n.:-J oc. To prove the IJ= 235 . cf. = ocn_ 1 *e. then A*f = (oc 0 *~h oc 1 *~ 2 . then flj. it is advisable to know what it is. ~ 1 . or by writing A as the produat of U and a diagonal operator and applying the known result for U*. To treat A*. = 1/n!. 2 be convergent. *~•+1 =A~" for all n. 1.=( +nhr 1 2 (=(::~r). then A must be in the approximate point spectrum.···). That can be learned by looking at matrices (the diagonal just below the main one flips over to the one just above). which in this case happens to be the same as the spectrum." 12 < OO.:-J a. Reason: consider the orthonormal basis {e. and therefore the power series converges in the closed unit disc.SPECTRAL RADIUS oc. The main idea in the proof is that if one of the weights is repeated very often. 1 (n 1. then it is very nearly an eigenvalue.It follows that A*f = Af if and only if a. The answer is that A*en = 0 if n = 0 and A*e. look separately at the cases A = 0 and A =I= 0." Ti 00 n 1 i=O OC. occurs infinitely often. if ()(m+1 94 = am+2 = ··· = am+k =A. 2. The condition is that a certain power series in A. it follows that A E II 0(A *) if and only if n~1 Ili'=J OC. 2. then ~ is the product of ~ 0 by A. by imitating the procedure used to find U* (Solution 82). Solution 82.

In many special cases that is quite easy to do.. 1Jo/JPo. Iff= (~ 0 . . consider a sequence {A.} dense in some interval.s. since the approximate point spectrum is closed and circularly symmetric. · · ·) = U(O. 1. then the supremum (over all n) of (fl~.1). '1z/~. 1's. If. · · ·) = (0. IJd)"i. There is a valuable discussion of many properties of weighted shifts in [133]. then spec A = {0} and II 0 (A) = 0. JPJP:IJz. on ZZ.. the proof is complete. all questions about weighted sequence spaces can be answered in terms of weighted shifts. 2's. Use Solution 91 to evaluate r(A).(k . !Jz/Jh. • • ·) E 1 . 96 Solution 96. ~ 0. then the shift S on l2 (p) is unitarily equivalent to the weighted shift A. occurs infinitely often in arbitrarily long blocks. and let {a. tt\ \t\'1 approximate point spectrum. JP.. = 1/2". 112..} is bounded. The spectral radius of S.. ~IJo.SOLUTIONS desired result. IJz.-~ 1j2"+i) 11k is attained when n = 0.1 (1J 0 . (Take one A.L i = 236 1 . etc. JPJi:111.} be a sequence in which each A. PROOF. If A is a unilateral weighted shift with positive weights an such that a. and three t. then two A. 2's. IJ 2 . a. then L:'=o Pnl~nl 2 = L:::'=o 1'1nl 2 . ~ 2 .\ The t:esutt\n~ we\~hted sh\ft wilt ha'le ea<. is limk supn(fl~. In view of this result. 1Jd)"i... · · ·).~1.. ~ 1 . It follows that that supremum is k-1 1)1/k = ( Il2• where 1 m = k- _!_ 2m' k-1 •=0 . · ·-> = A(!Jo. PROOF.-~ JPn+i+1/Pn+Y/k (see Solution 91). IJ 1 . 2 This implies that the supremum tends to 0 ask tends to oo. with weights {jPn+11Pn}.. three A.. · · ·). • • ·) E l 2 (p) write Uf = <JPo~o. If p = {p. '11. Computation: USU.. '11. then three A. Conclusion: the transform of the ordinary shift on a weighted sequence space is a weighted shift on the ordinary sequence space. 1's and two A.\1 t. The transformation U maps F(p) into l2 . this proves that U maps l2 (p) onto ZZ. · ·-) = US(!Jo/JPo. 2 If (IJ 0 . for instance. Assertion: U transforms S onto A. for instance. ~~2.} is a sequence of positive numbers such that {Pn+ dP. it is clearly linear and isometric. 95 Solution 95. and if ~n = IJn/JP:.

whose direct sum has spectral radius 1. 0 0 1 0 .:.. . e). then (0~. Consider the (unilateral) sequence {1. If n0 is such that ocn < e for n ~ n0 .-J ocn+i) 11 k < e just because k ~ k 0 (n.SPECTRAL RADIUS In the general case. e).. observe first that <Of. It is also easy to derive the multiplicative one from the additive one. i.1.-J ocn+... then supnCO~. To see that ll 0 (A) is empty.. A quasinilpotent operator is analytic nilpotent. each with spectrum {0}. . e). 2. and g(O) ¥= 0.-J oci-+ 0.. then max(k 0 (0. The last condition implies that g is invertible. k 0 (n. (This assertion is the multiplicative version of the one according to which convergence implies Cesaro convergence..) 11 k < e whenever k. g is analytic in a neighborhood of 0. <O~. 1. if k is greater than or equal to this maximum. 1. 0. 0. To prove "only if". that there exists a function h analytic in a neighborhood of 0 and such that g(z)h(z) 1. Indeed.. e)) is "large" enough.-J O(n+1) 11k < e because each factor in the product is less than e. There exists a countable set ofoperators.e. The O's in the sequence guarantee that A is the direct sum of the operators given by 0 0 1 0 ( 0 1 237 ( 0 0 0 0) 1 0 0 0 0 1 0 0 . Solution 98. then (1/k) D. and let A be the unilateral weighted shift with these weights. e) so that <07.-J ocn+i) 11k-+ 0 as k -+ oo for each n. if and only if it is 97 "If" is trivial. 1. if n < n 0 . 1. where f is analytic in a neighborhood of 0. it follows equally that <Of. Write f(z) = z"g(z).. then cn~. .. 1. The proofs are easy and similar. The additive version is that if ocn -+ 0.. 1. k 0 (n 0 . Solution 97.. Given e ( > 0) and given n ( = 0.···. 0. · · ·). find k 0 (n. Consequence: = 0 = f(A) = f(A)h(A) = A"g(A)h(A) = A". if n ~no. 1. If A is quasinilpotent but not nilpotent. PROOF.) Since ocn+ 1 -+ 0.-J oci) 11k-+ 0 as k-+ oo. e). apply Solution 93. 1. then A is not analytic. 1.. Conclusion: the answer to the question of Problem 97 is no.. more generally.. where n ~ 0. 98 Here is an example described in terms of weighted shifts. The problem is to prove that supn(O~. suppose that A is quasinilpotent and f(A) = 0. ·}.-J oc 1 +Y 1k-+ 0.. 1.-J ocn+i) 11 k is small when k is large. k 0 (1.-J an+Y 1k < e. 0.

What makes such examples possible is the misbehavior of the approximate point spectrum. 238 . Passage to adjoints implies that the same best possible assertion is true for the compression spectrum. For the point spectrum the best possible assertion is true (and easy to prove): the point spectrum of a direct sum is the union of the point spectra of the summands. however.SOLUTIONS and hence it is the direct sum of operators each of which has spectrum {0}. the formula for the spectral radius of a weighted shift (Solution 91) implies that r(A) = 1. Since. the sequence of weights has arbitrarily long blocks of 1's.

then Ill - AA 0 - 1 11 < 1. The set of invertible operators is open and inversion is continuous. II = 1. For an alternative example of the same thing. and let P 1 be the multiplication operator induced by cpt> 0 ~ t ~ 1.Alln 1 AA 0 1 00 1 = 1- Ill. let cp 1 be the characteristic function of [0. Solution 100. the existence of such a set is incompatible with separability.CHAPTER 12 Norm Topology Solution 99.P. and since every separable infinite-dimensional space is isomorphic to L 2 (0. If s < t. then A is invertible and A . t].1 11.A I < 1.1 = (1 . Since every infinite-dimensional Hilbert space has a separable infinite-dimensional subspace. The metric space of operators on an infinite-dimensional Hilbert space is not separable. Conclusion: there exists an uncountable set of operators such that the distance between any two of them is 1. it follows that if IIA 0 AI < l/IIA 0 . That granted.All. Since = (A 0 - A)A 0 - for each A. 1) to begin with. then P1 . is the multiplication operator induced by the characteristic function of (s. 99 PRooF. t]. 100 l::':o Recall first that if Ill . Solution 86). there is no loss of generality in assuming that the underlying Hilbert space is L 2 (0. and therefore II P1 . 1 Suppose now that A 0 is an invertible operator. 1). consider diagonal operators whose diagonals consist of O's and 1's only.At (cf. 239 . PROOF.P. it follows that IIA-111 ~ n~olll.

it follows that r(Ak 1 ) = limm k 11m = 1.1 /1. 1. 1. then Ak is invertible. (~)' 1. for some positive number e. since B is invertible. As for the puzzle: if A" . 1.1 is) and - AI/ < 1//IA 0 1 //.. Solution 84). but so long as an operator stays in a sufficiently small neighborhood of one of them. The supremum of the sliding products of length m is now equal to k. (1). The sequence of weights for Ak is { .1 1/ ~ I/A 0 /IA 0 A. k. 101 Solution 101. Consideration of the spectra of the two sides of this equation shows its impossibility. that the operator A is such that 1/X/1 < e implies that A + X is similar to A. 240 . Conclusion: the spectrum of Ak 1 is included in the closed unit disc.SOLUTIONS This implies that if /IA 0 AA 0 . it is not only invertible.. On a Hilbert space of positive dimension a conjugate class cannot even have a non-empty interior. · · ·. ' 1. 102 Solution 102. B. consequently. } Since 1/k ~ 1. · ··(but not for oo).1 )A 0 ).. 1. indeed. and hence that r(Ak) = 1. 1. and hence that s. 1. 1. in fact. If k < oo.. then multiply by A.1AS + e/2 = A for some invertible S. and this is true fork = 1. ·}. . but its inverse remains bounded. If A 0 is fixed and if A is sufficiently near to A 0 . infer that A"+ 1 --> AB. Conclusion: the spectrum of Ak is included in the closed unit disc.. . oo. and the other two factors remain bounded. 3.. the theory of weighted shifts is applicable to Ak 1 . then the middle factor on the right makes the outcome small..1 /I ~ /IA 0 - 1 /1 • 1/A.1 /I < = 1- /lAo /IA 0 - 1/1 A /1 · /IA 0 1 /1 · Conclusion: not only is the set of invertible operators open. Backwards and forwards are indistinguishable to within unitary equivalence (cf.1 en is en_ 1 or ken_ 1 according as n f= 1 or n = 1. and conclude that AB = B. 2. then A is invertible (because 1 /1 · 1/A. 1.1 is {. ' 1.. The result of the preceding paragraph makes the continuity proof accessible.1 1/ = I/((AA 0 . it follows that the supremum of the sliding products that enter the formula for the spectral radius of a weighted shift (see Solution 91) is equal to 1. and. 1. Suppose. 3. it follows that Ak 1 shifts the en's backwards (and weights them as just indicated). 2. it follows that A = B = 1.A 0 )A.1(A. and this is true for k = 1. It follows that A + e/2 is similar to A. Observe that /IA 0 1 - A.1 /1 = /IA 0 . The sequence of weights for Ak. Ak 1 itself is a weighted shift. Since Ak.A 0 /l · 1/A... and.

A. since A 00 e0 = 0. This.A 00 I -+ 0. there exists a positive number e such that cp(A. it is bounded (cf.¢ A0 . i. Conclusion: spec B is disjoint from .) Since .).) The spectrum of A 00 is. (This is an elementary fact about metric spaces.1) is included in the unit circle (perimeter).Bll < e. and given an operator A. put e = 1/cx. since it vanishes at co. then II(A. IIAk . T).A. equal to the unit disc. E . The spectrum of A 00 is clearly not the unit circle.)ll = IIA. cp(A.e. implies that the spectrum of Ak is equal to the unit circle (k = 1..) be the distance (in the metric space of operators) from A . . so that spec B c .A. E . The quickest way to prove this is to note that the span of the en's with n > 0 reduces A 00 (both it and its orthogonal complement are invariant under A 00 ). These two properties of spec B say exactly that spec B c A0 . and hence that A.A. At the same time. is not in spec B.) < a whenever A.) ~ e for all A. Problem 86).1 is the closed disc with center 0 and radius I + I A II. say.1 .A0 .111. if .Bll < e < II(A _A. It follows that if A. A. then 103 II(A . This example is due to G. i. If cp(A. T) = 0.) = II(A .A. (Note that Ak-+ Aoo.Bll < e and A. Since the spectrum of every operator includes the spectrum of each direct summand.) > 0. let cp(A. imply that the spectrum of Ak (and also the spectrum of Ak .)ll < e ~ d(A .. Suppose now that IIA . then cp(A.A..) 1ll' it follows as in Solution 100 that B . then A . in fact. there is clearly no loss of generality in assuming that e < I. it does not even depend on T being closed.Bll < 1 + IIAII. and if A. Lumer..A0 is compact.1 . as k-+ co.e. and that the restriction of A 00 to that span is the unilateral shift. If IIA . The function cp is continuous. 3. the spectrum of A 00 contains the origin.NORM TOPOLOGY The conclusions of the preceding two paragraphs.A.1 .) If A 0 is an open set that includes spec A. If.A0 . if A. A different proof can be based on the known properties of resolvents. together with the circular symmetry of the spectra of weighted shifts (see Problem 89).e.. the proof is complete.) . E T. if cp(A. the proof is complete..¢ A0 . together with the spectral mapping theorem for inverses. 2.A.I ~ IIBII ~ IIAII + IIA.A. to T. i.1 .) = 0. E spec A. This implies that B .1. d(A . This shows that the spectrum of A 00 is discontinuously different from the spectra of the other Ak's. is not in T.A. 241 1 . then cp is defined and continuous outside A0 . (This does depend on T being closed.A0 . Solution 103. fixed from now on. ·). Let T be the set of all singular operators (on a fixed Hilbert space).) - (B - A. is invertible.1 .A0 . in . then IA. E spec B.(B.

3.then ao . rxn Hence = e6 p-1 . the sequence of weights for A 2 is Two things are obvious from this construction: Ak is nilpotent of index 2k+ 1. for instance. the resolvent proof is due to E. Bo 2Bto Bo 4 2 81 82. A. ···)..1· or 2p p-1 1 6k 1 ( )lin+ 1 " og og rxo .nk ~ A. if n = 2P 2 (p 1. and so on ad infinitum.. The question of what the e's can be will be answered after the question of what they are expected to do. Nordgren. so that A. PROOF. for each non-negative integer k. All that remains is to prove that the e's can be chosen so as to make r(A) > 0. Thus.A. and the norm of Ak ..SOLUTIONS Perhaps the simplest proof is a sequential one. a 4 = e0 . The metric space proof is due to C. E spec A.e. Since Ank . The construction is based on a sequence {en} of positive numbers converging to 0.and.A. Begin by defining a sequence {an} as follows: every second a is equal to e0 (i. Bp..A (which is a weighted shift) is ek. a2 = e0 . a5 = e1... 2. is singular. a9 e1.. The sequence of a's looks like this: Let A be the weighted unilateral shift whose weights are the a's. rxn = 2P .. For this purpose note that ao ao al a2 ao a1 a2 a3 a4 as a6 = and. Suppose that An ~ A and A. There exists a convergent sequence of nilpotent operators such that the spectral radius of the limit is positive. · · ·).Ank is singular and Ank . choose A. a 0 e0 . E limsupn spec An.1 k~o 2k+ 1 • 242 . every second one of the still remaining a's is equal to e2 . every second one of the remaining a's is equal to 81 (i. 104 Solutions 104. it follows from Problem 100 that A . • · -). Wasiutynski.Ank ~ A . let Ak be the weighted unilateral shift whose weights are what the a's become when each ek is replaced by 0. a 1 = e1. in genera~ Bo...n in spec An so that for a suitable subsequence A.e.

NORM TOPOLOGY This implies that if the series ~ log ek L. and therefore liminf (ao n 0 0 0 an) 1 /n+ 1 > 0. to the set spec An does not tend to 0 as n --+ oo. Kakutani. k=O k L =o ~/2\ then log r = 00 -((log 2)/2)f'(l).1 loge lim 2v-1 I 2k+: - p-oo k=O = k=O I oo oo loge 2k+1k.z). not in liminfn spec An? Answer: exactly when the distance from A. then spec A c liminf spec An. exactly when there exists a positive number e such that 243 . n 105 (Note that A is necessarily normal. then an) 1in+ 1 > -00. The desired conclusion follows from Solution 91. Since 2P p. Here is how the computation goes.-2If f(z) = I 2k . and hence that r = 1. see [112. f'(z) = it follows thatf'(1) = 0. In some concrete cases the spectral radius can be computed. ek = 1/2\ then r = (r(A)) = 1. it is to be proved that if {An} is a sequence of normal operators and An --+ A. This example is due to S. 282]. if. The restriction of spec to the set of normal operators is continuous. To prove the statement. but that fact does not explicitly enter into the proof. for instance. p.) Question: ~when is a number A. it follows that log 2 log r = . Solutions 105. 2k+ 1 k=O is convergent (which happens if. for instance. ek liminf log(ao n 0 0 0 = 1/2k). in other words. Since -2log(2 .

A. . The way normality is used is via the observation that for normal operators the spectral radius is the same as the norm.Anii·II(Am.A.). lfe.:.). det(). This. Since n (A . is finite-dimensional and no otherwise.A.Bn). and hence that det(A. Accordingly. even the largest (in absolute value) of those eigenvalues tends to 0 as n tends to oo.) -J) ::.1 11 I ::.1 II·I1Am. then r(Bn)-+ 0 (and. then A . in fact. .Anll. so that An .A. .) = 1. ~ e (for some e). 2 e IIAm.A. .1 11 : . No use was made of normality so far.:. II(An.A. say. Everything is now prepared for the final argument.). of course.:.).A.Bn) -+ det(A. Since. the elementary symmetric functions of those eigenvalues determine the coefficients of the (monic) characteristic polynomial det(A. Since. in turn.1 ((Am .(An . on the other hand. finally. The answer is yes and no: yes if the underlying Hilbert space H .A).))(Am .).1 1 : .:. The inequality says that ---<- 1 lA. it follows that if Bn = B + Ajn. to and since An -+ A. There is no loss in simplifying the . and a subsequence {Ank} are such that II(Ank . it follows indeed that B is nilpotent. it follows that the elementary symmetric functions tend to 0.A. 1/efor all n. notation and assuming that II(An .)· lim (An . Bn-+ B).SOLUTIONS for infinitely many values of n.A. says that not only is A.:. and spec Ajn is included · in the disc with center 0 and radius IIAII/n.1 n and similarly. r((An .Bn) tends to A_dim". 106 Solution 106. B(A . the contrapositive of what is to be proved is that if a number A. by assumption.)B = lim (An .) .B).A) .A'l 1 = e whenever A' E spec An. is invertible.1 .A.A.A. Since II(An . is invertible. the proof is complete.1 11 = II(An . here it comes.A. spec(B + Ajn) =spec Ajn.). of course..A. it follows that the sequence some operator B. then spec Bn is a finite set of eigenvalues (each counted as often as its algebraic multiplicity requires). absent from spec An. {(An .A).(Am . Since.1 } converges. 244 . but...1 1 : .A.). If dim H < oo.

· · ·.) A counterexample in the infinite-dimensional case resembles the example in Solution 104 of a convergent sequence of nilpotent operators with a limit that has large spectrum. Cf. . 2. · · ·). k = 0. 1/n. . and. In the weight sequence of nB. Let A be a weighted shift with weight sequence 0. spec(A + nB) = {0} = spec A for all n. the weight 0 ( = -1 + I) must occur in the weight sequence of A + nB at least once. the weight 1/n occurs in the weight sequence of Bat position 2n-l the first time. and. -1. every second one of the remaining terms is 1. J. 0. 0.NORM TOPOLOGY So much for the finite-dimensional case. Problem 105. therefore. · · ·. 0. periodically with period 3 · 2n. Conclusion: A + nB is nilpotent. and so on (with ! . so that A is nilpotent of index 3. the weight 1 occurs at positions 2n-l + 2n · k. from then on. -1. Let B be the weighted shift in whose weight sequence every second term is 1. For each positive integer n. therefore. -1. Harrison. 245 . 0. 0. Consequence: B is not quasinilpotent. (What just happened was. Since 3 and 2n are relatively prime. These results are due to K. a proof that in the finite-dimensional case spectrum is continuous. 1. it follows that the spectral radius of B is greater than or equal to that of the Kakutani shift (and is therefore positive). and. Since each weight of B is greater than or equal to the corresponding weight of the Kakutani shift with Bk = 1/2k (Solution 104).. periodically with period 2n. from then on. in effect. every second one of the still remaining terms is t. t..

a fortiori. what follows is a proof of the second one.f I = 1 . The proof says nothing about the continuity of the norm in any other topology. and it implies that the norm is a uniformly continuous function in the norm topology. and. assume A = 0. if n is sufficiently large. other topologies take their chances. it is just a statement about uniform weak convergence of vectors on the unit sphere. It follows that if n is sufficiently large.IIBIII ~ IIA.CHAPTER 13 Operator Topologies 107 Solution 107. A norm is always continuous with respect to the topology it defines.Btl. For that purpose. 108 Solution 108. it is not continuous with 246 . this loses no generality. This is just a version of the subadditivity of the norm. for each positive number e. To show that the norm is not continuous with respect to the strong topology (not even sequentially). it applies to all nets. not only to sequences. The assumption in this case is that. then I An . then IIAnll ~e. the uniformity manifests itself in that the size of n does not depend on f. PROOF. The proofs of the two assertions are very similar. The norm is continuous with respect to the uniform topology and discontinuous with respect to the strong and weak topologies.f II < e whenever ll. The argument is general. The proof for the uniform topology is contained in the inequality IIIAII. The first assertion involving uniformity has nothing to do with operators.

indeed liPnil = 1 for all n. Cf. The proof for the uniform topology is contained in the identity IIA*. To prove the strong discontinuity of the adjoint.Bll. Solution 110.) If. (To see this. for each 8 109 > 0. and let {Pn} be the corresponding sequence of projections. !(Af. Assertion: Ak ~ 0 strongly. the adjoint changes from being continuous to being discontinuous. g)! over all unit vectors/and g is IIAII. and hence that !(Af. g) I ~ ~ I(Anf. The sequence {Pn} converges to 0 strongly. Solution 109. and therefore.B*ll = IIA. 110 Let 8 tend to 0. If a function from one space to another is continuous. let U be the unilateral shift. As the topologies march down (from norm to strong to weak). g) I + 8 for n sufficiently large. PRooF. g) I is small when n is large. g) . then there is no telling how it will behave when the topology is changed. · · · . In fact. It follows that IIAnll ·II !II ·llgll + 8 for n sufficiently large. form an orthonormal basis for M 1 \ one for M 1 n M 2 \ another for M 2 n M 3 \ etc. and write Ak = U*k. The adjoint is continuous with respect to the uniform and the weak topologies and discontinuous with respect to the strong topology. consider the following example. ~1• ~2• • • ·)11 = ll(~k> ~k+l• ~k+2• · · ·)11 2 2 247 . then it remains so if the topology of the domain is made larger. k = 1. for eachfand g. 3. I(Af.) The sequence {IIPnll} of norms does not converge to the number 0. Indeed: IIAk<~o. and it remains so if the topology of the range is made smaller. also Solution 120. and then recall that the supremum of !(Af. but the sequence {Ak *} is not strongly convergent to anything. Let {Mn} be a decreasing sequence of non-zero subspaces with intersection 0. g)!~ (li~inf IIAnll) ·11/11· llgll + 8. and back again. g) I I(Af.OPERA TOR TOPOLOGIES respect to the weak topology. every change works both ways.. If An~ A (weak). (This is the reason why the strong discontinuity of the norm implies its weak discontinuity. and string them together to make a basis for the whole space. 2.(A"f. everything can happen. then. a mapping from a space to itself is continuous. however. and the adjoint proves it.

and therefore every operator would be nilpotent of index 2. For each i ( = 1. otherwise replace them by a linearly independent (or even orthonormal) set with the same span.f.f)).A/. the tail of a convergent series. k) find a vector g. and. which is absurd.ll < e and such that the span of the g's has only 0 in common with the span of the f's. g)j 111 = j(.An*fll is nearly equal to J211.(. The proof for the uniform topology is contained in the inequalities IIAB. There is no loss of generality in assuming that the f's are linearly independent (or even orthonormal). the set of all operators A such that A 2 = 0) is strongly dense. . ~ An elegant counterexample for the strong topology depends on the assertion that the set of all nilpotent operators of index 2 (i. A belongs to the prescribed neighborhood.f. (For a different proof. Since IIU*m.AoBoll IIAB.l g. suppose that {A: IIA 0 / .SOLUTIONS so that II Ak . Let A be the operator such that (i = 1.Aoii·IIBoll.ABoll + IIAB 0 .Re(.U'ill 2 = IIUm.fll 2 2 2 = II um.f).f 11. Ag). for each .f 11 2 is.J)i.f. and h .(B*f. this is possible. k} is an arbitrary basic strong neighborhood. Indeed: IIAm+n*f.2 Re( um. see Problem 225.. ·. 248 . Bg)i = i(Ag.) To prove this. f) + III 11 = 2(11. in fact if m is large. .fll 2 .A0 Boll ~ IIAII·IIB..fll ~ 0 as m ~ oo..g.f. so long as the underlying Hilbert space is infinite-dimensional.. As for the weak continuity of the adjoint. U*m. · · ·. then IIAm+n*f. 'k). g).. (i = 1.Aoii·IIBoll ~ (IIA.Boll + IIA.fll 2 = IIUm+'i. make e as much smaller as is necessary.) If squaring were strongly continuous. then {Ak *f} is not a Cauchy sequence. at the same time. Clearly A is nilpotent of index 2. such that IIA 0 /.fll . The negative assertion about {Ak *} can be established by proving that if.Aoll + IIAoii)IIB.An*fll refuses to become small as m and n become large. that is implied by the identity I(A*.(Bg. ..f =I= 0. just as clearly.e. (The idea is due to Arnold Lebow.f. ' k) and Ah = 0 whenever h . i = 1.f.. it follows that IIAm+n*f. and. and therefore AJ ~ 0.l /.. Solution 111. .11 < e.f. .A/. then the set of nilpotent operators of index 2 would be strongly closed.Boll + IIA.

Bi-. multiplication is not jointly weakly continuous. that multiplication is not jointly strongly continuous.. and this settles strong continuity in A. For each k. ABf for each f.B)/II + II(An- IIAnBf . Weak continuity can be treated the same way. but the convergence of nets (generalized sequences) is good enough. so that IIA//11 = 1. Since. g) for each/and g. Since the strong topology is larger than the weak. f = 'Lr'= 1 (ljk)ek> then A/f = ek. Assume first that the sequence {IIAnll} of norms is bounded. it is also less computational. e 3 . in fact. in particular. that A J -> A! for each f.) It follows that (f. The easiest proof uses convergence. g) for each f and g.e. in particular.AnBfll ~ 112 113 + IIAnll · II(B". that AnBn f-> ABf. in particular.OPERA TOR TOPOLOGIES This result implies. of course. The idea is to use a badly unbounded set that contains 0 in its weak closure. Jknek)-> 0 for each f. i. Solution 112. it follows that IIAJII = J(f. jkek)ek. jkek)J. on the other hand. Conclusion: IIAk n2/11 cannot converge to 0. this settles strong continuity in B. It follows. then. A strongly. (Observe that Ak is a positive operator of rank l.ABfll ~ IIAnBnf . (It cannot be a sequence. it can be proved.g) for eachfand g. (ABif. g)-> (ABJ. Now what about the boundedness assumption? The answer is that it need not be assumed at all.e. jkek)(ek> jkek)ek = k(f. it is less geometric but.) Let {kn} be a net of positive integers such that Jknekn -> 0 weakly. ek)ek. (AiBf.g) -> (Af. As for the squares: = Ak(f. Solution 113. g) = (Bj. If. There is another way to see that squaring is not strongly continuous. If (Al. for each/. jkek)ek = (f. IIAnBnf. so that a strongly dense set is necessarily weakly dense. The convergence of sequences is sometimes misleading in general topology. define an operator Ak by AJ = (f. see Solution 28. B strongly. then apply A to conclude that ABi f -> AB f for each f. i. then.ABfll A)Bfll. (a) The crux of the matter is boundedness. Conclusion: squaring is not weakly continuous. Suppose therefore that A i-. the assumed boundedness implies. g) for eachfand g. in recompense. If. A*g) = (ABJ. so that squaring is not strongly continuous. that A iBf -. A*g)-> (BJ. the auxiliary assertion about nilpotent operators holds for the weak topology as well as for the strong. if (Bif. Jt is. If { e 1 . g)-> (Bf. if BJ -> Bf for each f. then the set E of all vectors of the form jkek will serve. and. an immediate 249 . e 2 . · · ·} is an orthonormal set. and therefore that I Ak n f I A/f -> 0. in particular. consequently. as desired.

n) are t. Indeed. The sequence {Pn} cannot. converge to anything strongly. (n. n) and (n. 3. 3. n = 1. 1). (Cf. This proves the weak sequential discontinuity of squaring at one point. 1). Since An ~ 0 strongly. · · ·). (1. A simple way of describing an example is to use matrices (as in Solution 114). A. I) are I.SOLUTIONS consequence of the principle of uniform boundedness for operators: if a sequence of operators is weakly convergent (and all the more if it is strongly convergent). Each P n is a projection (Hermitian and idempotent). Let U be the unilateral shift. then it is weakly bounded. n). Clearly Bn ~ B weakly. however. The same thing happens everywhere. 1). and therefore bounded. it follows that An ~ 0 weakly.) Since A =/:. put Bn = B + An. 2. 0. by a slight abuse of language. n). however. Problem 112 on separate continuity. namely 0. If A is the matrix whose only non-zero entry is t at the position (I. Since. AnBn = 1 for all n. (b) Multiplication is not weakly sequentially continuous. McCarthy. and whose remaining entries are all 0 (n = 2. 114 Solution 114. I) and (n. 3.) 115 Solution 115. then the limit would have to coincide with A (because strong convergence implies weak). and whose remaining entries are all 0 (n = 2. The matrix An 2 has its only two non-zero entries at (I. Let An be the infinite matrix (or. Let Pn be the infinite matrix whose entries at the positions (1. it follows that squaring is weakly sequentially discontinuous at B. · · · . 4. There exist weakly convergent sequences of projections that are not strongly convergent. (The latter part of the argument is due to C. The weak asymptotic properties of both sequences {An} and {A/} are easy to obtain: An ~ 0 weakly and A/ ~A weakly. for any operator B. where A is the matrix whose only non-zero entry is I at the position (I. 250 . PRooF. 4. and write An= U*n. and (n. then Pn ~A weakly. the operator) whose entries at the positions (I. 1). Solution 110). it is not true that AnBn ~ 0 weakly. · · -). and Bn 2 = B 2 + BAn + AnB + An 2 ~ B 2 +0+0+A weakly. and hence that Bn ~ 0 weakly (cf. Reason: if it did. and therefore A would have to be a projection (because a strong limit of projections is a projection). Bn = un.

pU*k tends strongly to 0 as k becomes large). For each n = (p. E) so that pU*k E E (which can be done because. · · -) is an operator such that u. The relation so defined makes N a directed set. and Unek = 0 when k ~ n. Suppose. e 2 . then An* --t A* (strong). Un *e 0 = en_ 1 . and suppose that Un (n = 1. that {e0 . 3. the sequence {Un *e0 } is not strongly convergent. 251 117 . then An* --t A*. The answer is no. Unen. What is easy and known (Solution 110) is that An* --t A* (weak). One good way to construct a counterexample is to use the unilateral shift U and exploit the fact that U* U f:.ek = ek+ 1 when 0 ~ k < n. E) ~ (q. E) in N. UU*.1 = e0 . however. Caution: the assertion does not mean that if {An} is a net of normal operators and An --t A (strong). the normality of A must be explicitly assumed. e 1 .I. The assertion is that if {An} is a net of normal operators and A is a normal operator such that An--t A (strong). indeed. for p fixed. Since (by the assumed normality and the assumed strong convergence) 116 IIAn*fll = IIAnf I --t IIAJII = IIA*fll for eachf. Problem 20 is applicable and yields the result. F) in case p ~ q and E :.CHAPTER 14 Strong Operator Topology Solution 116. write An = pU*k and Bn = (1/p)Uk. The details can be arranged as follows. The next step is to define a net of operators on N. Solution 117. Let N be the set of all pairs (p. Reference: [83]. ···}is an orthonormal basis. where U is the unilateral shift. It foll~ws that Un --t U (strong). find a positive integer k = k(p. Since. E) where pis a (strictly) positive integer and Eisa strong neighborhood of 0. 2. write (p. F.:.

1}. 2. in the sense of the strong topology. so that n = (p. The net {An} also converges to 0 strongly. and put n0 = <Po. 1. and r(P) = 1. let E 0 be an arbitrary strong neighborhood of 0. but in the norm. find p 0 so that l/p 0 < 8. Consequence: arbitrarily near to the operator 1. If n ~ n0 . so that n = (p. Proof: given a strong neighborhood E0 of 0. consider an orthonormal basis and write An= U*n (where U is the unilateral shift defined by Uen = en+l. Since en -+ 0 weakly and U*n -+ 0 strongly. the assumptions are satisfied. since. 2. then IIBnll < 8. spec Pn = {0. E) with p ~ 1 and E c E0 . · · ·. AnBn is identically equal to 1. 118 Solution 118. Could it be that despite all this the spectral radius is upper semicontinuous? No. E) with p ~ Po and E c E 0 . · · ·). IIBnll can be made arbitrarily small by making n sufficiently large. The product net {AnBn} does not converge to 0 strongly. that is. the conclusion is not. the spectral radius (and hence the spectrum) is not strongly lower semicontinuous. each An is nilpotent. U*nen = e0 . The powers U*n will do. Indeed: given 8 (>0). ···. there are operators with a much smaller spectral radius. 2. not only strongly. e2 . but to prove that a different example is needed. Since U*n -+ 0 strongly and r(U*n) = 1 for all n. Proof: Note that {Bn} is bounded. in fact. and write An= UPn. 1.en}. n = 0. It is easy to modify the example so as to prove that r is not continuous:just consider the corresponding unilateral shift U (Uen = en+l• n = 0. n Conclusion: the spectrum is not strongly upper semicontinuous. Clearly Pn-+ 1 strongly. For an example. In this example the spectral radius does not misbehave: r(Pn) = 1 for all n. If n ~ n0 .SOLUTIONS The net {Bn} converges to 0. · · ·). however. then An E E0 (because An E E. in the sense of the strong topology. E0 ). E 0 ). Suppose that {e 0 . No. Consequence: arbitrarily near to the operator U. by definition). e 1 . ···}is an orthonormal basis and let Pn be the projection onto the span of {e0 . The result is that An-+ U strongly. 119 Solution 119. it follows that 252 . More simply: limsup spec p n ¢ spec 1. there are operators with a (relatively) much larger spectrum. so that r(An) = 0 for all n. but r(U) = 1. and spec 1 = {1}. n = 1. write n0 = (1.

it follows that A frequently used consequence of the strong convergence theorem is about projections. these two comments together imply that there is a dense set on which P agrees with the projection onto M.An)f.STRONG OPERATOR TOPOLOGY arbitrarily near to 0. f)= ((A. by the Schwarz inequality for the inner product determined by the positive operator An .Am)f) 2 ~((An . then the sequence {Pn} cannot converge 253 .An is positive.Anll} is bounded. If {Mn} is an increasing sequence of subspaces. then the corresponding sequence {Pn} of projections is an increasing (and obviously bounded) sequence of Hermitian operators. nets would do just as well. Increasing sequences of projections serve also to show that the monotone convergence assertion is false for the uniform topology. Since {IIA. f)(( An . in the sense of the strong topology. The result just obtained can be proved without that theorem. but the proof with square roots shows better what really goes on. the sequence {Bn} tends strongly to 0.Amll ~ 1. then Pf = 0. say Bn (see Problem 121). and therefore it has a positive square root. but not necessarily with respect to the uniform topology. (An.Am~ 1. if it must be. so that IIAn . Assertion: Pis the projection onto the span. that A ~ 1. Bnf) = (B/J. Solution 120. Let {An} be a bounded increasing sequence of Hermitian operators. Since 11Bnfll 2 120 = (Bnf.Am. Solution 1) sequences play no essential role here. and let A be its weak limit. so is {IIBnll}. It follows that there exists a H~rmitian operator P such that Pn ~ P strongly. of all the Mn's (cf. the operator A . say M. there are operators with much larger spectral radius. Since An ~ A. Here is how a proof without square roots goes. (An -Am) f). Since An. and iff is orthogonal to all Mn's. As once before (cf. Reason: if f belongs to some Mn.f) ~ 0.Am)f. Indeed.Am) 2J. then II(An. One way to prove the assertion about the strong topology is to make use of the weak version. The asserted strong convergence now follows from the relation II(A .Am)fll 4 =((An.Am)J. PROOF. If m < n. Assume. if the sequence {Mn} is strictly increasing. then Pf = f. A bounded increasing sequence of Hermitian operators is always convergent with respect to the strong topology.An)fll = IIB/fll ~ f311Bnfll. with no loss of generality. Solution 108). say IIBnll ~ f3 for all n. There is sometimes a technical advantage in not using the theorem about the existence of positive square roots.

SOLUTIONS toP (or. The following quite pretty and geometrical consideration shows how topology comes in. it remains only to evaluate the limit.A with positive coefficients (by induction). f) = (A· Anf. It is convenient (for purposes of reference) to break up the proof into small steps. (Note: the assertion is a trivial consequence of the spectral theorem. and therefore it is strongly convergent to some (necessarily positive) operator B. for that matter. and hence (by (1)) each Bn is a positive operator.) (2) Each Bn is a polynomial in 1 . 254 . cf.. In fact. the sequence {Bn} is bounded. project it on M (i. and it follows that Bn+z. liPn . since Bn ~ B (strongly). it follows that B/ ~ B 2 (strongly). [114]. and the proof is complete. In the sequel the result is needed not for A but for 1 . f)= IIA'ill 2 and (A 2 n+ 1/. Convergence is proved.A.A with positive coefficients. {Bn} is a bounded increasing sequence of positive operators.P m I = 1 unless Pn = Pm. 121 Solution 121. project the result on N (FEJ).B) 2 . (1) All the positive integral powers of a positive operator are positive. and suppose that M and N are two distinct but not orthogonal lines through the origin. all the Bn's commute with one another. and hence positive. Since the point of what is now going on is to avoid square roots. and hence that This says that A = 1 . as follows. 122 Solution 122. because it is not even a Cauchy sequence.Bn)(Bn+t + Bn)· This implies (by (2) and induction) that Bn+ 1 . (4) The definition of Bn+ 1 in terms of Bn implies (induction) that IIBnll ~ 1 for all n.e. the former is positive because norms are. Take an arbitrary point fin H. form Ef). This is easy from Problem 113. Even a small amount of experience with non-commutative projections shows that the familiar algebraic operations are not likely to suffice to express E A F in terms of E and F. and the latter is positive because A is.Bn is a polynomial in 1 .2B + B2 = (1 . Suppose that the underlying Hilbert space H is twodimensional real Euclidean space. A'i). a monotone sequence of projections can be a Cauchy sequence in trivial cases only. it is necessary to use the version of Solution 120 that does not use square roots. to anything at all) in the norm. (3) By (2). The proof is standard.Bn+1 = !(Bn+1 2 - B/) = t(Bn+1. Indeed (A 2 Y. Note that the argument needs Solution 120. it follows that the sequence {Bn} is increasing. and motivates the actual proof. (5) By (3) and (4).

so that G0 ~ G. since in this case weak and strong convergence are equivalent. The assertion is that the projection E v F on the subspace M v N belongs to any von Neumann algebra containing E and F. is (E A F)f This suggests the formation of the sequence E. Clearly G is Hermitian.B)(1 . therefore GFG = G. FEFE. FE. 255 . G belongs to the given von Neumann algebra. EFEFE.F)G. EFE. .F)(1 .(1 . Since E v F = 1 .A) (for the theorem about E v F). an expression such as 1 .F)). If.A)(l ..F))*.(1 .GFG = G(l . The proof itself works with the subsequence EFE.) Since E(EFEt = (EFEt for all n. the proof is immediate. in this case. Since IIEFEII ~ 1. and (1 . whence G0 G = G0 . G. it looks plausible that the sequence so obtained converges to 0. · .. finally. . It follows that (EFEt is weakly convergent to. The proof is complete.B)(1 . Observe that even in the latter case it is not required that 1 belong to the set. so that G is a projection. the powers of EFE form a decreasing (and even commutative) sequence of positive operators.. Since (EFE)mFG = G for all m.A) is a convenient way of writing something that can obviously (though clumsily) be written without 1 if so desired. EFEFE.F)G = G(1 . this is a matter of merely technical convenience. Since (EFE)mG = G for all m. this implies that G ~F. then so is ABA (for the theorem about E A F) or 1 . which. all that was needed was a sequentially strongly closed set of operators such that if A and B are in the set.. The theorem has its own dual for an easy corollary.STRONG OPERATOR TOPOLOGY then project on M (EFEJ). then G0 (EFEt = G0 . (Proof: 0 = G. and continue so on ad infinitum.A)(1 . say. An examination of the proof shows that not all the defining properties of von Neumann algebras were used. therefore G2 = G. G0 is a projection such that G0 ~ E and G0 ~ F. Assertion: G = E A F. EFEFEFE.F)G = (G(l .((1 - E) A (1 . therefore EG = G or G ~ E.

If A E spec A and if N is an arbitrary neighborhood of F(A). A has a neighborhood disjoint from it. q. This completes the proof. it is empty. To prove the reverse inclusion is the same as to prove that if A¢ F(spec A). then F. That's the easier direction.1 (F. Solution 124. (It is the image under the continuous function F of the compact set spec A. f) ~ (f.1 (N)) = (F o q. and therefore not only does (Foq. The set F(spec A) is compact. 256 .1(N))) have measure zero.. then F.1 (N).1(N)) has positive measure. Since q.1(F. If 0 ~ A ~ 1.. but. then (Af.CHAPTER 15 Partial lsometries 123 Solution 123. Use the spectral theorem to represent A as a multiplication by. so that F(A) E spec F(A). say.1 (F. in fact. then A¢ spec F(A).. it follows that F(A) is in the essential range ofF o q... in particular.1(N) (=q. Consequence: A does not belQng to the spectrum of F(A). see Problem 121. This proves that F(spec A) c: spec F(A). 124 or this means that A 2 ~ A.1 (N) = 0.1 (N) is a neighborhood of A. The necessity for square roots is not surprising in a theorem that is explicitly about squares. and therefore q. and.).). If N is such a neighborhood. f) for allf. Perhaps one reason why the non-spectral proof is elusive here is that square roots seem to be needed. and square roots without the spectral theorem are known to require a little effort.) Since A is not in it.

eAo E. 0. The answer is the same even if the underlying Hilbert space is not assumed to be separable.. That is: E. 2 .. then 2 (Af.. is an orthonormal basis for L (X). Solution 126.. There is an intelligent way of choosing a basis E.1 {z:p(z). and the union of all the bases so obtained is a basis for the whole space.. and. span the whole space.· By assumption p(A) has a set A0 of eigenvalues A such that the corresponding eigenspaces M.) Question: which functions in L2 (X) belong to M. =f. note that if A is Hermitian. for L2 (E. Use the spectral theorem to represent A as multiplication by <p on L 2 (X) for some measure space X with a-finite measure fl.e.l. The sets {E. Choose E. that fl.. and hence that IIAII ~ 11J:411. indeed.. The fact is that if p(A) is diagonal for any non-zero polynomial p. but the proof involves some fuss necessitated by the possible pathology of large (meaning. 0 for only countably many values of A. it follows that A ~ 1. 125 that is. 3.).f). but the formulation of the problem is misleading..PARTIAL ISOMETRIES In the other direction...A= 0}...l M. Consequence: IIAII 2 ~ IIAII. then A also must be diagonal. The separability of the space implies that M. then IIA/11 2 = (Af. 2.A)f(x) = 0 a.? Answer: those functions f for which (p(<p(x)) .?. correspondingly. then M.)· Since the set {z: p(z). those whose support is a subset of E. . A2 . is finite. Suppose. 0 for only countably many values of A. Let c be a positive number such 126 257 .(E . is the union of a finite number of sets on each of which <p is a constant. that A and An are normal operators (n = 1. one for each of those subsets of E. in particular. A E A 0 . Solution 125. then 2 U. The answer is yes... or I All ~ 1. IfE. as the union of a corresponding finite number of orthonormal bases.....A = 0} is finite.. and hence. The answer is yes. and their union is (almost) equal to X. then llflfll 2 = It follows that if A ~ A. Each vector in each such basis is an eigenvector of A. · · -) with An --+ A. it follows that the essential range of the restriction of <p to E. is an orthonormal basis for L 2 (E.. if A .) (=the subspace of those functions in V(X) whose support is a subset of E. IIAfll ~ llflfll ~ llflll·llfll for all f..: A E A0 } are pairwise (almost) disjoint. non-a-finite) measure spaces.) =f.f). = ({f: p(A)f = Af}). Af) = (A 2f. since A is Hermitian. (Normality implies that if A1 =f. = <p.

.F(A)II for all m and n.. A typical and important example is the square root function on the positive half of the real line: the conclusion is that the mapping A 1-+ fl.f). A*)ll + IIPm(A. say. for m fixed. see Problem 51.. An *)II + IIPm(An. X*) . choose n so as to make the middle summand small. the norm IIPm(X.f) = 11Ufll 2 = 11!11 2 = (Ef. Now note that IIF(An). the spectral theorem implies that a Hermitian operator A with A 3 = A 2 is idempotent. (For a discussion of such transformations and their adjoints. A*). then the product of U and the 258 . then (U*Uf. Choose m so large that the first and last summands are small for all n (this is the uniformity mentioned above).Pm(An. The method of proof can be generalized so as to yield similar results for functions that are not defined on the entire plane. It follows that II Ufll 2 = (U*Uf. defined for positive operators.Pm(A. z*) . iff j_ M. Suppose that H and K are Hilbert spaces and suppose that U is a partial isometry from H into K with initial space M. UU* = kerJ.. that U is a bounded linear transformation from H into K such that U*U is a projection with domain Hand range M. and hence that II Ufll = 11!11 or Uf = 0 according as f EM or f l_M. Suppose. The assertion about initial and final spaces follows from the observation that kerJ.F(X)II is dominated by the supremum norm of Pm(z. If U*U is idempotent. is continuous...? c.) If E is the projection from H onto M.J) = (Ef.? IIF(An)..F(A)II .f). then (UU*) 3 = U(U*UU*U)U* = (UU*) 2 ..f) = 0 = (Ef.. As for Corollary 3: if U is a partial isometry. 127 Solution 127. and this implies that U*U =E.F(z) on the disc with center 0 and radius c.. F(z) uniformly in the disc with center 0 and radius c. U* = ran U (since ran U is closed). A/).? c). To prove Corollary 1. The proof of Corollary 2 is a trick... and iff EM. observe that ker U*U = ker U (this is true for every bounded linear transformation U). It follows that (U* Uf. z*) .SOLUTIONS that IIAnll . and then. X*)---> F(X) uniformly over all normal operators X with IIXII .. then (U*Uf.. for every normal operator X.f) = (Ef. f) for all f in H. By the Weierstrass polynomial approximation theorem in the plane there exists a sequence {pm} of polynomials in two variables such that Pm(z. conversely. Since. f)= 11Efll 2 for all f. it follows that Pm(X.? c for all n (and hence such that IIAII .

In that case UM = M. If U is an isometry and if UM = M. The first implication is false for co-isometries. The first implication is false if U is the adjoint of the unilateral shift and M is the (one-dimensional) subspace of eigenvectors belonging to a nonzero eigenvalue (see Solution 82). The second implication is false if U is the unilateral shift and M is the whole space.VII < 1. This comment is pertinent to the question concerning the connectedness of the set of all non-zero partial isometries. then f = 0. Solution 128. More precisely. if U is a co-isometry and if M reduces U. then UM = M. 1.Vll·llfll. which contradicts the assumption that II u . and if II U . then premultiply by U* and conclude that U*U is idempotent. then M reduces U. The fact is that if a partial isometry is sufficiently near to an isometry. since llfll = IIVfll = IIUf. Solution 130. then the set of all isometries is not connected. PROOF. One way to prove that the answer to that question is no is to prove that the set of all isometries is not only closed but also open in the set of all partial isometries (in the relative topology of the latter). It follows that the restriction of U to the initial space of V is one-to-one. Solution 129.. Reason: the set of all unitary operators is a non-empty proper ( !) subset that is simultaneously open and closed. Indeed. U = UU*U. then U*M = U*UM = M. The assertion about closure is obvious. The kernel of U and the initial space of V can have only 0 in common. In that case M reduces U. if. then II U . if Vis an isometry. then U is an isometry. conversely.VII < 1. then IIUf. and (2) the equation A = AA *A 129 characterizes partial isometries. but M does not reduce U. it follows that iff =I= 0. If U*U = 1 and UM = M. the second implication is false for isometries. and hence (Problem 56) the dimension of the initial space of V is less than or equal to the dimension of the entire range of U.::. An even easier version of the same proof shows that the set of all isometries is closed.VII . In other 130 259 .VII < 1. If UU* = 1 and both UM c M and U*M c M.Vfll = llfll. the reason is that (1) the mapping A H AA *A is continuous.Vfll ~ IIU. then apply U to the second 128 inclusion to obtain the reverse of the first. The same argument shows that if the underlying Hilbert space is infinitedimensional. but UM =I= M. Indeed. consider the mapping ( 1) A H A* A and the equation (2) A* A = 1. and this contradicts the hypothesis II U .PARTIAL ISOMETRIES projection U* U agrees with U on both ker U and its orthogonal complement. It is sufficient to prove that if Uf = 0. if U is a partial isometry. then it is an isometry. iff is a non -zero vector such that Uf = 0 and II Vf II = llfll.

Since it is easy to verify that this definition yields a linear transformation W that maps R 1 isometrically onto R 2 . Indeed.VII. then there exists at least one unit vector fin the kernel of V that is orthogonal to the kernel of U. 131 Solution 131. This proof is a simplification of the one in [65]. co-rank. Suppose that A and Bare unitarily equivalent. Suppose that V1 and V2 are partial isometries with the same rank. 260 . and t ~ Ut are continuous curves of unitary operators that join 1 to W and to U. and nullity. for definiteness. let N 1 and N 2 be their kernels.SOLUTIONS words. and the proof of the assertion about nullities is complete. it is due to R. then U transforms A* onto B*. but it exhibits a partial isometry that has the first for initial space and the second for final space. The assertion about co-ranks is an easy corollary: if I U .Vfll ~ I U . define Wf = V2 UV1*f. M 1 and M 2 their initial spaces. The result appears in [114. however. then I U . It follows that 1 = II !II = I Ufll = I Uf. for fin R 1. If g E N 1 . it follows that J J transforms M(A) onto M(B). Problem 57). onto R 2 J.AA * onto B' = 1 . as indicated. To say that f is orthogonal to the kernel of U is the same as to say that f belongs to the initial space of U. Let W be a linear transformation that maps R/ isometrically onto R/. if v(U) =1= v(V). G. Let U be an arbitrary unitary operator that maps N 1 onto N 2 and M 1 onto M 2 .V* I < 1. The generalization appears in [65]. more constructive. co-rank. and nullity. then II U* . that joins vl to v2. 132 Solution 132. say. The assertion about nullities can be phrased this way: if v(U) =1= v(V). and therefore p'(U) = v(U*) = v(V*) = p'(V). § 105] for the special case of projections (which is. the assertion about ranks follows by symmetry. and R 1 and R 2 their ranges. at the same time.BB*. it not only proves that two subspaces have the same dimension. in fact. If U is a unitary operator that transforms A onto B. then t ~ H-. It follows that WV1 = V2 U. it follows that there exists a unitary operator W that maps R 1 onto R 2 and R 1 J. The present statement is a generalization. v( U) < v(V). or WV1U* = V2. then WV1g = V2 UV1*V1g = V2 Ug. Douglas.VII < 1. V1 U/ is a continuous curve of partial isometries all with the same rank. and. the result is that p(V) ~ p(U).VII ~ 1. and therefore U transforms A' = 1 . If t ~ ft. the proof is a considerable simplification. The proof in [114] is.

then it transforms M(A)* onto M(B)*. # 0 and A . In the finite-dimensional case more can be said. and if each element of A is assigned a positive integral multiplicity. is invertible. 261 . a unitary operator matrix of size 2 transforms M(A) onto M(B). where A'= j l . which happens when A.) is invertible. then there exists a partial isometry with spectrum A.A. A unitary operator matrix that com- ~ ~)is necessarily of the form ( ~ ~). Solution 133. conversely.AA*. as in Problem 132. Let A be a normal contraction with spectrum A (see Problem 63). Conclusion: spec M = {0} u A = A. If A is a finite subset of the closed unit disc. is the operator matrix A') not invertible? By Problem 70.A. M .A. 133 The assumed relation between M(A) and M(B) implies that U* AU= B. then there exists a partial isometry with spectrum A whose eigenvalues have exactly the prescribed algebraic multiplicities. with 0 in A. see [ 65]. M= ( 0 A A') ' 0 -A. PROOF. If a compact subset A of the closed unit disc contains the origin. and therefore it transforms M(A)M(A)* ( mutes with ( = (~ ~))onto itself. what is its spectrum? The question reduces to this: for which values of A.PARTIAL ISOMETRIES If. If. then M is a partial isometry.(A . is invertible if and only if -A.with U and Vunitary.

multiply A = UP on the left by U*. a unique extension to a partial isometry from H to K with initial space R. Begin with the construction of P. 262 . cf. where E is the projection from H onto the initial space of U. This implies that ker U = ker P and completes the existence proof. it follows that U also is uniquely determined by the stated conditions. and. Since JJPfll 2 = (Pf. To prove uniqueness. and hence that A* A = P 2 • Since the equation U Pf = Af uniquely determines U for fin ran P. Pf) = (P 2f.L A. suppose that A = UP. it follows that the equation UPf = Af unambiguously defines a linear transformation U from the range R of P into the space K. and use the equation U*U = E. Since that initial space is equal to ker. it has a (unique) positive square root. call it P. Solution 127. f)= (A*Af.CHAPTER 16 Polar Decomposition 134 Solution 134.f) = l1Afll 2 for allfin H. To deduce Corollary 1. Pis positive. The equation A = UP holds by construction. it follows that EP = P. where U is a partial isometry. observe that ker U = ker P = ker A* A = ker A. and that U is isometric on R. and ker U = ker P. from there. It follows that A* = PU* and hence that A*A = PU*UP =PEP. and the orthogonal complement of the range of a Hermitian operator is its kernel. For Corollary 2.L U. Since A* A is a positive operator on H. and since Uf = 0 whenfis in ker P.L U = ran. The kernel of a partial isometry is the orthogonal complement of its initial space. and hence to ran P. and ker U* = ran. Since U is bounded on R. it has a unique bounded extension to the closure R.

and let R ( cK) be the range of U (or. the adjoint of each such isometry is a co-isometry from H into K that agrees with U on M. then there exist isometries from H into K that agree with U on M (many of them). the proof is complete. all that is needed is to map MJ.P2. equivalently. and to combine such a mapping with what U does on M. it follows that A = t{VW + VW*). it follows that V P = UP = A. Iff is a unit vector. since A = VP and P = 1(W + W*). then so is Uf. a+ {3 = 1. (The justification for the upper bound on Pis that IIAII ~ 1. where IIAfll ~ 1 and IIBfll ~ 1.i. IIAII ~ 1. If. Conclusion: isometries are extreme points. Since the range of P is included in M. with a> 0. Suppose first that U is an isometry and that U = o:A + [3B. isometrically into RJ. a consequence of it) is useful. The result for co-isometries is an immediate consequence. in case the underlying Hilbert space is one-dimensional. and Uf = rxAf + {3Bf. (The assertion is true and the proof below is valid whenever -1 ~ P ~ 1. and verify that everything works. it is possible to write A = VP.) Assertion: there exists a unitary operator W such that P = 1(W + W*). Solution 136. The converse can be proved by showing that every operator A. dim RJ. ~dim M\ then there exist isometries from K into H that agree with U* on R. 135 If dim MJ. Kadison [82] has proved that. Here the theory of polar decompositions (or. the closure of the range of A). is equal to a convex combination (in fact. on the other hand. just write w = p + ij1 . the extreme points in the unit ball of the algebra are those partial isometries U that satisfy the identity (1 . then both the assertion and its proof make simple geometric sense. it follows that Af = Bf = Uf. Now. Since the product of a maximal partial isometry and a unitary operator is a maximal partial isometry. with IIAII ~ 1. and IIBII ~ 1. PROOF. Suppose that A is a bounded linear transformation from a Hilbert space H to a Hilbert space K. {3 > 0. ~dim R. Since the closed unit ball of a Hilbert space is strictly convex (Problem 4). to the average) of two extreme points of the kind already found. let M ( c H) be the initial space of the partial isometry U.UU*) = 0 263 . for certain operator algebras.) To prove the assertion. In either case there exists a linear transformation V from H into K such that either V or V* is an isometry and such that V agrees with U on M. where V is a maximal partial isometry and 0 ~ P ~ 1.U*U)A(1 .POLAR DECOMPOSITION Solution 135. let A = UP be the polar decomposition of A.. The extreme points of the unit ball in the space ofoperators 136 on a Hilbert space are the maximal partial isometries. rather. and hence that A = B = U. By Problem 135.

If A = if each of X and four. f)· ( Y g. Since ker P = ker U. If U commutes with P. or.then (AJ. g) = 1. which shows that the positive square root of a positive operator is the weak limit of a sequence of polynomials in that operator. it is to be proved that 1 . That is easy: given g. inductively.PU = 0. apply the Weierstrass theorem on the approximation of continuous functions by polynomials to prove that "weak" can be replaced by "uniform". it is trivial that UP.f = (~). In other words. = UP 2 U*UP 2 U* = UP 4 U*. The converse is harder.SOLUTIONS for all A in the algebra. but Y is one of JA*A and JAA*. then the Kadison condition is satisfied.U* U = 0. that the condition is satisfied. since P also commutes with P 2. Q2 it follows that = UP 2 U*. Alternatively. 264 . ~). then. It is. clear that if either U or U* is an isometry. (Compare Problem 121. 138 Solution 138.UU* =F 0. g) = 0. then A commutes with P 2 (=A* A). For the algebra of all operators on a Hilbert space this is consistent with what was just proved. then U commutes with P 2 . If A is quasinormal. it is to be proved that if (1 . Q = UPU*. three times out of e JAA*. JA*A. with P = and put Q = Since AA* = (UP)(PU*) = U(A*A)U*. then (1 . let A = UP be the polar decomposition.PU)P = 0. (Xf. and assume that 1 . To prove it. Reason: note first that U*UP 2 = U*AP = P 2 (by Corollary 1 of Problem 134).and g = (~). and therefore Q4 and.) This says that (UP. indeed. Write UP= A.PU annihilates ker P also.UU*)f =g. so that UP. It follows from the most elementary aspects of the functional calculus that A commutes with P.PU annihilates ran P. in other words. The fourth possibility is a "mixed" Schwarz inequality that is always true. conversely.UU*)f =F 0 for some J. it follows that A (=UP) commutes with A* A ( =P 2 ).U*U)g = 0 for each g. Suppose. and it follows that UP. find an operator A such that A(l . 137 Solution 137.

it follows by induction that an+k = an for all positive k. Now. This implies that either there is a first n for which an -=1= 0.) It is to be proved that if IIA . Suppose that {en} is an orthonormal basis and Aen = anen+ 1 for all n. g)l 2 = I(Pf. U*g) = (Pf. in particular. the invertibility of both BT and T implies that of B. this implies that Tis left-invertible. then B is invertible and C = A = s. If an -=!= 0 for some n. Solution 140.Til < 1. If Tis rightinvertible also. find an invertible operator Q (which can be made positive. and then infer that F(Q 2 ) = UF(P 2 )U* for every polynomial F. and in that case the a's are constantly equal to that an from then on. an is never 0. then an+/ = a/. every operator has the form VP.) Assertion: there is a neighborhood of A that contains no right-invertible operators. with liB II ~ 1. for the proof. This elegant application of polar decomposition was shown me by John Duncan. observe first that B is right-invertible (with right inverse A). and hence BT is invertible. or else an = 0 for all n. gW = I(UPf.Til < 1. g) = (Pf.POLAR DECOMPOSITION for all n = 0. which is false. The proof of the density theorem for unilaterally invertible operators is completed by observing that if V is a maximal partial isometry. Solution 139. f)· (UPU*g.T)ll ~ IIA. then A is just a scalar multiple of the unweighted shift. U*g)l 2 ~ (Pf. (If both BA = 1 and CB = 1. then it is invertible. Given a positive number B.Qll <B. By Problem 135. 2. Assume (with no loss of generality) that A has a left inverse B. Assume with no loss that an~ 0 (see Problem 89). but not left-invertible.VQII <B. Is the question about unilateral or bilateral shifts? It turns out that it makes no difference. it follows that (A*A)Aen = anan+ 1 2 en+ 1 and A(A*A)en = a/en+ 1 • Consequence: A is quasinormal if and only if an an+ 1 2 = an 3 for all n. f)· (PU*g. In the presence of this normalization. (Example: the unilateral shift. then V is unilaterally invertible (left-invertible if Vis an isometry and right-invertible if V* is one). If. 265 139 140 . · · ·. Since A*Aen =a/en. consider an operator A that is leftinvertible but not right-invertible. the assertion can be made more precise: the open ball with center A and radius 1 contains no rightinvertible operators.BTII = IIB(A. and 0 before. if so desired) such that liP. Indeed: Ill.1 . where V is a maximal partial isometry and P is positive. (Compare Solution 137.) The rest is straightforward: I(Af. g). To obtain the negative conclusion. and that the product of a unilaterally invertible operator and an invertible operator is unilaterally invertible.f) · (Qg. and therefore an+ 1 = an. It follows that II VP . then T is not right-invertible. 1.

Join U to I by a continuous curve t ~---> U. of unitary operators (cf. so also are U and P. One way to approach the proof is to show that for each invertible operator A there is a continuous curve that connects it to the identity. consider tP + (I . (The latter does not even need the spectral theorem. consider the polar decomposition UP of A. similarly.SOLUTIONS 141 Solution 141.) If A. For this purpose. = U. is a continuous curve of invertible operators that joins A (=A 1) to 1 ( =A 0 ). It follows that U is unitary and P is strictly positive. join P to I by a continuous curve t ~---> P. .P. Problem 131).t). of strictly positive operators. 0 ~ t ~ 1. then t ~--->A. and. Since A is invertible.

For each Borel subset M of the complex plane.E(M) has finite rank..L are infinite-dimensional. the operator E(M) is the projection onto the subspace of functions that vanish outside cp 1 (M).1(M). let E(M) be the multiplication operator induced by the characteristic function of cp.CHAPTER 17 Unilateral Shift Solution 142. it is sufficient to prove that for each normal operator on a separable infinite-dimensional Hilbert space there exists a reducing subspace such that both it and its orthogonal complement are infinite-dimensional. Indeed.E(M) have infinite-dimensional ranges. If the intersection . i. If H is not separable. Proceed inductively to obtain an infinite sequence {Hn} of pairwise orthogonal infinite-dimensional reducing subspaces. It remains to prove the assertion italicized above. If.. and.L such that both Hz and H 1 . In other words. for some M. then there exists a reducing subspace H 1 of H such that both H 1 and H 1 . adjoin it to. Another application of the same result (consider the restriction of A to H/) implies that there exists a reducing subspace Hz of H 1 . In the contrary case what must happen is that for each M either E(M) or 1 . the assertion."= 1 H/ is not zero.e.L are infinite-dimensional. Draw a sequence of finer and finer square grids on n 267 . say.L 1"'1 H 2 . The spectral theorem shows that there is no loss of generality in restricting attention to a multiplication operator A induced by a bounded measurable function cp on some measure space. there is no loss of generality in assuming that H is separable in the first place. the range of each E(M) reduces A. is just that H is the direct sum of ~ 0 infinite-dimensional subspaces that reduce A. therefore. consequently. In a 142 separable Hilbert space all infinite-dimensional subspaces have the same dimension. the desired assertion is true. then it is the direct sum of separable infinite-dimensional subspaces that reduce A. if this is true. Clearly each E(M) commutes with A. H 1. It is sufficient to prove the assertion for 2 in place of ~ 0 . both E(M) and 1 .

it follows that no unitary operator with a non-real determinant can be the product of symmetries. for then they could be separated into two infinite subsets. The operator U belongs to the center of the group of all unitary operators. the point is that in no group can a central element of order 3 be the product of three Ln 268 . ± 1.2 "H" = H_"' so that PQf = PS. If S is an arbitrary right shift. it follows that every unitary operator is the product of two shifts. then H 0 and H 1 do everything that is required.2" on H" (n = 0. let P be the operator that is equal to S 1 .2"/ = S 1 2 <-nlS.2 " on H" and let Q be the operator that is equal to s. To prove that on every Hilbert space there exists a unitary operator that is not the product of three symmetries. Suppose then that S is a right shift. write T = S*U. let M be that range. The equi-dimensionality of all the H. ±2. Suppose that His an infinite-dimensional Hilbert space. There cannot be more than finitely many A's like that.2"/ =Sf The existence proof is complete. Conclusion: there exists at least one point A such that the dimension of the range of E({ll}) is infinite. and let each square in each grid play the role of M. It is convenient to let the index n run through all integers. and zero. the proof will be completed by showing that every shift is the product of two symmetries. · · ·. if H 0 = M 0 and H 1 = M 1 v M. If the underlying Hilbert space H is finite-dimensional. it follows that if E(M) has positive rank. negative. Since TH" = S*UH" = S*H" = Hn_ 1 for all n. and define a left shift as a unitary operator T such that THn = Hn_ 1. three is not always enough. Split M into two infinite-dimensional subspaces M 0 and M 1 . let w be a non-real cube root of unity. The restriction of A to M is a scalar and is therefore reduced by every subspace of M. the order of U in that group is exactly three. it follows that T is a left shift.2 '1E s. Every unitary operator on an infinite-dimensional Hilbert space is the product offour symmetries. it is sufficient to consider right shifts. If jE H". positive. and that would contradict the main assumption of this paragraph. and let U be wl.l. n = 0. Since U = ST. Since the inverse (equivalently. then M contains at least one point A such that E({A}) has positive rank. The remainder of the proof has nothing to do with operator theory. Relative to the fixed direct sum decomposition H = H". then the concept of determinant makes sense. 143 Solution 143. Since the determinant of a symmetry is ± 1.s guarantees the existence of shifts. ± 1. ±2. define a right shift as a unitary operator S such that SH" = Hn+ 1. then Qf = s. and begin by representing H as the direct sum of a sequence {H"} of equidimensional subspaces each of which reduces the given unitary operator U (Problem 142).SOLUTIONS the plane. the adjoint) of a left shift is a right shift. PROOF. · ·-).

and if A = t(U + U*). there exists a sequence Un} of unit vectors such that Ufn . and the proof is complete. and so therefore is U. · · ·.iU = B . Solution 145. Since U* = An* · · · A 1 *. (c) It is trivial that there is a normal operator (namely 0) within 1 of U.finite number of normal operators. Apply U* and change sign to get U*fn . then it has an inverse. (b) The norm of both the real and the imaginary part of the unilateral shift is 1. Since 1 is an approximate eigenvalue of U. This is a contradiction. it follows that u4 Reference: [62]. (Proof: for every operator. of course. Suppose indeed that u is central and that u = xyz. then it is clear that IIAII ~ 1. then .A I ~ 1. boundedness from below for a normal operator is the same as boundedness from below for its adjoint. PROOF. The unilateral shift has no square root. (c) The distance from the unilateral shift to the set of normal operators is 1. it follows that A/··· A1*A1 ···An= 1. In view of the preceding comments. If A is invertible. and therefore IIAII ~ 1. It turns out that U* is easier to treat than U.fn ~ 0. It follows by an obvious inductive repetition of the argument that each of the A's is invertible. it comes to the same thing. To get the result for the imaginary part. note that if U = A + iB. this follows from Solution 140. (a) The principal tool is the observation that if a normal operator has a one-sided inverse. and. Add and divide by 2 to get Afn . Problem 89). then II U . = uxuyuz = u(xu)y(uz) = u(yz)y(xy) = y(uz)y(xy) = yxy · yxy = 1. where U is the unilateral shift and A 1 . An are normal. Invertible operators can be peeled off either end of a product without altering its invertibility character. 144 Solution 144. the less trivial part of the assertion is that if A is normal.fn ~ 0. it follows that 269 145 .iU is unitarily equivalent to U (cf. indeed. the assertion there implies that the open ball with center U and radius 1 contains no invertible operators.) Suppose. Since ker V c: ker V 2 = N 0 . and let N 0 be the (onedimensional) kernel of U*.fn ~ 0. Conclusion: 1 is an approximate eigenvalue of A. and hence that An is left invertible. left invertibility is the same as boundedness from below. and therefore so is An*.UNILATERAL SHIFT elements of order 2.iA. where x 2 = y 2 = z 2 = 1. PROOF. The general case is now immediate: the set of invertible normal operators is dense in the set of all normal operators. this implies that An is invertible. and . that U = A 1 ···An. (b) If U is the unilateral shift. (a) The unilateral shift is not the product of a . Suppose therefore that V 2 = U*.

. in the usual sense of the functional calculus for normal operators. then W* ( = W. and hence. more "natural" proof. it follows that dim ker V = 1. in contradiction to the way f was chosen in the first place. = Ae 0 .) It is not necessary in this argument to use all bounded measurable functions. The first (and in fact the major) difficulty is to prove that q. Since N 0 is the kernel of V. ·1/J for all bounded measurable 1/J. Conclusion: there is no such V. · f for all f in L 2 . If the kernel of V were trivial (zero-dimensional). one way to do it is this. Do it again: since N 0 is the kernel of V.. (See [50. the same must be true of V. Further interesting contributions to the square root problem were made in [34] and [126]. this implies that V 2f = 0. and put q. Since U* maps the underlying Hilbert space onto itself. The very much simpler proof given above is due to J. finite linear combinations of the en's). Conclusion: q. is bounded. then Ae 0 = q. = BAe0 = ABe0 = A!/J. that U*f = 0. it must be Ae 0 . Since W commutes with A. That way the Fuglede theorem can be avoided. and hence that f E N 0 . G. The difference is large enough to invalidate one of the proofs that worked there. every function of W commutes with A. it would be sufficient to use trigonometric polynomials (i. See also Problem 151.. At this point Problem 65 is almost applicable. If A is the multiplication operator induced by a bounded measurable function q. but not large enough to invalidate the second. this implies that Vf = 0. it is the Fuglede commutativity theorem for normal operators. It follows in particular that N 0 is included in the range of V. B = 1/J(W). ·1/J = 1/J · q. If 1/J is an arbitrary bounded measurable function. then. This shows that in any attempt to prove that some operator A is a multiplication on L 2 there is no choice in the determination of the multiplier. Thompson. p. 270 .e. The hypothesis there was that A is an operator on L 2 such that Af = q.SOLUTIONS dim ker V ~ 1. and hence that ker V = N 0 . it follows that q. is bounded. the techniques used there would serve here too. and hence that there exists a vector f such that Vf is a non-zero element of N 0 . Suppose now that A W = W A. then the same would be true of U*. and if B is the multiplication operator it induces.e. all that is needed is to observe that if W commutes with A. The statement that every function of W commutes with A is not trivial. = Bq. the situation here is that A is an operator on L 2 such that AI/I = q. i. It is obvious that every multiplication operator on L 2 commutes with W. 68] and Problem 192. · e0 = q. 146 Solution 146.1) commutes with A. if there is one.. in particular. Similar negative results were first obtained in [64]. B commutes with A.

. 147 it follows that. at the same time. the last comment in Solution 65). p. for each n. There is no loss of generality in assuming that Pn --+ f almost everywhere and Apn --+ Af almost everywhere. Let U be the unilateral shift. it induces a multiplication operator. 2. The corollary is equivalent to this: if E is a projection that commutes with U. The least troublesome way seems to be to adapt (or. it agrees with A everywhere. then <p would be bounded (cf. it follows that cp · Pn --+ cp •f almost everywhere. it follows that cp · f = Af almost everywhere. where the multiplier itself is in H 00 • Since an idempotent multiplication on H 2 must be induced by an idempotent multiplier (apply to e0 ). The corollary. the product cp · p belongs to H 2 and cp · p = Ap. 1. Since at the moment neither of these ifs is known. To prove the corollary. and.UNILATERAL SHIFT The rest of the proof is trivial. does not have to be deduced from the main assertion. If cp were known to be bounded. see. then (by Problem 147) there exists a function cp in H 00 such that Af = cp · f for all fin H 2 . The crucial tool is that cp is the limit 271 148 . real. for each polynomial p. en E H 2 . . 41]. and hence. if it were known that <p ·f = Ajfor allfin H 2 . cp · Pn--+ Af almost everywhere. multiplication by en leaves H 2 invariant (n = 0. There are two ideas in this twice used proof: (1) if a closed transformation agrees with a bounded one on a dense set. ·). the desired conclusion follows from Problem 33. since. represented as the restriction to H 2 of the multiplication induced by e 1 . in particular. Since. Problem 147. and (2) multiplications are always closed. it is true for a suitable subsequence. if this is not true for the sequence {Pn}. the multiplier must be the characteristic function of a set. to repeat) the second proof used in Solution 66. for instance. Solution 147. moreover. then E = 0 or E = 1. that Apn --+ Af in H 2 . since that multiplication operator agrees with A on the dense set of all bounded functions. If f E H 2 . the proof would be over (the multiplication operator induced by cp agrees with A on a dense set). then there exist polynomials Pn such that Pn --+ f in H 2 . Since cp is bounded. for an easy direct proof see [50. there is nothing for it but to prove something. The result proved above implies that E is the restriction to H 2 of a multiplication. note that if a multiplication is a projection. then it is bounded. it follows. Since Pn --+ f almost everywhere. of course. it is inevitable to put cp = Ae0 and to try to prove that cp is the desired multiplier. As in Solution 146. it follows that ({J. then the multiplier is a characteristic function. to put it bluntly. If A commutes with U. incidentally. Since. Solution 148.

which is also invariant under V. its orthogonal complement is equal to V::O=o vnN.) This implies that yn+lu = P(VH) = vnN_L c (VnN)\ and that settles half the proof. and if gEM. (A picture helps. and it is the crux of the matter.. cf. iff E M_l_. A slightly sharper result is sometimes useful. The rest is easy. The restriction of V to the orthogonal complement V::'=o vnN is a direct sum of copies of the unilateral shift. then lvnCz)f(z)l.SOLUTIONS almost everywhere of a sequence {Pn} of polynomials such that liP nil 00 II cpll oo for every n. This suggests that the extent to which VH differs from His a useful measure of the non-unitariness of V. The only way an isometry V on a Hilbert space H can fail to be unitary is to map H onto a proper subspace of H. since V is an isometry. The incompressible core of H seems to be what is common to all the vnH's. since. llcplloo ·lf(z)l . almost everywhere. in terms of N the main result is that n=O nVnH n(VnN)l_. by the result just proved. Obviously n:=o vnu is invariant under V. This is true. the conclusion is that Pn · f ~ cp · f in H 2 .. the main thing to prove is that that incompressible core reduces the operator V.. the number of copies is dim N.l g.) Begin with the observation that VM_l_ c (VM)_l_ for all subspaces M. from f . For the reverse inclusion.. the Lebesgue dominated convergence theorem applies. it is good to know exactly what the orthogonal complement of that core is. The proof of the asserted equation is complete.l V g follows. and therefore g . n 272 . then Pg _L VnN (since f E (VnN)_l_). so that V g is a typical element of VM. and hence that f E vn +1H. this is trivial. One application of V compresses H into VH. The restriction of V to this reducing subspace is unitary (because it is an isometry whose range is equal to its domain). If n = 0.. it follows that ::'=o vnu reduces V.l N. (Indeed. the proof is complete. so that Vf is a typical element of VM_l_. It follows that iff E H 2 . assume that f E n:=o (VnN)l_ and prove by induction that f E V"H for all n. another application of V compresses VH into V 2 H. so thatf = Vng for some g. 149 Solution 149. 00 00 = n=O Both the statement (and the proof below) become intuitively obvious if orthogonal complements are replaced by ordinary set-theoretic complements. Since multiplication by Pn is a polynomial in U (namely Pn(U)).. Solution 41. Iff E VnH. Since Pn · f ~ cp · f almost everywhere. and so on. This implies that g E VH. as desired. Write N = (VH)_l_.. then Vf ..

2 null V. Since m = null U*. for every operator A.VII unitary operator V. whether n is finite or infinite. then the assertion of Sylvester's law of nullity is that null(AB) ::. B). null A (the nullity of A) is dim ker A (cf. It follows that if A is a non-normal (i.) 151 The "if" is obvious: if m = 2n. it follows that II U .. the reverse inequality is trivial. Since V* U is a non-normal isometry. and hence IIA . and the restriction of B to the second summand maps that summand one-toone into ker A. What was just proved can be expressed in geometric language by saying that if Vis unitary. non-unitary) isometry. and so also is each unitary operator. then II U . Indeed: B. For the proof it is convenient ro recall and generalize a classical and important finite-dimensional fact.UNILATERAL SHIFT Solution 150. recall that m = 2m.. If. then the square of a unilateral shift of multiplicity n is a unilateral shift of multiplicity m. What is peculiar is that this is true for every V. A unilateral shift of multiplicity m has a square root if and only if m is even. then -2 belongs to the spectrum of A .. Solution 151. = 2 for each 150 PROOF. null A + null B.1(ker A) = kerB + (B. then II U . they are as far from each other as if they were at the opposite ends of a diameter..111.1. This is a geometrically very peculiar result.VII = II V* U . The proof begins with the observation that if -1 belongs to the spectrum of an operator A.. then r(A .) If Vis unitary.. 273 . it is the negative assertion that if U is a unilateral shift of (finite) odd multiplicity. The proof can be arranged so as to work equally well whether the dimension of the underlying space is finite or infinite. The "only if" part of the statement is more delicate.VII ~ 2. If U is the unilateral shift.1) ~ 2..e. Suppose now that U is a unilateral shift of multiplicity m and that Vis a square root of U*. then U and V are diametrically opposite.1(ker A) f'l kerl. then U has no square root. and recall that the spectrum of the unilateral shift is the closed unit disc.111 ~ 2. Problem 130). the preceding paragraph implies that m::. (For the purpose of this assertion every infinite cardinal number m is even. (Use Problem 149. The unilateral shift is on the unit sphere of the space of operators.

SOLUTIONS The idea of the rest of the proof is to show that if m is finite.. even if it does. m).f. Af. then null A 2 . then equality must hold. then the sequence <f. i. That is: replace <. so that Tf.d1 + · · · + apply A. · · ·) may fail to be norm-preserving (even if L:'=o 11Anfll 2 converges. 274 . If A is an operator of finite nullity on a Hilbert space H. What is wanted is therefore a special case of the following statement. and hence that m cannot be odd. which is an odd kind of infinitedimensional reverse of Sylvester's law. Af. Lemma. the proof can be completed by proving that that set of 2m vectors is linearly independent. fJmgm = 0. for each j ( = 1. That does not work. and. · · ·) by an operator T so that the resulting series of square norms converges to I f 1 2 the easy way.e. it is replaced by <Af.. T A 2f. A 3 f. by telescoping. with ran A= H.f. If {. infer that r:t. etc. · · ·. 9m is in ker A 2 . · · ·. To construct: a Hilbert space Hand on it a shift U with the stated unitary equivalence property. The construction is partially motivated by the following observation: if a vector fin His replaced by Af. A 2 f. · · ·) should belong to the direct sum (the series L:'=o I1Anfll 2 need not converge).f1 . Since each of j~. A 2f. PROOF. then.mfm + fJ1g1 + · · · + = 0.f. There is no reason why the sequence <f. A. liTA 2fll 2 = IIA 2fll 2 . 2null A. 11Tfll 2 = 11!11 2 . To do that. A 2/. suppose that r:t. 152 Solution 152. j'..f 1 2 . Given: a Hilbert space H and on it a contraction A such that An-+ 0 strongly. its sum will be equal to I f 1 2 only in case Af = 0). • ·-) is shifted back by one step. Af. The inspiration that removes these difficulties is to transform each term of the sequence <f. · · ·) by < T Af. g 1 . • • ·.IIAYII 2 . the correspondence between f and <. A 2f.11Afll 2 .?.I A 2. and then use the linear independence of the fj's once more to conclude that rx 1 = · · · = rxm = 0. · · · ). I T Af 1 2 = I A. A 2f. A 2f.fll 2 . · · ·). there exists a vector gi with Agi = fj. Af. What this suggests is that H be something like the direct sum HEBHEBHEB···. · · ·. fm} is a linear basis for ker A. fJ1 /1 + · · · + /3m fm so that {3 1 = · · · = f3m = 0.

Let H 0 be the Hilbert space obtained from H by redefining the inner product. and if the mapping V is defined by Vf = <TJ. If f E H.A* A.A*A. Suppose that A is an operator on a Hilbert space H such that r( =r(A)) < 1. if required to hold for all f. The preceding paragraphs were intended as motivation. g)o = L (A'i. / 0 . if T*T= I .A* A.f) 0 .. is what guarantees that 275 . clearly. If U is the obvious shift on = <O. incidentally. TA 2f. then V is an isometric embedding of H into H. it follows that iff E H. where "rank" is interpreted to mean the dimension of the closure of the range. !1. Since A is a contraction. ·> Solution 153.UNILATERAL SHIFT The first of these equations alone. the proof is complete. Note that the multiplicity of the shift that the proof gives is equal to the rank of 1 .. implies that T*T= 1.A*A. . then. n=O Since I(Anf. Since r = limn IIAnll 11n. /1 . 1 . Let H write T = be the direct sum R E9 R E9 R E9 · · · .. H (U<fo.% 0 IIAnll 2 < oo. proceed as follows. all the more. For the proof itself.A*A)A'i. Ang)l ~ IIA'ill ·liAng II ~ 11Anll 2 • 11/11 · llgll.. there is no difficulty about convergence. the new inner product is given by 00 153 (f. 1 . · · -)). Ang). then 11/11 2 ~ 11/11/ ~ C~o 11Anll 2 ) ·11/11 2 . then all the equations hold. This implies that L. ). VAf = U*Vf for ·· all f Since the V image of H in H is invariant under U*. conversely. / 2 .% 0 IIAnll < oo. and let R be the closure of the range of T.Anf) n=O n=O = n=O L (IIA'ill 2 - k IIAn+l/11 2 ) 2 • = 11/11 2 11Ak+ 1 fll f Since IIAk+ 1 II -+ 0 by assumption.A* A is positive. then T Anf E R for all n. and hence. that L. and that implies that the identity mapping I from H to H 0 is an invertible bounded linear transformation. and. If 11!11 0 2 = (f. it follows that the power series L:'=o IIAnllzn converges in a disc with center 0 and radius (=1/r) greater than 1. and J k k L IITAnJIIZ = L((l. TAJ. (This.

if (1/e)A is similar to the contraction C._ 1 ) 11". 154 Solution 154.1ASII and therefore r(A) ~ inf8 IIS. and Corollary 2 is implied by the proof (given above) that IIAoll < 1. Clearly r(A) = r(S. is the spectral radius less than or equal to 1? The answer is known for all weight sequences (Solution 91). with weight sequence {o:0 . Under what conditions on a weighted shift A. To prove the reverse inequality. i. then A 0 is an operator on H 0 .SOLUTIONS H 0 is complete. but in the norm. and is especially simple in case the o:'s are strictly positive and monotone decreasing: in that case r(A) = lim (o: 0 n •• ·a. similar to A. o: 1.e.) Corollary 2 implies that IIS. then for every positive number e. the sequence of numbers IIA"II is bounded.1II·IISII. This implies that the powers of A 0 tend to zero not only strongly. apply Corollary 3 instead. and then let t tend to 1.1ASI\..1CS with IICII ~ 1.0. For Corollary 3: if A is quasinilpotent.1C"S and IIA"II ~ IIS. then it is power bounded. 2 L:'= 1 IIA~fll L:'= tCIIA'ill/11/11) 2 2 = 11!11 + L:'=1 IIA"/11 = 1 + L:':diiA"fll/11!11) 2 2 < L:'=1 IIA"II 1 = 1 + L:':1 IIA"II 2 < ' 2 so that A 0 is a strict contraction.1ASII < r(A). then A"= s. · · ·}. In view of the preceding paragraph the negative answer can be proved by exhibiting an operator A with spectral radius 1 that is not power bounded. If an operator A is similar to a contraction. so that t · IIS. not necessarily.1BSII < 1 for someS.1ASII ~ r(A). Reason: if A= s. Here is one simple example: let X be a non-zero operator with X 2 = 0 and put A= 1 +X. o: 2 . To understand power boundedness better it helps to look at the more complicated counterexample of weighted shifts.1AS) ~ IIS. then IIAo fllo 11!11/ 2 = JAr 1. 276 . The answer is no.) If A 0 Iff =1. r(A) t (If r(A) = 0. lett be a number in the open unit interval and write B=-A. Corollary 1 is immediate from Problem 152. Corollary 4 requires a little more argument. then A is similar to eC. Infer that t·infs\IS.

1 ~ oo? That is a question in elementary analysis.. That is: log ocn must be small.-) = log(n + ) = log(n + 2) . 0 unH 2 = 0. If N = M t"\ (UM)\ then N is the orthogonal complement of the range of that restriction.log(n + 1). n=O 00 Since.UNILATERAL SHIFT Under what conditions on the oc's (still assumed to be strictly positive and monotone decreasing) is A power bounded? The answer to this question too is known (Solution 91): since II A nil = oc 0 • · · ocn_ 1 .. The conditions are equivalent to 1 n-1 . the condition is that the sequence of partial products oc 0 • · · ocn_ 1 be bounded.L v V unN = H 2 .ij log oci are equal to log(n + 1). it follows that The span of M.L log oci ~ 0 n i=O and n-1 L log oci ~ oo j=O respectively. n+l n+l In that case the (telescoping) sums L}.L. n=O 00 It remains to prove that dim N = 1. unN c unM c M. and a few experiments will reveal that the answer is yes. Since n::. on the average. Can a bounded sequence of oc's be chosen so that limn (oc 0 • • · ocn_ 1 ) 11n = 1 and oc 0 · • · OCn. Conclusion: V UnN= M. it follows that M. Solution 155. For this purpose it is convenient to regard the unilateral shift U as the restriction to H 2 of the bilateral shift W 277 . The restriction of U to M is an isometry.L and a proper subspace of M can never be the whole space. on the other hand. Apply the result obtained in Solution 149 to that restriction to obtain 155 n=O n 00 unM = M t"\ n=O n 00 (UnN). but the sums of the log an's must be large. Both conditions are satisfied if 1 2 log ocn = log(1 + .

(A. = 0 for suitable scalars ai. = U*UUi. there exists a polynomial p of degree less thank such that p(U)f. if j ~ 1.. ±2.(A. and. = A. or. The dimension of N cannot be as great as 2. Vf= 1 M/(A. umcp) = enem* dJ... the proof is complete. is a linear combination of/.) If dim Mk J... in other words. It is geometric.. < uncp.. uk. ± 1.) This absurdity shows that f and g cannot co-exist.) contains all e/s.. Since /. enW + L l(g. this proves the "if". · · ·) is an orthonormal set in L 2 • (This assertion leans on the good behavior of wandering subspaces for unitary operators.A.1/. as soon as k > j + 1. cp) = o when n > o. For another proof of the same implication..A L An-len= eo. If N is the (one-dimensional) subspace spanned by cp. wmeo) L I(W*)'.l = o when n > o.1/.) = k. the reason is that (Uncp. This implies that U~ . Uk.. This is impossible. To prove equality.) under U. suppose that M is invariant under U and use Problem 155 to represent Min the form V':= 1 unN.nen. see [99. it follows (by the 2 formation of complex conjugates) that Jen Icp 1 df. then N is wandering.SOLUTIONS on the larger space L 2 • Iff and g are orthogonal unit vectors inN... See also [116]. U*U1. Since.108]. use the theory of wandering subs paces.) ~ k. wneo) + L (g.Uf.(A. for all n. since it cannot be 0 either. then UM = e 1 • M = e 1 • cp · H 2 = cp · e 1 • H 2 c cp · H 2 = M. it is clear that dim M/(A. (Indeed U*j.l = 0 when n < 0.(A.. eoW + L I(W*mg. note first that M/(A. 157 Solution 157. If M = cp · H 2 . n=O n=l 00 00 Mk J.) /. = Ui. the original proof in [54] was analytic.) is invariant under U*. and J 278 . where N is a wandering subspace for U. eo)l 11/11 2 + llgll 2 I = 2 n m 1 2 I 1 2 n m 2 n m (The inequality is Bessel's. then z:~. Note that this proves the in variance of Mk(A...) It follows that 2= = = L l(f.L(A. · · · .-J ai Ui.Ui+ 1 = U1e0 = ei E M/(A. 156 Solution 156. by assumption. = it follows that e0 E L A.) < k. and therefore dim Mk J... Halperin. consequently.) as soon as k > 1. This implies that U). and... em)l L (f.1/. Since Mk J.1 = bnm· To prove "only if"..1/. p. m = 0. and hence that Mk .) is invariant under U also.J. The last part of the proof is due to I. = 0. or Jen 1cp 12 df. then the set of all vectors of either of the forms wnj or wmg (n. Take a unit vector cp in N.. · · ·..f.) is spanned by /.

To prove the first statement of Corollary 1. it vanishes almost nowhere. and. Iff= l. so that f is an inner function. since JI <p 12 dJl = 1. 1. and. then. that every non-zero wandering subspace of U is onedimensional. Riesz theorem. · · ·) span M. by itself. For each measurable subset E of the circle. it follows that <p vanishes when f does. H 2 c tjJ. make the following definition. then() is a constant. (Note that the preceding argument contains a proof. Consider the functional representation of the two-sided shift. then <p = <p · e0 = t/J · f for some f in H 2 . and therefore (Problem 33) both Re () and Im () are constants. H 2 . and hence that M = <p. Suppose first that {an} is periodic of period p ( = 1. it is sufficient to prove that if both() and()* are inner functions. the constant modulus of <p must be 1.J::b Jj (with jj in M). 2. Since multiplication by <p (restricted to H 2 ) is an isometry. since M is the span of the image under that isometry of a dense set. the set of all functions of the form 4> · p. Conclusion: I<pI is constant almost everywhere. To prove the corollary. and M. either f = 0 or M contains a 158 function <p such that I<pI = 1 almost everywhere. As for Corollary 2:if M = <p · H 2 and N = t/J · H 2 . by the F. it follows that M is in fact equal to the range of that isometry. Equivalently. then <p · t/1 EM n N. Solution 158. then 159 = p-1 Af = and I ctjWjj j=O p-1 A*f = I IXj-1 W*jj.I and z ¢ E. To prove that. j=O 279 . · · ·. · · ·. and let Mi U = 0. Given fin H 2 . different in appearance from the one used in Solution 155. using it. spans N. observe that if <p. spans M. Since every vector of that form vanishes at least when f does. By Problem 157. H 2 . since f = <p • t/1*. it follows that <p is the limit in H 2 of a sequence of vectors of the form p(U)f. · · ·).UNILATERAL SHIFT hence that l<tJI 2 is a function in L 1 such that all its Fourier coefficients with non-zero index vanish.) Since <p. 3. and since the closure of the set of all vectors of the form p( U)f is a subspace of H 2 that contains f and is invariant under U. To prove the second statement.1) be the span of all those basis vectors en for which n j (mod p). its range is closed. and therefore g must vanish almost everywhere. 2. Since p(U)f EM for every polynomial p. Each vector f has a unique representation in the form j~ + · · · + fr 1 with jj in Mi. Solution 159. the functions <p ·en (n = 0. observe that both Re () and Im () are real functions in H 2 . let M (=ME) be the set of all those f's for which Jj(z) = 0 whenever j = 0. observe that iff does not vanish almost everywhere. let M be the least subspace of H 2 that contains f and is invariant under U. p . p . where p is a polynomial. it follows that If I = 1.

then fo vanishes on the complement of E.fjeMi. e.n. Assume now that the sequence {a.) = ( B IY. with measure strictly between 0 and 1/p./ IY. Consequence 2: if f3ii = 0 for some i and j. = a.i+k = 0 for all k. # r:t.. ei+l) Consequence 1: the main diagonal of <f3. and hence with A* A. with matrix <f3ii). where addition and subtraction are interpreted modulo p..j = m ..> A* Aei. it follows that f3ii = (Bei. 2 e. This completes the proof of the sufficiency of the condition. This says that the matrix of B is diagonal.i = 0.i+n.e. p . that commutes with A. i.. with measure strictly between 0 and 1.+t) = (B :i Aei. . . (Note that WMi = Mi+ 1 and W*Mi = Mi_ 1 . 280 .) = 2 r:t.i and i . The assumption implies that if m and n are distinct positive integers. let E0 be a measurable set.+t.j (B ei. Since A* Ae.i = /3. then there exist integers i and j such that IY. If. j = 0. then the same is true ofeachjj.. and/0 + . Necessity is the surprising part.i. then f3i+k.. jj(z) = zifo(z).SOLUTIONS this proves that M reduces A.. Clearly. A *Ae. # r:t. If B happens to be Hermitian. then /3.j.j+l = (Bei+l• e.i+n (by Consequence 3) (by Consequence 2). then it commutes with A* also. + fp-l is a typical non-trivial example of a vector in M..} is not periodic. To prove it.j f3 ii· Consequence 3: if IY.) To show that this construction does not always yield a trivial reducing subspace. If g is a function that vanishes on the complement of E 0 . It follows that 0 = /3. moreover.1. e) = 21 IY. that Pmn = 0 whenever m # n.. The set E is a measurable set. and let E be its inverse image under the mapping z ~ zP.i) is constant (put i = J). by Consequence 1 it follows that B is a scalar. Observe that /3. and if f 0 (z) = g(zP). suppose first that B is an operator. it is sufficient to prove that every Hermitian B that commutes with A is a scalar.

2 )11 =): !2 n= 1 Ll~n+kl -+0. · · · .~k-1eo = (0. U*f. ~k+ 2 . U* J.. ~ 1 . 3.) Assertion: f = (~ 0 . . · · ·. 281 .k 1 L l~n+k1 n = 1 00 2 = 0. o.eor = = II\ 1' ~e: ' ~e: ' 1 • 1 2 2 . An obvious inductive repetition of this twice-used argument proves that en belongs to the span of J.. This implies that 1 U*k.~k-1eo)- e1 1 2 = 1 ~ ~~k n~ 1~n+k 12 1 -+ 0' it follows that e 1 belongs to the span off. observe first that U*kf = <~k• ~k+ 1 . 2 Consequence: e0 belongs to the span off. and hence that f is cyclic.k ~e: 2 • .). ) belongs to that span (k = 1.. . • • . · · · for all n ( = 0. ~ 1 .k U*Y. U* 2f. U* 2f. Since Il l ~k (U *k-1f. I. ~ 2 . ~ 2 . ~k• ~k+1• ~k+2• .!. and hence that I . O. • • ·) is a cyclic vector for U*. ) 2 (l.. ~e: 1 00 1 <. 2. · · ·). ·>11 II\o. · · ·) be a sequence of complex numbers such that 1 160 limJ: !2 k <. · · ·). Consider first the simple unilateral shift U. 2. (Concrete example: ~n = 1/n!. Let (~ 0 . U*J.CHAPTER 18 Cyclic Vectors Solution 160.. For the proof. U*f.

the operator A is bounded from below. For multiplicity 2 consider the same sequence {c. Reason: since V* A is invertible. indeed. imitate this subsequence formation. These observations do not prove but they suggest that if an operator A is near to U. Thus. the most promising place to look for operators near to a translate of U* with nullity greater than 1 is on the unit circle.SOLUTIONS Once this is settled. equivalently. they behave as if they had infinite multiplicity.V*AII = IIV*V. the co-dimension of ran A must be at least 2. in fact.All < 1. The multiplicity of the number 0 as an eigenvalue of U* is 1. <~o' For higher finite multiplicities. no operator in the open ball with center V and radius 1 can be cyclic. each row is an infinite subsequence of {~n}. and hence A cannot be cyclic. the cases of higher multiplicity turn out to be trivial.) The point is that each 2 subseries of a series with the property that 2::'= 0 I'" 1 has (the ratio of terms to tails tends to 0) has the same property. that its adjoint has nullity 2 or more. it is necessary to observe that in the case at hand the linear manifold ran A is closed.V*AII ~ IIV. The complex numbers A with I A1 = 1 are not eigenvalues of U*. as such. that I V . As far as 282 . a cyclic vector for the shift of multiplicity N0 is the vector whose i-th component is the i-th row of the following array: ~0 ~1 0 ~2 0 0 0 0 0 0 '3 0 0 0 0 ~4 0 0 0 0 '5 0 0 0 '0 6 0 0 ~7 0 0 0 0 0 '8 0 0 0 0 '9 (Rule: lengthen the diagonals. and even for multiplicity N0 . The trouble is that there exist dense linear manifolds disjoint from subspaces of arbitrarily large finite dimension. Suppose. The usual way to prove that an operator is not cyclic is to prove that it has co-rank 2 or more. and hence that V* A is invertible. then the multiplicity of A as an eigenvalue of A* is likely to be 1 for all A with IAI < 1. The answer is yes. for instance. · ·-) ). Each column contains only one non-zero entry. and the same is true of A for all A in the open unit disc. · · -). and hence that ran A n ker V* = 0. then (Problem 160) Vis not cyclic. Caution: the preceding reasoning is incomplete. but they are approximate eigenvalues and. Since ran A is disjoint from the 2-dimensional space ker V*. ~3. ~4.n} and form the vector < ~2. 161 Solution 161. It follows that Ill. (~1. or. To avoid error. If U is the unilateral shift and V = U Ef:) U. 162 Solution 162. ~5.A I < 1.

(This elegant proof is due to N. Since the vectors of the form A"f. then (<U"f. g) is a cyclic vector of U EB U*.'s so that for distinct values of n they have disjoint supports (that is. For such a vector. · · ·..(1.(g. for each positive 2 integer n. 1. The answer is no.q(A *)f II < e.An --+ U. The proof itself begins with the observation that 1 is an approximate eigenvalue of U* in a strong sense: there exists an orthonormal infinite sequence Un} of vectors such that 2 11(1 . it follows that the co-rank of An is at least 2. and.U). Vasunin. then the coordinatewise product off. g*) = 0.An. If Pn is the projection onto the 2-dimensional space spanned by fn and fn+ band if An = (1 . consider U* as acting on 1 .(Unf. and. <g*. Nikolskii.An)ll = 11(1.U). Solution 163. g*). if n ¥:. -{n "n so that 1 . and let f* and g* be their coordinatewise complex conjugates. that f and g are in 1 .U *)Pn II ~ 2}2 .m.Ani!= 11(1. therefore. Choose the f.U)*(1 . The answer is yes.A. consider a vector fn in 1 with n consecutive coordinates equal to 1/Jn and all other coordinates equal to 0. Since ~ dim ker An* = dim ker(1 .U*)fnll --+ 0. 163 Conclusion: no <f. 2.Pn) 2. 164 283 .A/II = II(1 . K.U)*. and V.U*)fnll ~ 2/Jn. the norm llfnll is 1. g*) . neither is l . V. an operator A is just as cyclic or non-cyclic as A . and fm is 0). So much for the motivation of the proof.(U*ng. and 11(1 . span the space. then the set of non-cyclic operators on His dense [ 42]. it is sufficient to prove the assertion in case g = Anf. -/*)) = (U"f. Conclusion: An is not cyclic.CYCLIC VECTORS cyclicity is concerned.) Solution 164. U*ng). 2 Suppose. Indeed. A somewhat more sophisticated version of the same argument proves the general result: if dim H = ~ 0 . g*) . U"f*) = (U"f. f*) = (U"f. translation by a scalar makes no difference. I. V. n = 0. Peller. then IIU. The assertion is that to every vector g and to every positive number e there corresponds a polynomial q such that llg . indeed.Pn)(1 .

Another consequence of Solution 165 makes contact with the theory of total sets (cf.e. 3. Next: if g is an arbitrary bounded measurable function on [0. Solution 165. then so do all the A"f's. First step: if <pis a bounded measurable function on [0.p(A)fll < e. each <pn is uniformly arbitrarily near to a polynomial. so that Pnh E K. n = 0. There exists an operator A on a Hilbert space H. Each <p" is a bounded measurable function. as asserted. as asserted: g E K.(x) = g(x)/ f(x) otherwise. it follows. 1]. 1). If now hE K. To prove this. and. n = 1. 1) if and only if f(x) =F 0 almost everywhere. what 284 . so that iff vanishes on a set of positive measure.L is cyclic for A. 1]. and if K is the closure of the set of all vectors of the form p(A)f. then g E K. By the Weierstrass polynomial approximation theorem. then {pnh} converges dominatedly almost everywhere to <ph.p(A))fll. and by definition K is a subspace. that can be done because f is cyclic for A. A vector f is cyclic for the position operator on L 2 (0. The desired conclusion is now immediate: by the preceding paragraph K is dense in L 2 . If q(z) = (p(z*))*. so that K = L 2 . and there exists a subspace K ofH such that both K and K . PROOF. K = H 2 . and therefore <pn f E K by the preceding paragraph. 2. Problem 9). therefore. · · ·.. so does everything in their span. it follows that the sequence {<pnf} is uniformly bounded almost everywhere (by llglloo) and con verges to g almost everywhere. find a polynomial p such that IIA*"f . Since <pn f = 0 when If I < 1/n and <pn f = g otherwise. then <pK c K. Since pK c K for polynomials p.p(A))*fll = II(A*".q(A*)fll 165 = II(A*". and therefore. Put H = L 2 (0. then K = L 2 . PRooF. then q is a polynomial and IIA"f. · · ·) is total in V(O. 2. Corollary. It needs only easy analysis to prove that the set of all powers fn (i. fn(x) = x". Consequence. 1). Pnh-+ <ph in V. that <pK c K. It follows that there exists a sequence {Pn} of polynomials converging boundedly almost everywhere to <p. The non-trivial part of the problem is to prove that iff =F 0 almost everywhere.SOLUTIONS Given n and e. observe that by Fejer's theorem (see Solution 41) there exists a sequence {<pn} of trigonometric polynomials converging boundedly almost everywhere to <p. by Lebesgue. write <pn(x) = 0 when I f(x) I < 1/n and <p. then A"f vanishes at least as much as f does.L are infinite-dimensional and such that every non-zero vector in either K or K . The "only if" is obvious: if A is the position operator. and let A be the position operator on H. normality is needed for the last step. To prove this. where p varies over all polynomials. 1.

-l)P(n)fp.rxA). If rx = 0. e 1. 1. This implies (1. this surprising and satisfying statement is the Miintz-Szasz theorem [139]. Indeed: to prove that Un. The main step in the proof is to show that iff is cyclic for A. ···}is triangular + 1 and has no zero entries in the diagonal just 167 285 . the conclusion is trivial. The span of the set of all cyclic vectors of an operator on a Hilbert space His either 0 or H.1 (1 .rxA) that fP and hence that m = 0.. Geher [ 46] . note that = 166 L (rxAt(l 00 = 1. The proof of the principal assertion is to show that the fP's span H. Since the vectors Anf span H. I. every fP is cyclic. therefore.rxA)Pf. and fn+k = Akf. · · ·. Amfp = L ('J. The fact is that a set of powers is total if and only if the reciprocals of the exponents form a divergent series. · · ·.rxA). Both the result and the proof are due to L.J2 . 1.rxA) n=O converges (in the norm).fn+z. 2. p for n = 0.fn+ 1 . • • • span H. ···. For that purpose.provided only that rx is sufficiently small. k = 0. .. p = 0. it follows that the vectors fo . then so is fP = (1 . its sum is (1 . ·}is total for each n.n An+mfp+ 1' 00 (rxAt = (1 . 1. 2. then so is fp+ 1 .(1 .J1 . • · ·. It follows that the series .CYCLIC VECTORS takes a little more is that that set remains total after the omission of each finite subset. 2. n=O Consequence: if fP is cyclic. and. by induction. e 2 . and that. 2. therefore. Suppose now that 0 < II rxA II < 1. (k A'i = -~ p=O ±( Solution 167.rxA))n = p=O I (-1)P(n) (1 p rxA)P. note that they are valid for arbitrary Banach spaces. Solution 166. just observe that fn is cyclic for the position operator.1fp+ 1 = n=O L rxnA'1p+ 00 1. (a) If A is an operator whose matrix with respect to a basis {e0 . That all cofinite sets of powers are total is a very small fragment of the whole truth.

it is clear thatf E 1 (in fact. for convenience.j2Pn in f has exactly Pn predecessors. Ae 0 . / 2 . Ae 0 . p2 .fn-d 2 Since Ln llf. and it does not. let U be the unilateral shift. consider the total sequence {f. then e0 is a cyclic vector of A. A simple (but unnecessarily generous) way to determine the k's and the p's is as follows: choose k. (b) Yes. then A"+ 1f would belong to the span of {f. k3 . k 2 . followed by the coordinates of fd2P•. ···.'s and on two sequences (k1. The reason none of the entries just below the main diagonal is zero is that if en+ 1 belonged to the span of {e 0 . vectors in finite-dimensional Euclidean spaces) such that if each one is converted into an element of P by tacking on an infinite tail of O's.) Consider now a sequence <ft.. ···)of positive integers. (That is: the first coordinate of /. its proof is only a slight modification of what follows./ E P). followed by ftf2P• (i.) If /is a cyclic vector of A. kn and the lengths of ft. then the matrix of A with respect to that basis is triangular + 1. Let H be 1 . e 1 . ···}contains e2 .) If {e 0 .f begins with f. and e 2 . with the coefficient of e 2 different from zero.. These are the first steps of an induction. of length k2 . and so on ad infinitum. 168 2 Solution 168. the rest are just like them. with the coefficient of e 1 different from zero. so that llfnll/2kn < 1/2".) Next: the span of {e 0 . · ·-) and (Pt."'. (That coefficient is the first entry in the diagonal just below the main one. · • ·. in order). the converse is true. and the proof in the preceding paragraph applies.II/2Pn < Ln llfnll/2kn < 00. / 3 . · · ·} and note that it is linearly independent. of length k 1. The vector f to be constructed depends on the f. • • ·} is the basis obtained by orthonormalization. (This is where infinite-dimensionality is used. Alternatively: Pn is the sum of k 1 . · · ·. (There is nothing magic about 2.. e 2 .e. The reasoning goes as follows. A 2 e0 . The converse (b) is also true for finite-dimensional spaces.e. any number greater than 1 would do just as well. Af. p 3 . e 1 . A"f}. then the resulting sequence of vectors is dense in P. because Ae 0 is a linear combination of e0 and e 1. The density of the APnj's follows from the facts that AP". ···}contains e 1. (The result (a) is true for finite-dimensional spaces.SOLUTIONS below the main one. and write A = 2U*.f begins with /. Assume. followed by f 2 j2P2 . A 2f. A 2 e0 . because A 2 e0 is a linear combination of e0 . followed by a sequence of O's. and the norm of the tail of APnJ is the square root of 286 . · • ·) of finite sequences of complex numbers (i.en}. that the underlying Hilbert space is of dimension ~ 0 . The span of {e 0 . and it looks like this: a sequence of O's. and let Pn be the (unique) exponent such that AP".

287 .. and then approximate h within 8/2 by a vector of the form <fn.. that it is 1.~k•~k+b~k+2•··) 0. with n > n0 . then no vector of the form uy can be within 8 of e0 . Here is a pertinent question: does U have a dense orbit? The answer is no.···. . then the initial coordinate off cannot be 0. If 8 < min( 1. · · -).o. ). if n = 0.U"f I ~ 1 (look at the coordinate with index 0). Reason: if n > 0.O.o. then lleo .. . Reason: every orbit is bounded. U 2f.CYCLIC VECTORS Indeed: to approximate an arbitrary h in /2 by APnJ within 8. in [73]. 0. then consider the vector eo= <1. Uf. The merit of this technique is that a slight refinement of it can be used to show that even the set of scalar multiples of the vectors in the orbit off is not dense. If the next non-zero coordinate is the one with index k so that f = with ~k =1= (l. with no loss.o. The result here described is due to Rolewicz [117]. then I e0 . choose n0 so large that every tail beyond n0 has norm below 8/2. A related result (with scalar multiples allowed but sums still not) was discussed later. . 0.uy II ~ I~k I (look at the coordinate with index k). assume. A more complicated proof is worthwhile. I~k 1). · · ·} is dense.. If {f.o.O. in a different analytic context.o. 0.

jj --+ f (w). then. g) for all g. It follows that the inverse image under A of the open unit ball is a weak open set. If A is (s --+ s) continuous. and hence that it 288 . but heavily. --+ 0 n and hence that (w). a fortiori. This proves that A is (w --+ w) continuous. A*g) = (Af. assume the opposite. then (Ajj. and if {jj} is a net w-convergent to f. That implies the existence of a sequence{!.--+ 1 n 0 (s). then A is bounded.Af. This proves that A is (s --+ w) continuous.CHAPTER 19 Properties of Compactness 169 Solution 169." ~ n.fj} is a net s-convergent to f. If A is (w --+ w) continuous. g)= (jj.f. the assumption implies that 1 . Note that the assumption of (s --+ s) continuity was tacitly. this is contradicted by ~~~ Af. is a bounded sequence.// ~ n2 • Since . To prove that if A is (s --+ w) continuous. and the assumption implies that Ajj--+ Af (w). used via the existence of the adjoint A*. A*g)--+ (f. so that Ajj --+ Af (w). and if {. Suppose.} of unit vectors such that 1/Af. finally. that A is (w--+ s) continuous.

if A is compact.A. cf. it is to be proved that Ajj ---+ Af whenever {jj} is a bounded net converging weakly to f. Since the closed unit ball is weakly compact.Afll.J. and therefore strongly closed. consider the orthonormal basis {e. and.ll = supki1Xn+kl. The first term on the right is dominated by IIA -An II· lljjll. i = 1. fk includes a basic weak neighborhood of 0. This completes the proof that K is closed. there exist vectors and there exists a positive number 8 such that if j(f. a linear transformation that is continuous (s ---+ s)) is continuous (w ---+ w). Let A be an operator with diagonal {1Xn}. uniformly in j. it follows that A* E K. for each positive integer n. note that the range of A is equal to the image under A of the orthogonal complement of the kernel. since {lljjll} is bounded.. 1Xn. to prove the converse. so that P E K. Suppose now that A. if A E KandA = UP. then Aen . Since the limit (in the norm) of compact operators is compact. consider the diagonal operator An with diagonal { IXo . it too is small for large n.t 0. Since A . since A* = PU*.Anfll + IIA" f . consequently. 0. therefore weakly closed. · · ·}. · · ·. then certainly I(f.All . is a diagonal operator with diagonal {0. Solution 170. 171 1Xn _ 1. it follows that the first term is small for all large n. To prove the converse. so that IIA. it is clear that the assumption 1Xn . then the range of A is infinite-dimensional. (If there is an infinite-dimensional subspace on which A is one-to-one. • · ·.)I < 8.e. · · ·}.t 0 strongly (because en . · · ·. it follows that if 1Xn . it follows that A! must be 0. If A is compact.t 0 implies the conclusion IIA .A. In other words. The proof that K is an ideal is elementary.) To prove the corollary. E K and IIAn . the compactness of An implies that the middle term is small for "large" j. /. then P = U* A (see Corollary 1. 289 . i)(n+ 1.A. Problem 134).} that makes A diagonal.)I < 8. The proof that K is self-adjoint is easy via the polar decomposition.t 0. k. /. This proves that A annihilates a subspace of finite co-dimension.PROPERTIES OF COMPACTNESS / 1. 0. · · ·. • • • . J. and this says exactly that 1Xn ---+ 0. Note that IIAjj.lf f is in the orthogonal complement of the span of {!1. Fix some large n.Anll . and this is equivalent to the statement that A has finite rank.t 0. 0.t 0. Since this conclusion applies to all scalar multiples off too. k. Indeed. and.fll ~ IIAfj. Solution 19). then A is compact. In other words. then II A! II < l. use the result that an operator (i.A II · II fll. The last term is dominated by II An . 0. Solution 171. it follows that its image is weakly compact.t 0 weakly. then II1Xnen11 .jj. i = 1. · • ·.AnfJII 170 + IIA. and therefore IIA/11 < 1.

172 Solution 172. then A is compact. equivalently.(y). the operator A has finite rank. Suppose now that A is a compact normal operator. i= 1 n and where each g. and each h. (Proof: the image of the unit ball under an invertible operator is strongly compact if and only if the unit ball itself is strongly compact. belongs to L 2 (fl). is bounded from below. The spectrum of A is the essential range of cp.:. 290 .) Since the restriction of a compact operator to an invariant subspace is compact. and the restriction of A to M. The proof is immediate: iff E L 2(fl). it follows that M. is finite-dimensional. from this everything follows. reduces A. by the spectral theorem there is no loss of generality in assuming that A is a multiplication operator induced by a bounded measurable function cp on some measure space. it follows that if the restriction of a compact operator to an invariant subspace is invertible. The simple functions in L 2 (. Solution 173. Begin with the observation that a compact operator on an infinite-dimensional Hilbert space cannot be invertible.(y)f(y)dfl(y). and each h.. Recall that a simple function is a measurable function with a finite range. y) 173 = L g. which implies that A and SA are simultaneously compact or not compact. It is just as easy to prove something much stronger: as long as each g. It follows that the finite linear combinations of characteristic functions of measurable rectangles of finite measure are dense in L 2 (fl x fl). is a scalar multiple of a characteristic function of a measurable set of finite measure. Clearly each M. the part of the spectrum that lies outside the disc {A: IAI :.SOLUTIONS S*S If Sen = en+ 1 . a simple function is a finite linear combination of characteristic functions of measurable sets. With a sufficiently powerful tool (the spectral theorem) the proof becomes easy. then each of A and SA is a multiple of the other (recall that = 1). for each positive integer n. let M. 1/n} can contain nothing but a finite number of eigenvalues each of finite multiplicity.u) are dense in L 2 (fl). The preceding paragraph implies that.(x) J h.(x)h. an equivalent condition is that it is a finite linear combination of characteristic functions of measurable sets of finite measure. and let M. In view of these remarks it is sufficient to prove that if A is an integral operator with kernel K. This remark proves the corollary. then (Af)(x) = it g. In fact the range of A is included in the span of the g's. For each positive number e. where K(x. A simple function belongs to L 2 if and only if the inverse image of the complement of the origin has finite measure. be the set {x: lcp(x)l > e}. then the subspace is finite-dimensional. be the subspace of L 2 consisting of the functions that vanish outside M.

If A is a Hilbert-Schmidt operator. it consists of the functions gii defined by gii(x. To say that A is a Hilbert-Schmidt operator means. in the diagonal case. j j The proof is over. of course) in terms of that basis. Afj) = L (A* Afj. Since A* A is a compact normal operator. if the sum of the squares of the moduli of the entries is finite. write A* Ajj = A./x. The rest is simple computation: 174 IIKII 2 = 'L 'L I(K. say. Thus. that a slightly less sensible looking basis is algebraically slightly more convenient.(x)* di1(X) = r (by Parseval) j = L L I(Afj. in particular. if the matrix is diagonal. Consider an infinite matrix (i. by Parseval's equality. PROOF.. It turns out. the matrix defines a Hilbert-Schmidt operator. By definition.ijj. then the sum of the eigenvalues of A* A is finite. that A is an integral operator on.(x)* d11(x) 1 2 ~ ~ If(Ajj)(x)/. express the L 2 (11 x 11) norm of K (which is finite. This is true.(x)*fj(y)dl1(x)di1(Y) I 2 1 = = ~~I J(f K(x. there exists an orthonormal basis {fj} consisting of eigenvectors of A* A (Problem 172).fj) = L Aj. however.(x)jj(y). The theorem just proved implies that in that case the finiteness condition is not only sufficient but also necessary for the result to be a Hilbert-Schmidt operator. There is only one sensible looking basis in sight. y)}j(y)di1(Y) )/.(x)jj(y)*. /. L 2 (11). . the difference between compact and Hilbert-Schmidt is the difference between a sequence that tends to 0 and a sequence that is squaresummable. the one consisting of the functions 9ii• where 9. The useful way to put the preceding two statements together is to introduce a suitable basis for L 2 (11 x 11) and. y) = /. 9iiW (by Parseval) i j = ~ ~ JJK(x. induced by a kernel K in L 2 (11 x 11). a "kernel" on /2 ). y)/.e. of course.PROPERTIES OF COMPACTNESS Solution 174. y) = /.11 2 j i = L (Ajj.)1 2 j L IIA/. 291 . The construction of a concrete compact operator that does not satisfy the Hilbert-Schmidt condition is now easy.

and hence that A = U ·limn Pn =limn UPn. bye say..) The final step is to show that if I contains an operator A that is not compact. there exists an infinite-dimensional subspace M. By Problem 172. (Note that separability was not needed yet.f EM for all . observe that if u and v are non-zero vectors. If A is compact and UP is its polar decomposition.f. and. the proof is complete. then the operator A defined by Af = (f. To prove this. 177 Solution 177. Since P is Hermitian. Let B be the operator defined by Bf = (f. invariant under P. A simple one is given by the classically important Cesaro matrix 1 0 0 l_ C= 2 l_ 2 1 1 1 3 l_ 3 0 0 0 1 3 0 4 1 4 4 4. V. on which Pis bounded from below. there are many counterexamples. as promised. Solution 176. P is the direct sum of 0 and a diagonal operator on a separable space. The first step is to show that I contains every operator of rank 1.fll = IIPV. and every operator of rank 1 has this form.) Since PM = M. Since V* PV E I. it follows that V* PVH = V* PM = V*M = H. and P is not compact (because A = UP). Since I is an ideal. Since UPn has finite rank for each n. that CA 0 B = A. u)v = Af. and let u0 and v0 be non-zero vectors such that A 0 u0 = v0 .fll = ell/11· These two assertions imply that V* PV is invertible.fll ~ eiiV. To show that each such operator belongs to I.) Let V be an isometry from H onto M. (If not. u)v has rank 1. u)Cv 0 = (f. an ideal that contains an invertible element contains everything. Suppose that I is a non-zero closed ideal of operators. Since. then P E I (because P = U* A). This implies that P is the limit (in the norm) of a sequence {Pn} of operators of finite rank.SOLUTIONS 175 Solution 175. and the sequence of diagonal terms of the diagonal operator tends to 0. u)u 0 = (f. take a non-zero operator A 0 in I. it follows that A E I. it follows that I contains every compact operator. P would be compact. 1 292 . and let C be an arbitrary operator such that Cv0 = v. then P ( = U* A) is compact. (Here is where the separability of H comes in. Since I contains all operators of rank 1. it follows that IIV*PV. the proof is complete. i. u)u 0 .e. since I is closed. it contains also all operators of finite rank. If UP is the polar decomposition of A. then I contains every operator. The answer to the "some" question is no. moreover. It follows that 176 CA 0 Bj = CA 0 (f.

write an arbitrary operator as a scalar tC 293 . To prove (b) note that if A is an arbitrary Hermitian contraction. there is an almost trivial construction that settles everything. so that B is a co-isometry. necessary to prove that Cis bounded but not compact. Consider the collection J of all the operators that map every orthonormal basis onto a strong null sequence. then AU E J. and. of course. and U + V) = A.PROPERTIES OF COMPACTNESS To prove that it works it is. but for the present purpose even easy proofs are unnecessary effort. then both U = A + i Jl . The answer to the "every" question is yes. and therefore AU E J. in fact. the operator A obviously maps the natural coordinate basis onto a strong null sequence. 3 3 1 0 0 t t 3 Since the matrix A is the direct sum of an infinite sequence of projections. then BA E J for every operator B: in other words. if A E J. A E J. moreover. it is obviously bounded but not compact. 3 l. More is true: J is. the most efficient answer to Problem 177 is the projection A = B* B. a two-sided ideal.A 2 are unitary. and U is unitary. The reason is that (a) if A E J and U is unitary. 0 0 0 2 2 0 0 t t t 1 0 0 l. To prove (a). 2 2 0 0 0 l.A 2 and V = A . Note that BB* = 1. and (b) every operator is a linear combination of four unitary ones. then {Uen} is an orthonormal basis. The collection J is a norm-closed vector space. J is a closed left ideal. so that IIAUenll -+ 0. An alternative possibility is 1 0 1 0 1 0 0 1 0 0 1 0 0 1 0 0 0 0 0 j2j2 0 0 0 0 0 0 0 0 B= j3j3j3 0 0 0 0 0 0 0 0 0 In this case it is easy to prove directly that B is bounded but not compact. l.i Jl . note that if {en} is an orthonormal basis. Both those statements are true [21] but their proofs would be a digression that is not worth making here. 1 0 0 0 0 0 A= 0 0 0 0 0 1 l. since njn 2 -+ 0 as n-+ oo.

that basis and the number ej4 do what is required. -+ 0. 1 ..ll .s with ej2 < IIAe. To prove that. e for many n and then change both the basis and the number e.=. A.f. then A is compact.( Cfn) -+ 4[. If Cfn = A. ···}be an orthonormal sequence such that g. 2 . such that IIAe.e.ll .ll .nfn.I)IIfll. and the replacements together form an orthonormal basis. it is sufficient to prove that if A. Consider the indices k for which IIAekll < e/2. it follows that Cf = 4[. induced by a bounded measurable function cp on a suitable measure space.f. Conclusion (via Problem 138): an operator is in J if and only if it is compact. The preceding paragraph proves that J is a closed ideal of operators. e.)lj2. Since C(.=.} by {(ek + e.=.} does not converge strongly to 0.. This implies that A is the direct sum of 0 and a diagonal operator with diagonal terms converging to 0. and hence that A is compact. with fn # 0.f3 .. The corollary is the following strengthened statement: if A is not compact.. to f say. 3 . e.SOLUTIONS multiple of a contraction.. and there exists a positive number e. • • · are distinct eigenvalues of C. # 0 and Cfn . in terms of f 1. The answer is finished. The idea of the proof is to start with a basis {e.-+ 0. II = 1. and then express its real and imaginary parts as linear combinations of unitary operators. A. then there exists an orthonormal basis {e. fn} for each n.=. and therefore f # 0.ll . I . then A.. e for infinitely many values of n. Clearly J has non-zero elements (namely.. Expand Cg. Indeed: if A. the unpaired e.. If A is normal and integer n. so that llfnll-+ (1/IA. efor every value of n. g 2 . if A" is compact for some positive PROOF. With the approximate point spectrum TI(C) in place of spec C the proof is easy.)/j2}. then use compactness to drop down to a suitable subsequence so as to justify the assumption that the sequence {Cf. but it is worthwhile to call attention to an interesting and useful corollary. The next useful step is to observe that TI 0 ( C) is countable. 179 Solution 179.f2 . • • ·. it follows that the essential range of q.} converges.} such that I A e. Represent A as a multiplication operator. · · ·. by Problem 172. the e...} such that {A e.ll < e. (ek. with I f. It follows that ). 1 ¢ J). fn and V 294 . E {f1.}. Since. -+ f. then there exists an orthonormal basis {e.. and note that this automatically represents A" as the multiplication operator induced by cp".). and replace the pair {ek> e. Let {g 1 . • · ·} is linearly independent. What is now known is that if A is not compact.)-+ Cf and A." is a countable set that can cluster at 0 alone. g 3 . A" is the direct sum of 0 and a diagonal operator with diagonal terms converging to 0. That implies that there exists a positive number e such that IIAe. all compact operators). and J does not contain every operator (for example.s with IIAe.=. then the sequence {f1. Pair each such k with some n such that IIAe.lf.. 178 Solution 178.

and so is the intersection N n ker. is always a subspace.fll ~ IIBII· II A/II for allf.e. by invertibility.BA. Note that both P and Q have ker A. it includes a closed ball in M (i. and since AB = 1 . This implies that M is finite-dimensional.. and ran A is the sum A(ker BA) + A kerJ. of which one is finite-dimensional. It follows that both ker A* and ker A are finite-dimensional. the image under A of ker BA is finite-dimensional and hence closed.AB and D = 1 . BA. (See Solution 179.Q on both ran A and ran.e.f. The restriction of A to that intersection is a one-to-one bounded linear transformation from that intersection onto M.) 295 181 . (Recall that the sum of two subspaces.P on both ker.l A. and. In the present case ran A is closed. the asserted convergence follows from the compactness of C. 180 Solution 181.l A and ker A are finite-dimensional. and let Q be the projection on ran. by the preceding two paragraphs. the inverse mapping is bounded. it follows that both 1 . and therefore. Note that both P and Q have finite rank.lBA. it follows that A is bounded from below on kerJ. Suppose that A is compact and suppose that M is a subspace such that M c ran A. see Problem 13.AB have finite rank. The image of the closed unit ball of N (i. Trivial: an operator of finite rank is compact. BA. The only way a bounded set in the plane can have a countable boundary is to be included in its own boundary.). of the intersection of the closed unit ball with N) is a strongly compact subset ofM. then both ker B* A* and ker BA are finite-dimensional.BA and 1 . The inverse image of M under A is a subspace. it is the same as the coefficient of . This proof was shown to me by Peter Rosenthal. it includes the intersection of a closed ball with M). (2) implies (3). Let B be the operator that is equal to that inverse on ran A and equal to 0 on ran. by the closed graph theorem.1A. (3) implies (1). the boundary of spec C is countable. and. the boundary of spec Cis included in TI( C).l A. and therefore (Problem 52) it is invertible.l A one-to-one onto ran A. Since ker BA is finitedimensional.l A and ker A. the proof is complete.PROPERTIES OF COMPACTNESS observe the coefficient of fn in the expansion. Since BA = 1 . note first that BA is bounded from below on ker. and hence that the inverse mapping maps ran A one-to-one onto ker. therefore. say N.l BA is closed. An = (Cg. g. The proof of the full Fredholm alternative is now easy to complete. (1) implies (2).l A. with C and D compact.. If C = 1 . and hence that both ran. It is always true that A maps ker.l A. and hence that the image under A of ker.) Since IIBA. and therefore.. To prove that ran A is closed. in the expansion of An g•. Consequence: whenever n > 1. By Problem 78. Solution 180.

so that D is compact.1 (A .B) is invertible (in which case A is invertible).C. In other words.SOLUTIONS 182 Solution 182. but it is invariant under W . it follows (Problem 182) that every number in spec. What is the operator W . then B is not invertible. B* B. or 1 + B. This conclusion makes it easy to describe W. then U*U = B*B. it follows that the spectrum of B* B must be 296 . and let C be the operator defined by Cf = (J. consequently.C maps en onto en+ 1 or 0. The bilateral shift is an example.C. 0) is invariant under both Wand C.) Since a Hermitian operator on a separable Hilbert space can have only countably many eigenvalues.B.C. PROOF. according as n < -1 or n = -I.C? S'ince H 2 (the span of the en's with n . it agrees with the unilateral shift on H 2 and it agrees with the adjoint of the unilateral shift on the orthogonal complement of H 2 .Cis the direct sum U* EB U. ± 1. If U = B . assume that B is invertible and write A = B + (A . (Alternatively. Since U*U = 1. where D = C* B + B*C . must in fact be an eigenvalue of B* B. 184 Solution 184. W. · · ·} is the basis that is being shifted (Wen= en+ 1 ).B)). the one in row 0 and column -1.B) = B(1 + B.B) is compact along with A . 183 Solution 183.1 (A . (Reason: if n < 0. For the proof. Translate by A. It helps to look at all this via matrices.1(A . it is invariant under W . The orthogonal complement of H 2 (the span of the en's with n < 0) is invariant under neither W nor C (since We_ 1 = Ce _ 1 = e0 ). by 0.B) (in which case ker A =P 0). and it is therefore compact. The matrix of W (with respect to the shifted basis) has 1's on the diagonal just below the main one and O's elsewhere. Suppose that {en: n = 0. except possibly 1.D. Contrapositively: if B is invertible. the effect of subtracting Cis to replace one of the 1's.-. ±2. The operator B. and since spec 1 = {1}. No perturbation makes the unilateral shift normal. and.1(A . its spectrum is the union of the spectra of U* and U.C*C. then either ker A =P 0 or A is invertible.) Conclusion: H 2 reduces W . then W.C also. e _ 1 )e0 • The operator C has rank I (its range is the span of e0 ). and reduce the assertion to this: if A is not invertible but ker A = 0. It follows that either -1 is an eigenvalue of B.C. The technique is to examine the spectrum and to use the relative stability of the spectrum under perturbation. with B normal and C compact... use the Fredholm alternative.

C) II f. Problem 182 implies that spec(U . in fact. then II U . modulo countable sets. 1.011 = 1. 185 Use of a somewhat more powerful tool yields a somewhat simpler proof.C)*(U . If U is the unilateral shift. Conclusion: modulo countable sets. 1].C) includes the unit disc. Solution 185.C) = 1 . and since U has no eigenvalues. For the proof. A normal operator on a separable Hilbert space can have only countably many eigenvalues. and therefore II U . 297 . This contradicts the countability of spec B* B. Consequence: 1 E spec((U .C)*(U . 1.C)*(U .C' with C' compact. 1. spec B* B is the interval [0. and therefore the distance from U to the set of all compact operators is not more than L The distance is.C)*(U . Conclusion: II(U . suppose that Cis compact.C)) f. spec B is the unit disc. Since spec U is the closed unit disc. another consequence of Problem 182 is that the spectrum of B can differ from the disc by the set of eigenvalues of B only. and observe that (U .Cll f. equal to 1.PROPERTIES OF COMPACTNESS countable.C)). whence r((U . and therefore (Problem 123).

y)l ~ (n ~n 1)! (x. z)K(z. because unless z is between y and x one of H(x. y) = 0). (In this context symbols such as K 2 . y)dz I~ c 2 3 f(z. y)l 3 =I fK(x. 1K(x. y. with. y)dz. then IK (x.y)dz =~ (x. z)K(z. y) I = \ f K(x. 298 . Reason: (HK)(x. etc.y). y) 1 ~ c. y) must vanish. It follows that if K is a bounded Volterra kernel. etc.CHAPTER 20 Examples of Compactness 186 Solution 186. then IK 2 (x. If Hand K are Volterra kernels.) From this in turn it follows that if x 6:. then so is their "product" (matrix composition). z)K(z. then.. if x 6:. KKK.y) 2 • These are the first two steps of an obvious inductive procedure. 1 and x 6:. z) and K(z. the general result is that if n 6:. for all z. either x < z (in which case H(x.y)n-1. then (HK)(x. y) = 0 if x < y. refer to the "matrix products" KK. and if x 6:. In other words (HK)(x. y) = fH(x. or z < y (in which case K(z. y)dz I~ c 2 • (x . z)K (z. K 3 . and if x < y. then IKn(x. y. y)dz. y. y. y) = fH(x. z) = 0). say.

~ J:IK(x.1.. y)l 2 dy almost everywhere in (a. Since log g(1).1 log g(x{ = f1. note that 0 < g(l) < oo..1. Since =0 . Let a be the infimum of the support of g. A Volterra operator has no eigenvalue different from 0. y)l dy·J:If(y)l dy 2 2 almost everywhere. 187 Suppose that K is a Volterra kernel with associated Volterra operator V. so that Vf = ~f. y)f(y)dy almost everywhere.log 0 = oo and IlK II < oo. 1). g(a) and g(x) > 0 whenever a < x ~ 1. g(x) = f1 f(y) 12 dy. the proof is complete.1 2lf(xW ~ J:IK(x. By integration the last inequality becomes 2 2 1. so that l. a fortiori..1. Solution 187.. and suppose that ). If PROOF. then Since (en~ 1)!r/n --+0 asn-+ 00 (recall that the radius of convergence of the exponential series is oo ).1. = 0.f(x) = f K(x. y)l dy dx = IIKII .EXAMPLES OF COMPACTNESS This implies.e. that c" IK"(x..1. 299 . y)l ~ (n _ 1)!' and hence that if A is the induced integral operator. it follows that 1.1 ~~. dx ~ f 2 2 f1K(x. the proof that A is quasinilpotent is complete. this can happen only if).1 2 ~~. is an eigenvalue of V with associated (non-zero) eigenfunction f. then g is a monotone differentiable function and g'(x) = lf(x)l 2 almost everywhere.. i.

SOLUTIONS Corollary. There is no conceptual difficulty in filling in the steps. but is also Hermitian. with corresponding eigenvalue 1/(k + !)2 n 2 . . Since V is compact (Problem 173). is an eigenvector of V*V. The largest of these eigenvalues is the one with k = 0. and solving the resulting differential equation. is given by 1 -X K(x. so that its eigenvalues are positive. with the eigenvalues as described above. 1 foxJ(y)dy. differentiating (twice. the Fredholm alternative (Problem 179) implies that 0 is the only number that can possibly be in the spectrum of V. (Note that V*V is positive. y) = { _ 1 y It follows that (V*Vf)(x) = ff(y)dyX 1 ifO ~ y ~X~ 1. K say. IIVII = 2/n. and each ck .max(x. This suggests that the eigenvalues of V* V can be explicitly determined by setting V* Vf = A. The reason this is promising is that V*V is not only compact (as is V). A simple computation shows that that kernel.f < < 0 =x<y=. y) = 1 . whenever f E L 2 (0.) Since V* is given by (V*f)(x) = f f(y)dy. The outline above shows how the eigenvalues and eigenvectors can be discovered. It follows that V*V is diagonal. Every Volterra operator is quasinilpotent. 300 fork = 0. and that the ck's form an orthonormal basis for L2 . to get rid of all integrals). then the ck's form an orthonormal basis for L 2 . it is easy to find the integral kernel that induces V* V. PROOF. PROOF. possibly larger than any of the ones that go with the ck's. The second step is necessary in order to guarantee that V*V has no other eigenvalues. Poulsen. The first step is routine computation. If all that is wanted is an answer to the question (how much is IIVII ?). A direct attack on the problem seems not to lead anywhere. The outcome is that if ck(x) = ~ (ein(k+t)x + e-in(k+tl). 1. it is enough to verify that the ck's are eigenvectors of V*V. This elegant proof is due to Frode D. Here is a not unnatural indirect attack: evaluate II V* VII and take its square root. 1).fyf(y)dy for almost every x. the obvious way to find its norm is to find its largest eigenvalue. 188 Solution 188. · · · .f. 2.

301 . moreover. 1.x) whenever x E (0. 1). where g(x) = f(x) and h(x) = f(. where f(x) = h(x) + g(x) and f( -x) = h(x) . 1) EB L 2 (0. 1). . and it shows. ± 1. that V/ = 0. These two remarks imply that V 0 is nilpotent of index 2. 1) EB L 2 (0. -2. 2. but it is not the best thing to do. One natural way to identify L 2 ( -1. The most illuminating remark about V0 is that its range is included in the set of all odd functions in L 2 ( -1. . This gives something. 0). It is easy to verify that U is a unitary operator.. then V0 f = 0.) The second most illuminating remark (suggested by the first) is that iff is odd. then Uen = c 2 n for n = 0. and hence that the spectrum of V0 consists of 0 only. 1). From this form the matrix of V0 can be read off. it follows that if x E (0. that I V0 ll = 211 VII. This proves. 1) is to map f onto (g. ±2. and Uen = c-(2n+l) forn = -1. One way to try to find the norm of V0 is to identify L 2 ( -1. 1) is more pertinent. For each fin L2 write (Uf)(x) = fi (f(x)einx/2 + /(1 _ x)e-inx/2). For present purposes another identification of L 2 ( -1. h). + 1) with L 2 (0. h) onto f. The inverse map sends (g. 189 PRooF.. it is (2~ ~). then (V0 f)(x) = 2(Vh)(x) The conclusion can be expressed in the form V0 (g. ···. ·. where + f( -x)) whenever x E (0.f( -x). -3. If n = 0. determine the two-by-two operator matrix of V0 corresponding to such an identification.. IIVoll = 4/n. 1) EB L 2 (0. again. Solution 189. (Recall that f is even if f(x) = f( -x). and ~V0 f)( -x) = -2(Vh)(x). and f is odd if f(x) = .EXAMPLES OF COMPACTNESS Here is a way to prove that the ck's form a basis. h) = (2Vh. it is the one that maps f onto (g. and hope that the entries in the matrix are simple and familiar enough to make the evaluation of the norm feasible.g(x) g(x) = !(f(x). Since CVof)(x) = 2 ft<f(y) + f(-y))dy. + 1) with L 2 (0. 1). h). + 1) with 2 L (0. 1).f( -x)) and h(x) = !(f(x) whenever x E (0. + 1). spec V0 = {0}. .

. and hence that Mn c: M. it follows that spec(1 + V) = {1}. namely the largest index n (or oo if there is no largest) such that the Fourier coefficient (f. Since spec(1 + V) = {1}. 191 Solution 191.. is to prove that A. Anf are linearly independent.. n. then.. it follows that spec A= {1}.1 is bounded from below by 1. by the preceding paragraph. f)+ (V*f. II f II for all f. · · ·. This is true and already known. everything except the inequality IIAII . 1. that the real part of Vis positive)... · · ·. it is sufficient to prove that Mk c: M for all k. then M = H. reason alone does not seem to be enough. the operator V + V* is. Since Ai E Mn. the projection onto the (one-dimensional) space of constant functions (see Problem 188). f) . f + Vf) = 11!11 2 2 + (Vf. To prove that the example works. if A = PROOF. then M contains e0 • Assertion: the lemma implies the theorem.. and hence that the definition of A makes sense.. it follows that II A II . Consider the following lemma: ifM is a subspace invariant under A.. If they are... but it is the sort that takes either inspiration or experience to produce. Suppose that f EM and that deg f = n < oo.. n. begin by recalling that spec V = {0} (cf. and if M contains a vector of infinite degree. A'i} is Mn. · · · . Let {(l(n} be the weight sequence. f)+ IIV/11 2 . 1 is obvious.. /11 2 = 11(1 + V)/11 = (f + Vf. and (1 + v)-\ then spec A= {1} and JIAII = 1. the point is that the non-vanishing of the (l('s implies that deg A if = n . then M c: Mn. 1)( 1 . i. Since IIA 1 .. and if their maximum is n. The only remaining case is the one in which M contains a vector of infinite degree.. 1)( 2 . Mn c: M for infinitely many n. If M contains vectors of arbitrarily large finite degree. The degrees of the non-zero vectors in Mare either bounded or not. In precise language. · · ·. Since r(A) = 1. so that 1 + V is invertible.SOLUTIONS 190 Solution 190. 0 (i... The example is simple. the proof is induction on k.e. It is easy to see that the vectors . Each non-zero vector fin the given Hilbert space H has a "degree". so that 1)(0 . · • ·.. en) is not zero. the operator A is given by Aen = (l(n_ 1en_ 1 when n > 0 and Ae0 = 0. If V is the Volterra integration operator. Clearly A =I= 1. This settles all properties except one.. One way to prove that II Af II . 302 .. The idea of the proof is that nothing changes if the first few terms of the basis are omitted. it follows that the span of {f. i = 0.e. It remains to show that if M is a subspace invariant under A and if the degrees of the non-zero vectors in Mare not bounded... that A is bounded from above by 1.. it is sufficient to prove that ((V + V*)f. and the preceding paragraph implies that M = Mn. in fact. To prove this. (l(n =I= 0. and hence M = H. i = 0.i.. Problems 186 and 88 or else Problem 187)... and L:'=o ()(/ < oo.

e._ 1 .Lis invariant under Ak. Note also that if i ~ n + 1. __ L.1 · · · IY. ek+l> ···}).. Iff = L~o ~..n.. let Pk be the projection onto M/.-nei-n· i=n If n is such that ~n #.EXAMPLES OF COMPACTNESS The initial step is the lemma itself. and.0.the lemma is applicable.satisfying exactly the same coaditions as A on H and since the image under P k of a vector of infinite degree in H has infinite degree in Mk .·)2 -'i=n IY. first choose k so that L rt.!: i ~ k}.-1 IY.-n IY... i= n+ 1 ~ -._ 2 ~ a.. then I~R. i=n+ 1 00 = L (l/. IY./ i=k 00 < e2rt. ·IY.-1 · · · IY. if i ~ n. The induction hypothesis implies that Pk f EM for all f.n-1'''/Y.o )2 e2. Suppose now that Mk c M.. ~i IY. Turn now to the proof of the lemma. Since Ak is a weighted shift on M/ (with respect to the orthonormal basis {ek.IY.... then 00 A".L.o ~n ei-n· It is sufficient to prove that for each positive number e the integer n can be chosen so that llfnll <e.[ = L ~. Suppose that f EM and deg f = oo. ekE M. ···}). I ~ 1.1 ~max{ I~. and let Ak be the operator on M/ defined by Akf = Pk Af for each fin Mk . and hence that M n Mk .. ek+ 1. To do this.0.o < 303 .o2.o L: 00 ( IY. and the derivation of the theorem from the lemma is complete. Conclusion: 1Jf.i-1 . and then choose n so that n ~ k and so that 1~.-n)2 ~n IY. in particular. then where f.-1 · '· IY. then IY. whence Mk c M).-n ~ rt.L (with respect to the basis {ek. With this choice. 1 (here is where monotoneness is used). ~" #. • • ·. The conclusion is that M n M/ contains ek (so that.IJ2 = ~ I i=n+1 I ~i 12(/Y.

via operator matrices. as always.CHAPTER 21 Subnormal Operators 192 Solution 192. and (multiply the matrices A* and Bin both orders and compare corresponding entries) this implies the desired conclusion. and the proof of the corollary is complete. and if UP is its polar decomposition. Write ~ A = (A0 Az ) 0 1 ~ and B = (0 B) . and a straightforward verification proves that B commutes with it. there is a neat derivation. a variant is in [50] or 304 . The corollary takes a little more work. it follows that A1 U = UA 2 . Alternatively. If B is invertible. The Fuglede theorem implies that Bcommutes with A* also. 0 0 The operator A is normal. The original proof is in [ 45]. then U is unitary and Pis. of the statement for two normal operators from the statement for one. If A 1 and A 2 are normal and A 1 B = BA 2 . then AiB*B) = (A 2 B*)B = (B* A 1)B = B*(A 1 B) = B*(BA 2 ) = (B*B)A 2 . so that it follows that A 2 P = PA 2 • (Compare Solution 137.) Since A 1 UP = UPA 2 (by assumption)= UA 2 P (by what was just proved). Every known proof of Fuglede's theorem can be modified so as to yield this generalized conclusion. There is a breathtakingly elegant and simple proof of the PutnamFuglede theorem in [118]. the positive square root of B*B.

. Reference: [110]. If xlfJ-'<Dlf = f. there is no loss of generality in assuming that ker A = 0 in the first place. if C = (A + B)*(A + B) .1(D). and therefore {x: f(x) ¥= 0} c cp -t(D). then {x: f(x) ¥= 0} c cp. in that case. and if M.SUBNORMAL OPERATORS [56]..'(D) lcpYI 2 d. By Fuglede's theorem. since quasinormality implies that A* commutes with A* A. and (by Problem 137) UP = PU and 305 . The ingenious matrix derivation of Putnam from Fuglede is due to Berberian [9]. Solution 193.(A + B)(A + B)* and D = (A + iB)*(A + iB) . d. Reason: ker A = ker A* A for every operator A. The proof is based on a computational trick and Fuglede's theorem. and therefore 194 IIAYII 2 = J1cpYI 2 d.u ~ l r nlfl 2 M. then A + Band A + iB are in M. Multiply everything out and infer that 2(B* A . and if A and Bare in M.u If. then each pair of operators in the collection is commutative. then 11/11 2 ~ JlcpYI 2 d.e. that B* commutes with A. If. UP is the polar decomposition of A. then U is an isometry. The answer is surprising: if a collection of normal operators 193 is a vector space. Since the direct summands can be treated separately. It is convenient to begin with the observation that if A is quasinormal.(A + iB)(A + iB)*. and therefore both A + B and A + iB are normal.AB*) = 0. i. = {x: lcp(x)l 2 ~ r > 1}.u. IIAnfll ~ 11!11 for all n. Conclusion: f vanishes on M" for every r. The assumption that the collection is an algebra (is closed under multiplication) is irrelevant. 195 In view of the preceding paragraph every quasinormal operator is the direct sum of 0 and an operator with trivial kernel. and therefore C + iD = 0. then C = D = 0. In other words. conversely. Unless f vanishes on M.u = f lfJ. B commutes with A. it follows that ker A* A is invariant under A*. then ker A reduces A. the two-operator generalization first appeared in [109]. Solution 195. If M is a vector space of normal operators. Solution 194. the last written integral becomes infinite with n.

fj and gi in H). then (1 . that therefore U has a unique isometric extension that maps K 1 onto K 2 . (If H is identified with H EB 0. j k j k This computation accomplishes much more than the proof that U is unambiguously defined. Similarly. 2. Since H itself is included in M 1 . then AE(M) = E(M)A for every Borel subset M of the real line: this is the trivial. Ajk). and the desired conclusion is obviously true. A can be shown to be subnormal by explicitly constructing a normal extension for it. Hermitian. then there exists an M with E(M) i= 0 and E(M) i= 1. see Problem 149. then Vis unitary.fj = Li B1 *igi (with . This works. in that case the proof is over. It is tempting to try to complete the proof by setting U(Li B1 *1}) = Li B 2 *1J. the closure of M 1 reduces B1 . The isometric character of U implies that if E is the projection UU*. then B is Hermitian and AB = BA. · ·-). V = (u 1. special case of Fuglede's theorem. 197 Solution 197. the set M 2 of all finite . and 306 . If E is the spectral measure associated with B.fj = Li B 2 *jgi? Equivalently (subtract): if Li B1 *'ij = 0. The set M 1 is a linear manifold. 1.SOLUTIONS U* P = PU*. is in H. it implies that U is an isometry (from M 1 onto M 2 ).E)U = U*(1 -E)= 0. of course. t Bt*kh) L L (B/jj. If B is not a scalar. In view of these algebraic relations. If B is equal to the (positive) scalar /3. is dense in K 2 . does it follow that L B 2 *i. since 1 BtC[. 196 Solution 196. and therefore = VQ = (UP 0 (1 U*P E)P) is a normal extension of A.fj) and Bt*Cl:iB 1 *'iJ) = LiB 1 *i+ . where . sums of the form Li B 2 *1}. in particular. the minimality of B 1 implies that K1 = M 1 .) An operator on H EB H is given by a two-by-two matrix whose entries are operators on H.iBt*jj) = LiBt*i(B 1 . then a normal extension B can be constructed that acts on H EB H. then A/ J73 is an isometry. where each/.E) 0 U* B and Q = (P o) 0 p . Btfk) = L L (A~. V and Q commute. First: does this equation really define anything? That is: if Li B1 *i. Q is positive. If. the reason is contained in the following computation: 1 ~ Bt*tr = = (~ B~*t. but it takes a little care.fj. then His a subspace of H EB H. does it follow that Li B 2 *'lj = 0? The answer is yes. Let M 1 be the set of all finite sums ofthe form Li B1*-'l}. If A acts on H.fj E H for all j ( =0. If A is quasinormal and B = A* A.

·). however. defined by Wf(z) = zf(z). so that every polynomial is in M. Let B be the position operator on L 2(v) (that is. 3. Suppose now that a polynomial q of degree n (~I) is a power of p. It suffices to verify that UB 1 agrees with B 2 U on M 1 .. and this is implied by UB1(~B1*1J) = u(~B1*jB1JJ) = ~Bz*jAJ} = ~ Bz*jB J 2 fj = B 2 ~ B *1} = 2 .jej. Clearly B is normal and A is subnormal. and let A be its restriction to the closure H 2 ( v) of all polynomials. The operators p(W) and p(W)* are multiplication operators induced by the function p and its complex conjugate p*. In this context it is most convenient to consider Was the multiplication operator on L 2 of the unit circle. PRooF. see [104].. 3.I B 2 u(z. it is to be proved that if M is a subspace of L2 such that H 2 c M and M reduces p(W). it follows that q*en-1 eM. In other words. 1. assume (with no loss of generality) that q is monic.SUBNORMAL OPERATORS that U is the identity on H. H 2 c M. The result is a special case of a general theory of minimal normal extensions. 2. and U as its restriction to H 2 (see Problems 82 and 84).It follows from Problem 90 307 . The action of A on this basis is easy to describe: Ae0 = (1/ j2)e 1 and Aen = en+ 1 for n = 1. . 3. Consider the measure space consisting of the unit circle together with its center.-J r:t. · · ·. There exist two subnormal operators that are similar but not unitarily equivalent. it follows that e _ 1 eM. Since en-1 eM. A is a unilateral weighted shift.. so that q* eM. The invariance of M under p(W)* and the presence in M of e0 imply therefore the presence in M of the complex conjugates of polynomials of arbitrarily high degrees. and that is right. then M = L 2 . together with the function e0 . Solution 199. 1. (Bf)(z) = zf(z)). The proof that UB 1 = B 2 U is another computation. 2.. · · ·. An orthonormal basis for H 2 (v) consists of the functions 199 en (n = 1. ···}. J B 1 *~)· 198 Solution 198. Clearly p(W) is a normal extension of p(U). 1. The natural guess is p(W). defined by e0 (z) = 1/ J2. since. An inductive repetition of the argument shows that e-n eM for n = I. with weight sequence {1/fi. with measure v defined so as to be normalized Lebesgue measure in the circle and a unit mass at the center. so that q =en+ L}. 2. and hence that M = L 2 . where W is the bilateral shift. defined by en(z) = z".

1). Since A ELl and Ll is a hole of spec B. that). from this in turn (see Problem 194) it follows that B is invertible. and write E = {f: IIBnfll ~ enllfll. and ifjE E and g E H. (f. is to observe that U is an isometry and A is not. E . It is to be proved that if A is a complex number such that B . but it is quite easy to prove.1 1 = 1. Let e be an arbitrary number in the open interval (0. To prove this. fixed from now on. Assertion: Ll + is also open. however. By simple geometry (translate) and equally simple logic (form the contra positive). One way is to recall that two unilateral weighted shifts are unitarily equivalent only if corresponding weights have equal moduli (cf. consequently.A is not invertible. 200 Solution 200.is open. From this it follows again that U and A are not unitarily equivalent. n = 1. of course.SOLUTIONS (but it is just as easy to verify directly) that A is similar to the ordinary unweighted unilateral shift U. The proof depends on the trivial spectral inclusion IT(A) c IT(B) only.AnA-ng)l = l(f. This is not obvious at a glance. hence that A is not in IT( A). the simplest way.l H. This example is due to D. is not in IT(B). so that E = 0. Since Ll is open and spec A is closed. In other words.L. Since. Suppose therefore that A is invertible. Sarason. the point A cannot belong to spec B.g)J = l(f. ···}. the set Ll. Problem 90). however. The conclusion holds for a pair of operators A and B whenever their spectra and approximate point spectra are so related. consider an arbitrary point A in Ll +. Reason: their minimal normal extensions are not. A E Ll + and Ll + c spec A. There are several ways of proving that U and A are not unitarily equivalent. 3. and hence that A is not on the boundary of spec A (see Problem 78). the inter308 .= Ll . no deeper or more special properties of subnormal and normal operators are needed. Consider the sets Ll. and. it follows that E. This argument proves that Ll + is. without loss of generality normalize so that IIA. A-ng)l ~ IIB*nfii·IIA-ngll = IIB'ill · IIA -ngll ~en· 11!11 ·llgll for all n.If Hand K are the domains of A and B. in fact.BnA-ng)l = I(B*'i. It is worth noting that B is the minimal normal extension of A (see Problem 197). then l(f. it follows that the only place A can be is in the interior of spec A. then neither is A . the assertion reduces to this: if A is invertible.A.spec A and Ll + = Ll n spec A. g) = 0.L = K. E. and therefore H c E. Since (Problem 194) Eisa reducing subspace for B. then so is B. 2. 201 Solution 201. This implies.

and. If B (on K) is a normal extension of A (on H). Parrott. as asserted. that A+ is open.and A+. Reason: if B is a normal extension of A to K. This is easy: in fact each eigenvector of B is an eigenvector of B* too. The restriction of a normal operator to a reducing subspace is normal. therefore. Since A was assumed to be non-normal. Solution 202. H reduces B.)fll = II(B*. On finite-dimensional spaces every subnormal operator is normal.) = I I j i j i 309 . i = L: L: (B~.B*Yj. then IILi B*'lJII . The difficulty is to prove that something is not subnormal.A. if dim(K f'\ H. B*'l. PRooF. then.L is invariant under the normal operator B*. Since on a finite-dimensional space every operator has an eigenvalue. · · · .) J ' J ' = L L (B. and it follows. The simple proof above is due to S. what is wanted here is a non-existence theorem. 202 0 = II(B. From the result thus proved it follows that the answer to the dimension question is no. Since subnormality was defined by requiring the existence of something. Bij.*)fll.f. Every finite-dimensional subspace invariant under a normal operator B reduces B. and if fo.SUBNORMAL OPERA TORS section of A with the interior of spec A. as follows: 0 203 ~ (~ B*Yj. then K f'\ H. K. 0. this is impossible.:::.). Aij.A. are vectors in H. by normality. for a generalization see [77]. This triviality can be rewritten in a non-trivial way. it is sufficient to prove that each one-dimensional invariant subspace of B reduces B. Since A is the union of the disjoint open sets A. f. The result is due to Bram [16]. the connectedness of A implies that one of them is empty. Ji) j (because B is normal) (A~. Solution 203.) Corollary.L) is finite. PROOF. j fi) i = L L (B*iB~. derive a usable "constructive" necessary condition from it (with luck it will be sufficient as well). (If Bf = A. ~ B*~~) = ~ L (B*'lj. and then look for something that violates the condition. The best way to prove such a theorem (the only way?) is to assume existence.

U is normal. 1. · • ·}. A finite amount of experimentation might lead to the weighted shift S with weights {a.SOLUTIONS Replace each Ji by some scalar multiple i. If U is subnormal. then it is hyponormal. and the diagonal of SS* is {0.. where 0 < a < f3 < 1. with weights {a0 . then the initial space is equal to the range. 1. and conclude that L L (A~. (The point is that the typical non-unitary isometry. Aij.. and hence that it reduces U. j i i. 2. the unilateral shift. · • ·}. the construction of a counterexample along these lines (if it is possible at all) should be easy. · • ·}. the corresponding matrix ( (A1j. it is pertinent to comment that the condition is sufficient. 1.i* ~ 0. and UU* is the projection on the final space (the range).) To prove "only if". e 2 . examine the matrix ((Siei. The desired counterexample can be found among weighted shifts. however. 310 . Stampfli. • ··}is the orthonormal basis that S shifts. Sie. fn. is subnormal. then the operator A is subnormal. clearly the restriction of U to the initial space is an isometry. This is a "constructive" intrinsic necessary condition that follows from subnormality. and consequently the initial space includes the range. that the finite matrix ((A1. G. it will be used to exhibit a hyponormal operator that is not subnormal. I a1 12 • I a 2 12 . the fact will not be used in the sequel.{3 2 ) 2 .:o. as well as necessary. First. I a 1 12 . I a 2 12 . there is an easy answer in terms of the a's.)i. it follows that S is hyponormal if and only if the sequence { Ian I} is monotone increasing. 1. so that U*U is the projection on the initial space (the orthogonal complement of the kernel). The preceding paragraph implies that Sis hyponormal. This implies that the initial space is invariant under U. When is a weighted shift S. moreover. The proof is somewhat involved.a 2 (1 ·. a2 . The "if" for normal partial isometrics is trivial and for subnormal ones is a consequence of Problem 149. for subnormality.)) is positive definite. With this much information available. hyponormal? Since both S*S and SS* are diagonal.ji. · · ·.e. the matrix is 1 (X [32 ( af3 a af3) f3 . which is negative. 204 Solution 204. I a0 12 .)). If. {3. where {e 0 .)) is positive definite. for each finite set of vectors fo. e 1 . To put it precisely: if.. Written explicitly. suppose that U is a partial isometry. It is interesting to note (as a consequence of the proof) that a partial isometry is subnormal if and only if it is hyponormal. Examples of this type have been studied by J.Ji. The diagonal of S* S is { I a0 12 . Ai_{.. and where i andj take the values 0. Aij. f3 1 Its determinant is . a 1 . · · ·}. To prove that Sis not subnormal. and therefore the restriction of U to the initial space is unitary.

A"f) =(A* A}. The program is to prove that the span of the eigenvectors of A reduces A. IIA*A"fii·IIA"-tfll.. A*fz) = Az(ft. In the presence of compactness the orthogonal complement of that span becomes amenable. (3) If At-:/= A2 ..A is just as hyponormal as A. (1) For each complex number A. Solution 206. The reason is that A . 136].. proceed by induction.. . and that. Indeed: in variance under A is trivial. For n = 1.. IIA"+tii·IIA"-tll·ll/11 2 for every vector f. Since IIA"fll 2 = (A"f. the subspace {f: A! = Af} reduces A. A straightforward and often-used argument: if Aft = Adt and Af2 = A2 / then 2. Reference: [ 4.. AtCft. (4) The span of all the eigenvectors of A reduces A and the restriction of A to that span is normal. compactness does not enter here... Suppose that A is hyponormal. IIA"II 2 n). it follows that IIAn+tii·IIA"-tll· In view of the induction hypothesis (IIAkll = IIAIIk whenever 1 .. a necessary and sufficient condition that (A* ...... the equality is trivial. IIA"+tii·IIAII"-1....SUBNORMAL OPERATORS Solution 205.. on general grounds that have nothing to do with hyponormality. from which it follows that Since the reverse inequality is universal. then {f: Af =Ad} l. this can be rewritten as IIAII 2 ". fz). 311 . 206 {f: Af = Af} c {f: A*f = A*f}. the induction step is accomplished. k..A*)f = 0 is th~t (A. by (3). an application of Problem 205 will yield the conclusion... {f: Af = Azf}.. the restriction of A to each eigenspace is normal (in fact equal to a scalar).. fz) =(Aft. Proof: use (2) and observe that. (2) For each complex number A. The proof above is a slight simplification of the simple proof in [136].. fz) = Cft. and in variance under A* follows from (I). A"-tf) ..._ 0.A*)f .A)(A*..... IIAn+tfii·IIA"-tfll 205 .

Reference: [153]. ran Pis closed. 208 Solution 208.SOLUTIONS (5) Now assume that A is compact. M is a subspace. then there is a contradiction: the non-zero vectors in it both must be and cannot be eigenvectors of eigenvalue 0. Since the point spectrum of this compact operator is empty. and observe that because B IIAfll = II(B + iy)fll = II(B + iy)*fll. M then M contams every eigenvector of C. the result is of the form p It follows that = G~). in other words. and still compact.AA*. the same is true for quasinormal and subnormal idempotents). A = B + iC. = = IIAfll II(B. 312 . If A is hyponormal. 207 Solution 207. If the orthogonal complement on which all this action is taking place is not 0.iy)fll If. f)= 0 is equivalent to Df = 0. with Band C Hermitian and C compact. Conclusion: A is normal. and consider the restriction of A to the orthogonal complement of the span of all the eigenvectors. it is the same as to say (Df. Consider an eigenvalue y of C with corresponding eigenvector f. To say that IIA/11 = IIA*fll is the same as to say (A*Af. Since the eigenvectors of C span the whole space (Problem 172). It follows that IIA*/11is the set of all those vectors f for which IIAfll = IIA*fll. then A must be normal. observe first that if P 2 = P. Express Pas an operator matrix with respect to the decomposition ran P E13 ran. For a positive operator (Df. The resulting operator is still hyponormal (by the reduction assertion of (4)). consequently. where D is the positive operator A* A . f) = 0. then ran P = {f: Pf = f}. in particular. PP* = (~ A) ( 0 A* A* 1 0) = (1 +0AA * 0) ' 0 0 = P*P = ( l ~)(~ ~) (~* A~A). and. it is quasinilpotent (Problem 179). Consequence: M = ker D.f) = (AA*f). M = H. therefore.L P. in other words. so that. Every hyponormal idempotent is a projection (and. + iy is normal. For the proof. an application of Problem 205 implies that it must be 0. or.

One such vector f is e0 . the former always dominates. which does not shed any light at all. then nothing remains except numerical computation.AA* = 3. it is much simpler than the original one [ 49]. each of which is illuminating.SUBNORMAL OPERA TORS If now P is assumed to be hyponormal. and A*f = (2~1• ~o + 2~3• ~2 Each of ~b 2~ 1 .~ 1 .. What does not balance is 41~ 0 1 2 in IIA/11 2 and 1~ 0 1 2 in IIA*/11 2. + AA * ~ 1. The proof that A is hyponormal can be done in (at least) two ways. ~ 2 . then A is hyponormal but A is not. ···occurs in both Af and A*f exactly once. That implies Solution 209.UU*) ~ and therefore 0. The proof that A 2 is not hyponormal is less pleasant. their contributions to the sum of the squares of the moduli exactly balance. and the proof is complete. Numerically: since 0 0 0 2 0 1 0 0 2 0 0 0 2 0 0 2 0 0 1 0 2 0 0 0 2 0 0 1 0 A= and A*= itfollowsthatif/ = (~ 0 . Algebraically: A* A = (U AA* + 2U*)(U* + 2U) = UU* + 2U* 2 + 2U 2 + 4. 2~ 2 . This example is due to Ito and Wong [78]. = (U* + 2U)(U + 2U*) = 1 + 2U 2 + 2U* 2 + 4UU"'.~ 2 . ~ 3 .3UU* = 3(1. If U is the unilateral shift. The quickest way is 2 to exhibit a vector f such that I A 2/ I < I A * ! 11.then + ~2• 2~1 + + 2~2• ~1 ~3. ·) + 2~4• · · -)..2e 2 • Once that is said. and A= U* 2 209 + 2U. A*A. 2~2 Af = (~1• 2~o + ~4• .· . then 1 A = 0. The answer is that IIA 2/II 2 = 80 and IIA*2/II 2 = 89. 2~ 3 . ). 313 .

y = (Cf. f) = 0. al~l 2 1 and a I~ 1 + P111 1 = 0 is connected. If A = B + iC. Assume (with no loss) that Lis defined on C 2 by a diagonal matrix (~ I~ 1 2 ~} The desired conclusion now is that the set N of all 2 <~. then the set of all unit vectors f such that (Lf. f) with 11/11 = 1 and ((pB + qC + r)jJ) = 0. As I~ I goes from + Pl11l 2 goes from Pto a and can be equal to 0 at most once. (The ingenious idea of basing the proof on this observation is due to N. An efficient way to conclude the proof is to show that N itself is connected. f)+ i(Cf. Dekker [35]. f) = 0 is connected. y) is identified with the complex number x + iy). with Band C Hermitian. f) of the set N of all those unit vectors f for which ((pB + qC + r)f. Step 1. 1'/) with + II'/ 1 = 0 to 1. then the intersection in question consists of all those complex numbers x + iy for which px + qy + r = 0 and for which there exists a unit vector f such that x = (Bf. and is easy to prove directly. 2 2 314 . a slightly indirect two-step process gives more geometric insight.) Consider a line with the (real) equation px + qy + r = 0 (where the pair <x. it is sufficient to prove that the intersection of W with every straight line in the complex plane is connected. With superfluous notation discarded. In other words. P.CHAPTER 22 Numerical Range 210 Solution 210. To prove that the numerical range W of an operator A is convex. f) + i(Cf.f). the desired conclusion is just this: if L is a Hermitian operator. Step 2: reduce the general Toeplitz-Hausdorff theorem to its 2-dimensional special case. Step 1 : prove it for the special case when the dimension of the underlying Hilbert space is 2.f). There is still another way of describing the intersection: it is the image under the (continuous) mapping/ r---+ (Bf. the intersection is the set of all (Bf. Although this general statement is true.

gi)· 315 . follows by symmetry). For an alternative proof see [30]. · · ·.. and the desired conclusion degenerates to a triviality. f) = (Ag. k). that i # j. Assertion: the subspaces L. (If the span off and g is 1-dimensional. and T*g. For each i ( = 1. k.) Apply the Toeplitz-Hausdorff theorem (2-dimensional case) to the operator PAP acting on the space K to infer that to every complex number z on the segment joining (Af. g) ( = (PAPg. ak such that T/.. · ·. !. let UP be the polar decomposition of T.. gi) = (a. . it can be replaced by one (since dim M =dim N = k). k. To prove this. then the preceding lemma is applicable.g. .1 Jj and g. . · · ·.. · · ·. gk} for N.1 g.NUMERICAL RANGE To get the most general element of N.1 gi. fk} for M and {g 1 . with respective ranges M and N. I = 1. h)). • · ·. The proof is easy: (/. i = 1. That is what was wanted-plus the extra information that h is in K. g)) there corresponds a unit vector h (in K) such that z = (PAPh. i = 1.1 gi (for then jj . and reap the consequences: PROOF. Suppose that P and Q are projections of rank k.. = a. h) (=(Ah. Suppose. If T is the restriction of QP to M. 211 Suppose to begin with that M and N are k-dimensional Hilbert spaces and that T is a linear transformation from Minto N. Step 2. = U/. g). and therefore clearly connected. the Cartesian product of two copies of the unit circle. From this point of view N is a torus. multiply the two coordinates of the unique positive element by any two complex numbers of absolute value l. = P /.. indeed. f)) to (Ag. k. and T*g.. let L. are pairwise orthogonal.} for M and find positive scalars a 1. it is sufficient to prove that /. and there exist positive ( ~ 0) scalars a 1. = a. There is a useful sense in which T and T* (from N into M) can be simultaneously diagonalized./. and diagonalize P.g with IA. whence (Af. = a. • · ·./. be the span of/. and g. ak such that P/. That is a lemma. assume that that has been done... · · ·. then f = A. Solution 211. since/. The assertion is that there exist orthonormal bases {f1 . = a. Qg) = (QP/. If the partial isometry U is not an isometry from M onto N.g. Then put g. = PU*g. Suppose that A is an operator on a Hilbert space H and f and g are unit vectors in H. For every operator A and for every positive integer k. gj) = (Pfi. i = 1. · · ·./.(A) is convex. That is: find an orthonormal basis {/1 . the k-numerical range W. f) ( = (PAPf.. Let P be the projection from H onto the 2-dimensional space K spanned by f and g. now for the theorem. · · ·.

h. g.. This shows. The first solution. f) with llfll = 1. does not belong to W(A). E W(A). hk} is an orthonormal set..t)(Ag. then IIAII = sup{ IA. fi) i + (1 . then its numerical range will not be closed..) i = L: (Ah.t) · tr QAQ = t ·I (Afi. somewhat more complicated than the one above. then A. Another example.) If A is normal.. then (Af.f. f) = A.. 1]. In view of this comment it is easy to construct examples of operators whose numerical range is not closed.0 (!.. The problem was raised in [57].£ IIA II. so that (Ah. in W(A) such that IA. so that there always exists a A. then A is an eigenvalue of A. PROOF. is not an eigenvalue of A (and. t. in fact W(A) = (0. Since {h 1.Since A ~ 0 and ker A = 0. the projection R onto its span has rank k. 316 . It follows from this theorem that if A. i l'lhe proof of the theorem is complete. is a number in W(A) such that lA-I = IIAII and A. · · ·. I = IIA II. Berger. if A has no eigenvalues).. so that equality holds everywhere.) = t(Afi. I:A.£ IIAfll · IIIII . If 0 . in particular. (1) Observe that the eigenvalues of every operator A belong to W(A). in L. It follows that if a normal operator has sufficiently many eigenvalues to approximate its norm. use the classical Toeplitz-Hausdorfftheorem k times to obtain a unit vector h. Concrete example: a diagonal operator such that the modulus of the diagonal terms does not attain its supremum.0 /. but does not have one whose modulus is as large as the norm.£ 1. t. A.f) =A. If A = (Af. ···}. f) = (A.) + (1 .0 f for some A-0 . so that A is an eigenvalue of A.t) ·I (Ag. E W(A)}. that the numerical range may fail to be closed even for compact operators. then IIAII = lA-I = I(Af. g.). it follows that 0¢ W(A).0 = A. If A is an operator and A is a complex number such that lA-I = IIAII and A.. 212 Solution 212.SOLUTIONS The desired convexity proof is now near at hand. h. is due to C. (Proof: if Af =)/with llfll = 1.:. and this in turn implies that A. and t · tr PAP + (1 . along slightly different lines: take A to be the diagonal operator with diagonal {1. f) = (Af. t .f)l . The known facts about when the Schwarz inequality becomes an equation imply that Af = A.) = tr RAR. by the way.

Solution 213.} is a bounded net weakly convergent to f. if {f. f) = (Af. Conclusion: the 214 numerical range includes the compression spectrum. it follows that the open unit disc is included in W(A *). then W(A) is closed PROOF. f. the theorem proved above implies that W(A) cannot contain any number of modulus 1. Since every number in the open unit disc is an eigenvalue of A*. Solution 214. If A is compact and 0 E W(A).AI = I((A. Indeed. f) = t 2 (Af. If A is in the compression spectrum of A. so that W(A) is the image of a weakly compact set (the unit ball) under a mapping that is weakly continuous on that set (by the preceding paragraph). and the numerical range is convex.fn such that (A . tf) = t 2 (Af. and the second summand tends to 0 because .NUMERICAL RANGE (2) Take A to be the unilateral shift.. it follows that (Af. The next step is to show that in the presence of compactness the quadratic form f~ (Af. it follows that the open unit disc is included in W(A).fn --+ f weakly. f): llfll ~ 1} (by the first paragraph above). then (Af. then there exist unit vectors . These two paragraphs complete the proof. Since W(A *) is always (W(A))* (proof: (A *f. 317 . 0 E W(A). f) E W(A) for every vector fin the unit ball (and not only for the unit vectors). so that W(A) is equal to the open unit disc. A slightly different proof can be obtained by combining the fact just proved for the approximate point spectrum with two other facts: the boundary of the spectrum is included in the approximate point spectrum. f) is weakly continuous on bounded sets. finally. f) + (1 - t 2). Since. Reason: if llfll = 1 and 0 ~ t ~ 1. then 213 (A((f). If A is in the approximate point spectrum of A. Conclusion: the closure of the numerical range includes the approximate point spectrum. The argument of the preceding paragraph did not need compactness. Since I(AJ. it is valid for every operator.. f. then the first summand tends to 0 because (by compactness) {AJ. and therefore A E W(A). by convexity. Reference: [31]. so that A* E W(A*).ll. If both hypotheses are satisfied (0 E W(A) and A is compact).} is strongly convergent.A)j. f)*). --+ 0..A)f. f. Observe to begin with that if 0 E W(A). then A* is an eigenvalue of A*.) --+ A. then W(A) = {(Af.)l ~ II(A .). A has no eigenvalues.A)f.

/. PROOF. If V is the Volterra integration operator and if A= I -(I + V).1 = {1}. if dv = If /2 dfl. on a measure space with measure Jl. Since spec(1 + V). is included in the right half plane).) Use the spectral theorem to justify the assumption that A is a multiplication. In view of the characterization of convex hulls in terms of half planes. The closure of the numerical range ofa subnormal operator is the convex hull of its spectrum. that assumption enters in the proof of the reduced statement. for instance the right half plane.f.1). then (Af. f) = 0.)/f/ 2 dfl. finally. then f = 0.1 .f/ 2 d11 = J(Re q. It remains to prove the reverse inclusion. induced by a bounded measurable function q. then it includes W(A). then 0 ~ Re Jq. 217 Solution 217. If A is replaced by aA + f3 (where a and f3 are complex numbers). the reduced statement says that if 0 ~ Re q.f) ~ //(1 + V). (Observe that the reduction to this point did not use normality. Since W(A) is convex and spec A c W(A) (Problems 210 and 214). If .) It follows (by what is known about when the Schwarz inequality degenerates) that f must be an eigenvalue of (1 + V)./. (A/. the proof is complete.SOLUTIONS 215 Solution 215. it follows that conv spec A c W(A).1 (= V(l + V).1 has more than one application. 318 . This implies that f = 0 (see Problem 188). it follows that (1 + V) 1f = f. then A is quasinilpotent but W(A) does not contain 0. is obvious. then //f// = ((1 2 + V) 1 f. This remark makes it possible to" normalize" the problem. then spec and W are replaced by a spec + f3 and a W + {3. 216 Solution 216. This. then v is a positive measure. If.1 //·//f// 2 = //f// 2 • (See Solution 190. or f = (1 + V). then the same is true of the numerical range of A. In these terms. Its effect is to reduce the problem to the study of any one particular half plane. f) = 0. The desired result now is this: if every number in the spectrum of A has a positive ( ~ 0) real part. and the assertion is just that the integral of a positive function with respect to a positive measure is positive. almost everywhere (this says that the essential range of q. f) = Jq. or Vf = 0. Since the quasinilpotence of A is obvious (Problem 186). Observe that the operator A is compact.f E L 2 (J1).f /2 dfl. indeed. it is sufficient to prove that iff is a vector such that (Af. The trick of considering (1 + V) . Suppose that A is a normal operator. the desired result can be formulated this way: if a closed half plane includes spec A.

fn) ~ l. and. Next write A=(~~)· 319 . so that 1 is an approximate eigenvalue of A. so that l E spec(l . (a) If A is not invertible. f.A). (Multiply by a suitable positive constant. (b) Assume. then 0 E spec A. If A=(~~). and let N be a normal operator whose spectrum is the closed disc D with center 0 and radius !.) The hypothesis w(A) = I A I then guarantees the existence of a 218 sequence {f. then spec A = {0} u D = D..) Since I(Afn.) I ~ l. Solution 218. 219 PROOF. If A is subnormal and B is its minimal normal extension (see Problem 197). Write M = (~ ~). It follows that W(B) = conv spec B (Problem 216) c: conv spec A c: W(A) (Problems 210 and 214) c: W(B). with no loss of generality.} of unit vectors such that I(Af.fn) + 1 ~ 0. it follows that IIAfnll ~ l. Solution 219.fn)l ~ IIAfnll ~ land (Afn. as a corollary of the proof.fn) ~ l. that the closure of the numerical range of a subnormal operator is the same as the closure of the numerical range of its minimal normal extension. and W(A) = conv(W(M) u W(N)) = D. trivially. (Multiply by a suitable constant of modulus l.A) ~ w(l . then spec B c: spec A (Problem 200). it follows that l ~ r(l . assume with no loss of generality that (Afn. W(A) c: W(B).A). This implies that IIAf~- fnll 2 = IIAfnll 2 - 2 Re(Afn. Note. that IIA I = l. This shows that A is convexoid. and hence that all the sets that enter are the same. and therefore r(A) must be equal to 1. A is not normaloid. Since IIAII = 1 (in fact II Mil = 1).NUMERICAL RANGE PROOF. There exist convexoid operators that are not normaloid and vice versa.

(The proof is due to A.zA)f. The paper contains a small error. consider the unilateral shift U. f)l <a. but w(U*n) = 1 for all n. and iff is a unit vector.Re(za) ~ 1 . If II A . W(B) c W(A) + (a). in particular. 220 Solution 220. Many of these concepts were first studied by Wintner [159].tlal) ~ 0. If conversely the complex number a is such that Re(1 .) As for the second assertion. spec A = {0} u {1}.Bf. Indeed. therefore. the following assertions about A are pairwise equivalent: w(A) I(Af. PROOF. f) E W(B) + (a). 221 Solution 221. f) = (Bf. then this is true. and therefore 1 . 0 < t < 1. it asserts that every normaloid operator is convexoid.zA).f) ~ 0 If w(A) 1 = 1 and lzl < 1.zA) ~ 0. it follows that w(A) = 1 and hence that A is normaloid. however. whenever llfll = 1.B II < a.zA is invertible whenever z I < 1.SOLUTIONS Since IIAII = 1. The function W is continuous with respect to the uniform (norm) topology. then r(A) ~ 1. I .za) ~ 0 for each z with Iz I < 1. since. if za = t Ia I. The operator fact corresponding to (and implied by) this numerical fact is that w(A) ~ 1 if and only if Re(1 .B)f. then I((A . Brown. then itfollowsthatw(A) ~ 1ifandonlyifRe(1. and W(A) = conv(D u {1}). so that conv spec A is the closed unit interval. it follows (lett tend to I) that Ia! ~ I.tlal = Re(l . ~ 1. f) I~ 1 ~ 1.1 ~ Ointheunitdisc.A is not convexoid. The sequence {V*n} tends to 0 in the strong topology (more and more Fourier coefficients get lost as n increases). If a is a complex number with lal ~ 1 and if lzl < 1. symmetrically.Iz I > 0.za) = 1 . and therefore (Af. if the underlying Hilbert space is infinite-dimensional. Since an invertible operator has positive real part if and only if its inverse has positive real part (if B is invertible. Since. then the function w is discontinuous with respect to the strong topology (and hence with respect to the weak). then Re(1 . f) It follows that W(A) c W(B) + ((A . This proves the first assertion. + (a). whenever llfll (Re(1 . 320 .

z".1 1 for all z other than the powers of w. so that q> 1 (A) and q> 2 (A) make sense. q>z(A. The equivalence of w(A) ~ 1 and Re(1. 321 . That idea is visible in some form in all subsequent proofs. For a direct verification.z" = n k~o 1 . Since each summand on the right side has positive real part (because w(wk A) ~ 1). then 1 n-1 (1. observe that the right side becomes a polynomial of degree n . Explicitly: if q> 1 and q> 2 are rational functions whose poles are not in the spectrum of A. q> 2 (A)). then 1 . This identity is.)). One step in the proof might be unfamiliar enough to deserve a second look.zA).1 =(1.1 at most that is invariant under each of then substitutions (k = o.. it was Berger's main new idea. but basic for the argument. multiply through by 1 . and then evaluate the constant by setting z equal to 0. it follows that the left side has positive real part.z"A"). .NUMERICAL RANGE Observe next that if n is a positive integer and if w is a primitive n-th root of unity (i. The proof given above is a simplification of a simplification discovered by Pearcy [107].e.1) and is therefore constant. To prove an identity between operators by substitution into an identity between rational functions is to make use of the functional calculus for rational functions (cf. n is the smallest positive integer such that w" = 1).n. The proof is obvious. The identity of the preceding paragraph implies that if w(A) ~ 1..wkz 1 1 n. Problem 123).1 n k=o L whenever Iz I < 1.1 ~ 0 for lzl < 1 is elementary.) = p(q> 1 (A. if q>(A. . in fact. then q>(A) = p(q> 1 (A). the partial fraction expansion of the left side.. then the same is true of each polynomial p in q> 1 and q> 2 . and that implies that w(A") ~ 1.wkzA).).

followed by projection (back to the line). the assertion is that multiplication (on the line) by rx can be achieved by a suitable rotation (in the plane). of course. ~) -rx. J1 A* A. and.A* A are positive. and. The proof itself is the most direct possible imitation of the technique that worked for the plane. in particular. in geometric terms. consider the very special case in which the given Hilbert space H is one-dimensional real Euclidean space and the dilation space K is the plane. then each operator on K is a two-rowed matrix of operators on H. or sometimes one and sometimes another. In that case the given contraction A is a scalar rx (with jrxj ~ 1). write K = H EEl H and identify H with the first summand. Given H. (a) As a heuristic guide to the proof.AA * and 1 . The result can be described as follows. Given A. to see whether the role of rx 2 should be played by A 2 . A picture makes all this crystal clear. it follows that 1 . The desired dilation B can be defined by 322 . p = (~ ~).CHAPTER 23 Unitary Dilations 222 Solution 222. or A* A. simple analytic geometry shows that the matrix of the rotation is ( ~ IY. or AA *. A few experiments are needed. write S= J1- AA * and T = where the positive square roots are meant. note that since II All ~ 1.

J A(l - A)f). with 0 ~ A ~ 1. Given A. i.) Assertion: the image under P of the "x-axis". that A T 2 " = S2 " A for n = 0. Solution 137). belongs to ran A if and only if 00 L (n + 2) 2 n=O 1 ~n 1 < 2 00.A). Since 0 ~ ~ A 1. • · ·) 2 in 1 It is therefore pertinent to observe that a vector f = (~ 0 . A. BB - AA* + S T A*. 0). A standard example of an operator A with non-closed range is given by the matrix diag(t.TA*) S 2 + AA* ' AT. The answer is: not necessarily. Trivially A T 2 = S2 A. is not closed. Michael. let R be the positive square root of A(l . the operator matrix p _ ( . (The use of Pis the reason this discussion is here.J A(1 - A A) J A(1. see [49]. ~ 1 .e. (b) The proof is similar to that of (a).A*S 2 It remains only to prove that AT = SA.A is a projection. 323 . · · ·.SA) T 2 +A* A .A)) 1. 1. (The result (b) is due to Solution 223. The image in question is the set of all vectors of the form (Af. of the set of vectors of the form (f.. · · ·). Since A* B* = ( S it follows by direct computation that B B- * _ (A*A + *_( T SA. and write B = E. A good way to get a counterexample is to consider a suitable operator with non-closed range and make a projection out of it. 2.) (~ 1~A} 223 The verification that B is a projection is painless. see Solution 222. and it follows. i. and hence that AT= SA. and simpler. as desired. This implies that Ap(T2 ) = p(S2 )A for every polynomialp (cf.AT 2 A*S. by induction. ~ 2 .UNITARY DILATIONS That B is a dilation of A is clear. i.

coisometries. A ts a compressiOn of an Ao ( *) · 0 arbitrary operator A on H to a finite-dimensional subspace H 0 . f)~ 11/11 2 for each vector f. trivially. such as the sets of isometries. n +1 In view of these observations it is an easy exercise in elementary analysis to produce a sequence {fd of vectors such that the limit of the image sequence {P(j~.. it follows that the set of isometries is strongly closed. extend the operator B 0 to a normal operator B on H (for instance by defining B to be 0 on (H 0 EB H 0 ). It follows that the restriction of an isometry to a finitedimensional subspace of an infinite-dimensional space H always has a unitary extension to H.~. If. Consequence: every isometry is in the strong closure of U.A) if and only if I n=O (n + 2)2 1~. . . One possibility is A=(~.. . Every isometry has a unitary extension to a larger space (Problem 149). Since the set of all strict contractions is strongly dense in the set of all contractions. Since A is an isometry if and only if II A/ II = II f II for each vector f. f) is weakly continuous. As for co-isometries: recall that if A is a strict contraction. subnormal operators. ). and since A f-. It follows from Problem 222(b) and the projection character- ization of the weak topology that the weak closure of P includes the "interval" I of all positive contractions (i. A similar argument (based on Problem 222(a)) implies that the weak closure of U is the set of all contractions. then A has a co-isometric extension (Problem 152). if follows that I is weakly closed. m partlcuIar. and since A f-. and.o. then extend the embedding of H0 into H to an embedding of H 0 EB H0 into H. . 224 Solution 224. it follows that the weak closure of N contains every contraction.SOLUTIONS it belongs to ranJ A(1 . and thus infer that every compression of A to a finite-dimensional subspace has a dilation inN. . and hyponormal operators. (Similar use of the monotoneness of the closure operation yields. the weak closures of many other sets.e. it follows that P is strongly closed. as follows.? (Af. partial isometries.k. Bo = Ao * Ao ts normaI .. Caution: is that set weakly closed? Since U c: N. the set of all operators A such that 0 ~A ~ 1)..o. 225 Solution 225. therefore. For every operator A 0 the operator matrix Ao .12 < oo. Since the set P of projections can be characterized as the set of all idempotent contractions. 0)} is not in ran P. it follows that the strong closure 324 . Since A E I if and only if 0 ~ (Af.o.? II A/II is strongly continuous.) There is a longer way to get the same result. which also has some merit. that the weak closure of N is the set of all operators.L). .

Given g. Solution 226. The assertion is that if every such neighborhood of B contains a hyponormal operator. the behavior of the adjoint is what makes the result more interesting and the proof less trivial than for similar statements about the weak operator topology or the norm topology.BFII < e..? IIBgll.? ellgll + IIAgll = ellgll + IIAFgll . The set of hyponormal operators is strongly closed... PROOF.? ellgll + IIAFg.. To get an idea about how to use the neighborhood hypothesis. assume for a moment a much stronger hypothesis. F)) such that IIAF . of course. of course. 325 . let F be the projection onto the span of g and B*g. The assumption is that for every e > 0 there exists a hyponormal operator A ( =A(e.? 2ellgll + IIBFgll = 2ellgll + IIBgll.. e > 0) implies that IIB*gll Comparison ofthe first and last terms of this chain of relations (true for every .? IIFB*g. A general principle is at work here: in most statements about the strong operator topology. then IIB*gll . 226 A typical (basic) strong neighborhood of an operator B is the set {A: IIAF.BFII < e}.FB*II < e... as desired. and it follows that IIB*gll = IIFB*gll ..? IIA*gll . as follows. In that case. and it follows that IIB*gll = IIFB*gll = IIFA*gll . IIFA*.. The effect of the misbehavior of the adjoint is strikingly visible in the difference between the strong closure of the set of isometries and the strong closure of their adjoints.FA*gll + IIFA*gll . FA*= FB*. namely that there exists a hyponormal operator A such that AF = BF.? ellgll + IIA*gll .? IIBgll for every vector g.BFgll + IIBFgll .UNITARY DILATIONS of the set of all co-isometries is the set of all contractions.? IIAgll = IIAFgll = IIBFgll = IIBgllThe proof proper does all this again. In that case. where e > 0 and F is a projection of finite rank. carrying an e along.

L UJ = n-1 (1 - n-1 ni=o ni=O L <pi f. As for the theorem for unitary operators. Although this may not be the most revealing proof of the theorem. 185]). it is certainly the shortest. The proof that B is unitary is an obvious computation (which uses the results of Solution 222). p. indexed by all integers (positive. 228 Solution 228. The theorem can be proved directly. then each operator on K is an infinite operator matrix. but the parochial Hilbert space proof via the spectral theorem is more transparent. zero). Iff E H ( =L 2 (Jl)). and the diagonal entries of the powers are the corresponding powers of the diagonal entries of B. the projection P from K to H is given by P~o (~ ~) (The parentheses indicate the entry in position <O. This makes it obvious that B is a power dilation of A.SOLUTIONS 227 Solution 227. [60. and. The proof is constructive.L<fJj .) Given A. then the spectral theorem justifies the assumption that H = L 2 (Jl) for some measure Jl. it is due to Schaffer [125]. its powers are triangular. negative. If U is a unitary operator on H. it follows that . Since B is triangular. then 1 .£ 1 1 n i=O 1 n-1 I 326 . Given H. in particular. put 0 1 0 0 0 0 0 B= 0 0 1 0 0 0 0 0 0 0 s -A* 0 0 0 0 0 0 0 0 0 0 0 0 0 0 (A) 0 0 0 T 0 0 0 0 1 0 0 0 0 1 0 where Sand T are as in Solution 222. it can be proved by relatively ele· mentary and widely generalizable geometric methods (cf. 0). ) Since I<pI = 1 almost everywhere. let K be the direct sum of countably infinitely many copies of H. in such a way that U is the multiplication induced by a measurable function qJ of constant modulus I almost everywhere. on some suitable measure space. but the proof via unitary operators and dilation theory has an elegance that is hard to surpass.

the assumption that I<pI = 1 almost everywhere implies that the averages -I 1 n-1 <pj n i=O form a convergent sequence almost everywhere (whose limit is the characteristic function of the set where <p = 1). as stated. n = 0. the proof via dilation theory is due to Nagy [96]. then A"f = PU"f. Given A on H.L Aij = p (1 L Uij n -1 n-1 ) 0 n i=o n i=O Since ~ "f_1 Uij n i=O has a limit as n follows that ~ oo for each f. it . Relatively hard analytic proofs can be given. say. If p is a polynomial and iff is in H. If A is an arbitrary contraction on H.i=O Aij L n has a limit as n ~ oo for each f.UNITARY DILATIONS almost everywhere. it follows that the Lebesgue dominated convergence theorem (not necessarily the bounded convergence theorem) is applicable to the sequence This completes the proof of convergence.. Since. 2. and it follows that 1 . · · ·. The mean ergodic theorem for unitary operators was first proved by von Neumann [149]. then llp(A)/11 = IIPp(U)/11 ~ llp(U)II ·11/11 229 ~ IIPIIv·llfll (by the definition of power dilation) (because IIPII = 1) (because U is normal). 1. and it follows. a quick second glance will even reveal what the limit is. The extension to contractions is due to Riesz-Nagy [113]. bounded). and let P be the projection from K to H. let U on K be a unitary power dilation of it.e. and let P be the projection from K to H. then let U be a unitary power dilation of it on a Hilbert space K. 1 n-1 Solution 229. moreover. and since P is continuous (i. This means that iff E H. A good recent reference to ergodic theory in general is [152]. that llp(A)II ~ IIPIIv· 327 . with dilation theory all becomes simple [96].

1. (Consequence: the 2-variable von Neumann inequality is true. but the success of this "one-variable" theory made it tempting to look for "several-variable" extensions. k.···.) Next: is the commutative dilation theorem true fork = 3? Answer by Parrott [105] (seven years after Ando's proof for k = 2): no.. Ak) probably cannot be formulated in general. . and if Dis the closed unit disc. If ever the answer is no. Finally: is the k-variable von Neumann inequality fork ~ 3 true nevertheless? Answer by Varopoulos [145] and Crabb-Davie [28]: no.j = l. then p(A 1. . 328 .k}. · · · Akmk is the compression to the original space of U 1m. but the conclusion could be true just the same. A sensible definition of p(A 1 . Is it true that if At. however.. · • ·. · . assume that the A/s commute. the question about the inequality must be asked again. then llp(A1. if. counterexamples exist with k = 3 on spaces of dimension 5. mk} of exponents? Each time that the answer is yes. · · ·. Very well then. Ak)ll ~ IIPIIv for every polynomial p in k variables? What does the question mean? A sensible definition of IIPIIv is near the surface: IIPIIv = sup{lp(A. now is the k-variable von Neumann inequality true? A natural way to try to find the answer is first to look for a k-variable generalization of dilation theory.. the argument for k = 1 can be applied and yields the desired inequality. ·. j = 1. a natural attempt at a proof failed. Ak) makes unambiguous sense. the A /s are pairwise commutative. Ak are pairwise commutative contractions. .. · · · Ukm\ for every finite sequence {m 1.A_k)I:A-jED. . First: is the commutative dilation theorem true for k = 2? Answer by Ando [3] (eight years after Nagy's dilation proof for k = 1): yes. then there exist pairwise commutative unitary operators U 1. Is it true that if IIAjll ~ 1. The growth of the subject went through three periods of suspense. · · ·. · · · . U k on a suitable large space such that A 1m.SOLUTIONS That's all there is to the proof. · · ·.

If PQ . then nP". 2. · · ·.implies that nlexl ~ 211PII · IIQII.CHAPTER 24 Commutators Solution 230.II·Iexl ~ 2IIPII·IIQII· IIP". and more generally (induction) P"Q.1 ex If P is not nilpotent. The only translation that can leave a non-empty compact subset (such as spec PQ) of the complex plane invariant is the trivial translation (i. Wintner's proof. 3. again.. QP = P. where A is an arbitrary scalar. true for n = 1. 3. and hence that.QP 2 = P 2 Q .QP) + (PQ . consequently. no translation at all). There is.QP 2 = P(PQ .QP = ex. and therefore spec QP = spec PQ. and therefore ex = 0.QP)P = 2Pex. n = 1. If P is nilpotent.1ex. If PQ . ···. n11P". in that case. of index n. ex = 0. replace P by P + A. 2. the relation PQ = QP + ex implies that spec(PQ) 230 = spec(QP + ex) = spec(QP) + ex = spec(PQ) + ex.1(PQ)P.PQP + PQP .QP" = nP". then P 2 Q . in other words. Since. no loss of generality in assuming that P is invertible. 329 . then the inequality 1 = 0. Wielandt's proof.e. say.QP = ex. and observe that the new P satisfies the same commutation relation.1 11. ex must be 0.

SOLUTIONS

231

Solution 231. Given the Hilbert space H, let M be the normed vector space of all bounded sequences f = <f1, / 2 , / 3 , · · ·) of vectors in H (coordinatewise vector operations, supremum norm), and let N be the subspace of all null sequences in M (i.e., sequences f with limn llfnll = 0). The quotient space M = MjN is a normed vector space. Each bounded sequence A= <A 1, A 2 , A 3 , ···)of operators on H induces an operator on M; the image of <f1,f2 ,f3 , ··-)under <A 1, A 2 , A 3 , ··-)is <Atf1, A2fz, A3j3, · · -). Since the subspace N is invariant under each such induced operator, the sequence A also induces, in a natural manner, an operator on M; call it A. Bounded sequences of operators on H form a normed algebra (coordinatewise operations, supremum norm). The correspondence A ~---+ A from such bounded sequences to operators on M is a norm-decreasing homomorphism. If P = <P1, P 2 , P 3 , • ·-) and Q = <Qt> Q2 , Q3 , · ·-) are such that PnQn - Q"P n ~ C, then PQ - QP is a commutator on M; since that commutator cannot be equal to I ( = the identity operator on M), the proof is complete. Solution 232. Fix P and consider C = .t\Q = PQ - QP as a function of Q. The operation .t\ is obviously a linear transformation on the vector space of operators; since
ll.t\QII = IIPQ- QPII;;;; 2IIPII · IIQII, that linear transformation is bounded (on the Banach space of operators), and ll.t\11;;;; 211PII. Mappings such as .t\ often play an important algebraic role. The most important property of .t\ is that it is a derivation in the sense that .t\(QR)

232

= .t\Q·R + Q·.t\R.

Proof: PQR- QRP = (PQR - QPR) + (QPR - QRP). Derivations have many of the algebraic properties of differentiation, but, as is visible in the definition itself, they have them in a non-commutative way. First among those properties is the validity of the Leibniz formula for "differentiating" products with several factors. The assertion is that .t\(Q 1 · · · Q") is the sum of n terms; to obtain the j-th term, replace Qi by .t\Qi in the product Q 1 · · · Qn· The proof is an obvious induction. For n = I, there is nothing to do; for the step from n to n + I, write Q1 · · · Qn+ 1 as (Q 1 · · · Qn)Qn+ 1 and use the given (two-factor) product formula. The result is, of course, applicable to the special case in which all the Q/s coincide, but it does not become much more pleasant to contemplate. A special property of the derivation .t\ is that .t\ 2 Q = 0. Here .t\ 2 is, of course, the composition of .t\ with itself, so that
.t\zQ = .t\(.t\Q) = p . .t\Q- .t\Q. P;

the vanishing of .t\ 2 Q expresses exactly that .t\Q commutes with P.
330

COMMUTATORS

The Leibniz formula and the vanishing of .1 2 Q make it easy to evaluate higher order derivatives of higher powers of Q. The process begins with .1Q": it is equal to the sum of the n possible products each of which has one factor equal to .1Q and n 1 factors equal to Q. When -1 is applied to one of these summands, the result is the sum of only n - 1 products. (Reason: when -1 is applied to .1Q, the result is 0.) Each of then - 1 products so obtained has two factors equal to .1Q and n - 2 factors equal to Q. Consequence: .1 2 Q" is equal to the sum of the n(n - 1) possible products of that kind. The argument continues from here on with no surprises and yields a description of NQ". With k = n, the result is that .1"Q" is the sum of n! terms, each of which is (-1Qt; in other words -1"Q" = n !(.1Qt. The last equation is the crucial point of the proof; the desired result is a trivial consequence of it. Indeed, since 1 II(-1QYII = ~ II-1"Q"II ~ , 11-1"11 · IIQ"II ~ ~ 11-111" ·IIQII", n. n. n. it follows that II(-1Q)"II
11 "

~ (n\r

1 • ·

II-111·11QII,

and hence that .1Q is quasinilpotent. As a dividend, the equation for .1"Q" yields a proof of Jacobson's original algebraic result. Statement: if an element Q of an algebra over a field of characteristic greater than n! satisfies a polynomial equation of degree n, and if .1 is a derivation of that algebra such that .1 2 Q = 0, then .1Q is nilpotent of index n. Proof: from .1(.1Q) = 0 infer that .1(.1Q)k = 0 for every positive integer k, and hence, from the equation for .1"Q" infer that .1"+ 1Q" = 0. Consequence: .1"Q; = 0 whenever n > i. If Q" = 2.:7.:-J rx;Q; is the polynomial equation satisfied by Q, then it follows that .1"Q" = 0, and hence it follows, again from the equation for .1"Q", that n !(.1Q)" = 0. The conclusion follows from the assumption about the characteristic.

Solution 233. (a) The trick is to generalize the formula for the "derivative"
of a power to the non-commutative case; cf. Solutions 230 and 232. The generalization that is notationally most convenient here says that

233

L pn-i-1Cpi; i=O the proof is a straightforward induction. It follows that
P"Q- QP" =
n-1

n-1

P"Q- QP" = nPn-1and hence that

L pn-i-1(1- C)Pi,
n-1

i=O niiP"- 1 11 ~ 211PII · IIQII ·IIP"- 1 11

+

Ill -

C11 · 2.:

IIP"-i- 1 11 · IIPill.

i=O

331

SOLUTIONS

Up to now P could have been arbitrary. Since P was assumed hyponormal, the last written sum is equal to niiP"- 1 11 (see Problem 162). Divide through by niiP"- 1 11. (If P = 0, everything is trivial, and if P "I= 0, then pn- 1 "I= 0.) The result is that

I~

2 n

IIPII·IIQII + Ill- q,

and the conclusion follows. (b) If 111 - q < 1, then Cis invertible; since, by the Kleinecke-Shirokov theorem, C is quasinilpotent, that is impossible.

234

Solution 234. If A has a large kernel, then that kernel is the direct sum of ~ 0 subspaces all of the same dimension. The orthogonal complement of the kernel may or may not be large. If, however, one of the direct summands of the kernel is adjoined to that orthogonal complement, the result is a representation of the underlying Hilbert space in the form of an infinite direct sum H EB H EB H EB · · · in such a way that the direct sum of all the summands beginning with the second one is annihilated by A. If corresponding to this representation of the space the operator A is represented as a matrix, it will have the form

where each An (and each 0) is an operator on H. Write

P~(~ ~ ~ ~J
and

Q = A3

then P and Q are operators and (straightforward computation) PQ - QP = A. The proof of the main assertion is complete. To prove Corollary 1, suppose that {!1, · • ·, fn} is a finite set of vectors in an infinite-dimensional Hilbert space H, and let M be their span. If A is an
332

c
Az A4

-Ao 0 0 0

0 -Ao 0 0

-~,}

COMMUTATORS

operator on H, let C be the operator that is A on M and 0 on Mj_; by what was just proved, Cis a commutator, and C agrees with A on each/;, i = 1, · · ·, n. This implies that every basic strong neighborhood of A contains commutators. The proof of Corollary 2 is similar. Given H, let M be a large subspace with a large orthogonal complement; given A, define Cas in the preceding paragraph, and write B = A - C. Since A = B + C and both B and C are commutators, the proof is complete.
Solution 235. Since A is not a scalar, there exists a vector f such that f and

235

Af are linearly independent. Let T be an invertible operator such that Tf = f and T Af = - Af Since this implies that (A+ T- 1 AT)f = Af- Af = 0,

it follows that the direct sum
S =(A+ T- 1 AT)@ (A+ T- 1 AT)@ (A+ T- 1 AT)@ · ··

has a large kernel. (What really follows is that the kernel is infinite-dimensional; since the whole space is separable, this implies that the kernel is large.) By Problem 234, the direct sum S is a commutator. If
B=AEBAEBA@···

and

then
S=B+C.

The next step is the following somewhat surprising lemma: whenever B and C are operators such that B + C is a commutator, then B EB C is a commutator. The proof is an inspired bit of elementary algebra. If B + C = PQ - QP, then write R = C + QP = PQ - B, and compute the commutator of
0 (1 0

p)

an

d

= (PQ

0) 0 0 0 0)· 0 R _(R QP) = (B C
333

Consequence: B EB Cis a commutator. Since, however, Cis clearly similar to B, it follows that B EB B is a commutator; since, finally, B EB B is unitarily equivalent to B, the proof is complete.

SOLUTIONS

236

Solution 236. Suppose that C = A* A - AA * ~ 0. The problem is to show that 0 E spec C. It is sufficient to show that there exists a sequence {f,} of unit vectors such that Cfn -+ 0 (i.e., that 0 E IT( C)). For this purpose, take a complex number A. in the approximate point spectrum of A, and, corresponding to A., find a sequence { f,} of unit vectors such that (A - A.)fn -+ 0. Since the self-commutator of A - A. is equal to C, and since C ~ 0, it follows that

(A - A.)*(A - A.)

~

(A - A.)(A - A.)*.

Since (A - A.)fn -+ 0, it follows that (A - A.)*(A - A.)fn -+ 0. The last two remarks imply that (A - A.)(A - A.)*j~ -+ 0. Since, therefore, C is the difference of two operators, each of which annihilates { fn}, the operator C does so too.

237

Solution 237. (a) The program is to show that (1) A is quasinormal, (2) ker(l - A* A) reduces A, and (3) the orthogonal complement of ker(l - A* A) is included in ker(A* A - AA*).
(1)

Write P = A*A- AA*. Since, for allf,

11/11 2

~

IIA/11 2

= (A*Af, f)= (AA*f, f)+ (Pf, f)

= IIA*/11 2 + I1Pfll 2 ,
it follows that if Pf = f, then A* P = 0. (The norm condition was used at the first step.) This implies that A* P = 0, and hence that PA = 0, or, equivalently, that (A*A)A = A(A*A). (2) Write M = ker(l - A* A). Iff EM, then f - A* Af = 0. It follows that Af- (A*A)Af = Af- A(A*A)f = A(f- A*Af) = 0, so that M is invariant under A. Similarly (instead of replacing f by Af, replace f by A*f) M is invariant under A*. (Cf. Solution 195.) (3) Since P is idempotent, it follows that

A* A- AA*

= A*AA*A- AA*A*A- A*AAA* + AA*AA*.
A*A- AA*

Since A* A commutes with both A and A*, this can be rewritten as

= A*A(A*A- AA*).

In other words,

P = A*AP,
or (1- A*A)P = 0. It follows that ranP c M, or M_j_ c ker P. 334

COMMUTATORS

Now use the assumption that A is abnormal: it says exactly that ker P includes no non-zero subspace that reduces A. Conclusion: M.L = 0, and this means that A is an isometry.
(b) If A is the bilateral shift with weights {an} such that an = 1 or according as n ~ 0 or n > 0, then A is abnormal and the self-commutator of A is a projection. PROOF. The self-commutator of A is easy to compute; it turns out to be the projection of rank 1 whose range is spanned by e 1 . The abnormality of A follows from Problem 159: according to that result, A is irreducible, and

J2

hence as abnormal as can be.
Solution 238. An infinite-dimensional Hilbert space is the direct sum of

238

Hilbert spaces of dimension ~ 0 • To prove that every scalar of modulus 1 is a commutator, it is therefore sufficient to prove that on a Hilbert space of dimension ~ 0 every scalar of modulus 1 is the commutator of two unitary operators. (The unitary character of the factors guarantees that the possibly uncountable direct sum is bounded.) In a Hilbert space of dimension ~ 0 there always exists an orthonormal basis {en: n = 0, ± 1, ±2, · · ·}. Given a with lal = 1, let P be the diagonal operator defined by Pen = anen, and let Q be the bilateral shift, Qen = en+ 1 . Both P and Q are unitary; a straightforward computation shows that PQP- 1Q- 1 = a. The proof that if a = PQP- 1Q - 1 , then Ia I = 1 is an adaptation of the Wintner argument (Solution 230). Since PQ = aQP, it follows that spec PQ = a spec QP; since, however, PQ is similar to QP, it follows that spec QP = spec QP, and hence that ~pee QP = a spec QP. Since spec QP is a non-empty compact set different from {0} (remember that QP is invertible), and since the only homothety that can leave such a set fixed is a rotation, the result follows.
Solution 239. The proof is an adaptation of Solution 238. The first step is to use Problem 142 to represent the given space as the direct sum of ~ 0 subspaces, all of the same dimension, each of which reduces the given unitary operator U. The direct sum decomposition serves to represent U as a diagonal operator matrix whose n-th diagonal entry is Un, say, for n = 0, ± 1, ±2, · · ·. Solution 238 suggests that the multiplicative commutator of a diagonal operator and a bilateral shift may work here too. To avoid writing down large matrices, it is convenient to introduce some more notation. Think of the given Hilbert space as the set of all sequences f = Un: n = 0, ± 1, ± 2, · · ·} of vectors in some fixed Hilbert space (subject of course to the usual condition llfnll 2 < oo). A typical diagonal operator matrix Pis defined by

239

Ln

(Pf)n

= V,J,,

and the bilateral shift Q is defined by
(Qf)n = fn-1·

The commutator is easy to compute; the result is that
(PQP-1Q-1f)n

=

V, V,-1 -1!,.

335

SOLUTIONS

The equations

u. = v, V,-1

1

can be solved for the V's in terms of the U's. If, for instance, V0 is set equal to 1, then

v, = u•... u1
and
V-<n+ 1 > = U -n
1 · · ·

forn;?; 1,
1

U0 -

for n;?; 0.

The unitary character of the U's implies that the transformation P given by these V's is a unitary operator, and all is well.

240

Solution 240. On an infinite-dimensional Hilbert space, the commutator subgroup of the full linear group is the full linear group itself.
PROOF.

The assertion is that every invertible operator is the product of a finite (but not necessarily bounded) number of multiplicative commutators. The fact is that every invertible operator is the product of two commutators [23]. The proof of that fact takes more work than the present purpose is worth; it is sufficient to prove that every invertible operator is the product of three commutators, and that is much easier. Given an arbitrary invertible operator A on an arbitrary infinite-dimensional Hilbert space, consider the infinite operator matrices
.010 0 0 0 0 0010000 0001000 0 0 0 (0) A 0 0 0 0 0 OOAO 0 0 0 OOOA 000 0000

P

=

and
000 0 000 100 0 000 010 0 000 0 0 1 (0) 0 0 0 0001000 000 0100 000 0 010

Q=

336

Regard the direct sum on which these matrices act as the direct sum of the summand with index 0 and the others. With an obvious change of notation. Since every invertible operator is the product of a unitary operator and an invertible positive operator (polar decomposition).COMMUTATORS Routine computation proves that PQP-lQ-1 = 1 0 0 0 0 0 .0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 (A) 0 0 0 0 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 1. it follows. and identify the direct sum of the others with one of them. the result of the above computations becomes this: every operator matrix of either of the forms (~ ~) or (~ ~) is a multiplicative commutator (provided that the matrix entries operate on an infinite-dimensional space. that every invertible operator is the product of three multiplicative commutators. Consequence: every invertible normal operator is the product of two commutators. Every invertible normal operator on an infinite-dimensional Hilbert space has large reducing subspaces with large orthogonal complements (Problem 142).) 337 . and is therefore representable as the product of two matrices of the indicated forms. as stated. and that A is invertible). (Apply Problem 239 to dispose of the unitary factor.

CHAPTER 25 Toeplitz Operators 241 Solution 241. eJ = (PL<Pei. then (A Wei. eJ This implies that A commutes with W.j 0. 2. To prove sufficiency. 242 Solution 242. Consider for each nonnegative integer n the operator on L 2 given by An W*nAPWn 338 . 1. ···. ei. ei+ 1 ) = (Tcpei+ 1 . ei) = (~an en+ i' e) Lanbn+j. ei+ 1 ) (PL<Pei+ 1 . If qJ = Ln anen. ei) (i. and if W is the bilateral shift (multiplication by e 1). A is an operator on L 2 such that (Aei+ 1 . conversely. ei+ 1 ) (Aei. 2. The proof of necessity is a simple computation: if i. eJ = (Aei+l' eJ = (Aei. ei+ 1) = (Aei. it is to be proved that A is a Toeplitz operator. W*eJ = (W Aei.ij = (Lcpej. 1.then (T<Pei. eJ = (L<Pei+ 1 . and hence (Problem 146) that A is a multiplication.j = 0. ei+ 1). assume that A is an operator on H 2 such that (Aei+ 1. ei) for all i and j. then the matrix entries of L<P are given by A. · · ·).1) = (Aei.i n = ai-j· If. eJ = (L<Pei.

for n sufficiently large both j + nand i + n are positive.j ~ 0. by definition. then this implies that (<p. it is natural to wish for an answer expressed in terms of A instead of A 00 • That turns out to be easy. ± 1.TOEPLITZ OPERATORS (where W is. Iff and g are in H 2 . This is an answer. (A 00 ei. Conclusion: A is the compression to H 2 of a Laurent operator. and hence. ei) fori~ 0. ±2. and from then on (A 0 ei+n• ei+n) is independent of n. but not a satisfying one. A is a Toeplitz operator. then Something like this is true even for negative indices. · · ·). If i. ei) = lim (W*" APW"ei. If i.1· (W*" +1APW"+ 1e1· . g) = (A 00 f. Since IIAnll ~ IIAoll = IIAII. and therefore the Fourier coefficients of <pare the entries in the 0 column of the matrix of A"'. then the sequence {(An p. Consequence: if p and q are trigonometric polynomials (finite linear combinations of the e. for all i and j. ei) = (Ae 0 . then (PAoo f. g) = (Af.·) tm n = lim (W*"APW"ei+ 1 . g). e) n . as before. it follows on easy general grounds that the sequence {An} of operators on L 2 is weakly convergent to an operator A 00 on L 2 • Since. q)} is convergent. then A 00 = L'l'. e.j ~ 0. ei+ 1) n it follows that the operator A 00 has a Laurent matrix and hence that it is a Laurent operator (Problem 241). How can the function <p that induces A be recaptured from the matrix of A? If A = T'l'. the bilateral shift). n so that PAoof =A! for each fin H 2 . i = 0. Indeed. and 339 . g) = lim (W*" APW'1.'s.

if <p* is analytic.i+ 1 = Yii + et.-k/3k-i k=O If now t/J is analytic. it follows that ( <p. j ~ 0.-k/3k-j• k=O it follows that 00 li+1.+ 1-kf3k.). negative.). e.. ek) = 0 for all k (positive. and hence that <p = 0.e. Since 00 Yij = L CX._ i) and <!3. then Yi+ 1. Write C = T"' Tl/t and let <Yii) be the (not necessarily Toeplitz) matrix of C. then (<p. ek) = (Ttpen. observe that if <pis a bounded measurable function. The proof is straightforward. e. so that the matrices of T"' and Tl/t are <ex. then for all fin H 2 .j-1 whenever i. and t/1 = Li f3iei.j-1 + 2: ex. respectively. then 340 . cx. or zero).+ 1) = (AU ei. To prove Corollary 1. U e. 243 Solution 243.. observe that (Aei+ 1. en+k). If the Fourier expansions of <p and t/1 are <p = L.SOLUTIONS Conclusion: <pis the function whose forward Fourier coefficients (the ones with positive index) are the terms of the 0 column of the matrix of A and whose backward Fourier coefficients are the terms of the 0 row of that matrix.i+1-kf3k-i-1 k=O 00 = cxi+ 113. To prove Corollary 2.j+1= Let. If T"' is compact. then I T"'enll --+ 0 (since en--+ 0 weakly). and if both n and n + k are non-negative integers. so that T"' Tl/t = T"'l/t.) = (U* AUei.j-1 k= 1 00 = cxi+1f3-j-1 + L CX.+ 1P.

it follows that either cti+ 1 = 0 for all i ~ 0 or else fJ _j .1 = 0 for all j ~ 0. then q> = 0. the equation for Yi+ 1 . then t/1 = 0. 1) with the unit circle. Solution 244. via the mapping x 1--+ e2 "ix. the product T'P Tt/1 is a Toeplitz operator. If q> is the characteristic function of the closed upper semicircle and 244 t/1 is the characteristic function of the open lower semicircle. then.1 = 0 whenever i. n -:f 0. it follows from Problem 243 that either cp* or t/1 is analytic and that cpt/J = 0. since 0 is a Toeplitz operator. conversely.Tcpt/J is not compact. j ~ 0. and M. If. From this. T'P Tt/1 = 0. conversely. which is equivalent to the desired conclusion. if n is odd.j+ 1 then implies that r:xH 1 fJ _j . !J. nin 0 Consequence: the matrix of T'P is ni 2 -1 ni 2 0 -1 1 3 0 1 3 1 5 0 -3 0 0 1 0 - ni 2 -1 0 1 ni - 1 ni 2 3 -1 0 0 - 1 3 0 - ni 2 -1 1 5 0 1 3 0 1 ni 2 341 . then T'PTt/1 .TOEPLITZ OPERATORS This proves the sufficiency of the condition and the last assertion of the problem. sufficiency is trivial. For computational ease. The function q> becomes then the characteristic function of [0. The Fourier coefficients of q> are easy to determine: if n = 0. if n is even. then its matrix is a Toeplitz matrix (Problem 242). it's best to identify the interval [0. If. Riesz theorem applies (Problem 158) and proves that if cp* is analytic (and not zero). As for the corollary. The F. and if t/1 is analytic (and not zero). in turn.

therefore Tt/1 = 1 . +-z n I 4 + AUA . and the column index increases (from .SOLUTIONS The pattern suggests that the basis be split into two parts according to the parity of the subscripts. as usual. so that the row index increases (from .T'~'' and therefore T'~' Tt/1 = T'~'. 245 Solution 245. The answer is obviously no: the lower left corner is a non-zero Toeplitz operator. It is helpful to begin with some qualitative reflections.T/. the matrix of T'~' takes the form ni 2 0 0 0 0 0 -1 3 -1 5 1 3 ni 2 0 ni 2 - 3 -1 ni 1 3 - -1 1 3 ni 2 0 0 0 0 0 -1 1 3 ni 2 0 5 ni 2 T'~' If A is the operator corresponding to the lower left corner. Fix attention on any particular 342 . if all the even ones are written first. the relevant question is whether this is compact. The rest is an easy computation: the operator matrix of T'~' Tt/1 is ( -nz nr ( A -: I 2 2 n1 .cp. then described by the 2 x 2 operator matrix can be ~(~i nz A ni AU ) AU)· ni - 2 Since t/1 = 1 . followed by all the odd ones. Consider a Laurent matrix written.oo to + oo) as the columns go to the right.A niAU ) 2 • m A 2 U- r:_ 4 Since T'Pt/1 = 0.oo to + oo) as the rows go down.

Corollary 1 is now straightforward. For Corollary 2: if the spectrum ofT"' consists ofO alone.TOEPLITZ OPERA TORS entry on the main diagonal. Suppose therefore that to each positive number 8 there corresponds a unit vector f.T"'"' is compact.:. e0 . and it follows that qJ is real. the non-zero values are recaptured by an obvious translation argument. The preceding paragraph implies that W*"PW"J. it follows that W*"PLPW"--+ L (strongly) as n--+ oo. they all look like the associated Toeplitz matrix. 2. n = 1. and look at the unilaterally infinite matrix that starts there and goes down and to the right. then the same is true of L"'.T"'"' is 0. r(T). · • ·}.UiU*i)T"' Ui = T"' T"' + K (with K compact) = T"'"' + K (because T"' is analytic). therefore. It is instructive to compare this result with Solution 242 where weak convergence was enough. · · ·. then the same is true of L"'. r(L):. The proof of Corollary 3 is similar to that of Corollary 2: if the spectrum ofT"' is real. then W*"TPW"--+ L (strongly). All the matrices obtained in this way from one fixed Laurent matrix look the same. 3. If j ~ 0. The ground is now prepared for the proof of the spectral inclusion theorem for Toeplitz operators. If one of ({J and ljJ is continuous (and the other is arbitrary in L 00 ). Since Lis normal. e 2 . Solution 246. It follows that IlLII :. then T"' is coanalytic.···. It is sufficient to prove that if 0 is an approximate eigenvalue of L. and W*"TPW"fe--+ Lfe (strongly). An efficient non-matrix way of describing the situation might go like this. It follows. It follows that IIPW"fell-+ 1 and I TPW"fell--+ IILfell· The first of these assertions says that PW~f. by the result just proved. and. IlLII = r(L). then each Laurent operator Lis the strong limit of the Toeplitzlike operators PnLPn. for large n.: is. and the corollary follows from the known facts about the norm of a multiplication. If Pn is the projection onto the span of {e-n. and it follows that cp = 0. then T"'T"' .:. Intuition suggests that as the selected diagonal entry moves up and left. e_ 1. Since Pn = W*"PW". and. the second one says that T nearly annihilates it. that 0 is an approximate eigenvalue ofT. e 1 . This implies that if T is the Toeplitz operator corresponding to L. PROOF. the resulting Toeplitz matrices swell and tend to the original Laurent matrix.--+ f. If j < 0.(1 . and since W commutes with L (so that WLW* = L). as promised. the difference T"'T"' . then it is an approximate eigenvalue ofT also. such that IlL/ell < 8. nearly a unit vector. 343 246 . II Til. Suppose first that qJ = ej for some }. The proof of Corollary 4 is the same as Solution 217: W(L) = conv spec L c: conv spec T c: W(T) c: W(L). The reverse inequality was proved before. then T"' T"' = UjT"' = UjU*iT"' Uj (because T"' is a Toeplitz operator) = T"' Ui .

) is a bounded analytic function in the open unit disc. and let iJ be the function defined by iJ(z) = (cp(z). and H 00 + C is the set of all functions of the form e + y with ein H 00 and y in C). References: [8.T<Pl/1 compact?). T<P in norm. it follows that if (/Jn ___. is the set' of all continuous functions.. the (compression) spectrum of T<P.:.T<Pl/1 is compact if and only if 00 . Fix y. then T<Pn ___. The general case can be inferred from the Weierstrass approximation theorem. Conclusion: ip(D) c spec T<P. then](y) = (].AI . e H 00 [qJ*) 1"'1 H 00 [t/J) C H 00 +C (where C. it follows that iJ . as such. and if t/1 is analytic.T<Pl/1 in norm.= (j)(y). temporarily fix an element]in H2 .T<Pnl/1 ___. and therefore ij>(D) c spec T<P. b > 0 whenever Iz I < 1.e. then the limit is compact. It follows that its product with a function analytic in the disc and having a square-summable set of Taylor coefficients is another function with the same properties. That settles the matter when ({J is continuous. It is useful to remember that H2 is a functional Hilbert space. important to know what that kernel function is. that 344 .A.A. Since I T<PII = II({JII oo. ({J in L00 . put A. ({Jt/1 in L oo. then H 00 [t/J] = H 00 . Ky).. If either ({J or t/1 is in C. in this context. This implies that (T<P . 247 Solution 247..A. it follows that T<Pn Tl/1 .. therefore. and hence that A. it is not. The result for t/J follows from the consideration of adjoints.T<Pl/1 is compact whenever ({J is a trigonometric polynomial. and. and. For any e in L let H 00 [8] be the smallest closed subalgebra of L oo that contains and contains every function in H 00 • The ingenious unification goes as follows: T<PTI/I . T<Pl/1 in norm. it has a kernel function (Problem 37).)](y) = 0. If I(j)(z) . however. T<Pnl/1 ___. the condition is clearly satisfied. and. it follows from Solution 42 that 2 T<P]= ijJ }for each ]in H ..)](z). if ({J is co-analytic. The general answer had to take into account and to unify two special answers: the difference is compact if either of the given functions is continuous (a symmetric condition). as follows. in either case. so that it is a proper subspace of fP. T<P Tl/1 . If the terms of the approximating sequence are compact. and it is compact (in fact 0) if either the first function is co-analytic or the second one is analytic (an unsymmetric condition). Let T<P be what T<P becomes when it is transferred from H 2 to H2 . Since iJ(y) = ((j)(y).)H 2 is included in the orthogonal complement of Ky. belongs to.l Ky. then 1/({p .SOLUTIONS Consequence: T<P Tl/1 .l Ky.. If y is a complex(!) number.A. Problem 246 is a special case of a general question (when is T<P Tl/1 . with lyl < 1. Since at the same time ({Jn t/1 ___. therefore. and that took a long time to answer. then H 00 [({J*) = H 00 . the condition is satisfied again. this implies that](y) = 0 if and only if] . i. 147). The converse is even easier.

The F. as an apparently very small consequence of invertibility.A is invertible. then rp is constant. Since both rp · f* and fare in H 2 . there are Toeplitz zero-divisors modulo the compact operators.. It is sufficient to show that if rp is a real-valued bounded measurable function. 345 249 250 . rpljJ = 0. then f* can vanish on a set of measure 0 only. put rp = 0 on an arc that is slightly longer than the lower semicircle at both ends.. and if T"' · f = 0 for some fin H 2 . it follows that rp · f* · f is a constant (Solution 33). If ct. Conclusion: T"' . then spec T"' is the closed interval [ct. Since rp is real. and since f E H 2 .l H 2 . Riesz theorem (Problem 138) implies that either f = 0 or (P · f* = 0.A is not invertible. [3]. PROOF. and they can even be made positive..e. rp · f* · f is real. then either f = 0 or rp = 0. Problem 34 implies that rp · f* · f E H 1 .. Assume the contrary.TOEPLITZ OPERATORS it induces a bounded multiplication operator on is invertible. PROOF. it follows that rp · f* · f E H 1 (Problem 34). and therefore rp = 0. Yes. then both rp and ljJ are continuous. Since Jrp · f* · f dfl = (rp · . since. and consequently rp has almost everywhere the same sign as the constant value of rp · f · f*. assume that T"' . and therefore T"'Tt/1 . This means that rp · f . A Hermitian Toeplitz operator that is not a scalar has no eigenvalues. the constant must be 0. by an inessential change of notation. H2 . 248 Solution 249. More precisely: put rp = 1 on a subarc of the upper semicircle that reaches nearly down to the end points -1 and + 1.f. assume A = 0.f) = 0 (because T"'f = 0). Since rp · f* = rp* . Solution 33 becomes applicable and yields the information that rp · f* · f is a constant almost everywhere. A is not in spec T"'. < A < [3. Solution 250. Since f # 0. it follows that f is different from 0 almost everywhere (Problem 158). the next step is to deduce from this that sgn rp is constant (so that either rp > 0 almost everywhere or rp < 0 almost everywhere). however. If rp is a real-valued bounded measurable function. and [3. and.T"'t/1 is compact (Solution 246). however.A Solution 248. Equivalently (recall that e0 (z) = I for all z) the complex conjugate of rp · f is in H 2 . Since. f* E H 2 (because P(rp ·f) = 0).f) = (T"'.e0 . If ct. it is to be proved that T"' . and join the upper and the lower parts continuously (through positive values if so desired). Iff # 0. it follows that T"'Tt/1 must be compact.f. and M. If ljJ is the reflected version of rp (that is: 1/J(z) = rp(z*)). and if its essential lower and upper bounds are ct. that e0 belongs to the range ofT"'. i. and hence that there exists a (non-zero) function f in H 2 such that T"'f = e0 . It follows. i. Let rp be a smooth version of the characteristic function of the upper semicircle. = fJ. it follows that (rp ·f)* = rp · f*.e. and everything is trivial.

. since a < It < /3.... [3. 346 . T<P .f) = (L<Pf. This contradiction proves that [a.. /3.. [3]... [3.. and.. The reverse inclusion is easier..f). it follows that (T<Pf. it follows that a ...SOLUTIONS In the original notation the result just obtained is that sgn(<p ... This of course implies that spec T<P c: [a.... L<P. <p . /3] c: spec T<P. that is exactly what it is not..f) = (PL<Pf. since T<Pf = PL<Pf whenever f E H 2. and hence that a .. Since a .It) is constant.....

[27] 2. J.S. Bergman. Atkinson. Chang. 68 (1950) 337-404.M.-Y. [206] 11. I. Akhiezer. 60 (1954) 345-350. B. [36] 6. Note on a theoremofFugledeand Putnam. 1. [246] 9.M. [205. J. Sarason. T. T. Berberian. Smith. Paris. Invariant subspaces ofcompletely continuous operators.S. Ungar. 206] 5. [31] 347 . S. Ann. Aronszajn. 12 (1962) 1171-1175. The number 0 refers to the preface. A. New York. Pac. D. The normal solubility of linear equations in normed spaces. N. T. Math. 1961. Berberian. Acta Szeged 24 (1963) 88-90. Michigan Math. Ando. Aronszajn. I. Trans.References Each entry is followed by one or more numbers in brackets. K. V. K. 1947. A. Positive numbers refer to the problems or solutions in which the entry is cited. Muckenhoupt. N. 10 (1959) 175-182. Mat. Ando. Sur les fonctions orthogonales de plusieurs variables complexes avec les applications a Ia theorie des fonctions analytiques. [230] 3. S. Integral Equations and Operator Theory 1 (1978) 285-309. A note on hyponormal operators. A. On matrices of trace zero. N. Glazman. Gauthier-Villars. A. Products qf Toeplitz operators.M. K. On hyponormal operators. [229] 4. Theory of reproducing kernels. 4 (1957) 1-3. Theory of linear operators in Hilbert space. [192] 10. M. Proc. Axler. A. S. S. Sbornik 28 (1951) 3-14 [181] 8. Proc. F.S. S. On a pair of commutative contractions. Albert. A. [191] 7. Math. 14 (1963) 290-291.

Bull. [152] 348 . E. Acta Szeged 26 (1965) 125-137. A. Soc. Davie. J. A. [229] 29. 86 (1957) 414-445. Monthly 67 (1960) 835-841. London Math. Monthly 78 (1971) 536-537. De Barra.S. Shields.-C. B. Conway. de Branges.M. Davie. 7 (1975) 49-50. Bull. Sims. Duke Math. [210] 31. 17 (1965) 695-708. [0] 27. J.M. J.REFERENCES 12. Proc. P. C. Amer. [12] 15. J. B. Amer. Math Bull. Integral Equations and Operator Theory 2 (1979) 174-198. London Math. [173. Pac. Can. J. [231] 21. The approximation problem for Banach spaces. Von Neumann's inequality for Hilbert space operators.M. On a class of operators. 203] 17. 231 (1963) 89-102. P. [7] 26. R. M. Buddenhagen. M. [191] 18. Trans. J. A. L. Structure of commutators of operators. 235] 23. [213] 32. J. A. 14 (1971) 245-246. Subnormal operators. Math. Rota. The existence of invariant subspaces. Halmos. A. Can. On the completeness of Sturm-Liouville expansions. Multiplicative commutators of operators. [196] 25. Soc. B. R. Brown. [230. On the numerical range of compact operators on Hilbert space. L. New York. Brown. Bull. Springer. R. [240] 24. London Math. Brown. Pearcy. S. B. Halmos. Gelfand. Math. 1978. J. 16 (1966) 421-431. Bram. Davis. R. Solution of an invariant subspace problem of K. Halmos. J. Pearcy. Brown. Functions of one complex variable. Math. 82 (1965) 112-127. M. 5 (1973) 261-266. [191] 13. Some invariant subspaces for subnormal operators. J. [175] 30. C. Smiih and P.S. Pearcy. Algebraic properties of Toeplitz operators. A. Brown. Birkhoff. The Toeplitz-Hausdor. Can. J. On two problems concerning linear transformations in Hilbert space. Brown. Bishop. 18 (1966) 737-749. Conway. 81 (1949) 239-255. Math. Spectral theory for operators on a Banach space. A. Subsets of a countable set. 191] 14. A. Math. A. A. [203] 16. Ann. A. Robinson. Math. On a problem of I. C. G. Math. Uspekhi Mat. reine angew. A. [137] 19. Halmos.M.ff theorem explained. 177] 22. Brodskii. Integral Equations and Operator Theory 1 (1978) 310-333. 22 (1955) 75-94. Acta Math. A. Giles. A. M. S. P. R. R. [243] 20. B. W. Soc. Nauk 12 (1957) 129-132. Bernstein. G. 4 (1953) 723-728. 5 (1972) 704-706. A. Operators that are points of spectral continuity. J. [157. [201.S. Cesaro operators. Morrel. T. R. 70 (1964) 718-721. Brown. [105] 28. Rovnyak. Commutators of operators on Hilbert space. G. Beurling. J. Crabb. C.

[203. Dunford. and range inclusion of operators on Hilbert space. 76 (1954) 191-198. 74.S. [0. J. and a problem of H almos. Halmos. Halmos. D. lnterscience. J. Dekker. Normal dilations and extensions of operators.S. [200] 52. A commutativity theorem for normal operators. Acta Math. Enflo. Acta Szeged 33 (1972) 179-192.S. G. Linear operators. Commutators of operators. Amer. New York. 71 (1965) 396. 25 (1952) 43-49. W. The lattice of invariant subspaces of a completely continuous quasi-nilpotent transformation. J. Com. Foia~. Interscience. J. Proc. 147. Stampfli. 58. R. Joint numerical range and joint spectrum of Hilbert space operators. 218] 51. First course in functional analysis. [59] 39. 191] 41. [ 17 5] 42. New York. P. Fuglede. L. Pearcy. Monthly 81 (1974) 29-36. 192. A counterexample to the approximation problem in Banach spaces. 7 (1957) 1031-1035. On the essential numerical range. Perturbations of the shift operator. [234] 349 . 160. Rota. Soc. 14 (1963) 445-449. B. Romine 13 (1963) 349-352. Cyclic vectors of a cyclic operator span the space. 4 (1957) 261-263 [210] 37. Freeman. L. P. Michigan Math. Rovnyak. 74 (1952) 237-240. Inner product spaces. Linear operators. J.A. Math.M. T. J. Part II: Spectral theory. P. Thesis. Ann.S. [192] 46. A. Math. A remark on the universal model for contractions qf G. [152] 34. J. [210] 36. N. 1. Math. 1965. Part I: General theory. Halmos. R. Pedrick. Proc. Geher. 222] 50. C. N.M. the essential spectrum. [145] 35. [166] 47. A. Williams. A. W. [27. Trans. 146. Schwartz. Fillmore. R.P. Halmos. 74. Math. 114 (1965) 251-260. F. 130 (1973) 309-317. Pol. Gudder. Pac. 52.S. Introduction to Hilbert space and the theory of spectral multiplicity.M.orization.M. P. Spectra and spectral manifolds. M. Deckard. A. 53. factorization. R. 36 (1950) 35-40. P. 86. R. Dunford. Proc. Douglas. J. C. [54] 49. On the numerical range of a bounded operator. [184] 45. 209. F. Math. New York. A. A. Summa Brasil. R. S. [191] 38. [123.-C. [152] 44. 1958. Acad. Proc. II. 1969. Commutators of operators. [234] 53. G. 86. Schwartz. J. N. P. 3. PrenticeHall. 5. C. Halmos. Donoghue. Chelsea. J. P. 12. Bull. T. Free University of Amsterdam. 1951. G. R. On ma. P. 17 (1966) 413-415. Another class of invertible operators without square roots. [ 162] 43. [ 11] 48. Correction to "The existence of invariant subspaces". Amer. 2 (1950) 125-134. 170] 40. Amer. M. Math. N. P. 1963.REFERENCES 33. de Branges. Donoghue. 33 (1972) 109-110. Goffman.S. 123. 13. Englewood Cliffs.

R. 16 (1966) 433-437. Math. P. Amer. Z. 1. S. [230] 56. vol. Halmos. Functional analysis and semi-groups. 76 (1970) 887-933. R. Banach spaces of analytic functions. 1964. P. Measure theory. F. Proc. Bull. [168] 74. Math. 192] 57. Kakutani. 1934. P. Sunder. G. Halmos.S. Monthly 76 (1969) 40-41. Cambridge University Press. J. A. 64 (1958) 77-78. Amer. Math. [130. 13 (1962) 585-596. [103] 75. P. J. K. 157] 55. R. Hartman. Pac. Littlewood. Math. [210] 70. E. Proc. Ha1mos. 23 (1974) 557-565.M.. Helson..S. Hilden. Halmos.M. S. Hardy. Ein v. R. Halmos. [31.R. Polya. R. 3 (1919) 314-316. A glimpse into Hilbert space. Halmos. G. J. pp. P. A. 208 (1961) 102-112. R. J. New York. A. P. P. R. Lumer. Prentice-Hall. J. Halmos. 5 (1954) 589-595. Phillips. J. Bounded integral operators on L 2 spaces. V. Halmos. [155. H. Math. 131. [173] 67. Holland. A. Math. Halmos. Pac. P. Lectures on invariant subspaces. Springer. A. J. Hoffman. Some cyclic and non-cyclic vectors of certain operators. (1952) 5-6. 1974. Lectures on modern mathematics. [27] 350 . R. Indiana U. Halmos. [228] 61. What does the spectral theorem say?. Gottingen Nachr. 133] 66.S. S. [143] 63. [211] 58. Monthly 87 (1980) 217. R. [191] 59. 1957. M. II. 18] 62. Shifts on Hilbert spaces. Der Wertvorrat einer Bilinearform. A non topological proofofthe uniform boundedness theorem. S. [123.M. P. Springer. Heinz. Hennefeld. H. 191] 72. H. Neumannscher Satz iiber beschrankte Operatoren im Hilbertschen Raum. P. Providence. On the spectra of Toeplitz's matrices. Ten problems in Hilbert space. Math. A Hilbert space proof of the Banach-Steinhaus theorem. New York. Schaffer. R. [0. R. 1974. E. J. R.S. E.M. Amer. Hausdorff. E. Square roots of operators. 103] 64. Wallen. Acta Szeged 25 (1964) 1-5. Springer. Math.R. P. M. [245] 69. 1-22. G. Halmos. [56. Jr.S. S. L. Math. Academic Press. [299] 71. J. P. Berlin. Bull. Englewood Cliffs. [157. Lumer. Products of symmetries. J. G. Halmos. 1978. Partial isometries. reine angew. McLaughlin. New York. Invariant subspaces of polynomially compact operators. P.REFERENCES 54. J. [27] 73. Hille. 158] 76. Finite-dimensional vector spaces. A. [34. Halmos. Wiley. Amer. Square roots of operators. 1962. [154] 60. Wintner. Cambridge. 145] 65. Numerical ranges and normal dilations. Monthly 70 (1963) 241-247. 72 (1950) 359-366. 1963. P. [46] 68. 4 (1953) 142-149. Inequalities. New York.

REFERENCES

77. T. Ito, On the commutative family of subnormal operators, J. Fac. Sci. Hokkaido University 14 (1958) 1-15. [201] 78. T. Ito, T. K. Wong, Subnormality and quasinormality of Toeplitz operators, Proc. A.M.S. 34 (1972) 157-164. [209] 79. N. Jacobson, Rational methods in the theory ofLie algebras, Ann. Math. 36 (1935) 875-881. [232] 80. G. J. Johnson, A crinkled arc, Proc. A. M.S. 25 (1970) 375-376. [6] 81. G. J. Johnson, Hilbert space problem four, Amer. Math. Monthly 78 (1971) 525-527. [6] 82. R. V. Kadison, /sometries of operator algebras, Ann. Math. 54 (1951) 325-338. [136] 83. R. V. Kadison, Strong continuity of operator functions, Pac. J. Math. 26 (1968) 121-129. [116] 84. G. K. Kalisch, On similarity, reducing manifolds, and unitary equivalence of certain Volterra operators, Ann. Math. 66 (1957) 481-494. [191] 85. I. Kaplansky, Products of normal operators, Duke Math. J. 20 (1953) 257-260. [192] 86. T. Kato, Some mapping theorems for the numerical range, Proc. Japan Acad.41(1965)652-655.[221] 87. J. L. Kelley, General topology, Springer, New York, 1975. [0] 88. D. C. Kleinecke, On operator commutators, Proc. A.M.S. 8 (1957) 535-536. [232] 89. P. D. Lax, Translation invariant spaces, Acta Math.101 (1959) 163-178. [157] 90. A. Lebow, On von Neumann's theory of spectral sets, J. Math. Anal. Appl. 7 (1963) 64-90. [229] 91. V. I. Lomonosov, Invariant subspaces for the family of operators which commute with a completely continuous operator, Functional Anal. Appl. 7 (1973) 213-214. [191] 92. K. Lowner, Grundziige einer /nhaltslehre im Hilbertschen Raume, Ann. Math. 40 (1939) 816-833. [18] 93. A. J. Michaels, Hilden's simple proof of Lomonosov's invariant subspace theorem, Adv. in Math. 25 (1977) 56-58. [191] 94. W. Mlak, Unitary dilations of contraction operators, Rozprawy Mat. 46 (1965). [227] 95. B. Sz.-Nagy, Surles contractions de l'espace de Hilbert, Acta Szeged 15 (1953) 87-92. [227] 96. B. Sz.-Nagy, Prolongements des transformations de l'espace de Hilbert qui sortent de cet espace, Appendix to "Le9ons d'analyse fonctionelle" by F. Riesz, B. Sz.-Nagy, Akademiai Kiado, Budapest, 1955. (Extensions of linear transformations in Hilbert space which extend beyond this space, Appendix to "Functional analysis" by F. Riesz, B. Sz.-Nagy, Ungar, New York, 1960.) [227, 228, 229] 97. B. Sz.-Nagy, Surles contractions de l'espace de Hilbert. II, Acta Szeged 18(1957) 1-14. [227] 351

REFERENCES

98. B. Sz.-Nagy, Suites faiblement convergentes de transformations normales de l'espace Hilbertien, Acta Szeged 8 (1957) 295-302. [224] 99. B. Sz.-Nagy, C. Foia~, Surles contractions de l'espace de Hilbert. V. Translations bilaterales, Acta Szeged 23 (1962) 106-129. [155] 100. B. Sz.-Nagy, C. Foia~, Surles contractions de l'espace de Hilbert. IX. Factorisations de Ia fonction caracteristique. Sous-espaces invariants, Acta Szeged 25 (1964) 283-316. [191] 101. B. Sz.-Nagy, C. Foia~, On certain classes ofpower-bounded operators in Hilbert space, Acta Szeged 27 (1966) 17-25. [221] 102. J. D. Newburgh, The variation of spectra, Duke Math. J. 18 (1951) 165-176. [105]. 103. N. K. Nikolskii, Invariant subspaces of certain completely continuous operators, Vestnik Leningrad University Math. (1965) 68-77. [191] 104. R. F. Olin, Functional relationships between a subnormal operator and its minimal normal extension, Pac. J. Math. 63 (1976) 221-229. [198] 105. S. K. Parrott, Unitary dilations for commuting contractions, Pac. J. Math. 34 (1970) 481-490. [229] 106. C. Pearcy, On commutators of operators on Hilbert space, Proc. A. M.S. 16 (1965) 53-59. [234] 107. C. Pearcy, An elementary proofof the power inequality for the numerical radius, Michigan Math. J. 13 (1966) 289-291. [221] 108. C. R. Putnam, On commutators of bounded matrices, Amer. J. Math. 73 (1951) 127-131. [236] 109. C. R. Putnam, On normal operators in Hilbert space, Amer. J. Math. 73 (1951) 357-362. [192] 110. H. Radjavi, P. Rosenthal, On invariant subspaces and reflexive algebras, Amer. J. Math. 91 (1969) 683-692. [193] 111. H. Radjavi, P. Rosenthal, Invariant subspaces, Springer, Berlin, 1973. [191] 112. C. E. Rickart, General theory of Banach algebras, Krieger, Huntington, 1974. [103, 104] 113. F. Riesz, B. Sz.-Nagy, Ueber Kontraktionen des Hilbertschen Raumes, Acta Szeged 10 (1943) 202-205. [228] 114. F. Riesz, B. Sz.-Nagy, Le(:onsd'analysefonctionelle, Akademiai Kiad6, Budapest, 1952. (Functional analysis, Ungar, New York, 1955.) [12, 121, 130] 115. J. R. Ringrose, On the triangular representation of integral operators, Proc. London Math. Soc. 12 (1962) 385-399. [186] 116. J. B. Robertson, On wandering subspaces for unitary operators, Proc. A.M.S. 16 (1965) 233-236. [155] 117. S. Rolewicz, On orbits of elements, Studia Math. 32 (1969) 17-22. [168] 118. M. Rosenblum, On a theorem of Fuglede and Putnam, J. London Math. Soc. 33 (1958) 376-377. [192] 119. G.-C. Rota, On models for linear operators, Comm. Pure Appl. Math. 13(1960)469-472.[153] 352

REFERENCES

120. H. L. Royden, Real analysis, Macmillan, New York, 1968. [OJ 121. I. Z. Ruzsa, A pseudonorm for unbounded transformations, Acta Math. Acad. Sci. Hung. 34 (1979) 297-305. [50] 122. L. A. Sakhnovich, On the reduction of Volterra operators to the simplest form and on inverse problems, Izv. Akad. Nauk SSSR 21 (1957) 235-262. [191] 123. D. Sarason, A remark on the Volterra operator, J. Math. Anal. Appl. 12 (1965) 244-246. [191] 124. H. H. Schaefer, Eine Bemerkung zur Existenz invarianter Teilriiume linearer Abbildungen, Math. Z. 82 (1963) 90. [191] 125. J. J. Schaffer, On unitary dilations of contractions, Proc. A.M.S. 6 (1955) 322. [227] 126. J. J. Schaffer, More about invertible operators without roots, Proc. A. M.S. 16 (1965) 213-219. [145] 127. R. Schatten, Norm ideals of completely continuous operators, Springer, Berlin, 1960 [173]. 128. M. Schreiber, Unitary dilations of operators, Duke Math. J. 24 (1956) 579-594. [227] 129. J. Schur, Bemerkungen zur Theorie der beschriinkten Bilinearformen mit unendlich vie/en Veriinderlichen, J. reine angew. Math. 140 (1911) 1-28. [45] 130. J. Schur, Ober Potenzreihen, die im Innern des Einheitskreises beschriinkt sind, J. reine angew. Math. 147 (1917) 205-232. [71] 131. I. E. Segal, Equivalences of measure spaces, Amer. J. Math. 73 (1951) 275-313. [123] 132. I. E. Segal, A~qebraic integration theory, Bull. A. M.S. 71 (1965) 419-489. [18] 133. A. L. Shields, Weighted shift operators and analytic function theory, Mathematical Surveys 13 (1974) 49-128. (C. Pearcy, ed., Topics in operator theory, A.M.S., Providence, 1974.) [94] 134. F. V. Shirokov, Proof ofa conjecture of Kaplansky, Uspekhi Mat. Nauk 11 (1956) 167-168. [232] 135. K. Shoda. Ein(qe Siitze iiber Matrizen, Japanese J. Math. 13 (1936) 361-365. [230] 136. J. G. Stampfli, Hyponorrnal operators, Pac. J. Math. 12 (1962) 14531458. [205,206] 137. J. G. Stampfli, Perturbations of the shift, J. London Math. Soc. 40 (1965) 345-347. [184] 138. M. H. Stone, Linear transformations in Hilbert space and their applications to analysis, A.M.S., New York, 1932. [27, 53] 139. 0. Szasz, Ober die Approximation stetiger Funktionen durch lineare Aggregate von Potenzen, Math. Ann. 77 (1916) 482-496. [165] 140. R. C. Thompson, On matrix commutators, J. Wash. Acad. Sci. 48 (1958) 306-307. [236] 141. 0. Toeplitz, Zur Theorie der quadratischen Formen von unendlichvielen Veriinderlichen, Gottingen Nachr. (1910) 489-506. [44] 353

REFERENCES

142. 0. Toeplitz, Das algebraische Analogon zu einem Satze von Fejer, Math. z. 2 (1918) 187-197. [210] 143. N. Tsao, Approximate bases in a Hilbert space, Amer. Math. Monthly 75 (1968) 750. [12] 144. F. A. Valentine, Convex sets, McGraw-Hill, New York, 1964. [216] 145. N. T. Varopoulos, On an inequality of von Neumann and an application of the metric theory of tensor products to operator theory, J. Funct. Anal. 16 (1974) 83-100. [229] 146. R. A. Vitale, Representation of a crinkled arc, Proc. A.M.S. 52 (1975) 303-304. [6] 147. A. L. Volberg, Two remarks concerning the theorem of S. Axler, S.-Y. A. Chang, and D. Sarason, J. Operator Theory 7 (1982) 209-219. [246] 148. J. von Neumann, Zur Algebra der Funktionaloperationen und Theorie der norma/en Operatoren, Math. Ann. 102 (1929) 370-427. [27, 28] 149. J. von Neumann, Proof of the quasi-ergodic hypothesis, Proc. N.A.S. 18 (1932) 70-82. [228] 150. J. von Neumann, Functional operators, Princeton University Press, Princeton, 1950. [122] 151. J. von Neumann, Eine Spektraltheoriefur allgemeine Operatoren eines unitiiren Raumes, Math. Nachr. 4 (1951) 258-281. [229] 152. P. Walters, An introduction to ergodic theory, Springer, New York, 1981. [228] 153. R. J. Whitley, A note on hyponormal operators, Proc. A.M.S. 49 (1975) 399-400. [207] 154. H. Widom, Inversion of Toeplitz matrices, II, Illinois J. Math. 4 (1960) 88-99. [250] 155. H. Widom, On the spectrum of a Toeplitz operator, Pac. J. Math. 14 (1964) 365-375. [250] 156. N. A. Wiegmann, Normalproductsofmatrices, Duke Math. J.15 (1948) 633-638. [192] 157. N. A. Wiegmann, A note on infinite normal matrices, Duke Math. J. 16 (1949) 535-538. [192] 158. H. Wielandt, Ueber die U nbeschriinktheit der Operatoren des Quantenmechanik, Math. Ann. 121 (1949) 21. [230] 159. A. Wintner, Zur Theorie der beschriinkten Bilinearformen, Math. Z. 30 (1929) 228-282. [219, 247] 160. A. Wintner, The unboundedness of quantum-mechanical matrices, Phys. Rev. 71 (1947) 738-739. [230]

354

List of Symbols

Positive numbers refer to problems or solutions; the number 0 refers to the preface

I All A* A-t
A2
B(H) B 1 (H)

51 51 52 31 99 107 54 216 0, 70 61 0 35 97 123 0 0 0 33 33

uoo
Im

H2* :fJ:2

33 35 33 0 170 0 0 0 0 151 88 0 0 0 0 89 73 355

tr

0 218 210 211 0 0 0 0 73 130 73 73 130 0

w

w
J1.j,
{x} {xn}
x~x 2

K
ker ker.L

c conv
det diag dim

LP
M*
null

z*

r
v

1

f(A) F(A)

(f,g) <J,g)
fn~f

r ran ran.L Re sgn shift spec

rr
ITo
p 0

"2

nt

Index

Positive numbers refer to problems or solutions; the number 0 refers to the preface. Abnormal 237 Absolutely continuous function 69 Adams, G. T. 28 Adjoint continuity 110 cyclic vectors 164 linear transformation 51 strong normal continuity 116 weighted shift 93 Algebraic operator 97 Algebras of normal operators 193 Analytic characterization of H 2 35 element ofL 2 33 function 86, 97 Hilbert space 31 operator 97 resolvent 86, 87, 88 quasinilpotent operator 97 Toeplitz operator 242, 247 Approximate basis 12 point spectrum 73, 77, 78 point spectrum of weighted shift 94 Approximation theory, non-commutative 185 Atkinson's theorem 181 Atom of infinite measure 64, 68 Attainment of norm 170

Baire category 7, 27, 52, 161 Ball 4 Banach space 8, 23, 27, 52, 170 Basis 7 boundedness on 50 compactness on 177 for A2 32 Hamel 7 linear 7 orthonormal 7 Bastian, J. J. 132 Bendixson-Hirsch theorem 214 Berger, C. A. 211, 221 Bergman kernel 37 Bessel's inequality 12 Bilateral shift 84, 183, 198 commutant 146 Bilinear functional

357

52. 210. 148 of the unilateral shift as limit 148 Commutative operator determinants 70 ring 70 Commutator 230 direct sum 235 limit 231 multiplicative 238 subgroup 240 Compact diagonal operator 171 hyponormal operator 206 Toeplitz product 244 versus Hilbert-Schmidt 174 Compact operator 170 distance from shift 185 numerical range 213 range 180 square root 178 Compactness ofB 1 (H) 107 on bases 177 Complement of subspace 53 Complete continuity 170 Completeness of B(H) 99 Components of space of partial isometries 131 Compression 222 spectrum 73 Conjugate class 101 function 43 linearity 1 Conjugation continuity 33 in functional Hilbert spaces 38 Connectedness of invertible operators 141 of partial isometries 129 358 . 144 linear functional 3 linear transformation 51 matrix 47 real part 43 Volterra kernel 186 Roundedness of multiplications 66 of multipliers 65 on bases 50 weak 27 Brown.68.INDEX Borel set 18 Boundary of spectrum 78 Boundary point 4 Bounded approximation 41 from below 52. strong closure 225 Co-isometry 127 Column-finite matrix 44 Commutant 146 of the bilateral shift 146 of the unilateral shift 147. A. 220 Calkin algebra 0 Cauchy net and sequence 30 Cesaro convergence 41. 13 linear transformation 58 unit ball 4 vector sum 13. 181 ideal 170 linear manifold 0. 96 matrix 177 Characteristic polynomial 76 Characterization of isometries 149 Chord 5 Circular symmetry 89 Closed graph 29. 15 Closure of approximate point spectrum 77 of numerical range 212 of partial isometries 129 Co-analytic element of L 2 33 Toeplitz operator 243 Co-dimension 0 Co-isometries. 66. 59. 58.60. 73.

106 of square root 126 of squaring 111 Continuous curve 5 curve of unitary operators 131 spectrum 73 Toeplitz products 246 Continuum 54 Contraction 132. spectrum 63 Dilation 222 and weak closure 224 power 227 strong 227 unitary 222 Dim 0 Dimension 0. 11. 70 formal 70. 181 Co-subnormal operator 203 Counting measure 64. 173 Cramer's rule 70 Crinkled arc 6 Cyclic direct sum 163 matrix 167 operator 161 subspace 165 vector 160 vectors of adjoints 164 vectors of position operator vectors. 54 preservation 56 Diminishable complement 53 Direct sum 70 as commutator 235 spectrum 98 Dirichlet kernel 41 problem 43 Distance from commutator to identity 233 359 165 . 71. 105.INDEX Continuity of adjoint 110 of conjugation 33 of co-rank 130 of extension 39 of functional calculus 126 of inversion 100 of multiplication 111 ofnorm 108 of nullity 130 of numerical range 220 ofrank 130 of set-valued functions 105 of spectral radius 104 of spectrum 102. 153 similarity to 154 Co-nullity 130 Conv 216 Convergent sequences of sets 105 Convex hull 216 Convexity 4 Convexoid operator 219 Co-range 151 Co-rank 130. 72 Diag 61 Diagonal 61 compact operator 171 operator. 152. 71 operator 70. spectral parts 80 operator. 7. totality 166 Decreasing squares 124 Degree of a vector 191 Dense linear manifold 54 orbit 168 range 73 Density of cyclic operators 161 of invertible operators 140 of non-cyclic operators 162 Derivation 232 Det 0. 54 and local compactness 16 and separability 17 linear 7 orthogonal 7. 70 Determinant 0.

194 Full linear group 240 Functional calculus 97. 58 Greatest lower bound 14 Factorization 59 F. 34.250 55. 193. 184. 17. 158. 247 spectrum 0 supremum 64 Evaluation functional 36 Even function 189 Exponential Hilbert matrix 47 Extension and strong closure 225 continuity 39 into the interior 35 multiplicativity 42 of a function in H 2 39 of finite co-dimension 202 Extreme point 4. J. 126. 173. 53. 85 Gramian 48. 192.248. 54. P. 167 Graph 52. 138 Eigenspace 125 Eigenvalue of Hermitian Toeplitz operator 248 of weighted shift 93 Eigenvector 83 Elementary symmetric function 106 Ergodic theorem 228 Essential boundedness 66 range 67. 68. 54 Hamilton-Cayley equation 160 Hankel matrix 47 operator 0 Hardy spaces 33. 85 multiplication 68. 71 Fourier expansion 7 Fredholm alternative 179. 136 Fejer kernel 41 theorem 41. 85 Hamel basis 2. 123. 49 Gram-Schmidt process 7. 35 Harmonic function 38 Harrison. 165 Field of values 210 Filling in holes 201 Final space 127 Finite rank 169. W. K. 69. 106 Hausdorff metric 220 Heisenberg uncertainty principle 230 Hermitian operators. 171. increasing sequences 120 360 .INDEX from shift to compact operators 185 from shift to normal operators 144 from shift to unitary operators 150 Distributive lattice 14 Dominated convergence theorem 0. Riesz theorem 243. 228 Donoghue lattice 191 Douglas. J. and M. 148. R. 0 Halmos. 37. 69. 131 Duncan. G. 181 Form 1 Formal determinant 70. 175. 176. Hadwin. 0. 38. D. 30. 131 Hilbert space 36. 7.R. 187 operator 181 Fuglede theorem 146.

177 of operators 176 proper 176 Idempotent 208. 186. 191 Interior 4 of conjugate class 101 Internal direct sum 70 Invariant subspace of quasinormal operators 196 Kakutani. 232 Kelley. S. I.INDEX Hermitian Toeplitz operator 245. 104 Kakutani shift 106 Kaplansky. weak continuity 20 Integral operator 173. 174 Hole 201 Hyponormal operator 203. 225 Im 0 Image of subspace 223 Imaginary part of unilateral shift 144 Increasing sequence of Hermitian operators 120 Infimum 14 of two projections 122 Infinite matrix 44 Vandermonde 10 Initial space 127 Inner function 157 Inner product. 248 spectrum 250 Higher-dimensional numerical range 211 Hilbert matrix 46. 245 series 0 361 . 159 Ker 0 Kernel 0 function 37 of an integral operator 173 product 186 Volterra 186 Kleinecke-Shirokov theorem 232. 233 k-numerical range 211 Kronecker delta 37 Large subspace 142. 189 Integration operator 188. 206 compact imaginary part 207 idempotent 208 powers 209 quasinilpotent 205 spectral radius 205 strong closure 226 of the shift 156. 47. 90. R. 234 Lattice 191 distributive 14 modular 14 of subspaces 14 Laurent operator and matrix 241. 48 transform 43 Hilbert-Schmidt operator 173. L. 157 problem 191 Invertible function 67 modulo compact operators sequence 63 transformation 52 Invertible operators 100 connectedness 141 density 140 openness 100 Involution 143 Isometry 127 characterization 149 pure 149 181 Jacobi matrix 44 Joint continuity 111 Ideal 170.

K. E. 105 Lindelof space 0. 70 Mean ergodic theorem 228 Measure in Hilbert space 18 Metric space of operators 99 topology 19 Michael. 114 McLaughlin. 123 Linear basis 7 dimension 7 functional 3 functional on f2 29 manifold 13. 41 McCarthy. 70 bounded 47 products 186 Maximal orthonormal set 54. 104 strong density 111. J.148. 111 Left divisibility 59 invertibility 144 invertible operators 140 shift 143 Leibniz formula 232 Lim 105 Liminf 105 Limit of commutators 231. 55 partial isometry 128. G. 235 of operators of finite rank 17 5 of quadratic forms 1 Limsup 103. 225 362 .INDEX Law of nullity 151 Least upper bound 14 Lebesgue bounded convergence theorem 41 dominated convergence theorem 0. 69 of functional Hilbert space 69 Miintz-Szasz theorem 165 Net Nikolskii. 54 Liouville's theorem 87 Local compactness and dimension 16 Locally finite measure 64 Lower semicontinuity 21 of operator norm 109 Lumer. C. N. A. 222 Minimal normal extension 197 total set 11 Mixed continuity 169 Schwarz inequality 138 Modular lattice 14 law 15 Monotone shift 191 Multiple 157 Multiplication 64. 102 Majorization 59 Matrix 44. 66. E. A. 163 Nilpotent 97.228 Lebow. 136 polar representation 135 Maximum modulus principle 0. 123 continuity 111 in H 2 34 norm 64 on a functional Hilbert space 68 on f2 62 separate continuity 112 sequential continuity 113 spectral parts 81 strong bounded continuity 117 Multiplicative commutator 238 unitary 239 Multiplicativity of extension 42 Multiplicity of shift 151 theory 146 Multiplier 65. 67. A.

14. E. 103 Norm continuity I 08 of a linear transformation 51 of a multiplication 64 of a weighted shift 91 of the Volterra integration operator 188 power and power norm 205 topology 19. A.25 Paley. E. 72 space. 54 Orthonormal basis 7 set 55 sequences. A. 72 determinant with finite entry 72 determinant. minimal 197 Normal operators algebras 193 compact 172 distance from shift 144 partial isometries 204 reducing subspaces 142 spectral mapping theorem 123 versus unilateral shift 144 weak closure 224 Normality and numerical range 216 Normalized crinkled arc 6 Normaloid operator 219 Normed algebra 230 NOT 107 Null 151 Null point 68 Null-space 0 Nullity 130. R.230 Numerical radius 218 Numerical range 210 and normality 216 and quasinilpotence 215 and spectrum 214 and subnormality 217 closure 212 continuity 220 higher-dimensional 211 of a compact operator 213 of a Toeplitz operator 245 of the unilateral shift 212 of the Volterra integration operator 210 Odd function 189 One-point spectrum 96 Open unit ball 4 Openness of invertible operators I 00 Operator 0 determinant 70. 21 Norm 1. separability 99 topologies 107 with large kernel 234 Orbit 168 Order for partial isometries 128 Orthogonal dimension 7. 181. 52 Non-commutative approximation theory 185 Non-cyclic operator 162 Non-emptiness of spectrum 87 Non-overlapping chords 5 Non-separable Hilbert space 54 Non-tangential convergence 40 Nordgren. 0.20. 151. commutative 70 matrix 70. 71.INDEX Non-closed range 52 vector sum 13. C. 107 upper semicontinuity 21 weak continuity 20 weak discontinuity 20. weak convergence 19. 71. 12 Paley-Wiener 12 363 . spectrum {1} 190 Normal and subnormal partial isometries 204 continuity of spectrum I 05 dilation 222 extension.

134. 135. 88. 153. 195 Power idempotent 208 bounded 154 invariant subspaces 196 dilation 227 Quasinormal weighted shift 139 inequality 221 norm 92. 34. 60. G. Partial isometries 51. M. S. maximal 135 Putnam-Fuglede theorem 192 Polarization 1 Polya. Properideal 176 170. 187 Quasinormal operator 137. 35 Radiallimit 40 Powers of a hyponormal Radius of convergence 35 operator 209 Ran 0 Precompact 170 364 . 163 infimum 122 Projections Perturbation 182 of spectrum 106. 201 Principle of uniform Parseval's identity 12. 0 image of a subspace 223 Peller. Polynomials in the shift 198 186. C.INDEX Preservation of dimension 56 Parrott. V. 29. 97. 49. 136 of symmetries 143 normal 204 topology 23 spectrum 133 Projection subnormal 204 Part of an operator 152 as self-commutator 237 Parts of shifts 153 dilation 222 Pearcy.211 Pure isometry 149 Polar representation.205 of the continuum 54 r 88 series 10. 191 Quadratic form Polynomially diagonal operator 125 Quasinilpotence 96. 179. V. Frode D. 183 of equal rank 57 quasinilpotent 106 strong closure 225 Point evaluation 31 weak closure 224 Point spectrum 73 weak convergence 115 Polar decomposition 56. 199 and numerical range 215 Positive Quasinilpotent self-commutator 236 analytic 97 sesquilinear form 2 hyponormal 205 perturbation 106 Positivity of Hilbert matrix 48 Toeplitz operator 245 Poulsen. 175. K. 187. 113 closure 129 Product components 131 in H 2 34 connectedness 129 unitary equivalence 132 matrix 186 Partial isometry 127 of kernels 186 of normal operators 192 maximal 128. 30. 174 boundedness 27. 188 Position operator 165.

0 Schwarz inequality 2 mixed 138 Segment 4 Self-adjoint ideal 170 Self-commutator 236. 85 Shift 89. 157 modulo compact operators 184 monotone 191 multiplicity 151 polynomials in 198 simple 151 square root 145. 175. P. 158 Right-invertible operators 140 Right shift 143 Ring 70 of operators 122 Rosenthal. F. 69. and M. 152 finite 169.INDEX Range 0 inclusion 59 non-closed 52 of a compact operator 180 Rank 57. 88 Riesz. 179 Sarason. 22 Separate continuity of multiplication 112 Sequence space. A. 68. 151 unilateral 82 weighted 89. 2 Sgn 0 Shields. 139. 104. 87. E. F. 95. 147. 98. 36 Riesz theorem. 153 as universal operator 151 bilateral 84 invariant subspace 156. 92. 143. 27.171. 181 Re 0 Real function in H 2 33 Real part of unilateral shift 144 Reducible weighted shift 159 Reducing subspace of a normal operator 142 Region 31 Reparametrization 6 Representation of a linear functional 3 Reproducing kernel 37 Residual spectrum 73 ot a normal operator 79 Resolvent 86. D. 91. 176. 160. 199 Schur test 45 Schwab. 55. 93. 28. 90. 130. 109 of operator norm 109 of spectrum 103 Semilinear functional Separability and dimension 17 and weak metrizability 26 ofB(H) 99 Separable Hilbert space 17. 102. and M. 237 Self-conjugate functional Hilbert space 38 Semicontinuity 21 ofnorm 21. 158 Riesz-Fischer theorem 31 Riesz representation theorem 3. 159 a-field 0 a-finite measure space 29 Signum 0 Similar normal operators 192 operators 75 subnormal operators 199 Similarity and spectrum 75 of weighted shifts 90 to contractions 154 to parts of shifts 153 topological properties 101 Simple curve 5 365 .173. D. 94. weighted 95 Sequential continuity of multiplication 113 Sequential weak completeness 30 Series of vectors 49 Sesquilinear form 1. L. 96.

J. 201 for Toeplitz operators 245 Spectral mapping theorem 74. 105. 248. norm 1 190 Square root 121 continuity 126 of a compact operator 178 ofashift 145. 221 Strict contraction 153 convexity 4 positiveness 2 Strong bounded continuity of multiplication 117 boundedness 51 convergence 19 density of nilpotents 111. 203. 250 of the bilateral shift 84 one-point 96 perturbed 183 semicontinuity 103 strong semicontinuity 119 Spectrum {1}.151 Square-s'!lmmable Taylor series 35 Squaring continuity 111 weak sequential continuity 114 Stampfli. 225 dilation 227 normal continuity of adjoint 116 operator convergence 118 operator topology 107 semicontinuity of spectrum 119 topology 19 Strong closure and extension 225 of co-isometries 225 of hyponormal operators 226 of projections 225 of unitary operators 225 Structure of set of eigenvectors 83 Subadditivity of spectral radius 88 Submultiplicativity of spectral radius 88 366 . 123 avoidance 123. 153 continuity 104 of a weighted shift 91. 73 and numerical range 214 and similarity 75 continuity 102.147. 142 of the unit disc 194 Spectral parts of a diagonal operator 80 of a multiplication 81 Spectral radius 88. G.INDEX function 173 unilateral shift 151 Singular operator 100 Skew-symmetric Volterra operator 189 Sliding product 91 SOT 107 Span 0 Spec 73 Special invariant subspaces of the shift 156 Spectra and conjugation 73 Spectral inclusion theorem 200. 106 non-empty 87 normal continuity 105 of a commutator 235 of a diagonal operator 63 of a direct sum 98 of a functional multiplication 85 of a Hermitian Toeplitz operator 250 of a multiplication 67 of a partial isometry 133 of a product 76 of a Toeplitz operator 247. 121. 95 subadditivity 88 submultiplicativity 88 Spectral set 229 Spectral theorem 73. 124 Spectral theory 73 Spectraloid operator 219 Spectrum 67. 91. 123 Spectral measure 5.

151 versus normal operators 144 wandering subspaces 155 Unit 230 Unit ball 4 weak metrizability 24 Unit disc 4 spectral measure 194 367 .234.4~ 51. 128.•211. 88. G. 242 analytic 242. 162. 196 projection image 223 Subuniform convergence 0 Support 96 Supremum 14 Symmetric sesquilinear form 2 transformation 60 Symmetry 143 Szego kernel 37 Tr 0. 248. 248. 191. 168. 165 T -total set 11 Two-sided ideal 170. 30. 13 invariant 156. 198. 209 versus quasinormal 195 Subnormality and numerical range 217 Subspace 0. 149. J. 145 Toeplitz matrix 242 Toeplitz operator 0.113 Unilateral shift 82. 247 co-analytic 243 Hermitian 245. 160. 250 numerical range 245 quasinilpotent 245 spectral inclusion 245 spectrum 247.236 Triangular matrix 167 plus one 167 Trigonometric polynomial 33. 148 distance to normal operators 144 distance to unitary operators 150 imaginary part 144 invariant subspaces 156. 209 commutant 147.211 Trace 0. 147. 177 Tychonoff-Alaoglu theorem 23 Taylor series 0 Thompson.INDEX Subnormal idempotent 208 operator 195 operators. 250 Toeplitz products 243 compact 244 continuous 246 Toeplitz-Hausdorff theorem 210 Topological vector space 8 Topologies for operators 107 Total set 8 minimal 11 orthonormal 55 Totality of cyclic vectors 166 Totally total sets 11 Unbounded analytic function with bounded real part 43 symmetric transformation 60 vector 30 Volterra kernel 187 Uncertainty principle 230 Uniform strong convergence 107 topology 107 weak convergence 107 Uniform boundedness principle 2~29. 157. 60. 41. similarity 199 partial isometry 204 versus hyponormal 203. 230. 157 numerical range 212 of higher multiplicity 151 real part 144 simple 151 square root 145.

20. 8. 104.25 of projections 115 Weak discontinuity of norm 20. 157 Wasiutynski. 188 skew-symmetric 189 Von Neumann algebra 122 Von Neumann's inequality 229 w 218 w 210 Wallen. 159 approximate point spectrum 94 eigenvalues 93 norm 91 368 . L. 52 Volterra integration operator 188. 92 Wandering subspace 155. 107 · ·weak closure 25 Unit vector 4 Unitary dilation 222 multiplicative commutator 239 power dilations 227 Unitary equivalence 6 of partial isometries 132 ofSimilar normal operators 192 of similar subnormal operators 199 of weighted shifts 89. C. 103 of spectrum 103 V andermonde 10 Vasunin. 25. 137. 96. 163 Vector sum 13 closed 15 non-closed 13. I. 139. 90 Unitary operators distance from shift 150 strong closure 225 weak closure 224 Upper semicontinuity 21. 92. 14.INDEX Unit sphere 4. 98. 215 numerical range 210 Volterra kernel bounded 186 unbounded 187 Volterra operator 186. 90. 191. 102. 224 continuity of squaring 114 Weakly bounded set 51 closed linear manifold 19 complete Hilbert space 30 dense set 25 open set 19 separable Hilbert space 22 Weierstrass approximation theorem 0. V. 95. 246 Weighted sequence space 95 Weighted shift 89. 187. 126. 33.21 Weak metrizability and separability 26 of Hilbert space 28 of the unit ball 24 Weak sequential closure 25. 103 Weak analyticity 86 houndedness 27 Cauchy net and sequence 30 compactness of the unit ball 23 completeness 30 * compactness 23 operator topology 107 separability 22 topology 19 topology of Banach spaces 107 Weak closure and dilation 224 of a subspace 19 of normal operators 224 of projections 224 of unitary operators 224 of unit sphere 25 Weak continuity 19 of norm and inner product 20 Weak convergence of orthonormal sequences 19. 165. J. 94.

N. 56 Weyl's theorem 182. 249 modulo compact operators Zorn. L. wk 211 WOT 107 9 x-axis 52. A. M. 90 Weiss. 219 Wintner's proof 230 Wiser. 53 Zero-divisor 158. R. 70 249 369 . 95 Weighted shifts similarity 90 unitary equivalence 89. 243. A. 184 Wielandt's proof 230 Wiener. F.INDEX quasinormal 139 spectral radius 91. G. 12 Wintner.

Hilbert sr>~ee·theor:y presented in a way thatgi~~!tt'lS.5.re•der..~~==~~~For•. Each is pretEtded t}Jt the :'~efinitio..thaU!letP lSBN 0~387•9068.a}'ery:'J)r:oblem.~ost...ns ~ntfrrtott!-~t\ needed tQ· Understand it..ThJs revis'Sd and entarged.secc>nd edition c>f A Hll.fl Spac.t Probteml:l4l0ki~'atl'.~~~~~··:=~t~. •(lppof'b.ti1itv~~Q~l.· ··.toftoweiftiy~tiefitlori.tf\e.art..~ere··is first a aod.~·~ndfll$tt)rfcat·lemart<s.·doing·· msth'm•ti~: 'thebook~pre$Etntl'One•Eltl~ge aft•r another..tfttnathematias •by.exposition· of.:1 ISBN 3~54-0:-90$85~1 .

l!lf-ilii'~l~@~-~ ~~ ~~Ta .

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->