This textbook, based on three series of lectures held by the author at the University of Strasbourg, presents functional

*209*
*88*
*2MB*

*English*
*Pages 403
[417]*
*Year 2016*

- Author / Uploaded
- Vilmos Komornik

*Table of contents : Front Matter....Pages i-xxFront Matter....Pages 1-2Hilbert Spaces....Pages 3-54Banach Spaces....Pages 55-117Locally Convex Spaces....Pages 119-147Front Matter....Pages 149-149Monotone Functions....Pages 151-167The Lebesgue Integral in \(\mathbb{R}\) ....Pages 169-195Generalized Newton–Leibniz Formula....Pages 197-209Integrals on Measure Spaces....Pages 211-254Front Matter....Pages 255-256Spaces of Continuous Functions....Pages 257-304Spaces of Integrable Functions....Pages 305-340Almost Everywhere Convergence....Pages 341-362Back Matter....Pages 363-403*

Universitext

Vilmos Komornik

Lectures on Functional Analysis and the Lebesgue Integral

Universitext

Universitext Series Editors Sheldon Axler San Francisco State University Vincenzo Capasso Università degli Studi di Milano Carles Casacuberta Universitat de Barcelona Angus MacIntyre Queen Mary, University of London Kenneth Ribet University of California, Berkeley Claude Sabbah CNRS, École Polytechnique, Paris Endre Süli University of Oxford Wojbor A. Woyczy´nski Case Western Reserve University Cleveland, OH

Universitext is a series of textbooks that presents material from a wide variety of mathematical disciplines at master’s level and beyond. The books, often well classtested by their author, may have an informal, personal even experimental approach to their subject matter. Some of the most successful and established books in the series have evolved through several editions, always following the evolution of teaching curricula, to very polished texts. Thus as research topics trickle down into graduate-level teaching, first textbooks written for new, cutting-edge courses may make their way into Universitext. More information about this series at http://www.springer.com/series/223

Vilmos Komornik

Lectures on Functional Analysis and the Lebesgue Integral

123

Vilmos Komornik University of Strasbourg Strasbourg, France

Translation from the French language edition: Précis d’analyse réelle - Analyse fonctionnelle, intégrale de Lebesgue, espaces fonctionnels, vol - 2 by Vilmos Komornik Copyright © 2002 Edition Marketing S.A. www.editions-ellipses.fr/ All Rights Reserved ISSN 0172-5939 Universitext ISBN 978-1-4471-6810-2 DOI 10.1007/978-1-4471-6811-9

ISSN 2191-6675 (electronic) ISBN 978-1-4471-6811-9 (eBook)

Library of Congress Control Number: 2016941752 Mathematics Subject Classification: 46-01, 46E10, 46E15, 46E20, 28-01, 28A05, 28A20, 28A25, 28A35, 41A10, 41A36 © Springer-Verlag London 2016 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions that may have been made. Printed on acid-free paper This Springer imprint is published by Springer Nature The registered company is Springer-Verlag London Ltd.

Preface

This book is based on lectures given by the author at the University of Strasbourg. Functional analysis is presented first, in a nontraditional way: we try to generalize some elementary theorems of plane geometry to spaces of arbitrary dimension. This approach leads us to the basic notions and theorems in a natural way. The results are illustrated in the small `p spaces. The Lebesgue integral is treated next by following F. Riesz. Starting with two innocent-looking lemmas on step functions, the whole theory is developed in a surprisingly short and clear manner. His constructive definition of measurable functions quickly leads to optimal versions of the classical theorems of Fubini– Tonelli and Radon–Nikodým. These two parts are essentially independent of each other, and only basic topological results are used. In the last part, they are combined to study various function spaces of continuous and integrable functions. We indicate the original sources of most notions and results. Some other novelties are mentioned on page 375. The material marked by the symbol may be skipped during the first reading. Each chapter ends with a list of exercises. However, the most important exercises are incorporated in the text as examples and remarks, and the reader is expected to fill in the missing details. We list on p. xi some interesting papers of the general mathematical culture. We have put a great deal of effort into selecting the material, formulating aesthetic and general statements, seeking short and elegant proofs, and illustrating the results with simple but pertinent examples. Our work was strongly influenced by the beautiful lectures of Á. Császár and L. Czách at the Eötvös Loránd University, Budapest, in the 1970s, and more generally by the Hungarian mathematical tradition created by Leopold Fejér, Frédéric Riesz, Paul Turán, Paul Erd˝os, and others.

v

vi

Preface

We also thank C. Baud, B. Beeton, Á. Besenyei, T. Delzant, C. Disdier, O. Gebuhrer, V. Kharlamov, P. Loreti, C.-M. Marle, P. Martinez, P.P. Pálfy, P. Pilibossian, J. Saint Jean Paulin, Z. Sebestyén, A. Simonovits, Mrs B. Szénássy, J. Vancostenoble, and the editors of Springer for their precious help. This book is dedicated to the memory of my father. Strasbourg, France May 23, 2016

Vilmos Komornik

Contents

Part I 1

Functional Analysis

Hilbert Spaces .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.1 Definitions and Examples .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.2 Orthogonality .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.3 Separation of Convex Sets: Theorems of Riesz–Fréchet and Kuhn–Tucker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.4 Orthonormal Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.5 Weak Convergence: Theorem of Choice . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.6 Continuous and Compact Operators.. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.7 Hilbert’s Spectral Theorem . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.8 * The Complex Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.9 Exercises .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

3 3 11 16 24 29 35 39 45 47

2

Banach Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 55 2.1 Separation of Convex Sets . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 57 2.2 Theorems of Helly–Hahn–Banach and Taylor–Foguel .. . . . . . . . . . . . 65 2.3 The `p Spaces and Their Duals . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 69 2.4 Banach Spaces .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 76 2.5 Weak Convergence: Helly–Banach–Steinhaus Theorem .. . . . . . . . . . 79 2.6 Reflexive Spaces: Theorem of Choice. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 87 2.7 Reflexive Spaces: Geometrical Applications.. . .. . . . . . . . . . . . . . . . . . . . 91 2.8 * Open Mappings and Closed Graphs .. . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 96 2.9 * Continuous and Compact Operators.. . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 99 2.10 * Fredholm–Riesz Theory . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 103 2.11 * The Complex Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 112 2.12 Exercises .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 113

3

Locally Convex Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.1 Families of Seminorms.. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.2 Separation and Extension Theorems . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.3 Krein–Milman Theorem . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

119 120 123 126 vii

viii

Contents

3.4 3.5 3.6 3.7 3.8 Part II

* Weak Topology. Farkas–Minkowski Lemma .. . . . . . . . . . . . . . . . . . . . * Weak Star Topology: Theorems of Banach–Alaoglu and Goldstein .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . * Reflexive Spaces: Theorems of Kakutani and Eberlein–Šmulian .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . * Topological Vector Spaces . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . Exercises .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

130 135 140 144 146

The Lebesgue Integral

4

* Monotone Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.1 Continuity: Countable Sets . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.2 Differentiability: Null Sets . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.3 Jump Functions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.4 Proof of Lebesgue’s Theorem . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.5 Functions of Bounded Variation .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.6 Exercises .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

151 151 154 157 161 164 165

5

The Lebesgue Integral in R .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.1 Step Functions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.2 Integrable Functions.. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.3 The Beppo Levi Theorem .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.4 Theorems of Lebesgue, Fatou and Riesz–Fischer . . . . . . . . . . . . . . . . . . 5.5 * Measurable Functions and Sets . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.6 Exercises .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

169 170 174 177 181 187 194

6

* Generalized Newton–Leibniz Formula.. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.1 Absolute Continuity .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.2 Primitive Function .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.3 Integration by Parts and Change of Variable .. . .. . . . . . . . . . . . . . . . . . . . 6.4 Exercises .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

197 198 203 207 209

7

Integrals on Measure Spaces . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.1 Measures .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.2 Integrals Associated with a Finite Measure .. . . .. . . . . . . . . . . . . . . . . . . . 7.3 Product Spaces: Theorems of Fubini and Tonelli .. . . . . . . . . . . . . . . . . . 7.4 Signed Measures: Hahn and Jordan Decompositions . . . . . . . . . . . . . . 7.5 Lebesgue Decomposition . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.6 The Radon–Nikodým Theorem.. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.7 * Local Measurability .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.8 Exercises .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

211 211 217 224 229 235 239 247 251

Part III 8

Function Spaces

Spaces of Continuous Functions . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 257 8.1 Weierstrass Approximation Theorems . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 260 8.2 * The Stone–Weierstrass Theorem . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 265

Contents

8.3 8.4 8.5 8.6 8.7 8.8 8.9 8.10

ix

Compact Sets. The Arzelà–Ascoli Theorem . . . .. . . . . . . . . . . . . . . . . . . . Divergence of Fourier Series . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . Summability of Fourier Series. Fejér’s Theorem . . . . . . . . . . . . . . . . . . . * Korovkin’s Theorems. Bernstein Polynomials .. . . . . . . . . . . . . . . . . . . * Theorems of Haršiladze–Lozinski, Nikolaev and Faber . . . . . . . . . * Dual Space. Riesz Representation Theorem . .. . . . . . . . . . . . . . . . . . . . Weak Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . Exercises .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

268 270 275 279 284 289 299 300

Spaces of Integrable Functions. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.1 Lp Spaces, 1 p 1 . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.2 * Compact Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.3 * Convolution .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.4 Uniformly Convex Spaces . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.5 Reflexivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.6 Duals of Lp Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.7 Weak and Weak Star Convergence . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.8 Exercises .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

305 305 316 320 323 329 331 336 339

10 Almost Everywhere Convergence .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 10.1 Lp Spaces, 1 p 1 . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 10.2 Lp Spaces, 0 < p 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 10.3 L0 Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 10.4 Convergence in Measure .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

341 341 344 351 355

9

Hints and Solutions to Some Exercises . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 363 Teaching Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 375 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 377 Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 395 Name Index .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 401

Some Papers of General Interest

1. 2. 3. 4. 5. 6.

G.D. Birkhoff, What is the ergodic theorem? Am. Math. Mon. 49, 222–226 (1942) J.A. Clarkson, P. Erd˝os, Approximation by polynomials. Duke Math. J. 10, 5–11 (1943) R. Courant, Reminiscences from Hilbert’s Göttingen. Math. Intell. 3, 154–164 (1980/81) J.L. Doob, What is martingale? Am. Math. Mon. 78, 451–463 (1971) L.E. Dubins, E.H. Spanier, How to cut a cake fairly. Am. Math. Mon. 68, 1–4 (1961) P. Erd˝os, Beweis eines Satzes von Tschebyschef. Acta Sci. Math. (Szeged) 5, 194–198 (1930– 32) P 7. P. Erd˝os, Über die Reihe 1=p. Mathematica, Zutphen B. 7, 1–2 (1938) 8. L. Fejér, On some characterization of some remarkable systems of points of interpolation by means of conjugate points. Am. Math. Mon. 41, 1–14 (1934); see in Gesammelte Arbeiten von Leopold Fejér I-II (Akadémiai Kiadó, Budapest, 1970), II, pp. 527–539 9. W. Feller, The problem of n liars and Markov chains. Am. Math. Mon. 58, 606–608 (1951) 10. P.R. Halmos, The foundations of probability. Am. Math. Mon. 51, 493–510 (1944) 11. P.R. Halmos, The legend of John von Neumann. Am. Math. Mon. 80, 382–394 (1973) 12. P.R. Halmos, The heart of mathematics. Am. Math. Mon. 87, 519–524 (1980) 13. R.W. Hamming, An elementary discussion of the transcendental nature of the elementary transcendental functions. Am. Math. Mon. 77, 294–297 (1970) 14. G.H. Hardy, An introduction to the theory of numbers. Bull. Am. Math. Soc. 35, 778–818 (1929) 15. G.H. Hardy, The Indian mathematician Ramanujan. Am. Math. Mon. 44, 137–155 (1937) 16. D. Hilbert, Mathematische probleme. Göttinger Nachrichten, 253–297 (1900), and Arch. Math. Phys. 1(3), 44–63, 213–237 (1901). English translation: Mathematical problems. Bull. Am. Math. Soc. 8, 437–479 (1902) 17. H. Hochstadt, Eduard Helly, father of the Hahn–Banach theorem. Math. Intell. 2(3), 123–125 (1979) 18. J. Horváth, An introduction to distributions. Am. Math. Mon. 77, 227–240 (1970) 19. D.K. Kazarinoff, A simple derivation of the Leibnitz–Gregory series for =4. Am. Math. Mon. 62, 726–727 (1955) 20. K.M. Kendig, Algebra, geometry, and algebraic geometry: some interconnections. Am. Math. Mon. 90(3), 161–174 (1983) 21. J. Milnor, Analytic proofs of the “hairy ball theorem” and the Brouwer fixed-point theorem. Am. Math. Mon. 85, 521–524 (1978) 22. J. von Neumann, Zur Theorie der Gesellschaftsspiele. Math. Ann. 100, 295–320 (1928); [25] VI, 1–26. English translation: On the theory of game of strategy, in Contributions to the Theory of Games, vol. IV (AM-40), ed. by A.W. Tucker, R.D. Luce (Princeton University Press, Princeton, 1959), pp. 13–42.

xi

xii

Some Papers of General Interest

23. J. von Neumann, The mathematician, in The Works of the Mind, ed. by R.B. Heywood (University of Chicago Press, Chicago, 1947), pp. 180–196; [25] I, 1–9 24. J. von Neumann, The role of mathematics in the sciences and in society, in Address at the 4th Conf. of Assoc. of Princeton Graduate Alumni (1954); [25] VI, 477–490 25. J. von Neumann, Collected Works I-VI (Pergamon Press, Oxford, 1972–1979) 26. D.J. Newman, Simple analytic proof of the prime number theorem. Am. Math. Mon. 87, 693– 696 (1980) 27. B. Riemann, Ueber die Anzahl der Primzahlen unter einer gegebenen Grösse, Monatsberichte der Berliner Akademie (1859); in Gesammelte mathematische Werke (Teubner, Leipzig, 1876), pp. 135–144; H.M. Edwards, English translation: On the number of primes less than a given magnitude, in Riemann’s Zeta Function (Academic, New York, 1974), pp. 299–305 28. F. Riesz, Sur les valeurs moyennes des fonctions. J. Lond. Math. Soc. 5, 120–121 (1930); [31] I, 230–231 29. F. Riesz, L’évolution de la notion d’intégrale depuis Lebesgue. Ann. Inst. Fourier 1, 29–42 (1949); [31] I, 327–340 30. F. Riesz, Les ensembles de mesure nulle et leur rôle dans l’analyse. Az I. Magyar Mat. Kongr. Közl., Proceedings of the First Hungarian Mathematical Congress, pp. 214–224 (1952); [31] I, 363–372 31. F. Riesz, Oeuvres Complètes, I-II (Akadémiai Kiadó, Budapest, 1960) 32. C.A. Rogers, A less strange version of Milnor’s proof of Brouwer’s fixed-point theorem. Am. Math. Mon. 87, 525–527 (1980) 33. S. Russ, Bolzano’s analytic programme. Math. Intell. 14(3), 45–53 (1992) 34. A. Seidenberg, A simple proof of a theorem of Erdös and Szekeres. J. Lon. Math. Soc. 34, 352 (1959) 35. S. Smale, What is global analysis? Am. Math. Mon. 76, 4–9 (1969) 36. K. Stromberg, The Banach–Tarski paradox. Am. Math. Mon. 86, 151–161 (1979) 37. G. Szegö, Über eine Eigenschaft der Exponentialreihe. Sitzungsber. Berl. Math. Ges. 23, 50–64 (1924); see in The Collected Papers of Gábor Szegö I-III (Birkhäuser, Basel, 1982) 38. F. Tréves, Applications of distributions to PDE theory. Am. Math. Mon. 77, 241–248 (1970) 39. E.M. Wright, A prime-representing function. Am. Math. Mon. 58, 616–618 (1951) 40. F.B. Wright, The recurrence theorem. Am. Math. Mon. 68, 247–248 (1961) 41. D. Zagier, A one-sentence proof that every prime p 1.mod 4/ is a sum of two squares. Am. Math. Mon. 97, 144 (1990) 42. L. Zalcman, Real proofs of complex theorems (and vice versa). Am. Math. Mon. 81, 115–137 (1974)

Topological Prerequisites

We briefly recall some basic notions and results that we will use in this book. The proofs may be found in most textbooks on topology, e.g., in Kelley 1965.

Topological Spaces By a topological space we mean a nonempty set X endowed with a topology on X, i.e., a family T of subsets of X that contains ¿ and X and is stable under finite intersections and arbitrary unions. For example, the discrete topology contains all subsets of X, while the anti-discrete topology contains only ¿ and X. The elements of the topology are called the open sets and their complements the closed sets of the topological space. Given a set A in a topological space X, there exists a largest open set contained in A and a largest open set contained in X n A. They are called the interior and exterior of A and denoted by int A and ext A. The remaining set X n .int A [ ext A/ is called the boundary of A and denoted by @A. The three sets int A, ext A, and @A form a partition of X: they are pairwise disjoint, and their union is equal to X. If a 2 int A, then we also say that A is a neighborhood of a. The sets @A and A WD int A [ @A D X n ext A are closed; the latter is the smallest closed set containing A and is called the closure of A. A set D A is said to be dense in A if A D. A topological space X is called separable if it contains a countable dense set. A set K in a topological space X is called compact if every open cover of A has a finite subcover. For example, the finite subsets are compact. Theorem 1 (Cantor’s Intersection Theorem) If .Kn / is a decreasing sequence of nonempty compact sets, then \Kn is nonempty. Let X and Y be two topological spaces. We say that a function f W X ! Y is continuous at a 2 X if for every neighborhood V of f .a/ in Y there exists a

xiii

xiv

Topological Prerequisites

neighborhood U of a in X such that f .U/ V. Furthermore, we say that f is continuous if it is continuous at each point a 2 X. Theorem 2 (Hausdorff) Let X and Y be two topological spaces and f W X ! Y. (a) f is continuous ” the preimage f 1 .V/ of every open set V Y is open in X, or equivalently, if the preimage f 1 .F/ of every closed set F Y is closed in X. (b) If K X is compact and f is continuous, then f .K/ Y is compact, i.e., the continuous image of a compact set is compact. The last result implies another important theorem: Theorem 3 (Weierstrass) Let X be a compact topological space and f W X ! R a continuous function. Then f is bounded; moreover, it has maximal and minimal values. If Z is a nonempty subset of a topological space X, then there exists a smallest topology on Z such that the embedding1 of Z into X is continuous. This is called the subspace topology of Z. A nonempty set in a topological space X is compact ” the corresponding subspace topology is compact. A closed subspace of a compact space is also compact. A topological space X is called separated or a Hausdorff space if any two distinct points of X belong to two disjoint open sets. Hausdorff spaces have many open and closed sets; in particular, the compact sets of Hausdorff spaces are always closed. A topological space X is called connected if ¿ and X are the only sets that are simultaneously open and closed. A nonempty subset of a topological space X is called connected if it is connected as a subspace. The empty set is also considered to be connected. Theorem 4 (a) The closure of a connected set is also connected. (b) If a family of connected sets Ci has a nonempty intersection, then [Ci is also connected. (c) (Bolzano) The continuous image of a connected set is connected. If X is the direct product of an arbitrary nonempty family of topological spaces Xi , then there exists a smallest topology on X such that all projections X ! Xi are continuous. This is called the (Tychonoff ) product of the spaces Xi . Theorem 5 (a) (Tychonoff) The product of compact spaces is compact. (b) The product of connected spaces is connected. (c) The product of separated spaces is separated.

1

The embedding of Z into X is the function Z 3 z 7! z 2 X.

Topological Prerequisites

xv

Many topological properties may be conveniently characterized by a generalization of convergent sequences. By a net in a set X we mean a function x W I ! X where I is endowed with a partial ordering , i.e., a reflexive and transitive binary relation having the following extra property: for any i; j 2 I there exists a k 2 I satisfying k i and k j. We often write xi instead of x.i/ and .xi / instead of x. We say that a net .xi / converges to a point a in a topological space X if for each open set U X containing a, the net .xi / eventually belongs to U, i.e., there exists a j 2 I such that xi 2 U for all i j. Then we write xi ! a or lim xi D a, and a is called a limit of .xi /. Proposition 6 Let X and Y be topological spaces and a 2 A X. (a) a 2 A ” there exists a net in A converging to a. (b) A is closed ” no net in A converges to any point of X n A. (c) A function f W X ! Y is continuous at a ” lim f .xi / D f .a/ in Y for every converging net lim xi D a in X. (d) X is a Hausdorff space ” no net has more than one limit. In order to characterize compactness, we introduce accumulation points and subnets. By a subnet of a net x W I ! X, we mean a net x ı f W J ! X where f W J ! I is a function having the following property: for every i 2 I there exists a j 2 J such that k j H) f .k/ i. We say that a is an accumulation point of a net .xi / in a topological space X if for each open set U X containing a, the net .xi / often belongs to U, i.e., for every i 2 I there exists a j i such that xj 2 U. Proposition 7 Let X be a topological space and let a 2 A X. (a) a is an accumulation point of a net .xi / ” there exists a subnet converging to x. (b) A is compact ” each net in A has at least one accumulation point in A. (c) Equivalently, A is compact ” each net in A has a subnet converging to some point of A.

Metric Spaces By a metric on a nonempty set X, we mean a nonnegative and symmetric function d W X X ! R satisfying the relation d.x; y/ D 0 ” x D y, and the triangle inequality d.x; y/ d.x; z/ C d.z; y/ for all x; y; z 2 X. By a metric space we mean a nonempty set X endowed with a metric.

xvi

Topological Prerequisites

For example, the usual distance d.x; y/ WD jx yj between real numbers is a metric on R, and the Euclidean distance between the points of Rn is a metric on Rn . The discrete metric on an arbitrary nonempty set X is defined by d.x; x/ D 0 for all x 2 X, and d.x; y/ D 1 whenever x ¤ y. Every metric space has a natural topology as follows. By a ball of radius r > 0 centered at a 2 X, we mean the set Br .a/ WD fx 2 X W d.x; a/ < rg. A set U X is called open if for each a 2 U there exists an r > 0 such that Br .a/ U. Then the balls are open. In this way every metric space is a Hausdorff space. We define the diameter of a set A in a metric space by the formula diam A WD sup fd.x; y/ W x; y 2 Ag. A set A is called bounded if diam A < 1. If K is a nonempty set and X is a metric space, then the bounded functions f W K ! X form a metric space B.K; X/ with respect to the metric d1 .f ; g/ WD sup d.f .t/; g.t//: t2K

The boundedness of f means that its range (or image) is a bounded set in X. In metric spaces the convergence xi ! a is equivalent to d.xi ; a/ ! 0. The nets and subnets may be replaced by sequences (nets defined on I D N) and subsequences (subnets x ı f with an increasing function f W N ! N): Proposition 8 Let X and Y be metric spaces and a 2 A X. (a) a 2 A ” there exists a sequence in A converging to a. (b) A is closed ” no sequence in A converges to any point of X n A. (c) A function f W X ! Y is continuous at a ” lim f .xi / D f .a/ in Y for every converging sequence lim xi D a in X. (d) a is an accumulation point of a sequence ” there exists a subsequence converging to x. (e) A is compact ” each sequence in A has at least one accumulation point in A. (f) Equivalently, A is compact ” each sequence in A has a subsequence converging to some point of A. We will often use the following properties of compact sets: Proposition 9 Consider two nonempty compact sets K; L in a metric space. (a) The diameter of K is attained: there exist a; b 2 K such that diam K D d.a; b/. (b) The distance between K and L is attained: there exist a 2 K and b 2 L such that d.a; b/ d.x; y/ for all x 2 K and y 2 L. An important property of compact metric spaces is the following: Theorem 10 (Heine) Let .X; d/; .X 0 ; d0 / be two metric spaces and f W X ! X 0 a continuous function. If X is compact, then f is uniformly continuous, i.e., for each " > 0 there exists a ı > 0 such that x; y 2 X

and d.x; y/ < ı H) d0 . f .x/; f .y// < ":

Topological Prerequisites

xvii

Next we study the metric spaces for which the Cauchy criterion may be generalized. A sequence in a metric space is called a Cauchy sequence if diam fxk W k ng ! 0 as n ! 1. Every convergent sequence is a Cauchy sequence. A metric space is called complete if, conversely, every Cauchy sequence is convergent. For example, the discrete metric spaces are complete, and the spaces Rn are complete with respect to the Euclidean metrics. If X is a complete metric space, then the metric spaces B.K; X/ are complete. Cantor’s intersection theorem has a useful variant: Theorem 11 (Cantor’s Intersection Theorem) Let .Fn / be a decreasing sequence of nonempty closed sets in a complete metric space. If diam Fn ! 0, then \Fn is nonempty. Next we consider a strengthening of uniform continuity. Let .X; d/ and .X 0 ; d0 / be two metric spaces. A function f W X ! X 0 is Lipschitz continuous if there exists a constant L such that d 0 .f .x/; f .y// Ld.x; y/ for all x; y 2 X. If, moreover, L < 1, then f is called a contraction. Theorem 12 (Banach–Cacciopoli) In a complete metric space X, every contraction f W X ! X has a unique fixed point, i.e., a point a 2 X satisfying f .a/ D a. The following extension theorem is often applied in classical analysis, for example, to define integrals of continuous functions. Theorem 13 Let X; X 0 be two metric spaces, A X and f W A ! X 0 a uniformly continuous function. If X 0 is complete, then f may be extended in a unique way to a uniformly continuous function F W A ! X 0 . If, moreover, f is Lipschitz continuous, then F is Lipschitz continuous with the same constant L. Every metric space may be completed. More precisely: Theorem 14 For every metric space X, there exists a complete metric space X 0 and an isometry f W X ! X 0 such that f .X/ is dense in X 0 . The isometry means that f preserves the distances. This completion is essentially unique. A nonempty subset of a metric space may be considered as a metric subspace with respect to the restriction of the metric to this set. A set in a metric space is called complete if it is empty or if the corresponding metric subspace is complete. A complete set is always closed, and a closed subspace of a complete metric space is also complete. For example, if K is a topological space and X is a metric space, then the continuous functions in B.K; X/ form a closed subspace Cb .K; X/. If X is complete, then Cb .K; X/ is also complete. We end this section with another characterization of compactness.

xviii

Topological Prerequisites

A set A in a metric space is called totally (or completely) bounded if for each fixed " > 0 it has a finite cover by sets of diameter < " or, equivalently, if for each fixed r > 0 it has a finite cover by balls of radius r. Theorem 15 (a) A set A in a metric space is compact ” it is complete and totally bounded. (b) A set A in a complete metric space is compact ” it is closed and totally bounded.

Normed Spaces By a seminorm on a vector space X, we mean a nonnegative, positively homogeneous function p W X ! R satisfying p.0/ D 0 and the triangle inequality p.x C y/ p.x/ C p.y/ for all x; y 2 X. If we have also p.x/ > 0 for all x ¤ 0, then p is called a norm, and we often write kxk instead of p.x/. A normed space is a vector space X endowed with a norm. Every normed space is also a metric (and hence a topological) space with respect to the metric d.x; y/ WD kx yk. For example, Rn is a normed space with respect to each of the norms kxkp WD .jx1 jp C C jxn jp /1=p

.1 p < 1/

and kxk1 WD max fjx1 j ; : : : ; jxn jg : If I is a non-degenerate compact interval in R, then the vector space C.I; R/ of continuous functions f W I ! R is a normed space with respect to each of the norms Z kf kp WD

jf j

p

1=p

.1 p < 1/ and

kf k1 WD sup jf j :

I

If X is a normed space, then B.K; X/ is a normed space for every nonempty set K, and Cb .K; X/ is a normed space for every topological space X. If X; Y are normed spaces, then the continuous linear maps A W X ! Y form a normed space L.X; Y/ with respect to the norm kLk WD sup fkAxkY W x 2 X; kxkX 1g : More generally, for each positive integer k the continuous k-linear maps A W X k ! Y form a normed space Lk .X k ; Y/ with respect to the norm

Topological Prerequisites

xix

kLk WD sup fkA.x1 ; : : : ; xk /kY W xi 2 X

and

kxi kX 1;

i D 1; : : : ; kg :

Let X; Y be normed spaces, U X a nonempty open set, and k a positive integer, and consider the set Cbk .U; Y/ of Ck functions f W U ! Y for which f and its derivatives f .j/ W U ! Lj .X j ; Y/ are bounded for j D 1; : : : ; k. Then Cbk .U; Y/ is a normed space with respect to the norm kf k WD kf k1 C f 0 1 C C f .k/ 1 : By a scalar product on a vector space X, we mean a nonnegative, symmetric bilinear functional .; / W X X ! R satisfying .x; x/ > 0 whenever x ¤ 0. By a Euclidean space, we mean a vector space endowed with a scalar product. p Every Euclidean space is also a normed space with respect to the norm kxk WD .x; x/. Moreover, this norm satisfies the parallelogram identity kx C yk2 C kx yk2 D 2 kxk2 C 2 kyk2 and the Cauchy–Schwarz inequality j.x; y/j kxk kyk for all x; y 2 X. The balls of normed spaces are convex, i.e., if x; y 2 Br .a/, then the whole segment Œx; y WD ftx C .1 t/y W 0 t 1g lies in Br .a/. The connected open sets have a simple geometric characterization in normed spaces. By a broken line in a vector space, we mean a finite union of segments L WD [kiD1 Œxi1 ; xi . We say that it connects x0 and xk , and we say that it lies in a set U if L U. Proposition 16 An open set U in a normed space X is connected ” any two points a; b 2 U may be connected by a broken line lying in U. The theory of finite-dimensional normed spaces is considerably simplified by the following results: Theorem 17 (Tychonoff) (a) On a finite-dimensional vector space X, all norms are equivalent, i.e., for any two norms kk and kk0 there exist two positive constants c1 ; c2 such that c1 kxk kxk0 c2 kxk for all x 2 X. (b) Consequently, if X is a finite-dimensional normed space, then • X is complete. • Every bounded set in X is totally bounded.

xx

Topological Prerequisites

• A set in X is compact ” it is bounded and closed. • X is separable. • Every bounded sequence in X has a convergent subsequence. (c) Every linear map A W X ! Y, where X; Y are normed spaces and X is finitedimensional, is continuous. We emphasize that the Bolzano–Weierstrass theorem remains valid in every finite dimensional normed space.

Part I

Functional Analysis

Geometrical and physical problems led to the birth of functional analysis at the end of the nineteenth century. Following the works of Dini, Ascoli, Peano, Arzelà, Volterra, Hadamard and then the spectacular discoveries of Fredholm, Hilbert, Riesz, Fréchet and Helly, Banach laid the foundations of this new theory. It was later enriched by Hahn, von Neumann and many others. In addition to its inner beauty, it proved to be very useful in, among other areas, the calculus of variations, the theory of partial differential equations and in quantum mechanics. Instead of following the historical development,1 we will try to extend some wellknown results of Euclidean geometry to infinite-dimensional spaces: • if K is a non-empty convex, closed set in RN , then K has a closest point to each x 2 RN ; • for every proper subspace2 M of RN there exists a point x such that dist.x; M/ D jxj D 1; • two non-empty disjoint convex sets of RN may always be separated by an affine hyperplane; • every bounded convex polytope is the convex hull of its vertices; • every bounded sequence in RN has a convergent subsequence. This road will lead in a natural way to many deep theorems but also to surprising counterexamples. The more general the space, the more counter-intuitive the phenomena that appear. We start our investigations with Hilbert spaces, the closest to RN . We follow with the wider class of Banach spaces. Then we shortly investigate the still more general locally convex spaces: they play an important role in the theory of distributions, the basic framework for the study of linear partial differential

1

The last two chapters of this book are devoted mostly to the Lebesgue integral and its applications. In this book by a subspace without adjective we always mean a linear subspace. In case of metric or topological subspaces we will always write metric subspace or topological subspace. 2

2

I Functional Analysis

equations. We end our tour by exhibiting some strange properties of general topological vector spaces. From the immense literature we mention for further studies the classical monographs of Banach [24] and Riesz–Sz.-Nagy [394]: after many decades, they still keep their freshness and elegance. Many additional theoretical results can be found in [2, 32, 35, 40, 97, 117, 119, 254, 266, 285, 309, 321, 349, 367, 397, 403, 406, 411, 488], exciting historical aspects are given in [45, 106, 117, 144, 203, 316, 327, 367, 394, 431, 490], and many exercises are contained in [15, 117, 187, 249, 349, 367, 403, 406, 458].

Chapter 1

Hilbert Spaces

The infinite! No other question has ever moved so profoundly the spirit of man. –D. Hilbert

Stimulated by Fredholm’s discovery of an unexpectedly simple and general theory of integral equations in 1900, Hilbert developed a general theory of infinitedimensional inner product spaces between 1904 and 1906. This allowed him to solve several important problems of mathematical physics. His student Schmidt replaced his algebraic formulation by a more intuitive geometric language, making the theory accessible to a wider public. We may define the notion of orthogonality, and many results of plane geometry, such as Pythagoras’ theorem, remain valid. Hilbert spaces appear today in almost all branches of mathematics and theoretical physics: since the fundamental works of von Neumann,1 they have formed the mathematical framework of quantum mechanics. We give here an introduction to this theory.

1.1 Definitions and Examples Let X be a real vector space. We recall some basic definitions and properties. By a norm2 in X we mean a function kk W X ! R satisfying for all x; y; z 2 X and 2 R the following properties:

1 2

kxk 0;

kxk D 0 ” x D 0;

von Neumann [334, 337]. Riesz [383]. Notation of Schmidt [416].

© Springer-Verlag London 2016 V. Komornik, Lectures on Functional Analysis and the Lebesgue Integral, Universitext, DOI 10.1007/978-1-4471-6811-9_1

3

4

1 Hilbert Spaces

Fig. 1.1 Triangle inequality

x+y

y

x

kxk D jj kxk ;

kx C yk kxk C kyk :

The last property is called the triangle inequality; see Fig. 1.1. By a normed space we mean a vector space endowed with a norm. The norm is continuous with respect to the corresponding topology. By a scalar product in X we mean a function .; / W X X ! R satisfying for all x; y; z 2 X and ˛; ˇ 2 R the following properties:

.˛x C ˇy; z/ D ˛.x; z/ C ˇ.y; z/;

.x; y/ D .y; x/;

.x; x/ 0;

.x; x/ D 0 ” x D 0:

By a Euclidean or prehilbert space we mean a vector space endowed with a scalar product. Every Euclidean space has a natural norm: kxk WD .x; x/1=2 . This norm satisfies the Cauchy–Schwarz inequality: j.x; y/j kxk kyk and the parallelogram identity: kx C yk2 C kx yk2 D 2 kxk2 C 2 kyk2 : Finally, the scalar product is continuous with respect to the corresponding topology: if xn ! x and yn ! y, then .xn ; yn / ! .x; y/.

1.1 Definitions and Examples

5

Definition By a Hilbert space3 we mean a complete Euclidean space. Examples • We recall from topology that RN is a Euclidean space with respect to the natural scalar product .x; y/ WD x1 y1 C x2 y2 C C xN yN : Since every finite-dimensional normed space is complete, RN is a Hilbert space. • P The set `2 of sequences x D .xn / of real numbers satisfying the condition jxn j2 < 1 is a Hilbert space with respect to the scalar product .x; y/ WD

1 X

xn yn :

nD1

First of all, the inequalities 1 X nD1

1

jxn yn j

1

1X 1X jyn j2 < 1; jxn j2 C 2 nD1 2 nD1

and 1 X

j˛xn C ˇyn j2 2j˛j2

nD1

1 X

jxn j2 C 2jˇj2

nD1

1 X

jyn j2 < 1

nD1

(for arbitrary ˛; ˇ 2 R) imply that `2 is a vector space, and that .x; y/ is a correctly defined scalar product. Now let .x1n /, .x2n /, . . . be a Cauchy sequence in `2 . For every fixed " > 0 there exists a k0 such that 1 X

jxkn x`n j2 < "

(1.1)

nD1

for all k; ` k0 . In particular, .x`n / is a Cauchy sequence for every fixed n, and therefore converges to some real number xn . Letting ` ! 1 we deduce from (1.1) the inequality N X

jxkn xn j2 "

nD1

3

Hilbert [208], von Neumann [334], Löwig [312], and Rellich [368].

6

1 Hilbert Spaces

for every k k0 and N 1. Letting N ! 1 this yields .xn / 2 `2 and .xkn / ! .xn / in `2 . Many metric and topological properties of finite-dimensional normed spaces remain valid in all Hilbert spaces. But we have to be careful: there are important exceptions. Before giving some examples, we recall some compactness results in finite-dimensional spaces. We recall from topology that a subset K of a normed (or metric) space is compact if every sequence .xk / K has a subsequence, converging to some element of K. For example, every finite set is compact. Theorem 1.1 (a) (Kürschák)4 Every sequence of real numbers has a monotone subsequence. (b) (Bolzano–Weierstrass)5 Every bounded sequence of real numbers has a convergent subsequence. Proof (a) An element of the sequence .xk / is called a peak if it is larger than all later elements: xk > xm for all m > k. If there are infinitely many peaks, then they form a decreasing subsequence. Otherwise, there exists an index N such that no element xk with k N is a peak. This allows us to define by induction a non-decreasing subsequence. (b) There exists a bounded and monotone subsequence by (a). Its convergence follows from the axioms of real numbers. t u Corollary 1.2 Let X be a finite-dimensional normed space. (a) Every bounded sequence .xk / X has a convergent subsequence. (b) A subset of X is compact ” it is bounded and closed. (c) The distance between two non-empty bounded and closed sets of X is always attained. (d) The diameter of a non-empty bounded and closed set of X is always attained. (e) Every (linear) subspace of X is closed.6 (f) X is complete. Sketch of Proof (a) For X D RN endowed with the usual Euclidean norm the results easily follows from the one-dimensional case by observing that convergence in norm is equivalent to component-wise convergence.

4

Kürschák [275]. This elegant result and its combinatorial proof seems to be little known. Bolzano [54] and Weierstrass [482]. 6 We recall that, in this book, by a subspace without adjective we always mean a linear subspace. 5

1.1 Definitions and Examples

7

The general case hence follows by a theorem of Tychonoff7: on a finitedimensional vector space all norms are equivalent. (b)–(f) easily follow from (a). t u All these properties may fail in infinite dimensions: *Examples We show that properties (a)–(e) fail in H WD `2 . (a) The vectors k1

‚ …„ ƒ ek D .0; : : : ; 0; 1; 0; : : :/;

k D 1; 2; : : :

form a bounded sequence in `2 because kek k D 1 for all k. But this sequence has no convergent subsequence. Indeed, we have p 2 whenever k ¤ m, so that no subsequence satisfies the kek em k D Cauchy convergence criterion. (b) The previous example also shows that the closed unit ball of `2 , although bounded and closed, is not compact. (c) The subset 80 k1 9 1 < ‚ …„ ƒ k C 1 = F WD @0; : : : ; 0; ; 0; : : :A W k D 1; 2; : : : : ; k of `2 is non-empty, bounded and closed, but it has no element of minimal norm, i.e., its distance from 0 is not attained: we have dist.0; F/ D 1, but kyk > 1 for every y 2 F. (d) The subset 1 n o X 1 2 K WD x 2 `2 W 1C jxn j2 1 n nD1

of `2 is non-empty, convex, bounded and closed,8 but it has no element of maximal norm. Moreover, the diameter of K is not attained: we have diam K D 2, but kx yk < 2 for all x; y 2 K. (e) The proper subspace n

2

M WD x 2 ` W

1 X

xn D 0

o

nD1

of `2 is dense. 7 8

Tychonoff [454]. Observe that K is the inverse image of the closed unit ball by a continuous linear map.

8

1 Hilbert Spaces

For the proof we fix an arbitrary ball Br .x/. We choose first a large positive integer m such that k.0; : : : ; xmC1 ; xmC2 ; : : :/k < r=2; and then a large positive integer k such that jx1 C C xm j < vector k ‚ …„ ƒ y WD x1 ; : : : ; xm ; c; : : : ; c; 0; 0; : : : ;

cD

p

kr=2. Then the

x1 C C xm k

belongs to M, and kx yk k.0; : : : ; xmC1 ; xmC2 ; : : :/k m k ‚ …„ ƒ ‚ …„ ƒ C 0; : : : ; 0; c; : : : ; c; 0; 0; : : : < r:

Corollary 1.2 (f) may also fail in infinite dimensions: Examples (a) Consider the subspace X spanned by the vectors ek of the first example above: the elements .xn / ofPX have at most a finite number of non-zero components. The formula uk WD knD1 n1 en defines a Cauchy sequence .uk / in X because kuk um k2 D

1 X 1 1 !0 2 n n2 nDmC1 nDmC1 k X

as k > m ! 1. But .uk / does not converge to any point x 2 X. Indeed, each x D .xn / 2 X has a zero element xn D 0. Therefore kuk xk2

1 n2

for all k n, so that kuk xk 6! 0. (b) A more natural example is given if we take a non-degenerate compact interval I, and we endow the vector Rspace C.I/ of continuous functions x W I ! R with the scalar product .x; y/ WD I xy dt. To prove that this space is not complete, we assume for simplicity that I D Œ0; 2, and we consider the functions xn .t/ WD med f0; n.t 1/; 1g ;

0 t 2;

n D 1; 2; : : : ;

1.1 Definitions and Examples

9

Fig. 1.2 Graph of xn

1

0

1+ 1

1

2

(see Fig. 1.2), where med fx; y; zg denotes the middle number among x, y and z. For x z we have med fx; y; zg D max fx; min fy; zgg : If m > n ! 1, then 2

Z

kxm xn k D

.nC1/=n 1

jxm .t/ xn .t/j2 dt

1 ! 0; n

so that .xn / is a Cauchy sequence. Assume on the contrary that it converges to some x 2 C.I/. Since x is continuous, then we deduce from the estimate Z

1

0

jx.t/j2 dt D

Z

1 0

jx.t/ xn .t/j2 dt kx xn k2 ! 0

that x 0 in Œ0; 1; in particular, x.1/ D 0. On the other hand, for arbitrary integers n N 1 we have Z

2 .NC1/=N

jx.t/ 1j2 dt D

Z

2 .NC1/=N

jx.t/ xn .t/j2 dt kx xn k2 :

Letting n ! 1 and then N ! 1, we get Z

2 .NC1/=N

jx.t/ 1j2 dt D 0;

Z

2

and then 1

jx.t/ 1j2 dt D 0:

10

1 Hilbert Spaces

Hence x 1 in Œ1; 2, contradicting the previous equality x.1/ D 0. Our last examples show the importance of the following result: Proposition 1.3 Every Euclidean space E may be completed. More precisely, there exists a Hilbert space H and an isometry f W E ! H such that f .E/ is dense in H. First we recall for convenience the corresponding result for metric spaces: Proposition 1.4 (Hausdorff)9 For any given metric space .X; d/ there exists a complete metric space .X 0 ; d0 / and an isometry h W X ! X 0 . Remark The isometry h enables us to identify .X; d/ with the metric subspace h.X/ of .X 0 ; d0 /. Proof Consider the complete metric space .X 0 ; d0 / WD B.X/ of bounded functions f W X ! R with respect to the uniform distance d1 . f ; g/ WD sup j f .x/ g.x/j : x2X

Fix an arbitrary point a 2 X. For each x 2 X the formula hx .y/ WD d.x; y/ d.a; y/;

y2X

defines a function hx 2 B.X/, because jhx .y/j d.x; a/ for all y 2 X by the triangle inequality. Since jhx .z/ hy .z/j D jd.x; z/ d.y; z/j d.x; y/ for all z 2 X, we have d 0 .hx ; hy / d.x; y/ for all x; y 2 X. In fact, this is an equality, because for z D y we have jhx .y/ hy .y/j D d.x; y/: t u

9 Hausdorff [195]. The short proof given here, based on an idea of Fréchet [157, p. 161], is due to Kuratowski [273]. If the metric d is bounded, then the proof may be further shortened by simply taking hx .y/ WD d.x; y/.

1.2 Orthogonality

11

Proof of Proposition 1.3 Every Euclidean space E is a metric space with respect to the distance d.x; y/ WD kx ykE D .x y; x y/1=2 ; and thus it can be considered as a dense metric subspace of a suitable complete metric space .H; d/. For any fixed x; y 2 H and c 2 R we choose two sequences .xn / and .yn / in E such that d.x; xn / ! 0 and d.y; yn / ! 0, and then we set x C y WD lim.xn C yn /; cx WD lim cxn ; .x; y/ WD lim.xn ; yn /: One may readily check that • • • •

the limits exist; they do not depend on the particular choice of .xn / and .yn /; H is a Euclidean and thus a Hilbert space with respect to this scalar product; d.x; y/ D .x y; x y/1=2 for all x; y 2 H. t u

Definition We denote by L2 .I/ the Hilbert space obtained by the completion of C.I/.10 *Remark The Lebesgue integral will provide a more concrete interpretation of L2 .I/.11 Henceforth, until the end of this chapter the letter H always denotes a Hilbert space.

1.2 Orthogonality Definition Let x; y 2 H and A; B H. We say that • x and y are orthogonal if .x; y/ D 0; • x and A are orthogonal if .x; y/ D 0 for all y 2 A; • A and B are orthogonal if .x; y/ D 0 for all x 2 A and y 2 B. We express these relations by the symbols x ? y, x ? A and A ? B. Now we solve the first problem of the introduction. 10

As in the case of metric spaces, the proof shows that the completion is essentially (up to isomorphism) unique. 11 See Proposition 9.5 (b), p. 312.

12

1 Hilbert Spaces

Theorem 1.5 (Orthogonal Projection)12 Let K H be a non-empty convex, closed set, and x 2 H. (a) There exists in K a unique closest point y to x. It is characterized by the following properties: y 2 K;

and .x y; v y/ 0 for every v 2 K:

(1.2)

(b) The formula PK x WD y defines a Lipschitz continuous function PK W H ! K with some Lipschitz constant L 1. (c) If K is a subspace, then (1.2) is equivalent to the orthogonality property x y ? K;

(1.3)

and PK is a bounded linear map of norm 1. Definition The point y D PK .x/ is called the orthogonal projection of x onto K (see Fig. 1.3). Proof Existence. Set d D dist.x; K/, and consider a minimizing sequence .yn / K satisfying kx yn k ! d. This is a Cauchy sequence. Indeed, by the Fig. 1.3 Orthogonal projection

y

12

Levi [300], Schmidt [416], Nikodým [343] (statement), [344] (proof), and Riesz [389].

x

1.2 Orthogonality

13

parallelogram identity we have k.x yn / .x ym /k2 C k.x yn / C .x ym /k2 D 2 kx yn k2 C 2 kx ym k2 : Using the definition of d this implies 2 kym yn k2 D 2 kx yn k2 C 2 kx ym k2 4 x 21 .ym C yn / 2 kx yn k2 C 2 kx ym k2 4d 2 ; because 21 .ym C yn / belongs to the convex set K. It remains to observe that the right-hand side tends to zero as m; n ! 1. The limit y of the sequence belongs to K because K is closed, and we have kx yk D d by the continuity of the norm. Characterization and uniqueness. Let y 2 K be at a minimal distance d from x. For any fixed v 2 K the vectors .1 t/y C tv D y C t.v y/ belong to the convex set K for all 0 < t < 1, so that 0 t1 .kx yk2 kx y t.v y/k2 / D 2.x y; v y/ t kv yk2 : Letting t ! 0 this yields (1.2). Conversely, if (1.2) holds and v 2 K is different from y, then kx vk2 D kx yk2 C ky vk2 2.x y; v y/ kx yk2 C ky vk2 > kx yk2 : Lipschitz property. If x; x0 2 H, then writing y D PK .x/ and y0 D PK .x0 / we have .x y; y0 y/ 0

and .x0 y0 ; y y0 / 0:

Summing them we get .x x0 C y0 y; y0 y/ 0I hence 0 y y2 .x0 x; y0 y/ x0 x y0 y

14

1 Hilbert Spaces

and therefore 0 y y x0 x : The case when K is a subspace. Let w 2 K. Applying (1.2) with v D y ˙ w we obtain .x y; ˙w/ 0; and hence .x y; w/ D 0. Conversely, (1.3) implies .x y; v y/ D 0 because v y 2 K. The linearity of PK follows from its uniqueness. Indeed, if y D PK .x/, y0 D PK .x0 / and 2 R, then the relations x y ? K and x0 y0 ? K imply .x C x0 / .y C y0 / ? K

and x y ? K: t u

*Example The example of the set F in the preceding section shows that the convexity assumption is necessary also for the existence of the orthogonal projection. In order to state some corollaries we introduce two new notions: Definitions • The orthogonal complement of a set D H is defined by the formula13 D? WD fx 2 H W x ? Dg : • The closed subspace spanned by a set D H is by definition the intersection of all closed subspaces containing D.14 Observe that D? is a closed subspace of H, and that A B H) B? A? ;

.A [ B/? D A? \ B? :

Notice also that the closed subspace spanned by D is the closure of the set of all finite linear combinations formed by the points of D.

13 For instance, the orthogonal complement of a k-dimensional subspace in Rn is an .n k/dimensional subspace. 14 This is clearly the smallest closed subspace containing D.

1.2 Orthogonality

15

Part (b) of the following result solves the second problem of the introduction: Corollary 1.6 (a) (Riesz)15 Let M H be a non-empty closed subspace. Every x 2 H has a unique decomposition x D y C z with y 2 M and z 2 M ? . Consequently, M D M ?? . (b) Let M H be a non-empty proper closed subspace. There exists an x 2 H such that dist.x; M/ D kxk D 1: (c) The closed subspace spanned by D H is equal to D?? . Consequently, • if D? D f0g, then D spans H; • if M ? D f0g for some subspace M H, then M is dense in H. See Figs. 1.4 and 1.5. Proof (a) Existence. We have y WD PM x 2 M by definition, and z WD x y 2 M ? by (1.3). Uniqueness. If x D y C z and x D y0 C z0 are two decompositions with y; y0 2 M and z; z0 ? M, then w WD y y0 D z0 z 2 M \ M ? : Hence .w; w/ D 0, thus w D 0, and therefore x D x0 and y D y0 . Fig. 1.4 Orthogonal decomposition

M⊥

x

z

y

0

15

Riesz [389].

M

16

1 Hilbert Spaces

Fig. 1.5 dist.x; M/ D kxk

x

M 0

If x 2 M, then x is orthogonal to every z 2 M ? , i.e., x 2 M ?? . Conversely, if x 2 M ?? and x D y C z is its decomposition with y 2 M and z 2 M ? , then x y D z belongs to M ? but also to M ?? because M M ?? . Hence x y D z D 0, and therefore x D y 2 M. (b) Choosing y 2 H n M arbitrarily, x WD .y PM y/= ky PM yk has the required property. (c) The closed subspace M spanned by D satisfies D? D M ? and thus D?? D M ?? . Using (a) we conclude that D?? D M. t u

1.3 Separation of Convex Sets: Theorems of Riesz–Fréchet and Kuhn–Tucker In a finite-dimensional vector space X two disjoint non-empty convex sets may always be separated by an affine hyperplane, i.e., by a set of the form fx 2 X W '.x/ D cg ; where ' W X ! R is a non-zero linear functional, and c 2 R. More precisely, the following result holds: *Proposition 1.7 (Minkowski)16 Let A and B be two disjoint non-empty convex sets in a finite-dimensional vector space X. There exist a non-zero linear functional ' on X and a real number c such that '.a/ c '.b/ for every a 2 A 16

Minkowski [324, 325].

and b 2 B:

(1.4)

1.3 Separation of Convex Sets . . .

17

First we establish a weaker property that holds in all Hilbert spaces. We recall that we denote by X 0 the dual space of a normed space X, i.e., the space of continuous linear functionals on X.17 Theorem 1.8 (Tukey)18 Let A and B be two disjoint non-empty convex, closed sets in H. If at least one of them is compact, then there exist ' 2 H 0 and c1 ; c2 2 R such that '.a/ c1 < c2 '.b/ for all a 2 A

and b 2 B:

(1.5)

(See Fig. 1.6.) In particular, for two distinct points a; b 2 H there exists a ' 2 H 0 such that '.a/ ¤ '.b/.

Proof The set C WD B A D fb a W a 2 A; b 2 Bg is non-empty convex, closed, and 0 … C. The only nontrivial property is its closedness: we have to show that if a sequence of the form .bn an / converges Fig. 1.6 Separation of convex sets

B

A

17 The terminology of bounded linear maps and bounded linear functionals is frequently used instead of continuous linear maps and continuous linear functionals. 18 Tukey [460].

18

1 Hilbert Spaces

to some point x in H, then x 2 C. Assuming for example that A is compact, there exists a convergent subsequence ank ! a 2 A. Then we have bnk D .bnk ank / C ank ! x C a: Since B is closed, x C a 2 B, and therefore x D .x C a/ a 2 B A D C. Let us denote by y the orthogonal projection of 0 to C; then y ¤ 0 (because 0 … C), and .0 y; b a y/ 0

for all a 2 A

and b 2 B;

i.e., kyk2 C .a; y/ .b; y/ for all

a2A

and b 2 B:

The formula '.x/ WD .x; y/ defines a bounded linear functional ' 2 H 0 by the Cauchy–Schwarz inequality. Since A and B are non-empty, we infer from the just obtained inequality that c1 WD sup .a; y/; a2A

and c2 WD inf .b; y/ b2B

are finite numbers, and that (1.5) is satisfied. The last property corresponds to the special case A WD fag and B WD fbg.

t u

*Example The compactness assumption cannot be omitted.19 To see this we consider in H WD `2 the non-empty convex, closed sets ˚ j x1 A WD .xn / 2 `2 W njx2=3 n

for every n 2

and ˚ B WD .xn / 2 `2 W xn D 0

for every n 2 :

They are disjoint because a sequence .xn / 2 A \ B should satisfy the inequality x1 n1=3 for every n 2, while xn ! 0 and n1=3 ! 1. If A and B could be separated by a closed affine hyperplane, then A B would belong to a closed halfspace. This is, however, impossible, because A B is dense in `2 . This can be seen by using the relation ˚ A B D .xn / 2 `2 W x2=3 D O.1=n/ : n

19

Tukey [460].

1.3 Separation of Convex Sets . . .

19

For any fixed .zn / 2 `2 and " > 0 choose a large m such that X

X

jzn j2 < "2 =4 and

n>m

n4=3 < "2 =4:

n>m

Then the formula ( xn WD

if n m,

zn n

2=3

if n > m

defines a sequence .xn / 2 A B for which 1 X nD1

jxn zn j2

1=2

X

n4=3

1=2

C

n>m

X

jzn j2

1=2

< ":

n>m

The bounded linear functional ' obtained in the proof of Theorem 1.8 is represented by a vector y 2 H. Next we establish the very important fact that every bounded linear functional on H has this form. If y 2 H, then the formula 'y .x/ WD .x; y/ defines a bounded linear functional 'y 2 H 0 for which 'y kyk, because j'y .x/j kyk kxk for every x 2 H by the Cauchy–Schwarz inequality. Setting j.y/ WD 'y we obtain therefore a map j of H into H 0 . This map is linear by the bilinearity of the scalar product. Theorem 1.9 (Riesz–Fréchet)20 The map j is an isometric isomorphism of H onto H 0 . It follows from the theorem that H 0 is also a Hilbert space; using the theorem, H 0 is often identified with H. kyk for every y. The equality j'y .y/j D kyk2 Proof We already know that ' y implies the converse inequality 'y kyk. Hence j is an isometry; it remains to prove the surjectivity.

20

Riesz [373], Fréchet [155, 156] for L2 , Riesz [389] for the general case.

20

1 Hilbert Spaces

The kernel M D N.'/ WD fx 2 H W '.x/ D 0g of any ' 2 H 0 is a closed subspace. If M D H, then ' D 'y with y D 0. If M ¤ H, then applying Corollary 1.6 (p. 15) we may fix a unit vector e, orthogonal to M. We have '.e/x '.x/e 2 M for every x 2 H because ' .'.e/x '.x/e/ D '.e/'.x/ '.x/'.e/ D 0: By the choice of e this implies 0 D .'.e/x '.x/e; e/ D '.e/.x; e/ '.x/.e; e/ D .x; '.e/e/ '.x/; i.e., ' D 'y with y D '.e/e.

t u

Let us return to Minkowski’s theorem. Proof of Proposition 1.7 Let us endow X with a Euclidean norm. As a finitedimensional space, X is separable, hence the metric subspaces A and B are separable, too. We may therefore fix a dense sequence .an / in A and a dense sequence .bn / in B. Let us denote by An and Bn the convex hulls of a1 ; : : : ; an and b1 ; : : : ; bn , for n D 1; 2; : : : : The sets An , Bn are compact because they are the images of the compact21 simplex f.t1 ; : : : ; tn / 2 Rn W t1 0; : : : ; tn 0; t1 C C tn D 1g by the continuous (linear) maps f ; g W Rn ! X, defined by f .t1 ; : : : ; tn / WD t1 a1 C C tn an

and g.t1 ; : : : ; tn / WD t1 b1 C C tn bn :

Since An A and Bn B are disjoint, by Theorem 1.8 there exists a non-zero functional 'n 2 X 0 such that 'n .a/ 'n .b/ for all a 2 An

and b 2 Bn :

Multiplying by a suitable constant we may assume that k'n k D 1.

21

We recall that the finite-dimensional bounded closed sets are compact.

(1.6)

1.3 Separation of Convex Sets . . .

21

Since X 0 is finite-dimensional, there exists a convergent subsequence 'nk ! '. Then we have k'k D 1, so that ' is a non-zero functional. We claim that '.a/ '.b/ for all a 2 A

and b 2 BI

this will yield the proposition with c WD inf f'.b/ W b 2 Bg : Thanks to the density of the sequences .an /, .bn / it is sufficient to show that '.ak / '.bm / for all k; m D 1; 2; : : : : For any fixed k; m, we have 'n .ak / 'n .bm / for all n max fk; mg by (1.6). We conclude by letting n ! 1.

t u

22

*Example Proposition 1.7 does not hold in infinite dimensions. To show this we consider the vector space X of the polynomials and we denote by A the set of polynomials having a (strictly) positive leading coefficient. Then A and B WD f0g are disjoint non-empty convex sets in X. We claim that if (1.4) is satisfied for some linear functional ', then ' 0. Indeed, for any fixed polynomial x choose a positive integer k > deg x, and consider the polynomial ek .t/ WD tk . Then x C ek 2 A, and thus '.x/ C '.ek / c for all 2 R. Hence '.x/ D 0. As an application of Minkowski’s theorem we consider a finite number of convex functions f0 ; : : : ; fn W K ! R defined on a convex subset of a vector space X, and we investigate the minima of the restriction of f0 to the convex subset WD fx 2 K W fi .x/ 0;

i D 1; : : : ; ng :

We are going to prove the following version of the Lagrange multiplier theorem23:

22 23

Dieudonné [105]. See the books on differential calculus.

22

1 Hilbert Spaces

*Theorem 1.10 (Kuhn–Tucker)24 (a) If f0 j has a minimum in a,25 then there exist 0 ; : : : ; n 2 R, not all zero, such that the function 0 f0 C C n fn W K ! R has a minimum in aI

(1.7)

0 ; : : : ; n 0I

(1.8)

i fi .a/ D 0

(1.9)

for all i ¤ 0:

(b) Conversely, let a 2 and 0 ; : : : ; n satisfy (1.8)–(1.7). If 0 ¤ 0, then f0 j has a minimum in a. (c) If there exist a, b 2 K such that fi .b/ < 0 for all i ¤ 0;

(1.10)

then (1.7)–(1.9) imply that either 0 > 0 or 0 D D n D 0. Since a differentiable convex function has a minimum in a ” its derivative vanishes in a, hence we deduce the following *Corollary 1.11 Let K be a convex open subset of a normed space, and let f0 ; : : : ; fn W K ! R be convex, differentiable functions. Assume that there exist a, b 2 K satisfying (1.10). Then f0 j has a minimum at some point a ” there exist real numbers 1 ; : : : ; n 0 satisfying f00 .a/ C 1 f10 .a/ C C n fn0 .a/ D 0 and i fi .a/ D 0

for all i:

Proof of the Theorem We denote by x y the usual scalar product of RnC1 and we introduce the canonical unit vectors e0 D .1; 0; : : : ; 0/; e1 D .0; 1; 0; : : : ; 0/; : : : ; en D .0; : : : ; 0; 1/:

24

Karush 1939, Kuhn–Tucker 1951. We recall from differential calculus that every local minimum of a convex function is also a global minimum.

25

1.3 Separation of Convex Sets . . .

23

(a) The formula ˚ C WD c 2 RnC1 W 9x 2 K W f0 .x/ < f0 .a/ C c0 and fi .x/ ci ; i D 1; : : : ; n

defines a non-empty convex set in RnC1 with 0 … C. Applying Proposition 1.7 with A D f0g and B D C, there exists a non-zero vector D .0 ; : : : ; n / 2 RnC1 such that x 0 for all x 2 C. By the continuity of the scalar product this yields c 0 for all c 2 C:

(1.11)

Observe that ˚ c 2 RnC1 W 9x 2 K W f0 .x/ f0 .a/ C c0

and fi .x/ ci ; 8i 1 C:

(1.12)

Indeed, if c belongs to the first set, then .c0 C ı; c1 ; : : : ; cn / 2 C for every ı > 0, and we conclude by letting ı ! 0. For each fixed i, choosing x D a in (1.12) we get ei 2 C, whence i 0 by (1.11). For i 1 this choice also shows that ei 2 C, whence i fi .a/ 0 by (1.11). Since i 0 and fi .a/ 0 (because a 2 ), we conclude that in fact i fi .a/ D 0. Finally we observe that c WD .f0 .x/ f0 .a/; f1 .x/; : : : ; fn .x// 2 C for every x 2 K by (1.12). Applying (1.11) again, we get f .x/ 0 f0 .a/ D c 0: Since we already know that f .a/ D 0 f0 .a/, we conclude that f .x/ 0 f0 .a/ D f .a/ for all x 2 K. (b) For any fixed x 2 , applying consecutively (1.8)–(1.7) and the property fi .x/ 0 (i 1), we obtain that 0 f0 .a/ D f .a/ f .x/ 0 f0 .x/: Since 0 > 0, this implies f0 .a/ f0 .x/.

24

1 Hilbert Spaces

(c) If 0 D 0, then (1.9) and (1.7) imply n X

i fi .b/ D f .b/ f .a/ D 0 f0 .a/ D 0:

iD1

Since i 0 and fi .b/ < 0 for all i 1 by (1.8) and (1.10), hence we conclude that 1 D D n D 0. t u

1.4 Orthonormal Bases Hilbert spaces provide an ideal framework for the study of Fourier series. Definition By an orthonormal sequence we mean a sequence of pairwise orthogonal unit vectors.26 Examples • The vectors k1

‚ …„ ƒ ek D .0; : : : ; 0; 1; 0; : : :/;

k D 1; 2; : : :

form an orthonormal sequence in `2 . • (Trigonometric system) For any interval I of length 2 the functions 1 e0 D p ; 2

sin kt and e2k1 D p ;

cos kt e2k D p ;

k D 1; 2; : : :

form an orthonormal sequence in L2 .I/. p • The functions 2= sin kt (k D 1; 2; : : :) form an orthonormal sequence in L2 .0; /.27 p p • The functions 1= and 2= cos kt (k D 1; 2; : : :) form an orthonormal sequence in L2 .0; /. Lemma 1.12 If the vectors x1 ; : : : ; xn are pairwise orthogonal, then kx1 C C xn k2 D kx1 k2 C C kxn k2 :

26 27

Gram [173] and Schmidt [416]. We write L2 .0; / instead of L2 .Œ0; / for brevity.

1.4 Orthonormal Bases

25

Proof Since .xj ; xk / D 0 if j ¤ k, we have kx1 C C xn k2 D

n X n X

.xj ; xk / D

jD1 kD1

n X

.xj ; xj / D kx1 k2 C C kxn k2 :

jD1

t u Proposition 1.13 Let .ej / be an orthonormal sequence in H. (a) The orthogonal projection PMn onto Mn WD Vect fe1 ; : : : ; en g28 is given by the explicit formula PMn x D

n X

.x; ej /ej ;

x 2 H:

jD1

Consequently,29 n X dist.x; Mn / D x .x; ej /ej :

(1.13)

jD1

(b) (Bessel’s equality)30 The equality m m 2 X X .x; ej /ej D kxk2 j.x; ej /j2 x jD1

(1.14)

jD1

holds for all x 2 H and m D 1; 2; : : : : (See Fig. 1.7.) (c) (Bessel’s inequality)31 We have 1 X

j.x; ej /j2 kxk2

(1.15)

jD1

for all x 2 H. In particular, the series on the left-hand side is convergent. (d) If .cj / is a sequence of real numbers, then 1 X jD1

28

cj ej

is convergent in

H”

1 X

jcj j2 < 1:

jD1

The linear hull Mn is finite-dimensional, hence closed. Toepler [455]. 30 Bessel [41, 42]. Figure 1.7 shows that this is a generalization of Pythagoras’ theorem. 31 Bessel [41, 42]. 29

26

1 Hilbert Spaces

Fig. 1.7 Bessel’s equality for mD1

x x − (x, e1 )e1

Remarks • The case m D 1 of Bessel’s inequality follows from the Cauchy–Schwarz inequality. • The quantities .x; ej / are called the Fourier coefficients of x.32 Proof (a) It suffices to observe that the vector on the right-hand side belongs to Mn , and that the differences of the two sides is orthogonal to Mn , because it is orthogonal to each of the vectors e1 ; : : : ; en that span Mn : n n X X .x; ej /ej ; ek D .x; ek / .x; ej /.ej ; ek / x jD1

jD1

D .x; ek / .x; ek / D 0;

32

Clairaut [88, pp. 546–547], Euler [131], and Fourier [148].

k D 1; : : : ; n:

1.4 Orthonormal Bases

27

(b) Since x PMn x D x

n X

.x; ej /ej

jD1

is orthogonal to Mn by the properties of the orthogonal projection, the n C 1 vectors on the right-hand side of the equality n n X X xD x .x; ej /ej C .x; ej /ej jD1

jD1

are pairwise orthogonal. Applying the lemma, (1.14) follows. (c) By Bessel’s equality kxk2 is an upper bound of all partial sums of this series of nonnegative terms. (d) Since n n X 2 X cj ej D jcj j2 jDmC1

jDmC1

for all n > m, the Cauchy criteria are the same for the two series.

t u

Let us investigate the case of equality in Bessel’s inequality: Proposition 1.14 Let .ej / be an orthonormal sequence in H. The following four properties are equivalent: P (a) (Fourier series)33 we have 1 jD1 .x; ej /ej D x for all x 2 H; (b) the subspace34 M WD Vect fe1 ; eP 2 ; : : :g is dense in H; 2 2 (c) (Parseval’s equality)35 we have 1 jD1 j.x; ej /j D kxk for all x 2 H; (d) if y 2 H and .y; ej / D 0 for all j, then y D 0. Proof (a) ” (b). Setting Mm WD Vect fe1 ; : : : ; em g, (a) and (b) are equivalent to the conditions m X .x; ej /ej ! 0 and x

dist.x; Mm / ! 0

jD1

for all x 2 H. We conclude by applying the equality (1.13).

33

Fourier [148]. The linear hull M is by definition the set of all finite linear combinations of the vectors ej . 35 Parseval [352]. 34

28

1 Hilbert Spaces

(a) ” (c) follows from the Bessel equality because the two sides of (1.14) tend to zero at the same time. P P (a) H) (d). We have y D 1 .y; ej /ej D 1 jD1 0 D 0. PjD1 (d) H) (a). Set y WD x 1 .x; e /e 2 H: the series converges by parts (c) k k kD1 and (d) of the proposition. Since .y; ej / D .x; ej /

1 X

.x; ek /.ek ; ej / D .x; ej / .x; ej / D 0

kD1

for all j, using (d) we conclude that y D 0.36

t u

Definition An orthonormal sequence .ej / is complete if the equivalent conditions (a)–(d) are satisfied. In this case we also say that .ej / is an orthonormal basis. Examples • The orthonormal sequence e1 ; e2 ; : : : of `2 , given above, is complete because .x; ej / D xj for all j for every x D .xj / 2 `2 , so that Parseval’s equality follows from the definition of the norm. • The three other orthonormal sequences given above are complete as well.37 Applying Parseval’s equality for the trigonometric system on the interval I D Œ; and for the function x.t/ t we obtain by an easy computation a famous result of Euler38 : 1 X 2 1 : D 2 k 6 kD1

If .ej / is an orthonormal basis in H, then the finite linear combinations of the vectors ej with rational coefficients form a countable, dense set in H, so that H is separable. Conversely, we have the following Proposition 1.15 Every separable Hilbert space has an orthonormal basis. Proof Let .yn / be a dense sequence in a Hilbert space H. Let nk be the first index for which y1 ; : : : ; ynk span a k-dimensional subspace. Then the sequence yn1 , yn2 ,. . . is linearly independent; furthermore, y1 ; : : : ; ynk

and yn1 ; : : : ; ynk

span the same subspace Mk for each k. 36

The completeness of H was used only in this step, so that (a), (b) and (c) are equivalent in noncomplete Euclidean spaces as well. See Exercise 1.12. 37 See Corollary 9.6, p. 314. 38 Euler [128] (heuristic proof), [129] (§ 167).

1.5 Weak Convergence: Theorem of Choice

29

Writing xk WD ynk for brevity, the formulas39 e1 D

x1 kx1 k

and ek WD

xk PMk1 xk ; kxk PMk1 xk k

k D 2; 3; : : :

define a sequence of unit vectors satisfying e1 ; : : : ; ek1 2 Mk1 , ek ? Mk1 and Vect fe1 ; : : : ; ek1 g D Vect fx1 ; : : : ; xk1 g for all k 2. Hence .ek / is an orthonormal sequence, and Vect fe1 ; e2 ; : : :g D Vect fx1 ; x2 ; : : :g D Vect fy1 ; y2 ; : : :g D H: t u *Remark The convergence and the sum of an orthogonal series do not depend on the order of its terms. Therefore the results of this section may be extended to arbitrary non-separable Hilbert spaces, by considering orthonormal families instead of orthonormal sequences.40

1.5 Weak Convergence: Theorem of Choice The examples at the end of Sect. 1.1 show that the Bolzano–Weierstrass theorem fails in infinite-dimensional Hilbert spaces: bounded, closed sets are not always compact. A simple counterexample is provided by the closed balls of infinitedimensional Hilbert spaces41 : Example Every orthonormal sequence .en / is bounded, but it does not have any convergent subsequence because ken em k > 1 for all n ¤ m. However, Hilbert succeeded in generalizing the Bolzano–Weierstrass theorem for all Hilbert spaces by a suitable weakening of the notion of convergence. The idea comes from the following elementary observation: Proposition 1.16 Let e1 ; : : : ; ek be an orthonormal basis in a finite-dimensional Hilbert space H. Then the following properties are equivalent: (a) xn ! x; (b) .xn ; y/ ! .x; y/ for each fixed y 2 H; (c) .xn ; ej / ! .x; ej / for j D 1; : : : ; k.

39

Gram–Schmidt orthogonalization [173], [415]. See, e.g., Halmos [185]. 41 It suffices to consider unit balls by a similarity argument. 40

30

1 Hilbert Spaces

Proof The equivalence (a) ” (c) follows from the identity 2 X k k X ˇ ˇ ˇ.xn x; ej /ˇ2 : .x x; e /e D kxn xk2 D n j j jD1 jD1 Property (c) implies the formally stronger property (b) because we have y D P k jD1 cj ej with suitable coefficients cj , and then .xn ; y/ .x; y/ D

k X cj .xn ; ej / .x; ej / ! 0: jD1

t u Remark For the usual orthonormal basis of H D Rk the equivalence (a) ” (c) means that the convergence of a vector sequence is equivalent to its coordinate-wise or component-wise convergence. Definition The sequence .xn / converges weakly42 to x in H if .xn ; y/ ! .x; y/ for each fixed y 2 H.43 We express this by writing xn * x. Example In infinite dimensions every orthonormal sequence .en / converges weakly P to zero. Indeed, the numerical series j.y; en /j2 converges for each y 2 H by Bessel’s inequality (Proposition 1.13, p. 25), and therefore its general term tends to zero: .y; en / ! 0 D .y; 0/. We recall that .en / is not norm-convergent. Let us establish the basic properties of weak convergence: Proposition 1.17 (a) (b) (c) (d) (e) (f) (g)

A sequence has at most one weak limit. If xn * x, then xnk * x for every .xnk / subsequence, too. If xn * x and yn * y, then xn C yn * x C y. If xn * x in H and n ! in R, then n xn * x in H. Let K H be a convex closed set and .xn / K. If xn * x, then x 2 K. If kxn k L for all n and xn * x, then kxk L.44 The following equivalence holds: xn ! x

42

”

xn * x

and

kxn k ! kxk :

Hilbert [209]. We often write the last relation in the equivalent form .xn x; y/ ! 0. 44 Equivalently kxk lim inf kxn k. 43

1.5 Weak Convergence: Theorem of Choice

31

Proof (a) If xn * x and xn * y, then .xn ; x y/ ! .x; x y/ and .xn ; x y/ ! .y; x y/. By the uniqueness of the limit of numerical sequences we conclude .x; x y/ D .y; x y/, i.e., .x y; x y/ D 0, and thus x y D 0. (b), (c), (d) follow by definition from the corresponding properties of numerical sequences. For example, (d) may be shown in the following way: we have .n xn ; y/ D n .xn ; y/ ! .x; y/ D .x; y/ for each y 2 H, i.e., n xn * x. (e) Denoting by y the orthogonal projection of x onto K, we have .xn y; x y/ 0 for all n by Theorem 1.5 (p. 12). Since xn * x, taking the limit we find .x y; x y/ 0. Hence kx yk2 0 and therefore x D y 2 K. (f) We apply (e) with K WD fz 2 H W kzk Lg. (g) If xn ! x, i.e., if kxn xk ! 0, then j.xn ; y/ .x; y/j kxn xk kyk ! 0 for each y 2 H by the Cauchy–Schwarz inequality, and jkxn k kxkj kxn xk ! 0 by the triangle inequality. Conversely, if xn * x and kxn k ! kxk, then the right-hand side of the identity kxn xk2 D kxn k2 C kxk2 2.xn ; x/ tends to zero, so that xn ! x.

t u

Remarks • The convexity condition cannot be omitted in (e): every orthonormal sequence belongs to the closed unit sphere, but its weak limit, the null vector, does not. • Norm convergence is also called strong convergence because it implies weak convergence by (g).

32

1 Hilbert Spaces

Every weakly convergent sequence is bounded. For the proof of this deeper property we recall Baire’s lemma from topology45: Proposition 1.18 If a complete metric space is covered by countably many closed sets, then at least one of them has a non-empty interior. Proposition 1.19 (a) Every weakly convergent sequence is bounded. (b) If xn ! x and yn * y, then .xn ; yn / ! .x; y/. Example Part (b) expresses a strengthened continuity property of the scalar product. If .en / is an orthonormal sequence, then the example xn D yn WD en shows that it cannot be strengthened further: the relations xn * x and yn * y do not imply .xn ; yn / ! .x; y/ in general. Proof (a) If xn * x in H, then the numerical sequence n 7! .xn ; y/ is convergent for each y 2 H, and hence it is bounded. Consequently, the closed sets Fk WD fy 2 H W j.xn ; y/j k

for all ng ;

k D 1; 2; : : :

cover H. By Baire’s lemma, one of them, say Fk , contains a ball B2r .y/. If xn ¤ 0, then y C r kxn k1 xn 2 B2r .y/ Fk ; and hence j.xn ; y C r kxn k1 xn /j k: Since y 2 Fk , this yields r kxn k D j.xn ; r kxn k1 xn /j k C j.xn ; y/j 2k; i.e., the boundedness of .xn /.

45 Osgood [350], Baire [17], Kuratowski [272], Banach [23]. The usefulness of Baire’s lemma in functional analysis was recognized by Saks: see Banach and Steinhaus [28]. See also the selfcontained proofs of the more general Theorem 2.23 and Proposition 2.24 below (pp. 81–82), without using Baire’s lemma.

1.5 Weak Convergence: Theorem of Choice

33

(b) Since .yn / is bounded, we have j.xn ; yn / .x; y/j j.xn x; yn /j C j.x; yn y/j kxn xk kyn k C j.x; yn / .x; y/j !0 as n ! 1.

t u

The following lemma simplifies the verification of weak convergence: Lemma 1.20 Let .xn / be a bounded sequence in H and x 2 H. The set Y WD fy 2 H W .xn ; y/ ! .x; y/g is a closed subspace of H. Proof Y is a subspace by the linearity of the scalar product. For the closedness we show that if .yk / Y and yk ! y 2 H, then y 2 Y. Fixing " > 0 arbitrarily, we have to find an integer N such that j.xn x; y/j < " for all n N. Choose a large number L such that kxk < L, and kxn k < L for all n, and then choose a large index k satisfying kyk yk < "=3L. Since yk 2 Y, there exists an N such that j.xn x; yk /j < "=3 for all n N. Then the required inequality holds for all n N because j.xn x; y/j j.xn x; y yk /j C j.xn x; yk /j < kxn xk ky yk k C 2L

" " C D ": 3L 3

" 3

t u 1

.x1k /;

2

.x2k /; : : :

x D converges weakly to x D .xk / in Example The sequence x D `2 ” it is bounded, and xnk ! xk for each k (component-wise convergence). Indeed, writing xnk ! xk in the equivalent form .xn ; ek / ! .x; ek /, the necessity of this condition follows from the proposition. The sufficiency follows from Lemma 1.20 because .ek / spans `2 . Now we are ready to generalize the Bolzano–Weierstrass theorem: Theorem 1.21 (Theorem of Choice)46 In a Hilbert space every bounded sequence has a weakly convergent subsequence.

46

Hilbert [209], Schmidt [416], and von Neumann [336].

34

1 Hilbert Spaces

Proof Let .xn / be a bounded sequence in H, and fix a constant L such that kxn k < L for all n. Let us denote by M the closed linear hull of .xn /. Observe that M is separable. If M is finite-dimensional, then .xn / has even a strongly convergent subsequence by the classical Bolzano–Weierstrass theorem. Henceforth assume that M is infinitedimensional, and fix an orthonormal basis .ek / of M by Proposition 1.15 (p. 28). The numerical sequence n 7! .xn ; e1 / is bounded. By the Bolzano–Weierstrass theorem there exist a subsequence .x1n / .xn / and c1 2 R such that .x1n ; e1 / ! c1 . Next, since the numerical sequence n 7! .x1n ; e2 / is also bounded, there exist a subsequence .x2n / .x1n / and c2 2 R such that .x2n ; e2 / ! c2 . Continuing by recursion we construct an infinite sequence of subsequences .xn / .x1n / .x2n / and real numbers ck such that .xkn ; ek / ! ck 47 for each fixed k D 1; 2; : : : : Applying Cantor’s diagonal method, P1 the formula n zn WD xn defines a subsequence .zn / .xn / converging weakly to kD1 ck ek . For the proof first we notice that for each fixed k, the truncated subsequence zk ; zkC1 ; : : : of .zn / is also a subsequence of .xkn /1 nD1 P,1and hence .zn ; ek / ! ck . Next we claim that the orthogonal series kD1 ck ek converges strongly to some point z 2 M of norm L. For the convergence it suffices to check by P 2 2 Proposition 1.13 that m L for each fixed m. We have jc j k kD1 m X

j.zn ; ek /j2 kzn k2 < L2

kD1

for all n by Bessel’s inequality, and the required assertion follows by letting n ! 1. Finally, the inequality kzk L follows from the continuity of the norm. We already know that .zn ; ek / ! ck D .z; ek / for all k. Applying Lemma 1.20 we conclude that .zn ; y/ ! .z; y/ for all y 2 M, too. We prove finally that .zn ; y/ ! .z; y/ for all y 2 H. Denoting by u the orthogonal projection of y onto M, we already know that .zn ; u/ ! .z; u/. Furthermore, we have y u ? M, so that .zn z; y u/ D 0 for all n. We conclude that .zn ; y/ .z; y/ D .zn z; u/ C .zn z; y u/ D .zn z; u/ ! 0: t u

47

Cantor [75].

1.6 Continuous and Compact Operators

35

1.6 Continuous and Compact Operators For brevity a linear map A W H ! H is also called an operator. Its continuity may also be characterized by weak convergence: Proposition 1.22 For an operator A W H ! H the following properties are equivalent: (a) (b) (c) (d) (e) (f)

there exists a constant M such that kAxk M kxk for all x 2 H; A sends bounded sets into bounded sets; A sends totally bounded sets into totally bounded sets; xn ! x H) Axn ! Ax; xn * x H) Axn * Ax; xn ! x H) Axn * Ax.

Remark It suffices to check (d), (e) and (f) for x D 0 by linearity. The same remark applies to Proposition 1.24 below. For the proof we introduce adjoint operators: Proposition 1.23 For each operator A 2 L.H; H/ there exists a unique operator A 2 L.H; H/ such that .Ax; y/ D .x; A y/

for all x; y 2 H:

(1.16)

Definition A is called the adjoint of A.48 Remark It follows from the proposition that A D A for every A. Proof For any fixed y 2 H the formula y .x/ WD .Ax; y/ defines a bounded linear functional y 2 H 0 . Applying the Riesz–Fréchet theorem there exists a unique vector y 2 H satisfying .Ax; y/ D .x; y / for all x; y 2 H: Hence y is the unique possible candidate for A y. On the other hand, defining A y WD y the condition (1.16) is satisfied indeed. For any y1 ; y2 2 H and 2 R it follows from the definitions of y1 , y2 and from the bilinearity of the scalar product that .Ax; y1 C y2 / D .x; A y1 C A y2 /

and .Ax; y/ D .x; A y/

for all x; y 2 H. In view of the uniqueness of the vectors A .y1 C y2 / and A .y/ the linearity of A follows.

48

Lagrange [279, p. 471] and Riesz [379, 382] (in L2 and `2 ).

36

1 Hilbert Spaces

Applying (1.16) with x D A y we get for every y 2 H the estimate 2 kA yk D .AA y; y/ kAA yk kyk kAk kA yk kyk I

this shows that A continuous, and kA k kAk.

t u

Proof of Proposition 1.22 The implications (a) ” (b), (a) ” (c), (a) H) (d) and (e) H) (f) follows from the definitions. (d) H) (e). We have .Axn Ax; y/ D .xn x; A y/ ! 0 for any fixed y 2 H because xn * x. (f) H) (a). If (a) is not satisfied, then there exists a sequence .xn / such that kxn k D 1=n and kAxn k > n for every n. Then xn ! 0, while .Axn / is unbounded and hence does not converge weakly. t u Let us strengthen the continuity: Proposition 1.24 For an operator A W H ! H the following properties are equivalent: (a) .xn / is bounded H) .Axn / has a (strongly) convergent subsequence; (b) A sends bounded sets into totally bounded sets; (c) xn * x H) Axn ! Ax. For the proof we need the following result of Cantor: Lemma 1.25 (Cantor)49 In a topological space a sequence xn converges to x ” every subsequence .x0n / of .xn / has a subsequence .x00n / converging to x. Proof If xn ! x, then x0n ! x, so that we can choose x00n WD x0n . On the other hand, if xn 6! x, then there exist a neighborhood V of x and a subsequence .x0n / of .xn / such that x0n … V for all n. Then .x0n / has no subsequence converging to x. t u Proof of Proposition 1.24 (a) H) (b) If (b) does not hold, then there exists a bounded set B such that A.B/ is not totally bounded. It means that there exists an r > 0 such that A.B/ cannot be covered by finitely many balls of radius r. Using this property we may recursively construct a sequence .xn / B such that kAxn Axk k r for all n ¤ k. Then .xn / is a bounded sequence, but .Axn / has no convergent subsequence because the Cauchy criterion is not satisfied. Hence (a) does not hold either. (b) H) (c) In view of the lemma it is sufficient to show that every subsequence .Ax0n / of .Axn / has a subsequence .Ax00n / converging to Ax.

49

Cantor [69, p. 89]

1.6 Continuous and Compact Operators

37

Since the sequence .x0n / is weakly convergent and hence bounded, by property (b) the image sequence .Ax0n / belongs to a totally bounded set. Since the closure of a totally bounded set is compact,50 there exists a suitable subsequence Ax00n ! y. It remains to show that y D Ax. Since xn * x implies x00n * x, and since A is continuous by (b) and by Proposition 1.22, we have Ax00n * Ax. On the other hand, Ax00n ! y implies Ax00n * y, so that y D Ax by the uniqueness of the weak limit. (c) H) (a) Every bounded sequence .xn / has a weakly convergent subsequence x0n * x by Theorem 1.21. Then we have Ax0n ! Ax by (c). t u Definition An operator A W H ! H is compact or completely continuous,51 if it satisfies one of the equivalent properties of Proposition 1.24. Examples • If H is finite-dimensional, then every operator A W H ! H is continuous, and hence compact. • The identity map I W H ! H is not compact if H is infinite-dimensional. Indeed, we have en * 0 for every orthonormal sequence, but Ien D en 6! 0 in H. We establish some basic properties of compact operators: Proposition 1.26 (a) (b) (c) (d)

Every compact operator is continuous. Every continuous operator of finite rank52 is compact. If A; B 2 L.H; H/ and A is compact, then AB and BA are compact. The compact operators form a closed subspace in L.H; H/.

Proof (a), (b) and (c) follow from Propositions 1.22 and 1.24 and from the equivalence of weak and strong convergence in finite-dimensional spaces. (d) Only the closedness is not obvious. Let A1 , A2 , . . . be compact operators satisfying An ! A in L.H; H/. We have to show that A is compact. If .xk / is a bounded sequence in H, then repeating the proof of Theorem 1.21 we may construct a subsequence .zk / such that the image sequences .An zk / are convergent for each fixed n. It is sufficient to show that .Azk / is a Cauchy sequence. Fix a constant L such that kxn k < L for all n. For each fixed " > 0 choose n such that kA An k

0 choose N such that kAn Am k "

78

2 Banach Spaces

for all m; n N. Then kAn x Am xk " kxk for all m; n N and x 2 X. Letting m ! 1 we obtain kAn x Axk " kxk for all n N and x 2 X, i.e., An ! A in L.X; Y/.

t u

Corollary 2.19 All ` spaces are Banach spaces. p

Proof We have seen in the preceding section that all `p spaces are dual spaces, and hence complete by the preceding proposition. Alternatively, the completeness of `p for 1 p < 1 may be proved by a simple adaptation of the proof given for `2 in Sect. 1.1, by changing the exponents 2 to p everywhere. t u *Examples • If U is a non-empty open set in a normed space, Y a Banach space, and k a natural number, then the Ck functions f W U ! Y for which f , f 0 , . . . , f .k/ are all bounded form a Banach space Cbk .U; Y/ with respect to the norm k f k1 C f 0 1 C C f .k/ 1 ; because the derivative functions map into Banach spaces of the form L.X; Z/ by the proposition.32 • Let I D Œa; b be a non-degenerate compact interval and 1 p < 1. We know that C.I/ is a normed space with respect to the norm kxkp WD

Z

jx.t/jp dt

1=p

:

I

This norm is not complete. For p D 2 we have already proved this on page 10; the general case follows by changing every exponent 2 to p in that proof. An easy adaptation of the proof of Proposition 1.3 (p. 10) leads to the following result: Proposition 2.20 Every normed space may be completed, i.e., may be considered as a dense subspace of a Banach space. Definition We denote by Lp .I/, for 1 p < 1, the Banach space obtained by completion of C.I/ with respect to the norm kkp .

32

See any book on differential calculus.

2.5 Weak Convergence: Helly–Banach–Steinhaus Theorem

79

Remark Later we will give a concrete interpretation of these spaces.33 We end this section by giving another proof of the last proposition. Definition By the bidual of a normed space X we mean the Banach space X 00 WD .X 0 /0 .34 Example If x 2 X, then the formula ˆx .'/ WD '.x/;

' 2 X0

defines a continuous linear functional ˆx 2 X 00 , and kˆx k kxk because jˆx .'/j D j'.x/j j'j kxk for every ' 2 X 0 . Let us look more closely at the correspondence x 7! ˆx : Corollary 2.21 (Hahn)35 Let X be a normed space. (a) The formula J.x/ WD ˆx defines a linear isometry J W X ! X 00 . (b) X may be completed: there exist a Banach space Y and a linear isometry J W X ! Y such that J.X/ is dense in Y. Proof (a) The linearity of J is straightforward. The isometry follows from Corollary 2.13 (c): kJxk D sup j.Jx/.'/j D sup j'.x/j D kxk : k'k1

k'k1

(b) In view of (a) we may choose for Y the closure in X 00 of the range J.X/ of J: as a closed subspace of the Banach space X 00 , it is also a Banach space. t u

2.5 Weak Convergence: Helly–Banach–Steinhaus Theorem Weak convergence proved to be a useful tool in the study of Hilbert spaces. We generalize this notion to normed spaces. Definition A sequence .xn / in a normed space X converges weakly36 to x 2 X if '.xn / ! '.x/ 33

See Proposition 9.5 (b), p. 312. Hahn [182]. We will investigate these spaces in Sect. 2.6, p. 87. 35 Hahn [182]. 36 Riesz [380], Banach [24]. 34

80

2 Banach Spaces

for every ' 2 X 0 . We express this by writing xn * x. Remarks • For Hilbert spaces this reduces to the former notion by the Riesz–Fréchet theorem. • Norm convergence implies weak convergence by the continuity of the functionals of X 0 . Therefore norm convergence is also called strong convergence. • In finite-dimensional normed spaces the strong and weak convergences coincide. Let us collect the elementary properties of weak convergence: Proposition 2.22 (a) (b) (c) (d) (e)

A sequence has at most one weak limit. If xn * x, then xnk * x for every subsequence .xnk /. If xn * x and yn * y, then xn C yn * x C y. If xn * x in X and n ! in R, then n xn * x in X. Let K X be a convex closed set. If xn * x, and xn 2 K for every n, then x 2 K. ( f) If xn * x, and kxn k L for every n, then kxk L.37 (g) If xn ! x, then xn * x and kxn k ! kxk. *Remark In contrast to Hilbert spaces the relations xn * x and kxn k ! kxk do not imply xn ! x in general.38 If this holds, then X is said to have the Radon–Riesz property. Proof We may repeat the corresponding proofs given for Hilbert spaces (p. 30), except for (a) and (e); for the proof of (g) we now apply the continuity of ' 2 X 0 instead of the Cauchy–Schwarz inequality. (a) If xn * x and xn * y, then by Corollary 2.13 there exists a ' 2 X 0 satisfying '.x y/ D kx yk. Since '.xn / ! '.x/ and '.xn / ! '.y/ imply '.x/ D '.y/, hence kx yk D '.x y/ D '.x/ '.y/ D 0; and therefore x D y. (e) Instead of the orthogonal projection we use Tukey’s theorem (p. 61). Assume on the contrary that x … K; then there exist ' 2 X 0 and c1 ; c2 2 R such that '.x/ c1 < c2 '.y/

for every y 2 K:

Then '.xn / c2 for every n, so that '.xn / 6! '.x/, i.e., xn 6* x.

37 38

Equivalently, kxk lim inf kxn k. We give soon an example. See also Proposition 9.11, p. 328.

t u

2.5 Weak Convergence: Helly–Banach–Steinhaus Theorem

81

Every weakly convergent sequence is bounded. Before proving this deeper result, we establish another essential result of Functional Analysis: the uniform boundedness theorem: Theorem 2.23 (Helly–Banach–Steinhaus)39 Consider a family A L.X; Y/ of continuous linear maps where X is a Banach-space, and Y a normed space. If the sets A.x/ WD fAx 2 Y W A 2 Ag ;

x2X

are all bounded in Y, then A is bounded in L.X; Y/: sup fkAk W A 2 Ag < 1:

*Remark The idea of this theorem had already appeared in Riemann’s work.40 *Example The theorem fails in non-complete spaces X. Consider for example the subspace X of `2 formed by the sequences having at most finitely many non-zero elements. The formula 'n .x/ WD nxn defines a pointwise bounded but uniformly unbounded sequence of functionals in L.X; R/. Proof It suffices to prove41 that A is uniformly bounded in some ball, say kAxk C

for every A 2 A and x 2 B2r .x0 /:

This will imply for all A 2 A and x 2 X, kxk 1, the relations x0 ; x0 C rx 2 B2r .x0 /, and therefore the inequalities kAxk D

0 0 1 A.x0 C rx/ Ax0 kA.x C x/k C kAx k 2C ; r r r

whence kAk 2C=r for every A 2 A. 39

Helly [204], and Banach–Steinhaus [28]. See Hochstadt [215] on Helly’s contribution. See also Banach [19], Hahn [181], and Hildebrandt [211]. 40 Condensation of singularities, Riemann [371], and Hankel [190]. See also Gal [166]. 41 Following a suggestion of Saks, Banach and Steinhaus proved their theorem with the help of Baire’s lemma (p. 32). We prefer to adapt, following Riesz–Sz. Nagy [394], an argument of Osgood [350, pp. 163–164], that can also be used to prove Baire’s lemma.

82

2 Banach Spaces

Assume on the contrary that A is not uniformly bounded on any open ball, and fix an arbitrary ball B0 .42 By our assumption there exist A1 2 A and x1 2 B0 such that kA1 x1 k > 1. By the continuity of A1 the inequality remains valid in a small ball B1 centered at x1 . By choosing its radius sufficiently small, we may also assume that diam B1 < 1 and B1 B0 . Repeating these arguments, there exist A2 2 A and a ball B2 such that diam B2 < 1=2, B2 B1 , and kA2 xk > 2 for every x 2 B2 . Continuing by induction we obtain a sequence .Ak / A of maps and a sequence .Bk / of balls such that diam Bk < 1=k, Bk Bk1 , and kAk xk > k for every x 2 Bk , k D 1; 2; : : : : Applying Cantor’s intersection theorem we conclude that \k Bk ¤ ¿. If x is a common point of the balls Bk , then kAk xk k for every k, contradicting the boundedness of A.x/. t u Proposition 2.24 Let .xn / be a sequence in a normed space X. (a) If xn * x, then the sequence .xn / is bounded. (b) If xn * x in X and 'n ! ' in X 0 , then 'n .xn / ! '.x/. (c) If xn ! x in X and 'n * ' in X 0 , then 'n .xn / ! '.x/. Proof (a) We apply Theorem 2.23 for the family .ˆn / X 00 of the functionals ˆn ' WD '.xn /;

' 2 X 0 ; n D 1; 2; : : : ;

and we use the equalities kˆn k D kxn k from Corollary 2.21 (a) (p. 79). (b) The right-hand side of the identity 'n .xn / '.x/ D .'n '/.xn / C '.xn x/ tends to zero because xn * x implies '.xn x/ ! 0, and because .xn / is bounded by (a), so that j.'n '/.xn /j k'n 'k sup kxn k ! 0: (c) Writing ˆ. / WD

.x/ we have ˆ 2 X 00 , and the right-hand side of the identity

'n .xn / '.x/ D 'n .xn x/ C .'n '/x D 'n .xn x/ C ˆ.'n '/

42

As usual, all balls are considered to be open.

2.5 Weak Convergence: Helly–Banach–Steinhaus Theorem

83

tends to zero because 'n * ' implies ˆ.'n '/ ! 0, and because .'n / is bounded by (a), so that j'n .xn x/j kxn xk sup k'n k ! 0: t u A simple adaptation of the proof of Lemma 1.20 (p. 33) yields the following results: Lemma 2.25 Let .xk / be a bounded sequence in a normed space X. (a) For each x 2 X the set ˚ ' 2 X 0 W '.xk / ! '.x/ is a closed linear subspace of X 0 . (b) The set ˚ ' 2 X 0 W .'.xk // converges in

R

is a closed linear subspace of X 0 . *Examples • Let X D c0 or X D `p for some 1 < p < 1. Let k 7! .xkn / be a bounded sequence in X, and let .xn / 2 X. Lemmas 2.16 and 2.25 (pp. 73 and 83) yield the following characterizations of weak convergence (component-wise convergence): .xkn / * .xn / ” xkn ! xn

for each n:

• In particular, the sequence of the vectors k1

‚ …„ ƒ ek D .0; : : : ; 0; 1; 0; : : :/;

k D 1; 2; : : :

converges weakly to zero in the above spaces. • But this sequence does not converge weakly in `1 . Indeed, the formula '.x/ WD

1 X

.1/n xn ;

x D .xn / 2 `1

nD1

defines a functional ' 2 .`1 /0 for which the numerical sequence of numbers '.en / D .1/n is divergent.

84

2 Banach Spaces

• Let xn D e1 C en , then xn * e1 in c0 by the first example. Observe that kxn k1 ! ke1 k1 , but kxn e1 k1 6! 0. Hence c0 does not have the Radon–Riesz property. • Since c0 is a subspace of `1 , the relation xn * e1 also holds in `1 . Hence `1 does not have the Radon–Riesz property either. • On the other hand, it will follow from a later result43 that `p has the Radon–Riesz property for all 1 < p < 1. • Our next proposition will imply that `1 also has the Radon–Riesz property. The fact that component-wise convergence does not imply weak convergence in `1 also follows from the next surprising result: *Proposition 2.26 (Schur)44 In `1 the strong and weak convergences coincide. Proof It suffices to prove that if xk * x in `1 , then xk x1 ! 0. Changing xk to xk x we may assume that x D 0. Assume on the contrary that xk * 0 in `1 , but xk 1 6! 0. Denoting the elements of xk by xkn , we have xkn ! 0 for each fixed n by the definition of weak convergence. Set45 " WD lim sup xk 1 > 0 and k0 D n0 WD 0: Proceeding recursively, if km1 and nm1 have already been defined for some m, then choose a large index k D km > km1 such that k x m > " 1 2

and

nX m1

ˇ k ˇ ˇx m ˇ < " ; n 10 nD1

and then a large integer nm > nm1 such that Xˇ ˇ ˇxkm ˇ < " : n 10 n>n m

The formula yn WD sign xknm

43

if

nm1 < n nm

Proposition 9.11, p. 328. Schur [418]. 45 We apply the gliding hump method of Lebesgue [291]. 44

2.5 Weak Convergence: Helly–Banach–Steinhaus Theorem

85

defines a sequence .yn / 2 `1 of norm 1, satisfying the following inequalities for each m D 1; 2; : : : W 1 X

X ˇ ˇ Xˇ ˇ ˇ k ˇ ˇx m ˇ ˇxkm ˇ ˇxkm ˇ n n n

X

xknm yn

nm1 nm

Xˇ ˇ X ˇ ˇ ˇxkm ˇ 2 ˇxkm ˇ D xkm 1 2 n n nnm1

n>nm

" 4" 2 10 " : D 10

>

Hence xkm 6* 0, and thus xk 6* 0, contradicting our hypothesis.

t u

Finally we prove an interesting converse of Hölder’s inequality: *Proposition 2.27 (Hellinger–Toeplitz)46 LetP.yn / be a real sequence and p; q 2 Œ1; 1 two conjugate exponents. If the series xn yn converges for every .xn / 2 `p , then y 2 `q . Proof 47 The formula 'k .x/ WD

k X

xn yn ;

x 2 `p ;

k D 1; 2; : : :

nD1

defines a sequence .'k / in .`p /0 .48 By assumption the sequence .'k .x// is convergent, and hence bounded, for every x 2 `p . Applying the Banach–Steinhaus theorem there exists therefore a constant C such that k ˇ ˇX ˇ ˇ xn yn ˇ C kxkp ˇ

for every x 2 `p ;

k D 1; 2; : : : :

nD1

If q D 1 and thus p D 1, then choosing x D ek we deduce that jyk j C for all k, and hence y 2 `1 . If 1 q < 1, then introducing for each k the sequence ( xn WD

46

jyn jq1 sign xn

if

n k,

0

if

n > k,

Hellinger–Toeplitz [201] and Landau [282]. See also a short elementary proof of Riesz [382, pp. 47–48] by the gliding hump method. 48 The continuity of the functionals is evident because we have only finite sums here. 47

86

2 Banach Spaces

similarly to the proof of Proposition 2.15 we obtain that k X

jyn jq Cp :

nD1

Letting k ! 1 we conclude that y 2 `q and kykq C.

t u

Our next objective is to generalize the Bolzano–Weierstrass theorem to Banach spaces. Unfortunately, there are counterexamples even for the weak convergence: Examples • In `1 the bounded sequence .en / has no weakly convergent subsequence. Indeed, such a subsequence would also converge strongly by Schur’s theorem (p. 84). But this is impossible because no subsequence has the Cauchy property: kem en k D 2 for all m ¤ n. We can avoid the use of Schur’s theorem as follows. If .enk / is an arbitrary subsequence of .en /, then the formula '.x/ WD

1 X .1/k xnk ;

x D .xn / 2 `1

kD1

defines a functional ' 2 .`1 /0 . Since '.enk / D .1/k does not converge as k ! 1, the subsequence .enk / does not converge weakly. • In c0 the bounded sequence .e1 C Cen / has no weakly convergent subsequence. Indeed, if we had e1 C C enk * a for some subsequence, then we would also have '.e1 C C enk / ! '.a/ for every ' 2 c00 . Applying this for each fixed m D 1; 2; : : : to the functional '.y/ WD ym , we would get the equality a D .1; 1; : : :/. But this is impossible because the last sequence does not belong to c0 . • The bounded sequence .e1 C C en / has no weakly convergent subsequence in `1 either. Indeed, the previous reasoning shows again that the only possible weak limit is a D .1; 1; : : :/. But this is impossible because a does not belong to c0 , which is the closed subspace generated by the sequence .e1 C C en /: see Proposition 2.22 (e), p. 80. Nevertheless, we will see later49 that the above sequences converge in a natural, even weaker sense. In spite of these counterexamples, we prove in the next section that the weak convergence version of the Bolzano–Weierstrass theorem remains valid in a large class of Banach spaces.

49

See the examples on p. 136.

2.6 Reflexive Spaces: Theorem of Choice

87

2.6 Reflexive Spaces: Theorem of Choice Let X be a normed space. We recall from Corollary 2.21 (p. 79) that the formula ˆx .'/ WD '.x/;

' 2 X0

defines a functional ˆx 2 X 00 for each x 2 X, where X 00 denotes the bidual of X. In certain spaces every element of X 00 has this form: Definition A normed space X is reflexive50 if for each ˆ 2 X 00 there exists an x 2 X such that ˆ.'/ D '.x/ for all ' 2 X 0 : Before giving many examples, we discuss some consequences of the definition. We recall from Corollary 2.21 that the formula .Jx/.'/ WD '.x/;

x 2 X;

' 2 X0

defines a linear isometry J W X ! X 00 . Proposition 2.28 (Hahn)51 Let X be a normed space. (a) X is reflexive ” J is an isometric isomorphism between X and X 00 . (b) If X is reflexive, then it is complete, i.e., a Banach space. Proof (a) We already know that J is a linear isometry. By definition, J is surjective ” X is reflexive. (b) X is isomorphic to X 00 D .X 0 /0 , and every dual space is complete. t u Remark Reflexive Banach spaces are often identified with their bidual by the map J. Now we turn to the examples. Proposition 2.29 (a) Every finite-dimensional normed space is reflexive. (b) Every Hilbert space is reflexive. (c) The spaces `p spaces are reflexive for all 1 < p < 1.

50 51

Hahn [182]. Hahn [182].

88

2 Banach Spaces

Proof (a) We recall from linear algebra that dim X D dim X for every finite-dimensional vector space X. Hence we have dim X dim X 00 for every finite-dimensional normed space X.52 Therefore the linear isometry J W X ! X 00 must be onto (and dim X D dim X 00 ). (b) Let H be a Hilbert space and consider the Riesz–Fréchet isomorphism (Theorem 1.9, p. 19) j W H ! H 0 defined by the formula . jy/.x/ D .x; y/;

x; y 2 H:

(2.5)

For each ˆ 2 H 00 , ˆ ı j is a continuous linear functional on H. Applying the Riesz–Fréchet theorem again, there exists an x 2 H such that ˆ. jy/ D .y; x/

for all y 2 H:

Using (2.5) this implies ˆ. jy/ D . jy/.x/ for all y 2 H: Since j W H ! H 0 is onto, we conclude that ˆ.'/ D '.x/ for all ' 2 H 0 : (c) Consider the Riesz isomorphism j W `q ! .`p /0 (Proposition 2.15, p. 73) defined by the formula . jy/.x/ D

X

y n xn ;

x 2 ` p ; y 2 `q :

(2.6)

For each ˆ 2 .`p /00 , ˆ ı j is a continuous linear functional on `q . Applying Proposition 2.15 again, there exists an x 2 `p such that ˆ. jy/ D

X

xn yn

for all

y 2 `q :

Using (2.6) this implies ˆ. jy/ D . jy/.x/

52

for all y 2 `q :

In fact we have equality because every linear functional is continuous on finite-dimensional normed spaces.

2.6 Reflexive Spaces: Theorem of Choice

89

Since j W `q ! .`p /0 is onto, we conclude that ˆ.'/ D '.x/ for all ' 2 .`p /0 : t u Now we give some examples of non-reflexive Banach spaces. *Examples • c0 is not reflexive: the formula ˆ.'/ WD

1 X

'n ;

' D .'n / 2 `1

nD1

defines a functional ˆ 2 c000 D .`1 /0 which is not represented by any .xn / 2 c0 . Indeed, if such a sequence .xn / existed, then choosing ' WD ek in the corresponding equality 1 X

'n D

nD1

1 X

xn ' n

nD1

we would get xk D 1 for every k. But the constant sequence .1; 1; : : :/ does not belong to c0 . Let us give another proof. Since c00 is isomorphic to `1 , and .`1 /0 is isomorphic to `1 , c000 is isomorphic to `1 . Consequently, c000 is not separable. Since c0 is separable, it cannot be isomorphic to c000 . • `1 is not reflexive. For the proof we consider the subspace c of `1 formed by the convergent sequences. Applying Theorem 2.11 theorem we extend the continuous linear functional .yn / 7! lim yn , given on c, to a functional ˆ 2 .`1 /0 D .`1 /00 . We claim that ˆ is not represented by any sequence .xn / 2 `1 . Indeed, if such a sequence .xn / existed, then choosing y WD ek in the corresponding equality ˆ.y/ D

1 X

xn yn

nD1

we would get xk D 0 for every k, i.e., ˆ D 0. But this is impossible because for x D .1; 1; : : :/ we have ˆ.x/ D lim 1 D 1. • We will give further proofs for the non-reflexivity of c0 , `1 and `1 at the end of this section and in Sect. 3.6 (p. 144).

90

2 Banach Spaces

One of the most important properties of reflexive spaces is the following: Theorem 2.30 (Theorem of Choice)53 In a reflexive Banach space every bounded sequence has a weakly convergent subsequence.

Remark The converse of this theorem also holds: see Theorem 3.21, p. 140. Proof Let .xk / be a bounded sequence in a reflexive Banach space X. We identify X with its bidual X 00 , so that for every set X 0 we have ˚ ? W D ˆ 2 X 00 W ˆ.'/ D 0 for all ' 2 D fx 2 X W '.x/ D 0

for all ' 2 g :

Let us arrange the finite linear combinations of the vectors xk with rational coefficients into a sequence .yn /. Applying Corollary 2.13 (b) (p. 68) we fix for each n a functional 'n 2 X 0 satisfying k'n k 1 and j'n .yn /j D kyn k : Applying Cantor’s diagonal method similarly to the proof of Theorem 1.21 (p. 33), we obtain a subsequence .zk / of .xk / such that the numerical sequence k 7! 'n .zk / converges for each fixed n. Since for ' ? fzk g the numerical sequence .'.zk // vanishes identically, .'.zk // converges for every ' 2 WD f'n g [ fzk g? : Assume temporarily that generates X 0 . Then .'.zk // converges for every ' 2 X by Lemma 2.25 (p. 83), so that the formula 0

ˆ.'/ WD lim '.zk / defines a map ˆ W X 0 ! R. This map is clearly linear. Letting k ! 1 in the inequalities j'.zk /j kzk k k'k sup kzk k k'k k

53

Riesz [379, 380] and Pettis [357].

2.7 Reflexive Spaces: Geometrical Applications

91

we obtain jˆ.'/j sup kxk k k'k k

for every ' 2 X 0 . Since .xk / is bounded, we conclude that ˆ is continuous and kˆk supk kxk k. Since X is reflexive, ˆ 2 X 00 may be represented by a vector x 2 X: ˆ.'/ D '.x/ for all ' 2 X 0 . In view of the definition of ˆ this yields '.zk / ! '.x/ for all ' 2 X 0 , i.e., zk * x. It remains to show that generates X 0 . By Corollary 2.9 (p. 64) it is sufficient to show that ? D f0g. For any given y 2 ? we have 'n .y/ D 0 for all n by the definition of , and y belongs to the closed subspace fzk g?? generated by fzk g. (We apply Corollary 2.9 again.) Choose a subsequence ynk ! y, then kynk k D j'nk .ynk /j D j'nk .ynk y/j kynk yk : Letting k ! 1 we conclude that kyk 0, i.e., y D 0.

t u

Examples We have seen in the previous section that `1 , `1 and c0 have bounded sequences without convergent subsequences. Applying the theorem we conclude again that these spaces are not reflexive.

2.7 Reflexive Spaces: Geometrical Applications Using Theorem 2.30 (p. 90) we may generalize several results of plane geometry, mentioned in the introduction, to arbitrary reflexive Banach spaces. Proposition 2.31 If X is a normed space, then the properties below satisfy the following implications: .a/ H) .b/ H) .c/ H) .d/ H) .e/:

(a) X is reflexive. (b) (Tukey)54 Let A and B be disjoint non-empty convex, closed sets in X. If at least one of them is bounded, then there exist a functional ' 2 X 0 and real numbers

54

Tukey [460].

92

2 Banach Spaces

c1 , c2 such that '.a/ c1 < c2 '.b/ for all a 2 A and b 2 B:

(2.7)

(c) If K X is a non-empty convex, closed set and x 2 X, then there exists a point y 2 K at a minimal distance from x: kx yk kx zk

for all z 2 K:

(d) If M X is a proper non-empty closed subspace, then there exists an x 2 X satisfying kxk D 1

and

dist.x; M/ D 1:

(e) If ' 2 X 0 is a non-zero functional, then there exists an x 2 X satisfying kxk D 1 and

j'.x/j D k'k :

*Remarks • Let us compare property (b) with Theorem 2.5 (c) (p. 61): We recall55 that every infinite-dimensional normed space contains bounded and closed, but noncompact sets. • Klee56 proved the converse implication (b) H) (a): he constructed in every nonreflexive normed space two disjoint non-empty convex, bounded and closed sets, that cannot be separated in the sense of (2.7). • Property (c) is the generalization of the orthogonal projection Theorem 1.5 (p. 12). In strictly convex spaces57 the point y is unique. Indeed, if y1 ; y2 are two distinct points in K with c WD kx y1 k D kx y2 k, then c > 0 (for otherwise y1 D x D y2 ), and .y1 C y2 /=2 2 K is closer to x: x y1 C y2 D .x y1 / C .x y2 / < c: 2 2 See also Proposition 9.10, p. 326. • It is interesting to compare (d) with Proposition 2.1 (b), p. 55. • In Hilbert spaces property (d) is equivalent to the existence of a unit vector, orthogonal to M.

55

See Proposition 2.1, p. 55. Klee [250]. 57 See p. 67. 56

2.7 Reflexive Spaces: Geometrical Applications

93

• Property (e) shows that in a reflexive space X we have k'k D max j'.x/j kxk1

for every functional ' 2 X 0 , i.e., we may write max instead of sup. • James58 also established the implication (e) H) (a) so that the above five properties are in fact equivalent. Proof (a) H) (b). We may repeat the proof of Theorem 2.5 (c) (p. 61), except the proof of the inequality dist.A; B/ > 0. Now we can proceed as follows: If dist.A; B/ D 0, then there exist two sequences .an / A and .bn / B satisfying kan bn k ! 0. If for example A is bounded (the other case is analogous), then there exists a weakly convergent subsequence ank * a. Since an bn * 0, this implies that bnk * a. Since A and B are convex, closed sets, a 2 A and a 2 B, contradicting the disjointness of A and B. (b) H) (c). We may assume by translation that x D 0. It is sufficient to show that every non-empty convex, closed set K has an element of minimal norm. The case 0 2 K is obvious. Henceforth we assume that 0 … K; then r WD dist.0; K/ > 0 by the closedness of K. Assume on the contrary that K has no element of minimal norm. Then we may apply property (b) to the sets A WD fx 2 X W kxk rg and B WD K to get ' 2 X 0 and c1 ; c2 2 R satisfying (2.7). Let .yn / be a sequence in K satisfying kyn k ! r. Then c2 '.yn / D

kyn k ryn kyn k ' c1 ! c1 ; r r kyn k

contradicting the inequality c1 < c2 . (c) H) (d). For any fixed z 2 X n M there exists by (c) a closest point y 2 M to z: kz yk kz uk

for all u 2 M:

Since z y ¤ 0 (because z … M and y 2 M), this may be rewritten as zy uy 1 kz yk kz yk

58

James [226]. See, e.g., Diestel [103] or Holmes [216].

for all u 2 M;

94

2 Banach Spaces

or, using the unit vector x WD .z y/= kz yk, as uy 1 x kz yk

for all u 2 M:

uy If u runs over the subspace M, then kzyk also runs over M, so that dist.x; M/ 1. Since 0 2 M, the converse inequality is obvious. (d) H) (e). Applying (d) to the kernel M WD ' 1 .0/ of ', there exists an x 2 X satisfying

kxk D 1 D dist.x; M/: It suffices to show that j'.z/j j'.x/j kzk for all z 2 X, because this will imply k'k j'.x/j; since kxk D 1, the converse inequality is obvious. The required inequality is obvious if '.z/ D 0. If '.z/ ¤ 0, then the equality '.x/ '.x/ ' x z D '.x/ '.z/ D 0 '.z/ '.z/ implies x

'.x/ z '.z/

2 M, and hence '.x/ j'.x/j z D 1 x x kzk ; '.z/ j'.z/j

i.e., j'.z/j j'.x/j kzk.

t u

*Examples We show that properties (b)–(e) may fail in non-reflexive spaces. Let X D `1 , and fix a positive, strictly increasing sequence .˛n / converging to one, for example ˛n WD n=.n C 1/. P • The formula '.x/ WD ˛n xn defines a functional of norm 1. Indeed, on the one hand we have X j'.x/j jxn j D kxk1 for all x 2 `1 , whence k'k 1. On the other hand, we have k'k j'.en /j D j˛n j for all n, and j˛n j ! 1.

2.7 Reflexive Spaces: Geometrical Applications

95

But the norm k'k D 1 is not attained because j'.x/j < kxk1

x ¤ 0:

for all

Indeed, there is at least one non-zero component xk of x, and then j'.x/j j˛k j jxk j C

X

j˛n j jxn j

n¤k

j˛k j jxk j C

0 such that Br A.B1 /. Proof First we prove that there exists an r > 0 such that B2r A.B1 /:

(2.8)

Since A is onto, YD

1 [ kD1

59

See, e.g., Hörmander [218, 219]. Schauder [413]. 61 Banach [22]. 62 Banach [22]. 63 Banach [24]. 60

A.Bk / D

1 [ kD1

A.Bk /:

2.8 * Open Mappings and Closed Graphs

97

By Baire’s lemma (p. 32) at least one of the sets A.Bk / contains a ball, say Bs .y/ A.Bk /.64 Then we have Bs .y/ A.Bk / D A.Bk /: If x 2 Bs , then x ˙ y 2 Bs .˙y/ A.Bk /; using the convexity of A.Bk /, this yields xD

.x C y/ C .x y/ 2 A.Bk /: 2

We thus have Bs A.Bk /, and (2.8) follows by homogeneity with r WD s=2k. Now we fix an arbitrary point y 2 Br . We seek x 2 B1 satisfying Ax D y. For this we observe that (2.8) implies by similarity the more general relations B21n r A.B2n /;

n D 1; 2; : : : :

Using them we may construct recursively a sequence x1 ; x2 ; : : : in X such that kxn k

0 there exists a large integer N such that X

jJm j < ";

m>N

and the intervals JmC1 ; JmC2 ; : : : still cover A. Examples • (Harnack)8 Every countable set fan g of real numbers is a null set: for each " > 0: it is covered by the intervals .an "3n ; an C "3n / of total length ". • (Cantor’s ternary set)9 There exist uncountable null sets. Let us remove from the unit segment Œ0; 1 its middle third, i.e., the open interval .1=3; 2=3/. There

6

Hankel [190] (p. 86), Ascoli [11], Smith [423] (p. 150), du Bois-Reymond [52], Harnack [194]. By slightly enlarging them we may assume that all the intervals are open. 8 Harnack [194]. 9 Smith [423], Cantor [72] (p. 207). Many analogous sets appear “naturally” in combinatorial number theory, see, e.g., Erd˝os–Joó–Komornik [127], Komornik–Loreti [260], de Vries–Komornik [101], Komornik–Kong–Li [259], de Vries–Komornik–Loreti [102]. 7

4.2 Differentiability: Null Sets Fig. 4.2 The sets Cn

155

C0

C1

C2

0

1

1

remain two disjoint segments Œ0; 1=3 and Œ2=3; 1 of total length 2=3. Next remove from each of them their middle thirds: there remain four disjoint segments of total length .2=3/2 ; see Fig. 4.2. Continuing by induction, after n steps we obtain a set Cn , which is the union of 2n disjoint compact segments of length 3n each. The intersection C of this decreasing set sequence is a compact set, called Cantor’s ternary set. It is a null set. Indeed, for each " > 0 there is a large integer n such that .2=3/n < "; then the 2n disjoint segments of Cn form a finite cover of C with total length D .2=3/n < ". By construction C is formed by the real numbers x that may be written in base 3 in the form xD

1 X ci i 3 iD1

with .ci / f0; 2g, i.e., without using the digit ci D 1. Since all sequences .ci / f0; 2g occur here, the formula 1 1 X X ci ci ! 7 i iC1 3 2 iD1 iD1

defines a map of C onto Œ0; 1. The latter set is uncountable, hence C is also uncountable. • It follows from our next proposition that R is not a null set. Let us resume the basic properties of null sets. Proposition 4.3 (a) The empty set is a null set. (b) The subsets of a null set are null sets.

156

4 * Monotone Functions

(c) The union of countably many null sets is a null set. P (d) (Borel)10 If an interval sequence .Ik / covers an interval I then jIj jIk j. Consequently, non-degenerate intervals are not null sets. Proof (a) and (b) are obvious. (c) Given " > 0 arbitrarily, we cover the null set An by an interval set .Ink / of total length "2n , n D 1; 2; : : : : Then the union of all these intervals form a cover of [An of total length ". (d) We may assume that I is non-degenerate. First we consider the case where I D Œa; b is compact and the intervals Ik are open. Let .a1 ; b1 / be the first interval in .Ik / that contains the point a. Continuing by induction, if bn b for some n 1, then let .anC1 ; bnC1 / be the first interval in .Ik / that contains the point bn . The construction stops after a finite number of steps because bN > b for some N. For otherwise the bounded sequence .bn / would converge to some x b, and we would have x 2 I` for some `. Since I` is open, there would exist an index m such that bn 2 I` for all n m. By construction this would mean that the intervals .an ; bn / would precede I` in the sequence .Ik / for all n > m. But this is absurd because b1 < b2 < by construction, so that the intervals .an ; bn / are pairwise distinct. It follows that jIj D b a < bN a1 D

N X

.bi bi1 / C b1 a1

iD2

N X

.bi ai /

X

jIk j:

iD1

In the general case we fix a number ˛ > 1, a compact subinterval J I of length jIj=˛, and for each P n an open interval Jn In of length ˛jIn j. The sequence .Jn / covers J, so that jJn j jJj by the first part of the proof. In other words we have P ˛ jIn j jIj=˛, and we conclude by letting ˛ ! 1. t u Let us introduce a convenient terminology: Definition A property holds almost everywhere11 (shortly a.e.) if it holds outside a null set.

10 11

Borel [59]. His proof was based on a construction of Heine [200] (p. 188). Lebesgue [293] (p. 7).

4.3 Jump Functions

157

We may now state a deep theorem: Theorem 4.4 (a) (Lebesgue)12 Every monotone function f W I ! R is a.e. differentiable. (b) For each null set A there exists a non-decreasing, continuous function f W R ! R that is non-differentiable at the points of A. Part (a) of this theorem will be proved in the next two sections. *Proof of part (b) Choose a sequence .Jm / of open intervals, of finite total length, and covering each point of A infinitely many times. P Denoting the length of the interval Jm \ .1; x/ by fm .x/, the formula f WD fm defines a non-decreasing function f W R ! R. Since the series is uniformly convergent and each fm is continuous, f is also continuous. We complete the proof by establishing the relation f .a C h/ f .a/ D1 h h&a lim

for each a 2 A. Fix an arbitrarily large number N, and then choose a sufficiently small number ı > 0 such that at least N intervals Jm contain Œa; a C ı, say Jm1 ; : : : ; JmN . Then f .a C h/ f .a/

N X

fmk .a C h/ fmk .a/ D Nh

kD1

for all 0 < h < ı.

t u

4.3 Jump Functions Since every interval is the union of countably many compact intervals, it is sufficient to prove Lebesgue’s theorem for compact intervals I D Œa; b. In this section we follow an approach of Lipi´nski and Rubel13 to prove some special cases of the theorem.

12 Lebesgue [290], pp. 128–129. He considered only the case of continuous functions. Before him Weierstrass conjectured the existence of continuous and monotone, but nowhere differentiable functions; see Hawkins [198], p. 47. 13 Lipi´nski [307], Rubel [401].

158

4 * Monotone Functions

Fig. 4.3 Meaning of EC 2.5

2

1.5

1

0.5

0

1

2 x

3

4

We start with a lemma: Lemma 4.5 Let f W Œa; b ! R be a non-decreasing function. For each C > 0 we denote by EC the set of points a < x < b for which there exist numbers s D sx and t D tx satisfying s < x < t and f .t/ f .s/ > C.t s/:

(4.1)

Then EC is the union of countably many intervals .an ; bn / of total length 4C1 . f .b/ f .a//. Remark The set EC contains all points at which f has a derivative > C, but p it may contain other points as well. For example, consider the function f .x/ WD x in the p interval Œ0; 4. For C D 1= 2 we have ˚

f 0 > C D .0; 1=2/ and EC D .0; 2/:

(See Fig. 4.3: for 0 < x < 2 we may choose sx D 0 and tx D .x C 2/=2.) Proof The set EC is open by definition, hence it is the union of disjoint open intervals .an ; bn /. We also observe that if x 2 .an ; bn /, then .sx ; tx / .an ; bn / by definition. Fix for each n a compact subinterval Œa0n ; b0n .an ; bn / of length b0n a0n D .bn an /=2: It is covered by the intervals .sx ; tx /;

x 2 Œa0n ; b0n :

(4.2)

4.3 Jump Functions

159

Since Œa0n ; b0n is compact, there exists a finite subcover .s1 ; t1 /; : : : ; .sN ; tN /. Choose a finite subcover with N as small as possible. Then no point of [.sk ; tk / is covered more than twice, because if three intervals have a common point, then one of them belongs to the union of the other two. Consequently, using (4.1) and the relations .sk ; tk / .an ; bn /, we have b0n a0n

N N X X .tk sk / C1 . f .tk / f .sk // 2C1 . f .bn / f .an //: kD1

kD1

Using (4.2) this yields the required inequality: X X .bn an / 4C1 . f .bn / f .an // 4C1 . f .b/ f .a//: t u As a first application of this lemma, we prove that a non-decreasing function cannot have an infinite derivative at many points. More precisely, we have the Lemma 4.6 If f W Œa; b ! R is a non-decreasing function, then Df .x/ WD lim sup y!x

f .y/ f .x/ 0, so that the set of these points may be covered by a set of intervals of total length 4. f .b/f .a//=C. We conclude by letting C ! 1. t u As a second application we prove Lebesgue’s theorem in a special case.

P Definition By a jump function we mean a function P f W I ! R of the form f D fk where .ak / I is a given sequence of points, Sk is a nonnegative convergent numerical sequence, and fk .x/ D 0 if fk .x/ D Sk

x < ak ;

if

x > ak ;

0 fk .ak / Sk : Every jump function is non-decreasing. Proposition 4.7 If f W I ! R is a jump function , then f 0 D 0 a.e. Proof We may assume that I D Œa; b is compact. It suffices to show that Df C a.e. for every fixed C > 0.

160

4 * Monotone Functions

Fix an arbitrary " > 0,14 then choose a large N such that 1 X

Sk < ":

kDNC1

Then the function h WD

1 X

fk

kDNC1

is non-decreasing, and h.b/ h.a/ < ". By Lemma 4.5 we have Dh C outside a set of intervals of total length < 4"=C. Observe that the function f hD

N X

fk

kD1

has zero derivative everywhere, except a1 ; : : : ; aN . Hence Df C outside a set of intervals of total length < 4"=C. We conclude by letting " ! 0. t u Using jump functions we may isolate the discontinuous part of non-decreasing functions: Proposition 4.8 Every bounded non-decreasing function f W I ! R is the sum of a continuous non-decreasing function and a jump function. Proof Since f is bounded, extending f by constants we may assume that I D R. Let .ak / be the (finite or infinite) P sequence of discontinuities of f , and set Sk D f .ak C 0/ f .ak 0/. The series Sk is convergent because f is bounded. Introduce the functions fk as in thePdefinition of the jump functions, and set fk .ak / WD f .ak / f .ak 0/. Then h WD fk is a jump function by definition, while g WD f h is non-decreasing and continuous.15 t u

14 15

The following proof is due to Á. Császár; see Sz.-Nagy [448]. See Exercise 4.3 at the end of this chapter, p. 165.

4.4 Proof of Lebesgue’s Theorem

161

Fig. 4.4 Dini derivatives 0.2

0.1

–0.2

x 0.1

–0.1

0.2

0

–0.1

–0.2

–0.3

–0.4

4.4 Proof of Lebesgue’s Theorem In view of Propositions 4.7 and 4.8 it is sufficient to consider a non-decreasing and continuous function f W Œa; b ! R, defined on a compact interval. In this section we present an elementary proof due to F. Riesz.16 We introduce the Dini derivatives17 : D f .x/ WD lim sup

f .y/ f .x/ ; yx

DC f .x/ WD lim sup

d f .x/ WD lim inf

f .y/ f .x/ ; yx

dC f .x/ WD lim inf

yx y!x

f .y/ f .x/ ; yx

f .y/ f .x/ : yx

Since f is non-decreasing, they are all nonnegative. Example For f .x/ WD x C x sin.1=x/ we have D f .0/ D DC f .0/ D 1

and d f .0/ D dC f .0/ D 0I

see Fig. 4.4.

16 Riesz [386, 387]. The proof may be adapted to the discontinuous case: see Riesz and Sz.-Nagy [394], Sz.-Nagy [448]. See also other elementary proofs of Austin [14] and Botsko [63]. 17 Dini [109] (Sect. 145).

162

4 * Monotone Functions

Fig. 4.5 Invisible points from the right

a1

b1

a2

b2

Assume for a moment the following lemma: Lemma 4.9 The inequality DC f d f holds almost everywhere. Then applying this lemma to the function f .x/ we have also D f .x/ dC f .x/ a.e., and hence 0 DC f .x/ d f .x/ D f .x/ dC f .x/ DC f .x/ a.e. Since DC f .x/ < 1 a.e. by Lemma 4.6, we conclude that the four Dini derivatives are finite and equal a.e., proving Lebesgue’s theorem. The main tool for the proof of Lemma 4.9 is the “Rising sun lemma” of Riesz. We introduce the following notion: Definition Let g W Œa; b ! R be a continuous function on a compact interval. The point a < x < b is invisible (from the right) if there exists a y > x such that g.y/ > g.x/. (See Fig. 4.5.) Lemma 4.10 (“Rising sun lemma”) 18 The invisible points (from the right) form a union of disjoint open intervals .ak ; bk /, and g.ak / g.bk / for every k.19 Proof The set of invisible points is open by the continuity of g, hence a union of disjoint open intervals .ak ; bk /. Assume on the contrary that g.ak / > g.bk / for some k. Fix a number g.ak / > c > g.bk / and set x WD sup fak t bk W g.t/ cg : By the continuity of g we have g.x/ D c and thus ak < x < bk . Since x is invisible, there exists a y > x such that g.y/ > g.x/ D c. Since g < c on .x; bk by the choice of x, we have y > bk . But this contradicts the visibility of bk because g.y/ > c > g.bk /. t u Proof of Lemma 4.9 It suffices to show that for any fixed rational numbers c1 < c2 , E WD fx 2 .a; b/ W d f .x/ < c1 < c2 < DC f .x/g

18 19

Riesz [386, 387]. See the correspondence of Riesz in [443, 444] for the history of this result. It is easy to see that we even have g.ak / D g.bk / if ak ¤ a.

4.4 Proof of Lebesgue’s Theorem

163

is a null set. Indeed, then their (countable) union is also a null set, and d f .x/ DC f .x/ outside them. We are going to show that for any fixed open subinterval .a0 ; b0 / of .a; b/, we may cover E \.a0 ; b0 / by a (countable) set of open intervals of total length < .c1 =c2 /.b0 a0 /. Iterating this procedure we will get that E D E \ .a; b/ may be covered for each n D 1; 2; : : : by a set of open intervals of total length < .c1 =c2 /n .b a/. Since c1 =c2 < 1, letting n ! 1 we will conclude that E is a null set. If x 2 E \ .a0 ; b0 /, then f .y/ f .x/ < c1 yx for some a0 < y < x, i.e., f .y/ c1 y > f .x/ c1 x: In other words, x is invisible from the left20 for the function g.t/ WD f .t/ c1 t;

t 2 Œa0 ; b0 :

Applying Lemma 4.10 for the function t 7! g.t/, E \ .a0 ; b0 / may be covered by a countable set of disjoint open intervals .ak ; bk / such that g.ak / g.bk /, i.e., f .bk / f .ak / c1 .bk ak / for every k. Now consider one of these intervals .ak ; bk /. If x 2 E \ .ak ; bk /, then f .y/ f .x/ > c2 yx for some x < y < bk , i.e., f .y/ c2 y > f .x/ c2 x: In other words, x is invisible from the right for the function g.t/ WD f .t/ c2 t;

t 2 Œak ; bk :

20 We say that x is invisible from the left for a function g if x is invisible from the right for the function t 7! g.t/.

164

4 * Monotone Functions

Applying Lemma 4.10, E \ .ak ; bk / may be covered by a countable set of disjoint open intervals .akm ; bkm / such that g.akm / g.bkm /, i.e., f .bkm / f .akm / c2 .bkm akm / for every m. Consequently, the intervals .akm ; bkm / cover E \ .a0 ; b0 /, and X

.bkm akm /

k;m

1 X f .bkm / f .akm / c2 k;m

1 X f .bk / f .ak / c2 k c1 X .bk ak / c2 k

c1 0 .b a0 /: c2 t u

4.5 Functions of Bounded Variation The difference of two monotone functions is not necessarily monotone. However, it follows from Proposition 4.2 and Theorem 4.4 (pp. 153 and 157) that these functions still also have at most countably many discontinuities, and they are differentiable a.e. In this section we briefly discuss these functions. Definition A function f W I ! R is of bounded variation21 if there exists a number A such that n X

j f .xi / f .xi1 /j A

iD1

for every finite set of points x0 < < xn in I. The smallest such number A is called the total variation of f . Remarks • Every function of bounded variation is bounded.

21 Jordan [229]. He introduced this notion in order to give an elegant formulation of Dirichlet’s theorem on the convergence of Fourier series.

4.6 Exercises

165

• In the case of a bounded interval I, f has a bounded variation ” it is rectifiable, i.e, if its graph has a finite arc length. • Every monotone and bounded function has a bounded variation. • The functions of bounded variation form a vector space. Our last remarks imply that the difference of two monotone and bounded functions has a bounded variation. The converse also holds: Proposition 4.11 (Jordan)22 Every function of bounded variation is the difference of two non-decreasing and bounded functions. Proof If f W I ! R has bounded variation, then its restriction to any subinterval also has bounded variation. Let us denote by g.x/ the total variation of f on I \ .1; x/ for each x 2 I. Then 0 g T, where T denotes the total variation of f , so that g is a bounded function. If y 2 I and x < y, then g.x/ C jf .y/ f .x/j g.y/ by the definition of the total variation. It follows that g is non-decreasing, and then that g f is also nondecreasing because .g f /.y/ .g f /.x/ D g.y/ g.x/ f .y/ f .x/ g.y/ g.x/ jf .y/ f .x/j 0: Since f and g are bounded, h WD g f is bounded, too. Therefore the decomposition f D g h has the required properties. t u Remark It follows from the theorems of Jordan and Lebesgue that if f W I ! R has bounded variation and I D Œa; b, then f has a finite left limit at every a < x b, a finite right limit at every a x < b, and (applying Lebesgue’s theorem) that f is a.e. differentiable.

4.6 Exercises Exercise 4.1 Given an arbitrary null set D, does there exist a monotone function f W R ! R that is non-differentiable exactly at the points of D? Exercise 4.2 If C denotes Cantor’s ternary set, then C C D Œ0; 1. Exercise 4.3 Prove that the function g in the proof of Proposition 4.8 (p. 160) is non-decreasing and continuous.

22

Jordan [229].

166

4 * Monotone Functions

In the remaining exercises we consider bounded closed intervals. Exercise 4.4 (Lebesgue’s criterium)23 Let f W Œa; b ! R, then f

is Riemann integrable

”f

is bounded, and continuous a.e.

Exercise 4.5 Let f ; g W Œa; b ! R have bounded variations. (i) fg, max ff ; gg and min ff ; gg also have bounded variations. (ii) jf j has bounded variation. (iii) If moreover, inf jgj > 0, then f =g also has bounded variation. Exercise 4.6 If f W Œa; b ! R is continuous, then f and jf j have bounded variations at the same time. Is the continuity assumption necessary? Exercise 4.7 For which values of ˛; ˇ does f .x/ WD x˛ sin variation on Œ0; 1?

1 have bounded xˇ

Exercise 4.8 If f W Œa; b ! R has bounded variation, then f has finite left and right limits everywhere, and f has at most countably many discontinuities. Exercise 4.9 (i) If f W Œa; b ! R is Lipschitz continuous, then it has bounded variation. (ii) Construct a Hölder continuous function f W Œa; b ! R which is not of bounded variation. Exercise 4.10 Write the following functions as the difference of two nondecreasing functions: (i) f .x/ D sign x in Œ1; 1; (ii) f .x/ D sin x in Œ0; 2. Exercise 4.11 (Helly’s selection theorem)24 Let fn W Œa; b ! R, n D 1; 2; : : : be a uniformly bounded sequence of functions of bounded variation. Assume that their total variations are bounded by some constant. Prove the existence of an everywhere convergent subsequence by proving the statements below. (i) We may assume that all functions fn are non-decreasing. Henceforth we consider this special case. (ii) There exists a subsequence . fn1 / . fn / converging in a; b and in all rational points of .a; b/. Write .x/ WD lim fn1 .x/;

23

x 2 E WD fa; bg [ ..a; b/ \ Q/ :

Lebesgue [288], p. 29. Helly [204]. This is a weak compactness theorem in the space of functions of bounded variation. We follow Natanson [332].

24

4.6 Exercises

167

(iii) extends to a non-decreasing function W Œa; b ! R. (iv) fn1 .x/ ! .x/ at all points x 2 .a; b/ where is continuous. (v) There exists a second subsequence . fn2 / . fn1 / which also converges at the points of discontinuity of .

Chapter 5

The Lebesgue Integral in R

I turn with terror and horror from this lamentable scourge of continuous functions with no derivatives!—Letter of Hermite to Stieltjes, 1893 In former times when one invented a new function it was for a practical purpose; today one invents them purposely to show up defects in the reasoning of our fathers and one will deduce from them only that.—H. Poincaré

The Riemann integral has the drawback that many important functions are not integrable and the limiting processes are complicated: Examples • (Dirichlet function)1 The function ( f .x/ WD

1

if x is rational;

0

if x is irrational

Ris not Riemann integrable. However, since f D 0 a.e., it is tempting to define f dx WD 0. • Let us enumerate the rational numbers into a sequence .rn /. Then the functions ( fn .x/ WD

1

if x D r1 ; : : : ; rn ;

0

otherwise

R are Riemann integrable, R R fn dx D 0 for all n, and fn ! f a.e. We would like to conclude that fn dx ! f dx, but the last integral is not defined.

1

Dirichlet [112, pp. 131–132].

© Springer-Verlag London 2016 V. Komornik, Lectures on Functional Analysis and the Lebesgue Integral, Universitext, DOI 10.1007/978-1-4471-6811-9_5

169

5 The Lebesgue Integral in R

170

R • The formula kgk WD jgj dx defines a natural norm in the vector space of Riemann integrable functions. For this norm the above sequence . fn / satisfies the Cauchy criterion, but it is not convergent. The Lebesgue integral eliminates these difficulties: much more functions are integrable and they are easier to manipulate. One key of this theory is that we do not distinguish between two functions if they are equal outside some null set: Definition The functions f1 W D1 ! R and f2 W D2 ! R are equal almost everywhere (a.e.) if D 1 n D2 ;

D 2 n D1

and

fx 2 D1 \ D2 W f1 .x/ ¤ f2 .x/g

are null sets. This is an equivalence relation that is compatible with the usual algebraic operations: if f1 D g1 and f2 D g2 a.e., then j f1 j D jg1 j a.e.; f1 ˙ f2 D g1 ˙ g2 f1 f2 D g1 g2

a.e.;

a.e.;

min f f1 ; f2 g D min fg1 ; g2 g

a.e.;

max f f1 ; f2 g D max fg1 ; g2 g

a.e.

If, moreover, f2 ¤ 0 a.e., then f1 =f2 D g1 =g2 a.e. Finally, if fn ! f a.e., and fn D gn a.e. for every n, then gn ! f a.e.2 In view of these properties we often identify two functions if they are equal almost everywhere.3 Hence we often write f D g, f g, f > g instead of f D g a.e, f g a.e., f > g a.e., and a sequence . fn / is called simply nonnegative, non-decreasing or non-increasing if it is nonnegative a.e., non-decreasing a.e. or non-increasing a.e.

5.1 Step Functions Definition ' W R ! R is a step function if there exist finitely many points 1 < x0 < < xn < 1

2

It is essential here that we use countable covers in the definition of null sets. To be precise, we should use equivalence classes of functions but we follow the traditional, looser terminology. 3

5.1 Step Functions

171

Fig. 5.1 Step function

c2

c1

x0

c3

x1

x2

x3

and real numbers c1 ; : : : ; cn such that a.e., 8 ˆ if x < x0 , ˆ0 ˆ ˆ ˆ ˆ if x0 < x < x1 , ˆ < c1 '.x/ D : : : ˆ ˆ ˆ ˆ if xn1 < x < xn , cn ˆ ˆ ˆ :0 if x < x. n

See Fig. 5.1. The class of step functions is denoted by C0 . Remarks • We may always add to the definition a finite number of arbitrary points xi . Consequently, for finitely many given step functions we may always assume that they are defined by the same points xi . • Once the points xi are given, the corresponding numbers ci are uniquely determined because the non-degenerate intervals in the definition of '.x/ are not null sets. Definition By the integral of a step function we mean the number Z ' dx WD

n X

ck .xk xk1 /:

iD1

In order to show the correctness of this definition we introduce two useful notions:

5 The Lebesgue Integral in R

172

Definitions A vector space C of real functions is a vector lattice if ';

2 C H) max f'; g ; min f'; g 2 C:

A linear functional L W C ! R defined on a vector lattice C is positive if ' 0 H) L' 0: Remarks • Using the relations j'j D max f'; 'g and max f'; g D

'C

C j' j ; 2

min f'; g D

'C

j' 2

j

we see that a vector space C is a vector lattice ” ' 2 C H) j'j 2 C: • Every positive linear functional is monotone, i.e., '

H) L' L :

Using the remark following the definition of step functions the next result can be shown easily: Proposition 5.1 (a) C0 is a vector lattice. (b) The integral of a step function does not depend on the particular choice of the points xi . (c) The integral of step functions is a positive linear functional on C0 . The following two “innocent-looking” lemmas are due to Riesz. Almost the whole theory of Lebesgue integral will follow from them. The first one is a simple variant of a classical theorem of Dini4 : 5 RLemma 5.2 If a sequence .'n / of step functions satisfies 'n .x/ & 0 a.e., then 'n dx ! 0.

Proof Fix a compact interval Œa; b and a number M > 0 such that '1 D 0 outside Œa; b, and '1 < M on Œa; b. Changing the functions 'n on some null set if necessary, we may assume that they all vanish outside Œa; b.

4 5

See Proposition 8.24 below, p. 292. The notation means that the sequence is non-increasing and converges to zero for almost every x.

5.1 Step Functions

173

Fix an arbitrarily small number " > 0. Outside a suitable null set E all functions 'n are continuous, and the sequence tends to zero. Let us cover E by a countable open interval system fIg of total length < "=.2M/. If x0 … E, then 'n .x0 / ! 0, so that 'n0 .x0 /

0 there exists a ı > 0 such that X

jF.bk / F.ak /j < "

for every finite disjoint interval system f.ak ; bk /g of total length < ı. Remarks • Every Lipschitz continuous p function is absolutely continuous. On the other hand, the function F.x/ WD x is absolutely continuous on Œ0; 1, but not Lipschitz continuous.

2 Dini [110, p. 24], Harnack [193, p. 220], Lebesgue [290, pp. 128–129], Vitali [470]. We obtain an equivalent definition by using arbitrary intervals instead of open intervals.

6.1 Absolute Continuity

199

Fig. 6.1 The Cantor function

1

0

1

• (Cantor function)3 Every absolutely continuous function is uniformly continuous. On the other hand, consider Cantor’s ternary set C (p. 155), and define a function F W C ! Œ0; 1 by the formula 1 1 X X "i "i ! 7 : i iC1 3 2 iD1 iD1

Then F is surjective, non-decreasing and continuous. (See Fig. 6.1.) By construction the set Œ0; 1nC is a countable union of disjoint open intervals. If .a; b/ is one of these intervals, then F.a/ D F.b/ by the surjectivity of F. Set F.x/ WD F.a/ for a < x < b, then the extended function F W Œ0; 1 ! Œ0; 1 is continuous on a compact set, hence uniformly continuous. But F is not absolutely continuous. To see this we consider the sets Cn introduced during the construction of C. For each n, Cn is the union of 2n n n disjoint P intervals Œai ; bi of length 3 each, hence of total length .2=3/ . We have .F.bi / F.ai // D 1 for every n by the definition of F, although the total length .2=3/n tends to zero as n ! 1. • If I is bounded, then every absolutely continuous function f W I ! R has bounded variation.4 Applying Jordan’s Proposition 4.11 and Lebesgue’s Theorem 4.4 (pp. 157 and 165) it follows that every absolutely continuous function is a.e. differentiable.

3 4

Cantor [73], Lebesgue [290], Vitali [470]. The identity map of R shows that this is not necessarily true for unbounded intervals.

200

6 * Generalized Newton–Leibniz Formula

Proposition 6.1 An absolutely continuous function F W I ! R sends every null set of I into a null set. Proof Since F is uniformly continuous, it can be extended by continuity to I, and the extended function is still absolutely continuous. We may therefore assume that I is a closed interval. Fix a null set E I and a number " > 0 arbitrarily, and choose ı > 0 according to the definition of absolute continuity. We have to find an interval system of total length ", covering F.E/. Let us cover E with a sequence of half-open intervals Ik D Œak ; bk / I, k D 1; 2; : : :, of total length < ı.5 Replacing each Ik with the connected components of Ik n .I1 [ [ Ik1 / we may also assume that the intervals Ik are pairwise disjoint. Moreover, uniting the intervals having a common endpoint we may even assume that the closed intervals Ik are pairwise disjoint. Applying Weierstrass’s theorem we may choose in each interval Œak ; bk two points a0k ; b0k such that F.a0k / F.x/ F.b0k / for all

x 2 Œak ; bk :

Then the intervals ŒF.a0k /; F.b0k / cover F.E/, and their total length is at most ", because for each positive integer n we have n n X ˇ 0 ˇ X ˇb a0 ˇ jbk ak j < ı; k k kD1

kD1

whence n X ˇ ˇ ˇF.b0 / F.a0 /ˇ < " k

k

kD1

by the choice of ı.

t u

Proposition 6.2 If F is the indefinite integral of an integrable function f W Œa; b ! R, then6 (a) F is absolutely continuous; (b) F has bounded variation; (c) F 0 D f a.e. For the proof of (c) we temporarily admit the following

5 6

We may assume that E does not contain the right endpoint of I. Lebesgue [290], Vitali [470].

6.1 Absolute Continuity

201

P Proposition 6.3 (Fubini)7 If a series Gn of nonnegative, non-decreasing functions converges a.e. on some interval I, then X

0 Gn

D

X

G0n

a.e. on I:

(6.1)

Proof of Proposition 6.2 (a) Given any " > 0, by Proposition 5.14 (p. 185) we may choose a step function ' satisfying Z

b

j f 'j dx < "=2:

a

Fix a number A such that j'j < A. Consider a finite number of pairwise disjoint intervals .ak ; bk / Œa; b, of total length < ı WD "=2A. Then X

jF.bk / F.ak /j D

XˇˇZ ˇ XZ

bk

ˇ ˇ f dxˇ

ak bk

jf 'j dx C

ak

Z

b

jf 'j dx C A

XZ

bk

j'j dx

ak

X .bk ak /

a

" < C Aı 2 D ": This proves the absolute continuity of F. (b) The nonnegative functions fC WD max f f ; 0g

and f WD max ff ; 0g

are integrable, and f D fC f . Their indefinite integrals are bounded, nondecreasing functions, hence their difference F has a bounded variation. (c) The proposition is obvious for step functions. If f 2 C1 , then choose a nondecreasing sequence . fn / of step functions, converging a.e. to f . Their indefinite integrals Fn satisfy Fn0 D fn a.e. by our previous remark, and Fn ! F by the definition of the integral. Applying Proposition 6.3 with Gn WD FnC1 Fn we obtain that Fn0 F10 ! 0 F F10 a.e., i.e., fn ! F 0 a.e. On the other hand, we have fn ! f a.e., so that F 0 D f a.e.

7

Fubini [165].

202

6 * Generalized Newton–Leibniz Formula

The general case follows because every integrable function is the difference of two functions of C1 . t u Proof of Proposition 6.3 Since every interval is a countable union of compact intervals, we may assume that I D Œa; b is compact. P 0 (a) We prove Gn converges a.e. Let Sn D G1 C C Gn and P that the series S D Gn , then Sn ! S

on Œa; b

everywhere.

(6.2)

Since the functions Sn and S areP non-decreasing, apart from a null set they are differentiable in Œa; b. The series G0n .x/, i.e., the sequence .Sn0 .x// converges at each differentiability point x. Indeed, by the non-decreasingness of Gn we have SnC1 .x C h/ SnC1 .x/ S.x C h/ S.x/ Sn .x C h/ Sn .x/ h h h for all h satisfying x C h 2 Œa; b, and hence 0 .x/ S0 .x/ < 1 Sn0 .x/ SnC1

for every n. (b) For the proof of (6.1) it suffices to find a sequence n1 < n2 < of indices such that S0 Sn0 k ! 0 a.e.

(6.3)

By (6.2) we may choose n1 < n2 < satisfying S.b/ Snk .b/ < 2k for every k. Then the series X S.b/ Snk .b/ converges. Since 0 S.x/ Snk .x/ S.b/ Snk .b/ for all a x b, it follows that the series interval Œa; b.

P .S Snk / converges on the whole

6.2 Primitive Function

203

P The last series is of the same type asP Gn . Applying the already proved property (a), we conclude that the series .S0 Sn0 k / converges a.e. But then its general term tends to zero a.e., i.e., (6.3) holds. t u Using Proposition 6.2 we may investigate the density of sets: Definition A measurable set A set has density d at a point x 2 R if .A \ In / !d jIn j

(6.4)

for every sequence .In / of non-degenerate intervals, containing x and satisfying jIn j ! 0. We always have 0 d 1; for example a set has density one at each point of its interior. Much more is true: Proposition 6.4 (Lebesgue)8 Every measurable set A set has density one at a.e. point of A. Proof Since density is a local property, we may assume that A is bounded. Then A integrable, and its indefinite integral F satisfies F 0 D A a.e. by Proposition 6.2 (p. 200). The equality F 0 .x/ D A .x/ means that (6.4) holds with d D A .x/ if x is an endpoint of each interval In . The general case follows from the identity t F.x C t/ F.x/ s F.x/ F.x s/ F.x C t/ F.x s/ D C ; tCs tCs t tCs s valid for all t; s > 0, and from the equality s t C D 1: tCs tCs t u

6.2 Primitive Function Proposition 6.2 motivates the following Definition F W Œa; b ! R is a primitive function of f W Œa; b ! R if F is absolutely continuous, has bounded variation, and F 0 D f a.e.

8 Lebesgue [290, pp. 123–124]. See also Zajícek [491] for a direct proof using measure theory, and Riesz–Sz.-Nagy [394] for an extension to non-measurable sets A.

204

6 * Generalized Newton–Leibniz Formula

We have the following important generalization of the Newton–Leibniz formula: Theorem 6.5 (Lebesgue–Vitali)9 Let f W Œa; b ! R. (a) f has a primitive function ” f is integrable. (b) If F is a primitive function of f , then Z

b

f dx D F.b/ F.a/: a

First we complement Lebesgue’s differentiability theorem (p. 157): Proposition 6.6 (a) If F W Œa; b ! R has bounded variation, then F 0 is integrable. (b) If F W Œa; b ! R is non-decreasing, then10 Z

b

F 0 dx F.b/ F.a/:

a

Examples In the absence of absolute continuity the last inequality may be strict. • The simplest example is the discontinuous sign function: Z

1 1

sign0 dx D 0 < 2 D sign.1/ sign.1/:

• The Cantor function F W Œ0; 1 ! Œ0; 1 of the preceding section provides a more surprising example. We recall that F is continuous, non-decreasing and surjective. We also have F 0 .x/ D 0 a.e. because F is constant on each interval of the complement of C by construction. Hence11 Z

1 0

F 0 dx D 0 < 1 D F.1/ F.0/:

• There exist even continuous and strictly increasing functions F with F 0 D 0 a.e.12 9

Lebesgue [290], Vitali [466]. The theorem greatly extended former results of Darboux [95, pp. 111–112] and Dini [109, Sect. 197]. Denjoy [98–100] obtained even more complete results; see, e.g., Natanson [332], Bartle [30]. 10 Lebesgue [290]. 11 Lebesgue [290], Vitali [466]. The graph of F is often called the “Devil’s staircase”; see Fig. 6.1, p. 199. See a related, “natural” example in Komornik–Kong–Li [259]. 12 See, e.g., an example of F. Riesz in Sz.-Nagy [448].

6.2 Primitive Function

205

Proof We may assume by Jordan’s theorem (p. 165) that F is non-decreasing. Extending F as a constant to the left and to the right of its domain, we may also assume that Œa; b D R. Finally, by Propositions 4.7 and 4.8 we may assume that F is continuous. The formula Dn .x/ WD n.F.x C n1 / F.x//;

n D 1; 2; : : :

defines a sequence of nonnegative, continuous functions on R. Their integrals form a bounded sequence on each compact interval ŒN; N because by the continuity of F we have Z

Z

N N

Z

NCn1

Dn dx D n

NCn1

F dx

F dx ! F.N/ F.N/ N

N

as n ! 1. Since Dn ! F 0 a.e. on ŒN; N by Lebesgue’s theorem (p. 157), F 0 is integrable on ŒN; N by the Fatou lemma (p. 183), and Z

N

F 0 dx F.N/ F.N/:

N

Since F is non-decreasing, Z

F 0 ŒN;N dx F.1/ F.1/;

N D 1; 2; : : : :

Finally, F 0 ŒN;N % F 0 a.e., so that F 0 is integrable and Z

1

F 0 dx F.1/ F.1/

1

by the Beppo Levi theorem.

t u

Proof of Theorem 6.5 (a) If f is integrable, then its indefinite integral is a primitive function of f by Proposition 6.2. Conversely, if F is a primitive function of f , then f D F 0 a.e., and f integrable by the preceding proposition. t u For the proof of part (b) we need a lemma: Lemma 6.7 If H W I ! R is non-decreasing, absolutely continuous and H 0 D 0 a.e., then H is constant. Proof It is sufficient to consider the case where I D Œa; b is compact. Let us denote by E the null set of the points x 2 Œa; b where the property H 0 .x/ D 0 fails. By Proposition 6.1 its image H.E/ is also a null set. We are going to show that the image of the complementary set F WD Œa; b n E is a null set, too. Fix " > 0 arbitrarily. Since H 0 D 0 on F, for each x 2 F there exists

206

6 * Generalized Newton–Leibniz Formula

x < y < b such that H.y/ H.x/ < ": yx This means that x is invisible from the right with respect to the function g.t/ WD "t H.t/. Applying the “Rising Sun” lemma (p. 162), F has a countable cover by pairwise disjoint open intervals .ak ; bk / satisfying g.ak / g.bk /, i.e., H.bk / H.ak / ".bk ak /: Hence H.F/ may be covered by the system of intervals ŒH.ak /; H.bk / of total length ".b a/. Since " can be chosen arbitrarily small, this proves that H.F/ is a null set. We conclude from the preceding that the interval H.I/ D H.E/ [ H.F/ is a null set, so that it is a one-point set. In other words, H is constant. t u Proof of Theorem 6.5 (b) We have to show that if F W Œa; b ! R is absolutely continuous and has bounded variation, then Z

b

F 0 dx D F.b/ F.a/:

a

Observing that in the Jordan decomposition F D g h of F (Proposition 4.11) the functions g; h are also absolutely continuous, we may assume that F is nondecreasing. By Proposition 6.6 f WD F 0 is integrable, and by Proposition 6.2 the indefinite integral G of f is absolutely continuous, and Z

b

F 0 dx D G.b/ G.a/:

a

It suffices to show that H WD FG is constant. This readily follows from Lemma 6.7 because H is absolutely continuous, and H 0 D F 0 G0 D 0 a.e. t u Remark (Lebesgue Decomposition) 13 Let F W Œa; b ! R be a function of bounded variation, and denote by G the indefinite integral of F 0 . Then H WD F G has bounded variation, and H 0 D 0 a.e. Functions having this property are called singular. Thus every function F W Œa; b ! R of bounded variation is the difference of an absolutely continuous and a singular function.

13

Lebesgue [295, pp. 232–249].

6.3 Integration by Parts and Change of Variable

207

6.3 Integration by Parts and Change of Variable Proposition 6.8 If f ; g are integrable on Œa; b and F; G are their primitive functions, then fG and Fg are also integrable on Œa; b, and Z

Z

b

b

fG dx C a

a

Fg dx D F.b/G.b/ F.a/G.a/ DW ŒFGba :

Proof F and G are continuous functions on a compact interval, hence they are bounded by some constant M. It follows by applying Proposition 5.16 (b) and (e) (p. 187) that fG and Fg are integrable. Furthermore, using for the subintervals Œ˛; ˇ of Œa; b the estimates jF.ˇ/G.ˇ/ F.˛/G.˛/j D j.F.ˇ/ F.˛//G.ˇ/ F.˛/.G.ˇ/ G.˛//j MjF.ˇ/ F.˛/j C MjG.ˇ/ G.˛/j; we conclude that FG is absolutely continuous and has bounded variation. Since .FG/0 D F 0 G C FG0 D fG C Fg a.e., applying Theorem 6.5 (p. 204) we conclude that Z

b

Z

b

fG dx C

a

Z

b

Fg dx D

a

a

fG C Fg dx D ŒFGba : t u

Proposition 6.9 (de la Vallée-Poussin)14 Let x W Œ˛; ˇ ! R be an absolutely continuous, non-decreasing function. If f is integrable in Œx.˛/; x.ˇ/, then . f ı x/x0 is integrable in Œ˛; ˇ, and Z

x.ˇ/

Z f .x/ dx D

x.˛/

ˇ ˛

f .x.t//x0 .t/ dt:

(6.5)

Proof The statement is obvious if f is a step function. Since the general case may be reduced to the case of C1 functions by using the decomposition f D g h with g; h 2 C1 , it suffices to prove the proposition when f 2 C1 . Let f 2 C1 , and choose a non-decreasing sequence .'n / of step functions, converging a.e. to f . Set E WD fx 2 Œx.˛/; x.ˇ/ W 'n .x/ 6! f .x/g

14

de la Vallée-Poussin [465, p. 467].

(6.6)

208

6 * Generalized Newton–Leibniz Formula

and ˚ D WD t 2 Œ˛; ˇ W x.t/ 2 E

and x0 .t/ > 0 :

By assumption E is a null set. Assume temporarily that D is also a null set. Since x0 0, the sequence of measurable functions t 7! 'n .x.t//x0 .t/;

n D 1; 2; : : :

is non-decreasing. Furthermore, we have 'n .x.t//x0 .t/ ! f .x.t//x0 .t/ a.e. in Œ˛; ˇ because the exceptional points belong either to D or to the nondifferentiability set of x, both null sets. Finally, the corresponding integrals are uniformly bounded because using (6.5) for step functions we have Z

ˇ ˛

'n .x.t//x0 .t/ dt D

Z

Z

x.ˇ/

x.ˇ/

'n .x/ dx ! x.˛/

f .x/ dx: x.˛/

Applying the Beppo Levi theorem we conclude that . f ı x/x0 is integrable, and f satisfies (6.5). It remains to prove that D is a null set in Œ˛; ˇ. For this we consider a system fIk g of open intervals, of finite total length, covering each point of E infinitely many times. Then n X

Ik .x.t//x0 .t/;

n D 1; 2; : : :

kD1

is a non-decreasing sequence of functions whose integrals are uniformly bounded because using (6.5) for step functions we have Z 0

˛

ˇ

n X kD1

0

Ik .x.t//x .t/ dt D

n Z X kD1

x.ˇ/

Ik .x/ dx x.˛/

1 X

jIk j < 1:

kD1

The series converges a.e. by the Beppo Levi theorem. Since it tends to infinity for each t 2 D, D is a null set. u t Remark The formula (6.5) remains valid if f has an infinite integral. Considering the positive and negative parts of f , it suffices to study the case of nonnegative, measurable functions f . Choose a non-decreasing sequence .'n / of integrable functions, converging a.e. to f . Then we may repeat part (c) of the preceding proof by applying now the generalized Beppo Levi theorem, i.e., Proposition 5.17 (e) (p. 190).

6.4 Exercises

209

6.4 Exercises Exercise 6.1 Consider the Cantor function F W Œ0; 1 ! Œ0; 1, and set f .x/ WD x C F.x/, x 2 Œ0; 1. Prove the following15: (i) f is a homeomorphism between the intervals Œ0; 1 and Œ0; 2; (ii) f sends the null set C into a set of measure one; (iii) there exists a subset of C whose image by f is non-measurable. Exercise 6.2 (i) For each ˛ 2 Œ0; 1/ there exists a perfect nowhere dense set C˛ Œ0; 1 of measure ˛.16 (ii) Construct a set A Œ0; 1 of measure one and of the first category.17 (iii) Construct a null set B Œ0; 1 of the second category.18 Exercise 6.3 If f W Œa; b ! R is continuous, then f and jf j are absolutely continuous at the same time. Is the continuity assumption necessary? Exercise 6.4 Given an integrable function f W Œa; b ! R, x 2 .a; b/ is a Lebesgue point if 1 lim h!0 2h

Z

xCh

f .t/ dt D f .x/: xh

(i) If f is continuous at x, then x is a Lebesgue point. (ii) If f has different finite left and right limits at x, then x is not a Lebesgue point. (iii) Almost every x is a Lebesgue point.

15

See Gelbaum–Olmsted [167, 168] for other interesting properties. A perfect set is a closed set with no isolated points. A set is nowhere dense if its closure has no interior points. 17 A set A is of the first category (Baire [17]) if it is the countable union of nowhere dense sets. 18 A set A is of the second category (Baire [17]) if it is not of the first category. Baire’s theorem (see p. 32) states that every complete metric space and every compact Hausdorff space is of the second category. 16

Chapter 7

Integrals on Measure Spaces

In my opinion, a mathematician, in so far as he is a mathematician, need not preoccupy himself with philosophy – an opinion, moreover, which has been expressed by many philosophers. –H. Lebesgue

In Chap. 5 we defined the Lebesgue integral of functions defined on R. In this chapter we show that the theory remains valid in a much more general framework;1 moreover, almost all proofs can be repeated word for word. The results of this chapter include integrals of several variables and integrals on probability spaces.2

7.1 Measures In this section we generalize the notions of length, area and volume. We recall that by a disjoint set sequence we mean a sequence .An / of pairwise disjoint sets. To emphasize the disjointness we sometimes write [ An instead of [An . We denote by 2X the set of all subsets of a set X. The notation is motivated by the fact that if X has n elements, then 2X has 2n elements.

1 Radon [366], Fréchet [158], Daniell [93]. In this book we consider only real-valued functions, although Bochner [46] extended the theory to Banach space-valued functions, and this has important applications among others in the theory of partial differential equations. See, e.g., Dunford–Schwartz [117], Edwards [119], Yosida [488], and Lions–Magenes [305]. 2 Kolmogorov [252].

© Springer-Verlag London 2016 V. Komornik, Lectures on Functional Analysis and the Lebesgue Integral, Universitext, DOI 10.1007/978-1-4471-6811-9_7

211

212

7 Integrals on Measure Spaces

Definition By a semiring3 in a set X we mean a set system P 2X satisfying the following conditions: • ¿ 2 P; • if A; B 2 P, then A \ B 2 P; • if A; B 2 P, then there exists a finite disjoint sequence C1 ; : : : ; Cn in P such that A n B D C1 [ [ Cn : Remark It follows by induction on k that A1 \ \ Ak 2 P for every finite sequence A1 ; : : : ; Ak in P. Examples • Every -ring is a semiring. • The intervals of R form a semiring. The bounded intervals also form a semiring. • For any given set X and nonnegative integer k, the subsets of at most k elements of X form a semiring. • (Restriction) If P is a semiring in X, and Y X, then PY WD fP 2 P W P Yg is a semiring in Y. • (Direct product) If P is a semiring in X and Q is a semiring in Y, then P Q WD fP Q W P 2 P; Q 2 Qg is a semiring in X Y. Definitions By a measure4 on X we mean a nonnegative set function W P ! R, defined on a semiring P in X, satisfying .¿/ D 0, which is -additive in the following sense: if .An / P is a disjoint set sequence and A WD [ An 2 P, then5 .A/ D

X

.An /:

(7.1)

In this case the triplet .X; P; / is called a measure space.

3 Halmos [184] introduced a slightly more restricted notion, but the present definition has become standard by now. 4 Borel [59]. 5 Since .¿/ D 0, the equality (7.1) holds for finite disjoint sequences as well. Finitely additive set functions were studied before Borel by Harnack [192], Cantor [74, pp. 229–236], Stolz [437], Peano [353, pp. 154–158] and Jordan [231, pp. 76–79].

7.1 Measures

213

Examples • The length of bounded intervals is a measure on R: if a bounded interval I is P the union of a disjoint interval sequence .Ik /, then jIj D jIk j.6 Indeed, an elementary argument shows Pthat jI1 j C C jIn j jIj for every n; letting n ! 1 this yields the inequality jIk j jIj. The reverse inequality has been proved earlier in Proposition 4.3 (p. 155). • (Counting measure) Denoting by .A/ the number of elements of a set A X we get a measure on P WD 2X .7 • (Dirac measure) For any fixed point x 2 X the formula ( ıx .A/ WD

1 if x 2 A, 0 if x … A

defines a measure on P WD 2X . • (Zero measure) The formula .A/ WD 0 defines a measure on P WD 2X . • (Largest measure) The formula ( .A/ WD

0

if A D ¿,

1 otherwise

defines a measure on P WD 2X . • (Zero-one measure) Given an uncountable set X, the formula ( .A/ WD

0

if A is countable,

1

if X n A is countable,

defines a measure on the -ring formed by the countable subsets of X and their complements. • (Restriction) If is a measure on a semiring P and Y 2 P, then the restriction of to PY is a measure. • (Direct product) If W P ! R and W Q ! R are two measures, then the formula . /.P Q/ WD .P/ .Q/ defines a measure on P Q.

6

The statement and its proof remain valid for unbounded intervals, too. In this book we do not distinguish between different infinite cardinalities, except in an example on p. 243 and in some exercises. 7

214

7 Integrals on Measure Spaces

• (Finite part of a measure) For any given measure % W R ! R, P WD fA 2 R W %.A/ < 1g is a semiring, and the restriction of % to P is a measure. Now we prove that every measure may be extended uniquely to a measure defined on a set system which is easier to manipulate. This will enable us to establish various important features of the measures. Definition By a ring in a set X we mean a set system R 2X satisfying the following conditions: • ¿ 2 R; • if A; B 2 R, then A n B 2 R; • if A; B 2 R are disjoint sets, then A [ B 2 R. Remark If R is a ring, then the identity A [ B D .A n B/ [ B shows that the disjointness is not necessary in the last condition: if A; B 2 R, then A [ B 2 R. It follows by induction that A WD A1 [ [An 2 R for every finite sequence A1 ; : : : ; Ak in R. Using the identity \An D A n [.A n An / it follows that A1 \ \ Ak 2 R for every finite sequence A1 ; : : : ; Ak in R. In particular, every ring is also a semiring. Examples • Every -ring is also a ring. In particular, 2X is a ring in X. • The finite subsets of a set X form a ring in X. • The finite subsets of a set X and their complements8 form a ring in X. Given any set system A 2X , the intersection of all rings R satisfying A R 2X is a ring in X. It is called the ring generated by A. There is a simple construction of the rings generated by semirings: Lemma 7.1 The ring generated by a semiring P is formed by all finite disjoint unions of the form R D P1 [ [ Pn ;

P1 ; : : : ; Pn 2 P

n D 1; 2; : : : :

(7.2)

Proof Since every ring containing P contains the sets (7.2), it is sufficient to show that the system R of these sets is already a ring. We proceed in several steps. (a) We have ¿ 2 R because ¿ 2 P. (b) If R1 ; : : : ; Rm 2 R are pairwise disjoint sets for some positive integer m, then R WD R1 [ [ Rm 2 R. Indeed, if we decompose each Ri similarly to (7.2), then we obtain a decomposition of the same form of R.

8

The so-called co-finite sets.

7.1 Measures

215

(c) If P0 ; P 2 P, then P0 n P 2 R by the definition of the semiring and of R. (d) If R 2 R and P 2 P, then R n P 2 R. Indeed, considering a decomposition of the form (7.2) of R and using (b) and (c) we obtain that R n P D .P1 n P/ [ [ .Pn n P/ 2 R: (e) If R0 ; R 2 R, then R0 n R 2 R. Indeed, considering a decomposition of the form (7.2) of R and applying (d) n times we obtain that R0 n R D .: : : .R0 n P1 / n P2 / : : : / n Pn 2 R:

t u

Proposition 7.2 Every measure W P ! R may be extended to a unique measure defined on the ring R generated by the semiring P. Proof If there exists such an extension, then, still denoting it by , we must have .R/ D .P1 / C C .Pn / for every decomposition of the form (7.2). Since P1 ; : : : ; Pn 2 P, this proves the uniqueness. For the existence first we show that the above equality does indeed define an extension, i.e., if R D P01 [ [ P0m is another such decomposition of R, then .P1 / C C .Pn / D .P01 / C C .P0k /: This readily follows from the additivity of W P ! R because both sums are equal to n X k X

.Pj \ P0i /:

jD1 iD1

The extended set function is clearly nonnegative, it remains to prove its additivity. LetPR D [1 kD1 Rk be a disjoint union with R; Rk 2 R; we have to show that .R/ D .Rk /. Replacing each Rk with a decomposition of the form (7.2) and using the definition of .Rk / we may assume that Rk 2 P for every k. Now consider a decomposition of the form (7.2) of R; then we have Pj D

1 [ kD1

.Pj \ Rk /

216

7 Integrals on Measure Spaces

for each j. Since Pj ; Pj \ Rk 2 P, and since is -additive on P, this implies .Pj / D

1 X

.Pj \ Rk /:

kD1

Summing these equalities we obtain the required relation: .R/ D

n X

.Pj / D

jD1

n X 1 X

.Pj \ Rk /

jD1 kD1

D

1 X n X kD1

D

1 X

.Pj \ Rk /

jD1

.Rk /:

t u

kD1

Now we are ready to establish some basic properties of measures: Proposition 7.3 Every measure W P ! R (defined on a semiring) satisfies the following conditions: (a) (monotonicity) if A; B 2 P and A B, then .A/ .B/; (b) P (-subadditivity) if .An / P is a countable cover of A 2 P, then .A/ .An /; (c) (continuity) if .An / P is a non-decreasing set sequence and A WD [An 2 P, then .An / ! .A/; (d) (continuity) if .An / P is a non-increasing set sequence with .A1 / < 1 and A WD \An 2 P, then .An / ! .A/. Example The intervals An WD Œn; 1/ R show that the condition .A1 / < 1 in (d) cannot be omitted. Proof By the preceding proposition we may assume that P is a ring. (a) Using the nonnegativity and the additivity of the measures we have .B/ D .A/ C .B n A/ .A/: (b) Setting B1 WD A \ A1 and BnC1 WD .A \ AnC1 / n .A1 [ [ An /; n D 1; 2; : : :

7.2 Integrals Associated with a Finite Measure

217

we have A D [ Bn . Furthermore, Bn An and Bn 2 P for all n (because P is a ring). We conclude by using (a): .A/ D

X

.Bn /

X

.An /:

(c) Let A0 D ¿, then the sets Ak n Ak1 belong to the ring P. Since AD

1 [

.Ak n Ak1 / and An D

kD1

n [

.Ak n Ak1 /

kD1

for all n, we have .An / D

n X

.Ak n Ak1 / !

kD1

1 X

.Ak n Ak1 / D .A/:

kD1

(d) Since .An / is finite for all n, changing An to An nA we may assume that A D ¿. The sets Ak n AkC1 belong to the ring P. Since A1 D

1 [

.Ak n AkC1 /;

kD1

by the -additivity we have 1 X

.Ak n AkC1 / D .A1 /:

kD1

Since .A1 / < 1 by assumption, the series is convergent, and hence 1 X

.Ak n AkC1 / ! 0

kDn

as n ! 1. We conclude by noticing that the last sum is equal to .An / because 1 [

.Ak n AkC1 / D An :

kDn

7.2 Integrals Associated with a Finite Measure Definition A measure is finite if it takes only finite values.

t u

218

7 Integrals on Measure Spaces

Examples • The finite part of a measure is a finite measure. • Every bounded measure is finite. The length of bounded intervals shows that the converse is not always true. For the rest of this section we fix a semiring P in a set X and a finite measure W P ! R. Definition By a step function we mean a linear combination 'D

n X

ck Pk

kD1

of characteristic functions of sets in P. The integral of a step function is defined by the formula Z ' d WD

n X

ck .Pk /:

kD1

Proposition 5.1 (p. 172) remains valid: by the additivity of the measure the integral does not depend on the particular representation of the step function. Definition A set A is a null P set if for each " > 0 there exists a sequence .Pk / P satisfying A [Pk and .Pk / ". P Equivalently, A is a null set if there exists a sequence .Pk / P satisfying .Pk / < 1, and covering each point x 2 A infinitely many times. Proposition 4.3 (p. 155) takes the following form: Proposition 7.4 (a) (b) (c) (d)

The empty set is a null set. The subsets of a null set are null sets. The union of countably many null sets is a null set. P 2 P is a null set ” .P/ D 0.

Proof (a), (b) and (c) We may repeat the proof of Proposition 4.3. (d) If .P/ D 0, then P is null set: we may choose Pk D P for all k in the definition. Conversely, if P 2 P is a null set, P then for each " > 0 there exists a sequence .Pk / P satisfying A [Pk and .Pk / ". Using the subadditivity of the measure this implies .P/ " for every " > 0, and hence .P/ D 0. t u Chapter 5 was written in such a way that all theorems, propositions, corollaries and lemmas remain valid without any change. Moreover, the proofs also remain valid with three exceptions: • In the proof of Lemma 5.2 (p. 172) we have used the topological properties of intervals.

7.2 Integrals Associated with a Finite Measure

219

• In the proof of Proposition 5.16 (f) (p. 187) we have used the existence of an integrable, everywhere positive function. An example following Lemma 7.5 will show that such functions do not exist for all measures. • In the proof of Proposition 5.19 (p. 194) we have implicitly used that the constant functions are measurable.9 The just mentioned example will show that this is not always true either.10 The following alternative proofs are always valid: Proof of Lemma 5.2 We extend to the generated ring R by Proposition 7.2. Fix a null set Y X such that 'n .x/ ! 0 for every x 2 X n Y, and fix " > 0 arbitrarily. Choose a set sequence .Hi / P satisfying Y [Hi

and

X

.Hi / < ":

Then the sets Sn WD H1 [ [ Hn belong to R, S1 S2 ; and .Sn / < " for every n. Set K0 WD fx 2 X W '1 .x/ > 0g and Kn WD fx 2 X W 'n .x/ > "g ;

n D 1; 2; : : : I

they belong to R, and K0 K1 K2 : Setting M WD max '1 we have 'n M

9

on Kn ;

'n "

on K0 n Kn ;

'n D 0

on X n K0 :

The measurability of the constant functions is equivalent to the measurability of X. In this book, following F. Riesz, we adopt a more restrictive measurability notion than usual. See Sect. 7.7 on the advantages of this choice.

10

220

7 Integrals on Measure Spaces

Consequently, Z 0

'n d ".K0 n Kn / C M.Kn / D ".K0 n Kn / C M.Kn \ Sn / C M.Kn n Sn / ".K0 / C M" C M.Kn n Sn /:

The set sequence .Kn n Sn / is non-increasing and .K1 n S1 / .K0 / < 1: Furthermore, its intersection is empty. Indeed, if x 2 \Kn , then 'n .x/ 6! 0, so that x 2 Y; but then x 2 Sn for a sufficiently large n and therefore x 2 Sn and x … Kn n Sn . Applying Proposition 7.3 (d) we conclude that .Kn n Sn / ! 0. Consequently, we infer from the previous estimate that Z 0

'n d < .K0 / C M C 1 "

if n is sufficiently large.

t u

Proof of Proposition 5.16 (f) If there exists a set sequence .Pk / P such that each fn vanishes outside [Pk then we may repeat the proof of Chap. 5 by using the function X Pk h WD ; 2 k .1 C .Pk // k and defining the functions gn and g by zero outside [Pk . The existence of such a sequence .Pk / follows from the next lemma.11

t u

Lemma 7.5 To each measurable function f there exists a disjoint set sequence .Pk / P such that f D 0 outside [ Pk .12 Proof Choose a sequence .'n / of step functions converging to f a.e. By definition there exists a set sequence .A0j / P such that 'n ! f outside [A0j . Furthermore, by the definition of step functions there exists for each n a finite set sequence .Anj / P such that 'n D 0 outside [Anj . We may arrange all these sets A0j and Anj into a set sequence .Pk /. Furthermore, using the definition of a semiring we may replace each difference P2 n P1 , .P3 n

11

We apply the lemma for each fn , and we take the union of the corresponding set sequences. We sometimes express this property by saying that f has a -finite support. Using this terminology X is measurable ” X is -finite.

12

7.2 Integrals Associated with a Finite Measure

221

P2 / n P1 , . . . by a finite disjoint union of sets in P. Then the sequence .Pk / becomes disjoint, and f D 0 outside [ Pk . t u *Examples • Let be a finite measure on the ring of finite subsets of an uncountable set X. By Lemma 7.5 there is no measurable, strictly positive function for this measure. In particular, the non-zero constant functions are non-measurable. • Fix a non-empty set X and consider the measure .¿/ WD 0 on the ring P WD f¿g. Then only the zero function is measurable, and ¿ is the only measurable set. Proof of Proposition 5.19 Most of the former proof remains valid. The only property to check is that if f is a measurable function and c a positive constant, then the functions min f f ; cg and max f f ; cg are measurable. For the proof we consider the sets Pk of the preceding lemma. Then A WD [Pk is measurable, hence the functions c A and then the functions min f f ; cg D min f f ; c A g

and

max f f ; cg D max f f ; c A g t u

are also measurable.

Starting from an arbitrary finite measure defined on a semiring P, the theory of Chap. 5 leads to a measure defined on the system M of all measurable sets. Our next result states that this is the only possible extension of the original measure to M. *Proposition 7.6 Let W N ! R be another measure, defined on a semiring satisfying P N M. If D on P, then D on N , too. Proof (i) Every -null set is also a -null set. For, if a set may be covered by a set sequence .Pn / P of total -measure < ", then we have X

.Pn / D

X

.Pn / < ":

(ii) Now consider the two integrals associated with the measures and jP . We show that every -integrable function f is also -integrable, and the two integrals are equal: Z

Z f d D

f d :

(7.3)

Since Z

Z P d D .P/ D .P/ D

P d

222

7 Integrals on Measure Spaces

for every P 2 P by assumption, taking their linear combinations we obtain that (7.3) holds for all step functions. The equality holds for all functions f 2 C1 ./ as well.13 Indeed, consider a non-decreasing sequence .'n / of -step functions, converging -a.e. to f , and satisfying Z sup n

'n d < 1:

R R Then we have 'n d ! f d by definition. Furthermore, .'n / converges to f also -a.e. by (i), and Z sup n

Z 'n d D sup n

'n d < 1

because (7.3) has already been proved for step functions.RApplying theRBeppo Levi theorem we conclude that f is -integrable and 'n d ! f d ; hence (7.3) holds for f . Finally, if f is an arbitrary -integrable function, then we have f D g h with suitable functions g; h 2 C1 ./. We already know that (7.3) holds for g and h; taking the difference of these equalities we see that f satisfies (7.3) as well. (iii) It follows from (ii) that if A 2 N and .A/ < 1, then Z .A/ D

Z A d D

A d D .A/:

Consider finally an arbitrary set A 2 N . Then A 2 M, hence it is measurable, so that it may be covered by a disjoint sequence .Pn / P. Since P N , we have A \ Pn 2 N M

and .A \ Pn / < 1

for all n. Applying the preceding equality for A \ Pn instead of A we conclude that X X .A/ D .A \ Pn / D

.A \ Pn / D .A/: t u We end this section by characterizing the measures constructed via integrals. Definition A measure , defined on a semiring Q, is -finite if each set in Q has a countable cover by sets of finite measure.

13

The function class C1 was defined on p. 174.

7.2 Integrals Associated with a Finite Measure

223

Remark By the definition of a semiring in the -finite case each set in Q also has a countable disjoint partition by sets of finite measure. Examples • The usual Lebesgue measure in R is -finite. • Every finite measure is -finite. • Given a measure % on some semiring R, let us denote by Q the sets A 2 R having a countable cover by sets P 2 R of finite measure. Then Q is also a semiring. The restriction of % to Q is called the -finite part of %. • The counting measure on an uncountable set X is not -finite. Its -finite part is defined on the countable subsets of X. Consider again a finite measure defined on a semiring P, and let be its extension14 to the set system M of measurable sets. *Proposition 7.7 (a) M is a -ring. The extended measure W M ! R is -finite and complete. (b) Conversely, every -finite, complete measure, defined on a -ring may be obtained in this way. (c) More generally, every -finite measure, defined on a semiring, is a restriction of the measure W M ! R obtained by the extension of its finite part. Proof (a) This follows from Proposition 5.18 (p. 192) and Lemma 7.5. (c) Let N be a semiring and W N ! R a -finite measure. Consider the finite part of , i.e., the restriction of to the semiring P WD fA 2 N W .P/ < 1g ; and let W M ! R be the extension of jP to the -ring of jP -measurable sets. We have to show that N M and D jN . Fix an arbitrary set A 2 N . Since is -finite, there exists a disjoint set sequence .Pn / P satisfying A D [ Pn . Since P M and M is a -ring, A 2 M. Furthermore, since .Pn / D .Pn / for every n by the definition of , we conclude that X X

.A/ D

.Pn / D .Pn / D .A/: (b) Let N be a -ring and W N ! R a -finite, complete measure. By (c) we already know that N M and D jN . It remains to prove that M N . Fix an arbitrary A 2 M. Then A is a measurable function, so that there exists a sequence .'n / of P-step functions, converging to A -a.e. In other

14

We already know that this extension is unique.

224

7 Integrals on Measure Spaces

words, there exists a -null set P0 such that 'n ! A outside it. Observe that P0 is also a -null set.15 Then

1 2N An WD x 2 X W 'n .x/ > 2 for each n D 1; 2; : : : ; because An is a union of finitely many elements of P, P N , and N is a ring. Since N is also a -ring, the set 1 N WD lim sup An WD \1 kD1 [nDk Ak

also belongs to N . Now observe for each x 2 X n P0 the equivalences x 2 A ” x 2 An for infinitely many n ” x 2 lim sup An : It follows that .A n N/ [ .N n A/ P0 ; i.e., A differs from N 2 N on a -null set. Since is complete, we conclude that A 2 N . t u

7.3 Product Spaces: Theorems of Fubini and Tonelli In classical analysis the computation of double integrals may be reduced to that of simple integrals by using the formula16 Z f .x; y/ dx dy D XY

D

Z Z

f .x; y/ dy dx

X

Y

Y

X

Z Z

(7.4)

f .x; y/ dx dy:

In this section we prove that this formula remains valid for Lebesgue integrals as well. Consider two finite measures W P ! R and W Q ! R, where P is a semiring in X and Q is a semiring in Y. Then W P Q ! R is a finite measure on the

15 16

See the beginning of the proof of Proposition 7.6: we already know that P N M. Euler [130], Dirichlet [113], and Stolz [438, pp. 93–94].

7.3 Product Spaces: Theorems of Fubini and Tonelli

225

semiring P Q in X Y. In what follows we write Z

Z

Z

f .x; y/ dx dy; XY

g.x/ dx and

h.y/ dy

X

Y

instead of Z

Z f d. /;

Z g d and

h d :

The expressions null set and a.e. will refer to in X, to in Y, and to in X Y. The following theorem is a far-reaching generalization of the classical results: Theorem 7.8 (Fubini)17 If f is integrable in X Y, then the successive integrals in (7.4) exist, and the three expressions are equal. Remarks • By induction the theorem may be extended to arbitrary finite direct products of (finite) measures. • The existence of the successive integrals does not imply their equality. Moreover, their existence and equality does not imply the integrability of f . See the examples at the end of this section. We prepare the proof by clarifying the relationship among the null sets of the three spaces: Lemma 7.9 If E is a null set in X Y, then the “vertical sections” fy 2 Y W .x; y/ 2 Eg of E are null sets in Y for almost every x 2 X.

17

Lebesgue [288] (for bounded functions), Fubini [164]. Fubini’s proof was incorrect; the first correct proofs were given by Hobson [214] and de la Vallée-Poussin [464]. See Hawkins [198].

226

7 Integrals on Measure Spaces

Proof Fix a sequence of “rectangles” Rn D Pn Qn in P Q, covering each point of E infinitely many times, and satisfying 1 X . /.Rn / < 1: nD1

By the definition of the integral of step functions we have Z . /.Rn / D

Rn .x; y/ dx dy D

Z Z

XY

X

Rn .x; y/ dy dx Y

(their common value is .Pn / .Qn /), so that the series 1 Z Z X nD1

X

Rn .x; y/ dy dx Y

is convergent. Applying the Beppo Levi theorem we obtain that the series 1 Z X nD1

Rn .x; y/ dy Y

is convergent for a.e. x 2 X. If x0 is such a point, then another application of the Beppo Levi theorem implies that the series 1 X

Rn .x0 ; y/

nD1

is convergent for a.e.Py 2 Y. If y0 is such a point, then .x0 ; y0 / … E, because at the points of E we have Rn D 1. t u Proof of Theorem 7.8 By symmetry we prove only the equality Z f .x; y/ dx dy D XY

Z Z X

f .x; y/ dy dx:

(7.5)

Y

We have to show that R • the integral Y f .x;Ry/ dy is well defined for a.e. x 2 X; • the function x 7! Y f .x; y/ dy is integrable in X; • the two sides of (7.5) are equal. We have seen during the proof of the preceding lemma that these properties hold true if f is the characteristic function of a “rectangle”. Taking linear combinations we see that they hold for step functions as well. Since every integrable function is

7.3 Product Spaces: Theorems of Fubini and Tonelli

227

the difference of two step functions, it remains only to establish the three properties for functions belonging to the class C1 . Fix f 2 C1 arbitrarily. Choose a non-decreasing sequence .'n / of step functions and a null set E X Y such that 'n .x; y/ % f .x; y/

for each .x; y/ 2 .X Y/ n E;

(7.6)

and therefore Z

Z 'n .x; y/ dx dy !

f .x; y/ dx dy

XY

XY

by the definition of the integral. Since (7.5) is already known for step functions, the last relation may be rewritten in the form Z Z X

Z

'n .x; y/ dy dx !

f .x; y/ dx dy:

Y

(7.7)

XY

Applying the Beppo Levi theorem18 we obtain that the non-decreasing sequence of the functions Z x 7! 'n .x; y/ dy (7.8) Y

converges, and hence is bounded, for a.e. x 2 X. Fix a point x 2 X where the convergence holds, and for which the section fy 2 Y W .x; y/ 2 Eg is a null set. (By the preceding lemma a.e. x 2 X has this property.) Then 'n .x; y/ % f .x; y/ for a.e. y 2 Y by (7.6), so that, in view of the boundedness of the functions (7.8) we may apply the Beppo Levi theorem again: the function y 7! f .x; y/ is integrable, and Z

Z 'n .x; y/ dy %

Y

18

f .x; y/ dy: Y

In the proof of this theorem the application of Lemma 5.3 (p. 173) is sufficient because we consider only sequences of step functions.

228

7 Integrals on Measure Spaces

We recall that this convergence holds for a.e. x 2 X. Since the sequence of integrals Z Z X

'n .x; y/ dy dx Y

is bounded by (7.7) and the integrability of f , a third application of the Beppo Levi theorem shows that the function Z x 7! f .x; y/ dy Y

is integrable, and Z Z X

Z Z 'n .x; y/ dy dx ! f dy dx:

Y

X

(7.9)

Y

The equality (7.5) follows from (7.7) and (7.9).

t u

Fubini’s theorem remains valid for generalized (infinite-valued) integrals: Theorem 7.10 (Tonelli)19 If the integral of a function f exists in X Y, then the successive integrals in (7.4) also exist, and the three quantities are equal. Remarks • Like that of Fubini, Tonelli’s theorem holds for arbitrary finite direct products of measures as well. • We recall that every nonnegative, measurable function has an integral. Proof Considering the positive and negative parts of f , at least one of them is integrable, hence satisfies the assumptions of Fubini’s theorem. Therefore it is sufficient to investigate the case of a nonnegative, measurable function f . Applying Lemma 7.5 we fix a non-decreasing sequence .An / of sets of finite measure such that f D 0 outside [An . Then the functions 'n WD An min f f ; ng are integrable in X Y by Proposition 5.16 (e) (p. 187), and 'n % f a.e. by construction. We may therefore choose a null set E in X Y such that 'n .x; y/ % f .x; y/

19

Tonelli [457].

for each .x; y/ 2 .X Y/ n E:

7.4 Signed Measures: Hahn and Jordan Decompositions

229

Let us observe that formally this relation is identical with (7.6). We may therefore repeat the preceding proof with two small changes: • instead of the Beppo Levi theorem (or Lemma 5.3) we apply the generalized Beppo Levi theorem, i.e., Proposition 5.17 (e) (p. 190); • the validity of the equality for 'n (instead of f ) now follows from Fubini’s theorem. t u Examples The following examples show the optimality of the assumptions of Theorems 7.8 and 7.10.20 • The formula

f .x; y/ WD

8 ˆ ˆ 0g and f f < 0g, and at least one of the two measures is bounded.

22

See Gurevich–Silov [175, p. 180].

7.4 Signed Measures: Hahn and Jordan Decompositions

231

Fig. 7.1 Hahn decomposition

P

N

These properties remain valid for all signed measures, defined on -rings. Thanks to the following theorem many questions about signed measures may be reduced to the study of measures. Theorem 7.12 Let be a signed measure on a -ring M. (a) (Hahn decomposition)23 There exists a decomposition X D P [ N such that A \ P; A \ N 2 M, .A \ P/ 0 and .A \ N/ 0 for every A 2 M. (See Fig. 7.1.) (b) (Jordan decomposition)24 There exist two measures C ; on M, satisfying the equality D C , concentrated on disjoint sets, and such that at least one of them is bounded. Remarks • If D C is a Jordan decomposition, then at least one of the measures C and is bounded. For otherwise there would exist two disjoint sets A; B with C .A/ D .B/ D 1 and .A/ D C .B/ D 0, and then C .A [B/ .A [ B/ would not be defined.

23 24

Hahn [180, p. 404]. Jordan [229]. The decomposition is clearly unique.

232

7 Integrals on Measure Spaces

• The assumption that M is a -ring is important: for example, the signed measure of Smolyanov has no Hahn decomposition. Indeed, for such a decomposition we should have N D ¿,25 and then could not take negative values. • Smolyanov’s signed measure does not have a Jordan decomposition either. Indeed, if there were two measures C ; such that D C , then we would have C .X/ C .A/ .A/ D jAj and .X/ .X n A/ .X n A/ D jAj for each finite set A. This would imply C .X/ D .X/ D 1 and then C .X/ .X/ would not be defined. • The preceding remarks imply that Smolyanov’s finite signed measure cannot be extended to a signed measure defined on a -ring.26 This contrasts with the case of finite measures. The following lemma prepares the proof of the theorem. Definition Let be a signed measure on M. A set A 2 M is called negative if .B/ 0 for every subset of A, belonging to M. Lemma 7.13 Let W M ! R be a signed measure on a -ring M. (a) If A; B M and B A, then .A/ < 1 H) .B/ < 1

and .A/ > 1 H) .B/ > 1:

(b) If is finite, then it is bounded. (c) is bounded from below or from above. (d) If A 2 M and .A/ < 0, then there exists a negative subset A0 of A such that .A0 / .A/. We will often use property (b) in the sequel. Proof (a) This follows from the equality .A/ D .B/ C .A n B/ because the sum is defined by definition.

25 For otherwise we would have for every one-point set A N the impossible inequalities 1 D .A/ D .A \ N/ 0. 26 This also follows from Lemma 7.13 (c) below.

7.4 Signed Measures: Hahn and Jordan Decompositions

233

(b) If, for example, sup D 1, then we may define recursively a set sequence .An / satisfying .An / > 1 C

X

.Ak /;

n D 1; 2; : : : :

k 1 either because .C/ D 1. (d) If A is a negative set, then we may take A0 WD A. Otherwise let k1 be the smallest positive integer for which A has a subset A1 satisfying .A1 / 1=k1 . We have .A/ D .A1 / C .A n A1 /; whence .A n A1 / .A/.27 If A n A1 is a negative set, then we may take A0 WD A n A1 . Otherwise let k2 be the smallest positive integer for which A n A1 has a subset A2 satisfying .A2 / 1=k2 . Continuing we obtain either a suitable negative set of the form A0 WD A n .A1 [ [ An / after a finite number of steps, or an infinite disjoint sequence .An / M, satisfying .An / 1=kn for all n with suitable positive integers kn . In the latter case we have X 1 X .An / D .[ An / < 1I kn the last inequality follows by applying (a) with B WD [ An A. It follows that kn ! 1. Set A0 WD A n [ An , then A0 2 M and .A/ D .A0 / C .[ An /: Consequently, .A0 / .A/. It remains to show that B 2 M and B A0 imply .B/ 0. Since kn ! 1, we have kn 2 and (by construction) .B/ < 1=.kn 1/ for all sufficiently large n. Letting n ! 1 we conclude that .B/ 0. t u

27

We may have equality if .A/ D 1.

234

7 Integrals on Measure Spaces

Proof of Theorem 7.12 (a) By Lemma 7.13 (c) we may assume for example that does not take the value 1. Set a D inf .A/; where A runs over the negative sets in M; since ¿ is a negative set, a 0. Let .An / be a sequence of negative sets satisfying .An / ! a. Then N WD [An 2 M is also a negative set, and .N/ D a. Since does not take the value 1, this implies that a > 1, i.e., a is finite. Let P D X n N, then X D P [ N. Let A 2 M. Since N 2 M, we have A\N 2M

and A \ P D A n .A \ N/ 2 M;

and .A \ N/ 0 because N is negative. It remains to prove that .A \ P/ 0. Assume on the contrary that .A \ P/ < 0. Applying the preceding lemma, A \ P has a negative subset A0 satisfying .A0 / .A \ P/. But then N [ A0 is also negative, and the inequality .N [ A0 / D .N/ C .A0 / D a C .A0 / < a contradicts the definition of a. (b) Assume again that does not take the value 1, and consider the Hahn decomposition X D P [ N with N 2 M, obtained in (a). The formulas C .A/ WD .A \ P/ and .A/ WD .A \ N/

(7.10)

define two measures satisfying D C , and concentrated on the disjoint sets P and N. The measure is bounded because .A/ D .A \ N/ .N/ D a < 1 for all A 2 M.

t u

Remarks • We stress that at least one of the two sets of the Hahn decomposition X D P [ N belongs to M. • It follows from the formulas (7.10) that C .A/ WD max f.B/ W B 2 M; B Ag

7.5 Lebesgue Decomposition

235

and .A/ WD min f.B/ W B 2 M; B Ag : This alternative definition of the Jordan decomposition does not use the Hahn decomposition. Definition The measures C ; are called the positive and negative parts of . The measure jj WD C C is called the total variation measure of .

7.5 Lebesgue Decomposition We have seen at the end of Sect. 6.2 that every function of bounded variation is the sum of an absolutely continuous and a singular function. We generalize this result for measures. Similarly to the Hahn and Jordan decompositions, in this section we consider only measures defined on -rings. Hence the finite and bounded measures are the same. Definitions Let , and be three measures on a -ring N in X. • We say that is absolutely continuous with respect to , and we write , if .A/ D 0 H) .A/ D 0: • We say that and are singular, and we write ? , if there is a partition X D M [ S of X such that A2N A2N

and A S

H)

.A/ D 0;

and A M

H)

.A/ D 0:

Thus and are concentrated on the disjoint sets M and S. In some cases an equivalent "–ı definition holds: *Lemma 7.14 Let be absolutely continuous with respect to .28 If is finite, then for every " > 0 there exists a ı > 0, that .A/ < ı H) .A/ < ":

28

We recall that they are defined on a -ring.

236

7 Integrals on Measure Spaces

Example The indefinite integral29 of the function x 7! 1=x with respect to the usual Lebesgue measure in .0; 1/ shows that the boundedness assumption cannot be omitted in the lemma. Proof Assume on the contrary that there exist " > 0 and a sequence .An / satisfying .An / < 2n and .An / " for every n. Then 1 A WD lim sup An WD \1 mD1 [nDm An

satisfies .A/ D 0 and .A/ ", contradicting the relation . Indeed, the sets Bm WD Am [ AmC1 [ : : : form a non-increasing sequence such that .Bm /

gg and B WD f f < gg has a positive measure. If for example .A/ > 0, then Z Z Z f d g d D . f g/ d > 0 A

and therefore

A

A

Z

Z f d ¤

g d:

A

A

We prove a technical lemma: If ¤ 0, then there exist A 2 M and " > 0 such that .A/ > 0, and ".B/ .B/

for all measurable subsets B A:

For the proof we consider for each n D 1; 2; : : : the Hahn decomposition of the signed measure n1 , and we set P D [Pn ;

N D \Nn :

Since n1 is bounded from above, we have Pn 2 M for every n.36 It remains to find some n with .Pn / > 0 because then we may choose A WD Pn and " WD 1=n. We have .B/ D 0 for every measurable set B N because .B/ < 1, and 0 .B/

36

See the remark on p. 238

1 .B/ n

242

7 Integrals on Measure Spaces

for all n because N Nn . Since ¤ 0, .P/ > 0, and then .P/ > 0 by the absolute continuity of . Finally, since 0 < .P/

X

.Pn /

by -subadditivity, we have .Pn / > 0 for at least one n. Proof of the Existence. Let us denote by F the family of nonnegative, integrable functions f satisfying Z f d .A/ A

for all A 2 M. Since is bounded and 0 2 F , the formula Z ˛ WD sup f d f 2F

defines a finite, nonnegative number. The upper bound is attained. For the proof we choose a maximizing sequence . fn / 2 F satisfying Z fn d ! ˛: Then gn WD max f f1 ; : : : ; fn g 2 F for each n. Indeed, every set A 2 M has a partition A1 [ [ An such that gn D fj on each Aj , and then Z gn d D A

n Z X jD1

fj d Aj

n X

.Aj / D .A/:

jD1

Applying the Beppo Levi theorem, the functions gn converge a.e. to a nonnegative, integrable function f . Applying the Fatou lemma (or again R the Beppo Levi theorem) for the sequences . A gn /, we infer from the inequalities A gn d .A/ R R that f 2 F . Finally, the relations fn gn f and fn d ! ˛ imply the equality f d D ˛. To end the proof we show that the measure Z

0 .A/ WD .A/

A2M

f d; A

vanishes identically. Assume on the contrary that 0 ¤ 0. Then by the above lemma there exist A 2 M and " > 0 satisfying .A/ > 0, and Z ".A \ B/ .A \ B/

f d A\B

7.6 The Radon–Nikodým Theorem

243

for all B 2 M. Since f 2 F implies Z 0 .B n A/

f d; BnA

adding the two equalities we get Z ".A \ B/ .B/

f d; B

i.e., Z f C " A d .B/ B

for all B 2 M. Hence f C " A 2 F . This, however, is impossible because Z

Z f C " A d D

f d C ".A/ D ˛ C ".A/ > ˛:

t u

*Example We show37 that the strong -finiteness assumption cannot be omitted in Theorem 7.16 (p. 240). Let Z D X Y with two uncountable sets X, Y satisfying card X > card Y. A set L Z is called a vertical line if there exists an x 2 X such that both sets Ln.fxg Y/ and .fxg Y/ n L are countable. Similarly, a set L Z is called a horizontal line if there exists a y 2 Y such that both sets L n .X fyg/ and .X fyg/ n L are countable. The countable unions of vertical lines, horizontal lines and points form a -ring M. Denoting by .A/ the number of lines contained in A, we obtain a complete, -finite38 (but not strongly -finite) measure W M ! R, for which the null sets are the countable sets. Denoting by .A/ the number of vertical lines contained in A, we obtain another measure W M ! R, satisfying and hence . We claim that the Radon–Nikodým derivative @ =@ does not exist. Assume on the contrary that there exists a measurable function f W Z ! R satisfying Z

.L/ D

f d

for every line

L:

(7.13)

L

37

See Halmos [184, pp. 131–132]. In this example we use the notion of cardinality of infinite sets, but we need only the simplest results: see, e.g., Halmos [186] or Vilenkin [467, 468]. 38 Because every measurable set is covered by countably many lines.

244

7 Integrals on Measure Spaces

R By the measurability condition f is constant a.e. on each line L, and L f d is equal to this constant. Therefore we infer from (7.13) that f D 1 a.e. on every vertical line, and f D 0 a.e. on every horizontal line. This implies the inequalities card X card fx 2 Z W f .x/ D 1g card Y; contradicting the choice of X and Y. We may further generalize the preceding theorem for unbounded and even signed measures : *Theorem 7.17 If is strongly -finite, then the formula (7.11) establishes a one-to-one correspondence between • the functions f having an integral and • the absolutely continuous signed measures . Remark It is easy to see that

is a measure

”

f

is nonnegative.

Indeed, if f 0, Rthen is a measure because f A 0 for every A 2 M, and therefore .A/ D f RA d 0. Conversely, if f < 0 on some set A of positive measure, then .A/ D f A d < 0, and therefore is not a measure. Proof of Theorem 7.17 It follows again from Proposition 7.11 that if f has an integral, then the indefinite integral is an absolutely continuous signed measure. It remains to prove that each absolutely continuous signed measure is the indefinite integral of a unique measurable function f . Similarly to the preceding proof we may assume that .X/ < 1. Proof of the Uniqueness of f . Let f and g be two different functions whose integrals are defined. We have to find a set A such that .A/ > 0, and either f > g on A or f < g on A. Assume by symmetry that B WD f f > gg is not a null set. Since f > 1 and g < 1 on B, setting Ak WD fx 2 B W f .x/ > k we have [k Ak D B:

and g.x/ < kg

7.6 The Radon–Nikodým Theorem

245

Since 0 < .B/ .X/ < 1, there exists a k such that 0 < .Ak / < 1. Then Z

Z f d k.Ak / > 1

g d k.Ak / < 1:

and

Ak

Ak

Consequently, the integral

R

Ak .f

Z

g/ d exists,39 and hence Z

Z

f d Ak

g d D

.f g/ d > 0:

Ak

Ak

A technical lemma:40 if is an absolutely continuous measure, then there exists a disjoint sequence .Fn / of sets of finite -measure such that for each measurable set A, disjoint from F WD [Fn , we have either .A/ D 0 or .A/ D 1 (or both). For the proof we denote by A the -ring of measurable sets having a countable cover by sets of finite -measure. The upper bound ˛ WD sup f.B/ W B 2 Ag .X/ < 1 is attained on some set F 2 A because if .Bn / A and .Bn / ! ˛, then F WD [Bn 2 A and .Bn / .F/ for all n, i.e., .F/ D ˛. Consider a measurable set A, disjoint from F and satisfying .A/ < 1. Since F [ A 2 A, we have ˛ .F [ A/ D .F/ C .A/ D ˛ C .A/I since ˛ is finite, we conclude that .A/ D 0. Proof of the Existence When is a Measure. Consider the disjoint set sequence .Fn / of the previous step, and set E WD X n [Fn . Apply the already proved result for each Fn , and denote by fn the corresponding Radon–Nikodým derivatives. Setting f WD fn on each Fn and f WD 1 on E we get a nonnegative, measurable function. Each A 2 M is the disjoint union of the sets A \ Fn and A \ E, and Z

.A \ Fn / D

fn d A\Fn

for every n by the choice of fn . It remains to show that Z 1 d:

.A \ E/ D A\E

Indeed, then adding all these equalities we obtain (7.11).

39 40

See Proposition 5.17 (d), p. 190. See Hewitt–Stromberg [207, p. 317].

246

7 Integrals on Measure Spaces

R If .A \ E/ D 1, then .A \ E/ > 0 by the absolute continuity of , and hence A\E 1Rd D 1. Otherwise we have .A \ E/ D 0 by the definition of E; hence clearly A\E 1 d D 0, while .A \ E/ D 0 by the absolute continuity. Proof of Existence when is a Signed Measure. Applying the preceding result to the measures C , of the Jordan decomposition D C , we obtain two nonnegative, measurable functions fC ; f satisfying (7.11) with f˙ and ˙ instead of f and . Since at least one of the measures C and is bounded,41 at least one of the R functions fC ; f is integrable, so that the function f WD fC f and the integral f d are defined. Taking the difference of the equalities for C and we obtain (7.11) for f and . t u Using the Radon–Nikodým theorem we may greatly generalize the change of variable formula of integration42: Proposition 7.18 Assume that is strongly -finite, and let be an absolutely continuous measure. Then Z g

d d D d

Z g d

(7.14)

whenever the right-hand integral exists. Proof We may assume as usual that .X/ < 1. We write f WD d =d for brevity. (i) The set X0 WD fx 2 X W f .x/ D 0g satisfies the equality Z

Z

.X0 / D

f d D X0

0 d D 0 X0

and hence Z

Z

Z

gf d D X0

0 d D 0 D X0

g d X0

for all -measurable functions g.43 Therefore, changing X to X n X0 we may assume that f > 0. Then the -null sets and -null sets are the same by (7.11), so that we may use the expression a.e. without mentioning the corresponding measure or . (ii) Since is bounded, every -step function is also a -step function, and hence every -measurable function is also -measurable.

41

See Lemma 7.13 (b), p. 232. The proposition extends classical results of Euler [130, p. 303], Lagrange [280, p. 624] and Jacobi [224, p. 436]. 43 They are also -measurable because .X/ < 1. 42

7.7 * Local Measurability

247

If g is the characteristic function of a set A 2 M, then (7.14) reduces to the equality (7.11). Taking linear combinations it follows that (7.14) holds for all

-step functions. If .gn / is a sequence of -step functions satisfying gn % g a.e., then .gn f / is a sequence of -measurable functions satisfying gn f % g f a.e. Applying the generalized Beppo Levi theorem44 to the sequences .gn g1 / and .gn f g1 f / we get the equality (7.14). In the general case the equality (7.14) holds for gC and g instead of g. Taking the difference of these equalities we get (7.14) for g. t u

7.7 * Local Measurability As usual, we consider an integral associated with a finite measure defined on a semiring P. We denote by M the -ring of measurable sets and by W M ! R the extended measure. In the terminology of this chapter the constant functions are not necessarily measurable. In such cases the non-zero constant functions have no integral, and the measure of X is not defined either. We are going to extend the notions of the integral and the measure so as to deal with these cases in particular. Definition A function f is locally measurable if f P is measurable for every P 2 P. Remarks • Measurability implies local measurability. • The constant functions are locally measurable. If they are also measurable, then the notions of measurability and local measurability coincide. This is the case for X D R, studied in Chap. 5, more generally for X D RN , and for the probability measures. • If f is locally measurable, then the product fg is measurable for every measurable function g. For step functions g this follows at once from the definition. In the general case we choose a sequence .'n / of step functions converging to g a.e. Then the functions f 'n are measurable, and they converge to fg a.e., so that fg is measurable as well. An easy adaptation of the proof of Proposition 5.16 (p. 187) leads to the following Proposition 7.19 (a) The constant functions are locally measurable. (b) If f is locally measurable, and f D g a.e., then g is locally measurable.

44

Proposition 5.17 (e), p. 190.

248

7 Integrals on Measure Spaces

(c) If F W RN ! R is continuous, and f1 ; : : : ; fN are finite-valued, locally measurable functions, then the composite function h WD F. f1 ; : : : ; fN / is locally measurable. In particular, if f , g are finite-valued, locally measurable functions, then j f j, f C g, f g, fg, max f f ; gg and min f f ; gg are locally measurable as well. (d) If f is locally measurable and f ¤ 0 a.e., then 1=f is locally measurable. (e) If f is locally measurable, g is integrable, and j f j g a.e., then f is integrable. (f) If a sequence of locally measurable functions converges to f a.e., then f is also locally measurable. Next we generalize the integral: Definition Let f be a locally measurable function.

R • If f is nonnegative and non-integrable, then we define f dx WD 1. • If at least one of fC and f is integrable, then we define Z

Z

Z f dx WD

fC dx

f dx:

Remarks • If f is measurable, then the new definition reduces to the earlier one. • If neither fC nor f is integrable, then the right-hand sum is undefined. • We still keep the adjective “integrable” for the case where the integral is finite. Proposition 5.17 (p. 190) on the integration rules remains valid; we only have to use the local measurability of h in the proof of (d) instead of its measurability. After the integral we generalize the measure: Definition A set A is locally measurable if its characteristic function is locally measurable, i.e., if A \ P 2 M for every P 2 P. Remark The fundamental set X is always locally measurable.45 The following notion will be useful in the sequel: Definition A -algebra in X is a -ring containing X. Explicitly, a set system M in X is a -algebra if the following conditions are satisfied: • ¿ 2 M; • if A 2 M, then X n A 2 M; • if .An / is a disjoint sequence in M, then [ An 2 M. Examples • f¿; Xg and 2X are -algebras in X. 45

We recall from Lemma 7.5 (p. 220) that X is measurable ” it has a countable cover by sets of

P (and hence of finite measure).

7.7 * Local Measurability

249

• The usual Lebesgue measurable sets of R form a -algebra. • The countable subsets of an uncountable set X form a -ring, but not a -algebra. An easy adaptation of the proof of Proposition 5.19 (p. 194) leads to Proposition 7.20 (a) The locally measurable sets form a -algebra. (b) f is locally measurable ” the sets f f > cg ;

f f < cg ;

f f cg ;

f f cg

are locally measurable for all c 2 R. Remark The local measurability of f f > cg for all c 2 R already implies the local measurability of f . This follows from the relations f f > 1g D [1 nD1 ff > ng; f f > 1g D ¿ 2 M; f f cg D \1 nD1 f f > c 1=ng ; f f < cg D X n f f cg ; f f cg D X n f f > cg : Three similar statements are obtained by changing f f > cg to f f < cg, f f cg or f f cg. We extend the measure to the -algebra M of locally measurable sets by setting Z .A/ WD

A d:

Observe that .A/ D 1 for every A 2 M n M. Now we clarify the relationship between integrals and arbitrary measures. The following result complements Proposition 7.7 (p. 223): *Proposition 7.21 (a) M is a -algebra, and W M ! R is complete. (b) Every measure, defined on a semiring, is the restriction of the measure W M ! R associated with its finite part. Proof (a) We already know that M is a -algebra. The completeness of W M ! R follows from that of W M ! R because .A/ D 1 and thus .A/ ¤ 0 for all A 2 M n M.

250

7 Integrals on Measure Spaces

(b) Let W N ! R be a measure on a semiring, P WD fA 2 N W .P/ < 1g ; and W M ! R the measure obtained by the usual extension of WD jP . We have to show that N M and .A/ D .A/ for every A 2 N . First we observe the implication A2N

and P 2 P H) A \ P 2 P:

(7.15)

Indeed, since P N and N is a semiring, we have A \ P 2 N . Furthermore,

.A \ P/ .P/ < 1 and therefore A \ P 2 P. Since P M, (7.15) implies that every A 2 N is locally measurable, i.e., N M. It remains to show that .A/ D .A/ for every A 2 N . We distinguish the cases A 2 M and A 2 M n M. If A 2 N \ M, then A has a disjoint cover by sets Pn 2 P. Changing each Pn to A \ Pn by (7.15), we may also assume that A D [ Pn . Since .Pn / D .Pn / for every n by the definition of , it follows that .A/ D

X

.Pn / D

X

.Pn / D .A/:

If A 2 N and A 2 M n M, then .A/ D 1 by the definition of . Furthermore, A … P because P M, and therefore .A/ D 1 by the definition of P. Hence .A/ D .A/ again. t u Remark In view of part (b) of the proposition we may speak about the integral associated with an arbitrary measure, meaning the integral associated with its finite part. By the results of this section it is tempting to use local measurability and the measure W M ! R instead of measurability and the measure W M ! R.46 The following observations, however, convinced the author to return to the original definitions of Fréchet and Riesz47 : • Tonelli’s theorem on successive integration (p. 228) does not hold for locally measurable functions having an integral: the function f in the last example of Sect. 7.4 is locally measurable.

46

Indeed, this choice is taken by most contemporary textbooks by defining measurability using inverse images. While Hausdorff’s elegant characterization of continuous functions by inverse images of open or closed sets is extremely useful in topology, the analogous definition of measurability leads to several annoying counterexamples. 47 Fréchet [158] and Riesz–Sz.-Nagy [394].

7.8 Exercises

251

• Proposition 7.6 (p. 221) on the unique extension of measures does not remain valid for the -algebra M. To see this we consider the zero measure on the semiring P of finite subsets of an uncountable set X. Then M D 2X , and ( .A/ D

0

if A is countable,

1

if A is uncountable.

But the zero measure on 2X is also an extension of ! Moreover, the two measures already differ on the smallest -algebra N containing M, i.e., on the family of countable subsets and their complements. In fact,48 there are infinitely many other extensions of to N : the formula ( ˛ .A/ D

0

if A is countable,

˛

if X n A is countable

defines an extension of for each 0 ˛ 1. • The first part of the Radon–Nikodým theorem remains valid for locally measurable functions: if a locally measurable function has an integral, then the formula Z

.A/ WD

f d A

defines an absolutely continuous signed measure W M ! R, and even W M ! R. However, in the counterexample on p. 243 the Radon–Nikodým derivative f D d =d does not exist, even if we allow f to be only locally measurable.

7.8 Exercises Exercise 7.1 For each measure introduced in the examples on p. 213, determine its finite part, the -ring M of measurable sets, and the -algebra M of locally measurable sets. Exercise 7.2 Construct a nonnegative and additive, but not -additive function on the -algebra of all subsets of a countably infinite set X. Exercise 7.3 Construct a measurable set in R2 whose projections onto the coordinate axes are non-measurable.

48

L. Czách, private communication, 2005.

252

7 Integrals on Measure Spaces

Exercise 7.4 (Outer Measure)49 Given a finite measure on a semiring P in X, we set .A/ WD inf

1 X

.Pk /

kD1

for each A X where the infimum is taken over all sequences .Pk / P such that A [k Pk . (i) Show that is an outer measure: a nonnegative, -subadditive function on 2X , i.e, .A/

1 X

.An / whenever A

nD1

1 [

An :

nD1

(ii) Prove that .A [ B/ C .A \ B/ .A/ C .B/ for all A; B X. (iii) Prove that A X is measurable ” .B/ D .B \ A/ C .B n A/ for all B X. Exercise 7.5 (Riemann–Stieltjes Integral)50 Let us be given two functions f ; g W Œa; b ! R on a compact interval. For each finite subdivision I D fx0 ; 1 ; x1 ; : : : ; xn1 ; n ; xn g of the segment Œa; b, where a D x0 < 1 < x1 < < xn1 < n < xn D b; we set ı.I/ WD min.xk xk1 /; k

and we define the corresponding Riemann–Stieltjes sum by the formula S.I/ WD

n X

f . k / .g.xk / g.xk1 // :

kD1

49 50

Carathéodory [77]. See also Burkill [68], Halmos [184], and Natanson [332]. Stieltjes [435].

7.8 Exercises

253

If S.I/ converges to a finite limit L as ı.I/ ! 0, then we say that f is integrable with respect to g, and we write Z f 2 R.g/;

f dg D L:

Prove the following properties: (i) If f is continuous and g has bounded variation, then f 2 R.g/. (ii) If f 2 R.g/, then g 2 R. f /, and Z

Z g df D Œ fgba :

f dg C

Exercise 7.6 For which values of ˛ does the limit Z lim

1

x˛ d sin

h&0 h

1 x

exist? Exercise 7.7 Give an example of a strongly -finite measure that is not finite, and for which X is not measurable. Exercise 7.8 Construct measurable functions fi W R2 ! R with the following properties: (i) (ii) (iii) (iv)

The successive integrals of f1 in (7.4) exist, and are equal to zero. The successive integrals of f2 are equal to 0 and 1, respectively. The successive integrals of f3 are equal to 0 and 1, respectively. One of the successive integrals of f4 is equal to 0, and the other is undefined.

Taking linear combinations of the functions fi .x; y/ and fi .y; x/ show that no conclusion can be made of the successive integrals if f W R2 ! R is a measurable function whose integral is not defined. Exercise 7.9 (Hausdorff Dimension)51 Given a set A R and positive real numbers s; ı, let Hıs .A/ WD inf

1 nX

o jIi js ;

iD1

51 Hausdorff [196]. See, e.g., Falconer [134]. Some number-theoretical applications are given in de Vries–Komornik [101] and Komornik–Kong–Li [259].

254

7 Integrals on Measure Spaces

where the infimum is taken over the countable covers of A by intervals of length jIi j ı, and let H s .A/ WD sup Hıs .A/:52 ı>0

Prove the following results: (i) Hıs .A/ % H s .A/ as ı & 0. (ii) H s is an outer measure on R.53 (iii) There exists d 2 Œ0; 1 such that H s .A/ D 1 if s < d, and H s .A/ D 0 if s > d. It is called the Hausdorff dimension of A. (iv) Let Si W R ! R be a similarity with a scaling constant ci 2 .0; 1/, for i D 1; : : : ; m: If a non-empty compact set K is the disjoint union of S1 .K/; : : : ; Sm .K/, then the Hausdorff dimension d of K is the solution of the equation cd1 C C cdm D 1. (v) The Hausdorff dimension of Cantor’s ternary set is equal to ln 2= ln 3 0:63.

52 More generally, we may consider countable covers by sets of diameter diam Ii ı in a metric space. 53 Carathéodory’s construction (Exercise 7.4) yields the s-dimensional Hausdorff measure.

Part III

Function Spaces

We may resist everything, except temptation. –O. Wilde

Functional analysis started by studying (in today’s terminology) the space C.I/ of continuous functions defined on a compact interval. The idea of function spaces had already appeared in the doctoral dissertation of Riemann [370]. Dini [109] proved that for monotone sequences of continuous functions pointwise convergence is necessarily uniform. Ascoli [12] gave a sufficient condition for the compactness of a set in C.I/. This forms the basis for Peano’s theorem (1886) on the solvability of differential equations of the form x0 D f .t; x/ where f is merely continuous. (The Lipschitz condition serves only for the uniqueness of the solution.) Arzelà [8] proved that Ascoli’s condition is also necessary. Weierstrass [483] proved the density of polynomials in C.I/. Le Roux [299] and Volterra [472–475] obtained theorems of existence and uniqueness for a wide class of integral equations. Fredholm [150] discovered that the general theory of integral equations is much simpler than previously believed. Riesz [379] gave an elegant description of the dual space of C.I/ by using Stieltjes integrals. Cantor influenced Borel [58], Baire [17] and Lebesgue [287, 288] to widen the classes of sets and functions to be investigated. In his Ph.D. under the supervision of Hadamard, Fréchet [154] introduced the metric spaces and the notions of compactness, completeness and separability. Riesz [373, 374, 376] and Fischer [146] proved the completeness of the spaces of Lebesgue integrable functions, Riesz [375, 379] and Fréchet [155] characterized the duals of these spaces, and the discipline started to grow exponentially. The following works contain more complete studies of the historical development: [37, 45, 61, 106, 117, 203, 327, 365, 394, 421]. This last part of our book also serves as a synthesis: while Parts I and II are largely independent, here we build upon both.

256

III

Function Spaces

We did not resist the temptation to give multiple proofs of some theorems: either we could not choose among them or because they enlighten the problem from different angles, and thus contribute to the deeper understanding of the interconnections between different branches of analysis.

Chapter 8

Spaces of Continuous Functions

From the point of view of Mathematics the XIXth century could be called the century of the Theory of functions. . . . (V. Volterra, 1900)

In this chapter the letter K always denotes a compact Hausdorff space. We recall from topology that the continuous functions f W K ! R form a Banach space C.K/ with respect to the norm k f k1 WD max j f .t/j; t2K

and that norm convergence is uniform convergence on K. We will only present some basic results.1 Except for some uninteresting degenerate cases, the spaces C.K/ are not reflexive: Examples • Set I WD Œ0; 1, and consider in X WD C.I/ the closed affine subspace n

M WD f 2 C.I/ W f .0/ D 0 and

Z

1 0

o f .t/ dt D 1 :

We claim that M has no element of minimal norm, so that the distance dist.0; M/ is not attained.

1

Gillman–Jerison [169] and Semadeni [421] treat many further topics.

© Springer-Verlag London 2016 V. Komornik, Lectures on Functional Analysis and the Lebesgue Integral, Universitext, DOI 10.1007/978-1-4471-6811-9_8

257

258

8 Spaces of Continuous Functions

Fig. 8.1 Graph of fn

n+1 n 1

2 n+1

0

1

To prove this, first we observe that dist.0; M/ 1 because Z 1D

1 0

Z f .t/ dt

1 0

k f k1 dt D k f k1

(8.1)

for all f 2 M. Furthermore, the formula (see Fig. 8.1)

nC1 .n C 1/t min ;1 ; fn .t/ WD n 2

n D 1; 2; : : :

defines a sequence . fn / M satisfying k fn k1 D .n C 1/=n ! 1, so that in fact dist.0; M/ D 1. But this distance is not attained because the inequality in (8.1) is strict for every f 2 M because of the continuity of f and the condition f .0/ D 0. Applying Proposition 2.1 (p. 55) we conclude that C.I/ is not reflexive. • Set I D Œ1; 1, and consider on X WD C.I/ the linear functional Z '. f / WD

1 1

.sign t/f .t/ dt:

The obvious estimate Z j'. f /j

1 1

j f .t/j dt 2 k f k1

shows that ' is continuous, and k'k 2.

(8.2)

8 Spaces of Continuous Functions

259

Fig. 8.2 Graph of gn

1

−1

−1 n

1 n

1

−1

Furthermore, the formula2 (see Fig. 8.2) gn .t/ WD med f1; nt; 1g defines a sequence .gn / X satisfying kgn k1 D 1 for all n, and '.gn / ! 2; this implies that in fact k'k D 2. But the norm k'k is not attained, because j'. f /j < 2 k f k1 for all non-zero functions f 2 X. Indeed, we could have equality in (8.2) only if .sign t/f .t/ were constant in Œ1; 1, but this condition excludes all non-zero continuous functions. Applying Proposition 2.1 again, we conclude that C.I/ is not reflexive. • The spaces C.I/ are not only non-reflexive: they are not even dual spaces.3 Indeed, it follows from the Banach–Alaoglu and Krein–Milman theorems that the closed unit ball C of every dual Banach space is spanned by its extremal points. This is not satisfied for the closed unit ball C of C.I/: its only extremal points are the constant functions 1 and 1, and their closed convex hull contains only constant functions, while C contains non-constant functions as well. Later (on p. 298) we will also give a direct proof of the non-reflexivity. Despite their non-reflexivity, these spaces occur in many applications. This justifies their study in this chapter.

2 3

We recall that med fx; y; zg denotes the middle number among x, y and z. See Gelbaum–Olmsted [168]. The situation is similar to that of c0 ; see p. 140.

260

8 Spaces of Continuous Functions

8.1 Weierstrass Approximation Theorems The following theorem has countless applications: Theorem 8.1 (Weierstrass)4 Let Œa; b be a bounded, closed interval, and f W Œa; b ! R a continuous function. There exists a sequence . pn / of algebraic polynomials, converging uniformly to f on Œa; b. The theorem implies at once that C.Œa; b/ is separable: the polynomials with rational coefficients form a countable, dense set. The following proof is due to Landau.5 Fix a positive number R and define q W R ! R by the formula (see Fig. 8.3) ( q.t/ WD

R2 t 2

if jtj R,

0

if jtj R.

Lemma 8.2 For each fixed ı > 0 we have R jtj>ı

q.t/n dt

1

q.t/n dt

R1

!0

as n ! 1:

Proof The case ı R is obvious. Assuming henceforth that ı < R, we observe that q is a continuous even function, positive and decreasing in .0; R/, and vanishing Fig. 8.3 Graph of q for RD1

1 0.8 0.6 0.4 0.2 –1 –0.8 –0.6 –0.4 –0.2 0

4 5

0.2

0.4

t

0.6

0.8

Weierstrass [483], p. 5. Landau [283]. See Proposition 8.16 and Exercise 8.3 below (pp. 282,300) for other proofs.

1

8.1 Weierstrass Approximation Theorems

261

outside .R; R/. Therefore Z jtj>ı

q.t/n dt < .2R 2ı/q.ı/n < 2Rq.ı/n

and Z

Z

1

1

q.t/n dt >

jtjı=2

q.t/n dt > ıq.ı=2/n ;

so that R jtj>ı

q.t/n dt

1

q.t/n

0 R1

dt

2R q.ı/ n : ı q.ı=2/

Since 0 < q.ı/ < q.ı=2/, the last expression tends to zero as n ! 1.

t u

Proof of Theorem 8.1 By adding an affine polynomial if necessary, we may assume that f .a/ D f .b/ D 0. Then we may extend f by zero to a continuous function defined on R. The extended function is uniformly continuous, so that !. f ; ı/ WD sup fj f .x/ f .t/j W jx tj ıg ! 0 as ı & 0.6 Let us consider the function q of the preceding lemma with R to be chosen later, and set Z 1 n cn D q.t/n dt and Qn .t/ D c1 n q.t/ 1

for all n D 1; 2; : : : and t 2 R. Then we have Qn 0 in

R;

(8.3)

Qn .t/ D 0 if jtj R; Z 1 Qn .t/ dt D 1; Z

(8.5)

1

jtj>ı

Qn .t/ dt ! 0

as n ! 1;

see Fig. 8.4.

6

(8.4)

!. f ; ı/ is called the uniform continuity modulus of f .

for each ı > 0I

(8.6)

262

8 Spaces of Continuous Functions

Fig. 8.4 Graphs of Q1 , Q2 and Q3 for R D 1

1 0.8 0.6 0.4 0.2 –1 –0.8 –0.6 –0.4 –0.2 0

0.2

0.4

t

0.6

0.8

1

We claim that the functions Z pn .x/ WD

1 1

f .t/Qn .x t/ dt

converge to f uniformly in R. Indeed, applying (8.3) and (8.5) we have ˇZ ˇ j f .x/ pn .x/j D ˇ Z

1 1

ˇ ˇ . f .x/ f .t//Qn .x t/ dtˇ

jxtjı

(8.7)

j f .x/ f .t/jQn .x t/ dt Z C jxtj>ı

j f .x/ f .t/jQn .x t/ dt

Z

!. f ; ı/ C 2 k f k1

jsj>ı

Qn .s/ ds

for each x. For any fixed " > 0 choose ı > 0 such that !. f ; ı/ < "=2, and then using (8.6) choose N such that Z 2 k f k1 Qn .s/ ds < "=2 for all n N: jsj>ı

Then we conclude from (8.7) that j f .x/ pn .x/j < " for all x 2 R and n N.

8.1 Weierstrass Approximation Theorems

263

We complete the proof by showing that the restriction of pn to Œa; b is a polynomial if we choose R b a at the beginning of the proof. Applying (8.4), using the fact that f vanishes outside Œa; b, and taking into account that Œa; b Œx R; x C R for every a x b, we obtain the following equality for each a x b: Z 1 pn .x/ D f .t/Qn .x t/ dt Z

1

xCR

D Z

xR b

D a

2 2 n f .t/c1 n .R .x t/ / dt

2 2 n f .t/c1 n .R .x t/ / dt:

Since 2 2 n c1 n .R .x t/ / D

2n X

aj .t/xj

jD0

with suitable polynomials aj .t/, it follows that pn .x/ D

2n X jD0

Z b j xj

with

b

bj D

f .t/aj .t/ dt:

t u

a

Remark The above proof was perhaps the first example of regularization by convolution, a technique widely used today to establish density theorems in various functions spaces.7 Weierstrass also proved a similar result for periodic functions. The 2-periodic continuous functions form a closed subspace C2 in the Banach space B.R/, hence C2 is also a Banach space with respect to the norm kk1 .8 Definition A trigonometric polynomial is a finite linear combination of the functions 1; cos t; sin t; cos 2t; sin 2t; cos 3t; sin 3t; : : : :

7 8

See the references in the footnote of Sect. 9.3 below, p. 320. We recall that in this book by a subspace without adjective we always mean a linear subspace.

264

8 Spaces of Continuous Functions 2

Fig. 8.5 Graph of q for RD1

1.5 1 0.5 –3

–2

–1

0

1

t

2

3

Remark Using the three identities9 2 cos kt cos mt D cos.k m/t C cos.k C m/t; 2 sin kt sin mt D cos.k m/t cos.k C m/t; 2 sin kt cos mt D sin.k m/t C sin.k C m/t it is easy to show that the trigonometric polynomials form not only a vector space, but also an algebra: the product of two trigonometric polynomials is again a trigonometric polynomial. Theorem 8.3 (Weierstrass)10 For each f 2 C2 there exists a sequence .pn / of trigonometric polynomials converging uniformly to f on R. The following proof is due to de la Vallée-Poussin.11 Proof Introducing the function ( q.t/ WD

1 C cos t

if jtj ,

0

if jtj

(see Fig. 8.5), and repeating the preceding proof with R D we obtain that pn ! f uniformly in R.

9 Several proofs of this chapter could be simplified by adopting the complex framework, and using Euler’s formula eix D cos x C i sin x. For example, the trigonometric polynomials would be simply the algebraic polynomials of eit , and the single identity euCv D eu ev would suffice instead of these three real identities. 10 Weierstrass [483]. See Theorem 8.11 and a remark following Proposition 8.21 below (pp. 276, 288) for other proofs. 11 de la Vallée-Poussin [463]. His work was motivated by that of Landau.

8.2 * The Stone–Weierstrass Theorem

265

It remains to show that pn is a trigonometric polynomial. This follows from the following computation: Z pn .x/ D

1 1

D c1 n D c1 n D c1 n

f .t/Qn .x t/ dt

Z

xC

f .t/.1 C cos.x t//n dt

x Z

Z

D a0 C

f .t/.1 C cos.x t//n dt f .t/.1 C cos x cos t C sin x sin t/n dt

n X

ak cos kx C bk sin kx;

kD1

where ak and bk are suitable real numbers. The third equality follows from the 2-periodicity of the function under the integral sign, while the last one from the repeated application of the three trigonometric identities of the preceding remark. t u Remark Jackson [221], [222] investigated the error of the approximation as a function of the regularity of the approximated function. Müntz [329], Szász [445], Clarkson and Erd˝os [90] proved important generalizations of Theorem 8.1. See also Achieser [1], Cheney [85], Jackson [223], Natanson [333], Rudin [405].

8.2 * The Stone–Weierstrass Theorem Stone proved a far-reaching generalization of the Weierstrass approximation theorems. Definition A subspace M of C.K/ is a subalgebra if f ; g 2 M imply fg 2 M. Theorem 8.4 (Stone–Weierstrass)12 Let K be a compact topological space and M a subalgebra of C.K/. Assume that M contains the constant functions, and separates the points of K: for any two distinct points x; y 2 K there exists an h 2 M such that h.x/ ¤ h.y/. Then M is dense C.K/.

12

Stone [440], [441].

266

8 Spaces of Continuous Functions

Examples • Let K be a compact interval in R. The restrictions of the algebraic polynomials to K form a subalgebra M satisfying the conditions of Theorem 8.4. Hence Theorem 8.1 is a special case of Theorem 8.4. • More generally, if K is a compact set in RN , then the algebraic polynomials of N variables form a subalgebra M satisfying the conditions of Theorem 8.4. • Let K be the unit circle in R2 . Setting T.s/ WD .cos s; sin s/, the function f 7! f ı T establishes an isometric isomorphism between the Banach spaces C.K/ and C2 . Furthermore, the algebraic polynomials of two variables correspond to the trigonometric polynomials. Thus Theorem 8.3 also follows from Theorem 8.4. In the proof we use the notion of vector lattices (see p. 172). Proof of Theorem 8.4 First step. because

If fn ! f and gn ! g C.K/, then fn gn ! fg

k fg fn gn k1 k f fn k1 kgk1 C k fn k1 kg gn k1 ! 0: Hence the closure M of the subalgebra M is still a subalgebra of C.K/. Second step. We show that the closed subalgebra M is a vector lattice. Fix h 2 M arbitrarily and fix a number T > khk1 . By Theorem 8.1 there exist polynomials pn satisfying pn .x/ ! jxj uniformly in ŒT; T. Then pn ıh 2 M, and pn ıh ! jhj uniformly in K, so that jhj 2 M. The following proposition completes the proof of the theorem. t u Proposition 8.5 (Kakutani–Krein)13 Let K be a compact topological space and M C.K/ a vector lattice. Assume that 1 2 M, and that M separates the points of K. Then M is dense in C.K/. Proof Fixing f 2 C.K/ and " > 0 arbitrarily, we have to find g 2 M satisfying k f gk1 < ". First step.

For each fixed x 2 K there exists a function fx 2 M satisfying fx > f " on K;

and fx .x/ D f .x/:

Indeed, by our assumption for each y 2 K there exists a function fxy 2 M equal to f at x and y. Then the open sets ˚ Uy WD z 2 K W fxy .z/ > f .z/ " ;

13

Kakutani [240, pp. 1004–1005], Krein–Krein [268].

y2K

8.2 * The Stone–Weierstrass Theorem

267

cover the compact set K, because y 2 Uy for every y. If K D Uy 1 [ [ U y n is a finite subcover, then the function ˚ fx WD max fxy1 ; : : : ; fxyn has the required properties. Second step. There exists a function g 2 M satisfying f " < g < f C " on K; and hence the inequality k f gk1 < ". For the proof we consider the functions fx 2 M obtained in the first step. The open sets Vx WD fz 2 K W fx .z/ < f .z/ C "g ;

x2K

cover the compact set K, because x 2 Vx for every x. If K D V x1 [ [ Vxm is a finite subcover, then the function g WD min f fx1 ; : : : ; fxm g t u

has the required properties. 14

The following interesting application will be useful later : Proposition 8.6 (Stone)15 Let K be a compact set in a topological space X, and assume that the points of K may be separated by the continuous functions h W X ! R. Then every continuous function f W K ! R may be extended to a continuous function F W X ! R. Proof The restrictions of the continuous functions F W X ! R to K form a vector lattice M in C.K/, containing the constant functions. By our assumption M satisfies the conditions of the Kakutani–Krein theorem, and hence it is dense in C.K/. It remains to prove that M is closed. Let . fn / M converge uniformly on K to some function f . We have to find a continuous function F W X ! R such that F D f on K.

14 15

See the proof of Lemma 8.27, p. 297. Stone [441]. This is a version of similar theorems of Urysohn [461] and Tietze [453].

268

8 Spaces of Continuous Functions

Taking a subsequence if necessary, we may assume that j fnC1 fn j 2n

on K

for every n.16 By the definition of M the functions f1 and fnC1 fn have continuous extensions F1 and Gn to X. Furthermore, we may assume that jGn j 2n

on K

for every n: change Gn to med f2n ; Gn ; 2n g if necessary. Then the function series F1 C

1 X

Gn

nD1

converges uniformly to some function F W X ! R. We conclude that F is continuous, and F D f on K. u t

8.3 Compact Sets. The Arzelà–Ascoli Theorem In this section we characterize the compact sets of C.K/. Since in complete metric spaces the compact sets coincide with the totally bounded17 closed sets, it is sufficient to characterize the totally bounded sets. Definitions Consider a family of functions F C.K/. • F is pointwise bounded if ff .t/ W f 2 F g is bounded in R for each t 2 K. • F is equicontinuous if for each " > 0 and t 2 K there is a neighborhood V of t such that j f .s/ f .t/j < " for all s 2 V and f 2 F . Proposition 8.7 (Arzelà–Ascoli)18 A family of functions F C.K/ is totally bounded ” it is pointwise bounded and equicontinuous.

16

We have already used this technique when proving the Riesz Lemma 5.13, p. 184. We recall that a set A is totally bounded or precompact if for each r > 0 it has a finite cover by balls of radius r. 18 Ascoli [12] (pp. 545–549, sufficiency for K D Œ0; 1), Arzelà [8] (necessity), [9] (simplified treatment), [10], Fréchet [154] (general case). 17

8.3 Compact Sets. The Arzelà–Ascoli Theorem

269

Proof First let F be totally bounded. Then it is also bounded in norm, i.e., uniformly bounded on K, and hence pointwise bounded as well. To show the equicontinuity, it suffices to find for any fixed t 2 K and r > 0 a neighborhood V of t such that j f .t/ f .s/j < 3r

for all f 2 F

and s 2 V:

(8.8)

Let us cover F with finitely many balls of radius r: F Br . f1 / [ [ Br . fm / with f1 ; : : : ; fm 2 F. Since each fi is continuous at t, we may choose a neighborhood Vi of t such that j fi .t/ fi .s/j < r

for all s 2 Vi :

Then (8.8) is satisfied with V WD V1 \ \ Vm . Indeed, for any given f 2 F and s 2 V, choosing i such that k f fi k < r, we have j f .t/ f .s/j j f .t/ fi .t/j C j fi .t/ fi .s/j C j fi .s/ f .s/j < r C r C r: Conversely, if F is equicontinuous, then by the compactness of K we may find for each fixed r > 0 finitely many points t1 ; : : : ; tm 2 K and their neighborhoods V1 ; : : : ; Vm such that K D V1 [ [ Vm , and j f .t/ f .ti /j < r

whenever f 2 F

and t 2 Vi :

If, moreover, F is pointwise bounded, then the set f. f .t1 /; : : : ; f .tm // W f 2 F g is bounded Rm , and also totally bounded there.19 There exist therefore finitely many functions f1 ; : : : ; fn 2 F such that20 f. f .t1 /; : : : ; f .tm // W f 2 Fg

n [

Br . fj .t1 /; : : : ; fj .tm //:

jD1

19

We recall that the bounded and totally bounded sets are the same in all finite-dimensional normed spaces. 20 In this formula the balls are taken in Rm .

270

8 Spaces of Continuous Functions

We complete the proof by showing that21 F B3r . f1 / [ [ B3r . fn /: For any given f 2 F first we choose fj satisfying . f .t1 /; : : : ; f .tm // 2 Br . fj .t1 /; : : : ; fj .tm //: Next, for any given t 2 K we choose i such that t 2 Vi . Then we have ˇ ˇ ˇ ˇ ˇ ˇ ˇ f .t/ fj .t/ˇ j f .t/ f .ti /j C ˇ f .ti / fj .ti /ˇ C ˇ fj .ti / fj .t/ˇ < r C r C r; whence f 2 B3r . fj /.

t u

8.4 Divergence of Fourier Series By the Fourier series of a function f 2 C2 we mean the function series22 1

a0 X C ak cos kx C bk sin kx; 2 kD1 with the Fourier coefficients ak ; bk defined by the formulas 1 ak WD

Z

f .t/ cos kt dt

1 and bk WD

Z

f .t/ sin kt dt:

Remark C2 is a Euclidean space with respect to the scalar product . f ; g/ WD R fg dt. A simple computation shows that the mth partial sum of the Fourier series is the orthogonal projection of f onto the subspace Tm of the trigonometric polynomials of order m, spanned by the functions 1; cos t; sin t; cos 2t; sin 2t; cos 3t; sin 3t; : : : ; cos mt; sin mt: See Sect. 1.4, p. 24.

We recall that r > 0 was chosen arbitrarily at the beginning. Daniel Bernoulli P1 [38], Fourier [148]. Using complex numbers the Fourier series would take the simpler form kD1 ck eikx . 21 22

8.4 Divergence of Fourier Series

271

Following Fourier’s revolutionary treatise, many works were devoted to the convergence of Fourier series23 : • Dirichlet and Jordan24 proved (among others) that if f 2 C2 has bounded variation, then its Fourier series converges to f uniformly. • Lipschitz and Dini25 proved (among others) that if f 2 C2 , then its Fourier series converges to f .a/ at each point a where f is differentiable. It remained an open question for fifty years whether mere continuity already ensures the convergence of the Fourier series. Finally, a counterexample was found: Proposition 8.8 (du Bois-Reymond)26 There exists an f 2 C2 whose Fourier series does not converge pointwise to f . Remarks • However, Carleson proved that the Fourier series of each f 2 C2 converges to f a.e. everywhere.27 • On the other hand, Kahane and Katznelson28 proved that for each null set E there exists a function f 2 C2 that diverges at the points of E. First we establish two lemmas. Lemma 8.9 (Dirichlet)29 The partial sums a0 X C ak cos kx C bk sin kx 2 kD1 m

.Sm f /.x/ WD

of the Fourier series of a function f 2 C2 may be written in the closed form .Sm f /.x/ D

1 2

Z

Dm .x t/f .t/ dt;

with the Dirichlet kernel Dm 2 C2 defined by the formula30 Dm .2s/ WD

23

sin.2m C 1/s : sin s

A fascinating historical account is given by Kahane [237]. Dirichlet [112], Jordan [229]. 25 Lipschitz [308] and Dini [107], [110]. See a short proof in Exercise 8.5, p. 301. 26 du Bois-Reymond [49], [51]. A simpler explicit counterexample was given later by Fejér [139], [140]. We prove here the mere existence of such functions. 27 Carleson [78]. This was a long-standing open problem of Lusin [313]. See also the remark following Corollary 9.6 below (p. 314) concerning Lp convergence. 28 Kahane and Katznelson [238]. See also Edwards [120], Katznelson [245] and Zygmund [493] for many further results. 29 Dirichlet [112]. 30 For sin s D 0 we replace the right-hand side by its limit .2m C 1/. 24

272

8 Spaces of Continuous Functions 2

Fig. 8.6 Graph of D0

1.5

1

0.5

–3

–2

–1

Fig. 8.7 Graph of D1

0

1

t

2

3

2

3

3

2

1

–3

–2

1

–1

s

–1

See Figs. 8.6, 8.7, 8.8, and 8.9. Proof Since a0 X C ak cos kx C bk sin kx 2 kD1 m

.Sm f /.x/ D D

D

1 2 1 2

Z

Z

1C2

m X

cos kx cos kt C sin kx sin kt f .t/ dt

kD1

1C2

m X kD1

cos k.x t/ f .t/ dt;

8.4 Divergence of Fourier Series

273

5

4

3

2

1

–3

–2

–1

1

s

2

3

–1

Fig. 8.8 Graph of D2

it is sufficient to prove the identity 1C2

m X

cos 2ks D

kD1

sin.2m C 1/s : sin s

The case m D 0 is obvious. The general case follows by induction, using the trigonometric identities 2 sin s cos 2.m C 1/s D sin.2m C 3/s sin.2m C 1/s; Now we introduce the linear functionals 'm . f / WD .Sm f /.0/ on the Banach space C2 .

m D 0; 1; : : : :

t u

274

8 Spaces of Continuous Functions

6

4

2

–3

–2

–1

0

1

s

2

3

Fig. 8.9 Graph of D3

Lemma 8.10 The linear functionals 'm are continuous, and k'm k ! 1 as m ! 1. Proof Since jak j ; jbk j 2 k f k1 ; we deduce from the definition of Sm that 1 2 k f k1 D .4m C 1/ k f k1 I kSm f k1 2m C 2 hence k'm k 4m C 1 < 1. On the other hand, the formula f .2s/ WD .sign sin s/ sin.2m C 1/s

8.5 Summability of Fourier Series. Fejér’s Theorem

275

defines a function f 2 C2 satisfying k f k1 D 1 and 'm . f / D D > >

D

Z 1 Dm .t/f .t/ dt 2 Z 1 =2 sin2 .2m C 1/s ds =2 jsin sj Z 2 =2 sin2 .2m C 1/s ds 0 s m Z 2 X j sin2 s ds jD1 .j1/ s

1 D D D

2 2

2 >

Z

=2 =2

Z Z Z

=2 0

Dm .2s/f .2s/ ds sin2 .2m C 1/s ds sin s

.2mC1/=2 0

0

sin2 s ds s

m X sin2 s jD1

j

ds

m 1 X1 : jD1 j

Hence, k'm k 'm . f / >

m 1 X1 ! 1: jD1 j

t u

Remarks • We note for later reference that the test functions used in the proof are even. • Fejér31 has established the more precise asymptotic formulas k'm k D

4 log m C O.1/; 2

m ! 1:

Proof of Proposition 8.8 Assume on the contrary that 'm . f / ! f .0/ for each f 2 C2 . Then applying the Banach–Steinhaus theorem (p. 81) with X D C2 and Y D R we obtain sup k'm k < 1, contradicting the preceding lemma. t u

8.5 Summability of Fourier Series. Fejér’s Theorem Thought is only a flash in the middle of a long night, but this flash is everything. (H. Poincaré)

31

Fejér [141]. See also Edwards [120] or Zygmund [493].

276

8 Spaces of Continuous Functions

The counterexample of du Bois-Reymond made obvious the difficulties of representing continuous functions by Fourier series. Minkowski even asked whether the Fourier series of a continuous function may converge pointwise to another function.32 The long period of stagnation ended when Fejér discovered the following remarkable Theorem 8.11 (Fejér)33 Given any f 2 C2 , the mean values n f WD

n 1 X Sm f ; n C 1 mD0

n D 0; 1; : : :

converge to f uniformly on R. Remarks The theorem has important consequences: • It provides a new proof of the second approximation theorem of Weierstrass. • It implies that the Fourier series of f 2 C2 cannot converge at any point x to a value different from f .x/.34 Indeed, this follows from a classical result of Cauchy35 : if an ! a for a numerical sequence, then we also have .a1 C C an /=n ! a. First we prove a lemma: Lemma 8.12 We have .n f /.x/ D

1 2

Z

Fn .x t/f .t/ dt

with the Fejér kernel Fn 2 C2 defined by the formula36 Fn .2s/ WD

1 sin2 .n C 1/s : nC1 sin2 s

Let us compare Figs. 8.10, 8.11, 8.12, and 8.13 and Figs. 8.6, 8.7, 8.8, and 8.9 on p. 274: the positivity of the Fejér kernel has a great importance.

32 See Hawkins [198]. An analogous phenomenon for Taylor series has been known since Cauchy [80, p. 230]. 33 Fejér [137, 138]. He also investigated pointwise convergence for discontinuous functions f . Lebesgue [292] extended his results to Lebesgue integrable functions. 34 Thereby he has answered Minkowski’s question. Banach [20] has shown that Minkowski’s phenomenon occurs for a slight modification of the trigonometric system. 35 Cauchy [79]. 36 For sin s D 0 the right-hand side is replaced by its limit .n C 1/.

8.5 Summability of Fourier Series. Fejér’s Theorem

277

Fig. 8.10 Graph of F0

2

1.5

1

0.5

–3

–2

–1

Fig. 8.11 Graph of F1

0

1

s

2

3

2

3

2

1.5

1

0.5

–3

–2

–1

1

s

278

8 Spaces of Continuous Functions

Fig. 8.12 Graph of F2

3

2.5

2

1.5

1

0.5

–3

–2

–1

Fig. 8.13 Graph of F3

0

1

s

2

3

2

3

4

3

2

1

–3

–2

–1

1

s

8.6 * Korovkin’s Theorems. Bernstein Polynomials

279

Proof By the definition of the operators n it suffices to prove the equalities Fn D

D0 C C Dn ; nC1

or equivalently that n X sin2 .n C 1/s sin.2m C 1/s : D 2 sin s sin s mD0

They follow by a direct computation: n X

.sin s/ sin.2m C 1/s D

mD0

D

n 1 X cos 2ms cos.2m C 2/s 2 mD0

1 cos.2n C 2/s 2

D sin2 .n C 1/s:

t u

Proof of Theorem 8.11 We obtain the relations n 1 D 1;

n cos D

n cos nC1

and n sin D

n sin nC1

directly from the definitions. Hence k f n f k1 ! 0 for the three functions f D 1, cos and sin. If f 0, then n f 0 by the positivity of the Fejér kernels. Therefore we may conclude by applying Proposition 8.13 below. t u Definition A linear map L W C2 ! C2 is positive if f 0 H) Lf 0. Proposition 8.13 (Korovkin)37 Consider a sequence of positive linear maps Ln W C2 ! C2 . If k f Ln f k1 ! 0 for the three functions f D 1; cos; sin, then the relation k f Ln f k1 ! 0 holds in fact for all f 2 C2 . We prove a more general theorem in the next section.

8.6 * Korovkin’s Theorems. Bernstein Polynomials Let us investigate the positive linear maps L W C.K/ ! C.K/ for an arbitrary compact topological space.

37

Korovkin [263]. Many applications are given in Korovkin [264].

280

8 Spaces of Continuous Functions

Definition L is positive if f 0 H) Lf 0. Remarks If L is a positive linear map, then • L is monotone: Lf Lg whenever f g: this follows at once from the linearity of L; • L is continuous with kLk D kL1k1 . Indeed, using the monotonicity we infer from the inequalities k f k1 f k f k1 that k f k1 .L1/ Lf k f k1 .L1/; and hence kLf k1 kL1k1 k f k1 for all f . Since equality holds for f D 1, we conclude that kLk D kL1k1 . Let K be a compact topological space and h1 ; : : : ; hm 2 C.K/. Assume that the functions hj separate the points of K: for any two distinct points x; y 2 K there exists a j such that hj .x/ ¤ hj .y/. Consider a sequence of positive linear maps Ln W C.K/ ! C.K/. Proposition 8.14 (Freud)38 If k f Ln f k1 ! 0 for the functions f D 1; h1 ; : : : ; hm

and f D h21 C C h2m ;

(8.9)

then k f Ln f k1 ! 0 for all f 2 C.K/. Example If K is a compact set in Rm , then we may apply the proposition to the projections hj .x/ WD xj , j D 1; : : : ; m. Proof Fix f 2 C.K/ and " > 0 arbitrarily. First step. For each N D 1; 2; : : : ; let us denote by UN the set of pairs .x; y/ 2 K K satisfying the inequality j f .x/ f .y/j < " C N

m X ˇ ˇ ˇhj .x/ hj .y/ˇ2 :

(8.10)

jD1

These sets are open by the continuity of the functions f and hj , and they form an increasing set sequence. Furthermore, since m X ˇ ˇ ˇhj .x/ hj .y/ˇ2 > 0 jD1

38

Freud [153]. See Altomare and Campiti [5] for a very complete review of the subject.

8.6 * Korovkin’s Theorems. Bernstein Polynomials

281

whenever x ¤ y (by the separation condition), they cover K K. The latter space being compact, there exists a positive integer N such that (8.10) is satisfied for all x; y 2 K. Second step. For any fixed x 2 K, (8.10) implies the inequality j f .x/.Ln 1/.y/ .Ln f /.y/j ".Ln 1/.y/ CN

m X

h2j .x/.Ln 1/.y/ 2N

jD1

m X

hj .x/.Ln hj /.y/

jD1

C NLn

m X

h2j .y/

jD1

for all y 2 K. Choosing y D x and applying the triangle inequality this yields the following estimate: j f Ln f j j f j j1 Ln 1j C ".Ln 1/ CN

m X jD1

h2j .Ln 1/

2N

m X

hj .Ln hj / C NLn

jD1

m X

h2j :

jD1

Letting n ! 1, the right-hand side tends to " uniformly by our assumption, and hence k f Ln f k1 < 2" for all sufficiently large n.

t u

Corollary 8.15 (Bohman–Korovkin)39 Let I be a compact interval, and consider a sequence of positive linear maps Ln W C.I/ ! C.I/. If the relation k f Ln f k1 ! 0 holds for the three functions f .x/ D 1; x; x2 , then it holds in fact for all f 2 C.I/. Proof We apply the preceding example with K D I and m D 1. Now we return to the last statement of the preceding section.

39

Bohman [47], Korovkin [263].

t u

282

8 Spaces of Continuous Functions

Fig. 8.14 x21 C x22 D 1

x2

x1

Proof of Proposition 8.13 We apply the preceding example to the unit circle K of R2 . (See Fig. 8.14.) Since x21 C x22 D 1 on K, we have only three test functions instead of four. Hence, if a sequence of positive linear maps Ln W C.K/ ! C.K/ satisfies k f Ln f k1 ! 0 for the three functions f .x/ WD 1; x1 ; x2 , then the relation k f Ln f k1 ! 0 holds in fact for all f 2 C.I/. Now we recall (p. 266) that the map f 7! f ı T, where T.s/ WD .cos s; sin s/, is an isometric isomorphism between the Banach spaces C.K/ and C2 . Furthermore, f 0 ” f ı T 0, and the map transforms the functions f .x/ D 1; x1 ; x2 into f .T.s// D 1; cos s; sin s. Hence the result obtained for K is equivalent to Proposition 8.13. t u As another application of Korovkin’s theorems, we give a new proof of the first approximation theorem of Weierstrass.40 Let I D Œ0; 1 for simplicity, and introduce for each f 2 C.I/ the Bernstein polynomials41 ! n X n k k f x .1 x/nk ; .Bn f /.x/ WD k n kD0

x 2 I;

n D 1; 2; : : : :

Proposition 8.16 (Bernstein)42 The Bernstein polynomials Bn f converge uniformly to f on I for each f 2 C.I/.

40

Theorem 8.1, p. 260. Bernstein’s proof is probabilistic, based on the law of large numbers. 42 Bernstein [39]. His result answered a question of Borel [60, pp. 79–82]. 41

8.6 * Korovkin’s Theorems. Bernstein Polynomials

283

Proof The operators Bn are clearly positive linear on C.I/. Let us also observe that43 Bn 1 D 1 and Bn id D id for every n via the binomial theorem: ! n X n k x .1 x/nk .Bn 1/.x/ D k kD0 D .x C 1 x/n D1 and ! n X n k k x .1 x/nk .Bn id/.x/ D k n kD0 ! n X n1 k x .1 x/nk D k 1 kD1 D x.x C 1 x/n1 D x: In view of the Bohman–Korovkin theorem (p. 281) it suffices to show that Bn .id2 / converges uniformly to id2 on Œ0; 1. For this we first note that ! n X id n k.k 1/ k 2 .x/ D Bn id x .1 x/nk 2 k n n kD0 ! n n1 X n2 k x .1 x/nk D n kD2 k 2 D

n1 2 x : n

Hence Bn .id2 / D

n1 2 1 id C id n n

and therefore 2 id Bn .id2 /

1

43

We denote by id the identity map of I.

D

1 id2 id ! 0: 1 n

t u

284

8 Spaces of Continuous Functions

8.7 * Theorems of Haršiladze–Lozinski, Nikolaev and Faber The main theorem of this section reveals a deep common reason for many divergence theorems. As in Sect. 8.4, we denote by Tm the vector space of trigonometric polynomials of order m, and we denote by Sm f the mth partial sum of the Fourier series of f . Theorem 8.17 (Haršiladze–Lozinski)44 Consider a sequence of continuous linear maps Lm W C2 ! C2 . If Lm is a projection onto Tm for each m, then there exists a function f 2 C2 such that k f Lm f k1 6! 0. The main ingredient of the proof is an optimality property of Fourier series: Proposition 8.18 (Lozinski)45 If a continuous linear map Lm W C2 ! C2 is a projection onto Tm , then kLm k kSm k. Indeed, in view of the Banach–Steinhaus theorem (p. 81), Theorem 8.17 follows from this proposition and from the fact that kSm k ! 1, proved in Lemma 8.10 (p. 273). Proof of Proposition 8.18 For each real number s the formula .Ts f /.x/ WD f .x C s/ defines in C2 a continuous linear operator of norm one. It suffices to establish the following identity46 : .Sm f /.x/ D

1 2

Z

.Ts Lm Ts f /.x/ ds;

x 2 R;

f 2 C2 :

(8.11)

Indeed, since j.Ts Lm Ts f /.x/j kTs Lm Ts f k1 kTs k kLm k kTs k k f k1 D kLm k k f k1 for all f , s and x, (8.11) implies kSm f k1 kLm k k f k1 for all f , and hence kSm k kLm k.

44

Lozinski [311]. Lozinski [311]. 46 Marcinkiewicz [314], Lozinski [310]. 45

8.7 * Theorems of Haršiladze–Lozinski, Nikolaev and Faber

285

It is sufficient to prove (8.11) for the functions47 fk .x/ D cos kx

.k D 0; 1; : : :/ and gk .x/ D sin kx

.k D 1; 2; : : :/:

Indeed, then the identity will hold for all trigonometric polynomials by linearity, and then for all f 2 C2 by the Weierstrass approximation theorem because all operators occurring in (8.11) are continuous. If f 2 Tm , then Ts f 2 Tm . Hence Lm Ts f D Ts f and therefore Z Z 1 1 .Ts Lm Ts f /.x/ ds D f .x/ ds D f .x/ D .Sm f /.x/: 2 2 It remains to prove that Z Z .Ts Lm Ts fk /.x/ ds D

.Ts Lm Ts gk /.x/ ds D 0

for all k > m and x 2 R. We deduce from the identities cos k.x C s/ D cos ks cos kx sin ks sin kx and sin k.x C s/ D sin ks cos kx C cos ks sin kx that Ts fk D .cos ks/fk .sin ks/gk

and Ts gk D .sin ks/fk C .cos ks/gk :

Consequently, Z

.Ts Lm Ts fk /.x/ ds Z D

.cos ks/.Lm fk /.x s/ .sin ks/.Lm gk /.x s/ ds

and Z

.Ts Lm Ts gk /.x/ ds Z D

47

.sin ks/.Lm fk /.x s/ C .cos ks/.Lm gk /.x s/ ds:

The proof may be simplified by using complex numbers. See Exercise 8.10, p. 303.

286

8 Spaces of Continuous Functions

For any fixed x, .Lm fk /.x s/ and .Lm gk /.x s/ are trigonometric polynomials of order m in s. Since k > m, they are therefore orthogonal to the functions cos ks and sin ks, so that the right-hand side of both identities vanishes. t u Next we establish an algebraic variant of Theorem 8.17. For this we need a variant of Proposition 8.18, where we replace C2 and Tm by the subspaces CQ 2 and TQm formed by the even functions. Let us denote the restriction of Sm to CQ 2 by SQ m , and observe that SQ m W CQ 2 ! CQ 2 . Proposition 8.19 If a continuous linear map Lm W CQ 2 ! CQ 2 is a projection onto TQm , then kLm k QSm =2. Proof Using the notations of the preceding proof it suffices to prove the following identity: 1 .SQ m f /.x/ D 2

Z

.Ts Lm .Ts C Ts /f /.x/ ds

for all f 2 CQ 2 and x 2 R. Indeed, this will imply QSm f 2 kLm k k f k for all f 2 CQ 2 . Since the functions fk span CQ 2 , it suffices to prove the identity for these functions. We infer from the trigonometric identity cos k.x s/ C cos k.x C s/ D 2 cos ks cos kx that .Ts C Ts /fk D .2 cos ks/fk ; and hence 1 RQ m f .x/ WD 2

Z

.Ts Lm .Ts C Ts /f /.x/ ds D

1 2

Z

.2 cos ks/.Lm fk /.x s/ ds:

If k > m, then for each fixed x, .Lm fk /.x s/ is a trigonometric polynomial of order < k in s, and thus orthogonal to cos ks. Therefore RQ m fk D 0 D SQ m fk . If k m, then Lm fk D fk , so that .RQ m fk /.x/ D

1 2

D

1 2

Z

Z

2 cos ks cos k.x s/ ds cos kx C cos k.x 2s/ ds

8.7 * Theorems of Haršiladze–Lozinski, Nikolaev and Faber

287

D cos kx D fk .x/ D SQ m fk .x/ t u

again. Let us denote by Pm the vector space of algebraic polynomials of degree m.

Theorem 8.20 (Haršiladze–Lozinski)48 Consider a sequence of continuous linear maps Lm W CI ! CI , where I is a compact interval. If Lm is a projection onto Pm for each m, then there exists an f 2 CI such that k f Lm f k1 6! 0. Proof Let I D Œ1; 1 for simplicity of notation, and consider the isometric isomorphism T W f 7! f ı cos between the Banach spaces C.I/ and CQ 2 . Since f 2 Pm ” Tf 2 TQm ; we deduce from the preceding proposition that kLm k D TLm T 1 QSm =2: Let us observe that QSm ! 1 by the proof of Lemma 8.10 (p. 273), because in the proof only even test functions were used. Therefore we may conclude by applying the Banach–Steinhaus theorem (p. 81). t u We end this section with two further famous results. Given a compact interval I D Œa; b, we may ask the following natural questions: • Does there exist a weight function49 on some compact interval J I such that, considering P the corresponding orthonormal sequence of polynomials pn , the Fourier series . f ; pn /pn converges uniformly to f on I for every f 2 C.J/? • Given a system of points xm;0 < < xm;m in I for m D 0; 1; : : : ; we may define for each f 2 C.I/ a sequence of Lagrange interpolation polynomials Lm f such that Lm D f in the points xm;0 ; : : : ; xm;m . Is there a choice of points xm;k such that Lm f converges uniformly to f for every f 2 C.I/?

48

Lozinski [311]. By a weight function we mean a positive, integrable function. If w is a weight function on a compact interval J, then we may defineR a scalar product on the vector space P of algebraic polynomials by the formula .p; q/ WD I pqw dt, and we may apply the Gram–Schmidt orthogonalization (Proposition 1.15, p. 28) for the sequence of functions 1, id, id2 , . . . to obtain a sequence of orthogonal polynomials satisfying deg pk D k for every k D 0; 1; : : : : 49

288

8 Spaces of Continuous Functions

In case of a positive answer we would obtain a natural proof of the Weierstrass approximation theorem. But the answer is negative: Proposition 8.21 50 (a) (Nikolaev) For any given weight function there exists an f 2 C.J/ such that P . f ; pn /pn does not converge uniformly to f on I. (b) (Faber)51 For any given point system .xm;k / there exists an f 2 C.I/ such that Lm f does not converge uniformly to f on I.

Proof (a) The continuous linear projections Lm f WD

m X . f ; Pn /Pn nD0

satisfy the conditions of Theorem 8.20. (b) These operators Lm also satisfy the conditions of Theorem 8.20.

t u

Remarks Historically, the theorems of du Bois Reymond and Faber paved the way to the discovery of the Banach–Steinhaus theorem. Let us mention three further results related to Faber’s theorem. • (Fejér)52 Let us choose for xm;0 ; : : : ; xm;m 2 Œ1; 1 DW I the zeros of the corresponding Chebyshev polynomial, and for f 2 C.I/ let Hm f denote the Hermite interpolation polynomial of degree 2m C 1, satisfying the equalities .Hm f /.xm;k / D f .xm;k / and .Hm f /0 .xm;k / D 0. Then Hm f converges uniformly to f . • (Erd˝os–Turán)53 If w is a weight function on I and xm;0 ; : : : ; xm;m are the zeros of the corresponding mth orthogonal polynomial, then Lm Rf converges to f in the weaker norm associated with the scalar product .p; q/ WD I pqw dt. • (Erd˝os–Vértesi)54 For any given system of points xm;k there exists a function f 2 C.I/ such that lim sup jLn f .x/j D 1 for almost every x 2 I. Not only do we not have uniform convergence, but we even have divergence almost everywhere!

50

Nikolaev [346]. However, we will see later (Corollary 9.6, p. 314) that the answer is affirmative for the weaker norm associated with the scalar product. 51 Faber [133]. 52 Fejér [142]; see also Cheney [85]. In this way, Hermite interpolation can be used to prove the Weierstrass approximation theorem. 53 Erd˝os–Turán [124]. 54 Erd˝os–Vértesi [125].

8.8 * Dual Space. Riesz Representation Theorem

289

8.8 * Dual Space. Riesz Representation Theorem Let K be a compact Hausdorff space. Using measure theory we may characterize the dual of C.K/. Definition Let us denote by B the smallest -ring containing all sets of the form ff D 0g, where f runs over C.K/. The elements of B are called Baire sets.55 Remarks • B is even a -algebra. Moreover, if g 2 C.K/ and c 2 R, then the level sets fg D cg ;

fg cg ;

fg cg

fg ¤ cg ;

fg > cg ;

fg < cg

and their complements

are also Baire sets, because fg D cg D fg c D 0g ; ˚ fg cg D .g c/C D 0 and fg cg D f.g c/ D 0g : • In fact, B contains all open, closed or compact sets of K. This follows from the Tietze–Urysohn theorem of topology because every compact Hausdorff space is normal. See, e.g., Kelley [247]. Definition By a (signed) Baire measure we mean a finite (signed) measure defined on B. Examples For any fixed a 2 K the Dirac measure at a is a Baire measure. The Baire measures have an important regularity property: they may be well approximated by both open and closed sets: Proposition 8.22 Let be a Baire measure, A 2 B and " > 0. There exist a closed set F and an open set G in B such that F A G and .G n F/ < ":

55

Baire [17].

(8.12)

290

8 Spaces of Continuous Functions

Proof Let us denote temporarily by BQ the family of Baire sets having the property (8.12). We have to show that BQ is a -algebra containing all sets f f D 0g with f 2 C.K/. If A D f f D 0g for some f 2 C.K/, then the formulas F WD A;

Gn WD fj f j < 1=ng ;

n D 1; 2; : : :

define a closed set F 2 B and open sets Gn 2 B satisfying F A Gn for all n. Since the set sequence .Gn / is non-increasing and 1 \

.Gn n F/ D

nD1

1 \

f0 < j f j < 1=ng D ¿;

nD1

Proposition 7.3 (p. 216) implies that .Gn n F/ < " if n is sufficiently large. It remains to prove the -algebra property. Choosing the constant functions f D 0 and f D 1 we see that K and ¿ belong to B. Moreover, since they are both open and closed, they belong to BQ as well: we may choose F D G D ¿ and F D G D K. Q then K n A 2 B. Q Indeed, if F and G satisfy (8.12), then K n G is closed, If A 2 B, K n F is open, both belong to B, K nG K nA KnF

and .K n F/ n .K n G/ D .G n F/ < ":

Q then A WD [ An 2 B. Q For the proof, Finally, if .An / is a disjoint sequence in B, for any fixed " > 0 we choose closed sets Fn 2 B and open sets Gn 2 B such that F n A n Gn

and .Gn n Fn / < 2n1 "

N WD [NnD1 Fn are closed for all for all n. Then G WD [1 nD1 Gn is open, the sets F N N D 1; 2; : : : ; all belong to B, F A G, and

.G n F N /

N X

X " X .Gn n Fn / C .Gn / < C .Gn /: 2 n>N n>N nD1

Since 1 X nD1

.Gn /

0 we have to find a positive integer N such that k fn k1 < " for all n N. For each t 2 K there exists an index nt such that fnt .t/ < "; by continuity the inequality fnt < " remains valid in some open neighborhood Vt of t. Since K is compact, a finite number of such neighborhoods, say Vt1 ; : : : ; Vtm , already cover K. Choose N WD max fnt1 ; : : : ; ntm g, let n N, and consider a point s 2 K. Then s belongs to some neighborhood Vti , and therefore 0 fn .s/ fnti .s/ < " by the non-increasingness of the sequence . fn /.

t u

Dini [109, Sect. 99]. See the graphs of the functions fn .t/ WD tn for n D 1; 2; 3 in Fig. 8.15, and let K D Œ0; a, 0 < a < 1.

59

8.8 * Dual Space. Riesz Representation Theorem

293

Fig. 8.16 An “interval” Œ f ; g/

g

f

Lemma 8.25 For each positive linear functional ' W C.K/ ! R there exists a Baire measure 2 M.K/ such that ' D j. Proof Following Kindler60 we introduce the “intervals” Œ f ; g/ WD f.x; t/ 2 K R W f .x/ t < g.x/g for all functions f ; g 2 C.K/ satisfying f g.61 They form a semiring P in K R,62 and the formula

.Œ f ; g// WD '.g f / defines a finite, additive set function on P, satisfying .¿/ D 0. This set function is also -additive, and hence a measure. For the proof we consider an arbitrary countable decomposition Œ f ; g/ D [ Œ fn ; gn /. We have Œ f .x/; g.x// D [ Œ fn .x/; gn .x// for each x 2 K, and therefore g.x/ f .x/ D

1 X

gn .x/ fn .x/;

nD1

because the length of ordinary intervals is a measure.

60

Kindler [248]. See Fig. 8.16. 62 The proof is similar to that of ordinary intervals. 61

294

8 Spaces of Continuous Functions

Setting hm WD g f

m X .gn fn /;

m D 1; 2; : : :

nD1

we have hm & 0. By Dini’s theorem the convergence is uniform, and then '.hm / ! 0 by the continuity of '. This is equivalent to the -additivity relation

.Œ f ; g// D

1 X

.Œ fn ; gn //:

nD1

Applying Proposition 5.18 (p. 192) we extend to a measure defined on the -ring M of measurable sets, still denoted by .63 If f 2 C.K/ and c is a positive real number, then the set f f D 0g Œ0; c/ D

1 \

Œmin fn j f j ; cg ; c/

nD1

belongs to M. Since B is the smallest -algebra containing the sets f f D 0g, this implies that A 2 B H) A Œ0; 1/ 2 M: Consequently, the formula .A/ WD .A Œ0; 1// defines a Baire measure 2 M.K/.64 It remains to prove that '. f / D f 2 C.K/. Given f 2 C.K/, the continuous functions fn .x/ WD med f0; n. f .x/ 1/; 1g ;

x 2 K;

R

f d for all

n D 1; 2; : : :

form a non-decreasing sequence converging to the characteristic function f f >1g . Hence f f > 1g Œ0; c/ D

1 [

Œ0; cfn /

nD1

63

M is even a -algebra.

64

The finiteness follows from the relation .K/ D '.1/ < 1.

8.8 * Dual Space. Riesz Representation Theorem

295

for each positive number c, and therefore

.f f > 1g Œ0; c// D lim .Œ0; cfn // D lim '.cfn / n!1

n!1

D c lim '. fn / D c lim .Œ0; fn // n!1

n!1

D c .f f > 1g Œ0; 1// D c.f f > 1g/: By the additivity of the measures and this implies the more general relations

.fa < f bg Œ0; c// D c.fa < f bg/

(8.13)

for all numbers 0 < a < b.65 R Now we use (8.13) to prove the equalities '. f / D f d. Separating the positive and negative parts of f we may assume that f 0. Then the “interval” Œ0; f / is the union of the non-decreasing sequence of sets h n2

X i iC1 i 0; ; < f 2n 2n 2n iD1 n

Bn WD and therefore

'. f / D .Œ0; f // D lim .Bn / n!1

Z

iC1 i X i D lim n D f d: 0 arbitrarily, and consider the Hahn decomposition K D P [ N of . By Proposition 8.22 (p. 289) there exist two disjoint closed sets P0 P and N 0 N satisfying ˇ ˇ ˇ.P n P0 /ˇ < "

and

ˇ ˇ ˇ.N n N 0 /ˇ < ":

1

if t 2 P0 ,

1

if t 2 N 0

The function ( g.t/ WD

is clearly continuous on P0 [ N 0 . Applying Proposition 8.6 (p. 267), g may be extended to a function f 2 C.K/. Changing f to med f1; f ; 1g if necessary, we may also assume that j f j 1 on K.68 Then k f k 1, and kjk . j/. f / Z Z Z D f d C f d C P0

68

N0

Z PnP0

f d C

If K is metrizable, then we may define f explicitly by the formula f .t/ WD

dist.t; N 0 / dist.t; P0 / : dist.t; N 0 / C dist.t; P0 /

NnN 0

f d

298

8 Spaces of Continuous Functions

.P0 / .N 0 / 2" .P/ .N/ 4" D kk 4": Letting " ! 0 we conclude that kjk kk.

t u

Example Using Theorem 8.23 we may prove directly the non-reflexivity of C.Œ0; 1/.69 Given any 2 M.K/ with K WD Œ0; 1, the formulas m.t/ WD .Œ0; t/;

t 2 Œ0; 1

and ˆ./ WD

X

m.tC/ m.t/

0 0,

g.t/ " h.t/ WD "

is a finite linear combination of characteristic functions of measurable sets, satisfying the inequality kg hk1 ". t u Now we prove the L2 version of the Hilbert–Schmidt theorem (p. 38). Similarly to Sect. 7.3 (p. 224) we consider a product measure on X X. Proposition 9.4 (Hilbert–Schmidt)12 If a 2 L2 .X X/, then the formula Z a.t; s/f .s/ ds;

.Af /.t/ WD

t2X

X

defines a completely continuous operator in L2 .X/. Proof Using the Cauchy–Schwarz inequality and applying Tonelli’s theorem (p. 228), the following estimate holds for all f 2 L2 .X/: Z Z Z ˇZ ˇ2 Z ˇ ˇ 2 ja.t; s/j ds jf .s/j2 ds dt ˇ a.t; s/f .s/ dsˇ dt X

X

X

D

kak22

X

X

kf k22

:

Hence A is a continuous operator on L2 .X/, and kAk kak2 .13 To prove the compactness, in view of Proposition 2.37 (p. 101), it is sufficient construct a sequence .An / of continuous operators of finite rank on L2 .X/, satisfying kA An k ! 0.

12 13

Hilbert [209], Schmidt [415]. We even have equality here.

312

9 Spaces of Integrable Functions

Applying Proposition 9.3 we choose a sequence .an / of step functions satisfying an ! a in L2 .X X/, and we define Z .An f /.t/ WD

an .t; s/f .s/ ds;

f 2 L2 .X/;

t 2 X:

X

Repeating the above estimates with an and a an instead of a, we obtain that the operators An are continuous in L2 .X/, and that kA An k ka an k2 ! 0: It remains to show that each An has a finite rank. For this we observe that, by the definition of the product measure, each step function an on X X is of the form an .t; s/ D

N X

Ji .t/ Ki .s/

iD1

with some sets Ji ; Ki 2 M of finite measure, and hence the range of An is generated by the N functions K1 ; : : : ; KN . t u The rest of this section is devoted to the study of some important special cases. Let I be an open interval and w W I ! R a nonnegative measurable function with respect to the usual Lebesgue measure. Assume that w is integrable on every compact subinterval of I,14 and denote by P the Rsemiring of bounded intervals whose closures are in I. Then the formula .J/ WD J w dt defines a finite measure on P. Consider the corresponding integral, and denote by Lpw the corresponding Lp spaces. For w D 1 this reduces to the usual Lp .I/ spaces. We denote by Cc .I/ the vector space of continuous functions g W I ! R that vanish outside some compact subinterval of I, i.e., vanish in some neighborhood of the endpoints of I.15 Proposition 9.5 Let 1 p < 1. (a) Lpw is separable. (b) Cc .I/ is dense in Lpw .16 (c) If I is bounded and w is integrable on I, then the algebraic polynomials are dense in Lpw . (d) If jIj 2 and w is integrable in I, then the trigonometric polynomials are dense in Lpw .

14

We say in such cases that w is locally integrable. The compact subinterval may depend on g. 16 Moreover, the proof will show that for each f 2 Lpw there exists a function h 2 Lpw and a sequence .'n / Cc .I/ satisfying the relations (9.1) of Proposition 9.3. 15

9.1 Lp Spaces, 1 p 1

313

Fig. 9.1 Graph of gn

1

1 a a +− n

1 b−− n b

Proof We denote by kkp the norm of Lpw . (a) By Proposition 9.3 the characteristic functions of the intervals in P generate Lpw . If we consider only the intervals with rational endpoints, then we obtain countably many functions that still generates Lpw . (b) By Proposition 9.3 it is sufficient to find for each fixed compact interval J D Œa; b I a sequence of functions .gn / Cc .I/ converging to J in Lpw . The formulas 8 ˆ 0 if t a, ˆ ˆ ˆ ˆ ˆ n.t a/ if a t a C n1 , ˆ < gn .t/ WD 1 if a C n1 t b n1 , ˆ ˆ ˆ ˆ n.b t/ if b n1 t b, ˆ ˆ ˆ :0 if t b for n > 2=.b a/ yield such a sequence (see Fig. 9.1). Indeed, Z k J gn kpp D

b

j1 gn .t/jp w.t/ dt ! 0 a

by the dominated convergence theorem, because gn ! 1 a.e. in Œa; b, 0 j1 gn jp w w for all n, and w is integrable.

314

9 Spaces of Integrable Functions

(c) Given any f 2 Lpw and " > 0, using (b) we choose g 2 Cc .I/ such that kf gkp < "=2. Then applying the first approximation theorem of Weierstrass (p. 260) we choose a sequence .pn / of polynomials satisfying kg pn k1 ! 0. Since kf pn kp kf gkp C kg pn kp

0, we choose a step function ' satisfying kf 'kp < ". Then we also have kfh 'h kp < " for all h. If h is sufficiently close to zero, then k' 'h kp < ", and therefore kf fh kp kf 'kp C k' 'h kp C k'h fh kp < 3": Second step. If F is totally bounded, then for each fixed " > 0 it can be covered by finitely many balls of radius ". Let us denote by f1 ; : : : ; fm the centers of these balls. By the first step there exists R > 0 and ı > 0 such that kfi kLp .RnŒR;R/ < " and kfi fi;h kp < " if for i D 1; : : : ; m.

jhj < ı

318

9 Spaces of Integrable Functions

Each f 2 F belongs to one of the balls B" .fi /, so that kf kLp .RnŒR;R/ kf fi kLp .RnŒR;R/ C kfi kLp .RnŒR;R/ < 2" and kf fh kp kf fi kp C kfi fi;h kp C kfi;h fh kp < 3" if jhj < ı.

t u

Proof of the Sufficiency First step. Applying Steklov’s regularization method28 we reduce the problem to the case of continuous functions. Setting .Sr f /.t/ WD

1 r

Z

r 0

f .t C s/ ds;

f 2 Lp ; r > 0;

first we establish the following estimates: kSr f k1 r1=p kf kp I

(9.4)

j.Sr f /.t/ .Sr f /.t C h/j r1=p kf fh kp

(9.5)

kf Sr f kp sup kf fh kp :

(9.6)

for all t 2 R;

0 0 such that kf kLp .RnŒR;R/ < "

for all f 2 F :

Furthermore, using (9.3) and (9.6) we choose r > 0 such that kf Sr f kp < "

for all f 2 F :

Since F is bounded, by (9.4) and (9.5) the function system fSr f W f 2 F g is uniformly bounded and equicontinuous. Applying the Arzelà–Ascoli theorem (p. 268) on the interval ŒR; R, we obtain a finite number of continuous functions g1 ; : : : ; gm such that each f 2 F satisfies for some index i the inequalities jSr f gi j .2R/1=p "

in ŒR; R:

(9.7)

Extending the functions gi by zero to R, we obtain f1 ; : : : ; fm 2 Lp . To conclude we show that kf fi kp < 3" for every f 2 F , where the index i is the same as in (9.7). For the proof we use the triangle inequality, the definition of R and r, and finally the choice of i: kf fi kp D kf kLp .RnŒR;R/ C kf gi kLp .R;R/ < " C kf Sr f kLp .R;R/ C kSr f gi kLp .R;R/ < 2" C .2R/1=p kSr f gi kL1 .R;R/ 3":

t u

320

9 Spaces of Integrable Functions

9.3 * Convolution We have encountered integrals of the form Z f .s/g.t s/ ds many times: in the methods of Landau and de la Vallée-Poussin, in the closed forms of the Dirichlet and Fejér kernels in the preceding chapter, and in the Steklov functions in the preceding section. Such integrals often occur in the theory of partial differential equations and in harmonic analysis to prove density theorems.29 In this section we give only one basic result.30 Proposition 9.8 Let 1 p; q; r 1 satisfy the equality 1 1 1 C D C 1; p q r and let f 2 Lp .RN /, g 2 Lq .RN /. The formula Z .f g/.x/ WD

f .x y/g.y/ dy

defines a function f g 2 Lr .RN /, and kf gkr kf kp kgkq : If f vanishes outside A and g vanishes outside B, then f g vanishes outside ˚ A C B WD a C b 2 RN W a 2 A

and b 2 B :

Definition The function f g is called the convolution of f and g.31 Remarks • The definition shows that the convolution is commutative: f g D g f .

29

The latter applications are based on the celebrated Haar measure (Haar [178]), a natural generalization of the usual Lebesgue measure to topological groups. 30 There are many more results and applications in Brezis [65], Hörmander [218, 219], Katznelson [245], Pontryagin [364], Rudin [402, 405, 406], Schwartz [420], Weil [485]. 31 Fourier [148].

9.3 * Convolution

321

• It follows by induction on k that if f1 2 Lp1 .RN /; : : : ; fk 2 Lpk .RN / for some k 2, where 1 p1 ; : : : ; pk ; r 1 satisfy the equality 1 1 1 CC D C k 1; p1 pk r then g WD f1 . fk / / 2 Lr .RN / and kgkr kf1 kp1 kfk kpk : Moreover, the associativity relation .f g/ h D f .g h/ holds, so that we may remove the parentheses in the definition of g. The condition on the exponents is equivalent to the simpler relation 1 1 1 CC 0 D 0 p01 pk r where we use the conjugate exponents. Proof We proceed in several steps. (i) If the step functions 'n ; n converge a.e. to f and g, respectively in R, then the step functions 'n .xy/ n .y/ converge a.e. to f .xy/g.y/ in R2 ; the verification is left to the reader. Hence the function .x; y/ 7! f .x y/g.y/ is measurable. (ii) The case r D 1 of the theorem readily follows from Hölder’s inequality. Henceforth we assume that r < 1. Since p r and q r, then p and q are also finite. (iii) If f and g are nonnegative and integrable, then applying Tonelli’s theorem we obtain that Z Z Z .f g/.x/ dx D f .x y/g.y/ dy dx D D

Z Z

f .x y/g.y/ dx dy

Z Z

f .x y/ dx g.y/ dy

D kf k1 kgk1 < 1:

322

9 Spaces of Integrable Functions

Hence f g 2 L1 .RN /, and kf gk1 D kf k1 kgk1 :

(9.8)

(iv) Turning to the general case (f 2 Lp , g 2 Lq , r < 1), first we prove the following inequality: .jf j jgj/r kf krp kgkrq jf jp jgjq p q

a.e.

(9.9)

Introducing the conjugates p0 and q0 of p and q, we have 1 1 1 C 0 C D 1: p0 q r Since 1

1 1 1 p p Dp Dp 1 D 0 r p r q q

1

1 1 1 q q Dq Dq 1 D 0; r q r p p

and

the following equality holds a.e.: 1=q0 1=p0 1=r jg.y/jq jf .x y/jp jg.y/jq : jf .x y/g.y/j D jf .x y/jp Integrating with respect to y, applying Hölder’s inequality and using (iii) we obtain

ˇ1=r 0 0 ˇ kgkq=p ˇ jf jp jgjq .x/ˇ ; jf j jgj .x/ kf kp=q p q

or equivalently ˇ ˇ ˇ jf j jgj .x/ˇr kf krp=q0 kgkrq=p0 jf jp jgjq .x/: p q We conclude by observing that rp=q0 D r p and rq=p0 D r q. (v) The right-hand side of (9.9) is integrable by (iii). Hence jf j jgj 2 Lr .RN /, i.e., Z Z

jf .x y/g.y/j dy

r

dx < 1:

9.4 Uniformly Convex Spaces

323

Applying this to the positive and negative parts of f and g we conclude that the four functions y 7! f˙ .x y/g˙ .y/ are integrable for a.e. x. Hence their linear combination y 7! f .x y/g.y/ is also (measurable and) integrable for a.e. x. Therefore f g is well defined a.e. Next, applying (9.8) and (9.9) we obtain the following estimate: Z j.f g/.x/jr dx D

Z ˇZ ˇr ˇ ˇ ˇ f .x y/g.y/ dyˇ dx Z Z

jf .x y/g.y/j dy

r dx

D kjf j jgjkrr kf krp kgkrq kjf jp jgjq k1 p q D kf krp kgkrq : Hence f g 2 Lr .RN / and kf gkr kf kp kgkq . (vi) If .f g/.x/ is defined for some x … A C B, then x y … A for all y 2 B. Consequently, f .x y/g.y/ D 0 for a.e. y 2 RN , whence .f g/.x/ D 0. u t

9.4 Uniformly Convex Spaces The parallelogram identity is an important property of Euclidean spaces. For 1 < p < 1 the Lp spaces have a weaker, but still useful property: Definition A normed space X is uniformly convex32 if for each " > 0 there exists a ı > 0 such that if two vectors x; y 2 X satisfy the inequalities kxk 1; kyk 1 and

32

Clarkson [89].

kx C yk > 2 ı;

324

9 Spaces of Integrable Functions

x+y

Fig. 9.2 Uniform convexity

x y

then kx yk < ": (See Fig. 9.2.) It follows from the definition that every uniformly convex space is strictly convex (see p. 67). Examples • Every Euclidean space is uniformly convex. Indeed, since kx yk2 D 2 kxk2 C 2 kyk2 kx C yk2 < 4 .2 ı/2 < 4ı; we may choose ı WD "2 =4 for each ". • The space `1 is not uniformly convex, because ke1 k D ke2 k D 1

and

ke1 C e2 k D ke1 e2 k D 2;

so that for " < 2 there is no suitable ı > 0. • The space `1 is not uniformly convex either, because the vectors x WD e1 C e2 and y WD e1 e2 satisfy kxk D kyk D 1

and

kx C yk D kx yk D 2;

so that for " < 2 there is no suitable ı > 0.

9.4 Uniformly Convex Spaces

325

On the other hand, `p is uniformly convex if 1 < p < 1. More generally: Proposition 9.9 Let .X; M; / be an arbitrary measure space and 1 < p < 1. Then Lp .X; M; / is uniformly convex.33 Proof First step. If x and y are distinct real numbers, then ˇ x C y ˇp jxjp C jyjp ˇ ˇ ˇ < ˇ 2 2 by the strict convexity of the function t 7! jtjp . Second step. For each " 2 .0; 21p we denote by % D %."/ the minimum of the function jxjp C jyjp ˇˇ x C y ˇˇp ˇ ˇ 2 2 on the non-empty34 compact set n

.x; y/ 2 R2 W jxjp C jyjp D 2

ˇ x y ˇp o ˇ ˇ and ˇ ˇ " : 2

By the preceding step we have % > 0. By homogeneity it follows that if x; y 2 R satisfy the inequality ˇ x y ˇp jxjp C jyjp ˇ ˇ ; ˇ ˇ " 2 2 then %

jxjp C jyjp jxjp C jyjp ˇˇ x C y ˇˇp ˇ ˇ: 2 2 2

Third step. For any given " > 0 we have to find ı > 0 such that if two functions f ; g 2 Lp satisfy the inequalities Z

Z jf j dx 1; p

33 34

jgj dx 1 and p

Z ˇ ˇ ˇ f C g ˇp ˇ dx > 1 ı; ˇ 2

Clarkson [89]. The proof given here is due to McShane [320]. .21=p ; 0/ belongs to the set.

326

9 Spaces of Integrable Functions

then Z ˇ ˇ ˇ f g ˇp ˇ dx < 2": ˇ 2 We may assume that " 2 .0; 21p . Setting nˇ f g ˇp jf jp C jgjp o ˇ ˇ ; M WD ˇ ˇ " 2 2 applying the convexity of the function t 7! jtjp , and using the preceding step we obtain the following estimate: Z ˇ ˇ ˇ f g ˇp ˇ ˇ 2 X Z D

dx

Z ˇ ˇ ˇ f g ˇp ˇ ˇ ˇ f g ˇp ˇ dx C ˇ ˇ dx ˇ 2 2 XnM M Z Z jf jp C jgjp jf jp C jgjp dx C dx " 2 2 XnM M Z p Z 1 jf j C jgjp ˇˇ f C g ˇˇp jf jp C jgjp dx C ˇ " ˇ dx 2 % M 2 2 XnM Z Z 1 jf jp C jgjp ˇˇ f C g ˇˇp jf jp C jgjp dx C ˇ " ˇ dx 2 % X 2 2 X

"C

1 1ı % %

D"C

ı : %

We conclude by choosing ı < "%.

t u

The following variant of the orthogonal projection (p. 12) is valid in all uniformly convex Banach spaces: Proposition 9.10 (Sz.-Nagy)35 Let K be a non-empty convex closed set in a uniformly convex Banach space X. For each x 2 X there exists in K a unique closest point y to x.

35

Sz.-Nagy [447].

9.4 Uniformly Convex Spaces

327

Proof Existence. The result is obvious if x 2 K. Henceforth we assume that x … K, and we choose a minimizing sequence: .yn / K, and kx yn k ! d WD dist.x; K/: Setting tn WD 1= kx yn k

and zn WD tn .x yn /;

we have kzn k D 1 for every n. Furthermore, applying the convexity of K and the definition of d we obtain the following relation: kzn C zm k D ktn .x yn / C tm .x ym /k t tm n yn C ym D .tn C tm /x tn C tm tn C tm .tn C tm /d ! 2: By the uniform convexity this implies that .zn / is a Cauchy sequence; since, moreover, X is complete, it converges to some point z 2 X. Consequently, yn D x

zn ! x dz DW y: tn

Hence y 2 K because K is closed, and kx yk D lim kx yn k D d. Uniqueness. If y; y0 2 K and kx yk D kx y0 k D d, then the formulas y2n1 WD y and y2n WD y0 , n D 1; 2; : : : define a minimizing sequence. This sequence is convergent by the preceding step, but this is possible only if y D y0 . t u *Examples The spaces L1 and L1 do not always have the property of the last proposition, so they are not uniformly convex. • Consider in X D L1 .1; 1/ the closed subspace M formed by the functions having integral zero, and the constant function g D 1. If f 2 M, then Z kg f k1 D

1 1

Z j1 f .t/j dt

1 1

1 f .t/ dt D 2;

with equality for all f 2 M satisfying f 1. Therefore the distance dist.g; M/ D 2 is attained at infinitely many points.

328

9 Spaces of Integrable Functions

• Consider in X D L1 .1; 1/ the closed subspace M formed by the functions vanishing a.e. on Œ1; 0, and the constant function g D 1. We have kg f k1 kg f kL1 .1;0/ D 1 for all f 2 M, with equality whenever 0 f 2. Therefore the distance dist.g; M/ D 2 is attained at infinitely many points. In uniformly convex spaces we may complete Proposition 2.22 (p. 80) on the relation between strong and weak convergence: *Proposition 9.11 (Radon–Riesz)36 In uniformly convex spaces we have xn ! x ” xn * x

and

kxn k ! kxk :

Proof The implication H) holds in all normed spaces by Proposition 2.22 (p. 80). The converse implication is obvious if x D 0. Assume henceforth that kxk > 0, then kxn k > 0 for all sufficiently large n. The assumptions xn * x and kxn k ! kxk imply that xn x x C *2 : kxn k kxk kxk Since the norm of the limit is equal to 2, x x n C lim inf 2 kxn k kxk by Proposition 2.22 (f).37 By the definition of uniform convexity this implies that x x n ! 0: kxn k kxk Consequently, xn D kxn k

x xn ! kxk D x: kxn k kxk

t u

Hildebrandt [210] (`p ), Radon [366] (p. 1358: Lp ), Riesz [382] (pp. 58–59: `p ), Riesz [385] (simple proof for Lp ). 37 In fact, the left-hand norm converges to 2. 36

9.5 Reflexivity

329

*Remarks • We recall (p. 83) that the equivalence fails, for example, in c0 and `1 . • We also recall that `1 , although not uniformly convex, has the Radon–Riesz property: see Proposition 2.26, p. 84. • The preceding example is an exception: we will soon show (p. 338) that L1 .; / does not have the Radon–Riesz property. • By a theorem of Kadec38 every separable Banach space has an equivalent norm having the Radon–Riesz property.

9.5 Reflexivity Unlike the spaces C.K/, most Lp spaces are reflexive: Proposition 9.12 (Clarkson)39 For any given measure space .X; M; /, Lp .X; M; / is reflexive for all 1 < p < 1. In view of Proposition 9.9 it suffices to establish the following result: Proposition 9.13 (Milman–Pettis)40 Every uniformly convex Banach space is reflexive. *Remark This result clarifies the relationship between Proposition 2.31 (c) and Proposition 9.10 (pp. 91 and 326) on the distance from closed convex sets. Proof 41 Consider the canonical isometry J W X ! X 00 of Proposition 2.28 (p. 87). Since J is homogeneous, it is sufficient to show that if ˆ 2 X 00 and kˆk D 1, then there exists an x 2 X satisfying Jx D ˆ. Denote the closed unit balls of X and X 00 by B and B00 . By Goldstein’s theorem (p. 139) there exists a net .xn / in B such that J.xn / ! ˆ in the topology .X 00 ; X 0 /. It follows that the “doubled” net converges to 2ˆ: J.xm C xn / D J.xm / C J.xn / ! 2ˆ: Consequently, kxm C xn k ! k2ˆk D 2:

38

Kadec [234–236]. See also Bessaga–Pelczýnski [40]. Clarkson [89]. 40 Milman [322], Pettis [358]. 41 We follow Lindenstrauss–Tzafriri [303, p. 61]. See, e.g., Brezis [65] for a proof without using nets. 39

330

9 Spaces of Integrable Functions

Indeed, in the contrary case there would exist a subnet belonging to the ball ˛B00 for some 0 < ˛ < 2. This ball would be compact by the Banach–Alaoglu theorem (p. 139), and hence closed in the Hausdorff topology .X 00 ; X 0 /. This would imply k2ˆk ˛ < 2, contradicting the choice of ˆ. Since X is uniformly convex, the relation kxm C xn k ! 2 implies that .xn / is a Cauchy net in X. Since X is complete, it converges to some point x 2 X. Then J.xn / ! J.x/ in .X 00 ; X 0 / by the definition of this topology. But we also have J.xn / ! ˆ, so that ˆ D J.x/ by the uniqueness of the limit. t u The spaces L1 and L1 are not reflexive in general: *Examples • We have seen several proofs of the non-reflexivity of C.Œ0; 1/ in the preceding chapter. Since it is a closed subspace of L1 .0; 1/, by Proposition 3.23 (p. 143) L1 .0; 1/ cannot be reflexive either. • The space L1 .0; 1/ is not reflexive, because there exist linear functionals ' 2 .L1 .0; 1//0 whose norms are not attained.42 For example, let Z '.f / WD

1 0

f 2 L1 .0; 1/:

tf .t/ dt;

The inequalities Z j'.f /j

1 0

Z t jf .t/j dt

1 0

jf .t/j dt D kf k1

(9.10)

imply that k'k 1. Furthermore, the functions (see Fig. 9.3) fn WD n Œ1n1 ;1 have unit norm in L1 .0; 1/, and j'.fn /j ! 1, so that k'k D 1. But this norm is not attained, because the second inequality in (9.10) is strict for every non-zero function. • The non-reflexivity of L1 .X; M; / for most measure spaces also follows from the existence of bounded sequences with no weakly converging subsequences. (See Theorem 2.30, p. 90.) More precisely, if there exists a disjoint set sequence .An / such that 0 < .An / < 1 for all n, then the functions fn WD .An /1 An form a bounded sequence having no weakly converging subsequences.

42

See Proposition 2.1, p. 55.

9.6 Duals of Lp Spaces

331

Fig. 9.3 Graph of n Œ1n1 ;1

n

n−1 n

1

Indeed, for any given subsequence .fnk / consider the linear functional defined by the formula '.f / WD

1 X

Z .1/

k

f d: A nk

kD1

Then the numerical sequence .'.fnk // D ..1/k / is divergent. We return to the question of reflexivity at the end of the next section.

9.6 Duals of Lp Spaces In this section we generalize the relations .`p /0 D `q of Proposition 2.15, p. 73). If p; q 2 Œ1; 1 are conjugate exponents, then the formula Z .jg/.f / WD

fg d X

defines a continuous linear functional on Lp for each g 2 Lq . Indeed, the integrals are well defined by Hölder’s inequality, and j.jg/.f /j kgkq kf kp : Since jg is clearly linear, hence jg 2 .Lp /0

and

kjgk kgkq :

332

9 Spaces of Integrable Functions

This computation also shows that j W Lq ! .Lp /0 is a continuous linear map of norm 1. Theorem 9.14 Let .X; M; / be an arbitrary measure space, and p; q 2 Œ1; 1 two conjugate exponents. (a) The linear map j W Lq ! .Lp /0 is an isometry.43 (b) (Riesz)44 If 1 < p < 1, then j W Lq ! .Lp /0 is an isometric isomorphism. (c) (Steinhaus)45 If is strongly -finite, then j W L1 ! .L1 /0 is an isometric isomorphism. Proof (a) It remains only to prove the inequality kjgk kgkq .46 We may therefore assume that kgkq > 0. If 1 < p < 1, then the function f WD jgjq1 sign g satisfies the equalities Z kf kpp D

Z jf jp d D

Z jgjp.q1/ d D

jgjq d D kgkqq D kgkp.q1/ : q

Hence f 2 Lp ;

kf kp D kgkqq1 > 0;

and Z .jg/.f / D

jgjq d D kgkqq D kgkq kf kp :

Since kf kp > 0, we conclude that kjgk kgkq . If p D 1, then setting f WD sign g 2 L1 we have Z kgk1 D

jgj d D .jg/.f / kjgk kf k1 D kjgk :

In the case p D 1 it is essential for the existence of B that the functions in L1 are measurable by our definition, and not only locally measurable. It is instructive to consider on an uncountable set X the measure that is equal to zero on countable sets, and equal to 1 otherwise. This is another reason in favour of the constructive measurability definition adopted in this book. 44 Riesz [380] for X D Œ0; 1, Nikodým [343], McShane [320]. 45 Steinhaus [432] for X D Œ0; 1, Dunford [116]. 46 See also a direct proof for X D R in Riesz and Sz.-Nagy [394]. 43

9.6 Duals of Lp Spaces

333

Finally, if p D 1, then for any fixed number 0 < c < kgk1 the set A WD fx 2 X W jg.x/j cg has a positive measure. Applying Lemma 7.5 (p. 220) there exists a B A satisfying 0 < .B/ < 1. Then f WD B sign g 2 L1 , and Z c.B/

fg d D .jg/.f / kjgk kf k1 D kjgk .B/:

Hence c kjgk for all c < kgk1 , so that kgk1 kjgk. (b) We have to prove that j is onto. Since j is an isometry and Lq is complete, the range R.j/ of j is a closed subspace of .Lp /0 . It remains to show that it is dense in .Lp /0 . By Corollary 2.9 (p. 64) it suffices to show that if ˆ 2 .Lp /00 is orthogonal to R.j/ .Lp /0 , then ˆ D 0. Since Lp is reflexive, identifying .Lp /00 with Lp R p this is equivalent to the following property: if f 2 L and fg d D 0 every g 2 Lq , then f D 0. Setting g WD jf jp1 sign f and repeating the computation of (a), reversing the role of p and q, we obtain that Z Z q g 2 L and 0 D fg d D jf jp d: Hence f D 0 a.e. * (c) Given ' 2 .L1 /0 we have to find g 2 L1 satisfying Z '.f / D

fg d

(9.11)

X

for all f 2 L1 .47 First we assume that .X/ < 1. Then the formula

.A/ WD '. A / defines a set function on M. It is finitely additive by the linearity of '. Moreover, it is -additive. Indeed, if A D [ An with A; An 2 M, then P An D A in L1 by Corollary 5.9 (p. 180). Using the continuity of ' 2 .L1 /0

47

The following reasoning may be adapted for 1 < p < 1 as well: see Dunford–Schwartz [117].

334

9 Spaces of Integrable Functions

we conclude that

.A/ D '. A / D

X

'. An / D

X

.An /:

Observe that . Indeed, if .A/ D 0, then A D 0 a.e., and hence

.A/ D '. A / D 0: Applying the Radon–Nikodým theorem (p. 240) there exists a measurable function g such that Z

.A/ D

g d

(9.12)

A

for every set A of finite measure. We show that g 2 L1 . Given any number 0 < c < kgk1 , at least one of the two sets fx 2 X W g.x/ cg

fx 2 X W g.x/ cg

and

has a positive measure, and then (as in the proof of (a)) it contains a set B of finite positive measure. If for example g c on B (the other case is analogous), then Z c.B/ B

g d D .B/ D '. B / k'k k B k1 D k'k .B/:

Hence c k'k for all c < kgk1 , so that kgk1 k'k .< 1/. We deduce from (9.12) by linearity that (9.11) is satisfied for all step functions f . Since they are dense in L1 by Proposition 5.14 (p. 185), by continuity (9.11) holds for all f 2 L1 , too. In the general case there exists a finite or countable disjoint sequence .Pn / such that 0 < .Pn / < 1 for all n, and .A/ D 0 for all A 2 M satisfying A X n [ Pn .48 Applying the preceding result for each Pn we obtain a function g 2 L1 vanishing outside [ Pn and satisfying (9.11) for the functions f D h Pn , h 2 L1 , n D 1; 2; : : : : Using the dominated convergence theorem, the linearity and the continuity of ', (9.11) follows again: '.h/ D

X

'.h Pn / D

XZ

Z h Pn g d D X

48

We use the strong -additivity assumption.

hg d: X

t u

9.6 Duals of Lp Spaces

335

*Remarks • Hildebrandt and Fichtenholz–Kantorovich characterized .L1 /0 .49 • The map j W L1 ! .L1 /0 is onto only in degenerate cases, for example when is the counting measure on a finite set. We have already seen (p. 79) that j W `1 ! .`1 /0 is not onto. The map j W L1 .R/ ! L1 .R/0 is not onto either because L1 .R/ is not even a dual space.50 This follows (similarly to the analogous result on c0 on p. 140) from the theorems of Banach–Alaoglu and Krein–Milman, because the closed unit ball of L1 .R/ has no extremal points. R For the last property we show that if jf j dx D 1, then there exists a non-zero R function g 2 L1 .R/ satisfying jf C tgj dx D 1 for all t 2 Œ1; 1. For this we first choose a set A of finite positive measure and a number " > 0 suchRthat f > " or f < " on A. Then we choose any non-zero function g such that g dx D 0, g D 0 outside A, and jgj < " on A. • Let us also give a direct proof of the non-surjectivity of the map j W L1 .R/ ! L1 .R/0 . The Dirac functional, defined by the formula ı.g/ WD g.0/;

g 2 Cb .R/

is a continuous linear functional of norm one on Cb .R/. Applying the Helly– Hahn–Banach theorem (p. 65) it can be extended to a continuous linear functional on L1 .R/. We claim that no function f 2 L1 .R/ satisfies the equality Z fg dt D g.0/

(9.13)

for all g 2 Cb .R/.51 Assume on the contrary that there exists such a function f . The formula gn .x/ WD min fn jxj ; 1g defines a sequence of functions in Cb .R/ satisfying gn .0/ D 0, fgn ! f a.e., and jfgn j jf j for all n. Applying the dominated convergence theorem it follows that Z

Z f dt D lim

fgn dt D lim gn .0/ D 0:

But this is impossible because choosing g D 1 in (9.13) we get

49

R

f dt D 1.

Hildebrandt [213, p. 875], Fichtenholz–Kantorovich [145, p. 76]. See also Dunford–Schwartz [117], Kantorovich–Akilov [243]. 50 This property and the following proof remain valid for all measure spaces where each set A of positive measure has a subset B satisfying 0 < .B/ < .A/. 51 This is an important theorem in the theory of distributions, asserting that the Dirac functional is not a regular distribution. See Schwartz [420].

336

9 Spaces of Integrable Functions

• In the preceding remark we have found a linear functional in L1 .R/0 not represented by any f 2 L1 .R/. Since L1 .R/0 D L1 .R/00 , this proves directly the non-reflexivity of L1 .R/. • Since L1 .R/0 D L1 .R/, by Proposition 3.23 (p. 143) L1 .R/ is not reflexive either. *Example We show that the strong -finiteness assumption cannot be omitted in Part (c).52 Consider the measure space .X; M; / and the measure of the counterexample on page 243. Since , we have Z

Z jf j d

jf j d D kf k1

for all f 2 L1 , so that the formula Z '.f / WD

f d

defines an element ' of .L1 /0 . We claim that ' is not represented by any (measurable or locally measurable) function g 2 L1 . Indeed, if we had Z

Z f d D

gf d

for all f 2 L1 , then (taking f D A for A 2 M) g would be a (measurable or locally measurable) Radon–Nikodým derivative of with respect to , contradicting our results on pp. 243 and 251.

9.7 Weak and Weak Star Convergence The purpose of this section is to characterize the weak and weak star convergence of Lp spaces. Since all weakly convergent and weak star convergent sequences are bounded by Propositions 2.24 and 3.18 (pp. 82 and 138), it is sufficient to consider bounded sequences.

52

See Schwartz [419] and Ellis–Snow [123] for the characterization of .L1 /0 in the general case.

9.7 Weak and Weak Star Convergence

337

Let p; q 2 Œ1; 1 be conjugate exponents, and let us denote by .Lp ; Lq / the locally convex topology on Lp , defined by the family of seminorms ˇZ ˇ ˇ ˇ pg .f / WD ˇ fg dˇ;

g 2 Lq :

If 1 < p < 1, then this is the weak topology of Lp . If our measure space is strongly -finite, then .L1 ; L1 / is the weak topology of L1 , and .L1 ; L1 / is the weak star topology of L1 . Proposition 9.15 Let .fn / be a bounded sequence in Lp , and f 2 Lp . (a) (Riesz)53 If 1 < p 1, then fn ! f in .Lp ; Lq / ” Z

Z fn d ! A

f d

(9.14)

A

for each set A of finite measure. (b) If p D 1, then fn ! f in .L1 ; L1 / ” (9.14) holds for all measurable sets A. *Remarks • If 1 < p 1, then using Proposition 9.3 (p. 310) the proof below shows that it suffices to consider in (9.14) the sets A of the semiring at the origin of the definition of the integral. Consequently, for the usual Lebesgue measure on an interval I R the condition (9.14) is equivalent to the pointwise convergence Fn ! F, where Fn and F are some primitives of fn and f that coincide at some fixed point of I. • Let .In / be a sequence of disjoint subintervals of an interval I D Œa; b such that jIn j > 0 and In .a; a C 2n / for every n. The formula fn WD jI2n1 j1 I2n1 jI2n j1 I2n defines a bounded sequence in L1 .I/ satisfying the relation Fn ! F of the preceding remark with F D f D 0. But fn does not converge to f in .L1 ; L1 / because (9.14) fails for A WD [I2n . • The functions fn WD Œn;nC1 in R show that it is not sufficient to consider sets of finite measure in (9.14) when p D 1. Proof of Proposition 9.15 Let us rewrite (9.14) in the form Z

Z A fn d !

53

Riesz [380] (for finite p).

A f d:

(9.15)

338

9 Spaces of Integrable Functions

If fn ! f in .Lp ; Lq /, then (9.15) is satisfied for all sets A with the indicated properties because A 2 Lq . The converse implications hold because the characteristic functions A of the indicated sets A generate Lq in all cases 9.3 (b), (c) (p. 310), and R by Proposition R because the functions g 2 Lq satisfying gfn d ! gf d form a closed subspace of Lq by the boundedness of the sequence .fn / (see Lemma 2.25, p. 83). t u We end this section by presenting a basic example of weak convergence. Given a sequence .n / of real numbers, tending to infinity, we consider the functions fn .t/ WD sin n t

and gn .t/ WD cos n t:

*Proposition 9.16 (Riemann–Lebesgue)54 Given any conjugate exponents p; q 2 Œ1; 1, we have fn ! 0 and gn ! 0 in .Lp ; Lq / on each bounded interval I. Proof The sequences .fn /; .gn / are bounded in L1 and hence in all spaces Lp .I/. Since Lq L1 , it is sufficient to prove the convergences in the topology .L1 ; L1 /. For any fixed point a 2 I, the primitives of the functions fn ; gn vanishing at a converge pointwise to zero, because ˇZ x ˇ ˇ cos a cos x ˇ 2 ˇ n n ˇ ˇ ˇ ! 0; sin n t dtˇ D ˇ ˇ ˇ j n nj a and a similar estimate holds for cos n t as well. We conclude by applying the first remark on the preceding page. t u *Remark In the special case where jIj D 2, p D 2 and n D n, the proposition follows from the Bessel inequality for the trigonometric system55 : the Fourier coefficients of each f 2 L2 .I/ converge to zero. *Example We recall56 that `1 has the Radon–Riesz property. On the other hand, L1 .; / does not have this property. Indeed, the functions hn .t/ WD 1 C sin nt converge weakly to h.t/ WD 1 in L1 .; / by the Riemann– Lebesgue lemma. Furthermore, Z khn k1 D

54

1 C sin nt dt D 2 D khk1

Riemann [371], Lebesgue [289, p. 473] and [293, p. 61]. See an interesting application of Poincaré [363] to the distribution of small planets. 55 Halphén [188]. 56 See the example preceding Proposition 2.26, p. 84.

9.8 Exercises

339

for every n. Nevertheless, hn does not converge strongly to h because Z khn hk1 D

jsin ntj dt D 4

for all n.

9.8 Exercises In the first seven exercises we consider the Hilbert space H D L2 .0; 1/ with the R1 scalar product .f ; g/ WD 0 fg dt. Exercise 9.1 (i) Show that every uniformly convergent sequence .xn / H also converges in H. (ii) Set xn .t/ WD n2 tent . Show that .xn / converges pointwise to 0 but it does not converge in H. (iii) Construct a sequence of continuous functions converging in H but diverging at each point. Exercise 9.2 Consider the following sets in H: (i) The set of functions x 2 H vanishing a.e. on some neighborhood of t D 1=2.57 (ii) The set of functions x 2 H with values in Œ1; 1. Are they convex? Are they closed? Exercise 9.3 (i) For each 2 R we denote by M the set of all continuous functions x 2 H satisfying x.0/ D . Show that the sets are convex, dense and disjoint. (ii) Show that the set of polynomials P vanishing at 1 is convex and dense in H. Exercise 9.4 Show that n

2

M WD f 2 L .0; 1/ W

Z

1 0

f .t/ dt D 0

o

is a closed subspace of L2 .0; 1/. Determine M ? . Exercise 9.5 The formula .Af /.t/ WD tf .t/ defines a continuous self-adjoint operator on the Hilbert space H D L2 .0; 1/ which has no eigenvalues. Exercise 9.6 There is no translation invariant measure in L2 .0; 1/ such that 0 < .A/ < 1 for all open balls.

57

The neighborhood may depend on x.

340

9 Spaces of Integrable Functions

Exercise 9.7 There exists a continuous, injective function f W Œ0; 1 ! L2 .0; 1/ such that the vectors f .b/ f .a/ and f .d/ f .c/ are orthogonal whenever 0 a < b < c < d 1. What is the geometric meaning of this property of the “curve” f ? Exercise 9.8 (Haar System)58 Set 8 ˆ ˆ p

Lq D Lp

and

\

Lp D Lq :

p m ! 1. Considering the linear functionals ' 2 c00 associated with the sequences ej we obtain that the only possible weak limit of .xn / is the constant sequence .1; 1; : : :/. Since it does not belong to c0 , .xn / does not converge weakly. (vi) Argue as in the last example of Sect. 2.5, p. 79. Exercise 2.18. The linearly independent subsets of X satisfy the assumptions of Zorn’s lemma, hence there exists a maximal linearly independent subset B. This is necessarily a basis of the vector space X. Choose an infinite sequence . fn / B, define '. fn / WD n j fn k for n D 1; 2; : : : ; and define '.x/ arbitrarily for x 2 B n f f1 ; f2 ; : : :g. Then ' extends to a unique linear functional W X ! R, and is not continuous. Exercise 2.19. If a normed space X has a countably infinite Hamel basis f1 ; f2 ; : : : ; then X is the union of the (finite-dimensional and hence) closed subspaces Vect f f1 ; : : : ; fn g, n D 1; 2; : : : : Since none of them has interior points, by Baire’s theorem X cannot be complete. Exercise 2.20.6 (i) For each 2 Œ0; / let S be the intersection of Z2 with an infinite strip of inclination and width greater than one. Each S is infinite, but the intersection of two such sets belongs to a bounded parallelogram and hence is finite. Since

5 6

This was an application of the Helly–Hahn–Banach theorem in the course. We present the proofs of Buddenhagen [67] and Lacey [276], respectively.

Hints and Solutions to Some Exercises

369

.0; 1/ Œ0; / and since there is a bijection between N and Z2 , the desired result follows. (ii) By the Helly–Hahn–Banach theorem there exist two sequences .xn / X and .'n / X 0 satisfying 'n .xk / ¤ 0 ” n D k. Then .xn / is linearly independent; moreover, no xn belongs to the closed linear span of the remaining vectors xm . We may assume by normalization that the sequence .xn / is bounded. Then the vectors X xn ; t 2 .0; 1/ 2n n2N t

form a linearly independent set of vectors, having 2@0 elements. Exercise 2.21. (i) Consider the sets Nt of the preceding exercise. Setting ( xtn

D

1

if n 2 Nt ,

0

otherwise

we obtain 2@0 linearly independent functions xt 2 `1 . Since `1 itself has 2@0 elements, its Hamel dimension is 2@0 . (ii) Fix a sequence of vectors x1 ; x2 ; : : : satisfying kxn k D dist .xn ; Vect fx1 ; : : : ; xn1 g/ D 3n ;

n D 1; 2; : : : ;

and define Ac WD

1 X

cn xn 2 X

nD1

for all c 2 `1 . These vectors are well defined because X is complete and 1 X

kcn xn k kck1

nD1

1 X

kxn k < 1:

nD1

It remains to show that Ac D 0 implies c D 0. We have for each positive integer N the following estimate: N 1 X X cn xn cn xn kAck nD1

nDNC1

370

Hints and Solutions to Some Exercises

jcN j 3N

1 X

jcn j 3n

nDNC1 1 X

jcN j 3N kck1

3n :

nDNC1

If Ac D 0, then jcN j kck1

1 X nD1

3n D

1 kck1 2

1 kck1 and thus c D 0. 2 Exercise 4.1. The set of continuous functions f W R ! R has the power 2@0 of R because it is determined by its values at rational points. The set of jump functions also has the power 2@0 . Consequently, the set of monotone functions has the power 2@0 . @ On the other hand, the set of null sets has the power of 22 0 > 2@0 . Exercise 4.2. It suffices to prove that the line y D x C ˛ meets C C for each ˛ 2 Œ1; 1. We recall that C D \Cn where each Cn is the disjoint union of 2n intervals of length 3n . Hence each Cn Cn is the disjoint union of 4n squares of side 3n . Prove that the line y D x C ˛ meets at least one of the squares in C1 C1 , say S1 . Next prove that y D x C ˛ meets at least one of the squares in C1 C1 , lying in S1 , say S2 . Construct recursively a decreasing sequence of squares S1 ; S2 ; : : : ; each meeting the line y D x C ˛. Exercise 4.7. ˛ > ˇ or ˛ D ˇ 0. Exercise 4.11. Apply Jordan’s theorem in (i), Cantor’s diagonal method in (ii) and (v), and use Proposition 4.2 (a), p. 153. Exercise 5.6. (i) There is a compact subset of positive measure. Apply the Cantor– Bendixson theorem. (ii) All subsets of Cantor’s ternary set are measurable. (iii) For otherwise A is countable. (iv) Apply Vitali’s method modulo 1. Exercise 5.7. See Rudin [404]. Exercise 6.1. (i) f is continuous and strictly monotone. (ii) The image of its complement is a union of intervals of total length one. (iii) Consider the inverse image of a non-measurable subset of f .C/. Exercise 6.2. (i) For ˛ D 0 we can take Cantor’s ternary set. For ˛ 2 .0; 1/ modify the construction by changing the length of the removed open intervals. (ii) Take A D [C˛n with a sequence ˛n ! 1. (iii) Take the complement of A. Exercise 7.2. Let .A/ D 0 if A is finite, and .A/ D 1 otherwise. for all N; therefore kck1

Hints and Solutions to Some Exercises

371

Exercise 7.3. If A R is a non-measurable set, then ˚ .x; x/ 2 R2 W x 2 A

(10.1)

is a two-dimensional null set. Exercise 7.5. See, e.g., Riesz and Sz.-Nagy [394] and Sz.-Nagy [448] for detailed proofs and applications to Fourier series and to the Riesz representation theorem 8.23 (p. 291). Exercise 7.6. ˛ > 0. Exercise 7.7. Consider in R the measure generated by the length of bounded subintervals of Œ0; 1/. Exercise 7.8. For example, let

f1 .x; y/ WD

8 ˆ ˆ m and hk .x/ WD eikx , then .Ts hk /.x/ D eiks hk .x/, and therefore Z

Z .Ts Lm Ts hk /.x/ ds D

eiks .Lm hk /.x s/ ds D 0

because Lm hk has order < k and thus is orthogonal to hk . Exercise 8.11. P (iv) If cm is theP first non-zero coefficient in cn fn , then fn .xm / D 0 for all n > m, and hence cn fn .xm / D cm fm .xm / D cm ¤ 0. Exercise 9.1. (iii) Modify Fréchet’s example (p. 307) by making the functions continuous. Exercise 9.3. (i) For each n D 1; 2; : : : we define fn 2 M such that fn D f in Œ1=n; 1, and fn is affine in Œ0; 1=n with fn .0/ D . Then k f fn k 2

jj C k f k1 : p n

374

Hints and Solutions to Some Exercises

(ii) First solution. Given f 2 H and " > 0 arbitrarily, first we choose g 2 H satisfying k f gk < " and vanishing in a neighborhood of 1, and then we choose a polynomial p such that kg pk1 < ". Then jp.1/j < ", and hence the polynomial P WD p p.1/ satisfies P.1/ D 0 and k f Pk k f gkCkg pkCkp Pk k f gkCkg pk1 Cjp.1/j < 3": Second solution. The linear functional '.P/ WD P.1/, defined on the linear subspace P of the polynomials is not continuous, because idn ! 0 for the norm of X, but '.idn / D 1 does not converge to '.0/ D 0. Therefore its kernel N.'/ is dense in P. Since P is dense in X by the Weierstrass approximation theorem, N.'/ is dense in X. Exercise 9.4. We have M D 1? and hence M ? D 1?? D Vect f1g is the linear subspace of constant functions. p Exercise 9.6. If .ek / is an orthonormal sequence and 0 < r 2=2, then the pairwise disjoint balls Br .ek / belong to the ball B1Cr .0/. Exercise 9.7. Set f .t/ D .0;t/ . Exercise 9.9. (iii) Consider the functions x.t/ WD t1=p

and x.t/ WD t1=q jln tj2=q :

Teaching Remarks

Functional Analysis • Most results of functional analysis and their optimality may be and are illustrated by the small `p spaces. • Although we assume that the reader is familiar with the basic notions of topology, we could not resist presenting a little-known beautiful short proof of the classical Bolzano–Weierstrass theorem, based on an elementary lemma of a combinatorial nature, perhaps due to Kürschák (p. 6). • We have included in the English edition a transparent elementary proof of the Farkas–Minkowski lemma, of fundamental importance in linear programming (p. 133), the Taylor–Foguel theorem on the uniqueness of Hahn–Banach extensions, and the Eberlein–Šmulian characterization of reflexive spaces. • The simple proof of Lemma 3.24 (p. 144) may be new. • Chapter 1 and the first seven sections of Chap. 2 may be covered in a onesemester course if we omit the material marked by . Chapter 3 may be treated later, in a course devoted to the theory of distributions. • It seems to be a good idea to treat the `p spaces only for 1 < p < 1 in the lectures, and to consider `1 , `1 , c0 later as exercises.

The Lebesgue Integral • For didactic reasons Chap. 5 is devoted to the case of functions f W R ! R. However, it is shown subsequently in Chap. 7 that all results and almost all proofs remain valid word for word in arbitrary measure spaces. This approach may lead to a better understanding of the theory without loss of time.

© Springer-Verlag London 2016 V. Komornik, Lectures on Functional Analysis and the Lebesgue Integral, Universitext, DOI 10.1007/978-1-4471-6811-9

375

376

Teaching Remarks

• Applying Riesz’s constructive definition of measurable functions we quickly arrive at essentially the most general forms of the Fubini–Tonelli and Radon– Nikodým theorems. For strongly -finite measures this is equivalent to the familiar inverse image definition. Otherwise the latter definition is weaker (in this book it is called local measurability), and, as we explain at the end of Sect. 7.7, the usual unpleasant counterexamples to some important theorems appear because of this weaker measurability notion. • A one-semester course could start with the definition of null sets and with Proposition 4.3 (p. 155), followed by Chaps. 5 and 7, except Sect. 7.7. We suggest, however, to state without proof two further classical theorems of Lebesgue on the differentiability of monotone functions and on the generalized Newton–Leibniz formula (pp. 157, 204), and to treat briefly the Lp spaces by following Sect. 9.1 (p. 305) in Function spaces.

Function Spaces • In order to make our exposition of functional analysis more accessible, we have avoided the spaces of continuous and Lebesgue integrable functions. This was anachronistic, because it was precisely the investigation of these spaces that led to the first great discoveries of functional analysis. Since they continue to play an important role in mathematics and its applications, we devote the last part of the book to these spaces. • Contrary to the preceding parts, we give several different proofs of various important theorems, in order to stress the multiple interconnections among different branches of analysis. • We present a large number of important examples that are not easy to localize in the literature.

Bibliography

1. N.I. Achieser, Theory of Approximation (Dover, New York, 1992) 2. N.I. Akhieser, I.M. Glazman, Theory of Linear Operators in Hilbert Space I-II (Dover, New York, 1993) 3. L. Alaoglu, Weak topologies of normed linear spaces, Ann. Math. (2) 41, 252–267 (1940) 4. P.S. Alexandroff, Einführung in die Mengenlehre und in die allgemeine Topologie (German) [Introduction to Set Theory and to General Topology]. Translated from the Russian by Manfred Peschel, Wolfgang Richter and Horst Antelmann. Hochschulbücher für Mathematik. University Books for Mathematics, vol. 85 (VEB Deutscher Verlag der Wissenschaften, Berlin, 1984), 336 pp 5. F. Altomare, M. Campiti, Korovkin-Type Approximation Theory and its Applications (De Gruyter, Berlin, 1994) 6. Archimedes, Quadrature of the parabola; [199], 235–252 7. C. Arzelà, Sulla integrazione per serie. Rend. Accad. Lincei Roma 1, 532–537, 566–569 (1885) 8. C. Arzelà, Funzioni di linee. Atti Accad. Lincei Rend. Cl. Sci. Fis. Mat. Nat. (4) 5I, 342–348 (1889) 9. C. Arzelà, Sulle funzioni di linee. Mem. Accad. Sci. Ist. Bologna Cl. Sci. Fis. Mat. (5) 5, 55–74 (1895) 10. C. Arzelà, Sulle serie di funzioni. Memorie Accad. Sci. Bologna 8, 131–186 (1900), 701–744 11. G. Ascoli, Sul concetto di integrale definite. Atti Acc. Lincei (2) 2, 862–872 (1875) 12. G. Ascoli, Le curve limiti di una varietà data di curve. Mem. Accad. dei Lincei (3) 18, 521– 586 (1883) 13. G. Ascoli, Sugli spazi lineari metrici e le loro varietà lineari. Ann. Mat. Pura Appl. (4) 10, 33–81, 203–232 (1932) 14. D. Austin, A geometric proof of the Lebesgue differentiation theorem. Proc. Am. Math. Soc. 16, 220–221 (1965) 15. V. Avanissian, Initiation à l’analyse fonctionnelle (Presses Universitaires de France, Paris, 1996) 16. R. Baire, Sur les fonctions discontinues qui se rattachent aux fonctions continues. C. R. Acad. Sci. Paris 126, 1621–1623 (1898). See in [18] 17. R. Baire, Sur les fonctions à variables réelles. Ann. di Mat. (3) 3, 1–123 (1899). See in [18] 18. R. Baire, Oeuvres Scientifiques, (Gauthier-Villars, Paris, 1990) 19. S. Banach, Sur les opérations dans les ensembles abstraits et leur application aux équations intégrales. Fund. Math. 3, 133–181 (1922); [26] II, 305–348

© Springer-Verlag London 2016 V. Komornik, Lectures on Functional Analysis and the Lebesgue Integral, Universitext, DOI 10.1007/978-1-4471-6811-9

377

378

Bibliography

20. S. Banach, An example of an orthogonal development whose sum is everywhere different from the developed function. Proc. Lond. Math. Soc. (2) 21, 95–97 (1923) 21. S. Banach, Sur le problème de la mesure. Fund. Math. 4, 7–33 (1923) 22. S. Banach, Sur les fonctionnelles linéaires I-II. Stud. Math. 1, 211–216, 223–239 (1929); [26] II, 375–395 23. S. Banach, Théorèmes sur les ensembles de première catégorie. Fund. Math. 16, 395–398 (1930); [26] I, 204–206 24. S. Banach, Théorie des opérations linéaires (Monografje Matematyczne, Warszawa, 1932); [26] II, 13–219 25. S. Banach, The Lebesgue Integral in Abstract Spaces; Jegyzet a [409] könyvben (1937) 26. S. Banach, Oeuvres I-II (Pa´nstwowe Wydawnictwo Naukowe, Warszawa, 1967, 1979) 27. S. Banach, S. Mazur, Zur Theorie der linearen Dimension. Stud. Math. 4, 100–112 (1933); [26] II, 420–430 28. S. Banach, H. Steinhaus, Sur le principe de la condensation de singularités. Fund. Math. 9, 50–61 (1927); [26] II, 365–374 29. S. Banach, A. Tarski, Sur la décomposition des ensembles de points en parties respectivement congruentes. Fund. Math. 6, 244–277 (1924) 30. R.G. Bartle, A Modern Theory of Integration. Graduate Studies in Mathematics, vol. 32 (American Mathematical Society, Providence, RI, 2001) 31. W.R. Bauer, R.H. Brenner, The non-existence of a banach space of countably infinite hamel dimension. Am. Math. Mon. 78, 895–896 (1971) 32. B. Beauzamy, Introduction to Banach Spaces and Their Geometry (North-Holland, Amsterdam, 1985) 33. P.R. Beesack, E. Hughes, M. Ortel, Rotund complex normed linear spaces. Proc. Am. Math. Soc. 75(1), 42–44 (1979) 34. S.K. Berberian, Notes on Spectral Theory (Van Nostrand, Princeton, NJ, 1966) 35. S.K. Berberian, Introduction to Hilbert Space (Chelsea, New York, 1976) 36. S.J. Bernau, F. Smithies, A note on normal operators. Proc. Camb. Philos. Soc. 59, 727–729 (1963) 37. M. Bernkopf, The development of functions spaces with particular reference to their origins in integral equation theory. Arch. Hist. Exact Sci. 3, 1–66 (1966) 38. D. Bernoulli, Réflexions et éclaircissements sur les nouvelles vibrations des cordes. Hist. Mém Acad. R. Sci. Lett. Berlin 9, 147–172 (1753) (published in 1755) 39. S.N. Bernstein, Démonstration du théorème de Weierstrass fondée sur le calcul de probabilités. Commun. Kharkov Math. Soc. 13, 1–2 (1912) 40. C. Bessaga, A. Pelczýnski, Selected Topics in Infinite-Dimensional Topology. Monografje Matematyczne, Tome 58 (Pa´nstwowe Wydawnictwo Naukowe, Warszawa, 1975) 41. F.W. Bessel, Über das Dollond’sche Mittagsfernrohr etc. Astronomische Beobachtungen etc. Bd. 1, Königsberg 1815, [43] II, 24–25 42. F.W. Bessel, Über die Bestimmung des Gesetzes einer priosischen Erscheinung, Astron. Nachr. Bd. 6, Altona 1828, 333–348, [43] II, 364–368 43. F.W. Bessel, Abhandlungen von Friedrich Wilhelm Bessel I-III, ed. by R. Engelman (Engelmann, Leipzig, 1875–1876) 44. P. Billingsley, Van der Waerden’s continuous nowhere differentiable function. Am. Math. Mon. 89, 691 (1982) 45. G. Birkhoff, E. Kreyszig, The establishment of functional analysis. Hist. Math. 11, 258–321 (1984) 46. S. Bochner, Integration von Funktionen, deren Wert die Elemente eines Vektorraumes sind. Fund. Math. 20, 262–276 (1933) 47. H. Bohman, On approximation of continuous and analytic functions. Ark. Mat 2, 43–56 (1952) 48. H.F. Bohnenblust, A. Sobczyk, Extensions of functionals on complex linear spaces. Bull. Am. Math. Soc. 44, 91–93 (1938)

Bibliography

379

49. P. du Bois-Reymond, Ueber die Fourier’schen Reihen. Nachr. K. Ges. Wiss. Göttingen 21, 571–582 (1873) 50. P. du Bois Reymond, Versuch einer Classification der willkürlichen Functionen reeller Argumente nach ihren Aenderungen in den kleinsten Intervallen [German]. J. Reine Angew. Math. 79, 21–37 (1875) 51. P. du Bois-Reymond, Untersuchungen über die Convergenz und Divergenz der Fourierschen Darstellungsformeln (mit drei lithografierten Tafeln). Abh. Math. Phys. KI. Bayer. Akad. Wiss. 12, 1–103 (1876) 52. P. du Bois-Reymond, Die allgemeine Funktionentheorie (Laapp, Tübingen, 1882); French translation: Théorie générale des fonctions, Nice, 1887 53. P. du Bois-Reymond, Ueber das Doppelintegral. J. Math. Pures Appl. 94, 273–290 (1883) 54. B. Bolzano, Rein analytischer Beweis des Lehrsatzes, dass zwischen je zwei Werthen, die ein entgegengesetztes Resultat gewähren, wenigstens eine reelle Wurzel der Gleichung liege (Prag, Haase, 1817). New edition: Ostwald’s Klassiker der exakten Wissenschaften, No. 153, Leipzig, 1905. English translation: Purely analytic proof of the theorem the between any two values which give results of opposite sign there lies at least one real root of the equation. Historia Math. 7, 156–185 (1980); See also in [57]. 55. B. Bolzano, Functionenlehre, around 1832. Partially published in Königliche böhmische Gesellschaft der Wissenschaften, Prága, 1930; complete publication in [56] 2A10/1, 2000, 23–165; English translation: [57], 429–572. 56. B. Bolzano, Bernard Bolzano Gesamtausgabe (Frommann–Holzboog, Stuttgart, 1969) 57. B. Bolzano, The Mathematical Works of Bernard Bolzano, translated by S. Russ, (Oxford University Press, Oxford, 2004) 58. E. Borel, Sur quelques points de la théorie des fonctions. Ann. Éc. Norm. Sup. (3) 12, 9–55 (1895); [62] I, 239–285 59. E. Borel, Leçons sur la théorie des fonctions (Gauthier-Villars, Paris, 1898) 60. E. Borel, Leçons sur les fonctions de variables réelles (Gauthier-Villars, Paris, 1905) 61. E. Borel (ed.), L. Zoretti, P. Montel, M. Fréchet, Recherches contemporaines sur la théorie des fonctions: Les ensembles de points, Intégration et dérivation, Développements en séries. Encyclopédie des sciences mathématiques, II-2 (Gauthier-Villars/Teubner, Paris/Leipzig, 1912) German translation and adaptation by A. Rosenthal: Neuere Untersuchungen über Funktionen reeller Veränderlichen: Die Punktmengen, Integration und Differentiation, Funktionenfolgen, Encyklopädie der mathematischen Wissenschaften, II-9 (Teubner, Leipzig, 1924) 62. E. Borel, Oeuvres I-IV (CNRS, Paris, 1972) 63. M.W. Botsko, An elementary proof of Lebesgue’s differentiation theorem. Am. Math. Mon. 110, 834–838 (2003) 64. N. Bourbaki, Sur les espaces de Banach. C. R. Acad. Sci. Paris 206, 1701–1704 (1938) 65. H. Brezis, Functional Analysis, Sobolev Spaces and Partial Differential Equations (Springer, New York, 2010) 66. H. Brunn, Zur Theorie der Eigebiete. Arch. Math. Phys. (3) 17, 289–300 (1910) 67. J.R. Buddenhagen, Subsets of a countable set. Am. Math. Mon. 78, 536–537 (1971) 68. J.C. Burkill, The Lebesgue Integral (Cambridge University Press, Cambridge, 1951) 69. G. Cantor, Ueber trigonometrische Reihen. Math. Ann. 4, 139–143 (1871); [76], 87–91 70. G. Cantor, Über eine Eigenschaft des Inbegriffes aller reellen algebraischen Zahlen. J. Reine Angew. Math. 77, 258–262 (1874); [76], 115–118 71. G. Cantor, Über unendliche, lineare Punktmannigfaltigkeiten III. Math. Ann. 20, 113–121 (1882); [76], 149–157 72. G. Cantor, Über unendliche, lineare Punktmannigfaltigkeiten V. Math. Ann. 21, 545–586 (1883); [76], 165–208 73. G. Cantor, De la puissance des ensembles parfaits de points. Acta Math. 4, 381–392 (1884) 74. G. Cantor, Über unendliche, lineare Punktmannigfaltigkeiten VI. Math. Ann. 23, 451–488 (1884); [76], 210–244

380

Bibliography

75. G. Cantor, Über eine Frage der Mannigfaltigkeitslehre. Jahresber. Deutsch. Math. Verein. 1, 75–78 (1890–1891); [76], 278–280 76. G. Cantor, Gesammelte Abhandlungen (Springer, Berlin, 1932) 77. C. Carathéodory, Constantin Vorlesungen über reelle Funktionen [German], 3rd edn. (Chelsea, New York, 1968), x+718 pp. [1st edn. (Teubner, Leipzig, 1918); 2nd edn. (1927); reprinting (Chelsea, New York, 1948)] 78. L. Carleson, On convergence and growth of partial sums of Fourier series. Acta Math. 116, 135–157 (1966) 79. A.L. Cauchy, Cours d’analyse algébrique (Debure, Paris, 1821); [82] (2) III, 1–476 80. L.-A. Cauchy, Résumé des leçons données à l’École Polytechnique sur le calcul infinitésimal. Calcul intégral (1823); [82] (2) IV, 122–261 81. L.-A. Cauchy, Mémoire sur les intégrales définies. Mém. Acad. Sci. Paris 1 (1827); [82] (1) I, 319–506 82. A.L. Cauchy, Oeuvres, 2 sorozat, 22 kötet (Gauthier-Villars, Paris, 1882–1905) 83. P.L. Chebyshev [Tchebychef], Sur les questions de minima qui se rattachent à la représentation approximative des fonctions. Mém. Acad. Imp. Sci. St. Pétersb. (6) Sciences math. et phys. 7, 199–291 (1859); [84] I, 273–378 84. P.L. Chebyshev [Tchebychef], Oeuvres I-II (Chelsea, New York, 1962) 85. E.W. Cheney, Introduction to Approximation Theory (Chelsea, New York, 1982) 86. P.R. Chernoff, Pointwise convergence of Fourier series. Am. Math. Monthly 87, 399–400 (1980) 87. P.G. Ciarlet, Linear and Nonlinear Functional Analysis with Applications (SIAM, Philadelphia, 2013) 88. A.C. Clairaut, Mémoire sur l’orbite apparente du soleil autour de la Terre, en ayant égard aux perturbations produites par des actions de la Lune et des Planètes principales. Hist. Acad. Sci. Paris, 52–564 (1754, appeared in 1759) 89. J.A. Clarkson, Uniformly convex spaces. Trans. Am. Math. Soc. 40, 396–414 (1936) 90. J.A. Clarkson, P. Erd˝os, Approximation by polynomials. Duke Math. J. 10, 5–11 (1943) 91. R. Courant, D. Hilbert, Methoden de matematischen Physik I (Springer, Berlin, 1931) 92. Á. Császár, Valós analízis I–II [Real Analysis I–II] (Tankönyvkiadó, Budapest, 1983) 93. P. Daniell, A general form of integral. Ann. Math. 19, 279–294 (1917/18) 94. G. Dantzig, Linear Programming and Extensions (Princeton University Press/RAND Corporation, Princeton/Santa Monica, 1963) 95. G. Darboux, Mémoire sur les fonctions discontinues. Ann. Éc. Norm. Sup. Paris (2) 4, 57–112 (1875) 96. M.M. Day, The spaces Lp with 0 < p < 1. Bull. Am. Math. Soc. 46, 816–823 (1940) 97. M.M. Day, Normed Linear Spaces (Springer, Berlin, 1962) 98. A. Denjoy, Calcul de la primitive de la fonction dérivée la plus générale. C. R. Acad. Sci. Paris 154, 1075–1078 (1912) 99. A. Denjoy, Une extension de l’intégrale de M. Lebesgue. C. R. Acad. Sci. Paris 154, 859–862 (1912) 100. A. Denjoy, Mémoire sur la totalisation des nombres dérivées non sommables. Ann. Éc. Norm. Sup. Paris (3) 33, 127–222 (1916) 101. M. de Vries, V. Komornik, Unique expansions of real numbers. Adv. Math. 221, 390–427 (2009) 102. M. de Vries, V. Komornik, P. Loreti, Topology of univoque bases. Topol. Appl. (2016). doi:10.1016/j.topol.2016.01.023, 103. J. Diestel, Geometry of Banach Spaces. Selected Topics (Springer, New York, 1975) 104. J. Diestel, Sequences and Series in Banach Spaces (Springer, New York, 1984) 105. J. Dieudonné, Sur le théorème de Hahn–Banach. Revue Sci. 79, 642–643 (1941) 106. J. Dieudonné, History of Functional Analysis (North-Holland, Amsterdam, 1981) 107. U. Dini, Sopra la serie di Fourier (Nistri, Pisa, 1872) 108. U. Dini, Sulle funzioni finite continue de variabili reali che non hanno mai derivata. Atti R. Acc. Lincei, (3), 1, 130–133 (1877); [111] II, 8–11

Bibliography

381

109. U. Dini, Fondamenti per la teoria delle funzioni di variabili reali (Nistri, Pisa, 1878) 110. U. Dini, Serie di Fourier e altre rappresentazioni analitiche delle funzioni di una variabile reale (Nistri, Pisa, 1880); [111] IV, 1–272 111. U. Dini, Opere I-V (Edizioni Cremonese, Roma, 1953–1959) 112. G.L. Dirichlet, Sur la convergence des séries trigonométriques qui servent à représenter une fonction arbitraire entre des limites données. J. Reine Angew. Math. 4, 157–169 (1829); [114] I, 117–132 113. G.L. Dirichlet, Über eine neue Methode zur Bestimmung vielfacher Integrale. Ber. Verh. K. Preuss. Acad. Wiss. Berlin, 18–25 (1839); [114] I, 381–390 114. G.L. Dirichlet, Werke I-II (Reimer, Berlin, 1889–1897) 115. J.L. Doob, The Development of Rigor in Mathematical Probability (1900–1950), [361], 157–169 116. N. Dunford, Uniformity in linear spaces. Trans. Am. Math. Soc. 44, 305–356 (1938) 117. N. Dunford, J.T. Schwartz, Linear Operators I-III (Wiley, New York, 1957–1971) 118. W.F. Eberlein, Weak compactness in Banach spaces I. Proc. Natl. Acad. Sci. USA 33, 51–53 (1947) 119. R.E. Edwards, Functional Analysis. Theory and Applications (Holt, Rinehart and Winston, New York, 1965) 120. R.E. Edwards, Fourier Series. A Modern Introduction I-II (Springer, New York, 1979–1982) 121. D.Th. Egoroff, Sur les suites de fonctions mesurables. C. R. Acad. Sci. Paris 152, 244–246 (1911) 122. M. Eidelheit, Zur Theorie der konvexen Mengen in linearen normierten Räumen. Stud. Math. 6, 104–111 (1936) 123. H.W. Ellis, D.O. Snow, On .L1 / for general measure spaces, Can. Math. Bull. 6, 211–229 (1963) 124. P. Erd˝os, P. Turán, On interpolation I. Quadrature- and mean-convergence in the Lagrange interpolation. Ann. Math. 38, 142–155 (1937); [126] I, 50–51, 97–98, 54–63 125. P. Erd˝os, P. Vértesi, On the almost everywhere divergence of Lagrange interpolatory polynomials for arbitrary system of nodes. Acta Math. Acad. Sci. Hungar. 36, 71–89 (1980), 38, 263 (1981) 126. P. Erd˝os (ed.), Collected Papers of Paul Turán I-III (Akadémiai Kiadó, Budapest,P 1990) 127. P. Erd˝os, I. Joó, V. Komornik, Characterization of the unique expansions 1 D qni and related problems. Bull. Soc. Math. France 118, 377–390 (1990) 128. L. Euler, De summis serierum reciprocarum. Comm. Acad. Sci. Petrop. 7, 123–134 (1734/1735, appeared in 1740); [132] (1) 14, 73–86 129. L. Euler, Introductio in analysin infinitorum I (M.-M. Bousquet, Lausanne, 1748); [132] (1) 8 130. L. Euler, De formulis integralibus duplicatis. Novi Comm. Acad. Sci. Petrop. 14 (1769): I, 1770, 72–103; [132] (1) 17, 289–315 131. L. Euler, Disquisitio ulterior super seriebus secundum multipla cuius dam anguli progredientibus. Nova Acta Acad. Sci. Petrop. 11, 114–132 (1793, written in 1777, appeared in 1798); [132] (1) 16, Part 1, 333–355 132. L. Euler, Opera Omnia, 4 series, 73 volumes (Teubner/Füssli, Leipzig/Zürich, 1911) 133. G. Faber, Über die interpolatorische Darstellung stetiger Funtionen. Jahresber. Deutsch. Math. Verein. 23, 192–210 (1914) 134. K. Falconer, Fractal Geometry. Mathematical Foundations and Applications, 2nd edn. (Wiley, Chicester, 2003) 135. J. Farkas, Über die Theorie der einfachen Ungleichungen. J. Reine Angew. Math. 124, 1–27 (1902) 136. P. Fatou, Séries trigonométriques et séries de Taylor. Acta Math. 30, 335–400 (1906) 137. L. Fejér, Sur les fonctions bornées et intégrables. C. R. Acad. Sci. Paris 131, 984–987 (1900); [143] I, 37–41 138. L. Fejér, Untersuchungen über Fouriersche Reihen. Math. Ann. 58, 51–69 (1904); [143] I, 142–160

382

Bibliography

139. L. Fejér, Eine stetige Funktion deren Fourier’sche Reihe divergiert. Rend. Circ. Mat. Palermo 28, 1, 402–404 (1909); [143] I, 541–543 140. L. Fejér, Beispiele stetiger Funktionen mit divergenter Fourierreihe. J. Reine Angew. Math. 137, 1–5 (1910); [143] I, 538–541 141. L. Fejér, Lebesguesche Konstanten und divergente Fourierreihen. J. Reine Angew. Math. 138, 22–53 (1910); [143] I, 543–572 142. L. Fejér, Über Interpolation. Götting. Nachr. 66–91 (1916); [143] II, 25–48 143. L. Fejér, Leopold Fejér. Gesammelte Arbeiten I-II, ed. by Turán Pál (Akadémiai Kiadó, Budapest, 1970) 144. G. Fichera, Vito Volterra and the birth of functional analysis, [361], 171–183 145. G.M. Fichtenholz, L.V. Kantorovich [Kantorovitch], Sur les opérations linéaires dans l’espace des fonctions bornées. Stud. Math. 5, 69–98 (1934) 146. E. Fischer, Sur la convergence en moyenne. C. R. Acad. Sci. Paris 144, 1022–1024, 1148– 1150 (1907) 147. S.R. Foguel, On a theorem by A.E. Taylor. Proc. Am. Math. Soc. 9, 325 (1958) 148. J.B.J. Fourier, Théorie analytique de la chaleur (Didot, Paris 1822); [149] I 149. J.B.J. Fourier Oeuvres de Fourier I-II (Gauthier-Villars, Paris, 1888–1890) 150. I. Fredholm, Sur une nouvelle méthode pour la résolution du problème de Dirichlet. Kong. Vetenskaps-Akademiens Förh. Stockholm, 39–46 (1900); [152], 61–68 151. I. Fredholm, Sur une classe d’équations fonctionnelles. Acta Math. 27, 365–390 (1903); [152], 81–106 152. I. Fredholm, Oeuvres complètes de Ivar Fredholm (Litos Reprotryck, Malmo, 1955) 153. G. Freud, Über positive lineare Approximationsfolgen von stetigen reellen Funktionen auf kompakten Mengen. On Approximation Theory, Proceedings of the Conference in Oberwolfach, 1963 (Birkhäuser, Basel, 1964), pp. 233–238 154. M. Fréchet, Sur quelques points du calcul fonctionnel. Rend. Circ. Mat. Palermo 22, 1–74 (1906) 155. M. Fréchet, Sur les ensembles de fonctions et les opérations linéaires. C. R. Acad. Sci. Paris 144, 1414–1416 (1907) 156. M. Fréchet, Sur les opérations linéaires III. Trans. Am. Math. Soc. 8, 433–446 (1907) 157. M. Fréchet, Les dimensions d’un ensemble abstrait. Math. Ann. 68, 145–168 (1910) 158. M. Fréchet, Sur l’intégrale d’une fonctionnelle étendue à˘a un ensemble abstrait. Bull. Soc. Math. France 43, 248–265 (1915) 159. M. Fréchet, L’écart de deux fonctions quelconques. C. R. Acad. Sci. Paris 162, 154–155 (1916) 160. M. Fréchet, Sur divers modes de convergence d’une suite de fonctions d’une variable réelle. Bull. Calcutta Math. Soc. 11, 187–206 (1919–1920) 161. M. Fréchet, Les espaces abstraits (Gauthier-Villars, Paris, 1928) 162. G. Frobenius, Über lineare Substitutionen und bilineare Formen. J. Reine Angew. Math. 84, 1–63 (1878); [163] I, 343–405 163. G. Frobenius, Gesammelte Abhandlungen I-III (Springer, Berlin, 1968) 164. G. Fubini, Sugli integrali multipli. Rend. Accad. Lincei Roma 16, 608–614 (1907) 165. G. Fubini, Sulla derivazione per serie. Rend. Accad. Lincei Roma 24, 204–206 (1915) 166. I.S. Gál, On sequences of operations in complete vector spaces. Am. Math. Mon. 60, 527–538 (1953) 167. B.R. Gelbaum, J.M.H. Olmsted, Counterexamples in Mathematics (Holden-Day, San Francisco, 1964) 168. B.R. Gelbaum, J.M.H. Olmsted, Theorems and Counterexamples in Mathematics (Springer, New York, 1990) 169. L. Gillman, M. Jerison, Rings of Continuous Functions (D. Van Nostrand Company, Princeton, 1960) 170. I.M. Glazman, Ju.I. Ljubic, Finite-Dimensional Linear Analysis: A Systematic Presentation in Problem Form (Dover, New York, 2006) 171. H.H. Goldstine, Weakly complete Banach spaces. Duke Math. J. 4, 125–131 (1938)

Bibliography

383

172. D.B. Goodner, Projections in normed linear spaces. Trans. Am. Math. Soc. 69, 89–108 (1950) 173. J.P. Gram, Om Rackkendvilklinger bestemte ved Hjaelp af de mindste Kvadraters Methode, Copenhagen, 1879. German translation: Ueber die Entwicklung reeller Funktionen in Reihen mittelst der Methode der kleinsten Quadrate, J. Reine Angew. Math. 94, 41–73 (1883) 174. A. Grothendieck, Sur la complétion du dual d’un espace vectoriel localement convexe. C. R. Acad. Sci. Paris 230, 605–606 (1950) 175. B.L. Gurevich, G.E. Shilov, Integral, Measure and Derivative: A Unified Approach (PrenticeHall, Englewood Cliffs, NJ, 1966) 176. A. Haar, Zur Theorie der orthogonalen Funktionensysteme. Math. Ann. 69, 331–371 (1910); [179], 47–87 177. A. Haar, Reihenentwicklungen nach Legendreschen Polynomen. Math. Ann. 78, 121–136 (1918); [179], 124–139 178. A. Haar, A folytonos csoportok elméletér˝ol. Magyar Tud. Akad. Mat. és Természettud. Ért. 49, 287–306 (1932); [182], 579–599; German translation: Die Maßbegriffe in der Theorie der kontinuierlichen Gruppen, Ann. Math. 34, 147–169 (1933); [182], 600–622 179. A. Haar, Gesammelte Arbeiten, ed. by B. Sz.-Nagy (Akadémiai Kiadó, Budapest, 1959) 180. H. Hahn, Theorie der reellen Funktionen I (Springer, Berlin, 1921) 181. P. Hahn, Über Folgen linearer Operationen. Monatsh. Math. Physik 32, 3–88 (1922); [183] I, 173–258 182. P. Hahn, Über linearer Gleichungssysteme in linearer Räumen. J. Reine Angew. Math. 157, 214–229 (1927); [183] I, 259–274 183. P. Hahn, Gesammelte Abhandlungen I-III (Springer, New York, 1995–1997) 184. P.R. Halmos, Measure Theory (D. Van Nostrand Co., Princeton, NJ, 1950) 185. P.R. Halmos, Introduction to Hilbert Space and the Theory of Spectral Multiplicity (Chelsea, New York, 1957) 186. P.R. Halmos, L Naive Set Theory (Van Nostrand, Princeton, NJ, 1960) 187. P.R. Halmos, A Hilbert Space Problem Book (Springer, Berlin, 1974) 188. G. Halphén, Sur la série de Fourier. C. R. Acad. Sci. Paris 95, 1217–1219 (1882) 189. H. Hanche-Olsen, H. Holden, The Kolmogorov-Riesz compactness theorem. Expo. Math. 28, 385–394 (2010) 190. H. Hankel, Untersuchungen über die unendlich oft oszillierenden und unstetigen Funktionen. Math. Ann. 20, 63–112 (1882) 191. G. Hardy, J.E. Littlewood, G. Pólya, Inequalities, 2nd edn. (Cambridge University Press, Cambridge, 1952) 192. A. Harnack, Die Elemente der Differential- und Integralrechnung (Teubner, Lepizig, 1881) 193. A. Harnack, Die allgemeinen Sätze über den Zusammenhang der Functionen einer reellen Variabelen mit ihren Ableitungen II. Math. Ann. 24, 217–252 (1884) 194. A. Harnack, Ueber den Inhalt von Punktmengen. Math. Ann. 25, 241–250 (1885) 195. F. Hausdorff, Grundzüge der Mengenlehre (Verlag von Veit, Leipzig, 1914); [197] II 196. F. Hausdorff, Dimension und äusseres Mass. Math. Ann. 79, 1–2, 157–179 (1918); [197] IV 197. F. Hausdorff, Gesammelte Werke I-VIII (Springer, Berlin, 2000) 198. T. Hawkins, Lebesgue’s Theory of Integration. Its Origins and Development (AMS Chelsea Publishing, Providence, 2001) 199. T.L. Heath (ed.), The Works of Archimedes (Cambridge University Press, Cambridge, 1897) 200. E. Heine, Die Elemente der Funktionenlehre. J. Reine Angew. Math. 74, 172–188 (1872) 201. E. Hellinger, O. Toeplitz, Grundlagen für eine Theorie der unendlichen Matrizen. Nachr. Akad. Wiss. Göttingen. Math.-Phys. Kl., 351–355 (1906) 202. E. Hellinger, O. Toeplitz, Grundlagen für eine Theorie der unendlichen Matrizen. Math. Ann. 69, 289–330 (1910) 203. E. Hellinger, O. Toeplitz, Integralgleichungen une Gleichungen mit unendlichvielen Unbekannten. Encyklopädie der Mathematischen Wissenschaften, II C 13 (Teubner, Leipzig, 1927) 204. E. Helly, Über lineare Funktionaloperationen. Sitzber. Kais. Akad. Wiss. Math.-Naturwiss. Kl. Wien 121, 2, 265–297 (1912)

384

Bibliography

205. R. Henstock, The efficiency of convergence factors for functions of a continuous real variable. J. Lond. Math. Soc. 30, 273–286 (1955) 206. R. Henstock, Definitions of Riemann type of the variational integrals. Proc. Lond. Math. Soc. (3) 11, 402–418 (1961) 207. E. Hewitt, K. Stromberg, Real and Abstract Analysis (Springer, Berlin, 1965) 208. D. Hilbert, Grundzüge einer allgemeinen Theorie der Integralgleichungen I. Nachr. Akad. Wiss. Göttingen. Math.-Phys. Kl., 49–91 (1904) 209. D. Hilbert, Grundzüge einer allgemeinen Theorie der Integralgleichungen IV. Nachr. Akad. Wiss. Göttingen. Math.-Phys. Kl., 157–227 (1906) 210. T.H. Hildebrandt, Necessary and sufficient conditions for the interchange of limit and summation in the case of sequences of infinite series of a certain type. Ann. Math. (2) 14, 81–83 (1912–1913) 211. T.H. Hildebrandt, On uniform limitedness of sets of functional operations. Bull. Am. Math. Soc. 29, 309–315 (1923) 212. T.H. Hildebrandt, Über vollstetige, lineare Transformationen. Acta. Math. 51, 311–318 (1928) 213. T.H. Hildebrandt, On bounded functional operations. Trans. Am. Math. Soc. 36, 868–875 (1934) 214. E.W. Hobson, On some fundamental properties of Lebesgue integrals in a two-dimensional domain. Proc. Lond. Math. Soc. (2) 8, 22–39 (1910) 215. H. Hochstadt, Eduard Helly, father of the Hahn-Banach theorem. Math. Intell. 2(3), 123–125 (1979) 216. R.B. Holmes, Geometric Functional Analysis and Its Applications (Springer, Berlin, 1975) 217. O. Hölder, Ueber einen Mittelwerthsatz. Götting. Nachr. 38–47 (1889) 218. L. Hörmander, Linear Differential Operators (Springer, Berlin, 1963) 219. L. Hörmander, The Analysis of Linear Partial Differential Operators I (Springer, Berlin, 1983) 220. R.A. Hunt, On the convergence of Fourier series orthogonal expansions and their continuous analogues, in Proceedings of the Conference at Edwardsville 1967 (Southern Illinois University Press, Carbondale, IL, 1968), pp. 237–255 221. D. Jackson, Über die Genauigkeit der Annäherung stetiger Funktionen durch ganze rationale Funktionen gegebenen Grades und trigonometrische Summen gegebener Ordnung, Dissertation, Göttingen, (1911) 222. D. Jackson, On approximation by trigonometric sums and polynomials. Trans. Am. Math. Soc. 13, 491–515 (1912) 223. D. Jackson, The Theory of Approximation, vol. 11 (American Mathematical Society, Colloquium Publications, Providence, RI, 1930) 224. C.G.J. Jacobi, De determinantibus functionalibus. J. Reine Angew. Math. 22, 319–359 (1841); [225] III, 393–438 225. C.G.J. Jacobi, Gesammelte Werke I-VIII (G. Reimer, Berlin, 1881–1891) 226. R.C. James, Characterizations of reflexivity. Stud. Math. 23, 205–216 (1964) 227. V. Jarník, Bolzano and the Foundations of Mathematical Analysis (Society of Czechoslovak Mathematicians and Physicists, Prague, 1981) 228. I. Joó, V. Komornik, On the equiconvergence of expansions by Riesz bases formed by eigenfunctions of the Schrödinger operator. Acta Sci. Math. (Szeged) 46, 357–375 (1983) 229. C. Jordan, Sur la série de Fourier. C. R. Acad. Sci. Paris 92, 228–230 (1881); [232] IV, 393– 395 230. C. Jordan, Cours d’analyse de l’École Polytechnique I-III (Gauthier-Villars, Paris, 1883) 231. C. Jordan, Remarques sur les intégrales définies. J. Math. (4) 8, 69–99 (1892); [232] IV, 427– 457 232. C. Jordan, Oeuvres I-IV (Gauthier-Villars, Paris, 1961–1964) 233. P. Jordan, J. von Neumann, On inner products in linear, metric spaces. Ann. Math. (2) 36(3), 719–723 (1935) 234. M.I. Kadec, On strong and weak convergence (in Russian). Dokl. Akad. Nauk SSSR 122, 13–16 (1958)

Bibliography

385

235. M.I. Kadec, On the connection between weak and strong convergence (in Ukrainian). Dopovidi. Akad. Ukraïn RSR 9, 949–952 (1959) 236. M.I. Kadec, Spaces isomorphic to a locally uniformly convex space (in Russian). Izv. Vysš. Uˇcebn. Zaved. Matematika 13(6), 51–57 (1959) 237. J.-P. Kahane, Fourier Series; see J.-P. Kahane, P.-G. Lemarié-Rieusset, Fourier Series and Wavelets (Gordon and Breach, New York, 1995) 238. J.-P. Kahane, Y. Katznelson, Sur les ensembles de divergence des séries trigonométriques. Stud. Math. 26, 305–306 (1966) 239. S. Kakutani, Weak topology and regularity of Banach spaces. Proc. Imp. Acad. Tokyo 15, 169–173 (1939) 240. S. Kakutani, Concrete representations of abstract .M/-spaces. (A characterization of the space of continuous functions). Ann. Math. (2) 42, 994–1024 (1941) 241. N.J. Kalton, An F-space with trivial dual where the Krein–Milman theorem holds. Isr. J. Math. 36, 41–50 (1980) 242. N.J. Kalton, N. Peck, A re-examination of the Roberts example of a compact convex set without extreme points. Math. Ann. 253, 89–101 (1980) 243. L.V. Kantorovich, G.P. Akilov, Functional Analysis, 2nd edn. (Pergamon Press, Oxford, 1982) 244. W. Karush, Minima of Functions of Several Variables with Inequalities as Side Constraints. M.Sc. Dissertation. University of Chicago, Chicago (1939) 245. Y. Katznelson, An Introduction to Harmonic Analysis (Dover, New York, 1976) 246. J. Kelley, Banach spaces with the extension property. Trans. Am. Math. Soc. 72, 323–326 (1952) 247. J. Kelley, General Topology (van Nostrand, New York, 1954) 248. J. Kindler, A simple proof of the Daniell-Stone representation theorem. Am. Math. Mon. 90, 396–397 (1983) 249. A.A. Kirillov, A.D. Gvisiani, Theorems and Problems in Functional Analysis (Springer, New York, 1982) 250. V.L. Klee, Jr., Convex sets in linear spaces I-II. Duke Math. J. 18, 443–466, 875–883 (1951) 251. A.N. Kolmogorov, Über Kompaktheit der Funktionenmengen bei der Konvergenz im Mittel. Nachr. Ges. Wiss. Göttingen 9, 60–63 (1931); English translation: On the compactness of sets of functions in the case of convergence in the mean, in V.M. Tikhomirov (ed.), Selected Works of A.N. Kolmogorov, vol. I (Kluwer, Dordrecht, 1991), pp. 147–150 252. A.N. Kolmogorov, Grundbegriffe der Wahrscheinlichkeitsrechnung (Springer, Berlin, 1933); English translation: Foundations of the Theory of Probability (Chelsea, New York, 1956) 253. A. Kolmogorov, Zur Normierbarkeit eines allgemeinen topologischen Raumes. Stud. Math. 5, 29–33 (1934) 254. A.N. Kolmogorov, S.V. Fomin, Elements of the Theory of Functions and Functional Analysis (Dover, New York, 1999) 255. V. Komornik, An equiconvergence theorem for the Schrödinger operator. Acta Math. Hungar. 44, 101–114 (1984) 256. V. Komornik, Sur l’équiconvergence des séries orthogonales et biorthogonales correspondant aux fonctions propres des opérateurs différentiels linéaires. C. R. Acad. Sci. Paris Sér. I Math. 299, 217–219 (1984) 257. V. Komornik, On the equiconvergence of eigenfunction expansions associated with ordinary linear differential operators. Acta Math. Hungar. 47, 261–280 (1986) 258. V. Komornik, A simple proof of Farkas’ lemma. Am. Math. Mon. 105, 949–950 (1998) 259. V. Komornik, D. Kong, W. Li, Hausdorff dimension of univoque sets and Devil’s staircase, arxiv: 1503.00475v1 [math. NT] 260. V. Komornik, P. Loreti, On the topological structure of univoque sets. J. Number Theory 122, 157–183 (2007) 261. V. Komornik, M. Yamamoto, On the determination of point sources. Inverse Prob. 18, 319– 329 (2002) 262. V. Komornik, M. Yamamoto, Estimation of point sources and applications to inverse problems. Inverse Prob. 21, 2051–2070 (2005)

386

Bibliography

263. P.P. Korovkin, On the convergence of positive linear operators in the space of continuous functions (in Russian). Dokl. Akad. Nauk. SSSR 90, 961–964 (1953) 264. P.P. Korovkin, Linear Operators and Approximation Theory. Russian Monographs and Texts on Advanced Mathematics and Physics, vol. III (Gordon and Breach Publishers, Inc/Hindustan Publishing Corp., New York/Delhi, 1960) 265. C.A. Kottman, Subsets of the unit ball that are separated by more than one. Stud. Math. 53, 15–27 (1975) 266. G. Köthe, Topological Vector Spaces I-II (Springer, Berlin, 1969, 1979) 267. M.A. Krasnoselskii, Ya.B. Rutitskii, Convex Functions and Orlicz Spaces (P. Noordhoff Ltd., Groningen, 1961) 268. M.G. Krein, S.G. Krein, On an inner characteristic of the set of all continuous functions defined on a bicompact Hausdorff space. Dokl. Akad. Nauk. SSSR 27, 429–430 (1940) 269. M. Krein, D. Milman, On extreme points of regularly convex sets. Stud. Math. 9, 133–138 (1940) 270. P. Krée, Intégration et théorie de la mesure. Une approche géométrique (Ellipses, Paris, 1997) 271. H.W. Kuhn, A.W. Tucker, Nonlinear programming, in Proceedings of 2nd Berkeley Symposium. (University of California Press, Berkeley, 1951), pp. 481–492 272. K. Kuratowski, La propriété de Baire dans les espaces métriques. Fund. Math. 16, 390–394 (1930) 273. K. Kuratowski, Quelques problèmes concernant les espaces métriques non-séparables. Fund. Math. 25, 534–545 (1935) 274. J. Kurzweil, Generalized ordinary differential equations and continuous dependence on a parameter (in Russian). Czechoslov. Math. J. 7(82), 418–449 (1957) 275. J. Kürschák, Analízis és analitikus geometria [Analysis and Analytic Geometry] (Budapest, 1920). 276. H.E. Lacey, The Hamel dimension of any infinite-dimensional separable Banach space is c. Am. Math. Mon. 80, 298 (1973) 277. M. Laczkovich, Equidecomposability and discrepancy; a solution of Tarski’s circle-squaring problem. J. Reine Angew. Math. 404, 77–117 (1990) 278. M. Laczkovich, Conjecture and Proof (Typotex, Budapest, 1998) 279. J.L. Lagrange, Solution de différents problèmes de calcul intégral. Miscellanea Taurinensia III, (1762–1765); [281] I, 471–668 280. J.L. Lagrange, Sur l’attraction des sphéroïdes elliptiques. Nouv. Mém. Acad. Royale Berlin (1773); [281] III, 619–649 281. J.L. Lagrange, Oeuvres I-XIV (Gauthier-Villars, Paris, 1867–1882) 282. E. Landau, Über einen Konvergenzsatz. Nachr. Akad. Wiss. Göttingen. Math.-Phys. Kl. IIa, 25–27 (1907); [284] III, 273–275 283. E. Landau, Über die Approximation einer stetigen Funktion durch eine ganze rationale Funktion. Rend. Circ. Mat. Palermo 25, 337–345 (1908); [284] III, 402–410 284. E. Landau, Collected Works I-VIII (Thales, Essen, 1986) 285. R. Larsen, Functional Analysis. An Introduction (Marcel Dekker, New York, 1973) 286. H. Lebesgue, Sur l’approximation des fonctions. Bull. Sci. Math. 22, 10 p. (1898); [298] III, 11–20 287. H. Lebesgue, Sur une généralisation de l’intégrale définie. C. R. Acad. Sci. Paris 132, 1025– 1027 (1901); [298] I, 197–199 288. H. Lebesgue, Intégrale, longueur, aire. Ann. Mat. Pura Appl. (3) 7, 231–359 (1902); [298] I, 201–331 289. H. Lebesgue, Sur les séries trigonométriques. Ann. Éc. Norm. (3) 20, 453–485 (1903); [298] III, 27–59 290. H. Lebesgue, Leçons sur l’intégration et la recherche des fonctions primitives (GauthierVillars, Paris, 1904); [298] II, 11–154 291. H. Lebesgue, Sur la divergence et la convergence non-uniforme des séries de Fourier. C. R. Acad. Sci. Paris 141, 875–877 (1905) 292. H. Lebesgue, Leçons sur les séries trigonométriques (Gauthier-Villars, Paris, 1906)

Bibliography

387

293. H. Lebesgue, Sur les fonctions dérivées. Atti Accad. Lincei Rend. 15, 3–8 (1906); [298] II, 159–164 294. H. Lebesgue, Sur la méthode de M. Goursat pour la résolution de l’équation de Fredholm. Bull. Soc. Math. France 36, 3–19 (1908); [298] III, 239–254 295. H. Lebesgue, Sur l’intégration des fonctions discontinues. Ann. Éc. Norm. (3) 27, 361–450 (1910); [298] II, 185–274 296. H. Lebesgue, Notice sur les travaux scientifiques de M. Henri Lebesgue (Impr. E. Privat, Toulouse, 1922); [298] I, 97–175 297. H. Lebesgue, Leçons sur l’intégration et la recherche des fonctions primitives, 2nd edn. (Paris, Gauthier-Villars, 1928) 298. H. Lebesgue, Oeuvres scientifiques I-V (Université de Genève, Genève, 1972–1973) 299. J. Le Roux, Sur les intégrales des équations linéaires aux dérivées partielles du 2e ordre à 2 variables indépendantes. Ann. Éc. Norm. (3) 12, 227–316 (1895) 300. B. Levi, Sul principio di Dirichlet. Rend. Circ. Mat. Palermo 22, 293–360 (1906) 301. B. Levi, Sopra l’integrazione delle serie. Rend. Instituto Lombardo Sci. Lett. (2) 39, 775–780 (1906) 302. J.W. Lewin, A truly elementary approach to the bounded convergence theorem. Am. Math. Mon. 93(5), 395–397 (1986) 303. J. Lindenstrauss, L. Tzafriri, Classical Banach Spaces II : Function Spaces (Springer, Berlin, 1979) 304. J.-L. Lions, Quelques méthodes de résolution des problèmes aux limites non linéaires (Dunod–Gauthier-Villars, Paris, 1969) 305. J.-L. Lions, E. Magenes, Problèmes aux limites non homogènes et applications I-III (Dunod, Paris, 1968–1970) 306. J. Liouville, Troisième mémoire sur le développement des fonctions ou parties de fonctions en séries dont les divers termes sont assoujettis à satisfaire à une même équation différentielle du second ordre, contenant un paramètre variable. J. Math. Pures Appl. 2, 418–437 (1837) 307. J.S. Lipi´nski, Une simple démonstration du théorème sur la dérivée d’une fonction de sauts. Colloq. Math. 8, 251–255 (1961) 308. R. Lipschitz, De explicatione per series trigonometricas instituenda functionum unius variabilis arbitrariarum, etc. J. Reine Angew. Math. 63(2), 296–308 (1864) 309. L.A. Ljusternik, W.I. Sobolev, Elemente der Funtionalanalysis (Akademie-Verlag, Berlin, 1979) 310. S. Lozinski, On the convergence and summability of Fourier series and interpolation processes. Mat. Sb. N. S. 14(56), 175–268 (1944) 311. S. Lozinski, On a class of linear operators (in Russian). Dokl. Akad. Nauk SSSR 61, 193–196 (1948) 312. H. Löwig, Komplexe euklidische Räume von beliebiger endlicher oder unendlicher Dimensionszahl. Acta Sci. Math. Szeged 7, 1–33 (1934) 313. N. Lusin, Sur la convergence des séries trigonométriques de Fourier. C. R. Acad. Sci. Paris 156, 1655–1658 (1913) 314. J. Marcinkiewicz, Quelques remarques sur l’interpolation. Acta Sci. Math. (Szeged) 8, 127– 130 (1937) 315. A. Markov, On mean values and exterior densities. Mat. Sb. N. S. 4(46), 165–190 (1938) 316. R.D. Mauldin (ed.), The Scottish Book: Mathematics from the Scottish Café (Birkhäuser, Boston, 1981) 317. S. Mazur, Über konvexe Mengen in linearen normierten Räumen. Stud. Math. 4, 70–84 (1933) 318. S. Mazur, On the generalized limit of bounded sequences. Colloq. Math. 2, 173–175 (1951) 319. J. McCarthy, An everywhere continuous nowhere differentiable function. Am. Math. Mon. 60, 709 (1953) 320. E.J. McShane, Linear functionals on certain Banach spaces. Proc. Am. Math. Soc. 1, 402–408 (1950) 321. R.E. Megginson, An Introduction to Banach Space Theory (Springer, New York, 1998)

388

Bibliography

322. D.P. Milman, On some criteria for the regularity of spaces of the type (B). C. R. (Doklady) Acad. Sci. URSS (N. S.) 20, 243–246 (1938) 323. H. Minkowski, Geometrie der Zahlen I (Teubner, Leipzig, 1896) 324. H. Minkowski, Geometrie der Zahlen (Teubner, Leipzig, 1910) 325. H. Minkowski, Theorie der konvexen Körper, insbesondere Begründung ihres Oberflächenbegriffs, [326] II, 131–229 326. H. Minkowski, Gesammelte Abhandlungen I-II (Teubner, Leipzig, 1911) 327. A.F. Monna, Functional Analysis in Historical Perspective (Oosthoek, Utrecht, 1973) 328. F.J. Murray, On complementary manifolds and projections in spaces Lp and lp . Trans. Am. Math. Soc. 41, 138–152 (1937) 329. Ch.H. Müntz, Über den Approximationssatz von Weierstrass. Mathematische Abhandlungen H.A. Schwarz gewidmet, (Springer, Berlin, 1914), pp. 303–312 330. L. Nachbin, A theorem of Hahn-Banach type for linear transformations. Trans. Am. Math. Soc. 68, 28–46 (1950) 331. L. Narici, E. Beckenstein, Topological Vector Spaces (Marcel Dekker, New York, 1985) 332. I.P. Natanson, Theory of functions of a real variable I-II (Frederick Ungar Publishing, New York, 1955, 1961) 333. I.P. Natanson, Constructive Function Theory I-III (Ungar, New York, 1964) 334. J. von Neumann, Mathematische Begründung der Quantenmechanik. Nachr. Gesell. Wiss. Göttingen. Math.-Phys. Kl., 1–57 (1927); [340] I, 151–207 335. J. von Neumann, Zur allgemeinen Theorie des Masses. Fund. Math. 13, 73–116 (1929); [340] I, 599–643 336. J. von Neumann, Zur Algebra der Funktionaloperationen und Theorie der Normalen Operatoren. Math. Ann. 102, 370–427 (1929–1930); [340] II, 86–143 337. J. von Neumann, Mathematische Grundlagen der Quantenmechanik (Springer, Berlin, 1932) 338. J. von Neumann, On complete topological spaces. Trans. Am. Math. Soc. 37, 1–20 (1935); [340] II, 508–527 339. J. von Neumann, On rings of operators III. Ann. Math. 41, 94–161 (1940); [340] III, 161–228 340. J. von Neumann, Collected works I-VI (Pergamon Press, Oxford, 1972–1979). 341. M.A. Neumark, Normierte Algebren (VEB Deutscher Verlag der Wissenschaften, Berlin, 1959) 342. O. Nikodým, Sur une généralisation des intégrales de M. Radon. Fund. Math. 15, 131–179 (1930) 343. O. Nikodým, Sur le principe du minimum dans le problème de Dirichlet. Ann. Soc. Polon. Math. 10, 120–121 (1931) 344. O. Nikodým, Sur le principe du minimum. Mathematica (Cluj) 9, 110–128 (1935) 345. W.P. Novinger, Mean convergence in Lp spaces. Proc. Am. Math. Soc. 34(2), 627–628 (1972) 346. V.F. Nikolaev, On the question of approximation of continuous functions by means of polynomials (in Russian). Dokl. Akad. Nauk SSSR 61, 201–204 (1948) 347. W. Orlicz, Über eine gewisse Klasse von Räumen vom Typus B. Bull. Int. Acad. Polon. Sci. A (8/9), 207–220 (1932) 348. W. Orlicz, Über Räume .LM /, Bull. Int. Acad. Polon. Sci. A, 93–107 (1936) 349. W. Orlicz, Linear Functional Analysis, Series in Real Analysis, vol. 4 (World Scientific Publishing, River Edge, NJ, 1992) 350. W. Osgood, Non uniform convergence and the integration of series term by term. Am. J. Math. 19, 155–190 (1897) 351. J.C. Oxtoby, Measure and Category (Springer, New York, 1971) 352. M.-A. Parseval, Mémoire sur les séries et sur l’intégration complète d’une équation aux différences partielles linéaires du second ordre, à coefficients constants. Mém. prés. par divers savants, Acad. des Sciences, Paris, (1) 1, 638–648 (1806) 353. G. Peano, Applicazioni geometriche del calcolo infinitesimale (Bocca, Torino, 1887) 354. G. Peano, Sur une courbe qui remplit toute une aire. Math. Ann. 36, 157–160 (1890) 355. R.R. Phelps, Uniqueness of Hahn-Banach extensions and unique best approximation. Trans. Am. Math. Soc. 95, 238–255 (1960)

Bibliography

389

356. O. Perron, Über den Integralbegriff. Sitzber. Heidelberg Akad. Wiss., Math. Naturw. Klasse Abt. A 16, 1–16 (1914) 357. B.J. Pettis, A note on regular Banach spaces. Bull. Am. Math. Soc. 44, 420–428 (1938) 358. B.J. Pettis, A proof that every uniformly convex space is reflexive. Duke Math. J. 5, 249–253 (1939) 359. R.R. Phelps, Lectures on Choquet’s Theorem (Van Nostrand, New York, 1966) 360. J.-P. Pier, Intégration et mesure 1900–1950; [361], 517–564 361. J.-P. Pier (ed.), Development of Mathematics 1900–1950 (Birkhäuser, Basel, 1994) 362. J.-P. Pier (ed.), Development of Mathematics 1950–2000 (Birkhäuser, Basel, 2000) 363. H. Poincaré, Science and Hypothesis (The Walter Scott Publishing, New York, 1905) 364. L.S. Pontryagin, Topological Groups, 3rd edn., Selected Works, vol. 2 (Gordon & Breach Science, New York, 1986) 365. A. Pringsheim, Grundlagen der allgemeinen Funktionenlehre, Encyklopädie der mathematischen Wissenschaften, II-1 (Teubner, Leipzig, 1899); French translation and adaptation: A. Pringsheim, J. Molk, Principes fondamentaux de la théorie des fonctions, Encyclopédie des sciences mathématiques, II-1, (Gauthier-Villars/Teubner, Paris/Leipzig, 1909) 366. J. Radon, Theorie und Anwendungen der absolut additiven Mengenfunktionen. Sitsber. Akad. Wiss. Wien 122, Abt. II a, 1295–1438 (1913) 367. M. Reed, B. Simon, Methods of Modern Mathematical Physics I-IV (Academic Press, New York, 1972–1979) 368. F. Rellich, Spektraltheorie in nichtseparabeln Räumen. Math. Ann. 110, 342–356 (1935) 369. I. Richards, On the Fourier inversion theorem for R1 . Proc. Amer. Math. Soc. 19(1), 145 (1968) 370. B. Riemann, Grundlagen für eine allgemeine Theorie der Functionen einer veränderlichen complexen Grösse, Inaugural dissertation, Göttingen, 1851; [372], 3–45; French translation: Principes fondamentaux pour une théorie générale des fonctions d’une grandeur variable complexe, Dissertation inaugurale de Riemann, Göttingen, 1851; [372], 1–56 371. B. Riemann, Ueber die Darstellberkeit einer Function durch eine trigonometrische Reihe, Habilitationsschrift, 1854, Abhandlungen der Königlichen Gesellschaft der Wissenschaften zu Göttingen 13 (1867); [372], 213–251. French translation: Sur la possibilité de représenter une fonction par une série trigonométrique, Bull. des Sciences Math. et Astron. (1) 5 (1873); [372], 225–272. 372. B. Riemann, Werke (Teubner, Leipzig, 1876); French translation: Oeuvres mathématiques de Riemann (Gauthier-Villars, Paris, 1898). English translation: Collected Works of Bernhard Riemann (Dover Publications, New York, 1953) 373. F. Riesz, Sur les systèmes orthogonaux de fonctions. C. R. Acad. Sci. Paris 144, 615–619 (1907); [392] I, 378–381 374. F. Riesz, Sur les systèmes orthogonaux de fonctions et l’équation de Fredholm. C. R. Acad. Sci. Paris 144, 734–736 (1907); [392] I, 382–385 375. F. Riesz, Sur une espèce de géométrie analytique des systèmes de fonctions sommables. C. R. Acad. Sci. Paris 144, 1409–1411 (1907); [392] I, 386–388 376. F. Riesz, Über orthogonale Funktionensysteme. Göttinger Nachr. 116–122 (1907); [392] I, 389–395 377. F. Riesz, Sur les suites de fonctions mesurables. C. R. Acad. Sci. Paris 148, 1303–1305 (1909); [392] I, 396–397, 405–406 378. F. Riesz, Sur les opérations fonctionnelles linéaires à une. C. R. Acad. Sci. Paris 149, 974–977 (1909); [392] I, 400–402 379. F. Riesz, Sur certains systèmes d’équations fonctionnelles et l’approximation des fonctions continues. C. R. Acad. Sci. Paris 150, 674–677 (1910); [392] I, 403–404, 398–399 380. F. Riesz, Untersuchungen über Systeme integrierbar Funktionen. Math. Ann. 69, 449–497 (1910); [392] I, 441–489 381. F. Riesz, Integrálható függvények sorozatai [Sequences of integrable functions]. Matematikai és Physikai Lapok 19, 165–182, 228–243 (1910); [392] I, 407–440 382. F. Riesz, Les systèmes d’équations linéaires à une infinité d’inconnues (Gauthier-Villars, Paris, 1913); [392] II, 829–1016

390

Bibliography

383. F. Riesz, Lineáris függvényegyenletekr˝ol [On linear functional equations]. Math. Term. Tud. Ért. 35, 544–579 (1917); [392] II, 1017–1052; German translation: Über lineare Funktionalgleichungen. Acta Math. 41, 71–98 (1918); [392] II, 1053–1080 384. F. Riesz, Su alcune disuguglianze. Boll. Unione Mat. Ital. 7, 77–79 (1928); [392] I, 519–521 385. F. Riesz, Sur la convergence en moyenne I-II. Acta Sci. Math. (Szeged) 4, 58–64, 182–185 (1928–1929); [392] I, 512–518, 522–525 386. F. Riesz, A monoton függvények differenciálhatóságáról [On the differentiability of monotone functions]. Mat. Fiz. Lapok 38, 125–131 (1931); [392] I, 243–249 387. F. Riesz, Sur l’existence de la dérivée des fonctions monotones et sur quelques problèmes qui s’y rattachent. Acta Sci. Math. (Szeged) 5, 208–221 (1930–32); [392] I, 250–263 388. M. Riesz, Sur les ensembles compacts de fonctions sommables. Acta Sci. Math. Szeged 6, 136–142 (1933) 389. F. Riesz, Zur Theorie der Hilbertschen Raumes. Acta Sci. Math., 34–38 (1934–1935); [392] II, 1150–1154 390. F. Riesz, L’évolution de la notion d’intégrale depuis Lebesgue. Ann. Inst. Fourier 1, 29–42 (1949); [392] I, 327–340 391. F. Riesz, Nullahalmazok és szerepük az analízisben. Az I. Magyar Mat. Kongr. Közl., 204– 214 (1952); [392] I, 353–362; French translation: Les ensembles de mesure nulle et leur rôle dans l’analyse, Az I. Magyar Mat. Kongr. Közl., English translation: Proceedings of the First Hungarian Mathematical Congress, 214–224 (1952); [392] I, 363–372 392. F. Riesz, Oeuvres complètes I-II, ed. by Á. Császár (Akadémiai Kiadó, Budapest, 1960) 393. F. Riesz, B. Sz.-Nagy, Über Kontraktionen des Hilbertschen Raumes [German]. Acta Univ. Szeged. Sect. Sci. Math. 10, 202–205 (1943) 394. F. Riesz, B. Sz.-Nagy, Leçons d’analyse fonctionnelle (Akadémiai Kiadó, Budapest, 1952); English translation: Functional Analysis, (Dover, New York, 1990) 395. J. Roberts, Pathological compact convex sets in the spaces Lp , 0 < p < 1, The Altgeld Book 1975/76, University of Illinois 396. J. Roberts, A compact convex set without extreme points. Stud. Math. 60, 255–266 (1977) 397. A.P. Robertson, W. Robertson, Topological Vector Spaces (Cambridge University Press, 1973) 398. R.T. Rockafellar, Conves Analysis (Princeton University Press, New Jersey, 1970) 399. L.J. Rogers, An extension of a certain theorem in inequalities. Messenger Math. 17, 145–150 (1888) 400. S. Rolewicz, Metric Linear Spaces (Pa´nstwowe Wydawnictwo Naukowe, Varsovie, 1972) 401. L.A. Rubel, Differentiability of monotonic functions. Colloq. Math. 10, 277–279 (1963) 402. W. Rudin, Fourier Analysis on Groups (Wiley, New York, 1962) 403. W. Rudin, Principles of Mathematical Analysis, 3rd edn. (McGraw-Hill, New York, 1976) 404. W. Rudin, Well-distributed measurable sets. Am. Math. Mon. 90, 41–42 (1983) 405. W. Rudin, Real and Complex Analysis (McGraw-Hill, New York, 1987) 406. W. Rudin, Functional Analysis (McGraw-Hill, New York, 1991) 407. S. Russ, Bolzano’s analytic programme. Math. Intell. 14(3), 45–53 (1992) 408. S. Saks, On some functionals. Trans. Am. Math. Soc. 35(2), 549–556 and 35(4), 965–970 (1933) 409. S. Saks, Theory of The Integral (Hafner, New York, 1937) 410. S. Saks, Integration in abstract metric spaces. Duke Math. J. 4, 408–411 (1938) 411. H. Schaefer, Topological vector spaces, Second edition with M.P. Wolff, 1st edn. (Springer, Berlin, 1966), 2nd edn. (1999) 412. J. Schauder, Zur Theorie stetiger Abbildungen in Funktionenräumen. Math. Z. 26, 47–65 (1927) 413. J. Schauder, Über die Umkehrung linearer, stetiger Funktionaloperationen. Stud. Math. 2, 1–6 (1930) 414. J. Schauder, Über lineare, vollstetige Funktionaloperationen. Stud. Math. 2, 183–196 (1930) 415. E. Schmidt, Entwicklung willkürlicher Funktionen nach Systemen vorgeschriebener. Math. Ann. 63, 433–476 (1907)

Bibliography

391

416. E. Schmidt, Über die Auflösung linearer Gleichungen mit unendlich vielen Unbekannten. Rend. Circ. Mat. Palermo 25, 53–77 (1908) 417. I.J. Schoenberg, On the Peano curve of Lebesgue. Bull. Am. Math. Soc. 44(8), 519 (1938) 418. I. Schur, Über lineare Transformationen in der Theorie der unendlichen Reihen. J. Reine Angew. Math. 151, 79–111 (1920) 419. J.T. Schwartz, A note on the space L p . Proc. Am. Math. Soc. 2, 270–275 (1951) 420. L. Schwartz, Théorie des distributions (Hermann, Paris, 1966) 421. Z. Semadeni, Banach Spaces of Continuous Functions I (Pánstwowe Wydawnictvo Naukowe, Warszawa, 1971) 422. A. Shidfar, K. Sabetfakhiri, On the continuity of Van der Waerden’s function in the Hölder sense. Am. Math. Mon. 93, 375–376 (1986) 423. H.J.S. Smith, On the integration of discontinuous functions. Proc. Lond. Math. Soc. 6, 140– 153 (1875) 424. V.L. Šmulian, Linear topological spaces and their connection with the Banach spaces. Doklady Akad. Nauk SSSR (N. S.) 22, 471–473 (1939) 425. V.L. Šmulian, Über lineare topologische Räume. Mat. Sb. N. S. (7) 49, 425–448 (1940) 426. S. Sobolev, Cauchy problem in functional spaces. Dokl. Akad. Nauk SSSR 3, 291–294 (1935) (in Russian) 427. S. Sobolev, Méthode nouvelle à résoudre le problème de Cauchy pour les équations linéaires hyperboliques normales. Mat. Sb. 1(43), 39–72 (1936) 428. S.L. Sobolev, Some Applications of Functional Analysis in Mathematical Physics (American Mathematical Society, Providence RI, 1991) 429. R. Solovay, A model of set theory where every set of reals is Lebesgue measurable. Ann. Math. 92, 1–56 (1970) 430. G.A. Soukhomlinov, Über Fortsetzung von linearen Funktionalen in linearen komplexen Räumen und linearen Quaternionräumen (in Russian, with a German abstract). Mat. Sb. N. S. (3) 4, 355–358 (1938) 431. L.A. Steen, Highlights in the history of spectral theory, Am. Math. Mon. 80, 359–381 (1973) 432. H. Steinhaus, Additive und stetige Funktionaloperationen. Math. Z. 5, 186–221 (1919) 433. V.A. Steklov, Sur la théorie de fermeture des systèmes de fonctions orthogonales dépendant d’un nombre quelconque de variables. Petersb. Denkschr. (8) 30, 4, 1–86 (1911) 434. V.A. Steklov, Théorème de fermeture pour les polynômes de Tchebychev–Laguerre. Izv. Ross. Akad. Nauk Ser. Mat. (6) 10, 633–642 (1916) 435. T.J. Stieltjes, Recherches sur les fractions continues. Ann. Toulouse 8, 1–122 (1894), 9, 1–47 (1895); [436] II, 402–566 436. T.J. Stieltjes, Oeuvres complètes I-II (Springer, Berlin, 1993) 437. O. Stolz, Ueber einen zu einer unendlichen Punktmenge gehörigen Grenzwerth. Math. Ann. 23, 152–156 (1884) 438. O. Stolz, Die gleichmässige Convergenz von Functionen mehrerer Veränderlichen zu den dadurch sich ergebenden Grenzwerthen, dass einige derselben constanten Werthen sich nähern. Math. Ann. 26, 83–96 (1886) 439. M.H. Stone, Linear Transformations in Hilbert Space (Amer. Math. Soc., New York, 1932) 440. M.H. Stone, Application of the theory of Boolean rings to general topology, Trans. Am. Math. Soc. 41, 325–481 (1937) 441. M.H. Stone, The generalized Weierstrass approximation theorem. Math. Mag. 11, 167–184, 237–254 (1947/48) 442. R.S. Strichartz, How to make wavelets. Am. Math. Mon. 100, 539–556 (1993) 443. P.G. Szabó, A matematikus Riesz testvérek. Válogatás Riesz Frigyes és Riesz Marcel levelezéséb˝ol [The Mathematician Riesz Brothers. Selected letters from the correspondence between Frederic and Marcel Riesz] (Magyar Tudománytörténeti Intézet, Budapest, 2010) 444. P.G. Szabó, Kiváló tisztelettel: Fejér Lipót és a Riesz testvérek levelezése magyar matematikusokkal [Respectfully: Correspondence of Leopold Fejér and the Riesz Brothers with Hungarian Mathematicians] (Magyar Tudománytörténeti Intézet, Budapest, 2011)

392

Bibliography

445. O. Szász, Über die Approximation stetiger Funktionen durch lineare Aggregate von Potenzen. Math. Ann. 77, 482–496 (1915–1916) 446. G. Szeg˝o, Orthogonal Polynomials (American Mathematical Society, Providence, 1975) 447. B. Sz.-Nagy, Spektraldarstellung linearer Transformationen des Hilbertschen Raumes (Springer, Berlin, 1942) 448. B. Sz.-Nagy, Introduction to Real Functions and Orthogonal Expansions (Oxford University Press, New York, 1965) 449. T. Takagi, A simple example of a continuous function without derivative. Proc. Phys. Math. Japan 1, 176–177 (1903) 450. A.E. Taylor, The extension of linear functionals. Duke Math. J. 5, 538–547 (1939) 451. S.A. Telyakovsky, Collection of Problems on the Theory of Real Functions (in Russian) (Nauka, Moszkva, 1980) 452. K.J. Thomae, Ueber bestimmte integrale. Z. Math. Phys. 23, 67–68 (1878) 453. H. Tietze, Über Funktionen, die auf einer abgeschlossenen Menge stetig sind. J. Reine Angew. Math. 145, 9–14 (1910) 454. A. Tychonoff, Ein Fixpunktsatz. Math. Ann. 111, 767–776 (1935) 455. A. Toepler, Bemerkenswerte Eigenschaften der periodischen Reihen. Wiener Akad. Anz. 13, 205–209 (1876) 456. O. Toeplitz, Das algebrischen Analogen zu einem Satze von Fejér. Math. Z. 2, 187–197 (1918) 457. L. Tonelli, Sull’integrazione per parti. Atti Accad. Naz. Lincei (5), 18, 246–253 (1909) 458. V. Trénoguine, B. Pissarevski, T. Soboléva, Problèmes et exercices d’analyse fonctionnelle (Mir, Moscou, 1987) 459. N.-K. Tsing, Infinite-dimensional Banach spaces must have uncountable basis—an elementary proof. Am. Math. Mon. 91(8), 505–506 (1984) 460. J.W. Tukey, Some notes on the separation of convex sets. Port. Math. 3, 95–102 (1942) 461. P. Urysohn, Über die Mächtigkeit der zusammenhängenden Mengen. Math. Ann. 94, 262–295 (1925) 462. S. Vajda, Theory of Games and Linear Programming (Methuen, London, 1956) 463. Ch.-J. de la Vallée-Poussin, Sur l’approximation des fonctions d’une variable réelle et de leurs dérivées par des polynômes et des suites limitées de Fourier. Bull. Acad. R.. Cl. Sci. 3 (1908), 193–254 464. Ch.-J. de la Vallée-Poussin, Réduction des intégrales doubles de Lebesgue: application à la définition des fonctions analytiques. Bull. Acad. Sci. Brux, 768–798 (1910) 465. Ch.-J. de la Vallée-Poussin, Sur l’intégrale de Lebesgue. Trans. Am. Math. Soc. 16, 435–501 (1915) 466. G. Vitali, Sul problema della misura dei gruppi di punti di una retta (Gamberini e Parmeggiani, Bologna, 1905) 467. N.Ya. Vilenkin, Stories About Sets (Academic Press, New York, 1968) 468. N.Ya. Vilenkin, In Search of Infinity (Birkhäuser, Boston, 1995) 469. C. Visser, Note on linear operators. Proc. Acad. Amst. 40, 270–272 (1937) 470. G. Vitali, Sulle funzioni integrali. Atti Accad. Sci. Torino 40, 1021–1034 (1905) 471. G. Vitali, Sull’integrazione per serie. Rend. Circ. Mat. Palermo 23, 137–155 (1907) 472. V. Volterra, Sulla inversione degli integrali definiti I-IV. Atti Accad. Torino 31, 311–323, 400–408, 557–567, 693–708 (1896); [476] II, 216–254 473. V. Volterra, Sulla inversione degli integrali definiti. Atti R. Accad. Rend. Lincei. Cl. Sci. Fis. Mat. Nat. (5) 5, 177–185 (1896); [476] II, 255–262 474. V. Volterra, Sulla inversione degli integrali multipli. Atti R. Accad. Rend. Lincei. Cl. Sci. Fis. Mat. Nat. (5) 5, 289–300 (1896); [476] II, 263–275 475. V. Volterra, Sopra alcuni questioni di inversione di integrali definiti. Annali di Mat. (2) 25, 139–178 (1897); [476] II, 279–313 476. V. Volterra, Opere matematiche. Memorie e Note I-V (Accademia Nazionale dei Lincei, Roma, 1954–1962) 477. B.L. van der Waerden, Ein einfaches Beispiel einer nichtdifferenzierbaren stetiges Funktion. Math. Z. 32, 474–475 (1930)

Bibliography

393

478. S. Wagon, The Banach-Tarski Paradox (Cambridge University Press, Cambridge, 1985) 479. J.V. Wehausen, Transformations in linear topological spaces. Duke Math. J. 4, 157–169 (1938) 480. K. Weierstrass, Differential Rechnung. Vorlesung an dem Königlichen Gewerbeinstitute, manuscript of 1861, Math. Bibl., Humboldt Universität, Berlin 481. K. Weierstrass, Über continuirliche Functionen eines reellen Arguments, die für keinen Werth des letzteren einen bestimmten Differentialquotienten besitzen. Gelesen in der Königlich. Akademie der Wissenschaften, 18. Juli 1872; [484] II, 71–74 482. K. Weierstrass, Theorie der analytischen Funktionen, Vorlesung an der Univ. Berlin, manuscript of 1874, Math. Bibl., Humboldt Universität, Berlin 483. K. Weierstrsass, Über die analytische Darstellbarkeit sogenannter willkürlicher Funktionen reeller Argumente, Erste Mitteilung. Sitzungsberichte Akad. Berlin, 633–639 (1885); [484] III, 1–37. French translation: Sur la possibilité d’une représentation analytique des fonctions dites arbitraires d’une variable réelle, J. Math. Pures Appl. 2 (1886), 105–138 484. K. Weierstrsass, Mathematische Werke vol. I-VI, (Mayer & Müller, Berlin, 1894–1915), vol. VII, (Georg Olms Verlagsbuchhandlung, Hildesheim, 1927) 485. A. Weil, L’intégration dans les groupes topologiques et ses applications (Hermann, Paris, 1940) 486. R. Whitley, An Elementary Proof of the Eberlein–Šmulian Theorem. Math. Ann. 172, 116– 118 (1967) 487. N. Wiener, Limit in terms of continuous transformation. Bull. Soc. Math. France 50, 119–134 (1922) 488. K. Yosida, Functional Analysis (Springer, Berlin, 1980) 489. W.H. Young, On classes of summable functions and their Fourier series. Proc. R. Soc. (A) 87, 225–229 (1912) 490. W.H. Young, The progress of mathematical analysis in the 20th century. Proc. Lond. Math. Soc. (2) 24 (1926), 421–434 491. L. Zajícek, An elementary proof of the one-dimensional density theorem. Am. Math. Mon. 86, 297–298 (1979) 492. M. Zorn, A remark on a method in transfinite algebra. Bull. Am. Math. Soc. 41, 667–670 (1935) 493. A. Zygmund, Trigonometric Series (Cambridge University Press, London, 1959)

Subject Index

[ , 212 *, 136 *, 30, 79 a.e., 156 absolutely continuous function , 198 measure, 235 signed measure, 239 accumulation point of a net, xv of a sequence, xvi adjoint operator, 35, 99 affine hyperplane, 15, 16, 57 almost everywhere, 156 anti-discrete topology, xiii automorphism, 103 axiom of choice, 62 B.K/ space, 76 B.K; X/ space, 76

Baire measure, 289 set, 289 Baire’s lemma, 32 balanced set, 58 ball, xvi, 120 Banach algebra, 43 space, 55, 76 Bernstein polynomials, 282 Bessel equality, 25

inequality, 25 bidual space, 79 Bochner integral, 305 Borel set, 195 boundary, xiii bounded function, xvi set, xvi, 122 broken line, xix

c0 is not a dual space, 140 c0 space, 70, 77 C0 function class, 171 C1 function class, 174 C2 function class, 176 C.K/ space, 77 C.K; X/ space, 77 Cb .K/ space, 77 Cb .K; X/ space, 77 Cc .I/ space, 312 C2 space, 263 Cbk .I; Y/ space, 77 Cbk .U; Y/ space, 78 Cantor function, 199, 204 Cantor’s diagonal method, 34, 90 ternary set, 154, 209, 254, 303 Cauchy sequence, xvii Cauchy–Schwarz inequality, xix, 4, 46 change of variable in integrals, 246 characteristic function, 173 Chebyshev’s characterization of closest polynomials, 301 C.Œ0; 1/ is not a dual space, 259

© Springer-Verlag London 2016 V. Komornik, Lectures on Functional Analysis and the Lebesgue Integral, Universitext, DOI 10.1007/978-1-4471-6811-9

395

396 closed graph theorem, 96 set, xiii subspace spanned by a set, 14 compact, xiii, 6 operator, 37, 101 complete measure, 192 metric space, xvii orthonormal sequence, 28 set, xvii completely bounded set, xviii continuous operator, 37, 101 completion of a Euclidean space, 10 metric space, 10 normed space, 78, 79 complex antilinear map, 45 Euclidean space, 45 Hilbert space, 46 linear map, 45 norm, 45 normed space, 45 scalar product, 45 component-wise convergence, 6, 33, 83 condensation of singularities, 81 connected set, xiv connects, xix continuity of a measure, 216 continuous, xiii function of compact support, 312 contraction, xvii convergence in measure, 356 of Fourier series, 271, 301 convergence-preserving, 117 convergent net, xv convex, xix closed hull, 126 convolution, 320 countable set, 151 counting measure, 213

dense, xiii density at a point, 203 in .L1 ; L1 /, 310 in Lp , 310 in Lpw , 312 Devil’s staircase, 204

Subject Index diagonal method of Cantor, 34, 90 diameter, xvi Dini derivatives, 161 Dirac functional, 335 measure, 213, 298, 299 direct product of measures, 213 sum, 103 Dirichlet function, 169 kernel, 272, 303 discrete metric, xvi topology, xiii disjoint set sequence, 152 set system, 152 divergence of Fourier series, 270 dominated convergence theorem, 181

eigen-subspace, 40 eigenvalue, 40 eigenvector, 40 embedding, xiv equicontinuous, 268 equiconvergence, 316 equivalent norms, 7, 96 Euclidean space, xix, 4 exterior, xiii extremal point, 126

Farkas–Minkowski lemma, 133 Fatou lemma, 183 Fejér kernel, 276, 303 finite measure, 217 part of a measure, 214 fixed point, xvii Fourier coefficient, 26, 270 series, 27, 270, 301 Fredholm alternative, 44, 110 function of bounded variation, 164

generalized integral, 189, 248 Newton–Leibniz formula, 204 gliding hump method, 84, 85 Gram–Schmidt orthogonalization, 29, 287

Subject Index Haar measure, 320 system, 340 Hahn decomposition, 231 Hamel basis, 116 Hausdorff, 254 dimension, 254 measure, 254 space, xiv Hermite polynomials, 315 Hilbert space, 4 Hilbert’s spectral theorem, 39 Hilbert–Schmidt operator, 38, 39, 311 Hölder’s inequality, 70, 306 hyperplane, 57 jIj, 154 indefinite integral, 198 integrable function , 176 majorant, 181 integral, 171, 174, 176, 189, 248 depending on a parameter, 186 on an interval, 197 integration by substitution, 246 interior, xiii inverse mapping theorem, 96 invisible from the left, 163 from the right, 162 isometry, xvii, 43

Jordan decomposition, 165, 231 jump function, 159

kernel, 40 of a linear functional, 20 of an operator, 103 `2 space, 5 `p space, 1 p < 1, 70 `p space, 0 < p 1, 144 `1 space, 70, 77 L.X; Y/ space, 77 Lp .I/ space, 1 p < 1, 79 Lpw space, 312 L0 space, 351 L1 space, 184 L1 .R/ is not a dual space, 335 L2 space, 11

397 Lp space, 1 p 1, 306 Lp space, 0 < p 1, 344 L1 space, 306 Laguerre polynomials, 315 Lebesgue decomposition, 237 measure, 191 point, 209 Lebesgue’s proof of the theorem of Weierstrass, 300 left shift, 108 level set, 193 lies, xix limit of a net, xv of a sequence, xvi linear functional, 57 hull, 25, 27 Lipschitz continuous, xvii locally convex space, 121 integrable, 312 measurable function, 247 measurable set, 248

measurable function, 187 set, 191 measure, 212 space, 212, 305 med med fx; y; zg, 188, 194, 259, 268 metric, xv space, xv subspace, xvii minimizing sequence, 327 Minkowski’s inequality, 70, 306 monotonicity of a measure, 216

N.A/, 103 negative part of a function, 181 part of a signed measure, 235 set, 232 neighborhood, xiii non-degenerate interval, 152 non-reflexivity of C.Œ0; 1/, 258, 259, 298, 300 L1 .0; 1/, 330 L1 .X; M; /, 331 L1 .R/, 336

398 L1 .0; 1/, 330 L1 .R/, 336 `1 , 89, 91, 143 `1 , 91, 143 c0 , 89, 91, 143 non-separability of C.Œ0; 1/0 , 299 L1 .I/, 314 `1 , 74 norm, xviii, 3 equivalence, 7, 96 normal operator, 47 normed space, 4 nowhere dense, 209 null set, 154, 218

open mapping, 62 mapping theorem, 96 set, xiii operator, 35 orthogonal complement, 14, 64, 125, 137 decomposition, 15 polynomials, 287, 314 projection, 12 orthogonality, 11 orthonormal basis, 28 family, 29 sequence, 24 outer measure, 252

parallelogram identity, xix, 4, 46 Parseval’s equality, 27 partition, xiii peak of a sequence, 6 Peano curve, 303 perfect set, 209 pointwise bounded, 268 positive linear functional, 172, 293 linear map, 279, 280 part of a function, 175, 181 part of a signed measure, 235 prehilbert space, 4 primitive function , 203 product topology, xiv projection, 50, 107, 284 pseudonorm, 347

Subject Index quasi-uniform convergence, 361 quotient norm, 115 space, 103, 115

R.A/, 103 Radon–Nikodým derivative, 240 Radon–Riesz property, 80, 328 range of an operator, 103 rectifiable, 165 reflexive space, 87 reflexivity of Lp , 1 < p < 1, 329 `p , 1 < p < 1, 87 finite-dimensional normed spaces, 87 Hilbert spaces, 87 uniformly convex Banach spaces, 329 regular distribution, 335 resolvent set, 108 restriction of a measure, 213 reverse Hölder inequality, 345 Minkowski inequality, 345 Young inequality, 345 Riemann–Lebesgue lemma, 338 Riemann–Stieltjes integral, 253 Riesz lemma, 56, 184, 307, 351, 356, 357 right shift, 108 ring, 192, 214 generated by a set system, 214 Rising sun lemma, 162

scalar product, xix, 4 Schauder basis, 304 segment, xix self-adjoint operator, 39 seminorm, xviii, 120 semiring, 212 separability of X and X 0 , 75 `p , 1 p < 1, 73 separable, xiii separated, xiv sequence, xvi set of the first category, 209 second category, 209 side, 129 signed Baire measure, 289 measure, 230

Subject Index simplex, 20 singular function , 206 measure, 235 signed measure, 239 spectral radius, 51 theorem of Hilbert, 39 spectrum, 43, 108 step function, 170, 218 stochastic convergence, 356 strict convexity, 324 strictly convex, 67, 92, 126, 324 strong convergence, 31, 80 strongly -finite measure, 236 signed measure, 239 subalgebra, 265 subnet, xv subsequence, xvi subspace, 1 topology, xiv symmetric operator, 39 -additive, 212 -algebra, 248 -finite measure, 222 set, 220 support, 220 -ring, 191 -subadditivity, 216 .Lp ; Lq / topology, 337 .L1 ; L1 / topology, 310 .X; X 0 / topology, 130 .X 0 ; X/ topology, 136 theorem of Arzelà –Ascoli, 268 Ascoli, 61 Baire, 209 Banach–Cacciopoli, xvii Banach–Steinhaus, 79, 81 Beppo Levi, 178 Bohman–Korovkin, 281 Bolzano on continuous images of connected sets, xiv Bolzano–Weierstrass, xx, 6, 29, 33 Brunn, 61 Cantor on intersections, xiii, xvii choice, 33, 90, 138 completion of metric spaces, xvii Dini, 292 Eberlein–Šmulian, 140

399 Egorov, 361 Eidelheit, 61 Erd˝os–Vértesi, 288 Faber, 288 Fejér, 276 fixed points of contractions, xvii Freud, 280 Fubini, 201, 225 Hahn, 231 Haršiladze–Lozinski, 284, 287 Hausdorff on continuity, xiv Hausdorff on continuous images of compact sets, xiv Hellinger–Toeplitz, 85, 98 Helly–Banach–Steinhaus, 79, 81 Helly–Hahn–Banach, 65 Hilbert, 39 James, 93 Jordan, 165, 231 Jordan–von Neumann, 50 Kadec, 329 Kakutani–Krein, 266 Klee, 92 Korovkin, 279 Krein–Milman, 126, 129 Kuhn–Tucker, 22 Lebesgue on decomposition, 206, 237 Lebesgue on density, 203 Lebesgue on dominated convergence, 181 Lebesgue on the differentiability of monotone functions, 157 Lebesgue–Vitali, 204 Mazur, 61 Milman–Pettis, 329 Minkowski, 61 Nikolaev, 288 Radon–Nikodým, 240 Riesz, 73 Riesz on representation, 291, 332 Riesz–Fischer, 184, 306 Riesz–Fréchet, 19 Schauder, 102 Schur, 84 selection of Helly, 167 Steinhaus, 332 Steinhaus–Toeplitz, 117 Stone–Weierstrass, 265 Tonelli, 228 Tukey, 61 Tukey–Klee, 124 Tychonoff, 7 Tychonoff on finite dimensional normed spaces, xx Tychonoff on products of compact sets, xiv

400 Vitali, 357 Vitali–Hahn–Saks, 343, 357 Weierstrass, 260, 264, 276, 282 Weierstrass on continuous images of compact sets, xiv theory of distributions, 335 Tm , 270 topological group, 320 space, xiii vector space, 144 topology, xiii total variation measure, 235 of a function, 164 totally bounded set, xviii triangle inequality, xv, 4, 45 trigonometric polynomial, 263, 270 polynomial of order m, 270 system, 24, 315

uniform boundedness theorem, 81 continuity modulus, 261 convexity, 324 uniformly convex, 324 space, 323

Subject Index unit sphere, 126 unitary operator, 47 vector lattice, 172 vertex, 126 visible from the right, 6 wavelet, 340 weak Cauchy sequence, 116 completeness, 116 convergence, 30, 79 convergence Lp , 336 convergence in C.K/, 299 sequential completeness, 116 topology, 130 weak star convergence, 136 convergence L1 , 336 topology, 136 weight function, 287 Young’s inequality, 70 zero measure, 213 zero-one measure, 213 Zorn’s lemma, 62

Name Index

Alabau-Boussouira, 367 Alaoglu, 139 Alexander the Great, 341 Archimedes, 149 Arzelà, 181, 268 Ascoli, 61, 154, 268, 301 Austin, 161

Baire, 32, 149, 178, 209, 289 Banach, 1, 32, 65, 67, 76, 79, 81, 96, 99, 139, 140, 192, 276, 291 Bauer, 117 Benner, 117 Bernoulli, Daniel, 270 Bernstein, 282 Besenyei, 367 Bessel, 25 Bochner, 211, 305 Bohman, 281 Bohnenblust, 112 du Bois-Reymond, 154, 229, 270, 271 Bolzano, 29 Bolzano–Weierstrass, 6 Borel, 149, 156, 195, 212, 282 Botsko, 161 Bourbaki, 140 Brunn, 61 Buddenhagen, 369

Cantor, 34, 36, 90, 149, 151, 152, 199, 212 Carathéodory, 252 Carleson, 271, 315, 316 Cauchy, 4, 46, 149, 229, 275, 276

Chebyshev, 301 Chernoff, 301 Clairaut, 26 Clarkson, 265, 323, 329 Császár, 160 Czách, 251

Darboux, 204 Day, 350 Denjoy, 197, 204 Dieudonné, 1, 21 Dini, 161, 198, 204, 271, 292, 301 Dirac, 213 Dirichlet, 149, 169, 224, 271, 272, 303 Dunford, 332

Eberlein, 140 Egorov, 361 Eidelheit, 61 Ellis, 336 Erd˝os, 154, 265, 288 Euclid, 341 Euler, 26, 28, 224

Faber, 288 Farkas, 133 Fatou, 183 Fejér, 265, 271, 275, 276, 303 Fichtenholz, 67 Fischer, 149, 184, 306 Foguel, 65 Fourier, 26, 149, 270, 301, 320

© Springer-Verlag London 2016 V. Komornik, Lectures on Functional Analysis and the Lebesgue Integral, Universitext, DOI 10.1007/978-1-4471-6811-9

401

402 Fredholm, 1, 44, 103, 110 Freud, 280 Fréchet, 1, 10, 19, 76, 149, 211, 307, 356, 362 Frobenius, 47 Fubini, 149, 201, 225

Gebuhrer, 98 Goldstine, 139 Goodner, 67 Gram, 24 Grothendieck, 116

Haar, 315, 316, 320 Hahn, 1, 65, 76, 81, 87, 231, 343, 357 Halmos, 244, 336 Hamel, 116 Hanche-Olsen, 316 Hankel, 81, 154 Haršiladze, 284, 287 Hardy, 55, 305 Harnack, 149, 154, 198, 212 Hausdorff, 10, 192 Heine, 156 Hellinger, 85, 98 Helly, 1, 65, 81, 167 Henstock, 197 Hermite, 169, 315 Hilbert, 1, 3, 5, 30, 33, 37–40, 43, 46, 47, 101, 151, 311 Hildebrandt, 81, 110, 328 Hobson, 225 Holden, 316 Hölder, 70, 306, 345

Jackson, 265 James, 93 Joó, 154, 316 Jordan, 149, 164, 165, 212, 231, 271

Kadec, 329 Kahane, 271 Kakutani, 140, 266, 291 Kalton, 351 Kantorovich, 67 Katznelson, 271 Kelley, 67 Kindler, 293 Klee, 92, 124 Kolmogorov, 123, 145, 149, 211, 316 Komornik, 133, 154, 204, 253, 299, 316

Name Index Kong, 154, 204, 253 Korovkin, 279, 281 Kottman, 57 Krein, M.G., 126, 129, 266 Krein, S.G., 266 Kuhn, 22 Kuratowski, 10 Kurzweil, 197 Kürschák, 6

Lacey, 117, 369 Laczkovich, 192 Lagrange, 35 Laguerre, 315 Landau, 85, 260 Lebesgue, 81, 84, 149, 156, 157, 181, 198, 199, 203, 204, 206, 211, 225, 237, 265, 300, 303, 338, 356, 361 Leibniz, 149 Levi, 12, 178 Lewin, 181 Li, 154, 204, 253 Lindenstrauss, 329 Lions, 342 Liouville, 316 Lipi´nski, 157 Lipschitz, 271, 301 Loreti, 154 Lozinski, 284, 287, 303 Löwig, 5, 46 Lusin, 271, 316

Marcinkiewicz, 284 Markov, 291 Mazur, 61, 65, 67 McShane, 325, 332 Menaechmus, 341 Milman, 126, 129, 329 Minkowski, 17, 61, 70, 129, 133, 145, 275, 276, 306, 345 Murray, 67, 112 Müntz, 265

Nachbin, 67 von Neumann, 1, 3, 5, 33, 46, 47, 119, 121, 130, 192, 240, 315 Newton, 149 Nikodým, 12, 149, 240, 332, 355 Nikolaev, 288 Novinger, 342

Name Index Orlicz, 306 Osgood, 181 Ounaies, 367 Parseval, 27 Peano, 149, 212, 303 Peck, 351 Perron, 197 Pettis, 90, 143, 329 Picard, 197 Poincaré, 169, 275, 338 Radon, 149, 211, 240, 291, 328 Rellich, 5, 40, 46 Richards, 301 Riemann, 81, 149, 253, 338 Riesz, 1, 3, 12, 15, 19, 35, 37, 56, 70, 73, 76, 85, 90, 99, 101, 103, 108, 110, 149, 161, 162, 184, 291, 306, 307, 328, 332, 337, 351, 356, 357 Riesz, M., 316 Roberts, 146, 351 Rogers, 70 Rubel, 157 Saks, 32, 81, 291, 343, 357 Schauder, 96, 99, 102, 110, 304, 340 Schmidt, 3, 12, 24, 33, 38–40, 311 Schoenberg, 303 Schur, 84 Schwartz, Jacob T., 336 Schwartz, Laurent, 335 Schwarz, 4, 46 Sebestyén, 59 Smith, 154 Smolyanov, 230 Šmulian, 140 Snow, 336 Sobczyk, 112 Sobolev, 149 Solovay, 192 Steinhaus, 32, 79, 81, 117, 332 Steklov, 318 Stieltjes, 253, 315

403 Stobaeus, 341 Stolz, 149, 212, 224 Stone, 98, 265, 267 Sz.-Nagy, 1, 149, 326 Szász, 265 Szuhomlinov, 112

Tarski, 192 Taylor, 65 Thomae, 229 Toepler, 25 Toeplitz, 47, 85, 98, 117 Tonelli, 149, 228 Tsing, 117 Tucker, 22 Tukey, 17, 18, 61, 91, 124 Turán, 288 Tychonoff, 6 Tzafriri, 329

Urysohn, 267

de la Vallée-Poussin, 173, 207, 225, 264 Vértesi, 288 Visser, 300 Vitali, 192, 195, 198, 204, 343, 357 Voltaire, 119 Volterra, 257 de Vries, 154, 253

Wehausen, 131 Weierstrass, 6, 29, 157, 260, 264, 265, 276, 300 Wiener, 45, 76 Wilde, 255

Yamamoto, 299 Young, 70, 345

Zorn, 62, 63, 129