Representation Theorems in Hardy Spaces

London Mathematical Society Student Texts
Managing Editor: Professor D. Benson, Department of Mathematics, University of Aberdeen, UK

23 24 26 27 28 29 31 32 33 34 35 37 38 39 40 41 42 43 44 45 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 Complex algebraic curves, FRANCES KIRWAN Lectures on elliptic curves, J. W. S. CASSELS Elementary theory of L-functions and Eisenstein series, HARUZO HIDA Hilbert space, J. R. RETHERFORD Potential theory in the complex plane, THOMAS RANSFORD Undergraduate commutative algebra, MILES REID The Laplacian on a Riemannian manifold, S. ROSENBERG Lectures on Lie groups and Lie algebras, ROGER CARTER, et al A primer of algebraic D-modules, S. C. COUTINHO Complex algebraic surfaces: Second edition, ARNAUD BEAUVILLE Young tableaux, WILLIAM FULTON A mathematical introduction to wavelets, P. WOJTASZCZYK Harmonic maps, loop groups, and integrable systems, MARTIN A. GUEST Set theory for the working mathematician, KRZYSZTOF CIESIELSKI Dynamical systems and ergodic theory, M. POLLICOTT & M. YURI The algorithmic resolution of Diophantine equations, NIGEL P. SMART Equilibrium states in ergodic theory, GERHARD KELLER Fourier analysis on nite groups and applications, AUDREY TERRAS Classical invariant theory, PETER J. OLVER Permutation groups, PETER J. CAMERON Introductory lectures on rings and modules. JOHN A. BEACHY Set theory, ANDRAS HAJNAL & PETER HAMBURGER An introduction to K-theory for C -algebras, M. RRDAM, F. LARSEN & N. LAUSTSEN A brief guide to algebraic number theory, H. P. F. SWINNERTON-DYER Steps in commutative algebra: Second edition, R. Y. SHARP Finite Markov chains and algorithmic applications, OLLE HGGSTRM The prime number theorem, G. J. O. JAMESON Topics in graph automorphisms and reconstruction, JOSEF LAURI & RAFFAELE SCAPELLATO Elementary number theory, group theory and Ramanujan graphs, GIULIANA DAVIDOFF, PETER SARNAK & ALAIN VALETTE Logic, induction and sets, THOMAS FORSTER Introduction to Banach algebras, operators, and harmonic analysis, H. GARTH DALES et al Computational algebraic geometry, HAL SCHENCK Frobenius algebras and 2-D topological quantum eld theories, JOACHIM KOCK Linear operators and linear systems, JONATHAN R. PARTINGTON An introduction to noncommutative Noetherian rings, K. R. GOODEARL & R. B. WARFIELD, JR Topics from one-dimensional dynamics, KAREN M. BRUCKS & HENK BRUIN Singular points of plane curves, C. T. C. WALL A short course on Banach space theory, N. L. CAROTHERS Elements of the representation theory of associative algebras Volume I, IBRAHIM ASSEM, DANIEL SIMSON & ANDRZEJ SKOWRONSKI An introduction to sieve methods and their applications, ALINA CARMEN COJOCARU & M. RAM MURTY Elliptic functions, J. V. ARMITAGE & W. F. EBERLEIN Hyperbolic geometry from a local viewpoint, LINDA KEEN & NIKOLA LAKIC Lectures on Khler geometry, ANDREI MOROIANU Dependence logic, JOUKU VNNEN Elements of the representation theory of associative algebras Volume 2, DANIEL SIMSON & ANDRZEJ SKOWRONSKI Elements of the representation theory of associative algebras Volume 3, DANIEL SIMSON & ANDRZEJ SKOWRONSKI Groups, graphs and trees, JOHN MEIER
L O N D O N M AT H E M AT I C A L S O C I E T Y S T U D E N T TEXTS 74
Representation Theorems in Hardy Spaces

JAVAD MASHREGHI Universit Laval
c am b r id g e u n ive r s it y p r e s s Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, So Paulo, Delhi Cambridge University Press The Edinburgh Building, Cambridge CB2 8RU, UK Published in the United States of America by Cambridge University Press, New York www.cambridge.org Information on this title: www.cambridge.org/9780521517683 J. Mashreghi 2009 This publication is in copyright. Subject to statutory exception and to the provisions of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published 2009 Printed in the United Kingdom at the University Press, Cambridge A catalogue record for this publication is available from the British Library Library of Congress Cataloguing in Publication data ISBN 978-0-521-51768-3 hardback ISBN 978-0-521-73201-7 paperback Cambridge University Press has no responsibility for the persistence or accuracy of URLs for external or third-party internet websites referred to in this publication, and does not guarantee that any content on such websites is, or will remain, accurate or appropriate.
v To My Parents: Masoumeh Farzaneh and Ahmad Mashreghi
vi
Contents
Preface 1 Fourier series 1.1 The Laplacian . . . . 1.2 Some function spaces 1.3 Fourier coecients . 1.4 Convolution on T . . 1.5 Youngs inequality . x 1 1 5 8 13 16 21 21 25 32 39 43 47 49 55 55 59 60 65 66 70 77 81 81 84 88 90 95 97
. . . . . . . . and sequence . . . . . . . . . . . . . . . . . . . . . . . .
. . . . spaces . . . . . . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
2 AbelPoisson means 2.1 AbelPoisson means of Fourier series . . . . . . . 2.2 Approximate identities on T . . . . . . . . . . . . 2.3 Uniform convergence and pointwise convergence . 2.4 Weak* convergence of measures . . . . . . . . . . 2.5 Convergence in norm . . . . . . . . . . . . . . . . 2.6 Weak* convergence of bounded functions . . . . 2.7 Parsevals identity . . . . . . . . . . . . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
3 Harmonic functions in the unit disc 3.1 Series representation of harmonic functions . . . . . . 3.2 Hardy spaces on D . . . . . . . . . . . . . . . . . . . . 3.3 Poisson representation of h (D) functions . . . . . . . 3.4 Poisson representation of hp (D) functions (1 < p < ) 3.5 Poisson representation of h1 (D) functions . . . . . . . 3.6 Radial limits of hp (D) functions (1 p ) . . . . . 3.7 Series representation of the harmonic conjugate . . . . 4 Logarithmic convexity 4.1 Subharmonic functions . . . . . . . . . . . . 4.2 The maximum principle . . . . . . . . . . . 4.3 A characterization of subharmonic functions 4.4 Various means of subharmonic functions . . 4.5 Radial subharmonic functions . . . . . . . . 4.6 Hardys convexity theorem . . . . . . . . . . vii
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
viii 4.7
CONTENTS A complete characterization of hp (D) spaces . . . . . . . . . . . . 99 103 103 106 110 113 116 120 123 131 131 135 136 142 144 149 153 155 155 158 162 166 168 172 175 181 187 187 189 191 197 200 205 207 207 209 218 219 221
5 Analytic functions in the unit disc 5.1 Representation of H p (D) functions (1 < p ) 5.2 The Hilbert transform on T . . . . . . . . . . . 5.3 Radial limits of the conjugate function . . . . . 5.4 The Hilbert transform of C 1 (T) functions . . . 5.5 Analytic measures on T . . . . . . . . . . . . . 5.6 Representations of H 1 (D) functions . . . . . . . 5.7 The uniqueness theorem and its applications .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
6 Norm inequalities for the conjugate function 6.1 Kolmogorovs theorems . . . . . . . . . . . . . . . 6.2 Harmonic conjugate of h2 (D) functions . . . . . . . 6.3 M. Rieszs theorem . . . . . . . . . . . . . . . . . . 6.4 The Hilbert transform of bounded functions . . . . 6.5 The Hilbert transform of Dini continuous functions 6.6 Zygmunds L log L theorem . . . . . . . . . . . . . 6.7 M. Rieszs L log L theorem . . . . . . . . . . . . . . 7 Blaschke products and their applications 7.1 Automorphisms of the open unit disc . . . . . . . . 7.2 Blaschke products for the open unit disc . . . . . . 7.3 Jensens formula . . . . . . . . . . . . . . . . . . . 7.4 Rieszs decomposition theorem . . . . . . . . . . . 7.5 Representation of H p (D) functions (0 < p < 1) . . 7.6 The canonical factorization in H p (D) (0 < p ) 7.7 The Nevanlinna class . . . . . . . . . . . . . . . . . 7.8 The Hardy and FejrRiesz inequalities . . . . . . e 8 Interpolating linear operators 8.1 Operators on Lebesgue spaces . . . . . . . . 8.2 Hadamards three-line theorem . . . . . . . 8.3 The RieszThorin interpolation theorem . . 8.4 The HausdorYoung theorem . . . . . . . 8.5 An interpolation theorem for Hardy spaces 8.6 The HardyLittlewood inequality . . . . . . 9 The 9.1 9.2 9.3 9.4 9.5 Fourier transform Lebesgue spaces on the real line . . . The Fourier transform on L1 (R) . . The multiplication formula on L1 (R) Convolution on R . . . . . . . . . . . Youngs inequality . . . . . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
CONTENTS 10 Poisson integrals 10.1 An application of the multiplication formula on L1 (R) 10.2 The conjugate Poisson kernel . . . . . . . . . . . . . . 10.3 Approximate identities on R . . . . . . . . . . . . . . . 10.4 Uniform convergence and pointwise convergence . . . . 10.5 Weak* convergence of measures . . . . . . . . . . . . . 10.6 Convergence in norm . . . . . . . . . . . . . . . . . . . 10.7 Weak* convergence of bounded functions . . . . . . .
ix 225 225 227 229 232 238 241 243 247 247 248 250 252 253 255 258
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
11 Harmonic functions in the upper half plane 11.1 Hardy spaces on C+ . . . . . . . . . . . . . . . . . . . . 11.2 Poisson representation for semidiscs . . . . . . . . . . . 11.3 Poisson representation of h(C+ ) functions . . . . . . . . 11.4 Poisson representation of hp (C+ ) functions (1 p ) 11.5 A correspondence between C+ and D . . . . . . . . . . . 11.6 Poisson representation of positive harmonic functions . . 11.7 Vertical limits of hp (C+ ) functions (1 p ) . . . . . 12 The 12.1 12.2 12.3 12.4 12.5 12.6
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
Plancherel transform 263 The inversion formula . . . . . . . . . . . . . . . . . . . . . . . . 263 The FourierPlancherel transform . . . . . . . . . . . . . . . . . . 266 The multiplication formula on Lp (R) (1 p 2) . . . . . . . . . 271 The Fourier transform on Lp (R) (1 p 2) . . . . . . . . . . . . 273 An application of the multiplication formula on Lp (R) (1 p 2) 274 A complete characterization of hp (C+ ) spaces . . . . . . . . . . . 276 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279 279 284 286 287 289 293 294 298 301 301 303 305 308 311 321 329 336
13 Analytic functions in the upper half plane 13.1 Representation of H p (C+ ) functions (1 < p ) . . 13.2 Analytic measures on R . . . . . . . . . . . . . . . . 13.3 Representation of H 1 (C+ ) functions . . . . . . . . . 13.4 Spectral analysis of H p (R) (1 p 2) . . . . . . . . 13.5 A contraction from H p (C+ ) into H p (D) . . . . . . . 13.6 Blaschke products for the upper half plane . . . . . . 13.7 The canonical factorization in H p (C+ ) (0 < p ) 13.8 A correspondence between H p (C+ ) and H p (D) . . . 14 The 14.1 14.2 14.3 14.4 14.5 14.6 14.7 14.8
Hilbert transform on R Various denitions of the Hilbert transform . . . . . . 1 The Hilbert transform of Cc (R) functions . . . . . . . Almost everywhere existence of the Hilbert transform Kolmogorovs theorem . . . . . . . . . . . . . . . . . . M. Rieszs theorem . . . . . . . . . . . . . . . . . . . . The Hilbert transform of Lip(t) functions . . . . . . . Maximal functions . . . . . . . . . . . . . . . . . . . . The maximal Hilbert transform . . . . . . . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
x A Topics from real analysis A.1 A very concise treatment of measure theory A.2 Riesz representation theorems . . . . . . . . A.3 Weak* convergence of measures . . . . . . . A.4 C(T) is dense in Lp (T) (0 < p < ) . . . . . A.5 The distribution function . . . . . . . . . . A.6 Minkowskis inequality . . . . . . . . . . . . A.7 Jensens inequality . . . . . . . . . . . . . . B A panoramic view of the representation B.1 hp (D) . . . . . . . . . . . . . . . . . . . B.1.1 h1 (D) . . . . . . . . . . . . . . . B.1.2 hp (D) (1 < p < ) . . . . . . . . B.1.3 h (D) . . . . . . . . . . . . . . . B.2 H p (D) . . . . . . . . . . . . . . . . . . . B.2.1 H p (D) (1 p < ) . . . . . . . B.2.2 H (D) . . . . . . . . . . . . . . B.3 hp (C+ ) . . . . . . . . . . . . . . . . . . B.3.1 h1 (C+ ) . . . . . . . . . . . . . . B.3.2 hp (C+ ) (1 < p 2) . . . . . . . . B.3.3 hp (C+ ) (2 < p < ) . . . . . . . B.3.4 h (C+ ) . . . . . . . . . . . . . . B.3.5 h+ (C+ ) . . . . . . . . . . . . . . B.4 H p (C+ ) . . . . . . . . . . . . . . . . . . B.4.1 H p (C+ ) (1 p 2) . . . . . . . B.4.2 H p (C+ ) (2 < p < ) . . . . . . . B.4.3 H (C+ ) . . . . . . . . . . . . . Bibliography Index
CONTENTS 339 339 344 345 346 347 348 349 351 352 352 354 355 356 356 358 359 359 361 362 363 363 364 364 365 366 367 369
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
CONTENTS
xi
Preface
In 1915 Godfrey Harold Hardy, in a famous paper published in the Proceedings of the London Mathematical Society, answered in the armative a question of Landau [7]. In this paper, not only did Hardy generalize Hadamards three-circle theorem, but he also put in place the rst brick of a new branch of mathematics which bears his name: the theory of Hardy spaces. For three decades afterwards Hardy, alone or with others, wrote many more research articles on this subject. The theory of Hardy spaces has close connections to many branches of mathematics, including Fourier analysis, harmonic analysis, singular integrals, probability theory and operator theory, and has found essential applications in robust control engineering. I have had the opportunity to give several courses on Hardy spaces and some related topics. A part of these lectures concerned the various representations of harmonic or analytic functions in the open unit disc or in the upper half plane. This topic naturally leads to the representation theorems in Hardy spaces. There are excellent books [5, 10, 13] and numerous research articles on Hardy spaces. Our main concern here is only to treat the representation theorems. Other subjects are not discussed and the reader should consult the classical textbooks. A rather complete description of representation theorems of H p (D), the family of Hardy spaces of the open unit disc, is usually given in all books. To study the corresponding theorems for H p (C+ ), the family of Hardy spaces of the upper half plane, a good amount of Fourier analysis is required. As a consequence, representation theorems for the upper half plane are not discussed thoroughly in textbooks mainly devoted to Hardy spaces. Moreover, quite often it is mentioned that they can be derived by a conformal mapping from the corresponding theorems on the open unit disc. This is a useful technique in certain cases and we will also apply it at least on one occasion. However, in the present text, our main goal is to give a complete description of the representation theorems with direct proofs for both classes of Hardy spaces. Hence, certain topics from Fourier analysis have also been discussed. But this is not a book about Fourier analysis, and we have been content with the minimum required to obtain the representation theorems. For further studies on Fourier analysis many interesting references are available, e.g. [1, 8, 9, 15, 21]. I express my appreciation to the many colleagues and students who made valuable comments and improved the quality of this book. I deeply thank Colin Graham and Mostafa Nasri, who read the entire manuscript and oered several
xii
CONTENTS
suggestions and Masood Jahanmir, who drew all the gures. I am also grateful to Roger Astley of the Cambridge University Press for his great management and kind help during the publishing procedure. I have beneted from various lectures by Arsalan Chademan, Galia Dafni, Paul Gauthier, Kohur GowriSankaran, Victor Havin, Ivo Klemes and Paul Koosis on harmonic analysis, potential theory and the theory of Hardy spaces. As a matter of fact, the rst draft of this manuscript dates back to 1991, when I attended Dr Chademans lectures on the theory of H p spaces. I take this opportunity to thank them with all my heart. Thanks to the generous support of Kristian Seip, I visited the Norwegian University of Science and Technology (NTNU) in the fall semester of 2007 2008. During this period, I was able to concentrate fully on the manuscript and prepare it for nal submission to the Cambridge University Press. I am sincerely grateful to Kristian, Yurii Lyubarski and Eugenia Malinnikova for their warm hospitality in Trondheim. I owe profound thanks to my friends at McGill University, Universit Laval e and Universit Claude Bernard Lyon 1 for their constant support and encoure agement. In particular, Niky Kamran and Kohur GowriSankaran have played a major role in establishing my mathematical life, Thomas Ransford helped me enormously in the early stages of my career, and Emmanuel Fricain sends me his precious emails on a daily basis. The trace of their eorts is visible in every single page of this book. Qubec e August 2008
Chapter 1
Fourier series
1.1 The Laplacian
An open connected subset of the complex plane C is called a domain. In particular, C itself is a domain. But, for our discussion, we are interested in two special domains: the open unit disc D = { z C : |z| < 1 } whose boundary is the unit circle T = { C : || = 1 }
Fig. 1.1. The open unit disc D and its boundary T. 1
2 and the upper half plane C+ = { z C : z > 0}
Chapter 1. Fourier series
whose boundary is the real line R (see Figures 1.1 and 1.2). They are essential domains in studying the theory of Hardy spaces.
Fig. 1.2. The upper half plane C+ and its boundary R.
The notations D(a, r) = { z C : |z a| < r }, D(a, r) = { z C : |z a| r }, D(a, r) = { z C : |z a| = r } for the open or closed discs and their boundaries will be used frequently too. We will also use Dr for a disc whose center is the origin, with radius r. The Laplacian of a twice continuously dierentiable function U : C is dened by 2U 2U 2 U = + . 2 x y 2 If 0 and we use polar coordinates, then the Laplacian becomes 2 U = 1 2 U 2 U 1 U + 2 + . 2 r r r r 2
We say that U is harmonic on if it satises the Laplace equation 2 U = 0 at every point of . By direct verication, we see that U (rei ) = rn cos(n), (n 0), (1.1)
1.1. The Laplacian and
U (rei ) = rn sin(n),
(n 1),
are real harmonic functions on C. Since (1.1) is a linear equation, a complexvalued function is harmonic if and only if its real and imaginary parts are real harmonic functions. The complex version of the preceding family of real harmonic functions is U (rei ) = r|n| ein , (n Z).
A special role is played by the constant function U 1 since it is the only member of the family whose integral means 1 2

U (rei ) d
are not zero. This fact is a direct consequence of the elementary identity 1 if n = 0, 1 ein d = (1.2) 2 0 if n = 0, which will be used frequently throughout the text. Let F be analytic on a domain and let U and V represent respectively its real and imaginary parts. Then U and V are innitely continuously dierentiable and satisfy the CauchyRiemann equations V U = x y and U V . = y x (1.3)
Using these equations, it is straightforward to see that U and V are real harmonic functions. Hence an analytic function is a complex harmonic function. In the following we will dene certain classes of harmonic functions, and in each class there is a subclass containing only analytic elements. Therefore, any representation formula for members of the larger class will automatically be valid for the corresponding subclass of analytic functions.
Exercises
Exercise 1.1.1 Let F be analytic on and let U = that U and V are real harmonic functions on . Hint: Use (1.3). F and V = F . Show
Exercise 1.1.2 Let F : C be analytic on , and suppose that F (z) = 0 for all z . Show that log |F | is harmonic on .
4 Exercise 1.1.3
Chapter 1. Fourier series Let U be harmonic on the annular domain
A(R1 , R2 ) = { z C : 0 R1 < |z| < R2 }.

1 1 Show that U ( 1 ei ) is harmonic on the annular domain A( R2 , R1 ). r
Exercise 1.1.4 Let F : 1 2 be analytic on 1 , and let U : 2 C be harmonic on 2 . Show that U F : 1 C is harmonic on 1 . Exercise 1.1.5 Dene the dierential operators = 1 2 1 2 i x y +i x y , .
= Show that
2 = 4 .
Exercise 1.1.6 Let F = U + iV be analytic. Show that the CauchyRiemann equations (1.3) are equivalent to the equation F = 0. Remark: We also have F = F . Exercise 1.1.7 Let U1 and U2 be real harmonic functions on a domain . Under what conditions is U1 U2 also harmonic on ? Remark: We emphasize that U1 and U2 are real harmonic functions. The answer to this question changes dramatically if we consider complex harmonic functions. For example, if F1 and F2 are analytic functions, then, under no extra condition, F1 F2 is analytic. Note that an analytic function is certainly harmonic. Exercise 1.1.8 Let U be a real harmonic function on a domain . Suppose that U 2 is also harmonic on . Show that U is constant. Hint: Use Exercise 1.1.7. Exercise 1.1.9 Let F be analytic on a domain , and let be a twice continuously dierentiable function on the range of F . Show that 2 ( F ) = (2 ) F |F |2 .
1.2. Some function spaces and sequence spaces
Exercise 1.1.10 Let F be analytic on a domain , and let R. Suppose that F has no zeros on . Show that 2 (|F | ) = 2 |F |2 |F |2 . Hint: Apply Exercise 1.1.9 with (z) = |z| . Exercise 1.1.11 Let F be analytic on a domain . Under what conditions is |F |2 harmonic on ? Hint: Apply Exercise 1.1.10. Exercise 1.1.12 Let F be a complex function on a domain such that F and F 2 are both harmonic on . Show that either F or F is analytic on . Hint: Use Exercise 1.1.7.
1.2
Some function spaces and sequence spaces

1 p
Let f be a measurable function on T, and let f and f

p
1 2 = inf
|f (e )| dt
it
(0 0
M : |{ eit : |f (eit )| > M }| = 0 ,
where |E| denotes the Lebesgue measure of the set E. Then Lebesgue spaces Lp (T), 0 < p , are dened by Lp (T) = { f : f
p
< }.
If 1 p , then Lp (T) is a Banach space. In particular, L2 (T), equipped with the inner product f, g = 1 2

f (eit ) g(eit ) dt,
is a Hilbert space. It is easy to see that L (T) Lp (T) L1 (T) for each p (1, ). In the following, we will mostly study Lp (T), 1 p , and their subclasses and thus L1 (T) is the largest function space that enters our discussion. This simple fact has important consequences. For example, we will dene the Fourier coecients of functions in L1 (T), and thus the Fourier coecients of elements of Lp (T), for all 1 p , are automatically dened
too. We will appreciate this fact when we study the Fourier transform on the real line. Spaces Lp (R) do not form a chain, as is the case on the unit circle, and thus after dening the Fourier transform on L1 (R), we need to take further steps in order to dene the Fourier transform for some other Lp (R) spaces. A continuous function on T, a compact set, is necessarily bounded. The space of all continuous functions on the unit circle C(T) can be considered as a subspace of L (T). As a matter of fact, in this case the maximum is attained and we have f = max |f (eit )|. it
e T
On some occasions we also need the smaller subspace C n (T) consisting of all n times continuously dierentiable functions, or even their intersection C (T) consisting of functions having derivatives of all orders. The space C n (T) is equipped with the norm
n
C n (T)
=
k=1
f (k) k!
Lipschitz classes form another subfamily of C(T). Fix (0, 1]. Then Lip (T) consists of all f C(T) such that sup |f (ei(t+ ) ) f (eit )| < . | |
t, R =0
This space is equipped with the norm f Lip (T) = f
+ sup
t, R =0
|f (ei(t+ ) ) f (eit )| . | |
The space of all complex Borel measures on T is denoted by M(T). This space equipped with the norm = ||(T), where || denotes the total variation of , is a Banach space. Remember that the total variation || is the smallest positive Borel measure satisfying |(E)| ||(E) for all Borel sets E T. To each function f L1 (T) there corresponds a Borel measure 1 d(eit ) = f (eit ) dt. 2 Clearly we have = f 1 , and thus the map L1 (T) M(T) f f (eit ) dt/2
1.2. Some function spaces and sequence spaces
is an embedding of L1 (T) into M(T). In our discussion, we also need some sequence spaces. For a sequence of complex numbers Z = (zn )nZ , let
Z and
=
n=
|zn |p
1 p
(0 < p < ),
Z Then, for 0 < p , we dene

p
= sup |zn |.
nZ
(Z) = { Z : Z
< } |zn | = 0 }.
and c0 (Z) = { Z
(Z) :
|n|
lim
If 1 p , then p (Z) is a Banach space and c0 (Z) is a closed subspace of (Z). The space 2 (Z), equipped with the inner product
Z, W =
n=
zn w n ,
is a Hilbert space. The subspaces

p
(Z+ ) = { Z
(Z) : zn = 0, n 1 }
and
c0 (Z+ ) = { Z c0 (Z) : zn = 0, n 1 }
will also appear when we study the Fourier transform of certain subclasses of Lp (T).
Exercises
Exercise 1.2.1 Let f be a measurable function on T, and let f (eit ) = f (ei(t ) ). Show that
0
lim f f
=0
if X = Lp (T), 1 p < , or X = C n (T). Provide examples to show that this property does not hold if X = L (T) or X = Lip (T), 0 < 1. However, show that f X = f X in all function spaces mentioned above.
8 Exercise 1.2.2 Exercise 1.2.3 Exercise 1.2.4 Show that

p
Chapter 1. Fourier series (Z+ ), 0 < p , is closed in

p
(Z).
Show that c0 (Z) is closed in
(Z).
Show that c0 (Z+ ) is closed in c0 (Z).
Exercise 1.2.5 Let cc (Z) denote the family of sequences of compact support, i.e. for each Z = (zn )nZ cc (Z) there is N = N (Z) such that zn = 0 for all |n| N . Show that cc (Z) is dense in c0 (Z).
1.3
Fourier coecients
1 f (n) = 2

Let f L1 (T). Then the nth Fourier coecient of f is dened by f (eit ) eint dt, (n Z).
The two-sided sequence f = ( f (n) )nZ is called the Fourier transform of f . We clearly have |f (n)| 1 2

|f (eit )| dt,
(n Z),
which can be rewritten as f
(Z) with f
1.
Therefore, the Fourier transform L1 (T) f
(Z) f
is a linear map whose norm is at most one. The constant function shows that the norm is indeed equal to one. We will show that this map is one-to-one and its range is included in c0 (Z). But rst, we need to develop some techniques. The Fourier series of f is formally written as
n=
f (n) eint .
The central question in Fourier analysis is to determine when, how and toward what this series converges. We will partially address these questions in the following. Any formal series of the form
n=
an eint
1.3. Fourier coecients
is called a trigonometric series. Hence a Fourier series is a special type of trigonometric series. However, there are trigonometric series which are not Fourier series. In other words, the coecients an are not the Fourier coecients of any integrable function. Using Eulers identity eint = cos(nt) + i sin(nt), a trigonometric series can be rewritten as
0 +
n=1
n cos(nt) + n sin(nt).
It is easy to nd the relation between n , n and an . An important example which plays a central role in the theory of harmonic functions is the Poisson kernel Pr (eit ) = (See Figure 1.3.) 1 r2 , 1 + r2 2r cos t (0 r < 1). (1.4)
K3
K2
K1
0
x
Fig. 1.3. The Poisson kernel Pr (eit ) for r = 0.2, 0.5, 0.8.
Clearly, for each xed 0 r < 1, Pr C (T) L1 (T). Direct computation of Pr is somehow dicult. But, the following observation
10 makes its calculation easier. We have 1 r2 1 + r2 2r cos t = = = and thus, using the geometric series 1 + w + w2 + = we obtain Pr (eit ) = 1 , 1w
n=
1 r2 1 + r(eit + eit ) 1 r2 it )(1 reit ) (1 re 1 1 + 1 it 1 re 1 reit r2
(|w| < 1),
r|n| eint .
(1.5)
Moreover, for each xed r < 1, the partial sums are uniformly convergent to Pr . The uniform convergence is the key to this shortcut method. Therefore, for each n Z, Pr (n) = = =
m=
1 2 1 2
Pr (eit ) eint dt
m=
r|m| eimt

eint dt
r|m|
1 2
ei(mn)t dt (1.6)
= r (See Figure 1.4.)
|n|
Fig. 1.4. The spectrum of Pr .
1.3. Fourier coecients
11
We emphasize that the uniform convergence of the series enables us to change the order of and in the third equality above. This phenomenon will appear frequently in our discussion. The identity (1.5) also shows that Pr is equal to its Fourier series at all points of T. It is rather easy to extend the denition of Fourier transform for Borel measures on T. Let M(T). The nth Fourier coecient of is dened by (n) =
T
eint d(eit ),
(n Z),
and the Fourier transform of is the two-sided sequence = ((n))nZ . Con sidering the embedding M(T) L1 (T) f f (eit ) dt/2, if we think of L1 (T) as a subspace of M(T), it is easy to see that the two denitions of Fourier coecients are consistent. In other words, if d(eit ) = where f L1 (T), then we have (n) = f (n), Lemma 1.1 Let M(T). Then In particular, for each f L1 (T), f Proof. For each n Z, we have |(n)| = =
T
1 f (eit ) dt, 2
(n Z). (Z) and
1.
eint d(eit )
T
|eint | d||(eit ) d||(eit ) = ||(T) = .
The second inequality is a special case of the rst one with d(eit ) = f (eit ) dt/2. In this case, (n) = f (n) and = f 1 . It was also proved directly at the beginning of this section.
12
Based on the preceding lemma, the Fourier transform M(T)
(Z)
is a linear map whose norm is at most one. The Dirac measure 1 shows that the norm is actually equal to one. We will show that this map is also one-to-one. However, since 1 (n) = 1, (n Z), its range is not included in c0 (Z).
Exercises
Exercise 1.3.1 Exercise 1.3.2 Let f Lp (T), 1 p . Show that f Let z = rei D. Show that Pr (ei(t) ) = Let f L1 (T). Dene F (z) = Show that F (z) = f (0) + 2
n=1
p.
eit + z eit z
Exercise 1.3.3
1 2
eit + z f (eit ) dt, eit z
(z D).
f (n) z n ,
(z D).
Hint: Note that
eit + z =1+2 z n eint . eit z n=1 Let f L1 (T) and dene g(eit ) = f (ei2t ).
Exercise 1.3.4
Show that g (n) =
f(n) 2 0
if if
2|n, 2 |n.
Consider a similar question if we dene g(eit ) = f (eikt ), where k is a xed positive integer.
1.4. Convolution on T Exercise 1.3.5 Let f Lip (T), 0 < 1. Show that f (n) = O(1/n ), as |n| . Hint: If n = 0, we have 1 f (n) = 4

13
f (eit ) f (ei(t+/n) ) eint dt.
1.4
Convolution on T
Let f, g L1 (T). Then we cannot conclude that f g L1 (T). Indeed, it is easy to manufacture an example such that

|f (eit ) g(eit )| dt = .
Nevertheless, by Fubinis theorem,

|f (ei ) g(ei(t ) )| d

dt d < .
|f (ei )|
|g(ei(t ) )| dt

|f (ei )| d
|g(eis )| ds
Therefore, we necessarily have

|f (ei ) g(ei(t ) )| d <
for almost all eit T. This observation enables us to dene (f g)(eit ) =

f (ei ) g(ei(t ) ) d
for almost all eit T, and besides the previous calculation shows that f g L1 (T) with f g
1
g 1.
(1.7)
The function f g is called the convolution of f and g. It is straightforward to see that the convolution is (i) commutative: f g = g f ,
14 (ii) associative: f (g h) = (f g) h, (iii) distributive: f (g + h) = f g + f h, (iv) homogenous: f (g) = (f ) g = (f g),
for all f, g, h L1 (T) and C. In technical terms, L1 (T), equipped with the convolution as its product, is a Banach algebra. As a matter of fact, this concrete example inspired most of the abstract theory of Banach algebras. Let f, g L1 (T) and let C(T). Then, by Fubinis theorem,
T
(eis ) (f g)(eis )
ds 2
=
T T
(eis )
f (ei(s ) ) g(ei ) f (eit ) dt 2
d 2
=
T T
(ei(t+ ) )
ds 2 d g(ei ) 2
This fact enables us to dene the convolution of two Borel measures on T such that if we consider L1 (T) as a subset of M(T), the two denitions are consistent. Let , M(T), and dene : C(T) C by () =
T T
(ei(t+ ) ) d(eit ) d(ei ),
( C(T)).
The functional is clearly linear and satises | () | which implies . Therefore, by the Riesz representation theorem for bounded linear functionals on C(T), there exists a unique Borel measure, which we denote by and call the convolution of and , such that () =
T
( C(T)),
(eit ) d( )(eit ), = .
( C(T)),
and moreover, Hence, is dened such that

T
(eit ) d( )(eit ) =
(ei(t+ ) ) d(eit ) d(ei ),

T T
(1.8)
for all C(T), and it satises . (1.9)
The following result is easy to prove. Nevertheless, it is the most fundamental connection between convolution and the Fourier transform. Roughly speaking, it says that the Fourier transform changes convolution to multiplication.
1.4. Convolution on T Theorem 1.2 Let , M(T). Then (n) = (n) (n), In particular, if f, g L1 (T), then f g(n) = f (n) g (n), Proof. Fix n Z and put in (1.8). Hence, (n) =
T
15
(n Z).
(n Z).
(eit ) = eint
eint d( )(eit ) ein(t+ ) d(eit ) d(ei )
=
T T
=
T
eint d(eit )
ein d(ei ) = (n) (n).

T
We saw that if we consider L1 (T) as a subset of M(T), the two denitions of convolution are consistent. Therefore, M(T) contains L1 (T) as a subalgebra. But, we can say more in this case. We show that L1 (T) is actually an ideal in M(T). Theorem 1.3 Let M(T) and let f L1 (T). Let d(eit ) = f (eit ) dt. Then is also absolutely continuous with respect to Lebesgue measure and we have d( )(eit ) = f (ei(t ) ) d(ei )
T
dt.
Proof. According to (1.8), for each C(T) we have

T
(eit ) d( )(eit ) = =
(ei(t+ ) ) d(eit ) d(ei )

T T
(ei(t+ ) ) d(eit ) f (ei ) d

T T
=
T
(eis )
T
f (ei(st) ) d(eit )
ds.
Therefore, by the uniqueness part of the Riesz representation theorem, d( )(eis ) = f (ei(st) ) d(eit )
T
ds.
16
Let M(T) and f L1 (T). Considering f as a measure, by the preceding theorem, f is absolutely continuous with respect to Lebesgue measure and we may write ( f )(eit ) = f (ei(t ) ) d(ei ).
T it
(1.10)
Theorem 1.3 ensures that ( f )(e ) is well-dened for almost all eit T, f L1 (T) and, by (1.9), f
1
(1.11)
We will need a very special case of (1.10) where f C(T). In this case, (f )(eit ) is dened for all eit T.
Exercises
Exercise 1.4.1 Let n (eit ) = eint , n Z. Show that f n = f (n) n for any f L1 (T). Exercise 1.4.2 Show that M(T), equipped with convolution as its product, is a commutative Banach algebra. What is its unit? Exercise 1.4.3 Are you able to show that the Banach algebra L1 (T) does not have a unit? Hint: Use Theorem 1.2. Come back to this exercise after studying the Riemann Lebesgue lemma in Section 2.5.
1.5
Youngs inequality
Since Lp (T) L1 (T), for 1 p , and since the convolution was dened on L1 (T), then a priori f g is well-dened whenever f Lr (T) and g Ls (T) with 1 r, s . The following result gives more information about f g, when we restrict f and g to some smaller subclasses of L1 (T). Theorem 1.4 (Youngs inequality) Let f Lr (T), and let g Ls (T), where 1 r, s and 1 1 + 1. r s Let 1 1 1 = + 1. p r s p Then f g L (T) and f g p f r g s.
1.5. Youngs inequality
17
Fig. 1.5. The level curves of p.
Proof. (Figure 1.5 shows the level curves of p.) If p = , or equivalently 1/r + 1/s = 1, then f g is well-dened for all ei T and Youngs inequality reduces to Hlders inequality. Now, suppose that 1/r + 1/s > 1. We need a o generalized form of Hlders inequality. Let 1 < p1 , . . . , pn < such that o 1 1 + + = 1, p1 pn and let f1 , . . . , fn be measurable functions on a measure space (X, M, ). Then |f1 fn | d |f1 |p1 d
1 p1
|fn |pn d
1 pn
This inequality can be proved by induction and the ordinary Hlders inequality. o Let r and s be respectively the conjugate exponents of r and s, i.e. 1 1 + =1 r r and 1 1 + = 1. s s
Then, according to the denition of p, we have 1 1 1 + + = 1. r s p Fix ei T. To apply the generalized Hlders inequality, we write the integrand o |f (eit ) g(ei(t) )| as the product of three functions respectively in Lr (T), Ls (T)
18 and Lp (T). Write |f (ei ) g(ei(t ) )| =
|g(ei(t ) )|1 p |f (ei )|1 p |f (ei )| p |g(ei(t ) )| p

r s r
Hence, by the generalized Hlders inequality, o 1 2

|f (e ) g(e
i(t )
)| d
1 2 1 2 1 2
|g(e
i(t )
)|
s r (1 p )
1 r
d
1 s
|f (e )|
i
r s (1 p )
d )| d
s
1 p
|f (e )| |g(e
i(t )
But r (1 s/p) = s and s (1 r/p) = r. Thus |(f g)(eit )| g

s/r s
r/s r
1 2
|f (ei )|r |g(ei(t ) )|s d
1 p
for almost all eit T. Finally, by Fubinis theorem, f g

p
= = g g
1 2
s/r s s/r s
|(f g)(e )| dt
r/s r r/s r
it
1 p
f f
1 (2)2 g
s/p s
|f (e )| |g(e
r
i(t )
)| dt d
1 p
r/p r
= f
g s.
Another proof of Youngs inequality is based on the RieszThorin interpolation theorem and will be discussed in Chapter 8. The following two special cases of Youngs inequality are what we need later on. Corollary 1.5 Let f Lp (T), 1 p , and let g L1 (T). Then f g Lp (T), and f g p f p g 1. Corollary 1.6 Let f Lp (T), and let g Lq (T), where q is the conjugate exponent of p. Then (f g)(eit ) is well-dened for all eit T, f g C(T), and f g
g q.
19
Proof. As we mentioned in the proof of Theorem 1.4, Hlders inequality ensures o that (f g)(eit ) is well-dened for all eit T. The only new fact to prove is that f g is a continuous function on T. At least one of p or q is not innity. Without loss of generality, assume that p = . This assumption ensures that C(T) is dense in Lp (T) (see Section A.4). Thus, given > 0, there is C(T) such that f Hence |(f g)(eit ) (f g)(eis )| | (f ) g (eit )| + | (f ) g (eis )|
p p
< .
+ |( g)(eit ) ( g)(eis )| 2 f where () = sup |(eit ) (eis )|

|ts|
+ (|t s|) g q ,
is the modulus of continuity of . Since is uniformly continuous on T, () 0 as 0. Therefore, if |t s| is small enough, we have |(f g)(eit ) (f g)(eis )| 3 g q .
Exercises
Exercise 1.5.1 What can we say about f g if f Lr (T) and g Ls (T) with 1 r, s and 1 1 + 1? r s (See Figure 1.6.)
20
Fig. 1.6. The region
1 r
1 s
1.
Exercise 1.5.2 Show that Youngs inequality is sharp in the following sense. Given r, s with 1 r, s and 1 1 + 1, r s there are f Lr (T) and g Ls (T) such that f g Lp (T), where 1 1 1 = + 1, p r s but f g Lt (T) for any t > p. Hint: Start with the case r = s = 1 and a function L1 (T) such that Lt (T) for any t > 1.
Chapter 2
AbelPoisson means
2.1 AbelPoisson means of Fourier series
F (reit ) = Fr (eit ), (reit D).
Let {Fr }0r<1 be a family of functions on the unit circle T. Dene
Hence, instead of looking at the family as a collection of individual functions Fr which are dened on T, we deal with one single function dened on the open unit disc D. On the other hand, if F (reit ) is given rst, for each xed r [0, 1), we can dene Fr by considering the values of F on the circle {|z| = r}. This dual interpretation will be encountered many times in what follows. An important example of this phenomenon is the Poisson kernel which was dened as a family of functions on the unit circle by (1.4). This kernel can also be considered as one function 1 r2 P (reit ) = 1 + r2 2r cos t on D. Let M(T). Then, by (1.10), we have (Pr )(ei ) = r2 1 r2 d(eit ) 2r cos( t) (2.1)
1+
which is called the Poisson integral of . Moreover, by (1.6) and Theorem 1.2, the Fourier coecients of Pr are given by Pr (n) = r|n| (n), Thus the formal Fourier series of Pr is
n=
(n Z).
(n) r|n| ein . 21
22
Chapter 2. AbelPoisson means
These sums are called the AbelPoisson means of the Fourier series
n=
(n) ein .
The Fourier series of is not necessarily pointwise convergent. However, we show that its AbelPoisson means behave much better. The following theorem reveals the relation between the AbelPoisson means of and its Poisson integral. Theorem 2.1 Let M(T), and let U (rei ) = (Pr )(ei ) = Then U (re ) =
n= i T
1+
r2
1 r2 d(eit ), 2r cos( t) (rei D).
(rei D).
(n) r|n| ein ,
The series is absolutely and uniformly convergent on compact subsets of D, and U is harmonic on D. Proof. Since | (n) r|n| ein | r|n| ,
the series (n) r|n| ein is absolutely and uniformly convergent on compact subsets of D. Fix 0 r < 1 and . Then, by (1.5), U (rei ) =
T
1+
r2
1 r2 d(eit ) 2r cos( t) r|n| ein(t) d(eit ).
=
T n=
Since the series is uniformly convergent (as a function of eit ), and since || is a nite positive Borel measure on T, we can change the order of summation and integration. Hence, U (rei ) =
n= T
eint d(eit )
r|n| ein =
n=
(n) r|n| ein .
There are several ways to verify that U is harmonic on D. We give a direct proof. Fix k 0. Then the absolute and uniform convergence of k |n| in e on compact subsets of D enables us to change the order n= n (n) r of summation and any linear dierential operator. In particular, let us apply the Laplace operator. Hence, remembering that each term r|n| ein is a harmonic function, we obtain U =
2 2
(n) r
n=
|n| in
=
n=
(n) 2 (r|n| ein ) = 0.
2.1. AbelPoisson means of Fourier series
23
As a special case, if the measure in Theorem 2.1 is absolutely continuous with respect to the Lebesgue measure, i.e. d(eit ) = u(eit ) dt/2 with u L1 (T), then U (rei ) = =
n=
1 2
1+
r2
1 r2 u(eit ) dt 2r cos( t) (2.2)
u(n) r|n| ein ,
where the series is absolutely and uniformly convergent on compact subsets of D, and U represents a harmonic function there.
Exercises
Exercise 2.1.1 the series Let (an )n0 be a sequence of complex numbers. Suppose that
S=
n=0
an
is convergent. For each 0 r < 1, dene
S(r) =
n=0
an rn .
Show that S(r) is absolutely convergent and moreover

r1
lim S(r) = S.
Hint: Let Sm =
an ,
n=0
(m 0).
Then S(r) = S + (1 r)
(Sn S) rn .
n=0
Exercise 2.1.2 let
Let (an )n0 be a bounded sequence of complex numbers and
S(r) =
n=0
an rn ,
(0 r < 1).
Find (an )n0 satisfying the following properties: (i) the series
n=0
an is divergent;
(ii) for each 0 r < 1, S(r) is absolutely convergent;
24 (iii) limr1 S(r) exists. Exercise 2.1.3 the series
Let (an )n0 be a sequence of complex numbers. Suppose that
S=
n=0
an
n
is convergent. Let Sn =
ak
k=0
and dene Cn = Show that

n
S0 + S1 + + Sn = n+1
1
k=0
k n+1
ak .
lim Cn = S.
a Remark: The numbers Cn , n 0, are called the Ces`ro means of Sn . Exercise 2.1.4 Let (an )n0 be a sequence of complex numbers and dene
n
Cn =
k=0
k n+1
ak .
Find (an )n0 such that

n
lim Cn
exists, but the sequence
an
n=0
is divergent. Exercise 2.1.5 Let (an )n0 be a sequence of complex numbers and dene
n
Cn =
k=0
k n+1
ak .
Suppose that the series
n k=0
Ck is convergent and
n
k |ak |2 < .
k=0
ak is also convergent. Show that the series Remark: Compare with Exercises 2.1.3 and 2.1.4.
n k=0
2.2. Approximate identities on T
25
2.2
Approximate identities on T
We saw that L1 (T), equipped with convolution, is a commutative Banach algebra. This algebra does not have a unit element since such an element must satisfy f (n) = 1, (n Z), and we will see that the nth Fourier coecient of any integrable function tends to zero as |n| . To overcome this diculty, we consider a family of integrable functions { } satisfying lim (n) = 1 (2.3)
for each xed n Z. The condition (2.3) alone is not enough to obtain a family that somehow plays the role of a unit element. For example, the Dirichlet kernel satises this property but it is not a proper replacement for the unit element (see Exercise 2.2.2). We choose three other properties to dene our family and then we show that (2.3) is fullled. Let L1 (T), where the index ranges over a directed set. In the examples given below, it ranges either over the set of integers {1, 2, 3, . . . } or over the interval [0, 1). Therefore, in the following, lim means either limn or limr1 . Similarly, 0 means n > n0 or r > r0 . The family { } is called an approximate identity on T if it satises the following properties: (a) for all , 1 2 (b) C = sup

(eit ) dt = 1;
1 2
| (eit )| dt
< ;
(c) for each xed , 0 < < , lim

|t|
| (eit )| dt = 0.
The condition (a) forces C 1. If (eit ) 0, for all and for all eit T, then { } is called a positive approximate identity. In this case, (b) follows from (a) with C = 1. We give three examples of a positive approximate identity below. Further examples are provided in the exercises. Our main example of a positive approximate identity is the Poisson kernel Pr (eit ) = 1 r2 = r|n| eint , 1 + r2 2r cos n=
(0 r < 1).
26 It is easy to verify that the Fejr kernel e 1 Kn (e ) = n+1

it
sin( (n+1)t ) 2 t sin( 2 )
=
k=n
|k| n+1
eikt ,
(n 0),
is also a positive approximate identity. (See Figure 2.1.)
16
14
12
10
K3
K2
K1
Fig. 2.1. The Fejr kernel Kn (eit ) for n = 5, 10, 15. e
A less familiar example is the family Fr (eit ) = (1 + r)2 (1 r) t sin t , (1 + r2 2r cos t)2 (0 r < 1, t ).
We will apply Fr in studying the radial limits of harmonic functions. Note that the three examples given above satisfy the following stronger property: (c ) For each xed , 0 < < , lim
sup
|t|
| (eit )|
= 0.
We now show that an approximate identity fulls (2.3). Fix n Z. Given > 0, there is such that |eint 1| <
2.2. Approximate identities on T for all t [, ]. Hence, by (a), we have 1 (n) 1 = 2 and thus, by (b), | (n) 1| 1 2

27
(eit ) (eint 1) dt
+
<|t|<
| (eit )| |eint 1| dt
C +
<|t|<
| (eit )| dt.
The property (c) ensures that the last integral tends to zero as grows. More precisely, there is such that | (n) 1| (C + 1) for all > . Given f L1 (T) and an approximate identity { } on T, we form the new family { f }. In the rest of this chapter, assuming f belongs to one of the Lebesgue spaces Lp (T) or to C(T), we explore the way in which f approaches f as grows. Similarly, for an arbitrary Borel measure M(T), we nd the relation between the measures ( )(eit ) dt/2 and . According to Theorem 1.2, the Fourier series of and f are respectively
n=
(n) (n) ein
and
n=
(n) f (n) ein .
Hence, they are respectively the weighted Fourier series of and f . That is why we can also say that we study the weighted Fourier series in the following.
Exercises
Exercise 2.2.1 Let Fn (eit ) = 2Kn (eit ) 0 if 0 t , if < t < 0.
Show that (Fn )n0 is a positive approximate identity on T. Find Fn . Exercise 2.2.2 The Dirichlet kernel is dened by
n
Dn (eit ) =
k=n
eikt ,
(n 0).
28 (See Figure 2.2.) Show that

n
Dn (eit ) = 1 + 2
k=1
cos(kt) =
sin( (2n+1)t ) 2 t sin( 2 )
and that (Dn )n0 is not an approximate identity on T. Find Dn . (See Figure 2.3, which shows the spectrum of D4 .)
30
20
10
K3
K2
K1
Fig. 2.2. The Dirichlet kernel Dn (eit ) for n = 5, 10, 15.
Fig. 2.3. The spectrum of D4 .
Exercise 2.2.3
Lebesgue constants are dened by Ln = 1 2

|Dn (eit )| dt,
(n 0).
2.2. Approximate identities on T Show that Ln as n . Remark: By a more precise estimation one can show that 4 4 log n < Ln < 3 + 2 log n. 2 Exercise 2.2.4 Show that Kn = Exercise 2.2.5 Let (n 1). D0 + D1 + + Dn , n+1 (n 0).
29
n (eit ) = 2n Kn1 (eit ) sin(nt), Show that n

1
< 2n. Let p be a trigonometric polynomial of degree at most n. p
Exercise 2.2.6 Show that
2n p
Hint: Note that p = p n , where n is given in Exercise 2.2.5. Now, apply Corollary 1.6. Remark: This inequality is not sharp. Bernstein showed that p n p , where p is any trigonometric polynomial of degree at most n. Exercise 2.2.7 given by Show that the Fourier coecients of the Fejr kernel Kn are e 1 Kn (m) =
|m| n+1
if
|m| n,
if |m| n + 1.
(Figure 2.4 shows the spectrum of K4 .)
Fig. 2.4. The spectrum of K4 .
30 Exercise 2.2.8
Chapter 2. AbelPoisson means [de la Valle Poussins kernel] Let e Vn (eit ) = 2K2n+1 (eit ) Kn (eit ).
(See Figure 2.5.) Show that (Vn )n0 1 |m| Vn (m) = 2 n+1 0
is an approximate identity on T, and that if |m| n + 1,
if n + 2 |m| 2n + 1, if |m| 2n + 2.
50
40
30
20
10
K3
K2
K1
Fig. 2.5. The de la Valle Poussin kernel Vn (eit ) for n = 5, 10, 15. e (Figure 2.6 shows the spectrum of V3 .)
Fig. 2.6. The spectrum of V3 .
2.2. Approximate identities on T Exercise 2.2.9 Show that Kn

2 2
31
1 2
|Kn (eit )|2 dt =
2n2 + 4n + 3 . 3(n + 1)
Exercise 2.2.10 Jn (eit ) =
[Jacksons kernel] Show that the family K2 (eit ) 3 n = Kn 2 (2n2 + 4n + 3)(n + 1) 2 sin( (n+1)t ) 2 t sin( 2 )
4
is a positive approximate identity on T. (See Figure 2.7.)
25
20
15
10
K3
K2
K1
Fig. 2.7. The Jackson kernel Jn (eit ) for n = 5, 10, 15.
Exercise 2.2.11
Let Fn (eit ) =
n 0
if if
|t| <
, n
|t| . n
Show that (Fn )n1 is a positive approximate identity on T. Find Fn .
32 Exercise 2.2.12 Fr (eit ) = Let (1 + r)2 (1 r) t sin t , (1 + r2 2r cos t)2
(0 r < 1, t ).
Show that (Fr )0r<1 is a positive approximate identity on T. Exercise 2.2.13 Let Fr (eit ) = 2(1 r2 ) sin2 t . (1 + r2 2r cos t)2
Show that (Fr )0r<1 is a positive approximate identity on T.
2.3
Uniform convergence and pointwise convergence
In this section we study the relation between f and f itself. We start with the simple but very important case of a continuous function. The crucial property which is exploited below is that a continuous function on T, a compact set, is automatically uniformly continuous there. Theorem 2.2 Let { } be an approximate identity on T, and let f C(T). Then, for each , f C(T) with f and moreover lim f f

C f
= 0.
In other words, f converges uniformly to f on T. Proof. By Corollary 1.6, for each , we certainly have f C(T) and f
C f
Since f is uniformly continuous on T, given > 0, there exists = () > 0 such that | f (eit2 ) f (eit1 ) | < whenever |t2 t1 | < . Therefore, for all t, | ( f )(eit ) f (eit ) | = 1 2 f

1 2
(ei )
f (ei(t ) ) f (eit ) d
| (ei )| |f (ei(t ) ) f (eit )| d
| |
| (ei )| d + C .
2.3. Uniform convergence and pointwise convergence Pick () so large that 1

| |
33
| (ei )| d <
for > (). Thus, for > () and for all t, | ( f )(eit ) f (eit ) | < ( f
+ C ) .
As a special case, let { } be the Poisson kernel. By using this kernel we extend a continuous function on T into the unit disc D. The outcome is a function continuous on the closed unit disc D and harmonic on the open unit disc D. Corollary 2.3 Let u C(T), and let 1 r2 1 u(eit ) dt 2 2r cos( t) 2 1 + r i U (re ) = u(ei ) Then (a) U is continuous on D, (b) U is harmonic on D, (c) for each 0 r < 1, Ur
if 0 r < 1, if r = 1.
Proof. By Theorem 2.1, U is harmonic on D. Hence, U is at least continuous on D. On the other hand, Theorem 2.2 ensures that, as r 1, Ur converges uniformly to u, and Ur u . Therefore, U is also continuous at all points of T. By Theorem 2.1, we also have U (rei ) =
n=
u(n) r|n| ein , (0 r < 1).
(2.4)
Therefore, Corollary 2.3 says that the AbelPoisson means of the Fourier series of a continuous function u converge uniformly to u. Let P(T) denote the space of all trigonometric polynomials:
N
P(T) = { p : p(eit ) =
n=N
an eint , an C }.
Clearly P(T) C(T). A celebrated theorem of Weierstrass says that P(T) is dense in C(T). Using the Fejr kernel we are able to give a constructive proof of this result. e
34
Corollary 2.4 (WeierstrassFejr) Let f C(T), and let e

n
pn (eit ) =
k=n
|k| n+1
f (k) eikt ,
(eit T).
Then, for each n 0, pn and

n
lim
pn f
= 0.
Proof. It is enough to observe that pn = f K n , e where Kn is the Fejr kernel. Then apply Theorem 2.2. In the rest of this section we study a local version of Theorem 2.2 by assuming that f is continuous at a xed point on T. Thus we obtain some results about the pointwise convergence of Fourier series. Theorem 2.5 Let { } be an approximate identity on T. Suppose that, for each , L (T), and that { } satises the stronger property (c ): lim
sup
0<|t|
| (eit )|
= 0.
Let f L1 (T), and suppose that f is continuous at eit0 T. Then, given > 0, there exists (, t0 ) and = (, t0 ) > 0 such that | ( f )(eit ) f (eit0 ) | < , if > (, t0 ) and |t t0 | < . In particular, lim ( f )(eit0 ) = f (eit0 ).
Proof. Since L (T) and f L1 (T), by Corollary 1.6, ( f )(eit ) is well-dened at all eit T. By assumption, given > 0, there exists = (, t0 ) > 0 such that | f (ei ) f (eit0 ) | <
2.3. Uniform convergence and pointwise convergence whenever | t0 | < 2. Therefore, for |t t0 | < , | ( f )(eit ) f (eit0 ) | = 1 2 1 2
1
35
1 2
(ei )
f (ei(t ) ) f (eit0 ) d
| (ei )| |f (ei(t ) ) f (eit0 )| d
( f
| | it0
| (ei )| |f (ei(t ) )| + |f (eit0 )| d + C )|) sup

|t|
+ |f (e
| (eit )| + C . | (eit )| <
Pick (, t0 ) so large that sup

|t|
whenever > (, t0 ). Thus, for > (, t0 ) and for |t t0 | < , we have | ( f )(eit ) f (eit0 ) | < ( f
1
+ |f (eit0 )| + C ) .
The following result is a local version of Corollary 2.3. Note that the Poisson kernels fulls all the requirements of Theorem 2.5. Corollary 2.6 Let u L1 (T), and let U (rei ) = 1 2

1+
r2
1 r2 u(eit ) dt, 2r cos( t)
(rei D).
Suppose that u is continuous at eit0 T. Then U is harmonic on D and besides

zeit0 zD
lim U (z) = u(eit0 ).
In particular,
r1
lim U (reit0 ) = u(eit0 ).
In Theorem 2.5 and Corollary 2.6, among other things, we assumed that our function f is continuous at a xed point eit0 . Hence, we implicitly take it as granted that f (eit0 ) is a nite complex number and f (eit ) converges to this value as t t0 . The niteness of f (eit0 ) is crucial. If f is a complex-valued function and lim |f (eit )| = |f (eit0 )| = +,
tt0
we may still say that f is continuous at eit0 . However, the preceding results are not valid for this class of functions. Nevertheless, if our function is real-valued and lim f (eit ) = f (eit0 ) = +,
tt0
we are able to nd a proper generalization. As a matter of fact, in a very essential step in the theory of Hardy spaces we apply this result.
36
Theorem 2.7 Let { } be a positive approximate identity on T. Suppose that, for each , L (T), and that { } satises the stronger property (c ): lim
sup
0<|t|
(eit )
= 0.
Let f be a real function in L1 (T) such that

tt0
lim f (eit ) = +.
Then, given M > 0, there exists (M, t0 ) and = (M, t0 ) > 0 such that ( f )(eit ) > M if > (M, t0 ) and |t t0 | < . In particular, lim ( f )(eit0 ) = +.
Proof. As we mentioned before, since L (T) and f L1 (T), by Corollary 1.6, ( f )(eit ) is well-dened for all eit T. By assumption, given M > 0, there exists = (M, t0 ) > 0 such that f (ei ) > 2M whenever | t0 | < 2. Therefore, for |t t0 | < , ( f )(eit ) = = = 1 2 1 2 M

(ei ) f (ei(t ) ) d

(ei ) f (ei(t ) ) d (ei ) |f (ei(t ) )| d
(ei ) d
1 2
| |
2M
1 2
| | 1)
(ei ) (2M + |f (ei(t ) )|) d ( sup

| |
2M (2M + f Pick (M, t0 ) so large that (2M + f

1)
(ei ) ).
sup
| |
(ei ) < M
whenever > (M, t0 ). Thus, for > (M, t0 ) and for |t t0 | < , ( f )(eit ) > M.
2.3. Uniform convergence and pointwise convergence Corollary 2.8 Let u be a real function in L1 (T), and let U (rei ) = Suppose that 1 2

37
1 r2 u(eit ) dt, 1 + r2 2r cos( t) lim u(eit ) = +.
(rei D).
tt0
Then U is harmonic in D and besides

zeit0 zD
lim U (z) = +.
In particular,
r1
lim U (reit0 ) = +.
Let A be a subset of C and consider a function f : A [0, ]. We say that f is continuous at z0 A whenever
zz0 zA
lim f (z) = f (z0 ).
If f (z0 ) < , then there is nothing new in this denition. The other case should be familiar too. If f (z0 ) = , by continuity at z0 we simply mean that limzz0 f (z) = . The following result is an immediate consequence of Corollaries 2.6 and 2.8. It plays a vital role in Fatous construction and thereafter in F. and M. Rieszs theorem, a cornerstone of function theory (see Section 5.5). Corollary 2.9 Let u : T [0, ] be continuous on T, and suppose that u L1 (T). Let 1 r2 1 u(eit ) dt if 0 r < 1, 2 2r cos( t) 2 1 + r i U (re ) = if r = 1. u(ei ) Then U : D [0, ] is continuous on D and harmonic on D.
Exercises
Exercise 2.3.1 Use Corollary 2.3 and (2.4) to give another proof of Weierstrasss theorem.
38 Exercise 2.3.2
Chapter 2. AbelPoisson means Let f Lip (T), 0 < < 1. Show that Kn f f
f Lip (T) , n
where C is an absolute constant just depending on . Exercise 2.3.3 Let f Lip1 (T). Show that Kn f f where C is an absolute constant. Exercise 2.3.4 Let f Lip1 (T). Show that Jn f f where C is an absolute constant. Exercise 2.3.5 Let f L1 (T), and let
n
C f Lip (T) 1
log n , n
f Lip (T) 1 , n
pn (eit ) =
k=n
|k| n+1
f (k) eikt ,
(eit T).
Suppose that f is continuous at eit0 T. Show that

n
lim pn (eit0 ) = fn (eit0 ).
Hint: Apply Theorem 2.5. Exercise 2.3.6 Let f be a real function in L1 (T) such that
tt0
lim f (eit ) = +.
Let pn (eit ) =
1
k=n
|k| n+1
f (k) eikt ,
(eit T).
Show that Hint: Apply Theorem 2.7.
lim pn (eit ) = +.
2.4. Weak* convergence of measures
39
2.4
Weak* convergence of measures
According to Theorem 1.3, f , the convolution of an L1 (T) function f and a Borel measure , is a well-dened L1 (T) function. In this section, we explore the relation between f and where f ranges over the elements of an approximate identity. Theorem 2.10 Let { } be an approximate identity on T, and let M(T). Then, for all , L1 (T) with and sup 1 .
1
Moreover, the measures d (eit ) = ( )(eit ) dt/2 converge to d(eit ) in the weak* topology, i.e. lim
1 2
(eit ) ( )(eit ) dt =
(eit ) d(eit )
T
for all C(T). Proof. By Theorem 1.3 and (1.11), L1 (T) with
1
for all . Let C(T), and dene (eit ) = (eit ). Then, C(T) and by Fubinis theorem, 1 2

(eit ) ( )(eit ) dt
= =
1 2
T
(eit )
T
(ei(t ) ) d(ei )
dt
=
T
=
T
1 2 1 2 1 2
(ei(t ) ) (eit ) dt
d(ei ) d(ei ) d(ei )
(ei(s ) ) (eis ) ds (ei( s) ) (eis ) ds
=
T
( )(ei ) d(ei ).
Theorem 2.2 assures that ( )(ei ) converges uniformly to (ei ) on T. Since || is a nite positive Borel measure, we thus have lim
1 2
(eit ) ( )(eit ) dt
= =
lim
T
( )(ei ) d(ei )
(ei ) d(ei )
T
=
T
(ei ) d(ei ).
40 Since 1 2

(eit ) ( )(eit ) dt ( sup
the last identity implies (ei ) d(ei ) ( sup

1
for all C(T). Hence, by the Riesz representation theorem, sup 1 .
In the last theorem, if { } is a positive approximate identity on T then C = 1 and thus we necessarily have = sup 1 .
As a matter of fact, slightly modifying the proof of the theorem, we see that = lim 1 .
In particular, if we choose the Poisson kernel then, by Theorem 2.1, we are able to extend to a harmonic function U on D such that the measures U (reit ) dt/2 are uniformly bounded and, as r 1, converge to in the weak* topology. Corollary 2.11 Let M(T), and let U (rei ) =
T
1 r2 d(eit ), 1 + r2 2r cos( t)
(rei D).
(2.5)
Then U is harmonic on D, and = sup

0r<1
Ur
= lim Ur
r1
1.
Moreover, the measures dr (eit ) = U (reit )dt/2 converge to d(eit ), as r 1 , in the weak* topology, i.e. lim (eit ) dr (eit ) = (eit ) d(eit )
T
r1
for all C(T). The preceding result has interesting and profound consequences. For example, the identity = sup0r<1 Ur 1 immediately implies some uniqueness theorems. We give two slightly dierent versions below.
2.4. Weak* convergence of measures Corollary 2.12 (Uniqueness theorem) Let M(T). Suppose that 1 r2 d(eit ) = 0 1 + r2 2r cos( t)
41
for all rei D. Then = 0. Corollary 2.13 (Uniqueness theorem) Let M(T) and suppose that (n) = 0 for all n Z. Then = 0. Proof. Dene U by (2.5) and note that, by Theorem 2.1, U (rei ) = for all rei T. The last uniqueness theorem says that the map M(T)
n=
(n) r|n| ein = 0
(Z)
is one-to-one. In particular, if f L1 (T) and f (n) = 0, for all n Z, then f = 0.
Exercises
Exercise 2.4.1 Show that Let {n } be an approximate identity on T, and let M(T). lim inf n 1 .
n
Hint: By Theorem 2.10, for each N 1, sup n 1 .

nN
Exercise 2.4.2 Let {n } be a positive approximate identity on T, and let M(T). Show that lim n 1 = .
n
Hint: Use Exercise 2.4.1 and the fact that n
for all n 1.
Exercise 2.4.3 Let 0 r, < 1. Show that Pr P = Pr . Hint: Use (1.6), Theorem 1.2 and the uniqueness theorem.
42 Exercise 2.4.4 Show that the map M(T) is not surjective. Exercise 2.4.5 Let M(T), and let
n
(Z)
pn (eit ) =
k=n
|k| n+1
(k) eikt ,
(eit T).
Show that the measures dn (eit ) = pn (eit ) dt/2 converge to d(eit ), as n , in the weak* topology, i.e.
n
lim
(eit ) dn (eit ) =
(eit ) d(eit )
T
for all C(T). Hint: Apply Theorem 2.10. Exercise 2.4.6 Let {an }nZ be a sequence of complex numbers. Suppose that the measures dpn (eit ) = 1 2
n
1
k=n
|k| n+1
ak eikt
dt
are convergent in the weak* topology, say to the measure M(T). Show that an are in fact the Fourier coecients of . (This result can be regarded as the converse of Exercise 2.4.5.) Exercise 2.4.7 Let M(T) and let r0 (0, 1). Suppose that
T
1+
2 r0
2 1 r0 d(eit ) = 0 2r0 cos( t)
for all ei T. Show that = 0. Exercise 2.4.8 that Let M(T) and let r0 (0, 1). Suppose that (0) = 0 and 2r0 sin( t) d(eit ) = 0 2 1 + r0 2r0 cos( t)
for all ei T. Show that = 0. Hint: For a xed r, nd the Fourier series of 2r sin t . 1 + r2 2r cos t The relation 2r sin t i i = 2 2r cos t it 1+r 1 re 1 reit might be useful.
2.5. Convergence in norm
43
2.5
Convergence in norm
The space of continuous functions C(T) is dense in Lp (T), 1 p < . This assertion is not true when p = , since the uniform limit of a sequence of continuous functions has to be continuous and a typical element of L (T) is not necessarily continuous. We exploit this fact to study the behavior of f , where f Lp (T) and { } is an approximate identity. Theorem 2.14 Let { } be an approximate identity on T, and let f Lp (T), 1 p < . Then, for all , f Lp (T) with f and besides, lim f f
p p
C f
= 0.
Proof. By Corollary 1.5, f Lp (T) and f

p
C f
for all . Fix > 0. Given f Lp (T), pick C(T) such that f f f
p
< . Hence
p p
(f ) (f ) + ( )
(f ) p + f p + (1 + C ) f p + p (1 + C ) + . However, by Theorem 2.2, there is an () such that > (). Therefore f f p < (2 + C ) whenever > (). Since lim f f p = 0, we clearly have lim f { } is a positive approximate identity on T, lim f
p p
< , for
= f
p.
Hence, if
= sup f
= f
p.
Therefore, if we use the Poisson kernel to extend a function u Lp (T) to the open unit disc we obtain a harmonic function U whose mean values Ur p are uniformly bounded, with Ur converging to u in the Lp (T) norm. Corollary 2.15 Let u Lp (T), 1 p < , and let U (rei ) = 1 2

1+
r2
(rei D).
44 Then U is harmonic on D,
0r<1
sup
Ur
= lim Ur
r1
= u
and
r1
lim Ur u
= 0.
We can also exploit the Fejr kernel Kn in Theorem 2.14. The main advane tage is that f Kn is a trigonometric polynomial. Corollary 2.16 Let f Lp (T), 1 p < , and let
n
pn (eit ) =
k=n
|k| n+1 pn
f (k) eikt ,
(eit T).
Then, for each n 1,

p
f
p
and
n
lim
pn f
= 0.
Proof. It is enough to observe that pn = f K n , where Kn is the Fejr kernel. Now, apply Theorem 2.14. e The constructive method of Fejr shows that the trigonometric polynomials are e dense in Lp (T), 1 p < . On the other hand, in Lemma 1.1, we saw that the Fourier coecients of an integrable function are uniformly bounded. We are now in a position to improve this result by showing that the Fourier coecients actually tend to zero. Corollary 2.17 (RiemannLebesgue lemma) Let f L1 (T). Then
|n|
lim f (n) = 0.
Proof. Given > 0, by Corollary 2.16, there is n0 such that pn0 f where pn0 (eit ) = Hence, by Lemma 1.1, |n0 (n) f (n)| pn0 f p |f (n)| < for all |n| > n0 .
1 n0 1
< , f (k) eikt , (eit T).
1
k=n0
|k| n0 + 1
<
for all n Z. However, pn0 (n) = 0 if |n| > n0 . Therefore,
2.5. Convergence in norm We saw that the Fourier transform M(T)
45
(Z)
is injective but not surjective. As long as p (Z) spaces or their well-known subspaces are concerned, (Z) is the best possible choice in this mapping. However, if instead of M(T) we consider the smaller subclass L1 (T), by the RiemannLebesgue lemma (Corollary 2.17), we can slightly improve the preceding mapping and exhibit the Fourier transform on L1 (T) as L1 (T) c0 (Z) f f. We will see that this map is not surjective (see Exercise 2.5.6). As a matter of fact, there is no satisfactory description for the image of L1 (T ) under the Fourier transformation.
Exercises
Exercise 2.5.1 Let {an }nZ be a sequence of complex numbers, and let
n
pn (eit ) =
k=n
|k| n+1
ak eikt ,
(eit T).
Let 1 p < . Suppose that the sequence (pn )n0 is convergent in the Lp (T) norm to a function f Lp (T). Show that {an }nZ are in fact the Fourier coecients of f . Remark: This result can be regarded as the converse of Corollary 2.16. Exercise 2.5.2 Show that Lp (T), 1 p < , is separable. Is L (T) separable? Hint: Use Corollary 2.16. Exercise 2.5.3 [Fejrs lemma] Let f L1 (T) and let g L (T). Show that e 1 n 2 lim

f (eit ) g(eint ) dt = f (0) g (0).
Hint: First suppose that f is a trigonometric polynomial and apply Exercise 1.3.4. Then use Corollary 2.16 to prove the general case. Exercise 2.5.4 Let g L1 (T) and let 1 p . Consider the operator : Lp (T) Lp (T) f g f.
46
By Corollary 1.5, is well-dened. Show that = g 1. Hint: Apply Corollaries 1.5 and 2.16. Exercise 2.5.5 Let f L1 (T). Suppose that f (n) = f (n) 0 for all n 0. Show that f (n) < . n n=1 F (eit ) =
0 t
Hint: Let
f (ei ) d.
Then F C(T) and F (n) = f (n)/in, n = 0. Hence, by Corollary 2.4,

n
lim (Kn F )(1) = F (1) = 0.
Exercise 2.5.6
Show that the mapping L1 (T) c0 (Z) f f
is not surjective. Hint: Let an = an =
1 , log n
(n 2),
and a1 = a0 = 0. Use Exercise 2.5.5 to show that (an )nZ is not in the range. Exercise 2.5.7 Let (an )nZ be such that
(i) an 0, for all n Z, (ii) an = an , for all n 1, (iii) limn an = 0, (iv) an (an1 + an+1 )/2, for all n 1. Let f (e ) =
n=1 it
n(an1 + an+1 2an ) Kn1 (eit ).
2.6. Weak* convergence of bounded functions Show that f L1 (T) and that f (n) = an
47
for all n Z. Remark: This result convinces us that c0 (Z) is somehow optimal as the codomain of L1 (T) c0 (Z) f f.
2.6
Weak* convergence of bounded functions
Since C(T) is not uniformly dense in L (T), the results of the preceding section are not entirely valid if p = . However, a slightly weaker version holds in this case too. Theorem 2.18 Let { } be an approximate identity on T, and let f L (T). Then, for all , f C(T) with f and f

C f
sup f
Moreover, f converges to f in the weak* topology, i.e.
lim

(eit ) ( f )(eit ) dt =
(eit ) f (eit ) dt
for all L1 (T). Proof. By Corollary 1.6, f C(T) and f
C f
for all . Let L1 (T), and let (eit ) = (eit ). Then, by Fubinis theorem,

(eit ) ( f )(eit ) dt
(eit )

(ei(t ) ) f (ei ) d
dt
(ei(t ) ) (eit ) dt (ei( s) ) (eis ) ds
f (ei ) d f (ei ) d
( )(ei ) f (ei ) d.
48
Theorem 2.14 ensures that ( )(ei ) converges to (ei ) in L1 (T). Since f is a bounded function, we thus have
lim

(eit ) ( f )(eit ) dt
= =
lim

( )(ei ) f (ei ) d
(ei ) f (ei ) d (ei ) f (ei ) d.
Since
1 2
(eit ) ( f )(eit ) dt ( sup f
) 1,
the last identity implies

(ei ) f (ei ) d ( sup f
for all L1 (T). Hence, by the Riesz representation theorem, f
sup f
If { } is a positive approximate identity on T then C = 1 and thus we have f
= sup f
As a matter of fact, by slightly modifying the proof of the theorem, we obtain f
= lim f
Corollary 2.19 Let u L (T), and let U (rei ) = 1 2

1+
r2
(rei D).
Then U is bounded and harmonic on D and

0r<1
sup
Ur
= lim Ur
r1
= u
Moreover, as r 1, Ur converges to u in the weak* topology, i.e.

r1
lim
(eit ) U (reit ) dt =
(eit ) u(eit ) dt
for all L1 (T).
2.7. Parsevals identity
49
Exercises
Exercise 2.6.1 Let {n } be an approximate identity on T, and let f L (T). Show that f lim inf n f .
n
Exercise 2.6.2 Let {n } be a positive approximate identity on T, and let f L (T). Show that
n
lim
n f
= f
Exercise 2.6.3
Let f L (T) and let

n
pn (eit ) =
k=n
|k| n+1
f (k) eikt ,
(eit T).
Show that the sequence (pn )n0 converges to f in the weak* topology, i.e.
n
lim
(eit ) pn (eit ) dt =
(eit ) f (eit ) dt
for all L1 (T). Hint: Apply Theorem 2.18. Exercise 2.6.4 Let {an }nZ be a sequence of complex numbers, and let
n
pn (eit ) =
k=n
|k| n+1
ak eikt ,
(eit T).
Suppose that the sequence (pn )n0 is convergent in the weak* topology to a function f L (T). Show that {an }nZ are in fact the Fourier coecients of f. Remark: This result can be regarded as the converse of Exercise 2.6.3.
2.7
Parsevals identity
In the preceding sections we mainly studied the AbelPoisson means of the Fourier series of a measure or a function and, among other things, we saw how these means converge in an appropriate topology to the given measure or function. However, we have not yet considered the convergence of the Fourier series itself. The pointwise or uniform convergence of the Fourier series is a more subtle problem. The whole story is unveiled by a celebrated theorem of Carleson-Hunt, which says that the Fourier series of a function in Lp (T),
50
p > 1, converges almost everywhere to the function, and a dicult construction of Kolmogorov giving a function in L1 (T) whose Fourier series diverges almost everywhere. We do not need these results in the following. However, we explore further the Fourier series of L2 (T) functions. First of all, for the subclass L2 (T) L1 (T), the uniqueness theorem (Corollary 2.13) can be stated dierently. A family { } in L2 (T) is complete provided that
f (eit ) (eit ) dt = 0,
for all
holds only if f = 0. Therefore, the uniqueness theorem says that the sequence {eint }nZ is complete in L2 (T). Secondly, L2 (T) equipped with the inner product 1 f, g = f (eit ) g(eit ) dt 2 is a Hilbert space. Two functions f, g L2 (T) are said to be orthogonal if f, g = 0. A subset S L2 (T) is called an orthonormal set if every element of S has norm one and every two distinct elements of S are orthogonal. Using this terminology, the relation (1.2) along with the uniqueness theorem tells us that the sequence {eint }nZ is a complete orthonormal set. Lemma 2.20 (Bessels inequality) Let f L2 (T), and let { } be an orthonormal family in L2 (T). Then | f, |2 f
2 2.
Proof. Let = f, , where the sum is over a nite subset of indices {}. Then f
2 2
= =
f , f f, f f, , f + , .
But f, = = = = 1 2 1 2

f (eit ) (eit ) dt f (eit ) 1 2

f, (eit ) f (eit ) (eit ) dt | f, |2
dt
f,
f, f, =
2.7. Parsevals identity and similarly, , = = = = Hence f which gives | f, |2 f

2 2.
51
1 2 1 2
(eit ) (eit ) dt f, (eit ) f, f, 1 2

f,
(eit )
dt
(eit ) (eit ) dt
f, f, =
2 2 2 2
| f, |2 . | f, |2 ,
= f
Taking the supremum with respect to all such sums gives the required result. If we consider the orthonormal family {eint }nZ , then Bessels inequality is written as 1 |f (n)|2 |f (eit )|2 dt 2 n= or equivalently, f
2
(Z) with f
2
for each f L2 (T). Therefore, the mapping L2 (T) f

2
(Z) f
is well-dened and, by the uniqueness theorem (Corollary 2.13), is injective. Now, we show that it is also surjective. Theorem 2.21 (RieszFischer theorem) Let (an )nZ an f L2 (T) such that f (n) = an for all n Z. Proof. Let and let fn =
k=n 2
(Z). Then there is
k (eit ) = eikt ,
n
(k Z),
ak k .
52 Hence fn (k) = fn , k = Let m > n. Then fm fn

2 2
Chapter 2. AbelPoisson means ak 0

m
if n |k|, if n < |k|. |ak |2
=
|k|=n+1
and thus (fn )n1 is a Cauchy sequence in L2 (T). Hence it is convergent, say to f L2 (T). Therefore, for each k Z, we have f (k) = f, k = lim fn , k = ak .
n
The proof of the RieszFischer theorem contains more than what was stated n in the theorem. We saw that limn fn f 2 = 0, where fn = k=n ak k . Hence, f
2 2
= lim
fn
2 2
=
k=
|ak |2 .
Bessels inequality (Lemma 2.20) ensures that f 2 (Z), whenever f L2 (T). 2 (k), an appeal to the uniqueness Hence, given f L (T), if we pick ak = f theorem (Corollary 2.13) shows that
n
lim
fn f
n
= 0,
(2.6)
where fn (eit ) =
f (k) eikt .
k=n
This result is an improvement of Corollary 2.16 when p = 2 (see also Exercise 2.1.3). Moreover, f
2 2
=
k=
|f (k)|2 .
This last identity is very important and we state it as a corollary. Corollary 2.22 (Parsevals identity) Let f L2 (T). Then 1 2

|f (eit )|2 dt =
n=
|f (n)|2 .
Parsevals identity can be rewritten as f

2
= f 2.
2.7. Parsevals identity Hence the Fourier transform L2 (T) f

2
53
(Z) f
2
is bijective and shows that the Hilbert spaces L2 (T) and isometric.
(Z) are isomorphically
Exercises
Exercise 2.7.1 [Polarization identity] Let H be a complex inner product space and let x, y H. Show that 4 x, y = x + y
2
xy
+ i x + iy
i x iy 2 .
Exercise 2.7.2
Let f, g L2 (T). Show that 1 2

f (eit ) g(eit ) dt =
f (n) g (n).
n=
Hint: Use Parsevals identity and Exercise 2.7.1.
54
Chapter 3
Harmonic functions in the unit disc

3.1 Series representation of harmonic functions
Let U be a harmonic function on the disc DR = { |z| < R }. In the proof of the following theorem, we will see that there is another harmonic function V such that F = U + iV is analytic on DR . Such a function V is called a harmonic conjugate of U . It is determined up to an additive constant and we usually normalize it so that V (0) = 0. Remember that 1 if n > 0, 0 if n = 0, sgn(n) = 1 if n < 0. Theorem 3.1 Let U be harmonic on the disc DR . Then, for each n Z, the quantity |n| an = U ( eit ) eint dt, (0 < < R), (3.1) 2 is independent of , and we have U (rei ) = The function V (rei ) =
n= n=
an r|n| ein ,
(rei DR ).
(3.2)
i sgn(n) an r|n| ein ,
(rei DR ),
(3.3)
is the unique harmonic conjugate of U such that V (0) = 0. The series in (3.2) and (3.3) are absolutely and uniformly convergent on compact subsets of DR . 55
56
Chapter 3. Harmonic functions in the unit disc
Proof. Without loss of generality, assume that U is real. Let G(z) = U U (z) i (z). x y
Since U satises the Laplace equation (1.1), the real and imaginary parts of G satisfy the CauchyRiemann equations. Hence, G is analytic on DR . Let
z
F (z) = U (0) +
0
G(w) dw.
Since DR is simply connected, F is well-dened (the value of the integral is independent of the path of integration from 0 to z in DR ) and we have F = G. Let U = F and note that U(0) = U (0) and F (0) = 0. Now, on the one hand, F (z) = G(z) = U U (z) i (z) x y
and, on the other hand, by the CauchyRiemann equations F (z) = Hence (U U) (z) = 0 x U U (z) i (z). x y (U U) (z) = 0 y F,
and
on DR , which along with U(0) = U (0) imply U U. Thus, writing V for we have F = U + iV
with V (0) = 0. Since F is analytic on DR it has the unique power series representation F (rei ) = with
n=0
n rn ein ,
n=0
|n | rn <
(3.4)
for all r < R. Hence U (rei ) = F (rei )

n=1
=
n=0
n rn ein 1 2
n=1
1 = 0 + 2
n rn ein +
n rn ein
=
n=
an r|n| ein ,
3.1. Series representation of harmonic functions where n /2 if n > 0, 0 if n = 0, an = n /2 if n < 0.
57
Condition (3.4) ensures the absolute and uniform convergence of

n=
an r|n| ein
on compact subsets of DR . Moreover, for each m Z and 0 < r < R, r|m| 2

U (reit ) eimt dt
r|m| 2
n=
n=
an r|n| eint 1 2

eimt dt
= r|m| = am . Finally, we have V (rei ) = = =

n=
an r|n|
ei(nm)t dt
F (rei ) 1 2i
= 1 2i
n rn ein n rn ein
n rn ein
n=0 n=1
n=1
i sgn(n) an r|n| ein .
Condition (3.4) also implies the absolute and uniform convergence of this series on compact subsets of DR . Based on the content of the preceding theorem, the conjugate of any trigonometric series
S=
n=
an ein
is dened by S=
n=
i sgn(n) an ein .
A special case of (3.1), corresponding to m = 0, will be used often. We mention it as a corollary.
58
Corollary 3.2 Let U be harmonic on DR . Then U (0) = for all r, 0 r < R. Proof. It is enough to note that a0 = U (0). 1 2

U (rei ) d
Exercises
Exercise 3.1.1 Let U be harmonic on DR . Show that 1 r2 for all r, 0 < r < R. Hint: Use Corollary 3.2. Exercise 3.1.2 Suppose that Let U be a real harmonic function on C and let N 0. U (rei ) c rN + c (3.5)
0 r
U ( ei ) d d = U (0)
for all r 0 and all ; c and c are two positive constants. Show that
N
U (rei ) =
n=N
an r|n| ein ,
where the coecients an are given by (3.1). Hint: By Corollary 3.2 and (3.1) a0 r|n| an = 1 2

U (reit ) (1 cos(nt)) dt.
The advantage of this representation is that (1 cos(nt)) 0. Remark: We emphasize that in (3.5) we have U and not |U |. Not using the absolute values is crucial in some applications. Exercise 3.1.3 Let be a simply connected domain and let U be a harmonic function on . Show that there is a harmonic function V on such that F = U + iV is analytic over . Moreover, show that V is unique up to an additive constant. Hint: See the rst part of the proof of Theorem 3.1.
3.2. Hardy spaces on D
59
3.2
Hardy spaces on D
The family of all complex harmonic functions on the open unit disc D is denoted by h(D). Let U h(D) and write U
p
= sup
0r<1
Ur
= sup
0r<1
1 2
2 0
|U (r e )| d
1 p
if p (0, ), and U We dene hp (D) =
= sup | U (z) |.
zD
U h(D) :
< ,
where p (0, ]. It is straightforward to see that hp (D), 1 p , is a normed vector space and, by Hlders inequality, o h (D) hq (D) hp (D) if 0 < p < q < . We will see that h1 (D) and hp (D), 1 < p , are Banach spaces respectively isomorphic to M(T) and Lp (T), 1 < p . A complex harmonic function F is simply of the form F = U + iV where U and V are real harmonic functions and there is no other relation between U and V . However, if we assume that V is a harmonic conjugate of U , and thus F is analytic on D, a whole new family of functions with profound properties emerges. Let us denote the family of all analytic functions on D by H(D). Hence, H(D) h(D). Then, parallel to our previous denitions, we consider the Hardy classes of analytic functions on the unit disc H p (D) = for 0 < p . Clearly, F H(D) : F
p
<
H p (D) hp (D).
As a matter of fact, that is why we assumed that the elements of h(D) are complex-valued harmonic functions. As a consequence, any representation theorem for hp (D) functions is also automatically valid for the elements of the smaller subclass H p (D). The Hardy space H p (D), 1 p , is a normed vector space and, by Hlders inequality, o H (D) H q (D) H p (D) if 0 < p < q < . We will see that H p (D), 1 p , is also a Banach space isomorphic to a closed subspace of Lp (T) denoted by H p (T). Using these new notations, Corollaries 2.11, 2.15 and 2.19 can be rewritten as follows. Theorem 3.3 Let 1 p . If u Lp (T), then U = P u hp (D) and U p = u p . If M(T), then U = P h1 (D) and U 1 = .
60
In the following sections, we study the converse of this theorem. More precisely, we start with a harmonic function in hp (D) and show that it can be represented as P u or P with a suitable function u or measure .
Exercises
Exercise 3.2.1 Let F = U + iV be analytic on D, and let 0 < p . Show that F H p (D) if and only if U and V are real harmonic functions in hp (D). Exercise 3.2.2 Let 0 < p < q < . Show that h (D) and that H (D) hq (D) H q (D) hp (D) H p (D).
Exercise 3.2.3
Let u L2 (T). Dene F (z) = 1 2

eit + z u(eit ) dt, eit z
(z D).
Show that F H 2 (D) and F Deduce F

2 2 2
= |(0)|2 + 4 u
n=1
|(n)|2 . u
2 u
and show that 2 is the best possible constant. Hint: Use Exercise 1.3.3 and Parsevals identity (Corollary 2.22).
3.3
Poisson representation of h (D) functions
In studying the boundary values of harmonic functions on the unit disc, the best possible assumption is to consider harmonic functions which are actually dened on discs larger than the unit disc. Hence, let us consider h(D) = { U : 2 U = 0 on |z| < R, for some R > 1 }. The constant R is not universal and it depends on U . Putting such a strong assumption on the elements of h(D) makes it the smallest subclass of h(D) in our discussion.
3.3. Poisson representation of h (D) functions Lemma 3.4 Let U h(D). Then U (rei ) = 1 2

61
1 r2 U (eit ) dt, 1 + r2 2r cos( t)
(rei D).
(3.6)
Proof. Since U h(D), there exists R > 1 such that U is harmonic on DR . Hence, by Theorem 3.1, for all rei DR , we have U (rei ) = where an are given by an = 1 2
n=
an r|n| ein ,
U (eit ) eint dt,
(n Z).
In particular, for all rei D, we obtain U (rei ) =

n=
1 2
U (eit ) eint dt
r|n| ein .
Fix r and . Since U is bounded on T, the absolute and uniform convergence of the series n= r|n| ein(t) U (eit ), as a function of t, enables us to change the order of summation and integration. Hence, U (rei ) = 1 2
n=
r|n| ein(t)
U (eit ) dt,
(rei D).
But, as we saw in (1.5),

n=
r|n| ein(t) =
1+
r2
1 r2 . 2r cos( t)
In the following we show that the integral representation (3.6) is valid for some larger subclasses of h(D). Since h(D) h (D), the following result is the rst generalization of Lemma 3.4. Theorem 3.5 (Fatou [6]) Let U h (D). Then there exists a unique u L (T) such that U (rei ) = and U
1 2
1 r2 u(eit ) dt, 1 + r2 2r cos( t) = u

.
(rei D),
62
Remark: Compare with Corollary 2.19. Proof. The uniqueness is a consequence of Corollary 2.12. The rest of the proof is based on the Poisson representation of harmonic functions in h(D) and the following two facts: (i) L (T) is the dual of L1 (T); (ii) L1 (T) is a separable space (see Exercise 2.5.2). Step 1: Picking a family of bounded linear functionals on L1 (T). Put Un (z) = U 1 1 z , n
(n 2).
First of all, Un is dened on the disc {|z| < n/(n 1)}. Hence, Un h(D) and by Theorem 3.4, U 1 1 z n = 1 2

1 r2 Un (eit ) dt 1 + r2 2r cos( t)
(3.7)
for all rei D. Let n . The left side clearly tends to U (reit ). We show that the limit of the right side has an integral representation. Dene n : L1 (T) C by n (f ) = Since |n (f )| Un
1 2
f (eit ) Un (eit ) dt, U
(n 2). f
1,
(3.8)
each n is a bounded linear functional on L1 (T) with n U

.
Step 2: Extracting a convergent subsequence of n . At this point we can use the BanachAlaoghlu theorem and deduce that n has a convergent subsequence in the weak* topology. In other words, there is a bounded linear functional on L1 (T) and a subsequence nk such that
k
lim nk (f ) = (f )
for all f L1 (T). However, we give a direct proof of this fact. Take a countable dense subset of L1 (T), say {f1 , f2 , . . . }. By (3.8), there is a subsequence of {n }n2 , say {n1j }j1 , such that
j
lim n1j (f1 )
3.3. Poisson representation of h (D) functions
63
exists. Again by (3.8), there is a subsequence of {n1j }j1 , say {n2j }j1 , such that lim n2j (f2 )
j
exists. Continuing this process, for any i 1, we nd a subsequence {nij }j1 of {n(i1)j }j1 such that limj nij (fi ) exists. To apply Cantors method, consider the diagonal subsequence {nkk }k1 . Since {nkk }k1 is eventually a subsequence of {nij }j1 , the limit
k
lim nkk (fi )
exists for any i 1. Moreover, we show that this limit actually exists for all f L1 (T). To do so, we use the fact that n s are uniformly bounded and fi s are dense in L1 (T). Fix f L1 (T) and > 0. Hence, there is fi such that f fi 1 < . By (3.8), we have | nkk (f ) nll (f ) | | nkk (f ) nkk (fi ) | + | nkk (fi ) nll (fi ) | + | nll (fi ) nll (f ) | 2 U 2 U Picking k, l large enough, we obtain | nkk (f ) ll (f ) | (2 U Hence (f ) = lim nkk (f )
k
f fi
+ | nkk (fi ) nll (fi ) |
+ | nkk (fi ) nll (fi ) |.
+ 1) .
exists for all f L1 (T) and, again by (3.8), |(f )| U
1. .
In other words, is a bounded linear functional on L1 (T) with U Step 3: Appealing to Rieszs theorem. For a xed z = rei , fz (eit ) = Pr (ei(t) ) = 1 r2 , 1 + r2 2r cos( t)
as a function of t, is in L1 (T). Hence, by (3.7), we have (fz ) = = =

k
lim nkk (fz )

1 k 2 lim
k
Pr (ei(t) ) Unkk (eit ) dt 1 rei nkk = U (rei ).
lim U
64
On the other hand, by Rieszs theorem, there is a u L (T) such that (f ) = 1 2

f (eit ) u(eit ) dt
for all f L1 (T). First of all, by choosing f = fz , we obtain U (rei ) = 1 2

Pr (ei(t) ) u(eit ) dt.
Secondly, by Corollary 2.19, this identity implies U
= u
Exercises
Exercise 3.3.1 that U (a + rei ) = 1 2 Let U be harmonic on a domain and let D(a, R) . Show

R2 r 2 U (a + Reit ) dt, R2 + r2 2Rr cos( t)
(0 r < R).
Hint: Use Lemma 3.4 and make a change of variable. Exercise 3.3.2 Let 0 r0 < 1. Show that U (rei ) = and V (rei ) =
2 r(1 + r0 ) cos r0 (1 + r2 ) 2 1 + r0 r2 2r0 r cos 2 r(1 r0 ) sin 2 1 + r0 r2 2r0 r cos
are bounded harmonic functions on D. Find the Fourier series expansions of Ur and Vr . Hint: Besides the direct verication of each fact, it might be easier to show that U + iV is a bounded analytic function on D. Exercise 3.3.3 Find a bounded harmonic function on the unit disc such that its conjugate is not bounded. Exercise 3.3.4 Let U be harmonic on D. Show that there exists a unique u C(T) such that U (rei ) = 1 2

1 r2 u(eit ) dt, 1 + r2 2r cos( t)
(rei D),
if and only if the family (Ur )0r<1 is Cauchy in C(T) as r 1.
3.4. Poisson representation of hp (D) functions (1 < p < )
65
Exercise 3.3.5 [Harnacks theorem] Let (Un )n1 be a sequence of harmonic functions on a domain . Suppose that on each compact subset of , Un converges uniformly to U . Show that U is harmonic on . Hint: Use Corollary 2.3 and Exercise 3.3.1.
3.4
Poisson representation of hp (D) functions (1 < p < )
The proof of Theorem 3.5 can be modied slightly to give a representation formula for hp (D), 1 < p < , functions. The modication is based on the following two facts: (i) Lp (T) is the dual of Lq (T), where 1/p + 1/q = 1; (ii) Lp (T) is a separable space. Since h(D) h (D) hp (D), the following result is a generalization of Theorems 3.4 and 3.5. Theorem 3.6 Let U hp (D), 1 < p < . Then there exists a unique u Lp (T) such that U (rei ) = and U
p
1 2
1+
r2
1 r2 u(eit ) dt, 2r cos( t) = u p.
(rei D),
Remark: Compare with Corollary 2.15. If U h2 (D) this result, along with Theorem 2.1, gives us a unique u L2 (T) such that U (rei ) = 1 2

1 r2 u(eit ) dt = u(n) r|n| ein 2 2r cos( t) 1+r n=
for all rei D. Hence, by Parsevals identity (Corollary 2.22), 1 2

|U (rei )|2 d =
n=
|(n)|2 r2|n| . u
Thanks to the uniform and bounded convergence of the series n u(n) r|n| ein for each xed r < 1, and the fact that (ein )nZ is an orthonormal family, the preceding identity can also be proved by direct computation. Now, by the monotone convergence theorem, and again by Parsevals identity (this time we really need it), if we let r 1, we obtain
= u
=
n=
|n |2 u
1 2
66
There is no such relation between a function and its Fourier coecients in other Lp classes.
Exercises
Exercise 3.4.1 Show that U (rei ) = log(1 + r2 2r cos ) h2 (D). Hint: Study the analytic function f (z) = 2 log(1 z) to nd the Fourier series expansion of Ur . Exercise 3.4.2 Find V , the harmonic conjugate of U (rei ) = log(1 + r2 2r cos ). Do we have V h2 (D)?
3.5
Poisson representation of h1 (D) functions
In the proof of Theorems 3.5 and 3.6 we used the fact that Lp (T) is the dual of Lq (T), whenever 1 1 is essential and the suggested proof for 1 < p < does not work if p = 1. To overcome this diculty, we consider L1 (T) as a subset of M(T). By Rieszs theorem, M(T) is the dual of C(T). Now we proceed as in the proof of Theorem 3.5. The only dierence is that this time there is a unique Borel measure that represents our continuous linear functional on C(T). Therefore, an element of h1 (D) is represented by the Poisson integral of a measure (and not necessarily of an L1 (T) function). Since h(D) h (D) hp (D) h1 (D), 1 < p < , the following result is the last step in the generalization of Theorems 3.4, 3.5 and 3.6. Theorem 3.7 Let U h1 (D). Then there exists a unique M(T) such that U (rei ) =
T
1+
r2
1 r2 d(eit ), 2r cos( t) U
1
(rei D),
and = . Remark: Compare with Corollary 2.11. Let U be a positive harmonic function on D. As a convention in this context, positive means 0 (see Exercise 3.5.7). Then, by Corollary 3.2, 1 2

|U (rei )| d =
1 2
U (rei ) d = U (0),
3.5. Poisson representation of h1 (D) functions
67
and thus U h1 (D). Hence we can apply Theorem 3.7 and obtain a measure whose Poisson integral is U . But a closer look at its proof suggests an improvement. Here, the measures dn (eit ) = U ((1 1/n)eit ) dt/2 are positive and thus their weak limit d has to be positive too. Therefore, we arrive at the following special case of Theorem 3.7. Theorem 3.8 (Herglotz) Let U be a positive harmonic function on D. Then there exists a unique nite positive Borel measure on T such that U (rei ) =
T
1+
r2
1 r2 d(eit ), 2r cos( t)
(rei D).
Herglotzs theorem provides an easy proof of Harnacks inequality. The main feature of Harnacks inequality is that the constants appearing in the lower and upper bounds do not depend on the function U . Corollary 3.9 (Harnacks inequality) Let U be a positive harmonic function on D. Then, for each rei D, 1r 1+r U (0) U (rei ) U (0). 1+r 1r Proof. By Theorem 3.8, U is the Poisson integral of a positive measure . Since, for all and t, 1 r2 1+r 1r , 2 2r cos( t) 1+r 1+r 1r and since d 0, we have 1r 1+r
T
d(eit )
1+
r2
1 r2 1+r d(eit ) 2r cos( t) 1r
d(eit ).
T
But, again according to Theorem 3.8, d(eit ) = U (0).

T
Exercises
Exercise 3.5.1 Show that the Poisson kernel P (rei ) = 1 r2 , 1 + r2 2r cos
as a positive harmonic function on D, is in hp (D), 0 1.
68
Exercise 3.5.2 Let U be a real harmonic function on D. Show that U h1 (D) if and only if U is the dierence of two positive harmonic functions. Hint: Use Theorems 3.7 and 3.8 and the fact that each M(T) can be decomposed as = 1 2 where 1 and 2 are nite positive Borel measures.
Exercise 3.5.3 Let U be harmonic on D. Show that there exists a unique u L1 (T) such that U (rei ) = 1 2

1 r2 u(eit ) dt, 1 + r2 2r cos( t)
(rei D),
if and only if the family (Ur )0r<1 is Cauchy in L1 (T) as r 1. Exercise 3.5.4 [Generalized Harnacks inequality] Let U be a positive harmonic function on the disc D(a, R) = {z : |z a| < R}. Show that, for all 0 r < R and , Rr R+r U (a) U (a + rei ) U (a). R+r Rr Exercise 3.5.5 Let U be the collection of all positive harmonic functions U on D with U (0) = 1. Find sup U (1/2)
U U
and
U U
inf U (1/2).
Hint: Use Harnacks inequality (Corollary 3.9). Exercise 3.5.6 Let be a domain in C. Fix z and w in . Show that there exists 1 such that, for every positive harmonic function U on , 1 U (w) U (z) U (w). Remark: (z, w), the Harnack distance between z and w, is by denition the smallest satisfying the last relation.
Exercise 3.5.7 Let U be a positive harmonic function on a domain . Show that either U is identically zero on or it never vanishes there. Hint: Use Exercise 3.5.6.
3.5. Poisson representation of h1 (D) functions Exercise 3.5.8
69
Let f be a conformal mapping between and . Show that (z, w) = (f (z), f (w))
for all z, w . Exercise 3.5.9 Show that D (0, r) = 1+r , 1r (0 r < 1).
Then use the conformal mapping (z) = ei z , 1 z (|| < 1, R),
(an automorphism of the unit disc) to show that D (z, w) = 1+ 1

zw 1z w zw 1z w
for all z, w D. Remark: The content of Section 7.1 might help. Exercise 3.5.10 Let U be a positive harmonic function on D. Show that |U (0)| 2 U (0), where U = U/x + i U/y. Hint: Use Harnacks inequality and the fact that U (0) is the maximum directional derivative at zero, i.e. |U (0)| = sup U (rei ) U (0) . r0 r lim
Exercise 3.5.11 [Liouville] Use Harnacks inequality to show that every positive harmonic function on C is constant. Exercise 3.5.12 Let (Un )n1 be a sequence of positive harmonic functions on a domain , and let z0 . Suppose that
n
lim Un (z0 ) = .
Show that Un converges uniformly to innity on compact subsets of . Hint: Use Exercise 3.3.1.
70
Exercise 3.5.13 Let (Un )n1 be a sequence of positive harmonic functions on a domain , and let z0 . Suppose that
n
lim Un (z0 ) = 0.
Show that Un converges uniformly to zero on compact subsets of . Hint: Use Exercise 3.3.1. Exercise 3.5.14 [Harnacks theorem] Let (Un )n1 be an increasing sequence of harmonic functions on a domain . Show that either Un converges uniformly on compact subsets of to a harmonic function, or it converges uniformly to innity on each compact subset. Hint: Use Exercises 3.3.5, 3.5.12 and 3.5.6. Exercise 3.5.15 [Herglotz] Let f be an analytic function on the unit disc with values in the right half plane { z > 0}. Suppose that f (0) > 0. Show that there exists a positive Borel measure on T, say M(T), such that f (z) =
T
eit + z d(eit ) eit z
for all z D. Hint: Use Theorem 3.8 and Exercise 1.3.2. Exercise 3.5.16 Let U be a positive harmonic function on D. We know that there exists a positive Borel measure on T such that U (rei ) =
T
1 r2 d(eit ), 1 + r2 2r cos( t) U
1
(rei D).
Show that = U (0) = = (T).
3.6
Radial limits of hp (D) functions (1 p )
The following result is a generalization of Fatous theorem, which is about the boundary values of bounded analytic functions in the unit disc. However, his theorem works in a more general setting. Lemma 3.10 Let u L1 (T), and let U (rei ) = Then for almost all ei T. 1 2

1 r2 u(eit ) dt, 1 + r2 2r cos( t) lim U (rei ) = u(ei )
(rei D).
r1
(3.9)
3.6. Radial limits of hp (D) functions (1 p ) Proof. According to a classical result of Lebesgue, 1 t0 2t lim
+t t
71
u(eis ) ds = u(ei )
for almost all ei T. We prove that at such a point (3.9) holds. Without loss of generality, assume that = 0. Put
x
U(x) =
u(eit ) dt,
x [, ].
Then, doing integration by parts, we obtain
2 U (r)
1 r2 u(eit ) dt 1 + r2 2r cos t
t= 2
= = =
1 r2 U(t) 1 + r2 2r cos t 1r U() + 1+r

1 r2 1 + r2 2r cos t
U(t) dt
(1 r ) 2r sin t U(t) dt (1 + r2 2r cos t)2

2r 1r U() + 1+r 1+r Fr (eit ) =
(1 + r)2 (1 r) t sin t U(t) U(t) dt. 2 2r cos t)2 (1 + r 2t
But we know that
(1 + r)2 (1 r) t sin t (1 + r2 2r cos t)2

t t
is a positive approximate identity on T and, by assumption, U(t) U(t) 1 = lim t0 t0 2t 2t lim In other words, the function (eit ) = u(eis ) ds = u(1).
U(t)U(t) 2t
if if
0 < |t| , t=0
u(1)
is continuous at t = 0, and moreover, (eit ) = (eit ). Hence, by Theorem 2.5,

r1
lim U (r)
1r 2r 1 U(t) U(t) U() + dt Fr (eit ) r1 2 1+r 1 + r 2t 1 = lim Fr (eit ) (eit ) dt r1 2 = lim (Fr )(1) = (1) = u(1).
lim
r1
72
A slight modication of the preceding lemma yields the following result about harmonic functions generated by singular measures. Some authors prefer to combine both results and present them as a single theorem, as we do later in this section. Lemma 3.11 Let M(T) be singular with respect to the Lebesgue measure, and let 1 r2 U (rei ) = d(eit ). 2 2r cos( t) 1+r T Then
r1
lim U (rei ) = 0
(3.10)
for almost all ei T. Proof. Since is singular with respect to the Lebesgue measure, we have {eis : s ( t, + t)} t0 2t lim =0
for almost all ei T. We prove that at such a point (3.10) holds. Without loss of generality, assume that = 0. First, we extract the Dirac measure (if any) at the point 1. Hence, U (r) = = Since 1 r2 d(eit ) 2 T 1 + r 2r cos t 1 r2 ({1}) + 1 + r2 2r cos
T\{1}
1 r2 d(eit ). 1 + r2 2r cos t
1 r2 = 0, r1 1 + r 2 2r cos lim U(x) = {eis : s (, x)} ,
without loss of generality we may assume that has no point mass at 1. Put x (, ).
Then, doing integration by parts, U (r) = = = = 1 r2 d(eit ) 2 T 1 + r 2r cos t 1 r2 U(t) 1 + r2 2r cos t t= 1r U() + 1+r

1 r2 1 + r2 2r cos t
U(t) dt
(1 r2 ) 2r sin t U(t) dt (1 + r2 2r cos t)2

2r 1r U() + 1+r 1+r
(1 + r)2 (1 r) t sin t U(t) U(t) dt. (1 + r2 2r cos t)2 2t
3.6. Radial limits of hp (D) functions (1 p ) Since U(t) U(t) 1 = lim t0 t0 2t 2t lim
r1 t t
73
d(eis ) = 0,
by Theorem 2.5, lim U (r) = lim

r1
Fr (eit )
U(t) U(t) dt = 0. 2t
Let M(T). By Lebesgues decomposition theorem there are u L1 (T) and a measure M(T), singular with respect to the Lebesgue measure, such that d(eit ) = u(eit ) dt/2 + d(eit ). Moreover, (eit ) = lim {eis : s (t , t + )} 0 2
exists and equals u(eit )/2, for almost all eit T. Now, applying Lemmas 3.10 and 3.11, we obtain the following result about the radial limits of h1 (D) functions. Theorem 3.12 (Fatou) Let M(T) and let U (rei ) =
T
1 r2 d(eit ), 1 + r2 2r cos( t) lim U (rei ) = 2 (ei )
(rei D).
Then for almost all ei T.
r1
Exercises
Exercise 3.6.1 Show that Let u(ei ) = , for 0 < < 2, and let U (rei ) = (Pr u)(ei ). r sin 1 r cos (rei D).
U (rei ) = 2 arctan
Find all possible values of limn U (zn ), where zn is a sequence in D converging to 1. Exercise 3.6.2 Let u L1 (T), and let U (rei ) = (Pr u)(ei ). Suppose that lim0+ u(ei ) = L+ , lim2 u(ei ) = L and that L L+ .
74
(a) Let zn be a sequence in D such that limn zn = 1 and L = limn U (zn ) exists. Show that L [L , L+ ]. (b) Let L [L , L+ ]. Show that there exists a sequence zn D such that limn zn = 1 and limn U (zn ) = L. Hint: Consider v(ei ) = u(ei ) L+ + L+ L , 2 (0, 2),
and apply Exercise 3.6.1 and Corollary 2.6. Let be a real signed Borel measure in M(T). Let 1 2 1 r2 d(eit ), 1 + r2 2r cos( t) (rei D).
Exercise 3.6.3 U (rei ) =
For each ei0 T, dene (ei0 ) (ei0 ) Show that (ei0 ) lim inf U (rei0 ) lim sup U (rei0 ) (ei0 ).
r1 r1
= =
{eis : s (0 , 0 + )} , 2 0 {eis : s (0 , 0 + )} . lim inf 0 2 lim sup
Remark: This is a generalization of Theorem 3.12. Let be a real signed Borel measure in M(T). Let r2 1 r2 d(eit ), 2r cos( t) (ei0 ) = +. lim U (rei0 ) = +. (rei D).
Exercise 3.6.4
U (rei ) =
T
1+
Let ei0 T be such that Show that
r1
Hint: Apply Exercise 3.6.3.
3.6. Radial limits of hp (D) functions (1 p )
75
Exercise 3.6.5 Let M(T) be positive and singular with respect to the Lebesgue measure, and let U (rei ) =
T
1+
r2
1 r2 d(eit ), 2r cos( t) lim U (rei ) = +
(rei D).
Show that
r1
(3.11)
for almost all ei T (almost all with respect to ). Hint: Use Exercise 3.6.4. Remark 1: If = 0, then at least for one point (3.11) holds. Remark 2: Compare with Lemma 3.11 in which almost all is with respect to the Lebesgue measure. Exercise 3.6.6 Let U be a positive harmonic function on D such that
r1
lim U (rei ) = 0
for all ei T \ {1}. Show that U (rei ) = c 1 r2 , 1 + r2 2r cos
where c is a positive constant. Hint: Use Theorem 3.8 and Exercise 3.6.5. Exercise 3.6.7 Let M(T), and let 1+ r2 1 r2 d(eit ), 2r cos( t) lim U (rei ) = 0 (rei D).
U (rei ) =
T
Suppose that
r1
for all ei T. Show that U 0. Hint: Use Theorem 3.12 and Exercise 3.6.5. Exercise 3.6.8 suppose that Let U be a harmonic function on the open unit disc and
r1
lim U (rei ) = 0
for all ei T. Can we deduce that U 0? Hint: Consider U (rei ) = 2r(1 r2 ) sin P (rei ). = 2 2r cos )2 (1 + r
Remark: Compare with Exercise 3.6.7.
76 Exercise 3.6.9
Chapter 3. Harmonic functions in the unit disc Let M(T) and let 1 r2 d(eit ), 1 + r2 2r cos( t) (rei D).
U (rei ) =
T
Show that, for almost all 0 T,

zei0 zS (0 )
lim
U (z) = 2 (ei0 ),
where S (0 ) is the Stoltz domain S (0 ) = { z D : |z ei0 | C (1 |z|) }. (See Figure 3.1.) Remark 1: C > 1 is an arbitrary constant. Near the point ei0 , the boundaries of S (0 ) are tangent to a triangular-shaped region with angle 2 = 2 arccos(1/C ) and vertex at ei0 . Remark 2: We say that 2 (ei0 ) is the nontangential limit of U at ei0 .
Fig. 3.1. The Stoltz domain S (0 ).
Exercise 3.6.10 U (rei ) =
Let be a positive Borel measure in M(T). Let 1+ r2 1 r2 d(eit ), 2r cos( t) (rei D).
3.7. Series representation of the harmonic conjugate Let ei0 T be such that Show that, for each 0,
zei0 zS (0 )
77
(ei0 ) = +.
lim
U (z) = +.
Remark: This is a generalization of a particular case of Exercise 3.6.4. Exercise 3.6.11 Construct a real signed Borel measure M(T) such that (1) = +, but where U (rei ) =
T
t0
lim U (1 tei/4 ) = +, 1 r2 d(eit ), 2r cos( t)
1+
r2
(rei D).
Remark: Compare with Exercises 3.6.4 and 3.6.10.
3.7
Series representation of the harmonic conjugate

U (rei ) =
T
Let M(T), and let 1+ r2 1 r2 d(eit ), 2r cos( t) (rei D).
Then, by Theorem 2.1, U (rei ) =

n=
(n) r|n| ein ,
(rei D).
Hence, by Theorem 3.1, the harmonic conjugate of U is given by V (rei ) = Thus, V (rei )
n=
i sgn(n) (n) r|n| ein ,
(rei D).
=
n=
i sgn(n)
eint d(eit )
T
r|n| ein d(eit ).
=
T n=
i sgn(n) r|n| ein(t)
78
But, a simple calculation shows that
Qr (eit ) =
n=
i sgn(n) r|n| eint =
2r sin t . 1 + r2 2r cos t
(3.12)
(See Figure 3.2. Figure 3.3 shows the spectrum of 1 Qr .) This function is called i the conjugate Poisson kernel. Therefore, the harmonic conjugate of U is given by the conjugate Poisson integral
V (rei ) =
1 2
2r sin( t) d(eit ), 1 + r2 2r cos( t)
(rei D).
(3.13)
K3
K2
K1
K2
K4
Fig. 3.2. The conjugate Poisson kernel Qr (eit ) for r = 0.2, 0.5, 0.8.
3.7. Series representation of the harmonic conjugate
79
Fig. 3.3. The spectrum of
1 i
Qr .
We state the preceding result as a theorem which reveals the relation between the harmonic conjugate V and the Fourier coecients of when V is given by the conjugate Poisson integral of . Theorem 3.13 Let M(T), and let V (rei ) =
T
(rei D).
Then V (rei ) =
i sgn(n) (n) r|n| ein ,
(rei D).
n=
The series is absolutely and uniformly convergent on compact subsets of D. The function V is the unique harmonic conjugate of U (rei ) =
T
1 r2 d(eit ), 1 + r2 2r cos( t)
(rei D),
on D with V (0) = 0.
Exercises
Exercise 3.7.1 Let M(T), and let U = P and V = Q . Let F = U + iV . Show that F (z) =
T
eit + z d(eit ), eit z
(z D).
80 Exercise 3.7.2
Chapter 3. Harmonic functions in the unit disc Show that the conjugate Poisson kernel Q(rei ) = 2r sin , 1 + r2 2r cos
as a harmonic function on D, is not in h1 (D). However, Q hp (D), 0 < p < 1.
Chapter 4
Logarithmic convexity
4.1 Subharmonic functions
: [, ) is upper semicontinuous on if the set {z : (z) < c} is open for every c R. Note that is included as a possible value for (z). As a matter of fact, according to our denition, is an upper semicontinuous function. By a fundamental theorem of real analysis, a continuous function over a compact set is bounded and besides, it attains its maximum and its minimum. A similar, but certainly weaker, result holds for upper semicontinuous functions. Lemma 4.1 Let be a topological space, and let : [, ) be upper semicontinuous on . Let K be a compact subset of . Then there exists z0 K such that (z) (z0 ) for all z K. Proof. Clearly K
n=1
Let be a topological space. A function
{z : (z) < n}
and, by denition, each set {z : (z) < n} is an open subset of . Therefore, K has a nite subcover, say
N
K
n=1
{z : (z) < n},
which implies M = sup (z) N < .

zK
81
82
Chapter 4. Logarithmic convexity
On the other hand, each set Kn = {z K : (z) M 1/n} is compact, nonempty and Kn Kn+1 , for n 1. Therefore, by the nite intersection property for compact sets,
Kn = {z K : (z) = M }
n=1
is not empty. Any z0
n=1
Kn is a global maximum point.
Upper semicontinuous functions can be pointwise approximated by continuous functions on compact sets. This result plays a key role in developing the theory of subharmonic functions. In our applications, the compact set is usually a circle. Theorem 4.2 Let be an open subset of C, and let : [, ) be upper semicontinuous on . Let K be a compact subset of . Then there exist continuous functions n : K R, n 1, such that 1 2 on K and
n
lim n (z) = (z)
for each z K. Proof. If , simply take n n. Hence, assuming that , let n (z) = sup
wK
(w) n|w z| ,
(z K).
According to Lemma 4.1, for each n 1, n (z) M = sup (w) < ,

wK
(z K).
Moreover, by the triangle inequality, |n (z) n (z )| n|z z | for all z, z K, and clearly 1 (z) 2 (z) (z), (z K). It remains to show that n (z) converges to (z). Fix z0 K. Given > 0, choose R > 0 such that (z) < (z0 ) + whenever |z z0 | < R. Hence n (z0 ) max{ (z0 ) + , M nR }, Therefore, for all n > M/R, (z0 ) n (z0 ) (z0 ) + . (n 1).
4.1. Subharmonic functions
83
A function : [, ) is subharmonic on if it is upper semicontinuous and for each z there exists rz > 0 with D(z, rz ) , such that (z) 1 2
0 2
(z + r ei ) d
(4.1)
whenever r < rz . Our prototype of a subharmonic function is = log |F |, where F is an analytic function on . Indeed, is a continuous function on with values in [, ). If F (z) = 0, then (4.1) is trivial and if F (z) = 0, then log |F | is harmonic in an open neighborhood of z and thus, by Corollary 3.2, equality holds in (4.1) for suciently small values of r. The subharmonicity is a local property. In other words, the property (4.1) should be veried in a disc around z whose radius may depend on z. Nevertheless, we will see that (4.1) holds as long as D(z, r) . The passage from a local to a global property is a precious tool in studying subharmonic functions.
Exercises
Exercise 4.1.1 Let be an open subset of C. Show that : [, ) is upper semicontinuous if and only if lim sup (w) (z)
wz w=z
for all z . Hint: Remember that lim sup (w) = lim

wz w=z r0
sup
0<|wz|<r
(w) .
Exercise 4.1.2 Let be an open subset of C, let z0 , and let : [, ) be upper semicontinuous. Suppose that (z0 ) = . Show that is continuous at z0 , i.e. lim (z) = (z0 ).
zz0
Exercise 4.1.3 (Dinis theorem) Let K be a compact set. Let n : K R, n 1, and : K R be continuous functions on K. Suppose that 1 (z) 2 (z) (z) and that
n
lim n (z) = (z)
84
for each z K. Show that n converges uniformly to on K. Remark: Here we assumed that is continuous on K. Hence, in general, we cannot apply Dinis theorem to conclude that the sequence (n )n1 given in Theorem 4.2 converges uniformly to on K. Exercise 4.1.4 Let be an open subset of C and let U be harmonic on . Let 1 p < . Show that |U |p is subharmonic on . Hint: Use Corollary 3.2 and Hlders inequality. Jensens inequality can be o applied too. Exercise 4.1.5 Let be an open subset of C, let : [, ) be subharmonic and let : [, ) [, ) be nondecreasing, continuous and convex on (, ). Show that : [, ) is subharmonic. In particular, exp() and + are subharmonic functions. Hint: Use Jensens inequality. Remark: We assume exp() = 0. Exercise 4.1.6 Let be an open subset of C and let F be analytic on . Let 0 < p < . Show that |F |p and log+ |F | are subharmonic on . Hint: Use Exercise 4.1.5 and the fact that = log |F |p is subharmonic. Remark: The case 1 p < also follows from Exercise 4.1.4. Exercise 4.1.7 Let U be harmonic on D. Show that U hp (D), 1 p < , if and only if the subharmonic function |U |p has a harmonic majorant on D, i.e. there is a harmonic function V h(D) such that |U (z)|p V (z), (rei D). Remark: The function V is called a harmonic majorant of |U |p . Compare with Exercise 11.4.1.
4.2
The maximum principle
If F is analytic on and continuous on , the maximum principle for analytic functions says that max |F (z)| = max |F ()|.
z
In this section we show that this fundamental result is also fullled by subharmonic functions. Theorem 4.3 (Maximum principle for open sets) Let be a domain in C, and let be subharmonic on . Suppose that there is a z0 such that (z) (z0 ) for all z . Then is constant.
4.2. The maximum principle Proof. Let M = (z0 ), and let 1 = { z : (z) < M } and 2 = { z : (z) = M }.
85
Clearly = 1 2 , 1 2 = and 2 = . Hence, if we show that 1 and 2 are open subsets of , then the connectivity of forces 1 = , and thus M. Since is upper semicontinuous, by denition, 1 is open. On the other hand, for any z 2 , there exists rz > 0 with D(z, rz ) , such that M = (z) 1 2
2 0
(z + r ei ) d,
whenever r < rz . But, by assumption, we also have (z + r ei ) M for all . If for some 0 , (z + r ei0 ) < M holds, then, by upper semicontinuity, (z + r ei ) < M must hold on some open arc around 0 , which would imply 1 2
2 0
(z + r ei ) d < M.
Therefore, (z + r ei ) = M , for all and all 0 r < rz . In other words, D(z, rz ) 2 . Hence, 2 is also open. In the following, denotes the boundary of as a subset of C. Hence, if is unbounded, we do not assume that . Therefore, whenever a property has to be satised at boundary points of and also at innity, the latter requirement is explicitly expressed. Corollary 4.4 Let be a domain in C, and let be subharmonic on . Suppose that lim sup (z) 0
z z
for all . If is unbounded, we also assume that lim sup (z) 0.

z z
(4.2)
Then (z) 0, (z ). Proof. Let M = supz (z), and suppose that M > 0 (possibly +). Then there is a sequence (zn )n1 in such that limn (zn ) = M . Since we assumed that lim sup (z) 0
z z
86
for all , and also for = if is unbounded, the sequence (zn )n1 is bounded and all its accumulation points are inside . But, if z0 and (znk )k1 converges to z0 , then M = lim (znk ) lim sup (z) (z0 ) M.
k zz0 z
Thus (z0 ) = M , which means z0 is a global maximum point. Therefore, by Theorem 4.3, M > 0, which contradicts our assumptions about the behavior of as we approach the boundary points of . Hence, M 0. The condition (4.2) cannot be relaxed for unbounded domains. For example, (x + iy) = y is harmonic in the upper half plane C+ and lim sup (z) = 0
zt zC+
for all t R. However, > 0 on C+ . Nevertheless, we can sometimes replace (4.2) by another condition. The following corollary is a result of this type. Corollary 4.5 Let be a proper unbounded domain in C, and let be subharmonic and bounded above on . Suppose that lim sup (z) 0
z z
for all . Then
(z) 0,
(z ).
Proof. Without loss of generality, assume that 0 . Given > 0, there is 0 < R < 1/2 such that (z) for all z with |z| R. Since log(|z|/R) log(1/R) is a positive harmonic function on { |z| > R }, (z) = (z) log(|z|/R) log(1/R)
is subharmonic on R = { |z| > R }, and lim sup (z) 0

z zR
for all R . The term in the denition of is added to ensure that this property holds even if R and || = R. Moreover, since is bounded above on , lim sup (z) = 0.
z zR
4.2. The maximum principle Hence, by Corollary 4.4, (z) 0 for all z R , which gives (z) + log+ (|z|/R) log(1/R)
87
for all z R . Again thanks to the extra that we added from the beginning, the last inequality is actually valid for all z . Let 0 (which may force R 0). Nevertheless, we get (z) 0 for all z .
Exercises
Exercise 4.2.1 [Maximum principle for compact sets] Let be a bounded domain in C, and let be upper semicontinuous on and subharmonic on . Show that there is a 0 such that (z) (0 ) for all z . Hint: Apply Lemma 4.1 and Theorem 4.3. Exercise 4.2.2 Let (x + iy) =
if x 0,
x if x 0.
Show that is subharmonic on C. Hint: Use Exercise 4.3.1, part (a). Remark: Every point z = x + iy, x < 0, is a local maximum. Is that a contradiction to the maximum principle? Exercise 4.2.3 Let be a bounded domain in C, and let U be a real harmonic function on . Suppose that
z z
lim U (z) = 0
for all . Show that U 0. Hint: Apply Corollary 4.4 to U and U . Exercise 4.2.4 Let be a proper unbounded domain in C, and let F be a bounded analytic function on . Suppose that lim sup |F (z)| 1
z z
88 for all . Show that
|F (z)| 1
for all z . Hint: Apply Corollary 4.5 to log |F |. Exercise 4.2.5 [Gausss theorem] Let be an open subset of C, and let U be a real continuous function on . Suppose that U satises the mean value property, i.e. if D(z, r) , then U (z) = 1 2

U (z + reit ) dt.
Show that U is harmonic on . Hint: The proof of Theorem 4.3 might help. Exercise 4.2.6 [Schwarzs reection principle] Let be a domain in C symmetric with respect to the real axis. Suppose that F is analytic on C+ and, for each t R, lim F (z) = 0.
zt zC+
Show that there is a unique analytic function G on such that G = F on C+ . Hint: Put F (z) if z C+ , 0 if z R, V (z) = F () if z C , z and apply Exercise 4.2.5 to show that V is harmonic on .
4.3
A characterization of subharmonic functions
Corollary 4.4 enables us to provide an equivalent denition of subharmonic functions which justies the title subharmonic. Theorem 4.6 Let be an open subset of C, and let : [, ) be upper semicontinuous. Then the following are equivalent: (a) is subharmonic on ; (b) for any bounded subdomain with , and for any harmonic function U on , if lim sup
z z
(z) U (z)
4.3. A characterization of subharmonic functions for every , then we have (z) U (z), (c) if z , and D(z, R) , then (z + rei ) 1 2

89
(z );
R2 r 2 (z + Reit ) dt R2 + r2 2Rr cos( t)
for all 0 r < R and all . Proof. (a) = (b) : It follows from Corollary 4.4. (b) = (c) : By Theorem 4.2, there exist continuous functions n , n 1, dened on the circle CR = { : | z| = R }, so that n () n+1 () and limn n () = (), for all CR . By Corollary 2.3, if we also dene n (z + rei ) = 1 2

R2
r2
R2 r 2 n (z + Reit ) dt 2Rr cos( t)
for z + rei D(z, R), then n is continuous on the closed disc D(z, R), and besides it is harmonic inside that disc. Since is upper semicontinuous, for each DR = CR , we have lim sup
w wDR
(w) n (w)
() n () 0.
Hence, by assumption (b), (z + rei ) n (z + rei ), i.e. (z + rei ) 1 2

R2 r 2 n (z + Reit ) dt R2 + r2 2Rr cos( t)
for all 0 r < R and all . Let n and use the monotone convergence theorem to get (c). (c) = (a) : Put r = 0.
Exercises
Exercise 4.3.1 Show that Let 1 and 2 be subharmonic functions on an open set .
(a) max{ 1 , 2 } and (b) 1 1 + 2 2 , where 1 , 2 0, are also subharmonic on . Is min{ 1 , 2 } necessarily subharmonic on ?
90
Exercise 4.3.2 Let be an open subset of C, and let : R be twice continuously dierentiable on . Show that is subharmonic on if and only if 2 0 on . Hint: Use Theorem 4.6(b) with the upper semicontinuous function (z) U (z) + |z|2 if z , (z) = |z|2 if z . Exercise 4.3.3 Let 1 , 2 be open subsets of C, let F : 1 2 be analytic, and let : 2 [, ) be subharmonic on 2 . Show that F : 1 [, ) is subharmonic on 1 . Hint: Use Exercise 1.1.4 and Theorem 4.6, part (b).
4.4
Various means of subharmonic functions
Given a nite family of subharmonic functions, by taking positive linear combinations or by taking their supremum, we can create new subharmonic functions. In this section, we show that under certain conditions the preceding two procedures still create subharmonic functions even if the family is not nite. This topic is thoroughly treated in [17]. In the following theorem we assume that T is any compact topological space. But in our applications, it is always the unit circle T. Theorem 4.7 Let T be a compact topological space, and let be an open subset of C. Suppose that : T [, ) has the following properties: (i) is upper semicontinuous on T ; (ii) for each xed t T , t (z) = (z, t), as a function of z, is subharmonic on . Let (z) = sup t (z),
tT
(z ),
(z ).
Then is subharmonic on . Proof. First, let us show that is upper semicontinuous. Fix c R, and suppose that (z0 ) < c. Then, according to the denition of , we have (z0 , t) < c for all t T . Since is upper semicontinuous on T , for each t T , there exist rt > 0 and an open set Vt T with t Vt such that D(z0 , rt ) Vt { (z, ) : (z, ) < c }.
4.4. Various means of subharmonic functions Since T =

t
91
Vt , and T is a compact set, T has a nite subcover, say T = Vt1 Vtn
for some t1 , t2 , . . . , tn T . Let r = min{ rt1 , rt2 , . . . , rtn } and note that r > 0. Hence D(z0 , r) Vtk { (z, ) : (z, ) < c } for k = 1, 2, . . . , n. Therefore, taking the union over k, we obtain D(z0 , r) T { (z, ) : (z, ) < c }, which implies D(z0 , r) { z : (z) < c }. Note that {z} T is compact and thus, by Lemma 4.1, (z) = sup (z, ) < c
T
provided that (z, ) < c holds for all T . Hence is upper semicontinuous. Secondly, we show that is subharmonic. Fix z0 , and assume that D(z0 , R) . Since t is subharmonic, Theorem 4.6 implies t (z0 ) 1 2

t (z0 + rei ) d
for all r < R. Appealing to Theorem 4.6 is a crucial step in the proof. Since otherwise, based on the original denition of subharmonicity, the last inequality would be valid for r < R = R(t) and we have no control over R(t). Hence, t (z0 ) 1 2

(z0 + rei ) d
for all r < R. Now, once more, take the supremum over t T to get the required result. A subharmonic function on DR is called radial if (z) = (|z|) for all z DR . Given a subharmonic function on a disc, the preceding result enables us to create a radial subharmonic function by taking its supremum over each circle. Corollary 4.8 Let be subharmonic on the disc DR . Let (z) = sup (|z|ei ),
ei T
(z DR ).
Then is a radial subharmonic function on DR .
92 Proof. The function
DR T [, ) (z, ei ) (zei ) fulls all the requirements of Theorem 4.7. Hence, (z) = sup (zei ) = sup (|z|ei )
ei T ei T
is subharmonic on DR . The last identity also shows that (z) = (|z|). In the following result we assume that (T, M, ) is any measure space with a nite positive measure on T . But in our applications, T is either the unit circle T equipped with the normalized Lebesgue measure d/2, or is the unit disc D with the two-dimensional normalized Lebesgue measure rdrd/. Theorem 4.9 Let (T, M, ) be a measure space, 0, (T ) < , and let be an open subset of C. Suppose that : T [, ) has the following properties: (i) is measurable on T ; (ii) for each xed t T , (z, t), as a function of z, is subharmonic on ; (iii) for each z , there exists rz > 0 with D(z, rz ) such that sup
D(z,rz )T
(w, t) < .
Let (z) =
T
(z, t) d(t),
(z ).
Then is subharmonic on . Proof. Without loss of generality, assume that is a probability measure. We start by showing that is upper semicontinuous on . Fix z , and let Mz = sup
D(z,rz )T
(w, t).
First of all, assumption (iii) shows that (z) Mz < .
4.4. Various means of subharmonic functions Secondly, since (z, t) is subharmonic on , by Fatous lemma, Mz lim sup (w)
wz
93
= = =
lim inf Mz (w)

wz
lim inf
wz T wz
Mz (w, t) d(t)
lim inf Mz (w, t) d(t) Mz lim sup (w, t) d(t)

wz
Hence, we have
Mz (z, t) d(t) = Mz (z). (z ),
lim sup (w) (z),

wz
which ensures that is upper semicontinuous on . It remains to show that fulls the submean inequality. Let r < rz and let Iz (r) = Then, by Fubinis theorem, Mz Iz (r) = = =
T
1 2
(z + rei ) d.
1 2 1 2
Mz (z + rei ) d Mz (z + rei , t) d(t) Mz (z + rei , t) d d
1 2
d(t)
Mz (z, t) d(t) = Mz (z).
Note that the integrand is nonnegative and thus Fubinis theorem can be applied. Thus, for all r < rz , (z) 1 2

(z + rei ) d.
Given a subharmonic function on a disc, the preceding result enables us to create a radial subharmonic function by taking its integral means over each circle. Corollary 4.10 Let be subharmonic on the disc DR , and let (z) = 1 2

(zei ) d,
(z DR ).
Then is a radial subharmonic function on DR .
94 Proof. The function
DR T [, ) (z, ei ) (zei ) fulls all the requirements of Theorem 4.9. Hence, (z) = 1 2

(zei ) d =
1 2
(|z|ei ) d
is subharmonic on DR . The last identity also shows that (z) = (|z|).
Exercises
Exercise 4.4.1 Let be subharmonic on the disc DR . Let 1
0 1
(z) =
(zei ) dd,
(z DR ).
Show that is a radial subharmonic function on DR . Hint: Consider the function DR D [, ) (z, ei ) (zei ) and apply Theorem 4.9. Remark: By a simple change of variable, we also have (z) = if z = 0. 1 |z|2
0 |z|
(ei ) dd,
Exercise 4.4.2
Let be a domain in C and dene (z) = log dist(z, ), (z ),
where dist(z, ) = inf | z|.
Show that is subharmonic on . Hint: Use Theorem 4.7.
4.5. Radial subharmonic functions
95
4.5
Radial subharmonic functions

M (r) C (r) = = sup (z),
|z|=r
Let be subharmonic on DR , and let
1 2
(rei ) d
for 0 r < R. In Corollaries 4.8 and 4.10, we saw that (z) = M (|z|) and (z) = C (|z|) are radial subharmonic functions on DR . The following result characterizes all radial subharmonic functions. We say that (r) is a convex function of log r, if (e ) is a convex function of . In other words, satises (r) log r2 log r log r log r1 (r1 ) + (r2 ), log r2 log r1 log r2 log r1
whenever r1 < r < r2 , and of course r, r1 , r2 are in the domain of denition of . Theorem 4.11 Let be a radial subharmonic function on DR . Then (r), 0 < r < R, is an increasing convex function of log r, and limr0 (r) = (0). Proof. Let 0 < r1 < r2 < R. Since is subharmonic and radial, (0) 1 2

(r1 ei ) d = (r1 ).
(4.3)
Also, by the maximum principle, (r1 ) sup (z) = (r2 ).

|z|=r2
Thus is increasing. Moreover, by (4.3), and the fact that is upper semicontinuous on DR , we have (0) lim inf (r) lim sup (r) (0).
r0 r0
Therefore, is continuous at the origin. To prove that (r) is a convex function of log r, x r1 and r2 , and let (z) = (z) for z A(r1 , r2 ) = { z : r1 < |z| < r2 }. Then is a subharmonic function on the annulus A(r1 , r2 ), and
z zA(r1 ,r2 )
log r2 log |z| log |z| log r1 (r1 ) (r2 ) log r2 log r1 log r2 log r1
lim sup (z) 0
96
for all A(r1 , r2 ), where A(r1 , r2 ) consists of two circles {|| = r1 } and {|| = r2 }. Therefore, by the maximum principle (Corollary 4.4), (z) 0 for all z A(r1 , r2 ). This fact is equivalent to (r) log r log r1 log r2 log r (r1 ) + (r2 ), log r2 log r1 log r2 log r1
whenever r1 < r < r2 . Corollary 4.12 Let be a subharmonic function on DR . Then (a) M (r) and C (r) are increasing convex functions of log r, (b) limr0 M (r) = limr0 C (r) = (0), (c) (0) C (r) M (r), for all 0 < r < R. Proof. The rst two claims are direct consequences of Corollaries 4.8 and 4.10 and Theorem 4.11. Moreover, directly from the denition, C (r) M (r), and by (4.3), (0) C (r) for 0 < r < R. Let u Lp (T), 1 p , and let U = P u. According to Corollaries 2.11, 2.15 and 2.19, we have U hp (D) and U
p
= sup
0r<1
Ur
= lim Ur
r1
= u p.
p
However, based on Corollary 4.12, we can say more about the behavior of Ur as a function of r. Corollary 4.13 Let u Lp (T), 1 p , and let U = P u. Then Ur an increasing convex function of log r and
r1 p
is
lim Ur
= U
p.
Let F H p (D) with 0 < p . If 1 p then the preceding result applies and thus Fr p is an increasing function of r with
r1
lim Fr
= F
= f
p,
where f represents the boundary values of F . If 0 < p < 1, rst of all, we have not yet shown that F has boundary values almost everywhere on T. At this point, we do not have enough tools to prove this fact. However, since |F |p is a subharmonic function on D, even if 0 < p < 1, Corollary 4.12 applies and we obtain the following result. Corollary 4.14 Let F H p (D), 0 < p . Then Fr convex function of log r and
r1 p
is an increasing
lim Fr
= F
p.
4.6. Hardys convexity theorem
97
Exercises
Exercise 4.5.1 Let be subharmonic on DR and let B (r) = 1 r2
0 r
(ei ) dd
for 0 < r < R. Show that B (r) is an increasing convex function of log r, (0) B (r) C (r) M (r) for all 0 < r < R, and that
r0
lim B (r) = (0).
Exercise 4.5.2 Let be a radial function on DR . Suppose that (r), 0 < r < R, is an increasing convex function of log r and that limr0 (r) = (0). Show that is subharmonic on DR .
4.6
Hardys convexity theorem
G. H. Hardy, in a famous paper [7], showed that log( Fr p ) is a convex function of log r, whenever F is an analytic function on a disc. This result is considered as the starting point of the theory of Hardy spaces. Theorem 4.15 (Hardy) Let F be analytic on the disc DR , and let 0 < p < . Then log( Fr p ) is an increasing convex function of log r. Proof. First of all, note that Fr
p p
1 2
|F (rei )|p d = C|F |p (r)
and thus, by Corollary 4.12, Fr p is an increasing convex function of log r. Fix R. Since F is analytic on DR , (z) = |z| |F (z)|p is subharmonic there, and furthermore C (r) = r C|F |p (r). (4.4)
Fix r, r1 and r2 with 0 < r1 < r < r2 < R. By Corollary 4.12 and by (4.4), we have C (r) log(r2 ) log(r) log(r) log(r1 ) C (r1 ) + C (r2 ). log(r2 ) log(r1 ) log(r2 ) log(r1 ) (4.5)
98 The right side is the arithmetic mean of

1 = C (r1 ) = r1 C|F |p (r1 )
and
2 = C (r2 ) = r2 C|F |p (r2 )
with weights m1 = log(r2 ) log(r) log(r2 ) log(r1 ) and m2 = log(r) log(r1 ) . log(r2 ) log(r1 )
m m Note that r = r1 1 r2 2 . Hence, by (4.4), we can rewrite (4.5) as
r C|F |p (r) m1 1 + m2 2 . The arithmeticgeometric inequality says that

m m 1 1 2 2 m1 1 + m2 2
(4.6)
and that equality holds if and only if 1 = 2 . Thus we choose such that 1 = 2 , i.e. r1 C|F |p (r1 ) = r2 C|F |p (r2 ). Hence, with this particular choice of , we have m1 1 + m2 2
m m = 1 1 2 2 m1 m2
r1 C|F |p (r1 )
m1
r2 C|F |p (r2 )
= r Therefore, by (4.6), we obtain
m2
C|F |p (r1 )
C|F |p (r2 )
m1
m2
C|F |p (r) Since C|F |p (r) = Fr

p p,
C|F |p (r1 )
C|F |p (r2 )
taking the logarithm of both sides gives

p)
log( Fr
log(r2 ) log(r) log( Fr2 log(r2 ) log(r1 ) log(r) log(r1 ) log( Fr1 log(r2 ) log(r1 )
p)
p ),
which is the required result. Hardys convexity theorem is valid even for p = . In this case it is called Hadamards three-circle theorem [4, page 137]. If we use M|F | instead of C|F |p in the preceding proof, we obtain a proof of Hadamards theorem. However, we will also discuss a variation of this theorem in Section 8.2.
Exercises
4.7. A complete characterization of hp (D) spaces Exercise 4.6.1 F

p,r
99
Let F be analytic on the disc DR , and let 0 < p < . Dene 1 r2

0 r
|F (e )| dd
1 p
(0 < r < R).
Show that log( F p,r ) is an increasing convex function of log r. Hint: Repeat the proof of Theorem 4.15, replacing C|F |p by B|F |p . Exercise 4.6.2 [Hardy [7]] Let F be an entire function, and let 0 < p . Show that Fr p = O(rn ), as r , if and only if F is a polynomial of degree at most n. Otherwise, Fr p tends to innity more rapidly than any power of r. Exercise 4.6.3 Let F be analytic on the annular domain A(R1 , R2 ), and let 0 < p < . Show that log( Fr p ), R1 < r < R2 , is a convex function of log r. Remark: In this case, log( Fr p ) is not necessarily increasing. Exercise 4.6.4 [Littlewoods subordination theorem] Let : D D be analytic with |(z)| |z| for all z D. Let F and G be analytic on D. We say that F is subordinate to G if F (z) = G( (z) ) for all z D. Let F be subordinate to G. Show that M|F | (r) M|G| (r), and, for 0 < p < , C|F |p (r) C|G|p (r), (0 r < 1). (0 r < 1),
4.7
A complete characterization of hp (D) spaces
As far as the representation of harmonic functions on the open unit disc is concerned, our job is done. We summarize the results of the preceding sections to exhibit a complete characterization of hp (D) spaces. We only highlight the main points. Let U be a harmonic function on the open unit disc D. Case 1: p = 1. U h1 (D) if and only if there exists M(T) such that U (rei ) =
T
1+
r2
(rei D).
100 The measure is unique and U (rei ) =

n=
(n) r|n| ein ,
(rei D).
The family of measures (U (reit ) dt/2)0r<1 converges to d(eit ) in the weak* topology, as r 1 , i.e.
r1
lim
(eit ) U (reit )
T
dt = 2
(eit ) d(eit )
T
for all C(T), and we have U Case 2: 1 < p < . U hp (D), 1 < p < , if and only if there exists u Lp (T) such that U (rei ) = 1 2
1
= = ||(T).
1 r2 u(eit ) dt, 1 + r2 2r cos( t)
(rei D).
The function u is unique and U (rei ) = Moreover,

r1 n=
u(n) r|n| ein ,
(rei D).
lim Ur u U
p
= 0,
and thus = u p. Case 3: p = . U h (D) if and only if there exists u L (T) such that U (rei ) = 1 2

1 r2 u(eit ) dt, 1 + r2 2r cos( t)
(rei D).
The function u is unique and U (rei ) =

n=
u(n) r|n| ein ,
(rei D).
The family (Ur )0r<1 converges to u in the weak* topology, as r 1, i.e.

r1
lim
(eit ) u(eit ) dt
4.7. A complete characterization of hp (D) spaces for all L1 (T). Hence, U In all cases, the family Ur convex of log r.
p,
101
= u
0 r < 1, as a function of r, is increasing and a
The rst case shows that h1 (D) and M(T) are isometrically isomorphic Banach spaces. Similarly, by the other two cases, hp (D) and Lp (T), 1 < p , are isometrically isomorphic.
Exercises
Exercise 4.7.1 We saw that h1 (D) and M(T) are isometrically isomorphic Banach spaces. Consider L1 (T) as a subspace of M(T). What is the image of L1 (T) in h1 (D) under this correspondence? What is the image of positive measures? Hint: Use Exercise 3.5.3. Exercise 4.7.2 We know that h (D) and L (T) are isometrically isomorphic Banach spaces. Consider C(T) as a subspace of L (T). What is the image of C(T) in h (D) under this correspondence? Hint: Use Exercise 3.3.4.
102
Chapter 5
Analytic functions in the unit disc

5.1 Representation of H p (D) functions (1 < p )
Since we assumed that the elements of hp (D) are complex-valued harmonic functions, we necessarily have H p (D) hp (D). Hence, the facts we gathered in Section 4.7 are also valid for the analytic elements of hp (D), i.e. for elements of H p (T). As a matter of fact, we can say more in this case. If F H p (D), 1 < p , then there is f Lp (T) such that
F (z) =
n=
f (n) r|n| ein =
n=0
f (n) z n +
z f (n)n ,
(z = rei D).
n=
According to (3.1) in Theorem 3.1, we have r|n| f (n) = 2 Hence, for n 1, r|n| f (n) = 2i which implies f (1) = f (2) = = 0.
p
F (reit ) eint dt,
(n Z, 0 < r < 1).
F (z) z |n+1| dz,

|z|=r
(0 < r < 1),
(5.1)
On the other hand, if f L (T), 1 < p , satisfying (5.1) is given, and we dene F (z) = 1 2

1+
r2
1 r2 f (eit ) dt, 2r cos( t) 103
(z = rei D),
104
Chapter 5. Analytic functions in the unit disc
then, in the rst place, F hp (D). Secondly,
F (z) =
n=
f (n) r|n| ein =
n=0
f (n) z n
and thus it represents an analytic function. Hence, F H p (D). This observation leads us to dene the Hardy spaces on the unit circle by H p (T) = { f Lp (T) : f (1) = f (2) = = 0}, (1 p ). (5.2)
In the following, the content of Section 4.7 is rewritten for analytic functions. Theorem 5.1 Let F be analytic on the open unit disc. Then F H p (D), 1 < p < , if and only if there exists f H p (T) such that F (rei ) = 1 2

1+
r2
1 r2 f (eit ) dt, 2r cos( t)
(rei D).
The function f is unique and
F (z) =
n=0
f (n) z n ,
(z D).
Moreover,
r1
lim Fr f F
p
=0
p.
and = f
2
We have F H (D) if and only if f H (T), and in this case F

2
= f
=
n=0
|f (n)|2
1 2
Theorem 5.2 Let F be analytic on the open unit disc. Then F H (D) if and only if there exists f H (T) such that F (rei ) = 1 2

1+
r2
1 r2 f (eit ) dt, 2r cos( t)
(rei D).
The function f is unique and
F (z) =
n=0
f (n) z n ,
(z D).
Moreover, Fr converges to f in the weak* topology, as r 1, i.e.

r1 1
lim
(eit ) F (reit ) dt =
(eit ) f (eit ) dt
for all L (T), and F
= f
5.1. Representation of H p (D) functions (1 < p ) The restriction of the Fourier transform to H p (T) classes H p (T) c0 (Z+ ) f f
105
is not surjective. However, based on the contents of Section 2.7 and Theorem 5.1, we can say that the maps H 2 (D) H 2 (T) F f
2
(Z+ ) f
are bijective and preserve the inner product. The preceding two theorems show that H p (D) and H p (T), 1 < p , are isometrically isomorphic. One may naturally wonder why H 1 (D) was not treated in the same manner. Based on what we know about h1 (D), if we follow the same procedure, we conclude that H 1 (D) is isometrically isomorphic to the subclass { M(T) : (1) = (2) = = 0}, and every such measure creates a unique element of H 1 (D) given by F (z) =
T
1 r2 d(eit ) = (n) z n , 2 2r cos( t) 1+r n=0
(z = rei D).
This assertion is absolutely true and can be viewed as a characterization of H 1 (D). However, it does not reveal the whole truth. According to a celebrated result of F. and M. Riesz, such a measure is necessarily absolutely continuous with respect to the Lebesgue measure. Therefore, Theorem 5.1 is valid even for p = 1. In the rest of this chapter, we develop some tools to prove the F. and M. Riesz theorem and thus complete the characterization of H 1 (D) functions.
Exercises
Exercise 5.1.1 Show that H p (T), 1 p , is a closed subspace of Lp (T). Moreover, show that H (D) is weak* closed in L (T). Exercise 5.1.2 Let (an )n0
2
(Z+ ), and let an z n , (z D).
F (z) =
n=0
Show that F H 2 (D) and
F Hint: Apply Theorem 5.1.
=
n=0
|an |2
1 2
106
Exercise 5.1.3 We know that H (D) and H (T) are isometrically isomorphic Banach spaces. Let A(T) denote the closure of analytic polynomials in H (T). What is the image of A(T) in H (D) under this correspondence? Hint: Exercise 3.3.4 might help.
5.2
The Hilbert transform on T
Let F be analytic on a disc DR with R > 1 and suppose that F (0) = 0. Fix ei T. Let 0 < < R 1 and let be the curve shown in Figure 5.1. The radius of the small circle is r = 2 sin(/2).
Fig. 5.1. The curve . As Figure 5.2 shows, the curve T is parameterized by = eit , | t| , + + t . 2 2
and the rest of , according to the triangle , is given by = ei + r eit ,
5.2. The Hilbert transform on T
107
Fig. 5.2. The triangle .
Hence, by Cauchys theorem, F (ei ) = = 1 2i 1 2i F () d ei F (eit ) 1 ieit dt + eit ei 2

+ 2
|t|
+ 2
F (ei + r eit ) dt.
Since F (ei + r eit ) F (ei ), as 0, we obtain F (ei ) = lim 1 0 F (eit ) dt. 1 ei(t) (5.3)
|t|
By a similar argument we have 1 2i 1 2i ei F () d ( ei ) ei F (eit ) 1 ieit dt it (eit ei ) e 2

3 2 + 2
0 = =
|t|
ei F (ei + r eit ) dt, ei + r eit
where is the curve shown in Figure 5.3. Remember that F (0) = 0.
108
Fig. 5.3. The curve .
Hence, F (ei ) = lim Adding (5.4) to (5.3) gives F (ei ) = lim i 0 1 0

|t|
ei(t ) F (eit ) dt. 1 ei(t)
(5.4)
|t|
F (eit ) dt. 2 tan( t ) 2
(5.5)
The advantage of (5.5) over (5.4) and (5.3) is that the kernel appearing in it is a real function. Let us write u = F and v = F on the unit circle. Hence, (5.5) is equivalent to v(ei ) = lim and u(ei ) = lim 1 0
|t|
1 0
|t|
u(eit ) dt 2 tan( t ) 2 v(eit ) dt. 2 tan( t ) 2
The preceding argument highlights the importance of the transform 1 0 lim (eit ) dt 2 tan( t ) 2
|t|
5.2. The Hilbert transform on T
109
whenever is the real or imaginary part of a function analytic on a domain containing the closed unit disc. However, this transform is well-dened on other classes of functions and has far-reaching and profound properties. To have a meaningful Lebesgue integral we need at least to assume that L1 (T). Then the Hilbert transform of at the point ei T is dened by H(ei ) = (ei ) = = 1 0 lim
0
|t|
(eit ) dt 2 tan( t ) 2 (5.6)
lim
<|t|<
(ei(t) ) dt 2 tan(t/2)
wherever the limit exists. The notation

1 (ei(t) ) ( ei ) = dt 2 tan(t/2)
has exactly the same meaning as (5.6). Since around the origin 2 tan(t/2) behaves asymptotically like t, the transform H( ei ) = lim 1 (ei(t) ) dt t (5.7)
<|t|<
has close connections to . Indeed, some authors call the conjugate transform of u and keep the title Hilbert transform for H. However, since 1 1 1 , 2 tan(t/2) t (0 < |t| ), (5.8)
both transforms have the same behavior. More precisely, if one of them exists at ei T, the other one exists too and we have |H( ei ) H( ei ) | 2 . (5.9)
Since 2 tan(t/2) is an odd function, a more appropriate way to write is 1 ( ei ) = lim 0 Of course, if 1
0
(ei(t) ) (ei(+t) ) dt. 2 tan(t/2)
(5.10)
|(ei(t) ) (ei(+t) )| dt < , 2 tan(t/2)
(5.11)
then the Hilbert transform exists and is given by 1 ( ei ) =

0
(ei(t) ) (ei(+t) ) dt. 2 tan(t/2)
(5.12)
110
However, even if (5.11) does not hold, we will show that the limit (5.10) exists almost everywhere on T. To achieve this goal, we are led naturally to study the boundary behavior of the harmonic conjugate function.
Exercises
Exercise 5.2.1 Show that there is a C(T) such that 1
0
|(ei(t) ) (ei(+t) )| dt = 2 tan(t/2)
for all ei T. Hint: Use the Baire category theorem [13, page 25]. Exercise 5.2.2 Verify the following Hilbert transform table:
(ei )
(ei )
cos
sin cos
sin
5.3
Radial limits of the conjugate function
In Chapter 3 we saw that, under certain conditions, a harmonic function can be represented as U (rei ) = 1 2

1+
r2
(rei D),
where M(T). It is easy to verify that the harmonic conjugate of U is given by V (rei ) = 1 2

(rei D).
5.3. Radial limits of the conjugate function To verify this fact, simply note that (U + iV )(z) = 1 2

111
(z D),
which is an analytic function on the unit disc. According to Fatous theorem, we know that limr1 U (rei ) exists for almost all ei T. Hence, we naturally ask if the conjugate function behaves nicely too and limr1 V (rei ) exists for almost all ei T. Since the conjugate Poisson kernel is not an approximate identity, we are not able to apply the techniques developed in Chapter 3. Nevertheless, the result is true. Let us start with a special case. If is absolutely continuous with respect to the Lebesgue measure, then the harmonic conjugate of U is given by V (rei ) = 1 2

2r sin( t) u(eit ) dt, 1 + r2 2r cos( t)
(rei D),
where u L1 (T). Since the conjugate Poisson kernel is an odd function, the preceding formula for V (rei ) can be rewritten as V (rei ) = 1
0
r sin t 1 + r2 2r cos t
u(ei(t) ) u(ei(+t) ) dt.
(5.13)
This version is more appropriate in studying the boundary values of V . To save space, on some occasions we will write V = Q u. To evaluate limr1 V (rei ), let us formally change the order of the integral and limit in (5.13) to obtain lim V (rei ) = 1
0
r1
u(ei(t) ) u(ei(+t) ) dt. 2 tan(t/2)
(5.14)
We used the trigonometric identity sin t 1 = , 2(1 cos t) 2 tan(t/2) (5.15)
which will appear again several times in our discussion. The formula (5.14) also highlights the relation between limr1 V ( rei ) and u( ei ). But if 1
0
|u(ei(t) ) u(ei(+t) )| dt = , 2 tan(t/2)
(5.16)
the integral in (5.14) is meaningless. In the rst place, we show that if the integral in (5.16) is nite, then the radial limit limr1 V (rei ) and u(ei ) both exist and they are equal. Secondly, we will see that, even if (5.16) may hold, the radial limits and the Hilbert transform still exist and are equal almost everywhere on T. We start with the rst case. The general case will be treated at the end of Section 5.3.
112
Theorem 5.3 Let u L1 (T) and let V (rei ) = 1 2

(rei D).
Fix ei T and suppose that 1

0
|u(ei(t) ) u(ei(+t) )| dt < . 2 tan(t/2)
(5.17)
Then limr1 V (rei ) and u(ei ) exist and lim V (rei ) = u(ei ) = 1
0
r1
u(ei(t) ) u(ei(+t) ) dt. 2 tan(t/2)
(5.18)
Proof. The fact that u(ei ) exists and is given by u(ei ) = 1

0
u(ei(t) ) u(ei(+t) ) dt 2 tan(t/2)
is a direct consequence of the denition of u(ei ). We need to estimate r = V (rei ) u(ei ) . Hence, by (5.13) and (5.15), r = = Since 1 1 1
0
sin t r sin t |u(ei(t) ) u(ei(+t) )| dt 1 + r2 2r cos t 2(1 cos t) (1 r)2 sin t |u(ei(t) ) u(ei(+t) )| dt 2(1 cos t)(1 + r2 2r cos t) |u(ei(t) ) u(ei(+t) )| (1 r)2 dt. 2 2 tan(t/2) (1 r)2 + 4r sin (t/2) (1 r)2 1 (1 r)2 + 4r sin2 (t/2) (1 r)2 =0 r1 (1 r)2 + 4r sin2 (t/2) lim
and
for all t (0, ], by the dominated convergence theorem, the last integral tends to zero as r 1. We may make some stronger, but more familiar, assumptions on u at the point ei to ensure that (5.17) holds. Two such conditions are mentioned below.
5.4. The Hilbert transform of C 1 (T) functions
113
Corollary 5.4 Let u L1 (T) and let V = Q u. Fix ei T and suppose that, for some > 0, u(ei(t) ) u(ei(+t) ) is Lip at t = 0, i.e. for |t| t0 , | u(ei(t) ) u(ei(+t) ) | C |t| . Then limr1 V (rei ) and u(ei ) exist and (5.18) holds. Corollary 5.5 Let u L1 (T) and let V = Q u. Fix ei T and suppose that u(ei(t) ) u(ei(+t) ), as a function of t R, is dierentiable at t = 0. Then limr1 V (rei ) and u(ei ) exist and (5.18) holds.
5.4
The Hilbert transform of C 1 (T) functions
If the conjugate function V = Q u is produced by a continuously dierentiable function u, then, by Corollary 5.5, we know that u(ei ) and limr1 V (rei ) exist at all points of T and they are equal. We show that u is continuous on T. Theorem 5.6 Let u C 1 (T), and let V (rei ) = 1 2

(rei D).
Then, as r 1, Vr converges uniformly to u on T. Proof. Let r (ei ) = | V (rei ) u(ei ) |. 1

0
Hence, as we saw in the proof of Theorem 5.3, r (ei ) |u(ei(t) ) u(ei(+t) )| (1 r)2 dt. 2 2r cos t 1+r 2 tan(t/2)
.
According to the mean value theorem, we have |u(ei(t) ) u(ei(+t) )| 2|t| u Hence, there is an absolute constant C such that |u(ei(t) ) u(ei(+t) )| C 2 tan(t/2) for all t [0, ]. Therefore r (ei ) | which gives Vr u
0
1r (1 r)2 , dt = C 1 + r2 2r cos t 1+r C (1 r).
114
The uniform limit of a sequence of continuous functions is continuous. Hence, Theorem 5.6 implies immediately that u is continuous on T. Corollary 5.7 Let u C 1 (T). Then u C(T). Theorem 5.6 and its corollary can be generalized for dierentiable functions on an open arc I T. To prove this result we study a very special case in the rst step. Note that if I is an open arc on T and an integrable function u is identically zero on I, then Corollary 5.5 ensures that u(ei ) and limr1 V (rei ) exist at all ei I and they are equal. Lemma 5.8 Let u L1 (T), and let V (rei ) = 1 2

(rei D).
Let I be an open arc on T and suppose that u(eit ) = 0 for all eit I. Let J be an open arc such that J I. Then, as r 1, Vr converges uniformly to u on J. Proof. Without loss of generality assume that I = {eit : < t < } and J = {eit : < t < } with 0 < < < , since otherwise we can work with Vr (ei(+0 ) ) where 0 is an appropriate rotation. Let r (ei ) = | V (rei ) u(ei ) |, As in the proof of Theorem 5.6, r (ei ) 1
0
(ei J).
(1 r)2 sin t |u(ei(t) ) u(ei(+t) )| dt. 2(1 cos t)(1 + r2 2r cos t)
But, for all t [0, ] and all ei J, we have u(ei(t) ) = u(ei(+t) ) = 0. Hence, for each ei J, r (ei ) 1

(1 r)2 sin t |u(ei(t) ) u(ei(+t) )| dt 2(1 cos t)(1 + r2 2r cos t)
u 1 (1 r)2 , (1 cos( ))2
which is equivalent to Vr u
L (J)
u 1 (1 r)2 . (1 cos( ))2
5.4. The Hilbert transform of C 1 (T) functions Theorem 5.9 Let u L1 (T), and let V (rei ) = 1 2

115
(rei D).
Let I be an open arc on T and suppose that u C 1 (I). Let J be an open arc such that J I. Then, as r 1, Vr converges uniformly to u on J. Moreover, u C(I). Proof. Let K be an open arc such that J K K I. Then there exists a C (T) function such that 1 on K and 0 on T \ I. In this particular case, since I, J and K are open arcs, one can even provide an explicit formula for . Now, consider u1 = u and u2 = u (1 ) on T and, using the conjugate Poisson kernel, extend them to the open unit disc, say V1 = Q u1 and V2 = Q u2 . The rst observation is that u1 C 1 (T). Hence, by Theorem 5.6, u1 (ei ) and limr1 V1 (rei ) exist at all ei T, they are equal and, as r 1, V1,r converges uniformly to u1 on T. Secondly, u2 L1 (T) and u2 0 on K. Hence, by Lemma 5.8, u2 (ei ) and limr1 V2 (rei ) exist at all ei K, and as r 1, V2,r converges uniformly to u2 on J. Since, on D V = V1 + V2 , and on I u = u1 + u 2 , we immediately see that, as r 1, Vr = V1,r + V2,r converges uniformly to u on J. The uniform convergence ensures that u is continuous on J. Since J is an arbitrary open arc in I, with the only restriction J I, then u is continuous at all points of I.
Exercises
Exercise 5.4.1 Let u L1 (T) and suppose that u(eit ) = 0 for all eit I, where I is an open arc on T. Show that there is a harmonic function V on the domain (C \ T) I whose values on the unit disc D are given by the conjugate Poisson integral V = Q u. Remark 1: This is a generalization of Lemma 5.8. Remark 2: A similar result holds for the harmonic function U = P u.
116
Exercise 5.4.2 Let u L1 (T), and let V = Q u. Let I be a closed arc on T and suppose that u C 1 (I). Show that u C(I). Exercise 5.4.3 Let u C 2 (T). Show that u C 1 (T) and that
0
1 d i u (e ) = d 4 for all ei T.
u(ei(t) ) + u(ei(+t) ) 2u(ei ) dt sin2 (t/2)
5.5
Analytic measures on T
eint d(eit ) = 0
T
A measure M(T) is called analytic if (n) =
for all n 1. In other words, the negative part of the spectrum of vanishes completely. These measures appeared at the end of Section 5.1 in studying the Hardy space H 1 (D). According to F. and M. Rieszs theorem, such a measure is absolutely continuous with respect to the Lebesgue measure. To show this result we start with a construction due to Fatou. Let E be a closed subset of T with |E| = 0, where |E| stands for the Lebesgue measure of E. Fatou constructed a function F on the closed unit disc satisfying the following properties: (i) F is continuous on D; (ii) F is analytic on D; (iii) |F (z)| < 1, for all z D; (iv) |F ()| < 1, for all T \ E; (v) F () = 1, for all E. Such a function is called a Fatou function for E. Certainly F is not unique. For example, F 2 satises all preceding conditions if F does so. Without loss of generality, assume 1 E. Since E is closed, we have T\E =
n
In ,
where In are some open arcs on T, say In = { ei : 0 < n < < n < 2 }.
5.5. Analytic measures on T The condition |E| = 0 is equivalent to (n n ) = 2.

n
117
Hence, there is a sequence n > 0 such that

n
lim n = ,
but still
n (n n ) < . Dene + u(eit ) = n n 2 (t )2 n n
if
eit E,
if eit In ,
where n = (n n )/2 and n = (n + n )/2. This positive function has two important properties. First of all, u is integrable on T: |u(eit )| dt =
T
u(eit ) dt u(eit ) dt +
E n n In n
= =
n
u(eit ) dt n dt = 2 n (n n ) < .
n
2 n
(t n
)2
Secondly, u, as a function from T to [0, +], is continuous. As a matter of fact, u is innitely dierentiable on each In and tends to innity when we approach n or n from within In . Moreover, u(eit ) n on In and n tends to innity as n grows. That is why n entered our discussion. Hence, as eit T approaches any eit0 E, in any manner, u(eit ) tends to innity, i.e.
tt0
lim u(eit ) = +
(5.19)
for all eit0 E. Dene 1 2 i U (re ) =
1+
r2
1 r2 u(eit ) dt 2r cos( t) u(ei )
if if
0 r < 1, r = 1.
By Corollary 2.9, U is continuous on D with values in [0, +], and is harmonic on D. In particular, for each eit0 E,
zeit0 zD
lim U (z) = +.
(5.20)
118
Since u is continuously dierentiable on each In , by Theorem 5.9, the harmonic conjugate V (rei ) = 1 2

(rei D),
has continuous extension up to In , and its boundary values on In are given by the Hilbert transform u(ei ), ei In . Note that we did not dene u on E and moreover, for the rest of the discussion, we do not need to verify if u(ei ) exists or not whenever ei E. Finally, let F : D C be given by U (z) + i V (z) 1 + U (z) + i V (z) u(z) + i u(z) F (z) = 1 + u(z) + i u(z) 1 z D,
if
if z T \ E, if z E.
Clearly, F is analytic on D and continuous at all points of T \ E. Since u > 0 and U > 0, we also have |F (z)| < 1 for all z D and all z T \ E. Moreover, by (5.19) and (5.20), F is continuous at each point of E. Hence, F is continuous on D. Now, we have all the necessary tools to prove the celebrated theorem of the Riesz brothers. This fundamental result is actually a genuine application of Fatous functions for compact subsets of T. Theorem 5.10 (F. and M. Riesz) Let M(T) be an analytic measure. Then is absolutely continuous with respect to the Lebesgue measure. Proof. Put d(eit ) = eit d(eit ). Hence d(eit ) = eit d(eit ), and thus it is enough to show that is absolutely continuous with respect to the Lebesgue measure. The advantage of is that (n) = 0, for all n 0.
(The identity also holds for n = 0.) Thus, for any analytic function G on the unit disc, we have G(reit ) d(eit )
T
=
T n=0
an rn eint
d(eit )
=
n=0
an rn
eint d(eit )
T
=
n=0
an rn (n) = 0,
(0 r < 1).
5.5. Analytic measures on T
119
Suppose E T with |E| = 0. Let K be a compact subset of E and let F be a Fatou function for K. Hence, according to the preceding observation, F (reit )
T
d(eit ) = 0,
(0 r < 1),
for all integers
1. First, let r 1. Since F is bounded on D, we obtain F (eit )

T
d(eit ) = 0,
which is equivalent to (K) +

T\K
F (eit )
d(eit ) = 0.
Secondly, let . But F was constructed such that |F (eit )| < 1 for all eit T \ K. Thus, (K) = 0. By regularity, (E) = 0. In the proof of the theorem, we used the original denition of absolute continuity. The measure is absolutely continuous with respect to provided that (E) = 0 implies (E) = 0. However, in application, we will use the following characterization of absolute continuity: is absolutely continuous with respect to if and only if there is an f L1 () such that d = f d.
Exercises
Exercise 5.5.1 Let an 0 and n an < . Show that there are n 0 such that limn n = and still n n an < . Hint: Since the series n an is convergent, there are n1 < n2 < such that k . n>nk an < 2 Exercise 5.5.2 Let M(T). Suppose that has a one-sided spectrum, i.e. (n) =
T
eint d(eit ) = 0
either for n N or for n N , where N is an integer. Show that is absolutely continuous with respect to the Lebesgue measure. Exercise 5.5.3 Let M(T). Suppose that is analytic and also singular with respect to the Lebesgue measure. Show that = 0.
120
Exercise 5.5.4 Construct a nonzero measure M(T), singular with respect to the Lebesgue measure, such that (2n + 1) = 0, for all n Z. Hint: Put appropriate Dirac measures at 1 and 1.
5.6
Representations of H 1 (D) functions
At this point we have developed all the necessary tools to show that the representations of Section 5.1 for H p (D), 1 < p < , classes are valid even if p = 1. This fact constitutes the fundamental dierence between h1 (D) and H 1 (D). Let F H 1 (D) h1 (D). We saw that, by Theorems 2.1 and 3.7, F (rei ) =

1 r2 d(eit ) = (n) r|n| ein , (rei D), 2 2r cos( t) 1+r n=
where M(T). Moreover, the measures F (rei ) d/2 converge in the weak* topology to as r 1. Hence, for any n 1, (n) =
T
ein d(ei ) = lim
r1
1 2
F (rei ) ein d.
But, since F is analytic,

F (rei ) ein d = rn1

||=r
F () (n+1) d = 0,
which implies (n) = 0 for all n 1. Therefore, by the theorem of F. and M. Riesz, there exists a function f L1 (T) such that d(eit ) = f (eit ) dt/2 and f (n) = (n) = 0 for all n 1. In other words, f H 1 (T). On the other hand if f H 1 (T), then, by Theorem 2.1 and Corollary 2.15, the function F (rei ) = =
n=
1 2
1 r2 f (eit ) dt 1 + r2 2r cos( t)
n=0
f (n) r|n| ein =
f (n) z n ,
(z = rei D),
also represents an element of H 1 (D) and Fr converges to f in the L1 (T) norm as r 1. We state the preceding results as a theorem.
5.6. Representations of H 1 (D) functions
121
Theorem 5.11 Let F be analytic on the unit disc D. Then F H 1 (D) if and only if there exists a unique f H 1 (T) such that F (z) = 1 2

1 r2 f (n) z n f (eit ) dt = 1 + r2 2r cos( t) n=0
(5.21)
for all z = rei D. The series is uniformly convergent on compact subsets of D. Moreover, lim Fr f 1 = 0
r1
and F
1
= f
1.
Using the Cauchy integral formula and the norm convergence of Fr to f , we obtain another representation theorem for functions in H 1 (D). Corollary 5.12 Let F H 1 (D) and denote its boundary values by f H 1 (T). Then 1 f (eit ) F (z) = dt (5.22) 2 1 eit z and

z eit f (eit ) dt = 0 z eit 1
(5.23)
for all z D. Proof. Fix z D. By the Cauchy integral formula, if F (z) = Let r 1. Since 1 2i 1 F () d = z 2
1+|z| 2
< r < 1,
{||=r}
reit F (reit ) dt. reit z
reit 2 1 reit z r |z| 1 |z|
and, by Theorem 5.11, Fr converges to f in the L1 (T) norm, we have F (z) = lim 1 r1 2

reit 1 F (reit ) dt = reit z 2
eit f (eit ) dt. eit z
The second formula has a similar proof if we start with 1 2i z F () d = 0. z 1
{||=r}
122 Note that since 1+ r2
1 r2 eit z eit = it it , 2r cos( t) e z ze 1
(z = rei ),
the Poisson integral formula (5.21) can also be deduced from (5.22) and (5.23).
Exercises
Exercise 5.6.1 Let fn H 1 (T), n 1, and let f L1 (T). Suppose that
n
lim
fn f
= 0.
Show that f H 1 (T). Exercise 5.6.2 Let f, g H 2 (T). Show that f g H 1 (T).
Exercise 5.6.3 Let F H 1 (D) and denote its boundary values by f H 1 (T). Show that i 2r sin( t) F (z) = f (eit ) dt + F (0), (z = rei D). 2 1 + r2 2r cos( t) Hint: Use Corollary 5.12 and the identity eit z eit i2r sin( t) = it + it 1, 2 2r cos( t) 1+r e z ze 1 (z = rei ).
Exercise 5.6.4 Let F H 1 (D) and denote its boundary values by f H 1 (T). Show that F (z) = Hint: By (5.21), F (z) = Moreover, 1 r2 = 1 + r2 2r cos( t) eit + z eit z , (z = rei D). 1 2

1 2
eit + z eit z
f (eit ) dt + i F (0),
(z D).
1 r2 1 + r2 2r cos( t)
f (eit ) dt,
(z = rei D).
Exercise 5.6.5 Let F H 1 (D) and denote its boundary values by f H 1 (T). Show that F (z) = i 2

eit + z eit z
f (eit ) dt + F (0),
(z D).
Hint: Replace F by iF in Exercise 5.6.4.
5.7. The uniqueness theorem and its applications
123
5.7
The uniqueness theorem and its applications
An element f H 1 (T) is dened almost everywhere on T. Hence we are free to change its values on a set of Lebesgue measure zero. Therefore, the level set {ei : f (ei ) = 0} is also dened modulus a set of measure zero. Nevertheless, the Lebesgue measure of this set is well-dened. Amazingly, the size of this set is either 0 or 2. Dierent versions of the following theorem are stated by Fatou, F. Riesz, M. Riesz and Szeg. o Theorem 5.13 (Uniqueness theorem) Let F H 1 (D) with boundary values f H 1 (T). Let E T with |E| > 0. Suppose that f (ei ) = 0 for almost all ei E. Then F 0 on D, or equivalently f = 0 almost everywhere on T. Proof. If |E| = 2, then f = 0 almost everywhere on T and thus by the Poisson integral representation given in Theorem 5.11, F 0 on D. Hence suppose that 0 < |E| < 2. We also suppose that F 0 and then we obtain a contradiction. Without loss of generality assume that F (0) = 0, since otherwise we can work with F (z)/z m , where m is the order of the zero of F at the origin. Let 1 if eit E, |E| u(eit ) = 1 if eit T \ E 2 |E| and, for rei D, let U (rei ) V (re ) Clearly,
i
= =
1 2 1 2
1 r2 u(eit ) dt, 1 + 2r cos( t) 2r sin( t) u(eit ) dt. 1 + r2 2r cos( t) r2
1 1 U (rei ) 2 |E| |E| H (D),
for all rei D. Therefore, exp U + i V which implies
GN = F eN (U +i V ) H 1 (D)
for all integers N 1. Moreover, U (0) = V (0) = 0, which gives GN (0) = F (0). By Lemma 3.10, 1 (5.24) lim U (rei ) = r1 2 |E|
124
for almost all ei T \ E. Let gN denote the boundary values of GN . According to (5.24), N |gN (eit )| = |f (eit )| exp 2 |E| for almost all eit T \ E, and by our main assumption gN (eit ) = 0 for almost all eit E. Therefore, again by the Poisson integral representation given in Theorem 5.11, F (0) = GN (0) = Hence, for all N 1, |F (0)| 1 2 |gN (eit )| dt = 1 exp 2 N 2 |E| |f (eit )| dt. 1 2 gN (eit ) dt = 1 2 gN (eit ) dt.
T\E
T\E
T\E
Let N to get F (0) = 0, which is a contradiction. This result is not valid for the elements of h1 (D). For example, the Poisson kernel 1 r2 , (rei D), P (rei ) = 1 + r2 2r cos is in h1 (D) and
r1
lim P (rei ) = 0
for all ei T \ {1}. However, P 0. We know that elements of H (D) have radial boundary values almost everywhere on the unit circle. This result, along with the uniqueness theorem for H 1 (D) functions, enables us to show that analytic functions with values in a half plane also have radial boundary values. Lemma 5.14 Let F be analytic on the unit disc and suppose that for all z D. Then lim F (rei )
r1
F (z) > 0
exists and is nite for almost all ei T. Proof. We use a conformal mapping between the right half plane and the unit disc to obtain a function in H (D). Put G(z) = Hence |G(z)| < 1 (5.25) 1 F (z) , 1 + F (z) (z D).
5.7. The uniqueness theorem and its applications for all z D and thus, by Theorem 3.10, g(ei ) = lim G(rei )
r1
125
exists for almost all ei T. Moreover, g(ei ) = 1 at most on a set of Lebesgue measure zero. Since otherwise, according to Theorem 5.13, we necessarily would have G 1 on D, which contradicts (5.25). Therefore,
r1
lim F (rei ) = lim
1 g(ei ) 1 G(rei ) = i ) r1 1 + G(re 1 + g(ei )
also exists and is nite for almost all ei T. We are now in a position to prove that the conjugate function also has radial limits almost everywhere on T. We emphasize that Lp (T) L1 (T) M(T) if 1 p , and thus the following result applies to the conjugate of elements of all hp (D) classes. Theorem 5.15 Let M(T), and let V (rei ) = Then 1 2
T
2r sin( t) d(eit ), 1 + r2 2r cos( t) lim V (rei )
(rei D).
r1
exists and is nite for almost all ei T. Proof. Without loss of generality, assume that is a positive Borel measure on T. Since otherwise, using Hahns decomposition theorem, we can write = (1 2 ) + i(3 4 ), where k 0, and then consider each k separately. Let F (z) = Hence F (z) = 1 2
T
1 2
(z D).
(5.26)
1+
r2
(z = rei D),
and, according to our assumption on , we have F (z) > 0 for all z D. Thus F is an analytic function mapping the unit disc into the right half plane. Therefore, by Lemma 5.14, f (ei ) = lim F (rei )
r1
exists and is nite for almost all ei T. In particular, considering the imaginary parts of both sides of (5.26), we see that limr1 V (rei ) exists and is nite for almost all ei T.
126
In the following lemma, we show that the dierence of two functions of r tends to zero almost everywhere as r 1. According to Theorem 5.15, we know that the rst function has a nite limit almost everywhere. Hence, the same conclusion holds for the second one. Therefore, we will be able to conclude that the Hilbert transform of an integrable function exists almost everywhere. Lemma 5.16 Let u L1 (T), and let V ( rei ) = 1 2

(rei D).
Then, for almost all ei T,

r1
lim
V ( rei )
1r<|t|<
u(ei(t) ) dt 2 tan(t/2)
= 0.
Proof. We show that the theorem holds at points ei T where 1 0 lim

0
| u(ei(t) ) u(ei(+t) ) | dt = 0.
According to a well-known theorem of Lebesgue, u fulls this property at almost all points of T. First write V ( rei ) = Hence V ( rei ) where I1 = and I2 = 1
1r 0 1r
1r
+
0 1r 1r
u(ei(t) ) u(ei(+t) )
dt.
u(ei(t) ) u(ei(+t) ) dt = I1 + I2 , 2 tan(t/2) u(ei(t) ) u(ei(+t) ) dt
sin t r sin t 1 + r2 2r cos t 2(1 cos t)
u(ei(t) ) u(ei(+t) )
dt.
For the rst integral, based on our assumption at ei , we have | I1 |

1r 0
r sin t u(ei(t) ) u(ei(+t) ) dt 1 + r2 2r cos t t (1 r)2

1r 0
1r
u(ei(t) ) u(ei(+t) ) dt (as r 1).
1 1r
u(ei(t) ) u(ei(+t) ) dt = o(1),
5.7. The uniqueness theorem and its applications On the other hand, | I2 | = = 1 1 1 1
1r 1r 1r 1r
127
sin t r sin t |u(ei(t) ) u(ei(+t) )| dt 2 2r cos t 1+r 2(1 cos t) (1 r)2 sin t |u(ei(t) ) u(ei(+t) )| dt 2(1 cos t)(1 + r2 2r cos t) (1 r)2 cos(t/2) |u(ei(t) ) u(ei(+t) )| dt 2 sin(t/2)( (1 r)2 + 4r sin2 (t/2) ) (1 r)2 |u(ei(t) ) u(ei(+t) )| dt 8r sin3 (t/2)
1r
(1 r)2 2 8r
|u(ei(t) ) u(ei(+t) )| dt. t3
To show that I2 0, dene F : R R by x 1 |u(ei(t) ) u(ei(+t) )| dt x 0 F (x) = 0
if x = 0, if x = 0.
The function F is continuous at zero and bounded on R. Therefore, doing integration by parts, we obtain | I2 | = = = (1 r)2 2 8r (1 r) 8r
2 2 1r 1r
|u(ei(t) ) u(ei(+t) )| dt t3 d( tF (t) ) t3
(1 r)2 2 8r
2
tF (t) t3
2
+
t=1r
(1 r)2 2 8r
2 1
1r
3tF (t) dt t4
(1 r) F () F (1 r) 3 + 8r 8r 8r 2 3 F ((1 r) ) o(1) + d. 8r 3 1
/(1r)
F ((1 r) ) d 3
By the dominated convergence theorem, the last integral is also o(1) as r 1. Hence, V (rei ) 1
1r
u(ei(t) ) u(ei(+t) ) dt = o(1), 2 tan(t/2)
(as r 1).
The following result is a direct consequence of Theorem 5.15 and Lemma 5.16.
128
Theorem 5.17 Let u L1 (T), and let V (rei ) = Then both limits u(ei ) = lim and exist, are nite, and besides
r1
1 2
(rei D).
1 0
<|t|<
u(ei(t) ) dt 2 tan(t/2)
r1
lim V (rei )
lim V ( rei ) = u(ei )
for almost all ei T. Corollary 5.18 Let u L1 (T), and let F (z) = Then for almost all ei T. Remark: The assumption u L1 (T) is not enough to ensure that F H 1 (D). Proof. It is enough to note that 1 r2 = 1 + r2 2r cos( t) and that 2r sin( t) = 1 + r2 2r cos( t) eit + z eit z eit + z eit z 1 2

eit + z u( eit ) dt, eit z
(z D).
r1
lim F (rei ) = u(ei ) + i u(ei )
where z = rei D. Thus F (rei ) = (Pr u)(ei ) + i(Qr u)(ei ). The required result follows immediately from Lemma 3.10 and Theorem 5.17.
5.7. The uniqueness theorem and its applications
129
Exercises
Exercise 5.7.1 Let f, g H 1 (T). Let E T with |E| > 0. Suppose that f (ei ) = g(ei ) for almost all ei E. Show that f = g almost everywhere on T. Exercise 5.7.2 Let M(T) be an analytic measure. Prove that either 0 or its support is all of T. Exercise 5.7.3 Suppose that Let u, v L1 (T) be real and such that f = u + iv H 1 (T).

u( eit ) dt =
v( eit ) dt = 0. v = u,
Show that u=v or equivalently f = if. Hint: Use Corollary 5.18. Exercise 5.7.4 Let f H 1 (T), and let u = f . Suppose that

and
u( eit ) dt = 0.
Show that u = u. Hint: Use Exercise 5.7.3. Exercise 5.7.5 Let F H 1 (D) and denote its boundary values by f H 1 (T). Show that F (z) = i 2

eit + z eit z
f (eit ) dt + F (0),
(z D).
Hint: Use Exercises 5.6.5 and 5.7.3.
130
Chapter 6
Norm inequalities for the conjugate function

6.1 Kolmogorovs theorems
Let u L1 (T), and let U = P u. Then, by Corollary 2.15, U h1 (D) and, by Fatous theorem, lim U (rei ) = u(ei )
r1
for almost all e T. For the conjugate function V = Q u, according to Theorem 5.17, we also know that
r1
lim V (rei ) = u(ei )
for almost all ei T. However, the mere assumption u L1 (T) is not enough to ensure that u L1 (T) and V h1 (D). The best result in this direction is due to Kolmogorov, showing that u is in weak-L1 (T) and that V hp (D) for all 0 0, mu () 64 u 1 . 4 u 1+
Proof. (Carleson) Fix > 0. First suppose that u 0 and u 0. Therefore, the analytic function F (z) = U (z) + iV (z) = 1 2

(z D),
maps the unit disc into the right half plane and F (0) = 1 2

u(eit ) dt = u 1 .
131
132
Chapter 6. Norm inequalities for the conjugate function The conformal mapping z 2z z+
maps the right half plane into the disc {|z 1| 1} (See Figure 6.1). In particular, it maps {|z| , z 0} into {|z 1| 1, z 1}.
Fig. 6.1. The conformal mapping z Therefore, G(z) = 2F (z) , F (z) + (z D),
2z z+ .
is in H (D) and, by Lemma 3.10 and Theorem 5.17, its boundary values are given by 2(u(ei ) + i u(ei )) g(ei ) = lim G(rei ) = r1 u(ei ) + i u(ei ) + for almost all ei T. Hence, in the rst place, {g} = 1 + u2 + u 2 2 1 (u + )2 + u2
provided that u > . Secondly, by Theorem 3.5, G(0) = which implies 2 u 1 1 = u 1+ 2 Therefore, 4 u 1 u 1+ {g(eit ) } dt dt = mu (). 1 2

g(eit ) dt,
{ g(eit ) } dt.
{||>} u
{||>} u
6.1. Kolmogorovs theorems
133
For an arbitrary real u L1 (T), write u = u1 u2 , where uk 0 and u1 u2 0. Thus u 1 = u1 1 + u2 1 and u = u1 u2 . In this case, the inclusion {ei : |(ei )| > } {ei : |1 (ei )| > /2 } {ei : |2 (ei )| > /2 } u u u implies mu () mu1 (/2) + mu2 (/2) 4 u1 1 4 u2 1 + u1 1 + /2 u2 1 + /2 16 u 1 8 ( u1 1 + u2 1 ) . = ( u1 1 + u2 1 ) + /2 2 u 1+ For an arbitrary u L1 (T), write u = u1 + i u2 where u1 and u2 are real functions in L1 (T). Hence, u1
1
and
u2
u 1.
Since u = u1 + i u2 , again the inclusion {ei : |(ei )| > } {ei : |1 (ei )| > /2 } {ei : |2 (ei )| > /2 } u u u gives mu () mu1 (/2) + mu2 (/2), and thus mu () 16 u2 1 64 u 1 16 u1 1 + . 2 u1 1 + /2 2 u2 1 + /2 4 u 1+
Let u L1 (T), and let V = Q u. The condition u L1 (T) is not enough to conclude that V h1 (D). Since otherwise, by Fatous lemma, we would have u
1
lim inf Vr
r1
<
which is not true in the general case. However, we can show that V hp (D) for all 0 < p < 1. Corollary 6.2 (Kolmogorov) Let u L1 (T), and let V (rei ) = 1 2

(rei D).
Then, for each 0 < p < 1, V hp (D), u Lp (T) and V

p
= u
cp u 1 ,
where cp is a constant just depending on p.
134
Chapter 6. Norm inequalities for the conjugate function
Proof. Fix 0 < p < 1. Let U (rei ) = 1 2

1 r2 u(eit ) dt, 1 + r2 2r cos( t)
(rei D).
Clearly U h1 (D) hp (D), and F = U +iV is an analytic function on D. Based on the content of Section 5.2, for each xed 0 < r < 1, the function Vr (ei ) is the Hilbert transform of Ur (ei ). Hence, by Theorem 6.1 and Corollary 4.13, Vr
p p
= =
1 2 1 2
0
|Vr (ei )|p d p p1 mVr () d 32 Ur 1 d 4 Ur 1 + 32 u 1 d. 4 u 1+

1
p p1 p p1
See also Section A.5. Making the change of variable = u Vr

p
s gives us
32p
0
sp1 ds 4+s
1/p
u 1.
Hence V hp (D) and V
cp u 1 . Since
r1
for almost all ei T, an application of Fatous lemma implies that u

p
and thus u Lp (T). To prove that equality holds we need more tools, which will be discussed in Section 7.5. However, we do not use this fact. A function F = U + iV is in H p (D) if and only if U, V hp (D). If u L1 (T) is given and we dene F (z) = U (z) + iV (z) = 1 2

(z D),
then, by Corollary 2.15, U = P u h1 (D), which implies U hp (D) for all 0 < p < 1. On the other hand, Kolmogorovs theorem ensures that V hp (D) for all 0 < p < 1. Hence, we obtain the following result. Corollary 6.3 (Kolmogorov) Let u L1 (T), and let F (z) = 1 2

(z D).
6.2. Harmonic conjugate of h2 (D) functions Then F H p (D), for all 0 < p < 1, and F
p
135
cp u 1 ,
where cp is a constant just depending on p.
Exercises
Exercise 6.1.1 Exercise 6.1.2 Construct u L1 (T) such that u L1 (T). Construct u L1 (T) such that F H 1 (D), where 1 2

F (z) =
(z D).
Hint: Use Exercise 6.1.1.
6.2
Harmonic conjugate of h2 (D) functions
The following deep result is a special case of M. Rieszs theorem (Theorem 6.6). However, we provide an extra proof for this particular, but important, case. We call your attention to the number of results used in its proof. Theorem 6.4 Let u L2 (T), and let V (rei ) = 1 2

(rei D).
Then V h2 (D), u L2 (T) and V Moreover, V (rei ) = 1 2

2
= u
2.
1+
r2
(rei D).
Proof. Since u L2 (T), by Parsevals identity (Corollary 2.22),

n=
|(n)|2 = u u
2 2
< .
(6.1)
On the other hand, by Theorem 3.13, V (rei ) =
(i sgn n) u(n) r|n| ein .
n=
136 Hence, for 0 r < 1, Vr

2 2
1 2
|V (rei ) |2 dt =
n= n=0
|(n)|2 r2|n| . u
Thus, by the (discrete version) of the monotone convergence theorem, V

2 2
=
n= n=0
| u(n) |2 .
Comparing with (6.1), we obtain V

2 2
= u
2 2
| u(0) |2 .
Thus V h2 (D) and V 2 u 2 . Knowing that V h2 (D), by Theorem 3.6, there exists v L2 (T) such that V (rei ) = and V
2
1 2
1 r2 v( eit ) dt, 1 + r2 2r cos( t) = v 2.
(rei D),
According to Lemma 3.10, the radial limit limr1 V (rei ) exists and is equal to v(ei ) for almost all ei T. On the other hand, by Theorem 5.17,
r1
for almost all ei T. Hence u = v.
6.3
M. Rieszs theorem
To prove M. Rieszs theorem we consider two cases, 1 < p < 2 and 2 < p < , and for the second case we need the following lemma. Lemma 6.5 Let U1 and U2 be harmonic functions on D. Let V1 and V2 denote their harmonic conjugates satisfying V1 (0) = V2 (0) = 0. Then, for each 0 r < 1,

U1 (rei ) V2 (rei ) d =
V1 (rei ) U2 (rei ) d.
Proof. The function F (z) = U1 (z) + iV1 (z) is analytic on D. Hence, its imaginary part U1 (z) V2 (z) + V1 (z) U2 (z) U2 (z) + iV2 (z)
6.3. M. Rieszs theorem is harmonic on D and U1 (0) V2 (0) + V1 (0) U2 (0) = 0. Therefore, by Corollary 3.2, 1 2

137
U1 (rei ) V2 (rei ) + V1 (rei ) U2 (rei ) d = 0.
The following celebrated result is due to Marcel Riesz. According to this result an analytic function F = U + iV is in H p (D), 1 < p < , if and only if its real part U is in hp (D). It has several interesting proofs. The one given here is by P. Stein. In the proof, two elementary identities for the Laplacian 2 = 2 2 + 2 x2 y
of analytic functions are used. These identities are simple consequences of the CauchyRiemann equations. Moreover, we need Greens formula

i (re ) rd = r
r 0
2 (ei ) dd,
where is a twice continuously dierentiable function on the unit disc D. Theorem 6.6 (M. Riesz) Let u Lp (T), 1 < p < , and let V (rei ) = 1 2

(rei D).
Then V hp (D), u Lp (T) and V

p
= u
cp u
p,
where cp is a constant just depending on p. Moreover, V (rei ) = 1 2

1 r2 u(eit ) dt, 1 + r2 2r cos( t)
(rei D).
Proof. The case p = 2 was established in Theorem 6.4 (as a matter of fact, a slight modication of the following proof also works for p = 2). Case 1: Suppose that 1 0,
(z D).
Moreover, F (z) = U (z) + iV (z) is analytic on D with F (0) = U (0) and |F (z)| > 0, (z D).
Since U and |F | are strictly positive on D, they are innitely dierentiable at each point of D. Moreover, by the CauchyRiemann equations, 2 U p (z) = p(p 1) |F (z)|2 |U (z)|p2 and 2 |F |p (z) = p2 |F (z)|2 |F (z)|p2 , p 2 U p (z) p1
r
and thus (note that p 2 < 0) 2 |F |p (z) (6.2)
for all z D. Now, by Greens theorem, |F |p (rei ) rd r p U (rei ) rd r Therefore, by (6.2), we have d dr

=
0 r
2 |F |p (ei ) dd, 2 U p (ei ) dd.
=
0
|F (rei )|p d
d p p 1 dr
U p (rei ) d .
Integrating both sides over [0, r] gives us

|F (rei )|p d
|F (0)|p d
p p1
U p (rei ) d
U p (0) d .
Thus 1 2

|V (rei )|p d
1 |F (rei )|p d 2 1 p p U p (0). U p (rei ) d p 1 2 p1
Therefore, by Corollary 2.15, Vr Case 2: 1 < p < 2.

p
p p1
1 p
Ur
p p1
1 p
p.
(6.3)
6.3. M. Rieszs theorem
139
For a general u Lp (T), write u = (u(1) u(2) ) + i (u(3) u(4) ), where u(k) 0 (1) (2) and u(1) u(2) u(3) u(4) 0. Let V (k) = Q u(k) . Hence Vr = (Vr Vr ) + (3) (4) i (Vr Vr ), and by (6.3), Vr
p
Vr(1)
+ Vr(2)
1 p
+ Vr(3)
p
+ Vr(4)
p
p p
p p1 p p1 p p1
u(1)
+ u(2)
+ u(3)
+ u(4)
Therefore, Vr
p p
4p = 4p
u(1)
p p
+ u(2)
p p
p p
+ u(3)
p p
+ u(4)
p p
p p
u(1) u(2) u p. p
+ u(3) u(4)
2 4p
p p1
This nishes the case 1 < p < 2. Case 3: 2 < p < . Let q be its conjugate exponent of p (note that 1 < q < 2). Let
M
U (1) (rei ) =
n=N
an r|n| ein ,
(rei D),
be an arbitrary trigonometric polynomial. The harmonic conjugate of U (1) is

M
V (1) (rei ) =
n=N
i sgn(n) an r|n| ein ,
(rei D).
By Lemma 6.5, by Hlders inequality, and then by (6.3), we have o 1 2

V (rei ) U (1) (rei ) d
1 2 Ur u
p p
U (rei ) V (1) (rei ) d

q
Vr(1)
cq
(1) Ur q .
Since trigonometric polynomials are dense in Lq (T), if we take the supremum over all nonzero polynomials U (1) in the inequality 1 Ur we obtain Vr
p (1) q
1 2
V (rei ) U (1) (rei ) d cq u p ,
cq u p .
p
Therefore, we now know that V hp (D), 1 < p < , with V The rest of the proof is as given for Theorem 6.4.
cp u p .
140
Chapter 6. Norm inequalities for the conjugate function If u L1 (T) is real, then certainly F (z) = 1 2

(z D),
is an analytic function on the unit disc satisfying F (rei ) = 1 2

1+
r2
(rei D).
If we know that F H p (D), 1 p , then we necessarily have F hp (D), and thus, by Theorems 3.5, 3.6 and 3.7, u Lp (T). M. Rieszs theorem enables us to obtain the inverse of this result whenever 1 < p < . It is not dicult to construct u L1 (T) such that F H 1 (D), or similarly a function u L (T) which gives F H (D). Corollary 6.7 Let u Lp (T), 1 < p < , be real, and let F (z) = Then F H p (D) and u
p
1 2
eit + z u(eit ) dt, eit z F

p
(z D).
cp u p ,
where cp is a constant just depending on p. Moreover, the boundary values of F are given by f = u + i u H p (T). Proof. By Corollary 2.15, F hp (D) and by M. Rieszs theorem, F hp (D). Hence, F H p (D). Moreover, by Corollary 2.15 and Theorem 6.6, we have u
p
cp u p .
Then, by Corollary 5.18, f (ei ) = lim F (rei ) = u(ei ) + i u(ei )

r1
for almost all ei T. In other words, there is an f H p (T) such that u = f and u = f .
Exercises
Exercise 6.3.1 Find u L1 (T) such that F H 1 (D), where 1 2

F (z) =
(z D).
6.3. M. Rieszs theorem Exercise 6.3.2 Find u L (T) such that F H (D), where 1 2

141
F (z) =
(z D).
Exercise 6.3.3
Let u Lp (T ), 1 < p < . Show that u(n) = i sgn(n) u(n), (n Z).
Hint: Use both representations given in Theorem 6.6. Exercise 6.3.4 Let u Lp (T), 1 < p < . Let P u(eit ) = Show that P u H (T) and that Pu
p p n=0
u(n) eint .
Cp u p ,
where Cp is an absolute constant. Hint: Use Theorem 6.6 and Exercise 6.3.3. Note that 2P u = u(0) + u + i. u Remark: The operator P : Lp (T) H p (T) is called the Riesz projection. Exercise 6.3.5 Let u Lp (T), 1 < p < , with

u( eit ) dt = 0.
Show that u = u. Hint: By Corollary 6.7, there is an f H p (T) such that f = u + i. Replace f u by if . Remark: Compare with Exercise 5.7.4. A real u L1 (T) is not necessarily the real part of an f H 1 (T). Exercise 6.3.6 Let u Lp (T), 1 < p < . Suppose that

u( eit ) dt = 0.
Show that there are constants cp and Cp such that cp u

p
Cp u p .
Hint: Use Exercise 6.3.5 and Theorem 6.6.
142
6.4
The Hilbert transform of bounded functions

u
1<p<
If u L (T), then u Lp (T) for all 1 < p < , and thus by Theorem 6.6, Lp (T). (6.4)
However, by a simple example, we see that u is not necessarily in L (T) for an arbitrary u L (T). Let us consider the analytic function F (z) 1+z 1z 2r sin = arctan 1 r2 = i log
i log 2
1 + r2 + 2r cos 1 + r2 2r cos
where z = rei D. Hence

i
u(e ) = lim
r1
2 i F (re ) = 2
if if
0 < < , < < 0,
and
v(ei ) = lim
n=0
r1
F (rei ) = log | tan(/2)|.
Since F (z) = 2 Theorem 5.9,

i
z 2n+1 /(2n + 1), by Theorem 5.1, F H 2 (D). Thus, by u(ei ) = v(ei ) = log | tan(/2)| (6.5)
for e T \ {1, 1}. It is clear that u is unbounded. Nevertheless, u Lp (T) for all 1 < p < . Now we show that the Hilbert transform of a bounded function satises a stronger condition than (6.4). Theorem 6.8 (Zygmund) Let u L (T) be real with u Then 1 2

. 2
u e|(e
)|
cos(u(ei )) d 2.
Proof. Let F (z) = 1 2

eit + z u(eit ) dt = U (z) + i V (z), eit z
(z D).
The condition u
/2 implies U ( rei ) , 2 2
6.4. The Hilbert transform of bounded functions and thus cos( U ( rei ) ) 0, for all rei D. Fix r < 1. Then 1 2

143
ei F (re
d = ei F (0) .
Taking the real part of both sides gives 1 2

eV (re
cos( U (rei ) ) d = cos( U (0) ) 1.
Replace F by F and repeat the preceding argument to get 1 2

eV (re
cos( U (rei ) ) d 1.
The last two inequalities together imply 1 2 Now, by Fatous lemma, 1 2 1 2

r1 u e|(e
i
e|V (re
)|
)|
cos( u(ei ) ) d e|V ( re

i
lim inf
r1
)|
cos( U (rei ) )
lim inf
1 2
e|V (re
)|
Corollary 6.9 Let u L (T) be real. Then 1 2 for 0< Proof. By Theorem 6.8, 1 2
). u e|( e
i
u e |(e
)|
d 2 sec( u 2 u
)|
cos( u( ei ) ) d 2
provided that 0 < /(2 u u for all rei D, and
Since
,
U (rei ) u u < , 2
144 then Therefore,
cos( U (rei ) ) cos( u 1 2

u e |(e
i
) > 0. ).
)|
d 2/ cos( u
As (6.5) shows, the condition < /(2 u = /(2 u ) = 1 then
is sharp. In that example, if
exp |(ei )| = 1/| tan /2| u for (/2, /2), and this function is not integrable. Corollary 6.10 Let u L (T) be real. Then lim sup
n
2 u u n . n e
Proof. On the one hand, for > 0, 1 2

u e |(e
i
)|
d =
n=0
u n n n n!
and the series is convergent if and only if lim sup

n
u n < 1. (n!)1/n
On the other hand, by Corollary 6.9, 1 2 provided that 0 < /(2 u

). u e |(e
i
)|
d < ,
Therefore,
lim sup
n
u n 2 u (n!)1/n
n (n!)1/n
Finally, by Stirlings formula limn
= e and hence the result follows.
6.5
The Hilbert transform of Dini continuous functions
A continuous function on T is certainly bounded. Hence, at least its Hilbert transform is integrable as described in Corollary 6.9. But continuity on T and the fact that trigonometric polynomials are dense in C(T) enable us to get rid of the restriction < /(2 u ) in this case.
6.5. The Hilbert transform of Dini continuous functions Theorem 6.11 Let u C(T). Then
145
exp
|(ei )| u
d <
for all R. Proof. The assertion for 0 is trivial. Hence suppose that > 0. Fix > 0. N There is a trigonometric polynomial n=M an ein such that < , where
N
(ei ) = u(ei )
n=M
an ein ,
(ei T).
Hence u(e ) = (e ) +
i i
i sgn n an ein ,
n=M N
(ei T),
and thus |(ei )| |(ei )| + u
|an |,
n=M
(ei T).
By Corollary 6.9, we have

u e |(e
i
)|
d e
N n=M
|an |
e |(e
)|
d < ,
whenever 0<
< 2 2
u Since is arbitrary, e || is integrable for all .
The modulus of continuity of u C(T) is dened by (t) = u (t) = sup

| |<t
|u(ei ) u(ei )|.
(6.6)
One can easily show that is a positive decreasing function with

t0
lim (t) = 0
and (t1 + t2 ) (t1 ) + (t2 ) for all t1 , t2 > 0. Moreover, if we dene f for an arbitrary function f as above, then f is continuous on T if and only if limt0 f (t) = 0. The function u is called Dini continuous on T if
0
(t) dt < . t
146
Clearly, the upper bound can be replaced by any other positive number. If u is Dini continuous on T, then
0
lim
(t) dt = 0 t
(6.7)
and the inequality
(t) dt t2
(t) dt + t
(t) dt t
shows that we also have

0
lim
(t) dt = 0. t2
(6.8)
Theorem 6.12 Let u be Dini continuous on T. Then u exists at all points of T and moreover
u () C
u (t) dt + t
u (t) dt , t2
where C is an absolute constant. Proof. Since u(ei(t) ) u(ei(+t) ) 2 tan(t/2) u (t) t
then, by (5.10), Dini continuity ensures that u(ei ) exists at all ei T and it is given by u(ei(t) ) u(ei(+t) ) 1 u( ei ) = dt. (6.9) 0 2 tan(t/2) For a similar reason, Hu(ei ) also exists for all ei T and Hu( ei ) = 1
0
u(ei(t) ) u(ei(+t) ) dt. t
(6.10)
Representations (6.9) and (6.10), along with the property (5.8), imply that
uHu () C u () C at least for < /2. Since
u (t) dt t2
(6.11)
u uHu + Hu it is enough to prove the theorem for Hu. Fix > 0. Consider two arbitrary points 1 and 2 with |2 1 | < and let 3 = (1 + 2 )/2. Without loss of generality, assume that 1 < 2 . Since the
6.5. The Hilbert transform of Dini continuous functions
147
H transform of the constant function is zero, replacing u by u u(3 ), we also assume that u(3 ) = 0. According to the main denition (5.7), Hu(ei ) = lim
0
<|t|<
u(eit ) dt, t
and thus, for any > 0 and any , Hu(ei ) = = = lim 1 1 1 +

<|t|< <|t|<
u(eit ) dt t
<|t|<
<|t|<
u(eit ) 1 dt + lim 0 t 1 u(e ) dt + t

it
u(eit ) u(ei ) dt t
<|t|<
|t|<
u(eit ) u(ei ) dt. t
For the last integral we have u(eit ) u(ei ) dt 2 t

0
|t|<
u ( ) d.
Hence |Hu( ei1 ) Hu( ei2 )| where I1 I2 I3 = = = 1 1 1

2 + 1 + 2 1
u (t) dt + |I1 | + |I2 | + |I3 |, t
u(eit ) dt, 1 t u(eit ) dt, 2 t u(eit ) 1 1 1 t 2 t dt.
( ) + 2 2 1 <|t3 |<
We now estimate each of these integrals: |I1 | 1 1

2 + 1 + 2 + 1 +
u(eit ) u(ei3 ) 1 t u(eit ) u(ei3 ) t 3
dt dt
u (3/2) 3/2 dt t /2 3 ( log 3 ) u (). 2
148
A similar estimate holds for I2 and thus, by (6.11),
|Ik | C
(t) dt, t2
(k = 1, 2).
Finally, for the third integral we have |I3 | (2 1 ) 9 4 9 4

+
(2 1 ) <|t3 |< 2
u(eit ) u(ei3 ) dt (1 t)(2 t)
(2 1 ) <|t3 |< 2
u(eit ) u(ei3 ) dt (3 t)2
(t) dt. t2
Corollary 6.13 Let u be Dini continuous on T. Then u exists at all points of T and moreover u C(T). Proof. By (6.7) and (6.8) and by Theorem 6.12, we have
0
lim u () = 0.
Hence, u C(T). A function u C(T) is said to be Lip , 0 < 1, if u (t) = O( t ) as t 0. Clearly every Lip function is Dini continuous on T. Hence u exists at all points of T. Moreover, u is continuous on T. The following results provide more information about the modulus of continuity of u. Corollary 6.14 (Privalov) Let u be Lip on T with 0 < < 1. Then u is also Lip on T. Proof. By Theorem 6.12, we have

u () C C Hence, u is also Lip on T.
0 0
(t) dt + t c t dt + t
(t) dt t2 c t dt t2 C .
Corollary 6.15 Let u be Lip1 on T. Then u () = O( log 1/) as 0.
6.6. Zygmunds L log L theorem Proof. By Theorem 6.12, we have
149
u () C C
(t) dt + t
(t) dt t2
ct dt + t 0 C ( + log /).
ct dt t2
Exercises
Exercise 6.5.1 suppose that Let 0 < < 1, and let 1 , . . . , n R. Let u C(T) and
u (t) = O( t ( log 1/t )1 ( log log 1/t )2 ( log log log 1/t )n ) as t 0. Show that u (t) = O( t ( log 1/t )1 ( log log 1/t )2 ( log log log 1/t )n ) as t 0. Exercise 6.5.2 that Let 1 > 0, and let 2 , . . . , n R. Let u C(T) and suppose
u (t) = O( t( log 1/t )1 ( log log 1/t )2 ( log log log 1/t )n ) as t 0. Show that u (t) = O( t( log 1/t )1+1 ( log log 1/t )2 ( log log log 1/t )n ) as t 0.
6.6
Zygmunds L log L theorem
As we mentioned before, the assumption u L1 (T) is not enough to ensure that u is also in L1 (T). A. Zygmunds L log L theorem provides a sucient condition which ensures u L1 (T). Theorem 6.16 (Zygmund) Let u be real, and suppose that u log+ |u|
1
1 2
|u(eit )| log+ |u(eit )| dt < .
150 Let V (rei ) = 1 2
(rei D).
Then V h1 (D), u L1 (T) and V Moreover, V (rei ) = 1 2

1
= u
u log+ |u|
+ 3e.
1 r2 u(eit ) dt, 1 + r2 2r cos( t)
(rei D).
Proof. First suppose that u e, and let U (rei ) = 1 2

1+
r2
(rei D).
(6.12)
Thus U (z) e, and F (z) = U (z) + iV (z) is analytic on D with F (0) = U (0). The function t log t if t 1, (t) = 0 if t 1 is nondecreasing and convex on R. Thus, applying Jensens inequality to (6.12) gives us U (rei ) log U (rei ) 1 2

1 r2 u(eit ) log u(eit ) dt 1 + r2 2r cos( t)
for all rei D. Hence, by Fubinis theorem,

U (rei ) log U (rei ) d
u(eit ) log u(eit ) dt
(6.13)
for each 0 r < 1. Since |F | U > e, both U log U and |F | are innitely dierentiable on D. Moreover, by the CauchyRiemann equations, 2 (U log U ) (z) = and 2 (|F |) (z) = and thus we have |F (z)|2 U (z)
|F (z)|2 , |F (z)|
2 (|F |) (z) 2 (U log U ) (z)
6.6. Zygmunds L log L theorem for all z D. Now, Greens theorem says that

151
|F | (rei ) rd r (rei ) rd
=
0 r
2 |F | (ei ) dd, 2 (U log U ) (ei ) dd.
U log U
=
0
Therefore, d dr

|F (rei )| d
d dr
U (rei ) log U (rei ) d .
Integrating both sides over [0, r] gives us

|F (rei )| d
|F (0)| d
U (0) log U (0) d.
Hence, 1 2

|V (rei )| d
1 2 1 2 1 2
|F (rei )| d U (rei ) log U (rei ) d + U (0) (1 log U (0)) U (rei ) log U (rei ) d.
(Here we used the fact that U e.) Finally, by (6.13), 1 2

|V (rei )| d
1 2
u(eit ) log u(eit ) dt
(6.14)
for all 0 r < 1. This settles the case whenever u e. For an arbitrary u, write u = u(1) + u(2) u(3) , where u(ei ) > e, u(ei ) if e if e u(ei ) e, u(1) (ei ) = e if u(ei ) < e e u(ei ) e if u(ei ) > e,
and u(2) (ei ) =
if e u(ei ) e, if u(ei ) < e
152 and
e e u(ei )
if
u(ei ) > e,
u(3) (ei ) =
if e u(ei ) e, if u(ei ) < e.
The functions u(1) , u(2) and u(3) are dened such that u(1) e, u(3) e and (1) (2) (3) e u(2) e. Let V (k) = Q u(k) . Hence Vr = Vr + Vr Vr . Thus, by (6.14), by Corollary 2.19 and Theorem 6.4, we have Vr
1
Vr(1)
+ Vr(2)
+ Vr(3)
u(1) log u(1) 1 + Vr(2) 2 + u(3) log u(3) 1 (2) u log+ |u| 1 + 2e + Ur 2 + (2) u log |u| 1 + 2e + Ur u log+ |u|
1
+ 3e.
Therefore, V h (D) with V
u log |u|
1
+ 3e. Hence
F = U + iV H (D) and, by Lemma 3.10 and Theorem 5.17, f (ei ) = lim F (rei ) = u(ei ) + i(ei ) u
r1
for almost all e
T. Thus, by Theorem 5.11,

F (rei ) =
1 2 1 2
1 r2 f (eit ) dt, 1 + r2 2r cos( t) 1 r2 u(eit ) dt, 1 + r2 2r cos( t)

1
(rei D).
Taking the imaginary part of both sides gives V (rei ) =

(rei D).
Finally, by Corollary 2.15, V
= u 1.
Zygmunds theorem can easily be generalized for complex functions u satisfying u log+ |u| L1 (T). It also enables us to construct functions in H 1 (D). Corollary 6.17 Let

|u(eit )| log+ |u(eit )| dt < ,

and let F (z) = Then F H 1 (D). 1 2
(z D).
Proof. Without loss of generality, suppose that u is real. Since u L1 (T), by Corollary 2.15, F h1 (D), and since u log+ |u| L1 (T), by Zygmunds theorem, we also have F h1 (D). Hence F H 1 (D).
6.7. M. Rieszs L log L theorem
153
6.7
M. Rieszs L log L theorem
M. Riesz showed that Zygmunds sucient condition given in Theorem 6.16 is also necessary if u is lower bounded. However, this condition cannot be relaxed and the inverse of Zygmunds theorem, in the general case, is wrong. Theorem 6.18 (M. Riesz) Let u be real, u C > , and let u L1 (T). Let V (rei ) = 1 2

(rei D).
Suppose that V h1 (D). Then u log+ |u| L1 (T). Proof. Write u = u(1) + u(2) , where u(ei ) 1 (1) i u (e ) = 0 and u Since
(2)
if u(ei ) < 1, if u(ei ) 1
(e ) =
1 u(ei )
if u(ei ) < 1, if u(ei ) 1.
|u| log+ |u| |C| log+ |C| + |u(2) | log+ |u(2) |,
it is enough to show that |u(2) | log+ |u(2) | L1 (T). Therefore, without loss of generality, we assume that u 1. Let U (rei ) = 1 2

1 r2 u(eit ) dt, 1 + r2 2r cos( t)
(rei D),
and let F = U + i V . Then F (0) = U (0) and F maps the unit disc into the half plane z 1. Using the main branch of the logarithm on C \ (, 0], i.e. log z = log |z| + i arg z, where < arg z < , we have 1 2 Hence, 1 2

F (rei ) log F (rei ) d = F (0) log F (0) = U (0) log U (0).
U (rei )+i V (rei )
log |F (rei )|+i arg F (rei )
d = U (0) log U (0).
Taking the real part of both sides gives 1 2

U (rei ) log |F (rei )| V (rei ) arg F (rei )
d = U (0) log U (0).
154 Therefore,
1 U (rei ) log U (rei ) d 2 1 U (rei ) log |F (rei )| d 2 1 |V (rei )| | arg F (rei )| d + U (0) log U (0) 2 1 |V (rei )| d + U (0) log U (0) 2 2 V 1 + U (0) log U (0). 2
Hence, by Fatous lemma, 1 2

u(ei ) log u(ei ) d
lim inf
r1
1 2
1
V 2
+ U (0) log U (0) < .
Exercises
Exercise 6.7.1 Construct u 1, u L1 (T), such that u log u L1 (I) for any open arc I T. Exercise 6.7.2 Construct u L1 (T) such that u L1 (I) for any open arc I T. Hint: Use Exercise 6.7.1 and Theorems 6.16 and 6.18.
Chapter 7
Blaschke products and their applications

7.1 Automorphisms of the open unit disc
lim |zn | = 1.
Let (zn )n1 be a sequence in the open unit disc D with

n
Such a sequence has no accumulation point inside D. Therefore, according to a classical theorem of Weierstrass, there is a function G, analytic on the open unit disc, such that G(zn ) = 0 for all n 1. We are even able to give an explicit formula for G. If F is another analytic function with the same set of zeros, then K = F/G is a well-dened zero-free analytic function on D. In other words, we can write F = G K, where G is usually an innite product constructed using the zeros of F , and K is zero-free. This decomposition is ne and suitable for elementary studies of analytic functions. However, if we assume that F H p (D), then we are not able to conclude that G or K are in any Hardy spaces. Of course, the choice of G is not unique. F. Riesz observed that the zeros of a function F H p (D) satisfy the Blaschke condition (1 |zn |) < .
n
Hence he was able to take a Blaschke product in order to extract the zeros of F . More importantly, he showed that the zero-free function that we obtain stays in the same Hardy space and has the same norm as F . A bijective analytic function f : D D is called an automorphism of the open unit disc. Clearly, f 1 : D D, as a function, is well-dened and bijective. Moreover, based on elementary facts from complex analysis, we know 155
156
Chapter 7. Blaschke products and their applications
that f 1 is analytic too. In other words, f 1 is also an automorphism of the open unit disc. It is easy to see that the family of all automorphisms of the open unit disc equipped with the law of composition of functions is a group. As a matter of fact, the preceding assertion is true for the class of automorphisms of any domain in the complex plane C. What makes this specic case more interesting is that we are able to give an explicit formula for all its elements. Let w D and let wz bw (z) = . 1wz The function bw is called a Blaschke factor or a Mbius transformation for the o open unit disc. The rst important observation is that bw bw = id for any w D, where id is the identity element dened by id(z) = z. For each ei T and w = ei D, we have bw (ei ) = and thus 1 ei() ei ei = ei , 1 ei ei 1 ei() |bw (ei )| = 1. |bw (z)| < 1 (7.2) (7.1)
Hence, by the maximum principle, (7.3)
for all z D. The properties (7.1) and (7.3) show that each bw is an automorphism of the open unit disc. Similarly, (7.1) and (7.2) show that the restriction bw : T T is a bijective map. Let f be any automorphism of the open unit disc. Then, by denition, there is a unique w D such that f (w) = 0. Dene g = f bw . Then g is also an automorphism of the open unit disc and g(0) = 0. Hence, by Schwarzs lemma, we have |g(z)| |z|, (z D). The essential observation here is that exactly the same argument applies to g 1 . Since g 1 is an automorphism of the open unit disc and g 1 (0) = 0, then |g 1 (z)| |z|, (z D).
Replace z by g(z) in the last inequality to get |z| |g(z)|, and thus |g(z)| = |z|, (z D).
Therefore, there is a constant T such that g(z) = z. Hence (f bw )(z) = z, (z D).
7.1. Automorphisms of the open unit disc Finally, replace z by bw (z) in the last identity to obtain f = bw . Clearly, any such function is also an automorphism of the open unit disc.
157
Exercises
Exercise 7.1.1 (a) (b) U ei Let U h(D). Show that U (z n ) h(D), z 1 z (n 1).
h(D),
(|| < 1, R).
Hint: Use Exercise 1.1.4. Exercise 7.1.2 Show that and that |F (0)| 1. Moreover, if |F (z)| = |z| for one z D \ {0}, or if |F (0)| = 1, then F (z) = z, where is a constant of modulus one. Hint: Consider G(z) = F (z)/z and apply the maximum modulus principle. Exercise 7.1.3 Verify that bw (0) = (1 |w|2 ) and that bw (w) = 1 . 1 |w|2 [Schwarzs lemma] Let F H (D), F |F (z)| |z|, (z D),
1 and F (0) = 0.
Exercise 7.1.4
Let T, and let w D. Dene f (z) = bw (z), (z D).
Show that f is an automorphism of the open unit disc. Find T and w D such that f = bw .
158 Exercise 7.1.5 Exercise 7.1.6
Chapter 7. Blaschke products and their applications Let w, w D. Compute bw bw . Let , D, and let
A, = { F H (D) : F Find
F A,
1, F () = }.
sup |F ()|.
Hint: Consider G = b F b . Then G(0) = 0.
7.2
Blaschke products for the open unit disc
Let {zn }1 n N be a nite sequence of complex numbers inside the unit disc D and let R. Then N zn z B(z) = ei 1 zn z n=1 is called a nite Blaschke product for the unit disc. In the sequence {zn }1 n N , repetition is allowed. Based on the observation for Blaschke factors, we immediately see that |B(ei )| = 1 (7.4) for all ei T, and |B(z)| < 1 (7.5)
for all z D. But, if N 2, B is not one-to-one. More generally, consider an innite sequence of complex numbers in the unit disc {zn }n 1 , which is indexed such that 0 |z1 | |z2 |
and limn |zn | = 1. Let = {zn }n 1 denote the set of all accumulation points of {zn }n 1 . Since limn |zn | = 1, is a closed nonempty subset of T. Moreover, also coincides with the accumulation points of the set {1/n : n z 1}. Let = {zn }n
1
=C\
{1/n : n z
1} .
Note that we always have D for any possible choice of {zn }n 1 . Besides, T \ is also a subset of and if it is nonempty, being a countable union of disjoint open arcs of T, it provides the connection between the disjoint open sets D and \ D. Let N |zn | zn z BN (z) = . (7.6) z 1 zn z n=1 n
7.2. Blaschke products for the open unit disc
159
By convention, we put |zn |/zn = 1 whenever zn = 0. Under a certain condition on the rate of growth of {|zn |}n1 , the partial products BN converge uniformly on compact subsets of to a nonzero analytic function, which is denoted by B(z) = |zn | zn z . z 1 zn z n=1 n
(7.7)
The function B is called an innite Blaschke product for the open unit disc. Theorem 7.1 (Blaschke [3]) Let {zn }n 1 be a sequence in the open unit disc such that limn |zn | = 1. Let = {zn }n 1 denote the set of accumulation z 1} . Then the partial points of {zn }n 1 and let = C \ {1/n : n products N |zn | zn z BN (z) = z 1 zn z n=1 n are uniformly convergent on compact subsets of if and only if
( 1 |zn | ) < .
n=1
Remark: There is no restriction on arg zn . Proof. Suppose that the sequence {BN }N 1 is uniformly convergent to B on compact subsets of . Suppose that z1 = z2 = = zk = 0 and zk+1 = 0. By the maximum principle, {BN }N 1 is uniformly convergent on compact subsets of if and only if {BN (z)/z k }N 1 does the same. Hence, without loss of generality, we assume that |z1 | > 0. According to this assumption, we have
|B(0)| =
n=1
|zn | > 0.
On the other hand, the elementary inequality t implies

et1 ,
1,
|B(0)| =
n=1
|zn |
exp
n=1
(1 |zn |) ,
and thus
( 1 |zn | ) < .
n=1
Now, suppose that the last inequality holds. First of all, the identity |zn | zn z = 1 1 |zn | zn 1 z n z 1+ |zn | ( 1 + |zn | ) z zn ( 1 zn z ) (7.8)
160 implies 1 |zn | zn z zn 1 z n z
(1 |zn |)
1+
|z| (1 + 1/|z1 |) . | z 1/n | z
But, on a compact set K , we have 1+ |z| (1 + 1/|z1 |) | z 1/n | z CK ,
where CK is a constant independent of n. Therefore, for all z K, 1 |zn | zn z zn 1 z n z CK (1 |zn |), (n 1).
This inequality establishes the uniform convergence of BN on K. A sequence {zn } of complex numbers in the open unit disc satisfying
1 |zn | <
n=1
(7.9)
is called a Blaschke sequence. We already saw that the set of accumulation points of any Blaschke sequence is a subset of the unit circle T. Hence, under the conditions of Theorem 7.1, BN converges uniformly on each compact subset of D to the Blaschke product B(z) = |zn | zn z . z 1 zn z n=1 n
Since |BN (z)| < 1 on the open unit disc, we also have |B(z)| < 1, (z D). (7.10)
According to (7.10), a Blaschke product B, restricted to D, is in H (D). Hence, by Theorem 5.2, there is a unique function b L (T) such that B(rei ) = Moreover, B and, by Lemma 3.10,
H (D)
1 2
1+
r2
1 r2 b(eit ) dt, 2r cos( t) = b

L (T)
(rei D).
r1
lim B(rei ) = b(ei )
for almost all ei T. If B is a nite Blaschke product, then b is well-dened and analytic at all points of T. But, if B is an innite Blaschke product, then b, as a bounded
7.2. Blaschke products for the open unit disc
161
measurable function, is dened almost everywhere on T. Nevertheless, if an arc of T is free of the accumulation points of the zeros of B, then Theorem 7.1 ensures that b is a well-dened analytic function on this arc and |b| 1 there.
Exercises
Exercise 7.2.1 Let E be a closed nonempty subset of T. Construct a Blaschke sequence whose accumulation set is exactly E. In particular, construct a Blaschke sequence that accumulates at all points of T. Exercise 7.2.2 Let {zn }1
n N
D, let T, and let zn z . 1 zn z n=1

N
B(z) =
Show that, for any w D, the equation B(z) = w has exactly N solutions in D. Hint: Do it by induction. The function B bz1 might be useful too. Remark: Repetition is allowed among zn and also among the solutions of the preceding equation. Exercise 7.2.3 Construct an innite Blaschke product B with zeros on the interval [0, 1) such that lim sup |B(r)| = 1.
r1
Remark: Note that for any such product, lim inf r1 |B(r)| = 0. Exercise 7.2.4 Construct an innite Blaschke product B with zeros on the interval [0, 1) such that lim B(r) = 0.
r1
Hint: Take zn = 1 n Exercise 7.2.5 and let
, n 1.
Let B be a Blaschke product with zeros on the interval [0, 1), F (z) = (z 1)2 B(z),
(z D).
Show that F H (D).
162
7.3
Jensens formula
Let F H (D). Then, by Lemma 3.10, we know that f (ei ) = limr1 F (rei ) exists for almost all ei T. If |f (ei )| = 1, for almost all ei T, we say that F is an inner function for the open unit disc. A Blaschke product is an inner function. To establish this result we need Jensens formula. Jensens formula provides a connection between the moduli of the zeros of a function F inside a disc and the values of |F | on the boundaries of the disc. This result is indeed a generalization of the mean value property of harmonic functions (Corollary 3.2). Jensens formula is one of the most useful results of function theory. Theorem 7.2 (Jensens formula) Let F be analytic on a domain containing the closed disc Dr and let z1 , z2 , . . . , zn be the zeros of F inside Dr , repeated according to their multiplicities. Suppose that F (0) = 0. Then
n
log |F (0)| =
k=1
log
r |zk |
1 2
log |F (reit )| dt.
Proof. First, suppose that F has no zeros on the circle Dr . Let

n
B(z) =
k=1
r(zk z) , r 2 zk z
and let G = F/B. Then G is a well-dened analytic function with no zeros on the closed disc Dr . Hence, log |G| is harmonic there and, by Corollary 3.2, we have 1 log |G(0)| = log |G(reit )| dt. (7.11) 2 But, B is dened such that |B()| = 1 for all Dr . Therefore, |G(reit )| = |F (reit )| for all t. We also have G(0) = F (0)
k=1 n
r . zk
Plugging the last two identities into (7.11) gives Jensens formula whenever F has no zeros on the circle Dr . But both sides of the formula are continuous with respect to r, and the sequence {|zn |}n1 has no accumulation point inside the interval [0, 1). Hence, the equality holds for all r [0, 1).
7.3. Jensens formula
163
In the proof of Jensens formula, we used Corollary 3.2 to obtain (7.11). However, based on Lemma 3.4 and a simple change of variable, we obtain log |G(r0 ei0 )| = 1 2

r2
2 r0
2 r 2 r0 log |G(reit )| dt 2rr0 cos(0 t)
for all r0 ei0 Dr . This identity implies

n
log |F (z0 )| =
k=1
log

r 2 z k z0 r(z0 zk ) +
2 r0 2 r 2 r0 log |F (reit )| dt, 2r0 r cos(0 t)
1 2
r2
which is called the PoissonJensen formula. We state this formula slightly dierently in the following corollary. As a matter of fact, this result is a special case of the canonical factorization that will be discussed in Section 7.6. Corollary 7.3 Let F be analytic on a domain containing the closed disc Dr . Let z1 , . . . , zn denote the zeros of F in Dr repeated according to their multiplicities. Then, for each z0 = r0 ei0 Dr , we have
n
F (z0 ) =
k=1
r(z0 zk ) exp r 2 zk z0
1 2
reit + z0 log |F (reit )| dt , reit z0
where is a constant of modulus one. Now we are ready to show that each Blaschke product is an inner function. Note that this fact is trivial for nite Blaschke products. Theorem 7.4 (F. Riesz [18]) Let B be a Blaschke product for the open unit disc. Let b(ei ) = lim B(rei )
r1
wherever the limit exists. Then |b(ei )| = 1 for almost all ei T. Proof. First of all, by (7.10), we have 1 2

log B(rei ) d 0
(7.12)
for all r, 0 r < 1. Without loss of generality, we assume that B(0) = 0. Since otherwise, we can divide B by the factor z m and this modication does not change |b|. Then, by Jensens formula, log |B(0)| =
|zn |<r
log
r |zn |
1 2
log |B(rei )| d
164

n=1
for all r, 0 < r < 1. Since B(0) = 1 2

|zn |, we thus have log r |zn |
log B(rei ) d =
|zn |<r
n=1
log
1 . |zn |
Given > 0, choose N so large that
n=N +1
log 1/|zn | < , and let
R = max{ |z1 |, . . . , |zN | }. Therefore, for all r > R, we have 1 2

N
log B(rei ) d
n=1
log
r |zn |
n=1
log
1 |zn |
which immediately implies lim inf

r1
1 2
log B(rei ) d
Since is an arbitrary positive number and in the light of (7.12), the limit of integral means of |B| exists and 1 r1 2 lim By (7.10), we have | b(ei ) | Fatous lemma gives 1 2

log B(rei ) d = 0.
1 for almost all ei T. Now, an application of 1 r1 2

log |b(ei )| d lim

i
log B(rei ) d = 0.
Therefore, we necessarily have |b(e )| = 1 for almost all ei T. In proving Rieszs theorem, we established the following result which by itself is interesting and will be needed later on in certain applications. Corollary 7.5 Let B be a Blaschke product for the open unit disc. Then
r1
lim
log B(rei ) d = 0.
Exercises
Exercise 7.3.1 Show that (1 |z|2 ) (1 |z0 |2 ) , |1 z0 z|2 (z D).
|bz0 (z)|2 = 1
7.3. Jensens formula Exercise 7.3.2 Let z0 D, and let 0 r < 1. Show that
|z|=r
165
max
|z0 | + r z0 z . 1 z0 z 1 + r|z0 |
Exercise 7.3.3
Let z0 D, z0 = 0. Show that arg bz0 (z) = arcsin (z0 z ) (1 |z0 |2 ) . |z0 | |z0 z| |1 z0 z|
Exercise 7.3.4 are equivalent: (i) (ii) (iii)
Let zn D \ {0}, n 1. Show that the following conditions < ; ;
n=1 1 |zn | n=1 log |zn | > n=1 |zn | > 0.
Exercise 7.3.5 Let F be analytic on the open unit disc D and continuous on the closed unit disc D. Suppose that |F (ei )| = 1 for all ei T. Show that F is a nite Blaschke product. Remark: This exercise shows that a nite Blaschke product is the solution of an extremal problem. Exercise 7.3.6 Let F be meromorphic in the open unit disc D and continuous on the closed unit disc D. Suppose that |F (ei )| = 1 for all ei T. Show that F is the quotient of two nite Blaschke products. Exercise 7.3.7 Show that

log |1 ei | d = 0.
Remark: This fact was used implicitly in the proof of Jensens formula. Exercise 7.3.8 [Generalized PoissonJensen formula] Let F be a meromorphic function on a domain containing the closed disc Dr . Let z1 , . . . , zn and p1 , . . . , pm denote respectively the zeros and poles of F in Dr repeated according to their multiplicities. Show that, for each z0 = r0 ei0 Dr , we have
n
log |F (z0 )| =
k=1
log

r 2 z k z0 + r(z0 zk )
log
k=1
r 2 p k z0 r(z0 pk )
1 2
2 r 2 r0 log |F (reit )| dt. 2 + r 2 2r r cos( t) r 0 0 0
166
7.4
Rieszs decomposition theorem
Let F be analytic, with no zeros on the open unit disc. Let p be any positive number. Then we can properly dene the analytic function F p on D. This simple-looking property, which by the way is not fullled by harmonic functions, has profound applications in the theory of Hardy spaces. An element of H p (D) might have several zeros in the unit disc. F. Riesz showed that any such element can be written as the product of a Blaschke product and a nonvanishing element of H p (D). To prove this useful technique, we rst verify that the zeros of functions in any Hardy space form a Blaschke sequence. Lemma 7.6 Let F H p (D), 0 < p , F 0, and let {zn } denote the sequence of its zeros in D. Then (1 |zn |) < .
n
Proof. If F has a nite number of zeros, then the result is obvious. If it has innitely many zeros, since F 0, they necessarily converge toward some points in the unit circle. In other words, limr1 |zn | = 1. Without loss of generality assume that F (0) = 0. By Jensens formula, for each 0 < r < 1, we have log |F (0)| =
|zn |<r
log
1 r + zn 2
log |F (rei )| d.
If 0 < p < , then, by Jensens inequality, 1 2 and thus

log |F (rei )|p d log 1 2

1 2
|F (rei )|p d
log F
p p
log |F (rei )| d log F
p.
The last inequality clearly holds even if p = . Hence, log

|zn |<r
r log F zn
log |F (0)|.
Fix N and take R so large that the rst N zeros are in the disc DR . Therefore, if R < r < 1, we have
N
log
n=1
r log F zn
log |F (0)|.
First let r 1 to obtain

N
log 1/|zn | log F

n=1
log |F (0)|,
7.4. Rieszs decomposition theorem and then let N to get
167
log 1/|zn | log F

n=1
log |F (0)| < . |zn |) < .
But, the last inequality is equivalent to
n (1
Lemma 7.6 shows that we can form a Blaschke product with zeros of any nonzero element of a Hardy space. F. Riesz discovered that we can extract these zeros such that the remaining factor is still in the same space with the same norm. Theorem 7.7 (F. Riesz [18]) Let F H p (D), 0 < p , F 0, and let B be the Blaschke product formed with the zeros of F in D. Let G = F/B. Then G H p (D), G is free of zeros in D, and G
p
= F
p.
Proof. Clearly G is analytic and free of zeros on D. Moreover, since |B| 1, we necessarily have F p G p . Suppose that 0 < p < . Fix N 1 and 0 < r < 1. Let BN be the nite Blaschke product formed with the rst N zeros of F . Then GN = F/BN is an analytic function on D, and, by Corollary 4.14, 1 2

|GN (rei )|p d
1 2 1 2
|GN (ei )|p d
|FN (ei )|p d i p |BN (e )| 1 1 |FN (ei )|p d inf |BN (ei )| 2 F p p inf |BN (ei )|
for all , r < < 1. Let 1. Since |BN | uniformly tends to 1, then 1 2

|GN (rei )|p d
1 p
p.
Now, let N . On the circle {|z| = r}, BN tends uniformly to B. Hence, we obtain Gr p F p . Finally, take the supremum with respect to r to get G
p
p.
168
A small modication of this argument yields a proof for the case p = . As a matter of fact, the proof is even simpler in this case. First, by the maximum principle, F |GN (rei )| sup |GN (ei )| , inf |BN (ei )| and thus, by letting 1, we obtain |GN (rei )| F Then, if N , we get |G(rei )| F
.
. .
This inequality is equivalent to G
In the succeeding sections, we provide several applications of Rieszs decomposition theorem.
Exercises
Exercise 7.4.1 (G. Julia [11]) Let F H p (D) and F (0) = 0. Show that Fr for all 0 r < 1. Hint: Apply Theorem 7.7. Exercise 7.4.2 Let F , F 0, be analytic on the open unit disc and let {zn } denote the sequence of its zeros in D. Show that (1 |zn |) <
n p
r F
if and only if
0r<1
sup
log |F (rei )| d < .
7.5
Representation of H p (D) functions (0 < p < 1)
In this section, we show that certain results discussed in Sections 5.1 and 5.6 about H p (D) functions with 1 p are valid even if 0 < p < 1.
7.5. Representation of H p (D) functions (0 < p < 1) Theorem 7.8 Let F H p (D), 0 < p < 1. Then f (ei ) = lim F (rei )
r1
169
exists almost everywhere on T. Moreover f Lp (T),

r1
lim Fr f F = f
=0
and
p p.
Proof. By Theorem 7.7, F = BG, where G is free of zeros on D, G H p (D) and F p = G p . Pick a positive integer n such that np 1 and let K = G1/n . Then K H np (D) H 1 (D) and F = B K n. By Lemma 3.10, b(ei ) = limr1 B(rei ) and k(ei ) = limr1 K(rei ) exist almost everywhere on T. Since n is an integer, f (ei ) = limr1 F (rei ) also exist almost everywhere on T and f = b kn with b L (T) and k Lnp (T). Hence, in the rst place, by Theorem 7.4, |f |p = |k|np and thus f Lp (T). Secondly, to show that limr1 Fr f p = 0 we use the known fact that limr1 Kr k np = 0. Indeed, using the binomial theorem, we have |Fr f |p
n = |Br Kr b k n |p
Br [(Kr k) + k]n Br k n + (Br b) k n

n j=1
n |Kr k|jp |k|(nj)p + |Br b|p |k|np j
and thus, by Hlders inequality, o

n
Fr f
p p
j=1
n j
Kr k
j/n np
(nj)/n + np
1 2
|Br (ei )b(ei )|p |k(ei )|np d.
By Theorem 5.1, Kr k np 0 and, since |Br b|p |k|np 2 |k|np , the dominated convergence theorem ensures that
r1
lim
p
|Br (ei ) b(ei )|p |k(ei )|np d = 0.
Therefore, Fr f
0 as r 1. Finally, by Corollary 4.14, F

p
= lim Fr
r1
= f
p.
170
At this point we are able to complete the proof of Kolmogorovs theorem (Corollary 6.2). In the proof of the corollary, we showed that U, V hp (D) and thus F = U + iV H p (D) with boundary values f = u + i. By Theorem 7.8, u as r 1, Fr f in Lp (T). Hence, Vr u in Lp (T), which also implies V p = u p. The correspondence established between H p (D) and H p (T), 1 p , in Theorems 5.1, 5.2 and 5.11, along with Fatous theorem on radial limits (Lemma 3.10), enables us to give another characterization of H p (T), 1 p , classes. Functions in H p (T), 1 p , can be viewed as the boundary values of elements of H p (D). This interpretation is consistent with the original denition given in (5.2). However, the latter has one advantage. Based on Theorem 7.8, this point of view can equally be exploited as the denition of H p (T), 0 < p < 1, classes. The following result gives more information about the correspondence between H p (D) and H p (T). Theorem 7.9 Let F H p (D), 0 < p , F 0, and let f (eit ) = lim F (reit )
r1
wherever the limit exists. Then

r1
log |f (eit )| dt < ,
lim
log+ |F (reit )| log+ |f (eit )| dt = 0 1 r2 log |f (eit )| dt, 2r cos( t)
and log |F (rei )| 1 2 1+ r2 (rei D).
Proof. Without loss of generality, assume that p < and that F (0) = 0. Since log+ x x, x > 0, we have p log+ |F (rei )| |F (rei )|p , and thus

(rei D),
log+ |F (rei )| d
|F (rei )|p d 2 F
p p.
Hence, by Fatous lemma,

log+ |f (ei )| d 2 F
p p /p
< .
On the other hand, by Jensens formula (Theorem 7.2), log |F (0)| = 1 2

|zn |r
log
r |zk |
1 2
log |F (rei )| d
log |F (rei )| d
7.5. Representation of H p (D) functions (0 < p < 1) and thus 1 2

171
log |F (rei )| d
log |F (0)| + F p
p p
1 2
log+ |F (rei )| d
log |F (0)|.
Another application of Fatous lemma implies 1 2

log |f (ei )| d
F p
p p
log |F (0)| < .
Since log |f | = log+ |f | + log |f |, the rst assertion is proved. It is easy to verify that log x (x 1)p /p, for x 1 and p > 0. Hence, we immediately obtain | log+ a log+ b| |a b|p /p, (a, b > 0).
The second assertion is now a direct consequence of this inequality and Theorem 7.8. Let < 1 and consider G(z) = F (z). The function log |G| is subharmonic on a domain containing the closed unit disc. Hence, by Theorem 4.6, log |G(rei )| Thus, for rei D, log |F (rei )| 1 2

1 2
1 r2 log |G(eit )| dt. 1 + r2 2r cos( t) 1 r2 log+ |F (eit )| dt 1 + r2 2r cos( t)
1 r2 1 log |F (eit )| dt 2 1 + r2 2r cos( t) = I+ () I ().
Let 1. By the second assertion of the theorem, which was proved earlier, we have 1 r2 log+ |f (eit )| dt I+ () 2 2r cos( t) 1 + r and, by Fatous lemma,

1 r2 log |f (eit )| dt lim inf I (). 1 1 + r2 2r cos( t)
Taking the dierence of the last two relations gives the required inequality. Corollary 7.10 Let F H p (D), 0 < p < , and let f (eit ) = limr1 F (reit ) wherever the limit exists. Then |F (rei )|p 1 2

1 r2 |f (eit )|p dt, 1 + r2 2r cos( t)
(rei D).
172 Proof. By Theorem 7.9, log |F (rei )| Hence |F (rei )|p exp
1 2
1+
r2
1 r2 log |f (eit )| dt. 2r cos( t)
1 2
1 r2 log |f (eit )|p dt . 1 + r2 2r cos( t)
Now, apply Jensens inequality. In the second assertion of Theorem 7.9 we cannot replace log+ by log. For example, consider F (z) = exp 1+z 1z , (z D).
Then F H (D) with boundary values f (ei ) = ei cot(/2) , Hence log |F (rei )| log |f (ei )| = and thus, for each 0 r < 1,

(ei T). 1 r2 1 + r2 2r cos
log |F (rei )| log |f (ei )| d = 2.
7.6
The canonical factorization in H p (D) (0 < p )
Let F H p (D), 0 < p , F 0, and let f denote the boundary values of F on T. By Theorem 7.9, log |f | L1 (T). Hence, we are enable to dene OF (z) = exp 1 2

eit + z log |f (eit )| dt , eit z
(z D),
(7.13)
which is called the outer part of F . As a matter of fact, for any positive function h with log h L1 (T), the outer function Oh (z) = exp 1 2

eit + z log |h(eit )| dt , eit z
(z D),
(7.14)
as an analytic function on the open unit disc, is well-dened. However, OF has more interesting properties. First of all, |OF (rei )| = exp 1 2

1+
r2
1 r2 log |f (eit )| dt 2r cos( t)
7.6. The canonical factorization in H p (D) (0 < p ) and thus, by Fatous theorem (Lemma 3.10),
r1
173
lim |OF (rei )| = |f (ei )|
(7.15)
for almost all ei T. Secondly, assuming p < , |OF (rei )|p = exp 1 2

1+
r2
1 r2 log |f (eit )|p dt 2r cos( t)
and thus, by Jensens inequality, |OF (rei )|p 1 2

1 r2 |f (eit )|p dt. 1 + r2 2r cos( t)

Hence, by Fubinis theorem, 1 2

|OF (rei )|p d
1 2
|f (eit )|p dt = F
p p.
In other words, OF H p (D) and OF p F p . This inequality clearly holds even if p = . On the other hand, the inequality in Theorem 7.9 can be rewritten as |F (z)| |OF (z)|, (z D), (7.16) and thus we necessarily have OF The inner part of F is dened by IF (z) = F (z) , OF (z) (z D).
p
= F
p.
(7.17)
By (7.16), |IF (z)| 1, z D, and by (7.15), |IF (ei )| = 1 for almost all ei T. Hence IF is indeed an inner function. The factorization F = IF OF is called the canonical, or innerouter, factorization of F . We can say more about the inner factor IF . According to Rieszs theorem (Theorem 7.7), we can write IF = B S, where B is a Blaschke factor formed with the zeros of F and S is a zero-free inner function on D. Therefore, U = log |S| is a positive harmonic function on D satisfying lim U (rei ) = 0
r1
for almost all e T. Therefore, by Theorem 3.12, there is a positive measure M(T), singular with respect to the Lebesgue measure, such that log |S(rei )| = 1 2
T
1+
r2
(rei D).
174 Hence S(z) = S (z) = exp
1 2
eit + z d(eit ) , eit z
(z D).
(7.18)
As a matter of fact, the counterexample given after Theorem 7.9 is a prototype of these singular inner functions obtained by a Dirac measure at point 1. Note that for singular inner functions, we have

log |S (rei )| d =
d(eit ) = (T)
(7.19)
for each 0 < r 1. Compare with Corollary 7.5. The complete canonical factorization of F now becomes F = B S O, where B is the Blaschke product formed with the zeros of F , the singular inner function S is given by (7.18), and the outer function O by (7.13). There is also a constant of modulus one which was absorbed in B. Since is a positive Borel measure and singular with respect to the Lebesgue measure, we have (ei ) = 0 for almost all ei T, which in turn implies
r1
lim |S(rei )| = 1
for almost all ei T. What is implicit in this statement is that almost all is with respect to the Lebesgue measure. On the other hand, we also know that (ei ) = + for almost all ei T, which implies
r1
lim S(rei ) = 0
(7.20)
for almost all ei T. However, in the latter statement, almost all is with respect to the measure . In particular, if is nonnull, at least at one point on T, the radial limit of S is zero.
Exercises
Exercise 7.6.1 if and only if Let I be an inner function. Show that I is a Blaschke product
r1
lim
log |I(rei )| d = 0.
Hint: We have I = B S, where B is a Blaschke product and S is a singular inner function. Apply Corollary 7.5 and (7.19). Remark: Compare with Corollary 7.5. Exercise 7.6.2 Let I be an inner function. Suppose that for each ei T, either limr1 |I(rei )| does not exist or it exists but limr1 |I(rei )| > 0. Show that I is a Blaschke product. Hint: Use (7.20).
7.7. The Nevanlinna class
175
7.7
The Nevanlinna class
Let h be a positive measurable function with log h L1 (T), and let be a signed Borel measure on T. We emphasize that h is not necessarily in any Lebesgue space Lp (D) and is not necessarily positive. Let B be any Blaschke product for the unit disc. Let F = B S Oh , (7.21) where Oh and S are dened by (7.14) and (7.18). Write = + , where + and are positive Borel measures on T. Then log |F (rei )| + 1 2 1 2 1 r2 d (eit ) 2 T 1 + r 2r cos( t) 1 r2 log+ h(eit ) dt. 2 1 + r 2r cos( t)

Therefore, by Fubinis theorem,

log+ |F (rei )| d (T) +
log+ h(eit ) dt
for all 0 r < 1. The integral on the left side is an increasing function of r (Corollary 4.12). Hence, functions dened by (7.21) satisfy
0r<1
sup
log+ |F (rei )| d < .
(7.22)
The Nevanlinna class N is, by denition, the family of all analytic functions F on the open unit disc which satisfy (7.22). The following result shows that any such function also has a representation of the form (7.21). Note that, according to Theorem 7.9, H p (D) N .
p>0
Theorem 7.11 Let F N , F 0. Then (a) the zeros of F satisfy the Blaschke condition; (b) for almost all ei T, exists and is nite; (c) log |f | L1 (T); (d) there is a signed measure such that F = B S O|f | , where B is the Blaschke product formed with the zeros of F . f (ei ) = lim F (rei )
r1
176
Proof. Fix < 1. By Corollary 7.3, for each z D , we have

n
F (z) =
k=1
(z zk ) exp 2 z k z
1 2
eit + z log |F (eit )| dt , eit z
where is a constant of modulus one and z1 , . . . , zn denote the zeros of F in D repeated according to their multiplicities. Hence, for each z D,
n
F (z) =
k=1
z zk / exp 1 z zk /
1 2
eit + z log |F (eit )| dt . eit z
Dene
n
(z) (z)
=
k=1
z zk / exp 1 z zk / 1 2
it
1 2
eit + z log |F (eit )| dt , eit z
exp
e +z log+ |F (eit )| dt . eit z
Then and are analytic functions in the closed unit ball of H (D), i.e. | (z)| 1 and | (z)| 1 for all z D, and F (z) = (z) , (z)

(z D).
Based on our main assumption on F , there is a constant C > 0 such that (0) = exp 1 2 log+ |F (eit )| dt C (7.23)
for all < 1. Now take a sequence converging to 1, say n = 1 1/n, n 1. Hence, by Montels theorem, there is a subsequence (nk )k1 and functions and in the unit ball of H (D) such that, as k , nk (z) (z) and nk (z) (z)
uniformly on compact subsets of D. Therefore, F (z) = (z) , (z) (z D).
Moreover, by (7.23) and the fact that (z) = 0 for all z D, we deduce that (z) = 0 for all z D. This representation allows us to derive all parts of the theorem. The zeros of F are the same as the zeros of H (D). Hence (a) follows from Lemma 7.6. By Fatous theorem (Lemma 3.10), (ei ) (ei ) = =
r1 r1
lim (rei ),
lim (rei )
7.7. The Nevanlinna class
177
exist and are nite for almost all ei T, and by the uniqueness theorem (Theorem 5.13) the set { ei T : (ei ) = 0 } is of Lebesgue measure zero. Hence, part (b) follows. Since f = /, part (c) follows from Theorem 7.9. Finally, as discussed in Section 7.6, there are positive Borel measures 1 and 2 such that = B S1 O|| and = S2 O|| , where B is the Blaschke product formed with the zeros of . Hence, we obtain F = B S1 2 O||/|| = B S O|f | . In the proof of the preceding theorem it was shown that each element F N is the quotient of two bounded analytic functions. On the other hand, suppose that F = /, where and are bounded analytic functions on the open unit disc. Without loss of generality, we may assume that 1 and 1, and that (0) = 0. Hence, log+ |F (rei )| log |(rei )| and thus, by Jensens formula (Theorem 7.2),

log+ |F (rei )| d
log |(rei )| d 2 log |(0)|
for all r < 1. Therefore, F N . We thus obtain another characterization of the Nevanlinna class. Corollary 7.12 Let F be analytic on the open unit disc. Then F N if and only if F is the quotient of two bounded analytic functions. The measure appearing in the canonical decomposition of functions in the Nevanlinna class is not necessarily positive. Hence, we dene a subclass as follows: N + = { F N : F = B S O|f | with 0 }. Based on the canonical factorization theorem, we have H p (D) N + .
p>0
The main advantage of N + over N is the following result. Theorem 7.13 (Smirnov) Let F N + , and let f (ei ) = limr1 F (rei ), wherever the limit exists. Then F H p (D), 0 < p , if and only if f Lp (T). In particular, if F H s (D) for some s (0, ] and f Lp (T), 0 < p , then F H p (D). Proof. Since F N + we have the canonical factorization F = B S O|f | , where B is the Blaschke product formed with the zeros of F and is a positive singular measure. Therefore, I = IF = B S is an inner function. Hence, it is enough to show that O|f | H p (D). But, as discussed at the beginning of Section 7.6, O|f | H p (D) if and only if f Lp (T).
178 Let
F (z) = exp
1+z 1z
(z D).
Clearly F N with unimodular boundary values f (ei ) = ei cot(/2) , (ei T).
However, F H (D). Hence, in Theorem 7.13 we cannot replace N + by N . The following result provides a characterization of the elements of N + in the Nevanlinna class N . Theorem 7.14 Let F N , and let f (ei ) = limr1 F (rei ), wherever the limit exists. Then F N + if and only if
r1
lim
log+ |F (rei )| d =
log+ |f (ei )| d.
(7.24)
Proof. Suppose that F N + . Hence F = B S O|f | , where is a positive singular measure. Therefore, |F | |O|f | |, which implies log+ |F (rei )| 1 2

1+
r2
1 r2 log+ |f |(eit ) dt, 2r cos( t)

(rei D).
Thus, by Fubinis theorem,

log+ |F (rei )| d
log+ |f |(eit ) dt,
(0 r < 1).
The quantity on the right side is an increasing function of r (Corollary 4.12), and by Fatous lemma,

log+ |f (ei )| d lim inf

r1
log+ |F (rei )| d.
Hence, (7.24) holds. Now, suppose that (7.24) holds. Let F = B S O|f | be the canonical decomposition of F N . Write G = S O|f | and denote its boundary values by g. Since F = BG, it is enough to prove that G N + . We have |f | = |g| and log |B(z)| + log+ |G(z)| log+ |F (z)| log+ |G(z)|. Thus, by Corollary 7.5 and our assumption (7.24),
r1
lim
log+ |G(rei )| d =
log+ |g(ei )| d.
(7.25)
The main reason we replaced F by G is that it has the representation log |G(rei )| = 1 2 1+ r2 1 r2 d(eit ), 2r cos( t) (rei D),
7.7. The Nevanlinna class where
179
d(eit ) = log |g(eit )| dt d(eit ).
Hence, by Corollary 2.11 the measures log |G(reit )| dt converge in the weak* topology to d(eit ). Take any sequence rn > 0, n 1, such that rn 1. Since the sequences (log+ |G(rn eit )| dt)n1 and (log |G(rn eit )| dt)n1 are uniformly bounded in M(T), there is a subsequence (nk )k1 and two positive measures 1 , 2 M(T) such that log+ |G(rnk eit )| dt d1 (eit ) and log |G(rnk eit )| dt d2 (eit )
in the weak* topology. Hence, = 1 2 . Our main task is to show that 1 is absolutely continuous with respect to the Lebesgue measure. This fact is equivalent to saying that 0, and thus the theorem would be proved. Let E be any Borel subset of T. Then, by Fatous lemma,
E
log+ |g(eit )| dt lim inf

r1
log+ |G(reit )| dt
and
T\E
log+ |g(eit )| dt lim inf

r1
T\E
log+ |G(reit )| dt.
If, in any one of the last two inequalities, the strict inequality holds, then we add them up and obtain a strict inequality which contradicts (7.25). Hence lim inf
r1 E
log+ |G(reit )| dt =
log+ |g(eit )| dt
for all Borel subsets E of T. In particular, if we take r = rnk , k , then, by (7.25) and Theorem A.4, the measures log+ |G(rnk eit )| dt converge to log+ |g(eit )| dt in the weak* topology. Therefore, d1 (eit ) = log+ |g(eit )| dt.
Exercises
Exercise 7.7.1 Show that H p (D)
p>0
N+
N.
Exercise 7.7.2 Let F N , and let f (ei ) = limr1 F (rei ), wherever the limit exists. Show that the assumption f Lp (T), 0 < p , is not enough to conclude that F H p (D).
180
Exercise 7.7.3 Let F N . Show that there is a Blaschke product B and singular inner functions S1 , S2 and an outer function O such that F = B S1 O/S2 .
Exercise 7.7.4 [Generalized Smirnov theorem] Let Og and Oh be outer functions with g/h Lp (T), 0 < p . Show that Og H p (D). Oh Exercise 7.7.5 Let u, u L1 (T). Let V = Q u. Show that V h1 (D),
r1
lim Vr u
=0
and V (rei ) = 1 2

1 r2 u(eit ) dt, 1 + r2 2r cos( t)
(rei D).
Hint: Let U = P u, and let F = U + iV . By Kolmogorovs theorem (Corollary 6.2), F H 1/2 (D). Since the boundary values of F are given by f = u + i u L1 (T), the Smirnov theorem (Theorem 7.13) ensures that F H 1 (D). Exercise 7.7.6 Let u, u L1 (T). Show that u + i H 1 (T). u Hint: Use Exercise 7.7.5. Exercise 7.7.7 Let u, u L1 (T). Suppose that

u( eit ) dt = 0.
Show that u = u. Hint: Use Exercise 7.7.6. Exercise 7.7.8 Let u, u L1 (T). Show that u(n) = i sgn(n) u(n), Hint: Use Exercise 7.7.5. (n Z).
7.8. The Hardy and FejrRiesz inequalities e
181
7.8
The Hardy and FejrRiesz inequalities e
Another consequence of Theorem 7.7 is the following representation for the elements of H 1 (D). This result by itself is interesting and will lead us to Hardys inequality. Lemma 7.15 Let F H 1 (D). Then there are G, K H 2 (D) such that F = GK and F
1
= G
2 2
= K
2 2.
Proof. By Theorem 7.7, F = B , where B is a Blaschke product and H 1 (D) is free of zeros on D with 1 = F 1 . Let G = B 2
1
and K = 2 .
Clearly G K = F and K 2 = 1 = F 1 . Moreover, again by Theorem 7.7, 2 1 1 G 2 = B 2 2 = 2 2 = 1 = F 1. 2 2 2 An element F H 1 (D) corresponds to a unique f H 1 (T) and we have the power series representation
F (z) =
n=0
f (n) z n ,
(z D).
Since f H 1 (T) L1 (T), by the RiemannLebesgue lemma (Corollary 2.17), we know that f (n) 0 as n . Hardys inequality improves this result. Theorem 7.16 (Hardys inequality) Let F H 1 (D) and let f H 1 (T) denote its boundary values. Then |f (n)| F n n=1
1.
Proof. By Lemma 7.15, there are G, K H 2 (D) such that F = G K and F G 2 = K 2 . Moreover, if 2 2
G(z) =
n=0
g (n) z n ,
(z D),
and K(z) =
n=0
k(n) z n ,
(z D),
then we clearly have f (n) =

j+ =n j, 0
g (j)k( ).
182
By Parsevals identity (Corollary 2.22),

n=0
|(n)|2 = G g
2 2
= F
and
n=0
|k(n)|2 = K
2 2
= F
1.
Hence G(z) =
|(n)| z n , g
(z D),
n=0
and K(z) =
n=0
|k(n)| z n ,
2 2
(z D),
1.
are also in H 2 (D), with G
2 2
= K
= F
Therefore, if we dene (z D),
F(z) = G(z) K(z) =

n=0
An z n ,
then, by Hlders inequality, F H 1 (D) with o F and |f (n)|

j+ =n j, 0 1
= F
g (j)k( )
j+ =n j, 0
|(j)| |k( )| = An g
for all n 0. Hence, it is enough to prove that An F 1. n n=1 The main advantage of F over F is that all its coecients are positive. Let zn (z) = = log(1 z), (z D), n n=1 where log is the main branch of logarithm. Hence, Then, by Parsevals identity, we have 1 2

h (D) with
. 2
F(re
) (rei ) d
An 2n r n n=1
7.8. The Hardy and FejrRiesz inequalities e and

183
1 2 i

F(rei ) (rei ) d = 0.
Hence,
F(rei )
(rei ) d =
An 2n r . n n=1
Therefore, for each 0 r < 1, An 2n r 2 F n n=1 Now, let r 1. Let F H (D). Abusing the notation, let F also denote the restriction of F on the interval I = (1, 1). Then F is clearly bounded on I and F L (I) F H (D) . The FejrRiesz inequality is a generalization of this e simple observation for other H p (D) classes. Theorem 7.17 (FejrRiesz) Let F H p (D), 0 < p < . Then e
1 1 1
F 1.
|F (x)| dx
1 p
1/p F
p.
Proof. Case 1: F H 2 (D) and the Taylor coecients of F are real. In this case, F (x) is real for all x (1, 1). Fix 0 < r < 1. Let r be the curve formed with the interval [r, r] and the semicircle rei , 0 . Then F (z) dz = 0,
r
which implies
r r
F 2 (x) dx = ir
r r
0 0
F 2 (rei ) ei d,
and thus
F 2 (x) dx
|F (rei )|2 d.
(7.26)
A similar argument shows that

r r
F 2 (x) dx
|F (rei )|2 d.
(7.27)
To obtain this inequality we should start with the curve r formed with the interval [r, r] and the semicircle rei , 0. Adding (7.26) to (7.27) gives
r
2
r
F 2 (x) dx
|F (rei )|2 d 2 F
2 2.
184 Let r 1 to obtain
1 1
F 2 (x) dx
1 2
2 F
2.
Case 2: F H 2 (D). Write
F (z) =
n=0
(an + ibn ) z n =
n=0
an z n + i
n=0
bn z n = F1 (z) + iF2 (z),
where an and bn are real numbers. Then F1 and F2 are real on the interval (1, 1) and F1
2 2
+ F2
2 2
=
n=0
|an |2 +
n=0
|bn |2 =
n=0
|an + ibn |2 = F
2 2.
Hence, by Case 1,
1 1
|F (x)|2 dx
=
1
2 F1 (x) dx + 2 2
F1 Case 3: F H p (D).
+ F2
1 2 2 =
2 F2 (x) dx
2 2.
By Theorem 7.7, F = BG, where G H p (D), G is zero-free and F Let K = Gp/2 . Hence, K H 2 (D) and K Therefore, by Case 2,
1 1 2 2
= G p.
= G
p p
= F
p p.
|F (x)|p dx
1 1 1 1
|G(x)|p dx |K(x)|2 dx
2 2
= F
p p.
Exercises
7.8. The Hardy and FejrRiesz inequalities e
185
Exercise 7.8.1 Let F H 1 (D) and let f H 1 (T) denote its boundary values. Show that |f (n)| F 1. n+1 n=0 Exercise 7.8.2 Let G, K H 2 (D) and let F = G K. Show that F H 1 (D) and F
1
2.
Exercise 7.8.3
Let F H 1 (D). Show that there are G, K H 1 (D) such that F = G + K,
G and K are free of zeros on D, and G

1
1,
1.
Hint: If B is a Blaschke product, B 1, then 1 + B and 1 B are zero-free on D. Exercise 7.8.4 Let
F (z) =
n=2
zn , log n
(z D).
Show that F H 1 (D). Exercise 7.8.5 Give an example to show that the constant 1/p in Theorem 7.17 is sharp, i.e. it cannot be replaced by a smaller one. Hint: Consider F (z) = ( (z) )1/p , where is the conformal mapping from D to the rectangle (1, 1) (, ), and maps (1, 1) to itself.
186
Chapter 8
Interpolating linear operators

8.1 Operators on Lebesgue spaces
Let (X, M, ) be a measure space with 0, and let Lp (X), 0 0
M : { t : |f (t)| > M } = 0
< .
The family of Lebesgue spaces Lp (T), 0 < p , is an important example that we have already seen several times before. Let (X, M, ) and (Y, N, ) be two measure spaces. Let p, q (0, ]. If an operator : Lp (X) Lq (Y ) is bounded, i.e. with a suitable constant C, f
q
C f
(8.1)
for all f Lp (X), we say that is of type (p, q) and we denote its norm by
(p,q)
f Lp (X) f =0
sup
f q . f p
In other words,
(p,q)
is the smallest possible constant C in (8.1). 187
188
Chapter 8. Interpolating linear operators Suppose that the operator is dened on the set Lp0 (X) + Lp1 (X) = { f + g : f Lp0 (X) and g Lp1 (X)}
and is of types (p0 , q0 ) and (p1 , q1 ). More precisely, when we restrict to Lp0 (X) and Lp1 (X), both operators : Lp0 (X) Lq0 (Y ) and : Lp1 (X) Lq1 (Y )
are well-dened and bounded. Then if p is between p0 and p1 , we are faced with a natural question. Can we nd q between q0 and q1 such that : Lp (X) Lq (Y ) is also well-dened and bounded? The RieszThorin interpolation theorem (Theorem 8.2) provides an armative answer. Let us mention two relevant examples that have already been discussed. The Fourier transform F : L1 (T) f
(Z) f
was dened in Section 1.3. By Lemma 1.1, F is of type (1, ). Moreover, by Parsevals identity (Corollary 2.22), the operator F : L2 (T) f
2
(Z) f
is well-dened and has type (2, 2). As a matter of fact, it is not dicult to show that F (1,) = F (2,2) = 1. Therefore, if p [1, 2] is given, is there a proper q [2, ] such that F : Lp (T) f
q
(Z) f
is well-dened and bounded? Our second example is the convolution operator. Fix 1 r and x f Lr (T). Let : L1 (T) Lr (T) g f g. By Corollary 1.5, is of type (1, r). On the other hand, the same corollary implies : Lr (T) L (T) g f g
8.2. Hadamards three-line theorem
189
is of type (r , ), where 1/r +1/r = 1. Thanks to Youngs inequality (Theorem 1.4), we already know that if s [1, r ], and we pick p= then sr [r, ], r s
: Ls (T) Lp (T) g f g
is of type (s, p). However, using the RieszThorin interpolation theorem (Theorem 8.2), we will give a second proof of this fact.
Exercises
Exercise 8.1.1 Show that Let (X, M, ) be a measure space. Let 0 < p0 p p1 . Lp (X) Lp0 (X) + Lp1 (X).
Exercise 8.1.2 Let (X, M, ) be a measure space. When is the family Lp (X), 0 < p , monotone?
8.2
Hadamards three-line theorem
In Section 4.2 we discussed several versions of the maximum principle. In particular, we saw that if F is analytic on a proper unbounded domain , continuous on and |F ()| 1 at all boundary points , we still cannot deduce that F is bounded on . However, as Phragmen-Lindelf observed, under some extra o conditions, we can show that the maximum principle still holds even for unbounded domains. There are many types of Phragmen-Lindelf theorems with o dierent conditions for dierent domains. To prove the RieszThorin interpolation theorem, we need a very special case due to Hadamard. Theorem 8.1 (Hadamards three-line theorem) Let S be the strip S = { x + iy : 0 < x < 1, y R }. Suppose that F is continuous and bounded on S = [0, 1] R, and that F is analytic on S. Suppose that |F (iy)| m0 and |F (1 + iy)| m1 for all y R. Then |F (x + iy)| m1x mx 1 0
for all 0 < x < 1 and all y R.
190
Chapter 8. Interpolating linear operators
Proof. Without loss of generality assume that m0 , m1 > 0. Since otherwise, we can replace them by m0 + and m1 + , and after proving the inequality with these new constants, we let 0. Let F (z) G(z) = 1z z . m 0 m1 The function G is continuous and bounded on S = [0, 1] R, and analytic on S. Moreover, |G(iy)| 1 and |G(1 + iy)| 1 for all y R. If we succeed in proving that |G(z)| 1 inside the strip, then the required inequality follows immediately. However, this fact is a simple consequence of Corollary 4.5 applied to log |G|. Let Fx
= sup |F (x + iy)|.
yR
Then Hadamards three-line theorem can be rewritten as log Fx In other words, log Fx

(1 x) log F0
+ x log F1
is a convex function of x.
Exercises
Exercise 8.2.1 [Hadamards three-circle theorem] Let be the annulus = { z : 0 < r1 < |z| < r2 < }. Suppose that F is continuous on , and that F is analytic on . Suppose that |F (r1 ei )| m1 and for all . Show that
log log |F (rei )| m0 r2 log r1 m1 r2 log r1 log r2 log r log rlog r1
|F (r2 ei )| m2
for all r1 < r < r2 and all . Remark: Hadamards three-circle theorem says that log Fr is a convex function of log r. Hardys convexity theorem (Theorem 4.15) is a generalization of this result.
8.3. The RieszThorin interpolation theorem Exercise 8.2.2 Use Fn (z) = m1z 0
2 F (z) e(z 1)/n , z m1
191
(n 1),
and the ordinary maximum principle for analytic functions to give a direct proof of Hadamards three-line theorem. Hint: For each xed n, we have
z zS
lim Fn (z) = 0.
Hence, the maximum principle ensures that |Fn (z)| 1 for all z S. Now, let n .
8.3
The RieszThorin interpolation theorem

N
Let (X, M, ) be a measure space. A simple function is a nite linear combination cn An ,

n=1
where cn are complex numbers and An M with (An ) < . Let us recall that 1 if x A, A (x) = 0 if x A is the characteristic function of A. Clearly, we can assume that An are disjoints. Simple functions belong to all Lp (X) spaces and they are dense in Lp (X) whenever 0 < p < . Theorem 8.2 (RieszThorin interpolation theorem) Let (X, M, ) and (Y, N, ) be two measure spaces. Let p0 , q0 , p1 , q1 [1, ]. Suppose that : Lp0 (X) + Lp1 (X) Lq0 (Y ) + Lq1 (Y ) is a linear map of types (p0 , q0 ) and (p1 , q1 ). Let t [0, 1] and dene 1 1t t = + pt p0 p1 Then and 1 1t t = + . qt q0 q1
: Lpt (X) Lqt (Y )
is a linear map of type (pt , qt ), and

(pt ,qt )
1t (p0 ,q0 )
t (p1 ,q1 ) .
192
Remark: We consider 1/ = 0 and 1/0 = . Proof. Fix t (0, 1). Without loss of generality, we also assume that p0 p1 . We consider three cases. Case 1: Suppose that pt , qt (1, ). This assumption has two advantages. First, Lqt (Y ) is the dual of Lqt (Y ), where 1/qt + 1/qt = 1. Second, simple functions are dense in Lpt (X) and Lqt (Y ). Let be a simple function on X. Since Lp0 (X) + Lp1 (X), we have Lq0 (Y ) + Lq1 (Y ), and thus, by duality,
qt
= sup
Y
() d ,
where the supremum is taken over all simple functions Lqt (Y ) with
qt
1.
Y
We apply Hadamards three-line theorem (Theorem 8.1) to estimate Since and are simple functions we can write them as =
m
() d.
rm eim Am
and
=
n
n ein Bn ,
where rm , n > 0, Am M and Bn N, and each sum has nitely many terms. Dene z =
m
rm
pt
1z z p0 + p1
eim Am
and
z =
n
qt
1z z +q q0 1
ein Bn .
The denition is arranged such that t = and t = . Without loss of generality we can assume the sets {Am } and {Bn } are respectively pairwise disjoint. Hence, for all 0, we have |x+iy | =
m
rm
pt
1x x p0 + p1
Am
and
|x+iy | =
n
qt
1x x +q q0 1
Bn .
In particular, the identities |iy |p0 = |1+iy |p1 = ||pt and |iy |q0 = |1+iy |q1 = ||qt are essential for us. (8.3) (8.2)
8.3. The RieszThorin interpolation theorem Let F (z) =

Y
193
(z ) z d ei(m +n ) (Am ) Bn d rm
pt
1z z p0 + p1
=
m n
qt
1z z +q q0 1
Hence F is an entire function, and F (t) =

Y
() d.
Let S be the strip S = { x + iy : 0 < x < 1, y R }. Since rm

pt
1z z p0 + p1
qt
1z z +q q0 1
= rm
pt
1x x p0 + p1
qt
1x x +q q0 1
the entire function F is bounded on S. Moreover, by Hlders inequality, on the o vertical line z = 0 we have |F (iy)| = (iy ) iy d |iy |q0 d
1 q0
|iy |q0 d
1 p0
1 q0
(p0 ,q0 )
|iy |p0 d
qt
|iy |q0 d
1 q0
Hence, by (8.2)(8.3) and the fact that |F (iy)| Similarly, on the vertical line |F (1 + iy)| =
1, we obtain
pt /p0 . pt
(p0 ,q0 )
(8.4)
z = 1 we have
(1+iy ) 1+iy d |iy |q1 d

1 q1
Y p1
|iy |q1 d
1 p1
1 q1
(p1 ,q1 )
|iy |
d
Y
|iy | d
q1
1 q1
Hence, by (8.2)(8.3), we obtain |F (1 + iy)|

(p1 ,q1 )
pt /p1 . pt
(8.5)
194
Therefore, by Hadamards three-line theorem (Theorem 8.1) and by (8.4) and (8.5), we have |F (t)| = Thus
Y
(p0 ,q0 )
pt /p0 pt t (p1 ,q1 )
1t
pt .
(p1 ,q1 )
pt /p1 pt
1t (p0 ,q0 )
() d
1t (p0 ,q0 )
t (p1 ,q1 ) qt
pt .
Taking the supremum with respect to all with

qt
1 gives
pt .
1t (p0 ,q0 )
t (p1 ,q1 )
(8.6)
This is the required inequality, but just for simple functions Lpt (X). We use some tools from measure theory to show that (8.6) holds for all elements of Lpt (X). Let f Lpt (X). Then there is a sequence of simple functions n Lpt (X) such that lim n f pt = 0.
n
If we can show that

n
lim
n f
qt
= 0,
then immediately we see that (8.6) is also fullled by f and we are done. This fact is actually our main assumption whenever pt = p0 or pt = p1 . Hence we assume that p0 < pt < p1 . Since (n )n1 is convergent in Lpt (X), it is necessarily a Cauchy sequence, and thus, by (8.6), (n )n1 is a Cauchy sequence in Lqt (Y ). Therefore, there is g Lqt (Y ) such that lim n g qt = 0.
n
We show that n also converges in measure to f and this is enough to ensure that g = f . Let n = f n pt and let Sn = { x X : |f (x) n (x)| > n }. Clearly we can assume that n > 0. Since otherwise, f would be a simple function, and we saw that (8.6) is fullled by such functions. Now, on the one hand, by Chebyshevs inequality, (Sn ) = d f n n
pt
Sn
Sn
f n pt n
pt pt
= 1,
8.3. The RieszThorin interpolation theorem and thus, by Hlders inequality with = pt /p0 , we obtain o (f n )Sn
p0 p0
195
=
X
|f n |p0 Sn d |f n |
p0 pt . p0
1
d
X
|Sn |
f n
On the other hand, on X \ Sn we have |f n |/n 1, and thus f n n

p1
X\Sn
X\Sn
f n n
p1
pt
f n pt n
pt pt
= 1,
which gives (f n ) X\Sn Therefore, (f n )Sn

p0
f n
pt .
and
(f n ) X\Sn
p1
as n . Since is of type (p0 , q0 ) and (p1 , q1 ), we conclude (f n )Sn

q0
and
(f n ) X\Sn
q1
as n . In particular, (f n )Sn 0 and (f n ) X\Sn 0
in measure as n . Adding both together gives n f in measure, and we are done. Case 2: pt (1, ), but either qt = 1 or qt = . In this case, we necessarily have q0 = q1 = qt . Let T be a bounded linear functional on Lqt (Y ) and let F (z) = T (z ), where is a simple function on X. Hence |F (iy)| T and |F (1 + iy)| T 1+iy
q1
iy
q0
T T
(p0 ,q0 )
pt /p0 pt pt /p1 . pt
(p1 ,q1 )
Therefore, by Hadamards three-line theorem (Theorem 8.1), we have |T ()| = |F (it)| T

1t (p0 ,q0 )
t (p1 ,q1 )
pt .
196
Finally, the HahnBanach theorem implies that

qt
1t (p0 ,q0 )
t (p1 ,q1 )
pt .
The rest is exactly as in Case 1. Case 3: Either pt = 1 or pt = . In this case, we necessarily have p0 = p1 = pt . Let f Lpt (X). Then, by Hlders inequality, o Tf
qt
= =
|T f |1t |T f |t Tf
1t q0
Tf f
qt t q1 p0 1t
(p0 ,q0 ) 1t (p0 ,q0 )
f
p1 .
(p1 ,q1 )
p1
t (p1 ,q1 )
As a very special case, if r = p0 = q0 and s = p1 = q1 in the RieszThorin interpolation theorem, and, by assumption, the linear map is of types (r, r) and (s, s), i.e. f and f
s r
(r,r)
f f
r,
(f Lr (X)), (f Ls (X)),
(s,s)
s,
then, for each p between r and s with 1 1t t = + , p r s the map is of type (p, p), i.e. f and
(p,p) p
(0 t 1),
: Lp (X) Lp (Y ) (f Lp (X)),
t (s,s) .
(p,p)
p,
1t (r,r)
Exercises
Exercise 8.3.1 First, give a direct proof of Corollary 1.5. Then use this corollary and the RieszThorin interpolation theorem to obtain another proof of Youngs inequality (Theorem 1.4).
8.4. The HausdorYoung theorem
197
Exercise 8.3.2 Let (X, M, ) and (Y, N, ) be two measure spaces. Let p0 , q0 , p1 , q1 [1, ]. Suppose that is a linear map dened for all simple functions on X such that is a measurable function on Y and and Let t [0, 1] and dene 1t t 1 = + pt p0 p1 Show that
qt q1 q0
M0 M1
p0
p1 .
and
1 1t t = + . qt q0 q1
pt .
1t t M0 M1
Moreover, if pt < , then can be uniquely extended to the whole space Lpt ().
8.4
The HausdorYoung theorem
In this section we study the HousdorYoung theorem about the Fourier transform on T. There is another version with a similar proof about the Fourier transform of functions dened on R. Theorem 8.3 (The HausdorYoung theorem) Let f Lp (T), 1 p 2. Then f , the Fourier transform of f , belongs to q (Z), where q is the conjugate exponent of p, and f q f p. Proof. By Lemma 1.1, F : L1 (T) f is of type (1, ) with F
(1,)
(Z) f
1.
2
By Parsevals identity (Corollary 2.22), F : L2 (T) f is of type (2, 2) with F (As a matter of fact, we have F here.) Hence, with p0 = 1,
(2,2)
(Z) f
1.
2,2
1,
= F
= 1. But we do not need this q1 = 2,
q0 = ,
p1 = 2,
198
the RieszThorin interpolation theorem (Theorem 8.2) implies that F : Lpt (T) f is well-dened and has type (pt , qt ) with F Hence f
qt (pt ,qt ) qt
(Z) f
1t (1,)
t (2,2)
1.
pt .
To nd the relation between pt and qt , note that 1 1t t t = + = pt 2 2 and t t 1t 1 + =1 . = qt 1 2 2 1 1 + = 1. pt qt
Hence pt [1, 2] and
Corollary 8.4 Let f H p (T), 1 p 2. Then f , the Fourier transform of f , belongs to q (Z+ ), where q is the conjugate exponent of p, and f
q
p.
Exercises
Exercise 8.4.1 Let 1 p 2, and let q be the conjugate exponent of p. Let {an }nZ be a sequence in p (Z). Show that there is an f Lq (T) such that f (n) = an , and f
q
(n Z), f p.
Hint 1: Use duality and Theorem 8.3. Hint 2: Give a direct proof using the RieszThorin interpolation theorem.
8.4. The HausdorYoung theorem
199
Exercise 8.4.2 Let 1 p 2 and let q be the conjugate exponent of p. Let {an }nZ be a sequence in p (Z+ ). Show that there is an f H q (T) such that f (n) = an , and f Hint: Use duality and Corollary 8.4.
q
(n Z), f p.
Exercise 8.4.3 Let 1 < p < 2 and let q be the conjugate exponent of p. According to Theorem 8.3, the map Lp (T) f
q
(Z) f
is well-dened, and by the uniqueness theorem (Corollary 2.13) it is injective. However, show that it is not surjective.
Exercise 8.4.4 Let 1 < p < 2 and let q be the conjugate exponent of p. According to Corollary 8.4, the map H p (T) f
q
(Z+ ) f
is well-dened, and by the uniqueness theorem (Corollary 2.13) it is injective. However, show that it is not surjective. Hint: See Exercise 8.4.3. Exercise 8.4.5 According to Corollary 8.4 and by the RiemannLebesgue lemma (2.17), the map H 1 (T) c0 (Z+ ) f f is well-dened, and by the uniqueness theorem (Corollary 2.13) it is injective. However, show that it is not surjective. Hint: See Exercise 2.5.6. Exercise 8.4.6 [The generalized HausdorYoung theorem] Let (X, M, ) be a measure space. Let (n )nZ be an orthonormal family in L2 () such that |n (x)| M
200
for all n Z and all x X. In the rst place, show that n Lq () for any 2 q . Then, for each f Lp (), 1 p 2, dene f (n) =
X
f n d,
q
(n Z),
and f = (f (n))nZ . Show that f f

q
(Z), where 1/p + 1/q = 1 and

p.
M (2p)/p f
Hint: Prove the inequality for p = 1 and p = 2 and then use the RieszThorin interpolation theorem (Theorem 8.2).
8.5
An interpolation theorem for Hardy spaces

: V1 V2 Vn V
Let V, V1 , V2 , . . . , Vn be complex vector spaces. A map
is called a multilinear operator if it is linear with respect to each of its arguments while the others are kept xed. If these vector spaces are subspaces (not necessarily closed) of certain Lebesgue spaces, we say that is of type (p1 , p2 , . . . , pn , q) whenever (f1 , f2 , . . . , fn )
q
C f1
p1
f2
p2
fn
pn .
The following result is another version of the RieszThorin interpolation theorem (Theorem 8.2). More results of this type can be found in [2], [12] and [22]. Lemma 8.5 Let (Xi , Mi , i ), i = 1, 2, . . . , n, and (Y, N, ) be measure spaces. Let Vi represent the vector space of simple functions on Xi , and let V be the space of measurable functions on Y . Suppose that : V1 V2 Vn V is a multilinear operator of types (p1,0 , p2,0 , . . . , pn,0 , q0 ) and (p1,1 , p2,1 , . . . , pn,1 , q1 ), where pi,0 , q0 , pi,1 , q1 [1, ], with constants C0 and C1 . Let t [0, 1] and dene 1 1t t 1 1t t = + and = + . pi,t pi,0 pi,1 qt q0 q1
1t t Then is of type (p1,t , p2,t , . . . , pn,t , qt ) with constant C0 C1 , i.e.
(1 , 2 , . . . , n )
qt
1t t C0 C1 1
p1,t
p2,t
pn,t
for all simple functions 1 , 2 , . . . , n . Moreover, if p1,t , p2,t , . . . , pn,t are all nite, then can be extended to Lp1,t (1 ) Lp2,t (2 ) Lpn,t (n ), preserving the last inequality.
8.5. An interpolation theorem for Hardy spaces
201
Proof. The proof is similar to the proof of the RieszThorin interpolation theorem (Theorem 8.2). We will use the same notations applied there. Case 1: p1,t , p2,t , . . . , pn,t < and qt > 1. Fix simple functions 1 , 2 , . . . , n and with 1
p1,t
= 2
p2,t
= = n
pn,t
qt
= 1.
For the simple function i =

m
rm eim Am
on the measure spaces Xi , we dene i,z =

m
rm
pi,t
1z z pi,0 + pi,1
eim Am .
Similarly, for =
n
n ein Bn ,
qt
1z z +q q0 1
let z =
n
ein Bn .
Let F (z) =
Y
(1,z , 2,z , . . . , n,z ) z d.
Hence, F is an entire function, which is bounded on S and F (t) =

Y
(1 , 2 , . . . , n ) d.
By Hlders inequality, o |F (iy)| |(1,iy , 2,iy , . . . , n,iy )|

p1,0 q0
1 q0
d
Y pn,0
|iy |
q0
q0
1 q0
C0 1,iy = C0 1 and |F (1 + iy)|
2,iy 2
p2,0
n,iy
iy
p1,t /p1,0 p1,t
p2,t /p2,0 p2,t
pn,t /pn,0 pn,t
qt /q0 qt
= C0
|(1,1+iy , 2,1+iy , . . . , n,1+iy )|q1 d

p1,1
1 q1
Y pn,1
|1+iy |q1 d
q1
1 q1
C1 1,1+iy = C1
2,1+iy
p2,1
n,1+iy
1+iy
qt /q1 qt
1,t 1 p1,t /p1,1 p
2 p2,t /p2,1 p2,t
n pn,t /pn,1 pn,t
= C1 ,
202
for all y R. Hence, by Hadamards three-line theorem (Theorem 8.1), |F (t)| =

Y 1t t (1 , 2 , . . . , n ) d C0 C1 . qt
Taking the supremum with respect to with |(1 , 2 , . . . , n )|qt d

1 qt
= 1 gives
1t t C0 C1 ,
which is equivalent to the required result. Case 2: Any of p1,t = , p2,t = , . . . , pn,t = , or qt = 1 happens. If, for example, p1,t = , then we necessarily have p1,t = p1,0 = p1,1 = . Similarly, qt = 1 implies qt = q0 = q1 = 1. If any of these possibilities happens, in the denition of F (z) remove z from the corresponding function, e.g. F (z) =
Y
(1 , 2,z , . . . , n,z ) d
corresponds to the case p1,t = , qt = 1 and p2,t , . . . , pn,t < . Then follow the same procedure as in Case 1. It remains to show that if p1,t , p2,t , . . . , pn,t are all nite, then can be extended by continuity to Lp1,t (1 ) Lp2,t (2 ) Lpn,t (n ), preserving the inequality (f1 , f2 , . . . , fn )
qt 1t t C0 C1 f1 p1,t
f2
p2,t
fn
pn,t .
The proof is based on the following simple observation. Let i and i be simple functions on Xi . Let = (1 , 2 , . . . , n ) (1 , 2 , . . . , n ). To estimate , write = (1 , 2 , . . . , n ) (1 , 2 , . . . , n ) + (1 , 2 , . . . , n ) (1 , 2 , . . . , n ) . . . + (1 , . . . , n1 , n ) (1 , 2 , . . . , n ). Hence,
qt 1t t C0 C1 n1 n
max{ i
i
pi,t ,
pi,t } i=1
i i
pi,t
Therefore, by continuity, can be extended to Lp1,t (1 ) Lp2,t (2 ) Lpn,t (n ).
8.5. An interpolation theorem for Hardy spaces
203
The RieszThorin interpolation theorem (Theorem 8.2) shows that linear operators between Lebesgue spaces can be interpolated. Using Lemma 8.5, we show that linear operators on Hardy spaces enjoy a similar property. Theorem 8.6 Let (Y, N, ) be a measure space. Suppose that : H p0 (D) Lq0 (Y ) and : H p1 (D) Lq1 (Y ),
where p0 , p1 (0, ) and q0 , q1 [1, ], are bounded with constants C0 and C1 . Let t [0, 1] and dene 1 1t t = + pt p0 p1 Then is bounded and (F )
qt 1t t C C0 C1 F pt ,
and
1 1t t = + . qt q0 q1
: H pt (D) Lqt (Y ) (F H pt (D)),
where C = C(p0 , p1 ) is an absolute constant. Proof. Without loss of generality, assume that p0 p1 . By assumption, (F ) (F )
q0 q1
C0 F C1 F
p0 , p1 ,
(F H p0 (D)), (F H (D)).
p1
(8.7) (8.8)
For a function Lp (T), 1 1. Let 1 , 2 , . . . , n be functions in Lnp0 (T). By (8.10), i H np0 (D) and thus, by Hlders inequality, 1 2 n H p0 (D). Hence, we can dene o (1 , 2 , . . . , n ) = (1 2 n ) (8.11)
on Lnp0 (T) Lnp0 (T). Clearly, is linear with respect to each argument. Moreover, by (8.7)(8.8) and by Hlders inequality, o (1 , 2 , . . . , n ) and (1 , 2 , . . . , n )
q1 q0
C0 1 2 n
p0
C0 1
np0
np0
np0
C1 1 2 n
p1
C1 1
np1
np1
np1 .
204 Hence, by (8.10), (1 , 2 , . . . , n ) (1 , 2 , . . . , n ) Therefore, by Lemma 8.5, (1 , . . . , n )

qt q0 q1
C0 An 0 1 np C1 An 1 np 1
np0 np1
2 2
n np1 n
np0
np0 , np1 .
1t t C0 C1 (A1t At 1 )n 1 np0 np
npt
npt .
(8.12)
Let F H pt (D). Without loss of generality, assume that F (0) > 0. Since otherwise, either multiply F by a constant of modulus one or replace it by F + if F (0) = 0. By Theorem 7.7, F = Bn , where B is a Blaschke product and H npt (D) with B npt = npt = F pt and B(0) > 0, (0) > 0. Let npt npt pt 1 = B and 2 = 3 = = n = . i .
Since npt > 1, each i has a representation of the form (8.9) with i = Hence, by (8.11) and (8.12), (F ) But i Therefore, (F )
qt 1t t C C0 C1 F pt , npt qt 1t t C0 C1 (A1t At 1 )n 1 np0 np npt 1/n pt .
npt .
npt
= F
(F H pt (D)),
where C = (A1t At 1 )n . np0 np Using the well-known isometry between H p (D) and H p (T), we can replace D by T in the preceding theorem.
Exercises
Exercise 8.5.1 Let P be an analytic polynomial. Show that there is a nite Blaschke product B and an analytic polynomial Q with no zeros in the open unit disc D such that P = B Q. Exercise 8.5.2 Let (Y, N, ) be a measure space, let V be the space of measurable functions on Y , and let P denote the space of analytic polynomials on T. Suppose that : P V is a linear operator of types (p0 , q0 ) and (p1 , q1 ), where p0 , p1 (0, ) and q0 , q1 [1, ], with constants C0 and C1 . Let t [0, 1] and dene 1t t 1 = + pt p0 p1 and 1 1t t = + . qt q0 q1
8.6. The HardyLittlewood inequality Show that is of type (pt , qt ) and (P )

qt 1t t C C0 C1 P pt ,
205
(P P),
where C = C(p0 , p1 ) is an absolute constant. Moreover, can be extended to H pt (T), preserving the last inequality. Hint: Modify the proof of Theorem 8.6. Moreover, the ideas used at the end of the proof of Lemma 8.5 and the factorization given in Exercise 8.5.1 might be useful.
8.6
The HardyLittlewood inequality
The RiemannLebesgue lemma (Corollary 2.17) says that the Fourier coecients of a function in L1 (T) tend to zero as |n| grows. Parsevals identity (Corollary 2.22) and Hardys inequality (Theorem 7.16) provide further information if we consider smaller classes of functions. An easy argument based on Theorem 8.6 enables us to generalize these two results. Let Y = {0, 1, 2, . . . } be equipped with the measure (n) = 1 , (n + 1)2
(n 0).
Hence, Lp (Y ), 0 < p < , consists of all complex sequences (an )n0 such that (an )n0
p
|an |p (n + 1)2 n=0 L1 (Y )
1 p
< .
According to Hardys inequality, the operator : H 1 (T) f
( (n + 1)f (n) )n0
is of type (1, 1) with (1,1) . On the other hand, by Parsevals identity, the operator : H 2 (T) f is of type (2, 2) with
(2,2)
L2 (Y )
( (n + 1)f (n) )n0 = 1. To apply Theorem 8.6, note that p1 = q1 = 2,

1 p
p0 = q0 = 1, and thus p = pt = qt [1, 2], and that f

p
|f (n)|p (n + 1)2p n=0
(f H p (T)).
Hence we immediately obtain the following result.
206
Theorem 8.7 (HardyLittlewood) Let f H p (T), 1 p 2. Then |f (n)|p (n + 1)2p n=0
1 p
Cp f
p,
where Cp is a constant just depending on p. The HardyLittlewood inequality is valid even if 0 < p < 1. But we do not discuss this case here.
Exercises
Exercise 8.6.1 Let 2 q < and let (an )n0 be a sequence of complex numbers such that
(n + 1)q2 |an |q < .
n=0
Show that F (z) =
n=0
an z n is in H q (D) and
q
Cq
n=0
(n + 1)q2 |an |q
1 q
where Cq is a constant just depending on q. Hint: Use duality and Theorem 8.7. Exercise 8.6.2 Let f Lp (T), 1 < p 2. Show that |f (n)|p (|n| + 1)2p n= where Cp is a constant just depending on p. Hint: Use Theorem 8.7 and Exercise 6.3.4. Exercise 8.6.3 Let 2 q < . Let (an )nZ be a sequence of complex numbers such that

1 p
Cp f
p,
(|n| + 1)q2 |an |q < .
n=
Show that there is an f Lq (T) such that f (n) = an , n Z, and
Cq
(|n| + 1)
n=
q2
|an |
1 q
where Cq is a constant just depending on q. Hint: Use Exercise 8.6.2 and duality.
Chapter 9
The Fourier transform

9.1 Lebesgue spaces on the real line
From now on we study upper half plane analogues of the results obtained before for the unit disc. Since the real line R is not compact and the Lebesgue measure is not nite on R, we face several diculties. Using a conformal mapping between the upper half plane C+ and the open unit disc D, which also establishes a correspondence between R and T \ {1}, some of the preceding representation theorems can be rewritten for the upper half plane. Nevertheless, working with the Poisson kernel for C+ is somehow easier than its counterpart for D. That is why we mostly provide direct proofs in the following. Let f be a measurable function on R and let
f and
|f (t)| dt
1 p
(0 0
M : |{ t : |f (t)| > M }| = 0 .
The Lebesgue spaces Lp (R), 0 < p , are dened by Lp (R) = { f : f

p
< }.
For 1 p , Lp (R) is a Banach space and L2 (R) equipped with the inner product
f, g =
f (t) g(t) dt
is a Hilbert space. It is important to note that, contrary to the case of the unit circle, the Lebesgue spaces on R do not form a chain. In other words, for each p, q (0, ], p = q, we have Lp (R) \ Lq (R) = . 207
208
Chapter 9. The Fourier transform
We will have to cope with this fact later on. The space of all continuous functions on R is denoted by C(R). An element of C(R) is not necessarily bounded or uniformly continuous on R. For most applications it is enough to consider the smaller space C0 (R) = { f : f continuous on R and lim f (t) = 0 }.
|t|
Elements of C0 (R) are certainly bounded and uniformly continuous on R. However, the simple example f (t) = sin t shows that the inverse is not true. Since elements of C0 (R) are bounded, C0 (R) can be considered as a subspace of L (R). Indeed, in this case, we have f
= max |f (t)|
tR
and the maximum is attained. Another important subclass of C0 (R) is Cc (R), which consists of all continuous functions of compact support, i.e. there is an M = M (f ) such that f (t) = 0
n for all |t| M . The smooth subspaces C n (R), Cc (R), C (R) and Cc (R) are dened similarly. In spectral analysis of Hardy spaces, we will also need
Lp (R+ ) = { f Lp (R) : f (t) = 0 for almost all t 0 } and C0 (R+ ) = { f C0 (R) : f (t) = 0 for all t 0 }.
The space of all Borel measures on R is denoted by M(R). This class equipped with the norm = ||(R), where || is the total variation of , is a Banach space. Each function f L1 (R) corresponds uniquely to the measure d(t) = f (t) dt and thus we can consider L1 (R) as a subspace of M(R). Note that = f By a celebrated theorem of F. Riesz, the dual of C0 (R) is M(R).
1.
Exercises
Exercise 9.1.1 Let f : R C and dene its translations by f (t) = f (t ), Show that
0
( R). =0
lim f f
9.2. The Fourier transform on L1 (R)
209
if f X with X = Lp (R), 1 p < , or X = C0 (R). Provide an example to show that this property does not hold if X = L (R). However, show that f
X
= f
X,
( R),
in all spaces mentioned above. Exercise 9.1.2 Show that C0 (R) is closed in L (R).
Exercise 9.1.3 Show that each element of C0 (R) is uniformly continuous on R. More generally, suppose that f C(R) and that L1 = lim f (t)
t+
and L2 = lim f (t)

t
exist (we do not assume that L1 = L2 ). Show that f is uniformly continuous on R. Exercise 9.1.4 Show that Cc (R) is dense in Lp (R), 1 p < . Can you show that Cc (R) is dense in Lp (R), 1 p < ?
Exercise 9.1.5 Show that Cc (R) is dense in C0 (R). Can you show that Cc (R) is dense in C0 (R)? Remark: We will develop certain techniques later on which might be useful for the second part of this question and the previous one. See Sections 9.4 and 10.4.
Exercise 9.1.6 Show that Lp (R+ ) is a closed subspace of Lp (R). Similarly, show that C0 (R+ ) is a closed subspace of C0 (R).
9.2
The Fourier transform on L1 (R)

f (t) =

The Fourier transform of f L1 (R) is dened by f ( ) ei 2t d, (t R).
The Fourier integral of f is given formally by

f ( ) ei 2t d,
(t R).
210
One of our tasks is to study the convergence of this integral. The Fourier transform of a measure M(R) is dened similarly by ei 2t d( ),
R
(t) =
(t R).
However, if we consider L1 (R) as a subspace of M(R), the two denitions of Fourier transform are consistent. No wonder our most important example is the Poisson kernel for the upper half plane, which is dened by y 1 , t2 + y 2
Py (t) =
(y > 0).
(9.1)
(See Figure 9.1.)
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
Fig. 9.1. The Poisson kernel Py (t) for y = 0.8, 0.5, 0.2.
For a xed x + iy C+ , let us nd the Fourier transform of f (t) = y 1 . (x t)2 + y 2
9.2. The Fourier transform on L1 (R) Since Py (x) is an even function of x, we have f (t)
211
Py (x ) ei 2t d Py ( ) ei 2t(x ) d

= e
i 2xt
Py ( ) ei 2t d Py ( ) ei 2t d
= e and thus
i 2xt
f (t) = ei 2xt
Py ( ) ei 2|t| d.
Let R > y, and let R be the positively oriented curve formed with the interval [R, R] and the semicircle { Rei : 0 }. (See Figure 9.2.)
Fig. 9.2. The curve R . Then f (t) = ei 2xt lim = ei 2xt lim The only pole of our integrand F (w) = y ei 2|t|w (w2 + y 2 )
R
y ei 2|t| d ( 2 + y 2 ) R y ei 2|t|w dw. (w2 + y 2 ) R
212

1 2i
inside R is iy. Moreover, the residue of F at this point is fore, by the residue theorem, f (t) = ei 2xt2y|t| . In particular, the Fourier transform of the Poisson kernel Py (t) = is Py (t) = e2y|t| . (See Figure 9.3.) y 1 2 + y2 t
e2y|t| . There-
Fig. 9.3. The spectrum of Py .
Note that, for each xed t R, we have Py (t) 1 as y 0. We will see other families of functions having this behavior. This phenomenon will lead us to the denition of an approximate identity on R. We need the Fourier transform of four functions, which are gathered in Table 9.1. Since in each case our function is exponential, it is easy to calculate its Fourier transform. In all cases, z = x + iy C+ . Let us also remember that 1 if t > 0, 0 if t = 0, sgn(t) = 1 if t < 0. In Table 9.1, it is enough to verify the rst and second lines. The third and last lines are linear combinations of the rst two lines.
9.2. The Fourier transform on L1 (R) Table 9.1. Examples of the Fourier transform. f (t) f (t)
213
ei 2x t2y |t|
y 1 (x t)2 + y 2 xt 1 (x t)2 + y 2 1 1 2i t z 1 1 2i t z
i sgn(t) ei 2x t2y |t|
ei 2z t 0
if t > 0 if t < 0
0 ei 2z t
if t > 0 if t < 0
Lemma 9.1 Let M(R). Then C(R) L (R) and In particular, for each f L1 (R), f Proof. For all t R, we have |(t)| = =
R
1.
ei 2t d( )
R
|ei 2t | d||( ) d||( ) = ||(R) = ,
which is equivalent to . The second inequality is a special case of the rst one if we consider d(t) = f (t) dt and note that = f 1 .
214
Chapter 9. The Fourier transform To show that is (uniformly) continuous on R, let t, t R. Then |(t) (t )| = =
R R
ei 2t ei 2t
d( )
|ei 2t ei 2t | d||( ) |1 ei 2(tt ) | d||( ).
The integrand is bounded, |1 ei 2(tt ) | 2, and || is a nite positive Borel measure on R. The required result now follows from the dominated convergence theorem.
Exercises
Exercise 9.2.1 Let z = x + iy C+ . Show that Py (x t) = y 1 = (x t)2 + y 2 i 1 zt .
Exercise 9.2.2 Let f L1 (R) and let g(t) = ei20 t f (t 1 ), where 0 , 1 are xed real constants. Evaluate g in terms of f . Exercise 9.2.3 Let M(R) and let (E) = (E). Evaluate in terms of . Exercise 9.2.4 Let f L1 (R) and let g(t) = 2itf (t). Suppose that g L1 (R). Show that f C 1 (R) and df = g. dt What can we say if tn f (t) L1 (R)? Exercise 9.2.5 Let f L1 (R) and dene
t
F (t) =
f ( ) d,
(t R).
Suppose that F L1 (R). Evaluate F in terms of f .
9.2. The Fourier transform on L1 (R) Exercise 9.2.6 [GaussWeierstrass kernel] Let > 0 and let G (t) = et , Show that
2 1 G (t) = et / , 2
215
(t R). (t R).
Remark: Fix x R and > 0. A small modication of the preceding result shows that the Fourier transform of f (t) = ei 2xtt , is
2 1 f (t) = e(xt) / , 2
(t R), (t R).
(See Figure 9.4.)
2.5
2.0
1.5
1.0
0.5
Fig. 9.4. The spectrum of G for = 0.8, 0.5, 0.2.
Exercise 9.2.7
[Fejrs kernel] Fix > 0 and let e K (t) = sin(t) t

2
(t R).
(See Figure 9.5.) Show that K (t) = max 1 |t| ,0 , (t R).
216 (See Figure 9.6.)
10
0.4
0.3
0.2
0.1
0.0
0.1
0.2
0.3
0.4
Fig. 9.5. The Fejr kernel K (t) for = 4, 7, 10. e
Fig. 9.6. The spectrum of K .
Exercise 9.2.8
Fix x R and > 0. Let 1 |t| ei 2xt f (t) = 0 sin((x t)) (x t)

2
if |t| , if |t| .
Show that f (t) =
(t R).
9.2. The Fourier transform on L1 (R) Exercise 9.2.9
217
Let f C 1 (R) and suppose that f, f L1 (R). Show that f (t) = 2itf (t), (t R).
(n) n
What can we say if f C (R) and f, f , . . . , f Exercise 9.2.10
L1 (R)?
[de la Valle Poussins kernel] Fix > 0 and let e V (t) = 2K2 (t) K (t), (t R). |t| ,
(See Figure 9.7.) Show that
1 2 0
|t|
if
V (t) =
if |t| 2, if |t| 2.
(See Figure 9.8.)
30
20
10
0.2
0.1
0.1
0.2
Fig. 9.7. The de la Valle Poussin kernel V (t) for = 4, 7, 10. e
Fig. 9.8. The spectrum of V .
218
9.3
The multiplication formula on L1 (R)
The multiplication formula is an easy consequence of Fubinis theorem. However, it has profound implications in the spectral synthesis of harmonic functions in the upper half plane and also in extending the denition of the Fourier transform to other Lp (R) spaces. Lemma 9.2 (Multiplication formula) Let f L1 (R) and let M(R). Then f (t) d(t) =
R
f (t) (t) dt.
In particular, for f, g L1 (R),

f (t) g(t) dt =
f (t) g (t) dt.
Proof. By Fubinis theorem, we have f (t) d(t)

R
=
R
f ( ) ei 2t d ei 2 t d(t)
R
d(t) d
f ( )
f ( ) ( ) d.
Exercises
Exercise 9.3.1

Let > 0 and let f L1 (R). Show that

2 1 e(xt) / f (t) dt =
2 et f (t) ei 2xt dt
for all x R. Hint: Use the multiplication formula and Exercise 9.2.6. Exercise 9.3.2
Let > 0 and let f L1 (R). Show that sin((x t)) (x t)

2
f (t) dt =
1 |t|/ f (t) ei 2xt dt
for all x R. Hint: Use the multiplication formula and Exercise 9.2.8.
9.4. Convolution on R
219
9.4
Convolution on R
(f g)(t) =

As in the unit circle, we dene the convolution of two functions f, g L1 (R) by f ( ) g(t ) d.
By Fubinis theorem, we see that

|f ( ) g(t )| d
dt
|f ( )|
|g(t )| dt d

|f ( )| d
1
|g(t)| dt
< .
Hence, (f g)(t) is well-dened for almost all t R, and moreover f g L1 (R) with f g 1 f 1 g 1. (9.2) Lemma 9.3 Let f, g L1 (R). Then f g = f g. Proof. By Fubinis theorem, for each t R, we have f g(t)
(f g)( ) ei 2t d

f (s) g( s) ds

ei 2t d ds ds
f (s) f (s)

g( s) ei 2t d g( ) ei 2t( +s) d

= =
f (s) ei 2ts ds
g( ) ei 2t d
= f (t) g (t).
Using Rieszs theorem, we can also dene the convolution of two measures , M(R) by the relation (t + ) d(t) d( ) =
R R R
(t) d( )(t),
(9.3)
where C0 (R). It is rather easy to see that .
220
But, since the exponential functions x (t) = ei 2xt are not in C0 (R), it is slightly more dicult to show that = . However, we do not need this result in the following. If M(R) and is absolutely continuous with respect to the Lebesgue measure, i.e. d(t) = f (t) dt, where f L1 (R), then, for each C0 (R), we have
R
(t) d( )(t) = =
( + s) d( ) d(s)
R R
( + s) f (s) ds
R R
d( ) d( ) dt.
=
R R
(t) f (t ) dt
=
R
(t)
R
f (t ) d( )
Therefore, is also absolutely continuous with respect to the Lebesgue measure, and according to the previous calculation we may write d( )(t) = ( f )(t) dt, where ( f )(t) = for almost all t R.
R
f (t ) d( )
(9.4)
Exercises
Exercise 9.4.1

Let f, g L1 (R). Show that

(t + ) f (t) g( ) dt d =

(t) (f g)(t) dt
for all C0 (R). Remark: This exercise shows that the generalized denition (9.3) is consistent with the old one whenever both measures are absolutely continuous with respect to the Lebesgue measure. Exercise 9.4.2 Let f (t) = Evaluate f f . 1 0
if |t| 1, if |t| > 1.
221
9.5
Youngs inequality
On the unit circle the space L1 (T) contains all Lebesgue spaces Lp (T), 1 p . As a simple consequence, after dening f g over L1 (T), it was used even if f and g were from dierent subclasses Lp (T). Due to the fact that dt is not a nite measure on R, Lebesgue spaces Lp (R), 1 p , do not form a chain and thus f g is not yet dened if either f or g does not belong to L1 (R). However, Youngs theorem assures us that if f Lr (R) and g Ls (R), for certain values of r and s, then f g is a well-dened measurable function on R. We start with a special case of this result and then use the RieszThorin interpolation theorem (Theorem 8.2) to prove the general case. Lemma 9.4 Let f Lp (R), 1 p , and let g L1 (R). Then (f g)(t) is well-dened for almost all t R, f g Lp (R) and f g
p
g 1.
Proof. The case p = 1 was studied at the beginning of Section 9.4. If p = , the result is an immediate consequence of Hlders inequality. As a matter of o fact, in this case (f g)(t) is well-dened for all t R. Hence, suppose that 1 < p < . Let q be the conjugate exponent of p. Then, by Hlders inequality, o

|f ( ) g(t )| d
|f ( )| |g(t )| p |g(t )| q d
|f ( )| |g(t )| d |f ( )| |g(t )| d
p
1 p
|g(t )| d
1 q
1 p
1 q
Hence, by Fubinis theorem,

p
|f ( ) g(t )| d
dt
g g g
p q
|f ( )|p |g(t )| d

dt d
1 1
p q
|f ( )|p
p p
|g(t )| dt g p. 1
p q
f
1 p
= f
p p
Hence
|f ( ) g(t )| d
dt
g 1,
and this inequality ensures that (f g)(t) is well-dened for almost all t R, f g Lp (R) and that f g p f p g 1.
222
Lemma 9.5 Let 1 p and let q be its conjugate exponent. Let f Lp (R) and let g Lq (R). Then f g is well-dened at all points of R and f g is a bounded uniformly continuous function on the real line with f g
g q.
Proof. First, suppose that 1 < p < . Then, by Hlders inequality, we have o

|f ( ) g(t )| d

|f ( )| d
1 p
|g(t )| d
1 q
and thus
|f ( ) g(t )| d f
for all t R. Hence, f g is well-dened at all points of R and f g
g q.
To show that f g is uniformly continuous on the real line, note that |(f g)(t) (f g)(t )| =

|f ( )| |g(t ) g(t )| d |f ( )| |g(t + ) g(t + )| d gt gt

q,
where gt ( ) = g( + t). But the translation operator is continuous on Lq (R), 1 q < . The verication of this fact is similar to the one given for Lq (T) classes, except that on R we should exploit continuous functions of compact support. Fix > 0 and pick Cc (R) such that g
q
< .
Let supp [M, M ]. Hence, for all t, t R with |t t | < 1, gt gt

q
gt t
q q
+ t t + (2M + 2)
q 1/q
+ t gt (tt )
q .
2 g
Since is uniformly continuous on R, there is a < 1 such that |s s | < implies |(s) (s )| < /(2M + 2)1/q . Therefore, gt gt
q
provided that |t t | < . Hence, we have |(f g)(t) (f g)(t )| 3 f

p
if |t t | < . A small modication of this proof also works for p = 1 and p = .
223
Now we have all the necessary ingredients to apply the RieszThorin interpolation theorem (Theorem 8.2) in order to prove Youngs inequality on the real line. Theorem 9.6 (Youngs inequality) Let f Lr (R) and let g Ls (R) with 1 r, s 1 1 + 1. r s Then f g is well-dened almost everywhere on R and f g Lp (R), where 1 1 1 = + 1 p r s and f g
p
and
g s.
Proof. Fix f Lr (R). By Lemma 9.4, the convolution operator : L1 (R) Lr (R) g f g is of type (1, r) with
(1,r)
r.
On the other hand, Lemma 9.5 implies that : Lr (R) L (R) g f g is of type (r , ) with
(r ,)
r,
where 1/r + 1/r = 1. Hence, by the RieszThorin interpolation theorem (Theorem 8.2) with p0 = 1, q0 = r, p1 = r , q1 = , we conclude that : Lpt (R) Lqt (R) g f g
1t (1,r) t (r ,)
is well-dened and has type (pt , qt ) with Hence, f g

qt (pt ,qt )
r.
pt .
To nd the relation between pt and qt , note that 1 1t t t = + =1 pt 1 r r
224 and
Chapter 9. The Fourier transform 1 t 1 t 1t + = . = qt r r r 1 1 1 = + 1. qt r pt
Hence,
Exercises
Exercise 9.5.1 Write |g(t )|1 p
s
|f ( ) g(t )| =
|f ( )|1 p
|f ( )| p |g(t )| p
Now, modify the proof of Theorem 1.4 to give another proof of Youngs inequality on the real line. Exercise 9.5.2 Let 1 p and let q be its conjugate exponent. Let f Lp (R) and let g Lq (R). Show that f g C0 (R).
Chapter 10
Poisson integrals
10.1 An application of the multiplication formula on L1 (R)
Given a family of functions {Fy }y>0 on the real line R, we dene F (x + iy) = Fy (x), (x + iy C+ ),
and thus deal with one function dened in the upper half plane C+ . On the other hand, if F as a function on C+ is given, we can use the preceding identity as the denition of the family {Fy }y>0 . This dual interpretation will be encountered many times in what follows. The Poisson kernel Py (x) = P (x + iy) = y 1 2 + y2 x
is a prototype of this phenomenon. According to Lemma 9.4, if u Lp (R), 1 p , then Py u is well-dened almost everywhere on R and Py u
p
u p.
Looking at (Py u)(x) as a function dened in the upper half plane, we show that it represents a harmonic function there. Moreover, using the multiplication formula, we obtain another integral representation for (Py u)(x) in terms of the Fourier transform of u. We also consider some other convolutions K u, e.g. when K is the conjugate Poisson kernel. Let us start with a general case.
225
226 Theorem 10.1 Let M(R). Then we have 1 1 y (x t)2 + y 2 xt (x t)2 + y 2 1 2i R 1 2i

R
Chapter 10. Poisson integrals
d(t) d(t) d(t) tz d(t) tz
e2y|t| (t) ei 2xt dt, i sgn(t) e2y |t| (t) ei 2xt dt, (t) ei 2z t dt,
(t) ei 2z t dt,
=
0
for all z = x + iy C+ . Each integral is absolutely convergent on compact subsets of C+ . Moreover, each line represents a harmonic function on C+ . The third line gives an analytic function. Proof. All the required identities are special cases of the multiplication formula (Lemma 9.2) applied to the Fourier transforms given in Table 9.1. In Theorem 2.1, by applying the Laplace operator directly, we showed that the given function U is harmonic on the unit disc. This method applies here too. However, we provide a more instructive proof. Without loss of generality, assume that is real. Hence, U (z) = = = 1 1 y d(t) (t x)2 + y 2 i d(t) zt R i 1 d(t) . R zt
R
So U is the real part of an analytic function and thus it is harmonic in the upper half plane. Since | e2y|t| (t) ei2xt | e2y|t| , by the dominated convergence theorem, the integral is absolutely and uniformly convergent on compact subsets of C+ . The second integral is the imaginary part of the preceding analytic function and thus it is harmonic too. The last two lines are linear combinations of the rst two and, moreover, the third integral represents an analytic function. In Theorem 10.1, if the measure is absolutely continuous with respect to the Lebesgue measure, i.e. d(t) = u(t) dt, where u L1 (R), then we obtain the following corollary. One of our main goals is to show that this result also holds even if u Lp (R), 1 p 2. The great task is to dene u whenever 1 < p 2.
10.2. The conjugate Poisson kernel Corollary 10.2 Let u L1 (R). Then we have 1 1

227
y u(t) dt (x t)2 + y 2 xt u(t) dt (x t)2 + y 2 u(t) 1 dt 2i t z 1 2i

e2y|t| u(t) ei 2xt dt, i sgn(t) e2y |t| u(t) ei 2xt dt, u(t) ei 2z t dt,
u(t) ei 2z t dt,
=
0
u(t) dt tz
for all z = x + iy C+ . Each integral is absolutely convergent on compact subsets of C+ . Moreover, each line represents a harmonic function on C+ . The third line gives an analytic function.
10.2
The conjugate Poisson kernel

y = O(1/t2 ) (t x)2 + y 2
Let z = x + iy C+ . Since
as |t| , the function U (z) = 1

R
y d(t) (t x)2 + y 2
is well-dened whenever is a Borel measure on R such that

R
d||(t) < . 1 + t2
(10.1)
Clearly, (10.1) is fullled whenever M(R). Moreover, U represents a harmonic function in the upper half plane. The verication of this fact is exactly as given in the proof of Theorem 10.1. In particular, if d(t) = u(t) dt, where u Lp (R), 1 p , then U (z) = 1

y u(t) dt (x t)2 + y 2
is well-dened and represents a harmonic function in C+ . Moreover, by Lemma 9.4, Uy p u p for all y > 0. We already saw that i y = , zt (x t)2 + y 2
228 and since
Chapter 10. Poisson integrals i xt = , zt (x t)2 + y 2 1 xt d(t). (x t)2 + y 2
we are tempted to say that the harmonic conjugate of U is given by V (z) = However, since (10.2)
xt = O(1/|t|) (x t)2 + y 2

as |t| , (10.2) is well-dened whenever d||(t) < . 1 + |t|
For example, if u Lp (R), 1 p < , then

|u(t)| dt < , 1 + |t|
and thus the harmonic conjugate of U (z) = is given by V (z) = 1 1 y u(t) dt (x t)2 + y 2 xt u(t) dt. (x t)2 + y 2
(10.3)
For the general case, where satises (10.1), simply note that it i + z t 1 + t2 i it + z t 1 + t2 and that, as |t| , xt t + = O(1/t2 ). (x t)2 + y 2 1 + t2 Therefore, assuming (10.1), the harmonic conjugate of U (z) = is well-dened by the formula V (z) = 1

y , (x t)2 + y 2
xt t + (x t)2 + y 2 1 + t2
y d(t) (t x)2 + y 2
xt t + (x t)2 + y 2 1 + t2
d(t).
(10.4)
10.3. Approximate identities on R The family Qy (t) = t 1 , 2 + y2 t (y 0),
229
(10.5)
is called the conjugate Poisson kernel for the upper half plane. (See Figure 10.1.) However, as we will see, this kernel is more dicult to handle than the Poisson kernel.
0.8
0.6
0.4
0.2
K K K K
0.2
0.4
0.6
0.8
Fig. 10.1. The conjugate Poisson kernel Qy (t) for y = 0.8, 0.5, 0.2.
10.3
Approximate identities on R
Let { } be a family of integrable functions on R. In our examples below, we will use y, and instead of and in all cases they range over the interval (0, ). However, it should be noted that y and tend to zero while tends to innity. Therefore, depending on the example, in the following lim means limy0 , lim0 or lim . Similarly, 0 means y < y0 , < 0 or > 0 . The family { } of integrable functions is an approximate identity on R if it satises the following properties: (a) for all , C = sup

| (t)| dt < ;
(b) for all ,
(t) dt = 1;
230 (c) for each xed > 0, lim

|t|>
| (t)| dt = 0.
If, for all and for all t R,
(t) 0,
we say that { } is a positive approximate identity on R. In this case, (a) follows from (b) with C = 1. Our main example is of course the Poisson kernel Py (t) = y 1 , t2 + y 2 (t R),
which is a positive approximate identity. Given f Lp (R), 1 p , or f C0 (R), and an approximate identity { } on R, we form the new family { f }. With a measure M(R) we can also consider the family { }. Then, similar to the case of the unit circle, we will explore the way f and approach, respectively, f and as grows. In particular, we are interested in the families Py f and Py , which also represent harmonic functions in the upper half plane. By Theorem 10.1 and Corollary 10.2, the Fourier integrals of these two families are respectively

e2y|t| f (t) ei 2xt dt
and
e2y|t| (t) ei 2xt dt,
which are weighted Fourier integrals of f and .
Exercises
Exercise 10.3.1 Let f L1 (R) and dene (t) = f (t), (t R),
where > 0. Evaluate in terms of f . For each xed t R, what is the behavior of (t) if ? Exercise 10.3.2 Let f L1 (R) with
f (t) dt = 1.
10.3. Approximate identities on R Let > 0 and dene (t) = f (t), (t R).
231
Show that is an approximate identity on R. Under what condition is a positive approximate identity? Exercise 10.3.3 [GaussWeierstrass kernel] Let > 0 and let
2 1 G (t) = et / ,
(t R).
Show that {G } is a positive approximate identity on R. Exercise 10.3.4 Let p be a polynomial of degree n, and let f (t) = p(t) et /2 . 2 Show that f (t) = q(t) et /2 , where q is a polynomial of degree n. Moreover, show that q is odd (even) if p is odd (even). Exercise 10.3.5 [Fejr kernel] Let > 0 and let e K (t) = sin( t) t
2
2
(t R).
Show that K is a positive approximate identity on R. Exercise 10.3.6 [de la Valle Poussins kernel] Let > 0 and let e V (t) = 2K2 (t) K (t), Show that V is an approximate identity on R. Exercise 10.3.7 Show that

(t R).
K2 (t) dt =
2 . 3
Exercise 10.3.8
[Jacksons kernel] Let > 0 and let K2 (t) 3 = K 2 2 2 sin( t) t

4
J (t) =
(t R).
(See Figure 10.2.) Show that J is a positive approximate identity on R.
232
16
14
12
10
0.2
0.1
0.0
0.1
0.2
Fig. 10.2. The Jackson kernel J (t) for = 4, 7, 10.
10.4
Uniform convergence and pointwise convergence
To study the convergence of f toward f , we start with the simple case f C0 (R). Let us recall that the elements of C0 (R) are uniformly continuous and bounded on R. Theorem 10.3 Let be an approximate identity on R, and let f C0 (R). Then, for all , f C0 (R) with f
C f
Moreover, f converges uniformly to f on R, i.e. lim f f

= 0.
Proof. Fix . By Lemma 9.5, f is continuous on R. Since lim|t| f (t) = 0, there is an M > 0 such that |f (t)| < for all |t| M . Choose M = M () such that | (t)| dt < .
|t|M
10.4. Uniform convergence and pointwise convergence Hence, for |t| M + M , |( f )(t)| = f

233
(t ) f ( ) d + | (t )| |f ( )| d

| |M
| |M
|s|M
| (s)| ds +
| (s)| ds
( f
+ C ).
The previous two facts show that f C0 (R). Moreover, by Lemma 9.5, f
C f
To show that f converges uniformly to f on R, given > 0, take = () > 0 such that |f (t ) f (t)| < , for all | | < and for all t R. Therefore, we have | ( f )(t) f (t) | = 2 f 2 f

( ) f (t ) f (t) d
| ( )| f (t ) f (t) d

| ( )| d +
| ( )| d + 2 f
| ( )| d
| |>
| ( )| d + C .
Pick (, ) = () so large that | ( )| d < , () and for all t R,
| |>
whenever
(). Thus, for all
| ( f )(t) f (t) | < (2 f
+ C ) .
As a special case, by using the Poisson kernel we can extend an element of C0 (R) to the upper half plane C+ . The outcome is a function which is continuous on the closed upper half plane C+ and harmonic on the open half plane C+ . Moreover, due to special properties of the Poisson kernel, our function tends uniformly to zero as the argument goes to innity in C+ .
234
Corollary 10.4 Let u C0 (R), and let y 1 u(t) dt (x t)2 + y 2 U (x + iy) = u(x) Then (a) U is harmonic on C+ , (b) U is continuous on C+ , (c) for each y 0, Uy C0 (R) and Uy
if y > 0, if y = 0.
(d) as y 0, Uy converges uniformly to u on R, (e) given > 0, there is R > 0 such that |U (x + iy)| < , whenever |x + iy| R, y 0. Proof. In Theorem 10.1 we show that U is a harmonic function on C+ . Knowing that the Poisson kernel is an approximate identity on R, properties (b), (c) and (d) are a direct consequence of Theorem 10.3. To prove (e), rst suppose that u has a compact support, say [M, M ]. Fix > 0. Then |U (x + iy)| = Hence, pick y0 > 0 such that |U (x + iy)| < for all y y0 (no restriction on x). On the other hand, for 0 < y y0 and |x| M + 2y0 u /(), we have |U (x + iy)| < Hence it is enough to take R = y0 + M + 2y0 u
y u(t) dt (x t)2 + y 2
M 1 y |u(t)| dt M (x t)2 + y 2 2M u , (x + iy C+ ). y
|t|M
y |u(t)| dt (x t)2 + y 2
/()
| |2y0 u
y d . 2
10.4. Uniform convergence and pointwise convergence For the general case, take v Cc (R) such that u v Cc (R) is dense in C0 (R)). Then, by part (c),
235 < (remember that
|U (x + iy)| = |Py (u v)(x) + (Py v)(x)| u v + |(Py v)(x)| < + |(Py v)(x)|, (x + iy C+ ). As we saw in the previous paragraph, there is an R such that |(Py v)(x)| < if |x + iy| R and y 0. Hence |U (x + iy)| < 2 for all |x + iy| R and y 0. A small modication of the proof of Theorem 10.3 provides a local version for functions which are continuous at a xed point. Theorem 10.5 Let be an approximate identity on R, and let f L (R). Suppose that f is continuous at t0 R. Then, given > 0, there exist (, t0 ) and = (, t0 ) > 0 such that | ( f )(t) f (t0 ) | < , whenever (, t0 ) and |t t0 | < . In particular, lim ( f )(t0 ) = f (t0 ).
Proof. Given > 0, there exists = (, t0 ) > 0 such that |f () f (t0 )| < if | t0 | < 2. Therefore, for all t with |t t0 | < , | ( f )(t) f (t0 ) | =

( ) f (t ) f (t0 ) d
| ( )| f (t ) f (t0 ) d
| |>
| ( )| |f (t )| + |f (t0 )| d + | ( )| d + C .
| ( )| d
2 f
| |>
Pick (, t0 ) so large that

| |>
| ( )| d < (, t0 ) and for |t t0 | < ,
whenever
(, t0 ). Thus, for
|( f )(t) f (t0 )| < (2 f
+ C ) .
236
If u L (R) then the following result is a special case of Theorem 10.5. However, due to special properties of the Poisson kernel, we are able to relax the condition u L (R) slightly. Corollary 10.6 Suppose that u L1 (dt/(1 + t2 ), i.e.

|u(t)| dt < , 1 + t2
(10.6)
and that u is continuous at t0 R. Let U (x + iy) = 1

y u(t) dt, (x t)2 + y 2
(x + iy C+ ).
Then U is harmonic on C+ and besides

zt0 zC+
lim U (z) = u(t0 ).
Remark: The condition (10.6) is satised if u Lp (R), 1 p . Proof. Without loss of generality, suppose that t0 = 0. Since u is continuous at the origin, it is necessarily bounded in a neighborhood of the origin, say [2, 2]. Let 1 if |t| , (t) = 2 |t| if |t| 2, 0 if |t| 2. Now we write u = u1 + u2 , where u1 = u and u2 = u (1 ). The function u1 is bounded on R and continuous at the origin. Hence, by Theorem 10.5,
zt0 zC+
lim (P u1 )(z) = u1 (0) = u(0).
Since U = P u1 + P u2 , it is enough to show that

zt0 zC+
lim (P u2 )(z) = 0.
But, for |x| < /2 and y > 0, we have |(P u2 )(x + iy)| = y u2 (t) dt (x t)2 + y 2 |t| y |u(t)| dt (|t| /2)2 |t|
|t|
4y
|u(t)| dt t2

4(1 + 1/ 2 )
|u(t)| dt 1 + t2
y,
10.4. Uniform convergence and pointwise convergence which shows that (P u2 )(z) 0, as z 0.
237
Exercises
Exercise 10.4.1 f (t) = Let f C0 (R), and let

sin( (t )) (t )
f ( ) d,
(t R).
Show that, for each > 0, with
f C0 (R)
f and
lim
f f
= 0.
Hint: Use Exercise 10.3.5 and Theorem 10.3. Exercise 10.4.2 Let f C0 (R), and let

1 f (t) = Show that, for each > 0, with
e(t )
f ( ) d,
(t R).
f C0 (R)
f and
0
lim f f
= 0.
Hint: Use Exercise 10.3.3 and Theorem 10.3. Exercise 10.4.3 Let be an approximate identity on R, and let f be uniformly continuous and bounded on R. Show that, for each , f is also uniformly continuous and bounded on R, and moreover f converges uniformly to f on R. Remark: If f (x) = sin x, then (Py f )(x) = ey sin x. Clearly, for each y > 0, Py f is uniformly continuous and bounded on R, and Py f converges uniformly to f on R, as y 0. However, for all y > 0, Py f C0 (R). Compare with Corollary 10.4.
238
10.5
According to (9.4), , the convolution of an integrable function and a Borel measure , is a well-dened integrable function. We now explore the relation between and where ranges over the elements of an approximate identity. Theorem 10.7 Let be an approximate identity on R, and let M(R). Then, for all , L1 (R) with and sup 1 .
1
Moreover, the measures d (t) = ( )(t) dt converge to d(t) in the weak* topology, i.e.
lim

(t) ( )(t) dt =
(t) d(t)
R
for all C0 (R). Proof. The verication of the rst and second assertions is similar to the proof of Lemma 9.5. Indeed, according to (9.4), for each , we have |( )(t)|
R
| (t )| d||( )
almost everywhere on R, and thus, by Fubinis theorem,

|( )(t)| dt
| (t )| d||( ) | (t )| dt
dt
d||( )
d||( ) = C .
To show the weak* convergence, let C0 (R), and let (t) = (t). Then, C0 (R) and by Fubinis theorem,

(t) ( )(t) dt
(t)
R
(t ) d( )
dt
=
R
(t ) (t) dt
d( ) d( ) d( )
=
R
( t) (t) dt ( t) (t) dt
=
R
=
R
( )( ) d( ).
10.5. Weak* convergence of measures
239
Theorem 10.3 ensures that ( )( ) converges uniformly to ( ) on R. Since || is a nite Borel measure on R, we thus have
lim

(t) ( )(t) dt
= =
lim
R
( )( ) d( )
( ) d( )
R
=
R
( ) d( ).
Since
(t) ( )(t) dt (sup 1 )
and ( )(t) dt converges to d(t) in the weak* topology, we thus have

R
(t) d(t) (sup 1 )
for all C0 (R). Hence, by the Riesz representation theorem, sup 1 .
If { } is a positive approximate identity on R, then C = 1 and we obtain = sup 1 .
We can even show that sup can be replaced by lim in the preceding identity. As a special case, by Theorem 10.1, the Poisson kernel extends to a harmonic function U on C+ such that the measures U (t + iy) dt are uniformly bounded and converge to d(t), as y 0, in the weak* topology. This observation leads us to the uniqueness theorem, which is an essential result in harmonic analysis. Corollary 10.8 Let M(R), and let U (x + iy) = 1

y d(t), (x t)2 + y 2
(x + iy C+ ).
(10.7)
Then U is harmonic on C+ , and = sup Uy

y>0 1
= lim Uy
y0
1.
Moreover, the measures dy (t) = U (t + iy) dt converge to d(t), as y 0, in the weak* topology, i.e.
y0
lim
(t) dy (t) =
(t) d(t)
R
for all C0 (R).
240
We give two versions of the uniqueness theorem. However, both results are a direct consequence of the relation = supy>0 Uy 1 , where U is given by (10.7). Corollary 10.9 (Uniqueness theorem) Let M(R). Suppose that 1

y d(t) = 0 (x t)2 + y 2
for all x + iy C+ . Then = 0. Corollary 10.10 (Uniqueness theorem) Let M(R). Suppose that (t) = 0 for all t R. Then = 0. In particular, if f L1 (R) and f (t) = 0, for all t R, then we have f = 0. Proof. Dene U by (10.7) and note that, by Theorem 10.1,
U (x + iy) =
e2y|t| (t) ei 2xt dt = 0
for all x + iy C+ . The uniqueness theorem combined with Lemma 9.1 says that the map M(R) is one-to-one. C(R) L (R)
Exercises
Exercise 10.5.1 Let M(R) and let y0 > 0. Suppose that 1

y0 2 d(t) = 0 (x t)2 + y0
for all x R. Show that = 0. Exercise 10.5.2 Let M(R). Suppose that 1

xt d(t) = 0 (x t)2 + y 2
for all x + iy C+ . Can we conclude that = 0? What if our assumption only holds on the horizontal line y = y0 ?
10.6. Convergence in norm Exercise 10.5.3 Let M(R), and let
241
f (t) =
sin( (t )) (t ) f L1 (R)
d( ),
(t R).
Show that, for each > 0,
with f
1
and the measures f (t) dt converge to d(t) in the weak* topology, i.e.

lim
(t) f (t) dt =
(t) d(t)
R
for all C0 (R). Hint: Use Exercise 10.3.5 and Theorem 10.7. Let M(R), and let

Exercise 10.5.4
1 f (t) = Show that, for each > 0,
e(t )
d( ),
(t R).
f L1 (R) with f
1
and the measures f (t) dt converge to d(t) in the weak* topology, i.e.
0
lim
(t) f (t) dt =
(t) d(t)
R
for all C0 (R). Hint: Use Exercise 10.3.3 and Theorem 10.7.
10.6
Convergence in norm
In this section we use the fact that Cc (R), the space of continuous functions of compact support, is dense in Lp (R), 1 p < . This assertion is not true if p = and that is why the following result does not hold for the elements of L (R).
242
Theorem 10.11 Let be an approximate identity on R, and let f Lp (R), 1 p < . Then, for all , f Lp (R) and f Moreover, lim f f
p p
C f
p.
= 0.
Proof. The rst two assertions are proved in Lemma 9.4. To prove the last identity, we have ( f )(t) f (t) =

( ) (f (t ) f (t)) d,
(t R).
Hence, by Minkowkis inequality (see Section A.6), f f

p
| ( )| f f
d,
where f (t) = f (t ). It is enough to show that

0
lim f f
= 0,
since then we use the same technique applied in the proof of Theorem 10.5 to deduce that f f p 0. To do so, given > 0, pick Cc (R) such that f p < . Hence, f f
p
f p + p + f 2 f p + + p
p
2 + p . As we saw in the proof of Lemma 9.5, is > 0 such that, for | | < , f f
p
0, as 0. Hence, there
3.
As in the preceding cases, we use the Poisson kernel, to extend u Lp (R) to the harmonic function U = P u on C+ whose integral means Uy p are uniformly bounded and, y 0, Uy converges to u in Lp (R). Corollary 10.12 Let u Lp (R), 1 p < , and let U (x + iy) = 1

y u(t) dt, (x t)2 + y 2
(x + iy C+ ).
Then U is harmonic on C+ , sup Uy

y>0 p
= lim Uy
y0
= u
and
y0
lim Uy u
= 0.
10.7. Weak* convergence of bounded functions
243
Exercises
Exercise 10.6.1 f (t) = Let f Lp (R), 1 p < , and let

sin( (t )) (t )
f ( ) d,
(t R).
f Lp (R) f
p
f
p
and
lim
f f
= 0.
Hint: Use Exercise 10.3.5 and Theorem 10.11. Exercise 10.6.2 Let f Lp (R), 1 p < , and let

e(t )
f ( ) d,
(t R).
f Lp (R) f
p
f
p
and
0
lim f f
= 0.
Hint: Use Exercise 10.3.3 and Theorem 10.11.
10.7
Weak* convergence of bounded functions
Since Cc (R) is not dense in L (R), the results of Section 10.6 are not entirely valid if p = . Nevertheless, a slightly weaker version also holds in this case. Theorem 10.13 Let be an approximate identity on R, and let f L (R). Then, for all , f L (R) C(R) with f and f

C f
sup f
Moreover, f converges to f in the weak* topology, i.e. lim

(t) ( f )(t) dt =
(t) f (t) dt
for all L1 (R).
244
Proof. The rst two assertions are proved in Lemma 9.5. Let L1 (R), and let (t) = (t). Then, by Fubinis theorem,

(t) ( f )(t) dt
(t)

(t ) f ( ) d
dt
(t ) (t) dt
f ( ) d f ( ) d f ( ) d
( t) (t) dt ( t) (t) dt
( )( ) f ( ) d.
Theorem 10.11 ensures that ( )( ) converges in L1 (R) to ( ). Since f is a bounded function, we thus have
lim

(t) ( f )(t) dt
= =
lim

( )( ) f ( ) d
( ) f ( ) d ( ) f ( ) d.
= Since

(t) ( f )(t) dt (sup f
1,
and f converges to f in the weak* topology,

( ) f ( ) d (sup f
for all L1 (R). Hence, by Rieszs theorem, f
sup f
If { } is a positive approximate identity, then C = 1 and we have f
= sup f
= lim f
Corollary 10.14 Let u L (R), and let U (x + iy) = 1

y u(t) dt, (x t)2 + y 2
(x + iy C+ ).
10.7. Weak* convergence of bounded functions Then U is bounded and harmonic on C+ , and sup Uy
y>0
245
= lim Uy
y0
= u
Moreover, as y 0, Uy converges to u in the weak* topology, i.e.

y0
lim
(t) U (t + iy) dt =

(t) u(t) dt
for all L1 (R). Corollary 10.15 Let be continuous and bounded on C+ and subharmonic on C+ . Then (x + iy) Proof. Let U (x + iy) = 1

y (t) dt, (x t)2 + y 2
(x + iy C+ ).
y (t) dt, (x t)2 + y 2
(x + iy C+ ).
By Corollaries 10.6 and 10.14, U is continuous and bounded on C+ and harmonic on C+ with lim U (z) = (t), (t R).
zt zC+
Let = U . Then is bounded and subharmonic on C+ , and

zt zC+
lim (z) = 0,
(t R).
Hence, by Corollary 4.5, (z) 0 for all z C+ .
Exercises
Exercise 10.7.1 f (t) = Let f L (R), and let

sin( (t )) (t )
f ( ) d,
(t R).
f L (R) f
246
and, as , f converges to f in the weak* topology, i.e.

lim
(t) f (t) dt =
(t) f (t) dt
for all L1 (R). Hint: Use Exercise 10.3.5 and Theorem 10.13. Exercise 10.7.2 Let f L (R), and let

e(t )
f ( ) d,
(t R).
f L (R) f
and, as 0, f converges to f in the weak* topology, i.e.

0
lim
(t) f (t) dt =
(t) f (t) dt
for all L1 (R). Hint: Use Exercise 10.3.3 and Theorem 10.13.
Chapter 11
Harmonic functions in the upper half plane

11.1 Hardy spaces on C+
The family of all complex harmonic functions in the upper half plane is denoted by h(C+ ), and H(C+ ) is its subset containing all analytic functions on C+ . Let U h(C+ ), and write
U and
= sup
0<y<
|U (x + iy)| dx
1 p
(0 < p < ),
= sup | U (z) |.
zC+
Then, for 0 < p , we dene the harmonic family hp (C+ ) = and the analytic subfamily H p (C+ ) = F H(C+ ) : F
p
U h(C+ ) :
<
< .
These are Hardy spaces of the upper half plane. We will also need h(C+ ), the family of all functions U such that U is harmonic in some open half plane containing C+ . It is straightforward to see that hp (C+ ) and H p (C+ ), 1 p , are normed vector spaces. We will see that h1 (C+ ) and hp (C+ ), 1 < p , are actually Banach spaces respectively isomorphic to M(R) and Lp (R), 1 < p . Similarly, H p (C+ ), 1 p , is a Banach space isomorphic to a closed subspace of Lp (R) denoted by H p (R). 247
248
Chapter 11. Harmonic functions in the upper half plane
Using these new notations, Corollaries 10.8, 10.12 and 10.14 can be rewritten as follows: u Lp (R) = P u hp (C+ ), for each 1 p , and M(R) = P h1 (C+ ). In this chapter, we start with a harmonic function in hp (C+ ), 1 p , and show that it is representable as P u or P .
11.2
Poisson representation for semidiscs
To obtain the Poisson representation for hp (C+ ) classes, we need the Poisson integral formula for semidiscs. This formula recovers a harmonic function inside the open semidisc SR = { x + iy : y > 0 and x2 + y 2 < R } from its boundary values on the semicircle CR = { x + iy : y 0 and x2 + y 2 = R } and on the line segment [R, R]. Lemma 11.1 Let U be harmonic on a domain containing the closed semidisc S R . Then U (z) = + 1 1
R R 0
y xy t + y(R2 + x t) 2 (t x)2 + y 2 (R + x t)2 + (y t)2 R2 r 2 R2 + r2 2rR cos( ) R2 r 2 2 R + r2 2rR cos( + )
U (t) dt
U (Rei ) d,
for all z = x + iy = rei SR . Proof. Without loss of generality, assume that U is real. Let V be a harmonic conjugate of U , and let F = U + iV . Then, by the Cauchy integral formula, for each z SR , we have I1 I2 I3 I4 = = = = 1 i 2 1 2i 1 2i 1 2i F () d = F (z), SR z F () d = 0, SR z F () d = 0, 2 z SR R / F () d = 0, 2 SR R /z
11.2. Poisson representation for semidiscs where SR represents the boundary of SR . Hence, F (z) = I1 I2 I3 + I4 . Taking the real part of both sides gives the desired formula.
249
Exercises
Exercise 11.2.1 Let be a domain in C and x z0 . A function G(z, z0 ) dened on \ {z0 } and satisfying the following four properties is called the Green function for corresponding to the point z0 : (i) G(z, z0 ) > 0, for all z \ {z0 }; (ii) G is harmonic on \ {z0 }; (iii) G(z, z0 ) + log |z z0 | has a removable singularity at z0 (and thus it represents a harmonic function on ); (iv) G(z, z0 ) 0 as z tends to any boundary point of . First, verify that if G exists then it is unique. Then show that the Green function of the disc = DR is G(z, z0 ) = log R 2 z0 z , R(z z0 )
and for the semidisc SR , the Green function is given by G(z, z0 ) = log R2 z0 z R(z z0 ) . R(z z0 ) R2 z0 z (11.1)
Hint: To prove uniqueness, use the maximum principle. Exercise 11.2.2 [Greens formula] Let be a domain in C. Suppose that U is harmonic on a domain containing . Show that, for all z0 , U (z0 ) = 1 2 U ()
G (, z0 ) ds(). n
Remark: G/n represents the derivative of G in the direction of the outward normal vector n. Exercise 11.2.3 Theorem 11.1. Use Exercise 11.2.2 and (11.1) to provide another proof for
250
11.3
Poisson representation of h(C+ ) functions
In studying the boundary values of harmonic functions in the upper half plane C+ , the best possible assumption for the behavior on R is to assume that our function is actually dened on a half plane containing C+ . This class was already denoted by h(C+ ). Theorem 11.2 Let U h(C+ ) and suppose that U is bounded on C+ . Then U (x + iy) = 1

y U (t) dt, (x t)2 + y 2
(x + iy C+ ).
(11.2)
Proof. Since U h(C+ ), by denition, there exists Y < 0 such that U is harmonic on the half plane { z > Y }. Fix z = x + iy = rei C+ . By Lemma 11.1, for each R > r, we have U (z) = 1
0
R R
y (R2 y/r2 ) 2 + y2 2 x/r 2 ) t)2 + (R2 y/r 2 )2 (x t) ((R 2rR(R2 r2 ) sin sin
2
U (t) dt U (Rei ) d.
(R2 + r2 ) cos 2Rr cos
(R2 r2 ) sin
But, as R , (R2 y/r2 ) = O(1/R2 ) ((R2 x/r2 ) t)2 + (R2 y/r2 )2 and 2rR(R2 r2 ) sin sin (R2 + r2 ) cos 2Rr cos
2
(R2 r2 ) sin
= O(1/R).
Let R . Hence, by the dominated convergence theorem, we obtain U (z) = 1

y U (t) dt, (x t)2 + y 2
(z = x + iy C+ ).
To extend the previous representation theorem for other classes, we need to show that each element of hp (C+ ) is bounded if we stay away from the real line. Since the disc {|z| r < 1} is compact and continuous functions are automatically bounded on compact sets, the need for such a result was not felt in studying hp (D) spaces. Lemma 11.3 Let U hp (C+ ), 1 p < . Then |U (x + iy)| (2/)1/p U p , y 1/p (x + iy C+ ).
11.3. Poisson representation of h(C+ ) functions
251
Proof. Fix z = x + iy C+ and let 0 < < y. Then, for each r < , by Theorem 3.4, we have U (z) = Hence, by Hlders inequality, o |U (z)|p 1 2

1 2
U (z + reit ) dt.
|U (z + reit )|p dt.
(11.3)
(As a matter of fact, we just showed that |U |p is subharmonic.) Integrating both sides with respect to rdr from 0 to gives |U (z)|p 2 /2 1 2
0
|U (z + reit )|p rdrdt.
Since the disc D(z, ) is a subset of the strip (, ) [y , y + ], we thus have |U (z)|p 2 /2 = Therefore, |U (z)|p Now, let y. The following result is obvious if p = . For other values of p, it is a direct consequence of the preceding lemma. Corollary 11.4 Let U hp (C+ ), 1 p . Let > 0 and dene U (z) = U (z + i). Then U h(C+ ) and U is bounded on C+ . The main ingredient in the proof of Lemma 11.3 is the inequality (11.3), i.e. the fact that |U |p , 1 p < , is a subharmonic function. If F is analytic then |F |p is subharmonic for each 0 < p < . In other words, the inequality |F (z)|p 1 2

1 2 1 2
y+ y y+ y
|U (x + iy )|p dx dy U
p p
dy
Uy
p p
2 U p p .
|F (z + reit )|p dt
holds for all 0 < p < . Therefore, Lemma 11.3 can be generalized slightly for analytic functions. Lemma 11.5 Let F H p (C+ ), 0 < p < . Then |F (x + iy)| (2/)1/p F p , y 1/p (x + iy C+ ).
252
11.4
Poisson representation of hp (C+ ) functions (1 p )
Using Theorem 11.2 and Corollary 11.4 we are now able to obtain the Poisson representation for hp (C+ ) classes, 1 p . Theorem 11.6 Let U hp (C+ ), 1 < p . Then there exists a unique u Lp (R) such that U (x + iy) = and U
p
y u(t) dt, (x t)2 + y 2 = u p.
(x + iy C+ ),
Remark: Compare with Corollaries 10.6 and 10.14. Proof. The uniqueness is a consequence of Corollary 10.9. Let > 0 and dene U (z) = U (z + i). Then, by Corollary 11.4, U is bounded on C+ and U h(C+ ). Hence, by Theorem 11.2, U (x + iy + i) = 1

y U (t + i) dt (x t)2 + y 2
for all x + iy C+ . Let 0. The rest of the proof is exactly like that given for Theorem 3.5. A small modication of the preceding proof, as explained similarly before Theorem 3.7 for the case of the open unit disc, yields the Poisson representation for h1 (C+ ) functions. Theorem 11.7 Let U h1 (C+ ). Then there exists a unique M(R) such that 1 y U (x + iy) = d(t), (x + iy C+ ), (x t)2 + y 2 and U
1
= .
Remark: Compare with Corollary 10.8.
11.5. A correspondence between C+ and D
253
Exercises
Exercise 11.4.1 Let U hp (C+ ), 1 p < . Show that the subharmonic function |U |p has a harmonic majorant. Hint: Apply Theorem 11.7. Exercise 11.4.2 Find U h(C+ ) such that the subharmonic function |U |p has a harmonic majorant, but U hp (C+ ). Hint: Consider U (x + iy) = y h1 (C+ ). Remark: Compare with Exercises 4.1.7 and 11.4.1.
11.5
A correspondence between C+ and D

w= iz , i+z 1w , 1+w
The conformal mapping
or equivalently z=i gives the correspondence C+ z
D w
between the points z = x + iy of the upper half plane and w = rei of the unit disc. Moreover, on the boundary, the formulas ei = and t=i provide the correspondence R t T \ {1} ei it i+t
1 ei 1 + ei
between R and T \ {1}. The fact that 1 is not included is essential. Using the last two relations, we are able to establish the correspondence B(R) B( T \ {1} ) A B between the Borel subsets of R and the Borel subsets of T \ {1} by dening A= t : it B , i+t
254 or equivalently
B=
ei : i
1 ei A . 1 + ei
Based on this observation, we make a connection between Borel measures dened on R and Borel measures dened on T \ {1}. Let be a positive Borel measure on T \ {1}. Dene its pullback on R by (A) = (B), (11.4)
where A is an arbitrary Borel subset of R, and B is related to A as explained above. Then is a positive Borel measure on R, and the identity (A) = (B) can be rewritten as A (t) d(t) = A i 1 ei 1 + ei it i+t d(ei ),
T\{1}
or
T\{1}
B (ei ) d(ei ) =
d(t).
Taking positive linear combinations shows that both identities are valid for positive measurable step functions. Hence, by a standard limiting process, we obtain 1 ei d(ei ) (t) d(t) = i (11.5) 1 + ei R T\{1} and (ei ) d(ei ) =
T\{1} R
it i+t
d(t)
(11.6)
for all positive measurable functions and . We started with a measure on T \ {1} and then dened on R. Clearly, the procedure is symmetric and we can start with and then dene . The relations (11.5) and (11.6) still remain valid. A very special, but important, case is the Lebesgue measure d on T. Using dierential calculus techniques, the relation ei = (i t)/(i + t) immediately implies that 2 dt (11.7) d = 1 + t2 and thus (11.6) is written as

(ei ) d =
it i+t
2dt . 1 + t2
(11.8)
11.6. Poisson representation of positive harmonic functions
255
11.6
Poisson representation of positive harmonic functions
In Section 3.5, we saw that positive harmonic functions in the unit disc are those elements of h1 (D) which are generated by positive Borel measures. As the elementary example U (x + iy) = y (11.9) shows, in the upper half plane, positive harmonic functions are not necessarily in h1 (C+ ). Another observation is based on the contents of Section 10.2. If is a positive Borel measure on R such that d(t) < , 1 + t2 y d(t) (t x)2 + y 2 (11.10)
then U (z) =
R
is a well-dened positive harmonic function on C+ . In this section, we show that any positive harmonic function in the upper half plane is a positive linear combination of (11.9) and (11.10). Theorem 11.8 (Herglotz) Let U be a positive harmonic function on C+ . Then the limit U (iy) = lim y y exists and [0, ). Moreover, there exists a positive Borel measure on R satisfying d(t) < 2 R 1+t such that U (x + iy) = y + Proof. Let U(w) = U i 1
R
y d(t), (x t)2 + y 2 1w , 1+w
(x + iy C+ ).
(w D).
Then U is a positive harmonic function on D. Hence, by Theorem 3.8, there is a nite positive Borel measure on T such that U(rei ) = Thus, U(rei ) = 1 r2 ({1}) + 1 + r2 + 2r cos 1 r2 d(ei ). 1 + r2 2r cos( ) 1 r2 d(ei ), 1 + r2 2r cos( ) (rei D).
T\{1}
256
To apply (11.6), we need two elementary identities which are easy to verify. If z = x + iy C+ , t R, w = rei D and ei T, and they are related as explained in the preceding section, then we have 1 r2 =y 1 + r2 + 2r cos and 1 r2 y = (1 + t2 ). 1 + r2 2r cos( ) (x t)2 + y 2
Therefore, by (11.6), U (x + iy) = ({1}) y +

R
y (1 + t2 ) d(t), (x t)2 + y 2
where is dened by (11.4). Let = ({1}) and let d(t) = (1 + t2 ) d(t). Then clearly [0, ) and is a positive Borel measure on R such that 1 and U (x + iy) = y + Finally, since U (iy) =+ y
R
d(t) = (R) = (T \ {1}) < 1 + t2
y d(t), (x t)2 + y 2 t2 + 1 d(t), t2 + y 2
(x + iy C+ ).
and (R) < , by the dominated convergence theorem, we see that

y
lim
U (iy) = . y
Let be a Borel measure satisfying d||(t) < , 1 + t2
and let C. Let U (x + iy) = y + 1

R
y d(t), (x t)2 + y 2
(x + iy C+ ).
Then U is a harmonic function on the upper half plane. As we saw in Theorem 11.8, a positive harmonic function on C+ is a special case obtained by a positive
11.6. Poisson representation of positive harmonic functions
257
constant and a positive measure . The family discussed in Corollary 10.8 is an even smaller subclass. Nevertheless, we expect the horizontal sections Uy to somehow converge to as y 0. Since satises a weaker condition in this case, we have to consider a more restrictive type of convergence. This is done by considering Cc (R), the class of continuous functions of compact support. Theorem 11.9 Let C and let be a Borel measure on R satisfying
R
d||(t) < . 1 + t2
Let U (x + iy) = y + Then

y0
y d(t), (x t)2 + y 2
(x + iy C+ ).
lim
(t) U (t + iy) dt =
R R
(t) d(t)
for all Cc (R). Proof. Since has compact support, A=

R
(x) dx
is nite. Hence, (x) U (x + iy) dx

R
= A y +
R
(x)
R
Py (x t) d(t)
dx
= A y +
R
(Py )(t) d(t).
By Corollary 10.4, Py converges uniformly and boundedly to . Let supp [M, M ]. Then, for |x| 2M and 0 < y < 1, |(Py )(x)|
M M M
y dt (x t)2 + y 2
dt (x t)2 M 2M . x2 M 2 Hence, for all x R and all 0 < y < 1, |(Py )(x)| C , 1 + x2
where C = C() is an absolute constant. Therefore, by the dominated convergence theorem,

R
(Py )(t) d(t)
(t) d(t)
R
as y 0.
258
Exercises
Exercise 11.6.1 Let U be a real harmonic function in h1 (C+ ). Show that U is the dierence of two positive harmonic functions. Exercise 11.6.2 Let U1 and U2 be two positive harmonic functions on C+ . Let U = U1 U2 . Can we conclude that U h1 (C+ )? Remark: Compare with Exercises 3.5.2 and 11.6.1.
11.7
Vertical limits of hp (C+ ) functions (1 p )
Using the conformal mapping given in Section 11.5, we are able to exploit Fatous theorem (Theorem 3.12) and obtain a similar result about the vertical limits of harmonic functions dened in the upper half plane. The following result is rather general and applies for the elements of hp (C+ ), 1 p , as a special case. Theorem 11.10 Let be a Borel measure on R such that d||(t) < , 1 + t2 (11.11)
and let U (x + iy) = 1

R
y d(t), (x t)2 + y 2
(x + iy C+ ).
Let x0 R. Suppose that A = lim exists and is nite. Then

y0 s0
([x0 s, x0 + s]) 2s
lim U (x0 + iy) = A.
(11.12)
Proof. Without loss of generality, assume that 0, since otherwise, using Hahns decomposition theorem, we can write = (1 2 ) + i(3 4 ) where each k is a positive Borel measure satisfying (11.11). Without loss of generality, assume that x = 0. Let U(w) = U Then, on the one hand,
y0
1w , 1+w
(w D).
lim U (iy) = lim U(r)

r1
(11.13)
11.7. Vertical limits of hp (C+ ) functions (1 p ) and, on the other hand, since y 1 r2 = cos2 ( /2), 2 + y2 2 2r cos( ) (x t) 1+r we have U(rei ) = 1
T\{1}
259
1 r2 cos2 ( /2) d(ei ), 1 + r2 2r cos( )
where corresponds to as explained in (11.4). Let d(ei ) = 2 cos2 ( /2) d(ei ). Hence, U(rei ) = and 1 2s
s s
1 2 1 s 1 s 1 s
T s s s s s s
1 r2 d(ei ) 1 + r2 2r cos( ) cos2 ( /2) d(ei ) cos2 ( /2) 1 d(ei ) + cos2 ( /2) 1 d(ei ) + 1 s 1 s
s s s
d(ei )
= = =
d(ei ) d(t),
1 eis . 1 + eis Since cos2 ( /2) 1 = O( 2 ) and s /s 1/2, as s 0, we thus have s =i

s0
where
lim
1 2s
s s
d(ei ) = lim
s 0
1 2s
s s
d(t) = A.
s
Since, by Theorem 3.12,

r1
lim U(r) = lim
s0
1 2s
d(ei ),
the relation (11.13) implies that limy0 U (iy) = A. For each Borel measure satisfying (11.11), we know that, for almost all x R, ([x s, x + s]) lim = (x) s0 2s exists and (x) is nite. Hence, by Theorem 11.10,
y0
lim
y d(t) = (x) (x t)2 + y 2
for almost all x R. A special case of this fact is mentioned below.
260
Corollary 11.11 Let
|u(t)| dt < 1 + t2
(11.14)
and let U (x + iy) = Then

y0
y u(t) dt, (x t)2 + y 2 lim U (x + iy) = u(x)
(x + iy C+ ).
for almost all x R. Note that (11.14) is fullled if u Lp (R), 1 p .
Exercises
Exercise 11.7.1 Let be a real signed Borel measure on R such that d||(t) < , 1 + t2

R
y d(t), (x t)2 + y 2
(x + iy C+ ).
Let x0 R and suppose that (x0 ) = +. Show that

y0
lim U (x0 + iy) = +.
Exercise 11.7.2
Let be a Borel measure on R such that d||(t) < , 1 + t2

R
y d(t), (x t)2 + y 2
(x + iy C+ ).
Show that, for almost all x0 R,

zx0 zS (x0 )
lim
U (z) = (x0 )
11.7. Vertical limits of hp (C+ ) functions (1 p ) where S (x0 ), 0 < /2, is the Stoltz domain S (x0 ) = { x + iy C+ : |x x0 | (tan ) y }.
261
(See Figure 11.1.) Remark: We say that (x0 ) is the nontangential limit of U at the point x0 .
Fig. 11.1. The Stoltz domain S (x0 ).
Exercise 11.7.3
Let be a positive Borel measure on R such that d(t) < , 1 + t2

R
y d(t), (x t)2 + y 2
(x + iy C+ ).
Let x0 R and suppose that (x0 ) = +. Show that, for any 0,

zx0 zS (x0 )
lim
U (z) = +.
Exercise 11.7.4
Construct a real signed Borel measure such that d||(t) < 1 + t2
262 and
(0) = +. But
t0
lim U (t + it) = +,
where U (x + iy) = 1 y d(t), (x t)2 + y 2 (x + iy C+ ).
Remark: Compare with Exercises 11.7.1 and 11.7.3.
Chapter 12
The Plancherel transform

12.1 The inversion formula
The representation formulas given in Theorem 10.1 and Corollary 10.2 represent a harmonic function in h1 (C+ ) in terms of the Fourier transform of a boundary measure or function on R. Our goal is to generalize these results for other hp (C+ ) classes. Of course, the rst and most important step would be to generalize the denition of Fourier transform for other Lp (R) spaces. Then we will be able to obtain further representations of elements of hp (C+ ). The case p = 1 of Corollary 10.12 is of particular interest and helps us to say a bit more about the Fourier transform of functions in L1 (R). Theorem 12.1 (Inversion formula) Let f L1 (R) and suppose that f L1 (R). Then, for almost all t R,
f (t) =
f ( ) ei 2t d.
Proof. Let Fy (x) = 1

y f (t) dt, (x t)2 + y 2 lim Fy f
(x + iy C+ ).
By Corollary 10.12, for each y > 0, Fy L1 (R) and moreover

y0 1
= 0.
A convergent sequence in L1 (R) has a subsequence converging at almost all points of R. Hence, there is yn > 0, yn 0, such that
n
lim Fyn (x) = f (x)
for almost all x R. On the other hand, by Corollary 10.2, the identity Fy (x) = e2y|t| f (t) ei 2xt dt 263
264
Chapter 12. The Plancherel transform
holds for all x R and y > 0. Now, the assumption f L1 (R) together with the dominated convergence theorem imply that
y0
lim Fy (x) =
f (t) ei 2xt dt
for all x R. Therefore, we immediately obtain the inversion formula. Writing the inversion formula as
f (t) =
f ( ) ei 2(t) d,
the last integral can be interpreted as the Fourier transform of f , and we obtain f (t) = f (t), (t R), (12.1)
provided that f, f L1 (R). Suppose that f L1 (R), x y > 0, and consider
Fy (x) =
e2y|t| f (t) ei 2xt dt.
By Corollary 10.12, Fy L1 (R). Hence, according to (12.1), we have Fy (t) = e2y|t| f (t). Clearly Fy C0 (R), for all y > 0. On the other hand, Fy f by Lemma 9.1, lim Fy f = 0.
y0 1
0 and thus,
Therefore, f C0 (R). This is the RiemannLebesgue lemma on the real line. The RiemannLebesgue lemma combined with the uniqueness theorem (Corollary 10.10) say that the mapping L1 (R) C0 (R) f f is well-dened and one-to-one. The following result is a simple consequence of Corollaries 10.2, 10.6 and (11.11). However, this simple observation has an important application in extending the denition of Fourier transform to Lp (R), 1 < p 2, spaces. Corollary 12.2 Let f L1 (R). Suppose that f 0 and that f is continuous L1 (R) and at 0. Then f
f (t) =
f ( ) ei 2t d
12.1. The inversion formula for almost all t R. In particular,
265
f (0) =
f ( ) d.
Proof. By Corollary 10.2, 1

y f (t) dt = (x t)2 + y 2

e2y |t| f (t) ei 2xt dt
(12.2)
for all y > 0 and all x R. As a special case, on the imaginary axis we have 1 y f (t) dt = 2 + y2 t

e2y |t| f (t) dt
for all y > 0. Let y 0. Then the monotone convergence theorem ensures that the right side tends to f (t) dt,
and Corollary 10.6 says that the left side converges to f (0). Hence
f (0) =
f (t) dt,
which incidentally also implies that f L1 (R). Knowing that f L1 (R), the inversion formula (Theorem 12.1) applies and ensures that the rst identity holds for almost all t R.
Exercises
Exercise 12.1.1 Let f L1 (R), and let
f (t) =
1 | |/ f ( ) ei 2t d, f L1 (R) lim f
1
(t R).
1,
f (t) = and
1 |t|/ f (t) 0 f f
1
if |t| , if |t|
= 0.
266 Exercise 12.1.2
Chapter 12. The Plancherel transform Let f L1 (R), and let
2 e f ( ) ei 2t d,
f (t) =
(t R).
f L1 (R) f
1
f
2
1,
f (t) = et f (t) and

0
lim f f
= 0.
Hint: Use Exercises 9.3.1 and 10.6.2. Exercise 12.1.3 We saw that the map L1 (R) f C0 (R) f
is injective. However, show that it is not surjective.
12.2
The FourierPlancherel transform
Using Corollary 12.2, we show that the Fourier transform maps L1 (R) L2 (R) into L2 (R). This is the rst step in generalizing the denition of Fourier transform. Theorem 12.3 (Plancherel) Let f L1 (R) L2 (R). Then f L2 (R) and moreover f 2 = f 2. Proof. Let g(x) = f (x) and let h = f g. By Youngs inequality (Lemma 9.5), h L1 (R) C(R) and, by Lemma 9.3, h = f g = |f |2 0.
Hence, by Corollary 12.2, h L1 (R), which is equivalent to f L2 (R), and moreover h(t) dt. h(0) =
12.2. The FourierPlancherel transform
267
Let us calculate both sides according to their denitions. On the one hand, we have |f (t)|2 dt = f 2 h(t) dt = 2

and, on the other hand,
h(0) =
f (t) g(0 t) dt =
f (t) f (t) dt = f
2
2 2.
Comparing the last three identities implies f
= f
2.
Theorem 12.3 is a valuable result. Since Cc (R) L1 (R) L2 (R), the latter is also dense in L2 (R). This fact enables us to extend the denition of Fourier transform to L2 (R). Indeed, let f L2 (R). Pick any sequence fn L1 (R) L2 (R), n 1, such that lim fn f 2 = 0.
n
Then (fn )n1 is a Cauchy sequence in L2 (R) and, by Theorem 12.3, we have fn fm
2
= fn fm
= fn fm 2 ,
and thus (fn )n1 is also a Cauchy sequence in L2 (R). Since L2 (R) is complete,
n
lim fn
exists in L2 (R). If (gn )n1 is another sequence in L1 (R) L2 (R) satisfying similar properties, then limn gn also exists in L2 (R). However, again by Theorem 12.3, we have fn gn
2
fn gn fn f
= fn gn
2
2 + gn f
0,
and thus
n 2
lim gn = lim fn .
n
Therefore, as an element of L (R), the limit of fn is independent of the choice 1 of sequence as long as the sequence is in L (R) L2 (R) and converges to f in L2 (R). Hence, we dene the FourierPlancherel transform of f L2 (R) by Ff = lim fn ,
n
where fn is any sequence in L1 (R) L2 (R) satisfying

n
lim
fn f
= 0.
In particular, if f L1 (R) L2 (R), we can take fn = f , and thus we have Ff = f . (12.3)
268
In other words, the two denitions coincide on L1 (R) L2 (R). Therefore, with out any ambiguity we can also use the notation f for the FourierPlancherel transform of functions in L2 (R). Finally, by Theorem 12.3, f
2
= lim
fn
= lim
fn
= f
(12.4)
for all f L2 (R). Thus, the FourierPlancherel transformation F : L2 (R) L2 (R) f f is an operator on L2 (R) which preserves the norm. For an arbitrary function f L2 (R) we usually take f (t) if |t| n, fn (t) = 0 if |t| > n. Hence, we have f (t) = l.i.m. n
n n
f ( ) ei 2t d,
where l.i.m. stands for limit in mean and implies that the limit is taken in L2 (R). Let us discuss a relevant example. The conjugate Poisson kernel is in L2 (R). We evaluate its FourierPlancherel transform. Fix x + iy C+ , and let f (t) = 1 xt , (x t)2 + y 2 (t R).
To obtain the FourierPlancherel transform of f , we rst calculate the Fourier transform of 1 x if |t x| n, (x )2 + y 2 fn (t) = 0 if |t x| > n. Hence, fn (t) = = = 1 1 x ei 2t d 2 2 xn (x ) + y n ei 2t(x ) d 2 + y2 n
n n x+n
ei 2xt
ei 2t d. 2 + y2
12.2. The FourierPlancherel transform Suppose that t > 0 and n > y. Hence, by Cauchys integral formula, fn (t) ei 2xt ei 2t d 2 + y2
Cn
269
Cn
= i e2yti 2xt
ei 2xt
ei 2t d. + y2
Fig. 12.1. The curve n .
The curve Cn is the semicircle in n . (See Figures 12.1 and 12.2.) To estimate the last integral, note that n ei 2t 2 e2t 2 + y2 n y2 for all Cn , and write =
Cn
Cn { > n}
Cn { < n}
270
Fig. 12.2. The curve Cn .
Therefore, 2
n2 2n2 1 ei 2t d 2 e2t n + 2 arcsin , 2 2 +y n y n y2 n
Cn
which implies fn (t) = i e2yti 2xt + o(1), as n . A similar argument shows that fn (t) = i e2yti 2xt + o(1), (t < 0). (t > 0),
We know that fn converges to f in L2 (R). Hence a subsequence of fn converges . But, as the last two identities show, almost everywhere to f fn (t) i sgn(t) e2y|t|i 2xt at all t R \ {0}. Hence f (t) = i sgn(t) e2y|t|i 2xt , In particular, the FourierPlancherel transform of Qy (t) = is (See Figure 12.3.) t 1 t2 + y 2 (t R).
Qy (t) = i sgn(t) e2y|t| .
12.3. The multiplication formula on Lp (R) (1 p 2)
271
Fig. 12.3. The spectrum of
1 i
Qy .
Exercises
Exercise 12.2.1 Let f L1 (R) and suppose that f L2 (R). Show that we 2 also have f L (R). Exercise 12.2.2 Let f L1 (R) and g L2 (R). Show that f g = f g.
12.3
The multiplication formula on Lp (R) (1 p 2)
The identity f 2 = f 2 implies that the FourierPlancherel transform is 2 injective on L (R). To show that it is also surjective, and thus a unitary operator on L2 (R), we need the following generalization of the multiplication formula. Lemma 12.4 (Multiplication formula) Let f, g L2 (R). Then

f (t) g (t) dt =

f (t) g(t) dt.
272
Proof. Pick two sequences fn , gn L1 (R) L2 (R), n 1, such that

n
lim
fn f
= lim
gn g
= 0.
According to the denition of the FourierPlancherel transform, we also have

n
lim
fn f
= lim
gn g
= 0.
Thus, by Hlders inequality, o

n
lim
fn gn f g
= lim
fn gn f g
= 0.
But fn , gn L1 (R). Hence, by Lemma 9.2,

f (t) g (t) dt
= = =
lim
fn (t) gn (t) dt fn (t) gn (t) dt
lim
f (t) g(t) dt.
We are now able to show that the FourierPlancherel transform F is surjective on L2 (R). As a matter of fact, if g is orthogonal to the range of F, i.e. f (t) g(t) dt = 0
for all f L (R), then, by the multiplication formula, we have
f (t) g (t) dt = 0,
which immediately implies g = 0. Hence, by (12.4), g = 0. Thus F is surjective. Using (12.1), we can also provide an inversion formula for the Fourier Plancherel transform. Let T denote the unitary operator T : L2 (R) L2 (R) f (t) f (t). Hence T F F is also a unitary operator on L2 (R), and moreover, by (12.1), the identity T F F = id
holds at least on a dense subset of L2 (R), e.g. on Cc (R). Note that if f Cc (R), then certainly f L1 (R). Therefore, the identity holds on L2 (R) and thus F 1 = T F. (12.5)
12.4. The Fourier transform on Lp (R) (1 p 2)
273
Exercises
Exercise 12.3.1
Let f Cc (R). Show that
f C (R) Lp (R) for all 0 < p . Hint: Use Exercise 9.2.9. Let f, g L2 (R). Show that

Exercise 12.3.2
f (t) g (t) dt =
f (t) g(t) dt.
Hint: Use Lemma 12.4 and (12.5).
12.4
Let
The Fourier transform on Lp (R) (1 p 2)

f L1 (R) + L2 (R).
Then for each f = f1 + f2 , where f1 L1 (R) and f2 L2 (R), dene its Fourier transform by f = f1 + Ff2 . If f has another representation, say f = g1 + g2 , where g1 L1 (R) and g2 L2 (R), then f1 g1 = g2 f2 L1 (R) L2 (R), and thus, by (12.3), f1 g1 = F(g2 f2 ). Therefore, f1 + Ff2 = g1 + Fg2 , which implies that f is well-dened. Since, for all p with 1 p 2, Lp (R) L1 (R) + L2 (R), the Fourier transforms of elements of Lp (R), 1 p 2, are well-dened now. Theorem 12.5 (HausdorYoung theorem) Let f Lp (R), 1 p 2. Then f Lq (R), where q is the conjugate exponent of p, and f
q
p.
274 Proof. By Lemma 9.1,
F : L1 (R) L (R) f f is of type (1, ) with F

(1,)
and, by Plancherels theorem (see (12.4)), F : L2 (R) L2 (R) f f is of type (2, 2) with F
(2,2)
1.
The rest is exactly as in the proof of Theorem 8.3.
Exercises
Exercise 12.4.1 Let f L1 (R) and g Lp (R), 1 p 2. Show that f g = f g. Hint: Use Exercise 12.2.2. Exercise 12.4.2 Let 1 p 2, and let q be the conjugate exponent of p. Let g Lp (R). Show that there is an f Lq (R) such that f =g and that f
q
f p.
Hint 1: Use duality and Theorem 12.5. Hint 2: Give a direct proof using the RieszThorin interpolation theorem.
12.5
An application of the multiplication formula on Lp (R) (1 p 2)
The contents of this chapter are designed to obtain the following result, which reveals the spectral properties of elements in hp (C+ ), 1 p 2.
12.5. An application of the multiplication formula on Lp (R) (1 p 2) Theorem 12.6 Let u Lp (R), 1 p 2. Then we have 1 1

275
y u(t) dt (x t)2 + y 2 xt u(t) dt (x t)2 + y 2 u(t) 1 dt 2i t z 1 2i

e2y|t| u(t) ei 2xt dt, i sgn(t) e2y |t| u(t) ei 2xt dt, u(t) ei 2z t dt,
u(t) ei 2z t dt,
=
0
u(t) dt tz
for all z = x + iy C+ . The integrals are absolutely convergent on compact subsets of C+ . Moreover, each line represents a harmonic function on C+ . The third line gives an analytic function. Proof. The case p = 1 was considered in Corollary 10.2. We start by proving the rst identity for p = 2. If u L2 (R), pick a sequence un L1 (R) L2 (R), n 1, such that lim un u 2 = 0.
n
Then, again by Corollary 10.2, 1

y un (t) dt = (x t)2 + y 2
e2y |t| un (t) ei 2xt dt.
Let n . The CauchySchwarz inequality ensures that the left side converges to y 1 u(t) dt, (x t)2 + y 2 and the right side to

e2y |t| u(t) ei 2xt dt,
as n . Hence, we obtain the required equality. Other identities are proved similarly if p = 2. If u Lp (R) L1 (R) + L2 (R), write u = u1 + u2 , where u1 L1 (R) and u2 L2 (R). Each identity holds for u1 and u2 . Hence, taking their linear combination, each identity also holds for u. Since u = u1 + u2 C0 (R) + L2 (R), the presence of e2y |t| ensures that the right-side integral converges absolutely and uniformly on compact subsets of C+ . Finally, the same method used in Theorem 10.1 shows that each row represents a harmonic function. Corollary 12.7 (Uniqueness theorem) Let f Lp (R), 1 p 2. Suppose that f (t) = 0
276 for almost all t R. Then
f (t) = 0 for almost all t R. Proof. By Theorem 12.6, we have 1

y f (t) dt = (x t)2 + y 2
e2y|t| f (t) ei 2xt dt = 0
for all x + iy C+ . Hence, by Corollary 11.11, f = 0.
Exercises
Exercise 12.5.1 that [Multiplication formula] Let f, g Lp (R), 1 p 2. Show

f (t) g (t) dt =

f (t) g(t) dt.
Exercise 12.5.2 Let 1 < p < 2 and let q be the conjugate exponent of p. By Theorem 12.5 and Corollary 12.7, the map Lp (R) Lq (R) f f is well-dened and injective. However, show that it is not surjective.
12.6
A complete characterization of hp (C+ ) spaces
Similar to what we did in Section 4.7, we summarize the results of preceding sections to demonstrate a complete characterization of hp (C+ ) spaces. In the following, U is harmonic in the upper half plane C+ . Case 1: p = 1. U h1 (C+ ) if and only if there exists M(R) such that U (x + iy) = 1 y d(t), (x t)2 + y 2 (x + iy C+ ).
The measure is unique and
U (x + iy) =
(t) e2y|t|+i2xt dt,
(x + iy C+ ),
12.6. A complete characterization of hp (C+ ) spaces where C(R) L (R) and
277
The measures dy (t) = U (t + iy) dt converge to d(t) in the weak* topology, as y 0, i.e.
y0
lim
(t) dy (t) =
(t) d(t)
T
for all C0 (R), and we have U

1
= = ||(R).
Case 2: 1 < p 2. U hp (C+ ) if and only if there exists u Lp (R) such that U (x + iy) = 1
R
y u(t) dt, (x t)2 + y 2
(x + iy C+ ).
The function u is unique and
U (x + iy) =
u(t) e2y|t|+i2xt dt,
(x + iy C+ ),
where u Lq (R), with 1/p + 1/q = 1, and u Moreover,

y0 q
u p.
p
lim Uy u U
= 0,
and thus
p
= u p.
Case 3: 2 < p < . U hp (C+ ) if and only if there exists u Lp (R) such that U (x + iy) = 1
R
y u(t) dt, (x t)2 + y 2 lim Uy u U

p
(x + iy C+ ).
The function u is unique,

y0
= 0,
and thus
p
= u p.
278
Case 4: p = . U h (C+ ) if and only if there exists u L (R) such that U (x + iy) = 1 y u(t) dt, (x t)2 + y 2 (x + iy C+ ).
The function u is unique and U
= u
Moreover, Uy converges to u in the weak* topology, as y 0, i.e.

y0
lim
(t) U (t + iy) dt =
R R
(t) u(t) dt
for all L1 (R). The rst case shows that h1 (C+ ) and M(R) are isometrically isomorphic Banach spaces. Similarly, by the other three cases, hp (C+ ) and Lp (R), 1 < p , are isometrically isomorphic.
Exercises
Exercise 12.6.1 We saw that h1 (C+ ) and M(R) are isometrically isomorphic Banach spaces. Consider L1 (R) as a subspace of M(R). What is the image of L1 (R) in h1 (C+ ) under this correspondence? What is the image of positive measures? Do they give all the positive harmonic functions in the upper half plane? Exercise 12.6.2 We know that h (C+ ) and L (R) are isometrically isomorphic Banach spaces. Consider C0 (R) as a subspace of L (R). What is the image of C0 (R) in h (C+ ) under this correspondence? What is the image of Cc (R)?
Chapter 13
Analytic functions in the upper half plane

13.1 Representation of H p (C+ ) functions (1 < p )
Since H p (C+ ) hp (C+ ), the representation theorems that we have obtained for harmonic functions apply to analytic functions too. Some of these results have been gathered in Section 12.6. In this chapter we study the analytic Hardy spaces H p (C+ ) in more detail. Lemma 13.1 Let > 0. If F H p (C+ ), 1 p < , then

F (t + i) dt = 0, tz
(z C+ ).
If F H (C+ ), then

F (t + i) dt = 0, (t z )(t + i)
(z C+ ).
Proof. Fix z C+ and > 0. Let R > + |z| and let R, be the curve shown in Figure 13.1. Therefore, i + z is not in the interior of R, , and thus, by Cauchys integral formula, F () d = 0. i z
R,
279
280
Chapter 13. Analytic functions in the upper half plane
Fig. 13.1. The curve R, . Hence,

R cos R R cos R
F (t + i) dt = tz
R R
F (Rei ) iRei d, Rei i z
where, according to triangle R, , we have R = arcsin(/R). (See Figure 13.2.)
Fig. 13.2. The triangle R, . By Lemma 11.3, we now have

R R
F (Rei ) iRei d Rei i z
R (2/)1/p F R |i + z |
1/p
R R /2 R /2 R
1 d (R sin )1/p 1 d (2/)1/p 1 d 2/
F p 2R (2/) (R |i + z |) R1/p 2R (2/)1/p F p (R |i + z |) R1/p

1/p
R (2/) F p log(/2R ). (R |i + z |) R1/p
13.1. Representation of H p (C+ ) functions (1 < p ) But R sin R = /R, and thus
R R
281
F (Rei ) R (2/)1/p F p iRei d log(R/2). i i z Re (R |i + z |) R1/p

R cos R R cos R
Hence,
F (t + i) dt = O(log R/R1/p ). tz
Now let R . The proof for the case p = is even simpler. The reason we add the extra factor (t + i) in the denominator is that, for an arbitrary bounded analytic function F , F (t + i) tz is not necessarily integrable over R. No wonder we want to let 0 in the preceding lemma to obtain a representation theorem based on the boundary values of F on the real line R. Let us see what happens. If F H p (C+ ), 1 < p < , then, by Theorem 11.6, there is a unique f Lp (R) such that F (x + iy) = 1

y f (t) dt, (x t)2 + y 2
(x + iy C+ ).
(13.1)
Moreover, by Corollary 10.12, we have

0
lim F f
= 0.
Hence, by Lemma 13.1,

f (t) dt = lim 0 tz
F (t) dt = 0 tz
(13.2)
for all z C+ . Since i/2 i/2 y = 2 + y2 (x t) tz tz and (x t) 1/2 1/2 + , = tz tz (x t)2 + y 2
(13.3)
the representation (13.1) along with the property (13.2) imply F (z) = = 1 2i i f (t) dt tz xt f (t) dt, (x t)2 + y 2
282
for all z = x + iy C+ . On the other hand, if f Lp (R), 1 < p < , satisfying

f (t) dt = 0, tz
(z C+ ),
is given, we can dene F (x + iy) = 1

y f (t) dt, (x t)2 + y 2
(x + iy C+ ).
Then, on the one hand, by Corollary 10.12, F hp (C+ ), and on the other hand, by (13.3) and our assumption on f , we obtain F (z) = 1 2i

f (t) dt, tz
(z C+ ).
Thus F indeed represents an element of H p (C+ ). Finally, since Fy converges to f in Lp (R), we have F
p
= f
p.
The preceding discussion shows that H p (C+ ), 1 < p < , is isomorphically isometric to H p (R) = f Lp (R) :

f (t) dt = 0, for all z C+ , tz
(13.4)
which is a closed subspace of Lp (R). Theorem 13.2 Let F be analytic in the upper half plane C+ . Then F H p (C+ ), 1 < p < , if and only if there is f H p (R) such that F (x + iy) = 1

y f (t) dt, (x t)2 + y 2
(x + iy C+ ).
If so, f is unique and we also have F (z) = = Moreover,

y0
i 1 2i
xt f (t) dt (x t)2 + y 2 f (t) dt, (z = x + iy C+ ). t z lim Fy f

p
=0
and F
p
= f
p.
13.1. Representation of H p (C+ ) functions (1 < p )
283
A slight modication of the argument provided above for H p (C+ ), 1 < p < , gives a representation for H (C+ ). If F H (C+ ), then, by Theorem 11.6, there is a unique f L (R) such that F (x + iy) = and
0
y f (t) dt, (x t)2 + y 2
(x + iy C+ ),
(13.5)
lim
(t) F (t + i) dt =

(t) f (t) dt
for all L1 (R). In particular, with (t) = we obtain

1 , (t z )(t + i) f (t) dt = 0, (t z )(t + i)
(t R, z C+ ),
(z C+ ).
(13.6)
Since
y ( + i)/2i z (z + i)/2i , = (x t)2 + y 2 (t z)(t + i) (t z )(t + i) 1 2i

(13.7)
the representation (13.5) along with the properties (13.6) and (13.7) imply F (z) = 1 1 tz t+i f (t) dt, (z C+ ).
On the other hand, if f L (R) satisfying f (t) dt = 0, (t z )(t + i)

(z C+ ),
is given, we can dene F (x + iy) = 1 y f (t) dt, (x t)2 + y 2 (x + iy C+ ).
Then, on the one hand, by Corollary 10.14, F h (C+ ), and on the other hand, by (13.7) and our assumption on f , we obtain F (z) = 1 2i

1 1 tz t+i
f (t) dt,
(z C+ ).
Thus F represents an element of H (C+ ). Finally, by Corollary 10.14, we have F

= f
The preceding discussion shows that H (C+ ) is isomorphically isometric to H (R) = f L (R) : f (t) dt = 0, for all z C+ , (t z )(t + i)
which is a closed subspace of L (R).
284
Theorem 13.3 Let F be analytic in the upper half plane C+ . Then F H (C+ ) if and only if there is an f H (R) such that F (x + iy) = 1

y f (t) dt, (x t)2 + y 2
(x + iy C+ ).
If so, f is unique and we also have F (z) = = Moreover, F
i 1 2i
xt 1 f (t) dt + (x t)2 + y 2 t+i 1 1 f (t) dt, (z = x + iy C+ ). tz t+i = f

.
Exercises
Exercise 13.1.1 H (R) = Show that

f L (R) :
f (t) dt = 0, for all z, w C+ . (t z )(t w)
Exercise 13.1.2
Let f L (R) be such that

f (t) dt = 0 (t z )2
for all z C+ . Can we conclude that f H (R)?
13.2
Analytic measures on R
One of our main goals in this chapter is to show that Theorem 13.2 remains valid even if p = 1. To prove this result, we need to study a special class of Borel measures. A measure M(R) is called analytic if (t) = 0 (13.8)
for all t 0. We use Rieszs theorem (Theorem 5.10) to characterize these measures. Lemma 13.4 Let M(R). Then the following are equivalent: (a) is analytic;
13.2. Analytic measures on R (b) for all z C+ ,

R
285
d(t) = 0; tz
(13.9)
(c) d(t) = u(t) dt where u L1 (R) and u(t) = 0 for all t 0. Proof. (a) = (b) : According to Theorem 10.1, for any M(R), we have 1 2i d(t) = tz
0 (t) ei 2z t dt,
(z C+ ).
Hence, (a) clearly implies (b). (b) = (c) : Taking the successive derivative of (13.9) with respect to z gives d(t) =0 (t z )n R for all n 1 and z C+ . Since (t + z )n1 = (t z )n we thus have
R n1 k=0
n1 k
(2t)k (1)n1k , (t z )k+1
(t + z )n1 d(t) = 0, (t z )n
n
(n 1).
In particular, for z = i, we obtain ti t+i d(t) = 0, it

it i+t
(n 1). to write the preceding integral
Now we make the change of variable ei = over T. Therefore, by (11.6), we have ein d(ei ) = 0,
T
(n 1),
where d(ei ) is the pullback of d(t)/(i t) and we assume that ({1}) = 0. The celebrated theorem of F. and M. Riesz (Theorem 5.10) implies that d(ei ) is absolutely continuous with respect to the Lebesgue measure d . Hence, by (11.4) and (11.7), d(t) is absolutely continuous with respect to the Lebesgue measure dt, i.e. d(t) = u(t) dt, where u L1 (R). Moreover, based on the assumption, u satises

u(t) dt = 0, tz
(z C+ ).
286
Thus, by Corollary 10.2, we have

0 u(t) ei 2z t dt =
u(t) dt = 0, tz
(z C+ ).
In particular, with z = x + i, x R, we obtain

u(t) e2 t (,0) (t)
ei 2x t dt = 0
for all x R. Hence, by the uniqueness theorem (Corollary 12.7), u(t) = 0 for almost all t < 0. Since u is a continuous function, the last identity holds for all t 0. (c) = (a) : This part is obvious.
13.3
Representation of H 1 (C+ ) functions
The denition (13.4) is valid even if p = 1. Now, using Lemma 13.4, we show that H 1 (R) also deserves attention. If F H 1 (C+ ), then, by Theorem 11.7, there is a unique M(R) such that
0
lim
(t) F (t + i) dt =
R
(t) d(t)
for all C0 (R). In particular, if (t) = 1/(t z ), then, by Lemma 13.1,

d(t) = 0, tz
(z C+ ).
Hence, is an analytic measure and, by Lemma 13.4, there is an f L1 (R) such that d(t) = f (t) dt and

f (t) dt = 0, tz
(z C+ ).
The big step is now taken and the rest of the discussion is exactly like the one given for the case 1 < p < in Section 13.1. Theorem 13.5 Let F be analytic in the upper half plane C+ . Then F H 1 (C+ ) if and only if there is an f H 1 (R) such that F (x + iy) = 1

y f (t) dt, (x t)2 + y 2
(x + iy C+ ).
13.4. Spectral analysis of H p (R) (1 p 2) If so, f is unique and we also have F (z) = = Moreover,
y0
287
i 1 2i
xt f (t) dt 2 2 (x t) + y f (t) dt, (z = x + iy C+ ). t z lim Fy f = 0,
and F
1
= f
1.
13.4
Spectral analysis of H p (R) (1 p 2)
In the preceding sections we dened H p (R), 1 p , as a closed subspace of Lp (R), and showed that it is isomorphically isometric to H p (C+ ). If 1 p 2, the Fourier transforms of elements of H p (R) are well-dened. In this section we study the Fourier transform of these elements. Theorem 13.6 Let f H p (R), 1 p 2, and let F (x + iy) = 1

y f (t) dt, (x t)2 + y 2
(x + iy C+ ).
Then f (t) = 0, for almost all t 0, and moreover
F (z) =
0
f (t) ei2zt dt,
(z C+ ).
Proof. By Theorem 12.6, we have 1 2i

f (t) dt = tz
f (t) ei 2z t dt,
for all z = x + iy C+ and all f Lp (R), 1 p 2. If f H p (R), then, by denition, 1 f (t) dt = 0, (z C+ ). 2i t z Hence,
0 f (t) ei 2z t dt = 0,
(z C+ ).
Thus, by the uniqueness theorem (Corollary 12.7), we necessarily have f (t) = 0 for almost all t 0.
288
Chapter 13. Analytic functions in the upper half plane Finally, by Theorem 12.6, we have F (z) = = =
0
2y|t|
y f (t) dt (x t)2 + y 2 f (t) ei 2xt dt
e2yt f (t) ei 2xt dt f (t) ei2zt dt, (z = x + iy C+ ).
=
0
The case p = 1 of Theorem 13.6 is more interesting, since in this case f is continuous on R and we thus have f (t) = 0 for all t 0. In particular, for t = 0, we obtain the following result. Corollary 13.7 Let f H 1 (R). Then
f ( ) d = 0.
Theorem 13.6 has a converse which enables us to give another characterization of H p (R), 1 p 2. Suppose that f Lp (R), 1 p 2, satisfying f (t) = 0, for almost all t 0, is given. Hence, by Theorem 12.6, we immediately have 0 1 f (t) dt = f (t) ei 2z t dt = 0 2i t z for all z = x + iy C+ . Therefore, f H p (R). Based on this observation and Theorem 13.6, we can say H p (R) = { f Lp (R) : f (t) = 0 for almost all t 0 } for 1 p 2. Hence, by the HausdorYoung theorem (Theorem 12.5), the Hardy space H p (R), 1 p 2, is mapped into Lq (R+ ), where q is the conjugate exponent of p, under the Fourier transform. According to the uniqueness theorem (Corollary 12.7), the map is injective. Hence, we naturally ask if it is surjective too. Using Plancherels theorem we prove this assertion for p = 2. In other cases the map is not surjective. Theorem 13.8 Let L2 (R+ ). Then there is an f H 2 (R) such that f = .
13.5. A contraction from H p (C+ ) into H p (D) Proof. Let f (t) = (t). By Plancherels theorem we certainly have f L2 (R). Moreover, by (12.5), f (t) = (t) = (t). Hence f L2 (R+ ), which is equivalent to f H 2 (R).
289
Exercises
Exercise 13.4.1 12.7), the map By Theorem 13.6 and the uniqueness theorem (Corollary H 1 (R) C0 (R+ ) f f is injective. However, show that it is not surjective. Hint: Use Exercise 12.1.3.
Exercise 13.4.2 Let 1 < p < 2 and let q be the conjugate exponent of p. By Theorem 13.6 and the uniqueness theorem (Corollary 12.7), the map H p (R) Lq (R+ ) f f is injective. However, show that it is not surjective. Hint: See Exercise 12.5.2.
13.5
A contraction from H p (C+ ) into H p (D)
The following result gives a contraction from H p (C+ ) into H p (D). This is a valuable tool which enables us to easily establish some properties of Hardy spaces of the upper half plane from the known results on the unit disc. Lemma 13.9 Let F H p (C+ ), 0 0. By Lemma 11.5, |F (z + i)|p , as a function of z, is a bounded subharmonic function on C+ which is also continuous on C+ . Thus, by Corollary 10.15, |F (x + iy + i)|p 1

y |F (t + i)|p dt, (x t)2 + y 2
(x + iy C+ ).
By assumption, (|F |p )>0 is a uniformly bounded family in L1 (R). Hence, there is a measure M(R) and a subsequence (|Fn |p )n1 such that |F (t+in )|p dt converges to d(t) in the weak* topology. Hence, (R) = lim inf Fn
n p p
= F
p p.
(13.10)
Moreover, by letting n 0, we obtain |F (x + iy)|p 1

R
y d(t), (x t)2 + y 2
(x + iy C+ ).
Now we apply the change of variable z = i 1w , as discussed in Section 11.5, to 1+w get |G(rei )|p 1+ r2 1 r2 cos2 ( /2) d( ), 2r cos( ) (rei D),
where is the pullback of with the extra assumption ( {1} ) = 0. The identity 1 r2 y cos2 ( /2) = (x t)2 + y 2 1 + r2 2r cos( ) was also used implicitly in the preceding calculation. Therefore, by Fubinis theorem and (13.10), 1 2

|G(rei )|p d
cos2 ( /2) d( )
d( ) = (T) = (R) F
p p.
Lemma 13.9 says that the mapping H p (C+ ) F (z) H p (D)
1/p F i 1w 1+w
is a contraction. Even though this map is one-to-one, it is not onto. More precisely, if we start with G H p (D), 0 < p < , and dene F (z) = G iz i+z , (z C+ ),
then F is not necessarily in H p (C+ ). A simple counterexample is G(w) = w H 1 (D), for which F (z) = iz H 1 (C+ ). i+z
13.5. A contraction from H p (C+ ) into H p (D) Theorem 13.10 Let F H p (C+ ), 0 < p , F 0. Then (a) f (x) = lim F (x + iy)
y0
291
exists for almost all x R. (b)

log |f (t)|
dt < , 1 + t2
(c)
y0
lim
log+ |F (t + iy)| log+ |f (t)|
dt = 0, 1 + t2
(d) log |F (x + iy)| Proof. Dene G(w) = F i 1w 1+w , (w D). 1

y log |f (t)| dt, (x t)2 + y 2
(x + iy C+ ).
By Lemma 13.9, G H p (D). Hence, by Fatous theorem (Theorem 3.12), G has boundary values almost everywhere on T. Moreover, the results gathered in Theorem 7.9 hold for G. Applying again the change of variable z = i 1w , 1+w which was discussed in Section 11.5, implies all the parts. Note that d = 2dt/(1 + t2 ). The following result was partially obtained in the proof of Lemma 13.9. Knowing that F has boundary values almost everywhere on the real line, we obtain a candidate for the measure appearing there. Corollary 13.11 Let F H p (C+ ), 0 < p < , and let f (x) = lim F (x + iy)
y0
wherever the limit exists. Then f Lp (R), f and |F (x + iy)|p 1

p
= F
y |f (t)|p dt, (x t)2 + y 2
(x + iy C+ ).
292
Proof. The rst assertion is a direct consequence of Fatous lemma, which also implies f p F p . By Theorem 13.10, log |F (x + iy)| Hence |F (x + iy)|p exp 1

y log |f (t)| dt, (x t)2 + y 2
(x + iy C+ ).
y log |f (t)|p dt , (x t)2 + y 2
(x + iy C+ ).
Now apply Jensens inequality to obtain the last result. Moreover, by Fubinis theorem, we have

|F (x + iy)|p dx
p.
|f (t)|p dt,
(y > 0).
Hence, F
Exercises
Exercise 13.5.1 Let , n Lp (R), 0 < p < , n 1. Suppose that n (x) (x) for almost all x R, and that
n
lim
= p.
Show that
n
lim
= 0.
Hint: Apply Egorovs theorem [5, page 21]. Exercise 13.5.2 Let F H p (C+ ), 0 < p < , and let f (x) = lim F (x + iy)
y0
wherever the limit exists. Show that

y0
lim Fy f
= 0.
Hint: Use Corollary 13.11 and Exercise 13.5.1.
13.6. Blaschke products for the upper half plane
293
13.6
Blaschke products for the upper half plane

w0 w , 1 w0 w
In Section 7.1, we called bw0 (w) = (w D),
a Blaschke factor for the open unit disc. If we apply the change of variables given in Section 11.5, we obtain |w0 | w0 w = w0 1 w0 w where z C+ and w D are related by w= iz , i+z or equivalently z=i 1w . 1+w
iz0 i0 z iz0 i0 z
z z0 , z z0
(13.11)
The points z0 C+ and w0 D are related by the same equations. As usual, we assume that
iz0 i0 z iz0 i0 z
=1
if z0 = i. Moreover, we also have 1 |w0 |2 = Therefore, for each z0 C+ , we call bz0 (z) = z z0 z z0 4 z0 . |i + z0 |2 (13.12)
a Blaschke factor for the upper half plane. Either by direct verication or by (13.11) and our knowledge of the Blaschke factors for the open unit disc, it is easy to verify that (z C+ ), |bz0 (z)| < 1, and that |bz0 (t)| = 1, (t R). In the light of (13.12), a sequence (zn )n1 C+ is called a Blaschke sequence for the upper half plane if zn < . |i + zn |2 n=1 Hence, Theorem 7.1 is rewritten as follows for C+ .
(13.13)
294
Theorem 13.12 Let {zn }n 1 be a sequence in the upper half plane with no accumulation point in C+ . Let = {zn }n 1 denote the set of all accumulation points of {zn }n 1 which are necessarily on the real line R, and let = C \ {n : n 1} . Then the partial products z
N
BN (z) =
n=1
izn in z izn in z
z zn z zn
are uniformly convergent on compact subsets of if and only if zn < . |i + zn |2 n=1 The limit of these partial products is called an innite Blaschke product for the upper half plane and is denoted by

B(z) =
n=1
ein
z zn , z zn
where the real number n is chosen such that e

in
izn in z izn in z
Equivalently, we can say that n is a real number such that ein i zn 0. i zn
Clearly B H (C+ ) and B 1. However, in the same manner, Theorem 7.4 is written as follows for the upper half plane. This result implies that B = 1. Theorem 13.13 Let B be a Blaschke product for the upper half plane. Let b(x) = lim B(x + iy)
y0
wherever the limit exists. Then |b(x)| = 1 for almost all x R.
13.7
The canonical factorization in H p (C+ ) (0 < p )
In this section we show that the zeros of a function in H p (C+ ) satisfy the Blaschke condition. Hence we form a Blaschke product to extract these zeros. This technique was discussed for Hardy spaces of the open unit disc in Section 7.4.
13.7. The canonical factorization in H p (C+ ) (0 < p )
295
Lemma 13.14 Let F H p (C+ ), 0 < p , F 0. Let (zn ) be the sequence of zeros of F , counting multiplicities, in C+ . Then (zn ) is a Blaschke sequence for the upper half plane, i.e. zn < . |i + zn |2 1w 1+w
Proof. Let G(w) = F i
(w D).
By Lemma 13.9, G H p (D). Let wn = i zn . i + zn
The sequence (wn ) represents the set of zeros of G in D. Hence, by Lemma 7.6, (1 |wn |2 ) < .
n
But 1 |wn |2 =
4 zn , |i + zn |2
which implies that (zn ) satises the Blaschke condition for the upper half plane. Knowing that the zeros of a function F in the Hardy class H p (C+ ) full the Blaschke condition, we are able to construct a Blaschke product, say B. Then clearly G = F/B is a well-dened analytic function in the upper half plane. To be more precise, we should say that G has removable singularities at the zeros of F . Moreover, as we discussed similarly in Theorem 7.7, G stays in the same class and has the same norm. However, we go further and provide the complete canonical factorization for elements of H p (C+ ). A function F H (C+ ) is called inner for the upper half plane if |f (t)| = | lim F (t + iy)| = 1
y0
for almost all t R. According to Theorem 13.13, any Blaschke product B is an inner function for the upper half plane. Let be a positive singular Borel measure on R satisfying d(t) < . 2 R 1+t Then direct verication shows that S (z) = S(z) = exp i 1 t + z t 1 + t2 d(t) , (z C+ ),
296
is an inner function for C+ . For obvious reasons, S is called a singular inner function. Finally, E (z) = E(z) = eiz , (z C+ ),
where 0, is also an inner function for C+ . The product of two inner functions is clearly inner. Hence, I = EBS is inner. A part of the canonical factorization theorem says that each inner function has such a decomposition. Let h 0 and let | log h(t)| dt < . 2 1 + t Then Oh (z) = O(z) = exp i

t 1 + z t 1 + t2
log h(t) dt ,
(z C+ ),
is called an outer function for the upper half plane. According to Theorem 13.10, if F H p (C+ ), 0 < p , F 0 and f (x) = limy0 F (x + iy) wherever the limit exists, then dt log |f (t)| < . 1 + t2 Hence, we can dene the outer part of F by OF (z) = exp i

1 t + z t 1 + t2
log |f (t)| dt ,
(z C+ ).
Theorem 13.15 Let F H p (C+ ), 0 < p , F 0. Let B be the Blaschke product formed with zeros of F in C+ . Then there is a positive singular Borel measure satisfying d(t) < 1 + t2 R and a constant 0 such that F = IF OF , where the inner part of F is given by IF = E B S . Proof. Let G(w) = F i 1w 1+w
(w D).
By Lemma 13.9, G H p (D). Hence, according to the canonical factorization theorem for Hardy spaces on the open unit disc, we have G = B S OG ,
13.7. The canonical factorization in H p (C+ ) (0 < p )
297
where B is the Blaschke product formed with the zeros of G, is a nite positive singular Borel measure on T, S (w) = exp and OG (w) = exp is the outer part of G. Write S (w) = exp ({1}) 1 w 2 1+w exp 1 2 ei + w d(ei ) . ei w 1 2
T
1 2
ei + w d(ei ) ei w
ei + w log |G(ei )| d ei w
T\{1}
Let = ({1})/(2), and let be the pullback of to the real line. Now we i apply again the change of variables z = i 1w and t = i 1ei . Note that 1+w 1+e ei + w =i ei w 1 t + z t 1 + t2 (1 + t2 ).
Finally, let d(t) = (1 + t2 ) d(t). The canonical decomposition for the upper half plane follows immediately from the corresponding decomposition on the open unit disc. Exactly the same technique that was applied in the proof of Theorem 7.8 can be utilized again to obtain a similar result for the upper half plane. A dierent proof of this result was sketched in Exercise 13.5.2. Moreover, the result has already been proved when 1 p < . Corollary 13.16 Let F H p (C+ ), 0 < p < , and let f (t) = lim F (t + iy)
y0
wherever the limit exists. Then f Lp (R) and

r1
lim Fr f
= 0.
Exercises
Exercise 13.7.1 Let h 0, and let

| log h(t)| dt < . 1 + t2
Show that Oh H p (C+ ), 0 < p , if and only if h Lp (R). Moreover, Oh p = h p .
298 Exercise 13.7.2 satisfying
Chapter 13. Analytic functions in the upper half plane Let be a nonzero positive singular Borel measure on R d(t) < . 1 + t2
Let S(z) = exp i 1 t + z t 1 + t2 d(t) , (z C+ ).
Show that there is an x0 R such that

y0
lim S(x0 + iy) = 0.
Remark: Remember that limy0 |S(x + iy)| = 1 for almost all x R.
13.8
A correspondence between H p (C+ ) and H p (D)
In Section 11.5 we introduced a conformal mapping between the open unit disc and the upper half plane. However, if we apply this mapping, we do not obtain a bijection between the Hardy spaces of the upper half plane and those of the open unit disc. Nevertheless, these spaces are not completely irrelevant. Lemma 13.9 provides a contraction from H p (C+ ) into H p (D). However, the map is not surjective. The following result gives an isometric isomorphism between these two families of Hardy spaces. Theorem 13.17 Let 0 < p . Let F and G be analytic functions respectively on C+ and D which are related by the following equivalent equations: F (z) G(w) = = 1 (z + i) 2/p 2 i F 1+w
2/p
G i
iz i+z ,
(z C+ ), (w D).
1w 1+w
Then G H p (D) if and only if F H p (C+ ). Moreover, F

H p (C+ )
= G
H p (D) .
Proof. The result is clear if p = . Hence let 0 < p < . First, suppose that G H p (D). Thus G has boundary values almost everywhere on T, which we represent by G(ei ), and G = 1 2

H p (D)
|G(ei )|p d
1/p
13.8. A correspondence between H p (C+ ) and H p (D)
299
Hence, by the rst relation, F also has boundary values almost everywhere on R, which we denote by F (t), and F (t) = 1 + ei 2 i
2/p
G(ei ),
t=i
1 ei R . 1 + ei
We write this identity as Thus, by (11.7), (i + t)

2/p
F (t) = G(ei ).

|F (t)|p dt =
1 2
|G(ei )|p d.
(13.14)
By Corollary 7.10, we have |G(rei )|p 1 2

1+
r2
1 r2 |G(ei )|p d. 2r cos( )

iz i+z ,
Hence, making the change of variable w = |F (x + iy)|p Therefore, by Fatous lemma,

we obtain
y |F (t)|p dt. (x t)2 + y 2
|F (x + iy)|p dx
|F (t)|p dt,
(y > 0).
H p (D) .
This inequality shows that F H p (C+ ) and, by (13.14), F H p (C+ ) = G The reverse implication is similar, except that we need Corollary 13.11.
300
Chapter 14
The Hilbert transform on R

14.1 Various denitions of the Hilbert transform
= { x + iy : y > Y }, with Y > 0. Suppose that

Let F be analytic and bounded on some half plane
|F (t)| dt < 1 + |t|
(14.1)
and
z z0
lim F (z) = 0.
(14.2)
Fix x R. Let 0 < < Y and let ,R be the curve shown in Figure 14.1.
Fig. 14.1. The curve ,R . 301
302 Then, by Cauchys theorem, F (x) = = + 1 2i 1 2i 1 2
Chapter 14. The Hilbert transform on R
,R
F () d x 1 F (t) dt + tx 2
2
F (x + eit ) dt
|xt|R 0
F (x + Reit ) dt.
Let R . By (14.1) and (14.2) we obtain F (x) = Thus F (x) = lim Let us write u = to F and v =
0
1 2i
|xt|
1 F (t) dt + tx 2 i
F (x + eit ) dt.
|xt|
F (t) dt. xt
(14.3)
F on the real line. Hence, (14.3) is equivalent

0
v(x) = lim and
1 1
|xt|
u(t) dt xt v(t) dt. xt
u(x) = lim
|xt|
Similar to the case of the unit circle, the preceding argument highlights the importance of the transform
0
lim
|xt|
(t) dt xt
whenever is the real or the imaginary part of an analytic function satisfying certain properties. However, this transform is also well-dened on other classes of functions and has several interesting properties. To avoid any problem at innity, we impose the condition

|(t)| dt < . 1 + |t|
Then the Hilbert transform of at the point x R is dened by H(x) = (x) = lim
0
|xt|
(t) dt xt
(14.4)
whenever the limit exists. As we have seen before, in many cases we are faced with a function satisfying |(t)| dt < . 2 1 + t
1 14.2. The Hilbert transform of Cc (R) functions
303
In this case, the Hilbert transform is dened by H(x) = (x) = lim

0
|xt|>
t 1 + (t) dt. x t 1 + t2
(14.5)
The presence of t/(1 + t2 ) ensures that for each xed x R and > 0, 1 t (t) dt + x t 1 + t2
|xt|>
is a well-dened Lebesgue integral. This fact is a consequence of 1 t + = O(1/t2 ) x t 1 + t2 as |t| . If is a measurable function such that

|(t)| dt < , 1 + |t|n
for a certain integer n, we are still able to add a specic term to 1/(x t) such that the whole combination behaves like 1/|t|n , as |t| , and thus obtain a meaningful Lebesgue integral which depends on . Then, by taking its limit as 0, we dene the Hilbert transform of at x.
14.2
Let
1 The Hilbert transform of Cc (R) functions

|(t)| dt < . 1 + t2
Fix x R and suppose that

1 0
|(x t) (x + t)| dt < . t 1 t (t) dt + x t 1 + t2 t (t) dt+ 1 + t2
(14.6)
We rewrite
|xt|>
as
1
(x t) (x + t) dt+ t
x+1 x1
|xt|>1
t 1 + (t) dt. x t 1 + t2
Hence, (x), as dened in (14.5), exists and is given by (x) = +

x+1 (x t) (x + t) t (t) dt + dt 2 t 0 x1 1 + t t 1 + (t) dt. x t 1 + t2 |xt|>1 1
(14.7)
304
The condition (14.6) is fullled if is continuous on [x1, x+1] and dierentiable at x. Clearly, the interval [x1, x+1] can be replaced by any other nite interval 1 around x. Similarly, if Cc (R), the space of continuously dierentiable functions of compact support on R, then , as dened by (14.4), exists for all x R and is given by 1 (x) =
0
(x t) (x + t) dt. t
(14.8)
1 Theorem 14.1 Let Cc (R). Then C0 (R).
Proof. Let V (x + iy) = 1

xt (t) dt, (x t)2 + y 2

0
(x + iy C+ ).
This formula can be rewritten as Vy (x) = Hence, |Vy (x) (x)| However, 1 1
0
t ((x t) (x + t)) dt. t2 + y 2
t 1 |(x t) (x + t)| dt 2 +y t y2 |(x t) (x + t)| dt. 2 + y2 t t t2 < ,
|(x t) (x + t)| 2 t 2 |Vy (x) (x)|

0
which implies y2 dt = t2 + y 2
for all x R. In other words, Vy converges uniformly to on R. Since, for each is also continuous on R. y > 0, Vy is continuous, It remains to show that tends to zero at innity. Let supp [M, M ]. Without loss of generality, suppose that x > M . Hence 1 (x) = and thus |(x)| Therefore, we indeed have (x) = O(1/|x|) as |x| . (14.9)
x+M xM
(x t) dt, t
2M . xM
14.3. Almost everywhere existence of the Hilbert transform
305
14.3
Almost everywhere existence of the Hilbert transform
In this section we follow a similar path as in Section 5.7 to show that the Hilbert transform is well-dened almost everywhere on R. Lemma 14.2 Let

|u(t)| dt < 1 + t2
and let V (x + iy) = 1

xt t u(t) dt, + (x t)2 + y 2 1 + t2
(x + iy C+ ).
Then, for almost all t R,

y0
lim
V (x + iy)
|xt|>y
t 1 + u(t) dt x t 1 + t2
= 0.
Proof. By Lebegues theorem 1 0 lim

0
|u(x t) u(x + t) | dt = 0
for almost all x R. We show that at such a point the lemma is valid. Let 1 1 t = V (x + iy) u(t) dt. + |xt|>y x t 1 + t2 Hence, = I1 + I2 + I3 , where I1 = 1 xt u(t) dt, (x t)2 + y 2 t u(t) dt 1 + t2
|xt|y
I2 = and I3 = 1
|xt|y
|xt|>y
xt 1 u(t) dt. (x t)2 + y 2 xt
The rst two integrals are easier to estimate: |I1 | 1 1 y

y 0
|u(x ) u(x + )| d 2 + y2
y
|u(x ) u(x + )| d = o(1)
306 and |I2 | 1

y y
|xt|y
|t| |u(t)| dt = o(1) 1 + t2
as y 0. To estimate the third integral, rst note that I3 1 y2 1 |u(x ) u(x + )| d 2 + y2 |u(x ) u(x + )| d. 3
Now, as was similarly shown in the proof of Lemma 5.16, the last expression also tends to zero when y 0. In the preceding lemma, we saw that the dierence of two functions of y tends to zero, as y 0, at almost all points of the real line. Now, we show that one of them has a nite limit almost everywhere. Hence, the other one necessarily has the same nite limit at almost all points of the real line. Lemma 14.3 Let

|u(t)| dt < 1 + t2 (x + iy C+ ).
and let V (x + iy) = 1

xt t u(t) dt, + (x t)2 + y 2 1 + t2 lim V (x + iy)
Then, for almost all x R,

y0
exists. Proof. Without loss of generality assume that u 0, u 0. Since otherwise we can write u = (u1 u2 ) + i(u3 u4 ), where each uk is positive. Let F (z) = Then F (z) = 1

t 1 + u(t) dt, z t 1 + t2
(z C+ ).
y u(t) dt > 0, (x t)2 + y 2
(z = x + iy C+ ).
Hence, using a conformal mapping between the unit disc and the upper half plane and considering F , Lemma 5.14 ensures that
y0
lim F (x + iy)
exists and is nite for almost all x R. Since V = F , then

y0
lim V (x + iy)
also exists and is nite for almost all x R.
14.3. Almost everywhere existence of the Hilbert transform A similar reasoning shows that if u satises the stronger condition

307
|u(t)| dt < 1 + |t|
and we dene V (x + iy) = 1

xt u(t) dt, (x t)2 + y 2
(x + iy C+ ),
then limy0 V (x + iy) exists and is nite for almost all x R. Note that in this case i u(t) dt , (z = x + iy C+ ). V (x + iy) = z t Theorem 14.4 Let

|u(t)| dt < . 1 + t2
Then, for almost all x R, u(x) = lim and

y0 0
|xt|>
1 t u(t) dt + x t 1 + t2
lim
xt t u(t) dt + (x t)2 + y 2 1 + t2
exist and they are equal. Proof. Apply Lemmas 14.2 and 14.3. If

|u(t)| dt < 1 + |t|
and we dene V (x + iy) = 1

xt u(t) dt, (x t)2 + y 2
(x + iy C+ ),
then, for almost all x R,

y0
lim
V (x + iy)
|xt|>y
u(t) dt xt
= 0.
Therefore, we conclude that in this case, u(x) = lim for almost all x R.
0
|xt|>
u(t) dt = lim V (x + iy) y0 xt
308
14.4
Kolmogorovs theorem

Similar to the case of the unit circle, the assumption |u(t)| dt < 1 + t2
is not enough to ensure that

|(t)| u dt < . 1 + t2
Nevertheless, a slight modication of the Kolmogorov theorem for the unit circle gives the following result for the Hilbert transform of functions dened on the real line. Theorem 14.5 (Kolmogorov) Let u be a real function satisfying

|u(t)| dt < . 1 + t2
Then, for each > 0, dt 4 1 + t2

{||>} u
|u(t)| dt. 1 + t2
Proof. (Carleson) Fix > 0. First suppose that u 0 and u 0. Therefore, the analytic function F (z) = i

t 1 + z t 1 + t2
u(t) dt
maps the upper half plane into the right half plane, and moreover we have F (i) = 1

u(t) dt > 0. 1 + t2
(14.10)
Then, as y 0, by Corollary 11.11, F (x + iy) = and by Lemma 14.3, F (x + iy) = 1

y u(t) dt u(x) (x t)2 + y 2
xt t + (x t)2 + y 2 1 + t2
u(t) dt u(x),
for almost everywhere x R. Now, we apply the conformal mapping used in Theorem 6.1, and dene 2F (z) G(z) = . F (z) +
14.4. Kolmogorovs theorem Since G H (C+ ), g(t) = limy0 G(t + iy) exists and g(t) = for almost all t R. Hence g =1+ u 2 + u 2 2 1 (u + )2 + u2 2( u(t) + i u(t) ) u(t) + i u(t) +
309
provided that u > . On the other hand, by Theorem 11.6, 1 and, by (14.10), 2F (i) 2 F (i) + Therefore, dt 1 + t2 g(t) 2 dt 1 + t2

g(t) 2F (i) dt = G(i) = 1 + t2 F (i) +

|u(t)| dt. 1 + t2
{||>} u
{||>} u
|u(t)| dt. 1 + t2
For an arbitrary real u L1 (R), write u = u1 u2 , where uk 0 and u1 u2 0. The rest of the proof is similar to the one given for Theorem 6.1. Kolmogorovs theorem says dt = O(1/) 1 + t2
{||>} u
as . However, using the fact that in L1 (dt/(1 + t2 )), u can be approxi mated by functions in Cc (R), the space of innitely dierentiable functions of compact support, we slightly generalize this result. Corollary 14.6 Let u be a real function satisfying

|u(t)| dt < . 1 + t2 dt = o(1/) 1 + t2
Then
{||>} u
as .
Proof. Fix > 0. There is a Cc (R) such that
|u(t) (t)| dt < . 1 + t2
310
By Theorem 14.1, is bounded on R. Note that we used (14.5) as the denition If of . > 2 , then { || > } { | | > /2 }. u u Hence, by Theorem 14.5,
{||>} u
dt 1 + t2
{ ||>/2 } u
dt 8 2 1+t
|u(t) (t)| 8 . dt 2 1+t
Exercises
Exercise 14.4.1 Find a measurable function u such that

|u(t)| dt < , 1 + t2 |(t)| u dt = . 1 + t2 |u(t)| dt < . 1 + t2 x u(x)
but
Exercise 14.4.2
Let
Suppose that there is an > 0 such that
is increasing on R. Show that

x
lim
u(x) = 0. x dt = 0, 1 + t2
Hint: Fix > 0 and let = 1 + . By Corollary 14.6,

n
lim n
{||> n } u
for all large integers n. Hence, there is an xn [ n , n+1 ] such that |(xn )| n . u Note also that xn u(xn ) x u(x) xn+1 u(xn+1 ) for all x [xn , xn+1 ].
311
14.5
M. Rieszs theorem
1 1

Let u Lp (R), 1 < p < , and let U (x + iy) V (x + iy) = = y u(t) dt, (x t)2 + y 2 xt u(t) dt, (x t)2 + y 2 (x + iy C+ ), (x + iy C+ ).
Then, by Corollary 10.12, U hp (C+ ),

y0
lim Uy u
=0
and, as y 0, by Corollary 11.11, U (x + iy) u(x) for almost everywhere x R. For the conjugate function V , Theorem 14.4 says that lim V (x + iy) = u(x)
y0
for almost all x R. In this section our main goal is to show that V hp (C+ ), u Lp (R) and limy0 Vy u p = 0. Theorem 14.7 (M. Rieszs theorem) Let u Lp (R), 1 < p < , and let V (x + iy) = 1

xt u(t) dt, (x t)2 + y 2
(x + iy C+ ).
Then V hp (C+ ), u Lp (R) and there is a constant Cp , just depending on p, such that V p = u p Cp u p . Moreover, V (x + iy) = and
y0
y u(t) dt, (x t)2 + y 2 lim Vy u

p
(x + iy C+ ),
= 0.
Proof. Without loss of generality, assume that u is real. Let U = P u. The proof has several steps. Case 1: u Lp (R), 1 0, Uy
p
F > 0 in the upper half plane. By
u p.
312
On the other hand, if supp u [M, M ], then |V (x + iy)| and |V (x + iy)| which implies |V (z)| = for all z C+ , |z| > 2M . Put M = 1 2
M M
|x t| |u(t)| dt (x t)2 + y 2
1 2
M M
|u(t)| dt
1 y
|x t| |z| + M |u(t)| dt (x t)2 + y 2 (|z| M )2

M M
M M
|u(t)| dt,
|u(t)| dt
1 |z|
|u(t)| dt.
Fix > 0. Therefore, for all x + iy C+ ,

|V (x + iy + i)|p dx 4M (M /)p + 2(12M )p
2M
dx < . xp
In other words, V hp (C+ ). Dene G(z) = F p (z + i). Clearly, G is analytic in C+ and since |G(x + iy)| 2p (|U (x + iy + i)|p + |V (x + iy + i)|p ), we have G H 1 (C+ ). Thus, by Corollary 13.7,
G(x) dx = 0.
(14.11)
Choose (/2p, /2). Since 1 0, |y/x| tan } and = { x + iy : x < 0, |y/x| tan(p) }. (See Figures 14.2 and 14.3.)
313
Fig. 14.2. The domain S .
Fig. 14.3. The domain .
314 It is important to note that |x + iy| and x |x + iy| | cos(p)| Dene x cos
if
x + iy S
if
x + iy .
E = { x R : F (x + i) S }. (See Figure 14.4.)
Fig. 14.4. The set E + i.
Taking the real part of (14.11) gives us F p (x + i) dx +

E R\E
F p (x + i) dx = 0.
(14.12)
If x E, then F (x + i) is in the sector S and thus |F (x + i)| U (x + i) . cos
However, if x E, then F p (x + i) will be in and thus |F (x + i)|p (See Figure 14.5.) F p (x + i) . | cos(p)|
315
Fig. 14.5. The image under the pth power.
The rst inequality implies |F (x + i)|p dx 1 cosp U p (x + i) dx

E
and the second one, along with (14.12), gives |F (x + i)|p dx = Hence,

R\E
1 F p (x + i) dx | cos(p)| R\E 1 F p (x + i) dx | cos(p)| E 1 |F (x + i)|p dx | cos(p)| E 1 U p (x + i) dx. cosp | cos(p)| E
|F (x + i)|p dx
1 cosp
1+
1 | cos(p)|
U p (x + i) dx.
Since |V (x + i)| |F (x + i)|, the desired result follows for this case with the constant 1 Kp = cos 1 1+ | cos(p)|
1 p
Case 2: u Lp (R), 1 < p 2, and u has compact support.
316
We prove a weaker form of the required inequality. Write u = u1 u2 where u1 , u2 0, u1 u2 0 and they have compact support. Thus |u|p = |u1 |p + |u2 |p . Moreover, with obvious notations, V = V1 V2 and thus

|V (x + i)|p dx 2p
|V1 (x + i)|p dx +
|V2 (x + i)|p dx .
Hence, by the result of the preceding case, we obtain

|V (x + i)|p dx
p 2p Kp
|U1 (x + i)|p dx + |u1 (x)| dx +

p p
|U2 (x + i)|p dx
2 =
p Kp p Kp
|u2 (x)|p dx
|u(x)| dx.
Case 3: u Lp (R), 1 < p 2. First, we prove a weaker form of the required inequality. For an arbitrary u Lp (R), 1 0,

p |Vn (x + i)|p dx 2p Kp
|un (x)|p dx.
On the other hand, by Hlders inequality, for each x + i C+ , o Vn (x + i) V (x + i) as n . Hence, by Fatous lemma,

|V (x + i)|p dx
lim inf
n
|Vn (x + i)|p dx

p lim 2p Kp
|un (x)|p dx
p 2p Kp
|u(x)|p dx.
Second, we prove the general case. Fix x0 + iy0 C+ . Let (x + iy) = x0 x , (x0 x)2 + (y + y0 )2 (x + iy C+ ).
317
Clearly, is a bounded harmonic function on the closed upper half plane, and thus, by Theorem 11.2, (x + iy) = More explicitly, we have x0 x 1 = (x0 x)2 + (y + y0 )2

y (t) dt. (x t)2 + y 2 y x0 t 2 dt (x t)2 + y 2 (x0 t)2 + y0
for all x + iy C+ . Therefore, for all y > 0, V (x0 + iy0 + iy) = = = = 1 1 1 1 x0 x u(x) dx (x0 x)2 + (y + y0 )2 1 y x0 t u(x) dx 2 dt (x t)2 + y 2 (x0 t)2 + y0 1 y x0 t u(x) dx 2 dt (x t)2 + y 2 (x0 t)2 + y0 x0 t 2 U (t + iy) dt. (x0 t)2 + y0

Hence, by the result of Case 2,

p |V (x0 + iy0 + iy)|p dx0 2p Kp
|U (t + iy)|p dt.
Let y0 0 and apply Fatous lemma to get the required inequality. Case 4: u Lp (R), 2 < p < . The proof is by duality. Let q be the conjugate exponent of p. Hence 1 < q < 2. Let Lq (R) be of compact support and let (x + iy) = 1

xt (t) dt, (x t)2 + y 2

(x + iy C+ ).
Hence, by Fubinis theorem,
V (x + iy) (x) dx =
= =
xt u(t) dt (x) dx (x t)2 + y 2 1 tx (x) dx u(t) dt (t x)2 + y 2
(t + iy) u(t) dt.
Therefore, by Hlders inequality and the result of Case 3, o

V (x + iy) (x) dx y
Cq
u p.
318 Divide by
q
Chapter 14. The Hilbert transform on R and then take the supremum with respect to all to obtain Vy
p
Cq u p .
Now, the passage from this inequality to Vy

p
Cq Uy
is exactly as explained in Case 3 for the case 1 < p 2. Case 5: Poisson representation. Knowing that V hp (C+ ), 1 < p < , Theorem 11.6 ensures the existence of v Lp (R) such that V (x + iy) = and
y0
y v(t) dt, (x t)2 + y 2 lim Vy v p .
(x + iy C+ ),
On the other hand, by Corollary 11.11,

y0
lim V (x + iy) = v(x)
for almost all x R. Moreover, by Corollary 14.4, we also have

y0
lim V (x + iy) = u
for almost all x R. Hence, u = v and we are done. Corollary 14.8 Let u be real and let u Lp (R), 1 < p < . Let F (z) = Then F H p (C+ ) and F
p
1 i
u(t) dt, tz Kp u p ,
(z C+ ).
where the constant Kp just depends on p. Moreover, f (t) = lim F (t + iy) = u(t) + i(t) u
y0
for almost all t R. Proof. Write F = U + iV . Hence U = P u and V = Q u. By Corollary 10.12, U hp (C+ ). Moreover, the M. Riesz theorem ensures that V hp (C+ ) and Vy p Cp Uy p . Hence, Fy
p
Uy
+ Vy
(1 + Cp ) Uy
(1 + Cp ) u p .
Since U u and V u, almost everywhere on R, the last assertion imme diately follows.
14.5. M. Rieszs theorem Let u be real, u Lp (R), 1 < p < and let F (z) = 1 i

319
u(t) dt, tz
(z C+ ).
By the preceding result, F H p (C+ ), whose boundary values on the real line are given by f = u + i. Thus, by Theorem 13.2, we also have u F (z) = Therefore, F (z) = which is equivalent to F (z) = 1

1 2i 1
u(t) + i(t) u dt, tz
(z C+ ).
u(t) dt, tz
(z C+ ),
xt i u(t) dt + (x t)2 + y 2
y u(t) dt. (x t)2 + y 2
Hence, for almost all t R,

y0
u lim F (t + iy) = u(t) + i(t).
But, we also know that

y0
lim F (t + iy) = u(t) + i(t) u
for almost all t R. Therefore, we obtain the following result. Corollary 14.9 Let u Lp (R), 1 < p < . Then u = u. Combining this result with M. Rieszs theorem immediately implies the following. Corollary 14.10 Let u Lp (R), 1 < p < . Then cp u
p
Cp u p ,
where cp and Cp are constants just depending on p. Corollaries 14.9 and 14.10 show that the map Lp (R) Lp (R) u u is an automorphism of the Banach space Lp (R), 1 < p < . In the following we show that if p = 2, then this map is an isometry.
320
Corollary 14.11 Let 1 < p < , u Lp (R) and v Lq (R), where q is the conjugate exponent of p. Then

u(x) v (x) dx =

u(x) v(x) dx
and
u(x) v(x) dx =
u(x) v (x) dx.
In particular, if u L2 (R), then

|u(x)|2 dx
|(x)|2 dx, u
u(x) u(x) dx
0.
Proof. Without loss of generality, assume that u and v are real-valued. Let F (z) = and G(z) = 1 i 1 i

u(t) dt, tz v(t) dt, tz
(z C+ ),
(z C+ ).
By Corollary 14.8, F H p (C+ ) and G H q (C+ ). Hence F G H 1 (C+ ) with boundary values F (t) G(t) = u(t) + i(t) u Hence, by Corollary 13.7,
v(t) + i(t) . v
F (t) G(t) dt = 0.
The real and imaginary parts of this equality are the required identities.
Exercises
Exercise 14.5.1 U (x + iy) V (x + iy) = = Let u L2 (R), and let 1 1

y u(t) dt, (x t)2 + y 2 xt u(t) dt, (x t)2 + y 2
(x + iy C+ ), (x + iy C+ ).
14.6. The Hilbert transform of Lip(t) functions Show that, for all y > 0,

321
U 2 (x + iy) dx =
V 2 (x + iy) dx.
Exercise 14.5.2 Let u, u L1 (R). Show that u = u. 1 Hint: f = u + i H (R). u Exercise 14.5.3 Let f H p (R), 1 p < . Show that f = if and thus H p (R). f Hint: Use Exercise 14.5.2 and Corollary 14.9. Exercise 14.5.4 Let F H p (R), 1 p < , and let f denote its boundary value function on R. Show that 1 f (t) F (z) = dt, (z C+ ). 2 t z Hint: Use Exercise 14.5.3.
14.6
The Hilbert transform of Lip(t) functions
Let f : R C be a continuous function. The modulus of continuity of f is, as similarly dened in (6.6), f (t) = sup
|xx |t
|f (x) f (x )|.
Of course, the assumption f (t) 0, as t 0, implies that f is uniformly continuous on R. However, even for a bounded continuous function f on R, f (t) does not necessarily tend to zero as t 0. As in the unit circle, f is called Lip , for some with 0 < 1, if f (t) = O( t ) as t 0+ . In this section, our rst goal is to replace t by t(t) , where (t) is a function dened for small values of t, which properly tends to as t 0+ . Then we will study the Hilbert transform of these families. The contents of this section are from [14]. A test function (t) is a real continuous function dened in a right neighborhood of zero, say (0, t0 ), such that, as t 0+ , we have (t)
t 0 t0 t
= + o(1), = = t + o( t(t)+1 ), +1 t(t)+1 + o( t(t)+1 ), 1

(t)+1
( R), ( < + 1), ( > + 1).
( ) d ( ) d
322
The space of all continuous functions f : R C satisfying f (t) = O( t(t) ), (t 0+ ),
will be denoted by Lip(t) . Clearly, the classical space Lip is a special case corresponding to the test function (t) . Given a test function (t), let log (t) = +
t0 t
( )1 d log t .
We show that (t) is also a test function. We call it the test function associated with (t). In the denition of a test function, we put two conditions dealing with cases < + 1 and > + 1. The associated test function is introduced to deal with the troublesome case = + 1. Note that the associated test function is dened so that
t(t) = t t t0
( )1 d.
Lemma 14.12 Let (t) be a test function. Then (t) is also a test function. Proof. Clearly, (t) is a continuous function on (0, t0 ). Our rst task is to show that limt0+ (t) = . For each > 0, we have
t(t) t0
=
t
( )1 d =
t t0 t
t0
( )1 d t(t) + o(t(t) ) .
( )1 d = t
Hence, lim inf t0+ t(t)(t) 1 . Let 0 to get lim t(t)(t) = .
t0+
(14.13)
Let R and let Then, we have
(t) = t(t)+ . (t) = t1+(t) ( t(t)(t) 1 ).
If 0, then < 0 and thus is strictly decreasing. On the other hand, if > 0, by (14.13), (t) > 0 for small values of t and thus is strictly increasing. Fix > 0. For small enough , say 0 < < 0 < min{1, t0 }, we have ( ) + . Hence, 1 ( )1 1 .
14.6. The Hilbert transform of Lip(t) functions For 0 < t < 0 , we thus get
t 0 = 0 t
323
1 d
0 t
( )1 d
0 t
1 d =
t 0 .
Therefore, for 0 < t < 0 , log

t
log
0 t
( )1 d log t
log
log t which implies log lim inf

t0+ t0 t
log t
( )1 d log t lim sup

t0+
log
t0 t
( )1 d log t 0.
Now, let 0 to get log

t0+ t0 t
( )1 d log t = 0.
lim
Hence,
t0+
lim (t) = .
Let < + 1. Then

t 0 ( ) d t
=
0
( ) d t 0
t(t)
d =
t(t)+1 . +1
On the other hand, for 0 < < 1 + , we have

t 0 ( ) d t
=
0
( )+ d t 0
t(t)+
d =
t(t)+1 . +1
Hence,
0
( ) d =
t(t)+1 + o( t(t)+1 ). +1
Finally, let > + 1. Then

t0 t ( ) d t0
=
t
( ) d t0 t
(t)
t(t)+1 . 1
324
On the other hand, for each > 0, we have

0 t ( ) d 0
=
t
( )+ d 0 t
t(t)+
d =
t(t)+1 (1 (t/0 )+1 ). 1+
Hence,
t
t0
( ) d =
t(t)+1 + o( t(t)+1 ). 1
In the following, the notion f (t) constants c, C > 0 such that
g(t), as t 0+ , means that there are
c g(t) f (t) C g(t) in a right neighborhood of zero. Corollary 14.13 Let (t) be a test function with limt0+ (t) = . Let , , C be real constants such that C > 0, > 0 and > 1. Then (t)+1 if < + 1, t t0 ( ) d t(t) if = + 1, + C t 0 1 if > + 1, as t 0+ . Proof. We decompose the integral over two intervals (0, t) and (t, t0 ) and then we use our assumptions on (t). Hence,
t0 0
( ) d + C t
t0
=
t 0
+
t
( ) d + C t
t0 t
+ t( )+1 . t The rst estimate is true since > 1, and the second one holds since < + 1. Similarly, by (14.13), we have
t0 0
0 ( )+1
( ) d + t
( ) d
( )1+ d + C t
t0
=
t 0 0
+ t
t ( )1+
( )1+ d + C t
t0
d +
t t(t) .
( )1+ d
t(t) + t(t)
The last case is proved similarly.
14.6. The Hilbert transform of Lip(t) functions Let (t) = 1 logn+1 1/t log2 1/t log3 1/t n , 2 log 1/t log 1/t log 1/t
325
(14.14)
where , 1 , 2 , . . . , n R and logn = log log log (n times). For this function, we have t(t) = t ( log 1/t )1 ( log2 1/t )2 ( logn 1/t )n . We give a sucient and simple criterion showing at least that the function dened in (14.14) is a test function. Theorem 14.14 Let R and let (t) be a real continuously dierentiable function dened on (0, t0 ) with limt0+ (t) = . Suppose that
t0+
lim (t) t log t = 0.
Then (t) is a test function. Proof. Let d( t(t) ) = ( (t) + (t) t log t ) t(t)1 , dt (t) + (t) t log t as t 0+ . Hence, for < + 1, we have
t 0
and note that
( ) d
= = =
( ) d
t 0
1 t +1 +1 t(t)+1 + o(1) +1
t t 0
0 (t)+1
( ( ) + ( ) log ) ( ) d
( ) d.
Therefore, ( 1 + o(1) )
0
( ) d =
t(t)+1 . +1
Thus the second required condition is satised. For > + 1, we have

t0 t
( ) d
t0
= =
( ) d
=t0 =t
+1
t ( )+1
t0 t
( ( ) + ( ) log ) ( ) d
t0 t
t(t)+1 + o(1) = O(1) +1
( ) d.
326 Hence, ( 1 + o(1) )

t
t0
( ) d =
t(t)+1 . 1
Thus the third condition is also fullled. Corollary 14.15 Let , 1 , 2 , . . . , n be real constants, and let (t) = 1 logn+1 1/t log2 1/t log3 1/t 2 n . log 1/t log 1/t log 1/t
Then (t) is a test function. Moreover,

t(t)
t t ( log 1/t )1+1 ( log2 1/t )2 ( logn 1/t )n
if 1 < 1, if 1 > 1.
If 1 = 2 = = k1 = 1, then
t(t)
t ( logk 1/t )
1+k
t ( logk+1 1/t )k+1 ( logn 1/t )n
if k < 1, if k > 1.
Finally, if 1 = 2 = = n = 1, then
t(t)
t logn+1 1/t.
The following theorems are two celebrated results about the Hilbert transform of bounded Lipschitz functions. Theorem 14.16 (Privalov [16]) If u is a bounded Lip (0 < < 1) function on R, then u(x) exists for all x R and besides u is also Lip . Theorem 14.17 (Titchmarsh [20]) If u is a bounded Lip1 function on R, then u(x) exists for all x R and besides | u(x + t) u(x) | C t log 1/t for all x R and 0 < t < 1/2. We are now able to generalize both theorems. The following theorem is the main result of this section. Theorem 14.18 Let (t) be a test function with 0 < lim (t) 1. +
t0
Let u be a bounded Lip(t) function on R. Then u(x) exists for all x R, and besides Lip(t) if 0 < limt0+ (t) < 1, u is Lip(t) if limt0+ (t) = 1.
14.6. The Hilbert transform of Lip(t) functions Proof. Let V (x + iy) = 1

327
xt t + 2 + y2 (x t) 1 + t2
u(t) dt,
(x + iy C+ ).
To estimate | u(x + t) u(x) |, instead of taking the real line as our straight path to go from x to x + t, we go from x up to x + it, then to x + t + it and nally down to x + t: | u(x) u(x + t) | | u(x) V (x + it) | + | V (x + it) V (x + t + it) | + | V (x + t + it) u(x + t) |. (14.15)
Hence we proceed to study each term of the right side. By (14.5), we have | u(x) V (x + it) | = = = 1 0 lim 1

|x |>
1 + u( ) d x 1 + 2
lim
x u( ) d + (x )2 + t2 1 + 2 x 1 u( ) d x (x )2 + t2 |x |> u(x ) u(x + ) d ( 2 + t2 ) | u(x + ) u(x ) | d. ( 2 + t2 )
t2 0 lim t2
0
Since u is bounded and Lip(t) , there is a constant C such that | u(x + ) u(x ) | C (2 )(2 ) for all values of x R, and small values of , say 0 < < 1. Hence, by Corollary 14.13, we have | u(x) V (x + it) | t2
1 0
C (2 )(2 ) t2 d + ( 2 + t2 )
2 0
2 u d 3
4C t2
(t)
( )1 u 2 t d + 2 + 4t2
C t Hence, for each x R,
+ C t2 .
| u(x) V (x + it) | = O( t(t) ). By the mean value theorem, | V (x + it) V (x + t + it) | t sup
sR
(14.16)
V (s + it) , x
328 where V (s + it) x = = = = Hence, 1 1 1 1

x + 2 + t2 x (x ) 1 + 2 t2 (s )2 u( ) d ( t2 + (s )2 )2 t2 2 u(s + ) d ( t2 + 2 )2
u( ) d
x=s
t2 2 ( u(s + ) u(s) ) d. ( t2 + 2 )2
| V (x + it) V (x + t + it) | sup

sR
t
1 0
Ct The asymptotic behavior of

0 1
( )
| u(s + ) u(s) | d t2 + 2 d + t
1
t2 + 2
2 u 2
d.
t2
( ) d + 2
depends on limt0+ (t). According to Corollary 14.13, we have O( t(t) ) if limt0+ (t) < 1, | V (x + it) V (x + t + it) | = O( t(t) ) if limt0+ (t) = 1. Finally, (14.15), (14.16) and (14.17) give the required result.
(14.17)
In the light of Corollary 14.15 and Theorem 14.18, we get the following special result. Corollary 14.19 Let 1 , 2 , . . . , n R. Let u be a bounded function on R with u (t) = O t ( log 1/t )1 ( log2 1/t )2 ( logn 1/t )n . Then u(x) exists for all values of x R and besides u (t) = O(t) O t ( log 1/t )1+1 ( log2 1/t )2 ( logn 1/t )n if 1 < 1, if 1 > 1.
If 1 = 2 = = k1 = 1, then u (t) = O(t) O t ( logk 1/t )1+k ( logk+1 1/t )k+1 ( logn 1/t )n if k < 1, if k > 1.
Finally, if 1 = 2 = = n = 1, then u (t) = O( t logn+1 1/t ).
14.7. Maximal functions
329
Exercises
Exercise 14.6.1 Let (t) be a test function and let log (t) = (t) + Show that (t) is also a test function. Exercise 14.6.2 Let u(t) = and let (t) = 1 or equivalently, |t| log 1/|t| if 1/e if log2 1/t , log 1/t |t| 1/e, |t| 1/e, (0 < t < 1/e),
t0 t
( )1 d log t .
t(t) = t log 1/t.
Show that u is a bounded Lip(t) function on R and that u(t) u(0) behaves asymptotically like t(t) t (log 1/t)2 , as t 0+ . Hint: Write u(t) u(0) = and show that |(t) V (t + it)| = O( t(t) ), u and |V (it) u(0)| = O( t(t) ) u(t) V (t + it) + V (t + it) V (it) + V (it) u(0)
|V (t + it) V (it)| = O(t(t) ).
Remark: This exercise shows that Corollary 14.19 gives a sharp result.
14.7
Maximal functions
Let u be a measurable function on R. Then the HardyLittlewood maximal function of u is dened by M u(t) = sup 1 k+k
t+k tk
k,k >0
|u(x)| dx,
(t R).
If U is dened on the upper half plane, we dene a family of maximal functions related to U . Remember that a Stoltz domain anchored at t R is of the form (t) = { x + iy C+ : |x t| < y }, ( > 0).
330
Then the nontangential maximal function of U is given by M U (t) = sup and its radial maximal function by M0 U (t) = sup |U (t + iy)|,
y>0 z (t)
|U (z)|,
(t R),
(t R).
If u L1 (dt/(1 + t2 )), we rst apply the Poisson integral formula to extend u to the upper half plane U (x + iy) = 1

y u(t) dt, (x t)2 + y 2
(x + iy C+ ).
Then we will use M u and M0 u respectively instead of M U and M0 U . By denition, we clearly have M0 u(t) M u(t), (t R),
for all > 0. Moreover, since at almost all points of R, U (t + iy) tends to u(t) as y 0 (Lemma 3.10), we also have |u(t)| M0 u(t) for almost all t R. We now obtain a more delicate relation between M u and M u. Lemma 14.20 Let u L1 (dt/(1 + t2 )) and let 0. Then M u(t) for all t R. Proof. We give the proof for > 0. The case = 0 follows immediately. Fix t R. According to the denition of M u

1+
M u(t)
(s) |u(s)| ds 1 0
(s) ds M u(t),
(14.18)
where is the step function (s) =
if |t s| < k, if |t s| k.
By considering the linear combination of such step functions (with positive coefcients) we see that (14.18) still holds for any step function which is symmetric with respect to the vertical axis s = t and is decreasing on s > t. (See Figure 14.6.)
331
Fig. 14.6. A symmetric step function.
Finally, if is any positive and integrable function with the preceding properties, we can approximate it from below by a sequence of such step functions and thus a simple application of Fatous lemma shows that (14.18) indeed holds for any positive and integrable function which is symmetric with respect to the vertical axis s = t and is decreasing on s > t. Now let z = x + iy (t). Then |U (x + iy)| 1

y |u(s)| ds, (x s)2 + y 2
(x + iy C+ ).
But, if x = t, the kernel Py (x s) is not symmetric with respect to the axis s = t. (See Figure 14.7.) Nevertheless, we can overcome this diculty. Without loss of generality assume that x > t. Let (s) = z (s) = 1 y (x s)2 + y 2 if s x, 2t x < s < x, s 2t x.
1 y 1 y (2t x s)2 + y 2
if
if
(See Figure 14.8.) On the one hand, is positive and symmetric with respect to the axis s = t, and on the other hand it majorizes the Poisson kernel. Moreover,
(s) ds = 1 +
2 2|x t| 1+ . y
332
Fig. 14.7. The kernel Py (x s).
Fig. 14.8. The graph of z (s). Hence, for each z = x + iy (t), |U (x + iy)|

(s) |u(s)| ds
1+
M u(t).
Considering the preceding chain of inequalities, we see that, for any function u L1 (dt/(1 + t2 )), |u| M0 u M u 1+ 2 Mu (14.19)
almost everywhere on R. Hence, M u Lp (R) immediately implies that M u, M0 u and u itself are also in Lp (R). For the reverse implication (which holds for 1 < p < ) we need the following celebrated result of HardyLittlewood. A
333
measurable function u on R is locally integrable if it is integrable on each nite interval, i.e.

b a
|f (t)| dt <
for all < a 0, 2 |u(x)| dx. mM u () {M u>} Proof. (F. Riesz) Since M u = M |u|, we may assume that u 0. Let Mr u(x) = sup
k>0
1 k 1 k
x+k
u(t) dt
x x
and M u(x) = sup

k>0
u(t) dt.
xk
The notations Mr and M are temporary for this proof and they do not mean M , as dened before, for a specic value of . Fix > 0 and consider E Er E
b
= { x : M u(x) > }, = { x : Mr u(x) > }, = { x : M u(x) > }.
Since a u is a continuous function of a and b we have Er E and E E. On the other hand, the identity 1 h+k
x+h
u(t) dt =
xk
h h+k
1 h
x+h
u(t) dt
x
k h+k
1 k
u(t) dt
xk
implies E Er E . Hence, E = Er E , which yields mM u () mMr u () + mM u (). We show that mMr u () = and that mM u () = 1 1 |u(x)| dx (14.20)
Er
|u(x)| dx
(14.21)
334 and thus the theorem follows. For each xed k > 0, the function k (x) = 1 k
x+k
u(t) dt,
x
(x R),
is continuous on R, and thus Mr u = supk>0 k is lower semicontinuous on R. Hence, Er is an open set. Let
x
(x) =
0
u(t) dt,
(x R).
Thus is an increasing continuous function on R and the open set Er is a countable union of disjoint intervals Ik . As F. Riesz observed, if one shines light downward on the graph of with slope , then the Ik are the parts of the x-axis where the graph of remains in shadow. From this geometric observation we immediately have |Ik | = (bk ) (ak ) = u(t) dt,
Ik
where Ik = (ak , bk ). Summing up over all Ik gives (14.20). The proof of (14.21) is similar. Theorem 14.21 has several interesting consequences. The rst corollary immediately follows from the theorem. Corollary 14.22 Let u L1 (R). Then, for each > 0, mM u () 2 u 1.
Corollary 14.23 Let u Lp (R), 1 < p < . Then M u Lp and moreover Mu

p
2p u p. p1
Remark: The corollary clearly holds for p = with M u
Proof. Without loss of generality, assume that u 0. Let u(t) if |t| n and |u(t)| n, un (t) = 0 otherwise. Hence, 0 u1 u2 u and, for each x R, un (x) u(x) as n . Therefore, 0 M u1 M u2 M u
335
and, by the monotone convergence theorem, for each x R, M un (x) M u(x) as n . Thus, if the theorem holds for each un , an application of Fatous lemma immediately implies that it also holds for u. Based on the observation made in the last paragraph, we assume that u is bounded and of compact support, say supp u [A, A]. Hence, the rough estimate 1 k+k
x+k xk
u(t) dt
1 |x| A
A A
u(t) dt
4A u |x|
for |x| 2A implies that M u(x)
C , 1 + |x|
with an appropriate constant C. Hence, M u Lp . Therefore, by Theorem 14.21 and by Fubinis theorem, Mu
p p
= p
0
p1 mM u () d p2
{ M u> } M u(x) 0
2p = = 2p
|u(x)| dx
p2 d
|u(x)| dx
2p p1
(M u(x))p1 |u(x)| dx.
Finally, by Hlders inequality, o Mu Since M u

p p p
2p Mu p1
p1 p
u p.
p1 . p
< we can divide both sides by M u
Exercises
Exercise 14.7.1 Let E R with |E| < . Suppose that

|u(x)| log+ |u(x)| dx < .
Show that M u(x) dx 2|E| + 4

|u(x)| log+ |2u(x)| dx.
336 Exercise 14.7.2
Chapter 14. The Hilbert transform on R Let u L1 (R). Suppose that M u(x) dx <
for any measurable E with |E| < . Show that

|u(x)| log+ |u(x)| dx < .
14.8
The maximal Hilbert transform

1 u(t) dt . xt
The maximal Hilbert transform of a function u L1 (dt/(1 + |t|)) is dened by u(x) = sup
>0
|xt|>
According to the M. Riesz theorem (Theorem 14.7), u remains in Lp (R) when ever u Lp (R), 1 < p < . In this section, by showing that u Lp (R) we strengthen the M. Riesz result. Lemma 14.24 Let u L1 (dt/(1 + |t|)), and let V (x + iy) = Then |(x)| (1 + 1/) M u(x) + M0 V (x) u for all x R. Proof. Let u (x) = Fix y > 0 and write V (x + iy) = Hence, |y (x) V (x + iy)| u +
|xt|<y
(x t) u(t) dt, (x t)2 + y 2
(x + iy C+ ).
|xt|>
u(t) dt, xt
(x R).
+
|xt|<y |xt|>y
(x t) u(t) dt. (x t)2 + y 2
|xt|>y
y2 |u(t)| dt |x t|((x t)2 + y 2 ) |x t| |u(t)| dt (x t)2 + y 2 y 1 |u(t)| dt + 2 + y2 (x t) 2y
|xt|>y
|xt|<y
|u(t)| dt.
14.8. The maximal Hilbert transform Therefore, by Lemma 14.20, |y (x) V (x + iy)| M0 |u|(x) + M u(x) (1 + )M u(x). u Thus, |y (x)| (1 + 1/) M u(x) + |V (x + iy)|. u Taking the supremum with respect to y > 0 gives the required result. Corollary 14.25 Let u Lp (R), 1 < p < . Then |(x)| (1 + 1/) M u(x) + M u(x) u for all x R. Proof. By Theorem 14.7 the conjugate harmonic function V (x + iy) = 1

337
(x t) u(t) dt, (x t)2 + y 2 y u(t) dt, (x t)2 + y 2
(x + iy C+ ),
is also the Poisson integral of u, i.e. V (x + iy) = 1

(x + iy C+ ).
Hence M0 V = M0 u and, by Lemma 14.20, M0 u M u. Now, apply Lemma 14.24. Corollary 14.26 Let u Lp (R), 1 < p < . Then u Lp (R), and moreover u
p
Cp u p ,
where Cp is a constant just depending on p. Proof. By Corollary 14.23 Mu and

p
2p u p1
2p u p. p1 However, by the M. Riesz theorem (Theorem 14.7), Mu

p
Kp u
where Kp is a constant just depending on p. Hence, by Corollary 14.25, u

p
(1 + 1/) M u p + M u p 2p ((1 + 1/) u p + u p ) p1 2p ((1 + 1/) + Kp ) u p . p1
338
Appendix A
Topics from real analysis

A.1 A very concise treatment of measure theory
A measure space has three components: the ambient space X, a family M of subsets of X, and a measure . The family M satises the following properties: (a) X M; (b) if E M, then E c , the complement of E with respect to X, is also in M; (c) if En M, n 1, then En M. n=1 In technical terms, we say that M is a -algebra and its elements are measurable subsets of X. The measure is a complex function dened on M such that (a) () = 0, (b) ( En ) = n=1 (En ), for any sequence (En )n1 in M whose elements n=1 are pairwise disjoint. If is a positive measure, we usually allow + as a possible value in the range of . If is a complex measure, the smallest positive measure satisfying |(E)| (E), (E M),
is called the total variation and is denoted by ||. A function f : X C is measurable if f 1 (V ) M for any open set V in C. In particular, each simple function
N
=
n=1
an En ,
where an C and En M, is measurable. We dene

N
d =
X n=1
an (En ). 339
340
Appendix A. Topics from real analysis
For a positive measurable function f , we dene f d = sup

X X
d,
where the supremum is taken over all simple functions such that 0 f . A measurable function f is called integrable provided that |f | d < .
For such a function, we dene f d =

X X
u+ d
u d
X
+i
X
v + d
v d ,
X
where f = u + iv, and for a real function , we put + = max{, 0} and = max{, 0}. The family of all integrable functions is denoted by L1 (). Sometimes we also use L1 (X) if the measure is xed throughout the discussion. The great success of Lebesgues theory of integration is mainly based on the following three convergence theorems. Theorem A.1 (Monotone convergence theorem) Let fn , n 1, be a sequence of measurable functions on X such that 0 f1 f2 , and let f (x) = lim fn (x),
n
(x X).
Then f is measurable and

n
lim
fn d =
f d.
X
Proof. If fn = En , where (En )n1 is an increasing sequence of measurable sets in X and E = En , then the monotone convergence theorem reduces to n=1 (E) = lim (En ),
n
which is rather obvious. This special case in turn implies that

n
lim
En d =
E d,
for each simple function . Now we are able to prove the general case. Since fn fn+1 f ,
X
fn d
fn+1 d
f d.
X
Thus = limn
fn d exists and f d.
X
A.1. A very concise treatment of measure theory
341
If = , we are done. Hence, suppose that < . Let be any simple function such that 0 f and x the constant (0, 1). Let En = { x X : fn (x) (x) }. Clearly, (En )n1 is an increasing sequence of measurable sets with En = X. n=1 Moreover,
X
fn d
En fn d
En d.
Let n to obtain d.
X
Taking the supremum with respect to gives Finally, let 1. Theorem A.2 (Fatous lemma) Let fn , n 1, be a sequence of positive measurable functions on X. Then (lim inf fn ) d lim inf
n n
f d.
X
fn d.
Proof. Let g = lim inf n fn , and let gk = inf fn ,

nk
(k 1).
Hence, 0 g1 g2 and
k
lim gk (x) = g(x),
(x X).
Therefore, by the monotone convergence theorem, g d = lim

X k
gk d.
However, gk fk , which implies

k
lim
gk d lim inf
k
fk d.
Theorem A.3 (Dominated convergence theorem) Let fn , n 1, be a sequence of measurable functions on X such that f (x) = lim fn (x)
n
342
exists for every x X. Suppose that there is a g L1 () such that |fn (x)| g(x), Then f L1 () and
n
(x X, n 1). |fn f | d = 0.
lim
(A.1)
Proof. Since |f | g, we clearly have f L1 (). Let gn = 2g |fn f |, n 1. Hence, by Fatous lemma, 2g d
X
X n n
lim inf gn d gn d |fn f | d.
lim inf =
X
2g d lim sup
n
Since g L1 (), we conclude that lim sup

n X
|fn f | d 0,
and thus we obtain (A.1). Two more advanced results in measure theory are used in the text, which we explain heuristically here. The rst one that we have frequently applied is Fubinis theorem. Let (X, M, ) and (Y, N, ) be two measure spaces and let f : X Y [0, ] be a measurable function. Then the Fubini theorem says f (x, y) d(y)
X Y
d(x) =
Y X
f (x, y) d(x)
d(y).
However, if we only assume that f is a complex measurable function, then some extra assumption is needed. For example, the condition f L1 ( ) is enough to ensure that the preceding identity holds. Suppose that and are two nite positive measures dened on the same -algebra M of an ambient space X. We say that is absolutely continuous with respect to and write if (E) = 0 for every E M for which (E) = 0. At the other extreme, we say that and are mutually singular and write if there is a partition of X, say X = A B, with A, B M and A B = , such that (A) = (B) = 0. Naively speaking, this means that and are living on dierent territories; is concentrated on B and is concentrated on A. If and are complex measures, then, by denition, is absolutely continuous with respect to if || is absolutely continuous with respect to ||. Mutually singular complex measures are dened similarly. Given two complex measures and , the celebrated theorem of Lebesgue RadonNikodym says that there is a unique f L1 () and a unique measure such that d = f d + d.
A.1. A very concise treatment of measure theory This notation is a shorthand to say (E) =
X
343
E f d + (E)
for all E M. In particular, is absolutely continuous with respect to if and only if there is an f L1 () such that d = f d.
Exercises
Exercise A.1.1 Let (M )I be any family of -algebras on X. Put M=
I
M .
Show that M is a -algebra on X. Exercise A.1.2 Let T be any collection of subsets of X. Show that there is a smallest -algebra on X which contains T. Hint: Use Exercise A.1.1. Remark: If T is a topology on X, the elements of the smallest -algebra on X which contains T are called the Borel subsets of X. Exercise A.1.3 i.e. Suppose that the simple function has two representations,
N M
=
n=1
an En =
m=1 M
bm Fm .
Show that
an (En ) =
n=1 m=1
bm (Fm ).
Hint: Without loss of generality, assume that (En )1nN and (Fm )1mM are pairwise disjoint. Then
N N M
an En =
n=1 n=1 m=1
an En Fm
and similarly
M M N
bm Fm =
m=1 m=1 n=1 X
bm Fm En .
Remark: This fact shows that
d is well-dened.
344
A.2
Riesz representation theorems
Let X be a Banach space. The dual space X consists of all continuous linear functionals : X C. The norm of is dened by = sup
x=0
|(x)| . x xX
For each Banach space X, an important question is to characterize its dual X . F. Riesz answered this question for certain spaces of functions. We discuss two such theorems. Let (X, M, ) be a measure space and consider X = Lp (), 1 p < . Let q be the conjugate exponent of p, i.e. 1/p + 1/q = 1. Then the rst theorem says that Lp () = Lq (). This identity means that given any Lp () , there is a unique g Lq () such that (f ) =
X
f g d,
(f Lp ()),
and = g q. On the other hand, if g L () is given and we dene by the given formula, then clearly is a bounded linear functional on Lp () and its norm is equal to g q . Therefore, Rieszs theorem actually provides an isometric correspondence between Lp () and Lq (). This assertion is not valid if p = . As a matter of fact, using the above formula, L1 () can be embedded in L () , but in the general case, this is a proper inclusion. For the second theorem, let X be a locally compact Hausdor space. In our applications, X was either the real line R or the unit circle T. Put X = C0 (X), the space of all continuous functions vanishing at innity equipped with the supremum norm. Note that if X is compact, then we have C0 (X) = C(X). Then we have X = M(X), where M(X) is the space of all complex regular Borel measures on X. Hence, if C0 (X) is given, there is a unique M(X) such that (f ) =
X q
f d,
(f C0 (X)),
and = .
A.3. Weak* convergence of measures
345
On the other hand, if M(X) is given and we dene as above, then is a bounded linear functional on C0 (X) whose norm is equal to . Therefore, Rieszs second theorem provides an isometric correspondence between C0 (X) and M(X).
A.3
Let X be a locally compact Hausdor space. In the preceding section we saw that M(X) is the dual of C0 (X). Hence if and n , n 1, are measures in M(X), we say that n converges to in the weak* topology if
n
lim
dn =
d
X
for all C0 (X). For positive measures on separable spaces, there are several equivalent ways to dene the weak* convergence. The cumulative distribution function of the positive measure M(R) is dened by F (x) = ( (, x] ), (x R). Clearly, F is an increasing right continuous function on R. Moreover,
d =
R
(x) dF (x)
for all C0 (R). In the last identity, the left side is a Lebesgue integral and the right side is a RiemannStieltjes integral. We can also dene the cumulative distribution function for positive Borel measures on T and then obtain
d =
T
(x) dF (x)
for all C(T). Theorem A.4 Let X be a separable locally compact Hausdor space. Let , 1 , 2 , . . . be positive measures in M(X). Then the following are equivalent: (i) n converges to in the weak* topology; (ii) n (X) (X) and, for all open subsets U of X, lim inf n (U ) (U );
n
(iii) n (X) (X) and, for all closed subsets F of X, lim sup n (F ) (F );
n
346
(iv) for all Borel subsets E of X with (E) = 0, where E is the boundary of E in X, lim n (E) = (E);
n
(v) in the case X = R (or X = T), if Fn and F respectively denote the cumulative distribution functions of the measures n and , then
n
lim Fn (x) = F (x)
at all points x R where F is continuous.
A.4
C(T) is dense in Lp (T) (0 < p < )
The space of continuous functions C(T) is dense in Lp (T), 0 < p < . This assertion is not true if p = , since the uniform limit of a sequence of continuous functions has to be continuous and a typical element of L (T) is not necessarily continuous. As a matter of fact, we implicitly proved that C(T) is closed in L (T). If we know that each f Lp (T) can be approximated by step functions
N
an In ,
n=1
where In is an open interval, then we can approximate each In by a trapezoidshaped continuous function, and thus the result is immediate. Even though approximation by step function is possible, it does not come from the rst principles of measure theory. Based on the denition of a Lebesgue integral, each f Lp (T) can be approximated by simple functions
N
s=
n=1
an En ,
where En is a Lebesgue measurable set. At this point, we need two deep results from real analysis. First, given E, a Lebesgue measurable set on T, there is a compact set K and an open set V such that K E V , and the Lebesgue measure of V \ K is as small as we want. Second, by Urysohns lemma, there is a continuous function such that 0 1, |K 1 and supp V . Hence, | E |p d d = |V \ K|.
N
V \K
Given f Lp (T) and > 0, choose the simple function s = n=1 an En such that f s p < . Then, for each n, pick n such that n En p < /(N (1 + |an |)). Put
N
=
n=1
an n .
A.5. The distribution function Therefore, f

p
347
f s
N
+ s
p p
+
n=1
|an | n En
2.
Exercises
Exercise A.4.1 Show that Cc (R) is dense in Lp (R), 1 p < .
Exercise A.4.2 Show that Cc (R) is dense in Lp (R), 1 p < . Hint: Use Exercise A.4.1 and the convolution.
A.5
The distribution function
For a measurable set E T, let |E| denote its Lebesgue measure. Let be a measurable function on T. Then the distribution function of is dened by m () = |{ei : |(ei )| > }|, ( > 0).
Clearly, m is a decreasing function on (0, ). The space weak-L1 (T) consists of all measurable functions on T such that m () C ,
where C is a constant (not depending on , but it may depend on ). Since, for each L1 (T), 2 1 , m () weak-L1 (T) contains L1 (T) as a subspace. Let 0 < p < . On the measure space T [0, ) equipped with the product measure dt d, consider the measurable function 1 if 0 < |(eit )|, it (e , ) = 0 if |(eit )|. Now, using Fubinis theorem, we evaluate the integral I=
T[0,)
p p1 (eit , ) dt d
348 in two ways. On the one hand, we have

|(eit )| 0
I=
p p1 d
dt =
|(eit )|p dt
and on the other hand,
=
0 { ||> }
dt
p p1 d =
0
p p1 m () d.
Therefore, for all 0 < p < and for all measurable functions , we have

|(eit )|p dt =
p p1 m () d.
(A.2)
Exercises
Exercise A.5.1 Let (X, M, ) be a measure space with 0. Let : X C be a measurable function on X. Dene the distribution function of by m () = ( {x X : |(x)| > } ). Show that
X 0
|(x)|p d(x) =
p p1 m () d
for all 0 a}
min{ |(x)|p ap , bp ap } d(x) =
p p1 m () d.
A.6
Minkowskis inequality
Let (, ) and ( , ) be two measure spaces with positive measures and , and let f (x, t) be a positive measurable function on . Dene F (x) =
f (x, t) d(t).
A.7. Jensens inequality
349
Fix p, 1 p < , and let 1/p + 1/q = 1. Then, by Fubinis theorem, for all positive measurable functions G, we have F (x) G(x) d(x) =

f (x, t) d(t)
G(x) d(x) d(t) |G(x)| d(x) d(t).

q
1 q
f (x, t) G(x) d(x) |f (x, t)| d(x)

p
1 p
d(t)
1 p
|f (x, t)|p d(x)
In particular, for G = F p1 E , where E is any measurable set on which |F |p is summable, we have |F (x)|p d(x)
1 p
|f (x, t)|p d(x)
1 p
d(t).
Taking the supremum over E, we get F

p
|f (x, t)| d(x)
1 p
d(t).
Equivalently, for every measurable function f ,

p
|f (x, t)| d(t)
1 p
d(x)
|f (x, t)| d(x)
1 p
d(t).
This is the most generalized form of Minkowskis inequality. If = {1, 2} and is the counting measure, then the last inequality reduces to
p
|f (x, 1)| + |f (x, 2)|
1 p
d(x)
|f (x, 1)|p d(x) |f (x, 2)| d(x)

p
1 p
1 p
which is equivalent to f1 + f2
f1
+ f2
p,
where fn (x) = |f (x, n)|.
A.7
Jensens inequality
Let (a, b) be an interval on the real line R. A real function on (a, b) is convex if the inequality ( (1 t)x + ty ) (1 t)(x) + t(y)
350
holds for each x, y (a, b) and 0 t 1. It is easy to see that this condition is equivalent to (z) (x) (y) (z) (A.3) zx yz whenever a < x < z < y < b. Theorem A.5 (Jensens inequality) Let (X, M, ) be a measure space with 0 and (X) = 1. Let f L1 () and suppose that a < f (x) < b for all x X. Suppose that is convex on (a, b). Then
X
(A.4)
f d
( f ) d.
Proof. Let z = X f d. The condition (A.4) ensures that z (a, b). According to (A.3), there is a constant c such that (y) (z) (z) (x) c zx yz for all x (a, z) and all y (z, b). This implies (w) (z) + c(w z) for each w (a, b). Hence, ( f )(x) (z) + c( f (x) z ) for every x X. Therefore,
X
( f )(x) d(x)
(z) + c( f (x) z )
d(x)
= (z) + c
X
f (x) d(x) z f d .
= (z) =
X
Exercises
Exercise A.7.1 (a, b). Hint: Use (A.3). Let be convex on (a, b). Show that is continuous on
Exercise A.7.2 Let be convex on (a, b). Show that at each t (a, b), has a right and a left derivative. Hint: Use (A.3).
Appendix B
A panoramic view of the representation theorems

We gather here a rather complete list of representation theorems for harmonic or analytic functions on the open unit disc or on the upper half plane. These theorems have been discussed in detail throughout the text. Hence, we do not provide any denition of proof in this appendix. However, we remember that our main objects are elements of the following spaces: hp (D) H p (D) H p (T) and hp (C+ ) H p (C+ ) H p (R) = { U h(C+ ) : U = { F H(C+ ) : F < }, < }, = { U h(D) : U = { F H(D) : F < }, < },
p p
= sup
0r<1 0r<1 p
Ur Fr
p p
= sup
= { f Lp (T) : F H (D) such that Fr f }
p p
= sup Uy
y>0
p p
= sup Fy
y>0
= { f Lp (R) : F H p (C+ ) such that Fy f }.
351
352
Appendix B. A panoramic view of the representation theorems
B.1
hp (D)
h1 (D)
U (rei ) =
T
Let U be harmonic on the open unit disc D.
B.1.1
(a) U h1 (D) if and only if there exists M(T) such that 1 r2 d(eit ), 1 + r2 2r cos( t) (rei D).
(b) is unique and U (re ) =

n= i
(n) r|n| ein ,
(rei D).
The series is absolutely and uniformly convergent on compact subsets of D. (c)
(Z) and
(d) U is a positive harmonic function on D if and only if is a positive Borel measure on T. (e) As r 1, dr (eit ) = U (reit ) dt/2 converges to d(eit ) in the weak* topology, i.e.
r1
lim
(eit ) dr (eit ) =
(eit ) d(eit )
T
for all C(T). (f) U (g) decomposes as

1
= = ||(T).
d(eit ) = u(eit ) dt/2 + d(eit ),
where u L1 (T) and is singular with respect to the Lebesgue measure. Then we have lim U (rei ) = u(ei )
r1
for almost all e
T. 2r sin( t) d(eit ) 1 + r2 2r cos( t) i sgn(n) (n) r|n| ein , (rei D).
(h) The harmonic conjugate of U is given by V (rei ) = =

n=
The series is absolutely and uniformly convergent on compact subsets of D.
B.1. hp (D) (i) For almost all ei T,
353
r1
lim V (rei )
exists. (j) U h1 (D) does not imply that V h1 (D). (k) If U (rei ) = 1 2

1+
r2
(rei D),
with u L1 (T), then the following holds: (i)

r1
lim Ur u
= 0;
(ii) U (iii) u c0 (Z); (iv) u

1
= u 1;
u 1;
(v) the harmonic conjugate of U is given by V (rei ) = =

n=
1 2
2r sin( t) u(eit ) dt 1 + r2 2r cos( t) (rei D);
i sgn(n) u(n) r|n| ein ,
(vi) for almost all ei T,

r1
lim V (rei ) = u(ei );
(vii) for each 0 < p < 1, V hp (D), u Lp (T),

r1
lim Vr u
=0
and u
p
= V
cp u 1 ;
(viii) u L1 (T) does not imply that u L1 (T);
354
Appendix B. A panoramic view of the representation theorems (ix) if u, u L1 (T), then V h1 (D),
r1
lim Vr u
=0
and the harmonic conjugate is also given by V (rei ) = 1 2

1+
r2
(rei D);
(x) if u, u L1 (T), then u(n) = i sgn(n) u(n), (n Z).
B.1.2
hp (D) (1 < p < )

1 2

(a) U hp (D), 1 < p < , if and only if there exists u Lp (T) such that U (rei ) = 1 r2 u(eit ) dt, 1 + r2 2r cos( t) (rei D).
(b) u is unique and U (rei ) =

n=
u(n) r|n| ein ,
(rei D).
The series is absolutely and uniformly convergent on compact subsets of D. (c) If 1 < p 2, then u
q
(Z), where 1/p + 1/q = 1 and u

q
u p.
(d)
r1
lim Ur u
= 0.
(e) U (f) For almost all ei T,

r1 p
= u p.
lim U (rei ) = u(ei ).
(g) U h2 (D) if and only if u L2 (T). If so, U and if U (0) = 0, then U

2 2
= u
= u
= u 2.
B.1. hp (D) (h) u(n) = i sgn(n) u(n), (i) The harmonic conjugate of U is given by V (rei ) = = =
n=
355
(n Z).
1 2 1 2
2r sin( t) u(eit ) dt 1 + r2 2r cos( t) 1+ r2 1 r2 u(eit ) dt 2r cos( t) (rei D).
The series is absolutely and uniformly convergent on compact subsets of D. (j) For almost all ei T,
r1
lim V (rei ) = u(ei ).
(k) V hp (D), u Lp (T), and V
r1
lim Vr u
p
=0
p.
= u
cp U
B.1.3
h (D)
1 2

(a) U h (D) if and only if there exists u L (T) such that U (rei ) = 1+ r2 1 r2 u(eit ) dt, 2r cos( t) (rei D).
(b) u is unique and U (re ) =

n= i
u(n) r|n| ein ,
(rei D).
The series is absolutely and uniformly convergent on compact subsets of D. (c) As r 1, Ur converges to u in the weak* topology, i.e.
r1
lim
(eit ) u(eit ) dt
for all L1 (T). (d) U
= u
356
(e) But, we have

r1
lim Ur u
=0
if and only if u C(T). (f) For almost all ei T,

r1
lim U (rei ) = u(ei ).
(g) The harmonic conjugate of U is given by V (rei ) = = =

n=
1 2 1 2
2r sin( t) u(eit ) dt 1 + r2 2r cos( t) 1 r2 u(eit ) dt 1 + r2 2r cos( t) (rei D).
The series is absolutely and uniformly convergent on compact subsets of D. (h) For almost all ei T,
r1
lim V (rei ) = u(ei ).
(i) u(n) = i sgn(n) u(n), (j) U h (D) does not imply that V h (D). (k) u L (T) does not imply that u L (T). However, for real u, 1 2
u e |(e
it
(n Z).
)|
dt 2 sec( u
),
(0 < /2 u
).
B.2
H p (D)
H p (D) (1 p < )

Let F be analytic on the open unit disc D.
B.2.1
(a) F H p (D), 1 p < , if and only if there exists f H p (T) such that F (rei ) = 1 2 1 r2 f (eit ) dt, 1 + r2 2r cos( t) (rei D).
B.2. H p (D) (b) f is unique and F (z) = = = = =

n=0
357
1 2 i 2 1 2 i 2
f (eit ) dt 1 eit z 2r sin( t) f (eit ) dt + F (0) 1 + r2 2r cos( t) eit + z eit z eit + z eit z f (eit ) dt + i F (0) f (eit ) dt + F (0) (z D).
f (n) z n ,
The series is absolutely and uniformly convergent on compact subsets of D. (c) If f (0) = 0, then f = if or equivalently, f= f (d) If 1 p 2, then f
q
and
f = f.
(Z+ ), where 1/p + 1/q = 1 and f

q
p.
(e)
r1
lim Fr f
= 0.
(f) F (g) For almost all ei T,

r1 p
= f
p.
lim F (rei ) = f (ei ).
(h) F H 2 (D) if and only if f H 2 (T). If so, F and moreover, F provided that F (0) = 0.
2 2
= f
= f
= f
358
B.2.2
H (D)

(a) F H (D) if and only if there exists f H (T) such that F (rei ) = 1 2 1 r2 f (eit ) dt, 1 + r2 2r cos( t) (rei D).
(b) f is unique and F (z) = = = = =

n=0
1 2 i 2 1 2 i 2
f (eit ) dt 1 eit z 2r sin( t) f (eit ) dt + F (0) 1 + r2 2r cos( t) eit + z eit z eit + z eit z f (eit ) dt + i F (0) f (eit ) dt + F (0) (z D).
f (n) z n ,
The series is absolutely and uniformly convergent on compact subsets of D. (c) If f (0) = 0, then f = if or equivalently, f= f and f = f.
(d) As r 1, Fr converges to f in the weak* topology, i.e.

r1
lim
(eit ) F (reit ) dt =
(eit ) f (eit ) dt
for all L1 (T). (e) F (f) But, we have

r1
= f
lim Fr f
=0
if and only if f A(T) = C(T) H 1 (T). (g) For almost all ei T,

r1
lim F (rei ) = f (ei ).
B.3. hp (C+ )
359
B.3
hp (C+ )
h1 (C+ )
1 y d(t), (x t)2 + y 2
Let U be harmonic in the upper half plane C+ .
B.3.1
(a) U h1 (C+ ) if and only if there exists M(R) such that U (x + iy) = (b) is unique and
(x + iy C+ ).
U (x + iy) =
(t) e2y|t|+i2xt dt,
(x + iy C+ ).
The integral is absolutely and uniformly convergent on compact subsets of C+ . (c) C(R) L (R) and
(d) As y 0, dy (t) = U (t + iy) dt converges to d(t) in the weak* topology, i.e.,

y0
lim
(t) dy (t) =
(t) d(t)
R
for all C0 (R). (e) U (f) decomposes as d(t) = u(t) dt + d(t), where u L1 (R) and is singular with respect to the Lebesgue measure. Then we have lim U (x + iy) = u(x)
y0 1
= = ||(R).
for almost all x R. (g) The harmonic conjugate of U is given by V (x + iy) = =
xt d(t) (x t)2 + y 2 (x + iy C+ ).
i sgn(t) (t) e2y|t|+i 2xt dt,
Each integral is absolutely and uniformly convergent on compact subsets of C+ .
360
(h) For almost all x R,

y0
lim V (x + iy)
exists. (i) U h1 (C+ ) does not imply that V h1 (C+ ). (j) If U (x + iy) = 1

y u(t) dt, (x t)2 + y 2
(x + iy C+ ),
with u L1 (R), then the following holds: (i)

y0
lim Uy u U
= 0;
(ii)
1
= u 1;
(iii) u C0 (R); (iv) u
u 1;
(v) the harmonic conjugate of U is given by V (x + iy) = =
xt u(t) dt (x t)2 + y 2 (x + iy C+ ).
i sgn(t) u(t) e2y|t|+i 2xt dt,
Each integral is uniformly convergent on compact subsets of C+ ; (vi) for almost all x R,
y0
lim V (x + iy) = u(x).
(k) u L1 (R) does not imply that u L1 (R). (l) If u, u L1 (R), then V h1 (C+ ),
y0
lim Vy u
=0
and V (x + iy) = 1
R
y u(t) dt, (x t)2 + y 2
(x + iy C+ ).
(m) If u, u L1 (R), then u(t) = i sgn(t) u(t), (t R).
B.3. hp (C+ )
361
B.3.2
hp (C+ ) (1 < p 2)
1

(a) U hp (C+ ), 1 < p 2, if and only if there exists u Lp (R) such that U (x + iy) = (b) u is unique and
y u(t) dt, (x t)2 + y 2
(x + iy C+ ).
U (x + iy) =
u(t) e2y|t|+i2xt dt,
(x + iy C+ ).
The integral is absolutely and uniformly convergent on compact subsets of C+ . (c) u Lq (R), where 1/p + 1/q = 1 and u (d)
y0 q
u p.
lim Uy u
= 0.
(e) U (f) For almost all x R,

y0 p
= u p.
lim U (x + iy) = u(x).
(g) U h2 (C+ ) if and only if u L2 (R). If so, U

2
= u
= u
= u 2.
(h) The harmonic conjugate of U is given by V (x + iy) = = =
1 1
xt u(t) dt (x t)2 + y 2 y u(t) dt (x t)2 + y 2 (x + iy C+ ).
i sgn(t) u(t) e2y|t|+i 2xt dt,
Each integral is absolutely and uniformly convergent on compact subsets of C+ . (i) For almost all x R,
y0
362
(j) V hp (C+ ), u Lp (R),

y0
lim Vy u
=0
and V (k) u(t) = i sgn(t) u(t), (t R).

p
= u
cp U
p.
B.3.3
hp (C+ ) (2 < p < )

1

(a) U hp (C+ ), 2 < p < , if and only if there exists u Lp (R) such that U (x + iy) = y u(t) dt, (x t)2 + y 2 (x + iy C+ ).
(b) u is unique. (c)

y0
lim Uy u
= 0.
(d) U
p
= u p.
(e) The harmonic conjugate of U is given by V (x + iy) = = 1 1

xt u(t) dt (x t)2 + y 2 y u(t) dt, (x t)2 + y 2
(x + iy C+ ).
Each integral is absolutely and uniformly convergent on compact subsets of C+ . (f) For almost all x R,
y0
(g) V hp (C+ ), u Lp (R),

y0
lim Vy u
=0
and V
p
= u
cp U
p.
B.3. hp (C+ )
363
B.3.4
h (C+ )
1

(a) U h (R) if and only if there exists u L (R) such that U (x + iy) = (b) u is unique. (c) As y 0, Uy converges to u in the weak* topology, i.e.
y0
y u(t) dt, (x t)2 + y 2
(x + iy C+ ).
lim
(t) U (t + iy) dt =

(t) u(t) dt
for all L1 (R). (d) U (e) A sucient condition for

y0
= u
lim Uy u
=0
is u C0 (R). (f) The harmonic conjugate of U is given by V (x + iy) = 1

xt t + (x t)2 + y 2 1 + t2
u(t) dt,
(x + iy C+ ).
(g) For almost all x + iy R,

y0
(h) U h (C+ ) does not imply that V h (C+ ). Similarly, u L (R) does not imply that u L (R).
B.3.5
h+ (C+ )
(a) U is a positive harmonic function on C+ if and only if there exists a positive Borel measure on R with d(t) < 1 + t2
and a positive constant such that U (x + iy) = y + 1 y d(t), (x t)2 + y 2 (x + iy C+ ).
364 (b) If so,
lim
U (iy) = y (t) d(t)

R
and
y0
lim
(t) U (t + iy) d(t) =

R
for all Cc (R). (c) The harmonic conjugate of U is given by V (x + iy) = 1 xt t + (x t)2 + y 2 1 + t2 lim V (x + iy) d(t), (x + iy C+ ).
(d) For almost all x R,

y0
exists. (e) h+ (C+ ) is not a subclass of h1 (C+ ).
B.4
H p (C+ )
H p (C+ ) (1 p 2)
1

Let F be analytic in the upper half plane C+
B.4.1
(a) F H p (C+ ), 1 p 2, if and only if there exists f H p (R) such that F (x + iy) = (b) f is unique and F (z) = = = = =
0
y f (t) dt, (x t)2 + y 2
(x + iy C+ ).
i 1 2i 1 i 1
xt f (t) dt (x t)2 + y 2 f (t) dt tz f (t) dt tz

f (t) dt tz (z = x + iy C+ ).
f (t) ei2zt dt,
Each integral is absolutely and uniformly convergent on compact subsets of C+ .
B.4. H p (C+ ) (c) f = if , or equivalently f= f and f = f.
365
(d) f Lq (R+ ), where 1/p + 1/q = 1 and f (e)

y0 q
p.
lim Fy f
= 0.
(f) F (g) For almost all x R,

y0 p
= f
p.
lim F (x + iy) = f (x).
(h) F H 2 (C+ ) if and only if f H 2 (R). If so, F

2
= f
= f
= f 2.
B.4.2
H p (C+ ) (2 < p < )

1

(a) F H p (C+ ), 2 < p < , if and only if there exists f H p (R) such that F (x + iy) = (b) f is unique and F (z) = = = = i 1 i 1 1 2i xt f (t) dt 2 2 (x t) + y f (t) dt t z f (t) dt tz f (t) dt, tz

y f (t) dt, (x t)2 + y 2
(x + iy C+ ).
(z = x + iy C+ ).
(c) f = if , or equivalently f= f (d)

y0
and lim Fy f
f = f.
= 0.
366 (e)
F (f) For almost all x R,

y0
= f
p.
lim F (x + iy) = f (x).
B.4.3
H (C+ )
1

(a) F H (R) if and only if there exists f H (R) such that F (x + iy) = (b) f is unique and F (z) = = = = i 1 2i 1 i 1

y f (t) dt, (x t)2 + y 2
(x + iy C+ ).
xt t f (t) dt + F (i) + (x t)2 + y 2 1 + t2 1 1 f (t) dt tz t+i 1 1 f (t) dt + i F (i) tz t+i 1 1 f (t) dt + F (i), (z = x + iy C+ ). tz t+i
(c) If F (i) = 0, then f = if or equivalently, f= f and f = f.
(d) As y 0, Fy converges to f in the weak* topology, i.e.

y0
lim
(t) F (t + iy) dt =

(t) f (t) dt
for all L1 (R). (e) F (f) A sucient condition for

y0
= f
lim Fy f
=0
is f A0 (R) = C0 (R) H (R). (g) For almost all x R,

y0
lim F (x + iy) = f (x).
Bibliography
[1] N. Bary, A Treatise on Trigonometric Series, Macmillan, 1964. [2] J. Bergh, J. Lfstrm, Interpolation Spaces, An Introduction, Springero o Verlag, 1976. [3] W. Blaschke, Eine Erweiterung des Satzes von Vitali uber Folgen analytis cher Funktionen, Leipzig. Ber. 67 (1915). [4] J. Conway, Functions of One Complex Variable, Vol. I, Second Edition, Springer-Verlag, 1978. [5] P. Duren, Theory of H p Spaces, Academic Press, 1970. [6] P. Fatou, Sries trigonomtriques et sries de Taylor, Acta Math. 30 (1906) e e e 335400. [7] G. Hardy, The mean value of the modulus of an analytic function, Proc. London Math. Soc. 14 (1915) 269277. [8] G. Hardy, W. Rogozinski, Fourier Series, Cambridge Tracts No. 38, 1950. [9] G. Hardy, Divergent Series, Clarendon Press, 1949. [10] K. Homan, Banach Spaces of Analytic Functions, Prentice-Hall, 1962. [11] G. Julia, Sur les moyennes des modules de fonctions analytiques, Bull. Sci. Math., 2:51 (1927) 198214. [12] Y. Katznelson, An Introduction to Harmonic Analysis, Third Edition, Cambridge University Press, 2004. [13] P. Koosis, Introduction to H p Spaces, Second Edition, Cambridge Tracts No. 115, 1998. [14] J. Mashreghi, Generalized Lipschitz functions, Comput. Meth. Funct. Theor. 5:2 (2005) 431444. [15] R. Paley, N. Wiener, Fourier Transform in the Complex Domain, AMS College Publication No. 19, 1934. 367
368
BIBLIOGRAPHY
[16] I. Privalov, Intgrale de Cauchy, Bull. lUniv. Saratov (1918). e [17] T. Ransford, Potential Theory in the Complex Plane, London Mathematical Society, Student Texts No. 28, Cambridge University Press, 1995. [18] F. Riesz, Uber die Randwerte einer Analytischen Funktionen, Math. Z. 18 (1923) 8795. [19] K. Seip, Interpolation and Sampling in Spaces of Analytic Functions, University Lecture Series, Volume 33, AMS, 2004. [20] E. Titchmarsh, Introduction to the Theory of Fourier Integrals, Third Edition, Chelsea Publishing, 1986. [21] N. Wiener, The Fourier Integral and Certain of its Applications, Cambridge University Press, 2002. [22] A. Zygmund, Trigonometric Series, Cambridge, 1968.
Index
AbelPoisson means, 21, 22, 33, 49 accumulation point, 86, 155, 159162, 294 set, 158, 161, 294 analytic measure, 116, 118, 119, 129, 284, 286 approximate identity, 25, 27, 28, 30, 32, 34, 36, 39, 43, 47, 111, 229, 231, 232, 235, 237, 238, 242, 243 positive, 2527, 31, 32, 43, 71, 230, 231, 239 arithmeticgeometric inequality, 98 automorphism, 69, 155 Baire category theorem, 110 Banach algebra, 14, 16 space, 57, 59, 207, 208, 247 Bernstein, 29 Bessels inequality, 5052 binomial theorem, 169 Blaschke, 159, 294 factor, 156, 158, 293 product, 160, 162, 163, 166, 167, 181, 185, 293295, 297 nite, 158, 160, 165, 167 innite, 159, 160, 294 sequence, 160, 293, 295 Cantors method, 63 Carleson, 50, 131 Cauchy, 3, 56, 275 integral formula, 121, 248, 269, 279 Riemann equations, 3, 56, 137, 138, 150 369 Schwarz inequality, 275 sequence, 52, 194, 267 Ces`ro means, 24 a characteristic function, 191 Chebyshevs inequality, 194 compact set, 6, 22, 23, 32, 55, 57, 79, 81 83, 91, 118, 119, 121, 207, 226, 227, 250, 275, 352, 354 362, 364 support, 8, 208, 222, 234, 241, 257, 311, 315317 topological space, 90 complete space, 50, 267 conformal mapping, 69, 124, 132, 207, 253, 258, 306 conjugate exponent, 17, 18, 139, 197199, 221, 222, 224, 273, 274, 276, 288, 289, 317 function, 64, 66, 111, 123, 125, 131, 311 Poisson integral, 78, 115 Poisson kernel, 78, 80, 111, 115, 229, 268 series, 57 convergence, 8, 73, 119, 210, 257, 265 absolute, 23, 61, 226, 275 bounded, 257 dominated, 112, 214, 226, 250, 256, 257, 264 measure, 194 monotone, 136, 265 norm, 43, 45, 48, 52, 120, 121, 194, 241, 244, 267, 270, 282 pointwise, 22, 32, 34, 232
370 uniform, 10, 11, 22, 23, 32, 33, 39, 55, 57, 61, 79, 113115, 121, 167, 226, 227, 232, 234, 237, 257 weak*, 39, 40, 42, 4749, 120, 238, 239, 241, 243, 245, 246 convex function, 95, 97, 150, 190 convolution, 13, 14, 16, 39, 188, 219, 223, 238 cumulative distribution function, 345 de la Valle Poussins kernel, 30, 217, e 231 dense, 8, 19, 33, 43, 47, 62, 139, 191, 192, 209, 241, 243, 267 Dini continuous functions, 145 Dinis theorem, 84 Dirac measure, 120 Dirichlet kernel, 25, 27 distribution function, 347 dominated convergence theorem, 112, 214, 226, 250, 256, 257, 264 dual, 62, 65, 66, 192, 208, 317 embedding, 7 entire function, 193 Eulers identity, 9 extremal problem, 165 Fatous construction, 37, 116 function, 116, 118, 119 lemma, 93, 133, 154, 316, 317 theorem, 61, 70, 73, 123, 258 Fejrs e kernel, 26, 29, 33, 34, 44, 215, 231 lemma, 45 theorem, 34 FejrRiesz inequality, 183 e Fischer, 51 Fourier analysis, 8 coecients, 5, 8, 11, 21, 29, 42, 44, 45, 49, 66, 205 integral, 209, 230
INDEX Plancherel transform, 266268, 270272 series, 1, 8, 21, 22, 27, 33, 34, 49, 64, 66 transform, 68, 14, 105, 188, 197, 198, 207, 209, 218, 225, 226, 263, 264, 266, 267, 273, 287, 288 Fubinis theorem, 13, 18, 39, 47, 93, 150, 218, 219, 221, 238, 244, 317, 347 Gausss theorem, 88 GaussWeierstrass kernel, 215, 231 Green function, 249 Greens theorem, 137, 138, 151, 249 Hlders inequality, 17, 19, 59, 84, o 139, 169, 182, 193, 195, 196, 203, 221, 222, 251, 272, 316, 317 generalized, 17 Hadamards three-circle theorem, 98, 190 three-line theorem, 189, 190, 192, 194, 195, 202 Hahns decomposition theorem, 125, 258 HahnBanach theorem, 196 Hardy, 99 Hardy space, 59, 104, 116, 155, 166, 167, 200, 208, 247, 279, 288, 294, 298 Hardys convexity theorem, 97, 98, 190 Hardys inequality, 181, 205 HardyLittlewood inequality, 205, 206 harmonic analysis, 239 conjugate, 55, 59, 7779, 111, 113, 118, 136, 139, 228, 248, 352, 355, 356, 359, 361364 function, 2, 3, 9, 22, 23, 26, 33, 35, 37, 40, 43, 44, 48, 55, 5860, 62, 67, 72, 83, 84,
INDEX 88, 115, 218, 226, 233, 242, 247, 258, 275, 317 positive, 66, 68, 255 majorant, 84 Harnacks distance, 68 inequality, 6769 theorem, 65, 70 HausdorYoung theorem, 197, 199, 273, 288 Herglotzs theorem, 67, 70, 255 Hilbert space, 5, 7, 50, 53, 207 Hilbert transform, 118 Hunt, 50 ideal, 15 inner function, 162, 163 inner product, 5, 7, 50, 105, 207 inversion formula, 263, 265, 272 Jacksons kernel, 31, 231 Jensens formula, 162, 163, 165, 166, 170 inequality, 84, 150, 166 Julia, G., 168 Kolmogorovs example, 50 theorem, 131, 133, 134, 308, 309 Laplace equation, 2, 56 operator, 22, 90, 137, 226 Laplacian, 1, 2 Lebesgue, 44, 45, 126, 205, 264 constant, 28 decomposition theorem, 73 measure, 5, 15, 23, 72, 73, 75, 92, 105, 118, 119, 123, 207, 220, 226, 254, 285, 347, 352, 359 point, 71 space, 5, 27, 187, 207, 221 Liouville, 69 Lipschitz, 6, 321 Littlewoods subordination theorem, 99
371 Mbius transformation, 156 o maximum principle, 84, 95, 96, 156, 157, 159, 168, 191 measurable function, 5, 7, 17, 187, 207, 221, 254, 347 measure Borel, 6, 11, 14, 27, 39, 66, 67, 74, 77, 208, 227, 238, 239, 254, 256258, 260, 261 positive, 22, 39, 68, 70, 76, 214, 254256, 258, 261, 352, 363 Dirac, 120 Lebesgue, 5, 15, 23, 72, 73, 75, 92, 118, 119, 123, 207, 220, 226, 254, 285, 347, 352, 359 signed, 74, 77, 260, 261 singular, 72, 73, 119, 352, 359 space, 17, 92, 187, 191, 197, 199, 200, 347 Minkowkis inequality, 242 modulus of continuity, 19, 145, 148 monotone convergence theorem, 89, 136, 265 multiplication formula, 218, 225, 226, 271, 272, 274, 276 nontangential limit, 76, 261 norm, 6, 43, 45, 59, 120, 121, 187, 208, 241, 244, 247, 267, 268 operator, 187, 188 convolution, 45, 188, 223 dierential, 22 Fourier, 188, 268 Laplace, 22, 137, 226 translation, 222 type, 187, 188, 191, 197, 200, 204, 223, 274 unitary, 271, 272 orthonormal set, 50 Parsevals identity, 49, 52, 65, 135, 182, 188, 197, 205 Phragmen-Lindelf, 189 o Plancherel, 266268, 270272
372 theorem, 266, 274, 288, 289 Poisson integral, 21, 66, 67, 225 Jensen formula, 163, 165 kernel, 9, 21, 25, 33, 43, 67, 124, 207, 210, 225, 229, 230, 233, 236, 239, 242 representation, 60, 62, 65, 66, 123, 124, 248, 250, 252, 255 pullback, 254 Riemann, 3, 44, 45, 56, 138, 150, 264 RiemannLebesgue lemma, 16, 44, 45, 199, 205, 264 Riesz projection, 141 Rieszs theorem, 14, 37, 40, 48, 64, 66, 118, 120, 135, 140, 153, 163, 167, 219, 239, 244, 284, 285, 311 Riesz, F., 40, 48, 51, 64, 66, 105, 118, 120, 123, 155, 163, 167, 183, 208, 219, 239, 244 Riesz, M., 105, 118, 120, 123, 137, 140, 153, 188, 189, 191, 196, 198, 200, 221, 223 RieszFischer theorem, 51, 52 RieszThorin interpolation theorem, 188, 189, 191, 196, 198, 200, 221, 223 Schwarz, 275 Schwarzs lemma, 156, 157 reection principle, 88 simple function, 191, 192, 194, 339, 340, 343, 346 Smirnov, 177 space Lebesgue, 5, 27, 187, 207, 221 separable, 62, 65 sequence, 7 spectrum, 11, 29, 30, 79, 116, 119, 212, 271 Stein, P., 137 Stoltzs domain, 76, 261
INDEX subharmonic function, 8385, 87, 88, 9092, 96 radial, 91, 93, 95 subordinate, 99 Szeg, 123 o Taylor coecients, 155 Thorin, 188, 189, 191, 196, 198, 200, 221, 223 topological space, 81, 90 total variation, 6, 208 translation, 208, 222 triangle inequality, 82 trigonometric identity, 111 polynomial, 29, 33, 45, 139 polynomials, 44 series, 9, 57 uniqueness theorem, 41, 5052, 123, 240, 275, 287289 upper semicontinuous, 8183, 85, 87, 88, 90, 92, 93, 95 Weierstrass, 33, 37, 155, 231 Youngs inequality, 1618, 189, 196, 221, 223, 266 theorem, 197, 199, 273 Zygmunds theorem, 149, 152

Representation Theorems in Hardy Spaces

Transféré par

Informations du document

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Representation Theorems in Hardy Spaces

Transféré par

Droits d'auteur :

Formats disponibles

London Mathematical Society Student Texts

Managing Editor: Professor D. Benson, Department of Mathematics, University of Aberdeen, UK

Representation Theorems in Hardy Spaces

v To My Parents: Masoumeh Farzaneh and Ahmad Mashreghi

Fig. 1.1. The open unit disc D and its boundary T. 1

2 and the upper half plane C+ = { z C : z > 0}

Chapter 1. Fourier series

Fig. 1.2. The upper half plane C+ and its boundary R.

1.1. The Laplacian and

Chapter 1. Fourier series Let U be harmonic on the annular domain

A(R1 , R2 ) = { z C : 0 R1 < |z| < R2 }.

1.2. Some function spaces and sequence spaces

Some function spaces and sequence spaces

Let f be a measurable function on T, and let f and f

M : |{ eit : |f (eit )| > M }| = 0 ,

f (eit ) g(eit ) dt,

Chapter 1. Fourier series

This space is equipped with the norm f Lip (T) = f

1.2. Some function spaces and sequence spaces

Z Then, for 0 < p , we dene

is a Hilbert space. The subspaces

8 Exercise 1.2.2 Exercise 1.2.3 Exercise 1.2.4 Show that

Chapter 1. Fourier series (Z+ ), 0 < p , is closed in

Show that c0 (Z) is closed in

Show that c0 (Z+ ) is closed in c0 (Z).

which can be rewritten as f

Therefore, the Fourier transform L1 (T) f

1.3. Fourier coecients

Chapter 1. Fourier series

1 r2 1 + r(eit + eit ) 1 r2 it )(1 reit ) (1 re 1 1 + 1 it 1 re 1 reit r2

(|w| < 1),

= r (See Figure 1.4.)

Fig. 1.4. The spectrum of Pr .

1.3. Fourier coecients

(n Z). (Z) and

|eint | d||(eit ) d||(eit ) = ||(T) = .

Chapter 1. Fourier series

Based on the preceding lemma, the Fourier transform M(T)

eit + z f (eit ) dt, eit z

Hint: Note that

Show that g (n) =

f (eit ) f (ei(t+/n) ) eint dt.

Nevertheless, by Fubinis theorem,

Therefore, we necessarily have

|f (ei ) g(ei(t ) )| d <

for almost all eit T. This observation enables us to dene (f g)(eit ) =

14 (ii) associative: f (g h) = (f g) h, (iii) distributive: f (g + h) = f g + f h, (iv) homogenous: f (g) = (f ) g = (f g),

Chapter 1. Fourier series

f (ei(s ) ) g(ei ) f (eit ) dt 2

(ei(t+ ) ) d(eit ) d(ei ),

and moreover, Hence, is dened such that

(ei(t+ ) ) d(eit ) d(ei ),

for all C(T), and it satises . (1.9)

eint d( )(eit ) ein(t+ ) d(eit ) d(ei )

ein d(ei ) = (n) (n).

Proof. According to (1.8), for each C(T) we have

(ei(t+ ) ) d(eit ) d(ei )

(ei(t+ ) ) d(eit ) f (ei ) d

Chapter 1. Fourier series

1.5. Youngs inequality

Fig. 1.5. The level curves of p.

18 and Lp (T). Write |f (ei ) g(ei(t ) )| =

Chapter 1. Fourier series

|g(ei(t ) )|1 p |f (ei )|1 p |f (ei )| p |g(ei(t ) )| p