Introduction to Fourier Analysis and Wavelets Mark A. Pinsky Graduate Studies in Mathematics Volume 102 (G7) American Mathematical Society =This page intentionally left blankTitles in This Series 102 Mark A. Pinsky, Introduction to Fourier analysis and wavelets, 2009 101 Ward Cheney and Will Light, A course in approximation theory, 2009 100 I. Martin Isaacs, Algebra: A graduate course, 2009 99 Gerald Teschl, Mathematical methods in quantum mechanics: With applications to Schrédinger operators, 2009 98 Alexander I. Bobenko and Yuri B. Suris, Discrete differential geometry: Integrable structure, 2008 97 David C. Ullrich, Complex made simple, 2008, 96 N. V. Krylov, Lectures on elliptic and parabolic equations in Sobolev spaces, 2008 95 Leon A. Takhtajan, Quantum mechanics for mathematicians, 2008 94 James E. Humphreys, Representations of semisimple Lie algebras in the BGG category ©, 2008 93 Peter W. Michor, Topics in differential geometry, 2008 92. I. Martin Isaacs, Finite group theory, 2008 91 Louis Halle Rowen, Graduate algebra: Noncommutative view, 2008 90 Larry J. Gerstein, Basic quadratic forms, 2008 89 Anthony Bonato, A course on the web graph, 2008 88 Nathanial P. Brown and Narutaka Ozawa, C*-algebras and finite-dimensional approximations, 2008 87 Srikanth B. Iyengar, Graham J. Leuschke, Anton Leykin, Claudia Miller, Ezra Miller, Anurag K. Singh, and Uli Walther, Twenty-four hours of local cohomology, 2007 86 Yulij Ilyashenko and Sergei Yakovenko, Lectures on analytic differential equations, 2007 85 John M. Alongi and Gail S. Nelson, Recurrence and topology, 2007 84 Charalambos D. Aliprantis and Rabee Tourky, Cones and duality, 2007 83 Wolfgang Ebeling, Functions of several complex variables and their singularities (translated by Philip G. Spain), 2007 82 Serge Alinhac and Patrick Gérard, Pseudo-differential operators and the Nash—Moser ‘theorem (translated by Stephen S. Wilson), 2007 81 V. V. Prasolov, Elements of homology theory, 2007 80 Davar Khoshnevisan, Probability, 2007 79 William Stein, Modular forms, a computational approach (with an appendix by Paul E. Gunnells), 2007 78 Harry Dym, Linear algebra in action, 2007 77 Bennett Chow, Peng Lu, and Lei Ni, Hamilton’s Ricci flow, 2006 76 Michael E. Taylor, Measure theory and integration, 2006 75 Peter D. Miller, Applied asymptotic analysis, 2006 7A V. V. Prasolov, Elements of combinatorial and differential topology, 2006 73. Louis Halle Rowen, Graduate algebra: Commutative view. 2006 72 R. J. Williams, Introduction the the mathematics of finance, 2006 71 S. P. Novikov and I. A. Taimanov, Modern geometric structures and fields. 2006 70 Sedn Dineen, Probability theory in finance. 2005 69 Sebastian Montiel and Antonio Ros, Curves and surfaces. 2005 For a complete list of titles in this series, visit the AMS Bookstore at www.ams.org/bookstore/.This page intentionally left blankIntroduction to Fourier Analysis and WaveletsThis page intentionally left blankIntroduction to Fourier Analysis and Wavelets Mark A. Pinsky Graduate Studies in Mathematics Volume 102 American Mathematical Society Providence, Rhode IslandEDITORIAL COMMITTEE David Cox (Chair) Steven G. Krantz Rafe Mazzeo Martin Scharlemann 2000 Mathematics Subject Classification. Primary 42-02; Secondary 42C40. For additional information and updates on this book, visit www.ams.org/bookpages/gsm-102 Library of Congress Cataloging-in-Publication Data Pinsky, Mark A., 1940- Introduction to fourier analysis and wavelets / Mark A. Pinsky. p. cm. — (Graduate studies in mathematics ; v. 102), Originally published: Pacific Grove, CA : Brooks/Cole, 2002. Includes bibliographical references and index. ISBN 978-0-8218-4797-8 (alk. paper) 1. Fourier analysis. 2. Wavelets (Mathematics) I. Title, QA403.5.P56 2009 515/.2433—de22 2008047419 Copying and reprinting. Individual readers of this publication, and nonprofit libraries acting for them, are permitted to make fair use of the material, such as to copy a chapter for use in teaching or research. Permission is granted to quote brief passages from this publication in reviews, provided the customary acknowledgment of the source is given. Republication, systematic copying, or multiple reproduction of any material in this publication is permitted only under license from the American Mathematical Society. Requests for such permission should be addressed to the Acquisitions Department, American Mathematical Society, 201 Charles Street, Providence, Rhode Island 02904-2294, USA. Requests can also be made by e-mail to [email protected]. © 2002 held by the American Mathematical Society, All rights reserved. ‘The American Mathematical Society retains all rights ‘except those granted to the United States Government. Printed in the United States of America. @ The paper used in this book is acid-free and falls within the guidelines established to ensure permanence and durability. Visit the AMS home page at http: //uww.ams.org/ 10987654321 14181211 1009To my parents, Harry A. Pinsky and Helen M. Pinsky, who led me to the path of learningThis page intentionally left blankCONTENTS 1 FOURIER SERIES ON THE CIRCLE Ld 12 13 Motivation and Heuristics Lid 112 113 Lid LLS 1.1.6 Motivation from Physics 1.1.1.1 The Vibrating String 1.1.1.2 Heat Flow in Solids Absolutely Convergent Trigonometric Series *Examples of Factorial and Bessel Functions Poisson Kernel Example *Proof of Laplace’s Method *Nonabsolutely Convergent Trigonometric Series Formulation of Fourier Series 121 1.2.2 1.23 1.24 1.25 1.2.6 124 1.2.8 Fourier Coefficients and Their Basic Properties Fourier Series of Finite Measures *Rates of Decay of Fourier Coefficients 1.2.3.1 Piecewise Smooth Functions 1.2.3.2 Fourier Characterization of Analytic Functions Sine Integral 1.2.4.1 Other Proofs That Si(oo) = 1 Pointwise Convergence Criteria *Integration of Fourier Series 1.2.6.1 Convergence of Fourier Series of Measures Riemann Localization Principle Gibbs-Wilbraham Phenomenon 1.2.8.1 The General Case Fourier Series in L? 13.1 1.3.2 Mean Square Approximation—Parseval’s Theorem, * Application to the Isoperimetric InequalityX CONTENTS 1.3.3 *Rates of Convergence in L? 1.3.3.1 Application to Absolutely-Convergent Fourier Series 1.4 Norm Convergence and Summability 1.4.1 Approximate Identities 1.4.1.1 Almost-Everywhere Convergence of the Abel Means 14.2 Summability Matrices 1.4.3 Fejér Means of a Fourier Series 1.43.1 Wiener’s Closure Theorem on the Circle 1.4.4 *Equidistribution Modulo One 1.4.5 “Hardy's Tauberian Theorem 1.5 Improved Trigonometric Approximation 1.5.1 Rates of Convergence in C(T) 1.5.2 Approximation with Fejér Means 1.5.3 “Jackson's Theorem 1.5.4 “Higher-Order Approximation 1.5.5. *Converse Theorems of Bernstein 1.6 Divergence of Fourier Series 1.6.1 ‘The Example of du Bois-Reymond 1.6.2 Analysis via Lebesgue Constants 1.6.3 Divergence in the Space L! 1,7 *Appendix: Complements on Laplace’s Method 1.7.0.1 First Variation on the Theme-Gaussian Approximation 1.7.0.2 Second Variation on the Theme-Improved Error Estimate 1.7.1 *Application to Bessel Functions 1.7.2. *The Local Limit Theorem of DeMoivre-Laplace 1.8 Appendix: Proof of the Uniform Boundedness Theorem 1.9 *Appendix: Higher-Order Bessel functions 1.10 Appendix: Cantor’s Uniqueness Theorem 2 FOURIER TRANSFORMS ON THE LINE AND SPACE, 2.1 Motivation and Heuristics 2.2. Basic Properties of the Fourier Transform 2.2.1 Riemann-Lebesgue Lemma 2.2.2 Approximate Identities and Gaussian Summability 2.2.2.1 Improved Approximate Identities for Pointwise Convergence 2.2.2.2 Application to the Fourier Transform 2.2.2.3 The n-Dimensional Poisson Kernel 39 43 45 45 49 51 54 57 57 59 61 61 62 65 66 10 B 14 15 78 80 80 80 81 82 84 85 86 89 89 91 94 97 100 102 1062.3 24 25 2.6 27 contents 2.2.3. Fourier Transforms of Tempered Distributions 2.2.4 *Characterization of the Gaussian Density 2.2.5 *Wiener’s Density Theorem Fourier Inversion in One Dimension 2.3.1 Dirichlet Kernel and Symmetric Partial Sums 2.3.2. Example of the Indicator Function 2.3.3 Gibbs-Wilbraham Phenomenon 2.3.4 Dini Convergence Theorem 2.3.4.1 Extension to Fourier’s Single Integral 2.3.5 Smoothing Operations in R'—Averaging and Summability 2.3.6 Averaging and Weak Convergence 2.3.7 Cesiro Summability 2.3.7.1 Approximation Properties of the Fejér Kernel 2.3.8 Bernstein's Inequality 2.3.9 *One-Sided Fourier Integral Representation 2.3.9.1 Fourier Cosine Transform 2.3.9.2 Fourier Sine Transform 2.3.9.3. Generalized h-Transform L? Theory in R" 2.4.1 Plancherel’s Theorem 2.4.2. *Bernstein’s Theorem for Fourier Transforms 2.4.3. The Uncertainty Principle 2.4.3.1 Uncertainty Principle on the Circle 2.4.4 Spectral Analysis of the Fourier Transform 2.4.4.1 Hermite Polynomials 2.4.4.2 Eigenfunction of the Fourier Transform, 2.4.4.3 Orthogonality Properties 2.4.4.4 Completeness Spherical Fourier Inversion in R" 2.5.1 Bochner’s Approach 2.5.2 Piecewise Smooth Viewpoint 2.5.3 Relations with the Wave Equation 2.5.3.1 The Method of Brandolini and Colzani 2.5.4 Bochner-Riesz Summability 2.5.4.1 A General Theorem on Almost-Everywhere Summability Bessel Functions 2.6.1 Fourier Transforms of Radial Functions 2.6.2 L?-Restriction Theorems for the Fourier Transform 2.6.2.1 An Improved Result 2.6.2.2 Limitations on the Range of p The Method of Stationary Phase 2.7.1 Statement of the Result 2.7.2 Application to Bessel Functions 2.7.3 Proof of the Method of Stationary Phase 2.7.4 Abel’s Lemma xi 108 109 110 112 112 4 M5 AS 117 47 118 119 121 122 124 124 125 125 128 128 129 131 133 134 134 136 137 138 139 139 145 146 149 152 153 154 157 158 159 161 162 163 164 165 167xii contents 3 FOURIER ANALYSIS IN L? SPACES 3.1 3.2 3.3 34 35 3.6 37 3.8 3.9 Motivation and Heuristics The M. Riesz-Thorin Interpolation Theorem 3.2.0.1 Generalized Young’s Inequality 3.2.0.2. The Hausdorff-Young Inequality 3.2.1 Stein's Complex Interpolation Theorem ‘The Conjugate Function or Discrete Hilbert Transform 3.3.1 L? Theory of the Conjugate Function 3.3.2 L! Theory of the Conjugate Function 3.3.2.1 Identification as a Singular Integral The Hilbert Transform on R 3.4.1 L? Theory of the Hilbert Transform 3.4.2. L? Theory of the Hilbert Transform, 1 < p < 00 3.4.2.1 “Applications to Convergence of Fourier Integrals 3.4.3 _L! Theory of the Hilbert Transform and Extensions 3.4.3.1 Kolmogorov’s Inequality for the Hilbert Transform 3.4.4 Application to Singular Integrals with Odd Kernels, Hardy-Littlewood Maximal Function 3.5.1 Application to the Lebesgue Differentiation Theorem 3.5.2 Application to Radial Convolution Operators 3.5.3 Maximal Inequalities for Spherical Averages The Marcinkiewicz Interpolation Theorem. Calderén-Zygmund Decomposition A Class of Singular Integrals Properties of Harmonic Functions 3.9.1 General Properties 3.9.2 _ Representation Theorems in the Disk 3.9.3 Representation Theorems in the Upper Half-Plane 3.9.4 Herglotz/Bochner Theorems and Positive Definite Functions 4 POISSON SUMMATION FORMULA AND MULTIPLE FOURIER SERIES 4A 42 43 Motivation and Heuristics The Poisson Summation Formula in R! 4.2.1 Periodization of a Function 4.2.2. Statement and Proof 4.2.3 Shannon Sampling Multiple Fourier Series 43.1 Basic L' Theory 4.3.1.1 Pointwise Convergence for Smooth Functions 4.3.1.2 Representation of Spherical Partial Sums 169 169 169 174 174 175 176 7 179 183 184 185 186 187 188 192 194 197 200 202 203 206 209 210 212 212 214 216 219 222 222 223 223 225 228 230 231 233 23344 45 46 47 coNTENTS 4.3.2. Basic L? Theory 4.3.3. Restriction Theorems for Fourier Coefficients Poisson Summation Formula in R4 4.4.1 *Simultaneous Nonlocalization Application to Lattice Points 4.5.1 Kendall’s Mean Square Error 4.5.2. Landau’s Asymptotic Formula 4.5.3. Application to Multiple Fourier Series 4.5.3.1 Three-Dimensional Case 4.5.3.2. Higher-Dimensional Case ‘Schrédinger Equation and Gauss Sums 4.6.1 Distributions on the Circle 4.6.2 The Schrédinger Equation on the Circle Recurrence of Random Walk APPLICATIONS TO PROBABILITY THEORY 5.1 Motivation and Heuristics 5.2. Basic Definitions 5.2.1 The Central Limit Theorem 5.2.1.1 Restatement in Terms of Independent Random Variables 5.3. Extension to Gap Series 5.3.1 Extension to Abel Sums 5.4. Weak Convergence of Measures 5.4.1 An Improved Continuity Theorem 5.4.1.1 Another Proof of Bochner’s Theorem 5.5 Convolution Semigroups 5.6 The Berry-Esséen Theorem 5.6.1 Extension to Different Distributions 5.7 The Law of the Iterated Logarithm INTRODUCTION TO WAVELETS 6.1 Motivation and Heuristics 6.1.1 Heuristic Treatment of the Wavelet Transform 6.2 Wavelet Transform 6.2.0.1 Wavelet Characterization of Smoothness 6.3. Haar Wavelet Expansion 6.3.1 Haar Functions and Haar Series 6.3.2 Haar Sums and Dyadic Projections 6.3.3 Completeness of the Haar Functions 6.3.3.1 Haar Series in Cy and L,, Spaces 6.3.3.2 Pointwise Convergence of Haar Series xiii 235 236 238 239 241 241 243 244 245 247 247 248 250 252 256 256 256 260 261 262 266 268 269 270 272 276 279 280 284 284 285 286 290 291 291 292 295 296 298xiv conTeNts 63.4 63.5 6.3.6 63.7 *Construction of Standard Brownian Motion *Haar Function Representation of Brownian Motion *Proof of Continuity *Lévy's Modulus of Continuity 6.4 Multiresolution Analysis 64.1 6.4.2 6.4.3 644 64.5 Orthonormal Systems and Riesz Systems Scaling Equations and Structure Constants From Scaling Function to MRA 6.4.3.1 Additional Remarks Meyer Wavelets From Scaling Function to Orthonormal Wavelet 6.4.5.1 Direct Proof that V; © Vo Is Spanned by (WO —Dheex 6.4.5.2 Null Integrability of Wavelets Without Scaling Functions 6.5 Wavelets with Compact Support 65.1 6.5.2 65.3 6.5.4 From Scaling Filter to Scaling Function Explicit Construction of Compact Wavelets 6.5.2.1 Daubechies Recipe 6.5.2.2 Hernandez-Weiss Recipe ‘Smoothness of Wavelets, 6.5.3.1 A Negative Result Cohen’s Extension of Theorem 6.5.1 6.6 Convergence Properties of Wavelet Expansions 6.6.1 6.6.2 Wavelet Series in L” Spaces 6.6.1.1 Large Scale Analysis 6.6.1.2 Almost-Everywhere Convergence 6.6.1.3 Convergence at a Preassigned Point Jackson and Bernstein Approximation Theorems 6.7 Wavelets in Several Variables 6.7.1 6.7.2 6.7.3 References Notations Index ‘Two Important Examples 6.7.1.1 Tensor Product of Wavelets General Formulation of MRA and Wavelets in R¢ Notations for Subgroups and Cosets Riesz Systems and Orthonormal Systems in R* 6.7.2.3 Scaling Equation and Structure Constants 6.7.2.4 Existence of the Wavelet Set 6.7.2.5 Proof That the Wavelet Set Spans Vi © Vo 6.7.2.6 Cohen’s Theorem in R¢ Examples of Wavelets in R¢ 299 301 301 302 303 304 310 313 315 318 319 324 325 326 327 330 331 333 334 336 338 341 341 345 346 347 347 352 352 354 354 355 356 357 358 361 362 362 365 369 373Figure 1.1.1 Figure 1.1.2 Figure 1.2.1 Figure 1.2.2 Figure 1.2.3 Figure 1.4.1 Figure 1.4.2 Figure 1.5.1 Figure 2.5.1 Figure 2.5.2 Figure 6.1.1 Figure 6.1.2 LIST OF FIGURES Poisson kernel P, (9) with r = 0.8 [Page 8] Conjugate Poisson kernel Q,(9) with r = 0.8 [Page 9] The Dirichlet kernel Dy (u) with N = 5 [Page 16] Graphs of the partial sums Syf (x) for N = 1, 2, 3 of the Fourier series of f(x) =x, —1 An sin nx cos(nt — a). Functions of this form can be used to satisfy a variety of initial conditions, if we are given the values of f and the partial derivative df/8t when 1 = 0. This is possible whenever f(x,0) and af/at(x,0) are expressed as finite linear com- binations "a, sinnx. This may be less obvious in other cases; for example 12 INTRODUCTION TO FOURIER ANALYSIS AND WAVELETS f(x,0) = sin’x = (3sinx — sin3x)/4, whereas sin? x cannot be so expressed. In order to work with these trigonometric sums, we note the property of orthogonality, expressed as 1.3) fF sins sins ds = 0, mén. 0 If a function has the form f(x) F(x) sin nx dx = 0 forn > N. Dik axsin kx, then we must have Exercise 1.1.1. Show that if N is odd, si form Yt, ay sin kx. Exercise 1.1.2. Suppose that we have a convergent series expansion of sin’ x = DZ, ax sin kx on the interval 0 C,Dre' nee A large source of examples of absolutely convergent trigonometric series is obtained from power series. Consider a general Laurent series (1.1.22) $0 = Saget + bac Fa n=6 1TRODUCTION TO FOURIER ANALYSIS AND WAVELETS. assumed to be absolutely convergent in an annular region ry < |z| < ro. Then (1.1.23) Fore) = arte + Soba, or xr 0 and n=1,2,.... Then F(@) =e" and the Fourier coefficient formula (1.1.15) and Parseval identity (1.1.17) specialize to (1.1.24) Simoes = 2 Ses (ie eee terol (1.1.25) wan:=¥(5) ene 0 Equation (1.1.25) gives an integral formula for the modified Bessel function /9(2r) defined by the power series on the left side. In particular we will determine the asymptotic behavior of /o(2r) when r —> co by analysis of the integral on the right side of (1.1.25). Meanwhile (1.1.24) gives a useful representation of the factorial function. We will use this to present a self-contained treatment of Stirling’s formula in the form (1.1.26) lim nt/n"ien" = VI. ‘To obtain this result we take r = n in (1.1.24), to obtain 1 * F (8), +f erae, on fie where F (0) = e!” —1—i0. This is set up to apply the Laplace asymptotic method, whose proof will be given at the end of the section, and whose statement follows. (1.1.27) Proposition 1.1.10. Suppose that F(6), —x <6 < x isa continuous complex- valued function so that Re F() has a unique maximum at 6 = 0 with F(0) = 0, limp. F(0)/6? = —k with Rek > 0. Then (1.1.28) f eras = [+ 0( - nk In the present case F(@) = e"” — 1 — i6, so the conditions are satisfied with Re F (9) = cos@ — 1 |6| > Oand k = —F"(0)/2 = 1/2.FOURIER SERIES ON THECIRCLE 7 Exercise 1.1.11. Prove that the factorial function satisfies the two-sided system of inequalities C\n"e" (E02) en ne nat tek ‘We will prove later that the class of absolutely convergent trigonometric series contains the class of Hilder continuous functions of exponents greater than 3. 1.1.4. Poisson Kernel Example A second useful example of an absolutely convergent trigonometric series is generated by the function f(z) = 1/(1 — 2), defined in the unit disc |z| < 1. Thus we obtain the absolutely convergent series 1 (1.1.32) Oye ena Ore ee sl When we take the real and imaginary parts, we obtain the real series ‘cos 6 2 = Il gin icra eae 14° cosno pty é oat ne rsind Inj ind ees rere be sinnd = sgn(nje™, where the signum function sgn (n) is defined by setting sgn (0) = 0, sgn (n) = lifn > 1 and sgn (n) = —1 if n < —1. These can be rewritten as (1.1.33) pe esas seas Yolen wm 1+9? = 2reosé , nee 2rsind 3 (0) : =~ In git (1.1.34) 2,00) = Ta geass = sempre8 INTRODUCTION TO FOURIER ANALYSIS AND WAVELETS P, is the Poisson kernel and Q, is the conjugate Poisson kernel. See Figures 1.1.1 and 1.1.2. It is easily checked that both P, and Q, are solutions of Laplace's equation. Equations (1.1.15) and (1.1.17) yield 1; (1.1.35) zfe ™P.(0)d0 = rl", neZ, O on Je 4 Setting n = 0 in (1.1.35) yields eae 1-? (1.1.39) 1 ad, O —k when x — xo € (a, b) with Re(k) > 0. We will prove that when f > 00 > BO) dye = @fB00) z 1 (1.1.40) cw) [acve! dx = (seo/%+0(4)). Proof. Without loss of generality we can assume that xp = 0 and B(0) = 0.10 TRODUCTION T0 FOURIER ANALYSIS AND WAVELETS Step 1. Localizing the range of integration: For any 5 > 0 the maximum of Re B(x) when |x| > 4 is negative so that we can write C= J Awe de + O(€) acs for some D > 0, Step 2. Replacing B by its quadratic approximation: Now we write the expansion of B(x) around x = 0: BQ) =—k? +007), k= ke tik. For any two complex numbers 2) =x; + iy1, 22 = x2 + iy we have the inequality (141) le — e| < [zy —ale®, 23 = max(), 2). Applying this with z, = B(x), z2 = —kx?, we can take z; < —kex?/2 by taking 5 small enough. With this choice of 5 we can write je) 0 so that A(x) ~ A(O)| < € when |x| < 4. The error made in replacing A(x) by A(0) is bounded by const ef et de ce [lene axe Uy 7 fi also by making the substitution v = x./7, From Steps 1, 2 and 3 we have - 1 cw =a0) fe de+o( 4) (0) = AU » fie not Step 4. Integral over the real line: Finally we replace the integral on —3 < x < 6 by the integral over —o0 < x < 00 with an exponential error, since if em asl <5 f lente” dx Wi>8 8 Sinise 1 eer sss Big ssssssoe But the integral over —00 < x < co is (1.1.42) f om de= /F If k is complex, we must take the square root with positive real part, since the integral clearly has this property. Combining these steps completes the proof. . Exercise 1.1.13. If Re k > 0, show that the real part of f°, e-* dx is positive.FOURIER SERIES ON THE CIRCLE 11 Hint: First’ make the change of variable x°=y to reduce consideration to Ig (e-® /y') cos (by) dy, which is an infinite sum of integrals over (k7/b, (k + 1)x/b). Show that the first of these is positive and the remaining terms alternate in sign and decrease to zero. Alternatively, one may prove this by using complex analysis, considering the analytic function k > $(k) = Jy" e™ dx, defined for Re k > 0. On the real axis $(k) agrees with the branch of the function k > (7r/k)'?, which is positive and real on the positive real axis. Hence these two analytic functions agree on the entire halfplane Rek > 0. Exercise 1.1.14. Suppose, in addition, that A(x) is Lipschitz continuous: IAG) — AQ)| < Kx — y| and that the second derivative B"(x) exists and is also Lipschitz continuous. By going through the steps of the above proof, show that in this case the error term o(1/-/t) in Laplace's method can be replaced by O(1/t). Computation of (1.1.42). This is obtained by considering the square of 1:= Jze-™ de in polar coordinates. Thus 1.1.6 *Nonabsolutely Convergent Trigonometric Series Itis possible to deal with trigonometric series with monotonically decreasing coefficients by the method of summation by parts to produce convergent series. Given any sequence of complex numbers a,,n = 0, 1,2,..., define (Ad), = a, —a,-1 forn > 1. The basic identity is that for any two sequences ay, by ¥ ” (1.1.43) ayby ~ayby = > ax(Abe + D> b-(Aah, = M O and Ay > Any1 with lim, = 0. Then the trigonometric series 7°) Ane is convergent for x # 0.12. TRODUCTION TO FOURIER ANALYSIS AND WAVELETS, Proof. We apply (1.1.43) with a, = Ay, by = By(x) = Y>j_o €"". This is a finite geometric sum with eines 1 1.1.44) Bs una (1.1.44) 1B,(2)| | eae | = ait) Applying (1.1.43) shows that W y (1.145) DE aie = ayByx) — ay Bute) — Y> Brs()(Aas ats Sts ‘The first two terms tend to zero when M, N —+ oo. The sum is estimated by v v 1.1.46) By-\(x)(Aa),| < |Aag| = (ay — av) ———— d an |S SnD Dy Vail = Cau ~ av ee72) which tends to zero when M, N — 00. : Example 1.1.17. The trigonometric series ,..€/logn is convergent for x #0. By taking the real and imaginary parts, we see that the series Ye2 cos mx/logn is convergent for x # 0 and the series >, sinnx/logn is convergent for all x. Exercise 1.1.18. Prove that we have uniform convergence of ,.9€"*/logn on any closed interval not containing x = 0. Inan appendix to this chapter we prove the basic Cantor uniqueness theorem, which allows one to identify the coefficients A,, from the sum of the (conditionally convergent) trigonometric series. Summation by parts can also be used to estimate the modulus of continuity of an absolutely convergent trigonometric series in terms of the rail sum, defined as DY (Ad + Aad. To do this, let h > 0 and write Fe +H) — FQ) =P Ane™(e™ — 1), so LPG +H) — FOO S YO nha +257 lA mw ah The second sum is 2Ey. The first sum is rewritten as DY nha, = HY MAB, = —-hNEy +038, Wen n= so that oa fe WM) — FO] 1. Exercise 1.1.22. Illustrate the results of the previous exercise in the following three case Yo2 cos(3"2), S33" cost2"s), 22 cos(2"s) =o = = 1.2. FORMULATION OF FOURIER SERIES Armed with some motivation, we now begin the formal study of Fourier series. 1.2.1 Fourier Coefficients and Their Basic Properties We begin with an integrable function f on T = (—7, 77]. Any such function can naturally be identified with a 27-periodic function on the entire real line. This extension is helpful in many ways, especially in computing integrals since we may write /,, f(x +6) d0 = Jy £(0) a8. ‘The Fourier coefficients, or discrete Fourier transform, of f € L'(T) are defined by the formula hr 1 ind (1.2.1) j= 5 [se a8. This is a linear transformation from the space L'(T) to the space ["°(Z) of bounded bilateral sequences. We formalize this as a proposition. Proposition 1.2.1. The space of discrete Fourier transforms is an algebra, as expressed by (1.2.2) A) +h = (iF A) (1.2.3) A) xf) = ff) @)14 wTRODUCTION TO FOURIER ANALYSIS AND WAVELETS where the convolution of two L' functions is defined by (haAO= 2 | Herre -o4e at Sy and where the integral is convergent for almost every 9. The convolution product is commutative and associative: fi * fs = fr *fir (fi +f) *fs = fi * (h fa)» Furthermore the mapping f — f is a contraction, meaning that (124) ool = f iseniao. Proof. The first formula is immediate from the linearity of the integral. To prove the second formula, we note that the convolution is well defined by the Fubini theorem, since 21 fi *fill®) = fz \A@)@ — $)| dd. The latter integral has a finite integral over T, since by Fubini and a change of variable we find that [9 [i@nrne-orae= ff in@nreraeae = f ineiae [ rerde ‘Therefore the integral defining the convolution of |fi| * |/o| is finite almost everywhere, and dominates the convolution fi * f2. Having made these preparations, we can multiply the Fourier coefficients and transform by Fubini: anti xf = [roe [roe = [em ([ ion —4e) ay which was to be proved. The commutative and associative properties are most easily deduced from these properties of ordinary multiplication, once we have used Fubini to identify the convolution as the unique L' function F with the property that for any bounded function h, Sy Sg hx + fi @) dx dy = 20 f, h(@)F (2) dz. The contraction property is immediate from (1.2.1). . Exercise 1.2.2. Let fi,fo €L'(T). Show that for any bounded measurable function h we have ff [ret niconoracay=2n [nari * fa)(2) dz. tS IT One may be tempted to conclude that the convolution fi * f; is represented by the Fourier series 7,71 (m)f2(n)e”. This cannot be literally true in general, because the latter series does not converge pointwise or in L',, for a general L' function. Nevertheless the following special case of Fourier reciprocity is quite useful. Proposition 1.2.3. Suppose that S427 |Ay| < 00 and we have an absolutely convergent Fourier series K@) = Yonex Ane’. Then for any f € L\(T) we have (1.2) (F*O) = DAfne”. ned Proof. Begin with the identity £0 - KO) = TAne"*F6 - 0) nedFOURIER SERIES ONTHECIRCLE 15 By hypothesis, this is a series of L' functions that is convergent and bounded pointwise by the integrable function |f| x S,¢z [An So that by Lebesgue’s dominated convergence theorem we can integrate term-by-term to conclude [10 -oxcoae ™ ya. fewre —o)do oat 7 Da fewrronay sat =2n Asie” wet which was to be proved. . This can be immediately applied to the Poisson kernel K() = P,() = Lneg Me™. If Dynex f (nen? is the Fourier series of some L' function f, we may now assert that (fF PO) = forte. a Another example is obtained from the Dirichlet kernel, defined by the finite trigono- metric sum w (1.2.6) Dy() = Doe". The Fourier partial sum operator w Suf = Fe”? = Ym oof men? a nee is equivalently obtained as the convolution (1.2.7) Syf @) = (Dw *f)(0). Here 1, is the indicator function, defined by 14(n) = 1 if n © A and equals zero otherwise. From (1.2.6), it is clear that Dy is an even function and that Dy = 27. See Figure 1.2.1 for the Dirichlet kernel and Figure 1.2.2 for the Fourier partial sums of SQ) =x,-m K x f is self-adjoint, meaning that for any f € L'(T) and any bounded function g we have J(K *f)g = Jf (K * 8).16 _wTRODUCTION TO FOURIER ANALYSIS AND WAVELETS FIGURE 1.2.1 The Dirichlet kernel Dy (8) for N = 5. From M. Pinsky, Partial Differential Equations and Boundary-Value Problems with Applications. Reprinted by permission of The McGraw-Hill Companies, y FIGURE 1.2.2 ‘The graphs of the partial sums f(@) for N = 1, 2,3 of the Fourier series of f(@) = From M. Pinsky, Partial Differential Equations and Boundary-Value Problems with Applications. Reprinted by permission of The McGraw-Hill Companies, om cB en, Exercise 1.2.6. Suppose that K € L!(T) is an odd function: K(—0) = —K (0). Prove that the convolution operator is skew-adjoint, meaning that for any f €L'(Z) and any bounded function g we have [y(K * f)g = — [yf (K * 2). Atthis point we can formulate the uniqueness of the Fourier coefficients, following an elementary argument of Lebesgue. The result will also be deduced as a corollary of the summability of one-dimensional Fourier series in Section 1.3.FOURIER SERIESONTHE CIRCLE 17 Proposition 1.2.7. Suppose that, g € L'(T) have the property that f(n) = 3(n) foralln € Z. Thenf = ae. Proof. By the linearity of the map f — f it suffices to prove the result in case g = 0. We first prove the result in case f is a continuous function. Writing f = w+ iv, we have for any neZ, 0 = 2nf(n) = f (u(x) + iv@ye"™ de -[ (u(x) cos mx + v(x) sin nx) dx + if (u(x) cos mx — u(x) sin nx) dr For any m = 0, 1, 2, .... we apply this ton = m and n = ~m to conclude that [ sereosmeac= 0 f u(x) sin mx de, f vo) cosmdr=0= f u(x) sin mx de. ‘Therefore we are reduced to the case of a real-valued continuous function f for which Osis area (1.29) ff foscosmeds [reosinmsa, If f is not identically zero, there exists a point x where f(t) # 0. Replacing f by F(x + x9) /f 0), we may assume that f(x) = } in a closed interval J = [—5, 5], where 0 <5 < m. Let r(x) = 1 + cosx — cosd, Ty(x) = t(x)", Clearly r(x) = 1 on I while |1@)| < Lon, so that T,(x) > 1 on J while T,(x) + 0 on I . Now T, is a trigonometric polynomial of degree n, so that from the hypothesis (1.2.9) 0 f FO)T a(x) de = 0. ‘On the other hand, the dominated convergence theorem shows that lim, Se FO)T, 00) dx = 0, Subtracting these, we conclude that lim, f,f(x)T,(x) dx = 0, which contradicts the fact that f(x)T,(x) = } on J. Hence f(x) = 0. Applying this argument separately to f = u and (f =v proves the result for any complex-valued f € C(T). Now if f € L'(T) satisfies f(n) = 0, let F(x) = J", f() dt, a continuous function. Interchanging the orders of integration shows further that for n = +1, +2, .. fe Fae de = fe eae ( [’ soa) a = [10 CE a) dt “LoS )4 Hence F(n) = 0 for 0 #n € Z. Letting Ay = 1/2 /*, F(x) dx, we can apply the proof in the previous paragraph to the continuous function x -> F(x) — Ap, to conclude that F(x) ~Ap = 0. But from Lebesgue’s theorem on the differentiation of the integral, we have almost everywhere that f(x) = (d/dx)F(x) = 0, which completes the proof. .18 INTRODUCTION TO FOURIER ANALYSIS AND WAVELETS, The above properties of the mapping f — f are algebraic in nature. The following fundamental property is analytic. It is valid for an arbitrary approach to infinity, not restricting n to be integer-valued. Here we note that the Fourier coefficients (1.2.1) are well defined for any n € R. Theorem 1.2.8. Riemann-Lebesgue lemma: For any f € L'(T), Timyp)-+00(n) = 0 and the convergence is uniform on compact subsets of L!(T). We will give two separate proofs of the first statement, then deal with the compactness. First Proof. By making the change of variable ¢ = @ + x/n, the integral defining the Fourier coefficient is transformed into anfin) = [fo +ximenort ag = [re +m jnyen"* do. Ir Adding this to the definition of f(n), we obtain 4nf(n) = f LG + x/m) — Fedyle™* do. I; If is a continuous function, the integrand tends to zero uniformly when |n| —> 00, hence f(n) = 0. Incase f is an arbitrary L! function, we can find a continuous function g so that If ~ lly 00 whereas the second term is less than ¢, by virtue of (1.2.4). Thus lim sup, [/(n)| < €, which was arbitrary, completing the proof. . Second proof. The result is clearly true for the indicator function of an interval (a,b), since for n #0 (1.2.10) +0, Inf > oo. Hence it is also true for the indicator function of a finite union of intervals. Now if E is any measurable set, by the definition of outer measure, there exists a finite union of intervals so that the symmetric difference EAE has measure less than €. In terms of L' norms Selle — lel < €. By linearity, we have WTen)| < ITE@)1 + Trae ‘The first term tends to zero while the second term is less than ¢. Now we can extend to simple functions f = > a1, by linearity and finally to all L' functions by appealing to the density of simple functions. For example fy = D> .-yov K2-" Liaw eycursn-m is a simple function so that J-,|f — fi] > 0 when N — oo. To prove the uniform convergence, let K be a compact set in L'(T). We first cover K by a union of balls: K C UB(f,¢). By compactness, we can extract a finite set fi, .... iy so that K C Ul, B(f;, €). For each | < i < N we can apply the Riemann-Lebesgue lemma to conclude |/;(n)| < € for n > M and all i, 1 < i < N. Now any other member of K is included in one of the balls B,, so that by the contraction property, we must have lim sup, [f(n)| < €. This proves the uniform convergence. .FOURIER SERIES ON THECIRCLE 19 This property of uniform convergence is easily applied to show that iff ¢ L'(T) and g is a bounded function, then (L211 [eerre+ me aa +0 Ir uniformly for $1 < $ < $2 when |n| > 00. To see this, note that for any f € L'(T) the mapping x + f, from T to L'(T) is continuous. In detail 0. lim f(x + 8) —f (xo + 6)|d0 son J ‘Multiplication by the bounded function g preserves this continuity. The continuous image of a compact set is compact. Hence we can apply the uniform convergence on compacts to deduce that (1.2.11) holds uniformly on compact ¢ intervals. 1.2.2 Fourier Series of Finite Measures The concept of a Fourier series can be easily extended from the class of integrable functions to the class of finite signed measures. Recall that a signed measure on T is defined by a function of bounded variation, which can be represented as the differ- ence of two monotone functions. The sum of two signed measures is the setwise sum: (4 + H2)(A) = 11(A) + 42(A). The convolution of two signed measures 411, 42 is, by definition, the signed measure jz with the property that for every continuous function he C(t) h(x + y) dur (x) dur) = 20 f A(z) duc). TXT Ip ‘The Fourier coefficients of a signed measure are defined by : tf juny = fe" due TOIT Proposition 1.2.1 carries over with no essential change. Proposition 1.2.9. The space of discrete Fourier transforms is an algebra, as expressed by (1.2.12) fiy(n) + finn) = (ur F 2) (n), (1.2.13) fuln) x fi(n) = (Hi * 2) (n) The convolution product is commutative and associative: (1, * 12 = M2 * Mi, (uy H2)* M3 = 11 *(H2*H13). Furthermore the mapping 4 — jis a contraction, meaning that 1 (1.2.14) \(n)| = 5 Var(u). aT Exercise 1.2.10. Prove these statements.20. wTRODUCTION TO FOURIER ANALYSIS AND WAVELETS, ‘The Riemann-Lebesgue lemma is not true for signed measures. For example, the Dirac measure 49, for which 89(A) = 1 iff 0 € A, has 5o(n) = 1 for all n, One might ask if the Riemann-Lebesgue lemma holds for measures that are continuous, but this is not true either. The Fourier coefficients of the Cantor measure do not tend to zero. Exercise 1.2.11. Prove this statement Hint: For the Cantor measure, /i(n) = M172, cos 2rn/3*), with A(3") = fi(1) form = 1,2,.... For details, consult Zygmund, 1959, p. 196, 1.2.3. *Rates of Decay of Fourier Coet nts The Riemann-Lebesgue lemma provides no further quantitative information about the speed of convergence to zero for an arbitrary L' function. We can obtain a convenient upper bound from the representation a 1 cs ind (1.2.15) fa) = # [i e+*) —f0)] a6 Hence we have immediately Proposition 1.2.12. Proposition 1.2.12. Suppose that f € C(T) has a modulus of continuity @(6) := SUPj,-y125 Lf (X) — FO]. Then |f(n)| < 3@(r/n). More generally, if 1 < p < 00 and &y(8) °= Supyyycs Ifa fp is the L? modulus of continuity of f € L?(T), then \Fen)| < 42,(er/n). __For example, if f satisfies a Hélder condition with exponent a € (0, 1), we see that f(n) = O(n-*), |n| + 00. Exercise 1.2.13. Supppose that the L? modulus of continuity satisfies Qp(h) < Ch® for C > Oand a > 1. Prove that f is a constant, a.e. Hint: First show that 2,(h + k) = p(h) + 9) (k); then iterate this to obtain a contradiction, If we want to obtain a more precise estimation, we can assume that f is absolutely continuous and integrate by parts, as follows: Proposition 1.2.14, Suppose that f € C(T) is absolutely continuous. Then Pin) = (/inf'(n); in particular f(n) = o(1/\n!), \n| > 00. If in addition FH -+5f4 are absolutely continuous, then f(n) = (1/in)*f(n); in partic- ular f(n) = o(1/\n\*). If f satisfies a Hélder condition with exponent x, then Fn) = OC In|§**), |n| > 00. Exercise 1.2.15. Prove the above properties by integration by parts. It is difficult to characterize differentiability of a fixed degree in terms of the behavior of the Fourier coefficients. In order to obtain some simple characterizations of smoothness, we consider functions that are infinitely differentiable. This means that for each m € Z*, the derivative f exists and is a continuous function. Then we canFOURIER SERIES ON THE CIRCLE 21 integrate-by-parts for n 0 and write 2f(n) = (1/in)" fy e"*f™(t) dt, to conclude that the Fourier coefficients satisfy a system of estimates of the form Cn in| (1.2.16) Fol s i OfneZ, m=0,1, In other words f(n) tends to zero faster than any negative power when |n| > 00. Con- versely, if the Fourier coefficients of f ¢ L'(T) satisfy (1.2.16), then we can repeatedly differentiate the absolutely convergent Fourier series to conclude that f is a.e. equal to an infinitely differentiable function. This is summarized as follows. Proposition 1.2.16. f € L'(T) isa.e. equal to an infinitely differentiable function if and only if its Fourier coefficients are rapidly decreasing, according to (1.2.16). 1.2.3.1 Piecewise smooth functions The correspondence between smoothness of a fixed degree and decay of the Fourier coefficients is not sharp in general. More precisely, the converse of Proposition 1.2.14 is false: there exists a nonabsolutely continuous f € L'(T) for which nf(n) > 0 when In| — oo. In order to obtain sharp results, we consider functions that are piecewise smooth, described as follows. If there exists a subdivision —1 < 6 <6) < ++ < 6x. <2 so that f is absolutely continuous on each subinterval with a simple jump at the endpoint, denoted 4f(6;) := (6; + 0) — f(@ — 0), then we say that f is piecewise smooth of degree 0. In general we say that f is piecewise smooth of degree k if there exists such a subdivision so that f is absolutely continuous on T, together with its first k — | derivatives and that f is piecewise absolutely continuous as above, with jumps denoted 5f (6)) := f @; +0) —f G — 0). Proposition 1.2.17. Suppose that f is piecewise smooth of degree k. Then the Fourier coefficients satisfy the identity = LT Ey ng) P (1.2.17) j= £3(3) (0 (je )s jo Furthermore the coefficient of 1/n'*! tends to zero if and only if all of the jumps are zero, i.e. f is a continuous function. Note that, in case 6) = —z, we interpret 5f (60) = f(—=) — f(#). Proof. In case k = 0 we do an integration by parts on each interval of continuity [soem eo = ne ete pha i )-10 > -[ rose. " sora( , xin ‘When we sum the boundary terms and simplify, the sum is written in terms of the jumps and the Fourier coefficients of f", which proves the result for k = 0. If f has k — 1 absolutely continuous derivatives, we can iterate this to obtain the k terms displayed, together with the Fourier coefficient of f“. In order to prove the sharpness, we prove a separate lemma.22, INTRODUCTION TO FOURIER ANALYSIS AND WAVELETS Lemma 1.2.18, Suppose that —1 < 6) <6; < +++ 0, by hypothesis. Then Tye = Cy Den) im ee 1 wore = Cn + OG im GF) Pa Taking N — oo, both the left side and the sum on the right tend to zero, hence C,, = 0, as required proving the lemma and Proposition 1.2.17, . Corollary 1.2.19. Let f be piecewise smooth of degree k and let 0 < r < k. Then we have the asymptotic estimate f(n) = o(\n|-‘~"), |n| > 00 if and only if f €C(T), ie, f has r continuous derivatives. Proof. The Riemann-Lebesgue lemma implies that the last term in (1.2.17) is o(|n|~*'). ‘Therefore the asymptotic behavior off (n) is equivalent to that ofthe finite sum. Iff € C"(T), then all of the jump terms in (1.2.17) are zero for ! 1/8.FOURIER SERIES ON THECIRCLE 23 Equation (1.2.18) is also valid in an open interval |t — fo| < 8/3, by noting the following system of estimates: FM) = KF I + mat — 10) i= 0) fk oe ol Faxon ay" 10 rm) m! SCY 2ImRM MY — lt = __ on ~ 1 =2Rit — tl valid for |t — fo < 1/3R, which provides the desired uniform estimate. By covering T by a finite number of these intervals, we may assume that R, M are independent of 10. Conversely, if an infinitely differentiable function f(t), t € T satisfies (1.2.18), then the series 2% of (t0)(t— to)" /m! converges in the interval |t — fo| < 1/R andis therefore an analytic function, Summarizing, we see that f is analytic if and only if (1.2.18) holds. In order to characterize analyticity in terms of the Fourier coefficients, first suppose that f is analytic on TT. In particular f is infinitely differentiable and we can integrate- by-parts to write for n #0, 2nfin) = f em p(n dr 7 = (gy fe f(a) de Fal < min” = ine Since this is valid for all m, we choose the optimal value m = [n/R} and apply Stirling’s formula to conclude that |f(n)| < Me~‘"! for any c < 1/R. Conversely, suppose that fis an infinitely differentiable function on ‘T whose Fourier coefficients satisfy a system of inequalities of the form |/(n)| < Ae"! for some positive constants A, c. In particular, by modifying f on a null set, we have the convergent Fourier series f(t) = Dynex /(n)e”, which can be differentiated m times to obtain f(t) = Y,<2(in)"f(n)e”". Applying the hypothesis, we have Lf @| SA Dyez Ini". Comparing this sum with the integral {5° x"e~* dx shows that the successive derivatives satisfy the estimates | f)(1)| < Am!/cl Hence we obtain the following Fourier-analytic characterization of analytic functions on the circle. Proposition 1.2.20. f € L'(T) is a.e. equal to an analytic function if and only if its Fourier coefficients satisfy the system of inequalities (1.2.19) \f@)|< Ae" = neZ {for positive constants ¢, A.24 wTRODUCTION TO FOURIER ANALYSIS AND WAVELETS, Exercise 1.2.21. Carry out the details of the replacement of the sum by the integral. 1.2.4 Sine Integral In order to treat pointwise convergence of Fourier series and integrals in one dimension, a fundamental role is played by the function (1.2.20) Si(x) = zf la, OKx<00. ah ot Its basic properties are listed as follows: (Si) = 0, lim, 0 Si(x) = 1 Gi) Si(x) < Si(r) = 1.18... forallx > 0. (iii) x + Si(x) has relative maxima at the points x, 37, 57, ... and relative minima at the points 2x, 4zr, 6m, ... To prove these properties, we first note that Si((n + 1)m) —Si(nx) = 2/m ft" (sin t/t) dt and that these numbers alternate in sign and decrease to zero in absolute value. Hence the improper integral defining lim, Si (x) exists. To compute its value, we note that by the Riemann-Lebesgue lemma, i 1 2). 1 iin [ [som - 5 |sin(n 5)oae =0. But J; Da(o) do = 27, which shows that a 1 in f sin(n + 5)¢ mJy ~ sin(@/2) or equivalently Si(nx + $) > 1 when n -> 00. dp=a 1.2.4.1 Other proofs that Si(co) = 1 We offer two other proofs of this important improper integral. Proof using complex analysis. Consider the integral of the analytic function e/z on the (counterclockwise) contour defined by the two semicircular arcs z = ee", 0 < @ < x; z= Re", 0 < 6 < x; and the two segments of the real axis defined by € < |x| < R. By Cauchy's theorem the total line integral is zero, When R —> 00, € > 0, the integral on the large semicircle tends to zero, while the integral on the small semicircle is [ode in, «0 In Hence we have