This page intentionally left blank
Scaling, Fractals and Wavelets
This page intentionally left blank
Scaling, Fra...
106 downloads
2054 Views
4MB Size
Report
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form
This page intentionally left blank
Scaling, Fractals and Wavelets
This page intentionally left blank
Scaling, Fractals and Wavelets
Edited by Patrice Abry Paulo Gonçalves Jacques Lévy Véhel
First published in France in 2 volumes in 2002 by Hermes Science/Lavoisier entitled: Lois d’échelle, fractales et ondelettes © LAVOISIER, 2002 First published in Great Britain and the United States in 2009 by ISTE Ltd and John Wiley & Sons, Inc. Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced, stored or transmitted, in any form or by any means, with the prior permission in writing of the publishers, or in the case of reprographic reproduction in accordance with the terms and licenses issued by the CLA. Enquiries concerning reproduction outside these terms should be sent to the publishers at the undermentioned address: ISTE Ltd 27-37 St George’s Road London SW19 4EU UK
John Wiley & Sons, Inc. 111 River Street Hoboken, NJ 07030 USA
www.iste.co.uk
www.wiley.com
© ISTE Ltd, 2009 The rights of Patrice Abry, Paulo Gonçalves and Jacques Lévy Véhel to be identified as the authors of this work have been asserted by them in accordance with the Copyright, Designs and Patents Act 1988. Library of Congress Cataloging-in-Publication Data Lois d’échelle, fractales et ondelettes. English Scaling, fractals and wavelets/edited by Patrice Abry, Paulo Gonçalves, Jacques Lévy Véhel. p. cm. Includes bibliographical references. ISBN 978-1-84821-072-1 1. Signal processing--Mathematics. 2. Fractals. 3. Wavelets (Mathematics) I. Abry, Patrice. II. Gonçalves, Paulo. III. Lévy Véhel, Jacques, 1960- IV. Title. TK5102.9.L65 2007 621.382'20151--dc22 2007025119 British Library Cataloguing-in-Publication Data A CIP record for this book is available from the British Library ISBN: 978-1-84821-072-1 Printed and bound in Great Britain by CPI Antony Rowe Ltd, Chippenham, Wiltshire.
Table of Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
Chapter 1. Fractal and Multifractal Analysis in Signal Processing . . . . . Jacques L ÉVY V ÉHEL and Claude T RICOT
19
1.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2. Dimensions of sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.1. Minkowski-Bouligand dimension . . . . . . . . . . . . . . . . 1.2.2. Packing dimension . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.3. Covering dimension . . . . . . . . . . . . . . . . . . . . . . . . 1.2.4. Methods for calculating dimensions . . . . . . . . . . . . . . . 1.3. Hölder exponents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.1. Hölder exponents related to a measure . . . . . . . . . . . . . . 1.3.2. Theorems on set dimensions . . . . . . . . . . . . . . . . . . . 1.3.3. Hölder exponent related to a function . . . . . . . . . . . . . . 1.3.4. Signal dimension theorem . . . . . . . . . . . . . . . . . . . . . 1.3.5. 2-microlocal analysis . . . . . . . . . . . . . . . . . . . . . . . 1.3.6. An example: analysis of stock market price . . . . . . . . . . . 1.4. Multifractal analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.1. What is the purpose of multifractal analysis? . . . . . . . . . . 1.4.2. First ingredient: local regularity measures . . . . . . . . . . . . 1.4.3. Second ingredient: the size of point sets of the same regularity 1.4.4. Practical calculation of spectra . . . . . . . . . . . . . . . . . . 1.4.5. Refinements: analysis of the sequence of capacities, mutual analysis and multisingularity . . . . . . . . . . . . . . . . . . . 1.4.6. The multifractal spectra of certain simple signals . . . . . . . . 1.4.7. Two applications . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.7.1. Image segmentation . . . . . . . . . . . . . . . . . . . . . 1.4.7.2. Analysis of TCP traffic . . . . . . . . . . . . . . . . . . . 1.5. Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
19 20 21 25 27 29 33 33 33 36 42 45 46 48 48 49 50 52
. . . . . .
. . . . . .
60 62 66 66 67 68
6
Scaling, Fractals and Wavelets
Chapter 2. Scale Invariance and Wavelets . . . . . . . . . . . . . . . . . . . . Patrick F LANDRIN, Paulo G ONÇALVES and Patrice A BRY
71
2.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 2.2. Models for scale invariance . . . . . . . . . . . . . . . . . . . . . . . . . 72 2.2.1. Intuition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 2.2.2. Self-similarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 2.2.3. Long-range dependence . . . . . . . . . . . . . . . . . . . . . . . . 75 2.2.4. Local regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 2.2.5. Fractional Brownian motion: paradigm of scale invariance . . . . 77 2.2.6. Beyond the paradigm of scale invariance . . . . . . . . . . . . . . 79 2.3. Wavelet transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 2.3.1. Continuous wavelet transform . . . . . . . . . . . . . . . . . . . . 81 2.3.2. Discrete wavelet transform . . . . . . . . . . . . . . . . . . . . . . 82 2.4. Wavelet analysis of scale invariant processes . . . . . . . . . . . . . . . 85 2.4.1. Self-similarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 2.4.2. Long-range dependence . . . . . . . . . . . . . . . . . . . . . . . . 88 2.4.3. Local regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 2.4.4. Beyond second order . . . . . . . . . . . . . . . . . . . . . . . . . . 92 2.5. Implementation: analysis, detection and estimation . . . . . . . . . . . . 92 2.5.1. Estimation of the parameters of scale invariance . . . . . . . . . . 93 2.5.2. Emphasis on scaling laws and determination of the scaling range . 96 2.5.3. Robustness of the wavelet approach . . . . . . . . . . . . . . . . . 98 2.6. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 2.7. Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 Chapter 3. Wavelet Methods for Multifractal Analysis of Functions . . . . 103 Stéphane JAFFARD 3.1. Introduction . . . . . . . . . . . . . . . . . . . . 3.2. General points regarding multifractal functions 3.2.1. Important definitions . . . . . . . . . . . . 3.2.2. Wavelets and pointwise regularity . . . . 3.2.3. Local oscillations . . . . . . . . . . . . . . 3.2.4. Complements . . . . . . . . . . . . . . . . 3.3. Random multifractal processes . . . . . . . . . 3.3.1. Lévy processes . . . . . . . . . . . . . . . 3.3.2. Burgers’ equation and Brownian motion . 3.3.3. Random wavelet series . . . . . . . . . . . 3.4. Multifractal formalisms . . . . . . . . . . . . . 3.4.1. Besov spaces and lacunarity . . . . . . . . 3.4.2. Construction of formalisms . . . . . . . . 3.5. Bounds of the spectrum . . . . . . . . . . . . . 3.5.1. Bounds according to the Besov domain .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
103 104 104 107 112 116 117 117 120 122 123 123 126 129 129
Contents
7
3.5.2. Bounds deduced from histograms . . . . . . . . . . . . . . . . . . 132 3.6. The grand-canonical multifractal formalism . . . . . . . . . . . . . . . . 132 3.7. Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 Chapter 4. Multifractal Scaling: General Theory and Approach by Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 Rudolf R IEDI 4.1. Introduction and summary . . . . . . . . . . . . . . . 4.2. Singularity exponents . . . . . . . . . . . . . . . . . 4.2.1. Hölder continuity . . . . . . . . . . . . . . . . . 4.2.2. Scaling of wavelet coefficients . . . . . . . . . 4.2.3. Other scaling exponents . . . . . . . . . . . . . 4.3. Multifractal analysis . . . . . . . . . . . . . . . . . . 4.3.1. Dimension based spectra . . . . . . . . . . . . 4.3.2. Grain based spectra . . . . . . . . . . . . . . . 4.3.3. Partition function and Legendre spectrum . . . 4.3.4. Deterministic envelopes . . . . . . . . . . . . . 4.4. Multifractal formalism . . . . . . . . . . . . . . . . . 4.5. Binomial multifractals . . . . . . . . . . . . . . . . . 4.5.1. Construction . . . . . . . . . . . . . . . . . . . 4.5.2. Wavelet decomposition . . . . . . . . . . . . . 4.5.3. Multifractal analysis of the binomial measure . 4.5.4. Examples . . . . . . . . . . . . . . . . . . . . . 4.5.5. Beyond dyadic structure . . . . . . . . . . . . . 4.6. Wavelet based analysis . . . . . . . . . . . . . . . . . 4.6.1. The binomial revisited with wavelets . . . . . . 4.6.2. Multifractal properties of the derivative . . . . 4.7. Self-similarity and LRD . . . . . . . . . . . . . . . . 4.8. Multifractal processes . . . . . . . . . . . . . . . . . 4.8.1. Construction and simulation . . . . . . . . . . 4.8.2. Global analysis . . . . . . . . . . . . . . . . . . 4.8.3. Local analysis of warped FBM . . . . . . . . . 4.8.4. LRD and estimation of warped FBM . . . . . . 4.9. Bibliography . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . .
139 140 140 142 144 145 145 146 147 149 151 154 154 157 158 160 162 163 163 165 167 168 169 170 170 173 173
Chapter 5. Self-similar Processes . . . . . . . . . . . . . . . . . . . . . . . . . 179 Albert B ENASSI and Jacques I STAS 5.1. Introduction . . . . . . . . . . . . . . 5.1.1. Motivations . . . . . . . . . . . 5.1.2. Scalings . . . . . . . . . . . . . 5.1.2.1. Trees . . . . . . . . . . . . 5.1.2.2. Coding of R . . . . . . . 5.1.2.3. Renormalizing Cantor set
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
179 179 182 182 183 183
8
Scaling, Fractals and Wavelets
5.1.2.4. Random renormalized Cantor set . . . . . . . . . . . . . . . 5.1.3. Distributions of scale invariant masses . . . . . . . . . . . . . . . 5.1.3.1. Distribution of masses associated with Poisson measures . 5.1.3.2. Complete coding . . . . . . . . . . . . . . . . . . . . . . . . 5.1.4. Weierstrass functions . . . . . . . . . . . . . . . . . . . . . . . . 5.1.5. Renormalization of sums of random variables . . . . . . . . . . 5.1.6. A common structure for a stochastic (semi-)self-similar process 5.1.7. Identifying Weierstrass functions . . . . . . . . . . . . . . . . . . 5.1.7.1. Pseudo-correlation . . . . . . . . . . . . . . . . . . . . . . . 5.2. The Gaussian case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.1. Self-similar Gaussian processes with r-stationary increments . . 5.2.1.1. Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.1.2. Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.1.3. Characterization . . . . . . . . . . . . . . . . . . . . . . . . 5.2.2. Elliptic processes . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.3. Hyperbolic processes . . . . . . . . . . . . . . . . . . . . . . . . 5.2.4. Parabolic processes . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.5. Wavelet decomposition . . . . . . . . . . . . . . . . . . . . . . . 5.2.5.1. Gaussian elliptic processes . . . . . . . . . . . . . . . . . . 5.2.5.2. Gaussian hyperbolic process . . . . . . . . . . . . . . . . . 5.2.6. Renormalization of sums of correlated random variable . . . . . 5.2.7. Convergence towards fractional Brownian motion . . . . . . . . 5.2.7.1. Quadratic variations . . . . . . . . . . . . . . . . . . . . . . 5.2.7.2. Acceleration of convergence . . . . . . . . . . . . . . . . . 5.2.7.3. Self-similarity and regularity of trajectories . . . . . . . . . 5.3. Non-Gaussian case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.2. Symmetric α-stable processes . . . . . . . . . . . . . . . . . . . 5.3.2.1. Stochastic measure . . . . . . . . . . . . . . . . . . . . . . . 5.3.2.2. Ellipticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.3. Censov and Takenaka processes . . . . . . . . . . . . . . . . . . 5.3.4. Wavelet decomposition . . . . . . . . . . . . . . . . . . . . . . . 5.3.5. Process subordinated to Brownian measure . . . . . . . . . . . . 5.4. Regularity and long-range dependence . . . . . . . . . . . . . . . . . . 5.4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4.2. Two examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4.2.1. A signal plus noise model . . . . . . . . . . . . . . . . . . . 5.4.2.2. Filtered white noise . . . . . . . . . . . . . . . . . . . . . . 5.4.2.3. Long-range correlation . . . . . . . . . . . . . . . . . . . . 5.5. Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
184 184 184 185 185 186 187 188 188 189 189 189 189 190 190 191 192 192 192 193 193 193 193 194 195 195 195 196 196 196 198 198 199 200 200 201 201 201 202 202
Contents
9
Chapter 6. Locally Self-similar Fields . . . . . . . . . . . . . . . . . . . . . . 205 Serge C OHEN 6.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2. Recap of two representations of fractional Brownian motion . 6.2.1. Reproducing kernel Hilbert space . . . . . . . . . . . . . 6.2.2. Harmonizable representation . . . . . . . . . . . . . . . . 6.3. Two examples of locally self-similar fields . . . . . . . . . . . 6.3.1. Definition of the local asymptotic self-similarity (LASS) 6.3.2. Filtered white noise (FWN) . . . . . . . . . . . . . . . . . 6.3.3. Elliptic Gaussian random fields (EGRP) . . . . . . . . . . 6.4. Multifractional fields and trajectorial regularity . . . . . . . . . 6.4.1. Two representations of the MBM . . . . . . . . . . . . . . 6.4.2. Study of the regularity of the trajectories of the MBM . . 6.4.3. Towards more irregularities: generalized multifractional Brownian motion (GMBM) and step fractional Brownian motion (SFBM) . . . . . . . . . . . . . . . . . . . . . . . . 6.4.3.1. Step fractional Brownian motion . . . . . . . . . . . 6.4.3.2. Generalized multifractional Brownian motion . . . 6.5. Estimate of regularity . . . . . . . . . . . . . . . . . . . . . . . 6.5.1. General method: generalized quadratic variation . . . . . 6.5.2. Application to the examples . . . . . . . . . . . . . . . . . 6.5.2.1. Identification of filtered white noise . . . . . . . . . 6.5.2.2. Identification of elliptic Gaussian random processes 6.5.2.3. Identification of MBM . . . . . . . . . . . . . . . . . 6.5.2.4. Identification of SFBMs . . . . . . . . . . . . . . . . 6.6. Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
205 207 207 208 213 213 214 215 218 219 221
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
222 223 224 226 226 228 228 230 231 233 235
Chapter 7. An Introduction to Fractional Calculus . . . . . . . . . . . . . . 237 Denis M ATIGNON 7.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1.1. Motivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1.1.1. Fields of application . . . . . . . . . . . . . . . . . . . . . . . 7.1.1.2. Theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1.2. Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1.3. Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2. Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.1. Fractional integration . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.2. Fractional derivatives within the framework of causal distributions 7.2.2.1. Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.2.2. Fundamental solutions . . . . . . . . . . . . . . . . . . . . . . 7.2.3. Mild fractional derivatives, in the Caputo sense . . . . . . . . . . . 7.2.3.1. Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
237 237 237 238 238 239 240 240 242 242 245 246 246
10
Scaling, Fractals and Wavelets
7.2.3.2. Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247 7.2.3.3. Mittag-Leffler eigenfunctions . . . . . . . . . . . . . . . . . . 248 7.2.3.4. Fractional power series expansions of order α (α-FPSE) . . 250 7.3. Fractional differential equations . . . . . . . . . . . . . . . . . . . . . . 251 7.3.1. Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 7.3.1.1. Framework of causal distributions . . . . . . . . . . . . . . . 251 7.3.1.2. Framework of fractional power series expansion of order one half . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252 7.3.1.3. Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253 7.3.2. Framework of causal distributions . . . . . . . . . . . . . . . . . . 254 7.3.3. Framework of functions expandable into fractional power series (α-FPSE) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255 7.3.4. Asymptotic behavior of fundamental solutions . . . . . . . . . . . 257 7.3.4.1. Asymptotic behavior at the origin . . . . . . . . . . . . . . . 257 7.3.4.2. Asymptotic behavior at infinity . . . . . . . . . . . . . . . . . 257 7.3.5. Controlled-and-observed linear dynamic systems of fractional order 261 7.4. Diffusive structure of fractional differential systems . . . . . . . . . . . 262 7.4.1. Introduction to diffusive representations of pseudo-differential operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263 7.4.2. General decomposition result . . . . . . . . . . . . . . . . . . . . . 264 7.4.3. Connection with the concept of long memory . . . . . . . . . . . . 265 7.4.4. Particular case of fractional differential systems of commensurate orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265 7.5. Example of a fractional partial differential equation . . . . . . . . . . . 266 7.5.1. Physical problem considered . . . . . . . . . . . . . . . . . . . . . 267 7.5.2. Spectral consequences . . . . . . . . . . . . . . . . . . . . . . . . . 268 7.5.3. Time-domain consequences . . . . . . . . . . . . . . . . . . . . . . 268 7.5.3.1. Decomposition into wavetrains . . . . . . . . . . . . . . . . . 269 7.5.3.2. Quasi-modal decomposition . . . . . . . . . . . . . . . . . . 270 7.5.3.3. Fractional modal decomposition . . . . . . . . . . . . . . . . 271 7.5.4. Free problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272 7.6. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273 7.7. Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273 Chapter 8. Fractional Synthesis, Fractional Filters . . . . . . . . . . . . . . 279 Liliane B EL, Georges O PPENHEIM, Luc ROBBIANO and Marie-Claude V IANO 8.1. Traditional and less traditional questions about fractionals . . . . . . . . 279 8.1.1. Notes on terminology . . . . . . . . . . . . . . . . . . . . . . . . . 279 8.1.2. Short and long memory . . . . . . . . . . . . . . . . . . . . . . . . 279 8.1.3. From integer to non-integer powers: filter based sample path design 280 8.1.4. Local and global properties . . . . . . . . . . . . . . . . . . . . . . 281 8.2. Fractional filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282 8.2.1. Desired general properties: association . . . . . . . . . . . . . . . 282
Contents
8.2.2. Construction and approximation techniques . . . . . . . . . . . 8.3. Discrete time fractional processes . . . . . . . . . . . . . . . . . . . 8.3.1. Filters: impulse responses and corresponding processes . . . . 8.3.2. Mixing and memory properties . . . . . . . . . . . . . . . . . . 8.3.3. Parameter estimation . . . . . . . . . . . . . . . . . . . . . . . . 8.3.4. Simulated example . . . . . . . . . . . . . . . . . . . . . . . . . 8.4. Continuous time fractional processes . . . . . . . . . . . . . . . . . . 8.4.1. A non-self-similar family: fractional processes designed from fractional filters . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.2. Sample path properties: local and global regularity, memory . 8.5. Distribution processes . . . . . . . . . . . . . . . . . . . . . . . . . . 8.5.1. Motivation and generalization of distribution processes . . . . 8.5.2. The family of linear distribution processes . . . . . . . . . . . 8.5.3. Fractional distribution processes . . . . . . . . . . . . . . . . . 8.5.4. Mixing and memory properties . . . . . . . . . . . . . . . . . . 8.6. Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chapter 9. Iterated Function Systems and Some Generalizations: Local Regularity Analysis and Multifractal Modeling of Signals . Khalid DAOUDI 9.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2. Definition of the Hölder exponent . . . . . . . . . . . . . . . . 9.3. Iterated function systems (IFS) . . . . . . . . . . . . . . . . . . 9.4. Generalization of iterated function systems . . . . . . . . . . . 9.4.1. Semi-generalized iterated function systems . . . . . . . . 9.4.2. Generalized iterated function systems . . . . . . . . . . . 9.5. Estimation of pointwise Hölder exponent by GIFS . . . . . . . 9.5.1. Principles of the method . . . . . . . . . . . . . . . . . . . 9.5.2. Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.5.3. Application . . . . . . . . . . . . . . . . . . . . . . . . . . 9.6. Weak self-similar functions and multifractal formalism . . . . 9.7. Signal representation by WSA functions . . . . . . . . . . . . . 9.8. Segmentation of signals by weak self-similar functions . . . . 9.9. Estimation of the multifractal spectrum . . . . . . . . . . . . . 9.10. Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.11. Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
. . . . . . .
. . . . . . .
282 284 284 286 287 289 291
. . . . . . . .
. . . . . . . .
291 293 294 294 294 295 296 297
. . . . . 301 . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
Chapter 10. Iterated Function Systems and Applications in Image Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Franck DAVOINE and Jean-Marc C HASSERY 10.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2. Iterated transformation systems . . . . . . . . . . . . . . . . . . . . . . 10.2.1. Contracting transformations and iterated transformation systems 10.2.1.1. Lipschitzian transformation . . . . . . . . . . . . . . . . . .
301 303 304 306 307 308 311 312 314 315 318 320 324 326 327 329 333 333 333 334 334
12
Scaling, Fractals and Wavelets
10.2.1.2. Contracting transformation . . . . . . . . . . . . . . . . . . 10.2.1.3. Fixed point . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2.1.4. Hausdorff distance . . . . . . . . . . . . . . . . . . . . . . . 10.2.1.5. Contracting transformation on the space H(R2 ) . . . . . . 10.2.1.6. Iterated transformation system . . . . . . . . . . . . . . . . 10.2.2. Attractor of an iterated transformation system . . . . . . . . . . . 10.2.3. Collage theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2.4. Finally contracting transformation . . . . . . . . . . . . . . . . . 10.2.5. Attractor and invariant measures . . . . . . . . . . . . . . . . . . 10.2.6. Inverse problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3. Application to natural image processing: image coding . . . . . . . . . 10.3.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3.2. Coding of natural images by fractals . . . . . . . . . . . . . . . . 10.3.2.1. Collage of a source block onto a destination block . . . . . 10.3.2.2. Hierarchical partitioning . . . . . . . . . . . . . . . . . . . . 10.3.2.3. Coding of the collage operation on a destination block . . . 10.3.2.4. Contraction control of the fractal transformation . . . . . . 10.3.3. Algebraic formulation of the fractal transformation . . . . . . . . 10.3.3.1. Formulation of the mass transformation . . . . . . . . . . . 10.3.3.2. Contraction control of the fractal transformation . . . . . . 10.3.3.3. Fisher formulation . . . . . . . . . . . . . . . . . . . . . . . 10.3.4. Experimentation on triangular partitions . . . . . . . . . . . . . . 10.3.5. Coding and decoding acceleration . . . . . . . . . . . . . . . . . 10.3.5.1. Coding simplification suppressing the research for similarities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3.5.2. Decoding simplification by collage space orthogonalization 10.3.5.3. Coding acceleration: search for the nearest neighbor . . . . 10.3.6. Other optimization diagrams: hybrid methods . . . . . . . . . . . 10.4. Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
334 334 334 335 335 335 336 338 339 340 340 340 342 342 344 345 345 345 347 349 350 351 352 352 358 360 360 362
Chapter 11. Local Regularity and Multifractal Methods for Image and Signal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367 Pierrick L EGRAND 11.1. Introduction . . . . . . . . . . . . . . . . . . . . . 11.2. Basic tools . . . . . . . . . . . . . . . . . . . . . 11.2.1. Hölder regularity analysis . . . . . . . . . . 11.2.2. Reminders on multifractal analysis . . . . . 11.2.2.1. Hausdorff multifractal spectrum . . . 11.2.2.2. Large deviation multifractal spectrum 11.2.2.3. Legendre multifractal spectrum . . . . 11.3. Hölderian regularity estimation . . . . . . . . . . 11.3.1. Oscillations (OSC) . . . . . . . . . . . . . . 11.3.2. Wavelet coefficient regression (W CR) . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
367 368 368 369 369 370 371 371 371 372
Contents
11.3.3. Wavelet leaders regression (W L) . . . . . . . . . . . . . . . 11.3.4. Limit inf and limit sup regressions . . . . . . . . . . . . . . 11.3.5. Numerical experiments . . . . . . . . . . . . . . . . . . . . . 11.4. Denoising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.4.2. Minimax risk, optimal convergence rate and adaptivity . . . 11.4.3. Wavelet based denoising . . . . . . . . . . . . . . . . . . . . 11.4.4. Non-linear wavelet coefficients pumping . . . . . . . . . . . 11.4.4.1. Minimax properties . . . . . . . . . . . . . . . . . . . . 11.4.4.2. Regularity control . . . . . . . . . . . . . . . . . . . . 11.4.4.3. Numerical experiments . . . . . . . . . . . . . . . . . 11.4.5. Denoising using exponent between scales . . . . . . . . . . 11.4.5.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 11.4.5.2. Estimating the local regularity of a signal from noisy observations . . . . . . . . . . . . . . . . . . . . . . . . . 11.4.5.3. Numerical experiments . . . . . . . . . . . . . . . . . 11.4.6. Bayesian multifractal denoising . . . . . . . . . . . . . . . . 11.4.6.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 11.4.6.2. The set of parameterized classes S(g, ψ) . . . . . . . 11.4.6.3. Bayesian denoising in S(g, ψ) . . . . . . . . . . . . . 11.4.6.4. Numerical experiments . . . . . . . . . . . . . . . . . 11.4.6.5. Denoising of road profiles . . . . . . . . . . . . . . . . 11.5. Hölderian regularity based interpolation . . . . . . . . . . . . . . 11.5.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.5.2. The method . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.5.3. Regularity and asymptotic properties . . . . . . . . . . . . . 11.5.4. Numerical experiments . . . . . . . . . . . . . . . . . . . . . 11.6. Biomedical signal analysis . . . . . . . . . . . . . . . . . . . . . . 11.7. Texture segmentation . . . . . . . . . . . . . . . . . . . . . . . . . 11.8. Edge detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.8.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.8.1.1. Edge detection . . . . . . . . . . . . . . . . . . . . . . 11.9. Change detection in image sequences using multifractal analysis 11.10. Image reconstruction . . . . . . . . . . . . . . . . . . . . . . . . 11.11. Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
372 373 374 376 376 377 378 380 380 381 382 383 383
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
384 386 386 386 387 388 390 391 393 393 393 394 394 394 401 403 403 406 407 408 409
Chapter 12. Scale Invariance in Computer Network Traffic . . . . . . . . . 413 Darryl V EITCH 12.1. Teletraffic – a new natural phenomenon . . . . . . . 12.1.1. A phenomenon of scales . . . . . . . . . . . . . 12.1.2. An experimental science of “man-made atoms” 12.1.3. A random current . . . . . . . . . . . . . . . . . 12.1.4. Two fundamental approaches . . . . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
413 413 415 416 417
14
Scaling, Fractals and Wavelets
12.2. From a wealth of scales arise scaling laws . . 12.2.1. First discoveries . . . . . . . . . . . . . . 12.2.2. Laws reign . . . . . . . . . . . . . . . . . 12.2.3. Beyond the revolution . . . . . . . . . . 12.3. Sources as the source of the laws . . . . . . . 12.3.1. The sum or its parts . . . . . . . . . . . . 12.3.2. The on/off paradigm . . . . . . . . . . . 12.3.3. Chemistry . . . . . . . . . . . . . . . . . 12.3.4. Mechanisms . . . . . . . . . . . . . . . . 12.4. New models, new behaviors . . . . . . . . . . 12.4.1. Character of a model . . . . . . . . . . . 12.4.2. The fractional Brownian motion family 12.4.3. Greedy sources . . . . . . . . . . . . . . 12.4.4. Never-ending calls . . . . . . . . . . . . 12.5. Perspectives . . . . . . . . . . . . . . . . . . . 12.6. Bibliography . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
419 419 420 424 426 426 427 428 429 430 430 431 432 432 433 434
Chapter 13. Research of Scaling Law on Stock Market Variations . . . . . 437 Christian WALTER 13.1. Introduction: fractals in finance . . . . . . . . . . . . . . . . . . . . . . 13.2. Presence of scales in the study of stock market variations . . . . . . . 13.2.1. Modeling of stock market variations . . . . . . . . . . . . . . . . 13.2.1.1. Statistical apprehension of stock market fluctuations . . . . 13.2.1.2. Profit and stock market return operations in different scales 13.2.1.3. Traditional financial modeling: Brownian motion . . . . . . 13.2.2. Time scales in financial modeling . . . . . . . . . . . . . . . . . . 13.2.2.1. The existence of characteristic time . . . . . . . . . . . . . . 13.2.2.2. Implicit scaling invariances of traditional financial modeling 13.3. Modeling postulating independence on stock market returns . . . . . . 13.3.1. 1960-1970: from Pareto’s law to Lévy’s distributions . . . . . . . 13.3.1.1. Leptokurtic problem and Mandelbrot’s first model . . . . . 13.3.1.2. First emphasis of Lévy’s α-stable distributions in finance . 13.3.2. 1970–1990: experimental difficulties of iid-α-stable model . . . 13.3.2.1. Statistical problem of parameter estimation of stable laws . 13.3.2.2. Non-normality and controversies on scaling invariance . . 13.3.2.3. Scaling anomalies of parameters under iid hypothesis . . . 13.3.3. Unstable iid models in partial scaling invariance . . . . . . . . . 13.3.3.1. Partial scaling invariances by regime switching models . . 13.3.3.2. Partial scaling invariances as compared with extremes . . . 13.4. Research of dependency and memory of markets . . . . . . . . . . . . 13.4.1. Linear dependence: testing of H-correlative models on returns . 13.4.1.1. Question of dependency of stock market returns . . . . . . 13.4.1.2. Problem of slow cycles and Mandelbrot’s second model . .
437 439 439 439 442 443 445 445 446 446 446 446 448 448 448 449 451 452 452 453 454 454 454 455
Contents
13.4.1.3. Introduction of fractional differentiation in econometrics . 13.4.1.4. Experimental difficulties of H-correlative model on returns 13.4.2. Non-linear dependence: validating H-correlative model on volatilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13.4.2.1. The 1980s: ARCH modeling and its limits . . . . . . . . . 13.4.2.2. The 1990s: emphasis of long dependence on volatility . . . 13.5. Towards a rediscovery of scaling laws in finance . . . . . . . . . . . . 13.6. Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chapter 14. Scale Relativity, Non-differentiability and Fractal Space-time Laurent N OTTALE
15
455 456 456 456 457 457 458 465
14.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465 14.2. Abandonment of the hypothesis of space-time differentiability . . . . 466 14.3. Towards a fractal space-time . . . . . . . . . . . . . . . . . . . . . . . . 466 14.3.1. Explicit dependence of coordinates on spatio-temporal resolutions 467 14.3.2. From continuity and non-differentiability to fractality . . . . . . 467 14.3.3. Description of non-differentiable process by differential equations 469 14.3.4. Differential dilation operator . . . . . . . . . . . . . . . . . . . . 471 14.4. Relativity and scale covariance . . . . . . . . . . . . . . . . . . . . . . 472 14.5. Scale differential equations . . . . . . . . . . . . . . . . . . . . . . . . 472 14.5.1. Constant fractal dimension: “Galilean” scale relativity . . . . . . 473 14.5.2. Breaking scale invariance: transition scales . . . . . . . . . . . . 474 14.5.3. Non-linear scale laws: second order equations, discrete scale invariance, log-periodic laws . . . . . . . . . . . . . . . . . . . . . 475 14.5.4. Variable fractal dimension: Euler-Lagrange scale equations . . . 476 14.5.5. Scale dynamics and scale force . . . . . . . . . . . . . . . . . . . 478 14.5.5.1. Constant scale force . . . . . . . . . . . . . . . . . . . . . . 479 14.5.5.2. Scale harmonic oscillator . . . . . . . . . . . . . . . . . . . 480 14.5.6. Special scale relativity – log-Lorentzian dilation laws, invariant scale limit under dilations . . . . . . . . . . . . . . . . . . . . . . . 481 14.5.7. Generalized scale relativity and scale-motion coupling . . . . . . 482 14.5.7.1. A reminder about gauge invariance . . . . . . . . . . . . . . 483 14.5.7.2. Nature of gauge fields . . . . . . . . . . . . . . . . . . . . . 484 14.5.7.3. Nature of the charges . . . . . . . . . . . . . . . . . . . . . . 486 14.5.7.4. Mass-charge relations . . . . . . . . . . . . . . . . . . . . . 488 14.6. Quantum-like induced dynamics . . . . . . . . . . . . . . . . . . . . . 488 14.6.1. Generalized Schrödinger equation . . . . . . . . . . . . . . . . . 488 14.6.2. Application in gravitational structure formation . . . . . . . . . . 492 14.7. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493 14.8. Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495 List of Authors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
This page intentionally left blank
Preface
It is a common scheme in many sciences to study systems or signals by looking for characteristic scales in time or space. These are then used as references for expressing all measured quantities. Physicists may for instance employ the size of a structure, while signal processors are often interested in correlation lengths: (blocks of) samples whose distance is several times the correlation lengths are considered statistically independent. The concept of scale invariance may be considered to be the converse of this approach: it means that there is no characteristic scale in the system. In other words, all scales contribute to the observed phenomenon. This “non-property” is also loosely referred to as scaling law or scaling behavior. Note that we may reverse the perspective and consider scale invariance as the signature of a strong organization in the system. Indeed, it is well known in physics that invariance laws are associated with fundamental properties. It is remarkable that phenomena where scaling laws have been observed cover a wide range of fields, both in natural and artificial systems. In the first category, these include for instance hydrology, in relation to the variability of water levels, hydrodynamics and the study of turbulence, statistical physics with the study of long-range interactions, electronics with the so-called 1/f noise in semiconductors, geophysics with the distribution of faults, biology, physiology and the variability of human body rhythms such as the heart rate. In the second category, we may mention geography with the distribution of population in cities or in continents, Internet traffic and financial markets. From a signal processing perspective, the aim is then to study transfer mechanisms between scales (also called “cascades”) rather than to identify relevant scales. We are thus led to forget about scale-based models (such as Markov models), and to focus on models allowing us to study correspondences between many scales. The central notion behind scaling laws is that of self-similarity. Loosely speaking, this means that each part is (statistically) the same as the whole object. In particular, information gathered from observing the data should be independent of the scale of observation.
18
Scaling, Fractals and Wavelets
There is considerable variety in observed self-similar behaviors. They may for instance appear through scaling laws in the Fourier domain, either at all frequencies or in a finite but large range of frequencies, or even in the limit of high or low frequencies. In many cases, studying second-order quantities such as spectra will prove insufficient for describing scaling laws. Higher-order moments are then necessary. More generally, the fundamental model of self-similarity has to be adapted in many settings, and to be generalized in various directions, so that it becomes useful in real-world situations. These include self-similar stochastic processes, 1/f processes, long memory processes, multifractal and multifractional processes, locally self-similar processes and more. Multifractal analysis, in particular, has developed as a method allowing us to study complex objects which are not necessarily “fractal”, by describing the variations of local regularity. The recent change of paradigm consisting of using fractal methods rather than studying fractal objects is one of the reasons for the success of the domain in applications. We are delighted to invite our reader for a promenade in the realm of scaling laws, its mathematical models and its real-world manifestations. The 14 chapters have all been written by experts. The first four chapters deal with the general mathematical tools allowing us to measure fractional dimensions, local regularity and scaling in its various disguises. Wavelets play a particular role for this purpose, and their role is emphasized. Chapters 5 and 6 describe advanced stochastic models relevant in our area. Chapter 7 deals with fractional calculus, and Chapter 8 explains how to synthesize certain fractal models. Chapter 9 gives a general introduction to IFS, a powerful tool for building and describing fractals and other complex objects, while Chapter 10, of applied nature, considers the application of IFS to image compression. The four remaining chapters also deal with applications: various signal and image processing tasks are considered in Chapter 11. Chapter 12 deals with Internet traffic, and Chapter 13 with financial data analysis. Finally, Chapter 14 describes a fractal space-time in the frame of cosmology. It is a great pleasure for us to thank all the authors of this volume for the quality of their contribution. We believe they have succeeded in exposing advanced concepts with great pedagogy.
Chapter 1
Fractal and Multifractal Analysis in Signal Processing
1.1. Introduction The aim of this chapter is to describe some of the fundamental concepts of fractal analysis in view of their application. We will thus present a simple introduction to the concepts of fractional dimension, regularity exponents and multifractal analysis, and show how they are used in signal and image processing. Since we are interested in applications, most theoretical results are given without proofs. These are available in the references mentioned where appropriate. In contrast, we will pay special attention to the practical aspects. In particular, almost all the notions explained below are implemented in the FracLab toolbox. This toolbox is freely available from the following site: http://complex.futurs.inria .fr/FracLab/, so that interested readers may perform hands-on experiments. Before we start, we wish to emphasize the following point: recent successes of fractal analysis in signal and image processing do not generally stem from the fact that they are applied to fractal objects (in a more or less strict sense). Indeed, most real-world signals are neither self-similar nor display the characteristics usually associated with fractals (except for the irregularity at each scale). The relevance of fractal analysis instead results from the progress made in the development of fractal methods. Such methods have lately become more general and reliable, and they now allow to describe precisely the singular structure of complex signals,
Chapter written by Jacques L ÉVY V ÉHEL and Claude T RICOT.
20
Scaling, Fractals and Wavelets
without any assumption of “fractality”: as a rule, performing a fractal analysis will be useful as soon as the considered signal is irregular and this irregularity contains meaningful information. There are numerous examples of such situations, ranging from image segmentation (where, for instance, contours are made of singular points; see section 1.4.7 and Chapter 11) to vocal synthesis [DAO 02] or financial analysis. This chapter roughly follows the chronological order in which the various tools have been introduced. We first describe several notions of fractional dimensions. These provide a global characterization of a signal. We then introduce Hölder exponents, which supply local measures of irregularity. The last part of the chapter is devoted to multifractal analysis, a most refined tool that describes the local as well as the overall singular structure of signals. All the concepts presented here are more fully developed in [TRI 99, LEV 02]. 1.2. Dimensions of sets The concept of dimension applies to objects more general than signals. To simplify, we shall consider sets in a metric space, although the notion of dimension makes sense for more complex entities such as measures or classes of functions [KOL 61]. Several interesting notions of dimension exist. This might look like a drawback for the mathematical analysis of fractal sets. However, it is actually an advantage, since each dimension emphasizes a different aspect of an object. It is thus worthwhile to determine the specificity of each dimension. As a general rule, none of these tools outperform the other. Let us first give a general definition of the notion of dimension. DEFINITION 1.1.– We call dimension an application d defined on the family of bounded sets of Rn and ranging in R+ ∪ {−∞}, such that: 1) d(∅) = −∞, d({x}) = 0 for any point x; 2) E1 ⊂ E2 ⇒ d(E1 ) d(E2 ) (monotonicity); 3) if E has non-zero n-dimensional volume, then d(E) = n; 4) if E is a diffeomorphism T of Rn (such as, in particular, a similarity with non-zero ratio, or a non-singular affine application), then d(T (E)) = d(E) (invariance). Moreover, we will say that d is stable if d(E1 ∪ E2 ) = max{d(E1 ), d(E2 )}. It is said to be σ-stable if, for any countable collection of sets: d ∪n En = sup d En σ-stable dimensions may be extended in a natural way to characterize unbounded sets of Rn .
Fractal and Multifractal Analysis in Signal Processing
21
1.2.1. Minkowski-Bouligand dimension The Minkowski-Bouligand dimension was invented by Bouligand [BOU 28], who named it the Cantor-Minkowski order. It is now commonly referred to as the box dimension. Let us cover a bounded set E of Rn with cubes of side ε and disjoint interiors. Let Nε (E) be the number of these cubes. When E contains an infinite number of points (i.e. if it is a curve, a surface, etc.), Nε (E) tends to +∞ when ε tends to 0. The box dimension Δ characterizes the rate of this growth. Roughly speaking, Δ is the real number such that: Δ 1 , Nε (E) ε assuming this number exists. More generally, we define, for all bounded E, the number: Δ(E) = lim sup ε→∞
log Nε (E) |log ε|
(1.1)
A lower limit may also be used: δ(E) = lim inf ε→∞
log Nε (E) |log ε|
(1.2)
Note that some authors refer to the box dimension only when both indices coincide, that is, when the limit exists. Both indices Δ and δ are dimensions in the sense previously defined. However, Δ is stable, contrarily to δ, so that Δ is more commonly used. Let us mention an ¯ denotes the closure of E (the set of all limit points of important property: if E sequences in E), then: ¯ = Δ(E) Δ(E) This property shows that Δ is not sensitive to the topological type of E. It only characterizes the density of a set. For example, the (countable) set of the rational numbers of the interval [0, 1] has one dimension, which is the dimension of the interval itself. Even discrete sequences may have non-zero dimension: let, for instance, E be the set of numbers n−α with α > 0. Then Δ(E) = 1/(α + 1). Equivalent definitions It is not mandatory to use cubes to calculate Δ. The original definition of Bouligand is as follows:
22
Scaling, Fractals and Wavelets
– in Rn , let us consider the Minkowski sausage: E(ε) = ∪x∈E Bε (x) which is the union of all the balls of radius ε centered at E. Denote its volume by Voln (E(ε)). This volume is approximately of the order of Nε (E) εn . This allows us to give the equivalent definition: Voln E(ε) ; (1.3) Δ(E) = lim sup n − log ε ε→0 – we may also define Nε (E), which is the smallest number of balls of radius ε covering E; or Nε (E), the largest number of disjoint balls of radius ε centered on E. Replacing Nε (E) by any of these values in equation (1.1) still gives Δ(E). Discrete values of ε In these definitions, the variable ε is continuous. The results remain the same if we use a discrete sequence such as εn = 2−n . More generally we may replace ε with any sequence which does not converge too quickly towards 0. More precisely, we require that: lim
n→∞
log εn = 1. log εn+1
This remark is important, as it allows us to perform numerical estimations of Δ. Let us now give some well-known examples of calculating dimensions. EXAMPLE 1.1.– Let (an ) be a sequence of real numbers such that 0 < 2an+1 < an < a0 = 1. Let E0 = [0, 1]. We construct by induction a sequence of sets (En ) such that En is made of 2n closed disjoint intervals of length an , each containing exactly two intervals of En+1 . The sets En are nested, and the sequence (En ) converges to a compact set E such that: E = ∩ n En . Let us consider a particular case. When all the interval extremities En are also interval extremities of En+1 , E is called a perfect symmetric set [KAH 63] or sometimes, more loosely, a Cantor set. Assume that the ratio log an / log an+1 tends to 1. According to the previous comment on discrete sequences, we obtain the following values: δ(E) = lim inf n→∞
n log 2 , |log an |
Δ(E) = lim sup n→∞
n log 2 . |log an |
Fractal and Multifractal Analysis in Signal Processing
23
However, these results are true for any sequence (an ). Even more specifically, consider the case where an = an , with 0 < a < 12 . The ratios an /an+1 are then constant and dimensions take the common value log 2/|log a|. This is the case of the self-similar set which satisfies the following relation: E = f1 (E) ∪ f2 (E) with f1 (x) = a x and f2 (x) = a x + 1 − a. This set is the attractor of the iterated function system {f1 , f2 } (see Chapters 9 and 10). It is also called a perfect symmetric set with constant ratio. EXAMPLE 1.2.– We construct a planar self-similar curve with extremities A and B, A = B as follows: take N + 1 distinct points A1 = A, A2 , . . . , AN +1 = B, such that dist(Ai , Ai+1 ) < dist(A, B). For each i = 1, . . . , N , define a similarity fi (that is, a composition of a homothety, an orthogonal transformation and a translation), such that fi (AB) = Ai Ai+1 . The ratio of fi is ai = dist(Ai , Ai+1 )/ dist(A, B). Starting from the segment Γ0 = AB, define by induction the polygonal curves Γn = ∪i fi (Γn−1 ). This sequence (Γn ) converges to a curve Γ which satisfies the following relation: Γ = ∪i fi (Γ). In other words, Γ is the attractor of the IFS {f1 , . . . , fn }. When Γ is simple, the dimensions δ and Δ assume a common value, which is also the similarity dimension, i.e. the unique solution of the equation N
axi = 1.
i=1
In the particular case where all distances dist(Ai , Ai+1 ) are the same, the ratios ai are equal to a value a such that N a > 1 (necessary condition for the continuity of Γ) and N a2 < 1 (necessary condition for the simplicity of Γ). Clearly, δ(Γ) = Δ(Γ) = log N/|log a|.
0
1/3
2/3
1
Figure 1.1. Von Koch curve, the attractor of a system of four similarities with common ratio
1 3
24
Scaling, Fractals and Wavelets
Function scales The previous definitions all involve ratios of logarithms. This is an immediate consequence of the fact that a dimension is defined as an order of growth related to the scale of functions {tα , α > 0}. In general, a scale of functions F in the neighborhood of 0 is a family of functions which are all comparable in the Hardy sense, that is, for any f and g in F, the ratio f (x)/g(x) tends to a limit (possibly +∞ or −∞) when x tends to 0. Function scales are defined in a similar way in the neighborhood of +∞. Scales other than {tα } will yield other types of dimensions. A dimension must be considered as a Dedekind cut in a given scale of functions. The following expressions will make this clearer: Δ(E) = inf{α such that εα Nε (E) → 0} δ(E) = sup{α such that εα Nε (E) → +∞}
(1.4) (1.5)
these are equivalent to equations (1.1) and (1.2) (see [TRI 99]). Complementary intervals on the line In the particular case where the compact E lies in an interval J of the line, the complementary set of E in J is a union of disjoint open intervals, whose lengths will be denoted by cn . Let |E| be the Lebesgue measure of E (which means, for an interval, its length). The dimension of E may be written as: log E(ε) Δ(E) = lim sup 1 − log ε ε→0 If |E| = 0, the sum of the cn is equal to the length of J. The dimension is then equal to the convergence exponent of the series cn :
α Δ(E) = inf α such that cn < +∞ (1.6) n
Proof. This result may be obtained by calculating an approximation of the length of Minkowski sausage E(ε). Let us assume that the complementary intervals are ranked in decreasing lengths: c1 c2 · · · cn · · · If |E| = 0 and if cn ε > cn+1 , then: |E(ε)| nε +
in
ci
Fractal and Multifractal Analysis in Signal Processing
thus εα−1 L(E(ε)) nεα + εα−1
in ci .
It may be shown that both values
α
inf{α such that nε < +∞} and
25
inf
α such that ε
α−1
ci < +∞
in
are equal to the convergence exponent. It is therefore equal to Δ(E). EXERCISE 1.1.– Verify formula (1.6) for the perfect symmetric sets of Example 1.1. If |E| = 0, then the convergence exponent of cn still makes sense. It characterizes a degree of proximity of the exterior with the set E. More precisely, we obtain log|E(ε) − E| α (1.7) cn < +∞ = lim sup 1 − inf α such that log ε ε→0 n where the set E(ε) − E refers to the Minkowski sausage of E deprived of the points of E. How can we generalize the study of the complementary set in Rn with n 2? The open intervals must be replaced with an appropriate paving. The results connecting the elements of this paving to the dimension depend both on the geometry of the tiles and on their respective positions. The topology of the complementary set must be investigated more deeply [TRI 87]. The index that generalizes (1.7) (replacing the 1 of the space dimension by n) is the fact fractal exponent, studied in [GRE 85, TRI 86b]. In the case of a zero area curve in R2 , this also leads to the notion of lateral dimension. Note that the dimensions corresponding to each side of the curve are not necessarily equal [TRI 99].
1.2.2. Packing dimension The packing dimension is, to some extent, a regularization of the box dimension [TRI 82]. Indeed, Δ is not σ-stable, but we may derive a σ-stable dimension from any index thanks to the operation described below. PROPOSITION 1.1.– Let B be the family of all bounded sets of Rn and α : B −→ R+ . Then, the function α ˆ defined for any subsets of Rn as: α ˆ (E) = inf{sup α(Ei )/E = ∪Ei , Ei ∈ B} is monotonous and σ-stable.
26
Scaling, Fractals and Wavelets
Proof. Any subset E of Rn is a union of bounded sets. If E1 ⊂ E2 , then any covering of E1 may be completed with a covering of E2 . This entails monotonicity. Now, let ε > 0 and a sequence (Ek )k1 of sets whose union is E. For any k, there ˆ (Ek ) + ε2−k . Since exists a decomposition (Ei,k ) of Ek such that sup α(Ei,k ) α E = ∪i,k Ei,k , we deduce that: α ˆ (E) sup α ˆ (Ek ) + ε
2−k = sup α ˆ (Ek ) + ε
k
k
Thus, the inequality α ˆ (E) supk α ˆ (Ek ) holds. The converse inequality stems from monotonicity. The packing dimension is the result of this operation on Δ. We set ˆ Dim = Δ The term packing will be explained later. The new index Dim is indeed a dimension, and it is σ-stable. Therefore, contrarily to Δ, it vanishes for countable sets. The inequality: Dim(E) Δ(E) is true for any bounded set. This becomes an equality when E presents a homogenous structure in the following sense: THEOREM 1.1.– Let E be a compact set such that, for all open sets U intersecting E, Δ(E ∩ U ) = Δ(E). Then Δ(E) = Dim(E). Proof. Let Ei be a decomposition of E. Since E is compact, a Baire theorem entails that the Ei are not all nowhere dense in E. Therefore, there exist an index i0 and an ¯ ∩ U , which yields: open set U intersecting E such that Ei0 ∩ U = E ¯ ∩ U ) Δ(E ∩ U ) = Δ(E) Δ(Ei0 ) = Δ(Ei0 ) Δ(Ei0 ∩ U ) = Δ(E As a result, Δ(E) supi Δ(Ei ), and thus Δ(E) Dim(E). The converse inequality is always true. EXAMPLE 1.3.– All self-similar sets are of this type, including those presented above: Cantor sets and curves. For these sets, the packing dimension has the same value as Δ(E).
Fractal and Multifractal Analysis in Signal Processing
27
EXAMPLE 1.4.– Dense sets in [0, 1], when they are not compact, do not necessarily have a packing dimension equal to 1. Let us consider, for any real p, 0 < p < 1, the set Ep of p-normal numbers, that is, those numbers whose frequency of zeros in their dyadic expansion is equal to p. Any dyadic interval of [0, 1], however small it may be, contains points of Ep , so Ep is dense in [0, 1]. As a consequence, Δ(Ep ) = 1. In contrast, the value of Dim(Ep ) is: Dim(Ep ) =
1 p log p + (1 − p) log(1 − p). log 2
This result will be derived in section 1.3.2. 1.2.3. Covering dimension The covering dimension was introduced by Hausdorff [HAU 19]. Here we adopt the traditional approach through Hausdorff measures; a direct approach, using Vitali’s covering convergence exponent, may be used to calculate the dimension without using measures [TRI 99]. Covering measures Originally, the covering measures were defined to generalize and, most of all, to precisely define the concepts of length, surface, volume, etc. They constitute an important tool in geometric measure theory. Firstly, let us consider a determining function φ: R+ −→ R+ , which is increasing and continuous in the neighborhood of 0, and such that φ(0) = 0. Let E be a set in a metric space (that is, a space where a distance has been defined). For every ε > 0, we consider all the coverings of E by bounded sets Ui of diameter diam(Ui ) ε. Let Hεφ (E) = inf φ diam(Ui ) /E ⊂ ∪i Ui , diam(Ei ) ε . When ε tends to 0, this quantity (possibly infinite) cannot decrease. The limit corresponds to the φ-Hausdorff measure: H φ (E) = lim Hεφ (E) ε→0
In this definition, the covering sets Ui can be taken in a more restricted family. If we suppose that Ui are open, or convex, the result remains unchanged. The main properties are that of any Borel measure: – E1 ⊂ E2 =⇒ H φ (E1 ) H φ (E2 );
28
Scaling, Fractals and Wavelets
– if (Ei ) is a collection of countable sets, then H φ (∪Ei )
H φ (Ei )
i
– if E1 and E2 are at non-zero distance from each other, any ε-covering of E1 is disjoint from any ε-covering of E2 when ε is sufficiently small. Then H φ (E1 ∪ E2 ) = H φ (E1 ) + H φ (E1 ). This implies that H φ is a metric measure. The Borel sets are H φ -measurable and for any collection (Ei ) of disjoint Borel sets, φ H (∪i Ei ) = i H φ (Ei ). The scale of functions tα In the case where φ(t) = tα with α > 0, we use the simple notation H φ = H α . Consider the case α = 1. For any curve Γ the value H 1 (Γ) is equal to the length of Γ. Therefore H 1 is a generalization of the concept of length: it may be applied to any subset of the metric space. Now let α = 2. For any plane surface S, the value of H 2 (S) is proportional to the area of S. For non-plane surfaces, H 2 provides an appropriate mathematical definition of area – using a triangulation of S is not acceptable from a theoretical point of view. More generally, when α is an integer, H α is proportional to the α-dimensional volume. However, α can also take non-integer values, which makes it possible to define the dimension of any set. The use of the term dimension is justified by the following property: if aE is the image of E by a homothety of ratio a, then H α (aE) = aα H α (E)
Measures estimated using boxes If we want to restrict the class of sets from which coverings are taken even more, one option would be to cover E with centered balls or dyadic boxes. In each case, the result is a measure H ∗α which is generally not equal to H α (E); nevertheless, it is an equivalent measure in the sense that we can find two non-zero constants, c1 and c2 , such that for any E: c1 H α (E) H ∗α (E) c2 H α (E) Clearly the H ∗α measures give rise to the same dimension.
Fractal and Multifractal Analysis in Signal Processing
29
Dimension For every E, there exists a unique critical value α such that: – H α (E) > 0 and β < α =⇒ H β (E) = +∞; – H α (E) < +∞ and β > α =⇒ H β (E) = 0. The dimension is defined as
dim(E) = inf α such that H α (E) = 0
= sup α such that H α (E)= + ∞
(1.8)
NOTE 1.1.– This approach is not very different from the one which leads to the box dimension. Compare equation (1.8) with equations (1.4) and (1.5). Once again, it may be generalized by using other function scales than tα . Properties The properties of dim directly stem from those of H α measures. It is a σ-stable dimension, like Dim. To compare all these dimensions, let us observe that δ can be defined in the same manner as the covering dimension, by using coverings made up of sets of equal diameter. This implies the inequality dim(E) δ(E) for any E. The σ-stability property then implies the following: ˆ ˆ dim(E) δ(E) Δ(E) = Dim(E) Δ(E). These inequalities may be strict. However, the equality dim(E) = Dim(E), and even dim(E) = Δ(E), occur in cases where E is sufficiently regular. Examples include rectifiable curves and self-similar sets. Packing measures By considering packings of E, that is, families of disjoint sets at zero distance from E, and by switching inf and sup in the definitions, it is possible to define packing measures which are symmetric to the covering measures and whose critical index is precisely equal to Dim. This explains why Dim is called a packing dimension. 1.2.4. Methods for calculating dimensions Since a dimension is an index of irregularity, it may be used for the classification of irregular sets and for characterizing phenomenons showing erratic behaviors. Here we focus on signals. In practice, we may assume that the signal Γ is given in axes Oxy by discrete data (xi , fi ), with xi < xi+1 : for example, a time series. Notice that other sets, such as geographical curves obtained by aerial photography, are not of this type, so that the analysis tools can be different.
30
Scaling, Fractals and Wavelets
The algorithms used for estimating a dimension rely on the theoretical formula which defines the dimension, with the usual limitation of a minimal scale, roughly equal to the distances xi+1 − xi . Indeed, it is impossible to go further into the dataset structure without a reconstruction at very small scales, which leads to adding up arbitrarily new data. The evaluation of a dimension, whose theoretical definition requires the finest coverings possible, is therefore difficult to justify. This is why we do not propose algorithms for the estimation of σ-stable dimensions such as Hausdorff or packing dimensions. Calculations may be performed to estimate Δ or related dimensions that are naturally adapted to signal analysis. They are usually carried out using logarithmic diagrams. A quantity Q(f, ε) is estimated, which characterizes the irregularity for a certain number of values of the resolution ε between two limit values εmax and εmin . If Q(f, ε) follows a power law of the type cεΔ , then log Q(f, ε) is an affine function of log ε, with slope Δ. The idea is to seek functions Q(f, ε) which provide appropriate logarithmic diagrams, i.e. allow us to estimate the slopes precisely. Here are some examples. Boxes and disks After counting the number Nε of squares of a network of sides ε, we draw the diagram (|log ε|, log Nε ). Although it is very easy to program, the method presents an obvious disadvantage: the quantities Nε are integers, which makes the diagram chaotic for large values of ε. We could try a method using the Minkowski sausage of Γ: 1 |log ε|, log 2 Aire Γ(ε) ε However, this method is more difficult to program than the previous method and also lacks precision: the diagram shows, in general, a strong concavity – even for a curve as simple as a straight line segment! These methods are very popular, in spite of numerical difficulties. Unfortunately there is a major drawback for signals. The coordinates (xi , fi ) are not, in general, of the same nature. If they refer, for example, to stock exchange data, xi is a time value and fi an exchange value. In this case, it makes no sense to give the units the same value on the axis Ox and Oy. The covering of Γ by squares or disks is therefore meaningless. It is preferable to use algorithms which will provide a slope independent of changes of unit. For this purpose, the calculated quantity Q(f, ε) should satisfy the following properties: for any real a, there exists c(a), such that, for any ε: Q(af, ε) = c(a) Q(f, ε) as is the case for the methods which are described below.
(1.9)
Fractal and Multifractal Analysis in Signal Processing
31
Variation method Here we anticipate section 1.3.4. The oscillation of a function f on any set F is defined as β(f, F ) = sup{f (t) − f (t ) t, t ∈ F } The ε-variation on an interval J is the arithmetic mean of the oscillation over intervals of length 2ε: 1 β(f, [t − ε, t + ε] ∩ J) dt varε (f ) = |J| J The variation method [DUB 89, TRI 88] consists of finding the slope of the diagram: 1 |log ε|, log 2 varε (f ) . ε Since β(af, [x − ε, x + ε]) = a β(f, [x − ε, x + ε]) for all x and ε, we obtain through integration varε (af ) = a varε (f ), so equation (1.9) is satisfied. In this case we obtain diagrams which present an almost perfect alignment for a large class of known signals. Furthermore, this method presents the following advantages: ease of programming and speed of execution. Lp norms method Let us define the Lp norm of a function f : D ⊂ Rn −→ R by the relation: 1/p 1 p |f (x)| dx . Lp (f ) = Voln (D) D It is a functional norm when p 1. When p → +∞, the expression Lp (f ) tends to the norm: L∞ (f ) = sup |f (x)| x∈D
Given a signal f defined on [a, b] and the values x ∈ [a, b] and ε > 0, we apply this tool at any x to the local function difference defined by f (x) − f (x ) where x − x ε. Using the norm L∞ , this gives supx ∈[x−ε,x+ε] (|f (x) − f (x )|). This quantity is equivalent to ε-oscillation, since 1 β f, [x − ε, x + ε] sup (|f (x) − f (x )|) β f, [x − ε, x + ε] 2 x ∈[x−ε,x+ε]
32
Scaling, Fractals and Wavelets
It is therefore possible to replace the ε-variation of f by the integral over J of supx ∈[x−ε,x+ε] (|f (x) − f (x )|), without altering the theoretical result for Δ. However, it is also possible to use Lp norms. Indeed, the oscillation (or the local norm L∞ ) only takes into account the peaks of the function. In practice, it can happen that these peaks are measured with a significant error, or even destroyed in the process of acquisition of data (profiles of rough surfaces, for example). It is preferable to use all the intermediate values and replace the ε-variation with the quantity: J
1 2ε
x+ε
|f (x) − f (x )|p dx
1/p dt
x−ε
In this expression, large values of p allow us to emphasize the effect of local peaks, whereas if p = 1, all the values of function f have equal importance. These integrals make it possible to rectify the corresponding logarithmic diagram and to calculate the slope with precision. We can also replace the above integral on J by a norm Lq , with q > 1. If q is large, this will take into account the more irregular parts of the signal. We can also change the integral in the window [x−ε, x+ε] into a convolution product by a kernel of type K(x /ε), so that the results are even smoother. However, it should be noted that except for particular cases (Weierstrass functions, for example), we do not exactly calculate the dimension Δ with these methods, but rather an index smaller than Δ [TRI 99], which nevertheless remains relevant to the signal irregularity. Let us develop an example of the index just referred to. Let K be a kernel belonging to the Schwartz class, with integral equal to 1. Let Ka (t) = a1 K( at ) for a > 0. For a function f defined in a compact, let f a be the convolution of f with Ka . Since f a is regular, the length Λa of its graph is finite. We define the regularization dimension dimR (f ) of the graph of f as: dimR (f ) = 1 + lim
a→0
log(Λa ) . − log a
(1.10)
This dimension measures the speed at which the length of less and less regularized versions of the graph of f tend to infinity. It is easily proved that if f is continuous, the inequality dimR Δ is always true. An interesting aspect of the dimension of regularization is that it is a well-adapted estimation tool. Results obtained on usual signals (Weierstrass function, iterated function system and Brownian fractional motion) are generally satisfactory, even for small-sized samples (a few hundred points). Moreover, the simple analytical form of dimR allows us to easily obtain an estimator for data corrupted by an additional noise, which is particularly useful in signal processing (see [ROU 98] and the FracLab manual for more details).
Fractal and Multifractal Analysis in Signal Processing
33
1.3. Hölder exponents 1.3.1. Hölder exponents related to a measure The dimensional analysis of a set is related to its local properties. To go further into this study, it is convenient to use a measure μ supported by the set. In many cases (self-similar sets, for example), E is defined at the same time as μ. If E is a curve, constructing a measure on E is called a parameterization. Without a parameterization it is impossible to analyze the curve. However, a given set can support very different measures. Particularly interesting ones are the well-balanced measures, in a sense we will explain. Given a measure μ of Rn , let us first define the Hölder exponent of μ over any measurable set F by αμ (F ) =
log μ(F ) log diam(F )
By convention, 0/0 = 1 and 1/0 = +∞. Given a set E, we use this notion in a local manner, i.e. on arbitrarily small intervals or cubes intersecting E. A pointwise Hölder exponent is then defined using centered balls Bε (x) whose radius tends to 0: αμ (x) = lim inf αμ Bε (x) ε→0
The symmetric exponent can also be useful: αμ (x) = lim sup αμ Bε (x) ε→0
In addition, the geometric context sometimes induces a specific analysis framework. If a measure is defined by its value on the dyadic cubes, it will be easier to use the following Hölder exponents: αμ∗ (x) = lim sup αμ un (x) , α∗μ (x) = lim inf αμ un (x) n→+∞
n→+∞
N where un (x) is the cube of i=1 [ki 2−n , (ki + 1)2−n [ that contains x. Other covering nets can obviously be used, but dyadic cubes are well suited for calculations. 1.3.2. Theorems on set dimensions The first theorem can be used as a basis for understanding the subsequent more technical results.
34
Scaling, Fractals and Wavelets
THEOREM 1.2.– Let μ be a finite measure such that μ(E) > 0. Assume that there exists a real α, such that for any x ∈ E: αμ∗ (x) = α∗μ (x) = α Then α = dim(E) = Dim(E). Proof. Let ε > 0. Let En be the subset of E consisting of all points x such that, for k n: α − ε αμ uk (x) α + ε If ρ < 2−n , any ρ-covering {ui } of En by dyadic cubes of rank ≥ n is such that: |ui |α+ε μ(ui ) |ui |α−ε for all i. Therefore μ(En ) i |ui |α−ε and i |ui |α+ε μ . First, we deduce that H ∗(α−ε) (En ) μ(En ). If n is large enough then μ(En ) > 0. This gives dim(En ) α − ε. Since En ⊂ E, then dim(E) α − ε. Secondly, by using the covering of En formed by dyadic cubes of rank k n, we obtain: N2−k (En )2−k(α+ε) μ Therefore N2−k (En ) μ 2k(α+ε) , which implies Δ(En ) α + ε. As a consequence, Dim(E) α + ε by σ-stability. By making ε tend to 0, we obtain the desired result. An analogous theorem can be stated with balls Bε (x) centered at E. We can develop the arguments of the preceding proof to obtain more general results as follows. THEOREM 1.3.– Assume that μ is a finite measure such that μ(E) > 0. Then:
(1.11) inf αμ (x) dim(E) sup αμ (x) x∈E
x∈E
inf αμ (x) Dim(E) sup αμ (x)
(1.12)
inf αμ (x) Δ(E) lim sup sup αμ (x)
(1.13)
x∈E
x∈E
x∈E
ε→0
x∈E
Fractal and Multifractal Analysis in Signal Processing
35
Inequality (1.13) seems more complex than the others. Nevertheless, we can derive from it a simple result: if 0 < μ(E) < +∞ and if α(Bε (x)) converges uniformly on E to a number α, then α = Δ(E). The same results hold if, for example, we replace the network of balls centered at E with that of dyadic cubes. EXAMPLE 1.5.– A perfect symmetric set (see Example 1.1) is the support of a natural or canonical measure: each of the 2n covering intervals of rank n is associated with the weight 2−n . In the case where the set has constant ratio a, these intervals have rank n and their Hölder exponent assumes the value log 2/|log a| uniformly. This grid of intervals allows the computation of dimensions. Indeed dim(E) = Dim(E) = Δ(E) =
log 2 |log a|
By making the successive ratios an vary, it is also possible to construct a set such that: dim(E) = lim inf
n log 2 , |log an |
Dim(E) = Δ(E) = lim sup
n log 2 |log an |
EXAMPLE 1.6.– The set Ep of p-normal numbers (see Example 1.4) supports a measure which makes it possible to estimate its dimension and which is known as the Besicovitch measure. It is defined on the dyadic intervals [0, 1]. Set μ([0, 12 ]) = p and μ([ 12 , 1]) = 1 − p, p ∈ (0, 1). The weights p and 1 − p are then distributed similarly at the following stages: Since each dyadic interval un of rank n is the union of the intervals vn (on the left) and vn (on the right) of rank n + 1, we put μ(vn ) = p μ(un ),
μ(vn ) = (1 − p) μ(un )
It is easy to calculate the exact measure of each dyadic interval un (x) containing the point x by using the base 2 expansion of x. Denote: N0 (x, n) = number of 0 in the expansion of x between ranks 1 and n N1 (x, n) = number of 1 in the expansion of x between ranks 1 and n Thus, N0 (x, n) + N1 (x, n) = n and: μ un (x) = pN0 (x,n) (1 − p)N1 (x,n)
(1.14)
36
Scaling, Fractals and Wavelets
First, let us show that Ep has full measure. The easiest way to proceed is to use the language of probability. Each point x can be viewed as the result of a process which is a sequence of independent Bernoulli random variables Xi taking n the value 0 and 1 with probabilities p and 1 − p. The frequency N1 (x, n)/n = ( 1 Xi )/n has mean 1 − p and variance p(1 − p)/n. We may apply here the strong law of large numbers: with probability 1, N1 (x, n)/n tends to 1 − p when n → +∞, and N0 (x, n)/n tends to p. Coming back to the language of measure, this result tells us that the set of x for which N0 (x, n)/n tends to p has measure 1. Such x are in Ep . Thus μ(Ep ) = 1. Secondly, to compute the dimension, we first need to determine the value of the Hölder exponent. Equation (1.14) implies the following result: 1 1 |log p| |log(1 − p)| + N1 (x, n) α un (x) = N0 (x, n) n log 2 n log 2 for any x of [0, 1]. If x ∈ Ep , then α(un (x)) tends to the value: αμ∗ (x) = α∗μ (x) = p
|log(1 − p)| |log p| + (1 − p) log 2 log 2
(1.15)
Thus, the value is the same for dim(Ep ) and Dim(Ep ) according to equations (1.11) and (1.12). We observe that in equation (1.13), the left-hand side is also equal to this value. Moreover, supx∈Ep α(un (x)) is equal to the largest Hölder exponent of dyadic intervals of rank n. If, for example, p < 12 , this largest exponent is equal to |log p|/ log 2, which is larger than 1. Therefore, the right-hand side of (1.13) is larger than 1. In fact, equation (1.13) gives no indication on the value of Δ(E). An argument of density yields Δ(Ep ) = 1 for any value of p.
1.3.3. Hölder exponent related to a function The Hölder exponents of a function give much more information than those of measures. Firstly, let us generalize the notion of measure of a set F , by using the notion of oscillation of the function f in F : β(f, F ) = sup{f (t) − f (t ) for t, t ∈ F }
This allows us to define a Hölder exponent: α(F ) =
log β(f, F ) log diam(F )
Fractal and Multifractal Analysis in Signal Processing
37
Given an interval J and a function f : J → R, we may use this notion locally in arbitrarily small neighborhoods of t ∈ J. The pointwise Hölder exponent of f in t is obtained as αpf (t) = lim inf α [t − ε, t + ε] ∩ J ε→0
According to this definition, the exponent of an increasing function f is the same as that of the measure μ defined on any [c, d] ⊂ J by μ([c, d]) = f (d) − f (c). Indeed, f (d) − f (c) is also the oscillation value of f in [c, d]. However, in general, f is not monotonous and it is therefore necessary to carry out a more accurate analysis, as we will see below. As in the case of measures, we may also consider the “symmetric” exponent defined with an upper limit, and also the exponents obtained as lower and upper limits by using particular grids of intervals, like the dyadic intervals. Oscillation considered as a measurement of the local variability of a function possesses many advantages. In particular, it is closely related to the box dimension. However, there are some counterparts: it is not simple to use in a theoretical context, it is sometimes difficult to estimate with precision under experimental conditions and, finally, it is sensitive to various disturbances (discretization, noise, etc.). It is possible to replace the oscillation with other set functions v(f, F ) showing more robustness. However, most alternatives no longer verify the important triangle inequality: v(f, F1 ∪ F2 ) v(f, F1 ) + v(f, F2 ) for all F1 , F2 ⊂ J (see [TRI 99]). We can simplify the analysis by restricting the general F to the class of intervals and by setting: v(f, [a, b]) = |f (b) − f (a)|.
We may even consider only dyadic intervals, and take: v f, [k 2−n , (k + 1) 2−n ] = |cj,k | where cj,k is the wavelet coefficient of f at scale j and position k (see also Chapters 2, 8 and 9). Let us now give an alternate and useful definition of the Hölder exponent.
38
Scaling, Fractals and Wavelets
DEFINITION 1.2.– Let f : R → R be a function, s > 0 and x0 a real number. Then f ∈ C s (x0 ), if and only if there is a real number η > 0, a polynomial P of degree ≤ s and a constant C, such that ∀x ∈ B(x0 , η),
|f (x) − P (x − x0 )| C|x − x0 |s .
(1.16)
The pointwise Hölder exponent of f at x0 , denoted αpf (x0 ) or αp (x0 ), is given by
sup s/f ∈ C s x0 If 0 < s < 1, then the polynomial P is simply the constant f (x0 ) and the increments of f over [x0 − ε, x0 + ε] are indeed compared to ε. Considering P allows us to take into account the higher order singularities, i.e. in the derivatives of f . We remove the “regular part” of f to exhibit the singular behavior. Consider for example f (x) = x + |x|3/2 . Using P allows us to find αp = 32 , whereas the simple increment would give αp = 1, a non-significant value in this case. The pointwise exponent, whose definition is natural, has a geometric interpretation. To begin with, let us remove the signal’s regular part, thus performing the difference f (x) − P (x − x0 ). Around x0 , the signal thus obtained is entirely contained in a hull of the form C|x − x0 |αp +ε for any ε > 0, and this hull is optimal, i.e. any smaller hull C|x − x0 |αp −ε , does not contain an infinite number of points of the signal. We observe that the smaller the αp is, the more irregular f is in the neighborhood of x0 and vice versa. In addition, αp > 1 implies that f is derivable at x0 , and a discontinuous function at x0 is such that αp = 0. In many applications, the regularity of a signal is as important as, or more important, than its amplitude. For example, the edges of an image do not vary if an affine transformation of the gray-levels is carried out. This will modify the amplitude, but not the exponents. The pointwise Hölder exponent will thus be one of the main ingredients for the fractal analysis of a signal or image. Generally speaking, measuring and studying the pointwise exponent is particularly useful in the processing of strongly irregular signals whose irregularity is expected to contain relevant information. Examples includes biomedical signals and images (ECG, EEG, echography, scintigraphy), Internet traffic logs, radar images, etc. The pointwise exponent, however, presents some drawbacks. For example, it is neither stable by integral-differentiation nor, more generally, under the action of pseudo-differential operators. This means that it is not possible to predict the exponent at x of a primitive F of f knowing αpf (x). It is only guaranteed that αpF (x) αpf (x) + 1. In the same way, the exponent of the analytical signal associated with f is not necessarily the same as f . This is a problem in signal processing, since
Fractal and Multifractal Analysis in Signal Processing
39
this type of operator is often used. The second disadvantage, related to the first, is that αp (x) does not provide the whole information on the regularity of f in x. A common example is that of the chirp |x|γ sin(1/|x|δ ), with γ and δ positive. We verify that the pointwise exponent at 0 is equal to γ. In particular, it is independent of δ: αp (0) is not sensitive to “oscillating singularities”, i.e. to situations where the local frequency of the signal tends to infinity in the neighborhood of 0. It is therefore necessary to introduce at least one more exponent to fully describe the local irregularity. During the last few years, several quantities have been proposed, including the chirp exponent, the oscillation exponent, etc. Here we will focus on the local Hölder exponent, which we now define (for more details on the properties of this exponent, see [LEV 98a, SEU 02]). Let us first recall that, given a function f : Ω → R, where Ω ⊂ R is an open set, we say that f belongs to the global Hölder space Cls (Ω), with 0 < s < 1, if there is a constant C, such that for any x, y in Ω: |f (x) − f (y)| C|x − y|s
(1.17)
If m < s < m + 1 (m ∈ N), then f ∈ Cls (Ω) means that there exists a constant C, such that, for any x, y in Ω: |∂ m f (x) − ∂ m f (y)| C|x − y|s−m
(1.18)
Now let αl (Ω) = sup{s/f ∈ Cls (Ω)}. Notice that if Ω ⊂ Ω, then αl (Ω ) αl (Ω). To define the local Hölder exponent, we will use the following lemma. LEMMA 1.1.– Let (Oi )i∈I be a family of decreasing open sets (i.e. Oi ⊂ Oj if i > j), such that: ∩i Oi = {x0 }
(1.19)
αl (x0 ) = sup{αl (Oi )}
(1.20)
Let: i∈I
Then, αl (x0 ) does not depend on the choice of the family (Oi )i∈I .
This result makes it possible to define the local exponent by using any intervals family containing x0 .
40
Scaling, Fractals and Wavelets
DEFINITION 1.3.– Let f be a function defined in a neighborhood of x0 . Let {In }n∈N be a decreasing sequence of open intervals converging to x0 . By definition, the local Hölder exponent of f at x0 , noted αl (x0 ), is: αl (x0 ) = sup αl (In ) = lim αl (In ) n∈N
n→+∞
(1.21)
Let us briefly note that the local exponent is related to a notion of critical exponent of fractional derivation [KOL 01]. We may understand the difference between αp and αl as follows: let us suppose that there exists a single couple (y, z) such that β(f, B(x, ε)) = f (y)−f (z). Then αp results from the comparison between β(f, B(x, ε)) and ε, whereas for αl , we compare β(f, B(x, ε)) to |y − z|. This is particularly clear in the case of the chirp, where the distance between the points (y, z) realizing the oscillation tends to zero much faster than the size of the ball around 0. Accordingly, it is easy to demonstrate that αl (0) = γ/(1 + δ) for the chirp. The exponent αl thus “sees” oscillations around 0: for fixed γ, the chirp is more irregular (in the sense of αl ) when δ is larger. The local exponent possesses an advantage over αp : it is stable under the action of pseudo-differential operators. However, as well as αp , αl cannot by itself completely characterize the irregularity around a point. Moreover, αl is, in a certain sense, less “precise” than αp . The results presented below support this assertion. PROPOSITION 1.2.– For any continuous f and for all x: αlf (x) min αpf (x), lim inf αpf (t) t→x
The following two theorems describe the structure of the Hölder functions, i.e. the functions which associate with any x the exponents of f at x. THEOREM 1.4.– Let g : R → R+ be a function. The two assertions below are equivalent: – g is the lower limit of a sequence of continuous functions; – there exists a continuous function f whose pointwise Hölder function αp (x) satisfies αp (x) = g(x) for all x. THEOREM 1.5.– Let g : R → R+ be a function. The following two assertions are equivalent: – g is a lower semi-continuous (LSC) function; – there exists a continuous function f whose local Hölder function αl (x) satisfies αl (x) = g(x) for any x.
Fractal and Multifractal Analysis in Signal Processing
41
NOTE 1.2.– Let us recall that a function f : D ⊂ R → R is LSC if, for any x ∈ D and for any sequence (xn ) in D tending to x: lim inf f (xn ) f (x) n→∞
(1.22)
Figure 1.2 shows a generalization of the Weierstrass function defined on [0, 1] for which αp (x) = αl (x) = x for any x. This function is defined as ∞ Wg (x) = i=0 ω −nx cos(ω n x), with ω > 1.
Figure 1.2. Generalized Weierstrass function for which αp (x) = αl (x) = x
Since the class of lower limits of continuous function is much larger than that of lower semi-continuous functions, we observe that αp generally supplies more information than αl . For example, αp can vary much “faster” than αl . In particular, it is possible to construct a continuous function whose pointwise Hölder function coincides with the indicator function of the set of rational numbers. It is everywhere discontinuous, and thus its local Hölder function is constantly equal to 0. The following results describe more precisely the relations between αl and αp . PROPOSITION 1.3.– Let f : I → R be a continuous function, and assume that there exists γ > 0 such that f ∈ C γ (I). Then, there exists a subset D of I such that: – D is dense, uncountable and has Hausdorff dimension zero; – for any x ∈ D, αp (x) = αl (x).
42
Scaling, Fractals and Wavelets
Moreover, this result is optimal, i.e. there exists a function of global regularity γ > 0 such that αp (x) = αl (x) for all x outside a set of zero Hausdorff dimension. THEOREM 1.6.– Let 0 < γ < 1 and f : [0, 1] → [γ, 1] be a lower limit of continuous functions. Let g : [0, 1] → [γ, 1] be a lower semi-continuous function. Assume that for all t ∈ [0, 1], f (t) g(t). Then, there exists a continuous function F : [0, 1] → R such that: – for all x, αl (x) = g(x); – for all x outside a set of zero Hausdorff dimension, αp (x) = f (x). This theorem shows that, when the “compatibility condition” f (t) g(t) is satisfied, we can simultaneously and independently prescribe the local and pointwise regularity of a function outside a “small” set. These two measures of irregularity are thus to some extent independent and provide complementary information. 1.3.4. Signal dimension theorem Let us investigate the relationships between the dimension of a signal and its Hölder exponents. There is no general result concerning the Hausdorff dimension, apart from obvious upper bounds resulting from the inequalities dim(Γ) Dim(Γ) Δ(Γ). Here is a result for Dim(Γ) [TRI 86a]. THEOREM 1.7.– If Γ is the graph of a continuous function f , then: 2 − sup αμ (x) Dim(Γ) 2 − inf αμ (x) . x∈J
x∈J
The same inequalities are true if we use the grid of the dyadic intervals: 2 − sup α∗μ (x) Dim(Γ) 2 − inf α∗μ (x) . x∈J
x∈J
(1.23)
(1.24)
We do not provide the demonstration of these results, which requires an evaluation of the packing measure of the graph. In the same context, we could show that if the local Hölder exponents α(un (x)) tend uniformly to a real α, then this number is also equal to Δ(Γ). However, a much more interesting equality may be given for Minkowski-Bouligand dimension of the graph which is both simple and general. THEOREM 1.8.– Let f be a continuous function defined on an interval J, and non-constant on J. For any ε > 0, let us call ε-variation of f on J the arithmetic mean of ε-oscillations: 1 β f, [x − ε, x + ε] ∩ J dt varε (f ) = |J| J
Fractal and Multifractal Analysis in Signal Processing
43
then: log varε (f ) Δ(Γ) = lim sup 2 − log ε ε→0
(1.25)
The assumption that f is not constant is necessary, as otherwise the oscillations are all zero and varε (f ) = 0. In this case, the graph is a horizontal segment and the value of its dimension is 1. Proof. A proof [TRI 99] using geometric arguments consists of estimating the area of the Minkowski ε-sausage Γ(ε). We show that this is equivalent to that of the union of the horizontal segments ∪t∈J [t − ε, t + ε] × {z(t)} centered on the graph. This is equal to the variation varε (z). EXAMPLE 1.7.– The graph of a self-affine function defined on J = [a, b] may be obtained as the attractor of an iterated functions system, like the self-similar curves of Example 1.2 (see also Chapters 9 and 10). For this, it is sufficient to define: – an integer N 2; – N + 1 points in the plane A1 = (x1 , y1 ), . . . , AN +1 = (xN +1 , yN +1 ), such that x1 = a < · · · < xi < · · · < xN +1 = b; – N affine triangular applications of the plane T1 ,. . . ,TN , such that, for each Ti , the image of the segment A1 AN +1 is the segment Ai Ai+1 . These may be written as: Ti =
ρi hi
0 δi
+ i ηi
where 0 < ρi = (xi+1 − xi )/(b − a) < 1 and |δi | < 1. The five parameters of Ti are related by the relations Ti (A1 ) = Ai and Ti (AN +1 ) = Ai+1 . In the particular case where ρi = 1/N for any i and i |δ i | 1, we may verify that if ε = N −k , the quantity varε (f ) is of the order of (( i |δi |)/N )k . We then obtain: log i |δi | |log( i |δi |)/N | =1+ Δ(Γ) = 2 − log N log N (see the classical example of Figure 1.3 where N = 4 and δi = 12 for any i). The Hölder exponent is calculated using 4-adic intervals. Its uniform value is 12 . Therefore Δ(Γ) = Dim(Γ) = 32 . Let us note that the Hausdorff dimension, strictly lower than 3 2 , is much more difficult to estimate [MCM 84].
44
Scaling, Fractals and Wavelets
EXAMPLE 1.8.– Lacunary series, such as the Weierstrass function, provide other examples of signals: f (x) =
∞
ω −nH cos(ω n x + φn )
i=0
where ω > 1 and 0 < H < 1. The values of the “phases” φn are arbitrary. We can directly prove [TRI 99] that, restricted to any bounded interval J, the box dimension of the graph Γ is equal to 2 − H. By homogenity, Dim(Γ) = 2 − H. This confirms the fact – although it is more difficult to prove (see [BOU 00]) – that the ε-oscillations are uniformly of the order of εH . In other words, there exist two constants c1 and c2 > 0 such that, for any t ∈ J and for any ε, with 0 < ε < 1: c1 εH β f, [x − ε, x + ε] c2 εH
Figure 1.3. Graph of a nowhere derivable function, attractor of a system of four affine applications, such that ρi = 14 and δi = 12 . Dimensions Δ and Dim are equal to 32 . The Hausdorff dimension lies between 1 and 32
Fractal and Multifractal Analysis in Signal Processing
45
Finding the value of the Hausdorff dimension of such graphs is still an open problem. In the case where the φn are independent random variables with same distribution, it is known that dim(Γ) = 2 − H with probability 1. 1.3.5. 2-microlocal analysis In this section, we briefly present an analysis of local regularity which is much finer than the Hölder exponents described above. As was observed previously, neither αp nor αl allow us to completely describe the local irregularity of a function. To obtain an exhaustive description, we need to consider an infinite number of exponents. This is the purpose of 2-microlocal analysis. This powerful tool was defined in [BON 86], where it is introduced in the framework PDE in the frame of Littlewood-Paley analysis. A definition based on wavelets is developed in [JAF 96]. We present here a time-domain characterization [KOL 02].
DEFINITION 1.4.– A function f : I ⊂ R → R belongs to the 2-microlocal space Cxs,s 0 if there exist 0 < δ < 14 and C > 0 such that, for all (x, y) satisfying |x − x0 | < δ and |y − x0 | < δ: |f (x) − f (y)| −s /2 −s /2 |x − y| + |y − x0 | C|x − y|s+s |x − y| + |x − x0 |
Figure 1.4. Graph of the Weierstrass function for ω = 3 and H = 12 . Phases φn are random independent variables and are identically distributed. Dimensions Δ and Dim are equal to 32 . The Hausdorff dimension is almost surely close to 32
(1.26)
46
Scaling, Fractals and Wavelets
This definition is valid only for 0 < s < 1 and 0 < s + s < 1. The general case is slightly more complex and will not be dealt with here. 2-microlocal spaces, as opposed to Hölder space, require two exponents (s, s ) in their definition. While αp is defined as the sup of exponents α, such that f belongs to Cxα0 , we cannot proceed in the same way to define “2-microlocal exponents”. Instead, we define in the abstract }). plan (s, s ) the 2-microlocal frontier of f at x0 as the curve (s, sup{s , f ∈ Cxs,s 0 It is not hard to show that this curve is well defined, concave and decreasing. Its intersection with the s axis is exactly αl . Moreover, under the hypothesis that f possesses a minimum global regularity, αp is the intersection of the frontier with the line s + s = 0. The 2-microlocal frontier thus allows us to re-interpret the two exponents within a unified framework. The main advantage of the frontier is that it completely describes the evolutions of αp under integro-differentiation of arbitrary order: indeed, an integro-differentiation of order ε simply shifts the frontier by ε along the s axis. Thus, 2-microlocal analysis provides extremely rich information on the regularity of a function around a point. To conclude this brief presentation, let us mention that algorithms exist which make it possible to numerically estimate the 2-microlocal frontier. They often allow us to calculate the values of αp and αl more precisely than a direct method (various estimation methods for the exponents and the 2-microlocal frontier are proposed in FracLab). Furthermore, it is possible to develop a 2-microlocal formalism [LEV 04a] which presents strong analogies with the multifractal formalism (see below). 1.3.6. An example: analysis of stock market price As an illustration of some of the notions introduced above, we use this section to detail a simplified example of the fractal analysis of a signal based on Hölder exponents. Our purpose is not to develop a complete application (this would require a whole chapter) but instead to demonstrate how we calculate and use the information provided by local regularity analysis in a practical scenario. The signal is a stock market log. Financial analysis offers an interesting area of application for fractal methods (see Chapter 13 for a detailed study). We will consider the evolution of the Nikkei index between January 1, 1980 and November 5, 2000. These signals comprise 5,313 values and are presented in Figure 1.5. As with many stock market accounts, it is extremely irregular. We calculate the local Hölder exponent of the logarithm of this signal, which is the quantity on which financial analysts work. The exponent is calculated by a direct application of Definition 1.3: at each point x, we find, for increasing values of ε, one couple (y, z) for which the signal oscillation in a ball centered in x of radius ε is attained. A bilogarithmic regression between the vector of the values found for the oscillation and the distances |y − z| is then performed (see the FracLab manual for more details on the procedure). As Figure 1.6 shows, most local exponents are comprised between 0 and 1, with
Fractal and Multifractal Analysis in Signal Processing
47
Figure 1.5. The Nikkei index between 1st January 1980 and 5th November 2000
some peaks above 1 and up to 3. The exponent values in [0, 1], which imply that the signal is continuous but not differentiable, confirm the visual impression of great irregularity. We can go beyond this qualitative comment by observing that the periods where “something occurs” have a definite signature in terms of the exponents: they are characterized by a dramatic increase of αl followed by very small values, below 0.2. Let us examine some examples. The most remarkable point of the local regularity graph is its maximum at abscissa 2,018, with an amplitude of 3. The most singular points, i.e. those with the smallest exponent, are situated just after this maximum: the exponent is around 0.2 for the abscissae between 2,020 and 2,050, and of the order of 0.05 between points 2,075 and 2,100. These two values are distinctly below the exponent average, which is 0.4 (the variance being 0.036). Calculations show that less than 10% of the log points possess an exponent smaller than 0.2. This remarkable behavior corresponds to the crash of October 19, 1987 which occurs at abscissa 2,036, in the middle of the first zone of points with low regularity after the maximum: the most “irregular” days of the entire signal are thus, as expected, situated in the weeks which followed the crash. It is worthwhile noting that this fact is much more apparent on the regularity signal than on the original log, where only the crash appears clearly, with the subsequent period not displaying remarkable features. Let us now consider another area which contains many points of low regularity along with some isolated regular points (i.e. having αl > 1). It corresponds to the zone between abscissae 4,450 and 4,800: this period approximatively corresponds to the Asian crisis that took place between January 1997 and June 1998 (analysts do not agree upon the exact dates of the beginning and the ending of this crisis: some of them date its beginning in mid-1997 or its end towards the end of 1999, or much later). On the graph of the log, we can observe that this period seems slightly more
48
Scaling, Fractals and Wavelets
Figure 1.6. Local Hölder function of the Nikkei index
irregular than others. In terms of exponents, we notice that it contains two maxima, with values greater than 1, both followed by low regularity points: this area comprises a high proportion of irregular points, since 12% of its points have an exponent lower than 0.15. This proportion is three times higher than that observed in the whole log. The analysis just performed has remained at an elementary level. However, it has allowed us to show that major events have repercussions on the evolution of the local Hölder exponent and that the graph of αl emphasizes features not easily visible on the original log.
1.4. Multifractal analysis 1.4.1. What is the purpose of multifractal analysis? In the previous section, it has been observed that the Hölder functions provide precise information on the regularity at each point of a signal. In applications, this information is often useful as such, but there exists many cases where it is necessary to go further. Here are three examples that highlight this necessity. In image segmentation, we expect that edges correspond to low regularity points and hence to small exponents. However, the precise value of the Hölder exponents of contour points cannot be universal: non-linear transformations of an image, for instance, might preserve edges while modifying the exponent value. In this first situation, we see that the pointwise regularity does not provide all the relevant information and that it is necessary to supplement it with structural information. Let us now consider the issue of image texture classification. It is clear that classifying a pixel based on its
Fractal and Multifractal Analysis in Signal Processing
49
local exponent would not give satisfactory results. A more relevant approach would be to use the statistical distribution of exponents with zones. The same comment applies to the characterization of Internet traffic according to its degree of sporadicity. In this second situation, the Hölder function provides information that is too rich, and which we would like to balance in a certain sense. The last situation is when the exponents are too difficult to calculate: there exists, in particular, continuous signals, easy to synthesize, whose Hölder function is everywhere discontinuous. In this case, the pointwise regularity information is too complex to be used under its original form. In all these examples, we would like to use “higher level” information, which would be extracted from the Hölder function or would sum up, in some sense, its relevant features. Several ways of doing this exist. The idea that comes immediately to mind is simply to calculate histograms of exponents. This approach, however, is not adapted, both for mathematical reasons that go beyond the scope of this chapter and because, in fractal analysis, we always try to deal with quantities that are scale-independent. The most relevant way to extract information from the Hölder function and to describe it globally is to perform a multifractal analysis. There are many variants and we will concentrate on two popular examples: the first is geometric and consists of calculating the dimension of the set points possessing the same exponent. The second is statistical: we study the probability of finding, at a fixed resolution, a given exponent and how this probability evolves when the resolution tends to infinity. The following two sections are devoted to developing these notions. 1.4.2. First ingredient: local regularity measures Before giving a detailed description of the local regularity variations of a signal, it is necessary to determine what method will be used to measure this regularity. It has been explained in section 1.3 that many characterizations are equally relevant and that the choice of one or the other is dictated by practical considerations and the type of applications chosen. Likewise, we may base multifractal analysis on various measures of local regularity. However, there are a certain number of advantages in using the pointwise exponent, which is reasonably simple while leading to a rich enough analysis. Therefore, the following text will use this measure of regularity, which corresponds to the most common choice. Grain exponents For reasons which will be explained below, we need to define a new class of exponents called grain exponents. These are simply approximations, at each finite resolution, of the usual exponents. For simplicity, let us assume that our signal X is defined on [0, 1], and let u denote an interval in [0, 1]. We first choose a descriptor VX (u) of the relevant information of X in u: if X is a measure μ, we will most often take Vμ (u) = μ(u). If X is a function f , then Vf (u) may for instance be
50
Scaling, Fractals and Wavelets
the absolute value of the increment of f in u, that is |f (umax ) − f (umin )| (where u = [umin , umax ]). A more precise descriptor is obtained by considering the oscillation instead, and setting Vf (u) = β(f, u). Finally, in cases where the intervals u are dyadic, of the form Ink = [(k − 1) 2−n , k 2−n ], a third possibility is to choose Vf (Ink ) = |ckn |, where ckn is the wavelet coefficient of f on scale n and in position k (note that, in this case, most quantities defined below will additionally depend on the wavelet). Once VX (u) has been defined, the grain exponent may be calculated as follows: α(u) =
log VX (u) log|u|
Filtration It is then necessary to define a sequence of partitions (Ink )(n0,k=1...νn ) of [0, 1]. For each fixed n, the collection of the νn intervals (Ink )k constitutes a partition of [0, 1] and we require that, when n tends to infinity, maxk |Ink | tends to 0 (this implies, of course, that νn has to tend to infinity). A common (but not neutral) choice is to consider the dyadic intervals (and thus νn = 2n ). The grain exponent αnk is then defined by: log VX (Ink ) k αn = log(|Ink |)
Ink :
Intuitively, αnk does indeed measure the “singularity” of X in the (small) interval the smaller αnk , the greater the variation of X in Ink , and vice versa.
Abstract function A It is important to observe at this point that we can carry out a more general version of multifractal analysis by replacing the grain exponent α(u) with a function A defined on the metric space of closed intervals of [0, 1], and ranging in R+ ∪{+∞} [LEV 04b]. In this context, more general results can be obtained. 1.4.3. Second ingredient: the size of point sets of the same regularity In the same way as the local regularity can be defined through different approaches, there exist many ways of extracting high-level information. Geometric method Conceptually, the simplest method consists of considering the sets Eα of those points of [0, 1] possessing a given exponent α and then describing the size of Eα (to simplify notations and as the focus of this section is on the pointwise exponent,
Fractal and Multifractal Analysis in Signal Processing
51
we write α instead of αp ). In many cases of practical and theoretical interest, these sets have a zero Lebesgue measure for most of the values of α. In addition, they are often dense in [0, 1], with a box dimension equal to 1. To distinguish them, it is thus necessary to measure them either by their Hausdorff or packing dimension. Here, only the first of these dimensions will be considered. We set: fh (α) = dimH (Eα ) where Eα = {x : α(x) = α}. The function fh is called the Hausdorff multifractal spectrum of X. Since the empty set dimension is −∞, we see that fh will take values in {−∞} ∪ [0, n] for an n-dimensional signal. Even though a strict definition of a multifractal object does not exist, it seems reasonable to talk of multifractality when fh (α) is positive or zero for several values of α: indeed, this means that X will display different singular behaviors on different subsets of [0, 1]. Often, we will require that fh be strictly positive on an interval in order to consider X as truly multifractal. While the Hausdorff spectrum is simple to define, it requires an extremely delicate calculation in theoretical as well as numerical studies. The next subsection presents another multifractal spectrum, which is easier to calculate and which also serves to give an approximation by excess of fh . Statistical method The second method used to globally describe the variations of regularity is to adopt a statistical approach (as opposed to fh , which is a geometric spectrum): we first choose a value of α, and counts the number of intervals Ink , at a given resolution n, where X possesses a grain exponent approximately equal to α. We then let the resolution tend to infinity and observes how this number evolves. More precisely, the large deviation spectrum fg is defined as: fg (α) = lim lim inf ε→0 n→∞
log Nnε (α) log νn
where: Nnε (α) = #{k : α − ε αnk α + ε} We may heuristically understand this spectrum by letting Pn denote the uniform probability law on {1 . . . νn }, i.e. Pn (k) = 1/νn for k = 1, . . . , νn . Then, neglecting ε and assuming that the lower limit is a limit: Pn (αnk α) νnfg (α)−1
(1.27)
52
Scaling, Fractals and Wavelets
In other words, if an interval Ink at resolution n is drawn randomly (for a sufficiently large value of n), then the probability of observing a singularity exponent f (α)−1 . approximately equal to α is proportional to νng From the definition, it is clear that fg , as fh , ranges in {−∞} ∪ [0, 1] (in one dimension). As a consequence, whenever fg (α) is not equal to one (and thus is strictly smaller than 1), the probability of observing a grain exponent close to α decays exponentially fast to 0 when the resolution tends to infinity: only those α such that fg (α) = 1 will occur in the limit of infinite resolution. The study of this type of behavior and the determination of the associated exponential convergence speed (here, fg (α) − 1) are the topic of the branch of probability called “large deviation theory”, from which we derive the denomination of fg . Nature of the variable ε Let us consider the role of ε: at each finite resolution, the number of intervals Ink is finite. As a consequence, Nn0 (α) will be zero except for a finite number of values of α. The variable ε then represents a “tolerance” on the exponent value, which is made necessary by the fact that we work at finite resolutions. Once the limit on n has been taken, this tolerance is no longer needed and we can let ε tend to 0 (note that the limit in ε always exists, as we are dealing with a monotonic function of ε). From a more general point of view, we may understand the difference between fh and fg as an inversion of the limits: in the case of fh , we first let the resolution tend to infinity to calculate the exponents and then “count” how many points are characterized by a given exponent. In the case of fg , we start by counting, at a given resolution, the number of intervals having a prescribed exponent, and then let the resolution tend to infinity. This second procedure renders the calculation of the spectrum easier, but in general it will obviously lead to a different result. In section 1.4.7, two examples are given that illustrate the difference between fh and fg . NOTE 1.3.– Historically, the large deviation spectrum was introduced as an easy means to calculate fh through arguments developed in section 1.4.4. It was then gradually realized that fg contains information sometimes more relevant than fh , particularly in signal processing applications. For more details on this topic, see [LEV 98b], where the denominations “Hausdorff spectrum”, “large deviation spectrum” and “Legendre spectrum” were introduced. The next section tackles the problem of calculating the multifractal spectra. 1.4.4. Practical calculation of spectra Let us begin with the Hausdorff spectrum. It is clear that the calculation of the exponents in each point, and then of all the associated dimensions of the Hausdorff spectrum, is an extremely difficult task. It is thus desirable to look for indirect methods to evaluate fh . Two of them are described below.
Fractal and Multifractal Analysis in Signal Processing
53
Multifractal formalism In some cases, fh may be obtained as the Legendre transform of a function that is easily calculated. When this is the case, we say that the (strong) multifractal formalism holds. Define: n
Sn (q) =
2
VX (Ink )q
k=1
and set: τn (q) = −
1 log2 Sn (q) n
τ (q) = lim inf τn (q) n→∞
To understand the link between fh and τ heuristically, let us evaluate τn by grouping together the terms that have the same order of magnitude in the definition of Sn . In that view, fix a “Hölder exponent” α and consider those intervals for which, when n is sufficiently large, VX (Ink ) ∼ |Ink |α . Assume that this approximation is true uniformly in k. We may then roughly estimate the number of such intervals by 2nfh (α) . Then: Sn (q) = {2−nqα : VX (Ink ) ∼ |Ink |α } = 2−n(qα−fh (α)) α
k
α
Factorizing 2−n inf α (qα−fh (α)) , we obtain for τn : 1 2−n[qα−fh (α)−inf α (qα−fh (α))] τn (q) = inf qα − fh (α) − log α n α and, taking the “limit” when n tends to infinity: τ (q) := lim inf α (qα − fh (α)) =: fh∗ (α). The transformation fh → fh∗ is called the “Legendre transform” (for the concave functions). Provided we can justify all the manipulations above, we thus expect that τ will be the Legendre transform of fh . If, in addition, fh is concave, a well-known property of the Legendre transform guarantees that fh = τ ∗ . As announced above, fh is then obtained at once from τ . The important point is that τ is itself much easier to evaluate than fh : there is no need to calculate any Hausdorff dimension and furthermore the definition of τ involves only average quantities, and no evaluations of pointwise Hölder exponents. The only difficulty lies in the estimation of the limit. Therefore, the estimation of τ will be in general much easier and more robust than that of fh . For this reason, even though the multifractal formalism does not hold in
54
Scaling, Fractals and Wavelets
general, it is interesting to consider a new spectrum, called the Legendre spectrum and defined as1 fl = τ ∗ . The study of the validity domain of the equality fh = fl is probably one of the most examined issues in multifractal analysis, for obvious theoretical and practical reasons. However, this is an extremely complex problem, which has only been partially answered to this day, even in “elementary” cases such as the one of multiplicative processes (see section 1.4.6). It is easy to find counter-examples to the equality fh = τ ∗ . For instance, since a Legendre transform is always concave, the formalism certainly fails as soon as fh is not concave. There is no reason to expect that the spectrum of real signals will have this property. In particular, it is not preserved by addition, so that it will not be very stable in practice. Other dimensional spectra Instead of using the Legendre transform approach, another path consists of defining spectra which are close in spirit to fh , but are easier to calculate. Since the major difficulty arises from the estimation of the Hausdorff dimension, we may try to replace it with the box dimension, which is simpler to evaluate. Unfortunately, merely replacing the Haussdorff dimension by Δ in the definition of the spectrum leads to uninteresting results. Indeed, as mentioned above, the sets Eα have, in many cases of interest, a box dimension equal to 1 (see section 1.4.6 for examples). In this situation, the spectrum obtained by replacing dim by Δ, being constant, will not supply any information. A more promising approach consists of defining dimension spectra as in [LEV 04b, TRI 99]. To do so, let us consider a general function of intervals A, as in section 1.4.2. For any x in [0, 1], set: αn (x) = A un (x) where un (x) is the dyadic interval of length 2−n containing x (take for instance the right interval if two such intervals exist). Then let: Eα (, N ) = {x/n N ⇒ |αn (x) − α| } Eα (, N ) is an increasing function of N , and thus: ∪N Eα (, N ) supN Eα (, N ). Let:
=
Eα () = sup Eα (, N ) = {x/∃N such that n N ⇒ |αn (x) − α| } N
1. The inequalities fh g fl are always true (see below). Thus, it would be more accurate to say that the Legendre spectrum is an approximation (by excess) of the large deviation spectrum, rather than that of the Hausdorff spectrum.
Fractal and Multifractal Analysis in Signal Processing
55
Since the sets Eα () decrease with , we may define: Eα = lim Eα () = {x/αn (x)→n→∞ α} →0
DEFINITION 1.5.– Let d be any dimension. We define the following spectra: fd (α) = d(Eα ) = d lim sup Eα (, N ) →0 N
fdlim (α) = lim d Eα () = lim d sup Eα (, N ) →0
→0
fdlim sup (α) = lim sup d Eα (, N )
(1.28) (1.29)
N
→0 N
(1.30)
When d is the Hausdorff dimension, then fd is just the Hausdorff spectrum fh . Let D = Im (A) be the closure of the image of A. Then D is the support of the spectra. Indeed, for any α ∈ D, Eα (, N ) = ∅, and thus fd (α) = fdlim (α) = fdlim sup (α) = −∞ (obviously this also applies to fg ). The following inequalities are easily proved: fd fdlim
(1.31)
fdlim sup fdlim
(1.32)
There is no relation in general between fd and fdlim sup . However, if d is σ-stable: fd (α) fdlim (α) = fdlim sup (α)
(1.33)
Besides, if d1 and d2 have two dimensions such that, for any E ⊂ [0, 1]: d1 (E) d2 (E) then: sup sup (E) fdlim (E), fdlim (E) fdlim (E) fd1 (E) fd2 (E), fdlim 1 2 1 2
It is not hard to prove the below sequence of inequalities. PROPOSITION 1.4.– For any A: fh (α) fhlim sup (α) = fhlim (α) lim lim sup fΔ (α) min fΔ (α), fg (α)
(1.34)
56
Scaling, Fractals and Wavelets
This is an improvement on the traditional result fh (α) fg (α). In particular, lim . If, as soon as fh (α) = fg (α), all the above spectra coincide, except perhaps fΔ lim sup on the contrary, fh is smaller than fg , then we may hope that fΔ (α) will be a better approximation of fh than fg . The important point here is that the calculation of lim sup fΔ (α) only involves box dimensions and that it is of the same order of complexity lim sup as that of fg . The spectrum fΔ (α) is thus a good substitute when we want to lim sup numerically estimate fh . For example, the practical calculation of fΔ (α) on a multinomial measure (see section 1.4.6) yields good results. Since d([0, 1]) = 1, all the spectra have a maximum lower than or equal to 1. In certain cases, a more precise result concerning fdlim sup , fdlim and fg is available: PROPOSITION 1.5.– Let K be the set of x in [0, 1] such that the sequence (αn (x)) converges. Let us suppose that |K| > 0. Then, there exists α0 in D such that: fdlim sup (α0 ) = fdlim (α0 ) = fg (α0 ) = 1
(1.35)
Let us note that such a constraint is typically not satisfied by fh . This shows that fh contains, in general, more information. A more precise and more general result may be found in [LEV 04b]. The three spectra fdlim sup , fdlim and fg also obey a structural constraint, expressed by the following proposition. PROPOSITION 1.6.– The functions fg , fdlim and fdlim sup are upper semi-continuous. NOTE 1.4.– Recall that a function f : D ⊂ R → R is upper semi-continuous (USC) if, for any x ∈ D and for any sequence (xn ) of D converging to x: lim sup f (xn ) f (x)
(1.36)
n→∞
Let us define the upper semi-continuous envelope of a function f by: f˜(α) = lim sup{f (β)/|β − α| } →0
Then, the above results imply that: f˜d (α) fdlim (α)
(1.37)
Let us also mention the following result. PROPOSITION 1.7.– Let A and B be the two interval functions and C = max{A, B}. Let fg (α, A), fg (α, B) and fg (α, C) denote the corresponding spectra. Then, for any α: fg (α, C) max{fg (α, A), fg (α, B)}
(1.38)
Fractal and Multifractal Analysis in Signal Processing
57
PROPOSITION 1.8.– If d is stable, then inequality (1.38) is true for fd , fdlim and fdlim sup . A significant result concerning the “inverse problem” for the spectra serves as a conclusion for this section. PROPOSITION 1.9.– Let f be a USC function ranging in [0, 1] ∪ {−∞}. Then, there exists an interval function A whose fdlim sup or fdlim spectrum is exactly f as soon as d is σ-stable or d = Δ. Let us note that fd , when d is σ-stable, is not necessarily USC. This shows once more that this spectrum is richer than the other ones (see [LEV 98b] for a study of the structural properties of fd with d = h). Weak multifractal formalism Let us now consider the numerical estimation of fg . As in the case of fh , two approaches exist: either we resort to a multifractal formalism with fg as the Legendre transform of a simple function, or we analyze in detail the definition of fg and deduces estimation methods from this. In the first case, the heuristic justification is the same as for fh and we expect that, under certain conditions, fg = τ ∗ . Since we avoid an inversion of limits as compared to the case of fh , this formalism (sometimes called weak multifractal formalism, as opposed to the strong formalism which ensures the equality of τ ∗ and fh ) will be satisfied more often. However, the necessary condition that fg be concave, associated once again with the lack of stability, still limits its applicability. An important difference between the strong and weak formalisms is that, in the latter case, a precise and reasonably general criterion ensuring its validity is known. We are referring to a version of the Ellis theorem, one of the fundamental results in the theory of large deviations, which is recalled below in a simplified form. THEOREM 1.9.– If τ (q) = limn→∞ τn (q) exists as a limit (rather than a lower limit), and if it is differentiable, then fg = fl . When fg is not concave, it cannot equal fl , but the following result holds. THEOREM 1.10.– If fg is equal to −∞ outside a compact set, then fl is its concave hull, i.e.: fl = (fg )∗∗ This theorem makes it possible to measure precisely the information which is lost when fg is replaced by fl . See [LEV 98b] for related results.
58
Scaling, Fractals and Wavelets
Continuous spectra It is possible to prove that the previous relation is still valid in the more subtle case of continuous spectra [LEV 04b, TRI 99]. These continuous spectra constitute a generalization of fg that allow us to avoid choosing a partition. As already mentioned, this choice is not neutral and different partitions will, in general, lead to different spectra. To begin with, interval families are defined: Rη = {u interval of [0, 1] such that |u| = η} Rη (α) = {u : |u| = η, A(u) = α} Rεη (α) = {u : |u| = η, |A(u) − α| ε} DEFINITION 1.6 (continuous large deviation spectra).– log η1 |∪Rεη (α)| fgc (α) = lim lim sup ε→0 η→0 log η log η1 |∪Rη (α)| f˜gc (α) = lim sup log η η→0 Note that fgc (α) is defined similarly to fg , except that all the intervals of a given length η are considered where the variation of X is of the order of η α±ε , rather than only dyadic intervals. Since the number of these intervals may be infinite, we replace Nnε (α) with a measure of the average length, i.e. η1 |∪Rεη (α)|. Within this continuous framework, Rη (α) will, in general, be non-empty for infinitely many values of α and not only for at most 2n values, as is the case for Nn0 (α): it is thus possible to get rid of ε and define the new spectrum f˜gc . Legendre transform of continuous spectra For any family of R intervals, ∪R denotes the union of all the intervals of R. A packing of R is a sub-family composed of disjoint intervals. DEFINITION 1.7 (Legendre continuous spectrum).– For any real q, let:
q qA(u) |u| : R is a packing of R H (R) = sup u∈R
and: Jηq = sup H q Rη (α) α
Fractal and Multifractal Analysis in Signal Processing
Set:
59
log H q (Rη ) η→0 log η q log J (Rη ) τ˜c (q) = lim inf η→0 log η τ c (q) = lim inf
and finally: flc = (τ c )∗ and f˜lc = (˜ τ c )∗ . Here are the main properties of fgc , f˜gc , flc and f˜lc . PROPOSITION 1.10.– – flc and f˜lc are concave functions; – for any α, f˜c (α) f c (α) and fg (α) f c (α); g
g
g
– if μ is a multinomial measure (see section 1.4.6), then fgc (α) = f˜gc (α) = flc (α) = f˜lc (α) = fg (α). THEOREM 1.11.– If fgc (respectively f˜gc ) is equal to −∞ outside a compact interval, then, for any α: ∗∗ flc (α) = fgc (α) ∗∗ f˜lc (α) = f˜gc (α) PROPOSITION 1.11.– – τ c and τ˜c are increasing and concave functions; – τ c (0) = τ˜c (0) = −Δ(supp(X)); – if X is a probability measure, then τ c (1) = τ˜c (1) = 0; – τ c (q) = lim inf n→∞ log H q (Rηn )/ log ηn , where (ηn ) is a sequence tending to zero such that log ηn / log ηn+1 → 1 when n → ∞. The same is true for τ˜c . The last property is important in numerical applications: it means that τ c and τ˜c may be estimated by using discrete sequences of the type ηn = 2−n . Kernel method A second method to estimate fg , which does not assume that the weak formalism is true (and thus in particular allows us to obtain non-concave spectra), is based on the following. Let K denote the “rectangular kernel”, i.e. K(x) = 1 for x ∈ [−1, 1] and K(x) = 0 elsewhere. Let Kε (t) = 1ε K( εt ). Then, by definition, Nnε (α) = 2n+1 εKε ∗ ρn (α), where the symbol ∗ represents convolution. It is not hard to check that replacing K by any positive kernel with compact support whose
60
Scaling, Fractals and Wavelets
integral equals 1 in the definition of Nnε (α) will not change the value of fg . A basic idea is then to use a more regular kernel than the rectangular one to improve the estimation. A more elaborate approach is to use ideas from density estimation to try and remove the double limit in the definition of fg : this is performed by choosing ε to be a function of n in such a way that appropriate convergence properties are obtained [LEV 96b]. We may, for instance, show the following result: PROPOSITION 1.12.– Assume that the studied signal is a finite sum of multinomial measures (see section 1.4.6). Let: fnε (α) =
log Nnε (α) log νn
εn Then, if εn is a sequence such that limn→∞ log(ν = c, where c is a positive n) constant: lim sup fnεn (α) − fg (α) = 0 n→∞ α
Even without making such a strong assumption on the signal structure, it is still possible, in certain cases, to obtain convergence results, as above, with ε = ε(n), using more sophisticated statistical techniques. To conclude our presentation of multifractal spectra, let us emphasize that no spectrum is “better” than the others in all respects. All the spectra give similar but different information. As was already observed for the local regularity measures, each one has advantages and drawbacks, and the choice has to be made in view of the considered application. If we are interested in the geometric aspects, a dimension spectrum should be favored. The large deviation spectrum will be used in statistical and most signal processing applications. When the number of data is small or if it appears that the estimations are not reliable, we will resort to the Legendre spectrum. To be able to compare the different information and also to assert the quality of the estimations, it is important to dispose of general theoretical relations between the spectra. It is remarkable that, as seen above, such relations exist under rather weak hypotheses. 1.4.5. Refinements: analysis of the sequence of capacities, mutual analysis and multisingularity In this section, some refinements of multifractal analysis, useful in applications, will be briefly discussed. The first refinement stems from the following consideration. Assume we are interested in the analysis of road traffic and that the signal X(k) at hand is the flow
Fractal and Multifractal Analysis in Signal Processing
61
per hour, i.e. the number of vehicles crossing a given small section of road during a fixed time interval (often in the order a minute).2 In this case, each point of the signal is not a pointwise value but corresponds to a space-time integral. We would thus be inclined to model such data by a measure and carry out multifractal analysis of this measure. However, it appears that, if we want to anticipate congestions, the relevant quantity to consider for a given time interval Tn whose length n is large as compared to one minute is not the sum of the individual flows X(k) for k in Tn , but the maximum of these flows. Hence, instead of one signal, we would rather consider a sequence of signals, Xn , each yielding the maximum flow at the time scale n: Xn (j) = maxk∈Tn X(k). At each scale n, the signal Xn is a set function (i.e. it does not give pointwise values but averages). However, the Xn are not measures, as they are not additive: the maximum on two disjoint intervals T 1 and T 2 is not, in general, the sum of the maxima of T 1 and T 2. Nonetheless, each Xn possesses some regularity properties which allows us to model it as a Choquet capacity. Thus, we are led to generalize multifractal analysis so as to no longer process one signal, which would be a pointwise function or a measure, but a sequence of signals which are Choquet capacities. A number of other examples exist, particularly in image analysis, where such a generalization is found necessary (see [LEV 96a]). We will not give here the definition of a Choquet capacity, but we stress that the generalization of multifractal analysis to sequences of Choquet capacities, at first seemingly abstract, is in fact very simple. Indeed, nothing in the definition of multifractal analysis implies that the same signal has to be considered at each resolution, nor that the signal must be a function or a measure. In particular, the relations between the spectra are preserved in this generalization [LEV 98b]. A different generalization consists of noting that, in the usual formulation of multifractal analysis, the Lebesgue measure L plays a particular role which may not be desirable. Let us for instance consider the definition of the grain exponent. The logarithm of the interval measure Ink is compared to |Ink |, which is nothing but the Lebesgue measure of Ink . In the same way, when we define fh , the Hausdorff dimension is calculated with respect to L. However, it is a traditional fact that we may 3 define an s-dimensional Hausdorff with measure respects to an arbitrary non-atomic s measure μ by replacing the sum |Uj | with μ(Uj ) . As a matter of fact, once a non-atomic reference measure μ has been chosen, we may rewrite all the definitions of multifractal analysis (Hölder exponent, grain exponent, spectra, etc.) by replacing L with μ. If this analysis is applied to a signal X (which may be a function, a measure or a capacity), we obtain the multifractal analysis of X with respect to μ. This type of analysis is called mutual multifractal analysis. There are several benefits to this
2. A multifractal analysis of traffic in Paris is presented in [VOJ 94]. 3. A measure is non-atomic if it attributes zero mass to all singletons. The reasons for which one has to restrict to such measures go beyond the scope of this chapter; see [LEV 98b] for more details.
62
Scaling, Fractals and Wavelets
generalization. Let us briefly illustrate some of them through two examples. Assume that the signal to be studied X is equal to Y + Z, where Y and Z are two measures Y Z (α) fh,L (α) for supported respectively on [0, 12 ] and on [ 12 , 1]. Suppose that fh,L T all α, where fh,L is the Haussdorff spectrum of T with respect to L. Then, it is easy X Z to see that fh,L (α) = fh,L (α) for any α. The Y component is not detected by the analysis. On the contrary, if we calculate the spectra with respect to the reference X Z Z (α) = fh,Z (α), since fh,Z (α) will be equal to measure Z, then, in general, fh,Z 1 for α = 1 and to −∞ everywhere else. A change in the reference measure thus allows us to carry out a more accurate analysis, by shedding light on possibly hidden components. As a second example, consider an application in image analysis. It is possible to use mutual multifractal analysis to selectively detect certain types of changes in image sequences (for example, the appearance of manufactured objects in sequences of aerial images [CAN 96]). The idea consists of choosing the first image of the sequence as the reference measure and then analyzing each following image with respect to it. In the absence of any change, the mutual spectrum will be equal to 1 for α = 1 and to −∞ everywhere else. A change will lead to a spreading of this spectrum and it is possible to classify changes according to the corresponding values of the couple (α, f (α)) (see Chapter 11). Finally, a third refinement consists of carrying out a multifractal analysis “at a point”, using 2-microlocal analysis: the so-called 2-microlocal formalism allows us to define a function, corresponding to fg , which completely describes the singular behavior of a signal in the vicinity of any point. Such an analysis provides in particular ways to modify pointwise regularity in a powerful manner. 1.4.6. The multifractal spectra of certain simple signals The paradigmatic example of multifractal measures is the Besicovitch measure – often called, in this context, the “binomial measure”. For x in [0, 1], write: x=
∞
xi 2−i
where xi = 0, 1
1
1 1 − xi , n i=1 n
Φn0 (x) =
1 xi n i=1 n
Φn1 (x) =
According to (1.13): α un (x) = −Φn0 (x) log2 p − Φn1 (x) log2 (1 − p) Write lim inf n→∞ Φn0 (x) =: Φ0 (x) and Φ1 (x) := 1 − Φ0 (x). We obtain: α∗ (x) = −Φ0 (x) log2 p − Φ1 (x) log2 (1 − p)
Fractal and Multifractal Analysis in Signal Processing
63
The sets Eα are the sets of points of [0, 1] having a given proportion Φ0 (x) of 0 in their base-2 expansion. Calculations similar to those in Example 1.6 lead to: dim Eα = −Φ0 log2 Φ0 − Φ1 log2 Φ1 The Hausdorff spectrum is thus given in parametric form by: α(Φ0 ) = −Φ0 log2 p − Φ1 log2 (1 − p) fn (α) = −Φ0 log2 Φ0 − Φ1 log2 Φ1 Note also that the sets of points in (0, 1) for which Φn0 (x) does not converge has Hausdorff dimension equal to 1, although it has a zero Lebesgue measure: indeed, the strong law of large numbers entails that, for L-almost x, Φ0 (x) = Φ1 (x) = 12 . An immediate consequence is that, for αm := − 12 log2 p(1 − p), fh (αm ) = 1. Now consider the grain exponents. We observe that L-a.s., αnk → αm . The function fg (α) will measure the speed at which Pr (|αnk − αm | > ε) tends to 0 for ε > 0 when n → ∞. For fixed n there are exactly Cnk intervals Inj such that Φn0 (x) = nk for x ∈ Inj . This makes it possible to evaluate Nnε (α), which is equal to Cnk for ε sufficiently small and α close to α(Φ0 = nk ). Using Stirling’s formula to estimate Cnk , we can then obtain fg (α). However, these calculations are somewhat tedious, in particular due to the double limit, and it is much faster to evaluate fl (α) and use the general results on the relationships between the three spectra: By definition: τ (q) = lim inf n→∞
log2
2n −1 j q k=0 μ In −n
Now, μ(Inj ) = pk (1 − p)n−k exactly Cnk times, for k = 0 . . . n. As a consequence: n 2 −1
j=0
n n q μ Inj = Cnk pkq (1 − p)(n−k)q = pq + (1 − p)q k=0
and: τ (q) = − log2 pq + (1 − p)q A simple Legendre transform calculation then shows that fl (α) = τ ∗ (α) is equal to fh (α). Since fh fl fg is always true, it follows that the strong multifractal formalism holds in the case of the binomial measure: fh = fg = fl .
64
Scaling, Fractals and Wavelets
1
0.8
0.6
0.4
0.2
0
0.8
0.9
1
1.1
1.2
1.3
Figure 1.7. Spectrum of the binomial measure
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.2
0.4
0.6
0.8
1
1.2
1.4
Figure 1.8. Spectrum of the sum of two binomial measures
Let us note, however, that simple operations will break the formalism: for instance, the sum of two Besicovitch measures on [0, 1], or the lumping of two such measures with disjoint supports, will have fh = fg = fl (see Figures 1.7 and 1.8). There are many generalizations of the Besicovitch measure. We can replace base 2 by base b > 2, i.e. partition [0, 1] into b sub-intervals. We then speak of multinomial measures. We may also distribute the measure on b < b intervals, in which case the
Fractal and Multifractal Analysis in Signal Processing
65
support of the measure will be a Cantor set. Another variation is to consider stochastic constructions by choosing the partitioning and/or p as random variables following certain probability laws. An example of such a stochastic Besicovitch measure, where, at each scale, p is chosen as an iid lognormal random variable, is represented in Figure 1.9. This type of measure can serve as a starting point for the modeling of certain Internet traffic.
Figure 1.9. A random binomial measure
Essentially the same analysis as above applies to the self-affine functions defined in Example 1.7. Similar parametric formulae are obtained for the spectra, which also coincide in this case. Let us mention as a final note that such measures and functions have a Hölder function which is everywhere discontinuous and has level lines which are everywhere dense. We have αl (x) = α0 for all x, where α0 = mint αp (t). Let us now consider the Weierstrass function: W (x) =
∞
ω −nH cos(ω n x)
i=0
As was already mentioned, α(t) = H for any t for W . As a consequence: fh (H) = 1,
fh (α) = −∞ if α = H
The value of the large deviation spectrum depends on the choice of function VW (u). Taking VW (u) = β(W, u) leads to fg = fh . However, if VW (u) is defined to
66
Scaling, Fractals and Wavelets
be the increment of W in u, then, for certain values of ω: ⎧ ⎪ if α < H ⎨−∞ fg (α) = H + 1 − α if H α H + 1 ⎪ ⎩ −∞ if α > H + 1 The heuristic explanation of the fact that fg is positive for some values of α larger than H is as follows: at each finite resolution ε, the increments of W in intervals of size ε are at most of the order of εH , because the exponent is defined as a lim inf. However, there will also exist intervals for which the increments are smaller, yielding a larger observed grain exponent. The fg spectrum does in fact measure the speed at which the probability of finding one of these smoother increments tends to 0 when ε tends to 0. To conclude, we note that, in both cases (oscillations and increments), fg is concave and coincides with fl . Similar results hold for fractional Brownian motion. Of course, since fractional Brownian motion is a stochastic process, its spectrum is a priori a random function. However, we can show that, with probability 1, fh and fg (defined either using oscillations or increments) are exactly the same as those given above for the Weierstrass function. 1.4.7. Two applications To conclude this chapter, we briefly mention two applications of multifractal analysis to signal and image processing. Our intention is not to go into the details of these applications (they are developed in Chapters 11 and 12), but to indicate, in a simplified way, how the tools introduced above are put into practice. 1.4.7.1. Image segmentation The issue of edge detection in images allows us to illustrate in a concrete way the relevance of multifractal spectra, as well as the difference between fh and fg . We have observed above that edge points are irregular points, but that they cannot be characterized by a universal value of αp , since non-linear transformations of an image may leave the contours unchanged while modifying the exponents. To characterize edges, it is necessary to include higher level information. This information may be obtained from the following obvious comment: (smooth) edges of an image form a set of lines which is of one dimension. Looking for contours thus means looking for sets of points which are characterized by specific values of αp (local regularity criterion), and such that their associated dimension is 1 (global criterion). In other words, we will characterize edges as those points possessing an exponent which belongs to fh−1 (1). This provides a geometric characterization of edges: on the one hand, we rely on
Fractal and Multifractal Analysis in Signal Processing
67
pointwise exponents, which measure the regularity at infinite resolution; on the other hand, we use fh , which is a dimension spectrum. However, it is also possible to follow a statistical approach: assume we consider a very simple image which contains only a black line, the “edge”, drawn on a white background. If we draw randomly in a uniform way a point in the image at infinite resolution, the probability of hitting the line is zero. However, at any finite resolution, where the image is made of, say, 2n ×2n pixels, the probability of hitting the edge is of the order 2−n , since the edge contains 2n pixels. According to the definition of fg (recall in particular (1.27)), we see that, on the black line, fg (α) = 1. In this approach, we thus characterize an edge point as a point possessing a singularity α whose probability of occurrence decreases as 2−n when the resolution tends to infinity. When the multifractal formalism holds, the geometric and statistical approaches yield the same result. For more details, see Chapter 11 and [LEV 96a]. 1.4.7.2. Analysis of TCP traffic Our second application deals with the modeling and analysis of TCP traffic. Here the situation is a little bit different from that of the previous application: contrarily to typical images, TCP traffic possesses, under certain conditions, a multifractal structure. In the language of the discussion at the end of the introduction to this chapter, we here use fractal methods to study a fractal object. This will entail few changes as far as the analysis of the data is concerned. However, dealing with a multifractal signal brings up the question of the source of this multifractality, and thus of a model capable of explaining this phenomenon. This issue will not be not tackled here. See Chapter 12 and [LEV 97, LEV 01, RIE 97, BLV 01]. What is the use of carrying out a multifractal analysis of TCP? First of all, the range of values taken by the Hölder exponents provides important information on the small-scale behavior of traffic. The smaller α is, the more sporadic traffic will be, which means that variations on short time intervals will be significant. The spectrum also allows us to elucidate what is typical behavior, i.e. the value α0 such that fg (α0 ) = 1: with high probability, the variation of traffic between two close time instants (t1 , t2 ) will be of the order |t2 − t1 |α0 . While this typical behavior is important for the understanding and the management of the network, it is also useful to know which other variations may occur, and with what probabilities. This is exactly the information provided by fg . Thus, the whole large deviation spectrum is useful in this application. Let us note that, in contrast, the Hausdorff spectrum is probably less adapted here: first because the relevant physical quantities are increments at different time scales, small but finite; there is no notion of regularity at infinite resolution, as is the case with images. Second, the relevant information is statistical in nature rather than geometric. To conclude, let us mention that the large deviation spectrum of certain TCP traces, as estimated by the kernel method, displays a shape reminiscent of that of
68
Scaling, Fractals and Wavelets
the sum of two binomial measures. This provides useful information on the fine structure of these traces. In particular, fg is not concave. This shows the advantage of having procedures available to estimate spectra without making the assumption that a multifractal formalism is satisfied.
1.5. Bibliography [BLV 01] BARRAL J., L ÉVY V ÉHEL J., “Multifractal analysis of a class of additive processes with correlated nonstationary increments”, Electronic Journal of Probability, vol. 9, p. 508–543, 2001. [BON 86] B ONY J., “Second microlocalization and propagation of singularities for semilinear hyperbolic equations”, in Hyperbolic equations and related topics (Katata/Kyoto, 1984), Academic Press, Boston, Massachusetts, p. 11–49, 1986. [BOU 28] B OULIGAND G., “Ensembles impropres et nombre dimensionnel”, Bull. Soc. Math., vol. 52, p. 320–334 and 361–376, 1928 (see also Les définitions modernes de la dimension, Hermann, 1936). [BOU 00] B OUSCH T., H EURTEAUX Y., “Caloric measure on the domains bounded by Weierstrass-type graphs”, Ann. Acad. Sci. Fenn. Math., vol. 25, p. 501–522, 2000. [CAN 96] C ANUS C., L ÉVY V ÉHEL J., “Change detection in sequences of images by multifractal analysis”, in ICASSP’96 (Atlanta, Georgia), 1996. [DAO 02] DAOUDI K., L ÉVY V ÉHEL J., “Signal representation and segmentation based on multifractal stationarity”, Signal Processing, vol. 82, no. 12, p. 2015–2024, 2002. [DUB 89] D UBUC B., T RICOT C., ROQUES -C ARMES C., Z UCKER S., “Evaluating the fractal dimension of profiles”, Physical Review A, vol. 39, p. 1500–1512, 1989. [GRE 85] G REBOGI C., M C D ONALD S., OTT E., YORKE J., “Exterior dimension of large fractals”, Physics Letters A, vol. 110, p. 1–4, 1985. [HAR 16] H ARDY G., “Weierstrass non-differential function”, American Mathematical Society Translations, vol. 17, p. 301–325, 1916. [HAU 19] H AUSDORFF F., “Dimension und äusseres Mass”, Math. Ann., vol. 79, p. 157–179, 1919. [JAF 96] JAFFARD S., M EYER Y., “Wavelet methods for pointwise regularity and local oscillations of functions”, Mem. Amer. Math. Soc., vol. 123, 1996. [KAH 63] K AHANE J., S ALEM R., Ensembles parfaits et séries trigonométriques, Hermann, 1963. [KOL 61] KOLMOGOROV A., T IHOMIROV V., “Epsilon-entropy and epsilon-capacity of sets in functional spaces”, American Mathematical Society Translations, vol. 17, p. 277-364, 1961. [KOL 01] KOLWANKAR K., L ÉVY V ÉHEL J., “Measuring functions smoothness with local fractional derivatives”, Frac. Calc. Appl. Anal., vol. 4, no. 3, p. 285–301, 2001.
Fractal and Multifractal Analysis in Signal Processing
69
[KOL 02] KOLWANKAR K., L ÉVY V ÉHEL J., “A time domain characterization of the fine local regularity of functions”, J. Fourier Anal. Appl., vol. 8, no. 4, p. 319–334, 2002. [LEV 96a] L ÉVY V ÉHEL J., “Introduction to the multifractal analysis of images”, in F ISHER Y. (Ed.), Fractal Image Encoding and Analysis, Springer-Verlag, 1996. [LEV 96b] L ÉVY V ÉHEL J., “Numerical computation of the large deviation multifractal spectrum”, in CFIC (Rome, Italy), 1996. [LEV 97] L ÉVY V ÉHEL J., R IEDI R., “Fractional Brownian motion and data traffic modeling: The other end of the spectrum”, in L ÉVY V ÉHEL J., L UTTON E., T RICOT C. (Eds.), Fractals in Engineering, Springer-Verlag, 1997. [LEV 98a] L ÉVY V ÉHEL J., G UIHENEUF B., “2-Microlocal analysis and applications in signal processing”, in International Wavelets Conference (Tangier), 1998. [LEV 98b] L ÉVY V ÉHEL J., VOJAK R., “Multifractal analysis of Choquet capacities”, Advances in Applied Mathematics, vol. 20, no. 1, p. 1–43, 1998. [LEV 01] L ÉVY V ÉHEL J., S IKDAR B., “A multiplicative multifractal model for TCP traffic”, in ISCC’2001 (Tunisia), 2001. [LEV 02] L ÉVY V ÉHEL J., “Multifractal processing of signals”, forthcoming. [LEV 04a] L ÉVY V ÉHEL J., S EURET S., “The 2-microlocal formalism”, in Fractal Geometry and Applications: A Jubliee of Benoit Mandelbrot, Proc. Sympos. Pure Math., PSPUM, vol. 72, Part 2, p. 153–215, 2004. [LEV 04b] L ÉVY V ÉHEL J., “On various multifractal spectra”, in BANDT C., M OSCO U., Z ÄHLE M. (Eds.), Fractal Geometry and Stochastics III, Progress in Probability, Birtkhäuser Verlag, vol. 57, p. 23–42, 2004. [MCM 84] M C M ULLEN C., “The Hausdorff dimension of general Sierpinski carpets”, Nagoya Mathematical Journal, vol. 96, p. 1–9, 1984. [RIE 97] R IEDI R., L ÉVY V ÉHEL J., TCP traffic is multifractal: a numerical study, Technical Report RR-3129, INRIA, 1997. [ROU 98] ROUEFF F., L ÉVY V ÉHEL J., “A regularization approach to fractional dimension estimation”, in Fractals’98 (Malta), 1998. [SEU 02] S EURET S., L ÉVY V ÉHEL J., “The local Hölder function of a continuous function”, Appl. Comput. Hamron. Anal., vol. 13, no. 3, p. 263–276, 2002. [TRI 82] T RICOT C., “Two definitions of fractal dimension”, Math. Proc. Camb. Phil. Soc., vol. 91, p. 57–74, 1982. [TRI 86a] T RICOT C., “Dimensions de graphes”, Comptes rendus de l’Académie des sciences de Paris, vol. 303, p. 609–612, 1986. [TRI 86b] T RICOT C., “The geometry of the complement of a fractal set”, Physics Letters A, vol. 114, p. 430–434, 1986. [TRI 87] T RICOT C., “Dimensions aux bords d’un ouvert”, Ann. Sc. Math. Québec, vol. 11, no. 1, p. 205–235, 1987. [TRI 88] T RICOT C., Q UINIOU J., W EHBI D., ROQUES -C ARMES C., D UBUC B., “Evaluation de la dimension fractale d’un graphe”, Rev. Phys. Appl., vol. 23, p. 111–124, 1988.
70
Scaling, Fractals and Wavelets
[TRI 99] T RICOT C., Courbes et dimension fractale, Springer-Verlag, 2nd edition, 1999. [VOJ 94] VOJAK R., L ÉVY V ÉHEL J., DANECH -PAJOUH M., “Multifractal description of road traffic structure”, in Seventh IFAC/IFORS Symposium on Transportation Systems: Theory and Application of Advanced Technology (Tianjin, China), p. 942–947, 1994.
Chapter 2
Scale Invariance and Wavelets
2.1. Introduction
Processes presenting “power law” spectra (often regrouped under the restrictive, but generic, term of “1/f ” processes) appear in various domains: hydrology [BER 94], finance [MAN 97], telecommunications [LEL 94, PAR 00], turbulence [CAS 96, FRI 95], biology [TEI 00] and many more [WOR 96]. The characteristics of these processes are based upon concepts such as fractality, self-similarity or long-range dependence and, even though these different notions are not equivalent, they all possess a common characteristic: that of replacing the idea of a structure related to a preferred time scale with that of an invariant relationship between different scales.
The study of scale invariant processes presents several difficulties ranging from modeling to analysis and processing, for which few tools were available until recently [BER 94]. The effective possibility of appropriately manipulating these processes has recently been reinforced by the appearance of adequate multiresolution techniques: the tools which are referred to here have been developed for this purpose. These tools are explicitly based on the theoretical as well as algorithmic potentialities offered by wavelet transforms.
Chapter written by Patrick F LANDRIN, Paulo G ONÇALVES and Patrice A BRY.
72
Scaling, Fractals and Wavelets
2.2. Models for scale invariance 2.2.1. Intuition From a qualitative point of view, the idea beyond a 1/f spectrum involves situations where, given a signal observed in the time domain, its empirical power spectrum density behaves as S(f ) = C|f |−α with α > 0. From a practical viewpoint, it is evidently the equivalent form: log S(f ) = log C − α log |f | which is the most significant, since the “1/f ” character is translated by a straight line in doubly logarithmic coordinates. Generally, when dealing with physical observations, referring to some 1/f spectral behavior is only meaningful with respect to a frequency analysis band. Therefore, the introduction, on the half-line of positive frequencies, of two (adjustable) frequencies fbf and fhf such that 0 < fbf < fhf < +∞, will end up with a three regime classification, whether we study the 1/f behavior in (at least) one of the three domains that this partition defines. Let us consider each of these cases: 1) fbf f fhf : this is the context of a bandpass domain where we simply observe an algebraic decrease of the spectrum density, without a predominant frequency; 2) fhf f +∞: the 1/f character is dominant in the high frequency limit and highlights the local regularity of the sample paths, their variability and their fractal nature; 3) 0 f fbf : the power law of the spectrum intervenes here in the limit of low frequencies, resulting in a divergence at the origin of the spectrum density. If 0 < α < 1, this divergence corresponds to an algebraic decrease of the correlation function, which is slow enough for the latter not to be summable; there is long-range dependence or long memory. In fact, these three regimes represent three different properties, hence, they have no reason to exist at the same time. However, they possess the common denominator of being linked to an idea of scale invariance according to which – within a scale range and up to some renormalization – the properties of the whole are the same as those of the parts (self-similarity). Indeed, a power law spectrum belongs to the class of homogenous functions. Its form therefore remains invariant under scaling in the sense that, for any c ∈ + : S(f ) = C|f |−α
=⇒
S(cf ) = C|cf |−α = c−α S(f )
Scale Invariance and Wavelets
73
Given that, through Fourier transformation, a dilation or compression in the frequency domain is translated by a corresponding compression or dilation in the time domain, it is thus legitimate to expect “1/f ” processes to be closely coupled with self-similar processes. 2.2.2. Self-similarity To be more precise [BER 94, SAM 94] (see also Chapter 5), we introduce the following definition. DEFINITION 2.1.– A process X = {X(t), t ∈ } is said to be self-similar of index H > 0 if, for any c ∈ , it satisfies the following equality in a distributional sense: L {cH X(t/c), t ∈ }, {X(t), t ∈ } =
∀c > 0.
(2.1)
According to this definition, a self-similar process does not possess any characteristic scale, insofar as it remains (statistically) identical to itself after any scale change. If, from a theoretical point of view, self-similarity is likely to extend from the largest to the finest scales, the above-mentioned definition must, in general, go with a scale domain (i.e., with a variation domain of the factor c) for which the invariance has a meaning. For instance, the finite duration of an observation settles the maximum attainable scale, in the same way as the finite resolution of a sensor limits the finest scale. It is noteworthy that, if a (second-order) process is self-similar, it is necessarily non-stationary. Indeed, assuming that, at some arbitrary time instant t1 := 1, the condition var X(t1 ) = 0 holds, it stems from Definition 2.1 that var X(t) = varX(t × 1) = t2H var X(t1 ) and, as a consequence, the variance of the process depends on time. This behavior applies to all finite moments of X: E |X(t)|q = E |X(1)|q |t|qH .
(2.2)
Therefore, in a strict sense, an ordinary spectrum cannot be attached to a self-similar process. Nevertheless, there exists an interesting sub-class of the class of self-similar processes, which, in a sense, could be paralleled with that of stationary processes: it is that of processes with stationary increments defined as follows [BER 94, SAM 94]. DEFINITION 2.2.– A process X = {X(t), t ∈ } is said to have stationary increments if and only if, for any θ ∈ , the law of its increment process:
X (θ) := X (θ) (t) := X(t + θ) − X(t), t ∈ does not depend on t.
74
Scaling, Fractals and Wavelets
Figure 2.1. Path of a self-similar process. When we simultaneously apply to the sample path of a self-similar process a dilation of the time axis by a factor c and a dilation of the amplitude axis by a factor c−H , we obtain a new sample path that is (statistically) indistinguishable from its original
In this definition, the parameter θ plays the role of a time scale according to which we study the process X. Indeed, the self-similarity of the latter is translated on its increments by:
L H (θ/c) c X (t/c), t ∈ , X (θ) (t), t ∈ =
∀c > 0.
(2.3)
It is evidently possible to extend this definition to higher order increments (“increments of increments”). Coupling self-similarity of index H and the stationary increments property implies that the parameter H remains in the range 0 < H < 1. Moreover, the covariance function of a process (originally centered and zero at the origin) must be of the form: E X(t)X(s) =
σ 2 2H |t| + |s|2H − |t − s|2H 2
(2.4)
Scale Invariance and Wavelets
75
with the identification σ 2 := E |X(1)|2 . Indeed, if we adopt the convention according to which at time t = 0, X(0) = 0, it follows from the assumptions made that: 2 1 E X 2 (t) + E X 2 (s) − E X(t) − X(s) E X(t)X(s) = 2 2 1 E X 2 (t) + E X 2 (s) − E X(t − s) − X(0) = 2 E |X(1)|2 2H |t| + |s|2H − |t − s|2H , = 2 which explains the structure of relation (2.4). Moreover, the correlation function of the increment processes X (θ) reads: σ2 |τ + θ|2H + |τ − θ|2H − 2|τ |2H E X (θ) (t)X (θ) (t + τ ) = 2 2H 2H σ 2 2H θ 1 − θ − 2 . |τ | = 1 + + 2 τ τ
(2.5)
It is now possible to study in depth its asymptotic behaviors in both limits of large and small τ s. For instance, we show that in the limit τ → +∞ (i.e., τ θ), the autocorrelation function decreases asymptotically as τ 2(H−1) : E X (θ) (t)X (θ) (t + τ ) ∼
σ2 2H (2H − 1) θ2 τ 2(H−1) . 2
(2.6)
By Fourier duality, this behavior induces an algebraic spectral divergence with exponent 1 − 2H at the origin. Self-similar processes with stationary increments are hence closely related to long-range dependent processes. In the other limit, τ → 0 (i.e., τ θ), we show that for H > 12 : (2.7) E X (θ) (t)X (θ) (t + τ ) ∼ σ 2 θ2H 1 − θ−2H |τ |2H . This behavior characterizes the local regularity of each sample path of the process X. The following sections explain the notions associated with each of these limits: long-range dependence on the one hand and local regularity on the other hand. 2.2.3. Long-range dependence DEFINITION 2.3.– A second order stationary process X = {X(t), t ∈ } is said to be “long-range dependent” (or to have “long memory”) if its correlation function cX (τ ) := E X(t)X(t + τ ) is such that [BER 94, SAM 94]: cX (τ ) ∼ cr |τ |−β ,
τ −→ +∞
(2.8)
76
Scaling, Fractals and Wavelets
with 0 < β < 1. In the same way, the power spectrum density: +∞ cX (τ ) e−i2πf τ dτ ΓX (f ) := −∞
of a long-range dependent process is such that: ΓX (f ) ∼ cf |f |−γ ,
f −→ 0
(2.9)
with 0 < γ = 1 − β < 1 and cf = 2 (2π) sin((1 − γ)π/2) Γ(γ) cr , where Γ denotes the usual Gamma function. Under its form (2.8), long-range dependence is related to the fact that, for large lags, the (algebraic) decrease of the correlation function is so slow that it does not enable its summability: hence, there is a long memory effect, in the sense that significant statistical relations are maintained between very distant samples. Obviously, this situation is in contrast with that of Markovian processes with short memory, which are characterized by an asymptotic exponential reduction of the correlations. By definition, the existence of an exponential decrease involves a characteristic time scale, whereas this is no longer the case for an algebraic decrease: hence, it is a matter of scaling law behavior. By Fourier duality, long-range dependence implies that ΓX (0) = ∞, in accordance with the power law divergence expressed by (2.9). Finally, even if the property of long-range dependence exists and although its definition is independent from that of self-similarity, relation (2.6) demonstrates that a strong bond exists between these two notions, since it indicates that the increment process of a self-similar process with stationary increments presents, if H > 12 , long-range dependence. 2.2.4. Local regularity The main issue of this section, rather than the long-term behavior of the autocorrelation function, is its short-term behavior. Let X be a second order stationary random process, whose autocorrelation function is originally such that: E X(t)X(t + τ ) ∼ σ 2 (1 − C|τ |2h ),
τ −→ 0,
0 < h < 1.
(2.10)
Hence, it is easy to prove that this original covariance structure is equivalent to an algebraic behavior of the increments variance in the limit of short increments: E |X(t + τ ) − X(t)|2 ∼ C|τ |2h ,
τ −→ 0,
0 < h < 1.
This relation provides information on the local regularity of each sample path of the process X. For Gaussian processes, for instance, it indicates that these sample paths are continuous of order h < h. When 0 < h < 1, this means that these trajectories of X are everywhere continuous but nowhere differentiable.
Scale Invariance and Wavelets
77
To describe this local regularity more precisely, one can use the notion of Hölder exponent, according to the following definition. DEFINITION 2.4.– A signal X(t) is of Hölder regularity h 0 in t0 if there exists a local polynomial Pt0 (t) of degree n = h and a constant C such that: |X(t) − Pt0 (t)| C|t − t0 |h .
(2.11)
In the case where 0 h < 1, the regular part of X(t) is reduced to Pt0 (t) = X(t0 ), thus leading to the characterization of the Hölder regularity of X(t) in t0 by the relation: |X(t0 + θ) − X(t0 )| C|θ|h .
(2.12)
The Hölder exponent heuristically supplies a measure of the roughness of the sample path of X: the closer it is to 1, the softer and more regular the path; the closer it is to 0, the rougher and the more variable the path. The asymptotic algebraic behavior of the increments variance thus highlights a Hölder regularity h < h of the sample paths of the process X. This correspondence between the asymptotic algebraic behavior of increments and the local regularity remains valid even if the process X is no longer stationary, but only has stationary increments. The processes whose sample paths possess a uniform and constant local regularity are said to be monofractal. As far as self-similar processes with stationary increments are concerned, it is easy to observe that, on the one hand, starting from (2.12), the increments present an algebraic behavior for all the θ in general, hence in particular in the limit θ → 0: E |X(t + θ) − X(t)|2 = E |X(θ) − X(0) |2 = σ 2 |θ|2H
(2.13)
=0
whereas, on the other hand, relation (2.13) indicates that the increment process presents an autocovariance as in (2.10). Self-similar processes with stationary increments, as well as their increment processes, thus present uniform local regularities (i.e., on average and everywhere) h < H. 2.2.5. Fractional Brownian motion: paradigm of scale invariance The simplest and most commonly used model of a self-similar process is that of fractional Brownian motion (FBM) [MAN 68], which is characterized by its real exponent 0 < H < 1, called the Hurst exponent. DEFINITION 2.5.– FBM BH = {BH (t), t ∈ ; BH (0) = 0} is the only zero-mean Gaussian process which is self-similar and possesses stationary increments.
78
Scaling, Fractals and Wavelets
The self-similarity and the stationary nature of the increments guarantee that the covariance function of FBM is of the form (2.4). As regards the Gaussian character, it demands that the probability law of FBM must be entirely determined by this covariance structure. FBM can be considered as a generalization of ordinary Brownian motion. In the case of ordinary Brownian motion, we know that the increments possess the particularity of being decorrelated (and therefore independent because of Gaussiannity). The generalization offered by FBM consists of introducing a possibility of correlation between the increments. In fact, we show that: E BH (t + θ) − BH (t) BH (t) − BH (t − θ) = σ 2 22H−1 − 1 |θ|2H which confirms the decorrelation between the increments when H = 12 (i.e. ordinary Brownian motion) but induces a positive correlation (persistence) or a negative correlation (antipersistence) depending on whether H > 12 or H < 12 . DEFINITION 2.6.– We call fractional Gaussian noise (FGN) the increments process GH;θ := {GH;θ (t), t ∈ } defined by: GH;θ (t) :=
1 (θ) B (t), θ H
θ>0
(2.14)
where BH is FBM. It is, by construction, a stationary process, everywhere continuous but nowhere differentiable, that can be considered as an extension of white Gaussian noise. Hence, we must be very careful when we decide to take the limit of definition (2.14) when θ → 0. Nevertheless, if we are interested in the behavior of FGN with “small” increments, we observe according to (2.6) and (2.14) that: cGH;θ (τ ) := E GH;θ (t)GH;θ (t + τ ) ∼
σ2 2H(2H − 1)τ 2(H−1) , 2
τ θ.
On the one hand, this behavior highlights that FGN presents some long memory and, on the other hand, that the power spectrum density of FGN is proportional to |f |−(2H−1) . It is therefore possible to prove on the basis of several arguments (integration/differentiation type) [FLA 92] that the FBM itself possesses an “average spectrum” of the form ΓBH (f ) ∝ |f |−(2H+1) . Along with its role of spectral exponent, the parameter H also controls the Hölder regularity of the sample paths of FBM and FGN, which is h < H in any point. To this regularity (or irregularity), a notion of fractality is naturally associated with the Gaussian processes, since the Hausdorff dimension of the sample paths is equal to dimH graph(BH ) = 2 − H (for a precise definition of the Hausdorff dimension, see Chapter 1).
Scale Invariance and Wavelets
79
As a result, FBM presents the advantage (or the disadvantage) of being globally self-similar on the entire frequency axis, the only parameter H controlling, according to the requirements, one or other of the three regimes cited before: self-similarity, long memory and local regularity. In terms of modeling, FBM appears as a particularly interesting starting point (as can be the case for white Gaussian noise in stationary contexts). This simplicity (FBM is the only Gaussian process with stationary and self-similar increments and it is entirely determined by the single parameter H) is not of course without counterparts when it is comes to applications, i.e. as soon as it becomes necessary to consider real data. From this theme, numerous variations can be considered, which are not only mentioned here but are also studied in detail in the other chapters of this volume. In all cases, it is a matter of replacing the single exponent H by a collection of exponents. 2.2.6. Beyond the paradigm of scale invariance To begin with, we can consider modifying relation (2.13) by allowing the exponent to depend on time: E |X(t + θ) − X(t)|2 ∼ C(t)|θ|2h(t) ,
θ → 0.
When 0 < h(t) < 1 is a sufficiently regular deterministic function, we describe the process X as multifractional or, when it is Gaussian, as locally self-similar, i.e., that locally around t, X(t) possesses similarities with a FBM of parameter H = h(t) (for more details, see Chapter 6). The local regularity is no longer a uniform or global quantity along the sample path but, on the contrary, it varies in time, according to h(t), which therefore makes it possible to model time variations of the roughness. When h(t) is itself a strong irregular function, possibly a random process, in the sense that, with t fixed, h(t) depends on the observed realization of X, the process X is said to be multifractal. The variability fluctuations are no longer described by h(t), but by a multifractal spectrum D(h) which characterizes the Hausdorff dimension of the set of points t where h(t) = h (see Chapter 1 and Chapter 3). One of the major consequences of multifractality in the processes is the fact that quantities usually called partition functions, behave according to power laws in the small scale limit: 1 (τ ) |X (t + kτ )|q cq |τ |ζ(q) , n n
|τ | → 0
(2.15)
k=1
n For processes with stationary increments, the time averages (1/n) k=1 |X (τ ) (t+ kτ )|q can be regarded as estimations of the averages of the set E |X (τ ) (t)|q . Relation (2.15) thus recalls equation (2.2), which is a consequence of self-similarity. However, a fundamental difference exists: the exponents ζ(q) do not possess a priori any
80
Scaling, Fractals and Wavelets
reason to present a linear behavior qH. In other words, the description of scaling laws in data cannot be carried out with a single exponent but requires a whole collection of them. Measuring exponents ζ(q) represents a possibility, through a Legendre transform, of estimating the multifractal spectrum. However, a detailed discussion of the multifractal processes is beyond the scope of this chapter; to this end, see Chapter 1 and Chapter 3. Multifractal processes provide a rich and natural extension of the self-similar model insofar as a single exponent is replaced by a set; nevertheless, they are essentially related to the existence of power law behaviors. In the analysis of experimental data, such behaviors might not be observed. In order to illustrate these situations, the infinitely divisible cascades model exploits an additional degree of freedom: we relax the constraint of a proper power law behavior for the moments, and replace it with a simple behavior that has separable variables q (order of the moment) and τ (scale analysis). The equations below explain this behavior: self-similar multifractal inf. divisib. casc.
E |X (τ ) (t)|q = cq |τ |qH = cq exp(qH log τ ); E |X (τ ) (t)|q = cq |τ |ζ(q) = cq exp ζ(q) log τ ; E |X (τ ) (t)|q = cq exp H(q)n(τ ) .
(2.16) (2.17) (2.18)
In this scenario, the function n(τ ) is no longer fixed a priori to be log τ , as much as the function H(q) is no longer a priori linear according to qH. The concept of an infinitely divisible cascade was initially introduced by Castaing in the context of turbulence [CAS 90, CAS 96]. The complete definition of this notion is beyond the scope of this chapter and can be found in [VEI 00]. It is nonetheless important to indicate that a quantity, called the propagator of the cascade, plays an important role here: it links the probability densities of process increments with two different scales τ and τ . The infinite divisibility formally translates the notion of the absence of any preferred time scale and demands this propagator be constituted of an elementary function G0 , convoluted with itself a number of times dependent only on the scales τ and τ , and therefore with the following functional form: Gτ,τ (log α) = [G0 (log α)]∗(n(τ )−n(τ
))
.
A possible interpretation of this relation is to read the function G0 as the elementary step, i.e. the building block of the cascade, whereas the quantity n(τ ) − n(τ ) measures how many times this elementary step must be carried out to evolve from scale τ to scale τ . The derivative of n with respect to log τ thus describes, in a sense, the size of the cascade. The term of infinitely divisible cascade is ascribed to situations where the function n possesses the specific form n(τ ) = log τ ; otherwise, we only refer to a scaling law behavior. The infinitely divisible scale invariant cascades correspond to multiscaling or multifractality when the scaling law
Scale Invariance and Wavelets
81
exists in the small scale limit. The exponents ζ(q) associated with the multifractal spectrum are thus connected to the propagator of the cascade by ζ(q) = H(q). When the functions H and n simultaneously take the forms H(q) = qH and n(τ ) = log τ , the infinitely divisible cascades are simply reduced to the case of self-similarity, thus represented as a particular case. The propagator is hence written as a Dirac function, Gτ,τ (log α) = δ(log α − H log(τ /τ )). The fundamental characteristic of the infinitely divisible cascades – separation of variables q and τ – induces the following relations, which are essential for the analysis [VEI 00]: log E |X (τ ) |q = H(q)n(τ ) + Kq ; log E |X (τ ) |q =
H(q) log E |X (τ ) |p + κq,p . H(p)
(2.19) (2.20)
These equations indicate that the moments behave as power laws with respect to each other, this property being exploited in the analysis. Further definitions, interpretations and applications of the infinitely divisible cascades can be found in [VEI 00]. 2.3. Wavelet transform 2.3.1. Continuous wavelet transform The continuous wavelet decomposition of a signal X(t) ∈ L2 (; dt) is a linear transformation from L2 (; dt) to L2 (×+∗ ; dtada 2 ), defined by [DAU 92, MAL 98] u−t 1 X(u)ψ ∗ du. (2.21) TX (a, t) := √ a a This is the inner product between the analyzed signal X and a set of analyzing waveforms obtained from a prototype wavelet (or mother wavelet) ψ by dilations with a scale factor a ∈ +∗ and shifts in time t ∈ . In order for the wavelet transform to be a joint representation in time and frequency of the information contained in X, in other words, to be so that the coefficients TX (a, t) account for X around a given instant, in a given frequency range, the mother wavelet must be a function well localized in both time and frequency. In order to obtain the inverse of the wavelet transform, it is also necessary that the mother wavelet satisfies a closure relation: t − t du da = δ(t − t ), ψ(u)ψ u − a a2 which induces a condition called admissibility: ∞ cψ = |Ψ(f )|2 df /|f |. −∞
82
Scaling, Fractals and Wavelets
Given this condition, it is possible to reconstruct the signal X by inverting the wavelet transform according to: u − t dt da −1 TX (a, t)ψ . X(u) = cψ a a2 From the admissibility constraint, it also follows that ψ must satisfy: ψ(t) dt = 0. Such a waveform ψ is therefore an oscillating function, localized on a short temporal support, hence the name wavelet. This oscillating behavior indicates that the wavelet transform does not detect the DC component (average value) of the analyzed signal X. For certain mother wavelets, this property can be extended to higher orders: tk ψ(t) dt = 0, ∀0 k < nψ which means that the wavelet analyzing a signal X is orthogonal to the polynomial components of a degree lower than or equal to its number of vanishing moments nψ . In other words, the wavelet coefficients obtained from a mother wavelet characterized by nψ vanishing moments are insensitive to the behaviors of the signal, which are more regular, i.e. softer than the behavior of a polynomial of a degree strictly lower than nψ ; on the other hand, they account for the information relative to behaviors that are more irregular than such polynomial trends. 2.3.2. Discrete wavelet transform One of the fundamental characteristics of the continuous wavelet transform is its redundant character: the information contained in a signal, i.e. in a space of one dimension, is represented, through the wavelet transform, in a space of dimension 2, the time-scale plane (t, a) ∈ ( × ∗+ ); neighboring coefficients thus share some part of the same information. To reduce this redundance, we define the discrete wavelet transform by the set of coefficients: j/2 X(u) ψ(2j u − k) du (2.22) dX (j, k) := 2 defined using a critical discrete1 sampling of the time-scale plane, which is usually called the dyadic grid: t → 2−j k, a → 2−j , (k, j) ∈ Z × Z ,
1. In [DAU 92] a detailed study can be found of the frames or oblique bases which correspond to the sub-critical sampling of the time-scale plane.
Scale Invariance and Wavelets
83
thus ending up with the correspondence: dX (j, k) = TX (t = 2−j k, a = 2−j ). In this case, the collection of dilated and shifted versions of the mother wavelet {ψj,k (t), j ∈ Z, k ∈ Z} may constitute a basis for L2 (). Here, to simplify, we will suppose that this refers to orthonormal wavelet bases. However, discrete wavelet transforms are not necessarily or a priori equivalent to the existence of an orthonormal basis. The strict definition of the discrete wavelet transform leading to (orthonormal) bases goes through multiresolution analysis. A multiresolution analysis consists of a collection of nested subspaces of L2 (): . . . ⊂ Vj+1 ⊂ Vj ⊂ Vj−1 ⊂ . . . Each Vj , j ∈ Z possesses its own orthonormal basis {2j/2 φ(2j · −k), k ∈ Z} constructed, as for the wavelets, from a prototype scaling function2 (or father wavelet) φ0 onto which dyadic dilations and integer shifts are applied. The embedded structure demands that the function φ must satisfy a two-scale relation: √ un φ(t − n). φ(t/2) = 2 n
The projection of a signal X ∈ L2 () on this basis thus supplies the approximation coefficients at scale j: j/2 X(u) φ(2j u − k) du. aX (j, k) := 2 To complete these approximations, it is necessary to project the signal X onto the supplementary spaces of Vj in Vj+1 ; therefore, we define Wj by: Vj ⊕ Wj = Vj+1 ,
j−1
Wj = ∅,
Vj :=
Wj . j =−∞
j∈Z
For each fixed scale j, the wavelet family {2j/2 ψ(2j . − k), k ∈ Z} thus forms a (orthonormal) basis of the corresponding subspace:
+∞ j/2 j dX (j, k) ψ(2 t − k) . Wj := X : X(t) = 2 k=−∞
2. To define a multiresolution analysis, this function φ has to satisfy a certain number of constraints which are not detailed here [DAU 92].
84
Scaling, Fractals and Wavelets
There again, the embedded structure imposes a two-scale relation on the wavelet: √ vn φ(t − n). ψ(t/2) = 2 n
The wavelet coefficients or detail coefficients therefore correspond to the projections of X on Wj . The signal X can thus be represented as a sum of approximations and details: aX (j, k) 2j/2 φ(2j t − k) X(t) = k
+
−∞ j =j
(2.23)
j /2
dX (j , k) 2
j
ψ(2 t − k).
k
time
scale
signal
details approximation
high-pass filter + decimation
low-pass filter + decimation
Figure 2.2. Fast pyramidal algorithm with filter structure for discrete wavelet decompositions. An approximation aX (0, k) (at scale 0) of the continuous time signal X is initially calculated (only this stage involves a continuous time evaluation from X). The “signal” represented in the figure is made up by the sequence aX (0, k). In multiresolution analysis this approximation a0,k is decomposed in a series of details dX (−1, k) and a new and rougher approximation aX (−1, K). This procedure is then iterated from the sequence aX (−1, k). The impulse responses of the discrete-time filters depend on the generating sequences u and v which define the scaling function and the wavelet. In the case of orthonormal bases, they are exactly equal to them
Scale Invariance and Wavelets
85
Finally, thanks to the properties of embedded spaces specific to multiresolution analysis, there exist very fast algorithms with a pyramidal structure, which enables effective and efficient calculations of the discrete decomposition coefficients. From the sequences u and v, described as the generators of multiresolution analysis, we can prove that the approximation and detail coefficients at octave j can be calculated from those at octave j − 1: aX (j, k) = X(t)2j/2 φ(2j t − k) dt = =
√ X(t)2j/2 2 un φ 2(2j t − k) − n dt
n
un
X(t)2(j+1)/2 φ(2j+1 t − 2k − n) dt
n
=
un aj+1,2k+n
n
=
u∨ n aj+1,2k−n
n
= u∨ · ∗ aj+1,· (2k) and, in an identical manner: dX (j, k) = v·∨ ∗ aX (j + 1, ·) (2k) where ∗ denotes the discrete time convolution operator, i.e., (x· ∗ y· )(k) = ∨ ∨ n x(n)y(k − n), un = u−n and vn = v−n . The two previous relations can be rewritten by using the decimation operator ↓2 (y = ↓2 x means yk = x2k , i.e., that every other sample x is left out): ! " aX (j, k) = ↓2 (u∨ · ∗ aj−1,· ) (k); ! " dX (j, k) = ↓2 (v·∨ ∗ aj−1,· ) (k). Thanks to this recursive structure, the calculation cost of discrete wavelet decomposition of a signal uniformly sampled on N points is in O(N ). 2.4. Wavelet analysis of scale invariant processes The aim of this section is to study how these fundamental principles (scale changing operator) and essential properties (multiresolution structure, number of vanishing moments, localization) of wavelet decomposition can be exploited, in order to characterize and easily measure the scale invariance phenomena that have been previously described.
86
Scaling, Fractals and Wavelets
Let us note that the set of results mentioned below can be formulated in the same way as with the continuous wavelet decompositions. However, for the sake of simplicity and conciseness, we will only tackle the case of discrete random fields of orthogonal wavelet coefficients, arising from the decomposition of scale invariant processes. 2.4.1. Self-similarity PROPOSITION 2.1.– The wavelet coefficients resulting from the decomposition of a self-similar process of index H satisfy the equality: L −j(H+ 1 ) 2 dX (j, 0), . . . , dX (j, Nj − 1) = dX (0, 0), . . . , dX (0, Nj − 1) . 2 This result, initially demonstrated for the FBM [FLA 92] and then generalized to the set of self-similar processes [AVE 98], is based on the scale invariance principle stemming from the dilation/compression operator which defines the wavelet analysis. To outline the proof, it is only necessary to write down the main argument: when L 2jH X(u): X(2j u) = dX (j, k) = X(u)ψ(2j u − k)2j/2 du = L
=
2−j/2 X(2−j u)ψ(u − k) du 1
2−j(H+ 2 )
X(u)ψ(u − k) du
1
= 2−j(H+ 2 ) dX (0, k). The principal consequence of self-similarity is the fact that, when they do exist, the q-th order moment of the wavelet coefficients satisfy the equality: 1
E |dX (j, k)|q = 2−jq(H+ 2 ) E |dX (0, k)|q . PROPOSITION 2.2.– The wavelet coefficients resulting from the decomposition of a process with stationary increments are stationary at each scale 2j . To understand the origin of this result, let us note that the sampled process of −j increments X (θ=2 ) [2−j k] := X((k + 1)2−j ) − X(k 2−j ) can be identified with a wavelet decomposition (2.22) according to: −j X (2 ) [2−j k] = 2j X(u) [δ(2j u − k − 1) − δ(2j u − k)] du = 2j/2 dX (j, k),
Scale Invariance and Wavelets
87
with ψ(t) = δ(t−1)−δ(t) as the analyzing wavelet (an elementary wavelet sometimes referred to as the poor man’s # wavelet). In fact, it is the naturally admissible oscillating structure of the wavelets ( ψ(t) dt = 0) which guarantees this stationarity in the case of processes with stationary increments. Heuristically, and by underlining the main argument – the fact that the number of vanishing moments is at least greater than or equal to 1 – the proof reads (on the coefficients of the discrete decompositions and with j = 0 to simplify the writing): dX (0, k + k0 ) = X(u)ψ(u − k − k0 ) du = X(u + k0 )ψ(u − k) du = [X(u + k0 ) − X(k0 )]ψ(u − k) du L [X(u) − X(0)]ψ(u − k) du = = X(u)ψ(u − k) du = dX (0, k). This proof highlights the role played, for stationarization, by the fact that ψ is of zero-mean value (i.e. that its number of vanishing moments is at least 1). This result was obtained in the case of FBM, although directly from the covariance form, in [FLA 92, TEW 92], extended to stable cases, independently by different authors [DEL 00, PES 99] and proved in a general context in [CAM 95, MAS 93]. Given that we are dealing with processes with stationary increments of order p, the simple admissibility condition of the wavelets is no longer sufficient. Hence, it is necessary to choose a wavelet analysis ψ possessing nψ p vanishing moments so that the coefficient series dX (j, k) obtained are stationary at each scale. The complete proof of this result is given in [AVE 98]. However, a good way to make the issue clearer would be to argue here that the wavelet tool plays a role similar to that of a differentiation operator, insofar as the number of vanishing moments control, by time-frequency duality, the behavior of the spectrum magnitude |Ψ(f )| in the vicinity of the zero frequency. Indeed, for a wavelet ψ possessing nψ vanishing moments, we have |Ψ(f )| ∼ |f |nψ , f → 0, which at first approximation we can identify with the differentiation operator of order nψ . PROPOSITION 2.3.– The wavelet coefficients resulting from the decomposition of a process X which is zero-mean, self-similar of index H, of finite variance and with stationary increments (H − ASAS ) possesses, when they exist, moments of order q satisfying the following scaling law: 1
E |dX (j, k)|q = E |dX (0, 0)|q 2−jq(H+ 2 ) .
88
Scaling, Fractals and Wavelets
This last result stems directly from the coupling of the two previous propositions. For processes with finite variance (i.e., whose third order moment 2 exists) – Gaussian processes, just as the FBM, for instance – this relation takes on the following specific form: E |dX (j, k)|2 = E |dX (0, 0)|2 2−j(2H+1) .
(2.24)
Given that the latter are second order statistics, the particular form (2.4) of the covariance structure of a H − ASAS process makes it possible to deduce the asymptotic behavior of the dependence structure of the wavelet coefficients [FLA 92, TEW 92]. PROPOSITION 2.4.– The asymptotic covariance structure of the wavelet coefficients of a process X which is zero-mean, self-similar of index H, of finite variance and with stationary increments (H − ASAS ) takes on the form:
E dX (j, k)dX (j , k ) ≈ |2−j k − 2−j k |2(H−nψ ) ,
|2−j k − 2−j k | → ∞
which illustrates, on the one hand, that the larger the number of vanishing moments, the shorter the range of the correlation and on the other hand, that if H > nψ + 12 , the long-range dependence which exists for the increment process if H > 12 , is transformed into a short-range dependence [ABR 95, FLA 92, TEW 92]. The set of the results which have just been presented can be made more precise when we specify the distribution law which underlies the self-similar process with stationary increments. The Gaussian case, illustrated by the FBM, has been widely studied [FLA 92, MAS 93]. Its wavelet coefficients are Gaussian at all scales. More recently, interest in the non-Gaussian case has led to developments for self-similar α-stable processes (or α-stable motions) [ABR 00a, DEL 00]. Hence, we can deduce from the wavelet decomposition of such processes that the series of their coefficients dX (j, k), in addition to the above-mentioned properties, is itself α-stable with the same index. 2.4.2. Long-range dependence As specified in section 2.2.3, stationary processes with “long-range dependence” are characterized by a slow decrease of their correlation function cX (τ ) ∼ cr |τ |−β , 0 < β < 1. Thus, the strong statistical connections maintained even between distant samples, X(t) and X(t + τ ), make the study and analysis of such processes much more complex, by impairing, for example, the convergence of algorithms relying on empirical moment estimators. It will be shown hereafter that wavelet decomposition of a process with long-range dependence makes it possible to circumvent this difficulty since – under certain conditions – the series of coefficients dX (j, k) exhibit
Scale Invariance and Wavelets
89
short-term dependence. The covariance function of the wavelet coefficients possesses the following form: E dX (j, k)dX (j , k ) j+j −β 2 cr |τ | ψ(2j u − k) ψ 2j (u − τ ) − k du dτ ∼2 = 2−
j+j 2
cf
(2.25)
Ψ(2−j f )Ψ(2−j f ) −i2πf (2−j k−2−j k ) e df, |f |γ
indicating that its asymptotic behavior, i.e. for the large values of the interval |2−j k − 2−j k |, is equivalent to that of its original Fourier transform and hence to that of the relation:
2−(j+j )nψ 2(−j−j )nψ |f |2nψ |Ψ(2−j f )Ψ(2−j f )| ∼ = . −γ γ f →0 |f | |f | |f |γ−2nψ Thus, we can observe the effect of the number of vanishing moments nψ of the wavelet, which may compensate the original divergence of the spectrum density of the process. By choosing a wavelet such that nψ γ/2, the long-range dependence of the process is no longer preserved in the coefficient sequences of the decomposition. Hence, the bigger nΨ is, the faster the residual correlation decreases:
E dX (j, k)dX (j , k ) ≈ |2−j k − 2−j k |γ−2nψ −1 ,
|2−j k − 2−j k | → ∞.
From equation (2.25) we can also prove that the variance of the wavelet coefficients follows a power law behavior as a function of scales: E |dX (j, k)|2 = 2−j(1−β) cf
|Ψ(f )|2 df = c0 2−jγ . |f |γ
(2.26)
This relation will be at the core of the estimation procedure of the parameter γ (see the following section). Finally, it is important to specify that, since it is possible to invert the wavelet decompositions (see equation (2.23)), the non-stationarity of the studied processes does not disappear from the analysis (no more than the long-range dependence does); all the information is preserved but redistributed differently amongst the coefficients. Thus, long-range dependence and non-stationarity are related to the approximation coefficients aX (j, k) of the decomposition, whereas self-similarity is observed through the scales, by an algebraic progression of the moments of order q of the detail coefficients dX (j, k).
90
Scaling, Fractals and Wavelets
2.4.3. Local regularity The local regularity properties of process sample paths have been introduced in section 2.2.4. Their “wavelet” counterparts most often derive from the orthogonal discrete wavelet transform, given that they could be extended, normally quite easily, to the continuous (surfaces) varieties (see Chapter 3). THEOREM 2.1.– Let X be a signal with Hölder regularity h 0 in t0 and ψ a sufficiently regular wavelet (nψ h). Hence, there exists a constant c > 0 such that for any j, k ∈ Z × Z: 1 |dX (j, k)| c 2−( 2 +h)j 1 + |2j t0 − k|h . Conversely, if for any j, k ∈ Z × Z: 1 |dX (j, k)| c 2−( 2 +h)j 1 + |2j t0 − k|h for h < h, thus X has Hölder regularity h in t0 . The proof of the theorem was established independently by Jaffard [JAF 89] and Holschneider and Tchamitchian [HOL 90]. In the light of this result, we note once again that it is the decrease of wavelet coefficients through scales which characterizes the local regularity of the sample path of X. Furthermore, this result is not surprising since the Hölder regularity of a function is a particular cause for the 1/f spectral behavior at high frequencies. The second part of Theorem 2.1 also shows that knowledge of the coefficients located “vertically” to the singular point (|2j t0 − k| = 0) is itself not sufficient to determine the local regularity of X in t0 . Strictly speaking, it would be necessary to consider the decomposition in its entirety, thus implying that an isolated singularity can affect all the coefficients dX (j, k) inside a cone, called an influence cone. For a wavelet whose temporal support is finite, this cone is also limited at each scale. From the estimation point of view, the direct implication of Theorem 2.1 is to highlight the practical limits of (discrete) orthogonal wavelet transforms, because it is quite unlikely that the abscissa t0 of the singularity coincides with the coefficients line on the dyadic grid. Hence, in practice, it is more often a continuous analysis diagram which is preferred, for which we possess a less precise and incomplete version (direct implication) of Theorem 2.1 (see the following proposition). PROPOSITION 2.5.– If X is of Hölder regularity n < h < n + 1 in t0 , for a wavelet analysis ψ possessing nψ h vanishing moments, then we have the following asymptotic behavior: 1
TX (t0 , a) ∼ O(ah+ 2 ),
a −→ 0+ .
Scale Invariance and Wavelets
91
Proof. Let the continuous wavelet transform (2.21) constructed with nψ > n be: √ TX (t0 , a) = a ψ(u) X(t0 + au) du =
√
$
a
ψ(u) X(t0 + au) −
n
% r r
du,
cr a u
r=0
where cr represent the Taylor expansion coefficients of X in the vicinity of t0 . The signal X is of regularity n < h < n + 1 in t0 and ψ is a localized time function. Thus, in the limit of infinitely fine resolutions (a → 0+ ): √ lim+ TX (t0 , a) C a ψ(u)|au|h du a→0
h+ 12
a
1
|u|h ψ(u) du = Cψ ah+ 2 .
It is important to underline that if the wavelet ψ is not of sufficient regularity, it is the term of degree nψ in the Taylor polynomial which dominates at finite scales and it is thus the regularity of the wavelet which imposes the decrease of the coefficients through scales. However, one should not be misled by the interpretation of Proposition 2.5. It is only because we focus on the limited case of infinite resolution that the influence of the singularity seems to be perfectly localized in t = t0 . In reality, it is shown in [MAL 92] that, in the case of non-oscillating singularities (see Chapter 3), it is necessary and sufficient to consider the maximum local lines of the wavelet coefficients situated inside the influence cone, {TX (a, t) : |t − t0 | < c a}, to be able to characterize the local regularity of the process. In addition, the practical use of this property is made more difficult by the necessarily finite resolution imposed by the sampling of the data, which does not permit detailed scrutiny of the data beyond a minimum scale, which is noted by convention a = 1. Furthermore, the different aspects of the study of the local regularity of a function constitutes an important object in other chapters of this work. This is of true for Chapter 3, which tackles the issue of the characterization of functional regularity spaces, of Chapter 6 and Chapter 5 which expose the case of multifractional processes and their sample path regularity, and finally of Chapter 1, which presents the multifractal spectra as statistical and geometric measures of the distribution of pointwise singularities of a process. Finally, let us note that, as previously indicated in section 2.2.4, the increments of stochastic stationary processes with stationary increments for which the Hölder
92
Scaling, Fractals and Wavelets
exponent is constant throughout the sample paths satisfy the asymptotic relation: E |X(t + τ ) − X(t)|2 ∼ C|τ |2h ,
|τ | → 0.
The latter can be rewritten identically on the wavelet coefficients, which can be either continuous or discrete: E |TX (a, t)|2 ∼ a2h+1 ,
a −→ 0;
(2.27a)
E |dX (j, k)|2 ∼ 2−j(2h+1) ,
j −→ +∞.
(2.27b)
These relations should be compared with those obtained in the case of self-similarity (2.24) and long-range dependence (2.26), and will serve as a starting point in the construction of estimators (see the following section). 2.4.4. Beyond second order In this chapter, analysis is limited to the detailed presentation of the wavelet analysis of scaling laws existing in the second statistical order (self-similarity, long-range dependence, constant local regularity). Nevertheless, the study of scaling law models which involve all statistical orders (multifractal processes, infinitely divisible cascades) can be carried out from wavelet analysis in the same way and benefits from the same qualities and advantages. The wavelet analysis of multifractal processes is developed by S. Jaffard and R. Riedi in Chapter 3 and Chapter 4 respectively. The wavelet analysis of infinitely divisible cascades is detailed in [VEI 00]. 2.5. Implementation: analysis, detection and estimation This section is devoted to the implementation of wavelet analysis in the study of scale invariance phenomena, whether it is for detecting and highlighting them, or for estimating the parameters which describe them. The previous sections have outlined the power law behavior of the scale variance of wavelet coefficients: E |dX (j, k)|2 c C 2jα
(2.28)
see, for self-similarity, equation (2.24); for monofractality of long-range dependence, equation (2.26); for monofractality of sample paths, equation (2.27); these equations are represented in Table 2.1. This relation crystallizes the potential of the “wavelet” tool for the analysis of scaling laws and will hence be the core aspect of this section. In a real situation, we must begin by validating the relevance of a model before estimating its parameters. In other words, we must first highlight the existence of scale invariance phenomena and identify a scale range within which the above-mentioned
Scale Invariance and Wavelets
α
c
C
Self-sim. with stat. incr. 2H + 1 σ 2 = E |X(1)|2 Long-range dependence
γ
#
|Ψ(f )|2 /|f |2H+1 df #
cf
Uniform local regularity 2h + 1
93
|Ψ(f )|2 /|f |γ df
–
–
Table 2.1. Summary of scaling laws existing in the second statistical order
relation supplies an appropriate description of the data, then carry out the measure of exponent α. A simple situation serves as an introduction, where we suppose that there is an octave range j1 j j2 , already identified, for which the fundamental relation is satisfied in an exact manner: j1 j j2 ,
E |dX (j, k)|2 = cf C 2jα
and we concentrate on the estimation of the parameter α. 2.5.1. Estimation of the parameters of scale invariance To estimate the exponent α, a simple idea consists of measuring the slope of log2 E |dX (j, k)|2 against log2 2j = j. In practice of course, this implies that we have to estimate the quantity E |dX (j, k)|2 from a single observed realization of finite length. Given the properties of the wavelet coefficients (stationarity, weak statistical dependence) put forth in the previous section, we simply propose to perform the estimation of the ensemble average by the time average [ABR 95, ABR 98, FLA 92]: 1 dx (j, k)2 nj k
where nj designates the number of wavelet coefficients available at octave j. DEFINITION 2.7.– Let us begin by recalling the main characteristics of linear regression. Let yj be the random variables such that E yj = αj + b and let us define σj2 = var yj . The estimator by a weighted linear regression of α reads: j2 α ˆ=
j=j1
yj (S0 j − S1 )/σj2 S0 S2 − S12
≡
j2 j=j1
wj yj
(2.29)
94
Scaling, Fractals and Wavelets
with: Sp =
j2
j p /aj ,
p = 0, 1, 2
j=j1
where aj are arbitrary quantities, acting as weights associated with yj . With these definitions, the weights wj satisfy the usual constraints, i.e. j2 j2 j=j1 jwj = 1 and j=j1 wj = 0. We can also easily observe that the estimator is unbiased: α ˆ = α. Moreover, the variance of this estimator is written, in the case of uncorrelated variables yj : var α ˆ=
j2
aj σj2 .
j=j1
The choice of the weights remains to be specified. We know that the variance α ˆ is minimal if we take into consideration the covariance structure of yj , yj in the definition of aj . Once again, in the case of uncorrelated variables yj , this leads us to choose aj ≡ σj2 . In the case of scale invariance, we use the estimator defined earlier with the variables: 1 dX (j, k)2 − g(j), yj = log2 nj k
where g(j) are the correction terms aimed at taking into account the fact that E(log(·)) is not log(E(·)) and at ensuring that E yj = αj + b. Hence, this estimator simply consists of a weighted linear regression carried out in the diagram yj against j, referred to as the log-scale diagram [VEI 99]. In order to easily implement this estimator, it is necessary to further determine g(j) and σj2 and choose aj . To begin with, we assume that dx (j, k) are random Gaussian variables, i.e., that they result from the wavelet decomposition of a process which is itself jointly Gaussian. Moreover, if we idealize the weak correlation property of the wavelet coefficients in exact independence, then we can calculate g(j) and σj2 analytically: g(j) = Γ (nj /2)/ Γ(nj /2) log 2 − log2 (nj /2) ∼
−1 , nj log 2
nj → ∞;
(2.30a)
σj2 = ζ(2, nj /2)/ log2 2 ∼
2 , nj log2 2
nj → ∞,
(2.30b)
Scale Invariance and Wavelets
95
where Γ and Γ respectively designate the Gamma function and its derivative, and ∞ where ζ(2, z) = n=0 1/(z + n)2 defines a function called the generalized Riemann zeta function. Let us note that these analytical expressions, which depend on the known nj s alone, can hence be easily evaluated in practice. The numerical simulations presented in depth in [VEI 99] indicate that, for Gaussian processes, this analytical calculation happens to be an excellent approximation of reality, satisfying a posteriori the idealization of exact independence. Thus, for Gaussian processes, we obtain an estimator which is remarkably simple to carry out, since the quantities g(j) and σj2 can be analytically calculated and do not need to be estimated from data, and which gives excellent statistical performance. From these analytical expressions, we obtain: Eα ˆ = α, which indicates that the estimator is unbiased and this is also valid for observations of finite duration. With the choice aj = σj2 , its variance reads: Var α ˆ=
j2
σj2 wj2 =
j=j1
S0 S0 S2 − S12
and it attains the Cramér-Rao lower bound, which is calculated under the same hypotheses (Gaussian process and exact independence) [ABR 95, VEI 99, WOR 96]. In addition, if we add the form nj = 2−j n (with n as the number of coefficients in the initial process) induced by the construction of the multiresolution analysis, we obtain the following expression for the variance: Var α ˆ=
j2
σj2 wj2
j=j1
1 (1 − 2−J ) · 1−j1 , n 2 F
(2.31)
where F = F (j1 , J) = log2 2·(1−(J 2 /2+2) 2−J +2−2J ) and where J = j2 −j1 +1 denotes the number of octaves involved in the linear regression. This analytical result shows that the variance of the estimator decreases in 1/n, in spite of the possible presence of a long-range dependence in the analyzed process. It is noteworthy that, in practice, relation nj = 2−j n is not exactly satisfied because of boundary effects, which are systematically excluded a priori from the measures. In the case of non-Gaussian processes, the implementation of the estimator is more subtle, since we cannot use analytical expressions for g(j) and σj2 . Nevertheless, in the case of finite variance processes, the variables (1/nj ) k dX (j, k)2 are asymptotically Gaussian and we can show that correcting terms can be introduced [ABR 00b] to the Gaussian case: g(j) ∼ −
1 + C4 (j)/2 ; nj log 2
σj2 ∼ 2/ log2 2
1 + C4 (j)/2 , nj
96
Scaling, Fractals and Wavelets
where C4 denotes the normalized fourth-order cumulant: 2 2 C4 (j) = E dX (j, k)4 − 3 E dX (j, k)2 / E dX (j, k)2 of the wavelet coefficients at octave j. The practical use of these relations requires the estimation of C4 , which can be difficult, as well as the guarantee that, for each octave, a sufficient number of points exist, so that the above form, which results from an asymptotic expansion, is valid. An approximate yet simple practical choice, regularly implemented, consists of using: g(j) ≡ 0; σj2 ≡ 2/(nj log2 2); aj ≡ σj2 . The numerical simulations proposed in [ABR 98, VEI 99] show that the performance of these choices are very satisfying for the analysis of long-range dependence. Indeed, such a choice implies that the importance in linear regression of yj is twice less than that of yj−1 , which is a point of view a priori realistic for the study of long-range dependence (Gaussianization effect for large scales), but less obvious for the study of local regularities. This choice is all the more delicate as it induces an effect on the bias and the variance of the estimator at the same time; an alternative choice, aj constant, can also be considered. In equation (2.28), from which the study of the scaling law behavior stems, the focus has, until now, been on the exponent α of the power law, since it defines the phenomenon. However, the measure of the multiplicative parameter cf can be fruitful for certain applications. This estimation is detailed in [VEI 99]. Finally, the case of self-similar processes with stationary increments of infinite variance (and/or mean value) will not be tackled here. It is especially developed in [ABR 00a]. 2.5.2. Emphasis on scaling laws and determination of the scaling range The previous section relied on the hypothesis that the quantity E dX (j, k)2 was made on the basis of a power law. This ideal situation is rarely observed for two reasons. On one hand, the real experimental data are likely to be only approximately described by the proposed models of scale invariance. On the other hand, certain models themselves induce only an approximate or asymptotic behavior as a power law – as is the case for long-range dependence or processes with fractal sample paths; in fact, only the self-similar model induces a strictly satisfied power law: LRD
j → +∞ E dX (j, k)2 ∼ cf C 2jα ;
Fractal j → −∞ E dX (j, k)2 ∼ cf C 2jα ; H-ss
∀j
E dX (j, k)2 = cf C 2jα .
Scale Invariance and Wavelets
97
Hence, in the implementation of the estimator described earlier, it is necessary to choose an octave range on which the measure is carried out, i.e., to select the octaves j1 and j2 . Making this choice does not necessarily mean extracting the theoretical values of j1 and j2 from the data, since these are not always defined by the model, but rather optimizes the statistical performance of the estimator. Widening the octave range [j1 , j2 ] implies the use of a higher fraction of the available wavelet coefficients, resulting in a reduction of the estimation variance, as indicated by the above-mentioned relation (2.31); conversely, it can also mean an increase in the estimation bias if we carry out the measure on a range where the behavior is notably different from a power law behavior. The choice of the range is thus guided by the optimization of a bias-variance trade-off. The example of long-range dependence is used: according to the model, we wish to choose j2 = +∞, i.e., in practice, j2 as large as possible, since the maximum limit of the wavelet decomposition is fixed by the number of coefficients of the analysis process and the importance of boundary effects; as for j1 , it is not imposed by the model. Choosing the larger j1 makes it possible to work in a zone where the asymptotic behavior is satisfied and thus means a small estimation bias but a strong variance (this is essentially contained in 2j1 , see equation (2.31), and thus doubles each time we increase j1 of 1). Qualitatively, we are led to tolerate a little bias (i.e., widen the measure range towards the small octaves) to reduce the variance and thus minimize the average quadratic error (AQE = (bias)2 + variance). In the case of the local regularity measure, the situation is different. The model tends to impose small j1 (in practice, j1 = 1) without fixing j2 . As in the preceding case, increasing j2 may induce a gap in the ideal behavior of the bias, but will only slightly reduce the variance: indeed, it is sufficient to study the form of the variance dependence of α ˆ to note the reduced influence of j2 , in accordance with the fact that there exist fewer and fewer wavelet coefficients exist on the coarsest octaves. In practice, we are led to choose the narrowest possible range which is also limited on the lowest octaves. To move towards more quantitative arguments, it is necessary to completely specify the models of the processes studied. We will keep on using the method which imposes the least possible a priori assumptions on the model and consider that it is sufficient to postulate that scale invariance phenomena are present. We resort to the quantity: G(j1 , j2 ) =
j2 (yj − α ˆ j − ˆb).2 σj2 j=j
(2.32)
1
which conveys a usual measure of the mean-square error between data and model. For Gaussian yj , the variable G(j1 , j2 ) follows a Chi-2 law χ2J−2 with J − 2 degrees of freedom. The dependence in j1 , j2 of the quantity: G(j1 ,j2 ) χ2J−2 (u) du Q(j1 , j2 ) = 1 − 0
98
Scaling, Fractals and Wavelets Ŧ5 9
D = 0.55 8
Ŧ10
D = 2.57
cf = 4.7
7
1 d j d 10
Ŧ15
4 d j d 10 6
y
j
y
j
5
Ŧ20 4
3
Ŧ25
2
Ŧ30
1
0
Ŧ35
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
Octave j
Octave j
Figure 2.3. Examples of log-scale diagram. Right: second order log-scale diagram for a long-range dependent process, also possessing a highly pronounced correlation structure of short memory type (visible to the small octaves); practically, it refers to ARFIMA (0, d, 2) with d = 0.25 and a second order moving average Ψ(B) = 1 + 2B + B 2 , implying (γ, cf ) = (0.50, 6.38). The vertical error bars for each j carry out confidence intervals at (2) 95% of Yj . A linear behavior is observed between the octaves [j1 , j2 ] = (4, 10), which excludes the small octaves (short range memory) but includes the larger ones. A weighted linear regression enables, in spite of the strong presence of short-term dependencies, a precise estimation of γ: γˆ = 0.53 ± 0.07, cˆf = 6.0 with 4.5 < cˆf < 7.8. Left: second order log-scale diagram for a self-similar process (FBM) of parameter H = 0.8. The linear behavior spreads ˆ = 0.79 over all scales and allows a precise estimation of H: H
makes it possible to work upon the choice of the analysis range. A value of Q close to 1 indicates the adequacy of the model, as opposed to Q close to 0. An approach which consists of examining breaking points in the behavior of Q with j1 , j2 is proposed. 2.5.3. Robustness of the wavelet approach One of the great difficulties in the analysis of scale invariance phenomena is linked to the fact that their qualitative expressions are close to those induced by non-stationarities. For a long time, data modeling by scale invariance has been rejected because it was considered, sometimes correctly, as an artefact due to non-stationarity. The difficulties are of two types: on the one hand, as just mentioned above, identifying scale invariance when it refers, in fact, to non-stationarities; on the other hand, failing to detect scale invariance or to correctly estimate its parameters, when non-stationary effects are superimposed. Wavelet analysis has made it possible to find solutions to these two problems. For the first type of problem, it was proposed [VEI 01] to chop the signal under analysis into L segments which do not overlap. For each segment, we carry out an estimation α ˆ l of the scale invariance parameter. Then we validate the relevance of using a scale invariance model by testing the similarity between blocks of α ˆ l . Hence, it does not refer to a stationary test in the more general sense, but more simply to a
Scale Invariance and Wavelets
99
18
30
16
25
14
20 12
15
yj
10
10
8
5
6
0
4
2
Ŧ5
0
Ŧ10
2000
4000
6000
8000
10000
12000
14000
1
2
3
4
5
30
6
7
8
9
10
11
7
8
9
10
11
Octave j
16000
18
25
16
20
14
12
15
10
D = 0.59
yj
10
8
5 6
0 4
Ŧ5
Ŧ10
2
2000
4000
6000
8000
time
10000
12000
14000
16000
0
1
2
3
4
5
6
Octave j
Figure 2.4. Robustness with respect to superimposed trends. Left: fractional Gaussian noise with H = 0.80 (above) and with sinusoidal and superimposed linear trends (below). Right: log-scale diagrams of signal corrupted by the trends, as computed with a Daubechies 2 wavelet (i.e., N = 2) (above) and Daubechies 6 wavelet (i.e. N = 6) (below). We observe that increasing N cancels the effects of the superimposed trends ˆ = 0.795 and allows for a reliable estimation of H: H
test aimed at detecting abnormally large fluctuations of estimations α ˆ l which leads us to reject the presence of scale invariance. In practice, the properties of the wavelet coefficients (weak statistical dependence among coefficients) and the definition and the theoretical study of the estimator α ˆ , presented earlier, make it possible to conduct this test as the detection of n mean value change within independent Gaussian variables of unknown but identical mean values and of possibly different but known variances [VEI 01]. For the second type of problem, the number of vanishing moments of the wavelet plays a fundamental role. By definition, the wavelet coefficients of a polynomial p(t) of degree P strictly smaller than the number of vanishing moments of the mother wavelet, P < nψ , are exactly zero. This means that if the observed signal Z is made up of a signal to analyze X on which a polynomial trend is superimposed, the wavelet analysis of the scale invariance phenomena that are likely to be present in X will be, given its linear nature, insensitive to the presence of p as soon as
100
Scaling, Fractals and Wavelets
nψ is sufficiently large. In practice, we do not necessarily know a priori the order of the corrupting polynomial, if any; we can thus simply carry out a series of wavelet analyses by making N increase. When these results no longer change with nψ , this is an indication that P has been overtaken. It is noteworthy that this procedure is made fully practicable by the low calculation cost of the discrete wavelet decomposition. Certainly, in practice, the trends superimposed on the data are not polynomial. However, in the case where they possess a sufficiently regular behavior (e.g., quasi-sinusoidal oscillations) or slightly irregular (e.g., in t−β , β > 0), the preceding argument remains valid: when we make nψ increase, the magnitude of the wavelet coefficients of the trend decreases, whereas that of X remains identical; the effect of the trend thus becomes quasi-negligible [ABR 98]. The superposition of a deterministic trend to the process X can be interpreted as a non-stationarity of the mean; this situation can be complex considering that the variance of the process itself evolves. Hence, we can write the observation as: Z(t) = a(t) + b(t)X(t) where X(t) is a process presenting scale invariance under the form of one of the models referred to earlier, (self-similarity, long-range dependence, etc.) and where a(t) and b(t) are sufficiently regular deterministic functions. Thus, it has been shown that the variation in the number of vanishing moments of the mother wavelet makes it possible to overcome the drift effects of a and b and to carry out reliable estimations of the scale invariance parameters associated only with X [ROU 99]. Finally, the analysis of the signal plus noise situation, which is usual in signal processing when we write an observation Z = Y + X (where X is the process with scale invariance to be studied and Y some additive random noise) has been considered in [WOR 96] by maximum likelihood approaches and will not be developed here. 2.6. Conclusion In this chapter, we have focused on a qualitative description rather than a rigorous formalization of the concepts, models and analyses. The main concern was to offer the reader, who is not specialized in this field but is eager to implement, from real data, certain principles of the fractal analysis, some entry points that are as much theoretical as practical. Especially in the first part, emphasis has been put on the relations between the different models used to describe scaling laws, their similarities and differences, so that this notion become accessible. Similarly, in the second part, the presentation of wavelet tools enabling the characterization and analysis of scaling laws has been structured around the different practical aspects essential for their implementation (selection of the scale range, estimation of the parameters, robustness and algorithm). All the technical analysis described here, as well as some extensions and variations, have been put in practice in Matlab, with toolboxes that are freely
Scale Invariance and Wavelets
101
accessible on the websites http://www.ens-lyon.fr/pabry and http://perso.ens-lyon .fr/paulo.goncalves. 2.7. Bibliography [ABR 95] A BRY P., G ONÇALVES P., F LANDRIN P., “Wavelets, spectrum estimation, and 1/f processes”, in A NTONIADIS A., O PPENHEIM G. (Eds.), Wavelets and Statistics, Springer-Verlag, Lecture Notes in Statistics 103, New York, p. 15–30, 1995. [ABR 98] A BRY P., V EITCH D., “Wavelet analysis of long-range dependent traffic”, IEEE Trans. on Info. Theory, vol. 44, no. 1, p. 2–15, 1998. [ABR 00a] A BRY P., P ESQUET-P OPESCU P., TAQQU M.S., “Wavelet based estimators for self similar α-stable processes”, in International Conference on Signal Processing: Sixteenth World Computer Congress (Beijing, China, 2000), August 2000. [ABR 00b] A BRY P., TAQQU M.S., F LANDRIN P., V EITCH D., “Wavelets for the analysis, estimation, and synthesis of scaling data”, in PARK K., W ILLINGER W. (Eds.), Self-similar Network Traffic and Performance Evaluation, John Wiley & Sons, p. 39–88, 2000. [AVE 98] AVERKAMP R., H OUDRÉ C., “Some distributional properties of the continuous wavelet transform of random processes”, IEEE Trans. on Info. Theory, vol. 44, no. 3, p. 1111–1124, 1998. [BER 94] B ERAN J., Statistics for Long-memory Processes, Chapman and Hall, New York, 1994. [CAM 95] C AMBANIS S., H OUDRÉ C., “On the continuous wavelet transform of second-order random processes”, IEEE Trans. on Info. Theory, vol. 41, no. 3, p. 628–642, 1995. [CAS 90] C ASTAING B., G AGNE Y., H OPFINGER E., “Velocity probability density functions of high Reynolds number turbulence”, Physica D, vol. 46, p. 177, 1990. [CAS 96] C ASTAING B., “The temperature of turbulent flows”, Journal de physique II France, vol. 6, p. 105–114, 1996. [DAU 92] DAUBECHIES I., Ten Lectures on Wavelets, SIAM, 1992. [DEL 00] D ELBEKE L., A BRY P., “Stochastic integral representation and properties of the wavelet coefficients of linear fractional stable motion”, Stochastic Processes and their Applications, vol. 86, p. 177–182, 2000. [FLA 92] F LANDRIN P., “Wavelet analysis and synthesis of fractional Brownian motion”, IEEE Trans. on Info. Theory, vol. IT-38, no. 2, p. 910–917, 1992. [FRI 95] F RISCH U., Turbulence: The Legacy of A. Kolmogorov, Cambridge University Press, Cambridge, 1995. [HOL 90] H OLSCHNEIDER M., T CHAMITCHIAN P., “Régularité locale de la fonction non différentiable de Riemann”, in L EMARIÉ P.G. (Ed.), Les ondelettes en 1989, Springer-Verlag, 1990. [JAF 89] JAFFARD S., “Exposants de Hölder en des points donnés et coefficients d’ondelettes”, Comptes rendus de l’Académie des sciences de Paris, vol. 308, 1989.
102
Scaling, Fractals and Wavelets
[LEL 94] L ELAND W.E., TAQQU M.S., W ILLINGER W., W ILSON D.V., “On the self-similar nature of Ethernet traffic (extended version)”, IEEE/ACM Trans. on Networking, vol. 2, p. 1–15, 1994. [MAL 92] M ALLAT S.G., H WANG W.L., “Singularity detection and processing with wavelets”, IEEE Trans. on Info. Theory, vol. 38, no. 2, p. 617–643, 1992. [MAL 98] M ALLAT S.G., A Wavelet Tour of Signal Processing, Academic Press, San Diego, California, 1998. [MAN 68] M ANDELBROT B.B., VAN N ESS J.W., “Fractional Brownian motions, fractional noises, and applications”, SIAM Review, vol. 10, no. 4, p. 422–437, 1968. [MAN 97] M ANDELBROT B.B., Fractals and Scaling in Finance, Springer, New York, 1997. [MAS 93] M ASRY E., “The wavelet transform of stochastic processes with stationary increments and its application to fractional Brownian motion”, IEEE Trans. on Info. Theory, vol. 39, no. 1, p. 260–264, 1993. [PAR 00] PARK K., W ILLINGER W. (Eds.), Self-similar Network Traffic and Performance Evaluation, John Wiley & Sons (Interscience Division), 2000. [PES 99] P ESQUET-P OPESCU B., “Statistical properties of the wavelet decomposition of certain non-Gaussian self-similar processes”, Signal Processing, vol. 75, no. 3, 1999. [ROU 99] ROUGHAN M., V EITCH D., “Measuring long-range dependence under changing traffic conditions”, in IEEE INFOCOM’99 (Manhattan, New York), IEEE Computer Society Press, Los Alamitos, California, p. 1513–1521, March 1999. [SAM 94] S AMORODNITSKY G., TAQQU M.S., Stable Non-Gaussian Random Processes: Stochastic Models with Infinite Variance, Chapman and Hall, New York and London, 1994. [TEI 00] T EICH M., L OWEN S., J OST B., V IBE -R HEYMER K., H ENEGHAN C., “Heart rate variability: measures and models”, Nonlinear Biomedical Signal Processing, vol. II, Dynamic Analysis and Modeling (M. Akay, Ed.), Ch. 6, p. 159–213, IEEE Press, 2001. [TEW 92] T EWFIK A.H., K IM M., “Correlation structure of the discrete wavelet coefficients of fractional Brownian motions”, IEEE Trans. on Info. Theory, vol. IT-38, no. 2, p. 904–909, 1992. [VEI 99] V EITCH D., A BRY P., “A wavelet based joint estimator of the parameters of long-range dependence”, IEEE Transactions on Information Theory (special issue on “Multiscale statistical signal analysis and its applications”), vol. 45, no. 3, p. 878–897, 1999. [VEI 00] V EITCH D., A BRY P., F LANDRIN P., C HAINAIS P., “Infinitely divisible cascade analysis of network traffic data”, in Proceedings of the International Conference on Acoustics, Speech, and Signal Processing (Istanbul, Turkey), June 2000. [VEI 01] V EITCH D., A BRY P., “A statistical test for the constancy of scaling exponents”, IEEE Trans. on Sig. Proc., vol. 49, no. 10, p. 2325–2334, 2001. [WOR 96] W ORNELL G.W., Signal Processing with Fractals – A Wavelet-based Approach, Prentice-Hall, 1996.
Chapter 3
Wavelet Methods for Multifractal Analysis of Functions
3.1. Introduction A large number of signals are very irregular. In the most complex situations, the irregularity manifests itself in different forms and may change its form almost instantaneously. The most widely studied example in physics is the speed signal of a turbulent flow. During the 1980s, precise records of the speed of a turbulent flow were made in the ONERA wind tunnel at Modane (see Gagne et al. [GAG 87]). A thin wire, heated at one point, is placed in the flow of turbulent air; the rate at which the temperature decreases is directly proportional to the orthogonal component along the flow speed at the heated point. We obtain incomplete information as the signal is 1D; however, it is more precise than the best numerical simulations being carried out currently. The study of this signal showed that the signal recorded does not have statistical homogenity; its regularity varies a lot from one point to another [ARN 95b, FRI 95]. Such signals cannot be modeled by processes such as fractional Brownian motion, for example. The techniques of multifractal analysis were developed in order to model and analyze such behaviors. Originally introduced for fully developed turbulence, these techniques began to be used, within a few years, in various scientific fields: traffic analysis (road and Internet traffic) [ABR 98, LEV 97, TAQ 97, WIL 96, WIL 00], modeling of economic signals [MAN 97], texture analysis [BIS 98], electrocardiograms [AMA 00], etc. The mathematical theory of multifractal functions has expanded considerably: not only were numerous heuristic arguments used to numerically analyze the multifractal Chapter written by Stéphane JAFFARD.
104
Scaling, Fractals and Wavelets
signals studied and justified under certain assumptions within a context of limited validity but, most importantly, these mathematical results had various consequences on applications; they led researchers to introduce new tools that made it possible to refine and enrich the techniques of multifractal analysis. Inside mathematics, multifractal analysis acquired an extremely original position, since functions taken from very different domains prove to be multifractal: – in probabilities, the sample paths of the Lévy process [JAF 99]; – in analytical theory of numbers, trigonometric series linked to theta functions [JAF 96a]; – in analysis, the geometric constructions like “Peano functions” [JAF 97c, JAF NIC]; – in arithmetic, functions where diophantine approximation properties play a role [JAF 97b]; – etc. Multifractal analysis thus provides us with a vocabulary and a cluster of methods that help establish connections and find analogies between diverse fields of science and mathematics. Almost immediately since their appearance, wavelet analysis techniques were applied to multifractal analysis of signals by Arneodo and his team of CRPP of Bordeaux [ARN 95b]. These techniques were seen to be extremely powerful for mathematical analysis of problems, as well as for the construction of robust numerical algorithms. In this chapter, we explain certain fundamental results related to the wavelet methods of multifractal analysis in detail. We also briefly describe the vast scientific panorama of recent times and conclude with a review of specialized articles. In order to facilitate the presentation and the notations, all results will be set out in dimension l. Most of the results can easily be applied to functions defined in d (see [JAF 04b, JAF 06]). The reader can find a more detailed presentation of the different fields of application of multifractal analysis in [ARN 95a, JAF 01a, MAN 98, LAS 08, WEN 07]. 3.2. General points regarding multifractal functions 3.2.1. Important definitions Multifractal functions help in modeling signals whose regularity varies from one point to another. Thus, the first problem is to mathematically define a function’s regularity at every point. What is “pointwise regularity”? It is a way of quantifying,
Wavelet Methods for Multifractal Analysis of Functions
105
with the help of a positive real number α, the fact that the graph of a function is generally rough at a given point x0 (the picture is not simply superficial; the concepts we introduce have, in fact, been used in rough symmetry [DUB 89, TRI 97]). The Hölder regularity generalizes familiar concepts: the “minimum level” of regularity is continuity. A function f is continuous in x0 if we have |f (x) − f (x0 )| → 0 when x → x0 ; the continuity will correspond to a regularity index α = 0. Similarly, f is differentiable at x0 if there is a function P such that |f (x) − P (x − x0 )| → 0 which is faster than |x − x0 | when x → x0 ; the derivability will correspond to a regularity index α = 1. The following definition is a direct generalization of these two cases. DEFINITION 3.1.– Let α be a positive real number and x0 ∈ ; a function f : → is C α (x0 ) if there exists a polynomial P of degree less than α such that: |f (x) − P (x − x0 )| C|x − x0 |α
(3.1)
NOTE 3.1.– The polynomial P is unique (if Q was acceptable as well, by applying Definition 3.1 with P and Q, we would have |P (x) − Q(x)| C|x|α and, since P − Q is of a maximum degree [α], we would have P − Q = 0). The constant in the polynomial P (x − x0 ) is always f (x0 ); similarly, the first degree term, if present, is always (x − x0 )f (x0 ) (by the derivative’s definition). Also, if f is C [α] (x0 ) close to x0 , the comment that we just made concerning the uniqueness of P implies that the polynomial P is Taylor’s expansion of f in x0 of order α. However, as equation (3.1) can take place for large values of α, without f being twice more differentiable in x0 (in which case Taylor’s expansion stops at similar term), we will consider, for example, the “chirp” xn sin(x−n ) in 0 for big n (which does not have a second derivative in 0 since the first derivative is not continuous). We see that P gives a generalization of the notion of Taylor’s expansion (see also Chapter 4; we will consult [GIN 00, MEY 98] for extensions of these concepts in more general contexts). We will finally note that equation (3.1) implies that f is bounded in the vicinity of x0 ; therefore, we suppose that the functions we look at are locally bounded (see [MEY 98], where the exponent factor is introduced, which makes it possible to define a weak notion of Hölder regularity for functions that are not a priori locally bounded – and the same for distributions). DEFINITION 3.2.– The Hölder exponent of f in x0 is: hf (x0 ) = sup{α : f is C α (x0 )} The Hölder exponent is a function that is defined point by point and describes local variations of the regularity of f . Certain functions have a constant Hölder exponent. Thus, the Weierstrass series: b−Hj sin(bj x) Wb,H (x) = j
106
Scaling, Fractals and Wavelets
has a Hölder exponent that is constant and equal to H; similarly, the sample paths of the Brownian motion verify with near certainty that hB (x) = 12 for all x. In a more general manner, a fractional Brownian motion of exponent H has at every point a Hölder exponent equal to H. Such functions are irregular everywhere. Our objective is to study functions whose Hölder exponent can jump from one point to another. In such a situation, the numerical calculation of functions hf (x0 ) is completely unstable and of little significance. We are rather trying to extract less precise information: whether or not the function hf takes a certain given value H and, if it does, what is the size of the sets of points where hf takes on this value? Here we are faced with a new problem: what is the “right” notion of “size” in this context? We will not be able to fully justify the answer to this question because it is a result of the study of numerous mathematical examples. Let us just keep in mind that the term “size” does not signify “Lebesgue measure” because, in general, there exists a Hölder exponent that is the “most probable” and that appears almost everywhere. The other exponents thus appear on all zero sets and the “Lebesgue measure” does not make it possible to differentiate them. Besides, the “right” notion of size cannot be the box dimension because these sets are usually dense. In fact, we expect them to be fractal. A traditional mathematical method to measure the size of such dense sets of zero measure is to calculate their Hausdorff dimension. Let us recall its definition. DEFINITION 3.3.– Let A be a subset of . For ε > 0, let us note: εdi Mεd = inf R
i
where R signifies a generic covering of A by intervals ]xi , xi + εi [ of a length εi ε. The operator inf is thus taken on all these coverings. For all d ∈ [0, 1], Hausdorff d-dimensional measure of A is: mes d (A) = lim M d →0
This measure takes on a value of +∞ or 0 except for, at the most, a value of d and the Hausdorff dimension of A is: dim(A) = sup d: lim M d = +∞ = inf d: lim M d = 0 →0
→0
DEFINITION 3.4.– Let f be a function and H 0. If H is a value taken by function hf (x), let us note by EH the set of points x where we have hf (x) = H. Therefore, the singularity spectrum (or Hölder spectrum) of the signal being studied is: fH (H) = dim(EH ) (we use the convention fH (H) = −∞ if H is not a Hölder exponent of f ).
Wavelet Methods for Multifractal Analysis of Functions
107
The concept of multifractal function is not precisely defined (just like the concept of a fractal set). For us, a multifractal function is a function whose spectrum of singularities is “non-trivial”, i.e. unreduced to a point. In the examples, fH (H) takes on positive values on an entire interval [Hmin , Hmax ]. Its assessment thus requires a study of an infinity of fractal sets EH , hence the term “multifractal”. 3.2.2. Wavelets and pointwise regularity For many reasons that we shall gradually discover, wavelet methods of analysis are a favorite tool for studying multifractal functions. The first reason is that we have a simple criteria that allows us to characterize the value of the Hölder exponent by a decay condition of a given function’s wavelet coefficients. Let us begin by recapitulating certain points related to the wavelet analysis methods. An orthonormal base of wavelets of L2 () has a particularly simple algorithmic form: we start from a function ψ (the “mother” wavelet) that is regular and well-localized; the technical assumptions are: ∀i = 0, . . . , N,
|ψ (i) (x)|
∀m ∈ N,
C(i, m) (1 + |x|)m
for a relatively big N . We can choose such functions as ψ, such that, moreover, the translation-dilation of ψ: ψj,k (x) = 2j/2 ψ(2j x − k),
j, k ∈ Z
form an orthonormal base of L2 () (see [MEY 90]) (we will choose N to be bigger than the maximum regularity that we expect to find in the signal analyzed; we can also take N = +∞ and the wavelet ψ will thus belong to the Schwartz class). We verify that the wavelet has a corresponding number of zero moments: ψ(x)xi dx = 0 ∀i = 0, . . . , N,
Thus, every f ∈ L2 () function can be written as: f (x) =
ef (k, j) ψ(2j x − k)
j∈Z k∈Z
where df (k, j) are the wavelet coefficients of f : f (t)ψ(2j t − k) dt ef (k, j) = 2j
108
Scaling, Fractals and Wavelets
We should note that we do not choose an L2 normalization for the wavelet coefficients, but an L1 normalization which is better adapted to the study of Hölder regularity. Let us first study the characterization by wavelets of uniform regularity. Let us begin by defining it. A function f belongs to C α () if condition (3.1) takes place for all x0 , with the possibility of choosing C uniformly, i.e. independently of x0 . If we have α < 1, taking into account that P (x − x0 ) = f (x0 ), this condition can be rewritten as: f ∈ C α () ⇐⇒ ∀x, y ∈ |f (x) − f (y)| C|x − y|α This condition of uniform regularity is characterized by a condition of uniform decay of the wavelet coefficients of f (see [MEY 90]). PROPOSITION 3.1.– If α ∈ ]0, 1[ we thus have the following characterization: f ∈ C α () ⇐⇒ ∃C > 0 :
∀j, k
|df (k, j)| C 2−αj
Proof. Let us assume that f ∈ C α (); then, ∀j, k, we have: j j j f (x)ψ(2 x − k) dx = 2 f (x) − f (k 2−j ) ψ(2j x − k) dx ef (k, j) = 2 (because the wavelets are of zero integral) thus: C |ef (k, j)| C 2j |x − k 2−j |α dx C 2−αj (1 + |2j x − k|)2 (by the change of variable t = 2j x − k). Let us now prove the converse. Let us assume that we have |ef (k, j)| C 2−αj . Let j0 be defined by 2−j0 −1 |x − x0 | < 2−j0 and note that: fj (x) =
ef (k, j) ψ(2j x − k)
k
From the localization assumption of ψ, we deduce that we obtain: |fj (x)| C
k
2−αj C 2−αj (1 + |2j x − k|)2
Wavelet Methods for Multifractal Analysis of Functions
109
and similarly, using the localization of ψ , we deduce that we have |fj (x)| C 2(1−α)j . We obtain: |f (x) − f (x0 )| |fj (x) − fj (x0 )| + |fj (x)| + |fj (x0 )| j>j0
jj0
j>j0
Using finite increments, the first term is bounded by: |x − x0 | sup |fj (t)| C|x − x0 | 2(1−α)j C|x − x0 |2(1−α)j0 C jj0
[x,x0 ]
jj0
(because we have α < 1). Coming back to the definition of j0 , we see that the first term is bounded by C|x − x0 |α . The second and the third terms are bounded by: 2−αj C 2−αj0 C|x − x0 |α j>j0
thus, the converse estimate holds. The reader will easily be able to extend this result to a case where we have α > 1 and α ∈ N; see [MEY 90] in the case of α ∈ N. If a function f belongs to one of the C α spaces, for α > 0, we will say that f is uniformly Hölderian. The following theorem is similar to Proposition 3.1, but gives a result of pointwise regularity. THEOREM 3.1.– Let α ∈ ]0, 1[. If f is C α (x0 ), then we have: |ef (k, j)| C 2−αj (1 + |2j x0 − k|α )
(3.2)
conversely, if the wavelet coefficients of f verify (3.2) and if f is uniformly Hölderian, then, if we have |x − x0 | 1, we obtain: 2 α (3.3) |f (x) − f (x0 )| C|x − x0 | log |x − x0 | (f is “nearly” C α (x0 ) if we make a small logarithmic correction). Proof. Let us assume that f is C α (x0 ); then we have: j j j f (x)ψ(2 x − k) dx = 2 f (x) − f (x0 ) ψ(2j x − k) dx df (k, j) = 2 (because the wavelets are of zero integral). Thus, |df (k, j)| is bounded by: C|x − x0 |α |x − k 2−j |α + |k 2−j − x0 |α j j 2 dx 2 C 2 dx (1 + |2j x − k|)2 (1 + |2j x − k|)2
110
Scaling, Fractals and Wavelets
(because, for a, b > 0, we have (a + b)α 2aα + 2bα ). By once again changing the variable t = 2j x − k, we obtain |df (k, j)| C 2−αj (1 + |2j x0 − k|α ). Let us now prove the converse. Assuming that there exists an > 0 such that f ∈ C (), let j0 and j1 be defined by: 2−j0 −1 |x − x0 | < 2−j0
and
j1 =
α j0
From (3.2), we deduce that, for all x, we have: |ef (k, j)| C(2−αj + |x0 − k 2−j |α ) 2 C(2−αj + |x − x0 |α + |x − k 2−j |α ) and thus: |fj (x)| C
2−αj + |x − x0 |α + 2−αj |2j x − k|α (1 + |2j x − k|)2 k
We have: k
1 C (1 + |2j x − k|)2
and similarly: k
1 C (1 + |2j x − k|)2−α
because we have α < 1. We get: |fj (x)| C 2−αj + |x − x0 |α C 2−αj (1 + 2αj |x − x0 |α ) Using the localization of ψ , we obtain, in the same manner: |fj (x)| C 2(1−α)j (1 + 2αj |x − x0 |α ) in particular, if we have j j0 , we obtain |fj (x)| C2(1−α)j . We can write: |f (x) − f (x0 )|
jj0
|fj (x) − fj (x0 )| +
j>j0
|fj (x)| +
j>j0
|fj (x0 )|
(3.4)
Wavelet Methods for Multifractal Analysis of Functions
111
The first term is bounded by:
C|x − x0 |
sup |fj (t)| C|x − x0 |
jj0 [x,x0 ]
2(1−α)j
jj0
C|x − x0 |2(1−α)j0 C|x − x0 |α As far as the second term is concerned, if we have j > j0 , bound (3.4) becomes |fj (x)| C|x − x0 |α and we thus have:
|fj (x)|
j0 <j<j1
|x − x0 |α
j0 <j<j1
C(j1 − j0 )|x − x0 |α C|x − x0 |α log
2 |x − x0 |
because we have j1 − j0 (α/)j0 . Moreover, since f belongs to C (), we have:
|fj (x)|
jj1
2− j C 2− j1 C|x − x0 |α
jj1
(because of the choice of j1 ). As far as the third term is concerned, the bound (3.4) in x = x0 becomes |fj (x0 )| C 2−αj and thus:
|fj (x)|
j>j0
2−αj C|x − x0 |α
jj0
and the converse is proved in Theorem 3.1. Here again, the reader will easily be able to extend this result to exponents α > 1 (see [JAF 91]). As far as the second part of the theorem is concerned, we sometimes use the slight variation below. PROPOSITION 3.2.– Let α ∈ ]0, 1[. If the wavelet coefficients of f verify:
|df (k, j)| C 2−αj (1 + |2j x0 − k|α ) for an α < α, then f is C α (x0 ).
(3.5)
112
Scaling, Fractals and Wavelets
The demonstration of this proposition is very similar to that of the second part of the theorem and therefore we only outline it. We show this time that (3.5) implies that:
|fj (x)| C 2−αj (1 + 2α j |x − x0 |α ) and:
|fj (x)| C 2(1−α)j (1 + 2α j |x − x0 |α ) and we finish by using the estimation on fj for j j0 and on fj for j j0 . This proposition also extends to the case where α > 1 (see [JAF 91]). NOTE 3.2.– We could get the impression that Proposition 3.2, contrary to Theorem 3.1, does not make the assumption of global regularity. This does not matter because, if (3.5) is verified and if we have |x − x0 | 1, then we can deduce that |df (k, j)| C 2−(α−α )j , i.e. that we have f ∈ C α−α uniformly close to x0 . We could nevertheless refer to [JAF 00c] where pointwise Hölder estimations are obtained from wavelet coefficients, in the absence of any assumption of uniform regularity. From the theorem, we can immediately deduce the following corollary that characterizes the Hölder exponent by local decay of wavelet coefficients. COROLLARY 3.1.– If f is uniformly Hölder, the Hölder exponent of f at every point x0 is given by: log(|df (k, j)|) (3.6) hf (x0 ) = lim inf inf j→∞ k log(2−j + |k 2−j − x0 |) This corollary is used in all mathematical results where the singularity spectrum of a function f is derived from its wavelet coefficients. On the other hand, it can be used to obtain the numerical value of hf (x) only if function hf is constant, or if it varies slightly with the discretization scale of the signal. 3.2.3. Local oscillations Using the Hölder exponent as a means to measure pointwise regularity makes it a powerful signal and image analysis tool (see [DAO 95, LEV 95]). However, there are many disadvantages in characterizing the pointwise regularity by using only the Hölder exponent: – inability to measure the oscillatory character of the local behavior of f close to x0 ; – lack of stability while using “traditional” operators, such as differential operators, pseudo-differential operators or the Hilbert H transform (convolution
Wavelet Methods for Multifractal Analysis of Functions
113
operator with one dimension as x1 , which allows the transfer of the real signal to the associated analytic signal during signal analysis). So, we can create, for instance, functions f locally bounded with f ∈ C α (x0 ) and Hf ∈ C β (x0 ) for any β > 0 (see [JAF 91]).
spaces will provide a substitute to the pointwise regularity The 2-microlocal Cxs,s 0 notion which does not have any disadvantages; moreover, they are defined even if f is not locally bounded. We have already seen this condition in (3.5).
space if its DEFINITION 3.5.– An F distribution belongs to the 2-microlocal Cxs,s 0 wavelet coefficients satisfy, for sufficiently small |x0 − k 2−j |:
∃C > 0 |ef (k, j)| C 2−j(s+s ) (2−j + |k 2−j − x0 |)−s
(3.7)
This condition does not depend on the chosen wavelet base (sufficiently regular) (see [JAF 91]). DEFINITION 3.6.– The 2-microlocal domain of f in x0 , noted by E(f (x0 )) is the set of couples (s, s ) such that f ∈ C s,s (x0 ). By interpolation between conditions (3.7), we find that the 2-microlocal domain is a convex set and, moreover, by using the trivial lower bound 2−j +|k 2−j −x0 | 2−j , we can see that its boundary is a curve whose slope is everywhere larger than −1. Conversely, we can check that these conditions characterize the boundaries of the 2-microlocal domains in one point x0 (see [GUI 98, MEY 98]) (a more difficult and unresolved issue is to determine which are the compatibility conditions between the different 2-microlocal domains E(f (x)) of a function when the point x varies). The 2-microlocal domain provides very accurate information on the behavior of f close to x0 ; specifically, we can derive the Hölder exponent of primitives from it or fractional derivatives of f (see [ARN 97]). However, this complete information is superfluous when put into practice. In fact, we need to preserve few parameters – and not a convex function at each point (the boundary of the 2-microlocal domain): at least the Hölder exponent, and often a second parameter β, which will measure the oscillating character of f close to x0 . In fact, a similar Hölder exponent can, at any given point, show very different behavior, such as “cusps” |x − x0 |H or, on the contrary, high oscillating functions, such as “chirps”: 1 (3.8) gH,β (x) = |x − x0 |H sin |x − x0 |β for β > 0. In signal processing, the chirp notion models functions whose instantaneous frequency increases rapidly at a given time (see [CHA 99]). The β exponent measures the speed at which the instant frequency (3.8) diverges at x0 .
114
Scaling, Fractals and Wavelets
An additional motivation for the analysis of chirps is that this type of behavior incurs a failure in the first versions of the multifractal formalism, as we will see in section 3.4. We presently have two possible mathematical definitions for exponent β in a general framework. Of course, both result in the same value for functions (3.8). We will define them and discuss their respective advantages with regard to the analysis of the signal. Let f be locally bounded and let us note by f (−n) a primitive n times iterated from f . As shown by a sequence of integrations by parts, a consequence of the oscillations (−n) of (3.8) close to x0 is that gH,β belongs to C H+n(β+1) (x0 ) (the increase of the Hölder exponent in x0 is not 1 at each integration, as expected for an arbitrary function, but β + 1). This observation has led to the following definition given by Meyer [JAF 96b]. DEFINITION 3.7.– Let H 0 and β > 0. A function f ∈ L∞ () is a chirp (n) of type (H, β) in x0 if, for any n 0, f can be written as f = gn , with H+n(1+β) (x0 ). gn ∈ C The chirp type can be derived from the 2-microlocal domain in x0 . In fact, if we have f ∈ C H (x0 ), we also have f ∈ C H,−H (x0 ), etc. and, in general, (n) the condition f = gn with gn ∈ C H+n(1+β) (x0 ) implies that we have f ∈ H+nβ,−H−n(β+1) (x0 ). The other characterization of chirps, given below (see C [JAF 96b]), shows that their definition reflects appropriately the oscillatory phenomenon present in functions gH,β . PROPOSITION 3.3.– A function f ∈ L∞ is a chirp of type (H, β) in x0 if and only if a function r(x), C ∞ exists close to x0 and > 0 such that, if we have 0 < x < , then: f (x) = r(x − x0 ) + (x − x0 )H g+ (x − x0 )−β and if we have − < x < 0, then: f (x) = r(x − x0 ) + |x − x0 |H g− |x − x0 |−β since functions g+ and g− are permanently oscillating, i.e. they have bounded primitives of any order. It is very easy to check that the interior of the set of points (f, β) such that f is a chirp of the type (H, β) in x0 is always of the form H < hf (x0 ), β < βf (x0 ) [JAF 00a]. The positive number βf (x0 ) is called the chirp exponent in x0 .
Wavelet Methods for Multifractal Analysis of Functions
115
A highly oscillating local behavior such as (3.8) is notable and it was believed for a long time that this could be observed only in isolated points. This is why the Meyer result, proving that the Riemann series n−2 sin(πn2 x) has a dense set of chirps of type ( 32 , 1), was unexpected (see [JAF 96b, JAF 01a]). Since then, we know how to generate functions with chirps almost everywhere (see [JAF 00a]). Definition 3.7 has not been adapted to the analysis of signals. In fact, we will see that it is not stable by adding an arbitrary regular function (but not C ∞ ). We will also introduce another definition of the local exponent to measure the oscillation without this disadvantage. Let us consider the following example. Let B(x) be a Brownian motion and: 1 1/3 + B(x) (3.9) C(x) = x sin x The Hölder exponent of B(x) being everywhere 12 , the largest singularity at 0 is the chirp x1/3 sin( x1 ) and we effectively observe this oscillating behavior in the graph of C(x) expanded around 0. However, after integration, the random term becomes paramount (in fact, an integration by parts shows that the first term is O(|x|7/3 ) whereas a Brownian primitive has the Hölder exponent hC (x) = 32 everywhere); the oscillating behavior then disappears on the primitive. The oscillations of the graph of C(x), which do not exist in the primitive, are not taken into account in Definition 3.7: we see that C(x) is not a chirp at 0 (it is actually a chirp of exponents ( 13 , 0)). Nevertheless, this big oscillating behavior should be taken into account in the chirp exponent: C(x) “should” have exponents ( 13 , 1). Let us now show how to define an oscillating exponent taking the value β for (3.8), which would not be changed by adding a “regular noise” (it will take value 1 for C(x) and no longer 0). It is obvious, from the previous example, that the oscillating exponent should not be determined by taking into account primitives of F . An “infinitesimal” fractional integration should be used so that the order of importance of terms in (3.9) is not disturbed. Let ht (x0 ) be the Hölder exponent of a fractional integral of order t of the function f at x0 . To be more precise, if f is a locally bounded function, let us note by ht (x0 ) the Hölder exponent in x0 of: I t (f ) = (Id − Δ)−t/2 (φf )
(3.10)
where φ is a C ∞ function with compact support such that φ(x0 ) = 1 and (Id −Δ)−t/2 is the convolution operator which, in Fourier, is none other than multiplication by (1 + |ξ|2 )−t/2 (see Chapter 7 for more details). In the example of function C(x), we find, where x0 = 0, ht (x0 ) = 13 + 2t after a fractional integration of quite small order t, i.e., 13 + 2t < 12 + t. In this example, the increase of the Hölder exponent in x0 after a fractional integration of a quite small order t is 2t; we can therefore recover the β
116
Scaling, Fractals and Wavelets
oscillating exponent in this way. In general, the function t → ht (x0 ) is concave, so much so that its derivative at the right exists in 0 (with the possible value +∞). The following definition is derived from it. DEFINITION 3.8.– Let f : d → be a locally bounded function. The oscillating exponent of f in x0 is: ∂ −1 (3.11) β = ht (x0 ) ∂t t=0 This exponent belongs to [0, +∞]. We will notice that if we have ht (x0 ) = +∞, then β is not defined. The following proposition, extracted from [AUB 99], shows that this definition appropriately reflects the oscillating phenomenon present in (3.8). PROPOSITION 3.4.– If f is uniformly Hölderian, ∀H < h(x0 ) and ∀β < β(x0 ), f can be written as: 1 H + r(x) f (x) = |x − x0 | g |x − x0 |β with r(x) ∈ C α (x0 ) for a α > h(x0 ) and g is infinitely oscillating. 3.2.4. Complements We discuss here the known results relating to the construction of functions having a prescribed Hölder exponent (and possibly oscillating). This problem was first encountered in the speech simulation context. A speech signal possesses a Hölder exponent which varies drastically (particularly in the case of consonants) and this led to the idea of efficiently storing such a signal by keeping only the information contained in the Hölder exponent. The following theorem characterizes functions which are Hölder exponents (see [AYA 08]). THEOREM 3.2.– A positive function h(x) is the Hölder exponent of a bounded function f if and only if h can be written as lim inf of a sequence of continuous functions. When h(x) has a minimal Hölder regularity, a natural construction is provided by the multifractional Brownian motion (see [BEN 97, PEL 96]). Contrary to the previous result, a couple of functions (h(x), β(x)) should verify very specific conditions to be a couple (Hölder exponent, chirp exponent): β(x) should vanish on a dense set [GUI 98]. We also have a constructive result: if this condition is satisfied, we can prescribe this couple almost everywhere (see [JAF 05]) but unfortunately we do not know how to characterize couples which are couples of the form (Hölder exponent, chirp exponent).
Wavelet Methods for Multifractal Analysis of Functions
117
3.3. Random multifractal processes We present two examples of random multifractal processes. The first is the Lévy process. Their importance is derived, on the one hand, from the central place these processes have in determining the probability factor and, on the other hand, their increasing importance for physics or financial modeling, particularly in situations where Gaussian models are inadequate (see, for instance, [MAN 97, SCH 95] and Chapter 5 and Chapter 6). Our second example involves random wavelet series, that is, processes whose wavelet coefficients are independent and, at a particular scale, have the same laws (these laws can be set in an arbitrary manner at each scale). In addition to the intrinsic advantages of this model, it also makes it possible to introduce new concepts and it enriches the possible variants of the multifractal formalism. 3.3.1. Lévy processes A Lévy process Xt (t 0) with values in can be defined as a process with stationary independent expansions: Xt+s − Xt is independent on (Xv )0vt and has the same law as Xs . The function that characterizes the Lévy process is written as E(eiλXt ) = e−tφ(λ) , where: 2 1 − eiλx + iλx1|x|<1 π(dx) (3.12) φ(λ) = iaλ + Cλ +
where π(dx) is the Lévy measure of Xt , i.e. a positive Radon measure on − {0} verifying: inf(1, |x|2 ) π(dx) < ∞ (3.13) Or:
Cj =
2−j |x|2·2−j
π(dx)
We can measure the size of π close to the origin with the help of the “inferior” exponent of Blumenthal and Getoor, defined by: log Cj α = inf α 0 : |x| π(dx) < ∞ α = sup 0, lim sup j→∞ j log 2 |x|1 This exponent satisfies that 0 α 2 and, if Xt is a stable Lévy process, coincides with the stability index.
118
Scaling, Fractals and Wavelets
In general, the Lévy process has a dense set of discontinuities (it seems to skip everywhere!) and can be written as a superposition of particularly simple processes: the compensated composed Poisson processes, whose sample paths are piecewise linear with jumps. The sample paths of a composed Poisson process of Lévy measure π(dx) (which is finite) are generated in the following manner: X(t) remains at zero for 0 #t < t1 ; t1 is the time of the first jump, whose law is exponential of intensity C = π(dx) (i.e., the law of the first jump has Ce−Ct as density). In t1 , the process jumps, the amplitude of the jump is an independent random variable of the value taken by t1 and the probability measure of the jump is π(dx)/C. Then, we start again: if t2 is the second jump time, t2 − t1 does not depend on t1 and the jump value in t1 ; the t2 − t1 law is the same as t1 and, finally, the jump in t2 has the same law as in t1 and does not depend on the previous choices, etc. Thus, we can generate a process X(t) with independent stationary increments# which is piecewise constant; this is a composed Poisson process. If we have D = xπ(dx), the expectation of Y (t) = X(t) − Dt is zero for all t; this is the compensated composed Poisson process with measure π(dx). If a Lévy measure π(dx) verifies (3.13), then each measure π0 (dx) = 1|x|<1 π(dx) and πj (dx) = 12−j |x|2·2−j π(dx) is bounded and is the Lévy measure of a compensated composed Poisson process Xj (t). The Lévy process (without Brownian component) associated with π is then: X(t) =
+∞
Xj (t)
j=0
If X(t) has a Brownian component, we can add to the process that we have just constructed a Brownian motion originating from 0. The following theorem confirms that the Lévy process sample paths are multifractal. THEOREM 3.3.– Let Xt be a Lévy process without the Brownian component (C = 0 in (3.12)) which verifies: & (3.14) β > 0 and 2−j Cj log(1 + Cj ) < ∞ With probability 1, the singularity spectrum of Xt is: αH if we have H ∈ [0, 1/α] dα (H) = −∞ otherwise NOTE 3.3.– We notice that: – conditions (3.14) are verified as soon as we have 0 < α < 2 and particularly for all stable Lévy processes;
Wavelet Methods for Multifractal Analysis of Functions
119
– Theorem 3.3 can seem to infer the reverse of what we expect: we normally think of multifractal functions as functions whose “behavior” can vary greatly, which seems to contradict the stationary expansion hypothesis of Lévy processes. The paradox disappears if we remember that this hypothesis is related to the law of a Lévy process and thus it authorizes one particular trajectory presenting such “changes of behavior”. Thus, the Eh sets are, for a given trajectory, remarkable points, but the law of Eh remains invariant per translation; – the assertion in Theorem 3.3 is stronger than when we simply state that fH (H) has a given value with probability 1, which would not be enough in order to determine the singularity spectrum of almost any trajectory; – the Hölder exponent of a Lévy process without the Brownian component is α almost everywhere (see [PRU 81]), which, of course, complies with Theorem 3.3 (case H = α); – a lot of work has been devoted to the fractal nature of the set of values taken by a Lévy process (see [MAN 95] for “Lévy flights” and [BER 96] and its references). Let us now outline the demonstration of the theorem. The Lévy process Xt is written as the sum of a composed Poisson process of Lévy measure 1|x|>1 π(dx) (and we can “forget” this first process whose addition does not modify the spectrum) and the series j0 Xtj where Xtj are compensated and independent composed Poisson processes with Lévy measure πj (dx) = 12−j |x|<2·2−j π(dx) (the amplitude of jumps of Xj is therefore of the order of magnitude of Z −j ). The following comment enables us to bound the regularity of Xt . Let f be a function having a dense set of discontinuities and in each discontinuity having a limit on the right and on the left. Let rn be a sequence of discontinuities of f converging towards x. Then, the Hölder exponent of f verifies that: hX (x) lim inf
n→+∞
log|f (rn+ ) − f (rn− )| . log|rn − x|
(3.15)
The proof follows from the simple comment that one of the two numbers |f (x) − f (rn+ )| or |f (x) − f (rn− )| should be larger than |f (rn+ ) − f (rn− )|/2 due to the triangular inequality. The Lévy processes are of this type and thus we can apply this comment to them. We note by Ajδ the union of intervals of diameter 2−δj centered at the jumps of Xtj and Eδ = lim sup Ajδ ; if we have t ∈ Eδ there is a sequence jn of integers and instants tjn such that: |tjn − t| 2−δjn
and
|Xt+ − Xt− | 2−jn jn
jn
By applying (3.15), we then obtain: hX (t) 1/δ
(3.16)
120
Scaling, Fractals and Wavelets
We check that the accumulation of jumps close to t is the only cause of irregularity in a Lévy process and that we actually have the equality in (3.16) (see [JAF 99]). The proof of the theorem therefore consists of showing that the dimension of Eδ is α/δ if we have α δ. The dimension bound of Eδ is immediate: the average number of jumps of Xtj on [0, 1] is Cj and thus, it is almost sure that all Xtj have less than 2Cj jumps starting from a certain rank. Hence, Ajδ covered by 2Cj intervals of diameter 2−δj . The bound dimension bound is derived from it. The upper dimension bound is carried out using a traditional technique, by creating particular measures which “load” the Eδ sets as uniformly as possible (see [JAF 99]). We will also refer to [JAF 99] if Xt has a Brownian component. 3.3.2. Burgers’ equation and Brownian motion Multifractal analysis has been introduced within the frame of turbulence; however, mathematical equations governing the speed evolution of a turbulent flow (Navier-Stokes equations within the limit where the viscosity tends towards 0) are at present very little understood and hope for a mathematical result regarding the multifractal nature of the solutions is not in sight. At the most, we can anticipate precise results only for more simple non-linear evolution equations, hoping that they retain some physical characteristics of the fluid evolution. The simplest equation proposed for this objective is Burgers’ equation: ∂ u2 ∂2u ∂u + = ν 2 , x ∈ , t ∈ + ∂t ∂x 2 ∂t
(3.17)
which we shall consider within the limit of small viscosities ν → 0. One reason to study this equation is that we have explicit formulae providing the solution u(x, t) at any given moment t > 0 according to the initial condition u0 (x) = u(x, 0); in fact, if U is a primitive of u, then U verifies: 2 1 ∂U ∂2U ∂U + =ν 2 ∂t 2 ∂x ∂x Then, we carry out the Cole-Hopf transformation, which consists of supposing that φ = e−U/2ν ; φ then verifies the equation of the linear heat, which we explicitly resolve, hence the expression of U . By passing to the limit within this expression when ν → 0, we obtain using a standard technique (Laplace method) the following result (see [EVA 98] for the details of the calculations). Let us suppose, to simplify things, that the initial condition u0 is zero on ] − ∞, 0[ and that we have u0 (s) + s 0 for s large enough. To have an idea, let us have a look
Wavelet Methods for Multifractal Analysis of Functions
121
at the solution at this instant: t = 1. First, we consider for each x 0, the function of the variable s 0: s u0 (r) + r − x dr Fx (s) = 0
and we note by a(x, 1) the largest point where the minimum is attained: a(x, 1) = max{s 0 : Fx (s) Fx (s ); for all s }
(3.18)
The limit solution when ν → 0 with the time t = 1 is then given by: u(x, 1) = x − a(x, 1)
(3.19)
Formula (3.18) shows that the random process a(x, 1) obtained when the initial condition is a Brownian motion on + (and zero on − ) is a subordinator. We call a subordinator an increasing Lévy process (σx , x 0). A traditional example, which plays an important role later, is that of the first times of passage of a Brownian motion with derivative. More specifically, let us consider a real standard Brownian motion (Bs , s 0), let us note Xs = Bs + s and let us introduce for x 0: τx = inf{s 0, Xs > x} Because τx is a stopping time and since Xτx = x, the strong Markov property applied to the Brownian motion implies that the process Xs = Xs+τx − x is also a Brownian motion with derivative, which is independent of the portion of the trajectory before τx , (Xr , 0 r τx ). For any z ∈ [0, x], the first time of passage τz clearly only depends on the trajectory before τx and hence (Xs , s 0) is independent of (τz , 0 z x). The identification: τx+y − τx = inf{s 0, Xs+τx > x + y} = inf{s 0, Xs > y} then highlights the independence and the homogenity of the incrementation of τ , which is hence a subordinator. Close arguments apply to the increments of function (3.18) for Burgers’ equation (3.17) non-viscous with Brownian initial condition, i.e., when we have u(x, 0) = Bx for x 0. Indeed, we verify that (a(x, 1) − a(0, 1), x 0) and (τx , x 0) follow the same rule (see [BER 98]). The isolated example of the Burgers’ equation with initial Brownian data can make us hope that more general results are true – and in particular that large classes of non-linear partial differential equations generically develop multifractal solutions. However, at present, there are no proven results of this type (however, the reader can consult [VER 94] concerning Burgers’ equation in several space dimensions).
122
Scaling, Fractals and Wavelets
3.3.3. Random wavelet series Since we are interested in the local properties of the functions, it is equivalent, and also easier, to work with periodic wavelets that are obtained by periodization of a usual base of wavelets (see [MEY 90]) and are defined on the toric T = /Z. The periodic wavelets: ψ 2j (x − l) − k , j ∈ N, 0 k < 2j (3.20) ψj,k (x) = 2j/2 l∈Z
form an orthonormal base of L2 (T) (by adding the constant function equal to 1, see [MEY 90]; we use the same notation than for the wavelets on , which will not lead to any confusion). We also assume that ψ has enough regularity and zero moments. Any periodic function f is hence written as follows: ef (k, j) 2−j/2 ψj,k (x) (3.21) f (x) = j,k
where the wavelet coefficients of f are hence given by: 1 2j/2 ψj,k (t)f (t) dt ef (k, j) = 0
We assume that all the coefficients are independent and have the same law at each scale. Let ρj be the measure of common probability of 2j random variables Xj,k = −(log2 |df (k, j)|)/j (signs of wavelet coefficients do not have any influence on the Hölder regularity, which is why we are not making any assumptions on this subject). Thus, the measure ρj verifies: P |ef (k, j)| 2−aj = ρj (−∞, a] We will make the following assumption on ρj : ∃ > 0 :
supp(ρj ) ⊂ [, +∞]
It signifies that the sample paths of the process are uniformly Hölder. We need to define the logarithmic density ρ˜(α) of the coefficients; i.e.: log2 2j ρj ([α − , α + ]) ρ¯(α) = lim lim sup →0 j→+∞ j The reader should note that this density is, in fact, a spectrum of large deviation, but calculated from the wavelet coefficients (see Chapter 4 where general results concerning spectra of large deviation are established). Then, we note: ρ¯(α) if ρ¯(α) 0 ρ˜(α) = 0 otherwise
Wavelet Methods for Multifractal Analysis of Functions
123
or:
Hmax =
sup α>0
ρ˜(α) α
−1
THEOREM 3.4.– Let f be a random wavelet series verifying a uniform regularity assumption. The singularity spectrum of almost any trajectory of f is supported by [, Hmax ] and, within this interval: ρ˜(α) α∈[0,H] α
fH (H) = H sup
(3.22)
Function ρ˜(α) being essentially arbitrary, we notice that a singularity spectrum of a series of random wavelets is not necessarily concave. 3.4. Multifractal formalisms Even if the singularity spectrum of numerous mathematical functions can be determined by simply using Definition 3.4 with regard to the multifractal signals, it is not practical to determine their regularity at each point. This is because the Hölder exponent can be discontinuous everywhere, and it is even less practical to calculate the infinity of the corresponding Hausdorff dimensions! Frisch and Parisi have introduced a formula which allows us to deduce the singularity spectrum of a signal from quantities that are easily measurable. It is the first example of what we now call multifractal formalisms. Several variants have been proposed since then; we shall describe some of them and we shall compare their respective performances. Additional results are found in Chapter 4. The formula proposed by Frisch and Parisi is based on the knowledge of the Besov spaces to which the function belongs. This is why we begin with a few reminders regarding these spaces and their characterization by wavelets. 3.4.1. Besov spaces and lacunarity One of the reasons for the success of wavelet decomposition in applications is that they often provide representations of signals that are very lacunary (few coefficients are numerically not negligible). This lacunarity is often quantified by determining to which Besov spaces the considered function belongs to. For the characterization of the wavelet coefficients of Besov spaces see [MEY 90]: ∀s ∈ , p > 0, f ∈ B s,p () ⇐⇒
j,k
1
|ef (k, j) 2(s− p )j |p
1/p < +∞
(3.23)
124
Scaling, Fractals and Wavelets
When we have p 1, the Besov spaces are very close to the Sobolev spaces; indeed, if Lp,s is the space of the functions of Lp for which the fractional derivatives of order s still belong to Lp , we have the following injections: ∀ > 0,
∀p 1,
Lp,s+ → B s,p → Lp,s−
However, a determining advantage over Sobolev spaces is that the Besov spaces are defined for any p > 0. It is precisely these spaces for p close to 0 that enable us to measure the lacunarity of the representation in wavelets of f . We illustrate this point with an example. Let us consider the function:
H(x) =
1 if |x| 1 0 otherwise
and let us assume that the wavelet chosen has a compact support, with the interval [−A, A]. Because ψ is a zero integral, for each j, there is less than 4A non-zero wavelet coefficients, so much so that the decomposition in wavelets of f is very lacunary. Because H(x) is bounded, we have |df (k, j)| C for any j, k. By using (3.23), H(x) belongs to B s,p () as soon as we have s < 1/p. Let us show, at the same time, that such an assertion is a way to quantify the fact that the decomposition in wavelets of f is very lacunary. Let us suppose that f is a bounded function satisfying: ∀p > 0,
∀s <
1 , p
f ∈ B s,p ()
We will verify that for any D > 0 and for any > 0, at each scale j, there are −Dj . Indeed, if this was not the less than C(, D)2 j coefficients of a size larger than 2 case, by taking p = /(2D), we would obtain k |df (k, j)|p → +∞ when j → +∞, which is a contradiction. Here is another illustration of the relation between the lacunarity of the decomposition in wavelets and the Besov regularity. We assume that f belongs to ∩p>0 B 1/p,p . Coming back to (3.23), we observe that this condition means exactly that the sequence df (k, j) belongs to lp for any p > 0. Let us then note by dn the rearrangement in a decreasing order of the sequence of wavelet coefficient modules |df (k, j)|; hence, the sequence dn also belongs to lp for any p > 0. Thus: ∀p
∃Cp
such that
∞ n=1
dpn Cp
Wavelet Methods for Multifractal Analysis of Functions
125
Because the sequence dn is decreasing: ∀N,
N dpN
N n=1
dpn
∞
dpn Cp
n=1
and thus dN (Cp )1/p N −1/p . Since we can take p arbitrarily close to 0, we observe that the rearrangement in a decreasing order of the sequence |df (k, j)| has fast decay, which is, once again, a way to express the lacunarity (the converse is immediate: if the sequence dn has fast decay, it belongs to all lp and thus this is also the case for the sequence df (k, j)). The Besov space for p < 1 is not locally convex, which partly explains the difficulties in their utilization. Before the introduction of wavelets, these spaces were characterized either by order of approximation of f with rational fractions for which the numerator and the denominator have a fixed degree, or by an order of approximation with the splines “with free nodes” (which means that we are free to choose the points where the polynomials in parts are connected) (see [DEV 98, JAF 01a]). However, these characterizations are difficult to handle and hence do not have any real numerical applications. Characterization (3.23) shows that the knowledge of the Besov spaces to which f belongs is clearly linked to the asymptotic behavior (when j → +∞) of the moments of distribution of the wavelet coefficients of f ; see (3.26). Generally, more information is available; indeed, these moments are deduced from the histogram of the coefficients at each scale j. This is why it is normal to wonder which information regarding the pointwise regularity of f can be deduced from the knowledge of these histograms. We present a study of this problem below. We observe that the cascade type models for the evolution of the repartition function of wavelet coefficients through the scales have been proposed to model the speed of turbulent flows [ARN 98]. To start with, let us point out a limitation of the multifractal analysis: functions having the same histograms of wavelet coefficients at each scale can have singularity spectra that are completely different [JAF 97a]. In the multifractal analysis, it is not only the histogram of the coefficients which is important, but also their positions. This is why no formula deducing the singularity spectrum from the knowledge of the histograms can be valid in general. However, we can hope that some formulae are “more valid than others”. Indeed, we have observed that if the coefficient values are independent random variables, there is a spectrum which is almost sure; we shall notice that the formula that yields this spectrum differs from the formulae proposed until now. Another approach consists of specifying the information on the function and considering the functional spaces that take the positions of the large wavelet coefficients into consideration (see [JAF 05]).
126
Scaling, Fractals and Wavelets
3.4.2. Construction of formalisms The construction of a multifractal formalism can be based on two types of considerations: – counting arguments: we consider the increments (or wavelet coefficients) having a certain size; we estimate their number and deduce their contribution to some “calculable” quantities; – more mathematical arguments: we prove that a bound of the spectrum, according to the “calculable” quantities, is generally true and that this bound is “generically” an equality. The term “generically” is to be understood in the sense of “Baire classes” if the information at the start is of functional type, or as “almost sure” if the information at the start is a probability. We begin by describing the first approach; we do not exactly recapitulate the initial argument of Frisch and Parisi, but rather its “translation” in wavelets, as found in [ARN 95b, JAF 97a]. This approach admits two variants, according to the information on the function that we have. Indeed, we can start from: – the partition function τ (q) defined from knowledge of the Besov spaces to which f belongs: q log k |ef (k, j)| s/q,q } = 1 + lim inf τ (q) = sup{s : f ∈ B j→+∞ log 2−j – histograms of wavelet coefficients. Generally, for each j, let Nj (α) = #{k : |df (k, j)| 2−αj }. Thus, we have E(Nj (α)) = 2j ρj ([0, a]). If: log Nj (α + ) − Nj (α − ) ρ(α, ) = lim sup log(2j ) j→∞
(3.24)
then, we define: ρ(α) = inf ρ(α, ) >0
(3.25)
˜ (there is an order of 2ρ(α)j coefficients of size ∼ 2−αj ).
It is important to note that the information provided by ρ(α) is richer than that provided by τ (q); indeed, τ (q) can be deduced from the histograms with: −1 −j −αqj 2 log2 2 Nj (α) dα (3.26) τ (q) = lim inf j→+∞ j
Wavelet Methods for Multifractal Analysis of Functions
because, by definition of Nj , we have deduce that:
k |df
(k, j)|q =
#
τ (q) = inf αq − ρ(α) + 1 α0
127
2−αqj dNj (α). It is easy to (3.27)
On the other hand, we cannot reconstitute ρ(α) from τ (q); indeed, it is clear from (3.27) that the two functions ρ(α) and ρ (α) having the same concave envelope lead to the same function τ (q). Based on τ (q), we can thus only obtain the envelope of ρ(α), by carrying out a Legendre transformation once again. We will now describe the heuristic arguments based on the construction of these multifractal formalisms. We will divide them into four steps, highlighting the implicit assumptions that we make for each of them. S TEP 1. The first assumption, common to both approaches, is that the Hölder exponent at each point x0 is given by the order of magnitude of the wavelet coefficients of f in a cone |k2−j − x0 | C2−j . With respect to (3.6), if they are decreasing as 2−Hj , we then have hf (x0 ) = H. This assumption is verified if f does not have “cusp” type singularities ([MEY 98]), i.e., the oscillation exponent is zero everywhere. If we go from the data of ρ(H), we have, as an assumption, 2ρ(H)j wavelet coefficients of size 2−Hj . By using the supports of the corresponding wavelets to cover EH , we expect to obtain fH (H) = ρ(H). Thus, we also obtain a first form of multifractal formalism: the formalism said to be “of large deviation”, which simply affirms that: fH (H) = ρ(H) Let us briefly justify this name as well as that of the large deviation spectrum that we sometimes give to function ρ. The theory of large deviations takes care of the calculation of the probabilities, which are so small that we can only correctly estimate them on logarithmic scales. The basic example is as follows: if Xi are n independent reduced centered Gaussian, and if we have S˜n = n1 i=1 Xi , then we obtain n1 log P(|S˜n | δ) ∼ −δ 2 /2. The analogy with (3.24) and (3.25) is striking since, in the common law of wavelet coefficients, the parts of very small probabilities, which we measure with the help of a logarithmic scale, are those that provide the relevant information ρ(α) for calculating the spectrum (see Chapter 4 for a more detailed study). The effective calculation of function ρ(α) is numerically delicate because its definition leads to a double limit, which generally results into problems said to be “of finite size”. In theory, we must go “completely” to the limit in j in (3.24) before taking the limit in in (3.25). Practically speaking, the two limits must effectively be taken “together”. The problem is then to know how to take j sufficiently large according to , which creates significant numerical stability problems. In any case, a
128
Scaling, Fractals and Wavelets
calculation of ρ(α) which is numerically reliable requires us to know the signal on a large number of scales, i.e., with an excellent precision. This is why we often prefer to work from averages, such as k |df (k, j)|q , i.e., finally, based on the partition function for which the definition leads to only one limit. From now, this is the point of view that we shall adopt (however, let us note that the direct method introduced by Chhabra and Jensen in [CHH 89] is a method for calculating ρ(α) without going through a double limit, or through a Legendre transformation; we shall find a mathematical discussion of this method adapted to the framework of the wavelets in [JAF 04a]). S TEP 2. We will estimate, for each H, the contribution of Hölder singularities of exponent H at: |ef (k, j)|q (3.28) k
Each singularity of this type brings a contribution of C2−Hqj and there must be ∼ 2fH (H)j intervals of length 2−j to recover these singularities; the total contribution of the Hölder singularities of exponent H at (3.28) is thus as follows: 2fH (H)j 2−Hqj = 2−(Hq−fH (H))j
(3.29)
This is a critical step of reasoning; it contains an inversion of limits that implicitly assumes that all Hölder singularities have coefficients ∼ 2−Hj simultaneously from a certain scale J and that Hausdorff dimension can be estimated as if it were a box dimension. It is notable that the multifractal formalism leads to the correct singularity spectrum in several situations where these two assumptions are not verified. S TEP 3. The third step is an argument of the “Laplace method” type. When j → +∞, we note that, among the (3.29) terms, the one that brings the main contribution to (3.28) is that for which the exponent H carries out the minimum of Hq − fH (H), from which comes the heuristic formula: τ (q) − 1 = inf hq − fH (H) H
S TEP 4. If fH (H) is concave, then −fH (H) and −τ (q) + 1 are conjugate convex functions and each of them can be deduced from the other by a Legendre transformation. Thus, if we define the Legendre spectrum with: fL (H) = inf Hq − τ (q) + 1 q
we deduce a first formulation of the multifractal formalism: fH (H) = fL (H) = inf Hq − τ (q) + 1 q
(3.30)
Wavelet Methods for Multifractal Analysis of Functions
129
The assumption on which the concavity fH (H) depends has no need to be verified. We can then: – stop at step 3 and only affirm that τ (q) is the Legendre transformation of fH (H). However, this lower form of multifractal formalism is of little interest because the quantity that we would like to calculate is fH (H) and the quantity that we know is generally τ (q); – affirm that (3.30) provides, in fact, the concave spectrum envelope. This information is, however, particularly ambiguous when the function calculated in this manner contains segments at the right (it is often the case, see [JAF 97b, JAF 99]): do they correspond to effective values of the spectrum or only to its envelope in a region where it is not concave? 3.5. Bounds of the spectrum We will obtain bounds of fH (H) valid in general. Once again, we have here two points of view, depending on whether the information that we have concerns the functional spaces to which f belongs or histograms of the wavelet coefficients. We can make an observation before any calculation: since ρ(α) contains more information than τ (q), we expect that the bounds obtained from ρ(α) will be better; we shall see that this is indeed the case. 3.5.1. Bounds according to the Besov domain We start by bounding the singularity spectrum of functions belonging to a fixed Besov space. PROPOSITION 3.5.– Let s > 1/p, α ∈ [s − p1 , s] and d = 1 − p(s − α). Then, for any function f ∈ B s,p , all the points x0 where f ∈ C s−( d-dimensional measure equal to zero.
1−d p )
(x0 ) have a Hausdorff
Proof. Let s > 0 and p > 0. Let us note: ef (k, j) = df (k, j) 2j(s−1/p) With respect to (3.23), the condition f ∈ B s,p can then be rewritten: |ef (k, j)|p < ∞
(3.31)
j,k
Let us note by Ij,k the interval centered in k 2−j and length |ef (k, j)|p/d . We can now rewrite (3.31): ! "d diam(Ij,k ) < ∞
130
Scaling, Fractals and Wavelets
For all J,(Ij,k )jJ form a covering of all the points belonging to an infinity of intervals Ij,k , so much so that the Hausdorff d-dimensional measure of this set is zero. If a point x belongs to more than a finite number of intervals Ij,k , there exists J (= J(x)) such that: k ∀j J, ∀k j − x |ef (k, j)|p/d 2 and thus: d/p k 1 1 d |ef (k, j)| j − x 2−(s− p )j = 2−(s− p + p )j |2j x − k|d/p 2 1
d
Since we have s − 1/p > 0, Proposition 3.2 implies that f ∈ C s− p + p (x), from which we obtain Proposition 3.5, since: s−
1 d + = α. p p
NOTE 3.4.– The condition s > 1/p is necessary because there are functions of B 1/p,p that are nowhere locally bounded, (see [JAF 00c]). We will now deduce from Proposition 3.5 a bound of the spectrum fH (H). Let us assume that τ (q) is known. Let q and > 0 be fixed. By definition of τ (q), for any > 0, f belongs to B (τ (q)− )/q,q . We can then apply Proposition 3.5 for all q such that: ∃ > 0 :
1 τ (q) − > q q
which is equivalent to τ (q) > 1. If f is uniformly Hölder, τ (q) is increasing and continuous, and verifies: lim τ (q) = 0 and
q→0
τ (q) → +∞
when
q → +∞
There is a unique value qc such that τ (qc ) = 1 and, for any q > qc , we thus have: τ (q) − −H fH (H) 1 − q q thus: fH (H) qH − τ (q) + 1 Since this result is true for any q > qc , we have shown the following proposition.
Wavelet Methods for Multifractal Analysis of Functions
131
PROPOSITION 3.6.– If f is uniformly Hölder, the singularity spectrum of f verifies the bound: (3.32) fH (H) inf qH − τ (q) + 1 q>qc
The quasi-sure results that we now describe show that bound (3.32) is optimal. Proposition 3.6 suggests that, in formula (3.30), the domain of q on which the Legendre transform must be calculated is the interval [qc , +∞). Let us assume that τ (q) is a function of partition admissible, i.e., it is the partition function of a function f uniformly Hölder (this will be the case if s(q) = qτ (1/q) is concave and verifies 0 s (q) 1 and s(0) > 0; see [JAF 00b]). To say that f has as a partition function τ (q) implies, by definition, that f belongs to the space V (= Vτ (q) ) defined by: B (τ (q)− )/q,q (3.33) V = >0,q>0
Space V is a Baire space, i.e. it has the following property: any intersection of countable dense open subsets is dense. In a Baire space, a property which is true at least on one intersection countable dense open subsets is said to be quasi-sure. Hence, it is natural to wonder if the multifractal formalism occurs almost surely in V . The following theorem, taken from [JAF 00b], answers this question. THEOREM 3.5.– Let τ (q) be obtainable and V the space defined by (3.33). The definition domain of the singularity spectrum of almost any function of V is the interval [s(0), 1/qc ] and, on this interval, we have: (3.34) fH (H) = inf Hq − τ (q) + 1 qqc
Formula (3.34) affirms that the singularity spectrum of almost any function is made up of two parts: if we have H < τ (qc ), the infimum in (3.34) is reached for q > qc and the spectrum can be calculated using the “usual” Legendre transformation of τ (q): fH (H) = inf Hq − τ (q) + 1 q>0
If we have τ (qc ) H 1/qc , the infimum in (3.34) is reached for q = qc and the spectrum is the segment at the right fH (H) = Hqc . The study of regularity properties that are almost sure derives from Banach (see [BAN 31]). Buczolich and Nagy have shown in [BUC 01] that almost any monotone function is multifractal of spectrum fH (H) = H for H ∈ [0, 1].
132
Scaling, Fractals and Wavelets
3.5.2. Bounds deduced from histograms The following proposition provides the optimal of the Hölder spectrum, which can be deduced in general from the histograms of wavelet coefficients [AUB 00]. PROPOSITION 3.7.– If we have f ∈ C () for a > 0, then: fH (H) H sup α∈[0,H]
ρ(α) α
(3.35)
This becomes an equality for random wavelet series. Indeed, they verify ρ˜(α) = ρ(α), which shows that this bound is optimal. We can easily verify that it implies (3.32). However, (3.35) clearly provides a better bound if ρ(α) is not concave. Once again, we see that the histogram of the coefficients strictly contains more “useful” information than the partition function. We will note that, although (3.35) is more precise than (3.32), the fact still remains that (3.32) is optimal, when the only information available is the partition function (as shown by almost sure results). The optimal bounds (3.35) and (3.32) can propose variants of the multifractal formalism. We say that the almost sure multifractal formalism is verified if (3.35) is saturated, i.e. if: ρ(α) α∈[0,H] α
fH (H) = H sup
(3.36)
and the multifractal formalism almost sure is verified if (3.35) is saturated, i.e. if: (3.37) fH (H) = inf qH − τ (q) + 1 qqc
3.6. The grand-canonical multifractal formalism The aim of the grand-canonical multifractal formalism is to calculate the spectrum of oscillating singularities d(H, β) which, by definition, provides the Hausdorff dimension of all the points where the Hölder exponent is H and the oscillation exponent β. This formalism is based on new functional spaces. To define them, we will use the more geometric notations that follow: λ and λ will designate the dyadic intervals λj,k = k 2−j + [0, 2−j ] and λj ,k = k 2−j + [0, 2−j ] respectively, Cλ will designate the coefficient df (k, j) and ψλ the wavelet ψ(2j x − k).
DEFINITION 3.9.– Let p > 0 and s, s ∈ . A function f belongs to à Ops,s () if its wavelet coefficients satisfy: 1/p sj s j p sup |Cλ 2 | <∞ (3.38) sup 2 j∈Z
k
λ ⊂λ
Wavelet Methods for Multifractal Analysis of Functions
133
This definition is independent of the base of wavelets chosen [JAF 98]. We now derive a grand-canonical multifractal formalism, which will enable us to obtain d(H, β) from the knowledge of the wavelet coefficients of f . This formalism is based on a reasoning similar to that of section 3.4. It is motivated by two preoccupations: obtaining information that is more complete for the Hölder singularities of the signal and explicitly taking into consideration the behaviors of chirp type, which eliminates one of the causes of failure in the multifractal formalism. Indeed, we have seen, in step 1 of section 3.4.2, that the previous multifractal formalisms assume that the signal does not contain any chirps. Let: p s j log sup |C | 2 λ ⊂λ λ k ζ(p, s ) = lim inf j→+∞ log 2−j
= sup s : f ∈ Ops/p,s /p If f has a chirp with exponents (H, β) in x0 , its wavelet coefficients are of the order of magnitude of |k 2−j − x0 |H close to the curve 2−j ∼ |k 2−j − x0 |1+β . For each pair (H, β), we shall estimate the contribution of the chirp of exponents (H, β) to the quantity: sup |Cλ |p 2s j . (3.39) λ∈Λj
λ ⊂λ
Let us consider an interval λ of side 2−j that contains a chirp with exponent (H, β). The wavelet coefficients Cλ for λ ⊂ λ are negligible as long as we have 2−j (2−j )1+β , i.e., as long as we have j j(1 + β). When j ∼ j(1 + β), we have, for certain values of k : |Cλ | ∼ (2−j )H ∼ 2−j and hence:
H 1+β
Hp sup |Cλ |p 2s j ∼ 2−j ( 1+β −s ) ∼ 2−j(Hp−(1+β)s )
λ ⊂λ
(as long as we have s pH/(1 + β)). The contribution of the chirp of exponents (H, β) at (3.39) is thus:
2d(H,β)j 2−j(Hp−(1+β)s ) = 2−j(Hp−(1+β)s −d(H,β)) When j → +∞, the main contribution is provided by the pair (H, β) for which the inf of Hp − (1 + β)s − d(H, β) is attained, from which we obtain the heuristic formula: ζ(s , p) = inf Hp − (1 + β)s − d(H, β) H,β
134
Scaling, Fractals and Wavelets
If d(H, β) is a convex function, then: d(H, β) = inf Hp − (1 + β)s − ζ(s , p) s ,p
(3.40)
If we are interested in the Hölder spectrum fH (H), we can deduce it from the following argument: for a fixed H, the value of β which brings the largest contribution to (3.40) will provide the correct dimension fH (H), hence the formula: ' ( (1 + β)s fH (H) = sup d(H, β) = sup inf + Hp − ζ(s , p) (3.41) β
β
s ,p
Of course, this formula is not at all equivalent to the standard multifractal formalisms, as we can easily verify in the example of lacunary wavelet series. 3.7. Bibliography [ABR 98] A BRY P., V EITCH D., “Wavelet analysis of long-range-dependent traffic”, EEE Trans. Inf. Theory, vol. 44, no. 1, p. 2–15, 1998. [AMA 00] A MANN A., M AYR G., S TROHMENGER H.U., “N (α) histogram analysis of the ventricular fibrilation ECG-signal as predictor of countershock success”, Chaos, Solitons, and Fractals, vol. 11, p. 1205–1212, 2000. [ARN 95a] A RNEODO A., A RGOUL F., BACRY E., E LEZGARAY J., M UZY J.F., Ondelettes, multifractales et turbulence : de l’ADFN aux croissances cristallines, Diderot, 1995. [ARN 95b] A RNEODO A., BACRY E., M UZY J.F., “The thermodynamics of fractals revisited with wavelets”, Physica A, vol. 213, p. 232–275, 1995. [ARN 97] A RNEODO A., BACRY E., JAFFARD S., M UZY J.F., “Oscillating singularities on Cantor sets: A grand canonical multifractal formalism”, Journal of Statistical Physics, vol. 87, p. 179–209, 1997. [ARN 98] A RNEODO A., BACRY E., M UZY J.F., “Random cascades on wavelet dyadic trees”, Journal of Mathematical Physics, vol. 39, no. 8, p. 4142–4164, 1998. [AUB 99] AUBRY J.M., “Representation of the singularities of a function”, Appl. Comput. Harmon. Anal., vol. 6, no. 2, p. 282–286, 1999. [AUB 00] AUBRY J.M., JAFFARD S., “Random wavelet series”, Comm. Math. Phys., vol. 227, p. 483–514, 2002. [AYA 08] AYACHE A., JAFFARD S., Hölder Exponents of Arbitrary Functions (preprint), 2008. [BAN 31] BANACH S., Über die Baire’sche Kategorie gewisser Funktionenmengen, Studia Math., vol. 3, p. 174–179, 1931. [BEN 97] B ENASSI A., JAFFARD S., ROUX D., “Elliptic Gaussian random processes”, Revista Mathematica Iberoamericana, vol. 13, no. 1, p. 19–90, 1997. [BER 96] B ERTOIN J., An Introduction to Lévy Processes, Cambridge University Press, 1996.
Wavelet Methods for Multifractal Analysis of Functions
135
[BER 98] B ERTOIN J., “The inviscid Burgers equation with Brownian initial velocity”, Communications in Mathematical Physics, vol. 193, no. 2, p. 397–406, 1998. [BIS 98] B ISWAS M.K., G HOSE T., G UHA S., B ISWAS P.K., “Fractal dimension estimation for texture images: A parallel approach”, Pattern Recognition Letters, vol. 19, no. 3–4, p. 309–313, 1998. [BUC 01] B UCZOLICH Z., NAGY J., “Hölder spectrum of typical monotone continuous functions”, Real Anal. Exch., vol. 26, no. 1, p. 133–156, 2000–2001. [CHA 99] C HASSANDE -M OTTIN E., F LANDRIN P., “On the time-frequency detection of chirps”, Appl. Comput. Harmon. Anal., vol. 6, no. 2, p. 252–281, 1999. [CHH 89] C HHABRA A.B., J ENSEN R.V., “Direct determination of the f (α) singularity spectrum”, Physical Review Letters, vol. 62, p. 1327–1330, 1989. [DAO 95] DAOUDI K., L ÉVY V ÉHEL J., “Speech signal modeling based on local regularity analysis”, in IASTED/IEEE, International Conference on Signal and Image Processing (SIP’95, Las Vegas, New Mexico), 1995. [DEV 98] D E VORE R., “Nonlinear approximation”, Acta Numerica, p. 1–99, 1998. [DUB 89] D UBUC B., Z UCKER S.W., T RICOT C., Q UINIOU J.F., W EHBI D., “Evaluating the fractal dimension of surfaces”, Proceedings of the Royal Society of London A, vol. 425, p. 113–127, 1989. [EVA 98] E VANS L.C., “Partial differential equations”, American Mathematical Society, Graduate studies in mathematics, 1998. [FRI 95] F RISCH U., Turbulence, Cambridge University Press, 1995. [GAG 87] G AGNE Y., Etude expérimentale de l’intermittence et des singularités dans le plan complexe en turbulence développée, PhD Thesis, Grenoble University, 1987. [GIN 00] G INCHEV I., ROCCA M., “On Peano and Riemann derivatives”, Rend. Circ. Mat. Pal. Ser. 2, vol. 49, p. 463–480, 2000. [GUI 98] G UIHENEUF B., JAFFARD S., L ÉVY V ÉHEL J., “Two results concerning chirps and 2-microlocal exponents prescription”, Appl. Comp. Harm. Anal., vol. 5, no. 4, p. 487–492, 1998. [JAF 91] JAFFARD S., “Pointwise smoothness, two-microlocalization coefficients”, Publicacions Mathematiques, vol. 35, p. 155–168, 1991.
and
wavelet
[JAF 96a] JAFFARD S., “The spectrum of singularities of Riemann’s function”, Revista Mathematica Iberoamericana, vol. 12, no. 2, p. 441–460, 1996. [JAF 96b] JAFFARD S., M EYER Y., “Wavelet methods for pointwise regularity and local oscillations of functions”, Memoirs of the AMS, vol. 123, no. 587, 1996. [JAF 97a] JAFFARD S., “Multifractal formalism for functions”, SIAM Journal of Mathematical Analysis, vol. 28, no. 4, p. 944–998, 1997. [JAF 97b] JAFFARD S., “Old friends revisited. The multifractal nature of some classical functions”, J. Four. Anal. App., vol. 3, no. 1, p. 1–22, 1997.
136
Scaling, Fractals and Wavelets
[JAF 97c] JAFFARD S., M ANDELBROT B., “Peano-Polya motion, when time is intrinsic or binomial (uniform or multifractal)”, The Mathematical Intelligencer, vol. 19, no. 4, p. 21–26, 1997. [JAF 98] JAFFARD S., “Oscillation spaces: Properties and applications to fractal and multifractal functions”, Journal of Mathematical Physics, vol. 39, no. 8, p. 4129–4141, 1998. [JAF 99] JAFFARD S., “The multifractal nature of Lévy processes”, Probability Theory and Related Fields, vol. 114, no. 2, p. 207–227, 1999. [JAF 00a] JAFFARD S., “On lacunary wavelet series”, Ann. Appl. Proba., vol. 10, no. 1, p. 313–329, 2000. [JAF 00b] JAFFARD S., “On the Frisch-Parisi conjecture”, J. Math. Pures Appl., vol. 76, no. 6, p. 525–552, 2000. [JAF 00c] JAFFARD S., M EYER Y., “On the pointwise regularity of functions in critical Besov spaces”, J. Funct. Anal., vol. 175, p. 415–434, 2000. [JAF 01a] JAFFARD S., “Functions with prescribed Hölder and chirps exponents”, Revista Mathematica Iberoamericana, 2001. [JAF 01b] JAFFARD S., M EYER Y., RYAN R., Wavelets: Tools for Science and Technology, SIAM, 2001. [JAF 04a] JAFFARD S., “Beyond Besov spaces part 1: Distributions of wavelet coefficients”, J. Four. Anal. Appl., vol. 10, no. 3, p. 221–246, 2004. [JAF 04b] JAFFARD S., “Wavelet techniques in multifractal analysis”, in L APIDUS M., VAN F RANKENHUIJSEN M. (Eds.), Fractal Geometry and Applications: A Jubilee of Benoît Mandelbrot, Proceedings of Symposia in Pure Mathematics, AMS, vol. 72, Part 2, p. 91–152, 2004. [JAF 05] JAFFARD S., “Beyond Besov spaces part 2: Oscillation spaces”, J. Constr. Approx., vol. 21, no. 1, p. 29–61, 2005. [JAF 06] JAFFARD S., L ASHERMES B., A BRY P., “Wavelet leaders in multifractal analysis”, in Q IAN T. et al. (Eds.), Wavelet Analysis and Applications, Birkhäuser, vol. 72, Part 2, p. 219–264, 2006. [JAF NIC] JAFFARD S., N ICOLAY S., “Pointwise smoothness of space-filling functions”, to appear in Appl. Comp. Harmon. Anal. [LAS 08] L ASHERMES S., ROUX S., A BRY P., JAFFARD S., “Comprehensive multifractal analysis of turbulent velocity using wavelet leaders”, Eur. Phys. J. B., vol. 61, no. 2, p. 201–215, 2008. [LEV 95] L ÉVY V ÉHEL J., “Fractal approaches in signal processing”, Fractals: Symposium in Honor of Benoît Mandelbrot (Curaçao), vol. 3, no. 4, p. 755–775, 1995. [LEV 97] L ÉVY V ÉHEL J., R IEDI R., “Fractional Brownian motion and data traffic modeling: the other end of the spectrum”, in L ÉVY V ÉHEL J., L UTTON E., T RICOT C. (Eds.), Fractals in Engineering, Springer-Verlag, 1997. [MAN 95] M ANDELBROT B., Les Objets Fractals, Flammarion, 1995.
Wavelet Methods for Multifractal Analysis of Functions
137
[MAN 97] M ANDELBROT B., Fractals and Scalings in Finance, Springer, 1997. [MAN 98] M ANDELBROT B., Multifractals and 1/f Noise, Springer, 1998. [MEY 90] M EYER Y., Ondelettes et opérateurs, Hermann, 1990. [MEY 98] M EYER Y., “Wavelets, vibrations, and scalings”, in CRM Series AMS, University of Montreal Press, vol. 9, 1998. [PEL 96] P ELTIER R., L ÉVY V ÉHEL J., “Multifractional Brownian Motion: Definitions and Preliminary Results”, Technical Report, 1996. [PRU 81] P RUITT W., “The growth of random walks and Lévy processes”, Annals of Probability, vol. 9, no. 6, p. 948–956, 1981. [SCH 95] S CHLESINGER, Z ASLAVSKY, F RISCH (Eds.), Lévy Flights and Related Topics (Proceedings of the Nice Workshop, 1994), Springer-Verlag, 1995. [TAQ 97] TAQQU M., T EVEROVSKY V., W ILLINGER W., “Is network traffic self-similar or multifractal?”, Fractals, vol. 5, no. 1, p. 63–73, 1997. [TRI 97] T RICOT C., “Function norms and fractal dimension”, SIAM J. Math. Anal., vol. 28, p. 189–212, 1997. [VER 94] V ERGASOLA M., D UBRULLE B., F RISCH U., N OULLEZ A., “Burgers’ equation, devil’s staircase and the mass distribution for large-scale structures”, Astronomy and Astrophysics, vol. 289, p. 325–356, 1994. [WEN 07] W ENDT H., A BRY P., JAFFARD S., “Bootstrap for empirical multifractal analysis”, IEEE Signal Processing Magazine, vol. 24, no. 4, p. 38–48, 2007. [WIL 96] W ILLINGER W., TAQQU M., E RRAMILLI A., “A bibliographical guide to self-similar traffic and performance modeling for modern high-speed networks”, in K ELLY F.P. et al. (Eds.), Stochastic Networks: Theory and Applications (Selected Papers of the Royal Statistical Society Research Workshop, August 1995), Royal Statistical Society Lecture Notes Series, vol. 4, Clarendon Press, Oxford, p. 339–366, 1996. [WIL 00] W ILLINGER W., R IEDI R., TAQQU M., “Long-range dependence and data network traffic”, in D OUKHAN O., TAQQU M. (Eds.), Long-range Dependence: Theory and Applications, Birkhäuser, 2000.
This page intentionally left blank
Chapter 4
Multifractal Scaling: General Theory and Approach by Wavelets
4.1. Introduction and summary Fractal processes have been successfully applied in various fields such as the theory of fully developed turbulence [MAN 74, FRI 85, BAC 93], stock market modeling [EVE 95, MAN 97, MAN 99], and more recently in the study of network data traffic [LEL 94, NOR 94]. In networking, models using fractional Brownian motion (FBM) have helped advance the field through their ability to capture fractal features such as statistical self-similarity and long-range dependence (LRD). It has been recognized, however, that multifractal features need to be accounted for further, so as to gain a better understanding of network traffic, but also of stock exchange [RIE 97a, RIE 99, RIE 00, FEL 98, MAN 97]. In short, there is a call for more versatile models which can, for example, incorporate LRD and multifractal properties independently of each other. Roughly speaking, a fractal entity is characterized by the inherent, ubiquitous occurrence of irregularities, which govern its shape and complexity. The most prominent example is certainly FBM BH (t) [MAN 68]. Its paths are almost surely continuous but not differentiable. Indeed, the oscillation of FBM in any interval of size δ is of the order δ H where H ∈ (0, 1) is the self-similarity parameter: fd
BH (at) = aH BH (t).
Chapter written by Rudolf R IEDI.
(4.1)
140
Scaling, Fractals and Wavelets
Real world signals, on the other hand, often possess an erratically changing oscillation exponent, limiting the appropriateness of FBM as a model. Due to the various exponents present in such signals, they have been termed multifractals. This chapter’s main objective is to present the framework for describing and detecting such a multifractal scaling structure. In doing so we survey local and global multifractal analysis and relate them via the multifractal formalism in a stochastic setting. Thereby, the importance of higher order statistics will become evident. It might be especially appealing to the reader to see wavelets put to novel use. We focus mainly on the analytical computation of the so-called multifractal spectra, and on their mutual relations, dwelling extensively on variations of binomial cascades. Statistical properties of estimators of multifractal quantities, as well as modeling issues, are addressed elsewhere (see [GON 98, ABR 00, GON 99, MAN 97, RIE 99, RIB 06]). The remainder of this introduction provides a summary of the contents of the paper, roughly following its structure. 4.2. Singularity exponents For simplicity we consider processes Y over a probability space (Ω, F, PΩ ) and defined on a compact interval, which we assume without loss of generality to be [0, 1]. Generalization to higher dimensions is straightforward and extending this to processes defined on is simple and will be indicated. 4.2.1. Hölder continuity The erratic behavior or, more precisely, degree of local Hölder regularity of a continuous process Y (t) at a fixed given time t can be characterized to a first approximation by comparison with an algebraic function: Y is said to be in Cth if there is a polynomial Pt such that |Y (s) − Pt (s)| ≤ C|s − t|h for s sufficiently close to t. If Pt is a constant, i.e. Pt (s) = Y (t) for all s, then Y is in Cth for all h < h(t) and not in Cth for all h > h(t) where h(t) := lim inf ε→0
1 log sup |Y (s) − Y (t)|. log2 (2ε) 2 |s−t|<ε
(4.2)
On the other hand, it easy to prove the following LEMMA 4.1.– If h(t) ∈ / N then Pt is a constant, and h(t) = sup{h : Y ∈ Cth }. As the example Y (s) = s2 + s2.4 with t = 0 shows, the conclusion does not necessarily hold when h(t) ∈ N. Here, |Y (s) − Y (0)| ∼ s2 for s ∼ 0, thus h(0) = 2, while P0 (s) = s2 , Y (s) − P0 (s) = s2.4 , and thus sup{h : Y ∈ C0h } = 2.4.
Multifractal Scaling: General Theory and Approach by Wavelets
141
Proof. Assume there is h > h(t) and Pt (s) such that Y ∈ Cth . We will argue that h(t) must be an integer in this case. Note first that Pt is not constant, by definition of h(t), and we may write Pt (s) = Y (t) + (s − t)m · Q(s) for some integer m ≥ 1 and some polynomial Q without zero at t. Assume first that m < h(t) and choose h such that m < h < h(t). Writing Y (s) − Pt (s) = (Y (s) − Y (t)) − (Pt (s) − Y (t)), the first term is smaller than |s − t|h and the second term, decaying as C|s − t|m , governs. Whence h = m < h(t), against the assumption. Assuming m > h(t), choose h such that m > h > h(t) and a sequence sn such that |Y (sn ) − Y (t)| ≥ |sn − t|h , whence |Y (sn ) − Pt (sn )| ≥ (1/2)|sn − t|h for large n and h ≤ h . Letting h → h(t) we again obtain a contradiction. We conclude that h(t) equals m. For reasons of symmetry we define h(t) := lim sup ε→0
1 log sup |Y (s) − Y (t)|. log2 (2ε) 2 |s−t|<ε
(4.3)
If h(t) and h(t) coincide we denote the common value by h(t). We note first that the continuous limit in (4.2) may be replaced by a discrete limit. To this end we introduce kn (t) := t2n , an integer defined uniquely by t ∈ Iknn := [kn 2−n , (kn + 1)2−n [.
(4.4)
As n increases the intervals Ikn form a nested decreasing sequence (compare Figure 4.1). Provided n is chosen such that 2−n+1 ≤ ε < 2−n+2 we have [(kn − 1)2−n , (kn + 2)2−n [ ⊂ [t + ε, t − ε[ ⊂ [(kn−2 − 1)2−n+2 , (kn−2 + 2)2−n+2 [ from which it follows immediately that h(t) = lim inf hnkn n→∞
h(t) = lim sup hnkn n→∞
where hnkn := −
1 log2 sup |Y (s) − Y (t)| : s ∈ [(kn − 1)2−n , (kn + 2)2−n [ . n
(4.5)
It is essential to note that the countable set of numbers hnkn contains all the scaling information of interest to us. Being defined pathwise, they are random variables.
142
Scaling, Fractals and Wavelets
4.2.2. Scaling of wavelet coefficients A convenient tool for scaling analysis is found in the wavelet transform, both in its discrete and continuous forms. The discrete transform, for example, allows to represent a 1D process Y (t) in terms of shifted and dilated versions of a prototype bandpass wavelet function ψ(t), and shifted versions of a low-pass scaling function φ(t) [DAU 92, VET 95]. While such representations exist also in the framework of continuous wavelet transforms, we use the latter mainly as a “microscope” in this chapter. In the vocabulary of Hilbert spaces, the discrete wavelet and scaling functions ψj,k (t) := 2j/2 ψ 2j t − k , φj,k (t) := 2j/2 φ 2j t − k , j, k integer (4.6) form an orthonormal basis and we have the representations [DAU 92, VET 95] Y (t) =
DJ0 ,k φJ0 ,k (t) +
∞ j=J0
k
Cj,k ψj,k (t),
(4.7)
k
with Cj,k :=
∗ Y (t) ψj,k (t) dt,
Dj,k :=
Y (t) φ∗j,k (t) dt.
(4.8)
The wavelet coefficient Cj,k measures the signal content around time 2−j k and frequency 2j f0 , provided that the wavelet ψ(t) is centered at time zero and frequency f0 . The scaling coefficient Dj,k measures the local mean around time 2−j k. In the wavelet transform, j indexes the scale of analysis: J0 can be chosen freely and indicates the coarsest scale or lowest resolution available in the representation. The most simple example of an orthonormal wavelet basis are the Haar scaling and wavelet functions (see Figure 4.1a). Here, φ is the indicator function of the unit interval, while ψ = φ(2·) − φ(2 · −1). For a process supported on the unit interval a convenient choice is thus J0 = 0. The supports of the fine-scale scaling functions nest inside the supports of those at coarser scales; this can be neatly represented by the binary tree structure of Figure 4.1b. Row (scale) j of this scaling coefficient tree contains an approximation to Y (t) of resolution 2−j . Row j of the complementary wavelet coefficient tree (not shown) contains the details in scale j + 1 of the scaling coefficient tree that are suppressed in scale j. In fact, for the Haar wavelet we have Dj,k Cj,k
= =
2−1/2 (Dj+1,2k + Dj+1,2k+1 ), 2−1/2 (Dj+1,2k − Dj+1,2k+1 ).
(4.9)
Wavelet decompositions contain considerable information on the singularity behavior of a process Y . Indeed, adapting the argument of [JAF 95, p. 291] and
Multifractal Scaling: General Theory and Approach by Wavelets
2
Dj,k0
I j,k(t)
j/2
k’=0 0
0
k2
-j
k’=1 0
(k+1)2 -j
Dj+1,2k0 +1
Dj+1,2k
0
2
\ (t)
j/2
0
k’=0 1
j,k
k2 -j
143
k’=1 1
Dj+2,4k
0
(k+1)2 -j
(a)
k’=0 1
Dj+2,4k
0
+1
k’=1 1
Dj+2,4k
0
+2
Dj+2,4k0 +3
(b)
Figure 4.1. (a) The Haar scaling and wavelet functions φj,k (t) and ψj,k (t). (b) Binary tree of scaling coefficients from coarse to fine scales
correcting wavelet normalization used here for the L2 – as opposed to L1 in [JAF 95] – it is easily shown that |Y (s) − Y (t)| = O(|s − t|h ) implies that (4.10) 2n/2 |Cn,kn | = O 2−nh . This holds for any h > 0 and any compactly supported wavelet. As a matter of fact # only ψ = 0 is needed to obtain this result since the Taylor polynomial of Y is implicitly assumed to be constant. If we are interested in an analysis only, we may thus consider analyzing wavelets ψ such as derivatives of the Gaussian exp(−x2 ) which do not necessarily form a basis. To distinguish them from the orthogonal wavelets we will call them “analyzing wavelets”. In order to invert (4.10), however, we need representation (4.7) as well as some knowledge concerning the decay of the maximum of the wavelet coefficients in the vicinity of t and sufficient wavelet regularity. For a precise statement, see [JAF 95] and [DAU 92, Theorem 9.2]. All this suggests that replacing hnk (4.5) by the left hand side of (4.10) would produce an alternative description of the local behavior of Y . Consequently, we set w(t) := lim inf wknn n→∞
w(t) := lim sup wknn
(4.11)
n→∞
where wknn := −
1 log2 2n/2 Cn,kn . n
(4.12)
If w(t) and w(t) coincide we denote the common value by w(t). Using wavelets has the advantage of yielding an analysis which is largely # unaffected by polynomial trends in Y due to vanishing moments tm ψ(t)dt = 0 which are typically built into wavelets [DAU 92]. In this context recall Lemma 4.1. It has the disadvantage of complicating the analysis since the maxima of wavelet coefficients have to be considered for a reliable estimation of true Hölder continuity
144
Scaling, Fractals and Wavelets
[JAF 95, DAU 92, JAF 97, BAC 93]. In any case, the decay of wavelet coefficients is interesting in itself as it relates to LRD (compare [ABR 95]) and regularity spaces such as Besov spaces [RIE 99]. 4.2.3. Other scaling exponents Traditional multifractal analysis of a singular measure μ on the line constitutes a study of the singularity structure of its primitive M given by t μ(ds) = μ([0, t]), (4.13) M(t) = 0
Since M is an almost surely increasing process, the coarse exponents hnkn (see (4.5)) simplifying to hnkn = − n1 log2 |M((kn + 2)2−n ) − M((kn − 1)2−n )|, we are motivated (some would say “seduced”) to study an even simpler notion of a coarse exponent: α(t) := lim inf αknn n→∞
α(t) := lim sup αknn
(4.14)
n→∞
where αknn := −
1 1 log2 |M (kn + 1)2−n − M(kn 2−n )| = − log2 μ Iknn . n n
(4.15)
If α(t) and α(t) coincide we denote the common value by α(t). This exponent α(t) has attracted considerable attention in the multifractal community, with its potential due to its simplicity. In [LV 98] various examples of more general exponents were introduced, all of which are so-called Choquet capacities, a notion which is not needed to develop the multifractal formalism. As an interesting alternative, [PEY 98] considers an arbitrary function ξ(I) from the space of all intervals to + (instead of only the Ikn ) and develops a multifractal formalism similar to ours. There, it is suggested that we consider the oscillations of Y around the mean, i.e. # Y (s)ds (4.16) ξ(I) := Y (t) − I dt. |I| I Proceeding as with hn (t), we are led to the singularity exponent −(1/n) log2 (ξ(Ikn )) which is of particular interest since it can be used to define oscillation spaces, such as Sobolev spaces and Besov spaces. Another useful choice consists of interpolating Y in the interval I by the linear function aI + bI t and considering 1/2 2 (Y (t) − (aI + bI t)) dt . (4.17) ξ(I) := I
Multifractal Scaling: General Theory and Approach by Wavelets
145
This exponent measures the variability of Y and is related to the dimension of the paths of Y . Deducting constant, resp. linear terms in formulae (4.16) and (4.17) reminds us of the use of wavelets with one, resp. two vanishing moments. 4.3. Multifractal analysis Multifractal analysis has been discovered and developed in [MAN 74, FRI 85, KAH 76, GRA 83, HEN 83, HAL 86, CUT 86, CAW 92, BRO 92, BAC 93, MAN 90b, HOL 92, FAL 94, OLS 94, ARB 96, JAF 97, PES 97, RIE 95a, MAN 02, BAC 03, BAR 04, BAR 02, CHA 05, JAF 99] to give only a short list of some relevant work done in this area. The main insight consisted of the fact that local scaling exponents on fractals as measured by h(t), α(t) or w(t) are not uniform or continuous as a function of t, in general. In other words, h(t), α(t) and w(t) typically change in an erratic way as a function of t, thus imprinting a rich structure on the object of interest. This structure can be captured either in geometric terms, making use of the concept of dimensions, or in statistical terms based on sample moments. A useful connection between these two descriptions emerges from the multifractal formalism. As we will see, as far as the multifractal formalism is concerned there is no restriction in choosing a singularity exponent which seems fit for describing scaling behavior of interest. To express this fact we consider in this section the arbitrary scaling exponents s(t) := lim inf snkn n→∞
and
s(t) := lim sup snkn ,
(4.18)
n→∞
where snk (k = 0, . . . , 2n − 1, n ∈ N) is any sequence of random variables. To keep a connection with what was said before, think of snk as representing a coarse scaling exponent of Y over the dyadic interval Ikn . 4.3.1. Dimension based spectra A geometric description of the erratic behavior of a multifractal’s scaling exponents can be achieved using a quantification of the prevalence of particular exponents in terms of fractal dimensions as follows: We consider the sets Ka which are defined pathwise in terms of limiting behavior of snkn as n → ∞, as Ea := {t : s(t) = a},
E a := {t : s(t) = a},
Ka := {t : s(t) = a}
(4.19)
These sets Ka are typically “fractal”, meaning loosely that they have a complicated geometric structure and more precisely that their dimensions are non-integer. A compact description of the singularity structure of Y is therefore in terms of the following so-called Hausdorff spectrum d(a) := dim(Ka ), where dim(E) denotes the Hausdorff dimension of the set E [TRI 82].
(4.20)
146
Scaling, Fractals and Wavelets
The sets Ea (a ∈ ) – and also E a – form a multifractal decomposition of the support of Y . We will loosely address Y as a multifractal if this decomposition is rich, i.e. if the sets Ea (a ∈ ) are highly interwoven or even dense in the support of Y . However, the study of singular measures (deterministic and random) has often been restricted to the simpler sets Ka and their spectrum d(a) [KAH 76, CAW 92, FAL 94, ARB 96, OLS 94, RIE 98, RIE 95a, RIE 95b, BAR 97]. With the theory developed here (Lemma 4.2) it becomes clear that most of these results extend to provide formulae for dim(Ea ) and dim(E a ) as well. This aspect of multifractal analysis has been of much interest to the mathematical community. 4.3.2. Grain based spectra An alternative description of the prevalence of singularity exponents, statistical in nature due to the counting involved, is f (a) := lim lim sup ε↓0
n→∞
1 log2 N n (a, ε), n
(4.21)
where1 N n (a, ε) := #{k = 0, . . . , 2n − 1 : a − ε ≤ snk < a + ε}.
(4.22)
This notion has grown out of the difficulty faced by any real world application, that the calculation of actual Hausdorff dimensions is often hard, if not impossible. Using a mesh of given grain size as in (4.22) instead of arbitrary coverings as in dim(Ka ) leads generally to more simple notions. However, f should not be regarded as an auxiliary vehicle but recognized for its own merit, which will become apparent in the remainder of this section. Our first remark on f (a) concerns the fact that the counting used in its definition, i.e. N n (a, ε) may be used to estimate box dimensions. Based on this fact it was shown in [RIE 98] that dim(Ka ) ≤ f (a).
(4.23)
Here, we state a slightly improved version: LEMMA 4.2.– dim(Ea ) ≤ f (a)
dim(E a ) ≤ f (a)
(4.24)
1. More generally, using c-ary intervals in Euclidean space d kn will range from 0 to cnd − 1. Logarithms will have to be taken to the base c since we seek the asymptotics of N n (a, ε) in terms of a power law of resolution at stage n, i.e. N n (a, ε) cnf (a) . The maximum value of f (a) will be d.
Multifractal Scaling: General Theory and Approach by Wavelets
147
and dim(Ka ) ≤ f (a) := lim lim inf ε↓0 n→∞
1 log2 N n (a, ε). n
(4.25)
It follows immediately that dim(Ka ) ≤ dim(Ea ) ≤ f (a), but dim(Ea ) is not necessarily smaller than f (a). 4.3.3. Partition function and Legendre spectrum The second comment regarding the grain spectrum f (a) concerns its interpretation as a large deviation principle (LDP). We may consider N n (a, ε)/2n to be the probability of finding (for a fixed realization of Y ) a number kn ∈ κn := {0, . . . , 2n − 1} such that snkn ∈ [a − ε, a + ε]. Typically, there will be one value of s(t) that appears most frequently, denoted a ˆ, and f (a) will reach its maximum 1 at a=a ˆ. However, by definition, for a = a ˆ the chance to observe coarse exponents snkn which lie in [a − ε, a + ε] will decrease exponentially fast with a rate given by f (a). Appealing to the theory of LDP-s we consider the random variable An = −nsnK ln(2) where K is randomly picked from κn = {0, . . . , 2n − 1} with uniform distribution Un (recall that the we study one fixed realization or path of Y ) and define its “logarithmic moment generating function” or partition function τ (q) := lim inf − n→∞
1 log2 S n (q), n
(4.26)
where n
S (q) :=
n 2 −1
exp (−qnsnk
ln(2)) =
k=0
n 2 −1
' ( n n 2−nqsk = 2n En 2−nqsk .
(4.27)
k=0
Here, En stands for expectation with respect to Un . The Gärtner-Ellis theorem [ELL 84] then applies and yields the following result (see [RIE 95a] for a slightly stronger version): THEOREM 4.1.– If the limit τ (q) = lim − n→∞
1 log2 S n (q) n
(4.28)
exists and is finite for all q ∈ , and if τ (q) is a differentiable function of q, then the double limit f (a) = lim lim
ε↓0 n→∞
1 log2 N n (a, ε) n
(4.29)
148
Scaling, Fractals and Wavelets
exists, in particular f (a) = f (a), and f (a) = τ ∗ (a) := inf (qa − τ (q)) q∈
(4.30)
for all a. Proof. Applying [ELL 84, Theorem II] to our situation immediately gives lim sup n→∞
1 1 log2 N n (a, ε) ≤ lim sup log2 #{k : |snk − a| ≤ ε} ≤ sup τ ∗ (a ) n n→∞ n |a −a|≤ε
and lim inf n→∞
1 1 log2 N n (a, ε) ≥ lim inf log2 #{k : |snk − a| < ε} ≥ sup τ ∗ (a ). n→∞ n n |a −a|<ε
By continuity of τ ∗ (a) these two bounds coincide and (4.29) is established. Now, letting ε → 0 shows that f (a) = τ ∗ (a). Sometimes, the differentiability assumptions of this theorem are too restrictive. Before dwelling more on the relation between τ and f in section 4.4 let us note a simple fact, also providing a simple reason why the Legendre transform appears in this context. LEMMA 4.3.– We always have f (a) ≤ τ ∗ (a).
(4.31)
Proof. Fix q ∈ and consider a with f (a) > −∞. Let γ < f (a) and ε > 0. Then, we take n arbitrarily large such that N n (a, ε) ≥ 2nγ . For such n we bound S n (q) by noting n 2 −1
k=0
n
2−nqsk ≥
n
2−nqsk ≥ N n (a, ε)2−n(qa+|q|ε) ≥ 2−n(qa−γ+|q|ε) (4.32)
|sn k −a|<ε
and hence τ (q) ≤ qa − γ + |q|ε. Letting ε → 0 and γ → f (a), we find τ (q) ≤ qa − f (a). Since this is trivial if f (a) = −∞ we find τ (q) ≤ qa − f (a) and
f (a) ≤ qa − τ (q)
for all a and q.
From this it follows trivially that τ (q) ≤ f ∗ (q) and f (a) ≤ τ ∗ (a).
(4.33)
Multifractal Scaling: General Theory and Approach by Wavelets
149
With the special choice snk = αkn for the distribution function M of a measure μ, S (q) becomes n
S
n
α (q)
=
n 2 −1
n
2 −1 q M (k + 1)2−n − M(k2−n )q = (μ(Ikn )) .
k=0
(4.34)
k=0
This is the original form in which τ (q) has been introduced in multifractal analysis [HAL 86, HEN 83, FRI 85, MAN 74]. Note that there is a close connection to the thermodynamic formalism [TEL 88]. 4.3.4. Deterministic envelopes An analytical approach is often useful in order to gain an intuition on the various spectra of a typical path of Y , or at least some estimate of it. To establish such an approach, we consider the position, i.e. t or kn , as well as the path Y to be random simultaneously. We then apply the LDP to the larger probability space. More precisely, the exponents snK are now random variables over (Ω × κn , PΩ × Un ). The “deterministic partition function” corresponding to this setting reads as 1 T (q) := lim inf − log2 EΩ [S n (q)]. (4.35) n→∞ n NOTE 4.1 (Ergodic processes).– So far, we have assumed in the definitions of τ (q) and T (q) that Y is defined on a compact interval. Without loss of generality, this interval was assumed to be [0, 1]. In order to allow for processes defined on we modify S n (q) to N 2n −1 1 −nqsnk n 2 S (q) := lim N →∞ N k=0
n
and N (a, ε) similarly. For ergodic processes this becomes S n (q) = 2n EΩ [2−nqsk ] almost surely. Thus, EΩ [S n (q)] = S n (q) a.s. and n
a.s.
T (q) = τ (q, ω).
(4.36)
We refer to (4.74) for an account of the extent to which marginal distributions may be reflected in multifractal spectra in general. For processes on [0, 1] we cannot expect to have (4.36) in all generality. Nevertheless, we will point out scenarios where (4.36) holds. Notably T (q) always serves as a deterministic envelope of τ (q, ω): LEMMA 4.4.– With probability one2 τ (q, ω) ≥ T (q) for all q with T (q) < ∞.
2. For clarity, we make the randomness of τ explicit here.
(4.37)
150
Scaling, Fractals and Wavelets
Proof. Consider any q with finite T (q) and let ε > 0. Let n0 be such that EΩ [S n (q)] ≤ 2−n(T (q)−ε) for all n ≥ n0 . Then, n(T (q)−2ε) Sn (q, ω) ≤ E 2n(T (q)−2ε) Sn (q, ω) ≤ 2−nε < ∞. E lim sup 2 n→∞
n≥n0
n≥n0
Thus, almost surely lim supn→∞ 2n(T (q)−2ε) Sn (q, ω) < ∞, and τ (q) ≥ T (q)−2ε. Consequently, this estimate holds with probability one simultaneously for all ε = 1/m (m ∈ N) and some countable, dense set of q values with T (q) < ∞. Since τ (q) and T (q) are always concave due to Corollary 4.2 below (see section 4.4), they are continuous on open sets and the claim follows. Along the same lines we may define the corresponding deterministic grain spectrum. By analogy, we will replace probability over κn = {0, . . . , 2n − 1} in (4.21), i.e. N n (a, ε), by probability over Ω × κn , i.e. n 2 −1
! " PΩ [a − ε ≤ snk < a + ε] = 2n EΩ×κn 1[a−ε,a+ε) (snK )
k=0
(4.38)
= EΩ [N n (a, ε)] and define F (a) := lim lim sup ε↓0
n→∞
1 log2 EΩ [N n (a, ε)] n
(4.39)
Replacing N n (a, ε) with (4.38) in the proof of Theorem 4.1 and taking expectations in (4.32) we find properties analogous to the pathwise spectra τ and f : THEOREM 4.2.– For all a F (a) ≤ T ∗ (a).
(4.40)
Furthermore, under conditions on T (q) analogous to τ (q) in Theorem 4.1 F (a) = T ∗ (a) = F (a) := lim lim inf ε↓0 n→∞
1 log2 EΩ [N n (a, ε)]. n
(4.41)
It follows from Lemma 4.4 that with probability one τ ∗ (a, ω) ≤ T ∗ (a) for all a. Similarly, the deterministic grain spectrum F (a) is an upper bound to its pathwise defined random counterpart f (a, ω), however, only pointwise. On the other hand, we have here almost sure equality under certain conditions. NOTE 4.2 (Negative dimensions).– Defined through counting f (a) as always positive – or −∞. The envelopes T ∗ and F , being defined through expectations of
Multifractal Scaling: General Theory and Approach by Wavelets
151
counts and sums, may assume negative values. Consequently, the negative values of T ∗ and F are not very useful in the estimation of f ; however, they do contain further information and can be “observed”. Negative F (a) and T ∗ (a) have been termed negative dimensions [MAN 90b]. They correspond to probabilities of observing a coarse Hölder exponent a which decays faster than the 2n = #κn “samples” snk available in one realization. Oversampling the process, i.e. analyzing several independent realizations, will increase the number of samples more “rare” snk may be observed. In loose terms, in exp(−n ln(2)F (a)) independent traces we have a fair chance of seeing at least one snk of size a. Thereby, it is essential not to average the spectra f (a) of the various realizations but the numbers N n (a, ε). This way, negative “dimensions” F (a) become visible. 4.4. Multifractal formalism In the previous section, various multifractal spectra were introduced along with some simple relations between them. These can be summarized as follows: COROLLARY 4.1 (Multifractal formalism).– For every a a.s.
dim(Ka ) ≤ dim(Ea ) ≤ f (a) ≤ τ ∗ (a) ≤ T ∗ (a)
(4.42)
where the first relations hold pathwise and the last one (the two terms on both sides of the last inequality) with probability one. Similarly a.s.
dim(Ka ) ≤ f (a) ≤ f (a) ≤ F (a) ≤ T ∗ (a).
(4.43)
The spectra on the left end have stronger implications on the local scaling structure while the ones on the right end are more easy to estimate or calculate. This set of inequalities could fairly be called the “multifractal formalism”. However, in the mathematical community a slightly different terminology is already established which goes “the multifractal formalism holds” and means that for a particular process (or one of its paths, according to context) dim(Ka ) can be calculated using some adequate partition function (such as τ (q)) and taking its Legendre transform. Consequently, when “the multifractal formalism holds” for a path or process, then we often find that equality holds between several or all spectra appearing in (4.42), depending on the context of the formalism that had been established. This property (“the multifractal formalism holds”) is a very strong one and suggests the presence of one single underlying multiplicative structure in Y . This intuition is supported by the fact that the multifractal formalism in known to “hold” up to now only for objects with strong rescaling properties where multiplication is involved such as self-similar measures, products of processes and infinitely divisible
152
Scaling, Fractals and Wavelets
cascades (see [CAW 92, FAL 94, ARB 96, RIE 95a, PES 97, HOL 92], respectively [MAN 02, BAC 03, BAR 04, BAR 02, CHA 05] as well as references therein). A notable exception of processes without injected multiplicative structure are Lévy processes, the multifractal properties of which are well understood due to [JAF 99]. Though we pointed out some conditions for equality between f , τ ∗ and T ∗ we must note that in general we may have strict inequality in some or all parts of (4.42). Such cases have been presented in [RIE 95a, RIE 98]. There is, however, one equality which holds under mild conditions and connects the two spectra in the center of (4.42). THEOREM 4.3.– Consider a realization or path of Y . If the sequence snk is bounded, then τ (q) = f ∗ (q),
for all q ∈ .
(4.44)
Proof. Note that τ (q) ≤ f ∗ (q) from Lemma 4.3. Now, to estimate τ (q) from below, choose a larger than |snk | for all n and k and group the terms in S n (q) conveniently, i.e.
a/ε
S n (q) ≤
n
2−nqsk
i=− a/ε (i−1)ε≤sn k <(i+1)ε
(4.45)
a/ε
≤
N n (iε, ε)2−n(qiε−|q|ε) .
i=− a/ε
Next, we need uniform estimates on N n (a, ε) for various a. Fix q ∈ and let η > 0. Then, for every a ∈ [−a, a] there is ε0 (a) and n0 (a) such that N n (a, ε) ≤ 2n(f (a)+η) for all ε < ε0 (a) and all n > n0 (a). We would like to have ε0 and n0 independent from a for our uniform estimate. To this end note that N n (a , ε ) ≤ N n (a, ε) for all a ∈ [a − ε/2, a + ε/2] and all ε < ε/2. By compactness we may choose a finite set of aj (j = 1, . . . , m) such that the collection [aj − ε0 (aj )/2, aj + ε0 (aj )/2] covers [−a, a]. Set ε1 = (1/2) minj=1,...,m ε0 (aj ) and n1 = maxj=1,...,m n0 (aj ). Then, for all ε < ε1 and n > n1 , and for all a ∈ [−a, a] we have N n (a, ε) ≤ 2n(f (a)+η) and, thus,
a/ε
S n (q) ≤
2−n(qiε−f (iε)−η−|q|ε) (4.46)
i=− a/ε
≤ (2a/ε + 1) · 2−n(f
∗
(q)−η−|q|ε)
.
Letting n → ∞ we find τ (q) ≥ f ∗ (q) − η − |q|ε for all ε < ε1 . Now we let ε → 0 and finally η → 0 to find the desired inequality.
Multifractal Scaling: General Theory and Approach by Wavelets
153
Due to the properties of Legendre transforms3 it follows: COROLLARY 4.2 (Properties of the partition function).– If the sequence snk is bounded, then the partition function τ (q) is concave and monotonous. Consequently, τ (q) is continuous on , and differentiable in all but a countable number of exceptional points. In order to efficiently invert Theorem 4.3 we need: LEMMA 4.5 (Lower semi-continuity of f and F ).– Let am converge to a∗ . Then f (a∗ ) ≥ lim sup f (am )
(4.47)
m→∞
and analogous for F . Proof. For all ε > 0 we can find m0 such that a∗ − ε < am − ε/2 < am + ε/2 < a∗ + ε for all m > m0 . Then, N n (a∗ , ε) ≥ N n (am , ε/2) and E[N n (a∗ , ε)] ≥ E[N n (am , ε/2)]. We find lim sup n→∞
1 1 log2 N n (a∗ , ε) ≥ lim sup log2 N n (am , ε/2) ≥ f (am ) n n→∞ n
for any m > m0 (ε) and similar for F . Now, let first m → ∞ and then ε → 0. COROLLARY 4.3 (Central multifractal formalism).– We always have f (a) ≤ f ∗∗ (a) = τ ∗ (a).
(4.48)
Furthermore, denoting by τ (q±) the right- (resp. left-)sided limits of derivatives we have, f (a) = τ ∗ (a) = qτ (q±) − τ (q±)
at a = τ (q±).
(4.49)
Proof. The graph of f ∗∗ is the concave hull of the graph of f which implies (4.48). It is an easy task to derive (4.49) under assumptions suitable to make the tools of calculus available such as continuous second derivatives. To prove it in general let us first assume that τ is differentiable at a fixed q. In particular, τ (q ) is then finite for q close to q. Since τ (q) = f ∗ (q) there is a sequence am such that τ (q) = limm qam − f (am ). Since τ (q ) ≤ q a − f (a) for all q and a by (4.33), and since τ is differentiable
3. For a tutorial on the Legendre transform see [RIE 99, App. A].
154
Scaling, Fractals and Wavelets
at q this sequence am must converge to a∗ := τ (q). From the definition of am we conclude that f (am ) converges to qa∗ − τ (q). Applying Lemma 4.5 we find that f (a∗ ) ≥ qa∗ − τ (q). Recalling (4.33) implies the desired equality. Now, for an arbitrary q the concave shape of τ implies that there is a sequence of numbers qm larger than q in which τ is differentiable and which converges down to q. Consequently, τ (q+) = limm τ (qm ). Formula (4.49) being established at all qm Lemma 4.5 applies with am = τ (qm ) and a∗ = τ (q+) to yield f (τ (q+)) ≥ qτ (q+) − τ (q+). Again, (4.33) furnishes the opposite inequality. A similar argument applies to τ (q−). COROLLARY 4.4.– If T (q) is finite for an open interval of q-values then |snk | is bounded for almost all paths, and T (q) = F ∗ (q)
for all q.
(4.50)
Moreover, F (a) = T ∗ (a) = qT (±q) − T (±q)
at a = T (±q).
(4.51)
Proof. Assume for a moment that snk is unbounded from above with positive probability. Then, grouping (4.45) requires an additional term collecting the snk > a. In fact, for any number a we can find n arbitrarily large such that snk > a for some k. This implies that for any negative q we have S n (q) ≥ 2−nqa and τ (q) ≤ qa. Letting a → ∞ shows that τ (q) = −∞. By Lemma 4.4 we must have T (q) = −∞, a contradiction. Similarly, we show that snk is bounded from below. The remaining claims can be established analogously to those for τ (q) by taking expectations in (4.45). NOTE 4.3 (Estimation and unbounded moments).– In order to apply Corollary 4.4 in a real world situation, but also for the purpose of estimating τ (q), it is of great importance to possess a method to estimate the range of q-values for which the moments of a stationary process (such as the increments or the wavelet coefficients of Y ) are finite. Such a procedure is proposed in [GON 05] (see also [RIE 04]). 4.5. Binomial multifractals The binomial measure has a long-standing tradition in serving as the paradigm of multifractal scaling [MAN 74, KAH 76, MAN 90a, CAW 92, HOL 92, BEN 87, RIE 95a, RIE 97b]. We present it here with an eye on possible generalizations of use in modeling. 4.5.1. Construction To be consistent in notation we denote the binomial measure by μb and its distribution function by Mb (t) := μb (] − ∞, t[). Note that μb is a measure or
Multifractal Scaling: General Theory and Approach by Wavelets
155
(probability) distribution, i.e. not a function in the usual sense, while Mb is a right-continuous and increasing function by definition. In order to define μb we again use the notation (4.4): for any fixed t there is a unique sequence k1 , k2 , . . . such that the dyadic intervals Iknn = [kn 2−n , (kn +1)2−n [ contain t for all integer n. So, the Ikn form a decreasing sequence of half open intervals n+1 n+1 is the left subinterval of Iknn and I2k which shrink down to {t}. Moreover, I2k n n +1 the right subinterval (see Figure 4.1). Note that the first n elements of such a sequence, i.e. (k1 , k2 , . . . , kn ) are identical for all points t ∈ Iknn . We call this a nested sequence and it is uniquely defined by the value of kn . We set μb (Ikn ) = Mb ((kn + 1)2−n ) − Mb (kn 2−n ) · · · Mk11 · M00 . = Mknn · Mkn−1 n−1
(4.52)
In other words, the mass lying in Iknn is redistributed among its two dyadic n+1 n+1 n+1 n+1 subintervals I2k and I2k in the proportions M2k and M2k . For consistency n n +1 n n +1 n+1 n+1 we require M2kn + M2kn +1 = 1. Having defined the mass of dyadic intervals we obtain the mass of any interval ] − ∞, t[ by writing itas a disjoint union of dyadic intervals J n and noting Mb (t) = μb (] − ∞, t[= n μb (J n ). Therefore, integrals (expectations) with respect to μb can be calculated as g(t)μb (dt) = lim
n→∞
=
g(t)dMb (t) = lim
n→∞
n −1 2
g(k2−n )μb (Ikn )
(4.53)
k=0 n −1 2
g(k2−n ) Mb ((k + 1)2−n ) − Mb (k2−n ) (4.54)
k=0
Alternatively, the measure μb can be defined using its distribution function Mb . Indeed, as a distribution function, Mb is monotone and continuous from the right. Since (4.52) defines Mb in all dyadic points it can be obtained in any other point as the right-sided limit. Note that Mb is continuous at a given point t unless Mknn (t) = 1 for all n large. To generate randomness in Mb , we choose the various Mkn to be random variables. The above properties then hold pathwise. We will make the following assumptions on the multiplier distributions Mkn : i) Conservation of mass. Almost surely for all n and k Mkn is positive and n+1 n+1 + M2k = 1. M2k n n +1
(4.55)
156
Scaling, Fractals and Wavelets
0
M0
0.07
1
0
0.06 0.05
1 M0
.M
0.04
0 0
1
.
0.03
0
M1 M0
0
0.5
1
0.02 0.01 0
. .
2
1
0
M0 M0 M0 0
. .
2
1
Ŧ0.01
0
M1 M0 M0
0.25
. .
2
1
0
M2 M1 M0 0.5
. .
2
1
0
M3 M1 M0 0.75 1
Ŧ0.02 Ŧ0.03 0
2000
4000
6000
8000
Figure 4.2. Iterative construction of the binomial cascade
As we have seen, this guarantees that Mb is well defined. ii) Nested independence. All multipliers of a nested sequence are mutually independent. Analogously to (4.52) we have for any nested sequence EΩ [Mknn · · · M00 ] = EΩ [Mknn ] · · · EΩ [M00 ]
(4.56)
and similar for other moments. This will allow for simple calculations in what follows. iii) Identical distributions. For all n and k M0 if k is even, n fd Mk = M1 if k is odd.
(4.57)
A more general version of iii) was given in [RIE 99] to allow for more flexibility in model matching. The theory of cascades or, more properly, T -martingales4 [KAH 76, BEN 87, HOL 92, BAR 97], provides a wealth of possible generalizations. Most importantly, it allows us to soften the almost sure conservation condition i) to i’) Conservation in the mean EΩ [M0 + M1 ] = 1.
(4.58)
In this case, Mb is well defined since (4.52) forms a martingale due to the nested independence (4.56). The main advantage of such an approach is that we can use unbounded multipliers M0 and M1 such as log-normal random variables. Then, the marginals of the increment process, i.e. μb (Ikn ) are exactly log-normal on all scales. For general binomials, always assuming ii) it can be argued that the marginals μb (Ikn ) are at least asymptotically log-normal by applying a central limit theorem to the logarithm of (4.52).
4. For any fixed t sequence (4.52) forms a martingale due to the nested independence (4.56).
Multifractal Scaling: General Theory and Approach by Wavelets
157
4.5.2. Wavelet decomposition The scaling coefficients of μb using the Haar wavelet are simply
φ∗j,k (t) μb (dt) = 2n/2
Dn,k (μb ) =
(k+1)2−n
μb (dt) = 2n/2 μb (Ikn )
k2−n
(4.59)
from (4.8) and (4.53). With (4.9) and (4.52) we derive the explicit expression for the Haar wavelet coefficients: n+1 n+1 ) − μb (I2k ) 2−n/2 Cn,kn (μb ) = μb (I2k n n +1 n+1 n+1 − M2k ) = (M2k n n +1
n )
Mki i .
(4.60)
i=0
Similar scaling properties hold when using arbitrary, compactly supported wavelets, provided the distributions of the multipliers are scale independent. This comes about from (4.52) and (4.53), which give the following rule for substituting t = 2n t − kn −n/2 2 Cn,kn (μb ) = ψ(2n t − kn )μb (dt) Iknn
= Mknn · · · Mk11 ·
1 0
(4.61) (n,k ) ψ(t )μb n (dt ).
(n,k )
Here μb n is a binomial measure constructed with the same method as μb itself, however, with multipliers taken from the subtree which has its root at the node kn of level n of the original tree. More precisely, for any nested sequence i1 , . . . , im (n,kn )
μb
n+1 n+2 (Iim ) = M2k · M4k · · · M2n+m m k +i . m n +i1 n +i2 n m
(n,k )
From nested independence (4.56) we infer that this measure μb n is independent of Mki i (i = 1, . . . , n). Furthermore, the identical distributions of the multipliers iii) imply that for arbitrary, compactly supported wavelets 1 1 d (n,k ) ψ(t)μb n (dt) = C0,0 (μb ) = ψ(t)μb (dt) (4.62) 0
d
0
where = denotes equality in distribution. In particular, for the Haar wavelet we have 1 d (n,k ) n+1 n+1 Haar ψHaar (t)μb n (dt) = M2k − M2k = M0 − M1 = C0,0 (μb ) (4.63) n n +1 0
158
Scaling, Fractals and Wavelets
(the deterministic analog has also been observed in [BAC 93]). Finally, note that if ψ is supported on [0, 1], then ψ(2n (·) − k) is supported on Ikn . So, the tree of wavelet coefficients Cn,k of μb possess a structure similar to the tree of increments of Mb (compare (4.52)). With a little more effort we calculate the wavelet coefficients of Mb itself, provided ψ is admissible and supported on [0, 1]. Indeed, Mb (t) − Mb (kn 2−n ) = μb ([kn 2−n , t]) (n,kn )
= Mknn · · · Mk11 Mb (n,kn )
where Mb this yields
(n,kn )
(t ) := μb
2−n/2 Cn,kn (Mb ) =
([0, t ]). Using
Iknn
#
(2n t − kn ),
(4.64)
ψ = 0 and substituting t = 2n t − kn
ψ(2n t − kn ) Mb (t) − Mb (kn 2−n ) dt
= 2−n · Mknn · · · Mk11 ·
0
1
(4.65) (n,k ) ψ(t )Mb n (t )dt .
Again, we have 1 d (n,k ) ψ(t)Mb n (dt) = C0,0 (Mb ) = 0
1
ψ(t)Mb (dt)
(4.66)
0
LEMMA 4.6.– Let ψ be a wavelet supported on [0, 1]. Let Mb be a binomial with i)-iii). Then, Cn,kn (μb ) is given by (4.61), and if ψ is admissible then Cn,kn (Mb ) is given by (4.65). Furthermore, (4.62) and (4.66) hold. It is obvious that the dyadic structure present in both the construction of the binomial measure as well as in the wavelet transform are responsible for the simplicity of the calculation above. It is, however, standard by now to extend the procedure to more general multinomial cascades such as Mc , introduced in section 4.5.5 (see [ARB 96, RIE 95a]). 4.5.3. Multifractal analysis of the binomial measure In the light of Lemma 4.6 it becomes clear that the singularity exponent α(t) is most easily accessible for Mb while w(t) is readily available for both, Mb and μb . On the other hand, as increments appear in α(t) they are not well defined for μb . Thus, it is natural to calculate the spectra of both, Mb and μb , with appropriate singularity exponents, i.e. f α,Mb , f w,Mb and f w,μb .
Multifractal Scaling: General Theory and Approach by Wavelets
159
Now, Lemma 4.6 indicates that the singularity structures of μb and Mb are closely related. Indeed, μb is the distributional derivative of Mb in the sense of (4.52) and (4.54). Since taking a derivative “should” simply reduce the scaling exponent by one, we would expect that their spectra are identical up to a shift in a by −1. Indeed, this is true for increasing processes, such as Mb , as we will elaborate in section 4.6.2. However, it has to be pointed out that this rule cannot be correct for oscillating processes. This is effectively demonstrated by the example ta · sin(t−b ) with b > 0. Though this example has the exponent a at zero, its derivative behaves like ta−b−1 there. This is caused by the strong oscillations, also called chirp, at zero. In order to deal with such situations the 2-microlocalization theory has to be employed [JAF 91]. Let us first dwell on the well known multifractal analysis of Mb based on αkn . Recall that Mb ((kn + 1)2−n ) − Mb (kn ) is given by (4.52), and use the nested independence (4.56) and identical distributions (4.57) to obtain E[S
n
α,Mb (q)]
=
n 2 −1
E
'
Mknn
q
q 0 q ( M0 · · · Mk11
kn =0
n ( n i n−i E [M0q ] E [M1q ] · =E i i=0 ' ( n 0 q · (E [M0q ] + E [M1q ]) . = E M0
'
q M00
(4.67)
From this, it follows immediately that T α,Mb (q) = − log2 E [(M0 )q + (M1 )q ] .
(4.68)
Note that this value may be −∞ for some q. THEOREM 4.4.– Assume that i’), ii) and iii) hold. Assume furthermore that M0 and M1 have at least some finite moment of negative order. Then, with probability one dim(Ka ) = f (a) = τ ∗ (a) = T ∗ α,Mb (a)
(4.69)
for all a such that T ∗ α,Mb (a) > 0. Thereby, all the spectra are related to the singularity exponents αkn or hnk of Mb . NOTE 4.4 (Wavelet analysis).– In what follows we will show that we obtain the same spectra for Mb replacing αkn with wkn for certain analyzing wavelets. We will also mention the changes which become necessary when studying distribution functions of measures with fractal support (see section 4.5.5).
160
Scaling, Fractals and Wavelets
Proof. Inspection [BAR 97] we find that dim(Ka ) = T ∗ (a) for αkn under the given assumptions. Earlier results, such as [FAL 94, ARB 96], used more restrictive assumptions but are somewhat easier to read. Though weaker than [BAR 97] they are sufficient in some situations. 4.5.4. Examples Example 1 (β binomial).– Consider multipliers M0 and M1 that follow a β distribution, which has the density cp tp−1 (1 − t)p−1 for t ∈ [0, 1] and 0 elsewhere. Thus, p > 0 is a parameter and cp is a normalization constant. Note that the conservation of mass i) imposes a symmetric distribution since M0 and M1 are set to be equally distributed. The β distribution has finite moments of order q > −p which can be expressed explicitly using the Γ-function. We obtain β-Binomial: T α (q) = −1 − log2
Γ(p + q)Γ(2p) Γ(2p + q)Γ(p)
(q > −p),
(4.70)
and T (q) = −∞ for q ≤ −p. For a typical shape of these spectra, see Figure 4.3. 1.5
1.5
slope=D
slope=q
1 0.5
1
(q,T(q)) (1,T(1))=(1,0)
ŦT*(D)
T (D) o
(0,T(0))=(0,Ŧ1)
Ŧ1
0
q0=0
* (D,T (D))
0.5
*
T(q) o
0 Ŧ0.5
q1=1
ŦT(0)=1
ŦT(1)=0
Ŧ1.5 Ŧ2
Ŧ0.5
ŦT(q)
Ŧ2.5 Ŧ3 Ŧ1
0
1 qo
2
3
Ŧ1 Ŧ0.5
0
0.5
1
1.5
2
2.5
3
D o
Figure 4.3. The spectrum of a binomial measure with β distributed multipliers with p = 1.66. Trivially, T (0) = −1, where the maximum of T ∗ is 1. In addition, every positive increment process has T (1) = 0, where T ∗ touches the bisector. Finally, the LRD parameter is Hvar = (T (2) + 1)/2 = 0.85 (see (4.90) below)
An application of the β binomial for the modeling of data traffic on the Internet can be found in [RIE 99]. Example 2 (Uniform binomial).– As a special case of the β binomial we obtain uniform distributions for the multipliers when setting p = 1. Formula (4.70) simplifies
Multifractal Scaling: General Theory and Approach by Wavelets
161
to T α (q) = −1 + log2 (1 + q) for q > −1. Applying the formula for the Legendre transform (4.51) yields the explicit expression a (4.71) uniform binomial: T ∗ α (a) = 1 − a + log2 (e) + log2 log2 (e) for a > 0 and T ∗ α (a) = −∞ for a ≤ 0.
Example 3 (Log-normal binomial).– Another very interesting case is log-normal distributions for the multipliers M0 and M1 . Note that we have to replace i) with i’) in this case since log-normal variables can be arbitrarily large, i.e. larger than 1. Recall that the log-normal binomial enjoys the advantage of having exactly log-normal marginals μb (Ikn ) since the product of independent log-normal variables is again a log-normal variable. Having mass conservation only in the mean, however, may cause problems in simulations since the sample mean of the process μb (Ikn ) (k = 0, . . . , 2n − 1) is not M00 as in case i), but depends on n. Indeed, the negative (virtual) a appearing in the log-normal binomial spectrum reflects the possibility that the sample average my increase locally (see [MAN 90a]). The calculation of its spectrum starts by observing that the exponential M = eG of a N (m, σ 2 ) variable G, i.e. a Gaussian with mean m and variance σ 2 , has the q-th moment E[M q ] = E[exp(qG)] = exp(qm + q 2 σ 2 /2). Assuming that M0 and M1 are equally distributed as M their mean must be 1/2. Hence m + σ 2 = − ln(2), and σ2 q (4.72) log-normal binomial: T α (q) = (q − 1) 1 − 2 ln(2) for all q ∈ such that E[(Mb (1))q ] is finite. Note that the parabola in (4.72) has two zeros: 1 and qcrit = 2 ln(2)/σ 2 . It follows from [KAH 76] that E[(Mb (1))q ] < ∞ exactly for q < qcrit . Since T (q) is exactly differentiable for q < qcrit we may obtain its Legendre transform implicitly from (4.51) for a = T (q) with q < qcrit , i.e., for all a > acrit = T (qcrit ) = σ 2 /(2 ln(2)) − 1. Eliminating q from (4.51) yields the explicit form 2 σ2 ln(2) a − 1 − T ∗ α (a) = 1 − (a ≥ acrit ) (4.73) 2σ 2 2 ln(2) For a ≤ acrit the Legendre transform yields T ∗ (a) = a · qcrit . Thus, at acrit the spectrum T ∗ crosses over from parabola (4.73) to its tangent through the origin with slope qcrit (the other tangent through the origin is the bisector). It should be remembered that only the positive part of this spectrum can be estimated from one realization of Mb . The negative part corresponds to events so rare that they can only be observed in a large array of realizations (see Note 4.2).
162
Scaling, Fractals and Wavelets
The log-normal framework also allows us to calculate F (a) explicitly, demonstrating which rescaling properties of the marginal distributions of the increment processes of Mb are captured in the multifractal spectra. Indeed, if all ln(Mkn ) are N (m, σ 2 ) then − ln(2) · αkn is N (m, σ 2 /n). The mean value theorem of integration gives ln(2)(−a+ε) 1 (x − m)2 n dx exp − PΩ [|αk − a| < ε] = * 2σ 2 /n 2πσ 2 /n ln(2)(−a−ε) (− ln(2)xa,n − m)2 1 ln(2) · 2ε · exp − =* 2σ 2 /n 2πσ 2 /n with xa,n ∈ [a − ε, a + ε] for all n. Keeping only the exponential term in n and substituting m = −σ 2 − ln(2) we find ln(2) 1 log2 (2n PΩ [|αkn − a| < ε]) 1 − n 2σ 2
σ2 xa,n − 1 − 2 ln(2)
2 .
(4.74)
Comparing with (4.73) we see that T ∗ (a) = F (a), as stated in Theorem 4.2. The above computation shows impressively how well adapted a multiplicative iteration with log-normal multipliers is to multifractal analysis (or vice versa): F extracts, basically, the exponent of the Gaussian kernel. Since the multifractal formalism holds for Mb these features can be measured or estimated using the re-normalized histogram, i.e. the grain based multifractal spectrum f (a). This is a property which could be labeled with the term ergodicity. Note, however, that classical ergodic theory deals with observations along an orbit of increasing length, while f (a) concerns a sequence of orbits. 4.5.5. Beyond dyadic structure We elaborate here generalizations of the binomial cascade. Statistically self-similar measures: a natural generalization of the random binomial, denoted here by Mc , is obtained by splitting intervals Jkn iteratively n+1 n+1 n+1 n , . . . , Jck+c−1 with length |Jck+i | = Ln+1 into c subintervals Jck ck+i |Jk | and n+1 n+1 n mass μc (Jck+i ) = Mck+i μc (Jck ). In the most simple case, we will require mass n+1 n+1 n+1 conservation, i.e. Mck + · · · + Mck+c−1 = 1, but also Ln+1 ck + · · · + Lck+c−1 = 1 which guarantees that μc lives everywhere. Assuming the analogous properties of ii) and iii) to hold for both the length- as well as the mass-multipliers we find that T Mc (q) is the unique solution of ( ' (4.75) E (M0 )q (L0 )−T (q) + · · · + (Mc−1 )q (Lc−1 )−T (q) = 1.
Multifractal Scaling: General Theory and Approach by Wavelets
163
This formula of T (q) can be derived rigorously by taking expectations where appropriate in the proof of [RIE 95a, Prop 14]. Doing so shows, moreover, that T (q) assumes a limit in these examples. Multifractal formalism: it is notable that the multifractal formalism “holds” for the class of statistically self-similar measures described above in Theorem 4.4 (see [ARB 96]). n+1 n n However, if Ln+1 ck + · · · + Lck+c−1 = λ < 1, e.g. choosing Lk = (1/c ) almost surely with c > c, then the measure μc lives on a set of fractal dimension and its distribution function Mc (t) = μc ([0, t)) is constant almost everywhere. In this case, equality in the multifractal formalism will fail: indeed, unless the scaling exponents snk are modified to account for boundary effects caused by the fractal support, the partition function will be unbounded for negative q, e.g. τ α (q) = −∞ for q < 0 (see [RIE 95a]). As a consequence, T α (q) = −∞ and (4.75) is no longer valid for q < 0. Interestingly, the fine spectrum dim(Ka ) is still known, however, due to [ARB 96].
Stationary increments: however, an entirely different and novel way of introducing randomness in the geometry of multiplicative cascades which leads to perfectly stationary increments has been given recently in [MAN 02] and in [BAR 02, BAR 03, BAR 04, MUZ 02, BAC 03, CHA 02, CHA 05, RIE 07b, RIE 07a]. The description of these model is, unfortunately, beyond the scope of this work. Binomial in the wavelet domain: in concluding this section we should mention that, with regard to (4.61), we may choose to directly model the wavelet coefficients of a process in a multiplicative fashion in order to obtain a desired multifractal structure. Some early steps in this direction have been taken in [ARN 98]. 4.6. Wavelet based analysis 4.6.1. The binomial revisited with wavelets The deterministic envelope is the most simple wavelet-based spectra of μb to calculate. Taking into account the normalization factors in (4.12) when using Lemma 4.6, the calculation of (4.67) carries over to give n
S n w,μb (q) = 2nq E [|C0,0 |q ] · (EΩ [M0q ] + EΩ [M1q ]) , and similar for Mb . Provided E [|C0,0 |q ] is finite this immediately gives T w,μb (q) + q = T w,Mb (q) = T α,Mb (q), T
∗
w,μb (a
− 1) = T
∗
w,Mb (a)
=T
∗
α,Mb (a).
(4.76) (4.77)
164
Scaling, Fractals and Wavelets
Imposing additional assumptions on the distributions of the multipliers we may also control wkn (μb ) themselves and not only their moments. To this end, we should be able to guarantee that the wavelet coefficients do not decay too fast (compare (4.10)), i.e. the random factor (4.62) which appears in (4.61) does not become too small. Indeed, it is sufficient to assume that there is some ε > 0 such that |C0,0 (μb )| ≥ ε # (n,k ) almost surely. Then for all t, (1/n) log( ψ(t)μb n (dt)) → 0, and with (4.61) 1 (4.78) wμb (t) = lim inf − log2 2n/2 |Cn,kn | = αMb (t) − 1, n→∞ n and similarly wμb (t) = αMb (t) − 1. Observe that this is precisely the relation we expect between the scaling exponents of a process and its (distributional) derivative – at least in nice cases – and that it is in agreement with (4.77). In summary (first observed for deterministic binomials in [BAC 93]): COROLLARY 4.5.– Assume that μb is a random binomial measure satisfying i)-iii). # # (n,k) (n,k) Assume, that the random variables | ψ(t)μb (dt)| resp. | ψ(t)Mb (t)dt| are uniformly bounded away from 0. Then, the multifractal formalism “holds” for the wavelet based spectra of μb , resp. Mb , i.e. dim(Eaw,μb ) = f w,μb (a) = τ ∗ w,μb (a) = T ∗ w,μb (a),
(4.79)
dim(Eaw,Mb ) = f w,Mb (a) = τ ∗ w,Mb (a) = T ∗ w,Mb (a).
(4.80)
a.s.
a.s.
a.s.
respectively a.s.
a.s.
a.s.
# # (n,k) (n,k) Requiring that | ψ(t)μb (dt)| resp. | ψ(t)Mb (t)dt| should be bounded away from zero in order to insure (4.78), though satisfied in some simple cases, seems unrealistically restrictive to be of practical use. A few comments are in order here. First, this condition can be weakened to arbitrarily allow small values of these integrals, as long as all their negative moments exist. This can be shown by an argument using the Borel-Cantelli lemma. Second, the condition may simplify in two ways. For iid multipliers we know that these integrals are equal in distribution to C0,0 , thus only n = k = 0 has to be checked. Further, for the Haar wavelet and symmetric multipliers, it becomes simply the condition that M0 be uniformly bounded away from zero (see (4.60)), or at least that E[|M0 − 1/2|q ] < ∞ for all negative q. Third, if we drop iii) and allow the multiplier distributions to depend on scale (see # # (n,k) (n,k) [RIE 99]), then | ψ(t)μb (dt)| resp. | ψ(t)Mb (t)dt| has to be bounded away from zero only for large n. In applications such as network traffic modeling we find n+1 n+1 − M2k+1 is best modeled by discrete distributions on [0, 1] that on fine scales M2k with large variance, i.e. without mass around 1/2.
Multifractal Scaling: General Theory and Approach by Wavelets
165
Fourth, another way out is to avoid small wavelet coefficients entirely in a multifractal analysis. More precisely, we would follow [BAC 93, JAF 97] and replace Cn,kn in the definition of wknn (4.12) by the maximum over certain wavelet coefficients “close” to t. Of course, the multifractal formalism of section 4.4 still holds. [JAF 97] gives conditions under which the spectrum τ ∗ w,μb (a) based on this modified wkn agrees with the “Hölder” spectrum dim(Ea ) based on hnk (Mb ). 4.6.2. Multifractal properties of the derivative Corollary 4.5 establishes for the binomial what intuition suggests in general, i.e. that the multifractal spectra of processes and their derivative should be related in a simple fashion – at least for certain classes of processes. As we will show, increasing processes have this property, at least for the wavelet based multifractal spectra. However, the order of Hölder regularity in the sense of the spaces Cth (see Lemma 4.1) might decrease under differentiation by an amount different from 1. This is particularly true in the presence of highly oscillatory behavior such as “chirps”, as the example ta sin(1/t2 ) demonstrates. In order to assess the proper space Cth a 2-microlocalization has to be employed. For good surveys see [JAF 95, JAF 91]. In order to establish a general result on derivatives we place ourselves in the framework whereby we care less for a representation of a process in terms of wavelet coefficients and are interested purely in an analysis of oscillatory behavior. A typical example of an analyzing mother wavelet ψ are the derivatives of the Gaussian kernel exp (−t2 /2) which were used to produce Figure 4.4. The idea is to use integration by parts. For a continuous measure μ on [0, 1] with distribution function M(t) = μ([0, t)) and a continuously differentiable function g this reads as g(t)μ(dt) = lim
n→∞
= lim
n→∞
n −1 2
g(k2−n ) M((k + 1)2−n ) − M(k2−n )
k=0 n −1 2
M(k2−n ) g((k − 1)2−n ) − g(k2−n )
(4.81)
k=0
+ M(1)g(1 − 2−n ) − M(0)g(−2−n ) = M(1)g(1) − M(0)g(0) − M(t)g (t)dt where we alluded to (4.53) and regrouped terms. As a matter of fact, M(0) = 0 and M(1) = 1. A similar calculation can be performed for a more general, not necessarily increasing process Y , provided it has a derivative Y , by replacing μ(dt) with Y (t)dt.
166
Scaling, Fractals and Wavelets
Figure 4.4. Demonstration of the multifractal behavior of a binomial measure μb (left) and its distribution function Mb (right). On the top a numerical simulation, i.e. (4.52) on the left and Mb (k2−n ) on the right for n = 20. In the middle the moduli of a continuous wavelet transform [DAU 92] where the second Gaussian derivative was taken as the analyzing wavelet ψ(t) for μb , resp. the third derivative ψ for Mb . The dark lines#indicate the “lines of maxima” [JAF 97, BAC 93], i.e. the locations where the modulus of ψ(2j t − s)μb (dt) has a local maximum as a function of s with j fixed. On the bottom a multifractal analysis in three steps. First, a plot of log S n w (q) against n tests for linear behavior for various q. Second, the partition function τ (q) is computed as the slopes of a least square linear fit of log S n . Finally, the Legendre transform τ ∗ (a) of τ (q) is calculated following (4.49). Indicated with dashes in the plots of τ (q) and τ ∗ (a) of μb are the corresponding function for Mb , providing empirical evidence for (4.76), (4.77), and (4.83)
Now, setting g(t) = 2n/2 ψ(2n t − k) for a smooth analyzing wavelet ψ we have g (t) = 23n/2 ψ (2n t − k) and obtain Cn,k (ψ, μ) = 2n/2 ψ(2n − k) − 2n · Cn,k (ψ , M).
(4.82)
Estimating 2n − kn = 2n − t2n (1 − t)2n and assuming exponential decay of ψ(t) at infinity allows us to conclude w(t)ψ,μ = −1 + w(t)ψ ,M ,
(4.83)
and similarly to w(t). COROLLARY 4.6.– f ψ,μ (a) = f ψ ,M (a + 1)
τ ∗ ψ,μ (a) = τ ∗ ψ ,M (a + 1)
(4.84)
This is impressively demonstrated in Figure 4.4. We should note that ψ has one more vanishing moment than ψ which is easily seen by integrating by parts. Thus, it
Multifractal Scaling: General Theory and Approach by Wavelets
167
is natural to analyze the integral of a process, here the distribution function M of the measure μ, using ψ since the degree of the Taylor polynomials typically grows by 1 under integration. NOTE 4.5 (Visibility of singularities and regularity of the wavelet).– It is notable that the Haar wavelet yields the full spectra of the binomial Mb (and also of its distributional derivative μb ). This fact is in some discord with the folklore saying that a wavelet cannot detect degrees of regularity larger than its own. In other words, a signal will rarely be more regular than the basis elements it is composed of. To resolve the apparent paradox, recall the peculiar property of multiplicative measures which is to have constant Taylor# polynomials. So, it will reveal its # scaling structure to any analyzing wavelet with ψ = 0. No higher regularity, i.e. tk ψ(t)dt = 0 is required. The correct reading of the literature is indeed, that wavelets are only guaranteed to detect singularities smaller than their own regularity. 4.7. Self-similarity and LRD The statistical self-similarity as expressed in (4.1) makes FBM, or rather its increment process, a paradigm of long range dependence (LRD). To be more explicit let δ denote a fixed lag and define fractional Gaussian noise (FGN) as G(k) := BH ((k + 1)δ) − BH (kδ).
(4.85)
Possessing the LRD property means that the auto-correlation rG (k) := EΩ [G(n + k)G(n)] decays so slowly that k rG (k) = ∞. The presence of such strong dependence bears an important consequence on the aggregated processes G(m) (k) :=
1 m
(k+1)m−1
G(i).
(4.86)
i=km
They have a much higher variance, and variability, than would be the case for a short range dependent process. Indeed, if X is a process with iid values X(k), then X (m) (k) has variance (1/m2 ) var(X0 + · · · + Xm−1 ) = (1/m) var(X). For G we find, due to (4.1) and BH (0) = 0, H 1 m BH (mδ) = var BH (δ) var(G(m) (0)) = var m m (4.87) = m2H−2 var (BH (δ)) . Indeed, for H > 1/2 this expression decays much slower than 1/m. As is shown in [COX 84] var(X (m) ) m2H−2 is equivalent to rX (k) k 2H−2 and so, G(k) is indeed LRD for H > 1/2. Let us demonstrate with FGN how to relate LRD with multifractal analysis based only on the fact that it is a zero-mean processes, not (4.1). To this end let
168
Scaling, Fractals and Wavelets
δ = 2−n denote the finest resolution we will consider, and let 1 be the largest. For m = 2i (0 ≤ i ≤ n) the process mG(m) (k) becomes simply BH ((k + 1)mδ) − BH (kmδ) = BH ((k + 1)2i−n ) − BH (k2i−n ). However, the second moment of this expression – which is also the variance – is exactly what determines T α (2). More precisely, using stationarity of G and substituting m = 2i , we obtain ' ( ! " 2 −1 EΩ S n−i α (2) = EΩ |mG(m) (k)|2 n−i
−(n−i)T α (2)
2
= 2n−i 22i var G(2
i
)
(4.88)
k=0
.
This should be compared with the definition of the LRD parameter H using var(G(m) ) m2H−2
or
i
var(G(2 ) ) 2i(2H−2) .
(4.89)
At this point a conceptual difficulty arises. Multifractal analysis is formulated in the limit of small scales (i → −∞) while LRD is a property for large scales (i → ∞). Thus, the two exponents H and T α (2) can in theory only be related when assuming that the scaling they represent is actually exact at all scales, and not only asymptotically. In any real world application, however, we will determine both H and T α (2) by finding a scaling region i ≤ i ≤ i in which (4.88) and (4.89) hold up to satisfactory precision. Comparing the two scaling laws in i yields T α (2) + 1 − 2 = 2H − 2, or H=
T α (2) + 1 . 2
(4.90)
This formula expresses most pointedly how multifractal analysis goes beyond second order statistics: with T (q) we capture the scaling of all moments. The relation (4.90), here derived for zero-mean processes, can be put on more solid grounds using wavelet estimators of the LRD parameter [ABR 95] which are more robust than the estimators through variance. The same formula (4.90) also reappears for certain multifractals (see (4.100)). In this context it is worthwhile pointing forward to (4.96), from which we conclude that T BH (q) = qH − 1 if q > −1. The fact to note here is that FBM requires indeed only one parameter to capture its scaling while multifractal scaling, in principle, is described by an array of parameters T (q). 4.8. Multifractal processes The most prominent examples where we find coinciding, strictly concave multifractal spectra are the distribution functions of cascade measures [MAN 74,
Multifractal Scaling: General Theory and Approach by Wavelets
169
KAH 76, CAW 92, FAL 94, ARB 96, OLS 94, HOL 92, RIE 95a, RIE 97b, PES 97] for which dim(Ka ) and T ∗ (a) are equal and have the form of a ∩ (see Figure 4.3 and also 4.5(e)). These cascades are constructed through a multiplicative iteration scheme such as the binomial cascade, which is presented in detail earlier in this chapter with special emphasis on its wavelet decomposition. Having positive increments, this class of processes is, however, sometimes too restrictive. FBM, as noted, has the disadvantage of a poor multifractal structure and does not contribute to a larger pool of stochastic processes with multifractal characteristics. It is also notable that the first “natural”, truly multifractal stochastic process to be identified was the Lévy motion [JAF 99]. This example is particularly appealing since scaling is not injected into the model by an iterative construction (this is what we mean by the term natural). However, its spectrum is degenerative, though it shows a non-trivial range of scaling exponents h(t), in the sense that it is linear. 4.8.1. Construction and simulation With the formalism presented here, the stage is set for constructing and studying new classes of truly multifractional processes. The idea, to speak in Mandelbrot’s own words, is inevitable after the fact. The ingredients are simple: a multifractal “time warp’, i.e. an increasing function or process M(t) for which the multifractal formalism is known to hold, and a function or process V with strong monofractal scaling properties such as fractional Brownian motion (FBM), a Weierstrass process or self-similar martingales such as Lévy motion. We then form the compound process V(t) := V (M(t)).
(4.91)
To fix the ideas, let us recall the method of midpoint displacement which can be used to define a simple Brownian motion B1/2 which we will also call the Wiener motion (WM) for a clear distinction from FBM. This method constructs B1/2 iteratively at dyadic points. Having constructed B1/2 (k2−n ) and B1/2 ((k + 1)2−n ) we define B1/2 ((2k + 1)2−n−1 ) as (B1/2 (k2−n ) + B1/2 ((k + 1)2−n ))/2 + Xk,n . The offsets Xk,n are independent zero-mean Gaussian variables with variance such as to satisfy (4.1) with H = 1/2, hence the name of the method. One way to obtain Wiener motion in multifractal time WM(MF) is then to keep the offset variables Xk,n as they are but to apply them at the time instances tk,n defined by tk,n = M−1 (k2−n ), i.e. M(tk,n ) = k2−n : B1/2 (t2k+1,n+1 ) :=
B1/2 (tk,n ) + B1/2 (tk+1,n ) + Xk,n . 2
(4.92)
This amounts to a randomly located random displacement, the location being determined by M. Indeed, (4.91) is nothing but a time warp. An alternative construction of “warped Wiener motion” WM(MF) which yields equally spaced sampling, as opposed to the samples B1/2 (tk,n ) provided by (4.92), is
170
Scaling, Fractals and Wavelets
desirable. To this end, note first that the increments of WM(MF) become independent Gaussians once the path of M(t) is realized. To be more precise, fix n and let G(k) := B((k + 1)2−n ) − B(k2−n ) = B1/2 (M(k + 1)2−n ) − B1/2 (M(k2−n )).
(4.93)
For a sample path of G we start by producing first the random variables M(k2−n ). Once this is done, the G(k) are simply independent zero-mean Gaussian variables with variance |M((k + 1)2−n ) − M(k2−n )|. This procedure has been used in Figure 4.5. 4.8.2. Global analysis To calculate the multifractal envelope T (q) we need only to know that V is an H-sssi process, i.e. that the increment V (t + u) − V (t) is equal in distribution to uH V (1) (see (4.1)). Assuming independence between V and M, a simple calculation reads as EΩ
n −1 2
V (k + 1)2−n − V k2−n q
k=0
=
n 2 −1
k=0
=
n 2 −1
q ! E E V M (k + 1)2−n − V M k2−n M k2−n , " M (k + 1)2−n
(4.94)
' q " qH ( ! E V (1) . E M (k + 1)2−n − M k2−n
k=0
With little more effort the increments |V((k + 1)2−n ) − V(k2−n )| can be replaced n by suprema, i.e. by 2−nhk , or even certain wavelet coefficients under appropriate assumptions (see [RIE 88]). It follows that ! " T M (qH) if EΩ | sup0≤t≤1 V (t)|q < ∞ (4.95) Warped H-sssi: T V (q) = −∞ otherwise. Simple H-sssi process: when choosing the deterministic warp time M(t) = t we have T M (q) = q − 1 since S n M (q) = 2n · 2−nq for all n. Also, V = V . We obtain T M (qH) = qH − 1 which has to be inserted into (4.95) to obtain ! " qH − 1 if EΩ | sup0≤t≤1 V (t)|q < ∞ (4.96) Simple H-sssi: T V (q) = −∞ otherwise. 4.8.3. Local analysis of warped FBM Let us now turn to the special case where V is FBM. Then, we use the term FB(MF) to abbreviate fractional Brownian motion in multifractal time:
Multifractal Scaling: General Theory and Approach by Wavelets
171
B(t) = BH (M(t)). First, to obtain an idea of what to expect from the spectra of B, let us note that the moments appearing in (4.95) are finite for all q > −1 (see [RIE 88, lem 7.4] for a detailed discussion). Applying the Legendre transform easily yields ∗ (a/H). T ∗ B (a) = inf (qa − TM (qH)) = TM
(4.97)
q
(a)
(d)
50
1
40
0.8
30
0.6
20
0.4
10
0.2
0 0
0
0.2
0.4
0.6
0.8
1
Ŧ0.5
Ŧ0.4
Ŧ0.3
Ŧ0.2
Ŧ0.1
0
0.1
time lag
0.2
0.3
0.4
0.5
(b) (e)
1 0.8
0
0.6
Ŧ0.1
0.4
Ŧ0.2
0.2 0 0
0.2
0.4
0.6
0.8
1
Ŧ0.3 Ŧ0.4
(c)
Ŧ0.5
1.5
Ŧ0.6
1
Ŧ0.7
0.5
Ŧ0.8
0
Ŧ0.9
Ŧ0.5 0
0.2
0.4
0.6
time
0.8
1
Ŧ1 0
0.5
1
1.5
2
a
Figure 4.5. Left: simulation of Brownian motion in binomial time (a) sampling of Mb ((k + 1)2−n ) − Mb (k2−n ) (k = 0, . . . , 2n − 1), indicating distortion of dyadic time intervals, (b) Mb ((k2−n )): the time warp, (c) Brownian motion warped with (b): B(k2−n ) = B1/2 (Mb (k2−n )) Right: estimation of dim EaB using τ ∗ w,B , (d) empirical correlation of the Haar wavelet coefficients, (e) dot-dashed: T ∗ Mb (from theory), dashed: T ∗ B (a) = T ∗ Mb (a/H) Solid: the estimator τ ∗ w,B obtained from (c). (Reproduced from [GON 99])
Second, towards the local analysis we recall the uniform and strict Hölder continuity of the paths of FBM5 which reads roughly as sup |B(t + u) − B(t)| = sup |BH (M(t + u)) − BH (M(t))|
|u|≤δ
|u|≤δ
sup |M(t + u) − M(t)|H . |u|≤δ
5. For a precise statement see Adler [ADL 81] or [RIE 88, Theorem 7.4].
172
Scaling, Fractals and Wavelets
This is the key to concluding that BH simply squeezes the Hölder regularity exponents by a factor H. Thus, hB (t) = H · hM (t), etc. and M = KaB , Ka/H
and, consequently, analogous to (4.97), dB (a) = dM (a/H). Figure 4.5(d)-(e) displays an estimation of dB (a) using wavelets which agrees very closely with the form dM (a/H) predicted by theory (for statistics on this estimator see [GON 99, GON 98]). In conclusion: COROLLARY 4.7 (Fractional Brownian motion in multifractal time).– Let BH denote FBM of Hurst parameter H. Let M(t) be of almost surely continuous paths and independent of BH . Then, the multifractal warp formalism ∗ (a/H) dim(KaB ) = f B (a) = τ ∗ B (a) = T ∗ B (a) = TM
(4.98)
holds for B(t) = BH (M(t)) for any a such that the multifractal formalism holds for M ) = T ∗ M (a/H). M at a/H, i.e., for which dim(Ka/H This means that the local, or fine, multifractal structure of B captured in dim(KaB ) on the left can be estimated through grain based, simpler and numerically more robust spectra on the right side, such as τ ∗ B (a) (compare Figure 4.5 (e)). “Warp formula” (4.98) is appealing since it allows us to separate the LRD parameter of FBM and the multifractal spectrum of the time change M. Indeed, provided that M is almost surely increasing, we have T M (1) = 0 since S n (0) = M(1) for all n. Thus, T B (1/H) = 0 reveals the value of H. Alternatively, the tangent at T ∗ B through the origin has slope 1/H. Once H is known, T ∗ M follows easily from T ∗ B . Simple FBM: when choosing the deterministic warp time M(t) = t we have B = BH and T M (q) = q − 1 since S n M (q) = 2n · 2−nq for all n. We conclude that T BH (q) = qH − 1
(4.99)
for all q > −1. This confirms (4.90) for FGN. With (4.98) it shows that all spectra of FBM consist of the one point (H, 1) only, making the monofractal character of this process most explicit.
Multifractal Scaling: General Theory and Approach by Wavelets
173
4.8.4. LRD and estimation of warped FBM Let G(k) := B((k + 1)2−n ) − B(k2−n ) be FGN in multifractal time (see (4.93) for the case H = 1/2). Calculating auto-correlations explicitly shows that G is second order stationary under mild conditions with HG =
T M (2H) + 1 . 2
(4.100)
Let us discuss some special cases. For example, in a continuous, increasing warp time M, we have always T M (0) = −1 and T M (1) = 0. Exploiting the concave shape of T M we find that H < H G < 1/2 for 0 < H < 1/2, and 1/2 < H G < H for 1/2 < H < 1. Thus, multifractal warping cannot create LRD and it seems to weaken the dependence as measured through second order statistics. Especially in the case of H = 1/2 (“white noise in multifractal time”) G(k) becomes uncorrelated. This follows from (4.100). Notably, this is a different statement from the observation that the G(k) are independently conditioned on M (see section 4.8.1). As a particular consequence, wavelet coefficients will decorrelate fast for the entire process G, not only when conditioning on M (see Figure 4.5(d)). This is favorable for estimation purposes as it reduces the error variance. Of greater importance, however, is the warning that the vanishing correlations should not lead us to assume the independence of G(k). After all, G becomes Gaussian only lead us to assume that we know M. A strong, higher order dependence in G is hidden in the dependence of the increments of M which determine the variance of G(k) as in (4.93). Indeed, Figure 4.5(c) shows clear phases of monotony of B indicating positive dependence in its increments G, despite vanishing correlations. Mandelbrot calls this the “blind spot of spectral analysis”. 4.9. Bibliography [ABR 95] A BRY P., G ONÇALVES P., F LANDRIN P., “Wavelets, spectrum analysis and 1/f processes”, in A NTONIADIS A., O PPENHEIM G. (Eds.), Lecture Notes in Statistics: Wavelets and Statistics, vol. 103, p. 15–29, 1995. [ABR 00] A BRY P., F LANDRIN P., TAQQU M., V EITCH D., “Wavelets for the analysis, estimation and synthesis of scaling data”, Self-similar Network Traffic and Performance Evaluation, John Wiley & Sons, 2000. [ADL 81] A DLER R., The Geometry of Random Fields, John Wiley & Sons, New York, 1981. [ARB 96] A RBEITER M., PATZSCHKE N., “Self-similar random multifractals”, Math. Nachr., vol. 181, p. 5–42, 1996. [ARN 98] A RNEODO A., BACRY E., M UZY J., “Random cascades on wavelet dyadic trees”, Journal of Mathematical Physics, vol. 39, no. 8, p. 4142–4164, 1998.
174
Scaling, Fractals and Wavelets
[BAC 93] BACRY E., M UZY J., A RNEODO A., “Singularity spectrum of fractal signals from wavelet analysis: exact results”, J. Stat. Phys., vol. 70, p. 635–674, 1993. [BAC 03] BACRY E., M UZY J., “Log-infinitely divisible multifractal processes”, Comm. in Math. Phys., vol. 236, p. 449–475, 2003. [BAR 97] BARRAL J., Continuity, moments of negative order, and multifractal analysis of Mandelbrot’s multiplicative cascades, PhD thesis no. 4704, Paris-Sud University, 1997. [BAR 02] BARRAL J., M ANDELBROT B., “Multiplicative products of cylindrical pulses”, Probability Theory and Related Fields, vol. 124, p. 409–430, 2002. [BAR 03] BARRAL J., “Poissonian products of random weights: Uniform convergence and related measures”, Rev. Mat. Iberoamericano, vol. 19, p. 1–44, 2003. [BAR 04] BARRAL J., M ANDELBROT B., “Random multiplicative multifractal measures, Part II”, Proc. Symp. Pures Math., AMS, Providence, RI, vol. 72, no. 2, p. 17–52, 2004. [BEN 87] B EN NASR F., “Mandelbrot random measures associated with substitution”, C. R. Acad. Sc. Paris, vol. 304, no. 10, p. 255–258, 1987. [BRO 92] B ROWN G., M ICHON G., P EYRIERE J., “On the multifractal analysis of measures”, J. Stat. Phys., vol. 66, p. 775–790, 1992. [CAW 92] C AWLEY R., M AULDIN R.D., “Multifractal decompositions of Moran fractals”, Advances Math., vol. 92, p. 196–236, 1992. [CHA 02] C HAINAIS P., R IEDI R., A BRY P., “Compound Poisson cascades”, Proc. Colloque “Autosimilarité et Applications” Clermont-Ferrand, France, May 2002, 2002. [CHA 05] C HAINAIS P., R IEDI R., A BRY P., “On non-scale invariant infinitely divisible cascades”, IEEE Trans. Information Theory, vol. 51, no. 3, p. 1063–1083, 2005. [COX 84] C OX D., “Long-range dependence: a review”, Statistics: An Appraisal, p. 55–74, 1984. [CUT 86] C UTLER C., “The Hausdorff dimension distribution of finite measures in Euclidean space”, Can. J. Math., vol. 38, p. 1459–1484, 1986. [DAU 92] DAUBECHIES I., Ten Lectures on Wavelets, SIAM, New York, 1992. [ELL 84] E LLIS R., “Large deviations for a general class of random vectors”, Ann. Prob., vol. 12, p. 1–12, 1984. [EVE 95] E VERTSZ C.J.G., “Fractal geometry of financial time series”, Fractals, vol. 3, p. 609–616, 1995. [FAL 94] FALCONER K.J., “The multifractal spectrum of statistically self-similar measures”, J. Theor. Prob., vol. 7, p. 681–702, 1994. [FEL 98] F ELDMANN A., G ILBERT A.C., W ILLINGER W., “Data networks as cascades: Investigating the multifractal nature of Internet WAN traffic”, Proc. ACM/Sigcomm 98, vol. 28, p. 42–55, 1998. [FRI 85] F RISCH U., PARISI G., “Fully developed turbulence and intermittency”, Proc. Int. Summer School on Turbulence and Predictability in Geophysical Fluid Dynamics and Climate Dynamics, p. 84–88, 1985.
Multifractal Scaling: General Theory and Approach by Wavelets
175
[GON 98] G ONÇALVES P., R IEDI R., BARANIUK R., “Simple statistical analysis of wavelet-based multifractal spectrum estimation”, Proc. 32nd Asilomar Conf. on Signals, Systems and Computers, Pacific Grove, CA, Nov. 1998. [GON 99] G ONÇALVES P., R IEDI R., “Wavelet analysis of fractional Brownian motion in multifractal time”, Proceedings of the 17th Colloquium GRETSI, Vannes, France, September 1999. [GON 05] G ONÇALVES P., R IEDI R., “Diverging moments and parameter estimation”, J. Amer. Stat. Assoc., vol. 100, no. 472, p. 1382–1393, December 2005. [GRA 83] G RASSBERGER P., P ROCACCIA I., “Characterization of strange attractors”, Phys. Rev. Lett., vol. 50, p. 346–349, 1983. [HAL 86] H ALSEY T., J ENSEN M., K ADANOFF L., P ROCACCIA I., S HRAIMAN B., “Fractal measures and their singularities: the characterization of strange sets”, Phys. Rev. A, vol. 33, p. 1141–1151, 1986. [HEN 83] H ENTSCHEL H., P ROCACCIA I., “The infinite number of generalized dimensions of fractals and strange attractors”, Physica D, vol. 8, p. 435–444, 1983. [HOL 92] H OLLEY R., WAYMIRE E., “Multifractal dimensions and scaling exponents for strongly bounded random cascades”, Ann. Appl. Prob., vol. 2, p. 819–845, 1992. [JAF 91] JAFFARD S., “Pointwise smoothness, two-microlocalization coefficients”, Publicacions Mathematiques, vol. 35, p. 155–168, 1991.
and
wavelet
[JAF 95] JAFFARD S., “Local behavior of Riemann’s function”, Contemporary Mathematics, vol. 189, p. 287–307, 1995. [JAF 97] JAFFARD S., “Multifractal formalism for functions, Part 1: Results valid for all functions”, SIAM J. of M